text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'We develop a method for estimating well-conditioned and sparse covariance and inverse covariance matrices from a sample of vectors drawn from a sub-gaussian distribution in high dimensional setting. The proposed estimators are obtained by minimizing the quadratic loss function and joint penalty of $\ell_1$ norm and variance of its eigenvalues. In contrast to some of the existing methods of covariance and inverse covariance matrix estimation, where often the interest is to estimate a sparse matrix, the proposed method is flexible in estimating both a sparse and well-conditioned covariance matrix simultaneously. The proposed estimators are optimal in the sense that they achieve the minimax rate of estimation in operator norm for the underlying class of covariance and inverse covariance matrices. We give a very fast algorithm for computation of these covariance and inverse covariance matrices which is easily scalable to large scale data analysis problems. The simulation study for varying sample sizes and variables shows that the proposed estimators performs better than several other estimators for various choices of structured covariance and inverse covariance matrices. We also use our proposed estimator for tumor tissues classification using gene expression data and compare its performance with some other classification methods.'
author:
- |
Ashwini Maurya [email protected]\
Department of Statistics and Probability\
Michigan State University\
East Lansing, MI 48824, USA
bibliography:
- 'sample.bib'
title: 'A Well-Conditioned and Sparse Estimation of Covariance and Inverse Covariance Matrices Using a Joint Penalty'
---
Sparsity, Eigenvalue Penalty, Penalized Estimation
Introduction
============
With the recent surge in data technology and storage capacity, today’s statisticians often encounter data sets where sample size $n$ is small and number of variables $p$ is very large: often hundreds, thousands and even million or more. Examples include gene expression data and web search problems \[@Clark1:7, @pass:21\]. For many of the high dimensional data problems, the choice of classical statistical methods becomes inappropriate for making valid inference. The recent developments in asymptotic theory deal with increasing $p$ as long as both $p$ and $n$ tend to infinity at some rate depending upon the parameters of interest.
The estimation of covariance and inverse covariance matrix is a problem of primary interest in multivariate statistical analysis. Some of the applications include: **(i)** Principal component analysis (PCA) \[@Johnstone:14, @Zou:37\]:, where the goal is to project the data on “best" $k$-dimensional subspace, and where best means the projected data explains as much of the variation in original data without increasing $k$. **(ii)** Discriminant analysis \[@Mardia:19\]:, where the goal is to classify observations into different classes. Here estimates of covariance and inverse covariance matrices play an important role as the classifier is often a function of these entities. **(iii)** Regression analysis: If interest focuses on estimation of regression coefficients with correlated (or longitudinal) data, a sandwich estimator of the covariance matrix may be used to provide standard errors for the estimated coefficients that are robust in the sense that they remain consistent under mis-specification of the covariance structure. **(iv)** Gaussian graphical modeling \[@Mein:20, @Wainwright:27, @Yuan:34,@Yuan1:35\]:, the relationship structure among nodes can be inferred from inverse covariance matrix. A zero entry in the inverse covariance matrix implies conditional independence between the corresponding nodes.\
The estimation of large dimensional covariance matrix based on few sample observations is a difficult problem, especially when $n \asymp p$ (here $a_n \asymp b_n$ means that there exist positive constants $c$ and $C$ such that $c \le a_n/b_n \le C $). In these situations, the sample covariance matrix becomes unstable which explodes the estimation error. It is well known that the eigenvalues of sample covariance matrix are over-dispersed which means that the eigen-spectrum of sample covariance matrix is not a good estimator of its population counterpart \[@Marcenko:18, @Karoui1:16\]. To illustrate this point, consider $\Sigma_p=I_p$, so all the eigenvalues are $1$. A result from \[@Geman:12\] shows that if entries of $X_i$’s are i.i.d (let $X_i$’s have mean zero and variance 1) with a finite fourth moment and if $p/n \rightarrow \theta <1 $, then the largest sample eigenvalue $l_1$ satisfies: $$\begin{aligned}
l_1~ \rightarrow ~(1+\sqrt{\theta})^2, ~~~~~~ a.s\end{aligned}$$ This suggests that $l_1$ is not a consistent estimator of the largest eigenvalue $\sigma_1$ of population covariance matrix. In particular if $n=p$ then $l_1$ tends to $4$ whereas $\sigma_1$ is $1$. This is also evident in the eigenvalue plot in Figure 2.1. The distribution of $l_1$ also depends on the underlying structure of the true covariance matrix. From Figure 2.1, it is evident that the smaller sample eigenvalues tend to underestimate the true eigenvalues for large $p$ and small $n$. For more discussion on this topic, see @Karoui1:16.
To correct for this bias, a natural choice would be to shrink the sample eigenvalues towards some suitable constant to reduce the over-dispersion. For instance, @Stein:28 proposed an estimator of the form $\tilde{\Sigma}=\tilde{U} \Lambda (\tilde{\lambda}) \tilde{U}$, where $\Lambda (\tilde{\lambda})$ is a diagonal matrix with diagonal entries as transformed function of the sample eigenvalues and $\tilde{U}$ is the matrix of the eigenvectors. In another interesting paper @Ledoit:17 proposed an estimator that shrinks the sample covariance matrix towards the identity matrix. In another paper, @Karoui:15 proposed a non-parametric estimation of spectrum of eigenvalues and show that his estimator is consistent in the sense of weak convergence of distributions.
The covariance matrix estimates based on eigen-spectrum shrinkage are well-conditioned in the sense that their eigenvalues are well bounded away from zero. These estimates are based on the shrinkage of the eigenvalues and therefore invariant under some orthogonal group i.e. the shrinkage estimators shrink the eigenvalues but eigenvectors remain unchanged. In other words, the basis (eigenvector) in which the data are given is not taken advantage of and therefore the methods rely on premise that one will be able to find a good estimate in any basis. In particular, it is reasonable to believe that the basis generating the data is somewhat nice. Often this translates into the assumption that the covariance matrix has particular structure that one should be able to take advantage of. In these situations, it becomes natural to perform certain form of regularization directly on the entries of the sample covariance matrix.
Much of the recent literature focuses on two broad clases of regularized covariance matrix estimation. i) The one class relies on natural ordering among variables, where one often assumes that the variables far apart are weekly correlated and ii) the other class where there is no assumption on the natural ordering among variables. The first class includes the estimators based on banding and tapering \[@Bickel:3, @Cai1:6\]. These estimators are appropriate for a number of applications for ordered data (time series, spectroscopy, climate data). However for many applications including gene expression data, prior knowledge of any canonical ordering is not available and searching for all permutation of possible ordering would not be feasible. In these situations, an $\ell_1$ penalized estimator becomes more appropriate which yields a permutation-invariant estimate.
To obtain a suitable estimate which is both well-conditioned and sparse, we introduce two regularization terms: **i)** $\ell_1$ penalty for each of the off-diagonal elements of matrix and, **ii)** penalty propotional to the variance of the eigenvalues. The $\ell_1$ minimization problems are well studied in the covariance and inverse covariance matrix estimation literature \[@Freidman:11, @Banerjee:38, @Ravi1:24, @Bein:13, @Maurya:19 etc.\]. @Roth1:26 proposes an $\ell_1$ penalized log-likelihood estimator and shows that estimator is consistent in Frobenius norm at the rate of $O_P\Big(\sqrt{\{(p+s)~log~p\}/{n}}\Big)$, as both $p$ and $n$ approach to infinity. Here $s$ is the number of non-zero off-diagonal elements in the true covariance matrix. In another interesting paper @Bein:13 propose an estimator of covariance matrix as penalized maximum likelihood estimator with a weighted lasso type penalty. In these optimization problems, the $\ell_1$ penalty results in sparse and a permutation-invariant estimator as compared to other $l_q, q \neq 1$ penalties. Another advantage is that the $\ell_1$ norm is a convex function which makes it suitable for large scale optimization problems. A number of fast algorithms exist in the literature for covariance and inverse covariance matrix estimation \[(@Freidman:11, @Roth:25\]. The eigenvalues variance penalty overcomes the over-dispersion in the sample covariance matrix so that the estimator remains well-conditioned.
@Ledoit:17 proposed an estimator of covariance matrix as a linear combination of sample covariance and identity matrix. Their estimator of covariance matrix is well-conditioned but it is not sparse. @Roth:25 proposed estimator of covariance matrix based on quadratic loss function and $\ell_1$ penalty with a log-barrier on the determinant of covariance matrix. The log-determinant barrier is a valid technique to achieve positive definiteness but it is still unclear whether the iterative procedure proposed in @Roth:25 actually finds the right solution to the corresponding optimization problem. In another interesting paper, @Xue:31 proposed an estimator of covariance matrix as a minimizer of penalized quadratic loss function over set of positive definite matrices. In their paper, the authors solve a positive definite constrained optimization problem and establish the consistency of estimator. The resulting estimator is sparse and positive definite but whether it overcomes the over-dispersion of the eigen-spectrum of sample covariance matrix, is hard to justify. @Maurya:19 proposed a joint convex penalty as function of $\ell_1$ and trace norm (defined as sum of singular values of a matrix) for inverse covariance matrix estimation based on penalized likelihood approach.
In this paper, we propose the JPEN (Joint PENalty) estimators for covariance and inverse covariance matrices estimation and derive an explicit rate of convergence in both the operator and Frobenius norm. The JPEN estimators achieves minimax rate of convergence under operator norm for the underlying class of sparse covariance and inverse covariance matrices and hence is optimal. For more details see section $\S3$. One of the major advantage of the proposed estimators is that the proposed algorithm is very fast, efficient and easily scalable to a large scale data analysis problem.
The rest of the paper is organized as following. The next section highlights some background and problem set-up for covariance and inverse covariance matrix estimation. In section 3, we describe the proposed estimators and establish their theoretical consistency. In section 4, we give an algorithm and compare its computational time with some other existing algorithms. Section 5 highlights the performance of the proposed estimators on simulated data while an application of proposed estimator to real life data is given in section 6.
**Notation:** For a matrix $M$, let $\|M\|_1 $ denote its $\ell_1$ norm defined as the sum of absolute values of the entries of $M$, $\|M\|_F$ denote its Frobenius norm, defined as the sum of square of elements of $M$, $\|M\|$ denote its operator norm (also called spectral norm), defined as the largest absolute eigenvalue of $M$, $M^{-}$ denotes matrix $M$ where all diagonal elements are set to zero, $M^{+}$ denote matrix $M$ where all off-diagonal elements are set to zero, $\sigma_i(M)$ denote the $i^{th}$ largest eigenvalue of $M$, $tr(M)$ denotes its trace, $det(M)$ denote its determinant, $\sigma_{min}(M) $ and $\sigma_{max}(M)$ denote the minimum and maximum eigenvalues of $M$, $|M|$ be its cardinality, and let $\text{sign}(M)$ be matrix of s of elements of $M$. For any real $x$, let $\text{sign}(x) $ denotes of $x$, and let $|x|$ denotes its absolute value.
Background and Problem Set-up
=============================
Let $X=(X_1, X_2, \cdots, X_p) $ be a zero-mean p-dimensional random vector. The focus of this paper is the estimation of the covariance matrix $\Sigma:=\mathbb{E}(XX^T)$ and its inverse $\Sigma^{-1}$ from a sample of independently and identically distributed data $\{ X^{(k)} \}^{n}_{k=1}$. In this section we provide some background and problem setup more precisely.
The choice of loss function is very crucial in any optimization problem. An optimal estimator for a particular loss function may not be optimal for another choice of loss function. Recent literature in covariance matrix and inverse covariance matrix estimation mostly focuses on estimation based on likelihood function or quadratic loss function \[@Freidman:11, @Banerjee:38, @Bickel:3, @Ravi1:24, @Roth:25, @Maurya:19\]. The maximum likelihood estimation requires a tractable probability distribution of observations whereas quadratic loss function does not have any such requirement and therefore fully non-parametric. The quadratic loss function is convex and due to this analytical tractability, it is a widely applicable choice for many data analysis problems.
Proposed Estimators
-------------------
Let $S$ be the sample covariance matrix. Consider the following optimization problem. $$\hat{\Sigma}_{\lambda,\gamma}=\operatorname*{arg\,min}_{\Sigma=\Sigma^T,tr(\Sigma)=tr(S)}~~\Big[ ||\Sigma-S||^2_2 + \lambda \|{\Sigma^-}\|_1 + \gamma\sum_{i=1}^{p} \big \{\sigma_i(\Sigma)-\bar{\sigma}_{\Sigma} \big \}^2\Big],$$ where $\bar{\sigma}_\Sigma$ is the mean of eigenvalues of $\Sigma$, $\lambda$ and $\gamma$ are some positive constants. Note that by penalty function $\|{\Sigma^-}\|_1$, we only penalize off-diagonal elements of $\Sigma$. The eigenvalues variance penalty term for eigen-spectrum shrinkage is chosen from the following points of interest: i) It is easy to interpret and ii) this choice of penalty function yields a very fast optimization algorithm. By constraint $tr(\Sigma)=tr(S)$, the total variation in $\hat{\Sigma}_{\lambda,\gamma}$ is same as that in sample covariance matrix $S$, however the eigenvalues of $\hat{\Sigma}_{\lambda,\gamma} $ are well-conditioned than those of $S$. From here onwards we suppress the dependence of $\lambda, \gamma $ on $\hat{\Sigma }$ and denote $\hat{\Sigma}_{\lambda,\gamma} $ by $\hat{\Sigma}$.\
\
For $\gamma=0$, the solution to (2.1) is the standard soft-thresholding estimator for quadratic loss function and its solution is given by (see $\S4$ for derivation of this estimator): $$\begin{aligned}
\begin{split}
\hat{\Sigma}_{ii}& =s_{ii} \\
\hat{\Sigma}_{ij}& =\text{sign}(s_{ij})\max\Big (|s_{ij}|-\frac{\lambda}{2},0\Big), ~~~~~~~~~~~~~i \neq j.
\end{split}\end{aligned}$$ It is clear from this expression that a sufficiently large value of $\lambda$ will result in sparse covariance matrix estimate. But estimator $\hat{\Sigma}$ of (2.2) is not necesarily positive definite \[for more details here see @Xue:31\]. Moreover it is hard to say whether it overcomes the over-dispersion in the sample eigenvalues. The following eigenvalue plot (Figure (2.1)) illustrates this phenomenon for a neighbourhood type (see $\S5$ for details on description of neighborhood type of covariance matrix) covariance matrix. Here we simulated random vectors from multivariate normal distribution with sample size $n=50$ and number of covariates $~p=20$.
{width=".8\textwidth"}
As is evident from Figure 2.1, eigenvalues of sample covariance matrix are over-dispersed as most of them are either too large or close to zero. Eigenvalues of the proposed Joint Penalty (JPEN) estimator and PDSCE (Positive Definite Sparse Covariance matrix Estimator (@Roth1:26) of the covariance matrix are well aligned with those of true covariance matrix. See $\S 5$ for detailed discussion. Another drawback of the estimator (2.2) is that the estimate can be negative definite.\
As argued earlier, to overcome the over-dispersion in eigen-spectrum of sample covariance matrix, we include eigenvalues variance penalty. To illustrate its advantage, consider $\lambda=0$. After some algebra, let $\hat{\Sigma}$ be the minimizer of (2.1), then it is given by: $$\hat{\Sigma}= (S+\gamma~t~I)/(1+\gamma),$$ where $I$ is the identity matrix, and $t=\sum_{i=1}^{p}S_{ii}/p$. After some algebra, conclude that for any $\gamma>0$: $$\begin{aligned}
\sigma_{min} (\hat{\Sigma}) & = & \sigma_{min} (S+ \gamma~t~I)/(1+\gamma) \\
& \geq & \frac{\gamma~t}{1+\gamma}>0\end{aligned}$$ This means that the eigenvalues variance penalty improves $S$ to a positive definite estimator $\hat{\Sigma}$. However the estimator (2.3) is well-conditioned but need not be sparse. Sparsity can be achieved by imposing $\ell_1$ penalty on the entries of covariance matrix. Simulations have shown that, in general the minimizer of (2.1) is not positive definite for all values of $\lambda >0$ and $\gamma >0$. Here onwards we focus on correlation matrix estimation, and later generalize the method for covariance matrix estimation.\
To achieve both well-conditioned and sparse positive definite estimator we optimize the following objective function in $R$ over specific region of values of $(\lambda, \gamma)$ which depends upon sample correlation matrix $K$ and $\lambda,\gamma$. Here the condition $tr(\Sigma)=tr(S)$ reduces to $tr(R)=p$, and $t=1$. Consider the following optimization problem: $$\hat{R}_K=\operatorname*{arg\,min}_{R=R^T,tr(R)=p|(\lambda, \gamma) \in \hat{\mathscr{S}}^{K}_1}~~\Big[ ||R-K||^2_F + \lambda \|R^-\|_1 + \gamma\sum_{i=1}^{p} \big \{\sigma_i(R)-\bar{\sigma}_{R} \big \}^2\Big],$$ where
$$\begin{aligned}
\hat{\mathscr{S}}^{K}_1& = \Big \{(\lambda,\gamma): \lambda, \gamma >0, \lambda \asymp \gamma \asymp \sqrt{\frac{log~ p}{n}}, \forall \epsilon >0,\sigma_{min}\{(K+\gamma I)-\frac{\lambda}{2}*sign(K+\gamma I)\}>\epsilon \Big \},\end{aligned}$$
and $\bar{\sigma}_{R}$ is mean of the eigenvalues of $R$. For instance when $K$ is diagonal matrix, the set $\hat{\mathscr{S}}^{K}_1$ is given by:
$\hat{\mathscr{S}}^{K}_1 = \Big \{(\lambda,\gamma): \lambda, \gamma >0, \lambda \asymp \gamma \asymp \sqrt{\frac{log~ p}{n}}, \forall \epsilon >0,\lambda <2(\gamma-\epsilon) \Big \}$.
The minimization in (2.4) over $R$ is for fixed $ (\lambda, \gamma) \in \hat{\mathscr{S}}^{K}_1$. The proposed estimator of covariance matrix (based on regularized correlation matrix estimator $\hat{R}_K$) is given by $\hat{\Sigma}_K=({S^+})^{1/2}\hat{R}_K({S^+})^{1/2}$, where $S^+$ is the diagonal matrix of the diagonal elements of $S$. Furthermore Lemmas 3.1 and 3.2, respectively show that the objective function (2.4) is convex and estimator given in (2.4) is positive definite.
Our Contribution
----------------
The main contributions are the following:\
**i)** The proposed estimators are both sparse and well-conditioned simultaneously. This approach allows to take advantage of a prior structure if known on the eigenvalues of the true covariance and the inverse covariance matrices.\
**ii)** We establish theoretical consistency of proposed estimators in both operator and Frobenius norm. The proposed JPEN estimators achieves the minimax rate of convergence in operator norm for the underlying class of sparse and well-conditioned covariance and inverse covariance matrices and therefore is optimal.\
**iii)** The proposed algorithm is very fast, efficient and easily scalable to large scale optimization problems.
Analysis of JPEN Method
=======================
**Def:** A random vector $X$ is said to have sub-gaussian distribution if for each $t \ge 0$ and $y \in \mathbb{R}^p $ with $\|y\|_2=1$, there exist $0< \tau < \infty $ such that $$\mathbb{P}\{|y^T(X-\mathbb{E}(X))|>t\} \le e^{-t^2/2\tau}$$ Although the JPEN estimators exists for any finite $2 \le n<p<\infty$, for theoretical consistency in operator norm we require $s~log~p=o(n)$ and for Frobenus norm we require $(p+s) ~log~p=o(n)$ where $s$ is the upper bound on the number of non-zero off-diagonal entries in true covariance matrix. For more details, see the remark after Theorem 3.1.\
Covariance Matrix Estimation
----------------------------
We make the following assumptions about the true covariance matrix $\Sigma_0$.\
**A0.** Let $X:=(X_1,X_2,\cdots,X_p)$ be a mean zero vector with covariance matrix $\Sigma_0$ such that each $X_i/ \sqrt{\Sigma_{0ii}}$ has subgaussian distribution with parameter $\tau$ as defined in (3.1).\
**A1.** With $ E=\{(i,j): \Sigma_{0ij} \neq 0, i \neq j \}, $ the $|E| \le s $ for some positive integer $s$.\
**A2.** There exists a finite positive real number $\bar{k} >0$ such that $ 1/\bar{k} \le \sigma_{min}(\Sigma_0) \le \sigma_{max}(\Sigma_0) \le \bar{k}$.\
Assumption A2 guarantees that the true covariance matrix $\Sigma_0$ is well-conditioned (i.e. all the eigenvalues are finite and positive). A well-conditioned means that \[@Ledoit:17)\] inverting the matrix does not explode the estimation error. Assumption A1 is more of a definition which says that the number of non-zero off diagonal elements are bounded by some positive integer. Theorem 3.1 gives the rate of convergence of the proposed correlation based covariance matrix estimator (2.4). The following Lemmas show that optimization problem in (2.4) is convex and the proposed JPEN estimator (2.4) is positive definite.
The optimization problem in (2.4) is convex.
The estimator given by (2.4) is positive definite for any $2 \le n < \infty $ and $p <\infty$.
Let $(\lambda, \gamma) \in \hat{\mathscr{S}}^{K}_1$ and $\hat{\Sigma}_K$ be as defined in (2.4). Under Assumptions A0, A1, A2, $$\|\hat{R}_K-R_0\|_F=O_P \Big ( \sqrt{\frac{s~ log~p}{n}} \Big ) ~~~ \text{and} ~~~\|\hat{\Sigma}_K-\Sigma_0\|=O_P \Big ( \sqrt{\frac{(s+1) log~p}{n}} \Big ),$$ where $R_0$ is true correlation matrix.
**Remark: 1.** The JPEN estimator $\hat{\Sigma}_K$ is minimiax optimal under the operator norm. In (@Cai2:40), the authors obtain the minimax rate of convergence in the operator norm of their covariance matrix estimator for the particular construction of parameter space $\mathscr{H}_0(c_{n,p}):=\Big \{ \Sigma : max_{1 \le i \le p}\sum_{i=1}^{p}I\{\sigma_{ij}\neq 0\} \leq c_{n,p} \Big \}$. They show that this rate in operator norm is $c_{n,p} \sqrt{log~p/n}$ which is same as that of $\hat{\Sigma}_K$ for $1 \leq c_{n,p}=\sqrt{s}$.\
[**2.**]{} @Bickel1:4 proved that under the assumption of $\sum_{j=1}|\sigma_{ij}|^q \leq c_0(p)$ for some $ 0 \leq q \leq 1$, the hard thresholding estimator of the sample covariance matrix for tuning parameter $\lambda \asymp \sqrt{(log~p)/n}$ is consistent in operator norm at a rate no worse than $ O_P\Big ( c_0(p) \sqrt{p}(\frac{log ~p}{n})^{(1-q)/2} \Big ) $ where $c_0(p)$ is the upper bound on the number of non-zero elements in each row. Here the truly sparse case corresponds to $q=0$. The rate of convergence of $\hat{\Sigma}_K$ is same as that of @Bickel1:4 except in the following cases:\
[**Case (i)**]{} The covariance matrix has all off diagonal elements zero except last row which has $ \sqrt{p}$ non-zero elements. Then $c_0(p)=\sqrt{p}$ and $ \sqrt{s}=\sqrt{2~\sqrt{p}-1}$. The opeartor norm rate of convergence for JPEN estimator is $O_P \Big ( \sqrt{\sqrt{p}~(log~p)/n} \Big )$ where as rate of Bickel and Levina’s estimator is $O_P \Big (\sqrt{p~(log~p)/n} \Big )$.\
[**Case (ii)**]{} When the true covariance matrix is tridiagonal, we have $c_0(p)=2$ and $s=2p-2$, the JPEN estimator has rate of $\sqrt{p~log~p/n}$ whereas the Bickel and Levina’s estimator has rate of $\sqrt{log~p/n}$.\
For the case $\sqrt{s} \asymp c_0(p)$ and JPEN has the same rate of convergence as that of Bickel and Levina’s estimator.\
**3.** The operator norm rate of convergence is much faster than Frobenius norm. This is due to the fact that Frobenius norm convergence is in terms of all eigenvalues of the covariance matrix whereas the operator norm gives the convergence of the estimators in terms of the largest eigenvalue.\
**4.** Our proposed estimator is applicable to estimate any non-negative definite covariance matrix.\
Note that the estimator $\hat{\Sigma}_K$ is obtained by regularization of sample correlation matrix in (2.4). In some application it is desirable to directly regularize the sample covariance matrix. The JPEN estimator of the covariance matrix based on regularization of sample covariance matrix is obtained by solving the following optimization problem: $$\hat{\Sigma}_S=\operatorname*{arg\,min}_{\Sigma=\Sigma^T,tr(\Sigma)=tr(S)|(\lambda, \gamma) \in \hat{\mathscr{S}}^{S}_1}~~\Big[ ||\Sigma-S||^2_F + \lambda \|\Sigma^-\|_1 + \gamma\sum_{i=1}^{p} \{\sigma_i(\Sigma)-\bar{\sigma}_{\Sigma}\}^2\Big],$$ where $$\begin{aligned}
\hat{\mathscr{S}}^S_1& = \Big \{(\lambda,\gamma): \lambda,\gamma >0, \lambda \asymp \gamma \asymp \sqrt{\frac{log~ p}{n}}, \forall \epsilon >0, \sigma_{min}\{(S+\gamma t I)-\frac{\lambda}{2}*sign(S+\gamma t I)\}>\epsilon \},\end{aligned}$$ and $S$ is sample covariance matrix. The minimization in (3.3) over $\Sigma$ is for fixed $ (\lambda, \gamma) \in \hat{\mathscr{S}}^{S}_1$. The estimator $\hat{\Sigma}_S$ is positive definite and well-conditioned. Theorem 3.2 gives the rate of convergence of the estimator $\hat{\Sigma}_S$ in Frobenius norm.
Let $(\lambda, \gamma) \in \hat{\mathscr{S}}^S_1$, and let $\hat{\Sigma}_S$ be as defined in (3.3). Under Assumptions A0, A1, A2, $$\|\hat{\Sigma}_S-\Sigma_0\|_F=O_P \Big ( \sqrt{\frac{(s+p) log~p}{n}} \Big )$$
As noted in @Roth1:26 the worst part of convergence here comes from estimating the diagonal entries.
### Weighted JPEN Estimator for the Covariance Matrix Estimation
A modification of estimator $\hat{R}_{K}$ is obtained by adding positive weights to the term $(\sigma_i(R)-\bar{\sigma}_R)^2$. This leads to weighted eigenvalues variance penalty with larger weights amounting to greater shrinkage towards the center and vice versa. Note that the choice of the weights allows one to use any prior structure of the eigenvalues (if known) in estimating the covariance matrix. The weighted JPEN correlation matrix estimator $\hat{R}_A$ is given by : $$\hat{R}_A=\operatorname*{arg\,min}_{R=R^T,tr(R)=p|(\lambda, \gamma) \in \hat{\mathscr{S}}^{K,A}_1}~~\Big[ ||R-K||^2_F + \lambda \|R^-\|_1 + \gamma\sum_{i=1}^{p} a_i\{\sigma_i(R)-\bar{\sigma}_R\}^2\Big],$$ where $$\begin{aligned}
\hat{\mathscr{S}}^{K,A}_1& = \Big \{(\lambda,\gamma): \lambda \asymp \gamma \asymp \sqrt{\frac{log~p}{n}}, \lambda \le \frac{(2~\sigma_{min}(K))(1+\gamma ~max(A_{ii})^{-1})}{(1+\gamma~ min(A_{ii}))^{-1}p}+ \frac{\gamma ~min(A_{ii})}{p} \Big \},\end{aligned}$$ and $A=\text{diag}(A_{11},A_{22},\cdots A_{pp})$ with $A_{ii}=a_i$. The proposed covariance matrix estimator is $\hat{\Sigma}_{K,A}=(S^{+})^{1/2}\hat{R}_A (S^{+})^{1/2}$. The optimization problem in (3.5) is convex and yields a positive definite estimator for each $(\lambda,\gamma) \in \hat{\mathscr{S}}^{K,A}_1$. A simple excercise shows that the estimator $\hat{\Sigma}_{K,A}$ has same rate of convergence as that of $\hat{\Sigma}_{S}$.
Estimation of Inverse Covariance Matrix
---------------------------------------
We extend the JPEN approach to estimate a well-conditioned and sparse inverse covariance matrix. Similar to the covariance matrix estimation, we first propose an estimator for inverse covariance matrix based on regularized inverse correlation matrix and discuss its rate of convergence in Frobenious and operator norm.\
**Notation:** We shall use $Z$ and $\Omega$ for inverse correlation and inverse covariance matrix respectively.\
**Assumptions:** We make the following assumptions about the true inverse covariance matrix $\Omega_0$. Let $\Sigma_0=\Omega_0^{-1}$.\
**B0.** Same as the assumption $A0$.\
**B1.** With $ H=\{(i,j): \Omega_{0ij} \neq 0, i \neq j \}$, the $|H| \le s $, for some positive integer $s$.\
**B2.** There exist $ 0< \bar{k} < \infty $ large enough such that $ (1/{\bar{k}}) \le \sigma_{min}(\Omega_0) \le \sigma_{max}(\Omega_0) \le \bar{k}$.\
Let $\hat{R}_K$ be a JPEN estimator for the true correlation matrix. By Lemma 3.2, $\hat{R}_K$ is positive definite. Define the JPEN estimator of inverse correlation matrix as the solution to the following optimization problem, $$\hat{Z}_{K}=\operatorname*{arg\,min}_{Z=Z^T,tr(Z)=tr(\hat{R}_K^{-1})|(\lambda,\gamma) \in \hat{\mathscr{S}}^{{K}}_2}\Big [ \|Z-\hat{R}_K^{-1} \|^2~+~\lambda\|Z^-\|_1~+~\gamma\sum_{i=1}^{p} \{\sigma_i(Z)- \bar{\sigma}(Z)\}^2 \Big ]$$ where
$$\begin{aligned}
\hat{\mathscr{S}}^{{K}}_2& = \Big \{(\lambda,\gamma): \lambda,\gamma >0, \lambda \asymp \gamma \asymp \sqrt{\frac{log~ p}{n}}, \forall \epsilon >0, \\
& ~~~~~~~~~~~~ \sigma_{min}\{(\hat{R}_K^{-1}+\gamma t_1 I)-\frac{\lambda}{2}*sign(\hat{R}_K^{-1}+\gamma t_1 I)\}>\epsilon \Big \},\end{aligned}$$
and $t_1$ is average of the diagonal elements of $\hat{R}_K^{-1}$. The minimization in (3.6) over $Z$ is for fixed $ (\lambda, \gamma) \in \hat{\mathscr{S}}^{{K}}_2$. The proposed JPEN estimator of inverse covariance matrix (based on regularized inverse correlation matrix estimator $\hat{Z}_{K}$) is given by $\hat{\Omega}_{{K}}=(S^+)^{-1/2}{\hat{Z}}_{{K}}(S^+)^{-1/2}$, where $S^+$ is a diagonal matrix of the diagonal elements of $S$. Moreover (3.6) is a convex optimization problem and $\hat{Z}_K$ is positive definite.\
Next we state the consistency of estimators $\hat{Z}_{{K}}$ and $\hat{\Omega}_{{K}}$.
Under Assumptions B0, B1, B2 and for $(\lambda,\gamma) \in \hat{\mathscr{S}}^{{K}}_{2}$, $$\|\hat{Z}_{{K}}-R_0^{-1}\|_F=O_P \Big ( \sqrt{\frac{s~log~p}{n}} \Big ) ~~~\text{and} ~~~\|\hat{\Omega}_{{K}}-\Omega_0\|=O_P \Big ( \sqrt{\frac{(s+1)~log~p}{n}} \Big )$$ where $R_0^{-1}$ is the inverse of true correlation matrix.
**Remark:1.** Note that the JPEN estimator $\hat{\Omega}_{{K}}$ achieves minimax rate of convergence for the class of covariance matrices satisfying assumption $B0$, $B1$, and $B2$ and therefore optimal. The similar rate is obtained in @Cai2:40 for their class of sparse inverse covariance matrices.\
Next we give another estimate of inverse covariance matrix based on $\hat{\Sigma}_{S}$. Consider the following optimization problem: $$\hat{\Omega}_S=\operatorname*{arg\,min}_{\Omega=\Omega^T,tr(\Omega)=tr(\hat{\Sigma}_{S}^{-1})|(\lambda, \gamma) \in \hat{\mathscr{S}}^{{S}}_2}~~\Big[ ||\Omega- \hat{\Sigma}_{S}^{-1}||^2_F + \lambda \|\Omega^-\|_1 + \gamma\sum_{i=1}^{p} \{\sigma_i(\Omega)-\bar{\sigma}_{\Omega}\}^2\Big],$$ where
$$\begin{aligned}
\hat{\mathscr{S}}^{{S}}_2&= \Big \{(\lambda,\gamma): \lambda,\gamma >0, \lambda \asymp \gamma \asymp \sqrt{\frac{log~ p}{n}}, ~\forall \epsilon >0, \\
& ~~~~~~~~~~~~\sigma_{min}\{(\hat{\Sigma}_S^{-1}+\gamma ~t_2~I)-\frac{\lambda}{2}*sign(\hat{\Sigma}_S^{-1}+\gamma t_2 I)\}>\epsilon \Big \},
$$
and $t_2$ is average of the diagonal elements of $\hat{\Sigma}_S $. The minimization in (3.8) over $\Omega$ is for fixed $ (\lambda, \gamma) \in \hat{\mathscr{S}}^{S}_2$. The estimator in (3.8) is positive definite and well-conditioned. The consistency result of the estimator $\hat{\Omega}_S$ is given in following theorem.
Let $(\lambda, \gamma) \in \hat{\mathscr{S}}^{S}_2$ and let $\hat{\Omega}_S$ be as defined in (3.8). Under Assumptions B0, B1, and B2, $$\|\hat{\Omega}_S-\Omega_0\|_F=O_P \Big ( \sqrt{\frac{(s+p) log~p}{n}} \Big ).$$
### Weighted JPEN Estimator for The Inverse Covariance Matrix
Similar to weighted JPEN covariance matrix estimator $\hat{\Sigma}_{K,A}$, a weighted JPEN estimator of the inverse covariance matrix is obtained by adding positive weights $a_i $ to the term $(\sigma_i(Z)-1)^2$ in (3.8). The weighted JPEN estimator is $\hat{\Omega}_{{K},A}:=({S^{+}})^{-1/2}\hat{Z}_A ({S^{+}})^{-1/2}$, where $$\hat{Z}_A=\operatorname*{arg\,min}_{Z=Z^T,tr(Z)=tr(\hat{R}_K^{-1})|(\lambda, \gamma) \in \hat{\mathscr{S}}^{K,A}_2}~~\Big[ ||Z-\hat{R}_K^{-1}||^2_F + \lambda \|Z^-\|_1 + \gamma\sum_{i=1}^{p} a_i\{\sigma_i(Z)-1\}^2\Big],$$ with $$\begin{aligned}
\hat{\mathscr{S}}^{{K},A}_2& = \Big \{(\lambda,\gamma): \lambda \asymp \gamma \asymp \sqrt{\frac{log~p}{n}}, \lambda \le \frac{(2~\sigma_{min}({R}_K^{-1}))(1+\gamma t_1 max(A_{ii})^{-1})}{(1+\gamma~ min(A_{ii})^{-1}p}+ \frac{\gamma min(A_{ii})}{p} \Big \},\end{aligned}$$ and $A=\text{diag}(A_{11},A_{22},\cdots A_{pp})$ with $A_{ii}=a_i$. The optimization problem in (3.10) is convex and yields a positive definite estimator for $(\lambda,\gamma) \in \hat{\mathscr{S}}^{K,A}_2$. A simple excercise shows that the estimator $\hat{Z}_{A}$ has similar rate of convergence as that of $\hat{Z}_K$.
An Algorithm
============
Covariance Matrix Estimation:
-----------------------------
The optimization problem (2.4) can be written as:\
$$\begin{aligned}
\hat{R}_K=\operatorname*{arg\,min}_{R=R^T|(\lambda, \gamma) \in \hat{\mathscr{S}}^{K}_1 }~f(R),\end{aligned}$$ where $$\begin{aligned}
f(R)=||R-K||^2_F + \lambda \|{R}^-\|_1 + \gamma\sum_{i=1}^{p} \{\sigma_i(R)-\bar{\sigma}(R)\}^2.\end{aligned}$$ Note that $\sum_{i=1}^{p} \{\sigma_i(R)-\bar{\sigma}(R)\}^2=tr(R^2)-2~tr(R)+p$, where we have used the constraint $tr(R)=p$. Therefore, $$\begin{aligned}
f(R) & = & \|R-K\|_F^2+\lambda \|{R}^-\|_1 +\gamma~ tr(R^2)-2~ \gamma~tr(R) +p\\
& = & tr(R^2 (1+\gamma))-2tr\{R(K+\gamma I)\}+ tr(K^TK) +\lambda ~\|{R}^-\|_1 +p\\
& = & (1+\gamma)\{tr(R^2)-2/(1+\gamma) tr\{R(K+\gamma I)\}+ (1/(1+\gamma))tr(K^TK)\} \\
& &~~~~+\lambda ~\|{R}^-\|_1 +p \\
& = & (1+\gamma)\{\|R- (K+\gamma I)/(1+\gamma)\|^2_F+ (1/(1+\gamma))tr(K^TK)\} \\
& & ~~~~+\lambda ~\|{R}^-\|_1 +p.\end{aligned}$$ The solution of (4.1) is soft thresholding estimator and it is given by: $$\begin{aligned}
\hat{R}_K =\frac{1}{1+\gamma}~\text{sign}(K)*\text{pmax}\{\text{abs}(K+\gamma ~I)-\frac{\lambda}{2},0\}\end{aligned}$$ with $(\hat{R}_{K})_{ii}=(K_{ii}+\gamma)/(1+\gamma)$, $pmax(A,b)_{ij}:=max(A_{ij},b)$ is elementwise max function for each entry of the matrix $A$. Note that for each $(\lambda,\gamma) \in \hat{\mathscr{S}}^{K}_1$, $\hat{R}_{K}$ is positive definite.\
\
[**Choice of $\lambda$ and $\gamma$:**]{} For a given value of $\gamma$, we can find the value of $\lambda $ satisfying: $$\begin{aligned}
\sigma_{min}\{(K+\gamma I)-\frac{\lambda}{2}*sign(K+\gamma I)\}>0\end{aligned}$$ which can be simplified to $$\begin{aligned}
\lambda < \frac{\sigma_{min}(K+\gamma I)}{C_{12}~\sigma_{max}(\text{sign}(K))}. \end{aligned}$$ For some $C_{12} \ge 0.5$. Such choice of $(\lambda,\gamma)\in \hat{\mathscr{S}}_1^{{K}}$, and the estimator $\hat{R}_{K}$ is positive definite. Smaller values of $C_{12}$ yeild a solution which is more sparse but may not be positive definite.\
\
[**Choice of weight matrix A:**]{} For optimization problem in (3.5), the weights are chosen in following way:\
Let $\mathscr{E}$ be the set of sorted diagonal elements of the sample covariance matrix $S$.\
**i)** Let $k$ be largest index of $\mathscr{E}$ such that $k^{th}$ elements of $\mathscr{E}$ is less than $1$. For $i \le k, ~a_i=\mathscr{E}_{i}$. For $k<i \le p,~ a_i=1/\mathscr{E}_{i}.$\
**ii)** $A=\text{diag}(a_1,a_2,\cdots,a_p),~\text{where}~ a_j=a_j/\sum_{i=1}^p a_i.$ Such choice of weights allows more shrinkage of extreme sample eigenvalues than the ones in center of eigen-spectrum.
Inverse Covariance Matrix Estimation:
-------------------------------------
To get an expression of inverse covariance matrix estimate, we replace $K$ by $\hat{R}_K^{-1}$ in (4.2), where $\hat{R}_K$ is a JPEN estimator of correlation matrix. We chose $(\lambda,\gamma) \in \hat{\mathscr{S}}^{{K}}_2$. For a given $\gamma$, we chose $\lambda>0$ satisfying:\
$$\begin{aligned}
\sigma_{min}\{(\hat{R}_K^{-1}+\gamma t_1 I)-\frac{\lambda}{2}*sign(\hat{R}_K^{-1}+\gamma t_1 I)\}>0\end{aligned}$$ which can be simplified to $$\begin{aligned}
\lambda < \frac{\sigma_{min}(\hat{R}_K^{-1}+\gamma t_1 I)}{C_{12}~\sigma_{max}(\text{sign}(\hat{R}_K^{-1}))}. \end{aligned}$$
Computational Complexity
------------------------
The JPEN estimator $\hat{\Sigma}_{K}$ has computational complexity of $O(p^2)$ as there are at most $3p^2$ multiplications for computing the estimator $\hat{\Sigma}_{K}$. The other existing algorithm Glasso ([@Freidman:11]), PDSCE ([@Roth1:26]) have computational complexity of $O(p^3)$. We compare the computational timing of our algorithm to some other existing algorithms Glasso (@Freidman:11), PDSCE (@Roth1:26). The exact timing of these algorithm also depends upon the implementation, platform etc. (we did our computations in $R$ on a AMD 2.8GHz processor). Following the approach [@Bickel1:4], the optimal tuning parameter $(\lambda,\gamma)$ was obtained by minimizing the $5-$fold cross validation error $$(1/5) \sum_{i=1}^5\|\hat{\Sigma}_i^v-\Sigma_i^{-v}\|_1,$$ where $\hat{\Sigma}_i^v$ is JPEN estimate of the covariance matrix
[H]{}[0.5]{}
{width="40.00000%"}
based on $v=4n/5$ observations, $\Sigma_i^{-v}$ is the sample covariance matrix using $(n/5)$ observations. Figure 4.1 illustrates the total computational time taken to estimate the covariance matrix by $Glasso,~PDSCE$ and $JPEN$ algorithms for different values of $p$ for Toeplitz type of covariance matrix on log-log scale (see section $\S 5$ for Toeplitz type of covariance matrix). Although the proposed method requires optimization over a grid of values of $(\lambda,\gamma) \in \hat{\mathscr{S}}^{K}_1$, our algorithm is very fast and easily scalable to large scale data analysis problems.
Simulation Results
==================
#### •
We compare the performance of the proposed method to other existing methods on simulated data for five types of covariance and inverse covariance matrices.\
**(i) Hub Graph:** Here the rows/columns of $\Sigma_0$ are partitioned into J equally-sized disjoint groups: $ \{ V_1 \cup V_2~ \cup, ..., \cup~ V_J\} = \{1,2,...,p\},$ each group is associated with a $\bf pivotal$ row k. Let size $|V_1| = s$. We set $\sigma_{0i,j}=\sigma_{0j,i}=\rho $ for $ i \in V_k $ and $\sigma_{0i,j}=\sigma_{0j,i}=0$ otherwise. In our experiment, $J=[p/s], k =1,s+1, 2s+1,...,$ and we always take $\rho= 1/(s + 1)$ with J = 20.\
**(ii) **Neighborhood Graph:**** We first uniformly sample $(y_1,y_2,...,y_n)$ from a unit square. We then set $\sigma_{0i,j}=\sigma_{0j,i}=\rho $ with probability ${(\sqrt{2\pi})}^{-1}exp( -4\|y_i-y_j\|^2)$. The remaining entries of $\Sigma_0$ are set to be zero. The number of nonzero off-diagonal elements of each row or column is restricted to be smaller than $[1/\rho]$ where $\rho$ is set to be 0.245.\
**(iii) **Toeplitz Matrix:**** We set $\sigma_{0i,j}=2~$for$ ~i=j;~\sigma_{0i,j}=|0.75|^{|i-j|}~$ for $|i-j|=1,2;$ and $~\sigma_{0i,j}=0~$ otherwise.\
**(iv) **Block Diagonal Matrix:**** In this setting $\Sigma_0$ is a block diagonal matrix with varying block size. For $p=500$ number of blocks is 4 and for $p=1000$ the number of blocks is 6. Each block of covariance matrix is taken to be Toeplitz type matrix as in case (iii).\
**(v) Cov-I type Matrix:** In this setting, we first simulate a random sample $(y_1,y_2,...,y_p)$ from standard normal distribution. Let $x_i=|y_i|^{3/2}*(1+1/p^{1+log(1+1/p^2)})$. Next we generate multivariate normal random vectors $\mathscr{Z}=(z_1,z_2,...,z_{5p})$ with mean vector zero and identity covariance matrix. Let $U$ be eigenvector corresponding to sample covariance matrix of $\mathscr{Z}$. We take $\Sigma_0=UDU'$, where $D=\text{diag}(x_1,x_2,....x_p)$. This is not a sparse setting but the covariance matrix has most of eigenvalues close to zero and hence allows us to compare the performance of various methods in a setting where most of eigenvalues are close to zero and widely spread as compared to structured covariance matrices in **(i)-(iv)**.\
[|r|p[1.6cm]{}|p[1.6cm]{}|p[1.6cm]{}|p[1.6cm]{}|]{}\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Ledoit-Wolf & 1.54(0.102) & 2.96(0.0903) & 4.271(0.0394) & 2.18(0.11)\
Glasso & 0.322(0.0235) & 3.618(0.073) & 0.227(0.098) & 2.601(0.028)\
PDSCE & 3.622(0.231) & 4.968(0.017) & 1.806(0.21) & 2.15(0.01)\
BLThresh & 2.747(0.093) & 3.131(0.122) & 0.887(0.04) & 0.95(0.03)\
JPEN & 2.378(0.138) & 3.203(0.144) & 1.124(0.088) & 2.879(0.011)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Ledoit-Wolf & 2.13(0.103) & 2.43(0.043) & 1.07(0.165) & 3.47(0.0477)\
Glasso & 0.511(0.047) & 0.551(0.005) & 0.325(0.053) & 0.419(0.003)\
PDSCE & 0.735(0.106) & 0.686(0.006) & 0.36(0.035) & 0.448(0.002)\
BLThresh & 1.782(0.047) & 2.389(0.036) & 0.875(0.102)& 1.82(0.027)\
JPEN & 0.732(0.111) & 0.688(0.006) & 0.356(0.058) & 0.38(0.007)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Ledoit-Wolf & 1.36(0.054) & 2.89(0.028) & 1.1(0.0331) & 2.32(0.0262)\
Glasso & 0.608(0.054) & 0.63(0.005) & 0.428(0.047) & 0.419(0.038)\
PDSCE & 0.373(0.085) & 0.468(0.007) & 0.11(0.056) & 0.175(0.005)\
BLThresh & 1.526(0.074) & 2.902(0.033) & 0.870(0.028)& 1.7(0.026)\
JPEN & 0.454(0.0423) & 0.501(0.018) & 0.086(0.045) & 0.169(0.003)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Ledoit-Wolf & 1.526(0.074) & 2.902(0.033) & 1.967(0.041) & 2.344(0.028)\
Glasso & 2.351(0.156) & 3.58(0.079) & 1.78(0.087) & 2.626(0.019)\
PDSCE & 3.108(0.449) & 5.027(0.016) & 0.795(0.076) & 2.019(0.01)\
BLThresh & 0.858(0.040) & 1.206(0.059) & 0.703(0.039)& 1.293(0.018)\
JPEN & 2.517(0.214) & 3.205(0.16) & 1.182(0.084) & 2.919(0.011)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Ledoit-Wolf & 33.2(0.04) & 36.7(0.03) & 36.2(0.03) & 48.0(0.03)\
Glasso & 15.4(0.25) & 16.1(0.4) & 14.0(0.03) & 14.9(0.02)\
PDSCE & 16.5(0.05) & 16.33(0.04) & 16.9(0.03) & 17.5(0.02)\
BLThresh & 15.7(0.04) & 17.1(0.03) & 13.4(0.02) & 17.5(0.02)\
JPEN & 7.1(0.042) & 11.5(0.07) & 8.4(0.042) & 7.8(0.034)\
\[tab:addlabel\]
We chose similar structure of $\Omega_0$ for simulations. For all these choices of covariance and inverse covariance matrices, we generate random vectors from multivariate normal distribution with varying $n$ and $p$. We chose $n=50,100$ and $p=500,1000$. We compare the performance of proposed covariance matrix estimator $\hat{\Sigma}_K$ to graphical lasso \[@Freidman:11\], PDSC Estimate \[@Roth1:26\], Bickel and Levina’s thresholding estimator (BLThresh) \[[@Bickel1:4]\] and Ledoit-Wolf \[[@Ledoit:17]\] estimate of covariance matrix. The JPEN estimate $\hat{\Sigma}_{K}$ was computed using R software(version 3.0.2). The graphical lasso estimate of the covariance matrix was computed using R package “glasso" (http://statweb.stanford.edu/ tibs/glasso/).
[|r|p[1.6cm]{}|p[1.6cm]{}|p[1.6cm]{}|p[1.6cm]{}|]{}\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Glasoo & 4.144(0.523) & 1.202(0.042) & 0.168(0.136) & 1.524(0.028)\
PDSCE & 1.355(0.497) & 1.201(0.044) & 0.516(0.196) & 0.558(0.032)\
CLIME & 4.24(0.23) & 6.56(0.25) & 6.88(0.802) & 10.64(0.822)\
JPEN & 1.248(0.33) & 1.106(0.029) & 0.562(0.183) & 0.607(0.03)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Glasoo & 1.122(0.082) & 0.805(0.007) & 0.07(0.038) & 0.285(0.004)\
PDSCE & 0.717(0.108) & 0.702(0.007) & 0.358(0.046) & 0.356(0.005)\
CLIME & 10.5(0.329) & 10.6(0.219) & 6.98(0.237) & 10.8(0.243)\
JPEN & 0.684(0.051) & 0.669(0.003) & 0.34(0.024) & 0.337(0.002)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Glasoo & 1.597(0.109) & 0.879(0.013) & 1.29(0.847) & 0.428(0.007)\
PDSCE & 0.587(0.13) & 0.736(0.014) & 0.094(0.058) & 0.288(0.01)\
CLIME & 10.5(0.535) & 11.5(0.233) & 10.5(0.563) & 11.5(0.245)\
JPEN & 0.551(0.075) & 0.691(0.008) & 0.066(0.042) & 0.201(0.007)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Glasoo & 2.862(0.475) & 2.89(0.048) & 2.028(0.267) & 2.073(0.078)\
PDSCE & 1.223(0.5) & 1.238(0.065) & 0.49(0.269) & 0.473(0.061)\
CLIME & 4.91(0.22) & 7.597(0.34) & 5.27(1.14) & 8.154(1.168)\
JPEN & 1.151(0.333) & 2.718(0.032) & 0.607(0.196) & 2.569(0.057)\
\
& &\
& p=500 & p=1000 & p=500 & p=1000\
Glasoo & 54.0(0.19) & 190.(5.91) & 14.7(0.37) & 49.9(0.08)\
PDSCE & 28.8(0.19) & 45.8(0.32) & 16.9(0.04) & 26.9(0.08)\
CLIME & 59.8(0.82) & 207.5(3.44) & 15.4(0.03) & 53.7(0.69)\
JPEN & 26.3(0.36) & 7.0(0.07) & 15.7(0.08) & 23.5(0.3)\
\[tab:addlabel\]
The Ledoit-Wolf estimate was obtained using code from (http://econ.uzh.ch/faculty/wolf/publications.html\#9). The PDSC estimate was obtained using PDSCE package (http://cran. r-project. org/web/ packages/PDSCE/index.html). The Bickel and Levina’s estimator was computed as per the algorithm given in their paper. For inverse covariance matrix performance comparison we include glasso, CLIME (@Cai1:6) and PDSCE. For each of covariance and inverse covariance matrix estimate, we calculate Average Relative Error (ARE) based on 50 iterations using following formula, $$\begin{aligned}
ARE (\Sigma,\hat{\Sigma})= |log(f(S,\hat{\Sigma}))~-~log(f(S,\Sigma_0))|/|(log(f(S,\Sigma_0))|,\end{aligned}$$
{width="85.00000%"}
where $ f(S,\cdot) $ is multivariate normal density given the sample covariance matrix $S$, $\Sigma_0$ is the true covariance, $\hat{\Sigma}$ is the estimate of $\Sigma_0$ based on one of the methods under consideration. Other choices of performance criteria are Kullback-Leibler used by @Yuan:34 and @Bickel1:4. The optimal values of tuning parameters were obtained over a grid of values by minimizing $5-$fold cross-validation as explained in $\S4$. The average relative error and their standard deviations (in percentage) for covariance and inverse covariance matrix estimates are given in Table 5.1 and Table 5.2, respectively. The numbers in the bracket are the standard errors of relative error based on the estimates using different methods. Among all the methods JPEN and PDSCE perform similar for most of choices of $n$ and $p$ for all five type of covariance matrices. This is due to the fact that both PDSCE and JPEN use quadratic optimization function with a different penalty function. The behavior of Bickel and levina’s estimator is quite good in Toepltiz case where it performs better than the other methods. For this type of covariance matrix, the entries away from the diagonal decay to zero and therefore soft-thresholding estimators like BLThresh perform better in this setting. However for neighorhood and hub type covariance matrix which are not necessarily banded type, Bickel and Levina estimator is not a natural choise as their estimator would fail to recover the underlying sparsity pattern. The performance of Ledoit-Wolf estimator is not very encouraging for Cov-I type matrix. The Ledoit-Wolf estimator shrinks the sample covariance matrix towards identity and hence the eigenvalues estimates are highly shrunk towards one. This is also visible in eigenvalues plot in Figure 5.2 and Figure 5.3. For Cov-I type covariance matrix where most of eigenvalues are close to zero and widely spread, the performance of JPEN estimator is impressive. The eigenplot in Figure 5.3 shows that among all the methods, estimates of eigenvalues of JPEN estimator are most consistent with true eigenvalues. This clearly shows the advantage of JPEN estimator of covariance matrix when the true eigenvalues are dispersed or close to zero. The eigenvalues plot in Figure 5.2 shows that when eigen-spectrum of true covariance matrix are not highly dispersed, the JPEN and PDSCE estimates of eigenavlues are almost the same. This phenomenon is also apparent in Figure 2.1. Also Ledoit-Wolf estimator heavily shrinks the eigenvalues towards the center and thus underestimates the true eigen-spectrum.
{width=".8\textwidth"}
{width=".8\textwidth"}
For inverse covariance matrix, we compare glasso, CLIME and PDSCE estimates with proposed JPEN estimator. The JPEN estimator $\hat{\Omega}_{K}$ outperforms other methods for the most of the choices of $n$ and $p$ for all five types of inverse covariance matrices. Additional simulations (not included here) show that for $n \approx p$, all the underlying methods perform similarly and the estimates of their eigenvalues are also well aligned with true values. However in high dimensional setting, for large $p$ and small $n$, their performance is different as seen in simulations of Table 5.1 and Table 5.2. Figure 5.1 shows the recovery of non-zero and zero entries of true covariance matrix based on JPEN estimator $\hat{\Sigma}_{K}$ based on 50 realizations. The estimtor recovers the true zeros for about 90% of times for Hub and Neighborhood type of covariance matrix. It also reflect the recovery of true structure of non-zero entries and actual pattern among the rows/columns of covariance matrix. To see the implication of eigenvalues shrinkage penalty as compared to other methods, we plot (Figure 5.2) the eigenvalues of estimated covariance matrix for $n=100$, $p=50$ for neighborhood type of covariance matrix. The JPEN estimates of eigen-spectrum are well aligned with true ones and closest being PDSC estimates of eigenvalues. Figure 5.3 shows the recovery of eigenvalues based on estimates using different methods for Cov-I type covariance matrix. For this particular simulation, the eigenvalues are choosen differently than the one described in (v) of $\S5$. The eigenvalues of true covariance matrix are taken to be very diverse with maximum about $10^6$ and smallest eigenvalue about $10^{-6}$. For Cov-I type of matrix, JPEN estimates of eigenvalues are better than other methods.
Colon Tumor Classification Example
==================================
In this section, we compare performance of JPEN estimator of inverse covariance matrix for tumors classification using Linear Discriminant Analysis (LDA). The gene expression data (@Alan:2 consists of 40 tumorous and 22 non-tumorous adenocarcinoma tissue. After preprocessing, data was reduced to a subset of 2,000 gene expression values with the largest minimal intensity over the 62 tissue samples (source: *http://genomics-pubs.princeton.edu/oncology /affydata/index.html*). In our analysis, we reduced the number of genes by selecting $p$ most ificant genes based on logistic regression. We obtain estimates of inverse covariance matrix for $p=50,100,200$ and then use LDA to classify these tissues as either tumorous or non-tumorous (normal). We classify each test observation x to either class k = 0 or k = 1 using the LDA rule $$\begin{aligned}
\delta_k(x)=\operatorname*{arg\,max}_k \Big \{ x^T\hat{\Omega}\hat{\mu_k}~-~\frac{1}{2} \hat{\mu_k}\hat{\Omega}\hat{\mu_k} ~+~log(\hat{\pi}_k) \Big\},\end{aligned}$$ where $\hat{\pi}_k$ is the proportion of class $k$ observations in the training data, $\hat{\mu}_k$ is the sample mean for class k on the training data, and $\hat{\Omega}:=\hat{\Sigma}^{-1}$ is an estimator of the inverse of the common covariance matrix on the training data computed. Tuning parameters $\lambda$ and $\gamma$ were chosen using 5-fold cross validation. To create training and test sets, we randomly split the data into a training and test set of sizes 42 and 20 respectively; following the approach used by Wang et al. (2007), the training set has 27 tumor samples and 15 non-tumor samples. We repeat the split at random 100 times and measure the average classification error.
[l | [l]{} [l]{} [c]{}r]{} Method & p=50 & p=100 & p=200\
Logistic Regression & 21.0(0.84) & 19.31(0.89) & 21.5(0.85)\
SVM & 16.70(0.85) & 16.76(0.97) & 18.18(0.96)\
Naive Bayes & 13.3(0.75) & 14.33(0.85) & 14.63(0.75)\
Graphical Lasso & 10.9(1.3) & 9.4(0.89) & 9.8(0.90)\
Joint Penalty & **9.9(0.98)** & **8.9(0.93)** & **8.2(0.81)**
Since we do not have separate validation set, we do the 5-fold cross validation on training data. At each split, we divide the training data into 5 subsets (fold) where 4 subsets are used to estimate the covariance matrix and one subset is used to measure the classifier’s performance. For each split, this procedure is repeated 5 times by taking one of the 5 subsets as validation data. An optimal combination of $\lambda $ and $\gamma$ is obtained by minimizing the $5$-fold cross validation error.\
The average classification errors with standard errors over the 100 splits are presented in Table 6.1. Since the sample size is less than the number of genes, we omit the inverse sample covariance matrix as it is not well defined and instead include the naive Bayes’ and support vector machine classifiers. Naive Bayes has been shown to perform better than the sample covariance matrix in high-dimensional settings (@Bickel2:39. Support Vector Machine (SVM) is another popular choice for high dimensional classification tool. Among all the methods covariance matrix based LDA classifiers perform far better than Naive Bayes, SVM and Logistic Regression. For all other classifiers the classification performance deteriorates for increasing $p$. For larger $p$, i.e., when more genes are added to the data set, the classification performance of JPEN estimate based LDA classifier initially improves but it deteriorates for large $p$. For $p=2000$, the classifier based on inverse covariance matrix has accuracy of $30\%$. This is due to the fact that as dimension of covariance matrix increases, the estimator does not remain very informative.
Summary
=======
We have proposed and analyzed regularized estimation of large covariance and inverse covariance matrix using joint penalty. The proposed JPEN estimators are optimal under spectral norm for underlying classs of sparse and well-conditioned covariance and inverse covariance matrices. We also establish its theoretical consistency in Frobenius norm. One of its biggest advantage is that the optimization carries no computational burden and and the resulting algorithm is very fast and easily scalable to large scale data analysis problems. The extensive simulation shows that the proposed estimators performs well for a number of structured covariance and inverse covariance matrices. Also when the eigenvalues of underlying true covariance matrix are highly dispersed, it outperforms other methods (based on simulation analysis). The JPEN estimator recovers the sparsity pattern of the true covariance matrix and provides a good approximation of the underlying eigen-spectrum and hence we expect that PCA will be one of the most important application of the method. Although the proposed JPEN estimators of covariance and inverse covariance matrix do not require any assumption on the structure of true covariance and inverse covariance matrices respectively, any prior knowledge of structure of true covariance matrix might be helpful to choose a suitable weight matrix and hence improve estimation.\
0.2in
Appendix A. {#appendix-a. .unnumbered}
===========
[**Proof of Lemma 3.1**]{}\
Let $$f(R)=\|R-K\|^2+\lambda \|R^-\|_1+\gamma \sum_{i=1}^{p} \{\sigma_i(R)-\bar{\sigma}_R\}^2.$$ where $\bar{\sigma}_R$ is the mean of eigenvalues of $R$. Due to the constraint $tr(R)=p$, we have $\bar{\sigma}_R=1$. The third term of (.1) can be written as\
$$\sum_{i=1}^{p} \{\sigma_i(R)-\bar{\sigma}_R\}^2=tr(R^2)-2~tr(R)+p$$ We obtain, $$\begin{aligned}
\begin{split}
f(R)&= tr(R^2)-2~tr(RK)+tr(K^2)+\lambda\|R^-\|_1 + \gamma\{tr(R^2) -2~tr(R)+p\} \\
&= tr(R^2(1+\gamma))-2~tr(K+\gamma~I)+tr(K^2)+\lambda\|R^-\|_1 + p \\
&= (1+\gamma)\|R-(K+\gamma~I)/(1+\gamma)\|^2+tr(K^2)+\lambda \|R^-\|_1 + p
\end{split}\end{aligned}$$ This is quadratic in $R$ with a $\ell_1$ penalty to the off-diagonal entries of $R$, therefore a convex function in $R$.\
\
[**Proof of Lemma 3.2**]{} The solution to (.2) satisfies: $$\begin{aligned}
2(R-(K+\gamma I))(1+\gamma)^{-1}+ \lambda \frac{\partial{\|R^-\|_1}}{\partial{R}}=0\end{aligned}$$ where $\frac{\partial{\|R^-\|_1}}{\partial{R}}$ is given by: $$\frac{\partial{\|R^-\|_1}}{\partial{R}} = \left\{ \begin{array}{lr} 1 & : if~~~R_{ij}>0 \\ -1 &: if ~~~ R_{ij} <0 \\ \tau \in (-1,1)& : if~~~ R_{ij}=0 \end{array} \right.$$ Note that $\|R^-\|_1$ has same value irrespective of of $R$, therefore the right hand side of (.2) is minimum if :\
$$\begin{aligned}
\text{sign}(R)=\text{sign}(K+\gamma I)=\text{sign}(K)\end{aligned}$$ $\forall \epsilon >0$, using (.3), $\sigma_{min}\{ (K+\gamma I)-\frac{\lambda}{2} \text{sign}(K) \} >\epsilon $ gives a $(\lambda,\gamma) \in \hat{\mathscr{S}}^{K}_1$ and such a choice of $(\lambda,\gamma)$ gaurantees the estimator to be positive definite.\
[**Remark:**]{} Intuitively, a larger $\gamma$ shrinks the eigenvalues towards center which is 1, a larger $\gamma$ would result in positive definite estimator, whereas a larger $\lambda$ results in sparse estimate. A combination of $(\lambda, \gamma)$ results in a sparse and well-conditioned estimator. In particular case, when $K$ is diagonal matrix, the $\lambda<2*\gamma$.\
\
[**Proof of Theorem 3.1**]{} Define the function $Q(.)$ as following: $$Q(R)=f(R)-f(R_0)$$ where $R_0$ is the true correlation matrix and $R$ is any other correlation matrix. Let $R=UDU^T$ be eigenvalue decomposition of $R$, $D$ is diagonal matrix of eigenvalues and $U$ is matrix of eigenvectors. We have, $$\begin{aligned}
\begin{split}
Q(R)& = \|R-K\|_F^2+\lambda \|R^-\|_1+\gamma~ tr(D^2-2~D+p) \\
& ~~~-\|R_0-K\|_F^2 - \lambda\|R_0^-\|_1 -\gamma~ tr(D_0^2-2~D_0+p)
\end{split}\end{aligned}$$ $R_0=U_0D_0U_0^T$ is eigenvalue decomposition of $R_0$. Let $ \Theta_n(M):=\{\Delta : \Delta=\Delta^T , ~ \|\Delta\|_2=Mr_n,~ 0< M < \infty ~\}$. The estimate $\hat{R}$ minimizes the $Q(R)$ or equivalently $ \hat{\Delta}=\hat{R}-R_0$ minimizes the $G(\Delta)=Q(R_0+\Delta)$. Note that $G(\Delta)$ is convex and if $\hat{\Delta}$ be its solution, then we have $G(\hat{\Delta}) \le G(0) = 0$. Therefore if we can show that $G(\Delta)$ is non-negative for $\Delta \in \Theta_n(M)$, this will imply that the $\hat{\Delta}$ lies within sphere of radius $Mr_n$. We require $r_n=o\Big (\sqrt{(p+s)~log~p/n}\Big )$. $$\begin{aligned}
\|R-K\|_F^2-\|R_0-K\|_F^2 & =& tr(R^TR -2 R^TK+K^TK)-tr(R^T_0R_0-2R_0S+K^TK) \\
& = & tr(R^TR-R^T_0R_0)-2~tr((R-R_0)^TK) \\
& = & tr((R_0+\Delta)^T(R_0+\Delta)-R^T_0R_0)-2~tr(\Delta^TK)\\
& = & tr(\Delta^T\Delta)-2~tr(\Delta^T(K-R_0))\end{aligned}$$ Next, we bound term involving $K$ in above expression, we have $$\begin{aligned}
|tr(\Delta^T(R_0-K))| & \le & \sum_{i\neq j} |\Delta_{ij}({R_0}_{ij}-K_{ij})| \\
& \le & \max_{i\neq j}(|{R_0}_{ij}-K_{ij}|) \|\Delta^-\|_1 \\
& \le& C_0 (1+\tau) \sqrt{\frac{log~p}{n}} \|\Delta^-\|_1 \le C_1 \sqrt{\frac{log~p}{n}} \|\Delta^-\|_1\end{aligned}$$ holds with high probability by a result (Lemma 1) from @Ravi1:24 on the tail inequality for sample covariance matrix of sub-gaussian random vectors and where $C_1=C_0 (1+\tau), C_0 >0$. Next we obtain upper bound on the terms involving $\gamma$ in (.4). we have, $$\begin{aligned}
\lefteqn{tr(D^2-2D) - tr(D_0^2-2D_0) } \\
& = & tr\{R^2-R_0^2\} - 2~tr\{R-R_0)\} = tr(R_0+\Delta)^2-tr(R^2_0) \\
& = & 2~tr(R_0 \Delta) +tr(\Delta^T\Delta) \le 2~\sqrt{s}\|\Delta\|_F+ \|\Delta\|^2_F.
\end{aligned}$$ using Cauchy-Schwarz inequality. To bound the term $ \lambda(\|R^-\|_1-\|R_0^-\|_1) =\lambda(\|\Delta^-+R_0^-\|_1-\|R_0^-\|_1) $, let $E$ be index set as defined in Assumption A.2 of Theorem 3.2. Then using the triangle inequality, we obtain, $$\begin{aligned}
\lambda(\|\Delta^-+R_0^-\|_1-\|R_0^-\|_1) & = & \lambda(\|\Delta_E^-+R_{0}^-\|_1+\|\Delta_{\bar{E}}^-\|_1-\|R_0\|_1) \\
& \geq & \lambda(\|R_{0}^-\|_1-\|\Delta_E^-\|_1+\|\Delta_{\bar{E}}^-\|_1 -\|R_0^-\|_1) \\
& \geq & \lambda(\|\Delta_{\bar{E}}^-\|_1-\|\Delta_{E}^-\|_1)\end{aligned}$$ Let $ \lambda =(C_1/\epsilon)\sqrt{log~p/n}$, $\gamma=(C_1/\epsilon_1)\sqrt{log~p/n}, $ where $(\lambda, \gamma) \in \hat{\mathscr{S}}^{K}_1$, we obtain, $$\begin{aligned}
G(\Delta)& \geq & tr(\Delta^T\Delta)(1+\gamma)-2~C_1 \Big \{ \sqrt{\frac{log~p} {n}}(\|\Delta^{-}\|_1)+ \frac{1}{\epsilon_1}\sqrt{\frac{s~log~p} {n}} \|\Delta\|_F \Big \} \\
& & +\frac{C_1}{\epsilon} \sqrt{\frac{log~p}{n}} \big ( \|\Delta^-_{\bar{E}}\|_1 - \Delta^-_{E}\|_1 \big )\\
& \geq & \|\Delta\|_F^2(1+\gamma) -2C_1\sqrt{\frac{log~p}{n}} \big ( \|\Delta^-_{\bar{E}}\|_1 + \|\Delta^-_{{E}}\|_1 \big ) \\
& & \frac{C_1}{\epsilon} \sqrt{\frac{log~p}{n}} \big ( \|\Delta^-_{\bar{E}}\|_1- \Delta^-_{E}\|_1 \big ) - \frac{2C_1}{\epsilon_1}\sqrt{\frac{s~log~p} {n}} \|\Delta\|_F.\end{aligned}$$ Also because $\|\Delta^-_{E}\|_1=\sum_{(i,j) \in E, i \neq j } \Delta_{ij} \leq \sqrt{s} \|\Delta^-\|_F$, $$\begin{aligned}
-2C_1\sqrt{\frac{log~p}{n}} \|\Delta^-_{\bar{E}}\|_1 +\frac{C_1}{\epsilon} \sqrt{\frac{log~p}{n}} {\|\Delta^-_{\bar{E}}\|}_1 & \geq & \sqrt{\frac{log~p}{n}} \|\Delta^-_{\bar{E}}\|_1 \big ( -2 C_1 + \frac{C_1}{\epsilon} \big ) \\
& \geq & 0\end{aligned}$$ for sufficiently small $\epsilon$. Therefore, $$\begin{aligned}
G(\Delta) & \geq & \|\Delta\|_F^2 \big ( 1+\frac{C_1}{\epsilon_1}\sqrt{\frac{log~p}{n}} \big ) -C_1\sqrt{\frac{s~log~p}{n}} \|\Delta^+\|_F \{1+1/\epsilon+2/\epsilon_1\}\\
& \geq & \|\Delta\|_F^2\Big[1 +\frac{C_1}{\epsilon_1}\sqrt{\frac{log~p}{n}}-\frac{C_1}{M}\{1+1/\epsilon+2/\epsilon_1\} \Big] \\
& \geq & 0,\end{aligned}$$ for all sufficiently large $n$ and $M$. Which proves the first part of theorem. To prove the operator norm consistency, we have, $$\begin{aligned}
\|\hat{\Sigma}_K-\Sigma_0\| & = & \|\hat{W}\hat{R}\hat{W}-WK W\| \\
& \le & \|\hat{W}-W\| \|\hat{R}-K\| \|\hat{W}-W\| \\
& & + \|\hat{W}-W\| (\|\hat{R}\|\|W\| +\|\hat{W}\| \|K\|) + \|\hat{R}-K\| \|\hat{W}\| \|W\|.\end{aligned}$$ using sub-multiplicative norm property $\|AB\| \leq \|A\|\|B\|$. Since $\|K\|=O(1)$ and $\|\hat{R}-K\|_F=O(\sqrt{\frac{s~log~p}{n}})$ these together implies that $\|\hat{R}\|=O(1)$ . Also, $$\begin{aligned}
\|{\hat{W}}^2-W^2\| & = & \max_{\|x\|_2=1} \sum_{i=1}^p |({\hat{w}_i}^2-w_i^2)| x_i^2 \leq \max_{1 \leq i \leq p} |({\hat{w}_i}^2-w_i^2)| \sum_{i=1}^p x_i^2 \\
& = & \max_{1 \leq i \leq p} |({\hat{w}_i}^2-w_i^2)| = O \big ( \sqrt{\frac{log~p}{n}} \big ).\end{aligned}$$ holds with high probability by using a result (Lemma 1) from @Ravi1:24. Next we shall shows that $\|\hat{W}-W\|\asymp \|\hat{W}^2-W^2\|$, (where A$\asymp$B means A=$O_P(B)$ and B=$O_P(A)$). We have, $$\begin{aligned}
\|\hat{W}-W\| & = & \max_{\|x\|_2=1} \sum_{i=1}^p |({\hat{w}_i}-w_i)| x_i^2 = \max_{\|x\|_2=1} \sum_{i=1}^p | \big (\frac{{{\hat{w}_i}}^2-w_i^2}{\hat{w}_i+w_i} \big ) | x_i^2 \\
& \asymp & \sum_{i=1}^p |({\hat{w}_i}^2-w_i^2)| x_i^2 = C_3 \|\hat{W}^2-W^2\|.\end{aligned}$$ where we have used the fact that the true standard deviations are well above zero, i.e., $\exists~ 0 < C_3 < \infty $ such that $1/C_3 \leq w^{-1}_i \leq C_3 ~\forall ~i=1,2,\cdots, p$, and sample standard deviation are all positive, i.e, $\hat{w}_i > 0 ~\forall ~i=1,2,\cdots,p.$ Now since $\|\hat{W}^2-W^2\| \asymp \|\hat{W}-W\|$, this follows that $\|\hat{W}\|=O(1)$ and we have $\|\hat{\Sigma}_K-\Sigma_0\|^2=O \big(\frac{s~log~p}{n} +\frac{log~p}{n} \big )$. This completes the proof.\
\
**Proof of Theorem 3.2** Let $$\begin{aligned}
f(\Sigma)=||\Sigma-S||^2_F + \lambda \|\Sigma^-\|_1 + \gamma\sum_{i=1}^{p} \{\sigma_i(\Sigma)-\bar{\sigma}_{\Sigma}\}^2,\end{aligned}$$ Similar to the proof of theroem (3.1), define the function $Q_1(.)$ as following: $$Q_1(\Sigma)=f(\Sigma)-f(\Sigma_0)$$ where $\Sigma_0$ is the true covariance matrix and $\Sigma$ is any other covariance matrix. Let $\Sigma=UDU^T$ be eigenvalue decomposition of $\Sigma$, $D$ is diagonal matrix of eigenvalues and $U$ is matrix of eigenvectors. We have, $$\begin{aligned}
\begin{split}
Q_1(\Sigma)& = \|\Sigma-S\|_F^2+\lambda \|\Sigma^-\|_1+\gamma~ tr(D^2)-(tr(D))^2/p \\
& ~~~-\|\Sigma_0-S\|_F^2 - \lambda\|\Sigma_0^-\|_1 -\gamma~ tr(D_0^2)-(tr(D_0))^2/p
\end{split}\end{aligned}$$ where $A=diag(a_1,a_2,\cdots,a_p)$ and $\Sigma_0=U_0D_0U_0^T$ is eigenvalue decomposition of $\Sigma_0$. Write $\Delta=\Sigma-\Sigma_0$, and let $ \Theta_n(M):=\{\Delta : \Delta=\Delta^T , ~ \|\Delta\|_2=Mr_n,~ 0< M < \infty ~\}$. The estimate $\hat{\Sigma}$ minimizes the $Q(\Sigma)$ or equivalently $ \hat{\Delta}=\hat{\Sigma}-\Sigma_0$ minimizes the $G(\Delta)=Q(\Sigma_0+\Delta)$. Note that $G(\Delta)$ is convex and if $\hat{\Delta}$ be its solution, then we have $G(\hat{\Delta}) \le G(0) = 0$. Therefore if we can show that $G(\Delta)$ is non-negative for $\Delta \in \Theta_n(M)$, this will imply that the $\hat{\Delta}$ lies within sphere of radius $Mr_n$. We require $\sqrt{(p+s)~log~p}=o \Big (\sqrt{n} \Big ) $. $$\begin{aligned}
\|\Sigma-S\|_F^2-\|\Sigma_0-S\|_F^2 & =& tr(\Sigma ^T \Sigma -2 \Sigma^TS+S^TS)-tr(\Sigma_0^T\Sigma_0-2\Sigma_0S+S^TS) \\
& = & tr(\Sigma^T\Sigma-\Sigma_0^T\Sigma_0)-2~tr((\Sigma-\Sigma_0)S) \\
& = & tr((\Sigma_0+\Delta)^T(\Sigma_0+\Delta)-\Sigma_0^T\Sigma_0)-2~tr(\Delta^TS)\\
& = & tr(\Delta^T\Delta)-2~tr(\Delta^T(S-\Sigma_0))\end{aligned}$$ Next, we bound term involving $S$ in above expression, we have $$\begin{aligned}
|tr(\Delta(\Sigma_0-S))| & \le & \sum_{i\neq j} |\Delta_{ij}({\Sigma_0}_{ij}-S_{ij})| + \sum_{i=1} |\Delta_{ii}({\Sigma_0}_{ii}-S_{ii})| \\
& \le & \max_{i\neq j}(|{\Sigma_0}_{ij}-S_{ij}|) \|\Delta^-\|_1 + \sqrt{p} \max_{i=1} (|{\Sigma_0}_{ii}-S_{ii}|) \sqrt{\sum_{i=1} \Delta^2_{ii}} \\
& \le& C_0 (1+\tau) \max_{i}(\Sigma_{0ii}) \Big \{ \sqrt{\frac{log~p}{n}} \|\Delta^-\|_1+ \sqrt{\frac{p~log~p}{n}} \|\Delta^+\|_2 \Big \} \\
& \le & C_1 \Big \{ \sqrt{\frac{log~p}{n}} \|\Delta^-\|_1+ \sqrt{\frac{p~log~p}{n}} \|\Delta^+\|_2 \Big \}\end{aligned}$$ holds with high probability by a result (Lemma 1) from @Ravi1:24 where $C_1=C_0 (1+\tau) \max_{i}(\Sigma_{0ii}), C_0 >0$ and $\Delta^+$ is matrix $\Delta$ with all off-diagonal elements set to zero. Next we obtain upper bound on the terms involving $\gamma$ in (3.7). we have, $$\begin{aligned}
\lefteqn{tr(D^2)-(tr(D))^2/p - tr(D_0^2)-(tr(D))^2/p } \\
& = & tr(\Sigma^2)-tr(\Sigma_0^2) - (tr(\Sigma))^2/p + (tr(\Sigma_0))^2/p
\end{aligned}$$ $$\begin{aligned}
\textbf{(i)}~~~\lefteqn{tr(\Sigma^2)-\Sigma_0^2))} \\
& \leq & tr(\Sigma_0+\Delta)^2 -tr(\Sigma_0)^2 \\
& = & tr(\Delta)^2+ 2~tr(\Delta^2\Sigma_0) \leq tr(\Delta)^2 + C_1 \sqrt{s}\|\Delta\|_F\end{aligned}$$ $$\begin{aligned}
\textbf{(ii)}~~~\lefteqn{tr((\Sigma))^2 -(tr(\Sigma_0))^2}\\
& = & (tr(\Sigma_0+\Delta))^2 - (tr(\Sigma_0))^2 \\
& \leq & (tr(\Delta))^2 +2~tr(\Sigma_0)~tr(\Delta) \leq p~\| \Delta\|^2_F +2~\bar{k} p\sqrt{p} \|\Delta^+\|_F.\end{aligned}$$ Therefore the $\gamma$ term can be bounded by $2\|\Delta\|^2_F+(C_1\sqrt{s}+2\sqrt{p}\bar{k})\|\Delta\|_F$. We bound the term invloving $\lambda$ as in similar to the proof of Theorem 3.1. For $\lambda \asymp \gamma \asymp \sqrt{\frac{log~p}{n}}$, the proof follows very simialr to Therem 3.1.\
\
[**Proof of Theorem 3.3.**]{} To bound the cross product term involving $\Delta$ and $\hat{R}_K^{-1}$, we have, $$\begin{aligned}
|tr((R_0^{-1}-\hat{R}_K^{-1})\Delta)| & = & |tr(R_0^{-1}(\hat{R}_K-R_0)\hat{R}_K^{-1}{\Delta})| \\
& \leq & \sigma_1(R_0^{-1})|tr((\hat{R}_K-R_0)\hat{R}_K^{-1} \Delta)| \\
& \leq & \bar{k}\sigma_1(\hat{R}_K^{-1})|tr((\hat{R}_K-R_0) \Delta)| \\
& \leq & \bar{k} \bar{k}_1|tr((\hat{R}_K-R_0)\Delta)|.\end{aligned}$$ where $\sigma_{min}(\hat{R}_K) \geq (1/\bar{k}_1) >0 $, is a positive lower bound on the eigenvalues of JPEN estimate $\hat{R}_K$ of correlation matrix $R_0$. Such a constant exist by Lemma 3.2. Rest of the proof closely follows as that of Theorem 3.1.\
\
[**Proof of Theorem 3.4.**]{} We bound the term $tr((\hat{\Omega}_{S}-\Omega_0)\Delta)$ similar to that in proof of Theorem 3.3. Rest of the proof closely follows to that Theorem 3.2.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new approach to studying the evolution of massive black hole binaries in a stellar environment. By imposing conservation of total energy and angular momentum in scattering experiments, we find the dissipation forces that are exerted on the black holes by the stars, and thus obtain the decaying path of the binary from the classical dynamical friction regime down to subparsec scales. Our scheme lies between scattering experiments and $N$-body simulations. While still resolving collisions between stars and black holes, it is fast enough and allows to use a large enough number of particles to reach a smooth and convergent result. We studied both an equal mass and a 10:1 mass ratio binaries under various initial conditions. We show that while an equal mass binary stalls at a nearly circular orbit, a runaway growth of eccentricity occurs in the unequal mass case. This effect reduces the timescale for black hole coalescence through gravitational radiation to well below the Hubble time, even in spherical and gasless systems formed by dry mergers.'
author:
- |
Yohai Meiron[^1] and Ari Laor\
Department of Physics, Technion – Israel Institute of Technology, Haifa 32000, Israel
date: 'Accepted 2012 January 17. Received 2012 January 16; in original form 2011 October 28'
title: 'A conservation-based method for simulating the inspiral of binary black holes'
---
\[firstpage\]
black hole physics – stars: kinematics and dynamics – galaxies: nuclei.
Introduction
============
Supermassive black hole (BH) binaries are formed as a result of galactic mergers. The two BHs sink towards the bottom of the potential well and form a bound pair at the centre of the stellar distribution. According to the classical picture [@begelman80] the binary hardens until the loss cone is depleted (i.e. depleted of stars on low angular momentum orbits that pass close enough to the binary for a significant interaction), and stalls at a separation too large for the emission of gravitational radiation to cause rapid inspiral and coalescence (the ‘final parsec problem’). This picture has been tested extensively using $N$-body simulations and scattering experiments. Early numerical studies concentrated on the simple case of an equal mass binary on a circular orbit within a spherically symmetric stellar distribution. Those studies (e.g. @makino97; @qh97; @mm01; [-@mm03]), with a relatively small number of particles, were somewhat inconsistent with each other and could not confirm loss cone depletion due to spurious numerical relaxation over long time scales.
In recent years, more detailed investigations have been performed. Apart from the increased number of particles and longer integration times, various complications have been considered such as triaxiality and rotation (e.g. @berczik06; @khan11), post Newtonian corrections (e.g. @berentzen09) massive perturbers (e.g. @hagai08) and gas discs (e.g. @jorge09). These studies found that binary coalescence in less than the Hubble time is feasible even if only one of these factors is present. Thus, the final parsec is potentially not a problem in a realistic merger remnant.
More recent work (e.g. @iwasawa11; @sesana11) has focused on the evolution of the binary eccentricity rather than just the binary separation. For high mass ratios ($\sim$ 100:1), these authors showed that even in a spherically symmetric environment, eccentricity can increase to almost unity. Very high eccentricity means that in pericentre the two BHs can be close enough together that gravitational wave emission becomes efficient. This enhanced energy loss to gravitational radiation (as compared to the circular case with the same semi-major axis) can reduce coalescence timescale to well below the Hubble time.
Direct $N$-body simulations are the most accurate way to study binary BH (BBH) evolution, but they are computationally expensive. It is therefore difficult to perform diverse enough tests to cover the problem’s parameter space. By compromising for an unrealistically small number of particles, one introduces spurious relaxation. In spherical galaxies, this process drives unrealistic loss cone repopulation (@yu02; @mm03) and causes the hardening rate to be highly $N$-dependent (@makino04; @berczik05). @berczik06 followed the evolution of a BBH in triaxial and rotating galaxy models and found that the hardening rate was $N$-independent, implying a collisionless mode of loss cone repopulation.
Our previous work dealt with a possible stellar kinematical signature of stars around a BBH. In @meiron10 we produced kinematical maps (projected maps of mean stellar velocity, velocity dispersion and higher velocity moments); this was achieved by scattering stars on a BBH on a fixed circular orbit. To produce more realistic kinematical maps, we looked for a way to calculate a BBH path fast and accurately so it could be used for following scattering experiments. Good kinematical maps require a very large number ($N\lesssim 10^8$) of stars on small scales, otherwise the high moments of the line of sight velocity distribution are poorly resolved. $N$-body simulations cannot be made yet with such a large number of particles on small enough scales, but this is not a problem for scattering experiments which are performed on a precalculated BBH path. Since we were already working with scattering experiments, we derived a way to use the existing code base and adapt it to work more like an actual $N$-body simulation, where the effect of scatterings on the BBH orbit is taken into account, giving the orbital evolution.
In this paper we present a new scheme we developed to simulate the inspiral of the BBH from the galactic scale to the hard binary scale. Our method is based on imposing conservation of total energy and angular momentum, instead of directly summing the forces of individual stars on the BHs, and lies between scattering experiments and $N$-body simulations. Studying this type of systems with scattering experiments is significantly faster than a full $N$-body treatment and is appropriate when the bulge has relaxed. The phase space stellar distribution is followed accurately and a more realistic number of stars can be included. Thus, we are able to run many simulations and probe a large range of parameters under reasonable physical assumptions, as well as the convergence of the solution, and compare the inspiral timescale, the stalling radius, and the eccentricity evolution to earlier calculations. Since in our method we obtain the forces on the BHs, it is also possible to test analytical expressions for dynamical friction. The original treatment by @chandra is for a homogeneous background, therefore we compare our results to @just11 who investigate dynamical friction in power law cusps. Our solution extends beyond the range of validity of their formula and into the hard binary regime of the BHs.
We studied the evolution of both an equal mass binary and a 10:1 mass ratio binary under different initial conditions. We found that both cases presented stalling of the semi-major axis, but in the unequal mass case, eccentricity tended to grow towards unity on timescales well below the Hubble time.
In Section \[sec:math\] we give a general mathematical formulation of our simulation scheme. The model for the specific simulations we performed in this work is described in Section \[sec:model\] while a technical description of the scheme appears in Section \[sec:algorithm\]. In Section \[sec:results\] we present the results of all BBH simulations and discuss their physical implication. In Section \[sec:df\] we discuss dynamical friction, compare to earlier results, and show that our code is applicable in very large radii, thus potentially helpful in future studies of phenomena related to dynamical friction. Finally, we give a short summary in Section \[sec:summary\].
Mathematical Formulation {#sec:math}
========================
Motivation
----------
The two basic ideas of our solution are the separation of timescales and the balance of energy and angular momentum between the BHs and the stars. There are three timescales in the system corresponding to changes in:
1. stellar orbits following close encounters
2. the BHs’ orbits
3. the background stellar potential
In more detail: a star significantly changes its original orbit during a close encounter with a BH, but the BH’s path is only slightly perturbed as the force exerted on it is due to many small ‘scattering’ events; the change of the background stellar potential is due to the collective response of stars to the perturbation, which evolves on the dynamical timescale. A yet longer timescale would be of 2-body relaxation, which is longer by a factor of $\sim 0.1 N/\ln N$ [@bt] and is expected to be well above the Hubble time for a real galaxy core. However, relaxation time is potentially significantly shorter if the dominant relaxation mode is not 2-body relaxation.
The most basic kind of an $N$-body simulation uses very small steps, after each the vector forces exerted by all the field stars (particles) are summed up and applied to the BHs in the next step; each particle is propagated the same way. Energy and angular momentum are globally conserved (within a given numerical error tolerance) since it is a closed system. If $N$ is not large enough, noisy potential and unrealistically massive stars lead to spurious relaxation. If the time step is too large, close encounters cannot be resolved.
Our scheme uses a ‘large’ $\Delta t$, which we call an [*interval*]{}, within which there is one or more actual [*steps*]{} of the ODE solver. Let BH number $i$ (where $i\in\{1,2\}$) propagate one interval between times $t_0$ and $t_1=t_0+\Delta t$ from some vector position ${\bf s}_i(t_0)$ to ${\bf s}_i(t_1)$; we use the term [*segment*]{} for the path length. After each interval, the energies and angular momenta of all the particles are summed up. If no work was done and no torque exerted due to the background stellar potential, then the energy $\Delta E_\star$ (angular momentum $\Delta L_\star$) gained by the stars during some interval must be equal and opposite to the energy $\Delta E_{{\rm BH}}$ (angular momentum $\Delta L_{{\rm BH}}$) lost to the BHs during this interval. For simplicity, we use a static and spherically symmetric model for the stellar potential, so these requirements are automatically fulfilled. The scheme can accommodate more complicated models as well: the demand for spherical (or more generally, axial) symmetry, which is necessary if one assumes that change of a star’s angular momentum is only due to interaction with the BHs, can be relaxed if the torque component due to the stellar bulge is considered separately; a slowly varying potential can also be taken into account if the proper adjustments are made to the code.
The evolution of a BBH is dominated by close and fast encounters with the field stars and the evolution of the stellar gravity field is only a secular effect. Thus, the part of $\Delta E_\star$ due only to the change in stellar potential is negligible after each interval, justifying the use of a static model. However, at late times, the initial stellar potential is no longer consistent with the actual spatial distribution. The static potential assumption is still reasonable as far as the BHs are concerned, since by the time that any significant evolution of the background potential has taken place, the BHs will have fallen deep enough in the potential well, where the dominant force is the other BH’s gravity rather than the background stellar potential (so that the exact shape of the potential well does not matter). Stars further out orbit in a ‘wrong’ potential, but as long as approximate spherical symmetry is preserved, their interaction rate with the BHs should not be significantly influenced by this (and during a close encounter, the background stellar potential is of course unimportant).
Assuming additionally that the interval is short enough so that the forces on the BHs due to the background stars do not vary significantly, we perform scattering experiments in each interval to find $\Delta E_\star$ and $\Delta L_\star$. Using the simple algebra described below, we find the average forces on the BHs over this interval. We also assume a purely planar motion of the BHs. Thus, unless stated otherwise, by ‘angular momentum’ and ‘torque’ we mean only the $z$ component of these vectors.
Conservation laws
-----------------
The basic conservation equations are $$\begin{aligned}
\Delta E_{{\rm BH}} &= -\Delta E_{\star}=\int_{{\bf s}_1(t_{0})}^{{\bf s}_1(t_1)}{\bf F}_1\cdot{\bf ds}_1 + \int_{{\bf s}_2(t_{0})}^{{\bf s}_2(t_1)}{\bf F}_2\cdot{\bf ds}_2,\label{eq:cons-E}\\
\Delta L_{{\rm BH}} &= -\Delta L_{\star}=\int_{t_{0}}^{t_1}\tau_1{\rm d}t + \int_{t_{0}}^{t_1}\tau_2{\rm d}t,\label{eq:cons-L}\end{aligned}$$ where ${\bf F}_i$ and $\tau_i\equiv\left({\bf r}_i\times{\bf F}_i\right)\cdot\hat{{\bf z}}$ are the force and torque, respectively, exerted on BH number $i$ by the stellar population. Let us write ${\bf F}_i$ in the following form: $${\bf F}_i=-f_i\hat{{\bf v}}_i+\tilde{f}_i\hat{{\bf u}}_i,$$ where $\hat{{\bf v}}_i$ and $\hat{{\bf u}}_i$ are the unit vectors parallel and perpendicular to the velocity of the BH. Note that $f$ is ‘drag like’ and directed opposite to the velocity vector. If an object moves through a uniform background, symmetry dictates that the mean force would be antiparallel to the velocity vector. However in realistic environments there must be also a force due to the inhomogeneities of the background, so a perpendicular force component is required. In Cartesian coordinates: $$\begin{aligned}
\hat{{\bf v}}_i &= \frac{v_{ix}}{v_i}\hat{{\bf x}}+\frac{v_{iy}}{v_i}\hat{{\bf y}},\\
\hat{{\bf u}}_i &= \frac{v_{iy}}{v_i}\hat{{\bf x}}-\frac{v_{ix}}{v_i}\hat{{\bf y}}.\end{aligned}$$ To simplify equation (\[eq:cons-E\]) we write ${\bf ds}_i=\hat{{\bf v}}_i{\rm d}s_i$ and thus $$\begin{aligned}
\int_{{\bf s}_i(t_{0})}^{{\bf s}_i(t_1)}{\bf F}_i\cdot{\bf ds}_i &=
\int_{{\bf s}_i(t_{0})}^{{\bf s}_i(t_1)}\left(-f_i\hat{{\bf v}}_i+\tilde{f}_i\hat{{\bf u}}_i\right)\cdot\hat{{\bf v}}_i{\rm d}s_i\notag\\
{} &=-f_i\int_{{\bf s}_i(t_{0})}^{{\bf s}_i(t_1)}{\rm d}s_i\equiv-f_i\mathcal{S}_i.\label{eq:def-S}\end{aligned}$$ The perpendicular force component $\tilde{f}$ disappears due to the dot product, while $f$ is assumed constant along the path and can be taken out of the integral, which defines $\mathcal{S}$ (which is simply the path’s length). Finally, equation (\[eq:cons-E\]) for the energy becomes $$\Delta E_{\star}=f_1\mathcal{S}_1+f_2\mathcal{S}_2.\label{eq:forces-E}$$
To simplify equation (\[eq:cons-L\]) we write $\tau_1$ in Cartesian coordinates: $$\tau_i = -\frac{f_i}{v_i}(x_iv_{iy}-y_iv_{ix}) - \frac{\tilde{f}_i}{v_i}(x_iv_{ix}+y_iv_{iy}),$$ and thus $$\begin{aligned}
\int_{t_{0}}^{t_1}\tau_i{\rm d}t &= -f_i\int_{t_{0}}^{t_1}\frac{x_iv_{iy}-y_iv_{ix}}{v_i}{\rm d}t-\tilde{f_i}\int_{t_{0}}^{t_1}\frac{x_iv_{ix}+y_iv_{iy}}{v_i}{\rm d}t \notag \\
{} &\equiv -f_i\mathcal{P}_i-\tilde{f_i}\mathcal{Q}_i.\end{aligned}$$ The above integrals define $\mathcal{P}$ and $\mathcal{Q}$. Equation (\[eq:cons-L\]) for the angular momentum becomes $$\Delta L_{\star}=f_1\mathcal{P}_1+\tilde{f_1}\mathcal{Q}_1+f_2\mathcal{P}_2+\tilde{f_2}\mathcal{Q}_2.\label{eq:forces-L}$$
If we evolve the BHs between $t_0$ and $t_1$ under their mutual gravity alone, energy (and angular momentum) will be conserved along the produced orbital segment: $\Delta E_{\rm BH}=0$. Stars scattered on this orbital segment will not their conserve energy: $\Delta E_\star\neq 0$; so energy is also globally not conserved between $t_0$ and $t_1$. By solving equations (\[eq:forces-E\]) and (\[eq:forces-L\]) for the (non conserving) forces, the BHs can be evolved again in this time interval, under the additional forces, producing an orbital segment for which $\Delta E_{\rm BH}=-\Delta E_\star$. The revised orbital segment is only slightly different from the original, since the additional forces are much smaller than the forces exerted by the other BH and the background potential. Stars scattered on the revised orbit will have a slightly different $\Delta E_\star$. This process of alternatingly evolving the BHs and the scattering of stars can be repeated until converges is achieved.
Since $\Delta E_\star$ and $\Delta L_\star$ are directly obtained from the scattering experiments and the six integrals (calligraphic letters) are calculated from the orbital segments, equations (\[eq:forces-E\]) and (\[eq:forces-L\]) are a linear system of two equations with four variables: $f_1$, $f_2$, $\tilde{f}_1$ and $\tilde{f}_2$. It is worth noting that without the perpendicular force, which is expected to be negligible in the standard dynamical friction formalism, it is impossible to conserve energy and angular momentum simultaneously. In particular cases, additional constraints give an exact solution as explained in the following section.
Finding the forces {#sec:forces}
------------------
### Symmetric motion
In this case, the masses are equal and the initial conditions are symmetric with respect to the centre of the system. The motion of one BH mirrors that of its companion, so that the path integrals are equal for the two BHs (the index is therefore dropped) and the forces acting on them must also be equal due to symmetry. The solution is $$\begin{aligned}
f &= \frac{1}{2\mathcal{S}}\Delta E_\star,\label{eq:force-sym-par}\\
\tilde{f} &= \frac{1}{2\mathcal{SQ}}(\Delta L_\star \mathcal{S}-\Delta E_\star \mathcal{P}).\label{eq:force-sym-per}\end{aligned}$$ There is no solution for $\mathcal{Q}=0$ (but $\mathcal{S}\neq 0$ is guaranteed by equation \[eq:def-S\]). In points where the BHs’ velocity is purely tangential, $\tilde f$ is parallel to the radius vector and thus exerts no torque (and work is never done by $\tilde f$). At these points, $\tilde f$ is free but $f_i$ is overdetermined (must change both $E_{\rm BH}$ and $L_{\rm BH}$ by the specified amounts). Segments which are symmetric about such points have $\mathcal{Q}=0$.
### High mass ratio
In this case one mass is much larger that the other, and is assumed to sit motionless at the centre of the system. The forces $f$ and $\tilde{f}$ act on the secondary BH only. The integrals $\mathcal{S}$, $\mathcal{P}$ and $\mathcal{Q}$ are also calculated for the secondary only. The solution for $\mathcal{Q}\neq 0$ is $$\begin{aligned}
f &= \frac{1}{\mathcal{S}}\Delta E_\star,\\
\tilde{f} &= \frac{1}{\mathcal{SQ}}(\Delta L_\star \mathcal{S}-\Delta E_\star \mathcal{P}).\end{aligned}$$
### General solution
The results for the limiting cases described above motivate us to look for solutions of the form $$\begin{aligned}
f_1 & =\frac{\alpha}{\mathcal{S}_1}\Delta E_{\star}, & \tilde{f}_1 & =\frac{\alpha}{\mathcal{S}_1\mathcal{Q}_1}\left(\Delta L_{\star}\mathcal{S}_1-\Delta E_{\star}\mathcal{P}_1\right),\\
f_2 & =\frac{1-\alpha}{\mathcal{S}_2}\Delta E_{\star}, & \tilde{f}_2 & =\frac{1-\alpha}{\mathcal{S}_2\mathcal{Q}_2}\left(\Delta L_{\star}\mathcal{S}_2-\Delta E_{\star}\mathcal{P}_2\right).\end{aligned}$$ There is a mathematical solution for every $\alpha$, but we know from the limiting cases that $0\leq\alpha\leq\frac{1}{2}$. The value of $\alpha$ is a function of the mass ratio $q$, but it may also be dependent on other factors, such as the local stellar densities at the instantaneous position of either BH. In the case of a high mass ratio, one BH is almost stationary; therefore $\alpha$ must approach zero faster or at least as fast as $\mathcal{S}_{1}$. The force $f_{1}$ does not have to approach zero, but the acceleration $f_{1}/M_{1}$ does.
Model {#sec:model}
=====
Units {#sec:units}
-----
For reasons of consistency with our previous work [@meiron10], we use a unit system in which mass is measured in units of the primary BH’s mass, velocity is measured in units of $4\sigma$ and $G=1$ (where $\sigma$ and $G$ are the stellar velocity dispersion and the gravitational constant, respectively). The hard binary separation is defined as $$a_{\rm h} \equiv \frac{q}{1+q} \frac{GM_\bullet}{4\sigma^2},$$ where $q \leq 1$ is the mass ratio of the secondary and primary BHs and $M_\bullet$ is the mass of the primary. In our units, the hard binary separation of an equal mass binary ($q=1$) is 2.
The base units are therefore scalable by $M_\bullet$ and $\sigma$. Only one parameter is required if we also use the $M$–$\sigma$ relation (e.g. @m-sigma). The units of length, time and velocity units and their scaling, using $M_\bullet$ and $\sigma$, are $$\begin{aligned}
[{\rm L}] &= {\textstyle\frac{1}{16}} GM_\bullet\sigma^{-2} = 0.77\ M_8^{0.53} ~{\rm pc},\\
[{\rm T}] &= {\textstyle\frac{1}{64}} GM_\bullet\sigma^{-3} = 1\,000\ M_8^{0.29} ~{\rm yr},\label{eq:units-T}\\
[{\rm V}] &= 4\sigma = 750\ M_8^{0.24} ~{\rm km~s^{-1}},\end{aligned}$$ where $M_8$ is the physical mass of the primary BH in units of $10^8 ~{\rm M}_\odot$. Note that in all simulations we used a stellar velocity dispersion of 0.25 velocity units.
Bulge Properties {#sec:bulge}
----------------
In all our simulations, stars are initially distributed in a singular isothermal sphere and follow a Maxwell–Boltzmann distribution with 1D velocity dispersion $\sigma$. To avoid the non-physical divergence of the potential, we assume a core structure: $$\rho(r)=
\begin{cases}
\rho_0 & r\leq h\\
\rho_0\left(\frac{h}{r}\right)^{2} & r>h
\end{cases}\label{eq:isothermal-density}$$ where $h$ is an arbitrary break radius set to 1 and $\rho_0={\sigma^2}/{2\pi G h^2}$. The expression for the gravitational potential (or the bulge potential) derived from the above density is $$\Phi_{\rm bulge}(r)=
\begin{cases}
\frac{\sigma^2}{3Gh^2}r^2 & r\leq h\\
\frac{\sigma^{2}}{G}\left[\frac{4h}{3r}+2\ln\left(\frac{r}{h}\right)-1\right] & r>h
\end{cases}\label{eq:bulge-potential}$$ @koopmans09 found that massive elliptical galaxies, within their effective radii, are well approximated by a power law ellipsoid with an index of $-2$. @genzel03 also found that the density of the nuclear star cluster of the Milky way can be described by a broken power law with index of $-2.0 \pm 0.1$ down to 0.38 pc. This is however not appropriate within the BH sphere of influence, if dynamically relaxed, where the equilibrium distribution (radius independent mass and energy flow) is the famous @bw76 cusp of $\rho \propto r^{-7/4}$ (where only one mass species is present, cf. @bw77, @hopman-alexander06). Unrelaxed clusters around adiabatically growing BHs are expected to have shallower slopes [@young80] or steeper slopes in the case of rotating systems and non-isothermal clusters [@lee89].
In the case of a minor merger (equivalent to $q=0.1$), the structure of the more massive galaxy does not change significantly and equation (\[eq:isothermal-density\]) likely represents the stellar environment seen by the secondary BH after its parent galaxy is absorbed. This picture is somewhat naïve in the case of a major merger ($q=1$), but violent relaxation [@lynden-bell67] due to the rapidly varying potential in a newly merged galaxy causes widening of the stellar energy distribution and is analogous to relaxation by collisions in a gas. This process tends to drive galaxies towards a universal steady state [@syer98]. A nice demonstration of this appears in the $N$-body merger simulations of @mm01, who find an $r^{-2}$ density profile at the time the equal mass binary becomes hard, which extends down to the scale of the binary separation.
Throughout this work we assume that the BHs are ‘naked’, that is, do not carry clusters of bound stars. In the equal mass case, the stellar mass bound to a single BH once the binary becomes hard does not exceed 10 per cent of its mass (assuming that the cluster also has a power law density profile with index of -2 and normalization based on the $M$–$\sigma$ relation). This will most likely not affect the late time evolution, but will surely affect the early inspiral phase, and the exact inspiral time, but will probably not have a major effect.
Algorithm {#sec:algorithm}
=========
Equal masses {#sec:alg-equal}
------------
Below we provide a technical description of the application of the technique described in Section \[sec:math\] for an equal mass binary simulation. The slightly different procedure for the 10:1 mass ratio case is discussed in the following Section.
First, a realization of a singular isothermal sphere is produced up to a cutoff radius of ${R_{\rm max}}$, with two equal mass BHs placed on the $x$-axis at $x=\pm R_0$; at this initial distance the BHs are still unbound to each other and their inspiraling orbits are governed by the bulge’s gravity and dynamical friction. The simulation duration is divided into (equal) intervals $\Delta t$, after each the stellar force acting on the BHs is updated. The actual time [*steps*]{} of the ODE solver are smaller if necessary, so close encounters can be resolved. Within each interval we do the following:
1. Symmetrically advance[^2] the two BHs from $t_i$ to $t_{i+1}$, under their mutual gravity, and the effective forces exerted by the stars. The calculated antiparallel $f$ and perpendicular $\tilde{f}$ embody the stars’ pull on the BHs, so the bulge potential (equation \[eq:bulge-potential\]) should not be considered additionally. This gives a short orbital segment which is stored in memory.
2. Advance each star from $t_i$ to $t_{i+1}$, under the influence of the two BHs and the bulge potential. The motion of the BHs in this interval is already set from the previous stage, so this is essentially a short scattering experiment. This is the most computationally demanding stage of each iteration, but easily parallelized.
3. Sum up the energies and angular momenta of all stars and subtract the values from the previous iteration to obtain $\Delta E_\star$ and $\Delta L_\star$.
4. Use the BHs’ path to obtain $\mathcal{S}$, $\mathcal{P}$ and $\mathcal{Q}$.
5. Calculate $f$ and $\tilde{f}$ from equations (\[eq:force-sym-par\]) and (\[eq:force-sym-per\]), to be used in the next $\Delta t$ interval.
Unless otherwise indicated, our simulations end at $t=10\,000$ time units (or $10^7$ years for a $10^8~{\rm M}_\odot$ primary); this is equivalent to $\sim 1\,000$ revolutions after the binary becomes hard in most equal mass simulations. Integration of a single star is terminated prematurely in three cases: if it reaches a distance of ${R_{\rm max}}+10$ length units from the centre of the system, reaches $r_{\rm tidal}=10^{-3}$ from either BH or takes more than 35 integration time steps to complete the interval of $\Delta t=0.1$ (see Section \[sec:qa\]). In the first case the star is considered to have escaped the system (or [*diverged*]{}); the extra 10 length units beyond ${R_{\rm max}}$ are an arbitrary ‘padding’ required for technical reasons. The second case represents tidal disruption of the star (the orbit is then said to have [*crashed*]{}). The true tidal disruption radius is up to two orders of magnitude smaller than our $r_{\rm tidal}$, but this choice has a negligible effect on the BHs because the rate of crashing stars is negligible compared to the rate of diverging stars. The choice of ${R_{\rm max}}$ is rather arbitrary, and is chosen to be large enough to minimize its effect on the results, as discussed in Section \[sec:Rmax\].
Notably, the forces on the BHs are one interval retarded. Thus, the interval duration must be short enough in order not to break conservation of energy and angular momentum. In all simulations we chose to use equal intervals of $\Delta t=0.1$, and since simulations done with this choice both ran reasonably quickly and performed well in terms of conservation, we did not thoroughly investigate changing $\Delta t$. Certainly making the interval length longer or adaptive can significantly speed up the simulations.
The code outlined above is very simple in terms of decision making; the bulk of CPU effort is made to individually advance stars (scattering experiments) with only a small non-parallelizable overhead between intervals. A typical simulation with a large realization ($N=5\times 10^6$) would run for $\sim$ few days on a medium strength personal desktop computer.
Mass ratio {#sec:alg-ratio}
----------
As noted in Section \[sec:forces\], the force acting on the BHs can be uniquely found from the energy and angular momentum differences only in the cases of equal masses and high mass ratios (where it is possible to assume that the primary BH is fixed at the centre). However, when the velocity is perpendicular to the radius vector (i.e. at pericentre or apocentre), the values of $\mathcal{Q}_i$ approach zero, and thus $\tilde{f}_i$ are left out of the coupled equations (\[eq:forces-L\]) and (\[eq:forces-E\]), and cannot be solved for. At the same time, $f_i$ are overdetermined as they have to compensate for both energy and angular momentum changes. Even in intervals that contain an apsis, there is usually a solution unless the orbital segment happens to be symmetric about the apsis. Since only a few segments are affected, the effect on the BBH orbits is not significant.
In the case of equal masses, no special treatment of these segments was necessary; but in the case of very high mass ratios (i.e. $q \ll 1$) the finite numerical fluctuations are larger due to the fact that the secondary BH is less massive. In these cases the total angular momentum would discontinuously drop at an apsis, and the otherwise smooth path would suddenly break at this segment. We compromised on simulating a 10:1 binary, and utilized two techniques to improve accuracy. First, each step was performed twice: after calculating the frictional force, the stars and BH were reset to their original positions, and advanced again with the newly calculated force on the BH. Second, we attempted to compensate for the accumulated error: instead of calculating the force components using purely $\Delta E_\star$ and $\Delta L_\star$ of the last $\Delta t$ interval, we added their respective accumulated errors (the correction term had weight of 0.1 per cent). Those adjustments dramatically improved the accuracy of the 10:1 simulations, but unfortunately we did not yet overcome all the technical problems associated with reliably simulating a higher mass ratio inspiral from $\sim 50$ pc down to stalling separation.
Quality Assurance {#sec:qa}
-----------------
If the routine that advances the stars has a bug or is just not accurate enough, the values of $\Delta E_\star$ and $\Delta L_\star$ obtained in each interval will be faulty; but since the frictional force is calculated in such a way that would compensate for any change of $E$ and $L$ in the stars, the bug might remain undetected. Thus, an important validity test of $E$ and $L$ conservation in the code is inherently not available.
We can, however, test the ODE solver for a similar problem, and infer that our stellar orbits are at least [*well behaved*]{} in the original problem. Assume a BBH with a circular orbit of constant radius $R$; the BBH orbit is now decoupled from the stars and each stellar orbit can be integrated separately. This is the restricted 3-body problem plus a spherically symmetric potential; as in @meiron10, the energy in the rest frame of the BHs (the frame which rotates with the same angular frequency) is the only conserved quantity. The Jacobi integral is a constant of motion related to the rest frame energy by $C_J=-2E$.
We performed a number of tests where the BBH was in a circular orbit with a radius in the range of $1 \leq R \leq 40$ length units. The stellar model was the same as described in Section \[sec:bulge\] and the orbits were evolved for $t=10\,000$ time units. It was found that $C_J$ is well conserved for the great majority of stars: only 0.2 to 1.6 per cent of orbits were cast off as [*rogue*]{} orbits, exceeding 35 steps per $\Delta t=0.1$ interval; the rest had an average $|\Delta C_J|$ of (4 to 11)$\times10^{-5}$ energy units ($C_J$ is typically of order unity). When applying a short softening length of 0.04 (corresponding to @qh97), 0.2 to 1.7 per cent of orbits were eliminated by the same criterion and the rest had an average $|\Delta C_J|$ of (5 to 17)$\times10^{-5}$ energy units. Thus, we did not apply softening in the actual simulations. This type of number of steps filter against rogue orbits was found to work better than putting a lower limit on the allowed step size.
Accuracy may be improved simply by integrating close encounters with smaller error tolerance, or by employing more elaborate techniques such as regularization. The small number of stars which are lost due to the ODE solver’s inability to handle them does not have a significant systematic effect on the BBH orbital evolution.
Conservation of $E$ and $L$
---------------------------
As noted previously, total energy and angular momentum conservation is not a built-in requirement of the method, but rather indicates a successful transfer of energy and angular from the BBH to the stellar population. In the upper panels of Figure \[fig:conservation\] we show energy of the BBH (solid black line), the stellar population (solid red line) and their sum (dashed blue line); the lower panels is the same but for the angular momenta. The left panel is for an equal mass simulation while the right panel is for a 10:1 mass ratio simulation.
The quality of energy conservation is attested by the absolute difference of total energy between the beginning and the end of the simulation; it is also useful to look at the fluctuations in total energy, a very crude estimate of which is the amplitude of the largest fluctuation (this is not a very good measure since fluctuation amplitude can be high in some parts of the orbit mild in others). These two quantities have dimensions of energy, and it is most sensible to normalize them with respect to the absolute difference in the energy of stars (or the BHs). Thus, for a simulation ending at $t=T$: $$\begin{aligned}
\epsilon_1^E(T) &= [E_{\rm tot}(T) - E_{\rm tot}(0)] / |E_\star(T) - E_\star(0)|,\\
\epsilon_2^E(T) &= {\rm max}(|E_{\rm tot}(t) - \langle E_{\rm tot} \rangle|) / |E_\star(T) - E_\star(0)|.\end{aligned}$$ We similarly define $\epsilon_1^L$ and $\epsilon_2^L$ for the angular momentum. All these $\epsilon$ parameters must be very small. If the quantity that is supposed to be conserved has some trend, then usually $\epsilon_1=\epsilon_2$, otherwise it only fluctuates around its mean value and $\epsilon_1<\epsilon_2$.
For an equal mass simulation with $N=5\times 10^6$ (discussed in detail in Section \[sec:N\]) case at $T=43\,200$, we get $\epsilon_1^E, \epsilon_1^L < 10^{-6}$ and $\epsilon_2^E = \epsilon_2^L = 3\times 10^{-4}$. For a 10:1 mass ratio simulation (discussed in detail in Section \[sec:e-growth\]) at $T = 21\,800$, we get $\epsilon_1^L=4\times 10^{-6}$ and $\epsilon_2^L=6\times 10^{-5}$, with a decreasing trend in total energy: $\epsilon_1^E = \epsilon_2^E=0.008$. The trend in total energy begins at $t\sim 14\,000$ and is probably due to the fact that there is an accumulated inaccuracy in the solution of the BHs’ equations when they are so close together at pericentre. At $t = 14\,000$ we get $\epsilon_1^E = 8\times 10^{-5}$ and $\epsilon_2^E = 4\times 10^{-4}$ with values for the angular momentum similar to the end of the simulation. Thus, the numerical effect previously discussed is unrelated the eccentricity growth observed in this simulation.
 
Results {#sec:results}
=======
We performed a total of 58 simulations of an equal mass binary and a 10:1 mass ratio binary, varying $N$, $R_0$, ${R_{\rm max}}$ and the initial eccentricity $e_0$. In Figure \[fig:examples\] we shows $r(t)$, for selected simulations. In the 10:1 simulations, the primary BH is fixed at the centre, so $r$ is the BBH separation; in the equal mass simulations $r$ is half the separation. As can be seen in the figure, the semi-major axis stalls in all simulations but eccentricity (indicated by the thickness of the lines) does not reach a steady state in the unequal mass cases. These results are discussed in more detail below.
![A representative sample of simulation results showing the radius as a function of time. The red and blue lines are for an initially circular and eccentric equal mass binaries, respectively; magenta and cyan are the same for a mass ratio of 10:1; the green line is the same as the red line but with a larger cutoff of the stellar bulge. The thickness of the line is indicative of eccentricity. While the semi-major axis stalls in all simulations, eccentricity does not reach a steady state in the unequal mass cases.[]{data-label="fig:examples"}](examples.eps)
Number of Particles {#sec:N}
-------------------
Here we show two things: how the results converge with increasing number of stars $N$, and how the results depend on the specific realization of the stellar distribution. We present seven pairs of simulations with $N$ between $50\,000$ and $350\,000$; for each $N$ the two simulations have a different random seed, so that the stars have different initial positions and velocities, but are drawn from the same distribution. All simulations are of an equal mass binary starting at $R_0=60$ with the local circular velocity; the cutoff radius of the stellar sphere is $R_{\rm max}=200$.
In Figure \[fig:N\] we show the ‘final state’ (i.e. at $t=10\,000$) semi-major axes $a$ (equivalent to separation, in the equal mass case), and eccentricities $e$ as functions of $N$; these two numbers are the best way to appreciate differences between similar simulations. From this small sample, it is apparent that the effect of changing $N$ in this range is comparable in magnitude to that of changing the realization. In this set of simulations, the range of semi-major axis values is 0.07 length units or $\sim 5$ per cent of the sample’s average $a$; the eccentricities are small and in the range $0.03 < e < 0.1$. Thus, increasing $N$ beyond $100\,000$ (within $R_{\rm max}=200$) is unnecessary for this level of accuracy and following tests are made using this number.
The average of $a$ in these simulations is 1.43 (in physical units, for $10^8 ~{\rm M}_\odot$ BHs, this is equivalent to 1.1 pc); this is approximately 30 per cent below the hard binary separation of 2. @merritt06 suggested the following formula for the stalling separation: $$\frac{a_{\rm stall}}{r_h^\prime} = 0.2 \frac{q}{(1+q)^2}\label{eq:stall},$$ where ${r_h^\prime}$ is the radius containing a mass in stars equal to twice the combined mass, or $M({r_h^\prime})=2M_\bullet(1+q)$, at the time of stalling. We measured the accumulated mass in the above simulations and got that in all of them ${r_h^\prime}=37.5$ with a very small spread. By substituting this information and $q=1$ into equation (\[eq:stall\]), one gets $a_{\rm stall}=1.88$. It is important to note that in the @merritt06 simulation there was no actual stalling of the BBH, and $a_{\rm stall}$ was estimated as the value of $a$ in which a clear change in the hardening rate took place; ${r_h^\prime}$ was determined at the time when this change occurred. In our simulations both ${r_h^\prime}$ and $a_{\rm stall}$ were determined at $t=10\,000$, which is well after the hardening rate has dropped. Also, equation (\[eq:stall\]) was calibrated by @merritt06 using simulations with mass ratios $0.025 \leq q \leq 0.5$.
To test the convergence of the results, we also performed a single simulation with significantly more particles ($N = 5\times 10^6$) and longer duration ($t = 43\,200$); we show the inverse semi-major axis as a function of time in Figure \[fig:long\]. As seen in the figure, there is still some slow evolution of the semi-major after $t=10\,000$. At $t=40\,000$ the value of $a$ is 1.35, which is 5.6 per cent lower than the value at $t=10\,000$ and only one per cent lower than the value at $t=30\,000$; the decay rate at the end of this simulation is $\dot{a}=10^{-6}$ velocity units, equivalent to $\sim 10^{-9}~{\rm pc\ yr^{-1}}$ for a $10^8~{\rm M}_\odot$ primary. In this specific run, eccentricity was especially low at $e<0.01$.
The orbits produced in the other simulations in this set (with $N\leq 3.5\times 10^5$) are very similar to the orbit shown in Fig. \[fig:long\], regardless of the number of particle (cf. @berczik06, spherical and triaxial cases). Note that a large $N$ is required only to minimize the statistical fluctuations, our scheme is not subject to an artificial stellar relaxation mechanism, which requires a very large $N$ to overcome.
![The semi-major axes $a$ and eccentricities $e$ at $t=10\,000$ as functions of the number of particles $N$, for equal mass and initially circular binaries. For each $N$, the two circles represent two different realizations of the same stellar population. The effect of changing $N$ in the range tested in this set is comparable in magnitude to that of changing the realization, and no trend is seen. The simple averages for the entire set are $a = 1.43 \pm 0.02$, corresponding to a stalling separation of $\sim 1.1$ pc for two $10^8~{\rm M_\odot}$ BHs, and $e = 0.06 \pm 0.02$ which indicates nearly circular orbits. An additional run with $N = 5\times 10^6$ is not shown here (see text).[]{data-label="fig:N"}](N.eps)
![The inverse semi-major axis of an equal mass binary simulation that was performed with $N = 5\times 10^6$ particles (compared with $N \leq 3.5\times 10^5$ for the simulations in Fig. \[fig:N\]) and for a much longer duration. The value of $a$ at $t=40\,000$ is 1.35, which is 5.6 per cent lower that the value at $t=10\,000$ and only one per cent lower that the value at $t=30\,000$. Cf. figure 1 of @berczik06[]{data-label="fig:long"}](long.eps)
***R***${}_{\bf max}$ and ***R***${}_{\bf 0}$ {#sec:Rmax}
---------------------------------------------
Since we only simulate the spherical component and not a full merger, the initial and boundary conditions need be assumed: the initial distance of the BHs from the centre, $R_0$, and the cutoff radius of the stellar sphere, ${R_{\rm max}}$. Here we present six trios of simulations with ${R_{\rm max}}$ between 120 and 520, each three simulations are of an equal mass binary starting at $R_0=30$, $60$ and $90$ with the local circular velocity. The number of particles is $N = 500 \times {R_{\rm max}}$, so as to keep the particle density profiles of equal normalizations (but different cutoff radii) in all simulations; this is due to the fact that the number of particles (or mass) grows linearly with radius in an isothermal sphere.
In Figure \[fig:R\] we show the final state semi-major axes $a$ (stalling separation) as a function of ${R_{\rm max}}$. From the results discussed in Section \[sec:N\], a characteristic error of $\sim 0.04$ units on the semi-major axis can be attributed to a specific realization and number-of-particles statistical fluctuation. The simulations with $R_0=30$ give systematically lower values for $a$; the values for $R_0=60$ do appear to be systematically lower than the $R_0=90$ simulations, but the two sets are within the errors of each other. There appears to be a trend of decreasing values of $a$ with increasing ${R_{\rm max}}$ up to $\sim 300$, which is to be expected both because there is a larger supply of particles that can interact with the BHs, and because the potential well is deeper and stars that have already interacted with the BHs have a larger probability to fall back to the centre and interact again. Nevertheless, we see from this small sample that these effects are weak and comparable in magnitude to those discussed in the Section \[sec:N\]. In this set of simulations, the eccentricities are also small and in the range $0.02 < e < 0.11$; no correlation of eccentricity was observed with either $R_0$ or ${R_{\rm max}}$.
![The semi-major axes $a$ at $t=10\,000$ as a function of the cutoff radius ${R_{\rm max}}$, for equal mass and initially circular binaries. For each of ${R_{\rm max}}$, the tree data points represent BHs launched from different radii $R_0$. The increase of $a$ with $R_0$ seems to saturate, while the dependence on ${R_{\rm max}}$ is very weak in the tested range. Eccentricity in this simulation was not correlated with either of the tested parameters, its simple average for the entire set is $e = 0.06 \pm 0.03$ which indicates nearly circular orbits.[]{data-label="fig:R"}](RmaxR0.eps)
Mass Ratios {#sec:mass-ratios}
-----------
As noted in section \[sec:alg-ratio\], the simulation of an unequal mass binary is somewhat different in nature. Thus, when studying the evolution of a 10:1 binary, we followed the analysis of Section \[sec:N\] and performed a number of different tests with increasing number of particles and different realizations. Here we present four pairs of simulations with $N$ between $100\,000$ and $400\,000$; for each $N$ the two simulations have a different random seed, so that the stars have different initial positions and velocities, but are drawn from the same distribution. All simulations are of a binary with a 10:1 mass ratio starting at $R_0 = 60$ with the local circular velocity; the cutoff radius of the stellar sphere is ${R_{\rm max}}= 200$.
In Figure \[fig:M\] we show the final state semi-major axes $a$ and eccentricities $e$ as functions of $N$. As in Section \[sec:N\], changing $N$ in this range produces no apparent effect on the stalling radius. Within this set of simulations, the semi-major axes are consistent; the sample’s average $a$ is $0.241 \pm 0.005$. We measured the accumulated mass in the above simulations and got that in this case ${r_h^\prime}=16.8$ with a very small spread. By substituting this information and $q=0.1$ into equation (\[eq:stall\]), one gets $a_{\rm stall}=0.28$. The caveats of using this equation were explained in Section \[sec:N\]. If we recalibrate equation (\[eq:stall\]) using our two values of $q$, the prefactor is lowered from 0.2 to approximately 0.16; it is even somewhat smaller considering the fact that the true value of $a_{\rm stall}$ can be $\sim 6$ per cent lower than its measured value at $t=10\,000$ (see Figure \[fig:long\]).
In contrast with the equal mass case, the eccentricities do not reach a steady state value. In one of the $N=10^5$ simulations, the eccentricity increases very rapidly after the binary becomes hard. At $t=9\,400$ the pericentre distance reached $10^{-3}$ and the simulation is terminated. In the rest of the simulations, The eccentricity values range between 0.05 and 0.28 at $t=10\,000$, however, in all but one of these simulations eccentricity is slowly increasing. We calculate $T_{e\rightarrow 1}$, a very rough estimation for the time of gravitational wave regime, by fitting $e(t)$ with a linear function in the range $6\,000<t<10\,000$ and continue it to $e=1$. At this time span the rise in eccentricity is approximately linear, however this trend breaks at approximately $e=0.85$, so in fact $T_{e\rightarrow 1}$ can underestimate the time to the gravitational wave regime by some $\approx 20$ per cent. In Table \[tab:e\] we show the results for $T_{e\rightarrow 1}$ for this simulation set and also for the initially eccentric runs. The values range between approximately $4\times10^4$ and $3\times10^5$ time units, or on the order of $10^8$ years for a $10^7~{\rm M_\odot}$ secondary around a $10^8~{\rm M_\odot}$ primary.
![Same as Fig. \[fig:N\] but for a 10:1 binary. The semi-major axes are consistent within this sample, the simple average is $a = 0.241 \pm 0.005$, corresponding to a stalling separation of $\sim 0.19$ pc for for a $10^7~{\rm M_\odot}$ secondary around a $10^8~{\rm M_\odot}$ primary. Eccentricity varies greatly in this sample, see text.[]{data-label="fig:M"}](N_MR.eps)
$N$ \[$10^5$\] $e_0$ $e_{\rm f}$ $T_{e\rightarrow 1}$
---------------- ------- ------------- ----------------------
1 0 0.05 $2.7\times 10^5$
1 0 $\sim 1$ $\lesssim 10^4$
2 0 0.21 $4.8\times 10^4$
2 0 0.13 $2.2\times 10^5$
3 0 0.15 $1.4\times 10^5$
3 0 0.18 $5.9\times 10^4$
4 0 0.05 $4.4\times 10^7$
4 0 0.28 $3.5\times 10^4$
2 0.1 0.42 $2.0\times 10^4$
2 0.1 0.33 $2.5\times 10^4$
2 0.2 0.47 $2.1\times 10^4$
2 0.2 0.59 $1.7\times 10^4$
50 0.2 0.52 $1.8\times 10^4$
2 0.3 0.21 $9.2\times 10^4$
2 0.3 $\sim 1$ $\lesssim 10^4$
2 0.4 0.96 $\approx 10^4$
2 0.4 $\sim 1$ $\lesssim 10^4$
: A list of all 10:1 binary simulations. The columns from left to right are: the number of particles, initial eccentricity, eccentricity at $t=10\,000$ and the estimated timescale to reach $e=1$ (see text for definition).[]{data-label="tab:e"}
Eccentricity growth {#sec:e-growth}
-------------------
Here we tested how the initial eccentricity affects the results. We performed simulations with four different initial eccentricities between 0.1 and 0.4 with semi-major axes as in the circular simulations of Sections \[sec:N\] and \[sec:mass-ratios\]. The initial eccentricity $e_0$ is that of a BBH with the same initial conditions and that moves in the initial potential (equation \[eq:bulge-potential\]) with no friction; the orbit is not really an ellipse since the potential in not Keplerian, so $e_0$ corresponds to the mean orbital eccentricity, defined as the difference between the maximum and minimum separations divided by the major axis. For each $e_0$, there are two simulations (with two different realizations) for an equal mass binary and two for a 10:1 binary. The number of particles is $N = 2\times 10^5$ with cutoff radius ${R_{\rm max}}=200$; we performed a single simulation with significantly more particles ($N = 5\times 10^6$) for the $e_0=0.2$ case.
In Figure \[fig:E0\] we show the final state semi-major axes $a$ and eccentricities $e$ as functions of $e_0$. This figure also includes four data points with $e_0=0$ that have already been presented in Figs. \[fig:N\] & \[fig:M\]. The stalling separation is independent of initial eccentricity in the tested range, the sample’s average $a$ is $1.42 \pm 0.04$ for the equal mass simulations and $0.246 \pm 0.010$ for the unequal mass simulations. As in the other tests we performed for equal mass binaries, the final orbits are very much circular; the eccentricities are in the range $0.01 < e < 0.1$ despite the initially significant eccentricity. However, the final eccentricities in the 10:1 cases do appear to be generally correlated with $e_0$. As in Section \[sec:mass-ratios\], here too, eccentricities are still increasing when the simulations terminate at $t=10\,000$; the eccentricity timescales $T_{e\rightarrow 1}$ for the 10:1 simulations are shown in Table \[tab:e\], which shows that the eccentricity growth rate is related to $e_0$.
![The semi-major axes $a$ and eccentricities $e$ at $t=10\,000$ as functions of initial eccentricity $e_0$, for equal mass (blue circles) and 10:1 binaries (red squares). For each $e_0$, the two data points of each kind represent two different realizations of the same stellar population (all simulations have discrete values of $e_0$, but some overlapping data points were moved slightly to the left and right for graphical reasons). In the 10:1 case, the final eccentricities increase with $e_0$ while the semi-major axes are unaffected. The equal mass binaries seem to circularize independently of their initial eccentricities, as all results are very similar in both $a$ and $e$. The red stars represent a single 10:1 simulation with $N=5\times 10^6$; all others shown here have $N=2\times 10^5$.[]{data-label="fig:E0"}](InitialEccentricities.eps)
@sesana10 found that eccentricity growth is generally mild for equal mass binaries with very small initial eccentricities, but also that initial eccentricity $e_0 > 0.3$ leads to very high peak eccentricity almost regardless of the other system parameters. Although all our equal mass simulations end up in nearly circular orbit, this is not inconsistent with @sesana10: while in the latter work the BHs are launched from within their radius of influence, in our simulations the BHs are launched from much further out, in the dynamical friction regime. In our simulations, an equal mass binary becomes less eccentric as it inspirals from $r>60$. For one of the $e_0=0.4$ simulations, the rapid eccentricity decrease ceased at about $r=1.4$ (approximately twice the stalling radius) where the value was $e=0.035$. At $r=4$ (equivalent to the initial radius of @sesana10’s equal mass binaries) the eccentricity was 0.12; no significant eccentricity growth occurred in @sesana10’s equal mass simulation with initial eccentricity of 0.1. While it has been shown by @dotti06 that BBHs lose memory of their initial eccentricity if they corotate with a massive gaseous disc, studies of eccentric orbits of hard binaries is motivated for the purely stellar dynamical case by the theory of linear response for dynamical friction [@colpi99]. However, this theory is derived from a first order perturbative expansion and is not applicable when close encounters dominate the evolution, and the system is not well described by an analytical approximation.
It is important to note another major difference between our work and @sesana10 which greatly affects the evolution of the binary orbital parameters, mostly the semi-major axis: while in our work the loss cone empties, @sesana10 implicitly assumes that the loss cone is always full at $r > r_{\rm inf}$. This leads to a very rapid decay of the binary separation and quick coalescence due to gravitational wave emission. Similarly, @antonini12 calculated the eccentricity evolution in the case of a very small secondary BH that does not affect the stellar density profile. There, eccentricity grows because the orbit passes in and out of a flat core, where the star are fast and the drag force is much less efficient at pericentre than at apocentre. In our simulations, however, the much more massive secondary forms a cavity slightly larger than its apocentre, and eccentricity grows where the density is essentially zero.
In Figure \[fig:e-evolution\] we show the evolution of the semi-major axis for all the runs with $e_0=0.2$ and 10:1 mass ratio. The dotted green lines represent the two realizations with $2\times 10^5$ particles while the solid blue line represents the larger $N=5\times 10^6$ realization. The first two simulations are arbitrarily terminated at $t=10\,000$ while the latter is stopped only when the eccentricity reaches 0.99. If we scale to physical units for a primary BH of $10^8~{\rm M_\odot}$, the end of the simulation is 22 Myr from its beginning. For this mass scaling, using the @peters64 formula for orbital decay due to gravitational waves, the timescale for coalescence at the end of the simulation is just 1 Myr.
The rapid growth of the eccentricity while the semi-major axis remains fairly constant indicates a high value for the the dimensionless eccentricity growth, defined as: $$K = \frac{{\rm d}e}{{\rm d}\ln (1/a)}.$$ For single scattering of unbound stars from a fixed background, @quinlan96 derived a maximal value of $K$ of about 0.3 for mass ratio of 16:1, consistent with previous scattering experiments (@roos81; @mikkola92). The value of $K$ for the simulation presented in Fig. \[fig:e-evolution\] is at least an order of magnitude larger (we only roughly estimated the value from the results). The difference is probably due to the very different nature of the orbits in the restricted 3-body problem versus the more realistic model used here. We refer an in-depth study of the physical mechanism behind the eccentricity growth to a subsequent work.
It is interesting to compare our results to those of @iwasawa11, who also got ‘runaway’ growth of eccentricity while the semi-major axis stalled, but one must note the critical differences between the two studies. Most notably, @iwasawa11 used a mass ratio of 100:1 while we took only 10:1. Additionally, they used an initially very shallow central density profile, $\rho \propto r^{-3/4}$ while our bulge model (see Section \[sec:bulge\]) was an isothermal sphere, $\rho \propto r^{-2}$. More importantly, the mass of their entire stellar population was just $1.25\times 10^9~{\rm M_\odot}$, which is 8 times less than the primary BH’s mass and only 12.5 times more than the secondary’s mass. By comparison, the total stellar mass in our model is 25 times the mass of the primary and 250 times the mass of the secondary. Thus, their entire simulation is deep within the primary’s sphere of influence, where its gravity dominates, while our simulations started with the secondary well outside the primary’s radius of influence.
Let us scale our work to theirs by setting $M_8 = 100$ in the scaling equations of Section \[sec:units\]. @iwasawa11 start their [A32k]{} simulations (with $N=32\,768$) at $R_0 = 20~{\rm pc}$ within which there are less than $2\,500$ particles, their BH stalls at $a = 3.9~{\rm pc}$. We initiate the secondary BH at $r=637~{\rm pc}$ and get stalling at 2.1 pc; there are initially $40\,000$ particles enclosed within $r = 20~{\rm pc}$ in our large $N=5\times 10^6$ simulation (marked with a star in Fig. \[fig:E0\]). The secondary BH is $20\,000$ times more massive than a field star in our simulation, versus $2\,600$ in the @iwasawa11 simulation. Their stellar bulge model is very small compared to ours, with 90 per cent of their mass is within $R_0 = 190~{\rm pc}$, while our model is truncated at ${R_{\rm max}}= 1\,800~{\rm pc}$.
The semi-major axes in the two studies evolve at very different rates: while the @iwasawa11 binary takes 19 million years (Myr) to sink from $a_0 = 20~{\rm pc}$ to 10 pc in their shallow cusp, our binary does the same journey in only 0.18 Myr. However, when applying equation (6) of @iwasawa11 (derive from their numerical results) for our physical parameters, the timescale for significant eccentricity growth is extremely short at 0.78 Myr; the lifetime of the system from our simulation is approximately 85 Myr (scaled with $M_8 = 100$), which is two orders of magnitude longer. The difference might be due to the mass scaling assumed in their formula.
A final note about precession: in the unequal mass case, the BHs exhibit very small precession during the hard phase. In the large simulation of Fig. \[fig:e-evolution\], between $t_1=10\,000$ and $t_2=11\,000$ the semi-major axis precesses by $0.311^\circ$. During this period of 1 Myr (scaled with $M_8 = 1$), the average semi-major axis is 0.18 pc, and it drops by 0.002 pc; the average eccentricity is 0.559 and the growth is by 0.072. This precession can be produced, for example, by perturbing the Keplerian potential with a uniform density field of 20,000 solar masses per parsec cubed. Even with our large number of particles, this density corresponds to only four particles enclosed in the sphere with radius equals to the apocentre. In the snapshot taken at $t_1$, there was one particle inside this region, and it was probably a transient since there are no stable orbits there except those tightly bound to one of the BHs. In principle, a small flux of particles to this region can produce the measured precession, but torques due to the anisotropy of the potential at larger distances are more likely to cause the precession. For comparison, general relativistic precession of the orbit (not reproduced in the simulation) would be $56^\circ~\rm{Myr}^{-1}$. We cannot yet say whether this precession compromises resonances that possibly induce the eccentricity growth, but will refer to this point in a future paper.
![The semi-major axes $a$ and eccentricities $e$ at as functions of time, for three 10:1 binary simulations with initial eccentricity of 0.2. The dotted green lines are two different realizations with $2\times 10^5$ particles while the solid blue line is a run with $5\times 10^6$ particles. The first two simulations are arbitrarily terminated at $t=10\,000$ while the last is stopped only when the eccentricity reaches 0.99, this time for a $10^8~{\rm M_\odot}$ primary is 22 Myr from the beginning of the simulation. For this mass scaling, the timescale for coalescence due to gravitational waves at the end of the simulation is just 1 Myr.[]{data-label="fig:e-evolution"}](e.eps)
Dynamical friction {#sec:df}
==================
The orbital decay of a massive object within a galaxy down to its centre is well approximated by Chandrasekhar’s dynamical friction formula [@chandra]. However, the assumption of a uniform background in his classical treatment does not hold in real galaxies. Thus, corrections to the Coulomb logarithm are necessary to account for the changing background with radius. @just05 performed a detailed theoretical investigation of Chandrasekhar’s formula in the presence of a density gradient and gave an improved analytical formula for the Coulomb logarithm. @just11 took into account also self consistent velocity distribution functions, and made a comprehensive examination of the applicability of the new formula to sinking massive objects in a number of galaxy models, using high resolution $N$-body and particle-mesh codes. Their results suggest a delay in the orbital decay with respect to the standard formula, which quantitatively varies according to the studied case.
They notably give an explicit solution for the decay of a massive object moving on a ‘circular’ orbit. Their formula (equation 25 in their paper) is very general and holds for an arbitrary power law density profile (it does not hold for very flat cores where fast moving stars contribute to most of the frictional drag; see @antonini12). Here we bring their formula in our model’s units and adjusted the parameters for an isothermal sphere; the radial evolution is given implicitly by: $$\textstyle t = 52.918 \times \left\{{\rm Ei}\left[2\ln\left(\frac{1}{8}R_0\right)\right]-{\rm Ei}\left[2\ln\left(\frac{1}{8}r\right)\right]\right\}\label{eq:Just}$$ The prefactor is an exactly calculable number. The special function ${\rm Ei}(x)$ is called the exponential integral, it has real values only for $x>0$. Thus, the domain of definition of equation (\[eq:Just\]) is $r>8$. However, the assumption of a nearly circular orbit breaks well above $r=8$. The angle between the velocity vector and the tangent can be derived by finding the radial velocity from equation (\[eq:Just\]): $$\theta = 195.99^\circ \times \frac{1}{r}\ln\left({\textstyle\frac{1}{8}}r\right),\label{eq:spiral-angle}$$ where this approximation hold only for small angles or large $r$.
It is generally difficult to simulate a full infall of a compact object into a galaxy centre due to the collisional nature of the interactions and the low number density of particles which can generally be obtained at the outskirts of galaxy models. @just11 used the particle-mesh code [Superbox]{} [@superbox], which is collisionless and uses fixed time steps; This type of code, unlike direct $N$-body codes, allows a large particle number to be simulated in a relatively short time. Using the code we designed for BBHs, we can also study the early part of the inspiral, which is dominated by dynamical friction. Since our code resolves collisions between the BHs and stars, this study is complimentary to that of Just et al..
In Figure \[fig:df\] we show a simulation of an equal mass binary initiated from $R_0=500$ (solid blue line) and the theoretical curve (dashed black line), equation (\[eq:Just\]). The initial velocity is the local circular velocity, but the initial velocity vector is tilted by $1.62^\circ$; this angle is obtained be substituting $r=R_0$ in equation (\[eq:spiral-angle\]). If the initial velocity is purely tangential, then the spiral becomes slightly ‘eccentric’. A second run with a different realization (not shown) gave very similar results, including the position of the wiggles. It should be pointed out that this simulation is just a proof of concept; we do not expect the stellar distribution to be spherically symmetric at the early stages of a major merger.
The orbital decay in our simulation is initially very well described by equation (\[eq:Just\]); deviations become significant below $r \sim 300$, where mutual gravity of the two BHs is still negligible compared to the gravity of the bulge. These deviations are possibly due to the fact that the phase space distribution in the actual simulation (at the time and radius where the deviations occur) is no longer consistent with the assumption of isothermal sphere used to derive equation (\[eq:Just\]), in particular the velocity distribution might not be described well by a Gaussian.
![This simulation of an equal mass binary initiated from $R_0=500$ (solid blue line) focuses on the dynamical friction regime. The theoretical curve (dashed black line) is the analytical formula of @just11 given by equation (\[eq:Just\]). The orbital decay is initially very well described by the formula, and the deviations at $r\lesssim 300$ are possibly due to inconsistency of the phase space distribution in the actual simulation versus the assumption of isothermal sphere used to derive the formula[]{data-label="fig:df"}](DF.eps)
Summary {#sec:summary}
=======
Using a conservation-based scheme, we were able to follow the evolution of a BBH from a wide separation (enclosed stellar mass greater than the combined BH mass) down to sub-parsec scale. Our code resolves star-BH collisions and can run with $N>10^6$ stars on a desktop computer. We verified that our scheme yields convergent results which are independent of the number of particles, and the initial and boundary conditions. By performing scattering experiments on the inspiraling BBH, we will be able to extend @meiron10 and calculate the signature of the inspiral on the background stellar phase space distribution as a function of projected position. This calculation improves on $N$-body simulation by reducing statistical fluctuations and having no spurious relaxation (and thus no loss cone refilling). We performed calculations for both an equal mass binary and a 10:1 mass ratio. Our calculations reveal:
1. The inspiral from a radius scale of tens of parsecs to the hard binary radius occurs on a time scale of a few million years for a $10^8 ~{\rm M}_\odot$ primary, with only a weak dependence of the timescale on the mass ($\propto M_8^{0.29}$, equation \[eq:units-T\]).
2. The inspiral ends at a radius which is $\sim 30$ per cent smaller than the simple analytical estimate for the hard binary radius, and consistent with @merritt06.
3. An equal mass binary inspiral leads to a nearly circular final orbit, regardless of the initial eccentricity.
4. Eccentricity increases and coalescence due to gravitational wave emission will occur for a binary with a mass ratio of 10:1 in less than $10^8$ years ($\times M_8^{0.29}$). If the stellar distribution is triaxial or rotating the lifetime of such systems is potentially shorter.
While we used a static, spherically symmetric background potential to account for star-star interactions, it is straightforward to extended this method to treat more complicated cases such as an adiabatically evolving potential (e.g. due to core depletion during the BBH inspiral), non symmetric stellar models and perturbers within the scattering method.
This method can also be used to explore the time evolution of the statistical properties of the scattering events, and the extension of the dynamical friction formulation to the hard binary stage. This study will help understand the mechanism which leads to the growth of eccentricity for an $q \neq 1$ BBH, and its decay for the $q=1$ case.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank David Merritt, Eugene Vasiliev, Alberto Sesana, Fazeel Mahmood Khan and Fabio Antonini for helpful discussions and comments. We thank the referee for some helpful comments.
Antonini, F., & Merritt, D. 2012, [ApJ](http://dx.doi.org/10.1088/0004-637X/745/1/83), [ 745, 83](http://adsabs.harvard.edu/abs/2012ApJ...745...83A)
Bahcall, J. N., & Wolf, R. A. 1976, [ApJ](http://dx.doi.org/10.1086/154711), [ 209, 214](http://adsabs.harvard.edu/abs/1976ApJ...209..214B)
Bahcall, J. N., & Wolf, R. A. 1977, [ApJ](http://dx.doi.org/10.1086/155534), [ 216, 883](http://adsabs.harvard.edu/abs/1977ApJ...216..883B)
Binney, J., & Tremaine, S. 2008, Galactic Dynamics: Second Edition, by James Binney and Scott Tremaine. ISBN 978-0-691-13026-2 (HB).
Begelman, M. C., Blandford, R. D., & Rees, M. J. 1980, [Nature](http://dx.doi.org/10.1038/287307a0), [ 287, 307](http://adsabs.harvard.edu/abs/1980Natur.287..307B)
Berczik, P., Merritt, D., & Spurzem, R. 2005, [ApJ](http://dx.doi.org/10.1086/491598), [ 633, 680](http://adsabs.harvard.edu/abs/2005ApJ...633..680B)
Berczik, P., Merritt, D., Spurzem, R., & Bischof, H.-P. 2006, [ApJ](http://dx.doi.org/10.1086/504426), [ 642, L21](http://adsabs.harvard.edu/abs/2006ApJ...642L..21B)
Berentzen, I., Preto, M., Berczik, P., Merritt, D., & Spurzem, R. 2009, [ApJ](http://dx.doi.org/10.1088/0004-637X/695/1/455), [ 695, 455](http://adsabs.harvard.edu/abs/2009ApJ...695..455B)
Chandrasekhar, S. 1943, [ApJ](http://dx.doi.org/10.1086/144517), [ 97, 255](http://adsabs.harvard.edu/abs/1943ApJ....97..255C)
Colpi, M., Mayer, L., & Governato, F. 1999, [ApJ](http://dx.doi.org/10.1086/307952), [ 525, 720](http://adsabs.harvard.edu/abs/1999ApJ...525..720C)
Cuadra, J., Armitage, P. J., Alexander, R. D., & Begelman, M. C. 2009, [MNRAS](http://dx.doi.org/10.1111/j.1365-2966.2008.14147.x), [ 393, 1423](http://adsabs.harvard.edu/abs/2009MNRAS.393.1423C)
Dotti, M., Colpi, M., & Haardt, F. 2006, [MNRAS](http://dx.doi.org/10.1111/j.1365-2966.2005.09956.x), [ 367, 103](http://adsabs.harvard.edu/abs/2006MNRAS.367..103D)
Fellhauer, M., Kroupa, P., Baumgardt, H., Bien, R., Boily, C. M., Spurzem, R., & Wassmer, N. 2000, [New Astron.](http://dx.doi.org/10.1016/S1384-1076(00)00032-4), [ 5, 305](http://adsabs.harvard.edu/abs/2000NewA....5..305F)
Genzel, R., Sch[ö]{}del, R., Ott, T., et al. 2003, [ApJ](http://dx.doi.org/10.1086/377127), [ 594, 812](http://adsabs.harvard.edu/abs/2003ApJ...594..812G)
G[ü]{}ltekin, K., et al. 2009, [ApJ](http://dx.doi.org/10.1088/0004-637X/698/1/198), [ 698, 198](http://adsabs.harvard.edu/abs/2009ApJ...698..198G)
Hopman, C., & Alexander, T. 2006, [ApJ](http://dx.doi.org/10.1086/506273), [ 645, L133](http://adsabs.harvard.edu/abs/2006ApJ...645L.133H)
Iwasawa, M., An, S., Matsubayashi, T., Funato, Y., & Makino, J. 2011, [ApJ](http://dx.doi.org/10.1088/2041-8205/731/1/L9), [ 731, L9](http://adsabs.harvard.edu/abs/2011ApJ...731L...9I)
Just, A., & Pe[ñ]{}arrubia, J. 2005, [A&A](http://dx.doi.org/10.1051/0004-6361:20041175), [ 431, 861](http://adsabs.harvard.edu/abs/2005A%26A...431..861J)
Just, A., Khan, F. M., Berczik, P., Ernst, A., & Spurzem, R. 2011, [MNRAS](http://dx.doi.org/10.1111/j.1365-2966.2010.17711.x), [ 411, 653](http://adsabs.harvard.edu/abs/2011MNRAS.411..653J)
Khan, F. M., Just, A., & Merritt, D. 2011, [ApJ](http://dx.doi.org/10.1088/0004-637X/732/2/89), [ 732, 89](http://adsabs.harvard.edu/abs/2011ApJ...732...89K)
Koopmans, L. V. E., et al. 2009, [ApJ](http://dx.doi.org/10.1088/0004-637X/703/1/L51), [ 703, L51](http://adsabs.harvard.edu/abs/2009ApJ...703L..51K)
Lee, M. H., & Goodman, J. 1989, [ApJ](http://dx.doi.org/10.1086/167732), [ 343, 594](http://adsabs.harvard.edu/abs/1989ApJ...343..594L)
Lynden-Bell, D. 1967, MNRAS, [ 136, 101](http://adsabs.harvard.edu/abs/1967MNRAS.136..101L)
Makino, J. 1997, [ApJ](http://dx.doi.org/10.1086/303773), [ 478, 58](http://adsabs.harvard.edu/abs/1997ApJ...478...58M)
Makino, J., & Funato, Y. 2004, [ApJ](http://dx.doi.org/10.1086/380917), [ 602, 93](http://adsabs.harvard.edu/abs/2004ApJ...602...93M)
Meiron, Y., & Laor, A. 2010, [MNRAS](http://dx.doi.org/10.1111/j.1365-2966.2010.17031.x), [ 407, 1497](http://adsabs.harvard.edu/abs/2010MNRAS.407.1497M)
Merritt, D. 2006, [ApJ](http://dx.doi.org/10.1086/506139), [ 648, 976](http://adsabs.harvard.edu/abs/2006ApJ...648..976M)
Mikkola, S., & Valtonen, M. J. 1992, MNRAS, [ 259, 115](http://adsabs.harvard.edu/abs/1992MNRAS.259..115M)
Milosavljevi[ć]{}, M., & Merritt, D. 2001, [ApJ](http://dx.doi.org/10.1086/323830), [ 563, 34](http://adsabs.harvard.edu/abs/2001ApJ...563...34M)
Milosavljevi[ć]{}, M., & Merritt, D. 2003, [ApJ](http://dx.doi.org/10.1086/378086), [ 596, 860](http://adsabs.harvard.edu/abs/2003ApJ...596..860M)
Perets, H. B., & Alexander, T. 2008, [ApJ](http://dx.doi.org/10.1086/527525), [ 677, 146](http://adsabs.harvard.edu/abs/2008ApJ...677..146P)
Peters, P. C. 1964, [Phys. Rev. B](http://dx.doi.org/10.1103/PhysRev.136.B1224), [ 136, 1224](http://adsabs.harvard.edu/abs/1964PhRv..136.1224P)
Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. 1992, Numerical recipes in C. The art of scientific computing (2nd ed.; Cambridge: Cambridge Univ. Press)
Quinlan, G. D. 1996, [New Astron.](http://dx.doi.org/10.1016/S1384-1076(96)00003-6), [ 1, 35](http://adsabs.harvard.edu/abs/1996NewA....1...35Q)
Quinlan, G. D., & Hernquist, L. 1997, [New Astron.](http://dx.doi.org/10.1016/S1384-1076(97)00039-0), [ 2, 533](http://adsabs.harvard.edu/abs/1997NewA....2..533Q)
Roos, N. 1981, A&A, [ 104, 218](http://adsabs.harvard.edu/abs/1981A%26A...104..218R)
Sesana, A. 2010, [ApJ](http://dx.doi.org/10.1088/0004-637X/719/1/851), [ 719, 851](http://adsabs.harvard.edu/abs/2010ApJ...719..851S)
Sesana, A., Gualandris, A., & Dotti, M. 2011, [MNRAS](http://dx.doi.org/10.1111/j.1745-3933.2011.01073.x), [ 415, L35](http://adsabs.harvard.edu/abs/2011MNRAS.415L..35S)
Syer, D., & White, S. D. M. 1998, [MNRAS](http://dx.doi.org/10.1046/j.1365-8711.1998.01285.x), [ 293, 337](http://adsabs.harvard.edu/abs/1998MNRAS.293..337S)
Young, P. 1980, [ApJ](http://dx.doi.org/10.1086/158553), [ 242, 1232](http://adsabs.harvard.edu/abs/1980ApJ...242.1232Y)
Yu, Q. 2002, [MNRAS](http://dx.doi.org/10.1046/j.1365-8711.2002.05242.x), [ 331, 935](http://adsabs.harvard.edu/abs/2002MNRAS.331..935Y)
\[lastpage\]
[^1]: E-mail: <[email protected]> (YM); <[email protected]> (AL)
[^2]: The word ‘advance’ in this context means solve the equation of motion by means of Runge–Kutta method of order five with adaptive step size control (e.g. @nr).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The original canonical coherent states could be defined in several ways. As applications for other sets of coherent states arose, the rules of definition were correspondingly changed. Among such rule changes were a change of group and relaxation of the analytic nature of the labels. Recent developments have done away with the group connections altogether and thereby allowed sets of coherent states to be defined that are temporally stable for a wide variety of dynamical systems including the hydrogen atom. This article outlines some of the current trends in the definitions and properties of present-day coherent states.'
author:
- |
John R. Klauder\
Departments of Physics and Mathematics\
University of Florida\
Gainesville, FL 32611
title: 'The Current State of Coherent States[^1]'
---
[H]{} ø { }
Introduction {#introduction .unnumbered}
============
The modern reincarnation of what are now often called canonical coherent states began in 1960 [@kla2] (with a mathematical-physics application to define coherent state path integrals), in 1961 [@barg] (with a thorough mathematical study), and in 1963 [@gla] (with a physics application central to the new theory of quantum optics). Over the years, generalizations of the original family of canonical coherent states have been introduced based largely on mathematical or possibly mathematical-physics grounds. These generalizations have frequently involved one or another of the mathematical properties of the canonical coherent states and its elevation to the central concept in defining new sets of coherent states. As examples, we cite group-defined coherent states [@kla19; @per; @gil], annihilation-operator-eigenstate defined coherent states [@bar], and minimum-uncertainty-state defined coherent states [@nie]. Such generalizations typically lead to new sets of coherent states alright, but (apart perhaps from the group-defined coherent states) such rules for generating new sets of coherent states have always seemed to the present author to be overly mathematical and rather divorced from any specific physics. After all, what is the physics involved in choosing annihilation-operator eigenstates or in choosing minimum uncertainty states? What would be so wrong in choosing states for which the minimum uncertainty product was exceeded by a factor of three, for example?
These views have in recent years prompted the author to seek other generalizations of the canonical coherent states often with specific physical criteria chosen as the key factor involved in defining and obtaining such generalizations. Although other prescriptions exist, we shall, in the interests of brevity and consistency, pursue just one path among many in our discussion of new sets of coherent states.
A few introductory remarks are useful: For convenience, we denote each of the coherent states by $|l\>\in{\frak{H}}$, $|l\>\ne0$, where $l=(l^1,l^2,\ldots,l^L)\in{\cal L}$, $l^j\in{\mathbb R}$, denotes an $L$-dimensional (real) label lying in a label space $\cal L$ which locally is topologically equivalent to ${\mathbb R}^L$. This latter property means that we can identify continuous functions on $\cal L$. It is often useful to regard $l$ as a [*classical*]{} variable in a classical (phase) space $\cal L$. Although we shall not generally do so, it is often useful to group some or all of the real parameters by pairs and to form complex parameters. Throughout, we choose units so that $\hbar=1$.
With these remarks as background, we start with what we regard as the basic minimum properties for any set of states to be called a set of coherent states: .4cm
[**1. Continuity of Labeling:**]{} The map from the label space $\cal L$ into the Hilbert space $\frak H$ is strongly continuous.
[**Comment:**]{} Specifically, this condition requires that the expression $\| |l'\>-|l\>\|\ra0$ whenever $\l'\ra l$ in $\cal L$. This condition is equivalent to the joint continuity of the coherent state overlap function, $\<l''|l'\>$, in its two arguments. .4cm
[**2. Resolution of Unity:**]{} A positive measure $\mu(l)$ on $\cal L$ exists such that the unit operator $\one$ admits the representation $$\one = \int_{\cal L}|l\>\<l|\,d\mu(l)\;,$$ where $|l\>\<l|$ denotes the rank-one operator that takes an arbitrary vector $|\psi\>$ into a multiple (namely $\<l|\psi\>$) of the vector $|l\>$.
[**Comment:**]{} If $|l\>=0$ for some $l$, these vectors would make no contribution to the resolution of unity, and so we have already assumed that $|l\>\ne0$, i.e., $\<l|l\>>0$. If $d\mu(l)=0$ for a set of nonzero measure, then these vectors would also not contribute to the resolution of unity. Hence, there is no loss of generality to require that $\mu(l)$ is a strictly positive measure (up to sets of measure zero). In addition, it is often useful to assume that $\mu(l)$ is scaled (or rescaled, if necessary) so that $\<l|l\>=1$ for all $l\in{\cal L}$. If $\<l|l\>=1$, then it follows that $|l\>\<l|$ is a one-dimensional projection operator. (Ideally, $\mu(l)$ should be a countably additive measure, but a finitely additive measure is generally sufficient, which is a distinction for positive measures that may arise when an infinite number of degrees of freedom are involved.)
[**Remark:**]{} The two postulates about coherent states above were proposed in substantially this form nearly forty years ago [@kla18], even before such states were called “coherent states”. With very few exceptions, all states that have been so named have fulfilled these two postulates and for purposes of the present article we shall require that these two postulates hold. \[For a recent study of a case where a resolution of unity (Postulate 2) fails to hold, see [@kla222].\]
The generality of the first two postulates, and their mathematical specificity as well, has been done deliberately so that a vast catalog of sets of coherent states implicitly exists; ideally, it is the analysis of a specific physical problem which, whenever possible, puts on additional physical restrictions that singles out a subset of coherent-state sets—or even a single coherent-state set—tied to the specified physical problem.
A useful analogy to the present point of view lies in the mathematical concept of a set of [*orthonormal functions*]{}. Initially, one can define the properties that make a set of functions an acceptable set of orthonormal functions, i.e., completeness, orthogonality, and normalization. Finally, one can introduce criteria to select some sets or even one set of orthonormal functions relevant to some specific physical problem.
[**Remark:**]{} The philosophy of defining coherent states expressed here is, of course, just one of many possible choices. Others are free to choose alternative definitions, although it naturally diminishes the utility of the phrase when it is used too widely. The ultimate value of a definite rule of definition stems from its [*usefulness in applications*]{}$\s$; and applications generally arise for specific and concrete systems.
We now turn our attention to picking out suitable sets or even a single set of coherent states by adopting certain physical criteria rather than imposing selected mathematical requirements as discussed in the previous section.
Temporal Stability {#temporal-stability .unnumbered}
==================
For our first additional property we shall study time evolution as dictated by a specific Hamiltonian operator $\H$. The evolution of any coherent state $|l\>$ may always be captured by the relation $$e^{-i\H t}\,|l\>\equiv |l,t\>\;,$$ a definition that imposes no restriction whatsoever. However, we can ask for much more. Let us first restrict attention to normalized coherent states, $\<l|l\>=1$, for all $l\in{\cal L}$. Then we may ask that the following condition holds: .4cm
[**3. Temporal Stability:**]{} The time evolution of any coherent state always remains a coherent state. In symbols, $$e^{-i\H t}\,|l\>=|l(t)\>$$ for all $\l\in{\cal L}$ and all $t\in{\mathbb R}$, where $l(0)=l$.[^2]
[**Comment:**]{} In order to avoid any time-dependent scale factors, it has been useful to first assume that all coherent states are normalized. While the set of coherent states satisfies temporal stability, the same cannot be said for the temporal evolution of a general state $e^{-i\H t}\,|\psi\>\equiv|\psi,t\>$. Nevertheless, in a coherent state representation that enjoys temporal stability, [*dynamics becomes kinematics*]{}. In other words, l|,tl|e\^[-it]{}|l(-t)|, namely, the dynamical evolution of an arbitrary state $\psi(l)\equiv\<l|\psi\>$ in the coherent state representation simply amounts to a “reshuffling of the labels”, $\psi(l,t)\equiv\psi(l(-t))$.
Let us see how we can explicitly implement temporal stability. For convenience, we restrict attention to Hamiltonians with a discrete, nondegenerate spectrum and energy levels of the form $0=E_0<E_1<E_2<\cdots\,$. It follows that $\lim_{n\ra\infty}E_n=E^*$, and cases where $E^*=\infty$ and $E^*<\infty$ are both of interest. We set $e_n\equiv E_n/\omega$, for some convenient choice of $\omega$, to generate a sequence of dimensionless energy levels. If $E^*<\infty$, we can, without loss of generality, choose $\omega=E^*$ so that $\lim_{n\ra\infty}e_n=1$. Furthermore, we let $|n\>$, $n=0,1,2,\ldots\,$, be energy eigenvalues for $\H$, such that |n=E\_n|n=e\_n|n. We then define (see [@kla191]) coherent states asociated with this system by the expression |J,N(J)\^[-1/2]{}\_[n=0]{}\^|n, where $0\le J<J^*\le\infty$ and $-\infty<\gamma<\infty$, expressed with the aid of a set of positive weight factors $\{\rho_n\}$, with $\rho_0\equiv 1$ for convenience. Here normalization is achieved by setting N(J)=\_[n=0]{}\^, and where $J^*\equiv \liminf_{n\ra\infty}[\rho_n]^{1/n}$ denotes the radius of convergence of this series. We note first that &&e\^[-it]{}|J,N(J)\^[-1/2]{}\_[n=0]{}\^|n\
&&=|J,+t whatever the choice of the weight factors $\{\rho_n\}$. Thus by a careful choice of the phase factor we have ensured temporal stability.
Let us next discuss the freedom in the choice of the factors $\{\rho_n\}$ so that the coherent states fulfill Property 2 dealing with the resolution of unity. To that end, we assume there exists a nonnegative weight function $\rho(u)$, $\rho(u)\ge0$, $0\le u<U\le\infty$, with the property that \_n\_0\^Uu\^n(u)du;1.5cm \_0=1. Next we observe that &&|J,J,|d() \_(2)\^[-1]{}\_[-]{}\^ |J,J,|d\
&&3.5cm=N(J)\^[-1]{}\_[n=0]{}\^|nn|. Finally, if we introduce $k(J)\equiv N(J)\s\rho(J)$ and $U\equiv J^*$, then we find that &&-.8cm|J,J,|d(J,)\_0\^U k(J)dJd()|J,J,|\
&&=\_[n=0]{}\^\_0\^U J\^n(J)dJ\
&&=\_[n=0]{}\^|nn|.
As a result of this analysis, we learn that there are a vast number of coherent state sets, all of which fulfill temporal stability for a single Hamiltonian, and which are distinguished from each other by the presence of different weight factor sets $\{\rho_n\}$.
We now seek an additional physical criterion that picks out a [*single*]{} set of weights $\{\rho_n\}$ for a given Hamiltonian, thereby reducing the vast family of coherent states down to a single set.
The Action Identity {#the-action-identity .unnumbered}
===================
Let us return to the appropriate label map $l\ra l(t)$ for the set of coherent states under discussion. Specifically, the appropriate map in the present case is clearly given by $(J,\gamma)\ra(J,\gamma+\omega t)$. This temporal evolution is the most general solution of the two equations of motion =, 1.5cm [J]{}=0, which in turn arise, for example, from the “classical action functional” I=dt as the relevant Euler-Lagrange equations. In point of fact, other action functionals would work just as well, say, for instance, I’=dt . However, there is an additional sense in which $I$ is preferred since in that case $J$ and $\gamma$ can be said to be [*classical canonical coordinates*]{}; this interpretation is not supported by using $I'$ (or any other such form). Let us accept the physical notion that $J$ and $\gamma$ should represent classical canonical coordinates and thus $I$ corresponds to the appropriate classical action.
It is a longstanding proposal [@kla19; @kla37] that there is just [*one*]{} action principle in physics, and that in particular, [*the classical action principle is just the quantum action principle applied to a restricted set of Hilbert space vectors*]{}. We can illustrate this proposal as follows: Let I\_Q=dt denote the usual quantum action functional. Extremizing this functional over all bra vectors $\<\psi(t)|$ leads to Schrödinger’s equation i(d/dt)|(t)=|(t). Let us ask the question, however, what is the result if we extremize the quantum action functional over a [*limited set*]{} of vectors such as those in a set of coherent states. For example, consider states of the form |p,qe\^[-iqP]{}e\^[ipQ]{}|0, where $[Q,P]=i\one$ and $|0\>$, say, is a unit vector which satisfies $(Q+iP)\s|0\>=0$. It is then straightforward to show that &&I\_Q=dt\
&&.59cm=dt, where $H(p,q)\equiv\<p,q|\s\H\s|p,q\>$ is a classical Hamiltonian symbol asociated with the quantum Hamiltonian $\H$. Clearly, extremal variation of $I_Q$ within the limited set of coherent states, i.e., for general functions $p(t)$ and $q(t)$, leads to traditional classical equations of motion for the canonical variables $p$ and $q$. In this interpretation, classical dynamics is what remains of quantum dynamics when the latter is subject to a sufficiently large class of constraints that restrict possible variations. Stated otherwise, classical dynamics is quantum dynamics restricted to the only quantum degrees of freedom that may possibly be varied at a macroscopic level, namely, the mean position and the mean momentum (or velocity).
The foregoing discussion can be applied to the problem at hand as follows: If we seriously wish to identify the variables $J,\gamma$ of the coherent states $
|J,\gamma\>$ as canonical coordinates, then it is necessary that &&I=dt\
&&.35cm=dt. Consequently, we are led to the next, and last, postulate, namely.4cm
[**4. Action Identity:**]{} To ensure that the variables $J$ and $\gamma$ correspond to physical canonical coordinates, we require that J,||J,=J.
[**Comment:**]{} As easily seen, this last condition is equivalent to requiring that $i\<J,\gamma|\s d\s|J,\gamma\>=J\,d\gamma$. The action identity is a strong requirement, and we next show that it will uniquely specify the weight factors $\{\rho_n\}$ for a given Hamiltonian $\H$.
The action identity asserts, for all $J$, $0\le J< J^*$, that \_[n=o]{}\^=J\_[n=0]{}\^. Equating like powers of $J$, we are led to the condition $e_n/\rho_n=
1/\rho_{n-1}$, or $\rho_n=e_n\s\rho_{n-1}$. Choosing $\rho_0=1$ (as already noted), we find that \_ne\_ne\_[n-1]{}e\_1 =\_[l=1]{}\^ne\_l. The final result is, therefore, the set of coherent states introduced by Gazeau and Klauder [@kla207].
It is instructive to apply the final coherent-state prescription to a familiar example, namely, to the harmonic oscillator. In that case, $E_n=\omega\s n$, or $e_n=n$, and so $\rho_n=n!$. If we let $|z|\equiv J^{1/2}$ and set $z\equiv|z|\s e^{-i\gamma}$, then we find that |J,|z=e\^[-|z|\^2]{}\_[n=0]{}\^|n. Observe that in the present case it suffices that $\pi\le\gamma<\pi$ to achieve the needed orthogonality. Reassuringly, therefore, we have been able to deduce from our several postulates that the canonical coherent states are the unique family of coherent states associated with the harmonic oscillator dynamics.
Application to Hydrogen-like Spectrum {#application-to-hydrogen-like-spectrum .unnumbered}
=====================================
Finding coherent states for the bound state portion of the hydrogen atom has been a long-standing problem. Surely, various proposals for such coherent states have been made (see, e.g., [@help]), and just as surely they generally differ one from another. One means to gauge such proposals is how well they do in the semi-classical regime, namely, what is the spread in the energy levels for highly excited systems. Ideally, one would prefer that the spread decreases as the excitation level rises so that more nearly classical-like behavior is obtained. A measure of the spread is provided by the variance, and therefore it is appropriate to focus on the variance in the proposed coherent states. While the full hydrogen atom has been treated elsewhere, we content ourselves here with a simple one-dimensional model which serves to illustrate the principles involved in a clearer fashion.
We now turn our attention to a one-dimensional model problem with the hydrogen-like spectrum E\_n=. In this case \_n=\_[l=1]{}\^n= (), and thus the coherent states in question are defined by |J,=N(J)\^[-1/2]{}\_[n=0]{}\^J\^[n/2]{}e\^[-i]{}|n. Here N(J)=\_[n=0]{}\^() J\^n = +, provided that $0\le J<J^*=1$. As in the general case, these states clearly exhibit temporal stability, i.e., e\^[-it]{}|J,=|J,+t.
Variance {#variance .unnumbered}
--------
By design, of course, these states fulfill the condition J,||J,=J . A question of particular interest, however, refers to the [*variance*]{} of the energy in each of the given coherent states since this quantity serves to indicate how well the energy is peaked about its mean value in the coherent state $|J,\gamma\>$.
It may be shown (e.g., by direct computation) that for the hydrogen-like model under discussion the variance &&v(J)J,|\^2|J,-J,|| J,\^2\
&&.93cm (3\^2/4)J(1-J). It is noteworthy that the variance vanishes not only for $J=0$ but for $J=1$ as well. This fact implies that the state $|J,\gamma\>$ is peaked in its energy values about its mean value when $J\approx 0$ and when $J\approx 1$.
We now proceed to discuss the variance in a more general fashion.
Variances for more General Systems {#variances-for-more-general-systems .unnumbered}
==================================
Let us discuss the variance for rather general systems for which $J^*=1$. This analysis leads to further information about the hydrogen-like model as well as many other examples.
In the general case, the energy variance is defined by && v(J)=J,|\^2|J,-J,||J,\^2\
&&.91cm=-\
&&.91cm=. Let us examine $v(J)$ at the two extremes $J\approx0$ and $J\approx1$.
First, for $J\approx0$, we readily see that v(J)=e\_1J +O(J\^2) . In short, v(J)J near $J=0$.
For $J\approx1$ the analysis is somewhat more involved. We note that $v(J)$ may be written as v(J)=, where $\delta_n\equiv1-e_n$. Observe that $\delta_n\ra0$ as $n\ra\infty$. For the moment we assume even more, namely, that $\Sigma\s\delta_m^2<\infty$. Since large $n$ values dominate the $n$-sums in the numerator and the denominator, then near $J=1$ it suffices to consider &&v(J) =\
&&.85cm=(1-J)(\_\_m\^2/\_m) +O(\[1-J\]\^2). Roughly speaking, if $\delta_m^2\propto m^{-\tau}$, for large $m$, $1<\tau$, then we have shown to leading order that v(J)(1-J) near $J=1$. On the other hand, if $\delta_m^2\propto m^{-\tau}$, for large $m$, $0<\tau<1$, it follows to leading order that v(J)(1-J)\^near $J=1$.
Finally, we learn that the vanishing of the variance for large quantum numbers, i.e., when $J\approx 1$ is a rather general phenomena, given the choice of coherent states to which we have been led in the present article. This fact would seem to confirm their utility in semi-classical analyses rather generally.
Related work {#related-work .unnumbered}
============
Several other papers have recently appeared dealing with topics raised in this article, and the interested reader may wish to consult them directly. In [@ant] temporally stable coherent states are developed for the infinite square well and for the Pöschl-Teller potential. A review of various attempts to develop coherent states in general and hydrogen atom coherent states in particular is given in [@cra].
Acknowledgements {#acknowledgements .unnumbered}
================
I take this opportunity to thank several colleagues who have been recently involved with the author in one way or another regarding the subject of coherent states. These collaborators are: J.-P. Antoine, B. Bodmann, J.-P. Gazeau, P. Monceau, K.A. Penson, J.-M. Sixdeniers, S.V. Shabanov, and G. Watson.
[99]{} J.R. Klauder, “The Action Option and the Feynman Quantization of Spinor Fields in Terms of Ordinary C-Numbers", Annals of Physics [**11**]{}, 123-168 (1960). V. Bargmann, “On a Hilbert Space of Analytic Functions and an Associated Integral Transform, Part I”, Commun. Pure and Applied Math. [**14**]{}, 187-214 (1961). R.J. Glauber, “The Quantum Theory of Optical Coherence”, Phys. Rev. [**130**]{}, 2529-2539 (1963). J.R. Klauder, “Continuous-Representation Theory II. Generalized Relation Between Quantum and Classical Dynamics", J. Math. Phys. [**4**]{}, 1058-1073 (1963). A.M. Perelomov, “Coherent States for Arbitrary Lie Groups”, Commun. Math. Phys. [**26**]{}, 222-236 (1972). R. Gilmore, “On the Properties of Coherent States”, Revista Mexicana de Fisica [**23**]{}, 143-187 (1974). A.O. Barut and L. Girardello, “New ‘Coherent States’ Associated with Non-Compact Groups”, Commun. Math.Phys. [**21**]{}, 41-55 (1972). M.M. Nieto and L.M. Simmons, Jr., “Coherent States for General Potentials”, Phys. Rev. Lett. [**41**]{}, 207-210 (1978). J.R. Klauder, “Continuous-Representation Theory I. Postulates of Continuous Representation Theory", J. Math. Phys. [**4**]{}, 1055-1058 (1963). J.R. Klauder, “Coherent State Path Integrals [*without*]{} Resolutions of Unity”, Found. Phys. [**31**]{} 57-67 (2001). J.R. Klauder, “Coherent States for the Hydrogen Atom.” J. Phys. A: Math. Gen. [**29**]{}, L293-L298 (1996). J.R. Klauder, “Continuous-Representation Theory III. On Functional Quantization of Classical Systems", J. Math. Phys. [**5**]{}, 177-187 (1964). J.-P. Gazeau and J.R. Klauder, “Coherent States for Systems with Discrete and Continuous Spectrum”, J. Phys. A: Math. Gen. [**32**]{}, 123-132, (1999). L.S. Brown, Am. J. Phys. [**41**]{}, 525 (1973); J. Mostowski, Lett. Math. Phys. [**2**]{}, 1 (1977); J.C. Gay, D. Delande, and A. Bommier, Phys. Rev. A [**39**]{}, 6587 (1989); M. Nauenberg, Phys. Rev. A [**40**]{}, 1133 (1989); Z.D. Gaeta and C.R. Stroud, Jr., Phys. Rev. A [**42**]{}, 6308 (1990); J.A. Yeazell and C.R. Stroud, Jr., Phys. Rev. A [**43**]{}, 5153 (1991); M. Nauenberg, in [*Coherent States: Past, Present, and Future*]{}, Eds. D.H. Feng, J.R. Klauder, and M.R. Strayer (World Scientific, Singapore, 1994), p.345; I. Zlatev, W.-M. Zhang, and D.H. Feng, Phys. Rev. [**50**]{}, R1973 (1994); R. Bluhm, V.A. Kostelcky, and B. Tudose, (1995); J.R. Klauder, J. Phys. A: Math. Gen. [**29**]{}, L293-L298 (1996); P. Majumdar and H.S. Sharatchandra, Phys. Rev. A [**56**]{}, R3322 (1997); M.G.A. Crawford, Phys. Rev. A [**62**]{}, 012104, 1-7 (2000). J.-P. Antoine, J.-P. Gazeau, P. Monceau, J.R. Klauder, and K.A. Penson, “Temporally Stable Coherent States for Infinite Well and Pöschl-Teller Potentials”, J. Math. Phys. [**42**]{}, 2349-2386 (2001). M.G.A. Crawford, “Temporally Stable Coherent States in Energy-degenerate Systems: The Hydrogen Atom”, Phys. Rev. A [**62**]{}, 012104, 1-7 (2000).
[^1]: Contribution to the 7th ICSSUR Conference, June 2001.
[^2]: Although temporal stability refers to the [*quantum*]{} evolution of the coherent states, there is nonetheless an induced [*classical*]{} dynamics inherent in this concept that realizes the label-space map $l\ra l(t)$ for each $l\in{\cal L}$. We shall touch on this classical dynamics below.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, a new two-parameter model called generalized Ramos-Louzada (GRL) distribution is proposed. The new model provides more flexibility in modeling data with increasing, decreasing, j shaped and reversed-J shaped hazard rate function. Several statistical and reliability properties of the GRL model are also presented in this paper. The unknown parameters of the GRL distribution are discussed using eight frequentist estimation approaches. These approaches are important to develop a guideline to choose the best method of estimation for the GRL parameters, that would be of great interest to practitioners and applied statisticians. A detailed numerical simulation study is carried out to examine the bias and the mean square error of the proposed estimators. We illustrate the performance of the GRL distribution using two real data sets from the fields of medicine and geology and both data sets show that the new model is more appropriate as compared to the gamma, Marshall-Olkin exponential, exponentiated exponential, beta exponential, generalized Lindley, Poisson-Lomax, Lindley geometric and Lindley distributions, among others.\
**Keywords:** Cramér–von Mises Estimation; Maximum Likelihood Estimation; Maximum Product of Spacing Estimation; Right–Tail Anderson-Darling Estimation.\
**AMS subject classification:** 62E10, 60K10, 60N05.
author:
- 'Hazem Al-Mofleh[^1]'
- 'Ahmed Z. Afify'
title: '**A generalization of Ramos-Louzada distribution: Properties and estimation**'
---
Introduction {#sec1}
============
The probability distributions have great importance for modeling data in several areas such as medicine, engineering, and life testing, among others. Ramos and Louzada (2019) recently introduced the one-parameter distribution called Ramos-Louzada (RL) distribution with survival function (SF) given by $$\label{fdpnda}
S(t|\lambda) = \left(\frac{1}{\lambda-1}\right)
\left(\lambda-1+\frac{t}{\lambda} \right)e^{- \frac{t}{\lambda}}, \quad \quad t>0,$$ where $\lambda\geq 2$.
The two most common one-parameter distributions are the exponential and Lindley distributions. The important generalizations of the exponential distribution are the Weibull (Weibull, 1951) and exponentiated exponential (Gupta and Kundu, 2001) models. In the case of the Lindley distribution, the power Lindley (Ghitany et al., 2013) and generalized Lindley (Nadarajah et al., 2011) models have play an important role in survival analysis. These two generalizations are obtained by considering a power parameter in the exponential and Lindley distributions. Ramos and Louzada (2019) showed that (\[fdpnda\]) outperforms the common exponential and Lindley distributions in many situations. Therefore, we will propose a new two-parameter extension of the RL distribution by including a power parameter in the baseline model (\[fdpnda\]). The new proposed model is called a generalized Ramos-Louzada (GRL) distribution.
Let $T$ be a non-negative random variable follows the GRL model, the SF of random variable $T$ is given by $$\label{fdpnd}
S(t|\lambda,\alpha) = \left(\frac{1}{\lambda-1}\right)
\left(\lambda-1+\frac{t^\alpha}{\lambda} \right)e^{- \frac{t^\alpha}{\lambda}},$$ where $\lambda(\geq 2)$ and $\alpha (>0)$ are shape parameters.
Some mathematical properties, parameters estimation by eight different methods, simulations and applications are studied and proposed in this paper.
We can summarize the motivations of this proposed model as: (i) Its cumulative distribution function (CDF) and hazard rate function (HRF) have simple closed forms, hence, it can be utilized to analyze censored data; (ii) It can be represented as a mixture of Weibull distribution and a particular case of the generalized gamma distribution (Stacy, 1962) (See Section \[sec2\]); (iii) The GRL distribution exhibits increasing, decreasing, reversed-J shaped and J shaped hazard rates, whereas the RL model exhibits only increasing hazard rate; and (iv) The GRL distribution outperforms many of the well-known distributions namely: gamma, Marshall-Olkin exponential, exponentiated exponential, Beta exponential, generalized Lindley, Poisson-Lomax, Lindley geometric and Lindley distributions, using two real data sets from the fields of medicine and geology. Furthermore, another important goal of this paper is to show how several frequentist estimators of the GRL parameters perform to choose the best parameter estimation method for the proposed model, which would be a great interest to practitioners and applied statisticians.
The skewness of the GRL distribution varies within the interval (-0.68158, 5.17333), whereas the skewness of the RL distribution can only range in the interval (1.41421, 1.85648) when the parameter $\lambda$ takes values (2, 3.1, 4, 5.5). Furthermore, the spread of the kurtosis of the GRL distribution is much larger ranging, which is from 2.69447 to 52.6597, whereas the spread of the kurtosis of the RL distribution can only varies from 6.00 to 8.04 for the same values shown above for the parameter $\lambda$.
This paper is organized as follows: Section \[sec2\] introduces the GRL distribution and its properties such as: quantile function, moments, order statistics and HRF. Section \[sec3\] presents the estimators of the GRL unknown parameters based on eight classical estimation methods. Simulation study, to evaluate and compare the behavior of the eight classical estimation methods, is discussed in Section \[sec4\]. Section \[sec5\] illustrates the relevance of GRL model for two real lifetime data sets. Section \[sec6\] summarizes the present study.
Properties {#sec2}
==========
Let $T$ be a random variable follows the GRL model with SF given in (\[fdpnd\]), the probability density function (PDF) of the random variable $T$ is given by
$$\label{RLPDF}
\quad \quad f(t;\pmb \phi)=\frac{\alpha}{\lambda(\lambda-1)}t^{\alpha-1}\left(\lambda+\frac{t^\alpha}{\lambda} -2\right)e^{-\frac{t^\alpha}{\lambda} }, \quad t>0, \quad \lambda\geq 2, \quad \alpha>0,$$
where $\pmb \phi=(\lambda,\alpha)^{\intercal}$. Note that, the RL model can be obtained from (\[RLPDF\]) when $ \alpha=1$.
Shapes
------
The behavior of the PDF in (\[RLPDF\]) when $t\rightarrow0$ and $t\rightarrow\infty$ are, respectively, given by
$$\lim_{t\rightarrow 0}f(t;\pmb \phi)=
\begin{cases}
\infty, & \text{if }\alpha<1 \\
\dfrac{(\lambda-2)}{\lambda(\lambda-1)}, & \text{if }\alpha=1 \\
0, & \text{if }\alpha>1
\end{cases},$$
$$\lim_{t\rightarrow \infty} f(t;\pmb \phi)= 0.$$
In Figure \[density\], we present the shapes of the PDF for different values of the parameters $\lambda$ and $\alpha$. The shape of PDF of the GRL model can be right-skewed and reversed-J shaped.
[.5]{} ![PDF shapes for the GRL distribution considering different values of $\lambda $ and $\alpha$.[]{data-label="density"}](density1.pdf "fig:"){width="1\linewidth"}
[.5]{} ![PDF shapes for the GRL distribution considering different values of $\lambda $ and $\alpha$.[]{data-label="density"}](density2.pdf "fig:"){width="1\linewidth"}
The CDF of the GRL distribution is given by
$$\label{cdf}
F(t;\pmb \phi) = 1-\left(\frac{1}{\lambda-1}\right)
\left(\lambda-1+\frac{t^\alpha}{\lambda} \right)e^{- \frac{t^\alpha}{\lambda}}, \quad t>0, \quad \lambda\geq 2, \quad \alpha>0.$$
The HRF of $T$ is given by $$\label{fhwl}
h(t;\pmb \phi)=\frac{f(t|\alpha,\lambda)}{S(t|\alpha,\lambda)}=\frac{\alpha t^\alpha}{\lambda}\frac{(\lambda^2+t^\alpha-2\lambda)}{(\lambda^2+t^\alpha-\lambda)} .$$
Figure \[HRF\] displays some possible shapes of HRF of the GRL for some selected values of $\lambda$ and $\alpha$. The shape of HRF can be increasing, decreasing, reversed-J shaped and J shaped hazard rates.
![Hazard rat function shapes for the GRL considering different values of $\lambda $ and $\alpha$.[]{data-label="HRF"}](GRL_hrf.pdf)
The GRL distribution can be expressed as a two-component mixture $$\label{eqmix2}
f(t;\pmb \phi)= pf_1(t;\pmb \phi)+(1-p)f_2(t;\pmb \phi)\, ,$$ where $1-p=1/(\lambda-1)$ (or $p=(\lambda-2)/(\lambda-1)$) and $$\label{mixeq1} f_j(t;\pmb \phi)=\frac{\alpha}{\lambda^j} t^{j\alpha-1}e^{- \frac{t^\alpha}{\lambda}} \quad \mbox{for} \quad j=1,2.$$
Note that, $f_1(\cdot)$ is a Weibull distribution and $f_2(\cdot)$ is a particular case of the generalized gamma distribution (Stacy, 1962). Then, after some algebra, Equation (\[eqmix2\]) reduces to the PDF in (\[RLPDF\]).
Quantile function
-----------------
The quantile function (QF) of the GRL distribution defined in (\[RLPDF\]), say $Q(p)$ where $0<p<1$, it can be obtained by solving the equation $F(Q(p))=p$ in (\[cdf\]) for $Q(p)$ in terms of $p$, and this implies $$\begin{aligned}
\label{qf}
Q(p)=\left(-\lambda\left[ W_{-1} \left[(\lambda -1) (p-1) e^{1-\lambda }\right]+\lambda-1 \right ] \right)^{1/\alpha },\end{aligned}$$ where $W_{-1}(x)$ is the negative branch of the Lambert function.
Moments
-------
Moments play an important role in statistical theory, in this section we provide the $r-$th moment, the mean and variance for the GRL distribution.
\[moments1\] For the random variable $T$ follows the GRL distribution, the $r-$th moment is given by $$\label{rmNM}
\mu_r= E[T^r]=\frac{r\lambda^{\frac{r}{\alpha}}}{\alpha(\lambda-1)}\left(\lambda+\frac{r}{\alpha}-1 \right)\Gamma\left(\frac{r}{\alpha}\right) \, , \quad \text{for} \quad r\in\mathbb{N} .$$
Note that, the $r-$th moment for the random variable in (\[mixeq1\]) is given by $$E[T^r;\alpha,\lambda]=\lambda^{\frac{r}{\alpha}}\Gamma\left(\frac{r}{\alpha}+j\right), \ \mbox{ for } \ j=1,2 .$$
Since the GRL model can be expressed as a two-component mixture, as in (\[eqmix2\]), so we have $$\begin{aligned}
\mu_r&= E[T^r]= \int_{0}^{\infty}t^rf(t|\alpha,\lambda)dt
= pE[X^r;\alpha,\lambda]+(1-p)E[X^r;\alpha+1,\lambda] \\& = \left(\frac{\lambda-2}{\lambda-1}\right)\lambda^{\frac{r}{\alpha}}\Gamma\left(\frac{r}{\alpha}+1\right)+\frac{1}{(\lambda-1)}\lambda^{\frac{r}{\alpha}}\Gamma\left(\frac{r}{\alpha}+2\right)\\&= \frac{r\lambda^{\frac{r}{\alpha}}}{\alpha(\lambda-1)}\left(\lambda+\frac{r}{\alpha}-1 \right)\Gamma\left(\frac{r}{\alpha}\right) \cdot
\end{aligned}$$
The random variable $T$ follows the PDF in (\[RLPDF\]), its mean and variance, respectively, are given by
$$\label{lem2}
\mu=\frac{\lambda^{\frac{1}{\alpha}}}{\alpha(\lambda-1)}\Gamma\left(\frac{1}{\alpha}\right)\left(\lambda+\frac{1}{\alpha}-1 \right) \ \ \text{ and }$$
$$\sigma^2=\frac{\lambda^{\frac{2}{\alpha}}}{\alpha^2(\lambda-1)^2}\left[2\alpha(\lambda-1)\Gamma\left(\frac{2}{\alpha}\right)\left(\lambda+\frac{2}{\alpha}-1 \right)-\Gamma^2\left(\frac{1}{\alpha}\right)\left(\lambda+\frac{1}{\alpha}-1 \right)^2\right].$$
From (\[rmNM\]) and considering $r=1$, it follows that $\mu_1=\mu$. The second result can be obtained by using $\sigma^2=E[T^2]-\mu^2$ and with some algebra the proof is completed.
The $r-$th central moment for the GRL distribution is given by $$\label{rcmNMp}
\begin{aligned}
M_r&= E[T-\mu]^r= \sum_{i=0}^{r}\binom{r}{i}(-\mu)^{r-i}E[T^i] \\ &= \sum_{i=0}^{r}\binom{r}{i}\left[-\frac{\lambda^{\frac{1}{\alpha}}\Gamma\left(\alpha^{-1}\right)}{\alpha(\lambda-1)}\left(\lambda+\frac{1}{\alpha}-1 \right) \right]^{r-i}\left[\frac{i\lambda^{\frac{i}{\alpha}}}{\alpha(\lambda-1)}\left(\lambda+\frac{i}{\alpha}-1 \right)\Gamma\left(\frac{i}{\alpha}\right) \right] .
\end{aligned}$$
The result follows directly from the Proposition \[moments1\].
The mean, variance, skewness and kurtosis of the GRL distribution are computed numerically for different values of the parameters $\lambda$ and $\alpha$, using the R software. Table \[tab:tab1\] displays these numerical values. From Table \[tab:tab1\] we can indicate that the skewness of the GRL distribution can range in the interval $(-0.68158, 5.17333)$. The spread for the GRL kurtosis is much larger ranging from 2.69447 to 52.6597. Further, the GRL model can be left skewed or right skewed. Hence, the GRL distribution is a flexible distribution which can be used in modelling skewed data.
[crrrr]{}\
\
${\pmb \phi}^{\intercal}$ & Mean &Variance & Skewness & Kurtosis\
\
\
$(\lambda=2.0,\alpha=0.5)$ & 24.00 & 1344.00 & 4.30 & 37.41\
$(\lambda=2.0,\alpha=0.7)$ & 8.27 & 72.44 & 2.39 & 12.65\
$(\lambda=2.0,\alpha=2.5)$ & 1.64 & 0.23 & 0.20 & 2.89\
$(\lambda=2.0,\alpha=3.5)$ & 1.41 & 0.09 & -0.04 & 2.88\
$(\lambda=3.1,\alpha=0.5)$ & 37.52 & 5030.15 & 5.17 & 52.66\
$(\lambda=3.1,\alpha=0.7)$ & 10.71 & 186.23 & 2.85 & 16.24\
$(\lambda=3.1,\alpha=2.5)$ & 1.66 & 0.42 & 0.18 & 2.73\
$(\lambda=3.1,\alpha=3.1)$ & 1.49 & 0.23 & -0.03 & 2.71\
$(\lambda=4.0,\alpha=1.5)$ & 2.78 & 3.19 & 0.92 & 3.89\
$(\lambda=4.0,\alpha=3.5)$ & 1.46 & 0.20 & -0.07 & 2.71\
$(\lambda=4.0,\alpha=5.0)$ & 1.29 & 0.08 & -0.35 & 2.95\
$(\lambda=4.0,\alpha=10)$ & 1.13 & 0.02 & -0.73 & 3.74\
$(\lambda=5.5,\alpha=3.5)$ & 1.56 & 0.23 & -0.03 & 2.69\
$(\lambda=5.5,\alpha=5.0)$ & 1.35 & 0.09 & -0.30 & 2.90\
$(\lambda=5.5,\alpha=5.5)$ & 1.31 & 0.07 & -0.36 & 2.99\
$(\lambda=5.5,\alpha=10)$ & 1.15 & 0.02 & -0.68 & 3.64\
\
Order statistics
----------------
Let $X_{1}, X_{2},\ldots X_{n}$ be a random sample from (\[RLPDF\]) and $X_{1:n}\leq X_{2:n}\leq \ldots \leq X_{n:n}$ denote the the corresponding order statistics. It is well known that the PDF and the CDF of the of $r$-th order statistic say $X_{r:n}$; $1\leq r \leq n$, respectively, are given by $$\label{orderapdf}
\begin{aligned}
f_{r:n}(x)&=\frac{n!}{(r-1)!(n-r)!}[F(x)]^{r-1}[1-F(x)]^{n-r}f(x) \\
&=\frac{n!}{(r-1)!(n-r)!}\sum_{u=0}^{n-r}(-1)^{u}\left(\begin{array}{c}n-r\\u\end{array}\right)[F(x)]^{r-1+u}f(x)
\end{aligned}$$ and $$\label{orderacdf}
\begin{aligned}
F_{r:n}(x)=\sum_{l=k}^{n}\left(\begin{array}{c}n\\l\end{array}\right)[F(x)]^{l}[1-F(x)]^{n-l}=\sum_{l=k}^{n}\sum_{u=0}^{n-r}(-1)^{u}\left(\begin{array}{c}n\\l\end{array}\right)\left(\begin{array}{c}n-r\\u\end{array}\right)[F(x)]^{l+u},
\end{aligned}$$ for $k=1, 2, \ldots, n$. It follows, from (\[orderapdf\]) and (\[orderacdf\]), that the PDF and CDF of the $r$-th order statistic of the GRL reduce to
$$\begin{aligned}
f_{r:n}(x)=&\frac{n!}{(\lambda-1)(r-1)!(n-r)!}\left(\lambda+\frac{t}{\lambda}-2\right)e^{-\frac{t}{\lambda}}\sum_{u=0}^{n-r}(-1)^{u}\left(\begin{array}{c}n-r\\u\end{array}\right)\times \\ & \left[1-\left(\frac{1}{\lambda-1}\right)
\left(\lambda-1+\frac{t^\alpha}{\lambda} \right)e^{- \frac{t^\alpha}{\lambda}} \right]^{r-1+u}
\end{aligned}$$
and $$F_{r:n}(x)=\sum_{l=k}^{n}\sum_{u=0}^{n-r}(-1)^{u}\left(\begin{array}{c}n\\l\end{array}\right)\left(\begin{array}{c}n-r\\u\end{array}\right)\left[1-\left(\frac{1}{\lambda-1}\right)
\left(\lambda-1+\frac{t^\alpha}{\lambda} \right)e^{- \frac{t^\alpha}{\lambda}}\right]^{l+u}.$$
Inference {#sec3}
=========
In this section, we estimate of the GRL parameters $\lambda$ and $\alpha$ using eight frequentist approaches. These methods are: the weighted least-squares (WLSE), ordinary least squares (OLSE), maximum likelihood (MLE), maximum product of spacing (MPSE), Cramér–von Mises (CVME), Anderson–Darling (ADE), Right-tail Anderson–Darling (RADE) and percentile based (PCE) estimators.
Maximum likelihood estimation
-----------------------------
In this sub-section we present the maximum likelihood estimator (MLE) of the parameters $\lambda$ and $\alpha$ of the GRL distribution.
Let $T_1,\ldots,T_n$ be a random sample such that $T$ has PDF given in (\[RLPDF\]). In this case, for $\pmb {\phi}=(\lambda,\alpha)^{\intercal}$, the likelihood function from (\[RLPDF\]) is given by $$\label{veroiNM}
L(\pmb {\phi};\pmb {t})=\frac{\alpha^{n}}{{\lambda^{n+1}(\lambda-1)}^n}\prod_{i=1}^n{t_i^{\alpha-1}}\prod_{i=1}^n\left(\lambda^2+t_i^\alpha-2\lambda\right)\exp\left(-\frac{1}{\lambda}\sum_{i=1}^n t_i^\alpha\right).$$
The log-likelihood function $l(\pmb {\phi};\pmb{t})=\log{L(\pmb {\phi};\pmb{t})}$ is given by $$\label{verogNM2}
\begin{aligned}
l(\pmb {\phi};\pmb {t})=&\ n\log\left(\alpha\right)-(n+1)\log(\lambda)-n\log(\lambda-1)-\frac{1}{\lambda}\sum_{i=1}^n t_i^\alpha +(\alpha-1)\sum_{i=1}^{n}\log(t_i) \\ &+ \sum_{i=1}^n\log\left(\lambda^2+t_i^\alpha-2\lambda\right).
\end{aligned}$$
From the expressions $\frac{\partial}{\partial \lambda}l(\pmb {\phi};\pmb {t})=0$, $\frac{\partial}{\partial \alpha}l(\pmb {\phi};\pmb {t})=0$, we get the likelihood equations $$\label{verogg21}
-\frac{n+1}{\lambda}-\frac{n}{\lambda-1}+\frac{1}{\lambda^2}\sum_{i=1}^n t_i^\alpha + \sum_{i=1}^n\frac{2(\lambda-1)}{\lambda^2+t_i^\alpha-2\lambda}=0$$ and $$\label{verogg22}
\frac{n}{\alpha}-\frac{1}{\lambda}\sum_{i=1}^n t_i^\alpha\log(t_i)+\sum_{i=1}^{n}\log(t_i)+ \sum_{i=1}^n\frac{t_i^\alpha\log(t_i)}{\lambda^2+t_i^\alpha-2\lambda}=0 .$$ Under mild conditions (Migon, 2014) the ML estimates are asymptotically normal distributed with a bivariate normal distribution given by $$(\hat\lambda,\hat\alpha) \sim N_2[(\lambda,\alpha),H^{-1}(\lambda,\alpha)] \mbox{ for } n \to \infty ,$$ where the elements of the observed Fisher information matrix $H(\lambda,\alpha)$ are given by $$h_{11}(\lambda,\alpha)=-\frac{n+1}{\lambda^2}-\frac{n}{(\lambda-1)^2}+\frac{2}{\lambda^3}\sum_{i=1}^n t_i^\alpha - \sum_{i=1}^n\frac{2(t_i^\alpha-\lambda^2+2\lambda-2)}{\left(\lambda^2+t_i^\alpha-2\lambda\right)^2},$$ $$h_{12}(\lambda,\alpha)=h_{21}(\alpha,\lambda)=-\frac{1}{\lambda^2}\sum_{i=1}^n t_i^\alpha\log(t_i)+\sum_{i=1}^n\frac{2(\lambda-1)t_i^\alpha\log(t_i)}{\left(\lambda^2+t_i^\alpha-2\lambda\right)^2},$$ $$h_{22}(\lambda,\alpha)=+\frac{n}{\alpha^2}+\frac{1}{\lambda}\sum_{i=1}^n t_i^\alpha\log(t_i)^2 -\sum_{i=1}^n\frac{\lambda(\lambda-2)t_i^\alpha\log(t_i)^2}{\left(\lambda^2+t_i^\alpha-2\lambda\right)^2}.$$
This can also be done by using different programs namely `R` (`optim` function), `SAS` (`PROC NLMIXED`) or by solving the nonlinear likelihood equations obtained by differentiating $\ell $.
Ordinary and weighted least-square estimators
---------------------------------------------
Let $x_{(1)},x_{(2)},\cdots ,x_{(n)}$ be the order statistics of the random sample of size $n$ from $F\left( \mathbf{x};\lambda,\alpha \right) $ in (\[cdf\]). The ordinary least square estimators (OLSEs) (Swain et al., 1988). $\widehat{\lambda }_{LSE}$ and $\widehat{\alpha }_{LSE}$ can be obtained by minimizing $$V\left(\lambda,\alpha \right) =\sum_{i=1}^{n}\left[ F\left(
x_{(i)}|\lambda,\alpha \right) -\frac{i}{n+1}\right] ^{2},$$with respect to $\lambda$ and $\alpha$. Or equivalently, the OLSEs follow by solving the non-linear equations$$\sum_{i=1}^{n}\left[ F\left( x_{(i)}|\lambda,\alpha \right) -\frac{i}{n+1}%
\right] \Delta _{s}\left( x_{(i)}|\lambda,\alpha \right) =0,~\ s=1,2,$$where $$\begin{aligned}
\Delta _{1}\left( x_{(i)}|\lambda,\alpha \right) &=&\frac{\partial }{%
\partial \lambda }F\left( x_{(i)}|\lambda,\alpha \right) \text{ and } ~\Delta
_{2}\left( x_{(i)}|\lambda,\alpha \right) =\frac{\partial }{\partial
\alpha }F\left( x_{(i)}|\lambda,\alpha \right). \label{LSEs}\end{aligned}$$Note that the solution of $\Delta _{s}$ for $s=1,2$ can be obtained numerically.
The weighted least-squares estimators (WLSEs) (Swain et al., 1988), $%
\widehat{\lambda }_{WLSE}$ and $\widehat{\alpha }_{WLSE}$, can be obtained by minimizing the following equation $$W\left(\lambda,\alpha \right) =\sum_{i=1}^{n}\frac{\left( n+1\right)
^{2}\left( n+2\right) }{i\left( n-i+1\right) }\left[ F\left( x_{(i)}|\lambda,\alpha \right) -\frac{i}{n+1}\right] ^{2}.$$Further, the WLSEs can also be derived by solving the non-linear equations $$\sum_{i=1}^{n}\frac{\left( n+1\right) ^{2}\left( n+2\right) }{i\left(
n-i+1\right) }\left[ F\left( x_{(i)}|\lambda,\alpha \right) -\frac{i}{n+1}%
\right] \Delta _{s}\left( x_{(i)}|\lambda,\alpha \right) =0,~\ s=1,2,$$where $\Delta _{1}\left( \cdot |\lambda,\alpha \right) $ and $\Delta
_{2}\left( \cdot |\lambda,\alpha \right)$ are provided in (\[LSEs\]).
Method of maximum product of spacing
------------------------------------
The maximum product of spacings (MPS) method (Cheng and Amin, 1979 and Cheng and Amin, 1983 and Ranneby, 1984), as an approximation to the Kullback-Leibler information measure, is a good alternative to the MLE method.
Let $D_{i}(\lambda,\alpha)=F\left( x_{(i)}|\lambda,\alpha \right)
-F\left( x_{(i-1)}|\lambda,\alpha \right) $, for $i=1,2,\ldots ,n+1,$ be the uniform spacing of a random sample from the [GRL]{} distribution, where $F\left( x_{(0)}|\lambda,\alpha \right) =0$, $F\left( x_{(n+1)}|\lambda,\alpha \right) =1$ and $\sum_{i=1}^{n+1}D_{i}(\lambda,\alpha)=1$. The maximum product of spacing estimators (MPSEs) for $\widehat{\lambda }%
_{MPSE}$ and $\widehat{\alpha }_{MPSE}$ can be obtained by maximizing the geometric mean of the spacing $$G\left(\lambda,\alpha \right) =\left[ \prod\limits_{i=1}^{n+1}D_{i}(%
\lambda,\alpha)\right] ^{\frac{1}{n+1}}$$with respect to $\lambda$ and $\alpha$, or, equivalently, by maximizing the logarithm of the geometric mean of sample spacings $$H\left(\lambda,\alpha \right) =\frac{1}{n+1}\sum_{i=1}^{n+1}\log
D_{i}(\lambda,\alpha).$$The MPSEs of the GRL parameters can be obtained by solving the nonlinear equations defined by$$\frac{1}{n+1}\sum\limits_{i=1}^{n+1}\frac{1}{D_{i}(\lambda,\alpha)}\left[
\Delta _{s}(x_{(i)}|\lambda,\alpha)-\Delta _{s}(x_{(i-1)}|\lambda,\alpha)\right] =0,\quad s=1,2,$$where $\Delta _{1}\left( \cdot |\lambda,\alpha \right) $ and $\Delta
_{2}\left( \cdot |\lambda,\alpha \right) $ are defined in (\[LSEs\]).
The Cramér–von Mises minimum distance estimators
------------------------------------------------
The Cramér–von Mises estimators (CVMEs) as a type of minimum distance estimators have less bias than the other minimum distance estimators (Macdonald, 1971). The CVMEs are obtained based on the difference between the estimates of the CDF and the empirical distribution function (Luceño, 2006). The CVMEs of the GRL parameters are obtained by minimizing $$C(\lambda,\alpha)=\frac{1}{12n}+\sum_{i=1}^{n}\left[ F\left(
x_{(i)}|\lambda,\alpha \right) -{\frac{2i-1}{2n}}\right] ^{2},$$with respect to $\lambda$ and $\alpha$. Further, the CVMEs follow by solving the non-linear equations $$\sum_{i=1}^{n}\left[ F\left( x_{(i)}|\lambda,\alpha \right) -{\frac{2i-1}{%
2n}}\right] \Delta _{s}\left( x_{(i)}|\lambda,\alpha \right) =0,\quad
s=1,2,$$where $\Delta _{1}\left( \cdot |\lambda,\alpha \right) $ and $\Delta
_{2}\left( \cdot |\lambda,\alpha \right)$ are provided in (\[LSEs\]).
The Anderson-Darling and right-tail Anderson-Darling estimators
---------------------------------------------------------------
The Anderson-Darling statistic or Anderson-Darling estimator is another type of minimum distance estimators. The Anderson-Darling estimators (ADEs) of the GRL parameters are obtained by minimizing $$A(\lambda,\alpha)=-n-\frac{1}{n}\sum_{i=1}^{n}\left( 2i-1\right) \,\left[
\log F\left( x_{(i)}|\lambda,\alpha \right) +\log S\left( x_{(i)}|\lambda,\alpha \right) \right] ,$$with respect to $\lambda$ and $\alpha$. These ADEs can also be obtained by solving the non-linear equations $$\sum_{i=1}^{n}\left( 2i-1\right) \left[ \frac{\Delta _{s}\left(
x_{(i)}|\lambda,\alpha \right) }{F\left( x_{(i)}|\lambda,\alpha \right) }%
-\frac{\Delta _{j}\left( x_{(n+1-i)}|\lambda,\alpha \right) }{S\left(
x_{(n+1-i)}|\lambda,\alpha \right) }\right] =0,\ \ s=1,2.$$The right-tail Anderson-Darling estimators (RADEs) of the GRL parameters are obtained by minimizing $$R(\lambda,\alpha)=\frac{n}{2}-2\sum_{i=1}^{n}F\left( x_{i:n}|\lambda,\alpha \right) -\frac{1}{n}\sum_{i=1}^{n}\left( 2i-1\right) \log S\left(
x_{n+1-i:n}|\lambda,\alpha \right) ,$$with respect to $\lambda$ and $\alpha$. The RADEs can also be obtained by solving the non-linear equations $$-2\sum_{i=1}^{n}\Delta _{s}\left( x_{i:n}|\lambda,\alpha \right) +\frac{1}{%
n}\sum_{i=1}^{n}\left( 2i-1\right) \frac{\Delta _{s}\left(
x_{_{n+1-i:n}}|\lambda,\alpha \right) }{S\left( x_{n+1-i:n}|\lambda,\alpha \right) }=0,\ \ s=1,2.$$where $\Delta _{1}\left( \cdot |\lambda,\alpha \right) $ and $\Delta
_{2}\left( \cdot |\lambda,\alpha \right)$ are defined in Equation (\[LSEs\]).
Method of percentile estimation
-------------------------------
This method was originally suggested by Kao (1958, 1959). Let $ u_{i}=i/\left( n+1\right) $ be an unbiased estimator of $F\left(x_{(i)}|\lambda,\alpha \right) $. Then, the PCE of the parameters of GRL distribution are obtained by minimizing the following function$$P(\lambda,\alpha)=\sum_{i=1}^{n}\left( x_{(i)}-\left[-\lambda\left(W_{-1} \left((\lambda -1) (u_{i}-1) e^{1-\lambda }\right)+\lambda-1 \right) \right]^{1/\alpha } \right) ^{2},$$with respect to $\lambda$ and $\alpha$, where $W_{-1}(x)$ is the negative branch of the Lambert function.
Simulation analysis {#sec4}
===================
A simulation study is conducted to evaluate and compare the behavior of the estimates with respect to their: average of absolute value of biases ($|Bias(\widehat{\pmb \phi})|$), $|Bias(\widehat{\pmb \phi})|=$ $\frac{1}{N}\sum_{i=1}^{N}|\widehat{\pmb \phi}-\pmb \phi|$, the average of mean square errors (MSEs), $MSEs=\frac{1}{N}\sum_{i=1}^{N}(\widehat{\pmb \phi}-\pmb \phi)^2$, and average of mean relative errors (MREs), $MREs=\frac{1}{N}\sum_{i=1}^{N}|\widehat{\pmb \phi}-\pmb \phi|/\pmb \phi$.\
We generate $N=5000$ random samples: ${X_1,X_2,\ldots,X_N}$ of sizes $n=30, 50, 80,100$ and $200$ from GRL model by using equation (\[qf\]) by choosing $\lambda=\{2.0,4.5\}$ and $\alpha=\{0.5,2.5,0.7,3.5\}$, we used R software (version 3.6.1) (R Core Team, 2019) For each parameters combination and each sample, we estimate of the GRL parameters $\lambda$ and $\alpha$ using eight frequentist approaches: WLSE (weighted least-squares), OLSE (ordinary least squares), MLE (maximum likelihood), MPSE (maximum product of spacing), CVME (Cramer-von Mises), ADE (Anderson-Darling), RADE (right-tail Anderson-Darling) and PCE (percentile based). Then, the MSEs and MREs of the parameter estimates are computed. Simulated outcomes are listed in Tables \[tab:tab2\]-\[tab:tab9\] (see Appendix A). Furthermore, these tables show the rank of each of the estimators among all the estimators in each row, which is the superscript indicators, and the $\sum Ranks$, which is the partial sum of the ranks for each column in a certain sample size. Table \[tab:tab11\] shows the partial and overall rank of the estimators. From tables \[tab:tab2\]-\[tab:tab9\], we observe that:
- All Estimation methods show the property of consistency i.e., the MSEs and MREs decrease as sample size increase, for all parameter combinations, except the estimator method WLSE.\
- WLSE Estimation method shows the property of consistency for all parameter combinations, except the combinations $\phi=(\lambda=3.1,\alpha=0.7)^T$ and $\phi=(\lambda=3.1,\alpha=3.5)^T$, for the parameter $\lambda$.
Form Table \[tab:tab10\], and for the parameter combinations, we can conclude that the MPSE estimator method outperforms all the other estimator methods (overall score of 62). Therefore, depends on our study, we can confirm the superiority of MPSE and ADE estimator methods for the GRL distribution.
Data analysis {#sec5}
=============
In this section, we illustrate the importance of the GRL distribution in modelling skewed data using two real data sets from the medicine and geology fields. The first data represent the survival times, in weeks, of 33 patients suffering from acute myelogeneous leukaemia (Feigl and Zelen, 1965). These data have been analyzed by Abouelmagd et al. (2018), Nassar et al. (2018) and Sen et al. (2019). The second data is used to evaluate the risks associated with earthquakes occurring close to the central site of a nuclear power plant. This data set refers to the distances, in miles, to the nuclear power plant of the most recent 8 earthquakes of intensity larger than a given value (Castillo, 2012) and it consists of 60 observations.
We consider some measures of goodness-of-fit namely, minus maximized log-likelihood ($-\widehat{\ell }$), Cramér-Von Mises ($W^{\ast }$), Anderson-Darling ($A^{\ast }$) and Kolmogorov Smirnov (KS) statistics with its bootstrapped $p$-value (PV), to compare the fits of the GRL distribution with other competitive models given in Tables \[tab:tab12\] and \[tab:tab13\]. We draw $999999$ bootstrap samples to obtain the KS bootstrapped PV.
The fitted competitive models are namely given in \[tab:tab11\], and their densities (for $x>0$) are given by:
MOEx: $f(x)=\alpha \lambda \mathrm{\exp }(-\lambda x)\left[ 1-\left(
1-\alpha \right) \mathrm{\exp }(-\lambda x)\right] ^{-2}.$
BEx: $f(x)=\frac{\lambda }{B(a,b)}\mathrm{\exp }(-b\lambda x)\left[ 1-\mathrm{\exp }(-\lambda x)\right] ^{a-1}.$
EEx: $f(x)=\alpha \lambda \mathrm{\exp }(-\lambda x)\left[ 1-\mathrm{\exp }(-\lambda x)\right] ^{\alpha -1}.$
Ga: $f(x)=\frac{b^{-a}}{\Gamma \left( a\right) }x^{a-1}\exp \left(
-x/b\right) .$
GLi: $f(x)=\frac{\alpha \lambda ^{2}}{\lambda +1}\left( 1+x\right) \mathrm{\exp }(-\lambda x)\left[ 1-\frac{1+\lambda +\lambda x}{\lambda +1}\mathrm{\exp }(-\lambda x)\right] ^{\alpha -1}.$
TTLi: $f\left( x\right) =\left(\frac{a^{2}}{\alpha +a}(1+\alpha x)\exp (-ax)\right)
\left(1+\lambda-2 \lambda \left(
1-\frac{\alpha +a+\alpha ax}{\alpha +a}\exp (-ax)\right) \right).$
PLx: $f\left( x\right) =\alpha \beta \lambda \left( 1+\beta x\right)
^{-\alpha -1}\mathrm{\exp }\left[ -\lambda \left( 1+\beta x\right) ^{-\alpha
}\right] \left[ 1-\mathrm{\exp }(-\lambda )\right] ^{-1}.$
LiGc: $f\left( x\right) =\left[ 1-\left( 1+\frac{ax}{a+1}\right) \exp (-ax)\right] /\left[ 1-\alpha \left( 1+\frac{ax}{a+1}\right) \exp (-ax)\right] .$
Li: $f\left( x\right) =\frac{\lambda ^{2}}{\lambda +1}\left( 1+x\right)
\mathrm{\exp }(-\lambda x).$
The parameters of the above densities are all positive real numbers except $\left\vert \lambda \right\vert \leq 1$ for the TTLi distribution and $\alpha
\in \left( 0,1\right) $ for the LiGc distribution.
The numerical values of $-\widehat{\ell }$, $W^{\ast }$, $A^{\ast }$, KS, Bootstrapped PV, the MLEs and their corresponding standard errors (SEs) (given in parentheses) of the fitted models are listed in Tables \[tab:tab12\] and \[tab:tab13\], for both data sets, respectively. The figures in Tables \[tab:tab12\] and \[tab:tab13\] show that the GRL distribution has the lowest values for all goodness-of-fit statistics among all fitted models.\
Tables \[tab:tab14\] and \[tab:tab15\] display the parameter estimates under various estimation methods and the goodness-of-fit statistics for both data sets, respectively. From Table \[tab:tab14\] and based on the $K-S$ bootstrapped PV, we recommend to use the MPSE method to estimate the parameters of the GRL distribution for leukaemia data, while the OLS method is recommended to estimate the GRL parameters for epicenter data, based on the $K-S$ bootstrapped PV in Table \[tab:tab15\].
The histogram of the fitted GRL distribution and the other distributions are displayed in Figures \[fig:fig3\] and \[fig:fig4\] for the two data sets, respectively. Figures \[fig:fig3\] and \[fig:fig4\] show the plots of PDFs and CDFs of the fitted models for leukaemia and epicenter data. The HRF plot of the GRL distribution and the TTT plot of leukaemia data are displayed in Figure \[fig:fig5\], whereas the HRF plot of the GRL distribution and the TTT plot of epicenter data are displayed in Figure \[fig:fig6\]. It is shown that, the HRF is decreasing for leukaemia data, whereas the HRF is increasing for epicenter data. Furthermore, the scaled TTT plot for the leukemia data is convex which indicates a decreasing HRF and it is concave for epicenter data which indicates an increasing HRF. Then, the GRL distribution is a suitable for modeling leukaemia and epicenter data.
[.5]{} ![PDFs and CDFs of the fitted models for leukaemia data[]{data-label="fig:fig3"}](app_pdf.pdf "fig:"){width="1\linewidth"}
[.5]{} ![PDFs and CDFs of the fitted models for leukaemia data[]{data-label="fig:fig3"}](app_cdf.pdf "fig:"){width="1\linewidth"}
[.5]{} ![PDFs and CDFs of the fitted models for epicenter data[]{data-label="fig:fig4"}](app_2_pdf.pdf "fig:"){width="1\linewidth"}
[.5]{} ![PDFs and CDFs of the fitted models for epicenter data[]{data-label="fig:fig4"}](app_2_cdf.pdf "fig:"){width="1\linewidth"}
[.5]{} ![The HRF plot of the GRL distribution and TTT plot for leukaemia data[]{data-label="fig:fig5"}](app_hrf.pdf "fig:"){width="1\linewidth"}
[.5]{} ![The HRF plot of the GRL distribution and TTT plot for leukaemia data[]{data-label="fig:fig5"}](TTT_app.pdf "fig:"){width="1\linewidth"}
[.5]{} ![The HRF plot of the GRL distribution and TTT plot for epicenter data[]{data-label="fig:fig6"}](app_2_hrf.pdf "fig:"){width="1\linewidth"}
[.5]{} ![The HRF plot of the GRL distribution and TTT plot for epicenter data[]{data-label="fig:fig6"}](TTT_app_2.pdf "fig:"){width="1\linewidth"}
Concluding remarks {#sec6}
==================
In this paper, we introduce a new two-parameter distribution called generalized Ramos-Louzada (GRL) distribution. Further, the mathematical properties of the GRL model are studied in detail. The GRL parameters are estimated by eight estimation methods namely: the weighted least-squares, ordinary least squares, maximum likelihood, maximum product of spacing, Cramér–von Mises, Anderson–Darling, Right-tail Anderson–Darling and percentile based estimators. The simulation study illustrates that the maximum product of spacing estimation method outperforms all other estimation methods. Therefore, depends on our study, we can confirm the superiority of the maximum product of spacing method for the GRL distribution. Finally, the practical importance of GRL model was reported in two real applications. The goodness of fit for the proposed data sets showed that our model returned better fitting in comparison with other well-known distributions. Further, the two real data applications show that the maximum product of spacing estimator for leukemia data and least-square estimator for epicenter data return the best estimates for the parameters of the GRL distribution.
[999]{} Abouelmagd, T. H. M., Hamed, M. S. , Afify, A. Z., Al-Mofleh, H. and Iqbal, Z. (2018). The Burr X Fréchet distribution with its properties and applications. Journal of Applied Probability and Statistics, 13, 23–51 Al-Zahrani, B. and Sagor, H. (2014). The poisson-lomax distribution. Revista Colombiana de Estad stica, 37, 225-245. Castillo, E. (2012). Extreme value theory in engineering. Elsevier.
Cheng, R. and Amin, N. (1979). Maximum product of spacings estimation with application to the lognormal distribution (mathematical report 79-1). University of Wales IST. Cheng, R. and Amin, N. (1983). Estimating parameters in continuous univariate distributions with a shifted origin. Journal of the Royal Statistical Society. Series B (Methodological), 45, 394-403. Feigl, P. and Zelen, M. (1965). Estimation of exponential probabilities with concomitant information. Biometrics, 21, 826-838. Ghitany, M., Al-Mutairi, D., Balakrishnan, N. and Al-Enezi, L. (2013). Power lindley distribution and associated inference. Comput. Stat. Data Anal., 64, 20–33. Gupta, R. D. and Kundu, D. (2001). Exponentiated exponential family: an alternative to gamma and Weibull distributions. Biometrical Journal: Journal of Mathematical Methods in Biosciences, 43, 117-130. Jones, M. C. (2004). Families of distributions arising from distributions of order statistics. Test, 13, 1-43. Kao, J. (1958). Computer methods for estimatingweibull parameters in reliability studies. IRE Reliability Quality Control, 13, 15-22. Kao, J. (1959). A graphical estimation of mixed weibull parameters in life testing electron tube. Technometrics, 1, 389-407. Kemaloglu, S. A. and Yilmaz, M. (2017). Transmuted two-parameter Lindley distribution. Commun. Stat. Theory Methods, 46, 11866-11879. Lindley, D. V. (1958). Fiducial distributions and Bayes theorem. Journal of the Royal Statistical Society, 20, 102-107. Luceño, A. (2006). Fitting the generalized pareto distribution to data using maximum goodness-of-fit estimators. Computational Statistics and Data Analysis, 51, 904-917. Macdonald, P. (1971). Comment on “an estimation procedure for mixtures of distributions” by choi and bulgren. Journal of the Royal Statistical Society. Series B (Methodological), 33, 326-329. Marshall, A. W. and Olkin, I. (1997). A new method for adding a parameter to a family of distributions with application to the exponential and Weibull families. Biometrika, 84, 641-652. Migon, H. S., Gamerman, D., and Louzada, F. (2014). Statistical Inference: An Integrated Approach. New York: CRC press. Nadarajah, S., Bakouch, H. S. and Tahmasbi, R. (2011). A generalized Lindley distribution. Sankhya B, 73, 331-359. Nassar, M., Afify, A. Z., Dey, S. and Kumar, D. (2018). A new extension of Weibull distribution: properties and different methods of estimation. Journal of Computational and Applied Mathematics, 336, 439-457. Ramos, P. L. and Louzada, F. (2019). A distribution for instantaneous failures. Stats, 2, 247-258. Ranneby, B. (1984). The maximum spacing method. an estimation method related to the maximum likelihood method. Scandinavian Journal of Statistics, 11, 93-112. Sen, S., Afify, A. Z., Al-Mofleh, H. and Ahsanullah, M. (2019). The quasi Xgamma-geometric distribution with application in medicine. Filomat, 33, 5291–5330. Stacy, E. W. (1962). A generalization of the gamma distribution. The Annals of mathematical statistics, pages 1187-1192. Swain, J. J., Venkatraman, S., and Wilson, J. R. (1988). Least-squares estimation of distribution functions in johnson’s translation system. Journal of Statistical Computation and Simulation, 29, 271-297. Weibull, W. (1951). A statistical distribution function of wide applicability. Journal of Applied Mechanics,18, 293-297. Zakerzadeh, H. and Mahmoudi, E. (2012). A new two parameter lifetime distribution: model and properties. arXiv preprint arXiv:1204.4248.
Appendix A: Tables {#appendix-a-tables .unnumbered}
==================
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We thoroughly investigate vibrational quantum dynamics of dimers attached to He droplets motivated by recent measurements with K$_2$ [@Claas_1151_2006]. For those femtosecond pump-probe experiments, crucial observed features are not reproduced by gas phase calculations but agreement is found using a description based on dissipative quantum dynamics, as briefly shown in [@Schle_245_2010]. Here we present a detailed study of the influence of possible effects induced by the droplet. The helium droplet causes electronic decoherence, shifts of potential surfaces, and relaxation of wave packets in attached dimers. Moreover, a realistic description of (stochastic) desorption of dimers off the droplet needs to be taken into account. Step by step we include and study the importance of these effects in our full quantum calculation. This allows us to reproduce and explain all major experimental findings. We find that desorption is fast and occurs already within $2-10$ ps after electronic excitation. A further finding is that slow vibrational motion in the ground state can be considered frictionless.'
address: 'Institut für Theoretische Physik, Technische Universität Dresden, D-01062 Dresden, Germany'
author:
- 'Martin Schlesinger, Walter T. Strunz'
title: ' Detailed study of dissipative quantum dynamics of K$_2$ attached to helium nanodroplets'
---
\[sec:introduction\]Introduction
================================
Helium nanodroplet isolation (HENDI) spectroscopy allows to study atoms, molecules and clusters embedded in an ideal cryogenic environment [@Toenn_2622_2004]. Ultracold helium droplets ($0.38\mathrm{\,K}$) provide a gentle, since weakly disturbing host for embedded species, which can be studied with high resolution spectroscopy [@Stien_127_2006]. However, spectra of immersed species are slightly broadened and shifted away from their gas phase value due to their interaction with the surrounding He droplet. Inhomogeneous broadening in pure rotational spectra is ascribed to the motion of the purity inside the droplet [@Calle_4636_2000; @Lehma_645_1999] or to coupling to collective degrees of freedom of the droplet [@Calle_10090_2001]. The overall spectral features, which are unseen in gas phase measurements, have been used to study the weak interaction between dopant and helium droplet [@Reho_161_1997; @Grebe_2083_1998; @Calle_5058_1999; @Calle_4636_2000; @Nauta_45084514_2001; @Dick_10206_2001]. Attached species allow to probe the peculiar properties of the superfluid He droplet itself [@Hartm_4560_1996]. Further insight into the interaction and the quantum properties of the droplet is gained from recently obtained time-resolved studies [@Claas_1151_2006; @Schle_245_2010; @Mudri_42512_2009; @Grune_6816_2011].
Helium nanodroplets, typically consisting of several thousand $^4\mathrm{He}$ atoms, are ideally suited to study relaxation (cooling) of embedded species [@Koch_35302_2008; @Braun_253401_2004; @Przys_21202_2008; @Korni_1437_2010; @Dropp_233402_2004]. Whether dissipation plays a role depends on the involved energy scales, the coupling strengths, or on typical time scales in the system and “bath” [@Nitza_200_1975; @Ewing_4662_1987]. The group of Miller has studied relaxation of the HF system inside He droplets [@Nauta_9466_2000] . They report on ineffective vibrational relaxation due to a mismatch in energy scales. However, rotational relaxation in the immersed HF system is fast and appears as Lorentzian line broadening in the rotational transition spectrum. (see also [@Calle_10090_2001] and references therein).
Alkali metal atoms and molecules are known to reside in bubble-like structures on the surface of He droplets [@Dalfo_61_1994; @Ancil_16125_1995; @Stien_3592_1995]. Attached Rb$_2$ dimers in the triplet state reveal the presence of vibrational relaxation on the measurement timescale [@Higgi_4952_1998; @Grune_6816_2011]. For Lithium dimers interacting with a He environment, vibrational relaxation has been investigated by means of Monte Carlo calculations [@Bovin_224312_2008], including a few He atoms. The corresponding relaxation rates depend on whether the dimer is orientated in-plane or out-of-plane with respect to the He surface [@Bovin_224903_2009].
Femtosecond pump-probe techniques are established tools to analyze the ultrafast vibrational motion in molecular systems [@Zewai_12427_1993; @Grueb_459_1990; @Baume_639_1992]. A first laser pulse excites a coherent wave packet, which is allowed to freely evolve on the respective energy surface. The WP is probed by a time-delayed pulse. Since the early studies by the Zewail group [@Bowma_297_1989; @Grueb_883_1993], vibrational wave packets in various diatomic systems have been studied, such as $\mathrm{\,Na}_2$ [@Baume_8103_1991] or $\mathrm{\,K}_2$ [@Vivie_7789_1996; @Rutz_9_1997; @Nicol_7857_1999]. When molecules are located in a solid rare-gas matrix [@Karav_814_2003; @Guhr_5353_2004], a suppression of revival structures in the pump-probe signal indicate a loss of vibrational coherence. Meier et. al. have thoroughly investigated such decoherence in molecular systems placed in a rare-gas environment [@Meier_4_2004], motivated by seminal experiments in the Zewail group [@Liu_18666_1996]. For experiments at room temperature with a significant thermal occupation of rotational states, unavoidable coupling between internal vibrational and rotational degrees of freedom may lead to sufficient decoherence to suppress (fractional) revivals [@Schle_12111_2008]. Experiments with Rb$_2$ dimers on He droplets reveal an ongoing decay of the pump-probe ion yield [@Mudri_42512_2009]. The decay has been ascribed to damping and accompanying decoherence of vibrational wave packets [@Grune_6816_2011]. In this work we investigate potassium dimers on He droplets, studied experimentally with the pump-probe technique [@Claas_1151_2006]. A brief account of a theoretical description based on dissipative quantum dynamics was given in [@Schle_245_2010]. Experimental spectra show significant deviations from corresponding gas phase calculations for the unperturbed dimer.
In this paper, in a phenomenological approach, we investigate in detail how the helium influence may be described effectively. First, we see that the helium environment destroys electronic coherence, which is imprinted by the exciting laser pulse. As possible causes for electronic decoherence, we consider a distribution of shifts of electronic surfaces. Electronic decoherence alone cannot account for the decay of the signal at certain excitation wavelengths. Therefore, in a next step, we include a general damping of vibrational wave packets on each electronic surface. The effective dynamics is described by means of a quantum optical master equation.
Moreover, it is important to take into account the desorption of dimers from the droplet properly. No general rule can be given when attached atoms or molecules leave the droplet. It is known that lighter alkali metal atoms leave the droplet upon electronic excitation [@Stien_10119_2001] or form a bound exciplex [@Reho_161_1997a]. The exciplex tends to desorb off the droplet surface during the formation process or several picosecond thereafter [@Schul_153401_2001]. On the other hand, desorption of Rb atoms may be completely inhibited upon electronical excitation in a certain laser wavelength range [@Auboc_35301_2008]. Recent measurements indicate that K$_2$ molecules desorb several picoseconds after laser excitation [@Claas_1151_2006]. Indeed, (stochastic) desorption of dimers are a crucial ingredient to explain spectral features with a theoretical model [@Schle_245_2010].
We here report on a more realistic description of desorption - we use a model which only allows electronically excited molecules to leave the droplet. We thus extend our previous, state-independent desorption scheme [@Schle_245_2010]. Together with dissipation and shifts of potential energy surfaces, one can explain experimental findings over the full laser excitation range reported in [@Claas_1151_2006].
The article is organized as follows: In section \[sec:pump-probe-signal-1\] we review the calculation of the pump-probe signal for free dimers. In the following section \[sec:he-influence\], the influence of the helium environment is considered in a phenomenological way. In section \[sec:state-indep-desorpt\] we explain how we treat desorption of dimers. Finally, for an even better agreement with experiment, we consider undamped motion in the electronic ground state. A thorough comparison between theoretical and experimental findings is given at every step in the model. Section \[sec:conclusion\] is devoted to the conclusions.
\[sec:pump-probe-signal-1\]Pump-probe signal
============================================
![\[fig:exitationscheme\] Excitation schemes in the potassium dimer for two distinct laser wavelengths $\lambda=833\mathrm{\, nm}$ (scheme I) and $\lambda=800\mathrm{\, nm}$ (scheme II). While the first scheme exclusively maps the WP in the $\mathrm{\,A}\, ^1\Sigma^+_u$ state, the second scheme allows to observe the WP in the ground state $\mathrm{\,X}\, ^1\Sigma^+_g$. ](fig1.jpg){width=".49\textwidth"}
A first “pump” laser pulse creates a vibrational wave packet (WP) $| \psi_i \rangle$ in some electronic state $|i\rangle$. The excited WP oscillates in a region between classical inner and outer turning point of that surface. It may periodically enter and leave a transition region, where a resonance condition with higher lying states is met and the potential energy difference matches the second, “probe” pulse energy. This region defines the so-called Franck-Condon (FC) window. The time-delayed probe pulse leads to a significant number of ions only when the WP is located in the FC window. By varying the time delay $\tau$ between pump-and probe pulse, one obtains an oscillatory ion yield. A typical excitation scheme is depicted in fig. \[fig:exitationscheme\](a), where one probes the dynamics of the WP on the $\mathrm{\,A}\, ^1\Sigma^+_u$ surface.
Wave packets may be excited in several involved electronic surfaces of the dimer. More specifically, the pump pulse prepares the dimer in a superposition of electronic states, such that the full state vector $| \Psi\rangle$ takes the form $$\label{eq:16}
| \Psi(t_{\mathrm{pump}}) \rangle =\sum_i |\psi_i \rangle |i \rangle .$$ Here, $\psi_i$ denotes the WP on a specific electronic state $i$ and $$\label{eq:20}
p_i= \langle \psi_i | \psi_i \rangle$$ is the probability that the electronic state $i$ is excited.
For the pump-probe signal, we fully numerically solve the time-dependent Schrödinger equation (TDSE)
$$\begin{aligned}
\label{eq:9}
i \hbar \frac{\partial} {\partial t} \left (
\begin{array}{ccccc } \psi_0 \\ \psi_1 \\ \psi_2 \\ \psi_3(E)
\end{array}
\right) \!\!
= \!\! \left( \!\!
\begin{array}{ccccc }
H_0 & J_{01} & 0 & 0 \\
J_{10} & H_1 & J_{12} & 0 \\
0 & J_{21} & H_2 & J_{23}^{I} \\
0 & 0& J_{32}^{I} & H_{3,E}
\end{array} \!\! \right) \!\! \left (
\begin{array}{ccccc } \psi_0 \\ \psi_1 \\ \psi_2 \\ \psi_3(E)
\end{array}
\right), \end{aligned}$$
for the full state vector $| \Psi\rangle$ (see also [@Vivie_16829_1995; @Vivie_7789_1996]). In the matrix equation, the diagonal elements denote the molecular Hamiltonian $H_\mathrm{mol}=\sum_i H_i= T + \sum_i V_{i}$, which involves kinetic energy $T = P^2/2\mu$ (reduced mass $\mu$) and adiabatic potentials $V_{i}$. In the final ionic state, the energy $E$ of the ejected electron is included in the diagonal entry of the Hamiltonian $H_{3,E}$, such that $V_{i=final}=V_3+E$. The coupling to the laser field is described by the matrix elements $J_{ij}=-\vec{\mu}_{ij}\cdot \vec{E}(t)$, where $\vec{\mu}_{ij}$ denotes the transition dipole moment. Both pump- and probe pulse have the form $\vec{E}_\mathrm{pump/probe}(t) =
\vec{\epsilon}_0\varepsilon(t)\cos{(\omega_L t)}$ and $\omega_L$ is the respective laser frequency. Moreover, $\vec{\epsilon}_0$ is the polarization and $\varepsilon(t)$ the shape function of the field, which is assumed to be Gaussian. For the field parameters we use a full width at half maximum of the laser pulse of $110 \mathrm{\, fs}$ and an intensity of $1.2 \mathrm{\, GW}/\mathrm{cm}^2$. The employed intensity is higher than the experimentally estimated value (I=$0.5 \mathrm{\, GW}/\mathrm{cm}^2$ [@Claas_1151_2006]), but, according to [@Vivie_7789_1996], still located in the moderate power regime. The ionic state $\psi_{3}(E)$ also depends on the energy $E$ of the ejected electron. We use discretization of the (electronic) continuum, a technique successfully employed earlier [@Vivie_16829_1995]. One determines the final state probability $|\psi_{3} (E_k)|^2$ after both pulses have passed for distinct electronic energies $E_k$. The pump-probe signal is proportional to the sum over different electronic contributions, $$\label{eq:13}
S(\tau) = \lim_{t\rightarrow \infty} \sum_{E_k} |\psi_3 (\tau,E_k)|^2.$$ In the limit $t \rightarrow \infty$ (upon complete decay of the second pulse), it only depends on the delay $\tau$ between the pulses. For the propagation of the wave function, we use the split-operator method [@Feit_412_1982].
The ion signal is composed of beat frequencies $\omega_{v v'} \equiv (E_v - E_{v'})/ \hbar$ between all pairs of energy levels that contribute to the WP [@Grune_6816_2011]. The most prominent oscillation originates from the energy spacing between central and neighboring vibrational levels $\bar{v}$ and $\bar{v} \pm 1$. This oscillation has the frequency $$\label{eq:12}
\omega_i \equiv \omega_{\bar{v},\bar{v}+1}$$ and is characteristic for the electronic surface $i$. In an anharmonic potential, the level spacing and therefore also $\omega_i$ decreases as $\bar{v}$ increases. Higher-order frequency components $\omega_{\bar{v},\bar{v}+\Delta v}$ with $\Delta v > 1$ are visible in the Fourier transform (FT) of the signal.
The laser wavelength $\lambda$ determines which wave packets $\psi_i$ can be mapped to the final state. In the one-color pump-probe setup, we consider two different excitation schemes.
For $820\mathrm{\, nm} \lesssim \lambda \lesssim 840\mathrm{\, nm}$ (scheme I), one exclusively follows the vibrational dynamics in the excited state $\mathrm{\,A}\, ^1\Sigma^+_u$. The WP in that state can be probed at the outer turning point (see fig. \[fig:exitationscheme\](a)). Contributions from other surfaces are negligible. For this scheme we concentrate on an excitation wavelength $\lambda=833\mathrm{\, nm}$.
![\[fig:PP\_833\_gphase\_exp\]Pump-probe ion yield at $\lambda=833\mathrm{\, nm}$. (a) Numerical gas phase calculation. (b) Experimental result (from [@Claas_1151_2006]). The oscillation is exclusively attributed to the circulating WP in the $\mathrm{\,A}\, ^1\Sigma^+_u$ state. ](fig2.jpg){width=".49\textwidth"}
Fig. \[fig:PP\_833\_gphase\_exp\](a) shows the theoretical gas phase calculation. Fig. \[fig:PP\_833\_gphase\_exp\](b) shows experimental spectra at this wavelength obtained from dimers attached to helium nanodroplets [@Claas_1151_2006]. In the experiment, the signal amplitude significantly decreases, but oscillates with nearly constant amplitude at later delay times.
![\[fig:FT\_800nm\_gphase\_exp\] Spectra $\mathcal{F}(\omega,\tau)$ in the time-frequency domain at $\lambda=800\mathrm{\, nm}$. (a) Numerical gas phase calculation, where electronic interferences lead to a spectrum that is difficult to interpret. (b) Experimental HENDI result (from [@Claas_1151_2006]). Electronic interferences are absent, allowing for a clear identification of the structures of the spectrum. ](fig3.jpg){width=".49\textwidth"}
For $800 \lesssim \lambda \lesssim 820\mathrm{\, nm}$ (scheme II), transitions preferably take place at the inner turning point of the excited WP. The wave packet in the electronic ground state $\mathrm{\,X}\, ^1\Sigma^+_g$ is excited through resonant impulsive stimulated Raman scattering (RISRS). It can be probed through a 3-photon process, see fig. \[fig:exitationscheme\]. On the other hand, simultaneous and coherent contributions from WPs in electronic excited states lead to constructive or destructive interferences and thus to an unstructured ion yield [@Nicol_7857_1999]. We use sliding window Fourier transforms $\mathcal{F}(\omega, \tau$) to follow the evolution of respective beat frequency components as a function of the delay time [@Fisch_331_1996; @Vrakk_37_1996]. Fig. \[fig:FT\_800nm\_gphase\_exp\](a) shows a window transform (spectrogram) for excitation at $\lambda=800\mathrm{\, nm}$. A frequency component $\omega_A \approx 63 \mathrm{\,cm}^{-1}$ can be ascribed to the WP in the state $\mathrm{\,A}\, ^1\Sigma^+_u$. Moreover, a higher-order frequency component around $\approx 2\omega_A$ is visible as is a significant contribution at $\omega \approx 85\mathrm{\, cm}^{-1}$ after about 15 ps that can be traced back to originate from the $2 ^1\Pi_g$ surface. A contribution from the ground state and resulting frequency component $\omega_X$, however, is missing. In agreement with earlier findings [@Nicol_7857_1999; @Vivie_7789_1996], the mapping of the ground state WP is not possible at low to intermediate intensities in the gas phase. This is because, as explained below, contributions from different potential energy surfaces interfere destructively. In contrast, for dimers attached to He droplets, the contribution from the state $\mathrm{\,A}\, ^1\Sigma^+_u$ is suppressed after about 8 ps, while on the other hand the vibrational ground state WP is clearly resolved, see fig. \[fig:FT\_800nm\_gphase\_exp\](b).
Qualitatively, the difference between fig. \[fig:FT\_800nm\_gphase\_exp\](a) (gas phase) and fig. \[fig:FT\_800nm\_gphase\_exp\](b) (experiment) can be explained by the loss of electronic coherence alone: In fig. \[fig:proj\_scheme\], we construct an artificial signal from the incoherent sum of contributions. The full coherent wave vector after decay of the pump pulse at $t\equiv t_\mathrm{pump}$, eq. (\[eq:16\]), serves as a starting point. In order to determine the contribution of a single WP $\psi_i$, we project the fully coherent wave vector according to $\tilde{\Psi}_i(t_\mathrm{pump})= P_i| \Psi(t_\mathrm{pump})\rangle $ with the projector $P_i=|i \rangle \langle i|$ on one of the electronic states. The electronic occupation $p_i$ is not altered and the usual probe scheme is employed after projection. The resulting spectrogram $\mathcal{F}^i(\omega, \tau)$ after projection on the state $\mathrm{\,A}\, ^1\Sigma^+_u$ and $\mathrm{\,X}\, ^1\Sigma^+_g$, respectively, is shown in fig. \[fig:proj\_scheme\](a)/(b). The incoherent sum of contributions is given by $$\label{eq:17}
\mathcal{F}^{\mathrm{inc}}(\omega,\tau)=\sum_i \mathcal{F}^i(\omega, \tau).$$ Contrary to the coherent signal, fig. \[fig:FT\_800nm\_gphase\_exp\](a), in eq. (\[eq:17\]) interferences are removed by hand and contributions from ground- and excited state WPs are clearly visible. Obviously, the incoherent sum fig. \[fig:proj\_scheme\](c) already reproduces important features of the experimental data.
[ ![\[fig:proj\_scheme\] Shown is a the spectrogram $\mathcal{F}^i(\omega,\tau$), which is obtained upon projecting the created full wave function $\Psi(t)$ on the electronic ground state ($i=0$) and (b) first excited state ($i=1$). (c) In the incoherent sum, interferences are excluded. In particular, the ground state component is clearly resolved. ](fig4.jpg "fig:"){width=".49\textwidth"}]{}
In the experiment (fig. \[fig:FT\_800nm\_gphase\_exp\](b)), however, the component $\omega_A$ fades out at later delay times, while the component $\omega_X$ becomes dominant. Therefore, a model based on pure loss of electronic coherence alone cannot explain the measured spectrum. One has to consider additional influences of the He environment on the dimer dynamics.
He influence and results {#sec:he-influence}
========================
In a phenomenological approach, we take into account three possible effects:
1. A He-induced energetic shift of electronic potential energy surfaces, with possibly small fluctuations.
2. Damping of vibrational wave packets. This effect is treated fully quantum mechanically within the master equation approach. The WP $\psi_i$ dissipates energy with a certain damping rate $\gamma_i$. This rate is here seen as a fit parameter and is adjusted to experimental observations.
3. Desorption of dimers off the droplet. After desorption, the influence of the helium droplet (shift/damping) vanishes.
A description based on damping of vibrational wave packets has been applied to HENDI studies with spin triplet Rb$_2$ dimers [@Grune_6816_2011]. There, slow vibrational decoherence (as a consequence of very weak dissipation) is most relevant for the decay of the revival amplitude. In particular, as witnessed by the ongoing decay, desorption of dimers seems to be inhibited.
In the following, step by step, we include shift, damping and finally desorption in the calculation of pump-probe spectra. It is found that all model “ingredients” 1. - 3. are crucial to find agreement with obtained experimental spectra.
Electronic shifts {#sec:electronic-shifts}
-----------------
The He environment may lead to shifts of electronic surfaces in attached dimers. Such shifts are common for alkali atoms and molecules on He droplets [@Higgi_4952_1998]. Spectra from attached species are shifted relative to what one expects from gas-phase potential energy curves [@Stien_253_1996; @Allar_1169_2006; @Bruhl_10275_2001; @Auboc_54304_2010; @Buener_12684_2007]. For certain weakly coupled species, spectra are only shifted by a few wavenumbers [@Higgi_4952_1998]. Denoting the shift of surface $i$ with $\Delta_i$, the full state vector is propagated according to $$\label{eq:7}
|\dot{\Psi}(t) \rangle = -\frac{i}{ \hbar} H |\Psi(t) \rangle + \frac{i}{ \hbar} \sum_i \Delta_i | \psi_i \rangle$$ Both surfaces $\mathrm{\,A}\, ^1\Sigma^+_u$ and 2$^1 \Pi_g$ may be affected. Only relative shifts are relevant for the signal, such that we cannot differentiate between a shift of the electronic ground - or first excited state. Also, a shift of the ionic surface would change the energy distribution of the ejected electrons, but not the overall calculated ion yield. This is because we sum over all electronic energies in eq. (\[eq:13\]). We find that even large, but fixed shifts of up to $\pm 100 \mathrm{\, cm}^{-1}$ have a very small influence and resemble the gas phase calculation.
However, it is reasonable to assume that the shifts $\Delta_i$ fluctuate for the following reason: The number of He atoms of a droplet is not fixed but varies according to a log-normal distribution [@Lewer_381_1993]. Consequently, we have to average over a distribution of shifts $\Delta_i$ [@Login__2008]. Also, for the considered laser intensities, electronic surfaces are slightly Stark-shifted. Depending on their position in the laser beam, dimers are exposed to a distribution of laser intensities [@Mudri_priv]. The beam width leads to a distribution of Stark shifts, which in turn has to be treated as random distribution of electronic shifts $\Delta_i$.
Considering fluctuating potential energies, eq. (\[eq:7\]) represents only a single realization $|\Psi_j(t)\rangle$ with shifts $\Delta_i(j)$ in the ensemble. The pump-probe signal is proportional to the ensemble average $$\label{eq:14}
\langle S(\tau) \rangle = \frac{1}{N} \sum_j \left [\lim_{t\rightarrow \infty} \sum_{E_k} |\psi_{j,3} (\tau,E_k)|^2\right ],$$ where $N$ realizations have been taken into account. In this expression, the final state $ |\psi_{j,3} \rangle$ is obtained from propagation of the realization $|\Psi_j(t)\rangle$, in which $\Delta_i(j)$ is chosen randomly. Through the ensemble average in eq. (\[eq:14\]), one obtains an incoherent mixture of electronic contributions in the final state. The spectrogram of $\langle S(\tau)\rangle$ nearly perfectly resembles the incoherent sum of contributions, see eq. (\[eq:17\]) and fig. \[fig:proj\_scheme\](c). It does not contain any electronic coherence, such that the frequency components $\omega_i$ are clearly resolved. For full decoherence fig. \[fig:proj\_scheme\](c), random shifts in the range of $\pm 5\mathrm{\, cm}^{-1}$ around the average shifts $\overline{\Delta_1}=0\mathrm{\, cm}^{-1}$ and $\overline{\Delta_2}=-50\mathrm{\, cm}^{-1}$ are sufficient. Note that this decoherence is due to the inhomogeneous size distribution and not due to entanglement between system and environment [@Schlo__2007; @Helm_42108_2009].
To conclude, randomly distributed energy shifts can explain the visibility of the ground state component $\omega_X$. However, this effect lacks an explanation for the decay of the frequency component $\omega_A$ of the excited state, which is observed in the HENDI experiment. Therefore, we consider vibrational damping next.
Damped vibrational wave packets {#sec:damp-vibr-wave}
-------------------------------
Dissipation can be treated fully quantum mechanically by using approximate master equations for the density operator of the (reduced) system. The overall dissipation of an excited WP originates in our case from the interaction of vibrational degrees of freedom of the molecule with collective degrees of freedom of the helium droplet. Note that in recent studies [@Bovin_224903_2009] vibrational relaxation rates for alkali dimers on $^4 \mathrm{He}$ clusters are estimated from full quantum Monte Carlo calculations for a few He atoms. These rates turn out to be roughly of the same order of magnitude as our phenomenologically chosen damping rates below.
In our approach, the density $\rho$ of the damped WP in an electronic state $|i \rangle$ evolves according to the master equation $$\label{eq:3}
\dot{\rho} = - \frac{i}{\hbar} [H_i,\rho]
+\gamma \left( a \rho a^{\dagger}
-\frac{1}{2} a^\dagger a \rho - \frac{1}{2} \rho a^\dagger a \right ),$$ which is of Lindblad form [@Lindb_147_1975]. It describes friction for near-harmonic systems at effectively zero temperature in the rotating wave approximation (see, for instance, [@Scull__1997]). The first term in eq. (\[eq:3\]) contains the molecular Hamiltonian $H_i$ of a specific electronic state $|i\rangle$ and determines the unitary evolution of the WP $| \psi_i \rangle$. Relaxation of the vibrational WP is achieved through the second, irreversible contribution. Any initial state approaches the ground state on a time scale $\gamma^{-1}$, i.e.$\gamma$ denotes the damping rate. In eq. (\[eq:3\]), $a, a^\dagger$ are the creation/annihilation operators of a harmonic oscillator, defined through $$\label{eq:10}
a = \frac{1}{\sqrt{2}}\left ( \sqrt{\frac{\hbar}{m\omega_e }}\hat{X}
+ i\frac{1}{\sqrt{\hbar m \omega_e }} \hat{P} \right).$$ Here, $\hat{X}$ and $\hat{P}$ are the usual position and momentum operator w.r.t. the harmonic oscillator minimum and $\omega_e$ its frequency. The quantum optical master equation (\[eq:3\]) can be derived from a von-Neumann equation for the full system and is valid only for weak couplings between system and “bath” (which is the helium droplet here). Also, in the derivation one makes use of the Markov and rotating wave approximation. For a significant temperature, additional terms that describe thermal excitations from the environment have to be taken into account [@Breue__2002]. The damping constant $\gamma$ may in principle be derived from a microscopic description of interaction between system and bath (Fermi’s Golden Rule). We do not specify this interaction, but use the damping rate $\gamma$ as a fit parameter to obtain agreement with experimental data. The master equation (\[eq:3\]) induces the evolution of a pure initial state into a state mixture.
[ ![\[fig:WPs\_833nm\] (a) Shown is single realization of a wave function in position space. The WP circulates in the A state and is damped with a rate $\gamma_1=0.3/\mathrm{ps}$. (b) Coordinate expectation value for that single realization. The noise is due to the stochastic propagation. (c) Potential curves A (black) and 2$\Pi$ (red). The latter is shifted by one photon energy ($\lambda=800\mathrm{\, nm}$) to clarify possible transitions to the ionic state (arrows). A fully damped WP can be mapped to the final state. ](fig5.jpg "fig:"){width=".49\textwidth"}]{}
For the numerical propagation, we return from the density matrix description to a Schrödinger-type equation for the state vector. It is not possible to evolve a pure state into a mixed state with a deterministic Schrödinger equation. One therefore considers a stochastic differential equation for a state vector, quantum state diffusion (QSD) [@Gisin_5677_1992]. The density is recovered from the average over several realizations of state vectors. Given a master equation in “Lindblad form”, as in our case, it is straightforward to state the corresponding quantum state diffusion (QSD) Ito stochastic Schrödinger equation for a state vector [@Gisin_5677_1992]: $$\begin{aligned}
\label{eq:4}
|d\psi(t) \rangle = & -\frac{i}{\hbar} H_i |\psi \rangle dt \\ \nonumber
& + \frac{1}{2} \gamma \left(2 \langle a^\dagger \rangle a
- a^\dagger a - |\langle a \rangle|^2 \right)
|\psi \rangle dt
+ \sqrt{\gamma}(a- \langle a \rangle ) |\psi \rangle d\xi(t). \end{aligned}$$ The left hand side means $|d\psi(t) \rangle = |\psi(t+dt) \rangle -
|\psi(t)\rangle$, i.e. the change of the state after a time increment $dt$. The second term induces transitions to lower lying vibrational states. The third term is stochastic and contains complex normalized Ito increments $d\xi(t)$, which satisfy $$\begin{aligned}
\label{eq:5}
d\xi^2 & = (d\xi^*)^2 = 0 \nonumber \\
d\xi d\xi^* & = dt .\end{aligned}$$ The increment of the density up to second order in $dt$ is given through $ d \rho = | d \psi \rangle \langle \psi| +
|\psi \rangle \langle d \psi| + | d \psi \rangle \langle d \psi |$. Since eq. (\[eq:4\]) is written in Ito form, one can easily take the average with respect to (\[eq:5\]) and prove the equivalence with the given master equation (\[eq:3\]). Recovering the density $\rho=\overline{ | \psi(t) \rangle \langle \psi(t) | }$ from several realizations of $|\psi\rangle$ amounts to obtaining expectation values through $\langle A \rangle = \mathrm{Tr} ( A \rho ) = \overline{ \langle \psi | A | \psi \rangle} $. In fig. \[fig:WPs\_833nm\] we show the coordinate expectation value $\langle R \rangle$, obtained from a single realization of $| \psi \rangle$. The norm of the state in eq. (\[eq:4\]) is conserved, i.e.$d (\langle \psi | \psi \rangle) = 0$. However, due to the finite time step $\Delta t$, the norm can slightly fluctuate. Therefore, for the numerics, we impose norm preservation by renormalizing the state vector after every time step. After replacing the ladder operators in eq.(\[eq:4\]) through their definition in (\[eq:10\]), the r.h.s. is strictly separable in operators, which act in either momentum or coordinate space. Therefore, the split operator method [@Feit_412_1982] for the propagation of the WP can still be used, which is an advantage of this approach.
Helium induced damping of vibrational wave packets on a [*single*]{} electronic surface is described through eq. (\[eq:4\]), which is equivalent to the master equation eq. (\[eq:3\]) on average. In order to obtain damping on several electronic surfaces, we propagate a full state vector $|\Psi(t)\rangle $ according to $$\label{eq:15}
|d\Psi(t) \rangle = -\frac{i}{ \hbar} H |\Psi(t) \rangle dt + \underbrace{\sum_j\left[ D(\gamma_j) + \frac{i }{ \hbar} \Delta_j \right] | \psi_j \rangle}_{\mathrm{\,coupling\,to\,He\,bath}}$$ The average is taken over several realizations of state vectors to recover the density via $ \rho(t)= \overline{ | \Psi(t) \rangle \langle \Psi(t) | }$. In eq. (\[eq:15\]), $D(\gamma_j)$ is the generalization of the r.h.s. of eq. (\[eq:4\]) to obtain damping and accompanying fluctuations on an arbitrary surface, $$\label{eq:1}
D(\gamma_j) = \frac{1}{2} \gamma_j \left(2 \langle a_j^\dagger \rangle a_j
- a_j^\dagger a_j - |\langle a_j \rangle|^2 \right) dt
+ \sqrt{\gamma_j}(a_j- \langle a_j \rangle ) d\xi_j(t)$$
In eq. (\[eq:1\]), all stochastic differential Wiener increments $d \xi_j$ are taken independently of each other. In the ladder operator of surface $j$, one uses the the position w.r.t. the respective harmonic oscillator minimum and its frequency. For most parts of the following, for simplicity, we set damping constants to be equal $\gamma_j=\gamma$. However, we hasten to add that an interesting exception is provided in section \[sec:undamp-ground-state\]. The agreement with experiment improves, if one allows for [*undamped*]{} vibrational motion in the electronic ground state, $\gamma_0=0$.
The damping model provides an explanation for the signal decrease at $\lambda=833\mathrm{\, nm}$ (excitation scheme I). After several periods, vibrationally damped WPs on the A surface no longer enter the initial FC region and therefore the ion yield decreases.
[ ![\[fig:FT\_833nm\_SD\] Spectra $\mathcal{F}(\omega,\tau)$ in the time-frequency domain at $\lambda=833\mathrm{\, nm}$. (a) Full damping. (b) Damping and state-dependent desorption. Also shown is the frequency upon averaging in the lined frequency interval. Explanation see text. ](fig6.jpg "fig:"){width=".49\textwidth"}]{}
[ ![\[fig:PP\_FT\_SD\_800nm\] Calculated spectra in the time-frequency domain at $\lambda=800\mathrm{\, nm}$. (a) Full damping: The WP in the A state leaves the initial transition region, consequently the component $\omega_A$ decreases. The fully damped WP, however, reaches a transition region at the equilibrium distance, such that $\omega_A$ returns. See also fig. \[fig:WPs\_833nm\]. (b) State-dependent desorption model. (c) State-dependent desorption model together with undamped motion in the ground state. ](fig7.jpg "fig:"){width=".49\textwidth"}]{}
In fig. \[fig:FT\_833nm\_SD\] we depict the result of the damping model eq. (\[eq:15\]), using $\gamma_1=0.15/\mathrm{ps}$. Clearly visible is a shift of the central frequency $\omega_A\rightarrow \omega_A'>\omega_A$, since the WP relaxes to lower vibrational levels. A frequency shift is also observed experimentally [@Claas_1151_2006]. In the full damping model, however, the signal decays to zero, since at this wavelength vibrational relaxation leads to a complete “closing” of the initial FC window. This result is in contrast to the experimental observation, where a pronounced oscillation is also present at later delay times, see [@Claas_1151_2006] and fig. \[fig:PP\_833\_gphase\_exp\](b), and requires further studies below.
At $\lambda=800\mathrm{\, nm}$ (excitation scheme II), full damping leads to the result shown in fig. \[fig:PP\_FT\_SD\_800nm\](a). Through dissipation, the WP in the state $\mathrm{\,A}\, ^1\Sigma^+_u$ leaves the initial FC region. However, the decelerated WP approaches another FC window around the equilibrium distance, cf. fig. \[fig:WPs\_833nm\], after several circulations. The signal therefore shows a massive increase of the (shifted) frequency component $\omega_A'$.
For both excitation schemes, the inclusion of damping improves agreement with experiment for the first $\simeq 10\mathrm{\, ps}$. However, at later times, there are still significant discrepancies. These can be removed by taking into account desorption of dimers off the droplet.
The clear visibility of both frequency components $\omega_A'$ and $\omega_X$ is again attributed to electronic decoherence: Since the damping scheme is carried out (stochastically) independent on all involved electronic states, damping leads to electronic decoherence similar to the fluctuating shifts we assumed previously. The final signal is an incoherent mixture of electronic contributions, as previously considered through eq. (\[eq:17\]). However, a massive increase of the component $\omega_A'$ is not observed experimentally (compare fig. \[fig:FT\_800nm\_gphase\_exp\](b) and fig. \[fig:PP\_FT\_SD\_800nm\](a)).
Desorption of the dimer off the droplet prevents the full vibrational damping in the electronic ground or excited state. Indeed, as we will show, desorption implies the disappearance of component $\omega_A$ and simultaneous ongoing presence of $\omega_X$. For excitation scheme I, desorption explains the observed signal oscillation at later delay times.
\[sec:state-indep-desorpt\] Desorption
---------------------------------------
The release of a dimer into the gas phase takes place at some random time $t'$. In our model, a state vector evolves according to eq. (\[eq:15\]) up to time $t'$. As the He influence vanishes at $t'$, shift/damping terms in eq. (\[eq:15\]) are set to zero. We have to consider the ion yield at delay time $\tau$, which is changed because of the dissapearance of the He influence at time $t'$. If the desorption occurs before the decay of the probe pulse ($t' \lesssim \tau$), the resulting signal is denoted as $S(\tau,t')$. If the desorption occurs after decay of the probe pulse ($t' \gtrsim \tau$), the ion yield is unaffected by the desorption process. Anything that happens after the probe pulse, will not be mapped to the ion yield. The resulting signal is denoted with $S(\tau,\tau)$.
Note that in an ensemble of attached dimers, the desorption time $t'$ will be distributed according to some probability distribution $P(t')$. We therefore have to calculate the pump-probe signal for various desorption times $t'$ and consider the averaged signal.
We first assume that the electronic state occupation is not relevant in the desorption process. A dimer therefore has a certain probability to stay on or leave (still being in the superposition) the droplet, but this probability is independent of the specific electronic occupation. We assume a constant probability $$\label{eq:6}
p_\mathrm{off}(t', t' + \Delta t')=\Delta t' R_D$$ that the dimer leaves the droplet within a small time interval $(t', t' + \Delta t)$. Here, $R_D$ marks the constant desorption rate. The probability to find the dimer attached at arbitrary time $t'$ is hence given through $$P_\mathrm{on}(t') =e^{-t'R_D}.$$ The signal contains weighted contributions and is obtained from $$\begin{aligned}
\label{eq:2}
\langle S(\tau) \rangle& = R_D \int_0^\infty P_\mathrm{on}(t') S(\tau, t')dt' \\ \nonumber
& = P_\mathrm{on}( \tau )S(\tau,\tau) + R_D \int_0^\tau P_\mathrm{on}(t') S(\tau, t')dt' .\end{aligned}$$ In the second line we have used that for $t'\gtrsim \tau$ (desorption after decay of the probe pulse) the ion yield does not change. In this case, the dimer is fully damped until the probe pulse has passed.
In a previous study, we made use of the state-independent desorption scheme eq. (\[eq:2\]), see Ref. [@Schle_245_2010]. Although it is possible to find reasonable agreement with experimental findings, it is an oversimplification to not take into account the electronic state for desorption. It is likely that only electronically excited dimers leave the droplet. The electronic excitation implies a larger degree of distortion of the helium enviroment [@Stien_10119_2001]. On the other hand, it is reasonable to assume that (slowly moving) ground state systems stay attached. Therefore, we consider an alternative desorption scheme and only allow excited dimers to leave the droplet. Desorption is again described in terms of a constant in time desorption rate $R_D$. The probability for the molecule to desorb, however, is now proportional to the excitation probability $p_\mathrm{e}=\sum_{i \neq 0} p_i$, such that eq. (\[eq:6\]) is replaced by $$\label{eq:11}
p_\mathrm{off}(t', t' + \Delta t')=p_\mathrm{e}R_D\Delta t'.$$ Note that in this scheme, the state after desorption does not contain any electronic ground state component (see later). Those dimers that do not desorb evolve according to the dissipative dynamics eq. (\[eq:15\]). In order to compensate for apparent loss of ground state dimers, we need a third possible channel: With probability $p_\mathrm{g} R_D \Delta t'=(1-p_\mathrm{e}) R_D \Delta t'$ the dimer remains on the droplet and is projected onto its electronic ground state. For the full averaged signal, we obtain $$\begin{aligned}
\label{eq:8}
\langle S_\mathrm{SD}(\tau) \rangle= & P_\mathrm{on}( \tau ) S(\tau,\tau) + p_\mathrm{e} R_D \int_0^{\tau} \!\! P_\mathrm{on}(t') S_\mathrm{e}(\tau, t' ) dt' +\\ \nonumber
& p_\mathrm{g} R_D \int_0^{\tau} \!\! P_\mathrm{on}(t') S_{\mathrm{g}} (\tau, t' ) dt' .
\end{aligned}$$ $S_\mathrm{e}(\tau, t' )$ is the signal obtained upon removal of the helium influence at time $t'\lesssim \tau$ and subsequent projection on the excited superposition of electronic state. Likewise, $S_{\mathrm{g}}(\tau, t' )$ is obtained upon projection on the ground state. Note that a renormalization of the full wavefunction $|\Psi \rangle $ is required after projection, such that $\sum p_i=1$ is always valid.
In fig. \[fig:FT\_833nm\_SD\](b), we show the spectrogram for $\langle S_{SD}(\tau) \rangle$ at $\lambda=833\mathrm{\, nm}$ (scheme I). We use a desorption rate $R_D=0.1/\mathrm{\,ps}$, while the damping/shift parameters are not changed (as before , $\gamma=0.15/\mathrm{ps}$ and $\Delta_{2}=-50\mathrm {cm}^{-1}$). In the spectrogram, a (small) frequency shift is $\omega_A \rightarrow \omega_A'$ is observable. The shift is mainly due to contributions of vibrationally damped dimers by means of the first term in eq. (\[eq:8\]). We obtain a frequency shift $\omega_A \rightarrow \omega_A'$ in the model, but the shift is more pronounced in the experiment [@Claas_1151_2006]. The second term in eq. (\[eq:2\]) marks contributions from dimers which are damped up to time $t'\lesssim \tau$, but are not damped afterwards, i.e.upon release from the droplet. Early desorbing dimers are not vibrationally relaxed and the WP in the $\mathrm{\,A}\, ^1\Sigma^+_u$ state continous to reach the initial FC region. From this point of view, undamped dimers may well contribute to the ion yield. Indeed, contributions to the ion yield at later delay times $t \gtrsim 20\mathrm{\, ps}$ are exclusively attributed to these undamped dimers. There, the oscillation frequency is near the initial gas phase value $\omega_A=\omega_{\bar{v},\bar{v}+1}$.
As a final note, we find that the observed frequency shift $\omega_A \rightarrow \omega_A'$ becomes negligible if we do not include the electronic shift $\Delta_2$ in the model.
As the laser frequency is further increased, the vibrational energy of the WP in the $\mathrm{\,A}\, ^1\Sigma^+_u$ state increases (scheme II). Due to the larger elongation and faster dynamics, one may think of larger damping of vibrational motion and/or faster desorption of dimers off the droplet. In fig. \[fig:PP\_FT\_SD\_800nm\](b) we show the spectrogram at $\lambda=800\mathrm{\, nm}$, as obtained for $\gamma_i=0.15/\mathrm{ps}$ and $R_D=0.5/\mathrm{ps}$, i. e. we assume a faster desorption as before for scheme I. In fact, this value for the desorption rate means that dimers quickly leave the droplet after the pulse excitation. Until desorption, electronically excited dimers are fully damped. During that time, they approach an intermediate transition region with smaller overlap with higher electronic states, see fig. \[fig:WPs\_833nm\](c). As a consequence, the component $\omega_A$ fades away after several picoseconds. Also, vibrational WPs of desorbed dimers do not reach the FC window at the equilibrium distance, which is also marked in fig. \[fig:WPs\_833nm\](c). Therefore, an increase of the component $\omega_A$ is excluded, compare fig. \[fig:PP\_FT\_SD\_800nm\](a) and (b). To conclude, damping in connection with fast desorption of dimers explains the experimental result, see fig. \[fig:FT\_800nm\_gphase\_exp\](b), where the component $\omega_A$ is only visible in the beginning of the measurement and then disappears. Note that dimers which remain on the droplet in the ground state, are fully damped. Therefore, the ground state component $\omega_X$ decreases in this model.
Undamped ground state WPs {#sec:undamp-ground-state}
-------------------------
Agreement with experiment can be further improved, if one allows for undamped motion of the vibrational WP in the ground state. This is of particular relevance for excitation scheme II, for which the result upon leaving the model parameters for damping, desorption and shift unchanged, but setting $\gamma_0=0$, is shown in fig. \[fig:PP\_FT\_SD\_800nm\](c). Frequency components at later delay times are exclusively attributed to the ground state motion, i. e. the component $\omega_X$ is clearly visible. Undamped vibrational wave packet motion has been discussed in terms of a critical Landau velocity $v_\mathrm{crit}$ in Ref.[@Schle_245_2010]. The existence of $v_\mathrm{crit}$ in the superfluid nanodroplet may allow for frictionless motion of slowly moving ground state WPs. In fig. \[fig:PP\_FT\_SD\_800nm\](c) the ground state motion is assumed frictionless; good agreement with the experimental result, see fig. \[fig:FT\_800nm\_gphase\_exp\](b), is also obtained for nearly frictionless motion $\gamma_0 \approx 0$.
Note that for the observation of the ground state WP, the electronic shift $\Delta_2$ is less important. It only leads to a slightly different ratio between the ground-and excited state frequency component.
\[sec:conclusion\]Conclusions
=============================
We consider vibrational wave packet dynamics of dimers attached to He nanodroplets. It is found that (calculated) gas phase spectra and spectra from dimers attached to He droplets are markedly different. The interaction between droplet and dimer influences the vibrational dimer dynamics in three ways: Shifts, damping, and desorption. All three ingredients are taken into account in a phenomenological manner and each contributes to characteristic changes of the resulting spectra.
First, we study electronic decoherence, occurring for instance due to slightly fluctuating shifts of electronic surfaces. We find that, indeed, resulting spectra do not show electronic interferences such that contributions from several electronic states, in particular the ground state, are clearly resolved. In this way, the modeled spectrum is already similar to the experimental finding.
However, pure electronic decoherence cannot explain a decreasing contribution to the signal from wave packets in the first excited $\mathrm{\,A}\, ^1\Sigma^+_u$ state. Consequently, vibrational damping of wave packets is taken into account, which improves agreement with experiment for short delay times. We find that damping is not present over the full observation timescale, probably due to desorption of dimers from the droplet. We use a desorption scheme, which takes into account the occupation of electronic levels. We find that the desorption rate depends on the mean vibrational energy of the wave packet. At $\lambda=800\mathrm{\, nm}$, desorption is fast, taking place on average several picosecond after excitation ($R_D=0.5/\mathrm{ps}$).
Note that in this study, all involved electronic states are spin singlet states. For the WP in these singlet states, we find a relaxation rate which is significantly higher than in previously considered spin triplet Rb$_2$ dimers attached to He droplets. The smaller rate in the latter may be ascribed to the orientation of the dialkali axis relative to the droplet surface. Recent calculations show that the axis of spin triplet dimers is oriented parallel to the droplet surface while singlet states are assumed to be oriented perpendicularly [@Guill_6918_2011; @Bovin_224903_2009]. Upon the perpendicular orientation in our case, the dimer might interact with the droplet more efficiently, such that relaxation is faster.
Damping appears to be absent for (slowly moving) wave packets in the electronic ground state. There, vibrational motion is found to be nearly frictionless. A potential energy surface for the dimer-droplet system, which is currently underway for Rb$_2$ [@Guill_6918_2011], should give additional insight on dissipation rates of the wave packet.
The authors would like to thank Marcel Mudrich and Frank Stienkemeier for providing experimental data, for fruitful discussions and helpful comments. Further, we thank Alexander Eisfeld for valuable remarks. Support by the Deutsche Forschungsgemeinschaft through the research grant “Control and Coherence of the Few-Particle Continuum” is gratefully acknowledged. Computing resources have been provided by the Zentrum für Informationsdienste und Hochleistungsrechnen (ZIH) at the TU Dresden. M. S. is a member of the IMPRS Dresden.
[10]{} url \#1[[\#1]{}]{}urlprefix\[2\]\[\][[\#2](#2)]{} Claas P, Droppelmann G, Schulz C P, Mudrich M and Stienkemeier F 2006 [*J. Phys. B*]{} [**39**]{} S1151
Schlesinger M, Mudrich M, Stienkemeier F and Strunz W T 2010 [*Chem. Phys. Lett.*]{} [**490**]{} 245–248
Toennies J P and Vilesov A F 2004 [*Angew. Chem., Int. Ed.*]{} [**43**]{} 2622–2648
Stienkemeier F and Lehmann K K 2006 [*J. Phys. B*]{} [**39**]{} R127
Callegari C, Reinhard I, Lehmann K K, Scoles G, Nauta K and Miller R E 2000 [*J. Chem. Phys.*]{} [**113**]{} 4636–4646
Lehmann K K 1999 [*Mol. Phys.*]{} [**97**]{} 645–666
Callegari C, Lehmann K K, Schmied R and Scoles G 2001 [*J. Chem. Phys.*]{} [**115**]{} 10090–10110
Reho J, Callegari C, Higgins J, Ernst W E, Lehmann K K and Scoles G 1997 [ *Faraday Discussions*]{} [**108**]{} 161–174
Grebenev S, Toennies J P and Vilesov A F 1998 [*Science*]{} [**279**]{} 2083–2085
Callegari C, Conjusteau A, Reinhard I, Lehmann K K and Scoles G 1999 [*Phys. Rev. Lett.*]{} [**83**]{} 5058–5061
Nauta K and Miller R E 2001 [*J. Chem. Phys.*]{} [**115**]{} 45084514
Dick B and Slenczka A 2001 [*J. Chem. Phys.*]{} [**115**]{} 10206–10213
Hartmann M, Mielke F, Toennies J P and Vilesov A F 1996 [*Phys. Rev. Lett.*]{} [**76**]{} 4560–4563
Mudrich M, Heister P, Hippler T, Giese C, Dulieu O and Stienkemeier F 2009 [ *Phys. Rev. A*]{} [**80**]{} 042512
Gr[ü]{}ner B, Schlesinger M, Heister P, Strunz W T, Stienkemeier F and Mudrich M 2011 [*Phys. Chem. Chem. Phys.*]{} [**13**]{} 6816–6826 ISSN 1463-9076 <http://dx.doi.org/10.1039/C0CP02355H>
Koch M, Aub[ö]{}ck G, Callegari C and Ernst W E 2008 [*Phys. Rev. Lett.*]{} [**103**]{} 035302
Braun A and Drabbels M 2004 [*Phys. Rev. Lett.*]{} [**93**]{} 253401
Przystawik A, G[ö]{}de S, D[ö]{}ppner T, Tiggesb[ä]{}umker J and Meiwes-Broer K H 2008 [*Phys. Rev. A*]{} [**78**]{} 021202
Kornilov O, Wang C C, B[ü]{}nermann O, Healy A T, Leonard M, Peng C, Leone S R, Neumark D M and Gessner O 2010 [*J. Chem. Phys. A*]{} [**114**]{} 1437
Droppelmann G, Bünermann O, Schulz C P and Stienkemeier F 2004 [*Phys. Rev. Lett.*]{} [**93**]{} 0233402
Nitzan A, Mukamel S and Jortner J 1975 [*J. Chem. Phys.*]{} [**63**]{} 200
Ewing G E 1987 [*J. Phys. Chem. A*]{} [**91**]{} 4662–4671
Nauta K and Miller R E 2000 [*J. Chem. Phys.*]{} [**113**]{} 9466–9469
Dalfovo F 1994 [*Z. Phys. D*]{} [**29**]{} 61–66
Ancilotto F, DeToffol G and Toigo F 1995 [*Phys. Rev. B*]{} [**52**]{} 16125–16129
Stienkemeier F, Higgins J, Ernst W E and Scoles G 1995 [*Phys. Rev. Lett.*]{} [**74**]{} 3592–3595
Higgins J, Callegari C, Reho J, Stienkemeier F, Ernst W E, Gutowski M and Scoles G 1998 [*J. Phys. Chem. A*]{} [**102**]{} 4952–4965
Bovino S, Bodo E, Yurtsever E and Gianturco F A 2008 [*J. Chem. Phys.*]{} [**128**]{} 224312
Bovino S, Coccia E, Bodo E, Lopez-Duran D and Gianturco F A 2009 [*J. Chem. Phys.*]{} [**130**]{} 224903
Zewail A H 1993 [*J. Phys. Chem.*]{} [**97**]{} 12427
Gruebele M, Roberts G, Dantus M, Bowman R M and Zewail A H 1990 [*Chem. Phys. Lett.*]{} [**166**]{} 459–469
Baumert T, Engel V, R[ö]{}ttgermann C, Strunz W T and Gerber G 1992 [*Chem. Phys. Lett.*]{} [**191**]{} 639–644
Bowman R, Dantus M and Zewail A 1989 [*Chem. Phys. Lett.*]{} [**161**]{} 297
Gruebele M and Zewail A H 1993 [*J. Chem. Phys.*]{} [**98**]{} 883
Baumert T, Bühler B, Grosser M, Weiss V and Gerber G 1991 [*J. Phys. Chem.*]{} [**8103**]{} 8103–8110
de Vivie-Riedle R, Kobe K, Manz J, Meyer W, Reischl B, Rutz S, Schreiber E and Wöste L 1996 [*J. Phys. Chem.*]{} [**100**]{} 7789–7796
Rutz S and Schreiber E 1997 [*Chem. Phys. Lett.*]{} [**269**]{} 9
Nicole C, Bouchène M A, Meier C, Magnier S, Schreiber E and Girard B 1999 [*J. Chem. Phys.*]{} [**111**]{} 7857
Karavitis M, Segale D, Bihary Z, Pettersson M and Apkarian V A 2003 [*Low Temp. Phys.*]{} [**29**]{} 814
G[ü]{}hr M, Ibrahim H and Schwentner N 2004 [*Phys. Chem. Chem. Phys.*]{} [**6**]{} 5353–5361
Meier C 2004 [*Phys. Rev. Lett.*]{} [**93(17)**]{} 173003(4)
Liu Q, Wan C and Zewail A H 1996 [*J. Phys. Chem. A*]{} [**100**]{} 18666
Schlesinger M and Strunz W T 2008 [*Phys. Rev. A*]{} [**77**]{} 012111
Stienkemeier F and Vilesov A F 2001 [*J. Chem. Phys.*]{} [**115**]{} 10119–10137
Reho J, Callegari C, Higgins J, Ernst W E, Lehmann K K and Scoles G 1997 [ *Faraday Discuss.*]{} [**108**]{} 161–174
Schulz C P, Claas P and Stienkemeier F 2001 [*Phys. Rev. Lett.*]{} [**87**]{} 153401
Aub[ö]{}ck G, Nagl J, Callegari C and Ernst W E 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 035301
de Vivie-Riedle R, Reischl B, Rutz S and Schreiber E 1995 [*J. Phys. Chem.*]{} [**99**]{} 16829–16834
Feit M, Fleck F and Steiger A 1982 [*J. Comp. Phys.*]{} [**47**]{} 412–433
Fischer I, Vrakking M J J, Villeneuve D M and Stolow A 1996 [*Chem. Phys.*]{} [**207**]{} 331–354
Vrakking M J J, Villeneuve D M and Stolow A 1996 [*Phys. Rev. A*]{} [**54**]{} R37
Stienkemeier F, Higgins J, Callegari C, Kanorsky S I, Ernst W E and Scoles G 1996 [*Zeitschrift Fur Physik D-Atoms Molecules And Clusters*]{} [**38**]{} 253–263
Allard O, Nagl J, Aubock G, Callegari C and Ernst W E 2006 [*J. Phys. B*]{} [**39**]{} S1169–S1181
Br[ü]{}hl F R, Miron R A and Ernst W E 2001 [*J. Chem. Phys.*]{} [**115**]{} 10275–10281
Aub[ö]{}ck G, Aymar M, Dulieu O and Ernst W E 2010 [*J. Chem. Phys.*]{} [ **132**]{} 054304
B[ü]{}nermann O, Droppelmann G, Hernando A, Mayol R and Stienkemeier F 2007 [*J. Phys. Chem. A*]{} [**111**]{} 12684
Lewerenz M, Schilling B and Toennies J P 1993 [*Chem. Phys. Lett.*]{} [ **206**]{} 381–387
Loginov E 2008 [*Photoexcitation and photoionization dynamics of doped liquid helium-4 nanodroplets*]{} Ph.D. thesis Lausanne : EPFL
Mudrich M private communication
Schlosshauer M 2007 [*Decoherence And The Qantum-To-Classical Transition*]{} (Berlin: Springer)
Helm J and Strunz W T 2009 [*Phys. Rev. A*]{} [**80**]{} 042108
Lindblad G 1975 [*Commun. Math. Phys.*]{} [**40**]{} 147–151
Scully M O and Zubairy M S 1997 [*Quantum Optics*]{} (Cambridge: Cambridge University Press)
Breuer H and Petruccione F 2002 [*The theory of open quantum systems*]{} (Oxford: Oxford University Press)
Gisin N and Percival I C 1992 [*J. Phys. A: Math. Gen.*]{} [**25**]{} 5677–5691
Guillon G, Zanchet A, Leino M, Viel A and Zillich R E 2011 [*J. Phys. Chem. A*]{} [**115**]{} 6918–6926
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We prove that the inhomogeneous estimate of vector fields $u\in{\operatorname{C}}^\infty({\operatorname{B}},{\mathbb{R}}^N)$ $$\begin{aligned}
\left(\int_{{\operatorname{B}}}|{\operatorname{D}\!}^{k-1}u|^{n/(n-1)}{\operatorname{d}\!}x\right)^{(n-1)/n}{\leqslant}c\left(\int_{{\operatorname{B}}}|{\mathbb{A}}u|+|u|{\operatorname{d}\!}x\right)\end{aligned}$$ holds if and only if the linear, constant coefficient differential operator ${\mathbb{A}}$ of order $k$ has finite dimensional null-space (FDN). This generalizes the Gagliardo–Nirenberg–Sobolev inequality on domains and provides the local version of the analogous homogeneous embedding in full–space $$\begin{aligned}
\left(\int_{{\mathbb{R}}^n}|{\operatorname{D}\!}^{k-1}u|^{n/(n-1)}{\operatorname{d}\!}x\right)^{(n-1)/n}{\leqslant}c\int_{{\mathbb{R}}^n}|{\mathbb{A}}u|{\operatorname{d}\!}x\qquad\text{ for all }u\in{\operatorname{C}}^\infty_c({\mathbb{R}}^n,{\mathbb{R}}^N),\end{aligned}$$ proved by <span style="font-variant:small-caps;">Van Schaftingen</span> in [@VS Thm 1.3] precisely for elliptic and cancelling (EC) operators, building on fundamental ${\operatorname{L}}^1$–estimates from the works of <span style="font-variant:small-caps;">Bourgain</span> and <span style="font-variant:small-caps;">Brezis</span>, e.g., in [@BB07 Thm. 25]. We prove that FDN *strictly* implies EC, effectively showing how the contrast between homogeneous and inhomogeneous estimates can be seen at the level of algebraic properties of the operators ${\mathbb{A}}$.
author:
- Franz Gmeineder
- Bogdan Raita
title: 'Embeddings for ${\mathbb{A}}$–weakly differentiable functions on domains'
---
[^1]
Introduction {#sec:intro}
============
Let ${\mathbb{A}}$ be a linear, homogeneous differential operator with constant coefficients on ${\mathbb{R}}^n$ from $V$ to $W$, i.e., $$\begin{aligned}
\label{eq:A}
{\mathbb{A}}u=\sum_{|\alpha|=k} A_\alpha\partial^\alpha u,\qquad u\colon{\mathbb{R}}^{n}\to V,\end{aligned}$$ where ${\mathbb{A}}_\alpha\in{\mathscr{L}}(V,W)$ are fixed linear mappings between two finite dimensional real vector spaces $V$ and $W$. In this respect, we recall the (Fourier) symbol map $$\begin{aligned}
{\mathbb{A}}[\cdot]\colon {\mathbb{R}}^{n}\to \mathscr{L}(V,W),\;\;\;{\mathbb{A}}[\xi] v=\sum_{|\alpha|=k} \xi^\alpha A_\alpha v,\end{aligned}$$ defined for $\xi\in{\mathbb{R}}^n$, $v\in V$. It is a well–known fact that a Korn–type inequality in full–space, by which we mean that for all $u\in{\operatorname{C}}_c^\infty({\mathbb{R}}^n,V)$ we have $$\begin{aligned}
\|{\operatorname{D}\!}^k u\|_{{\operatorname{L}}^p({\mathbb{R}}^{n},V\odot^{k}{\mathbb{R}}^{n})}{\leqslant}c\|{\mathbb{A}}u\|_{{\operatorname{L}}^p({\mathbb{R}}^{n},W)},\end{aligned}$$ for $1<p<\infty$, is equivalent to *ellipticity* of ${\mathbb{A}}$; this is to say that the symbol map ${\mathbb{A}}[\xi]\in{\mathscr{L}}(V,W)$ is injective for any $\xi\in{\mathbb{R}}^n\setminus\{0\}$. It is also well–known by the so–called <span style="font-variant:small-caps;">Ornstein</span>’s *Non–inequality* that no such estimate can hold in the critical case $p=1$ (see [@Ornstein; @CFM] for the classical statement and [@KirKri Thm. 1.3] for a general form). However, it is natural to ask whether a weaker coercive estimate holds in this borderline case. This has been established by <span style="font-variant:small-caps;">Van Schaftingen</span> in the seminal paper [@VS]. Namely, it is proved in [@VS Thm. 1.3] that a Sobolev–type inequality of the form $$\begin{aligned}
\label{eq:VS}
\|{\operatorname{D}\!}^{k-1}u\|_{{\operatorname{L}}^{\frac{n}{n-1}}({\mathbb{R}}^{n},V\odot^{k-1}{\mathbb{R}}^{n})}{\leqslant}c\|{\mathbb{A}}u\|_{{\operatorname{L}}^1({\mathbb{R}}^{n},W)}\end{aligned}$$ for $u\in{\operatorname{C}}^\infty_c({\mathbb{R}}^n,V)$ holds if and only if ${\mathbb{A}}$ is elliptic and *cancelling* (EC). The latter condition states that the intersection of ${\mathbb{A}}[\xi](V)$ for all non–zero $\xi$ is trivial. This sharp result generalizes the proof of for a large class of operators ${\mathbb{A}}$ given by <span style="font-variant:small-caps;">Bourgain</span> and <span style="font-variant:small-caps;">Brezis</span> in [@BB07 Thm. 25], building on their fundamental work on critical case estimates in [@BB02; @BB04; @BBCR; @BB07]. We remark that in the case $p=1$, the critical estimate cannot be achieved by standard potential estimates. With such means, only weak–type estimates can be obtained (see [@Stein Ch. V.1]), and, in turn, one needs to employ the vectorial structure of ${\mathbb{A}}$.
The aim of this paper is to give precise conditions on ${\mathbb{A}}$ under which a Sobolev–type embedding holds on bounded domains, for which we consider the unit ball ${\operatorname{B}}\subset{\mathbb{R}}^n$, thereby generalizing the well–known Gagliardo–Nirenberg–Sobolev inequality $$\begin{aligned}
\|u\|_{{\operatorname{L}}^{\frac{n}{n-1}}({\operatorname{B}},V)}{\leqslant}c\left(\|{\operatorname{D}\!}u\|_{{\operatorname{L}}^1({\operatorname{B}},V\times{\mathbb{R}}^n)}+\|u\|_{{\operatorname{L}}^{1}({\operatorname{B}},V)}\right)\end{aligned}$$ for $u\in{\operatorname{C}}^\infty(\bar{{\operatorname{B}}},V)$ to arbitrary differential operators ${\mathbb{A}}$. Our result also covers the Korn–Sobolev inequality $$\begin{aligned}
\label{eq:ST}
\|u\|_{{\operatorname{L}}^{\frac{n}{n-1}}({\operatorname{B}},{\mathbb{R}}^n)}{\leqslant}c\left(\|\mathcal{E} u\|_{{\operatorname{L}}^1({\operatorname{B}},{\mathbb{R}}^{n\times n}_{{\operatorname{sym}}})}+\|u\|_{{\operatorname{L}}^{1}({\operatorname{B}},{\mathbb{R}}^n)}\right)\end{aligned}$$ proved by <span style="font-variant:small-caps;">Strang</span> and <span style="font-variant:small-caps;">Temam</span> in [@ST Prop. 1.2], building on the homogeneous inequality of <span style="font-variant:small-caps;">Strauss</span> [@Strauss]. Our local version of <span style="font-variant:small-caps;">Van Schaftingen</span>’s Theorem is particularly relevant to variational problems of minimizing energy functionals of the type, e.g., $$\begin{aligned}
\label{eq:varprob}
\mathfrak{F}\colon u\mapsto\int_{{\operatorname{B}}}f({\mathbb{A}}u){\operatorname{d}\!}x,\end{aligned}$$ over Dirichlet classes of mappings $u\colon{\operatorname{B}}\rightarrow V$, for $f\colon W\rightarrow{\mathbb{R}}$ of linear growth, i.e., there exist constants $c,C>0$ such that $c|w|{\leqslant}f(w){\leqslant}C(1+|w|)$ for $w\in W$. Such problems arise for example in plasticity theory ([@FS Ch. 1-2]), which is the original motivation for the study in [@ST]. In this respect, we mention the connection with the recent paper [@BDG], where existence of generalized minimizers of was established for first order operators ${\mathbb{A}}$. To be precise, we introduce the spaces ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ (resp. ${\operatorname{BV}}^{\mathbb{A}}({\operatorname{B}})$) as the space of $u\in{\operatorname{L}}^1({\operatorname{B}},V)$ such that the distribution ${\mathbb{A}}u$ is an integrable function (resp. a Radon measure) on ${\operatorname{B}}$, which are complete with respect to the obvious norms. Assuming $k=1$, it is proved in [@BDG Thm. 1.1] that the trace embedding ${\operatorname{BV}}^{\mathbb{A}}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^1(\partial{\operatorname{B}},V)$ holds if and only if ${\mathbb{A}}$ has *finite dimensional null–space* (FDN). Under this assumption, it is shown that the infimum of $\mathfrak{F}$ over a Dirichlet class in ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ is attained in ${\operatorname{BV}}^{\mathbb{A}}({\operatorname{B}})$ by a minimizer of the lower–semicontinuous envelope $\bar{\mathfrak{F}}$ of $\mathfrak{F}$ (see [@BDG Thm. 5.3], cp. [@ADM; @GMS]).
The main result of this paper complements the study of by showing that FDN is also equivalent with the critical Sobolev–type embedding ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{\frac{n}{n-1}}({\operatorname{B}},V)$:
\[thm:main\] Let ${\mathbb{A}}$ be as in , $k=1$, $n>1$. The following are equivalent:
1. \[it:main\_a\] ${\mathbb{A}}$ has FDN.
2. \[it:main\_b\] ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{\frac{n}{n-1}}({\operatorname{B}},V)$.
3. \[it:main\_c\] ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{p}({\operatorname{B}},V)$ for some $1<p{\leqslant}\frac{n}{n-1}$.
4. \[it:main\_d\] ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow\hookrightarrow{\operatorname{L}}^{q}({\operatorname{B}},V)$ for all $1{\leqslant}q<\frac{n}{n-1}$.
5. \[it:main\_e\] ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow\hookrightarrow{\operatorname{L}}^{1}({\operatorname{B}},V)$.
The same holds true with ${\operatorname{W}}^{{\mathbb{A}},1}$ being replaced by ${\operatorname{BV}}^{\mathbb{A}}$ and for bounded domains that are star–shaped with respect to a ball. The compact embedding \[it:main\_d\] generalizes the well–known result for ${\operatorname{BD}}$ (i.e., for ${\mathbb{A}}=\mathcal{E}$; see [@FS; @ST; @Suquet]). In Theorem \[thm:main\_k\] below, we remove the restriction $k=1$, which we temporarily keep for simplicity of exposition. The novelty of Theorems \[thm:main\] and \[thm:main\_k\] comes from the fact that, up to our knowledge, except for a few examples of operators ${\mathbb{A}}$, in the literature there are no similar ${\operatorname{L}}^1$–estimates on bounded domains (without additional assumptions of zero or periodic boundary values). We do not include the case $n=1$, which is not covered by our methods, but turns out to be a simple exercise.
We pause to compare the embedding \[it:main\_b\] in Theorem \[thm:main\] with <span style="font-variant:small-caps;">Van Schaftingen</span>’s homogeneous embedding $\dot{{\operatorname{W}}}^{{\mathbb{A}},1}({\mathbb{R}}^n)\hookrightarrow{\operatorname{L}}^{\frac{n}{n-1}}({\mathbb{R}}^n,V)$. For elliptic ${\mathbb{A}}$, this embedding is equivalent to $$\begin{aligned}
\label{eq:zerotraceemb}
{\operatorname{W}}^{{\mathbb{A}},1}_0({\operatorname{B}})\hookrightarrow{\operatorname{L}}^\frac{n}{n-1}({\operatorname{B}},V)\end{aligned}$$ (see Lemma \[lem:embimpliesEC\] for a scaling argument), and it can easily be shown that, in the absence of cancellation, we can still prove by means of a Green’s formula and boundedness of Riesz potentials that ${\operatorname{W}}^{{\mathbb{A}},1}_0({\operatorname{B}})\hookrightarrow{\operatorname{L}}^p({\operatorname{B}},V)$ for any $1{\leqslant}p < n/(n-1)$ (see Lemma \[lem:zerotraceemb\]). Here ${\operatorname{W}}^{{\mathbb{A}},p}_0({\operatorname{B}})$ is defined as the closure of ${\operatorname{C}}_c^\infty({\operatorname{B}},V)$ in the (semi–)norm $u\mapsto\|{\mathbb{A}}u\|_{{\operatorname{L}}^p}$. The situation is dramatically different as far as ${\operatorname{L}}^p$–embeddings of ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ are concerned. By Theorem \[thm:main\], if the critical embedding $$\begin{aligned}
\label{eq:ourembedding}
{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^\frac{n}{n-1}({\operatorname{B}},V)\end{aligned}$$ fails, then no uniform higher integrability estimate is possible. The difference can be even sharper: in Section \[sec:ECvsFDN\], we give an example of first order differential operator ${\mathbb{A}}$ of the form that is elliptic but does not have FDN such that there is a map in ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ that has no higher integrability. Moreover, this operator can be chosen such that it satisfies the cancelling condition and even the more particular condition of [@BB07 Thm. 25], so the homogeneous embedding can hold even if the inhomogeneous fails. We remark that the main difference between ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ and ${\operatorname{W}}^{{\mathbb{A}},1}_0({\operatorname{B}})$ lies in the traces, which are integrable if and only if ${\mathbb{A}}$ has FDN [@BDG]. Another way to look at this discrepancy is to note that, for elliptic ${\mathbb{A}}$, the only solution of ${\mathbb{A}}u=0$ in ${\operatorname{W}}^{{\mathbb{A}},1}_0({\operatorname{B}})$ is the trivial one, which can be seen, e.g., from . In the case of , if ${\mathbb{A}}$ is elliptic but does not have FDN, the space $\{u\in{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\colon {\mathbb{A}}u=0\}$ contains maps that are not integrable at the boundary (see [@BDG Sec. 4.3]) and maps that are not ${\operatorname{L}}^{n/(n-1)}$–integrable (see Section \[sec:ECvsFDN\]). Both examples use the lack of boundary regularity of the solutions of ${\mathbb{A}}u=0$ in the absence of FDN, a phenomenon which is not relevant for , where zero boundary data is implicit. If, in turn, ${\mathbb{A}}$ has FDN, all solutions of ${\mathbb{A}}u=0$ are polynomials. From this point of view, we can heuristically say that FDN is a canonical condition for Dirichlet problems/inhomogeneous estimates on bounded domains, whereas EC is a canonical condition for problems/homogeneous estimates in full–space.
In this respect, it is of particular interest to compare the conditions EC and FDN. We will prove in Lemma \[lem:FDNimpliesEC\] that FDN implies EC. In Section \[sec:ECvsFDN\], we complete the comparison of these conditions, showing that the implication is strict in general. We write $N:=\dim V$ and summarize our findings in the table below:
[|c|c|]{} &
------------------ ----------------------
$N=1\qquad\quad$ $\qquad\quad N\geq2$
------------------ ----------------------
\
$n=2$&
[c|c]{}
$k=1$:
-----------
$k=2$:
$k\geq3$:
&
$k=1$:
-----------
$k\geq2$:
\
\
$n\geq3$&
[c|c]{}
$k=1$:
-----------
$k\geq2$:
&
-----------------------------
$\qquad\quad\hspace{0.4mm}$
-----------------------------
\
The streamline is that for first order operators acting on scalar fields, ellipticity is equivalent to both conditions, whereas for large values of $n,N,k$, the implication of EC by FDN is strict. Interestingly, two cases remain, which match the case of the canonical elliptic, non–FDN operators $\bar{\partial}$ and $\partial_1^2+\partial_2^2$ (see Section \[sec:examples\]). In these cases, we show that EC and FDN are equivalent. Our comparison is completely elementary and we hope that it will have some impact in understanding the nature of and the differences between ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ and ${\operatorname{W}}^{{\mathbb{A}},1}({\mathbb{R}}^n)$ or its homogeneous version.
We move on to briefly describing the proof of Theorem \[thm:main\]. As mentioned above, we show that the FDN assumption (strictly) implies the cancelling condition. This is coupled with the construction of a suitable extension operator to full–space which enables us to use [@VS Thm. 1.3]. To prove these, we rely on the known fact that a homogeneous operator ${\mathbb{A}}$ has FDN if and only it is *${\mathbb{C}}$–elliptic*, i.e., the map ${\mathbb{A}}[\xi]\in{\mathscr{L}}(V+{\operatorname{i}}V,W+{\operatorname{i}}W)$ is injective for any $\xi\in{\mathbb{C}}^n\setminus\{0\}$. This was noticed in [@BDG] for first order operators. We recently became aware of the work [@Smith], where the case of operators of arbitrary order is proved. We record these auxiliary facts below:
\[thm:tools\] Let ${\mathbb{A}}$ be as in , $n>1$. Then ${\mathbb{A}}$ has FDN if and only if ${\mathbb{A}}$ is ${\mathbb{C}}$–elliptic. Moreover, if ${\mathbb{A}}$ has FDN, then ${\mathbb{A}}$ is cancelling and there exists a bounded, linear extension operator $E_B\colon{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\rightarrow{\operatorname{W}}^{{\mathbb{A}},1}({\mathbb{R}}^n)$.
We remark that the same holds true for any $1{\leqslant}p{\leqslant}\infty$. To construct the extension operator we use the technique introduced by <span style="font-variant:small-caps;">Jones</span> in [@Jones]. Although introduced to deal with rough domains, our reason for resorting to this rather involved method is that we could not otherwise circumvent lack of boundedness of singular integrals on ${\operatorname{L}}^1$ (see Lemma \[lem:extp>1\] for a simple proof if $1<p<\infty$). Of the modifications required to adapt <span style="font-variant:small-caps;">Jones</span>’ technique, we single out as a novelty the Poincaré–type inequality for FDN operators (Proposition \[prop:poinc\]), namely that for $1{\leqslant}l<k$, $1{\leqslant}p{\leqslant}\infty$ we have $$\begin{aligned}
\label{eq:poinc}
\inf_{P\in\ker{\mathbb{A}}}\|{\operatorname{D}\!}^l (u-P)\|_{{\operatorname{L}}^p({\operatorname{B}},V\odot^l{\mathbb{R}}^n)}{\leqslant}c{\operatorname{diam}}({\operatorname{B}})^{k-l}\|{\mathbb{A}}u\|_{{\operatorname{L}}^p({\operatorname{B}},W)}\end{aligned}$$ for $u\in{\operatorname{C}}^\infty(\bar{{\operatorname{B}}},V)$. Interestingly, ${\mathbb{A}}$ having FDN is not necessary for the estimate to hold, as can be seen from [@Fuchs1]. We believe that ellipticity alone is sufficient for the estimate to hold and intend to pursue this in future work.
Using the tools from Theorem \[thm:tools\], we can refine our result on fractional scales, thereby obtaining the local versions of the embeddings in [@VS Thm. 8.1, Thm. 8.4]:
\[thm:main\_k\] Let ${\mathbb{A}}$ be as in , $s\in[k-1,k)$, $q\in(1,\infty)$. Then ${\mathbb{A}}$ has FDN if and only if there exists $c>0$ such that $$\begin{aligned}
\|u\|_{{\operatorname{B}}^{s,\frac{n}{n-k+s}}_q ({\operatorname{B}},V)}{\leqslant}c\left(\|{\mathbb{A}}u\|_{{\operatorname{L}}^1({\operatorname{B}},W)}+\|u\|_{{\operatorname{L}}^1({\operatorname{B}},V)}\right)\end{aligned}$$ for all $u\in{\operatorname{C}}^\infty(\bar{{\operatorname{B}}},V)$.
We obtain the embeddings ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{W}}^{s,n/(n-k+s)}({\operatorname{B}},V)$ if we choose $q=n/(n-k+s)$ (cp. [@VS Thm. 8.1]) and ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{W}}^{k-1,n/(n-1)}({\operatorname{B}},V)$ if we further choose $s=k-1$ (cp. [@VS Thm. 1.3]).
In view of Theorem \[thm:main\_k\], it is natural to ask what is the generalization of [@BDG Thm. 4.17-18] to operators of arbitrary order. If ${\mathbb{A}}$ has FDN, we can use Theorem \[thm:main\_k\] to give a sub–optimal trace embedding $$\begin{aligned}
\label{eq:weaktrace}
{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{B}}_q^{s-\frac{1}{p},p}(\partial{\operatorname{B}},V)\quad\text{ for }s\uparrow k\text{, so }p=\frac{n}{n-k+s}\downarrow1\text{, and }q\downarrow1,\end{aligned}$$ using standard trace theory for Besov spaces. The optimal embedding for ${\mathbb{A}}=\nabla^k$ was only recently proved by <span style="font-variant:small-caps;">Mironescu</span> and <span style="font-variant:small-caps;">Russ</span> in [@MR], building on the $k=2$ case proved by <span style="font-variant:small-caps;">Uspenskiĭ</span> in [@Uspenskii]. They proved that the trace operator is continuous onto ${\operatorname{B}}^{k-1,1}_1$, which is in general strictly smaller than the quick guess ${\operatorname{W}}^{k-1,1}$ (see [@BP Rk. A.1]). Coupling this with the trace theorem in [@BDG], it is natural to make the following:
An operator ${\mathbb{A}}$ as in has FDN if and only if there exists a continuous, linear, surjective trace operator $\mathrm{Tr}\colon{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\rightarrow{\operatorname{B}}^{k-1,1}_1(\partial{\operatorname{B}},V)$.
A few remarks are in order. Necessity of FDN can be proved by a modification of the arguments in [@BDG Sec. 4.3]. Surjectivity is obvious, using [@MR Thm. 1.3-4] and ${\operatorname{W}}^{k,1}({\operatorname{B}},V)\hookrightarrow{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$. The difficulty stems from proving boundedness (hence, well-definedness) of the trace operator, which cannot be reduced to the situation in [@MR] by <span style="font-variant:small-caps;">Ornstein</span>’s Non–inequality, or to by strict inclusion of Besov spaces. We do not see a way to merge the techniques in [@BDG; @MR] and intend to tackle the problem in the future.
This paper is organized as follows: In Section \[sec:prel\] we collect preliminaries on function spaces, multi–linear algebra, harmonic analysis and give examples of operators. In Section \[sec:ECvsFDN\] we give the proof of the first two statements in Theorem \[thm:tools\] and complete the comparison between EC and FDN. In Section \[sec:proof\] we construct the Jones–type extension and prove Theorems \[thm:main\] and \[thm:main\_k\].
Acknowledgement {#acknowledgement .unnumbered}
---------------
The authors wish to thank Jan Kristensen for reading a preliminary version of the paper.
Preliminaries {#sec:prel}
=============
Throughout this paper we assume that $n>1$.
Function spaces {#sec:prelfspaces}
---------------
We define, reminiscent of [@Mazya], for $1{\leqslant}p{\leqslant}\infty$ and open $\Omega\subset{\mathbb{R}}^n$ $$\begin{aligned}
{\operatorname{W}}^{{\mathbb{A}},p}(\Omega)&:=\{u\in{\operatorname{L}}^p(\Omega,V)\colon {\mathbb{A}}u\in{\operatorname{L}}^p(\Omega)\},\\
{\operatorname{BV}}^{{\mathbb{A}}}(\Omega)&:=\{u\in{\operatorname{L}}^1(\Omega,V)\colon {\mathbb{A}}u\in\mathcal{M}(\Omega,W)\},\\
{\operatorname{V}}^{{\mathbb{A}},p}(\Omega)&:=\{u\in{\operatorname{W}}^{{\mathbb{A}},p}(\Omega)\colon\nabla^l u\in{\operatorname{L}}^p(\Omega,V\odot^l{\mathbb{R}}^n),l=1\ldots k-1\},\end{aligned}$$ and the homogeneous spaces $\dot{{\operatorname{W}}}^{{\mathbb{A}},p}$ as the closure of ${\operatorname{C}}_c^\infty({\mathbb{R}}^n,V)$ in the semi–norm $|u|_{{\mathbb{A}},p}:=\|{\mathbb{A}}u\|_{{\operatorname{L}}^p}$. In the case ${\mathbb{A}}=\nabla^k$, we write ${\operatorname{W}}^{k,p}(\Omega,V)$, ${\operatorname{V}}^{k,p}(\Omega,V)$. When it is clear from the context what the target space is, we abbreviate the ${\operatorname{L}}^p$–norm of maps defined on $\Omega$ by $\|\cdot\|_{p,\Omega}$. We denote the space of $V$–valued polynomials of degree at most $d$ in $n$ variables by ${\mathbb{R}}_d[x]^V$. We recall the weighted Bergman spaces $A^p_\alpha(\mathbb{D})$ of holomorphic maps defined on the open unit disc $\mathbb{D}\subset{\mathbb{C}}$, that are $p$–integrable with weight $w_\alpha(z)=(1-|z|^2)^\alpha$. It is well–known that these are Banach spaces under the ${\operatorname{L}}^p_{w_\alpha}$–norm for $1{\leqslant}p<\infty$ and $-1<\alpha<\infty$. We also recall, for $s>0$, $1{\leqslant}p,q<\infty$, the Besov space $$\begin{aligned}
{\operatorname{B}}^{s,p}_q(\Omega):=\{u\in{\operatorname{L}}^p(\Omega)\colon |u|_{{\operatorname{B}}^{s,p}_q(\Omega)}<\infty\},\end{aligned}$$ with an obvious choice of norm. Here, the Besov semi–norm is defined (see, e.g., [@devoresharpley Sec. 2]) for integer $r>s$ by $$\begin{aligned}
|u|_{{\operatorname{B}}^{s,p}_q(\Omega)}=\|u\|_{\dot{{\operatorname{B}}}^{s,p}_q(\Omega)}:=\left(\int_0^\infty \dfrac{\sup_{|h|<t}\|\Delta^r_h u\|^q_{{\operatorname{L}}^p(\Omega)}}{t^{1+sq}}{\operatorname{d}\!}t\right)^\frac{1}{q},\end{aligned}$$ where the $r$-th finite difference $\Delta^r_h u$ is defined to be zero if undefined, i.e., if at least one of $x+jh$, $j=1\ldots r$, falls outside $\Omega$. We also define the homogeneous space $\dot{{\operatorname{B}}}^{s,p}_q({\mathbb{R}}^n)$ as the closure of ${\operatorname{C}}^\infty_c({\mathbb{R}}^n)$ in the Besov semi–norm.
We also collect the assumptions on our operators. As in Section \[sec:intro\], we say that ${\mathbb{A}}$ is (${\mathbb{C}}$–)elliptic if and only if the linear map ${\mathbb{A}}[\xi]\colon V(+{\operatorname{i}}V)\rightarrow W(+{\operatorname{i}}W)$ is injective for all non–zero $\xi\in{\mathbb{R}}^n(+{\operatorname{i}}{\mathbb{R}}^n)$. Trivially, ${\mathbb{C}}$–elliptic operators are elliptic. We say that ${\mathbb{A}}$ has FDN (finite dimensional null–space) if and only if the vector space $\{u\in\mathscr{D}'({\mathbb{R}}^n,V)\colon{\mathbb{A}}u=0\}$ is finite dimensional. Finally, ${\mathbb{A}}$ is cancelling if and only if $\bigcap_{\xi\in S^{n-1}}{\mathbb{A}}[\xi](V)=\{0\}$.
Multi–linear Algebra
--------------------
Let $U,V$ be finite dimensional vector spaces and $l\in\mathbb{N}$. We write ${\mathscr{L}}(U,V)$ for the space of linear maps $U\rightarrow V$ and $V\odot^l U$ for the space of $V$–valued symmetric $l$–linear maps on $U$, a subspace of $V\otimes^l U$, the $V$–valued $l$–linear maps on $U$. This is naturally the space of the $l$–th gradients, i.e., ${\operatorname{D}\!}^l f(x)\in V\odot^l U$ for $f\in{\operatorname{C}}^l(U,V)$, $x\in U$. For more detail, see [@Federer Ch. 1]. We also write $a\otimes b=(a_i b_j)$ (the usual tensor product) and $\otimes^l a:=a\otimes a\otimes\ldots \otimes a$, where $a$ appears $l$ times on the right hand side. We single out the standard fact that $\widehat{\nabla^l f}(\xi)=\hat{f}(\xi)\otimes^l \xi\in V\odot^l U$ for $f\in\mathscr{S}(U,V)$, $\xi\in U$. We recall the pairing introduced in [@BDG], $v\otimes_{\mathbb{A}}\xi:={\mathbb{A}}[\xi]v$, which is reminiscent of the tensor product notation, i.e., if ${\mathbb{A}}={\operatorname{D}\!}$, we have $\otimes_{\mathbb{A}}=\otimes$. We have the following calculus rules if $k=1$: $$\begin{aligned}
{\mathbb{A}}(\rho u)&=\rho{\mathbb{A}}u+u\otimes_{\mathbb{A}}\nabla\rho\qquad\text{for }u\in{\operatorname{C}}^1({\mathbb{R}}^n,V), \rho\in{\operatorname{C}}^1({\mathbb{R}}^n),\\
{\mathbb{A}}(\phi(w))&=\phi^\prime(w)\otimes_{\mathbb{A}}\nabla w\qquad\text{for }\phi\in{\operatorname{C}}^1({\mathbb{R}},V),w\in{\operatorname{C}}^1({\mathbb{R}}^n).\end{aligned}$$ The above can easily be checked by direct computation and will be used without mention.
Harmonic Analysis {#sec:harmonic}
-----------------
Let ${\mathbb{A}}$ as in be elliptic and $u\in\mathscr{S}({\mathbb{R}}^n,V)$. We Fourier transform ${\mathbb{A}}u$ and apply the one–sided inverse $m_{\mathbb{A}}(\xi):=({\mathbb{A}}^*[\xi]{\mathbb{A}}[\xi])^{-1}{\mathbb{A}}^*[\xi]\in{\mathscr{L}}(W,V)$ of ${\mathbb{A}}[\xi]$ to get that $\hat{u}(x)=m_{\mathbb{A}}(\xi)\widehat{{\mathbb{A}}u}(\xi)$ for $\xi\in{\mathbb{R}}^n$ (we omitted the complex multiplicative constant arising from Fourier transforming, as it can be absorbed in the definition of $m_{\mathbb{A}}$). We define the $(k-n)$–homogeneous map $\textbf{G}_{\mathbb{A}}$ as the inverse Fourier transform of the $k$–homogeneous map $m_{\mathbb{A}}$. Thus we have the Green’s function representation $u=\textbf{G}_{\mathbb{A}}*{\mathbb{A}}u$. Moreover, we can extrapolate the following:
Let ${\mathbb{A}}$ as in be elliptic. Then there exists a $(1-n)$–homogeneous map $\mathbf{K}_{\mathbb{A}}\in{\operatorname{C}}^\infty({\mathbb{R}}^n\setminus\{0\},{\mathscr{L}}(W,V\otimes^k{\mathbb{R}}^n))$ such that $$\begin{aligned}
\label{eq:representation}
{\operatorname{D}\!}^{k-1}u(x)=\int_{{\mathbb{R}}^n}\mathbf{K}_{\mathbb{A}}(x-y){\mathbb{A}}u(y){\operatorname{d}\!}y\end{aligned}$$ for all $u\in{\operatorname{C}}^\infty_c({\mathbb{R}}^n,V)$.
We also record standard facts regarding Riesz potentials and singular integrals (see [@Stein Ch. II.4, Ch. V.1] and [@GT Lem. 7.2]), which we define by $$\begin{aligned}
I_\alpha f:=|\cdot|^{n-\alpha}*f\end{aligned}$$ for $\alpha\in[0,n)$ and measurable $f$. If $\alpha=0$, the convolution is understood in the sense of a principal value integral.
\[thm:anal\_harm\] We have that
1. $I_0$ is bounded on ${\operatorname{L}}^p({\mathbb{R}}^n)$ for $1<p<\infty$,
2. $I_\alpha$ is bounded ${\operatorname{L}}^p({\mathbb{R}}^n)\rightarrow{\operatorname{L}}^q({\mathbb{R}}^n)$ for $0<\alpha<n$, $p>1$, $q= np/(n-\alpha p)$,
3. \[itm:riesz\_domains\] $I_\alpha$ is bounded ${\operatorname{L}}^p(\Omega)\rightarrow{\operatorname{L}}^q(\Omega)$ for $0<\alpha<n$, $1{\leqslant}p{\leqslant}q<np/(n-\alpha p)$ with $$\begin{aligned}
\|I_\alpha(u)\|_{{\operatorname{L}}^q(\Omega)}{\leqslant}c({\operatorname{diam}}\Omega)^{\alpha-n(1/p-1/q)}\|u\|_{{\operatorname{L}}^p(\Omega)}\end{aligned}$$ for all $u\in{\operatorname{L}}^p(\Omega)$.
Examples {#sec:examples}
--------
We give examples of operators arising in conductivity, elasticity, plasticity and fluid mechanics ([@FM; @FS; @Milton]). Let ${\mathbb{A}}$ be as in . The facts that we use without mention are the main Theorems \[thm:main\], \[thm:tools\], and \[thm:main\_k\].
1. If ${\mathbb{A}}=\nabla^k$, we have that $\ker{\mathbb{A}}={\mathbb{R}}_{k-1}[x]^V$, so ${\mathbb{A}}$ has FDN, hence is EC. This, of course, corresponds to the case of classical Sobolev spaces, but we highlight it here to stress that our generalization brings a new perspective on their study.
2. If ${\mathbb{A}}u=\mathcal{E}u:=(\nabla u+(\nabla u)^\mathsf{T})/2$ is the symmetrized gradient, it is easy to see that $\ker{\mathbb{A}}$ is the space of rigid motions, i.e., affine maps of anti–symmetric gradient, so ${\mathbb{A}}$ has FDN, hence is EC. In this case, we recover the inequality in .
3. \[it:delbar\] Let ${\mathbb{A}}u=\mathcal{E}^D u:=\mathcal{E}u-({\operatorname{div}}u/n) \textbf{I}$, where $n\geq2$ and $\textbf{I}$ is the identity $n\times n$ matrix. If $n\geq3$, we have from [@Reshet] that $\ker{\mathbb{A}}$ is the space of conformal Killing vectors, so ${\mathbb{A}}$ has FDN, hence is EC. If $n=2$, we show in Counterexample \[ex:EC>FDN\] that ${\mathbb{A}}$ is elliptic. However, under the canonical identification ${\mathbb{R}}^2\cong{\mathbb{C}}$, we can also identify $\mathcal{E}^D$ with the anti–holomorphic derivative $\bar{\partial}$, so that we can further identify $\ker{\mathbb{A}}$ with the space of holomorphic functions, so ${\mathbb{A}}$ does not have FDN. Neither is ${\mathbb{A}}$ cancelling: by ellipticity, we have that $\mathcal{E}^D[\xi]({\mathbb{R}}^2)={\mathbb{R}}^2$. No critical embedding , can hold in this case.
4. Recall from [@FS A.2 (2.2)] that ${\operatorname{W}}^{{\operatorname{div}},1}\cap{\operatorname{W}}^{\mathcal{E}^D,1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{n/(n-1)}({\operatorname{B}},{\mathbb{R}}^n)$. By \[it:delbar\], if $n\geq3$ we can simplify and extend the embedding, whereas if $n=2$ the intersection is necessary.
5. If ${\mathbb{A}}=\Delta$, which is clearly elliptic, we have that $\ker{\mathbb{A}}$ is the space of all harmonic functions, so ${\mathbb{A}}$ does not have FDN and since ${\mathbb{A}}[\xi](V)=\left(\sum_{j=1}^n\xi_j^2\right){\mathbb{R}}^N={\mathbb{R}}^N$ for $\xi\in{\mathbb{R}}^n\setminus\{0\}$, neither is ${\mathbb{A}}$ cancelling.
6. One can use Lemma \[lem:FDNimpliesEC\] to prove non–rigidity. If ${\mathbb{A}}$ is elliptic, one can consider minimizers of the ${\mathbb{A}}$–Dirichlet energy $u\mapsto\int_{{\operatorname{B}}}|{\mathbb{A}}u|^2{\operatorname{d}\!}x$, which has Euler–Lagrange system ${\mathbb{A}}^*{\mathbb{A}}u=0$. Then $\Delta_{\mathbb{A}}:={\mathbb{A}}^*{\mathbb{A}}$ is elliptic, as $\langle({\mathbb{A}}^*{\mathbb{A}})[\xi] v,v\rangle=|{\mathbb{A}}[\xi]v|^2\gtrsim|\xi|^{2k}|v|^2$, where the last inequality follows from $|{\mathbb{A}}[\xi]v|>0$ on $\{|\xi|=1,|v|=1\}$ and homogeneity. Therefore $({\mathbb{A}}^*{\mathbb{A}})[\xi](V)=V$ for all $\xi\neq0$, so the Euler–Lagrange system above has infinite dimensional solution space.
EC Versus FDN {#sec:ECvsFDN}
=============
We begin by proving the first two statements in Theorem \[thm:tools\]. Throughout, $n>1$.
\[prop:FDNiffTypeC\] Let ${\mathbb{A}}$ be as in . Then ${\mathbb{A}}$ has FDN if and only if ${\mathbb{A}}$ is ${\mathbb{C}}$–elliptic.
From Theorem \[thm:Ka\], we have that if ${\mathbb{A}}$ is ${\mathbb{C}}$–elliptic, then $\ker{\mathbb{A}}$ consists of polynomials of fixed maximal degree. Suppose now that ${\mathbb{A}}$ is not ${\mathbb{C}}$–elliptic, so that there exist non–zero $\xi\in{\mathbb{C}}^n$, $v\in V+{\operatorname{i}}V$ such that ${\mathbb{A}}[\xi]v=0$. We define $u_f(x)=f(x\cdot\xi)v$, for holomorphic $f\colon{\mathbb{C}}\rightarrow{\mathbb{C}}$. It can be shown by direct real differentiation of real and imaginary parts and use of the Cauchy–Riemann equations for $f$ that ${\operatorname{D}\!}u_f(x)=(\partial_1 f)(x\cdot\xi)v\otimes\xi$. Since $\partial_1 f$ is itself holomorphic, inductively we get that ${\operatorname{D}\!}^l u_f(x)=(\partial^l_1f)(x\cdot\xi)v\otimes^l\xi$. We make the simple observation that there exists a linear map $A\in{\mathscr{L}}(V\otimes^k{\mathbb{R}}^n,W)$ such that ${\mathbb{A}}u=A({\operatorname{D}\!}^k u)$, so that by standard properties of the Fourier transform we get ${\mathbb{A}}[\eta]w=A(w\otimes^k\eta)$ for $\eta\in{\mathbb{R}}^n$, $w\in V$. It is then easy to see that ${\mathbb{A}}u_f(x)=(\partial_1^k f)(x\cdot\xi)A(v\otimes^k\xi)=0$. In particular, $\Re u_f,\Im u_f\in\ker{\mathbb{A}}$, so ${\mathbb{A}}$ has infinite dimensional null–space.
In light of this result, we we will henceforth use FDN and ${\mathbb{C}}$–ellipticity interchangeably. We next proceed to an instrumental ingredient for proving sufficiency of FDN for the embedding Theorem.
\[lem:FDNimpliesEC\] Let ${\mathbb{A}}$ be as in . If ${\mathbb{A}}$ has FDN, then ${\mathbb{A}}$ is cancelling.
We use Lemma \[lem:canc\]. Let $u\in{\operatorname{C}}^\infty({\mathbb{R}}^n,V)$ be such that $K:={\operatorname{spt}}{\mathbb{A}}u$ is compact. Consider an open ball ${\operatorname{B}}$ containing $K$. Cover the complement of ${\operatorname{B}}$ with an increasing chain of pairwise overlapping open balls $B_j$ such that ${\operatorname{B}}^c\subset \bigcup_j B_j\subset K^c$. In particular, we have ${\mathbb{A}}u=0$ in each $B_j$, so by Theorem \[thm:Ka\], $u$ must be a polynomial of degree at most $d({\mathbb{A}})$ in each $B_j$. Since the pairs of balls overlap on a set of positive measure, we get that $u$ equals a polynomial $P$ in ${\operatorname{B}}^c$ such that ${\mathbb{A}}P=0$ in ${\mathbb{R}}^n$. To conclude, we elaborate on the notation introduced in the proof of Proposition \[prop:FDNiffTypeC\]. Put $m:=\dim W$, so that we can write $(A\mathscr{V})_{l}=A^l\cdot\mathscr{V}$ for fixed $A^l\in V\otimes^k{\mathbb{R}}^n$, $l=1\ldots m$, and all $\mathscr{V}\in V\otimes^k{\mathbb{R}}^n$. For $l=1\ldots m$, we integrate by parts to get $$\begin{aligned}
\int_{{\operatorname{B}}}({\mathbb{A}}u)_l{\operatorname{d}\!}x&=\int_{{\operatorname{B}}}A^l\cdot {\operatorname{D}\!}^k u{\operatorname{d}\!}x=
\int_{\partial{\operatorname{B}}}A^l\cdot({\operatorname{D}\!}^{k-1}u\otimes\nu) {\operatorname{d}\!}\mathcal{H}^{n-1}\\
&=\int_{\partial{\operatorname{B}}}A^l\cdot({\operatorname{D}\!}^{k-1}P\otimes \nu){\operatorname{d}\!}\mathcal{H}^{n-1}=\int_{{\operatorname{B}}}A^l\cdot{\operatorname{D}\!}^k P{\operatorname{d}\!}x=\int_{{\operatorname{B}}}({\mathbb{A}}P)_l{\operatorname{d}\!}x=0,\end{aligned}$$ where $\nu$ denotes the unit normal to $\partial{\operatorname{B}}$. The proof is complete.
The converse of Lemma \[lem:FDNimpliesEC\], however, is not true in general. In what follows, we complete the comparison of the FDN condition and <span style="font-variant:small-caps;">Van Schaftingen</span>’s EC condition. We write $N:=\dim V$. The streamline here is that for $N=k=1$, ellipticity alone implies FDN (rendering these cases rather uninteresting), whereas in high dimensions, there are EC operators that are not FDN. Somewhat surprisingly, there are also a few instances in which ellipticity and ${\mathbb{C}}$–ellipticity differ, but EC implies FDN. We give the details below.
\[lem:EimpliesFDN\] Let ${\mathbb{A}}$ as in be elliptic, $N=k=1$. Then ${\mathbb{A}}$ has FDN.
Since $N=1$, it is clear that ${\mathbb{A}}$ is $\mathbb{F}$–elliptic, $\mathbb{F}\in\{{\mathbb{R}},{\mathbb{C}}\}$, if and only if the polynomials $({\mathbb{A}}[\xi])_l$, $l=1\ldots m$, have no common non–trivial zeroes in $\mathbb{F}$. Since we also assume $k=1$, we have ${\mathbb{A}}[\xi]=A\xi$ for some ${\mathbb{A}}\in{\mathbb{R}}^{m\times n}$. It is clear that all roots of the polynomials thus arising are real (in fact, ${\mathbb{A}}$ is elliptic if and only if $A\in\mathrm{GL}_n$).
If $n\geq3$, EC turns out to be insufficient for FDN, even for scalar fields or first order operators.
\[ex:EC>FDN\] Consider the operators $$\begin{aligned}
{\mathbb{A}}_{k,n} u&:=\nabla^{k-1}\left(\partial_1 u_1-\partial_2 u_2, \partial_2 u_1+\partial_1 u_2, \partial_j u_i\right)_{(i,j)\notin\{1,2\}\times\{1,2\}} \text{ for }N\geq2\\
\mathbb{B}_{k,n}u&:=\nabla^{k-2}\left(\partial^2_1 u+ \partial^2_2 u, \partial^2_j u\right)_{j=3\ldots n}\text{ for }N=1,k\geq2.\end{aligned}$$ If $n\geq3$ or $k\geq2$, then ${\mathbb{A}}_{k,n}$ is elliptic and cancelling, but has infinite dimensional null–space. The same is true of $\mathbb{B}_{k,n}$ if $n\geq3$ or $k\geq3$.
The failure of FDN is clear: simply take $$\begin{aligned}
u_{\mathbb{A}}(x)&:=\left(\Re f\left(x_1+{\operatorname{i}}x_2\right), \Im f\left(x_1 + {\operatorname{i}}x_2\right), 0,\ldots,0\right)^\mathsf{T}\\
u_\mathbb{B}(x)&:=g(x_1,x_2)\end{aligned}$$ for holomorphic $f$ and (scalar) harmonic $g$. We next show that ${\mathbb{A}}_{k,n}=\nabla^{k-1}{\mathbb{A}}_{1,n}$ is elliptic if $n,N\geq2$. We can reduce to ellipticity of ${\mathbb{A}}_{1,n}$, since for non–zero $\xi$, we have that $0={\mathbb{A}}_{k,n}[\xi]v=({\mathbb{A}}_{1,n}[\xi]v)\otimes^{k-1}\xi$, so ${\mathbb{A}}_{1,n}[\xi]v=0$. Let $1{\leqslant}j{\leqslant}n$ be such that $\xi_j\neq0$. If $j\geq3$, we clearly get $v=0$. If $1{\leqslant}j{\leqslant}2$, we get that $v_i=0$ for $3{\leqslant}i{\leqslant}N$. The remaining equations are $\xi_1v_1-\xi_2v_2=0=\xi_2v_1+\xi_1v_2$, with determinant $\xi_1^2+\xi_2^2>0$, so $v_1=0=v_2$. It remains to check that, under our assumptions, ${\mathbb{A}}_{k,n}$ is cancelling. The case $k>1$ is easier, since the composition of operators $\mathbb{L}_1\circ\mathbb{L}_2$ is cancelling if $\mathbb{L}_1$ is. This is simply due to the fact that $\textrm{im}(\mathbb{L}_1\circ\mathbb{L}_2)[\xi]=\mathbb{L}_1[\xi](\textrm{im}\mathbb{L}_2[\xi])\subseteq\textrm{im}\mathbb{L}_1[\xi]$. If $k=1$ and $n\geq3$ we can make a straightforward computation. Write $(w_l)_{l=1\ldots Nn-2}:={\mathbb{A}}_{1,n}[\xi]v$. For $w\in\bigcap_{\xi\neq0}{\mathbb{A}}_{1,n}[\xi](V)$, we can essentially test with different values of $\xi\neq0$. By choosing $\xi$ to have exactly one non–zero entry, we obtain that $w_l=0$ for $3{\leqslant}l{\leqslant}Nn-2$. Incidentally, when testing with $\xi$ such that $\xi_1=0=\xi_2$, we also obtain $w_1=0=w_2$, so all properties are checked for ${\mathbb{A}}_{k,n}$. Ellipticity of $\mathbb{B}_{k,n}$ is obvious, whereas cancellation is established analogously.
The two specific cases that are not covered by Lemma \[lem:EimpliesFDN\] and Counterexample \[ex:EC>FDN\] reveal that the classes EC and FDN can coincide even if they are strictly smaller than the class of elliptic operators.
\[lem:EC=FDN\] Let $n=2$ and ${\mathbb{A}}$ be as in $\eqref{eq:A}$ be elliptic but not ${\mathbb{C}}$–elliptic. If any of the following hold,
1. \[it:ECimpliesFDN\_a\] $N=1$, $k=2$,
2. \[it:ECimpliesFDN\_b\] $N\geq2$, $k=1$,
then ${\mathbb{A}}$ is not cancelling.
Suppose \[it:ECimpliesFDN\_a\]. Since $N=1$ and ${\mathbb{A}}$ is not ${\mathbb{C}}$–elliptic, the homogeneous, quadratic, scalar polynomials $({\mathbb{A}}[\xi])_l$, $l=1\ldots m$, must have a common complex root. This root cannot be real, as ${\mathbb{A}}$ is real–elliptic. It follows that $({\mathbb{A}}[\xi])_l$ are all multiples of the same quadratic polynomial $P:{\mathbb{R}}\rightarrow{\mathbb{R}}$, so that ${\mathbb{A}}[\xi]v=vP(\xi)w_0$ for all $v\in V\simeq{\mathbb{R}}$ and some $w_0\in W\setminus\{0\}$. It is clear then that ${\mathbb{A}}[\xi](V)={\mathbb{R}}w$ for all $\xi\neq0$. We next assume \[it:ECimpliesFDN\_b\]. Since ${\mathbb{A}}$ is elliptic, there exist linearly independent $\xi,\eta\in{\mathbb{R}}^2$, $v,w\in{\mathbb{R}}^N$ such that ${\mathbb{A}}[\xi]v={\mathbb{A}}[\eta]w$ and ${\mathbb{A}}[\xi]w=-{\mathbb{A}}[\eta]v$. We also have that any $\zeta\in{\mathbb{R}}^2$ can be written as $\zeta=a\xi+b\eta$. We put $v_\zeta:=a v+b w$. It follows that $$\begin{aligned}
{\mathbb{A}}[\zeta]v_\zeta={\mathbb{A}}[a\xi+b\eta](a v+b w)=(a^2+b^2){\mathbb{A}}[\xi]v,\end{aligned}$$ so that $\bigcap_{\zeta\in{\mathbb{R}}^2\setminus\{0\}}{\mathbb{A}}[\zeta](V)\ni{\mathbb{A}}[\xi]v\neq0$.
We remark that we can append the proof above by taking $w_\zeta:=b v-a w$ and obtain $\bigcap_{\zeta\neq0}{\mathbb{A}}[\zeta](V)\supset\{{\mathbb{A}}[\xi]v,{\mathbb{A}}[\xi]w\}$. Therefore, if $n=2$, $k=1$, and ${\mathbb{A}}$ is elliptic but not cancelling, then $\dim\bigcap_{\zeta\neq0}{\mathbb{A}}[\zeta](V)\geq2$. We do not know of any elliptic, non–cancelling, first order operator in higher dimensions for which this intersection is one–dimensional.
Insufficiency of EC {#sec:EC>emb}
-------------------
We next give examples of first order EC operators and domains $\Omega\subset{\mathbb{R}}^n$ for which the Sobolev–type embedding fails. This follows from Counterexample \[ex:EC>FDN\] above and from the next lemma, which is a strengthened version of the strict inclusion of (weighted) Bergman spaces in the language of non–FDN operators.
\[lem:EC>emb\] Let $k=1$ and ${\mathbb{A}}$ as in be elliptic but *not* have FDN, so there exist linearly independent $\eta_1,\eta_2\in{\mathbb{R}}^n$ such that ${\mathbb{A}}[\eta_1+{\operatorname{i}}\eta_2]$ has non–trivial kernel in $V+{\operatorname{i}}V$. Assume that $\eta_1,\eta_2$ are orthonormal. If any of the following holds:
1. \[itm:cylinder\] $\Omega:={\operatorname{B}}_{{\operatorname{span}}\{\eta_1,\eta_2\}}\times[0,1]^{n-2}$,
2. \[itm:ball\] $\Omega:={\operatorname{B}}$,
then there exists smooth $u\in{\operatorname{L}}^1\setminus\bigcup_{p>1}{\operatorname{L}}^p(\Omega,V)$ such that ${\mathbb{A}}u=0$.
We write $\xi=\eta_1+{\operatorname{i}}\eta_2$, and write $D$ for the unit disc in ${\operatorname{span}}\{\eta_1,\eta_2\}$. We stress that each $\eta_j$ must be non–zero by ellipticity of ${\mathbb{A}}$, so $D$ is indeed a non–degenerate disc. We also know from the proof of Proposition \[prop:FDNiffTypeC\] that there exist non–zero $v\in V+{\operatorname{i}}V$ such that ${\mathbb{A}}[\xi]v=0$, and one can show by direct computation that for any holomorphic function $f$ we can define $u_f(x):=f(x\cdot\xi)v$, for which ${\mathbb{A}}\Re u_f=0={\mathbb{A}}\Im u_f$. We have that $$\begin{aligned}
\int_{\Omega}|u_f(x)|^p{\operatorname{d}\!}x&=\int_D\int_{(\eta+\{\eta_1,\eta_2\}^\perp)\cap\Omega}|f(\eta\cdot\xi)|^p|v|^p{\operatorname{d}\!}\mathcal{H}^{n-2}{\operatorname{d}\!}\mathcal{H}^2(\eta)\\
&=|v|^p\int_D |f(\eta\cdot\xi)|^p\mathcal{H}^{n-2}\left((\eta+\{\eta_1,\eta_2\}^\perp)\cap\Omega\right){\operatorname{d}\!}\mathcal{H}^2(\eta)\end{aligned}$$ We now make the case distinction. Assume \[itm:cylinder\] holds, so $$\begin{aligned}
\int_{\Omega}|u_f(x)|^p{\operatorname{d}\!}x&=|v|^p\int_D |f(\eta\cdot\xi)|^p{\operatorname{d}\!}\mathcal{H}^2(\eta)=\int_{\mathbb{D}}|f(z)|^p{\operatorname{d}\!}\mathcal{L}^2(z).\end{aligned}$$ Assume \[itm:ball\], so $$\begin{aligned}
\int_{\Omega}|u_f(x)|^p{\operatorname{d}\!}x&=c(n)|v|^p\int_D |f(\eta\cdot\xi)|^p(1-|\eta|^2)^\frac{n-2}{2}{\operatorname{d}\!}\mathcal{H}^2(\eta)\\
&=c(n)|v|^p\int_{\mathbb{D}}|f(z)|^p(1-|z|^2)^\frac{n-2}{2}{\operatorname{d}\!}\mathcal{L}^2(z),\end{aligned}$$ where $c(n)$ denotes the volume of the $(n-2)$–dimensional ball. By Lemma \[lem:baire\] below, we can choose $f\in A^1_{\alpha}(\mathbb{D})\setminus\bigcup_{p>1}A^p_\alpha(\mathbb{D})$ for $\alpha=0$ and $\alpha=(n-2)/2$ respectively, so that both $\Re u_f$ and $\Im u_f$ are in ${\operatorname{L}}^1({\operatorname{B}},V)$, but one of them is in not in any other ${\operatorname{L}}^p$. This proves the claim.
The following Lemma is also feasible by direct computation, but we prefer to give an abstract argument for the sake of brevity.
\[lem:baire\] For all $1{\leqslant}p<\infty$, $\alpha\geq0$ the set $A^p_\alpha(\mathbb{D})\setminus\bigcup_{q>p}A^q_\alpha(\mathbb{D})$ is non–empty.
We abbreviate $A^p:=A_\alpha^p(\mathbb{D})$. The proof relies on the strict inclusion $A^q\subsetneq A^p$ for $1{\leqslant}p<q<\infty$ proved in [@ZZ Cor. 68] and a Baire category argument. Assume that the result is false, so that by Hölder’s Inequality we can find a sequence $q_j\downarrow p$ such that $A^p=\bigcup_j A^{q_j}$. For natural $l$, we define the sets $F_l^j:=\{f\in A^{q_j}\colon\|f\|_{A^{q_j}}^{q_j}{\leqslant}l\}$, which we claim to be closed in $A^p$. Let $f_m\in F^j_l$ converge to $f$ in $A^p$. By completeness of $A^p$, we have, by Fatou’s Lemma on a pointwise convergent, not relabelled subsequence that $$\begin{aligned}
\int_{\mathbb{D}}|f|^{q_j}w_\alpha{\operatorname{d}\!}\mathcal{L}^2{\leqslant}\liminf_{m\rightarrow\infty}\int_{\mathbb{D}}|f_m|^{q_j}w_\alpha{\operatorname{d}\!}\mathcal{L}^2{\leqslant}l,\end{aligned}$$ so that indeed $f\in F_l^j$. Since $A^{q_j}$ is a proper subspace of $A^p$, it follows that the sets $F^j_l$ are nowhere dense in $A^p$. It remains to notice that then $A^p=\bigcup_{j,l}F^j_l$, which contradicts completeness of $A^p$ by Baire’s Theorem.
Comparison to the <span style="font-variant:small-caps;">Bourgain</span>–<span style="font-variant:small-caps;">Brezis</span> condition
---------------------------------------------------------------------------------------------------------------------------------------
We recall here the assumptions on ${\mathbb{A}}$ (sufficient for EC) under which a general inequality of the type was first proved in [@BB07], in the case $k=1$ and $V={\mathbb{R}}^n$. In their notation, we write $({\mathbb{A}}u)_s=\langle L^{(s)},\nabla u\rangle$ for matrices $L^{(s)}\in{\mathbb{R}}^{n\times n}$, $s=1\ldots m$. It is shown in [@BB07 Thm. 25], that if an operator ${\mathbb{A}}$ is elliptic such that $\det L^{(s)}=0$ for $s=1\ldots m$, then holds. It is clear (either by [@VS Thm. 1.3] or by direct computation) that such operators are cancelling. By Lemma \[lem:EC=FDN\], if $n=2$, we have that such ${\mathbb{A}}$ also has FDN, and thus satisfies . However, if $n\geq3$, we show that ${\mathbb{A}}_{1,n}$ as in Counterexample \[ex:EC>FDN\] with $N=n$ satisfy the <span style="font-variant:small-caps;">Bourgain</span>–<span style="font-variant:small-caps;">Brezis</span> condition, but do not have FDN. We explicitly write down the matrices $L^{(s)}$ if $n=3$, the general case being a simple exercise: $$\begin{aligned}
&\left(\begin{array}{ccc}1 & 0 & 0 \\ 0 & -1 & 0\\ 0 & 0&0\end{array}\right),
\left(\begin{array}{ccc}0 & 1 & 0 \\ 1 & 0 & 0\\ 0 & 0&0\end{array}\right),
\left(\begin{array}{ccc}0 & 0 & 0 \\ 0 & 0 & 0\\ 1 & 0&0\end{array}\right),
\left(\begin{array}{ccc}0 & 0 & 0 \\ 0 & 0 & 0\\ 0 & 1&0\end{array}\right),\\
&\left(\begin{array}{ccc}0 & 0 & 0 \\ 0 & 0 & 0\\ 0 & 0&1\end{array}\right),
\left(\begin{array}{ccc}0 & 0 & 1 \\ 0 & 0 & 0\\ 0 & 0&0\end{array}\right),
\left(\begin{array}{ccc}0 & 0 & 0 \\ 0 & 0 & 1\\ 0 & 0&0\end{array}\right).\end{aligned}$$ By the reasoning in Section \[sec:EC>emb\], with ${\mathbb{A}}={\mathbb{A}}_{1,n}$, we have that $\dot{{\operatorname{W}}}^{{\mathbb{A}},1}({\mathbb{R}}^n)\hookrightarrow{\operatorname{L}}^{n/(n-1)}({\mathbb{R}}^n)$, but there are maps in ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ that have no higher integrability.
The Sobolev–type Embedding on Domains {#sec:proof}
=====================================
Jones–type Extension
--------------------
In this section we complete the proof of Theorem \[thm:tools\] with the following generalization:
\[thm:extension\] Let ${\mathbb{A}}$ as in have FDN, $1{\leqslant}p {\leqslant}\infty$, $\Omega\subset{\mathbb{R}}^n$ be a star–shaped domain with respect to a ball. Then there exists a bounded, linear extension operator $E_\Omega\colon{\operatorname{W}}^{{\mathbb{A}},p}(\Omega)\rightarrow{\operatorname{V}}^{{\mathbb{A}},p}({\mathbb{R}}^n)$.
To prove this result we use <span style="font-variant:small-caps;">Jones</span>’ method of extension developed in the celebrated paper [@Jones]. Recall that <span style="font-variant:small-caps;">Jones</span>’s original idea was to decompose a small neighbourhood of $\partial\Omega$ into small cubes and assign suitable polynomials of degree at most $k-1$ to each cube. Inspired by [@BDG Sec. 4.1-2], we assign polynomials in $\ker{\mathbb{A}}$ on such cubes, as explained below. With this crucial modification, the streamlined proof that we include below mostly follows the same lines as in [@Jones Sec. 2-3], where all the details we omit can be found. What deserves some special attention is a Poincaré–type inequality, which is interesting in its own right. We present it below and mention that it is a generalization of the results in [@BDG Sec. 3]. We extend the notation presented in Theorem \[thm:Ka\] by $\pi_\Omega u:=\Pi \mathcal{P} u$, where $\Pi$ denotes the ${\operatorname{L}}^2$–orthogonal projection of ${\mathbb{R}}_d[x]^V$ onto $\ker{\mathbb{A}}$.
\[prop:poinc\] Let ${\mathbb{A}}$ as in have FDN, $1{\leqslant}p{\leqslant}\infty$, $0{\leqslant}l<k$, and $\Omega\subset{\mathbb{R}}^n$ be a star–shaped domain with respect to a ball. Then there exists $c>0$ such that $$\begin{aligned}
\|\nabla^l(u-\pi_\Omega u)\|_{p,\Omega}{\leqslant}c({\operatorname{diam}}\Omega)^{k-l}\|{\mathbb{A}}u\|_{p.\Omega}\end{aligned}$$ for all $u\in{\operatorname{C}}^\infty(\bar{\Omega},V)$.
We start with $\|\nabla^l(u-\pi_\Omega u)\|_{p,\Omega}{\leqslant}\|\nabla^l(u-\mathcal{P} u)\|_{p,\Omega}+\|\nabla^l(\mathcal{P}u-\pi_\Omega u)\|_{p,\Omega}$, and estimate both terms. We have by the growth conditions on $K$ from Theorem \[thm:Ka\] that $$\begin{aligned}
\|\nabla^l(u-\mathcal{P} u)\|_{p,\Omega}&=\left(\int_{\Omega}\left\vert\int_{\Omega}\nabla^l_x K(x-y){\mathbb{A}}u(y){\operatorname{d}\!}y\right\vert^p{\operatorname{d}\!}x\right)^{\frac{1}{p}}\\
&\lesssim\left(\int_{\Omega}\left(\int_{\Omega}\dfrac{|{\mathbb{A}}u(y)|}{|x-y|^{n+l-k}}{\operatorname{d}\!}y\right)^p{\operatorname{d}\!}x\right)^{\frac{1}{p}},\end{aligned}$$ and we obtain the estimate by standard boundedness of Riesz potentials (see Theorem \[thm:anal\_harm\]\[itm:riesz\_domains\] for the precise scaling if $n+l-k>0$; the case $n+l-k{\leqslant}0$ follows by Hölder’s Inequality). We then note that $p\mapsto\|p-\Pi p\|_{p,\Omega}$ and $p\mapsto\|{\mathbb{A}}p\|_{p,\Omega}$ respectively define a semi–norm and a norm on the finite dimensional vector space ${\mathbb{R}}_d[x]^V/\ker{\mathbb{A}}$, so that $\|\nabla^l(\mathcal{P}u-\pi_\Omega u)\|_{p,\Omega}\lesssim\|{\mathbb{A}}\mathcal{P}u\|_{p,\Omega}$, with a domain dependent constant. We recall from the original proof of Theorem \[thm:Ka\] that $\mathcal{P}u$ is the averaged Taylor polynomial $$\begin{aligned}
\mathcal{P}u(x)=\int_{\Omega}\sum_{|\alpha|{\leqslant}d} \frac{\partial^\alpha_y\left((y-x)^\alpha w(y)\right)}{\alpha!}u(y){\operatorname{d}\!}y=\int_{\Omega}\sum_{|\alpha|{\leqslant}d}\frac{(x-y)^{\alpha}}{\alpha!} w(y)\partial^\alpha u(y){\operatorname{d}\!}y,\end{aligned}$$ where the weight $w$ is a smooth map supported in the ball with respect to which $\Omega$ is star–shaped such that $\int w=1$. One can show by direct computation that averaged Taylor polynomials “commute” with derivatives, in the sense that $$\begin{aligned}
{\mathbb{A}}\mathcal{P}u=\int_{\Omega} \sum_{|\beta|{\leqslant}d-k} \frac{\partial^\beta_y\left((y-\cdot)^\beta w(y)\right)}{\alpha!}{\mathbb{A}}u(y){\operatorname{d}\!}y.\end{aligned}$$ It is then obvious that $\|{\mathbb{A}}\mathcal{P}u\|_{p,\Omega}\lesssim\|{\mathbb{A}}u\|_{p,\Omega}$. The precise dependence of the constant on the domain follows by standard scaling arguments.
We next introduce the framework required to prove Theorem \[thm:extension\]. We use the same Whitney coverings as in [@Jones], which we recall for the reader’s convenience. Firstly recall the Decomposition Lemma introduced in [@Whitney], that any open subset $\Omega\subset{\mathbb{R}}^n$ can be covered with a countable collection $\mathcal{W}_1:=\{S_j\}$ of closed dyadic cubes satisfying
1. $\ell(S_j)/4{\leqslant}\ell(S_l){\leqslant}4 \ell(S_j)$ if $S_j\cap S_l\neq\emptyset$,
2. ${\operatorname{int}}S_j\cap{\operatorname{int}}S_l=\emptyset$ if $j\neq l$,
3. $\ell(S_j){\leqslant}\mathrm{dist}(S_j,\partial\Omega){\leqslant}4\sqrt{n}\ell(S_j)$ for all $j$,
where $\ell(Q)$ denotes the side–length of a cube $Q$. We henceforth assume that $\Omega$ is as in the statement of Theorem \[thm:extension\], so in particular $\Omega$ is an $(\varepsilon,\delta)$–domain. We also consider a Whitney decomposition $\mathcal{W}_2:=\{Q_l\}$ of ${\mathbb{R}}^n\setminus\bar{\Omega}$, and further define $\mathcal{W}_3:=\{Q\in\mathcal{W}_2\colon \ell(Q){\leqslant}\varepsilon\delta/(16n)\}$. We reflect each cube $Q\in\mathcal{W}_3$ to a non–unique interior cube $Q^*\in\mathcal{W}_1$ such that
1. $\ell(Q){\leqslant}\ell(Q^*){\leqslant}4\ell(Q)$,
2. $\mathrm{dist}(Q,Q^*){\leqslant}C\ell(Q)$,
where above and in the following $C$ denotes a constant depending on $k,p,n,\varepsilon,\delta$ only; additional dependencies will be specified. The non–uniqueness causes no issues, as one can show that
1. For any two choices $S_1,S_2$ of $Q^*$, we have $\mathrm{dist}(S_1,S_2){\leqslant}C\ell(Q)$,
2. For any $S\in\mathcal{W}_1$, there are at most $C$ cubes $Q\in\mathcal{W}_3$ such that $S=Q^*$,
3. For any adjacent $Q_1,Q_2\in\mathcal{W}_3$, we have $\mathrm{dist}(Q_1^*,Q_2^*){\leqslant}C\ell(Q_1)$.
For detail on theses basic properties of the reflection see [@Jones Lem. 2.4-7]. We conclude the presentation of the decomposition with the following simple modification of [@Jones Lem. 2.8]:
For any adjacent cubes $Q_1,Q_2\in\mathcal{W}_3$, there is a chain $\mathcal{C}(Q^*_1,Q^*_2):=\{Q_1^*=:S_1,S_2,\ldots S_m:=Q^*_2\}$ of cubes in $S_j\in\mathcal{W}_1$, i.e., such that $S_j$ and $S_{j+1}$ touch on an $(n-1)$–dimensional hyper–surface for all $j$, and $m{\leqslant}C$.
We proceed to define the extension operator $$\begin{aligned}
E_\Omega u:=
\begin{cases}
u&\text{in }\Omega\\
\sum_{Q\in\mathcal{W}_3}\varphi_Q \pi_{Q^*}u&\text{in }{\mathbb{R}}^n\setminus\bar{\Omega},
\end{cases}\end{aligned}$$ where $\{\varphi_Q\}_{Q\in\mathcal{W}_3}\subset{\operatorname{C}}^\infty({\mathbb{R}}^n)$ is a partition of unity such that for all $Q\in\mathcal{W}_3$ we have
1. $0{\leqslant}\varphi_Q{\leqslant}1$ and $\sum_{Q\in\mathcal{W}_3}\varphi_Q=1$ in $\bigcup\mathcal{W}_3$,
2. ${\operatorname{spt}}\varphi_Q\subset17/16Q$, where $\lambda Q$ denotes the homothety of $Q$ by $\lambda$ about its centre,
3. $|\nabla^l\varphi_Q|{\leqslant}C\ell(Q)^{-l}$ for all $0{\leqslant}l{\leqslant}k$.
Our proof mostly follows the lines of the original proof. We first prove an estimate on chains in $\mathcal{W}_1$, then suitably bound the norms of the derivatives in the exterior domains, and we conclude by showing that the extension has weak derivatives in full–space. We warn the reader that in the remainder of this section we may use the properties of the decomposition, reflection and partition of unity without mention.
\[lem:chain\] Let $\mathcal{C}:=\{S_1,\ldots S_m\}\subset\mathcal{W}_1$ be a chain. Then for $0{\leqslant}l<k$ we have $$\begin{aligned}
\|\nabla^l(\pi_{S_1}u-\pi_{S_m}u)\|_{p,S_1}{\leqslant}C(m)\ell(S_1)^{k-l}\|{\mathbb{A}}u\|_{p,\cup\mathcal{C}}\end{aligned}$$ for all $u\in{\operatorname{C}}^\infty(\bar{\Omega},V)$.
We remark that ${\operatorname{L}}^p$–norms of polynomials of degree at most $d$ on adjacent cubes in $\mathcal{W}_1$ are comparable (see, e.g., [@Jones Lem. 2.1]). We get $$\begin{aligned}
\mathrm{LHS}&{\leqslant}\sum_{j=1}^{m-1}\|\nabla^l(\pi_{S_{j+1}}u-\pi_{S_j}u)\|_{p,S_1}\\
&{\leqslant}C(m) \sum_{j=1}^{m-1}\|\nabla^l(\pi_{S_{j+1}}u-\pi_{S_j\cup S_{j+1}}u)\|_{p,S_{j+1}}+\|\nabla^l(\pi_{S_j\cup S_{j+1}}u-\pi_{S_j}u)\|_{p,S_j}\\
&{\leqslant}C(m)\sum_{j=1}^{m-1}\left(\|\nabla^l(\pi_{S_{j+1}}u-u)\|_{p,S_{j+1}}+2\|\nabla^l(u-\pi_{S_j\cup S_{j+1}}u)\|_{p,S_j\cup S_{j+1}}\right.\\
&\left.+\|\nabla^l(u-\pi_{S_j}u)\|_{p,S_j}\right)\end{aligned}$$ and we can use the Poincaré–type inequality, Proposition \[prop:poinc\], to conclude.
\[lem:localbounds\] For $1{\leqslant}p{\leqslant}\infty$, we have $\|E_\Omega u\|_{{\operatorname{V}}^{{\mathbb{A}},p}({\mathbb{R}}^n\setminus\bar{\Omega})}{\leqslant}C \|u\|_{{\operatorname{W}}^{{\mathbb{A}},p}(\Omega)}$ for all $u\in{\operatorname{C}}^\infty(\bar{\Omega},V)$.
We estimate on each cube in $\mathcal{W}_2$, distinguishing between small and large cubes. We also distinguish between ${\mathbb{A}}$ and the derivatives of order less than $k$. Let $Q_0\in\mathcal{W}_3$. Then, since $\varphi_Q$ sum to one in $Q_0$ and ${\mathbb{A}}\pi_{Q_0^*}u\equiv0$, we have $$\begin{aligned}
\|{\mathbb{A}}E_\Omega u\|_{p,Q_0}&=\left\|{\mathbb{A}}\sum_{Q\in\mathcal{W}_3}\varphi_Q( \pi_{Q^*}u-\pi_{Q_0^*}u)\right\|_{p,Q_0}{\leqslant}\left\|\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}{\mathbb{A}}(\varphi_Q( \pi_{Q^*}u-\pi_{Q_0^*}u))\right\|_{p,Q_0}\\
&{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=0}^{k-1}\||\nabla^{k-j}\varphi_Q| |\nabla^j(\pi_{Q^*}u-\pi_{Q_0^*}u)|\|_{p,Q_0}\\
&{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=0}^{k-1}\ell(Q_0)^{j-k}\|\nabla^j(\pi_{Q^*}u-\pi_{Q_0^*}u)\|_{p,Q_0^*}\\
&{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\|{\mathbb{A}}u\|_{p,\cup\mathcal{C}(Q_0^*,Q^*)},\end{aligned}$$ where the last inequality follows from Lemma \[lem:chain\]. With a similar reasoning we obtain, for $0{\leqslant}l{\leqslant}k-1$, that $$\begin{aligned}
\|\nabla^l E_\Omega u\|_{p,Q_0}{\leqslant}C\left(\|\nabla^l u\|_{p,Q_0^*}+\ell(Q_0)^{k-l}\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\|{\mathbb{A}}u\|_{p,\cup\mathcal{C}(Q_0^*,Q^*)}\right).\end{aligned}$$ We move on to the case $Q_0\in\mathcal{W}_2\setminus\mathcal{W}_3$, so if $Q\cap Q_0\neq\emptyset$, then $\ell(Q)\geq\ell(Q_0)/4\geq\varepsilon\delta/(64n)\geq C$. Let $0{\leqslant}l{\leqslant}k-1$, so that $$\begin{aligned}
\|\nabla^l E_\Omega u\|_{p,Q_0}&{\leqslant}\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\|\nabla^l(\varphi_Q \pi_{Q^*}u)\|_{p,Q_0}{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=1}^l \ell(Q_0)^{j-l}\|\nabla^{j} \pi_{Q^*}u\|_{p,Q_0}\\
&{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=1}^l \ell(Q_0)^{j-l}\|\nabla^{j} \pi_{Q^*}u\|_{p,Q^*}\\
&{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=1}^l \ell(Q_0)^{j-l}(\|\nabla^{j}( \pi_{Q^*}u-u)\|_{p,Q^*}+\|\nabla^{j}u\|_{p,Q^*})\\
&{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\|u\|_{{\operatorname{V}}^{l,p}(Q^*,V)}+\ell(Q_0)^{k-l} \|{\mathbb{A}}u\|_{p,Q^*}.\end{aligned}$$ As, above, we similarly show that $\|{\mathbb{A}}E_\Omega u\|_{p,Q_0}{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3} \|u\|_{{\operatorname{V}}^{{\mathbb{A}},p}(Q^*)}$. There is no loss in assuming that $\ell(Q_0){\leqslant}1$ for any $Q_0\in\mathcal{W}_2$, so that we can collect the estimates to obtain $$\begin{aligned}
\|E_\Omega u\|_{{\operatorname{V}}^{{\mathbb{A}},p}(Q_0)}{\leqslant}C \sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3} \|u\|_{{\operatorname{V}}^{{\mathbb{A}},p}(\mathcal{C}(Q_0^*,Q^*))}.\end{aligned}$$ It remains to use local finiteness of the partition of unity (see, e.g., [@Jones Eqn. (3.1-4)]) and Lemma \[lem:sob\_variants\] to conclude.
It remains to show that $E_\Omega u$ has weak derivatives in ${\mathbb{R}}^n$, for which it suffices to show that $E_\Omega$ maps ${\operatorname{V}}^{k,\infty}(\bar{\Omega},V)$ functions to ${\operatorname{V}}^{k,\infty}({\mathbb{R}}^n,V)$ functions. This we do in two steps. First, we show that the obvious candidate $(\nabla^l u)\chi_{\bar{\Omega}}+(\nabla^l E_\Omega u)\chi_{{\mathbb{R}}^n\setminus\bar{\Omega}}$ is bounded for all $0{\leqslant}l{\leqslant}k$. We need only prove this for $l=k$, the other cases being dealt with in Lemma \[lem:localbounds\] for $p=\infty$. As before, we first take $Q_0\in\mathcal{W}_3$, where $$\begin{aligned}
|\nabla^l E_\Omega u|&{\leqslant}|\nabla^k\pi_{Q_0^*}u|+\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}|\nabla^k(\varphi_Q(\pi_{Q^*} u-\pi_{Q_0^*}u))|\\
&{\leqslant}C\left( |\nabla^k\pi_{Q_0^*}u|+\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\|\nabla^k u\|_{\infty,\mathcal{C}(Q_0^*,Q^*)}\right).\end{aligned}$$ Clearly, $p\mapsto\|\nabla^k p\|_{\infty,Q_0^*}$ is a norm on ${\mathbb{R}}_d[x]^V/{\mathbb{R}}_{k-1}[x]$, whereas $p\mapsto\|\nabla^k \Pi p\|_{\infty,Q_0^*}$ is a semi–norm. We therefore get that $\|\nabla^k\pi_{Q_0^*}u\|_{\infty,Q_0^*}{\leqslant}C \|\nabla^k\mathcal{P}_{Q_0^*}u\|_{\infty,Q_0^*}{\leqslant}C\|\nabla^k u\|_{\infty,Q_0^*}$, where the latter inequality is given by stability of averaged Taylor polynomials. Now consider the other case, whence $Q_0\in\mathcal{W}_2\setminus\mathcal{W}_3$, and recall that then $\ell(Q_0)\geq C$. We have $$\begin{aligned}
|\nabla^l E_\Omega u|&{\leqslant}\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}|\nabla^k(\varphi_Q\pi_{Q^*})|{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=1}^k\ell(Q)^{j-k}|\nabla^j\pi_\Omega u|\\
&{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=1}^k\ell(Q_0)^{j-k}|\nabla^j\pi_\Omega u|{\leqslant}C\sum_{\emptyset\neq Q_0\cap Q\in\mathcal{W}_3}\sum_{j=1}^k|\nabla^j\pi_\Omega u|,\end{aligned}$$ so we can conclude as in the previous step. The second and final step is to show that $\nabla^l u$ is continuous for $0{\leqslant}l < k$. The proof of this fact can be found in [@Jones Lem. 3.5].
Proofs of the main results
--------------------------
We now begin the proof of Theorem \[thm:main\]. It is clear that \[it:main\_b\] implies \[it:main\_c\] and that \[it:main\_d\] implies \[it:main\_e\]. We first prove that \[it:main\_a\] implies \[it:main\_b\] in full generality.
Since ${\mathbb{A}}$ has FDN, by Theorem \[thm:tools\], there exists a bounded, linear extension operator $E_{{\operatorname{B}}} \colon{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\rightarrow{\operatorname{V}}^{{\mathbb{A}},1}({\mathbb{R}}^n)$. A close inspection of the proof of Theorem \[thm:extension\] reveals that $E_{{\operatorname{B}}}$ maps restrictions to ${\operatorname{B}}$ of ${\operatorname{C}}^\infty({\mathbb{R}}^n,V)$–functions into ${\operatorname{C}}^\infty_c(\tilde{{\operatorname{B}}},V)$ for a larger ball $\tilde{{\operatorname{B}}}\Supset{\operatorname{B}}$, which depends on ${\operatorname{B}}$ only. We write $p:=n/(n-k+s)$ and use Hölder’s Inequality to get that $$\begin{aligned}
\|u\|_{{{\operatorname{B}}}^{s,p}_q({\operatorname{B}},V)}&{\leqslant}\|E_{{\operatorname{B}}}u\|_{{{\operatorname{B}}}^{s,p}_q({\mathbb{R}}^n,V)}=\|E_{{\operatorname{B}}}u\|_{{\operatorname{L}}^p(\tilde{{\operatorname{B}}},V)}+\|E_{{\operatorname{B}}}u\|_{\dot{{\operatorname{B}}}^{s,p}_q({\mathbb{R}}^n,V)}\\
&\lesssim\|E_{{\operatorname{B}}}u\|_{{\operatorname{L}}^{\frac{n}{n-1}}(\tilde{{\operatorname{B}}},V)}+\|E_{{\operatorname{B}}}u\|_{\dot{{\operatorname{B}}}^{s,p}_q({\mathbb{R}}^n,V)}\\
&\lesssim\|\nabla^{k-1}E_{{\operatorname{B}}}u\|_{{\operatorname{L}}^{\frac{n}{n-1}}(\tilde{{\operatorname{B}}},V)}+\|E_{{\operatorname{B}}}u\|_{\dot{{\operatorname{B}}}^{s,p}_q({\mathbb{R}}^n,V)}\end{aligned}$$ where the last estimate follows from Poincaré’s Inequality with zero boundary values. We conclude by [@VS Thm. 1.3, Thm. 8.4] and boundedness of $E_{{\operatorname{B}}}$.
We will complete the proof of Theorem \[thm:main\_k\] at the end of this section. Returning to Theorem \[thm:main\], to see that \[it:main\_b\] implies \[it:main\_d\], we prove the following:
\[thm:compactness\] Let ${\mathbb{A}}$ be as in with $k=1$. Suppose that ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{\frac{n}{n-1}}({\operatorname{B}},V)$. Then ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow\hookrightarrow{\operatorname{L}}^{q}({\operatorname{B}},V)$ for all $1{\leqslant}q<\frac{n}{n-1}$.
The proof of Theorem \[thm:compactness\] relies on the Riesz-Kolmogorov criterion and the following Nikolskiĭ–type estimate:
\[lem:Besov\] Let ${\mathbb{A}}$ be an elliptic operator of the form , $k=1$. Fix $R>0$. Then for every $0<s<1$ there exists a constant $c=c(s,R)>0$ such that if $u\in{\operatorname{W}}^{{\mathbb{A}},1}({\mathbb{R}}^{n})$ vanishes identically outside ${\operatorname{B}}(0,R)$, then there holds $$\begin{aligned}
\int_{{\mathbb{R}}^{n}}|u(x+y)-u(x)|^{p}{\operatorname{d}\!}x {\leqslant}c\|{\mathbb{A}}u\|_{{\operatorname{L}}^{1}({\operatorname{B}}(0,R),W)}^{p}|y|^{sp}.\end{aligned}$$ whenever $p<n/(n-1+s)$.
Note that by <span style="font-variant:small-caps;">Ornstein</span>’s Non–Inequality, $s=1$ is not allowed in the lemma. A more general version, showing in addition that ellipticity is also necessary for the estimate, can be found in [@VS Prop. 8.22]. We include an elementary proof.
Fix $0<s<1$. By smooth approximation (see [@BDG]), it is no loss of generality to assume that $u\in{\operatorname{C}}_{c}^{\infty}({\mathbb{R}}^{n},V)$ and ${\operatorname{spt}}(u)\subset{\operatorname{B}}(0,R)$. Let $x\in{\mathbb{R}}^{n}$ be arbitrary but fixed and note that there exists a constant $c=c(s)>0$ such that for all $z,z',z''\in{\mathbb{R}}^{n}$ with $z\neq z',z\neq z''$ there holds $$\begin{aligned}
\label{eq:potentialest}
\left\vert \frac{1}{|z-z'|^{n-1}}-\frac{1}{|z-z''|^{n-1}}\right\vert {\leqslant}c |z'-z''|^{s}\Big(\frac{1}{|z-z'|^{n-1+s}}+\frac{1}{|z-z''|^{n-1+s}} \Big).\end{aligned}$$ Since ${\mathbb{A}}$ is elliptic, the representation formula yields by use of $$\begin{aligned}
\int_{{\mathbb{R}}^{n}}|u(x+y)-u(x)|^{p}{\operatorname{d}\!}x &\;\sim \int_{{\mathbb{R}}^{n}}\left\vert\int_{{\operatorname{B}}(0,R)}\frac{{\mathbb{A}}u(z)}{|x+y-z|^{n-1}}-\frac{{\mathbb{A}}u(z)}{|x-z|^{n-1}}{\operatorname{d}\!}z\right\vert^{p}{\operatorname{d}\!}x\\
& \stackrel{\eqref{eq:potentialest}}{\lesssim} |y|^{sp}\int_{{\mathbb{R}}^{n}}\left\vert\int_{{\operatorname{B}}(0,R)}\frac{{\mathbb{A}}u(z)}{|x+y-z|^{n-1+s}}-\frac{{\mathbb{A}}u(z)}{|x-z|^{n-1+s}}{\operatorname{d}\!}z\right\vert^{p}{\operatorname{d}\!}x\end{aligned}$$ and since for every $y\in{\mathbb{R}}^{n}$ by Young’s convolution inequality $$\begin{aligned}
\| (\mathbbm{1}_{{\operatorname{B}}(0,R)}|{\mathbb{A}}u|)*\tfrac{1}{|\cdot|^{n-1+s}}\|_{{\operatorname{L}}^{p}({\mathbb{R}}^{n})}{\leqslant}\|\mathbb{A} u\|_{{\operatorname{L}}^{1}({\mathbb{R}}^{n},W)}\|\tfrac{1}{|\cdot|^{n-1+s}}\|_{{\operatorname{L}}^{p}({\operatorname{B}}(0,3R))}, \end{aligned}$$ we conclude with the observation that $\|\tfrac{1}{|\cdot|^{n-1+s}}\|_{{\operatorname{L}}^{p}({\operatorname{B}}(0,3R))}<\infty$ if and only if $p<n/(n-1+s)$. The proof is complete.
Recall that by the Riesz–Kolmogorov Theorem on relatively compact subsets of ${\operatorname{L}}^{p}$–spaces [@BrezisFA Thm. 4.26] on $\Omega\subset{\mathbb{R}}^{n}$ open and $1{\leqslant}p<\infty$, a subset $\mathcal{F}\subset{\operatorname{L}}^{p}(\Omega,V)$ is relatively compact in ${\operatorname{L}}^{p}(\Omega,V)$ if and only if (i) $\mathcal{F}$ is a bounded set in ${\operatorname{L}}^{p}(\Omega,V)$ and (ii) for all $\varepsilon>0$ there exists $\delta>0$ such that for all $f\in\mathcal{F}$ and all $y\in{\mathbb{R}}^{n}$ with $|y|<\delta$ there holds $$\begin{aligned}
\|\overline{f}(\cdot+y)-\overline{f}(\cdot)\|_{{\operatorname{L}}^{p}({\mathbb{R}}^{n},V)}<\varepsilon, \end{aligned}$$ where $\overline{f}$ is the trivial extension of $f\in\mathcal{F}$ to ${\mathbb{R}}^{n}$.
Let $1{\leqslant}q<1^{*}$ and $\mathcal{F}$ be the unit ball in ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$. The embedding ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{1^{*}}({\operatorname{B}},V)$ implies that $\mathcal{F}$ is bounded in ${\operatorname{L}}^{q}({\operatorname{B}},V)$ which shows condition (i) of the Riesz–Kolmogorov criterion. As to (ii), let $\varepsilon>0$ be arbitrary. Given $\varrho>0$ sufficiently small (to be determined later on), let $\widetilde{\rho}_{\varrho}\colon [0,1]\to[0,1]$ be the Lipschitz function given by $$\begin{aligned}
\rho_{\varrho}(t):=\begin{cases} 1&\;\text{if}\;0<t<1-2\varrho,\\
-\frac{1}{\varrho}t+\frac{1-\varrho}{\varrho}&\;\text{if}\;1-2\varrho<t<1-\varrho,\\
0&\;\text{if}\;1-\varrho<t{\leqslant}1,
\end{cases}\end{aligned}$$ and put $\rho_{\varrho}(x):=\widetilde{\rho}_{\varrho}(|x|)$, $x\in{\mathbb{R}}^{n}$, and finally set, for given $f\in\mathcal{F}$, $f_{\varrho}:=\rho_{\varrho}f$. Denoting ${\operatorname{B}}_{t}:={\operatorname{B}}(0,t)$ for $t>0$, we note that if $|y|<\varrho$, then $f(\cdot+y)-f(\cdot)$ and $f_{\varrho}(\cdot+y)-f_{\varrho}(\cdot)$ coincide on ${\operatorname{B}}_{1-3\varrho}$. Let $f\in{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$ be arbitrary. We split $$\begin{aligned}
\int_{{\mathbb{R}}^{n}}|\overline{f}(x+y)-\overline{f}(x)|^{q}{\operatorname{d}\!}x = \left(\int_{{\mathbb{R}}^{n}\setminus{\operatorname{B}}_{1-3\varrho}}+\int_{{\operatorname{B}}_{1-3\varrho}}\right)|\overline{f}(x+y)-\overline{f}(x)|^{q}{\operatorname{d}\!}x =:\mathbf{I}_{\varrho}+\mathbf{II}_{\varrho},\end{aligned}$$ with an obvious definition of $\mathbf{I}_{\varrho}$ and $\mathbf{II}_{\varrho}$.
Ad $\mathbf{I}_{\varrho}$. As $|y|<\varrho$, if $x\in{\mathbb{R}}^{n}\setminus{\operatorname{B}}_{1-3\varrho}$, then $x+y\in{\mathbb{R}}^{n}\setminus{\operatorname{B}}_{1-4\varrho}$. Therefore, we obtain with a constant $c>0$ independent of $f\in\mathcal{F}$ $$\begin{aligned}
\mathbf{I}_{\varrho} {\leqslant}c\int_{{\operatorname{B}}_{1}\setminus{\operatorname{B}}_{1-4\varrho}}|f(z)|^{q}{\operatorname{d}\!}x & {\leqslant}c\left(\int_{{\operatorname{B}}}|f|^{\frac{n}{n-1}}{\operatorname{d}\!}x\right)^{\frac{(n-1)q}{n}}\mathscr{L}^{n}({\operatorname{B}}_{1}\setminus{\operatorname{B}}_{1-4\varrho})^{\frac{n-q(n-1)}{n}}\\
& {\leqslant}c\mathscr{L}^{n}({\operatorname{B}}_{1}\setminus{\operatorname{B}}_{1-4\varrho})^{\frac{n-q(n-1)}{n}}\end{aligned}$$ and we may hence record that there exists $\delta_{1}>0$ such that if $0<\varrho<\delta_{1}$, then $\mathbf{I}_{\varrho}<\varepsilon/3$.
Ad $\mathbf{II}_{\varrho}$. Firstly, since $1{\leqslant}q<n/(n-1)$, we find and fix $0<s<1$ such that $q<n/(n-1+s)$. By Lemma \[lem:nec\_ell\], ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{1^{*}}({\operatorname{B}},V)$ implies that ${\mathbb{A}}$ is elliptic so that we are in position to suitably apply Lemma \[lem:Besov\]. Since $f(\cdot+y)-f(\cdot)$ equals $f_{\varrho}(\cdot+y)-f_{\varrho}(\cdot)$ on ${\operatorname{B}}_{1-3\varrho}$ and $f_{\varrho}$ is compactly supported in ${\mathbb{R}}^{n}$ with supports in a sufficiently large fixed ball, we find with a constant $c>0$ independent of $f\in\mathcal{F}$ $$\begin{aligned}
\label{eq:kolmogorovest}
\begin{split}
\mathbf{II}_{\varrho} & = \int_{{\operatorname{B}}_{1-3\varrho}}|f_{\varrho}(x+y)-f_{\varrho}(x)|^{q}{\operatorname{d}\!}x {\leqslant}\int_{{\mathbb{R}}^{n}}|f_{\varrho}(x+y)-f_{\varrho}(x)|^{q}{\operatorname{d}\!}x \\
& {\leqslant}c\|{\mathbb{A}}f_{\varrho}\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}|y|^{sq}\\
& {\leqslant}c\big(\|\rho_{\varrho}{\mathbb{A}}f\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}+\|f\otimes_{{\mathbb{A}}}{\operatorname{D}\!}\rho_{\varrho}\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}\big)|y|^{sq}\\
& {\leqslant}c\big(\|{\mathbb{A}}f\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}+\|f\otimes_{{\mathbb{A}}}{\operatorname{D}\!}\rho_{\varrho}\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}\big)|y|^{sq}
\end{split}\end{aligned}$$ Pick $\delta_{2}>0$ such that if $|y|<\delta_{2}$, then $c\sup_{f\in\mathcal{F}}\|{\mathbb{A}}f\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}|y|^{sq}<\varepsilon/3$. Finally, we note because of $|{\operatorname{D}\!}\rho_{\varrho}|{\leqslant}4/\varrho$ by definition of $\rho_{\varrho}$, $$\begin{aligned}
\left(\int_{{\mathbb{R}}^{n}}|{\operatorname{D}\!}\rho_{\varrho}|^{n}{\operatorname{d}\!}x\right)^{\frac{1}{n}} {\leqslant}\frac{c}{\varrho}\big( (1-2\varrho)^{n}-(1-3\varrho)^{n}\big)^{\frac{1}{n}}=1+\mathcal{O}(\varrho)\end{aligned}$$ and so, by ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^{1^{*}}({\operatorname{B}},V)$ and since $0<\varrho{\leqslant}1$, $$\begin{aligned}
\|{\operatorname{D}\!}\rho_{\varrho}\otimes_{{\mathbb{A}}}f\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}|y|^{sq} & {\leqslant}\left(\int_{{\mathbb{R}}^{n}}|{\operatorname{D}\!}\rho_{\varrho}|^{n}{\operatorname{d}\!}x\right)^{\frac{q}{n}}\sup_{f\in\mathcal{F}}\|f\|_{{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})}^{q}|y|^{sq} \\ & {\leqslant}C\sup_{f\in\mathcal{F}}\|f\|_{{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})}^{q}|y|^{sq}, \end{aligned}$$ and from here it is evident that there exists $\delta_{3}>0$ such that if $|y|<\delta_{3}$, then $\|{\operatorname{D}\!}\rho_{\varrho}\otimes_{{\mathbb{A}}}f\|_{{\operatorname{L}}^{1}({\operatorname{B}}_{R},W)}^{q}|y|^{sq}<\varepsilon/3$ and so, by , $\mathbf{II}_{\delta}<2\varepsilon/3$ for all $f\in\mathcal{F}$. Now let $0<\delta<\varrho:=\min\{\delta_{1},\delta_{2},\delta_{3}\}$. Collecting estimates, we see that (ii) is satisfied and thus ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow\hookrightarrow{\operatorname{L}}^{q}({\operatorname{B}},V)$.
With an inexpensive modification of the proof of Theorem \[thm:compactness\], one can show that \[it:main\_c\] implies that ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow\hookrightarrow{\operatorname{L}}^q({\operatorname{B}},V)$ for all $1{\leqslant}q<p$, which trivially then implies \[it:main\_e\].
It remains to see that \[it:main\_e\] implies \[it:main\_a\], which is now a simple consequence of the Equivalence Lemma \[lem:equivalencelemma\]. We choose $E_{1}={\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})$, $E_{2}={\operatorname{L}}^{1}({\operatorname{B}},W)$, $E_{3}={\operatorname{L}}^{1}({\operatorname{B}},V)$, and $A:={\mathbb{A}}\in\mathscr{L}({\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}}),{\operatorname{L}}^{1}({\operatorname{B}},W))$, whereas $B:=\iota$ is the embedding operator $\iota\colon{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow\hookrightarrow {\operatorname{L}}^{1}({\operatorname{B}},V)$. It is then clear that $\|u\|_{{\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})}= \|u\|_*$, so the equivalence lemma yields that ${\mathbb{A}}$ has finite dimensional null–space.
Assume that the embedding holds. By standard embeddings of Besov spaces, we have that ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{W}}^{k-1,p}({\operatorname{B}},V)$ for some $p>1$. If $k=1$, we use Theorem \[thm:main\], \[it:main\_c\] implies \[it:main\_a\], to see that ${\mathbb{A}}$ has FDN. Otherwise, we give the following simple argument: assume that ${\mathbb{A}}$ is not FDN, so that the maps $u_j(x)=\exp(jx\cdot\xi)v$ lie in $\ker{\mathbb{A}}$ for some non–zero complex $\xi,v$. We traced this example back to [@Smith], but it was likely known before. The assumed embedding and Hölder’s Inequality give $$\begin{aligned}
j^{k-1}\left(\int_{{\operatorname{B}}}|\exp(jx\cdot\xi)|^p{\operatorname{d}\!}x\right)^\frac{1}{p}&\lesssim \|u_j\|_{{\operatorname{W}}^{k-1,p}({\operatorname{B}},V)}\lesssim\|u_j\|_{{\operatorname{L}}^1({\operatorname{B}},V)}\\
&\lesssim\left(\int_{{\operatorname{B}}}|\exp(jx\cdot\xi)|^p{\operatorname{d}\!}x\right)^\frac{1}{p},\end{aligned}$$ which leads to a contradiction. Here constants depend on ${\operatorname{diam}}{\operatorname{B}},p,n$ only.
Appendix
========
Miscellaneous background
------------------------
The following relevant facts we quote without proof:
\[lem:canc\] Let ${\mathbb{A}}$ as in be elliptic. Then ${\mathbb{A}}$ is cancelling if and only if we have that $$\begin{aligned}
\int_{{\mathbb{R}}^n}{\mathbb{A}}u{\operatorname{d}\!}x=0\end{aligned}$$ for all $u\in{\operatorname{C}}^\infty({\mathbb{R}}^n,V)$ such that the support of ${\mathbb{A}}u$ is compact.
\[lem:equivalencelemma\] Let $E_{1}$ be a Banach space and let $E_{2},E_{3}$ be two normed spaces (with corresponding norms $\|\cdot\|_{i}$, $i\in\{1,2,3\}$) and let $A\in\mathscr{L}(E_{1},E_{2})$ and $B\in\mathscr{L}(E_{1},E_{3})$ be two bounded linear operators such that $B$ is compact and the norms $\|\cdot\|_{1}$ and $\|\cdot\|_{*}:=\|A\cdot\|_{2}+\|B\cdot\|_{3}$ are equivalent on $E_{1}$. Then $\dim(\ker A))<\infty$.
\[thm:Ka\] Let ${\mathbb{A}}$ as in be ${\mathbb{C}}$–elliptic, and $\Omega\subset{\mathbb{R}}^n$ be a star–shaped domain with respect to a ball. Then there exist an integer $d:=d({\mathbb{A}})$, a linear map $\mathcal{P}\in{\mathscr{L}}({\operatorname{C}}^\infty(\bar{\Omega},V),{\mathbb{R}}_d[x]^V)$ and a smooth map $K\in{\operatorname{C}}^\infty({\mathbb{R}}^n\setminus\{0\},{\mathscr{L}}(W,V))$ such that $|\partial^\alpha K|\sim|\cdot|^{k-n-|\alpha|}$ and $$\begin{aligned}
u(x)=\mathcal{P}u(x)+\int_{\Omega}K(x-y){\mathbb{A}}u(y){\operatorname{d}\!}y\end{aligned}$$ for all $x\in\Omega$ and $u\in{\operatorname{C}}^\infty(\bar{\Omega},V)$. Therefore $\ker{\mathbb{A}}\subseteq{\mathbb{R}}_d[x]^V$.
Other facts about ${\operatorname{W}}^{{\mathbb{A}},p}$
-------------------------------------------------------
We collect some complementary results that explain, e.g., our choice of definition for the ${\mathbb{A}}$–Sobolev spaces and of extension technique for $p=1$.
\[lem:sob\_variants\] Let ${\mathbb{A}}$ be as in have FDN. Then ${\operatorname{W}}^{{\mathbb{A}},p}({\operatorname{B}},V)\simeq{\operatorname{V}}^{{\mathbb{A}},p}({\operatorname{B}},V)$, for $1{\leqslant}p {\leqslant}\infty$.
One embedding is clear. Let $u\in{\operatorname{W}}^{{\mathbb{A}},p}({\operatorname{B}},V)$. We recall from Theorem \[thm:Ka\] that $u$ can be represented as $u(x)=\mathcal{P}u(x)+\int_{{\operatorname{B}}}K(x-y){\mathbb{A}}u(y){\operatorname{d}\!}y$, where $\mathcal{P}u$ is a polynomial of degree at most $d({\mathbb{A}})$ and $|\partial^\alpha K|\sim|\cdot|^{k-n-|\alpha|}$. Let $1{\leqslant}l{\leqslant}k-1$. Then $$\begin{aligned}
\|\nabla^l u\|_{{\operatorname{L}}^p}{\leqslant}\|\nabla^l(u-\mathcal{P}u)\|_{{\operatorname{L}}^p}+\|\nabla^l\mathcal{P}u\|_{{\operatorname{L}}^p}.\end{aligned}$$ The first term can easily be controlled by the ${\operatorname{L}}^p$–norm of ${\mathbb{A}}u$ by Theorem \[thm:anal\_harm\]\[itm:riesz\_domains\]. The latter term defines a semi–norm on the space of polynomials of degree at most $d({\mathbb{A}})$, so it can be controlled by the ${\operatorname{L}}^p$–norm. We get $$\begin{aligned}
\|\nabla^l u\|_{{\operatorname{L}}^p}{\leqslant}\|{\mathbb{A}}u\|_{{\operatorname{L}}^p}+\|\mathcal{P}u\|_{{\operatorname{L}}^p}\lesssim \|{\mathbb{A}}u\|_{{\operatorname{L}}^p}+\|\mathcal{P}u-u\|_{{\operatorname{L}}^p}+\|u\|_{{\operatorname{L}}^p}\lesssim \|{\mathbb{A}}u\|_{{\operatorname{L}}^p}+\|u\|_{{\operatorname{L}}^p},\end{aligned}$$ which concludes the proof. Here constants depend on the domain.
\[lem:extp>1\] Let ${\mathbb{A}}$ as in have FDN, $1< p <\infty$, and $\Omega\subset{\mathbb{R}}^n$ be a star–shaped domain with respect to a ball. Then there exists a bounded, linear extension operator $E_\Omega\colon{\operatorname{W}}^{{\mathbb{A}},p}(\Omega)\rightarrow{\operatorname{V}}^{k,p}({\mathbb{R}}^n,V)$.
We use the extension suggested in [@Kalamajska94], namely, in the notation of Theorem \[thm:Ka\], $$\begin{aligned}
E_\Omega u(x):=\eta(x)\left(\mathcal{P}u(x)+\int_{\Omega}K(x-y){\mathbb{A}}u (y){\operatorname{d}\!}y\right)\end{aligned}$$ for $u\in{\operatorname{C}}^\infty(\bar{\Omega},V)$ and $x\in{\mathbb{R}}^n$. Here $\eta\in{\operatorname{C}}^\infty_c({\mathbb{R}}^n)$ is a smooth cut–off that equals 1 in a neighbourhood of $\Omega$. We abbreviate $\mathcal{K}u:=K*({\mathbb{A}}u\chi_\Omega)$. Let $0{\leqslant}l{\leqslant}k$, and let ${\operatorname{B}}$ be a ball containing the support of $\eta$. Then, with domain dependent constants, $$\begin{aligned}
\|\nabla^l E_\Omega u\|_{p,{\operatorname{B}}}\lesssim \sum_{j=0}^l \|\nabla^j(\mathcal{P}u+\mathcal{K}u)\|_{p,{\operatorname{B}}}{\leqslant}\|\mathcal{P}u\|_{{\operatorname{V}}^{l,p}({\operatorname{B}},V)}+\sum_{j=0}^l \|\nabla^j\mathcal{K}u\|_{p,{\operatorname{B}}}.\end{aligned}$$ We note that $\|\cdot\|_{{\operatorname{V}}^{l,p}({\operatorname{B}},V)}$ and $\|\cdot\|_{{\operatorname{L}}^p(\Omega,V)}$ both define norms on ${\mathbb{R}}_d[x]^V$, hence they are equivalent. We also remark that $\nabla^j\mathcal{K}u=(\nabla^j K)*({\mathbb{A}}u\chi_\Omega)$, so that $\|\nabla^j\mathcal{K}u\|_{p,{\operatorname{B}}}{\leqslant}\|\nabla^j\mathcal{K}u\|_{p,{\mathbb{R}}^n}\lesssim\|{\mathbb{A}}u\|_{p,\Omega}$, where we use the growth bounds on the derivatives of $K$ and boundedness of Riesz potentials, and, in the case $j=l=k$, of singular integrals (Theorem \[thm:anal\_harm\]). Collecting, we get $$\begin{aligned}
\|\nabla^l E_\Omega u\|_{p,{\operatorname{B}}}\lesssim\|\mathcal{P}u\|_{p,\Omega}+\|{\mathbb{A}}u\|_{p,\Omega}{\leqslant}\|\mathcal{P}u+\mathcal{K}u\|_{p,\Omega}+\|\mathcal{K}u\|_{p,\Omega}+\|{\mathbb{A}}u\|_{p,\Omega}\lesssim\|u\|_{{\operatorname{W}}^{{\mathbb{A}},p}(\Omega)},\end{aligned}$$ where the last inequality is obtained from the representation formula and, again, boundedness of Riesz potentials.
\[lem:nec\_ell\] Let ${\mathbb{A}}$ be as in , $k=1$. If ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{L}}^p({\operatorname{B}},V)$ for some $p>1$, then ${\mathbb{A}}$ is elliptic.
The proof is similar to that of Lemma \[lem:EC>emb\]. Suppose ${\mathbb{A}}$ is not elliptic, such that there exist $\xi\in S^{n-1}$, $v\in V\setminus\{0\}$ such that ${\mathbb{A}}[\xi]v=0$. We put $u(x)=f(x\cdot\xi)v$ for some measurable $f\colon(-1,1)\rightarrow{\mathbb{R}}$. We have that $$\begin{aligned}
\int_{{\operatorname{B}}} |u|^q{\operatorname{d}\!}x=\int_{-1}^{1}\int_{\{\xi\}^\perp\cap{\operatorname{B}}}|f(y)|^q|v|^q{\operatorname{d}\!}\mathcal{H}^{n-1}{\operatorname{d}\!}y
=c(n)|v|^q\int_{-1}^{1}|f(y)|^q(1-y^2)^{\frac{n-1}{2}}{\operatorname{d}\!}y,\end{aligned}$$ so that $$\begin{aligned}
c(n,q)\int_{-1/2}^{1/2}|f(y)|^q{\operatorname{d}\!}y{\leqslant}\|u\|_{{\operatorname{L}}^q({\operatorname{B}},V)}{\leqslant}C(n,q)\int_{-1}^{1}|f(y)|^q{\operatorname{d}\!}y.\end{aligned}$$ We now let $f\in{\operatorname{L}}^1(-1,1)$ such that $\int_{-1/2}^{1/2}|f|^p=\infty$, let ${\operatorname{C}}^\infty_c(-1,1)\ni\varphi_j\rightarrow f$ in ${\operatorname{L}}^1$, and put $u_j(x)=\varphi_j(x\cdot\xi)v$ for $x\in{\operatorname{B}}$. It is then clear that ${\mathbb{A}}u_j=0$ in the classical sense, and the assumed estimate applied to $u_j$ gives $$\begin{aligned}
\int_{-1/2}^{1/2}|\varphi_j(y)|^p{\operatorname{d}\!}y\lesssim\int_{-1}^{1}|\varphi(y)|{\operatorname{d}\!}y,\end{aligned}$$ which contradicts the choice of $f$.
\[lem:embimpliesEC\] Let ${\mathbb{A}}$ be as in . If ${\operatorname{W}}^{{\mathbb{A}},1}({\operatorname{B}})\hookrightarrow{\operatorname{W}}^{k-1,n/(n-1)}({\operatorname{B}},V)$, then ${\mathbb{A}}$ is elliptic and cancelling.
Necessity of ellipticity follows via Lemma \[lem:nec\_ell\] for $k=1$ or via simplifying the arguments for necessity of ${\mathbb{C}}$–ellipticity in the proof of Theorem \[thm:main\_k\] for $k>1$. We leave the details to the interested reader.
We next show that our assumed embedding implies $\dot{{\operatorname{W}}}^{{\mathbb{A}},1}({\mathbb{R}}^n)\hookrightarrow\dot{{\operatorname{W}}}^{k-1,n/(n-1)}({\mathbb{R}}^n,V)$ by a scaling argument and use the necessity part of [@VS Thm. 1.3]. Let $u\in{\operatorname{C}}^\infty_c({\mathbb{R}}^n,V)$ be such that ${\operatorname{spt}}u\subset{\operatorname{B}}_r:={\operatorname{B}}(0,r)$. Then $u_r(x):=u(rx)$ for $x\in{\mathbb{R}}^n$ is also a test function, with ${\operatorname{spt}}u_r\subset{\operatorname{B}}:={\operatorname{B}}(0,1)$. We estimate, with constants independent of $r$: $$\begin{aligned}
\|{\operatorname{D}\!}^{k-1} u\|_{{\operatorname{L}}^{\frac{n}{n-1}}}&=\left(\int_{{\operatorname{B}}_r}|{\operatorname{D}\!}^{k-1}u(x)|^{\frac{n}{n-1}}{\operatorname{d}\!}x\right)^{\frac{n-1}{n}}=\left(\int_{{\operatorname{B}}}r^{\frac{n(k-1)}{n-1}}|{\operatorname{D}\!}^{k-1}u_r(y)|^{\frac{n}{n-1}}r^n{\operatorname{d}\!}y\right)^{\frac{n-1}{n}}\\
&=r^{n-k}\left(\int_{{\operatorname{B}}}|{\operatorname{D}\!}^{k-1}u_r(y)|^{\frac{n}{n-1}}{\operatorname{d}\!}y\right)^{\frac{n-1}{n}}{\leqslant}cr^{n-k}\int_{{\operatorname{B}}}|{\mathbb{A}}u_r(y)|+|u_r(y)|{\operatorname{d}\!}y\\
&{\leqslant}c\int_{{\operatorname{B}}_r}|{\mathbb{A}}u(x)|{\operatorname{d}\!}x=c\|{\mathbb{A}}u\|_{{\operatorname{L}}^1},\end{aligned}$$ where the last inequality follows from a change of variable and the Poincaré–type inequality with zero boundary values (for elliptic operators) $$\begin{aligned}
\label{eq:zerotracepoinc}
\|v\|_{{\operatorname{L}}^1(\Omega,V)}{\leqslant}c({\operatorname{diam}}\Omega)^{n-k}\|{\mathbb{A}}v\|_{{\operatorname{L}}^1(\Omega,W)}\end{aligned}$$ for all $v\in{\operatorname{C}}^\infty_c(\Omega,V)$.
The inequality follows from the Green–type Formula and Theorem \[thm:anal\_harm\]\[itm:riesz\_domains\]. So does the inequality in Lemma \[lem:zerotraceemb\] below.
\[lem:zerotraceemb\] Let ${\mathbb{A}}$ as in be elliptic, $k=1$. Then for each $1{\leqslant}p<n/(n-1)$, there exists $c>0$ such that $$\begin{aligned}
\|u\|_{{\operatorname{L}}^p({\operatorname{B}},V)}{\leqslant}c\|{\mathbb{A}}u\|_{{\operatorname{L}}^1({\operatorname{B}},W)}\end{aligned}$$ for all $u\in{\operatorname{C}}^\infty_c({\operatorname{B}},V)$.
[99]{} <span style="font-variant:small-caps;">Ambrosio, L.</span> and <span style="font-variant:small-caps;">Dal Maso, G.</span>, 1992. *On the relaxation in ${\operatorname{BV}}(\Omega,{\mathbb{R}}^m)$ of quasi–convex integrals*. Journal of functional analysis, **109**(1), pp.76-97. <span style="font-variant:small-caps;">Bourgain, J.</span> and <span style="font-variant:small-caps;">Brezis, H.</span>, 2002. *Sur l’équation ${\operatorname{div}}u=f$*. Comptes Rendus Mathematique, **334**(11), pp.973-976. <span style="font-variant:small-caps;">Bourgain, J.</span> and <span style="font-variant:small-caps;">Brezis, H.</span>, 2003. *On the equation ${\operatorname{div}}Y=f$ and application to control of phases*. Journal of the American Mathematical Society, **16**(2), pp.393-426. <span style="font-variant:small-caps;">Bourgain, J.</span> and <span style="font-variant:small-caps;">Brezis, H.</span>, 2004. *New estimates for the Laplacian, the div–curl, and related Hodge systems*. Comptes Rendus Mathematique, **338**(7), pp.539-543. <span style="font-variant:small-caps;">Bourgain, J.</span> and <span style="font-variant:small-caps;">Brezis, H.</span>, 2007. *New estimates for elliptic equations and Hodge type systems*. Journal of the European Mathematical Society, **9**(2), pp.277-315. <span style="font-variant:small-caps;">Breit, D.</span>, <span style="font-variant:small-caps;">Diening, L.</span> and <span style="font-variant:small-caps;">Gmeineder, F.</span>, 2017. *Traces of functions of bounded ${\mathbb{A}}$-variation and variational problems with linear growth*. arXiv preprint arXiv:1707.06804. <span style="font-variant:small-caps;">Brezis, H.</span>, 2010. *Functional analysis, Sobolev spaces and partial differential equations*. Springer Science & Business Media. <span style="font-variant:small-caps;">Brezis, H.</span> and <span style="font-variant:small-caps;">Ponce, A.C.</span>, 2008. *Kato’s inequality up to the boundary*. Communications in Contemporary Mathematics, **10**(06), pp.1217-1241. <span style="font-variant:small-caps;">Conti, S.</span>, <span style="font-variant:small-caps;">Faraco, D.</span>, and <span style="font-variant:small-caps;">Maggi, F.</span>, 2005. *A new approach to counterexamples to ${\operatorname{L}}^1$ estimates: Korn’s inequality, geometric rigidity, and regularity for gradients of separately convex functions.* Archive for rational mechanics and analysis, **175**(2), pp.287-300. <span style="font-variant:small-caps;">DeVore, R.A.</span> and <span style="font-variant:small-caps;">Sharpley, R.C.</span>, 1993. *Besov spaces on domains in ${\mathbb{R}}^d$*. Transactions of the American Mathematical Society, **335**(2), pp.843-864. <span style="font-variant:small-caps;">Federer, H.</span>, 2014. *Geometric measure theory*. Springer. <span style="font-variant:small-caps;">Fonseca, I.</span> and <span style="font-variant:small-caps;">Müller, S.</span>, 1999. *$\mathcal{A}$–Quasiconvexity, Lower Semicontinuity, and Young Measures*. SIAM journal on mathematical analysis, **30**(6), pp.1355-1390. <span style="font-variant:small-caps;">Fuchs, M.</span>, 2011. *An estimate for the distance of a complex valued Sobolev function defined on the unit disc to the class of holomorphic functions*. Journal of Applied Analysis, **17**(1), pp.131-135. <span style="font-variant:small-caps;">Fuchs, M.</span> and <span style="font-variant:small-caps;">Seregin, G.</span>, 2000. *Variational methods for problems from plasticity theory and for generalized Newtonian fluids*. Springer Science & Business Media. <span style="font-variant:small-caps;">Giaquinta, M.</span>, <span style="font-variant:small-caps;">Modica, G.</span>, and <span style="font-variant:small-caps;">Souček, J.</span>, 1979. *Functionals with linear growth in the calculus of variations, I*. Commentationes Mathematicae Universitatis Carolinae, **20**(1), pp.143-156. <span style="font-variant:small-caps;">Gilbarg, D.</span> and <span style="font-variant:small-caps;">Trudinger, N.S.</span>, 2015. *Elliptic partial differential equations of second order.* Springer. <span style="font-variant:small-caps;">Jones, P.W.</span>, 1981. *Quasiconformal mappings and extendability of functions in Sobolev spaces*. Acta Mathematica, **147**(1), pp.71-88. <span style="font-variant:small-caps;">Ka[ł]{}amajska, A.</span>, 1994: *Pointwise multiplicative inequalities and Nirenberg type estimates in weighted Sobolev spaces*. Studia Math, **108**(3), pp.275–290. <span style="font-variant:small-caps;">Ka[ł]{}amajska, A.</span>, 1993. *Coercive inequalities on weighted Sobolev spaces*. Colloquium Mathematicae, Vol. **66**(2), pp.309-318. <span style="font-variant:small-caps;">Kirchheim, B.</span> and <span style="font-variant:small-caps;">Kristensen, J.</span>, 2016. *On rank one convex functions that are homogeneous of degree one*. Archive for Rational Mechanics and Analysis, **221**(1), pp.527-558. <span style="font-variant:small-caps;">Maz’ya, V.</span>, 2013. *Sobolev spaces*. Springer. <span style="font-variant:small-caps;">Milton, G.W.</span>, 2002. *The theory of composites*. Cambridge University Press. <span style="font-variant:small-caps;">Mironescu, P.</span> and <span style="font-variant:small-caps;">Russ, E.</span>, 2015. *Traces of weighted Sobolev spaces. Old and new*. Nonlinear Analysis: Theory, Methods & Applications, **119**, pp.354-381. <span style="font-variant:small-caps;">Ornstein, D.</span>, 1962. *A non-inequality for differential operators in the ${\operatorname{L}}^1$ norm*. Archive for Rational Mechanics and Analysis, **11**(1), pp.40-49. <span style="font-variant:small-caps;">Reshetnyak, Y.G.</span>, 1970. *Estimates for certain differential operators with finite-dimensional kernel*. Siberian Mathematical Journal, **11**(2), pp.315-326. <span style="font-variant:small-caps;">Smith, K.T.</span>, 1970. *Formulas to represent functions by their derivatives*. Mathematische Annalen, **188**(1), pp.53-77. <span style="font-variant:small-caps;">Stein, E.M.</span>, 2016. *Singular integrals and differentiability properties of functions*. Princeton university press. <span style="font-variant:small-caps;">Strang, G.</span> and <span style="font-variant:small-caps;">Temam, R.</span>, 1980. *Functions of bounded deformation*. Archive for Rational Mechanics and Analysis, **75**(1), pp.7-21. <span style="font-variant:small-caps;">Strauss, M.J.</span>, 1973. *Variations of Korn’s and Sobolev’s inequalities*. Proceedings of Symposia in Pure Mathematics **23**, D. Spencer (ed.), American Mathematical Society, pp.207-214. <span style="font-variant:small-caps;">Suquet, P.M.</span>, 1979. *Un espace fonctionnel pour les équations de la plasticité*. Annales de la Faculté des sciences de Toulouse: Mathématiques **1**(1), pp.77-87. <span style="font-variant:small-caps;">Tartar, L.</span>, 2007. *An introduction to Sobolev spaces and interpolation spaces (Vol. 3)*. Springer Science & Business Media. <span style="font-variant:small-caps;">Uspenskiĭ, S.V.</span>, 1961. *Imbedding theorems for classes with weights*. Trudy Matematicheskogo Instituta imeni VA Steklova, **60**, pp.282-303. <span style="font-variant:small-caps;">Van Schaftingen, J.</span>, 2013. *Limiting Sobolev inequalities for vector fields and canceling linear differential operators*. Journal of the European Mathematical Society, **15**(3), pp.877-921. <span style="font-variant:small-caps;">Whitney, H.</span>, 1934. *Analytic extensions of differentiable functions defined in closed sets*. Transactions of the American Mathematical Society, **36**(1), pp.63-89. <span style="font-variant:small-caps;">Zhao, R.</span> and <span style="font-variant:small-caps;">Zhu, K.</span>, 2008. *Theory of Bergman spaces in the unit ball of ${\mathbb{C}}^n$*. Mémoire de la Société mathématique de France, (115).
[^1]: <span style="font-variant:small-caps;">Authors’ Address: University of Oxford, Andrew Wiles Building, Woodstock Rd, Oxford OX2 6GG, United Kingdom</span>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new insight on NGC 6034 and UGC 842, two groups of galaxies previously reported in the literature as being fossil groups. The study is based on optical photometry and spectroscopy obtained with the CTIO Blanco telescope and Sloan Digital Sky Survey archival data. We find that NGC 6034 is embedded in a large structure, dominated by three rich clusters and other small groups. Its first and next four ranked galaxies have magnitude differences in the $r$ band and projected distances which violate the optical criteria to classify it as a fossil group. We confirm that the UGC 842 group is a fossil group, but with about half the velocity dispersion that is reported in previous works. The velocity distribution of its galaxies reveals the existence of two structures in its line of sight, one with $\sigma_v$ $\sim$ 223 kms$^{-1}$ and another with $\sigma_v$ $\sim$ 235 kms$^{-1}$, with a difference in velocity of $\sim$ 820 kms$^{-1}$. The main structure is dominated by passive galaxies, while these represent $\sim$ 60% of the second structure. The X-ray temperature for the intragroup medium of a group with such a velocity dispersion is expected to be $kT$ $\sim$ 0.5–1 keV, against the observed value of $kT$ $\sim$ 1.9 keV reported in the literature. This result makes UGC 842 a special case among fossil groups because (1) it represents more likely the interaction between two small groups, which warms the intragroup medium and/or (2) it could constitute evidence that member galaxies lost energy in the process of spiraling toward the group center, and decreased the velocity dispersion of the system. As far as we know, UGC 842 is the first low-mass fossil group studied in detail.'
author:
- 'R. Lopes de Oliveira, E. R. Carrasco, C. Mendes de Oliveira, D. R. Bortoletto, E. Cypriano, L. Sodré Jr., G. B. Lima Neto'
title: Revisiting the fossil group candidates UGC 842 and NGC 6034
---
Introduction
============
Fossil groups are galaxy systems optically dominated by an elliptical galaxy immersed in an extended and luminous X-ray halo (L$_{X, \rm bol}$ $>$ 10$^{42}$ $h^{-2}_{50}$ ergs$^{-1}$), in which the magnitude gap between the two brightest galaxies within half of the virial radius is greater than 2 in the $r$ band [@Ponman94]. The scarcity of L$^*$ galaxies and the evidence for the existence of a “massive" structure traced by the hot intragroup gas support the hypothesis of their central galaxies being formed by merging of luminous galaxies, most likely by dynamical friction [@Cypriano06; @MdO06; @MdO09].
The estimated time scales for dynamical friction and results of numerical simulations are consistent with an old age for fossil groups. For example, $N$-body/hydrodynamical simulations carried out by @DOnghia05 suggest a correlation between the magnitude gap between the two brightest galaxies and the formation time of the group. More recently several studies of simulated fossil groups were made using the Millennium simulations [@Dariush07; @Sales07; @Diaz08] which corroborated the idea that fossil groups/clusters assembled most of their masses much earlier than non-fossil systems. These studies also predicted a fraction of 3%–13% of groups with masses greater than 10$^{13}$ M$_{\odot}$ being fossil groups, depending on the exact range of masses considered.
The true nature of fossil groups has promoted a lively debate in the past decade. The absence of a conclusive explanation for the nature of these objects is mainly due to (1) the lack of a proper sample for statistical studies, (2) the lack of X-ray data with sufficient resolution and signal-to-noise (S/N) for a proper study of the group properties, and (3) the lack of optical spectroscopy of the group members, for membership confirmation and a study of the kinematic properties of the known groups. A growing number of fossil group candidates has been claimed in the recent literature. For example, @Santos07 identified 34 potential candidate fossil groups in a systematic search carried out using the [*Sloan Digital Sky Survey*]{} (SDSS). More recently, @LaBarbera09 identified 25 other fossil groups. However, the nature of these newly identified groups still need to be confirmed. We investigate here two groups pointed in the literature as being fossil groups: UGC 842 and NGC 6034 [@Voevodkin08; @Yoshioka04].
UGC 842 is a bright elliptical galaxy with a heliocentric radial velocity of 13,556$\pm$32 kms$^{-1}$ [@Huchra99]. It is a weak radio source detected by the Very Large Array (VLA) telescope with flux level of S(6 cm) $\sim$ 1.1 mJy [@Gioia83]. The UGC 842 group is immersed in an extended halo with a radius of $\sim$ 4’ in the sky, which corresponds to $\sim$ 300 kpc at its redshift. @Gastaldello07 reported an X-ray observation of UGC 842 system carried out by the [*Chandra*]{} satellite. They derived a virial radius and mass of 1272$\pm$220 $h^{-1}_{70}$kpc and 12.8$\times$10$^{13}$ $h^{-1}_{70}$$M_{\odot}$, and $r_{500}$ and $M_{500}$ of 634$\pm$85 $h^{-1}_{70}$kpc and 7.54$\pm$3.41$\times$10$^{13}$ $h^{-1}_{70}$$M_{\odot}$, respectively. About 1 year later, @Voevodkin08 reported a combined optical and X-ray analysis of this group from SDSS and [*XMM-Newton*]{} data. According to these authors, galaxies in the UGC 842 group have $\sigma_v$ $\sim$ 439 kms$^{-1}$ (from 16 galaxies inside $r$ = 509 $h^{-1}_{71}$kpc), its intragroup gas has a temperature $kT$ of 1.90$\pm$0.30keV and metallicity of 0.34$\pm$0.12Z$_{\odot}$, and displays a bolometric X-ray luminosity of $\sim$ 1.63$\times$10$^{43}$$h^{-2}_{71}$ergs$^{-1}$. From the [*XMM-Newton*]{} data, @Voevodkin08 also inferred $r_{500}$ $\sim$ 509 $h^{-1}_{71}$kpc, M$_{\rm gas,500}$ = (3.5$\pm$1.1)$\times$10$^{12}$ $h^{-1}_{71}$M$_{\odot}$, and M$_{\rm total,500}$ = (4.03$\pm$0.69)$\times$10$^{13}$ $h^{-1}_{71}$M$_{\odot}$. The UGC 842 group is classified as a fossil group by these authors.
[cccccccrrrl]{} UGC 842 & J011853.62–010007.2 & 13428$\pm$40 kms$^{-1}$ & 13.47 & -23.19\
NGC 6034 & J160332.08+171155.2 & 10162$\pm$48 kms$^{-1}$ & 13.54 & -22.49\
NGC 6034 is a bright E/S0 radio galaxy member of Abell 2151 and immersed in the Hercules supercluster [@Corwin71; @Val78], with a heliocentric radial velocity of $\sim$ 10,112 kms$^{-1}$ [@Tarenghi79]. VLA observations at $\lambda$21-cm (H[I]{}) reported by @Dickey97 show that NGC 6034 has a head-tail morphology in the continuum, with two jets, and evidence of gas falling toward the nucleus with a velocity of about 70 kms$^{-1}$ as derived from the detection of shifted narrow absorption line at 10,226$\pm$15 kms$^{-1}$. An X-ray observation of the NGC 6034 group with the [*Einstein*]{} Observatory shows that $L_{\rm X,0.5-4.5
keV}$ $<$ 1.3$\times$10$^{42}$ $h_{50}^{-2}$ ergs$^{-1}$ [@Canizares87]. ASCA observation (1999 August 24) reveals that this group has a relatively cold intragroup medium with $kT$ = 0.67$\pm$0.09 keV and abundance $Z$ = 0.08$\pm$0.05 $Z_{\odot}$, affected by a photoelectric absorption equivalent to $N_{H}$ $\sim$ 3.4$\times$10$^{20}$ cm$^{-2}$, and displaying a bolometric X-ray luminosity of about 2.8$\times$10$^{43}$ $h_{50}^{-2}$ ergs$^{-1}$ [@Fukazawa04]. Curiously, also from ASCA observation, @Yoshioka04 derived $kT$ = 1.29(+0.48/-0.36) keV, $Z$ = 0.11(+0.39/-0.10) $Z_{\odot}$, $N_{H}$ $\sim$ 3.4$\times$10$^{20}$ cm$^{-2}$, and $L_{\rm X, 0.1-2.4 keV}$ $\sim$ 7.5$\times$10$^{41}$ $h_{75}^{-2}$ ergs$^{-1}$. The main properties of the UGC 842 and NGC 6034 groups are described in Table \[tbl:sample\].
We report on new optical photometry and spectroscopy of the groups UGC 842 and NGC 6034 (hereafter UGC 842 and NGC 6034, for simplicity) and their “neighborhoods", carried out at the CTIO-Blanco 4.0 m telescope, which are combined with data from the SDSS Data Release 6 [DR6; @Adelman08]. These data are used here to perform an analysis of the structure and kinematics of these two systems and also allowed an investigation about the nature of the systems as fossil groups (confirmed in the case of UGC 842 and not confirmed for NGC 6034). When needed, we adopt a lambda cold dark matter ($\Lambda$CDM) cosmology with $H_{0}$ = 70 kms$^{-1}$Mpc$^{-1}$, $\Omega_{M}$ = 0.3, and $\Omega_{\Lambda}$ = 0.7.
Observations: Photometry and Spectroscopy
=========================================
Images in the $B$ and $R$ bands of UGC 842 and NGC 6034 were obtained at the CTIO-Blanco 4.0 m telescope on 2005 August 25. The images were taken with the mosaic camera, covering a region (after trimming of the edges) of approximately $\sim$ 38$\times$38 arcmin$^2$ (equivalent to about 2$\times$2 Mpc$^2$ and 1.5$\times$1.5 Mpc$^2$ at the redshift of each of these groups, respectively).
A total of five mosaic images were obtained for UGC 842, in the $B$ band, with exposure times of 600s each, and other five images of 360s each in the $R$ band, with a seeing of about 1 arcsec in the $R$ band and 1.1 arcsec in the $B$ band. For NGC 6034, only one image was taken for each filter, with exposures times of 200s and seeing of 1.6 arcsec in the $R$ band and 2.1 arcsec in the $B$ band.
The galaxies in the area around NGC 6034 and UGC 842 were observed spectroscopically on 2006 August 13 UT with the Hydra-CTIO multi-object spectrograph [@Barden98] at the CTIO Blanco 4 m telescope in Chile. NGC 6034 was also observed on 2007 April 14, with the same equipment. The observations were performed during dark/gray times with a good atmospheric transparency in 2007 April, and under bright sky with some cirrus in 2006 August, in both epochs with a seeing that on average varied between 0.9” and 1.2” (DIMM monitor). The observation of UGC 842 had in general poor S/N due to the proximity of the Moon in the field – we were able to extract only 20 out of a total of 84 spectra observed.
The spectra were acquired using two different setups. For the data taken on 2006 August, we used the KPGL3 grating over the wavelength range 3960–6960Å, centered on 5278Å, which provided a spectral resolution of $\sim$ 4Å, and a dispersion of 1.39Åpixel$^{-1}$. In 2007 April, we used the KPGL2 grating over the wavelength range 3450–8242Å, centered on 5845Å, which provided a spectral resolution of $\sim$ 6.5Å, and dispersion of 2.33Åpixel$^{-1}$. To avoid second order contamination above 8000Å, the blocking filter GG385 was used. All spectra were imaged with the 400 nm Bench Schmidt camera onto a SITe 2k$\times$4k CCD, with a binning of 2 pixels in the spectral direction. Total exposure times of 2 hr (4$\times$30 minutes) and 1.5 hr (3$\times$30 minutes) were used for the objects observed in 2006 August and 2007 April, respectively.
Data Reduction and Analysis
---------------------------
### Photometry
The images were processed using standard reduction techniques within IRAF[^1]. After correcting them by bias and flat field, we derived positions and magnitudes for all identified objects using Sextractor. Flux calibration was performed using magnitudes of common stars in the field with the SDSS database, given that our nights were not photometric. The probable galaxies were identified based on the [*class\_star*]{} parameter (less than 0.7). From the number counts of galaxies in the whole field we estimate that the photometry is complete down to about 21.5 mag in the $R$ band for UGC 842 and NGC 6034.
### Spectroscopy
The spectroscopic data were reduced using standard procedures in IRAF. In summary, all science exposures, comparison lamps (He/Ne/Ar), spectroscopic flats (taken before or after each object spectrum) and the so-called “milk flats”[^2] were bias/overscan subtracted and trimmed using the IRAF task CCDPROC. The “milk flats" were then combined and spectral shapes in $x$- and $y$-direction were removed using the IRAF task FIT1D. The resultant image was then filtered by using a median filter and normalized to one. The science exposures and spectroscopic flats were then divided by the processed “milk flats" in order to reduce spectral noise in the images.
The spectra were extracted with the IRAF DOHYDRA task. Dome flats were used to flat field the individual fibers, while twilight flats were used for fiber-to-fiber throughput correction. The spectra were then wavelength calibrated. The residual values in the wavelength solution for 20–30 points using a 4th or 5th-order Chebyshev polynomial typically yielded $rms$ values of $\sim$ 0.20–0.30 Å. Finally, the average sky spectrum was subtracted from each object spectrum using typically 12 sky fiber spectra per field. The final one-dimensional object spectra were then combined by their average value.
### Radial velocities {#sct:rv}
In order to measure radial velocities we first inspected the spectra to search for obvious absorption and emission features characteristic of early- and late-type galaxy populations. For galaxies with clear emission lines, the IRAF task RVIDLINE was used employing a line-by-line Gaussian fit to measure the radial velocity (this was done for only one galaxy belonging to UGC 842 and 11 galaxies in NGC 6034). The residuals of the average velocity shifts of all measurements were used to estimate the errors. The velocities for absorption-line and for emission-line systems with clear absorption lines (6 out of 11 galaxies belonging to NGC 6034), were calculated using the cross-correlation technique [@Tonry79]. The spectra were cross-correlated with high S/N templates using the task FXCOR inside IRAF. The detected narrow cross-correlation peaks were fitted by a Gaussian, with errors given by the $R$-statistic in which $\sigma_v$ = (3/8)($w$/(1 + $R$)) [@Tonry79], where $w$ is the FWHM of the correlation peak and $R$ is the ratio of the correlation peak height to the amplitude of the antisymmetric noise. This $R$ value was used as a reliability factor of the quality of the measured velocity. For $R$ $>$ 3.5, the resulting velocity was that associated to the template which produced the lowest error. For galaxies with $R$ $<$ 3.5, we looked for absorption features like Ca[II]{} and $G$ band in the spectra, and performed a line-by-line Gaussian fit using the package RVIDLINE. The resulting values were then compared with the velocities given by cross-correlation. In all cases the agreement between the two procedures was good. The measured velocities are shown in Tables \[tbl:ugcvel\] and \[tbl:ngcvel\] (see Section \[sct:veldist\]) for galaxies in the field of UGC 842 and NGC 6034, respectively.
The SDSS catalog {#sct:sdsscat}
----------------
We used the SDSS DR6 to search for galaxies in the region around NGC 6034 and UGC 842 with three objectives: (1) to find missing galaxies that were not observed spectroscopically in our observations in the CTIO; (2) to study the environment in a large area around the groups and to minimize the contamination produced by other nearby structures in our analysis (in Section \[sct:envngc\]); and (3) to find any systematic differences that may exist between SDSS and our data.
From the SDSS DR6 archival we obtain magnitudes and spectroscopic redshifts of all objects classified as galaxies (class 3 in the SDSS) with $r$’-magnitude $<$ 21 in a 3$\times$3 deg$^{2}$ area around NGC 6034 and in a 2.3$\times$1.7 deg$^{2}$ area around UGC 842. The SDSS DR6 catalog in the field of UGC 842 ends at declination $>$ -1.25 deg (J2000), and for this reason the analysis performed on SDSS data in this work does not reflect a symmetric field around UGC 842. However, our results are not affected by such limitation. The galaxy counts calculated using the objects classified as galaxies reach their maximum at $r$’ $\sim$ 21 mag. Using this estimate value for the completeness limit and the usual uncertainties in the galaxy classification above $r$’ $\sim$ 20 mag, we have adopted this latter value as a conservative upper limit for the magnitude. This catalog was then used for the subsequent analysis.
Analysis and Results
====================
The environment around NGC 6034 and UGC 842 {#sct:envngc}
-------------------------------------------
We used the entire SDSS DR 6 galaxy catalog described in Section \[sct:sdsscat\] to inspect the environment around UGC 842 and NGC 6034. In addition, we used the NASA/IPAC Extragalactic Database (NED) to search for known groups and clusters of galaxies that could be present in the regions of UGC 842 and NGC 6034.
The positions of all SDSS objects photometrically classified as galaxies in the fields of UGC 842 and NGC 6034 were used to plot the density of objects ($\rho$) as a function of distance to the brightest galaxy in each group. These are shown in Figure \[fig:dens\]. As can be noticed, over densities of galaxies are present within circles of projected radii of about 400 $h^{-1}_{70}$kpc from the centers of UGC 842 and NGC 6034 (7.4 and 9.8 arcmin, respectively).
### UGC 842
Figure \[fig:densmap\_UGC\] shows the projected galaxy-density map for the field of UGC 842. It is clear that the UGC 842 group is well isolated (at least to the north), with no major or massive galaxy clusters at the same redshift. There are two galaxy clusters that lie closer in redshift but far in projected distance to the group: Abell 181 at $z$ = 0.072 and Abell 168 at $z$ = 0.045. Both clusters are located more than 2 deg ($\sim$ 6.4 $h^{-1}_{70}$Mpc at $z$ = 0.045) north from UGC 842. Other major structures are also indicated in the figure, but they all are background clusters. Figure \[fig:densmap\_UGC\] (right panel) shows a zoom of $\sim 80 \times 80$ arcmin$^2$ around the UGC 842 group. The meaning of the symbols are the same as in the previous (left panel) figure except that we divided the sample of member galaxies in passive (squares) and blue, star forming (triangles). The selection of the two samples was performed using the SDSS Database by inspecting visually each spectrum.
{width="80mm"} {width="80mm"}
{width="80mm"} {width="80mm"}
### NGC 6034
It became clear from our analysis that NGC 6034 is located in a much richer neighborhood than UGC 842. Figure \[fig:densmap\_NGC\] (on the left panel) shows the projected galaxy-density map of all galaxies with $r$’$<$ 20 mag inside an area of 3$\times$3 deg$^2$, centered around NGC 6034. In this figure one can see that the group is embedded in a region dominated by three major, rich, galaxy clusters: Abell 2151 (Hercules), Abell 2152, and NSC J160216+162033 (big filled circles in the figure). Several galaxy groups are also present in the area: WBL 601, WBL 603, WBL 606, WBL 607, and USGC U741 (big filled diamonds in the figure). Table \[tbl:aroundngc\] summarizes the main parameters of the groups and clusters in the region. Note also that the structure named NGC 6034 group is not listed in NED and its position corresponds to the position of the group WBL 604.
[llccrl]{} Abell 2151 & Cluster & 16:05:25.9 & $+$17:47:50 & 11070 & @Bird95\
Abell 2151E & Cluster & 16:06:51.9 & $+$17:46:51 & 9623 & @Ebeling98\
Abell 2151C & Cluster & 16:05:15.5 & $+$17:39:45 & 10650 & @Bird95\
Abell 2151N & Cluster & 16:05:55.0 & $+$18:08:27 & 11445 & @Bird95\
Abell 2152 & Cluster & 16:05:22.4 & $+$16:26:55 & 12291 & @Struble99\
NSC J160216+162033 & Cluster & 16:02:16.9 & $+$16:20:32 & 11422 & @Gal03\
WBL 601 & Group & 16:01:40.7 & $+$16:20:41 & 10637 & @White99\
WBL 603 & Group & 16:03:34.6 & $+$16:20:48 & 11212 & @White99\
WBL 606 & Group & 16:04:40.4 & $+$16:32:52 & 9539 & @White99\
WBL 607 & Group & 16:05:26.8 & $+$17:53:55 & 10936 & @White99\
UCG U741 & Group & 16:03:58.8 & $+$16:45:33 & 10821 & @Ramella00\
In order to investigate how many galaxy groups belong to the large structure in which NGC 6034 seems to be embedded we show in Figure \[fig:densmap\_NGC\] (on the right side) a zoom of the 40$\times$40 arcmin$^2$ field around this group. From the figure it is clear that a number of galaxies with velocities in the interval 9,000 and 12,000 kms$^{-1}$ may in fact belong to other nearby structures. The (red) dashed contours in the figure represent the projected density of galaxies at the distance of the group. We can see that the peak of the distribution is offset by 3’–4’ from the high density peak given by the projected density map obtained using all galaxies with $r$’ $<$ 20 mag. The reason of this displacement is due to the existence of two massive structures along the line of sight at $z$ $\sim$ 0.11 and $z$ $\sim$ 0.14 (see Figure \[fig:veldist\], left).
Color-Magnitude Diagrams \[sct:cmd\]
------------------------------------
A red-sequence of galaxies is clearly seen in the color-magnitude diagram (CMD) of both groups (Figure \[fig:cmd\]). Group membership is confirmed by spectroscopy for all galaxies with $r$ $<$ 17.7 mag inside 500 $h_{70}^{-1}$kpc around UGC 842 and NGC 6034, with two exceptions for NGC 6034.
The UGC 842 group obeys the optical criteria to be classified as a fossil group, in which the brightest galaxy dominates the group within R$_{vir}$/2. The two closest “bright” galaxies are yet relatively far from UGC 842 (J011913.49-010839.9: $d_{proj}$ $\sim$ 528 $h_{70}^{-1}$kpc, $r$ = 14.58 mag; and J011832.25-011150.4: $d_{proj}$ $\sim$ 688 $h_{70}^{-1}$kpc, $r$ = 14.66 mag). The same is not true for the NGC 6034 group. Six galaxies in the field of NGC 6034, shown in Figure \[fig:cmd\], violate the magnitude criteria adopted to classify a group as a fossil group within R$_{vir}$/2 (see parameters in Table \[tab:vel\]). One of them, J160348.24+171426.2, has a $\Delta r$ $\sim$ 0.9 mag, $d_{proj}$ $\sim$ 188 $h_{70}^{-1}$kpc, and $\Delta$$V$ $\sim$ 800 kms$^{-1}$, with respect to the brightest galaxy. It is not completely clear if this galaxy, with such a relatively high $\Delta$$V$ is or is not a group member – note that $\sigma_v$ for this group is $\sim$ 230 kms$^{-1}$. The other five identified members which also break the fossil group definition have 300 $h_{70}^{-1}$kpc $<$ $d_{proj}$ $<$ 460 $h_{70}^{-1}$kpc, and $\Delta$V $<$ 240 kms$^{-1}$. They are: J160402.75+171656.6, J160419.55+171049.4, J160249.22+171002.6, J160314.11+172202.4, and J160356.65+171818.4.
SDSS DR6 Versus CTIO Observations {#sct:sdsshydra}
---------------------------------
From the initial SDSS catalog described in Section \[sct:sdsscat\], we concentrated on all galaxies inside a region of radius $<$ 25’ around UGC 842 and NGC 6034 (the region covered by our Hydra-CTIO observations) and with velocities between 9,000 kms$^{-1}$ and 60,000 kms$^{-1}$. In this velocity interval, we have 55 and 9 galaxies that were observed in common with SDSS DR6 and our observations in the area of UGC 842 and NGC 6034, respectively. We used this information to search for any possible systematic effects and to estimate the true errors in our measurements. Figure \[fig:resvel\] shows the residuals of the velocities (our velocities, V$_{H}$, minus velocities from the SDSS catalog, V$_S$) as a function of the quadratic errors ($\sigma_H$ and $\sigma$$_S$) for the 64 galaxies in common with the SDSS DR6 data set. The average difference between the two data sets is 35 kms$^{-1}$ with and $rms$ of 63 kms$^{-1}$. Thus, in order to minimize the errors in the determination of the average velocity and the velocity dispersion, we applied this small velocity correction to all galaxies in the SDSS DR6 catalog.
![ Residual of the heliocentric velocity as a function of the internal quadratic errors for 64 galaxies (NGC 6034 and UGC 842) in common with SDSS DR6. The mean shift of the data is 35 kms$^{-1}$ with a [*rms*]{} of 63 kms$^{-1}$. []{data-label="fig:resvel"}](fig5.eps){width="75mm"}
For NGC 6034, we selected all galaxies with radial velocities in the interval 9,000–12,000 kms$^{-1}$. A total of 35 galaxies are in common with the SDSS DR6. For five of the galaxies observed with Hydra-CTIO, the velocities were obtained for the first time. In addition, there are six galaxies listed in the SDSS DR6 and inside the above interval that were not observed by us. In total, we have 46 galaxies in the area of NGC 6034 and within the inverval of radial velocities of 9,000 to 12,000 kms$^{-1}$.
For UGC 842, we selected all galaxies with radial velocities in the interval 10,000–17,000 kms$^{-1}$. A total of nine galaxies are in common with the SDSS DR6. For three of the galaxies observed with Hydra-CTIO, the velocities were obtained for the first time.
Velocity distributions and mass determinations {#sct:veldist}
----------------------------------------------
The new galaxy radial velocities in the fields of UGC 842 and NGC 6034 are listed in Tables \[tbl:ugcvel\] and \[tbl:ngcvel\], respectively. The radial velocities and the associated errors derived using cross-correlation (absorption-line systems) are given in the column named V$^{\rm CC}_{\rm hel}$. The radial velocities from emission lines are given in the column named V$^{\rm EM}_{\rm hel}$. All radial velocities are corrected to the heliocentric reference frame. The galaxy identifications and their coordinates, total magnitudes in the $g$’ band and $r$’ band, $R$ values, and the number of emission lines ($\#$el) used when it is possible for velocity calculation (see Section \[sct:rv\]) are also given in the tables. We used the robust bi-weight estimators CBI and SBI of @Beers90 to calculate a reliable value for the average velocity and the velocity dispersion for each group. The adopted radial velocities in Tables \[tbl:ugcvel\] and \[tbl:ngcvel\] for the determination of the average velocity and velocity dispersion of the groups are always the velocities calculated by cross-correlation, except in the cases when the S/N of the absorption line spectra was too low (in one case for the UGC 842 group and five cases for the NGC 6034 group). We use the relation of @Carlberg97 in order to estimate the virial radius of the systems: $r_{vir}$ = $\sqrt{3}$/10$\times$ $\sigma_v$/H($z$), where H($z$)=H$_0 \times$\[$\Omega_M$(1+$z$)$^3$+(1-$\Omega_M$-$\Omega_{\Lambda}$)(1+$z$)$^{2}$+$\Omega_{\Lambda}$\]$^{1/2}$. Then the mass was estimated by $M_{virial}$ = (2/G)$\sigma_v^2$$r_{virial}$.
[rrcccccrrrr]{}
248 & 66 & J011726.91-004700.1 & 18.81 & 18.41 & 18.08 & 19.10 & 17339$\pm$69 & 2.42 & &\
1309 & 21 & J011743.74-004953.0 & 16.00 & 15.18 & 14.93 & 16.48 & 15930$\pm$35 & 10.07 & &\
3134 & 5 & J011745.47-010606.2 & 19.02 & 18.54 & 18.24 & 19.38 & 44194$\pm$84 & 2.58 & &\
3079 & 79 & J011745.62-010641.5 & 19.66 & 19.04 & 18.77 & 20.07 & 6928$\pm$113& 2.62 & &\
3755 & 114 & J011750.18-010727.2 & 19.60 & 18.78 & 18.88 & 20.35 & 51098$\pm$82 & 4.14 & &\
4780 & 54 & J011758.75-010112.8 & 18.76 & 18.47 & 18.20 & 19.03 & 14190$\pm$101& 3.06 & &\
5256 & 11 & J011803.03-010258.1 & 17.12 & 16.21 & 16.03 & 17.68 & 13192$\pm$67 & 5.24 & &\
7155 & 52 & J011814.40-005256.3 & 16.85 & 16.43 & 16.13 & 17.20 & & &14781$\pm$85& 6\
8131 & 67 & J011818.87-003619.6 & 19.53 & 19.05 &&& 26548$\pm$60 & 4.93 & &\
9106 & 12 & J011827.75-010131.4 & 17.83 & 17.59 & 17.24 & 18.18 & & & 5868$\pm$72& 6\
7894 & 108 & J011832.25-011150.4 & 15.61 & 14.66 & 14.39 & 16.07 & 13873$\pm$89 & 4.20 & &\
1002 & 43 & J011837.30-005924.2 & 16.70 & 15.85 & 15.60 & 17.19 & 13987$\pm$66 & 5.56 & &\
1126 & 15 & J011840.90-004021.7 & 17.44 & 16.62 & 16.37 & 17.92 & 14627$\pm$59 & 6.76 & &\
1187 & 18 & J011852.99-003935.3 & 16.08 & 15.20 & 14.94 & 16.58 & 14311$\pm$33 & 2.07 & &\
8901 & 39 & J011853.62-010007.2 & 14.35 & 13.47 & 13.21 & 14.85 & 13428$\pm$40 & 10.09 & &\
13007 & 37 & J011854.47-004059.5 & 15.93 & 15.04 & 14.78 & 16.44 & 13511$\pm$35 & 11.30 & &\
1597 & 93 & J011913.43-010503.4 & 17.56 & 16.70 & 16.47 & 18.14 & 13771$\pm$58 & 4.94 & &\
1480 & 124 & J011913.49-010839.9 & 15.55 & 14.58 & 14.30 & 16.08 & 13789$\pm$44 & 8.70 & &\
1662 & 31 & J011916.05-011040.2 & 19.17 & 18.81 & 18.49 & 19.53 & 8358$\pm$77 & 3.89 & &\
1821 & 128 & J011932.22-010607.9 & 18.16 & 17.39 & 16.94 & 18.48 & 61195$\pm$42 & 4.34 & &\
[rrccccccrrrr]{}
1 & 87 & 1 & J160152.09+171139.8 & 17.83 & 16.96 & ... & ... & 10499$\pm$33 & 10.68 & &\
333 & 63 & 1 & J160204.26+170433.0 & 14.47 & 13.60 & ... & ... & 10957$\pm$16 & 14.68 & &\
780 & 129 & 1 & J160206.23+171345.4 & 18.08 & 16.99 & ... & ... & 42264$\pm$66 & 8.42 & &\
857 & 137 & 1 & J160206.90+171819.8 & 18.80 & 18.38 & ... & ... & 49763$\pm$61 & 3.98 & &\
885 & 72 & 1 & J160208.02+172708.4 & 18.62 & 17.50 & 17.35 & 19.29 & 42339$\pm$38 & 9.10 & &\
1042 & 104 & 1 & J160210.90+170925.4 & 19.45 & 18.00 & 16.82 & 19.14 & 69098$\pm$72 & 5.56 & &\
1076 & 130 & 1 & J160210.98+172707.1 & 18.00 & 17.25 & 17.05 & 18.48 & 32522$\pm$43 & 8.26 & &\
1278 & 49 & 2 & J160214.96+171557.4 & 18.49 & 17.88 & 17.94 & 19.22 & && 42010$\pm$23 & 5\
1363 & 48 & 1 & J160218.39+171421.3 & 17.84 & 16.72 & 16.54 & 18.56 & 42384$\pm$48 & 9.12 & &\
1473 & 115 & 1 & J160219.00+172938.4 & 20.51 & 18.78 & 17.65 & 20.21 & 92751$\pm$87 & 7.40 & &\
1438 & 132 & 1 & J160219.24+171134.7 & 19.25 & 18.13 & 17.91 & 20.01 & 42523$\pm$53 & 7.60 & &\
1426 & 19 & 1 & J160221.89+170949.3 & 18.65 & 17.52 & ... & ... & 42061$\pm$56 & 8.23 & &\
1830 & 81 & 2 & J160227.29+170845.9 & 18.27 & 17.17 & 16.85 & 18.80 & 42394$\pm$68 & 6.64 & &\
1873 & 8 & 2 & J160227.48+171313.6 & 18.35 & 17.51 & ... & ... & 42429$\pm$77 & 2.80 & &\
1888 & 111 & 1 & J160227.52+172232.3 & 18.78 & 18.06 & 17.80 & 19.23 & 41834$\pm$48 & 5.12 & &\
1908 & 23 & 1 & J160229.59+172700.3 & 19.11 & 18.35 & 18.06 & 19.51 & 11034$\pm$53 & 5.81 & &\
1994 & 6 & 1 & J160230.54+170035.6 & 17.82 & 17.06 & 16.81 & 18.24 & 9842$\pm$31 & 9.58 & &\
& 114 & 2 & &&&&& 9808$\pm$41 & 8.98 & 9922$\pm$49 & 10\
2144 & 82 & 1 & J160233.45+170754.4 & 18.57 & 18.11 & 17.82 & 18.93 & &&40973$\pm$12 & 5\
2265 & 100 & 1 & J160235.55+170136.9 & 19.02 & 18.24 & 17.97 & 19.48 & 36300$\pm$96 & 2.40 & &\
2588 & 57 & 1 & J160242.80+170614.3 & 19.19 & 17.92 & 17.71 & 19.98 & 52668$\pm$40 & 7.59 & &\
2677 & 92 & 1 & J160244.90+172159.4 & 18.62 & 18.17 & 17.97 & 19.07 & 18319$\pm$80 & 3.44 & &\
2580 & 66 & 1 & J160245.04+171815.4 & 16.45 & 15.82 & 15.58 & 16.83 & 10678$\pm$51 & 6.90 & &\
2720 & 12 & 1 & J160246.95+165916.5 & 19.30 & 18.30 & 17.85 & 19.67 & 50271$\pm$65 & 5.97 & &\
2746 & 22 & 2 & J160247.36+171056.6 & 17.52 & 16.71 & 16.77 & 18.30 & 10282$\pm$24 & 12.84 & &\
2716 & 8 & 1 & J160247.54+171454.9 & 17.16 & 16.41 & ... & ... & 13051$\pm$48 & 7.01 & 13160$\pm$63 & 5\
2801 & 54 & 1 & J160249.11+171120.4 & 18.21 & 17.17 & 16.87 & 18.65 & 32196$\pm$46 & 6.03 & &\
2691 & 79 & 1 & J160249.22+171002.6 & 15.48 & 14.63 & 14.41 & 16.01 & 10356$\pm$27 & 19.49 & &\
2858 & 103 & 1 & J160249.29+172228.4 & 20.89 & 20.53 & 17.62 & 18.61 & 15987$\pm$100& 3.53 & &\
3033 & 61 & 1 & J160254.86+173106.4 & 17.74 & 16.71 & 16.59 & 18.47 & 30006$\pm$23 & 13.37 & &\
3158 & 116 & 1 & J160255.92+170348.3 & 18.08 & 17.65 & 17.35 & 18.42 & && 13810$\pm$10 & 6\
3344 & 69 & 1 & J160301.29+171609.5 & 17.98 & 17.32 & 17.10 & 18.46 & 33203$\pm$77 & 3.20 & &\
3524 & 46 & 1 & J160303.39+171429.1 & 18.73 & 17.72 & 17.47 & 19.24 & 40356$\pm$34 & 8.86 & &\
& 85 & 2 & &&&&& 40335$\pm$88 & 3.99 & &\
3615 & 25 & 1 & J160305.24+171136.1 & 17.67 & 16.88 & 16.74 & 18.32 & 9952$\pm$38 & 10.62 & &\
3379 & 86 & 1 & J160305.72+171020.3 & 17.11 & 16.59 & 16.26 & 17.36 & && 10050$\pm$40 & 7\
3651 & 84 & 1 & J160306.68+173136.4 & 17.61 & 16.95 & 16.88 & 18.05 & 29801$\pm$23 & 10.56 & &\
3686 & 33 & 1 & J160307.22+172749.4 & 17.36 & 16.88 & 16.61 & 17.69 & 10776$\pm$51 & 4.35 & 10835$\pm$36 & 9\
3580 & 67 & 1 & J160307.58+172412.9 & 18.39 & 17.87 & 17.69 & 18.77 & 33031$\pm$82 & 3.41 & 32942$\pm$40 & 6\
3910 & 68 & 1 & J160307.88+171719.9 & 18.67 & 17.65 & 17.35 & 19.07 & 40339$\pm$57 & 5.36 & &\
3998 & 16 & 1 & J160310.23+172819.0 & 18.25 & 17.58 & 17.38 & 18.68 & 33765$\pm$38 & 5.99 & &\
4168 & 15 & 1 & J160314.11+172202.4 & 15.66 & 14.76 & 14.49 & 16.18 & 9943$\pm$41 & 12.47 & &\
4221 & 90 & 1 & J160314.95+171421.3 & 18.20 & 17.45 & 17.05 & 18.51 & 10487$\pm$37 & 7.84 & &\
4411 & 108 & 1 & J160317.65+171101.7 & 17.68 & 16.83 & 16.76 & 18.32 & 10435$\pm$53 & 5.77 & &\
4527 & 99 & 1 & J160322.65+170326.0 & 17.10 & 16.34 & ... & ... & 11783$\pm$25 & 12.06 & &\
& 45 & 2 & &&&&& 11805$\pm$48 & 11.39 & 11717$\pm$52 & 10\
4729 & 42 & 1 & J160324.62+171648.8 & 17.92 & 16.74 & 16.40 & 18.40 & 40197$\pm$30 & 13.00 & &\
4930 & 59 & 1 & J160327.88+171628.7 & 20.16 & 19.06 & 17.57 & 19.40 & 40581$\pm$95 & 3.60 & &\
4920 & 124 & 1 & J160328.00+171146.2 & 17.84 & 16.93 & 16.76 & 18.38 & 10314$\pm$34 & 12.36 & &\
& 29 & 2 & &&&&& 10339$\pm$35 & 11.63 & &\
4913 & 91 & 2 & J160328.00+171619.8 & 18.03 & 16.91 & 16.89 & 18.83 & 40418$\pm$79 & 5.79 & &\
4749 & 91 & 1 & J160328.02+172559.4 & 16.09 & 15.24 & 15.03 & 16.65 & 10568$\pm$22 & 24.08 & &\
4989 & 31 & 1 & J160328.93+170930.2 & 18.40 & 17.62 & 17.58 & 19.06 & 10563$\pm$32 & 8.74 & &\
& 14 & 2 & &&&&& 10638$\pm$74 & 4.19 & &\
5099 & 93 & 2 & J160330.86+171056.4 & 18.05 & 17.24 & 16.96 & 18.41 & 10003$\pm$32 & 12.22 & &\
5065 & 20 & 1 & J160330.94+172226.0 & 17.75 & 16.94 & 16.58 & 18.11 & 10307$\pm$28 & 9.30 & &\
4090 & 29 & 1 & J160332.08+171155.2 & 14.44 & 13.54 & 13.08 & 14.69 & 10162$\pm$48 & 10.75 & &\
5204 & 18 & 1 & J160332.41+173051.9 & 18.53 & 17.65 &&& 29715$\pm$21 & 18.43 & &\
4972 & 37 & 1 & J160332.58+172845.9 & 15.70 & 15.01 & 14.78 & 16.06 & 13197$\pm$22 & 14.83 & &\
5232 & 9 & 2 & J160333.18+170947.1 & 17.80 & 17.01 & 16.95 & 18.43 & 9985$\pm$31 & 12.02 & &\
4846 & 64 & 1 & J160333.42+171421.6 & 17.35 & 16.22 & 15.45 & 17.47 & 40951$\pm$45 & 10.05 & &\
5091 & 135 & 1 & J160334.27+165725.7 & 16.26 & 15.48 & 15.06 & 16.49 & 10064$\pm$31 & 13.64 & &\
5269 & 117 & 1 & J160334.38+172817.5 & 17.76 & 17.46 & 16.98 & 17.95 & && 11384$\pm$71 & 6\
5306 & 34 & 2 & J160334.45+172936.3 & 19.57 & 18.11 & 17.74 & 20.15 & 67955$\pm$80 & 5.33 & &\
5348 & 9 & 1 & J160335.72+165841.7 & 17.87 & 17.22 & 16.86 & 18.41 & 14316$\pm$51 & 5.88 & 14376$\pm$49 & 9\
5493 & 18 & 2 & J160337.57+172527.3 & 19.53 & 18.25 & 18.06 & 20.07 & 68403$\pm$63 & 5.47 & &\
5522 & 126 & 1 & J160339.26+171105.4 & 16.85 & 15.97 & 15.77 & 17.42 & 10194$\pm$50 & 8.97 & &\
5446 & 110 & 1 & J160340.50+172016.6 & 16.50 & 15.93 & 15.56 & 16.59 & 12748$\pm$44 & 6.34 & &\
& 47 & 2 & &&&&& 12784$\pm$56 & 5.05 & &\
5543 & 39 & 1 & J160340.81+170409.8 & 18.14 & 17.00 & 16.64 & 18.71 & 41139$\pm$47 & 9.36 & &\
5688 & 56 & 1 & J160342.84+171812.9 & 18.51 & 17.49 & 17.10 & 18.96 & 29005$\pm$18 & 15.66 & &\
5770 & 35 & 2 & J160343.72+171450.8 & 19.50 & 18.59 & 18.07 & 19.80 & 42185$\pm$54 & 6.66 & &\
5801 & 106 & 1 & J160344.27+171750.8 & 18.37 & 17.60 & 17.55 & 19.10 & 10694$\pm$42 & 5.93 & &\
5827 & 106 & 2 & J160344.58+171719.5 & 18.06 & 17.64 & 17.49 & 18.45 & && 29142$\pm$18 & 6\
5833 & 95 & 1 & J160344.60+172826.4 & 18.27 & 17.75 & 17.44 & 18.64 & && 11099$\pm$12 & 7\
5785 & 17 & 1 & J160344.88+170808.1 & 18.21 & 17.60 & 16.53 & 18.10 & && 20677$\pm$50 & 5\
5855 & 75 & 1 & J160345.04+171347.0 & 17.94 & 17.09 & 16.83 & 18.45 & 9923$\pm$26 & 16.41 & &\
5651 & 120 & 1 & J160348.24+171426.2 & 15.31 & 14.47 & 14.04 & 15.63 & 10974$\pm$31 & 12.08 & &\
6110 & 83 & 1 & J160350.67+171529.8 & 19.46 & 17.97 & 16.96 & 19.74 & 69091$\pm$48 & 9.80 & &\
6065 & 126 & 2 & J160350.71+171421.8 & 17.65 & 16.90 & 15.97 & 17.64 & 12767$\pm$41 & 9.87 & 12855$\pm$58 & 10\
6296 & 73 & 1 & J160354.42+170308.2 & 17.49 & 16.55 & 16.39 & 18.08 & 33321$\pm$33 & 9.11 & &\
6331 & 93 & 1 & J160354.55+170435.2 & 18.11 & 17.36 & 17.44 & 18.91 & 10430$\pm$43 & 6.35 & &\
6418 & 128 & 1 & J160356.30+170639.3 & 18.46 & 17.70 & 17.48 & 18.96 & 10366$\pm$43 & 8.18 & &\
6313 & 122 & 1 & J160356.65+171818.4 & 16.16 & 15.33 & 15.10 & 16.68 & 10004$\pm$45 & 9.80 & &\
6472 & 138 & 2 & J160357.52+171419.5 & 18.80 & 17.72 & 17.53 & 19.33 & 41861$\pm$54 & 6.97 & &\
6563 & 78 & 1 & J160359.24+173209.4 & 18.46 & 17.43 & 17.13 & 18.96 & 33584$\pm$32 & 13.42 & &\
6589 & 102 & 1 & J160359.79+172952.2 & 18.25 & 17.18 & 17.00 & 18.97 & 33642$\pm$45 & 12.30 & &\
6569 & 47 & 1 & J160400.07+173104.0 & 17.71 & 16.64 & 16.33 & 18.30 & 33693$\pm$20 & 18.02 & &\
6580 & 95 & 2 & J160400.17+173115.6 & 18.85 & 18.00 & 17.38 & 19.27 & 33534$\pm$47 & 9.06 & &\
6501 & 20 & 2 & J160400.31+173317.5 & 17.25 & 16.24 & 16.15 & 17.85 & 33757$\pm$41 & 10.55 & &\
6678 & 78 & 2 & J160401.12+172839.4 & 18.67 & 17.92 & 17.72 & 19.18 & 34908$\pm$81 & 4.57 & &\
6692 & 28 & 1 & J160401.42+170820.4 & 18.55 & 17.91 & 17.68 & 18.95 & 29743$\pm$50 & 4.43 & 29776$\pm$46 & 5\
6512 & 65 & 1 & J160402.75+171656.6 & 15.21 & 14.33 & 14.14 & 15.77 & 9956$\pm$41 & 13.41 & &\
7347 & 36 & 1 & J160405.47+171314.7 & 19.36 & 17.70 & 17.61 & 20.02 & 88536$\pm$73 & 5.37 & &\
7055 & 131 & 2 & J160407.17+171410.6 & 17.39 & 16.84 & 16.69 & 17.72 & 83790$\pm$118& 2.02 & &\
7244 & 96 & 1 & J160407.55+171224.8 & 18.29 & 17.90 & 17.59 & 18.59 & && 10546$\pm$41 & 7\
7100 & 58 & 1 & J160408.21+173051.1 & 18.58 & 17.75 & 17.61 & 19.11 & 11006$\pm$37 & 9.83 & &\
8528 & 40 & 1 & J160418.17+171627.8 & 18.90 & 18.17 &&& 10368$\pm$65 & 4.06 & &\
8413 & 51 & 1 & J160420.36+172611.2 & 16.67 & 15.90 & 15.65 & 17.07 & 10628$\pm$41 & 8.49 & 10787$\pm$51 & 10\
& 110 & 2 & &&&&& 10805$\pm$68 & 5.51 & 10753$\pm$50 & 10\
8144 & 134 & 1 & J160425.95+172543.1 & 18.86 & 17.85 & 17.60 & 19.43 & 32435$\pm$26 & 13.33 & &\
8143 & 17 & 2 & J160427.24+170203.4 & 18.68 & 17.68 & 17.73 & 19.45 & 41147$\pm$83 & 4.41 & &\
7584 & 71 & 1 & J160427.99+172148.0 & 17.89 & 16.95 & 16.84 & 18.54 & 32229$\pm$28 & 10.89 & &\
8264 & 2 & 2 & J160428.03+171317.1 & 19.20 & 18.38 & 17.74 & 19.08 & 41823$\pm$64 & 2.86 & &\
8134 & 2 & 1 & J160428.44+171319.8 & 17.96 & 17.50 & 17.23 & 18.31 & && 10301$\pm$39 & 6\
7994 & 13 & 1 & J160429.13+172116.8 & 19.11 & 18.32 & 17.58 & 18.96 & && 53349$\pm$24 & 6\
7895 & 59 & 2 & J160430.42+172809.0 & 15.52 & 14.92 & 14.95 & 16.04 & 11923$\pm$53 & 7.39 & 11831$\pm$27 & 10\
7600 & 51 & 2 & J160434.80+172638.0 & 18.97 & 18.21 & 18.02 & 19.46 & 9692$\pm$63 & 6.94 & &\
6780 & 133 & 1 & J160436.01+171739.9 & 18.79 & 17.80 &&& 32297$\pm$34 & 10.73 & &\
7506 & 44 & 1 & J160437.44+171825.0 & 17.51 & 16.59 &&& 32196$\pm$17 & 19.33 & &\
5682 & 24 & 1 & J160354.19+173242.7 &&&&& 15506$\pm$61 & 5.08 & &\
3474 & 63 & 2 & J160304.42+171127.6 &&& 16.88 & 18.10 & 10871$\pm$75 & 3.00 & 10956$\pm$45 & 9\
{width="75mm"}
[lccccccc]{} UGC 842 (all) & 35 & 13,808$\pm$81 & 471$\pm$37 & ... & ...\
UGC 842 (early-type) & 28 & 13,682$\pm$80 & 419$\pm$51 & ... & ...\
UGC 842-S1 (all) & 20 & 13,467$\pm$51 & 223$\pm$29 & 509 & 1.1\
UGC 842-S1 (early-type) & 19 & 13,458$\pm$53 & 226$\pm$30 & 515 & 1.1\
UGC 842-S2 (all) & 15 & 14,283$\pm$63 & 235$\pm$35 & ... & ...\
UGC 842-S2 (early-type) & 9 & 14,204$\pm$53 & 231$\pm$67 & ... & ...\
NGC 6034 (all) & 44 & 10,444$\pm$62 & 410$\pm$39 & 954 & 7.0\
NGC 6034 (non-emission) & 31 & 10,347$\pm$66 & 361$\pm$40 & 840 & 4.7\
### UGC 842
The velocity distribution of all known galaxies with velocities between 10,000 kms$^{-1}$ and 17,000 kms$^{-1}$ and within a region of $\sim$80$\times$50 arcmin${^2}$ is shown in Figure \[fig:veldist\_ugc\] (left). The right histogram in Figure \[fig:veldist\_ugc\] shows the velocity distribution of the member galaxies of UGC 842 (with projected distances from the central galaxy within $d_{proj}$ $<$ 1 h$^{-1}_{70}$Mpc, or $\sim$ 20 arcmin). It suggests that UGC 842 has at least a bi modal distribution. In order to investigate its structure, we use the KMM-test [@Ashman94], which is appropriate to detect the presence of two or more components in an observational data set. First we consider whether the data are consistent with a single component. The results of applying the test in the homoscedastic mode (common covariance) yields strong evidence that the velocity distribution of galaxies in the velocity interval above is at least bimodal, rejecting a single Gaussian model at a confidence level of 99.5% ($P$-value of 0.005). The $P$-value is another way to express the statistical significance of the test, and is the probability that a likelihood test statistic would be at least as large as the observed value if the null hypothesis (one component in this case) were true. Assuming two components, the KMM-test estimated a mean value for each component of 13,477 kms$^{-1}$ and 14,331 kms$^{-1}$. The two components correspond to the structures S1 and S2 in Figure \[fig:veldist\_ugc\] (right). The average velocities and the velocity dispersions of S1 and S2 calculated using the robust bi-weight estimator are listed in Table \[tab:vel\]. The two structures are separated by $\sim$ 820 km s$^{-1}$ in the group rest frame. The first structure includes UGC 842 itself, which has a radial velocity of $\sim$ 13,428 kms$^{-1}$, close to the peak of its distribution. In total, 20 and 15 galaxies are members of the structures S1 and S2, respectively. It is interesting to note that the first structure is mainly formed by passive galaxies (19 out of 20 galaxies) while for the second one, the number of passive galaxies is $\sim$ 60% (9 galaxies) of the total population. Similar double peaked galaxy distribution is also found when only passive galaxies are considered (the shaded region in Figure \[fig:veldist\_ugc\], right), which shows that the result is not biased by emission-line objects. The groups S1 and S2 seem to occupy the same projected area although S1 seems more centrally concentrated. Although the interaction between two small groups each with $\sigma_v$ $\sim$ 230 kms$^{-1}$ cannot be conclusively discarded, a radial velocity difference of $\sim$ 820 kms$^{-1}$ between the peaks of the distributions S1 and S2 is an evidence that the observed bimodality is most probably due to a superposition of structures in the line of sight. If this is true, and assuming a condition of equilibrium, the group S1 – the UGC 842 group – has a virial mass and radius of 1.1$\times$10$^{13}$ $h^{-1}_{70}$M$_{\odot}$ and 509 $h^{-1}_{70}$kpc, respectively. The high concentration of late-type galaxies in the group S2 strongly suggests that it is not in equilibrium.
### NGC 6034
Figure \[fig:veldist\] (right) shows the velocity distribution of all known member galaxies (with measured velocities) of the NGC 6034 group. The distribution is well represented by a Gaussian, although there is a high velocity tail. The best estimates for its parameters are V$_{avg}$ = 10,444$\pm$62 kms$^{-1}$ and $\sigma_v$ = 410$\pm$39 kms$^{-1}$, from 44 member galaxies (Table \[tab:vel\]). This velocity dispersion implies a virial radius of $\sim$ 954 $h^{-1}_{70}$kpc and mass of $\sim$ 7$\times$10$^{13}$ $h^{-1}_{70}$M$_{\sun}$ for this group. The shaded histogram in Figure \[fig:veldist\] (right) shows the distribution of the 31 non-emission-line (early-type) galaxies of the sample. The fraction of emission-line galaxies in the group is relatively high and represents about $\sim$ 30% of the known members, and they have preferentially higher velocities than the system, populating the right tail of the distribution. For this reason, we suspect that the tail is due to spiral galaxies falling onto the principal group. Consequently, we derive a slightly lower dispersion from the non-emission galaxies ($\sim$ 361 kms$^{-1}$) than that of the whole sample ($\sim$ 410 kms$^{-1}$). If we do not consider the emission-line galaxies, the calculated average velocity, velocity dispersion, virial radius and virial mass for the non-emission-line population of NGC 6034 are: $V_{avg}$ = 10,347$\pm$66 kms$^{-1}$, $\sigma_v$ = 361$\pm$40 kms$^{-1}$, $R_{virial}$ = 840 $h^{-1}_{70}$kpc, and $M_{virial}$ = 4.7$\times$10$^{13}$ $h^{-1}_{70}$M$_{\odot}$.
{width="80mm"} {width="80mm"}
In order to check the extent of the NGC 6034 group and to minimize the contamination of galaxies that could be part of other structures, we analyzed the cumulative and differential distributions of velocities in the group as a function of radius (Figure \[fig:cumlvel\]). For the cumulative distribution, we calculated the average velocity and the velocity dispersion by including all galaxies inside a given radius, in steps of 100 kpc (denoted in arcmin in the figures). Thus, for a radius of 400 kpc, we included all galaxies from the center of the group up to this radius. With this analysis we can see the general behavior of the group from the central part to the outskirt regions. For the differential distribution, we calculated the average velocity and the velocity dispersion for those galaxies inside concentric rings of 200 kpc from the center of the group. If the ring had less than 10 galaxies, then we increased the width of the ring until we reached this minimum value. This analysis may reveal the degree of anisotropy present in the group.
In both figures we can see an increase in the average velocity and velocity dispersion for radii larger than 16 arcmin ($\sim$ 650 $h_{70}^{-1}$kpc). This increment probably is due to the presence of the three major galaxy clusters in the region. We calculated the average velocity and the velocity dispersion for all galaxies inside this radius. The results are $V_{avg}$ = 10,392$\pm$65 kms$^{-1}$ and $\sigma_v$ = 379$\pm$52 kms$^{-1}$, which results in $R_{virial}$ $\sim$ 882 $h_{70}^{-1}$kpc and $M_{virial}$ $\sim$ 5.5$\times$10$^{13}$ $h_{70}^{-1}$M$_{\odot}$. These are the values we would favor for describing the group NGC 6034. Even with this revised smaller $R_{virial}$ for NGC 6034, three galaxies (J160348.24+171426.2, J160402.75+171656.6, and J160356.65+171818.4) populate the $\Delta r$ $<$ 2 magnitude gap, within $R_{virial}$/2 of the group center.
### Optical Versus X-ray Properties {#sct:cons}
The measured value of the intracluster gas bolometric X-ray luminosity of NGC 6034 is $\sim$ 2.8$\times$10$^{43}$ $h_{50}^{-2}$ ergs$^{-1}$ [@Fukazawa04] and for UGC 842 it is $\sim$ 1.63$\times$10$^{43}$$h^{-2}_{71}$ergs$^{-1}$ [@Voevodkin08]. According to the relations for groups and clusters from @Xue00 and @Mahdavi01, we find that for the X-ray luminosity of NGC 6034 group it is expected a velocity dispersion of about 425 kms$^{-1}$, in good agreement with our direct measurement of 410 kms$^{-1}$ (when we include all galaxies in the group; or 360 kms$^{-1}$ when only the non-emission-line galaxies are included). On the other hand, a velocity dispersion of $\sim$ 450 kms$^{-1}$ is expected for a group with the X-ray luminosity of UGC 842, which is much higher than its measured velocity of 223 kms$^{-1}$ even if we assume that each substructure seen in its line of sight contributes with one-half of the observed L$_{X}$. The first consequence of this divergence between X-rays and optical properties is the determination of the true virial radius for UGC 842. From dynamical relations of velocity distribution of galaxies we determine a virial radius of 509 $h^{-1}_{70}$kpc, while the intragroup X-ray gas implies a virial radius of 1272 $h^{-1}_{70}$kpc, according to @Gastaldello07. This apparent inconsistency is discussed in Section \[sct:disc\].
Summary and Discussion {#sct:disc}
======================
We summarize below the main findings of this paper:
About NGC 6034:
- NGC 6034 is a group of $\sim$ 4 L$^{*}$ galaxies with a mass of $\sim$ 7$\times$10$^{13}$ $h^{-1}_{70}$M$_{\odot}$ and $R_{virial}$ of $\sim$ 954 $h^{-1}_{70}$kpc (or $\sim$ 4.7$\times$10$^{13}$ $h^{-1}_{70}$M$_{\odot}$ and $R_{virial}$ of $\sim$ 840 $h^{-1}_{70}$kpc if only non-emission-line galaxies are considered). It is not a fossil group given that the magnitude difference in the $r$ band between the first and second (J160402.75+171656.6) ranked galaxies is only 0.79 mag and the projected distance between these two galaxies is $\sim$ 360 $h_{70}^{-1}$kpc. Other four (or five; see Section \[sct:cmd\]) galaxies also violate the optical criteria to classify it as a fossil group.
- NGC 6034 is clearly part of a much larger structure that includes at least three clusters (among them the Hercules cluster) and several other groups.
- The velocity distribution of NGC 6034 is fairly well represented by a Gaussian except for a high velocity tail composed of spiral galaxies. These are most probably objects falling onto the system. The velocity dispersion of only the non-emission-line galaxies (31 members with measured redshifts) is 361$\pm$40 kms$^{-1}$ and including the 13 emission-line objects it increases slightly to 410$\pm$39 kms$^{-1}$.
About UGC 842:
- The system referred to as UGC 842 in the literature is in fact a superposition of two groups, S1 and S2, with a velocity difference of about 820 kms$^{-1}$.
- UGC 842/S1 is dominated by passive galaxies while there is a high fraction (40%) of emission galaxies in S2.
- UGC 842/S1 is dominated by a bright elliptical galaxy and it is a low mass fossil [*group*]{}. The large content in passive galaxies suggests equilibrium, and the estimated virial mass and radius are 1.1$\times$10$^{13}$ $h^{-1}_{70}$M$_{\odot}$ and 509 $h^{-1}_{70}$kpc, respectively.
- There is a large discrepancy between the expected temperature and X-ray luminosity expected for a group with such a low sigma and its observed values of $kT$ of 1.90$\pm$0.30keV and L$_{X,\rm bol}$ $\sim$ 1.63$\times$10$^{43}$$h^{-2}_{71}$ ergs$^{-1}$ by @Voevodkin08, suggesting that we have a case of interaction of the two subclumps S1 and S2 or we are whitenessing the decrease of sigma due to the central merger event.
Other previous papers have also studied UGC 842. However, it has not been realized in these works that what is referred to as UGC 842 is in fact two systems. For example, a velocity dispersion of 439 kms$^{-1}$ was derived by @Voevodkin08 from the redshifts of 16 galaxies within a radius of 509 $h^{-1}_{71}$kpc around UGC 842. This value is in good agreement with our value of 471$\pm$37 kms$^{-1}$ determined from the redshifts of 35 potential member galaxies, if we ignore the double-peaked velocity distribution. However, the KMM-test [@Ashman94] rejects a Gaussian velocity distribution at the 99.5% level, and the distribution reveals clearly two structures – S1, the UGC 842 group, and S2 – with a velocity dispersion of about 230 kms$^{-1}$ each (see Figure \[fig:veldist\_ugc\] and Table \[tab:vel\]). The fact that the bolometric X-ray luminosity of UGC 842 fits well in the L$_X$-T$_X$ relation, the measured sigma for the group is too low in the L$_X$–$\sigma_v$ and T$_X$–$\sigma_v$ group relations [e.g., @Xue00]: UGC 842 seems to be too luminous and too hot, with a $kT$ of $\sim$1.9 keV [@Voevodkin08], for its sigma. The expected X-ray temperature for S1 and S2 from their velocity dispersion is around 0.5–1 keV, and the superposition of both plasmas in the line of sight does not imply in the detection of a temperature as high as $kT$$\sim$1.9keV. Although not too pronounced, similar behavior has also been observed in other fossil groups [@Khosroshahi07]. In fact, this high value for the temperature of UGC 842 is in line with the suggestion of @Khosroshahi07 that both temperature and X ray luminosity has been boosted for fossil groups compared to the values for “normal” groups with similar velocity dispersions, given their early times of formation. Another possibility we envisage is that the galaxies have lower relative velocities in fossil groups given that they have lost energy by dynamical friction in the process of spiraling towards the group center for interacting and finally merging. The radial velocity difference between the peaks of the velocity distribution of S1 and S2 ($\sim$ 820 kms$^{-1}$; see Section \[sct:veldist\]) suggests that the two groups are simply overlapping in the line of sight. Although the smooth appearance of the X-ray image [@Voevodkin08] agrees with this hypothesis, we think the SNR of the X-ray observation cannot rule out the presence of X-ray substructures that could indicate a recent interaction of the two sub-groups. Better S/N and spatial resolution X-ray data may reveal if there is an interaction between the two groups (S1 or S2) or not. Thus, the measured temperature of $kT$ $\sim$ 1.9 keV for UGC 842 may either represent the intragroup medium of the most massive group, S1, or it could be the result of an interaction between S1 and S2. Both possibilities make UGC 842 especially interesting in the study of formation and evolution of fossil groups.
We thank financial support from the Brazilian agency FAPESP (Fundação de Amparo à Pesquisa do Estado de São Paulo) and CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico). R.L.O.: FAPESP Postdoctoral Research Fellow grant – number 2007/04710-1. E.R.C. is supported by the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., on behalf of the international Gemini partnership of Argentina, Australia, Brazil, Canada, Chile, the United Kingdom, and the United States of America. We thank Renato Dupke for valuable discussions on X-ray properties of groups/clusters of galaxies. This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration, and use of data of the Sloan Digital Sky Survey. Funding for the SDSS and SDSS-II was provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society, and the Higher Education Funding Council for England. The SDSS was managed by the Astrophysical Research Consortium for the Participating Institutions.
Adelman-McCarthy, J. K., et al. 2008, ApJS, 175, 297
Ashman, K. M., Bird, C. M., & Zepf, S. E. 1994, AJ, 108, 2348
Barden, S. C., Ingerson, T. E. in ASP Conf. Ser. 152, Fiber Optics in Astronomy III, ed. S. Arribas, E. Mediavilla, and F. Watson (San Francisco, CA: ASP), 60
Beers, T. C., Flynn, K., & Gebhardt, K. 1990, AJ, 100, 32
Bird, C. M., Davis, D. S. & Beers, T. C. 1995, AJ, 109, 920
Canizares, C. R., Fabbiano, G., & Trinchieri, G. 1987, ApJ, 312, 503
Carlberg, R. G., et al. 1997, ApJ, 485, 13
Corwin, H. G., Jr. 1971, PASP, 83, 320
Cypriano, E. S., Mendes de Oliveira, C., & Sodré, L., Jr. 2006, , 132, 514
Dariush, A., Khosroshahi, H. G., Ponman, T. J., Pearce, F., Raychaudhury, S., Hartley, W. 2007, MNRAS, 382, 433
Díaz-Giménez, E., Muriel, H., Mendes de Oliveira, C. 2008, A&A, 490, 965
Dickey, J. M. 1997, AJ, 113, 1939
D’Onghia, E., Sommer-Larsen, J., Romeo, A. D., Burkert, A., Pedersen, K., Portinari, L., & Rasmussen, J. 2005, ApJ, 630, 109
Ebeling, H., et al. 1998, MNRAS, 301, 881
Fukazawa, Y., Makishima, K., & Ohashi, T. 2004, PASJ, 56, 965
Gal, R. R., et al. 2003, AJ, 125, 2064
Gastaldello, F., et al. 2007, ApJ, 669, 158
Gioia, I. M., Feigelson, E. D., Maccacaro, T., Schild, R., & Zamorani, G. 1983, ApJ, 271, 524
Goto, T., et al. 2002, AJ, 123, 1807
Huchra, J. P., Vogeley, M. S., & Geller, M. J. 1999, ApJS, 121, 287
Khosroshahi, H. G., Ponman, T. J., & Jones, L. R. 2007, MNRAS, 377, 595
La Barbera, F., de Carvalho, R. R., de la Rosa, I. G., Sorrentino, G., Gal, R. R., & Kohl-Moreira, J. L. 2009, AJ, 137, 3942
Mahdavi, A., & Geller, M. J. 2001, ApJ, 554, 129
Mendes de Oliveira, C. L., Cypriano, E. S., Dupke, R. A., & Sodré Jr., L. 2009, AJ, 138, 502
Mendes de Oliveira, C., Cypriano, E. S., & Sodré Jr., L. 2006, , 131, 158
Ponman, T. J. et al. 1994, Nature, 369, 462
Ramella, M.,Pisani, A. & Geller, M. J. 2000, AJ, 113, 483
Sales, L. V., Navarro, J. F., Lambas, D. G., White, S. D. M., Croton, D. J. 2007, MNRAS, 382, 1901
Santos, W. A., Mendes de Oliveira, C., & Sodré, L., Jr. 2007, AJ, 134, 1551
Struble, M. F. & Rood, H. J. 1999, ApJS, 125, 35
Tarenghi, M., Tifft, W. G., Chincarini, G., Rood, H. J., & Thompson, L. A. 1979, ApJ, 234, 793
Tonry, J., & Davis, M. 1979, AJ, 84, 1511
Valentijn, E. A., & Perola, G. C. 1978, A&A, 63, 29
Voevodkin, A., Miller, C. J., Borozdin, K., Heitmann, K., Habib, S., Ricker, P., & Nichol, R. C. 2008, ApJ, 684, 204
Xue, Y., & Wu, X. 2000, ApJ, 538, 65
White, R., et al. 1999, AJ, 118, 2014
Yoshioka, T., Furuzawa, A., Takahashi, S., Tawara, Y., Sato, S., Yamashita, K., & Kumai, Y. 2004, Adv. Space Res., 34, 2525
[^1]: 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.
[^2]: Milk flats are calibrations obtained once per run, during the afternoon, by opening the dome and placing a dispersing filter in front of the row of fibers, in the spectrograph.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the application of the Deift-Zhou steepest descent method to the Riemann-Hilbert problem for orthogonal polynomials, a model Riemann-Hilbert problem that appears in the multi-cut case is solved with the use of hyperelliptic theta functions. We present here an alternative approach which uses meromorphic differentials instead of theta functions to construct the solution of the model Riemann-Hilbert problem. By using this representation, we obtain a new and elementary proof for the solvability of the model Riemann-Hilbert problem.'
author:
- 'Arno Kuijlaars[^1] and Man Yue Mo[^2]'
date:
title: 'The global parametrix in the Riemann-Hilbert steepest descent analysis for orthogonal polynomials'
---
The global parametrix
=====================
Introduction
------------
The Deift-Zhou steepest descent method is a powerful technique in the asymptotic analysis of Riemann-Hilbert problems that has been successfully applied to numerous problems in integrable systems, random matrix theory and orthogonal polynomials, see e.g.[@Deift], [@DIZ], [@DKMVZ], [@DZ1], [@KMM]. When applying the steepest descent method as in [@DKMVZ], one performs a series of transformations to the $2\times 2$ matrix-valued Riemann-Hilbert problem for orthogonal polynomials to eventually approximate it by a “model Riemann-Hilbert problem” which is also known as the “global parametrix” or the “outer parametrix”. The solution of the model Riemann-Hilbert (RH) problem in [@Apt], [@DIZ], [@DKMVZ], [@DIKZ], [@KMM Section 4.3] and [@KK] uses Riemann theta functions on hyperelliptic Riemann surfaces. In [@Kor04], such approach was generalized to higher dimensional Riemann-Hilbert problems with quasi-permutation jump matrices and the solutions were expressed in terms of Riemann theta functions together with the Szegö kernel. These constructions use various notions from Riemann surfaces, which, although classical, require a fair amount of background in algebraic geometry.
In this paper, we present an alternative approach to the solution of the model RH problem. The approach in this paper is less constructive since it does not lead to explicit formulas. However, in many applications, such as the universality results of random matrix theory, one is merely interested in the existence of a global parametrix, rather than its explicit form. In such cases, the lack of explicit formulas is not an issue. The present approach also enables us to obtain a new elementary proof for the solvability of the model Riemann-Hilbert problem.
On the other hand, our approach is conceptually rather simple (in our opinion), and generalizes without too much effort from hyperelliptic (i.e., two-sheeted) Riemann surfaces to multi-sheeted Riemann surfaces which arise in the steepest descent analysis of larger size RH problems associated with multiple orthogonal polynomials [@VAGK]. In [@DuKu], the RH steepest descent analysis of a 4 x 4 matrix-valued Riemann-Hilbert problem was done with the help of an associated Riemann surface that is a four-sheeted cover of the Riemann sphere. The analysis in [@DuKu] was restricted to the one-cut case (i.e., genus zero). The extension to the multi-cut case was done in [@Mo] where both meromorphic differentials and Riemann theta functions are used to solve the model RH problem. Analyzing the approach in [@Mo] we found that it is also possible to avoid the use of Riemann theta functions completely and to use meromorphic differentials only. For the sake of clarity we present this approach here for the case of the $2\times 2$ matrix-valued model RH problem as it arises in the steepest descent analysis for orthogonal polynomials. We will use this construction in the forthcoming work [@DKM] for a $4 \times 4$ matrix-valued RH problem. See also [@BDK] for a similar situation in a $3 \times 3$ context.
One of the main problems in solving the model Riemann-Hilbert problem is the proof of its solvability. As pointed out in [@DIKZ] and [@KK], the model Riemann-Hilbert problem can be represented as the monodromy problem of the Schlesinger equation. To see whether the monodromy problem is solvable, one can construct an isomonodromic tau function [@JM], [@Mal], corresponding to the Schlesinger equation and the monodromy problem will be solvable if the value of the tau function is non-zero. However, the determination of the zeroes of the tau function, known as the Malgrange divisor, is often a difficult task. In [@KK] and [@Kor04], it was shown that for the type of model Riemann-Hilbert problem considered here, the isomonodromic tau function is zero if and only if a certain theta function is zero. Therefore to prove the solvability of the model Riemann-Hilbert problem, one would need to study the theta divisor, which is a highly transcendental object. By using the approach in this paper, we were able to use much more elementary arguments to show the existence of the global parametrix in the hyperelliptic case, which allows us to avoid the theta divisor completely.
The model Riemann-Hilbert problem
---------------------------------
The model RH problem that arises in the application of the Deift-Zhou steepest descent method to orthogonal polynomials is the following.
We are given $N$ intervals $[a_k,b_k]$, $k=1,\ldots, N$ on the real line ordered so that $b_k < a_{k+1}$ for $k=1, \ldots, N-1$. We also have $N-1$ real numbers $\alpha_k$ for $k=1, \ldots, N-1$, and an integer $n$. The aim is then to construct a solution of the following RH problem.
\[def:model\] The model Riemann-Hilbert problem is the following RH problem for a $2 \times 2$-matrix valued function $M: \mathbb C \setminus [a_1, b_N] \to \mathbb C^{2\times 2}$:
1. $M$ is analytic on $\mathbb C \setminus [a_1,b_N]$,
2. $M$ has jumps $M_+(x) = M_-(x) J_M(x)$ for $x \in [a_1, b_N]$ where $$\begin{aligned}
\label{eq:JM1}
J_M(x) & = \begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix}, \qquad
\text{for } x \in (a_k,b_k),\end{aligned}$$ for $k=1, \ldots, N$, $$\begin{aligned}
\label{eq:JM2}
J_M(x) & = \begin{pmatrix} e^{-2\pi i n \alpha_k} & 0 \\
0 & e^{2\pi i n \alpha_k} \end{pmatrix},
\qquad \text{for } x \in (b_k, a_{k+1}),
\end{aligned}$$ for $k=1, \ldots, N-1$,
3. $M(z) = I + O(1/z)$ as $z \to \infty$,
4. $M$ has at most fourth-root singularities near the endpoints $a_k$ and $b_k$.
Clearly the jump condition in the model RH problem only depends on the value of the numbers $n \alpha_k$ modulo the integers, and so we may (and usually do) consider them to belong to $\mathbb{R} \slash \mathbb{Z}$.
The model RH problem was stated and solved in [@DKMVZ]. The goal of this paper is to present an alternative construction and to show that away from the endpoints $a_k$ and $b_k$, the solution and its inverse are bounded in $n$.
As already mentioned, we use the hyperelliptic Riemann surface with cuts along the intervals $[a_k,b_k]$. The main role is played by the meromorphic differentials $\omega_P^{(\nu)}$, $\nu = 1,2$, introduced in Definition \[def:1form\]. The meromorphic differential depends on $N-1$ points $P_1, \ldots, P_{N-1}$ on the Riemann surface. The heart of the matter is Theorem \[thm:Psibijection\] which states that a suitably defined mapping from $P_1, \ldots, P_{N-1}$ to a vector of $B$-periods is bijective. This result allows us to take the points so that the $B$-periods are exactly the numbers $2 \pi i n \alpha_k$, $k=1, \ldots, N-1$, that appear in the jump condition . In Section \[section3\] we define the corresponding Abelian integrals, which after exponentation lead to functions $v_j^{(\nu)}$, $j,\nu=1,2$, that are used in the Definition \[def:M\] of the solution of the model RH problem.
Some further properties of $M$ are discussed in Section \[section4\], including the fact that $M$ and $M^{-1}$ are uniformly bounded in $n$ if we stay away from the endpoints $a_k$ and $b_k$. In the final Section \[section5\] we present an alternative construction for the second row of $M$, assuming that we know the first row of $M$.
Meromorphic differentials {#section2}
=========================
The construction will be based on meromorphic differentials (Abelian differentials of the third kind) on the two-sheeted Riemann surface $\mathcal R$ for the equation $$\begin{aligned}
\label{eq:SurfaceEquation}
w^2 = \prod_{k=1}^N (z-a_k)(z-b_k)
\end{aligned}$$ which is obtained by gluing together two copies of $\overline{\mathbb C} \setminus \bigcup_{k=1}^N [a_k,b_k]$ along the cuts $[a_k,b_k]$ in the usual crosswise manner. The surface is compact (we add a point at infinity to each sheet) and has genus $N-1$.
We need a few standard facts about Riemann surfaces. Our main reference is [@FK].
[CanonicalHomology.eps]{} (18,14)[$a_1$]{} (28,14)[$b_1$]{} (34,14)[$a_2$]{} (47,14)[$b_2$]{} (50.5,14)[$a_3$]{} (61,14)[$b_3$]{} (66,14)[$a_4$]{} (76,14)[$b_4$]{} (35,11)[$A_1$]{} (29,22)[$B_1$]{} (52,11)[$A_2$]{} (44,24)[$B_2$]{} (68,11)[$A_3$]{} (55,26)[$B_3$]{} (23,22.7)[$\blacktriangleleft$]{} (30,25.3)[$\blacktriangleleft$]{} (37,27.5)[$\blacktriangleleft$]{} (32,19.5)[$\blacktriangleright$]{} (49.5,19.5)[$\blacktriangleright$]{} (64,19.5)[$\blacktriangleright$]{}
We will now define the canonical homology basis on this Riemann surface. We choose the cycles $A_j$ and $B_j$ for a canonical homology basis $$(A_1, \ldots, A_{N-1}; B_1, \ldots, B_{N-1})$$ as indicated in Figure \[fig:canonicalhomology\] for the case $N=4$. That is, $B_j$ is a cycle on the first sheet that encircles the interval $[a_1, b_j]$ once in the counterclockwise direction. We choose $B_j$ to be symmetric with respect to the real axis on the first sheet. The cycle $A_j$ has a part in the upper half-plane of the first sheet, a part in the lower half-plane of the second sheet, and passes through the cuts $[a_j,b_j]$ and $[a_{j+1},b_{j+1}]$ with an orientation also indicated in Figure \[fig:canonicalhomology\].
There is an anti-holomorphic involution $\phi$ on $\mathcal R$ which maps $z$ to $\overline{z}$ on the same sheet. The set of fixed points of $\phi$ are the real ovals, which are $N$ closed contours $\Gamma_j$, $j=0, \ldots, N-1$, where for $j=1, \ldots, N-1$, $\Gamma_j$ is the union of the two intervals $[b_j,a_{j+1}]$ from both sheets and $\Gamma_0$ is unbounded and contains the intervals from $a_1$ and $b_N$ to the two points at infinity. Thus $$\begin{aligned}
\label{def:gammaj}
\Gamma_j & = \{ (z,w) \in \mathcal R \mid b_j \leq z \leq a_{j+1} \},
\quad j=1, \ldots, N-1.
\end{aligned}$$ The cycle $A_j$ is homotopic to $\Gamma_j$ but we choose $A_j$ to be disjoint from $\Gamma_j$.
For each $j = 1, \ldots, N-1$ we choose a point $P_j \in \Gamma_j$. We are going to associate with the $N-1$ points $(P_1,\ldots,P_{N-1})$ and an index $\nu \in \{1,2\}$ a meromorphic differential $\omega = \omega_P^{(\nu)}$.
\[def:1form\] The meromorphic differential $\omega =\omega_P^{(\nu)}$ associated with $(P_1,\ldots,P_{N-1}) \in \Gamma_1 \times \cdots \times
\Gamma_{N-1}$ and $\nu \in \{1,2\}$, is defined by the following properties:
- $\omega_P^{(\nu)}$ has simple poles at the points $a_1, b_1, \ldots, a_N, b_N$, $P_1, \ldots, P_{N-1}$ with residues $$\begin{aligned}
\label{eq:res1}
{\mathrm{Res}}(\omega_P^{(\nu)}, a_j) & = {\mathrm{Res}}(\omega_P^{(\nu)}, b_j) = -\tfrac{1}{2},
\qquad j=1,\ldots, N, \\
\label{eq:res2}
{\mathrm{Res}}(\omega_P^{(\nu)}, P_j) & = 1, \qquad j=1, \ldots, N-1,
\end{aligned}$$
- $\omega_P^{(1)}$ has a simple pole at $\infty_2$ (the point at infinity on the second sheet) and $\omega_P^{(2)}$ has a simple pole at $\infty_1$ (the point at infinity on the first sheet) with residue $$\begin{aligned}
\label{eq:res3}
{\mathrm{Res}}(\omega_P^{(1)}, \infty_2) & = {\mathrm{Res}}(\omega_P^{(2)}, \infty_1) =1.
\end{aligned}$$
- The differential $\omega_P^{(\nu)}$ is holomorphic elsewhere.
- The $A$-periods satisfy: $$\begin{aligned}
\label{eq:Aperiods}
\int_{A_j} \omega_P^{(\nu)} = 0, \qquad j=1, \ldots, N-1.
\end{aligned}$$ Note that the points $P_j$ are not on any of the $A$-cylces.
The meromorphic differential $\omega_P^{(\nu)}$ exists and is uniquely defined by the properties –. Indeed, a simple count shows that the sum of the residues in – is equal to zero which is a necessary and sufficient condition for the meromorphic differential to exist, see for example Theorem II.5.3 in [@FK]. The residue conditions determine the meromorphic differential up to a holomorphic differential. The vector space of holomorphic differentials has dimension $N-1$, and the $N-1$ conditions in determine the meromorphic differential uniquely.
If one or more of the $P_j$’s coincide with a branch point, then the residue conditions – have to be modified appropriately. For example, if $P_j = b_j$ then $${\mathrm{Res}}(\omega_P^{(\nu)}, P_j) = \tfrac{1}{2}.$$ In this way, the meromorphic differential $\omega_P^{(\nu)}$ depends continuously on the $P_j$’s. This fact will play a role in the proof of Proposition \[prop:Psicontinuous\] below.
The anti-holomorphic involution $\phi$ can be used to map $\omega_P^{(\nu)}$ to a meromorphic differential $\phi^{\#}(\omega_P^{(\nu)})$ in an obvious way. If $\omega_P^{(\nu)}$ is equal to $f_j(z) dz$ for some meromorphic function $f_j$ on sheet $j$ for $j=1,2$, then $\phi^{\#}(\omega_P^{(\nu)})$ is equal to $$\overline{f_j(\overline{z})} \, dz$$ on sheet $j$. A crucial property is that $\omega^{(\nu)}_P$ is invariant under the map $\phi^{\#}$.
\[lem:omegasymmetry\] For every $(P_1, \ldots, P_{N-1}) \in \Gamma_1 \times \cdots \times
\Gamma_{N-1}$ and $\nu \in \{1,2\}$, we have $$\label{eq:phiomegaP}
\omega_P^{(\nu)} = \phi^{\#}(\omega_P^{(\nu)}).$$
Since all poles of $\omega_P^{(\nu)}$ are invariant under $\phi$, the meromorphic differential $\phi^{\#}(\omega_P^{(\nu)})$ has the same (simple) poles as $\omega_P^{(\nu)}$ with the same residues. For the $A$-periods we have by definition of $\phi$, $$\int_{A_j} \phi^{\#}(\omega_P^{(\nu)}) = \overline{\int_{\phi(A_j)} \omega_P^{(\nu)}},
\qquad j =1, \ldots, N-1.$$ The cycle $\phi(A_j)$ is homotopic to $A_j$ in $\mathcal R$. In the process of deforming $\phi(A_j)$ to $A_j$ we pick up residue contributions from the poles of $\omega_P^{(\nu)}$ at $b_j$, $P_j$ and $a_{j+1}$. Since the combined residue is $-\tfrac{1}{2} + 1 -
\tfrac{1}{2} = 0$, it follows that $$\int_{A_j} \phi^{\#}(\omega_P^{(\nu)}) = \overline{\int_{A_j} \omega_P^{(\nu)}} = 0.$$ Therefore $\phi^{\#}(\omega_P^{(\nu)})$ has all the properties that characterize $\omega_P^{(\nu)}$ and the lemma follows.
The map $\Psi^{(\nu)}$ from $(P_1, \ldots, P_{N-1})$ to the $B$-periods
-----------------------------------------------------------------------
The differential $\omega_P^{(\nu)}$ has a vector of $B$-periods and it is convenient for us to divide by $2 \pi i$. So we define $(\beta_1, \ldots, \beta_{N-1})$ with $$\label{eq:betak}
\beta_k = \frac{1}{2\pi i} \int_{B_k} \omega_P^{(\nu)}.$$ Then by mapping the points $(P_1,\ldots,P_{N-1})$ to the $B$-periods of the differential $\omega_P^{(\nu)}$, we obtain a map from $\Gamma_1\times\cdots\times\Gamma_{N-1}$ into $\mathbb{C}^{N-1}$. We will show that this map is well-defined from $\Gamma_1\times\cdots\times\Gamma_{N-1}$ to $\left(\mathbb{R} \slash
\mathbb{Z}\right)^{N-1}$.
\[prop:Psicontinuous\] For $\nu =1,2$, the map $$\begin{gathered}
\label{def:Psi}
\Psi^{(\nu)} : \Gamma_1 \times \cdots \times \Gamma_{N-1} \to \left( \mathbb R \slash \mathbb Z \right)^{N-1} : \\
(P_1, \ldots, P_{N-1}) \mapsto
(\beta_1, \ldots, \beta_{N-1}) =
\frac{1}{2\pi i} \left(\int_{B_1} \omega_P^{(\nu)}, \ldots, \int_{B_{N-1}} \omega_P^{(\nu)} \right)
\end{gathered}$$ is well-defined and continuous.
Since the meromorphic differential $\omega_P^{(\nu)}$ depends continuously on the $P_j$’s, we then also have that the $\beta_k$ from is well-defined and continuous in the $P_j$’s, unless $P_k$ is on $B_k$.
Recall that $B_k$ intersects the interval $(a_k,
b_{k+1})$ in one point on the first sheet. The value of $\beta_k$ then makes a jump when $P_k \in \Gamma_k$ passes through this intersection point. As the residue of $\omega_P^{(\nu)}$ at the pole $P_k$ is an integer (in fact, it is $1$, see ), the jump in $\beta_k$ is by an integer value. Since we consider values modulo $\mathbb Z$, the map $\Psi^{(\nu)}$ is thus well-defined and continuous from $\Gamma_1\times\cdots\times\Gamma_{N-1}$ into $\left( \mathbb C
\slash \mathbb Z \right)^{N-1}$.
Now let us show that the $\beta_k$ are real and hence $\Psi^{(\nu)}$ is really a map into $\left( \mathbb R \slash \mathbb Z
\right)^{N-1}$. Recall that the cycle $B_k$ is chosen to be symmetric with respect to the real axis. Therefore $\phi(B_k) = -
B_k$. Since $\omega_P^{(\nu)} = \phi^{\#}(\omega_P^{(\nu)})$ by Lemma \[lem:omegasymmetry\], we then have $$\begin{aligned}
\int_{B_k} \omega_P^{(\nu)} & = \int_{B_k} \phi^{\#}(\omega_P^{(\nu)})
= \overline{\int_{\phi(B_k)} \omega_P^{(\nu)}}
= \overline{\int_{-B_k} \omega_P^{(\nu)}}
= - \overline{\int_{B_k} \omega_P^{(\nu)}}.
\end{aligned}$$ Thus $\int_{B_k} \omega_P^{(\nu)}$ is purely imaginary and so $\beta_k$ is real indeed.
The main result of this section is that the map is a bijection. This would imply that there exists a unique set of points $(P_1^{(\nu)},\ldots,P_{N-1}^{(\nu)})$ such that the $B$-periods of the differential $\omega_P^{(\nu)}$ are given by $2\pi i n
\alpha_j$. The bijectivity proof relies on the fact that the divisor corresponding to any choice of points $P_1, \ldots, P_{N-1}$ with $P_j \in \Gamma_j$ is non-special.
We use additive notation for divisors and we write $$\label{eq:divisorD}
D = \sum_{j=1}^{N-1} P_j.$$ A divisor is special if there exists a non-constant holomorphic function on $\mathcal R \setminus \{ P_1,
\ldots, P_{N-1} \}$ with at most simple poles at the points $P_j$. By the Riemann-Roch theorem, the divisor is special if and only if there exists a non-zero holomorphic differential with zeros at each of the points $P_j$ for $j=1,
\ldots, N-1$.
\[lem:nonspecial\] (see also Statement 1 in [@Kor]) If $P_j \in \Gamma_j$ for each $j=1, \ldots, N-1$, then the divisor is non-special.
The holomorphic differentials on the hyperelliptic Riemann surface defined by are of the form $$\frac{p(z)}{w} dz,$$ where $p$ is a polynomial of degree $\leq N-2$. Therefore the zeros of a non-zero holomorphic differential project onto at most $N-2$ points in the complex $z$-plane. The points $P_j \in \Gamma_j$, $j=1, \ldots, N-1$, project onto $N-1$ distinct points, and so there can be no non-zero holomorphic differential with a zero at each of the $P_j$’s.
In the proof of Theorem \[thm:Psibijection\], which is the main result of this section, we also need the invariance of domain theorem of Brouwer [@Bro], which is a classical result from topology. See e.g. [@Dug section XVII 3] or [@Ful] for more recent accounts. We state the theorem here for the reader’s convenience.
\[thm:invariance\] **(invariance of domain)** If $U$ is an open subset of $\mathbb R^n$ and $f : U \to \mathbb R^n$ is an injective continuous map, then $f$ is open (i.e., $f$ maps open subsets of $U$ to open subsets of $\mathbb R^n$).
Of course, the theorem readily extends to injective continuous maps between manifolds of the same dimension, which is what we will use in the proof of Theorem \[thm:Psibijection\].
\[thm:Psibijection\] The map $\Psi^{(\nu)}$ defined in is a bijection from $\Gamma_1 \times \cdots \times \Gamma_{N-1}$ to $\left( \mathbb
R \slash \mathbb Z \right)^{N-1}$.
We first prove that $\Psi^{(\nu)}$ is injective. Suppose $(P_1,
\ldots, P_{N-1})$ and $(Q_1, \ldots, Q_{N-1})$ are in $\Gamma_1
\times \cdots \times \Gamma_{N-1}$ so that $$\Psi^{(\nu)}(P_1, \ldots, P_{N-1}) = \Psi^{(\nu)}(Q_1, \ldots, Q_{N-1}).$$ Let $\omega_P^{(\nu)}$ and $\omega_Q^{(\nu)}$ be the corresponding meromorphic differentials. Then $\omega_Q^{(\nu)} -
\omega_P^{(\nu)}$ has poles with residues $\pm 1$ at the points $Q_j$ and $P_j$ only and all periods are integer multiples of $2\pi
i$. Let $P_0$ be a given base point different from any of the $P_j$’s and $Q_j$’s. It then follows that $$\exp\left(\int_{P_0}^z \left(\omega_Q^{(\nu)} - \omega_P^{(\nu)}\right)\right),
\qquad z \in \mathcal R,$$ is a meromorphic function on $\mathcal R$ with only possible poles at $P_1, \ldots, P_{N-1}$. Since the divisor $D = \sum_{j=1}^{N-1}
P_j$ is non-special, see Lemma \[lem:nonspecial\], the meromorphic function is a constant, which implies that $\omega_Q^{(\nu)} = \omega_P^{(\nu)}$. Hence the $Q_j$’s and the $P_j$’s coincide: $(Q_1, \ldots, Q_{N-1})
= (P_1, \ldots, P_{N-1})$ and therefore $\Psi^{(\nu)}$ is injective.
To prove surjectivity we now note that $\Psi^{(\nu)}$ is an injective and continuous (by Proposition \[prop:Psicontinuous\]) map from the $N-1$-dimensional manifold $\Gamma_1 \times \cdots \times \Gamma_{N-1}$ to the $N-1$-dimensional manifold $(\mathbb R \slash \mathbb Z)^{N-1}$. Thus $\Psi^{(\nu)}$ is open by Theorem \[thm:invariance\]. It follows that $\Psi^{(\nu)}(\Gamma_1 \times \cdots \times \Gamma_{N-1})$ is a subset of $(\mathbb R \slash \mathbb Z)^{N-1}$ that is both open (since $\Psi^{(\nu)}$ is open) and compact (since $\Psi^{(\nu)}$ is continuous and $\Gamma_1 \times \cdots \times \Gamma_{N-1}$ is compact). Since $(\mathbb R \slash \mathbb Z)^{N-1}$ is connected it follows that $\Psi^{(\nu)}$ is surjective.
Construction of $M$ {#section3}
===================
It will be a consequence of Theorem \[thm:Psibijection\] that we can construct the matrix $M$ that solves the model Riemann-Hilbert problem.
By Theorem \[thm:Psibijection\] there exist $P_j^{(\nu)} \in
\Gamma_j$ for $j=1, \ldots, N-1$, $\nu =1,2$, so that $$\label{eq:Psiimage}
\Psi^{(\nu)}(P_1^{(\nu)}, \ldots, P_{N-1}^{(\nu)}) = \left(n \alpha_1, \ldots, n \alpha_{N-1}\right)$$ where each $n \alpha_j$ is considered modulo $\mathbb Z$. Let $\omega_P^{(\nu)}$ be the corresponding meromorphic differential. We then have that $$\label{eq:Bperiods}
\frac{1}{2\pi i} \int_{B_k} \omega_P^{(\nu)} \equiv n \alpha_k \quad \mod \mathbb Z,
\qquad \text{for } k = 1, \ldots, N-1.$$
Abelian integrals
-----------------
For $\nu =1,2$, we choose the base point $$\label{eq:basepoint}
P_0 = \infty_{\nu}$$ and we define the functions $u_1^{(\nu)}(z)$, $u_2^{(\nu)}(z)$ of a complex variable $z$ as follows.
\[def:abel\] Let $z \in \mathbb C \setminus \mathbb R$.
1. To define $u_{j}^{(\nu)}$ with $j=\nu$, we consider $z$ as a point on the $j$th sheet of the Riemann surface. We define $$\label{def:u1z}
u_{j}^{(\nu)}(z) = \int_{P_0}^z \omega_P^{(\nu)}, \qquad z \in \mathbb C \setminus \mathbb R,
\quad j = \nu,$$ where the path of integration is on the $j$th sheet of the Riemann surface and it does not intersect the real line, except for the initial point $P_0$.
2. To define $u_{j}^{(\nu)}(z)$ with $j \neq \nu$, we consider $z$ as a point on the $j$th sheet. We define $$\label{def:u2z}
u_j^{(\nu)}(z) = \int_{P_0}^z \omega_P^{(\nu)}, \qquad z \in \mathbb C \setminus \mathbb
R,\quad j \neq \nu,$$ where now the path of integration is as follows. If $\impart z > 0$ ($ \impart z < 0$) then the path starts in the lower (upper) half-plane of the $\nu$th sheet and passes to the $j$th sheet via a cut $(a_k,b_k)$. It then stays in the upper (lower) half-plane of the $j$th sheet.
Since $\omega_P^{(\nu)}$ has vanishing $A$-periods, as well as vanishing $\phi^{\#}(A)$ periods, it does not matter which cut $(a_k,b_k)$ is taken, and so $u_j^{(\nu)}(z)$ in is uniquely defined.
The functions $u_j^{(\nu)}$ are analytic on $\mathbb C \setminus \mathbb
R$ with the following jumps on $\mathbb R$.
\[lem:u1u2jumps\] The functions $u_j^{(\nu)}$, $j,\nu =1,2$ satisfy the following jump conditions for $x \in \mathbb R$.
1. For $x \in (a_k, b_k)$ with $k=1, \ldots, N$, we have $$\begin{aligned}
\label{jumpu1a}
u_{1,+}^{(\nu)}(x) & = u_{2,-}^{(\nu)}(x), \\ \label{jumpu1b}
u_{2,+}^{(\nu)}(x) & = u_{1,-}^{(\nu)}(x).\end{aligned}$$
2. For $x < a_1$ or $x > b_N$ we have $$\begin{aligned}
\label{jumpu2a}
u_{j,+}^{(\nu)}(x) & = u_{j,-}^{(\nu)}(x), \qquad j = \nu, \\ \label{jumpu2b}
u_{j,+}^{(\nu)}(x) & \equiv u_{j,-}^{(\nu)}(x) + \pi i \quad \mod 2 \pi i \, \mathbb Z,\quad j\neq \nu.\end{aligned}$$
3. For $z \in \mathbb C \setminus \bigcup_{k=1}^N [a_k, b_k]$, we use $P_j(z)$ to denote the point on the $j$th sheet of $\mathcal R$ that corresponds to $z$. Then we have, for $x \in (b_k, a_{k+1})$ with $k=1,\ldots, N-1$, $$\begin{aligned}
\label{jumpu3a}
u_{1,+}^{(1)}(x) & \equiv u_{1,-}^{(1)}(x) - 2\pi i n \alpha_k,
\qquad \text{if } P_1(x) \neq P_k^{(1)}, \\
\label{jumpu3b}
u_{2,+}^{(1)}(x) & \equiv u_{2,-}^{(1)}(x) + 2 \pi i n \alpha_k + \pi i
\qquad \text{if } P_2(x) \neq P_k^{(1)}, \\
\label{jumpu3c}
u_{1,+}^{(2)}(x) & \equiv u_{1,-}^{(2)}(x) - 2\pi i n \alpha_k + \pi i
\qquad \text{if } P_1(x) \neq P_k^{(2)}, \\
\label{jumpu3d}
u_{2,+}^{(2)}(x) & \equiv u_{2,-}^{(2)}(x) + 2 \pi i n \alpha_k,
\qquad \text{if } P_2(x) \neq P_k^{(2)},\end{aligned}$$ The equalities – are valid modulo $2\pi i \, \mathbb Z$.
The properties and follow immediately from the definition of $u_j^{(\nu)}$.
Let $\Delta^{(\nu)}$ be the set of poles for $\omega_P^{(\nu)}$, that is, $$\Delta^{(\nu)}=\{a_1,b_1,\ldots,a_N,b_N,P_1^{(\nu)},\ldots,P_{N-1}^{(\nu)},
\infty_j\},\quad j\neq \nu.$$
For $x < a_1$ and $x > b_N$, we have $$u_{j,+}^{(\nu)}(x) - u_{j,-}^{(\nu)}(x) = \oint_{C} \omega_P^{(\nu)}$$ where $C$ is a closed contour on $\mathcal R \setminus\Delta^{(\nu)}$. For $j=\nu$, the contour $C$ is contractible in $\mathcal{R} \setminus \Delta^{(\nu)}$ and follows. When $j\neq \nu$, we choose the contour to pass through the cut $[a_1,b_1]$ in case $x < a_1$, and through the cut $[a_N,b_N]$ in case $x > b_N$. Then $C$ is contractible to a small loop around $a_1$ (in case $x < a_1$) or around $b_N$ (in case $x > b_N$). Since the residues of $\omega_P^{(\nu)}$ at $a_1$ and $b_N$ are $-\frac{1}{2}$ we have in either case $$\oint_{C} \omega_P^{(\nu)} = \pi i \qquad \mod 2 \pi i,$$ and follows.
Let $x \in (b_k, a_{k+1})$, $x \neq z(P_k^{(\nu)})$, for some $k=1,
\ldots, N-1$. Then we have $$u_{j,+}^{(\nu)}(x) - u_{j,-}^{(\nu)}(x) = \oint_{C} \omega_P^{(\nu)}$$ where $C$ is again a closed contour on $\mathcal R \setminus \Delta^{(\nu)}$. When $j = \nu = 1$, then $C$ is on the first sheet and can be deformed into $-B_k$ and follows because of .
When $j= \nu = 2$, then $C$ is a closed contour on the second sheet, and it is homotopic to $B_k$ in $\mathcal R$, but the deformation will pick up residue contributions from the poles at $a_1,\ldots,a_k, b_1,\ldots,b_k$, $P_1^{(\nu)},\ldots,P_{k-1}^{(\nu)}$ and possibly $P_k^{(\nu)}$ depending on its position. Since the combined residues of $\omega_P^{(\nu)}$ at these poles is an integer number, the poles do not contribute (modulo $2\pi i$) and we obtain from .
When $j\neq \nu$, the closed loop $C$ is on both sheets, For $j=2$, $\nu=1$, we choose $C$ so that it passes through the cut $[a_k,b_k]$. Then $C$ is homotopic to $B_k$ in $\mathcal R$, but a deformation from $C$ to $B_k$ will pick up a residue contribution from $b_k$ and possibly from $P_k^{(\nu)}$. The combined residue is $-1/2$ or $+1/2$, and this leads to .
Finally, for $j=1$, $\nu=2$, we choose $C$ so that it passes through the cut $[a_1,b_1]$. Then $C$ is homotopic to $-B_k$ in $\mathcal
R$, and the deformation from $C$ to $-B_k$ picks up a residue contribution at $a_1$ and possibly $P_k^{(\nu)}$. The combined residue is $-1/2$ or $+1/2$, and we obtain .
Note that in part (c) of Lemma \[lem:u1u2jumps\] we excluded the case $P_j(x) = P_k^{(\nu)}$ since $P_k^{(\nu)}$ is a pole of $\omega_P^{(\nu)}$ and so the limiting values $u_{j,+}^{(\nu)}(x)$ and $u_{j,-}^{(\nu)}(x)$ do not exist if $P_k^{(\nu)}$ is on the $j$th sheet, see also part (b) of Lemma \[lem:u1u2behavior\].
The behavior near all poles of $\omega_P^{(\nu)}$ is stated in the following lemma. We use $z(P)$ to denote the $z$-coordinate of a point $P=(z,w)$ on the Riemann surface .
\[lem:u1u2behavior\] We have
1. for $j, \nu = 1,2$ and $k=1, \ldots, N$, $$\begin{aligned}
u_j^{(\nu)}(z) & = -\tfrac{1}{4} \log(z-a_k) + O(1), \qquad \text{as } z \to a_k, \\
u_j^{(\nu)}(z) & = -\tfrac{1}{4} \log(z-b_k) + O(1), \qquad \text{as } z \to b_k,\end{aligned}$$
2. if $P_k^{(\nu)}$ is on the $j$th sheet of the Riemann surface, then $$\begin{aligned}
\label{eq:ujatPknu}
u_j^{(\nu)}(z) = \log(z- z(P_k^{(\nu)})) + O(1) \qquad \text{as } z \to z(P_k^{(\nu)}),
\end{aligned}$$
3. as $z \to \infty$ $$\begin{aligned}
u_j^{(\nu)}(z) & = O(1/z), \qquad \text{if } j = \nu \\
u_j^{(\nu)}(z) & = - \log z + O(1), \quad \text{if } j \neq \nu.
\end{aligned}$$
The fact that $u_j^{(\nu)}(z) = O(1/z)$ as $z \to \infty$ in case $j = \nu$ follows directly from since $P_0 = \infty_{\nu}$. The other statements of the lemma follow from – and the residue conditions in –.
In Lemma \[lem:u1u2behavior\] we implicitly assumed that the point $P_k^{(\nu)}$ is different from $b_k$ and $a_{k+1}$. If for example, $P_k^{(\nu)} = b_k$, then the residue of $\omega_P^{(\nu)}$ at $b_k$ is equal to $+\tfrac{1}{2}$, and in part (a) of Lemma \[lem:u1u2behavior\] we get $$u_j^{(\nu)}(z) = \tfrac{1}{4} \log(z-b_k) + O(1),
\qquad \text{as } z \to b_k.$$ The modifications that are needed when one or more of the $P_k^{(\nu)}$ coincide with an endpoint are obvious, and we will not specify them in the rest of the paper.
Exponential of the Abelian integrals
------------------------------------
Now we exponentiate the functions $u_j^{(\nu)}$.
\[def:v1v2\] We define $v_j^{(\nu)}$, $j,\nu=1,2$ by $$\label{eq:v1v2}
v_j^{(\nu)}(z) = \exp \left( u_j^{(\nu)}(z) \right), \qquad z \in \mathbb C \setminus \mathbb R.$$
Then the functions $v_j^{(\nu)}(z)$ are analytic in $\mathbb C \setminus
\mathbb R$ with the following jumps on $\mathbb R$.
\[cor:v1v2jumps\] The vectors $(v_1^{(\nu)},v_2^{(\nu)})$ satisfy the following jump conditions on $\mathbb{R} \setminus \{a_1,b_1, \ldots, a_N, b_N\}$, $$\label{eq:vjump}
(v_1^{(\nu)},v_2^{(\nu)})_+ = (v_1^{(\nu)},v_2^{(\nu)})_- J_v^{(\nu)}, \qquad \nu =1,2,$$ where $$\begin{aligned}
\label{eq:Jvnu1}
J_v^{(\nu)}(x) & = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix} \qquad \text{for } a_k < x < b_k,
\quad k = 1, \ldots, N, \\
\label{eq:Jvnu2}
J_v^{(\nu)}(x) & = (-1)^{\nu-1} \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix} \qquad
\text{for } x < a_1 \text{ or } x > b_N, \\
\label{eq:Jvnu3}
J_v^{(\nu)}(x) & = (-1)^{\nu-1} \begin{pmatrix} e^{-2\pi i n \alpha_k} & 0 \\
0 & - e^{2\pi in \alpha_k} \end{pmatrix} \qquad
\text{for } b_k < x < a_{k+1},\end{aligned}$$ for $k=1, \ldots, N-1$.
The jumps follow directly from Lemma \[lem:u1u2jumps\] and Definition \[def:v1v2\] in case $x \neq z(P_k^{(\nu)})$.
If $x = z(P_k^{(\nu)})$ and $P_k^{(\nu)}$ is on the $j$th sheet, then Lemma \[lem:u1u2jumps\] does not apply to $u_j$. However, in that case we find by part (b) of Lemma \[lem:u1u2behavior\] and Definition \[def:v1v2\] that $$v_{j,+}^{(\nu)}(x) = v_{j,-}^{(\nu)}(x) = 0,$$ and , is also valid.
From Lemma \[lem:u1u2behavior\] and we find the following behavior near the poles of $\omega_P^{(\nu)}$.
\[cor:v1v2behavior\] We have
1. for $j, \nu = 1,2$ and $k=1, \ldots, N$, $$\begin{aligned}
v_j^{(\nu)}(z) & = O\left((z-a_k)^{-1/4}\right), \qquad \text{as } z \to a_k, \\
v_j^{(\nu)}(z) & = O\left((z-b_k)^{-1/4}\right), \qquad \text{as } z \to b_k,\end{aligned}$$
2. if $P_k^{(\nu)}$ is on the $j$th sheet of $\mathcal R$, then $$\label{eq:vjatPk}
v_j^{(\nu)}(z) = O (z- z(P_k^{(\nu)})) \qquad \text{as } z \to z(P_k^{(\nu)}),$$
3. as $z \to \infty$ $$\begin{aligned}
v_j^{(\nu)}(z) & = 1 + O(1/z), \quad \text{if } j = \nu \\
v_j^{(\nu)}(z) & = O(1/z), \qquad \text{if } j \neq \nu.
\end{aligned}$$
Note that by the function $v_j^{(\nu)}$, has a zero at $z(P_k^{\nu})$ in case $P_k^{(\nu)}$ is on the $j$th sheet.
The parametrix $M$
------------------
To construct the solution $M$ of the model Riemann-Hilbert problem in Definition \[def:model\], we only need a trivial modification of the functions $v_j^{(\nu)}$.
\[def:M\] We define $$\label{eq:defM}
M(z) = \left\{ \begin{aligned}
\begin{pmatrix} v_1^{(1)}(z) & v_2^{(1)}(z) \\
-v_1^{(2)}(z) & v_2^{(2)}(z) \end{pmatrix}, \qquad \text{if } \impart z > 0, \\
\begin{pmatrix} v_1^{(1)}(z) & -v_2^{(1)}(z) \\
v_1^{(2)}(z) & v_2^{(2)}(z) \end{pmatrix}, \qquad \text{if } \impart z < 0.
\end{aligned} \right.$$
From –, it is easy to verify that $M(z)$ does indeed satisfy the jump conditions – of the RH problem in Definition \[def:model\]. The asymptotic condition $M(z) = I + O(1/z)$ as $z \to \infty$ holds because of part (c) of Corollary \[cor:v1v2behavior\]. Part (a) of Corollary \[cor:v1v2behavior\] shows that $M$ has at most fourth root singularities at the endpoints $a_k$ and $b_k$. So we have proved the main result of this paper:
\[thm:Msolution\] The matrix-valued function $M(z)$ defined by satisfies the model Riemann-Hilbert problem of Definition \[def:model\].
Properties of $M$ {#section4}
=================
We collect here some further properties of $M$ that are useful in applications.
Uniqueness of the solution
--------------------------
The first two properties are standard, see e.g. [@Deift]
\[prop:detM\] For every $z \in \mathbb C \setminus [a_1,b_N]$ we have $$\det M(z) = 1.$$
Since the jump matrices in the model Riemann-Hilbert have determinant $1$, the function $z \mapsto \det M(z)$ has no jump discontinuities in $\mathbb{C}$, and so it has an analytic extension to $\mathbb C \setminus \{a_1, b_1, \ldots, a_N, b_N\}$. From part (d) in Definition \[def:model\], we conclude that $\det M(z)$ can have at most square-root singularities at the endpoint $a_j, b_j$, and therefore these isolated singularties are removable. Thus $z \mapsto \det M(z)$ is an entire function, which by the asymptotic condition satisfies $\det M(z) = 1 + O(z^{-1})$ as $z \to \infty$. Then the proposition follows, by Liouville’s theorem.
\[prop:Munique\] The solution $M$ of the model RH problem is unique.
Let $\widetilde{M}$ be a second solution of the model RH problem. By Proposition \[prop:detM\], we have that $M(z)$ is invertible for every $z \in \mathbb C \setminus [a_1,b_N]$. Then $$H(z) = \widetilde{M}(z) M(z)^{-1}, \qquad z \in \mathbb C \setminus \mathbb [a_1,b_N]$$ is well-defined and analytic. Since $\widetilde{M}$ and $M$ satisfy the same jump conditions, it follows that $H_+(x) = H_-(x)$ for every $x \in [a_1,b_N] \setminus
\{ a_1, b_1, \ldots, a_N,b_N \}$. The entries of $\widetilde{M}$ and $M^{-1}$ have at most fourth root singularities at the endpoints $a_j, b_j$. Thus $H$ has at most square root singularities, and it follows that the singularities are removable. Thus $H$ has an extension to an entire function. Since $H(z) = I + O(1/z)$ as $z \to \infty$, we find by Liouville’s theorem that $H(z) = I$ for every $z \in \mathbb C$. The proposition follows.
Zeros
-----
By and the entries of $M(z)$ do not vanish at any point $z \in \mathbb C \setminus [a_1,b_N]$. Across each cut $(a_k,b_k)$, $k=1, \ldots, N$, and each gap $(b_k,
a_{k+1})$, $k=1, \ldots, N-1$ the entries have an analytic continuation given by the RH problem. By we have that $v_j^{(\nu)}(z) \to 0$ as $z \to z(P_k^{(\nu)})$ if $P_k^{(\nu)}$ is on the $j$th sheet. By this translates into the following statement about the zeroes of the entry $M_{\nu,j}$ of $M$.
\[prop:Mzeros\] If $P_k^{(\nu)}$ is on the $j$th sheet then $M_{\nu,j}(z)$ has a simple zero at $z = z(P_k^{(\nu)})$, in the sense that the restriction of $M_{\nu,j}$ to the upper (lower) half-plane has an analytic continuation across $(b_k,a_{k+1})$ into the lower (upper) half-plane which has a simple zero at $z=z(P_k^{(\nu)})$.
The points $z=z(P_k^{(\nu)})$ are the only possible zeros of the (analytic continuations of the) entries of $M$.
Everything is already proved, except for the fact that the zero is simple. This follows from the fact that $M_{\nu,j} = \pm
e^{u_j^{(\nu)}}$ where $u_j^{(\nu)}(z)$ has the behavior as $z \to z(P_k^{(\nu)})$.
Uniform boundedness
-------------------
The solution $M$ clearly depends on $n$. The following uniform boundedness property is needed in the construction of a local parametrix in the steepest descent analysis, see [@DKMVZ].
\[prop:uniformbound\] For every $\varepsilon > 0$, we have that $M(z)$ and $M^{-1}(z)$ are uniformly bounded in $n$ for $z$ in the set $$\label{eq:setepsilon}
\{ z \in \mathbb C \setminus [a_1,b_N] \mid |z-a_j| \geq \varepsilon, |z-b_j| \geq \varepsilon
\textrm{ for all } j=1, \ldots, N \}.$$
It follows easily from our construction that for every $(\beta_1,
\ldots, \beta_{N-1}) \in (\mathbb R \slash \mathbb Z)^{N-1}$ there is a unique solution to the model RH problem where the jump on $(a_j, b_{j+1})$ is replaced by $$M_+(x) = M_-(x) \begin{pmatrix} e^{-2\pi i \beta_j} & 0 \\ 0 & e^{2\pi i \beta_j}
\end{pmatrix}, \qquad x \in (a_j, b_{j+1})$$ for $j=1, \ldots, N-1$. If we denote the solution of this RH problem by $$\label{eq:Mwithbetas}
M(z; \beta_1, \ldots, \beta_{N-1})$$ then the solutions we are interested in are $$M(z; n \alpha_1, \ldots, n \alpha_{N-1}), \qquad n \in \mathbb N,$$ and so they are part of this family . It is therefore enough to show that the solutions are uniformly bounded on the set .
The map $$(\beta_1, \ldots, \beta_{N-1}) \mapsto M(z; \beta_1, \ldots, \beta_{N-1})$$ is continuous as a map from $(\mathbb R \slash \mathbb Z)^{N-1}$ to the $2 \times 2$-matrix valued analytic functions on $\mathbb C
\setminus [a_1, b_N]$ provided with the topology of uniform convergence on compact subsets of $\mathbb C \setminus \{a_1, b_1,
\ldots, a_N, b_N \}$. Since $(\mathbb R \slash \mathbb Z)^{N-1}$ is compact, we then have that the functions $M(z; \beta_1, \ldots,
\beta_{N-1})$ are also compact in this space, which in particular implies that the functions are uniformly bounded on every set of the form . So $M$ is uniformly bounded in $n$ on .
Since $\det M \equiv 1$, the entries of $M^{-1}$ are, up to a sign, the same as those of $M$, and so $M^{-1}$ is also uniformly bounded in $n$ on .
Alternative construction for the second row {#section5}
===========================================
The above construction of the solution $M$ of the model RH problem is done row by row. Indeed, the case $\nu = 1$ leads to the first row, and the case $\nu=2$ leads to the second row of $M$. The difference between the two cases lies in the condition that specifies which point at infinity is a pole of the meromorphic differential. Otherwise the two cases are similar, and we treated them simultaneously.
The fact that these two cases are similar is also related to the hyperelliptic Riemann surface, which possesses hyperelliptic involution interchanging the sheets. In other situations related to multi-sheeted Riemann surfaces as in [@DuKu], [@Mo], there is no simple symmetry between the sheets. In addition, the point at infinity is a branch point in [@DuKu], [@Mo] that connects all but one of the sheets. For the construction of a model RH problem in such situations, it may be of interest to realize that the construction of one row of the model RH problem can help to construct the other rows.
We illustrate this here for the hyperelliptic case, and so we give an alternative way to construct the second row, on the assumption that we know the first row of $M$. We thank Alexander Aptekarev for this remark.
Recall that the construction of the first row of $M$ is based on the points $P_k^{(1)} \in \Gamma_k$, $k=1, \ldots, N-1$ satisfying (\[eq:Psiimage\]). In Lemma \[lem:nonspecial\] we proved that the divisor $$D = \sum_{k=1}^{N-1} P_k^{(1)}$$ is non-special.
\[lem:dimLD\] The vector space of meromorphic functions on $\mathcal{R}$ (including constant functions) with divisor greater than or equal to $-\sum_{k=1}^{N-1} P_k^{(1)} - \infty_2$ is of dimension $2$.
Let us denote, for a positive divisor $D'$, the space of meromorphic functions on $\mathcal R$ whose divisor is greater than or equal to $-D'$ by $L(D')$.
Since $D = \sum_{k=1}^{N-1} P_k^{(1)}$ is non-special, the only functions in $L(D)$ are constant functions, so that $$\dim L(D) = 1.$$
Also, $L(D)$ is the kernel of the linear functional from $L(D +
\infty_2)$ to $\mathbb C$ that maps a function in $L(D+\infty_2)$ to its residue at $\infty_2$. Thus by the dimension theorem for linear functionals, $$\dim L(D + \infty_2) \leq \dim L(D) + 1 = 2.$$ On the other hand, the divisor $D +\infty_2$ is of degree $N$, so that by the Riemann-Roch theorem, see [@FK], $$\dim L(D + \infty_2) \geq 2.$$ This proves the lemma.
Suppose now that we have the first row of $M$. We can then construct the second row of $M$ by modifying the first row $(M_{11}, M_{12})$ by a meromorphic factor as follows.
By Lemma \[lem:dimLD\], the space of meromorphic functions $F$ on $\mathcal R$ with divisor $\geq -\sum_{k=1}^{N-1} P_k^{(1)} -
\infty_2$ is two-dimensional. The following two conditions on $F$ (recall that $M_{12}$ is analytic in a neighborhood of $\infty_2$ with a simple zero at $\infty_2$) $$\begin{aligned}
F(\infty_1) = 0, \qquad \text{and} \qquad
\lim_{z \to \infty_2} M_{12}(z) F(z) = 1,\end{aligned}$$ determine $F$ uniquely.
We use $F_1$ and $F_2$ to denote the restrictions of $F$ to the first and second sheet, respectively. Then the two functions defined by $$\label{def:M21M22}
M_{21}(z) = M_{11}(z) F_1(z), \qquad M_{22}(z) = M_{12}(z) F_2(z)$$ satisfy all conditions that we need for the entries in the second row of $M$. Note also that the poles of $F_1$ and $F_2$ at $P_1^{(1)}, \ldots, P_{N-1}^{(1)}, \infty_2$ are cancelled by the zeros of $M_{11}$ and $M_{12}$ at these points, see also Proposition \[prop:Mzeros\].
Acknowledgement {#acknowledgement .unnumbered}
===============
We thank Alexander Aptekarev for helpful discussions.
[99]{} A.I. Aptekarev, Analysis of the matrix Riemann-Hilbert problems for the case of higher genus - asymptotics of polynomials orthogonal on a system of intervals, KIAM Preprint No 28, Keldysh Institute of Appl. Math., Moscow, 2008.
P. Bleher, S. Delvaux, and A.B.J. Kuijlaars, Random matrix model with external source and a constrained vector equilibrium problem, to appear in Comm. Pure Appl. Math., Preprint arXiv:1001.1238.
A.A. Bolibruch, The Riemann–Hilbert problem, Russian Math. Surveys 45 (2) (1990), 1–58.
L.E.J. Brouwer, Zur Invarianz des $n$-dimensionalen Gebiets, Math. Ann. 72 (1912), 55–56.
P. Deift, Orthogonal Polynomials and Random Matrices: a Riemann-Hilbert approach. Courant Lecture Notes in Mathematics Vol. 3, Amer. Math. Soc., Providence R.I. 1999.
P. Deift, A.R. Its and X. Zhou, A Riemann-Hilbert approach to asymptotic problems arising in the theory of random matrix models, and also in the theory of integrable statistical mechanics, Ann. of Math. 146 (1997), 149–235.
P. Deift, T. Kriecherbauer, K.T-R McLaughlin, S. Venakides, and X. Zhou, Uniform asymptotics for polynomials orthogonal with respect to varying exponential weights and applications to universality questions in random matrix theory, Comm. Pure Appl. Math. 52 (1999), 1335–1425.
P. Deift and X. Zhou. A steepest descent method for oscillatory Riemann-Hilbert problems. Asymptotics for the MKdV equation, Ann. of Math. (2) 137 (1993), no. 2, 295–368.
P. Deift, A. Its, A. Kapaev, and X. Zhou, On the algebro-geometric integration of the Schlesinger equations, Comm. Math. Phys. 203 (1999), no. 3, 613–633.
J. Dugundji, Topology, Allyn and Bacon, Inc., Boston, Mass., 1966.
M. Duits and A.B.J. Kuijlaars, Universality in the two matrix model: a Riemann-Hilbert steepest descent analysis, Comm. Pure Appl. Math. 62 (2009), 1076–1153.
M. Duits, A.B.J. Kuijlaars, and M.Y. Mo, The Hermitian two matrix model with an even quartic potential, in preparation.
H. Farkas and I. Kra, Riemann Surfaces, Graduate Texts in Mathematics 71, Springer-Verlag, New York-Berlin, 1980.
W. Fulton, Algebraic Topology, a First Course, Springer-Verlag, New York, 1995.
M. Jimbo and T. Miwa, Monodromy preserving deformation of linear ordinary differential equations with rational coefficients II, Physica D 2 (1981), 407–448.
S. Kamvissis, K.D.T-R McLaughlin, and P.D. Miller, Semiclassical soliton ensembles for the focusing nonlinear Schrödinger equation Ann. Math. Studies 154, Princeton University Press, Princeton, NJ, 2003.
A. V. Kitaev and D. Korotkin, On solutions of the Schlesinger equations in terms of $\Theta$-functions, Internat. Math. Res. Notices (1998), 877–905.
D. Korotkin, Introduction to the functions on compact Riemann surfaces and theta-functions, in “Nonlinearity and Geometry” (D. Wojcik and J. Cieslinski, eds.) Polish Scient. Publ. PWN, Warsaw, 1998, pp. 109–139. Preprint arXiv:solv-int/9911002
D. Korotkin, Solution of matrix Riemann-Hilbert problems with quasi-permutation monodromy matrices, Math. Ann. 329 (2004), 335–364.
B. Malgrange, Sur les déformations isomonodromiques I, singularités régulières, in “Mathematics and Physics” (Paris, 1979/1982), Progress in Mathematics 37, Birkhäuser, Boston, 1983, pp. 401–426.
M.Y. Mo, Universality in the two matrix model with a monomial quartic and a general even polynomial potential, Commun. Math. Phys. 291 (2009), 863–894.
W. Van Assche, J. Geronimo, and A.B.J. Kuijlaars, Riemann-Hilbert problems for multiple orthogonal polynomials, in “Special Functions 2000: Current Perspectives and Future Directions” (J. Bustoz et al., eds.), Kluwer, Dordrecht, 2001, pp. 23–59.
[^1]: Department of Mathematics, Katholieke Universiteit Leuven, Celestijnenlaan 200B, B-3001 Leuven, Belgium. email: [email protected]. The work of the first author is supported by FWO-Flanders project G.0427.09, by K.U. Leuven research grant OT/08/33, by the Belgian Interuniversity Attraction Pole P06/02, by the European Science Foundation Program MISGAM, and by grant MTM2008-06689-C02-01 of the Spanish Ministry of Science and Innovation.
[^2]: Department of Mathematics, University of Bristol, Bristol BS8 1TW, U.K. email: [email protected]. The second author is supported by the EPSRC grant EP/G019843/1.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present results of our spectroscopic campaign dedicated to the ultracompact binary ES Ceti. On the nights 2002 Oct. 27-28, 528 spectra were taken with the 6.5-meter telescope in Las Campanas Observatory in Chile. The averaged spectrum shown the double-peaked helium emission lines which imply the presence of an accretion disk in this system.'
author:
- Karolina Bąkowska
- 'Thomas R. Marsh'
bibliography:
- 'ptapapdoc.bib'
title: Spectroscopy of the cataclysmic variable ES Ceti
---
Introduction
============
Ultracompact binaries consist of a primary white dwarf with a companion component also at least partially degenerated. It is suggested that close double-degenerate binaries are possible progenitor populations of Type Ia supernovae and among detectable sources of gravitational wave radiation.
The AM CVn stars are helium-rich binaries with their orbital periods ranging from 5 to 65 min. Almost all of them do not show traces of hydrogen in their spectra. They evolve through one or two common envelope (CE) events. Therefore, the AM CVn family is important for our understanding of the binary formation and the CE phase [for review see: @2010PASP..122.1133S].
Among the AM CVn stars there are two systems with the shortest known orbital periods: HM Cnc [324s, @2010ApJ...711L.138R] and V407 Vul . The exact nature of these objects is unknown. ES Ceti shows the 620s orbital period, which is only 51s greater than the one of V407 Vul. Hence, it is the perfect subject for a timing study based on photometric observations [@2011MNRAS.413.3068C] and preliminary analysis of the disk structure (this work).
Averaged and Trailed Spectra
============================
![The mean spectrum of ES Ceti taken with Magellan on 27-28 Oct. 2002. All lines can be identified with HeII, HeI or NIII.[]{data-label="fig:ESCet_Average"}](Fig1.eps){width="75.00000%"}
The averaged Magellan spectrum of ES Ceti is shown in Fig.\[fig:ESCet\_Average\]. The spectrum is dominated by ionized helium emission lines. Lines of neutral helium and nitrogen NIII are also present.
![The phase-folded, continuum-normalised and subtracted trailed spectra of ES Ceti.[]{data-label="fig:ESCet_Trail"}](4100.eps){width="\textwidth"}
![The phase-folded, continuum-normalised and subtracted trailed spectra of ES Ceti.[]{data-label="fig:ESCet_Trail"}](4339.eps){width="\textwidth"}
![The phase-folded, continuum-normalised and subtracted trailed spectra of ES Ceti.[]{data-label="fig:ESCet_Trail"}](4471.eps){width="\textwidth"}
![The phase-folded, continuum-normalised and subtracted trailed spectra of ES Ceti.[]{data-label="fig:ESCet_Trail"}](4541.eps){width="\textwidth"}
![The phase-folded, continuum-normalised and subtracted trailed spectra of ES Ceti.[]{data-label="fig:ESCet_Trail"}](4686.eps){width="\textwidth"}
![The phase-folded, continuum-normalised and subtracted trailed spectra of ES Ceti.[]{data-label="fig:ESCet_Trail"}](4860.eps){width="\textwidth"}
![The phase-folded, continuum-normalised and subtracted trailed spectra of ES Ceti.[]{data-label="fig:ESCet_Trail"}](5411.eps){width="\textwidth"}
Panels of Fig.\[fig:ESCet\_Trail\] show the time-resolved (trailed spectra) of ES Ceti, folded on the ephemeris given by [@2011MNRAS.413.3068C]. These data are the first spectroscopic observations confirming that the observed period is indeed orbital. However, the variations seen in trailed spectra are not classical, they are very clear and “S-wave”-like. Worth noting is their double-peaked structure, e.g. in the HeII 5411 line.
Summary and Future Work
=======================
Based on the presented spectroscopic data, we concluded that the accretion in ES Ceti is via a disk, and we excluded the direct-impact scenario proposed by [@2005PASP..117..189E]. Probably, the hot gas produces a “disk-like” signature similar to the one observed in HM Cnc system [@2010ApJ...711L.138R].
The strong emission lines are among of the hallmarks of an accreting binaries. Hence, based on the trailed spectra of ES Ceti, we plan to create the equivalent Doppler maps. The method of Doppler tomography [see: @1988MNRAS.235..269M] is a perfect tool which allows to track asymmetric structures in accretion disks and reveals details of the gas flow in a variety of systems.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Bei dieser Seminarausarbeitung handelt es sich um eine Ausarbeitung, die für einen 3D-Modellierungskurs (Masterstudiengang Angewandte Informatik) an der Universität Duisburg-Essen im Jahr 2011 geschrieben wurde. Ich veröffentliche diese Ausarbeitung, damit interessierte Studierende oder generell an Bildverarbeitung/Rendering Interessierte sich einen ersten Eindruck über Raycasting erwerben können. Neben der Ausarbeitung wurde ein funktionierender OpenCL Raycaster entwickelt. Ein Video ist unter [@Kopal] verfügbar. Der vollständige Source-Code des Raycasters ist im Google Code Archive [@RechKopal] archiviert. Falls dies nicht mehr der Fall ist, können Interessierte mir auch gerne eine Email schreiben, damit ich den Source-Code zur Verfügung stellen kann.
Die Seminarausarbeitung bietet eine Einführung und einen Überblick über das Thema “Volume Raycasting mit OpenCL”. Es wird gezeigt, wie mittels moderner Grafikprozessoren Volumendatensätze in Echtzeit geladen, angezeigt und manipuliert werden können. Außerdem werden grundlegende Algorithmen und Datenstrukturen, die für dieses Thema notwendig sind, vorgestellt. Es wird gezeigt, wie ein rudimentärer Raycaster mittels OpenCL aufgebaut werden kann. Desweiteren werden verschiedene Gradientenoperatoren (CentralDifference, Sobel3D und Zucker-Hummel) vorgestellt, implementiert und evaluiert. Abschließend werden noch Beschleunigungsmöglichkeiten für das Raycasting vorgestellt.
author:
-
bibliography:
- 'IEEEabrv.bib'
- 'references.bib'
title: Volume Raycasting mit OpenCL
---
Einleitung
==========
Volume Raycasting findet heutzutage besonders in der medizinischen Informatik aber auch im Bereich des Entertainments, hier in Computerspielen, seinen Einsatz. Beim Raycasting werden mittels Strahlenverfolgung Volumendatensätze abgetastet und aus diesen Abtastdaten ein Bild errechnet. Diese Volumendaten enthalten z.B. mittels 3D-Scanner erfasste Objekte. Im medizinischen Bereich werden mit Hilfe von Raycasting mittels Computertomographie, Magnetresonanztomographie oder ähnlichen erzeugte Volumendaten, visualisiert. In Computerspielen werden Raycastingverfahren genutzt, um realistischere Rauch-, Nebel- und Feuereffekte zu erzeugen. In dieser Ausarbeitung wird näher auf die medizinische Anwendung, also das Visualisieren von medizinischen Daten eingegangen. So werden sowohl der berühmte “Stanford Bunny” als auch drei medizinische Volumendatensätze, die jeweils mittels Computertomographie/Magnetresonanztomographie erstellt wurden, durch einen auf der GPU (*Graphics Processing Unit* ) ausgeführten Raycaster visualisiert.
Medizinische Geräte zur Untersuchung menschlichen Gewebes, wie z.B. Computertomographen oder Magnetresonanztomographen, liefern uns eine schichtweise Darstellung des menschlichen Körpers. So erzeugen solche Geräte eine Vielzahl von Schnittbildern, die aneinandergereiht, einen vollständigen menschlichen Körper modellieren. Weist man den einzelnen Messpunkten innerhalb dieser “Bilder” Graustufen oder Farbwerte hinzu, so erhält man eine relativ genaue Visualisierung, jedoch nur in Schichten. Diese Schichten lassen sich vorwärts und rückwärts durchlaufen, bieten aber kein Gesamtbild des Körpers oder einzelner Teile. Nutzt man diese Daten jedoch als Gesamtes, so ermöglicht das Raycasting eine dreidimensionale originalgetreue Abbildung des gesamten Körpers oder, wenn notwendig, sogar nur einzelner Bereiche wie Organe, Knochen etc. Medizinischem Personal ist es so möglich, eine 360 Grad Ansicht der abgescannten Person einzusehen, “unter die Haut” zu dringen um so medizinische Maßnahmen ohne direkten Eingriff am Menschen zu planen.
Dafür notwendige Raycasting-Verfahren sind relativ Prozessorlastig, jedoch einfach zu parallelisieren. Das ermöglicht es Informatikern die Algorithmen auf moderne Grafikkarten zu portieren um so 3D-Anwendungen zu Erstellen, die den menschlichen Körper in Echtzeit darstellen.
Grafikkarten waren in der Vergangenheit zunächst als Entlastung für die CPU (*Central Processing Unit*) des Computers gedacht. Sie boten fest eingebaute Funktionen um zweidimensionale, gerasterte Grafiken zu manipulieren. Im Zuge der Entwicklung von 3D-Grafiken verfügten Grafikkarten mehr und mehr über Funktionen um auch diese zu Erzeugen. So dienen Grafikkarten dazu, den Prozessor bei der Projektion und Texturierung sowie Beleuchtung von 3D-Szenen zu unterstützen oder ihm diese Arbeit sogar gänzlich abzunehmen. Im Laufe dieser Entwicklung verfügten Grafikkarten über immer mehr Operationen welche in Standards wie OpenGL oder DirectX zusammengefasst wurden. Grafikkarten trugen zu einer deutlichen Steigerung der Rendergeschwindigkeit bei und ermöglichten somit immer umfangreichere und komplexere Szenen.
Bald schon ermöglichten sogenannte *Shader* die Manipulation der Grafikpipeline in bestimmten Abschnitten. Bei einem Shader handelt es sich um ein kleines Programm das direkt in der Grafikkarte ausgeführt wird. Dies ermöglicht dem Entwickler den Renderingprozess für seine Anforderungen individuell anzupassen.
Der neueste Trend bei den Grafikkarten ist heute die vollständige Programmierung dieser. So ist es nicht mehr nur möglich, einzelne Schritte der Renderpipeline anzupassen, sondern mittels Grafikprogrammierung die komplette Grafikkarte zu steuern. Hier sind vor allem NVidias CUDA und der offene Standard OpenCL zu nennen, die diese direkte Programmierung ermöglichen. Mittlerweile ermöglicht dies auch Algorithmen auf Grafikkarten zu portieren, die eigentlich mit Grafikprogrammierung überhaupt nichts mehr zu tun haben (wie z.B. das Brechen von Verschlüsselungen mit Hilfe der Grafikkarte). Nun können aber auch Raytracing- und Raycasting-Verfahren vollständig in einer dieser Sprachen umgesetzt und vollständig von der Grafikkarte berechnet werden. Da Grafikkarten für Berechnungen dieser Art hoch optimiert sind erreichen diese Algorithmen auf der Grafikkarte bisher nie dagewesene Rechengeschwindigkeiten und ermöglichen sogar deren Darstellung in Echtzeit.
Der Rest der Ausarbeitung ist wie folgt aufgebaut: Im Grundlagen Kapitel werden zunächst Volumendatensätze vorgestellt und es wird gezeigt, wie diese generiert und verarbeitet werden können. Die Grundlagen und die Unterschiede zwischen Raycasting und Raytracing, welche für die Anzeige der Volumendatensätze genutzt werden, werden in diesem Kapitel ebenfalls vorgestellt. Abschließend wird die Programmierplattform OpenCL (*Open Computing Language*), mit der man z.B. auf NVIDIA Grafikkarten programmieren kann, kurz vorgestellt. Danach wird ein rudimentärer Raycaster vorgestellt, der im Zuge dieser Ausarbeitung entwickelt worden. Dieser wird daraufhin kurz evaluiert. Hier werden die verschiedenen implementierten Gradientenoperatoren bezüglich ihrer Leistung (in Bildern pro Sekunde) evaluiert. Dann werden verschiedene Beschleunigungsmöglichkeiten für das Raycasting vorgestellt. Den Abschluss dieser Ausarbeitung bildet eine Zusammenfassung sowie ein kurzer Ausblick in mögliche Folgearbeiten.
Grundlagen
==========
Dieses Kapitel bietet eine kleine Einführung in die Grundlagen des Raycastings. So werden zunächst Volumendatensätze erläutert. Im zweiten Teil wird kurz auf Raytracing und Raycasting und deren Unterschiede eingegangen. Abschließend wird auf den von der Khronos Group entwickelten Standard OpenCL eingegangen, welcher benutzt werden kann, um Programme direkt für die Grafikkarte zu entwickeln.
![Schematische Darstellung eines Volumendatensatzes[]{data-label="label:Volumendatensatz"}](images/volumen.pdf){width="0.75\columnwidth"}
Volumendatensätze
-----------------
Ein Volumendatensatz ist eine dreidimensionale Funktion, die jedem $(x,y,z)$-Tupel einen Funktionswert zuordnet:
$$V: \mathrm{N}^3 \rightarrow \mathrm{R}$$
Man stelle sich einen Quader vor, der aus senkrecht und waagerecht angeordneten Gitternetzpunkten besteht. Jeder Gitterpunkt beinhaltet genau einen Funktionswert. Diese Punkte werden auch als Voxel bezeichnet, wobei sich der Name aus Volumen und Pixel zusammensetzt. Da Volumendatensätze nur aus diesen Punkten bestehen, müssen Funktionswerte, die zwischen mehreren Punkten liegen, interpoliert werden. Dazu wird lineare oder trilinieare Interpolation zwischen zwei oder acht Gitterpunkten verwendet. Gerade im medizinischen Umfeld werden Volumendatensätze zur Visualisierung des menschlichen Körpers genutzt. So wird zunächst eine Aufnahme eines Körpers mittels eines Computertomographen, eines Magnetresonanztomographen oder ähnlichen Geräten erstellt. Diese Geräte speichern ihre gewonnen Daten innerhalb eines Volumendatensatzes ab (im einfachsten Fall als ein dreidimensionales Array von Abtastwerten). Diese Datensätze können dann später am Computer geladen und in einer dreidimensionale Abbildung dargestellt werden. Von derlei Geräten erzeugte Volumendatensätze sind häufig sehr groß. So ist ein würfelförmiger Datensatz mit einer Kantenlänge von 512 Punkten, wobei jedes Datum in einem Short (2 byte) gespeichert würde, bereits 256 Megabyte groß ($512^3 * 2 $ byte $ = 256 $ Megabyte).
Raytracing und Raycasting
-------------------------
Raytracing ist ein Verfahren der Computergrafik, mit dem sehr realistische Computergrafiken von dreidimensionalen Szenen gerendert werden können. Beim Raytracing werden, ausgegangen vom Betrachter einer Szene, für jeden Bildpunkt Strahlen, die sogenannten Rays, losgeschickt. Trifft ein Strahl auf ein Objekt, dass sich in der Szene befindet, so werden Oberflächeneigenschaften, Beleuchtung der Szene und gegebenenfalls Reflektionen und Refraktionen für den Schnittpunkt berechnet. Aus diesen Informationen lässt sich dem Bildpunkt ein Farbwert zuweisen, der dann dem Betrachter angezeigt wird. Treffen Strahlen auf reflektierende oder refraktierende Oberflächen, werden Schnittwinkel bestimmt und Folgestrahlen ausgesandt, um festzustellen, ob Informationen weiterer Objekte mit in den aktuellen Bildpunkt eingerechnet werden müssen. Um dies zu vereinfachen, sind Raytracer häufig mittels rekursiver Algorithmen implementiert. Da ein einfaches Bild mit einer Auflösung von 640 \* 480 Bildpunkten bereits 307.200 Strahlen vom Betrachter aussendet, sind Raytracer in ihrer Berechnungsgeschwindigkeit häufig sehr langsam. Die Hauptrechenlast bei Raytracern wird durch die Schnittpunktberechnungen zwischen Strahlen und in der Szene angeordneten Objekten erzeugt. So muss im allgemeinen Fall, ohne jede Optimierung, für alle Strahlen ein möglicher Schnittpunkt mit jedem Objekt der Szene berechnet werden. Ist ein Schnittpunkt gefunden, wird die Beleuchtung des Punktes, z.B. mittels Phong Shading berechnet. Hierfür berechnet man auch den Normalenvektor im Schnittpunkt der getroffenen Oberfläche und nutzt diesen um besagtes Phong Shading durchzuführen.
Im Gegensatz zum Raytracing werden beim Raycasting Strahlen nicht gegen Objekte, wie z.B. Kugeln, Würfel oder Ebenen geleitet. Beim Raycasting werden Strahlen auf einen der oben beschriebenen Volumendatensätze geleitet. Trifft ein solcher Strahl auf einen Voxel, so wird aus dem Schnittpunkt, dem Auftreffwinkel und dem Wert an entsprechende Volumenposition ein Farbwert bestimmt, der dem Betrachter dann angezeigt wird. Im Gegensatz zum Raycasting wird eine einfache Beleuchtung, bestehend aus nur einer Lichtquelle, genutzt. Außerdem werden keine Reflektions- und auch keine Refraktionsstrahlen ausgesandt. Häufig werden auch Strahlen, nach dem Aufprall auf ein Voxel, weiter innerhalb des Volumendatensatz verfolgt. Die “tiefer” liegenden Voxel werden so auch mittels eines Emissions-Absorptions-Modell in den Farbwert mit eingerechnet. Dies ermöglicht dem Betrachter, z.B. in medizinischen Anwendungen, auch innerhalb eines Körpers liegende Regionen des Volumendatensatzes, z.B. Adern oder Knochen, die unter der Haut liegen, zu betrachten. Hierfür werden physikalische Grundlagen genutzt. So wird die Energie des Lichtes auf seinem Weg durch den Volumendatensatz “immer mehr aufgebraucht” und das Licht schließlich von Knochen vollständig absorbiert. So kann man durch dünne Hautschichten hindurch sehen, Knochen stoppen die Strahlen jedoch.
Da ein Voxel genau genommen keine räumliche Ausdehnung besitzt und auch keine Oberfläche, ist die Bestimmung des Auftreffwinkels eines Strahles nicht direkt möglich. Abhilfe schaffen hier Bildableitungsfunktionen die den Gradienten des Voxels bestimmen. Im nächsten Kapitel wird auf diese Gradientenfunktionen näher eingegangen.
Der große Vorteil sowohl von Raytracing als auch Raycasting sind ihre einfache parallelisierbarkeit. Da für jeden Bildpunkt ein eigener Strahl verfolgt wird, der absolut unabhängig von denen andere Bildpunkte ist, kann man jeden Strahl separat berechnen. So lässt sich ein Raytracer/Raycaster schon auf einem Multiprozessor-System deutlich beschleunigen, indem man die Berechnungen auf mehrere der Prozessoren des Systems verteilt. So halbiert sich die Berechnungsgeschwindigkeit bereits beim hin zuschalten eines zweiten Prozessors. Moderne Grafikkarten verfügen Heutzutage zwischen 16 (Geforce 8400 GS) und 128 (Geforce 8800 Ultra) und sogar bis zu 480 (Geforce GTX 480) sogenannter Stream Prozessoren. So können parallele Algorithmen massiv parallel ausgeführt werden. Im Fall von Raytracing/Raycasting werden nun Bildwiederholraten durch Portierung der Algorithmen (mittels CUDA oder OpenCL) auf die Grafikkarte von mehren Bildern pro Sekunde möglich.
OpenCL
------
OpenCL, auch Open Computing Language, ist ein von der Khronos Group entwickelter Standard. Mit diesem und der Sprache OpenCL C können Programme sowohl für CPUs als auch für GPUs (Grafikprozessoren) entwickelt werden. OpenCL Programme (auch Kernels genannt) können während der Ausführung auf mehrere OpenCL-fähige Geräte verteilt werden. Ein OpenCL-System besteht immer aus zwei oder mehreren Komponenten. Zum einem dem sogenannten Host-Programm und aus mehreren OpenCL fähigen Devices. Ein Device kann z.B. eine CPU oder eine GPU darstellen. Die Devices wiederum bestehen aus einer oder mehreren Computing Units. Bei CPUs sind dies die Kerne und bei Grafikkarten die Streamprozessoren. Zur Laufzeit werden die Kernels (OpenCL Programme) auf diese Streamprozessoren durch den Host verteilt und können somit parallel abgearbeitet werden. OpenCL Kernel werden erst zur Laufzeit in ausführbaren Code übersetzt. Somit kann ein in OpenCL C geschriebenes Programm auf unterschiedlichen Zielplattformen ausgeführt werden und es muss zur Entwicklungszeit nicht feststehen, was die Zielplattform ist. So übersetzt der NVidia Compiler das OpenCL C Programm zunächst in CUDA-Code um es dann mittels CUDA auszuführen. OpenCL wurde für die Entwicklung von massiv parallelen Anwendungen konzipiert. So kann z.B. ein Computerbild der Auflösung 1024 x 1024 komplett parallel verarbeitet werden:
{Klassisches C Programm}
void trad_mul(int n,
const float *a,
const float *b,
float *c)
{
int i;
for(i = 0; i < n; i++)
{
c[i] = a[i] * b[i];
}// klassische schleife
}
{OpenCL C Programm}
kernel void trad_mul(global const float *a,
global const float *b,
global float *c)
{
int id = get_global_id(0);
c[id] = a[id] * b[id];
} // fuehre n work items parallel aus
Wie in den Codebeispielen \[label:klassischC\] und \[label:openCL\] (entnommen aus [@openclintro]) zu erkennen, kann man klassische Schleifen relativ einfach in einen OpenCL Kernel umschreiben um diesen dann parallel auszuführen. Im Codebeispiel \[label:klassischC\] werden zwei Computerbilder multipliziert und das Ergebnis in einem dritten Bild gespeichert. Dies geschieht mit Hilfe einer einfachen for-Schleife die alle Operationen sequenziell abarbeitet. Im Codebeispiel \[label:openCL\] wird das selbe Ergebnis mit Hilfe eines OpenCL Kernels erzielt. Mittels des Aufrufs *get\_global\_id(0);* erhält das aktuell ausgeführte Work Unit seine ID und kann daraus die Position in den Bildarrays errechnen, die es zu bearbeiten hat. Im Gegensatz zum klassischen C-Programm kann der OpenCL Kernel von einer Vielzahl von Work Units parallel abgearbeitet werden.\
\
OpenCL verfügt über eine Vielzahl von Datentypen und Funktionen, die Grafikprogrammierung unterstützen und vereinfachen. So existieren die Datentypen *float2, float3, float4, float8* und *float16* im OpenCL Standard. Diese entsprechen n-dimensionalen Vektoren deren Komponenten *float* Werte beinhalten. So kann eine Variable x des Typs *float3* z.B. über seine Komponenten *x.S0*, *x.S1* und *x.S2* angesprochen werden. Arithmetische Operationen auf Vektor-Datentypen werden von den GPUs unterstützt und entsprechend schnell berechnet. So ist eine Addition zweier *float4*- Werte auf einer GPU ebenso schnell oder gar schneller wie die Addition zweier reiner *float*-Werte mittels einer CPU. OpenCL liefert, wie bereits erwähnt, auch mehrere nützliche Funktionen für die Grafikprogrammierung. So kann mittels *fast\_normalize(x)* ein *float-n*-Wert schnell normalisiert werden. Ebenso werden Funktionen zur Berechnung des Skalarprodukts (*dot(v1,v2)*) oder des Kreuzprodukts (*cross(v1,v2)*) zweier Vektoren mitgeliefert, welche durch die GPU ebenfalls besonders schnell ausgeführt werden können.
Entwurf eines Raycasters mit OpenCL
===================================
In diesem Kapitel wird ein einfacher Raycaster mit OpenCL und .NET C\# (Hostprogramm) entworfen. Zu diesem Raycaster gehört unter anderem das Laden und Anzeigen bereitgestellter Volumendatensätze, die durch einen Computertomographen erstellt wurden. Diese Volumendatensätze sollen im Raum positioniert und mittels einer virtuellen, positionierbaren Kamera dargestellt werden. Der Raycaster ermöglicht das Manipulieren dieser Datensätze. Darunter fällt das Vergrößern und Verkleinern des Bildausschnitts. Außerdem wird es möglich sein, den Datensatz um eine Achse zu Drehen, um ihn so von mehreren Seiten betrachten zu können. Die Lichtquelle, die den Volumendatensatz erhellt, wird im Raum bewegt werden können. Außerdem kann man die sogenannte Bounding-Box um den Datensatz einstellen um diesen sozusagen zu “zerschneiden”. Damit das Volumen auch auf schwächeren Rechnern dargestellt werden kann, wird die Render-Qualität (die Auflösung und die sogenannte “Gradientenfuntkion”) einstellbar sein. Der Raycaster wird nur die Oberflächen der sich innerhalb des Volumen befindlichen Objektes abtasten, jedoch nicht tiefer in das “Material” eindringen. Die Implementierung des hier vorgestellten Raycasters kann von [@RechKopal] heruntergeladen werden.
Hostprogramm
------------
Als Hostprogramm für den Raycaster dient ein .NET C\# Programm das die Volumendatensätze einließt und in den Grafikkartenspeicher kopiert. Die zur Verfügung stehenden Volumendatensätze bestehen aus n Binärdateien, die 512 X 512 große Rohdaten von Messwerten enthalten, die ein Computertomograph erzeugt hat. Diese werden zunächst in ein dreidimensionales Array und dann mit Hilfe der OpenCL API in die Grafikkarte kopiert. Sobald das Volumen vollständig geladen ist, werden mehrere Instanzen eines OpenCL Kernels erstellt. Diese arbeiten jeder für sich genau einen Strahl in das Volumen ab und speichern ihre Ausgabe in einem großen Bildbuffer der dann final von dem C\#-Programm wieder angezeigt wird.
Laden eines Volumendatensatzes
------------------------------
Die in diesem Raycaster verwendeten Volumendatensätze stammen aus dem [@TS3DSR] “Stanford 3D Scanning Repository”. Bei den Datensätzen handelt es sich zum einem um einen CT-Scan des berühmten “Stanford Bunny”. Dieser Porzellanhase dient in der 3D Grafikszene einer Vielzahl von Wissenschaftlern als Datenvorlage. In diesem Raycaster stellt er den ersten von drei Volumendatensätzen dar. Der “Hase” besteht aus 360 Binärdateien, die jeweils ein 512 X 512 großes “Bild”, eine Scheibe, des Hasen darstellen. Im Raycaster wird ein kubischer Volumendatensatz der Kantenlänge 512 erstellt in dem die Scheiben an die Koordinaten $Z=0$ bis $Z=360$ geladen werden. Jeder Voxel besteht aus einem 16bit breiten short-Value, der jedoch nur 12bit Daten enthält. Nur 12bit aus dem Grund, dass CT-Scanner Daten, basierend auf der Hounsfield-Skala, liefern. Für diese Skala reichen 12bit vollkommen aus. Neben dem Bunny ermöglicht der Raycaster das Laden und Anzeigen eines CT-Scans eines menschlichen Kopfes sowie eines CT-Scans eines weiteren Kopfes, dem die Schädeldecke medizinisch entfernt wurde, damit das Gehirn frei liegt. Der letzte Volumendatensatz ist ein menschlicher Oberkörper, bei dem sowohl die Knochen als auch die Organe “freigelegt” werden können.
![Volumenschnittpunkte/Abtastung[]{data-label="label:VolumenschnittpunktAbtastung"}](images/Abtastung_Volumen.pdf){width="0.75\columnwidth"}
Schnittpunktbestimmung/Abtastung im Volumen
-------------------------------------------
Zunächst wird der Schnittpunkt des zu berechnenden Strahls und einer Bounding Box berechnet. Die Bounding Box umschließt den Volumendatensatz. Gibt es keinen Schnittpunkt, so wird als Farbe Schwarz ausgeben. Existieren zwei Schnittpunkte, so wird zwischen diesen beiden Punkten entlang des Strahls “abgetastet”. Der Abstand zwischen den Abtastpunkten spiegelt sich später sowohl in der Qualität das erzeugten Bildes als auch in der Berechnungsgeschwindkeit wieder. Da Volumendatensätze aus einzelnen Punkten bestehen und die Wahrscheinlichkeit, dass ein Strahl exakt auf einen Punkt trifft, sehr gering ist, müssen die Werte zwischen mehreren Punkten interpoliert werden. Als einfachstes Verfahren (aber auch qualitativ nicht sehr hochwertig) lassen sich die Koordinaten auf ganzzahlige Werte Runden. Alternativ kann man auch zwischen zwei Punkten linear oder zwischen 8 Punkten trilinear interpolieren. Dies liefert qualitativ höherwertige Bilder.
In Abbildung \[label:VolumenschnittpunktAbtastung\] ist eine Abtastung eines Volumendatensatzes schematisch dargestellt. Der schwarze Pfeil stellt den Strahl dar, der durch das Volumen gecastet wird. Zwischen dem Eintrittspunkt und dem Austrittspunkt des Strahls und der Bounding-Box (hier in rot dargestellt) werden in regelmäßigen Abständen Punkte abgetastet (hier in grün dargestellt). Hier wird auch deutlich, warum zwischen den eigentlichen Voxeln (blaue Punkte) weitere Funktionswerte interpoliert werden müssen, da der Strahl genau genommen keinen der Voxel genau trifft.
![Trilineare Interpolation[]{data-label="label:TrilineareInterpolation"}](images/Lineare_Interpolation.pdf){width="0.75\columnwidth"}
![Trilineare Interpolation[]{data-label="label:TrilineareInterpolation"}](images/Trilineare_Interpolation.pdf){width="0.75\columnwidth"}
Lineare und trilineare Interpolation
------------------------------------
Um ein Datum aus dem Volumensatz zu berechnen, das zwischen einzelnen Voxeln liegt, wird lineare und trilineare Interpolation genutzt. Bei der linearen Interpolation wird eine einfache Gerade $g$ zwischen zwei Funktionswerte $f(x0)$ und $f(x1)$ gelegt um Zwischenwerte zu berechnen. So kann mittels Einsetzen eines Wertes in die Geradengleichung ein Zwischenwert interpoliert werden. Mittels der Formel
$$g(x) = f(x0) + \frac{f(x1)-f(x0)}{x1 - x0} ( x - x0)$$
kann die Geradengleichung bestimmt werden. In Abbildung \[label:LineareInterpolation\] ist das Verfahren als Skizze verdeutlicht. Im dreidimensionalen wird die trilineare Interpolation bei den Voxeln angewandt. Hierfür werden die diskreten Werte von acht Voxeln in Zweierpaaren linear interpoliert. In Abbildung \[label:TrilineareInterpolation\] werden zunächst die Paare $(p1,p2), (p3,p4), (p5,p6), (p7,p8)$ mittels linearer Interpolation miteinander kombiniert. Anschließend werden deren Ergebnispaare $(I1, I3), (I2, I4)$ nochmals linear interpoliert. Abschließend erhält man den gesuchten Funktionswert $I7$ nach dritter linearer Interpolation des Paares $(I5,I6)$.
Gradientenfunktion
------------------
Hat der Strahl im Volumen einen Punkt erreicht, dessen Wert ungleich 0 (= leerer Raum) ist, so kann diesem Punkt ein Farbwert zugewiesen werden (Shading). Da für die Beleuchtung im dreidimensionalen Raum (mittels Phong-Shading) allerdings ein Normalenvektor notwendig ist, muss dieser zunächst berechnet werden. Um aus einem Volumendatensatz Normalenvektoren zu berechnen bieten sich, aus der klassischen Bildverarbeitung bekannte, Verfahren an. Anstelle des Normalenvektors nutzt man die Ableitung der Bildfunktion bzw in unserem Fall der Volumen-Funktion, was dem Gradienten entspricht. Hierfür nutzt man einfache oder auch komplexere Gradientenfunktionen die sich der Bildableitung relativ genau annähern. Als sehr einfache Gradientenfunktion ist der *Central-Difference-Operator* zu gebrauchen. Dieser bestimmt die lokalen Ableitungen in X, in Y und in Z-Richtung indem er folgende Rechnungen vornimmt:
$$\begin{split}
\delta_X = V(P.X + 1,P.Y,P.Z) - V(P.X - 1,P.Y,P.Z) \\
\delta_Y = V(P.X,P.Y + 1,P.Z) - V(P.X,P.Y - 1,P.Z) \\
\delta_Z = V(P.X,P.Y,P.Z + 1) - V(P.X,P.Y,P.Z - 1) \\
G = normalize( (\delta_X, \delta_Y, \delta_Z) )
\end{split}$$
![Beispielrechnung Central-Difference-Operator[]{data-label="label:gradient"}](images/gradient.png){width="0.5\columnwidth"}
Zunächst berechnet man die jeweiligen Funktionswerte des Volumens $V(x,y,z)$ in allen drei Achsenrichtungen vor dem Punkt P minus den entsprechenden Funktionswerten in gleicher Achsenrichtung nach dem Punkt P. Der neu berechnete Vektor $(\delta_X, \delta_Y, \delta_Z)$ wird abschließend noch normalisiert (durch seinen Betrag geteilt) und ergibt so den Gradienten. Der Gradient zeigt somit immer in die Richtung, in der die größten Funktionswerte stehen. Als Beispiel im zweidimensionalen ist die Abbildung \[label:gradient\] beigefügt, die eine Gradientenbestimmung in X und in Y-Richtung zeigt.\
\
Mit Hilfe des *Central-Difference-Operator* lassen sich relativ einfach und vor allem relativ schnell Gradienten im Volumen bestimmen. Um qualitativ bessere Ergebnisse zu erhalten bietet sich bessere Operatoren wie z.B. der *Zucker-Hummel-Operator* oder der *Sobel 3D Operator* an. Sowohl der Zucker-Hummel als auch der Sobel3D-Operator nutzen für die Gradientenberechnung, im Gegensatz zum Central-Difference-Operator nicht nur 6 Voxel sondern insgesamt 54 Voxel. Dadurch fallen die Gradienten deutlich weicher und klarer aus als dies beim Central-Difference-Operator der Fall ist. Der qualitative Unterschied zwischen dem *Central-Difference-Operator* und dem *Zucker-Hummel-Operator* wird in den Abbildungen \[label:kadaver\_central\_difference\] und \[label:kadaver\_zucker\_hummel\] deutlich.\
\
![Kopf Zucker-Hummel-Operator[]{data-label="label:kadaver_zucker_hummel"}](images/kadaver_central_difference.png){width="0.5\columnwidth"}
![Kopf Zucker-Hummel-Operator[]{data-label="label:kadaver_zucker_hummel"}](images/kadaver_zucker_hummel.png){width="0.5\columnwidth"}
Als komplexeres Beispiel für einen Gradientenoperator ist in Formel \[Sobel3D\] der Sobel3D-Operator mittels folgender Operatormasken dargestellt (entnommen von [@S01Project]).
$$\begin{split}
x-1:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
-1 & -3 & -1\\
-3 & -6 & -3\\
-1 & -3 & -1
\end{bmatrix}
x:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
\ 0 & \ 0 & \ 0\\
\ 0 & \ 0 & \ 0\\
\ 0 & \ 0 & \ 0
\end{bmatrix}
x+1:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
\ 1 & \ 3 & \ 1\\
\ 3 & \ 6 & \ 3\\
\ 1 & \ 3 & \ 1
\end{bmatrix}
\\
y-1:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
\ 1 & \ 3 & \ 1\\
\ 0 & \ 0 & \ 0\\
-1 & -3 & -1
\end{bmatrix}
y:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
\ 3 & \ 6 & \ 3\\
\ 0 & \ 0 & \ 0\\
-3 & -6 & -3
\end{bmatrix}
y+1:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
\ 1 & \ 3 & \ 1\\
\ 0 & \ 0 & \ 0\\
-1 & -3 & -1
\end{bmatrix}
\\
z-1:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
-1 & \ 0 & \ 1\\
-3 & \ 0 & \ 3\\
-1 & \ 0 & \ 1
\end{bmatrix}
z:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
-3 & \ 0 & \ 3\\
-6 & \ 0 & \ 6\\
-3 & \ 0 & \ 3
\end{bmatrix}
z+1:
\begin{bmatrix}
\hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}} & \hphantom{\hspace{20pt}}\\[-15pt]
-1 & \ 0 & \ 1\\
-3 & \ 0 & \ 3\\
-1 & \ 0 & \ 1
\end{bmatrix}
\end{split}
\label{Sobel3D}$$
Der Operator wird ähnlich wie der Central-Difference-Operator angewandt. Die Operatormaske wird jeweils in der X-, in der Y- und in der Z-Ebene entsprechend einen Voxel vor dem Abtastpunkt, genau auf dem Abtastpunkt und einen Voxel nach dem Abtastpunkt angelegt. Die Zahlen in den Operatormasken geben Koeffizienten an, mit denen der Wert des Voxels an der entsprechenden Position multipliziert werden muss. Alle so errechneten Werte einer jeden Maske werden für jede Ebene danach addiert und ergeben so eine Komponente des Gradienten. Da sowohl der Sobel3D-Operator als auch der Zucker-Hummel-Operator deutlich mehr Voxel in die Gradientenberechnung mit einbeziehen, gibt der errechnete Gradient eine deutlich bessere Beschreibung der lokalen Oberflächenstruktur im Volumendatensatz wieder als es der Central-Difference-Operator vermag.
Bounding Box
------------
Wie bereits erwähnt wird eine Bounding Box genutzt, um den Strahleintritt in den Volumendatensatz und den Strahlaustritt aus dem Volumendatensatz zu bestimmen. Dies beschleunigt zum einen die Berechnung, da Strahlen die “am Volumen vorbeigehen” nicht abgetastet werden müssen. Zum anderen ermöglicht die Bounding Box auch das “Zerschneiden” des Volumens. So kann z.B. in einen Datensatz “hineingeschaut” werden. In Abbildung \[label:stanford\_bunny\_aufgeschnitten\] sieht man ein Computertomographie-Bild des berühmten *Stanford Bunny* dessen Bounding Box so gesetzt wurde, dass man in den Volumendatensatz hinein und sogar hindurch schauen kann. Hier kann man erkennen, dass der original Porzellanhase einen Hohlraum beinhaltet. Die scheinbar nicht vorhandene Dicke der “Hülle” das Hasen liegt an der Berechnung der Gradienten. Die Strahlenberechnung beginnt zwar “innerhalb” des Hasenkörpers, jedoch werden für die Gradientenbestimmung auch Daten vor Strahlbeginn genutzt. Somit ist der Gradient bei Strahlbeginn 0 da der Strahl “mitten im Material” beginnt.
![Bounding Box Stanford Bunny[]{data-label="label:stanford_bunny_aufgeschnitten"}](images/stanford_bunny_aufgeschnitten.png){width="0.5\columnwidth"}
Farbbestimmung
--------------
Um den Schnittpunkten nicht nur eine Schattierung (Shading) sondern auch eine natürliche Farbe zu geben müssen den Volumendaten Farbwerte zugeordnet werden. Entweder beinhaltet das Volumen bereits Farbwerte die jedem Voxel mitgegeben werden. Oder, wie im Fall von Computertomographie, beinhalten die Volumendaten allerdings andere Werte, wie z.B. die ortsabhängige Abschwächung der Röntgenstrahlung durch Absorption im Gewebe. Allerdings kann man aus diesen Werten Farbwerte ableiten/zuweisen. So lässt sich mit Hilfe der sogenannten *Hounsfield-Skala* [@hounsefield] einem jeden Voxelpunkt auch ein Farbwert zuweisen:
$$Hounsfield(\mu_{Gewebe}) = \frac{\mu_{Gewebe} - \mu_{Wasser}}{\mu_{Wasser}} * 1000 HU$$
Mit Hilfe dieser Formel kann aus dem Abschwächungskoeffizienten des Gewebes ($\mu_{Gewebe}$) ein Hounsfield-Wert bestimmt werden. Diesen Hounsfield-Werten wiederum lassen sich Farbwerte zuweisen. Wasser besitzt, durch die Normierung in der Formel, einen HU-Wert von 0. Luft besitzt einen Wert von -1000 HU. Gewebe besitzt einen Wert um die -100 HU und Knochen einen zwischen 500 und 1500HU. Hier kann man Gewebe rötliche Töne, Knochen weiße und Luft schwarze Werte zuweisen. Auch ist es dank dieser Skala möglich, bestimmte Gewebetypen (z.B. Krebsgewebe) hervorzuheben (einzufärben) oder ganz auszublenden. Diese Farbskalen werden häufig als Lookup Tabelle implementiert, in der zu einem HU-Wert ein Farbwert ausgelesen werden kann.
Beleuchtung
-----------
Damit die Oberflächen des Volumendatensatzes verschiedene Schattierungen erhalten (Shading), wird ein einfaches Beleuchtungsmodell mit nur einer einzigen Lichtquelle benutzt. Die Lichtquelle wird als einfache Punkt-förmige Quelle modelliert, die durch den Vektor $lightPosition$ beschrieben wird. Um den Beleuchtungswert im Schnittpunkt zwischen Strahl und Oberfläche zu bestimmten wird außerdem der Gradient $norm$ im Schnittpunkt benötigt. Mittels des Skalarprodukts zwischen Normalenvektor im Schnittpunkt (Gradienten) und der normalisierten Differenz aus Lichtposition und Schnittpunkt wird die Stärke der Beleuchtung bestimmt. Im Codebeispiel \[label:shading\] ist ein einfacher Algorithmus für die Beleuchtung in OpenCL dargestellt.
{Volume Shading}
float4 Shade(float4 pos,
float4 norm,
float4 lightPosition)
{
float4 color =
(float4)(1.0f,1.0f,1.0f,0.0f);
//Farbe der Lichtquelle: Weiss
float4 livec =
fast_normalize(lightPosition - pos);
//Differenz aus Lichtposition und Schnittpunkt
float illum =
dot(livec,norm);
//Grad der Beleuchtung; Skalarprodukt
return illum * color;
//Farbe zurueckgeben
}
Indem die Position der Lichtquelle bewegt wird (verändern des Vektors $lightPosition$) lässt sich das zu raycastende Bild unterschiedlich ausleuchten. In Abbildungen \[label:beleuchtung1\] und \[label:beleuchtung2\] ist eine Abbildung eines menschlichen Kopfes mit freigelegten Gehirn mit zwei unterschiedlichen Beleuchtungspositionen ($lightPosition = (-1000,-1000,-1000)$ und $lightPosition = (1000,1000,1000)$) abgebildet.
![Beleuchtetes Gehirn 2[]{data-label="label:beleuchtung2"}](images/beleuchtung1.png){width="0.5\columnwidth"}
![Beleuchtetes Gehirn 2[]{data-label="label:beleuchtung2"}](images/beleuchtung2.png){width="0.5\columnwidth"}
Transparenz-Formel
------------------
[@Pawasauskas] zeigt eine Transparenz-Formel, mit der sich Voxel, die hintereinander liegen und mittels eines Strahls abgetastet werden, so verrechnet werden können, das auch in das Volumen hinein geschaut werden kann:
$$C_out = C_in (1 - \alpha(X_i )) + C(X_i )\alpha(X_i)$$
Wobei $C(x)$: eine Schattierung, errechnet aus dem lokalen Gradienten ist, $\alpha(X)$ die Lichtundurchlässigkeit, berechnet aus dem CT-Wert des Voxels ist, $C_out$ die ausgehende Intensität/Farbe des aktuellen Voxels ist und $C_in$ die eingehende Intensität in den Voxel ist. Diese Werte errechnen sich zum einen aus den Daten der Voxel bzw des Volumendatensatzes und zum anderen aus der Lichtquelle, mit der der Volumendatensatz beleuchtet werden soll. Diese Formel wird rekursiv so lange entlang des Strahls angewandt, bis auf ein Alpha-Wert innerhalb des Volumens getroffen wird, der 0 ist (z.B. durch Treffen auf lichtundurchlässiges Gewebe wie Knochen). Der in dieser Ausarbeitung entwickelte Raycaster implementiert die Transparenz-Formel jedoch nicht und bietet somit nur Bilder ohne Transparenz.
Gesamtaufbau eines Raycasters
-----------------------------
Der grundlegende Raycasting-Algorithmus wird mittels der oben beschriebenen Techniken implementiert und lässt sich wie folgt darstellen:
1. **Raycasting:** Ausgehend vom Auge des Betrachters wird durch jeden Bildpunkt des zu rendernden Bildes ein Strahl geleitet. Der Strahl trifft auf das Volumen, dass abgetastet werden soll und verlässt dieses wieder. Eine BoundingBox kann als einfaches Mittel dienen, um die Berechnungsdauer zu verkleinern. So wird das Volumen innerhalb der Box platziert um so den Startpunkt und Endpunkt der folgenden Abtastung zu bestimmen. So erspart man sich die Abtastung von “leeren” Raum.
2. **Sampling:** Entlang des im ersten Schritt bestimmten Strahls werden Sampling-Punkte gewählt und das Volumen wird “abgetastet”. Da Voxel nicht exakt getroffen werden dienen Techniken wie lineare oder trilineare Interpolation dazu, bessere Abtastergebnisse zu erhalten.
3. **Shading:** Für jeden bestimmten Abtastpunkt wird mittels Gradientenoperator ein Gradient bestimmt. Dieser Gradient gibt die lokale Struktur der Oberfläche des im Volumendatensatz befindlichen Objektes wieder Mittels dieses Gradienten und einer gegebenen Lichtquelle lassen sich die Beleuchtung und Farbe der Punkte ermitteln.
4. **Compositing:** Nachdem alle Abtastpunkte berechnet und ihre Beleuchtung bestimmt wurde lässt sich mittels eines Transmissions-Emmissions-Modells der Einfluss eines jeden Punktes auf den zu berechnenden Bildpunkt bestimmen um so eine resultierende Farbe zu bestimmen. Diese wird letzendlich an die zu berechnende Stelle im Bild gezeichnet.
Implementierung in .NET und OpenCL
----------------------------------
Der für diese Ausarbeitung implementierte Raycaster ist mit Hilfe von Microsoft .NET C\# und OpenCL unter Microsoft Windows implementiert. Ein Video, dass alle Funktionen des Raycasters demonstriert kann unter [@Kopal] betrachtet werden. Das Host Programm besteht zum einen aus vier Laderoutinen, die vier unterschiedliche Volumendatensätze in den Speicher der Grafikkarte laden können (Stanford Bunny, Head, Brain, Stent).\
\
Der Benutzer hat jederzeit die Möglichkeit, den darzustellenden Volumensatz zu ändern aber auch die Darstellung selbst zu manipulieren. Dafür gibt es auf der linken Seite der Applikation ein Settings-Fenster. Hier kann aber auch die BoundingBox abgeändert werden, um so Einblicke innerhalb des Volumens zu erhalten. Neben der BoundingBox können auch zwei Schwellenwerte festgelegt werden, die angeben, von welchem Startwert bis zu welchem Zielwert der Hounsefield Skala Voxel überhaupt dargestellt werden sollen. So kann z.B. beim “Head” das Fleisch entfernt werden, um so den freigelegten Schädel darzustellen.\
\
![Implementierung eines Raycasters in .NET und OpenCL[]{data-label="label:Implementierung"}](images/raycaster_implementierung.jpg){width="0.8\columnwidth"}
\
Mittels weiterer Regler kann der Beleuchtungsvektor verändert werden. Er gibt die Position der einzigen Lichtquelle an, welche den Volumendatensatz ausleuchtet.\
\
Um auch auf langsameren Grafikkarten noch annehmbare Resultate zu erhalten, ermöglicht die Implementierung ein Umschalten der Render-Auflösung (512X384, 640X480, 800X600, 1024X768) und ein Umschalten des Gradienten Operators (Central-Difference-Operator, Zucker-Hummel-Operator, Sobel3D-Operator).\
\
Mit Hilfe der Maus sowie der Pfeiltasten der Tastatur lässt sich das dargestellte Model drehen sowie vergrößern und verkleinern (Zoomen).\
\
In der Kopfleiste der Applikation lassen sich die aktuelle Bildrate (Frames per Second) sowie die Berechnungsdauer eines einzelnen Bildes nachvollziehen.\
\
Der Raycaster ist, bis auf die Nutzung der BoundingBox, wenig optimiert worden. Die Abtastung innerhalb des Volumendatensatzes trickst ein wenig, indem sie zunächst in großen Schritten entlang des Strahls vorwärts abtastet, bis sie auf “feste Materie” trifft. Um den Eintrittspunkt genauer zu bestimmen wird ab diesem Punkt rückwärts in kleineren Schritten abgetastet, um so den Eintrittspunkt in die feste Materie zu bestimmen. Obwohl der Raycaster wenig optimiert ist, liefern selbst ältere Computer ( hier getestet mit einem Intel Core 2 Duo 6420 mit 2,1GHz und einer NVidia GForce 8800 GT ) annehmbare Bildraten im Bereich zwischen fünf und 50 Bildern pro Sekunde. Im nachfolgendem Kapitel sind einige Messungen aufgeführt, die die Performance des Raycasters darstellen und einige der implementierten Gradientenoperatoren miteinander vergleichen. Weitere Optimierungsmöglichkeiten für einen Raycaster zeigt das Kapitel “Beschleunigungsmöglichkeiten für Raycasting”.
Raycaster Evaluation
====================
In diesem Kapitel wird der in dieser Ausarbeitung entwickelte Raycaster bezüglich seiner Rendergeschwindigkeit untersucht. Hierzu wurde gemessen, wie viele Bilder pro Sekunde mittels des Raycasters gerendert werden können. Vergleichend sind nachfolgend unterschiedliche Auflösungen sowie unterschiedliche Gradientenoperatoren dargestellt.
**512X384** **640X480** **800X600** **1024X768**
----------- ------------- ------------- ------------- --------------
**Bunny** 20,00 fps 16,13 fps 12,65 fps 8,92 fps
**Head** 27,02 fps 24,39 fps 20,40 fps 14,08 fps
**Brain** 30,30 fps 27,02 fps 23,25 fps 16,39 fps
**Stent** 15,87 fps 12,56 fps 9,90 fps 7,46 fps
: Messung Zucker-Hummel-Operator:
**512X384** **640X480** **800X600** **1024X768**
----------- ------------- ------------- ------------- --------------
**Bunny** 10,10 fps 7,57 fps 5,84 fps 4,09 fps
**Head** 18,51 fps 15,38 fps 11,62 fps 7,87 fps
**Brain** 23,25 fps 19,60 fps 16,39 fps 11,76 fps
**Stent** 6,53 fps 4,90 fps 3,62 fps 2,69 fps
: Messung Zucker-Hummel-Operator:
Die Messungen wurden mittels eines Intel Core 2 Duo 6420 mit 2,1GHz und einer NVidia GForce 8800 GT vorgenommen. Die erste Tabelle zeigt die Bildwiederholfrequenz (fps) gemessen beim Central-Difference-Operator, die zweite Tabelle beim Zucker-Hummel-Operator. Bildauflösungen sind Spaltenweise aufgetragen, die gerenderten Modelle Zeilenweise. In den Abbildungen \[label:MessungCentralDifference\] und \[label:MessungZuckerHummel\] sind die Daten als Graph dargestellt. Hier lässt sich ein linearer Zusammenhang zwischen Renderauflösung und Bildwiederholfrequenz erkennen. Dies ist aber auch nicht verwunderlich, da beim Erhöhen der Renderauflösung auch die Anzahl der zu berechnenden Strahlen steigt. Da pro neu zu berechnendem Pixel ein neuer Strahl entsteht und der Raycaster weder Reflexion noch Refraktion berechnet, steigt der Rechenaufwand linear. In der Abbildung \[label:MessungVergleich\] sind beide Messungen in einer Zeichnung als Vergleich dargestellt. Der Central-Difference-Operator liefert eine deutlich höhere Bildwiederholfrequenz als der Zucker-Hummel-Operator. Dies lässt sich einfach dadurch erklären, dass der Zucker-Hummel-Operator deutlich komplexer in seiner Berechnung ist als der Central-Difference-Operator. Während der Central-Difference-Operator nur 6 Sampling-Punkte für seine Berechnung benötigt sind es beim Zucker-Hummel-Operator 54. Der Performance-unterschied des Operators wird aber wieder durch eine deutlich gesteigerte Bildqualität wett gemacht.
![Direkter Vergleich Central-Difference und Zucker-Hummel - Messungen[]{data-label="label:MessungVergleich"}](images/CentralDifferenceMessung.pdf){width="1\columnwidth"}
![Direkter Vergleich Central-Difference und Zucker-Hummel - Messungen[]{data-label="label:MessungVergleich"}](images/ZuckerHummelMessung.pdf){width="1\columnwidth"}
![Direkter Vergleich Central-Difference und Zucker-Hummel - Messungen[]{data-label="label:MessungVergleich"}](images/OperatorVergleichMessung.pdf){width="1\columnwidth"}
Beschleunigungsmöglichkeiten für Raycasting
===========================================
In diesem Kapitel werden einige Techniken vorgestellt, mit denen Raycaster optimiert werden können. Mittels dieser Optimierung lassen sich die Rendergeschwindigkeit und damit die Bildwiederholfrequenz des Raycasters deutlich verbessern.
Hitpoint-Refinement
-------------------
Beim *Hitpoint-Refinement* wird ohne Einbruch der Rendering-Geschwindigkeit eine höhere Qualität im Detektieren von Oberflächen (Schnittpunkt zwischen Strahl und Oberfläche) erreicht. Zunächst wird entlang eines Strahls abgetastet. Diese Abtastung wird mit einer relativ kleinen Abtastrate durchgeführt (große Schritte zwischen den Abtastpunkten). Trifft man auf einen Schnittpunkt mit einem Objekt wird zunächst ein halber Schritt zurück gemacht. Nun wird überprüft, ob man sich näher am gewünschten Zielwert befindet oder sich von diesem entfernt hat. Befindet man sich näher am Zielwert, so geht man einen halben Schritt der letzten Schrittweite weiter zurück (also ein Viertelschritt). Ansonsten geht man einen solchen Schritt wieder nach vorne. Diese Vorgehen wiederholt sich 5 bis 6 mal und führt laut [@Scharsach] zu einem 64 fach besseren Schnittpunkt als der zuerst gewählte. Wird Hitpoint-Refinement eingesetzt kann die Abtastdistanz auf 400% oder 500% der originalen angehoben werden und so die Rendering-Geschwindigkeit deutlich erhöhen.
Adaptives Sampling
------------------
Beim *Adaptiven Sampling* unterteilt man den zu Grunde liegenden Volumendatensatz in unterschiedliche Regionen. Es gibt solche Regionen die Detailarm sind, wie z.B. großflächige Oberflächen. Dann gibt es andere Regionen die besonders Detailreich sind, wie z.B. Regionen mit sehr vielen kleinen Objekten. In den Detailarmen Regionen wird die Abtastrate des Volumendatensatz herunter skaliert, da eine gröbere Abtastung in diesen ausreichend ist. In den Detailreichen Regionen wiederum wird die Abtastrate entsprechend erhöht, um auch filigrane Details in die Berechnung einfließen zu lassen. Da Volumendatensätze aber häufig zumeist aus Detailarmen Regionen (im einfachsten Fall leerer Raum) bestehen, führt das adaptive Sampling zu einer deutlichen Steigerung der Rendergeschwindigkeit.
Empty-Space-Skipping
--------------------
Das *Empty-Space-Skipping* versucht, wie der Name bereits vermuten lässt, Empty-Space, also freien Raum, zu überbrücken. Die größte Rechenzeitverschwendung bei einem Raycaster ist das Abtasten “im Leeren”. Der Raycaster verfolgt einen Strahl und tastet diesen sukzessiv ab. Abtastpunkte werden in immer gleichen Abständen gesetzt und der abgetastete Wert wird für den resultierenden Pixel ein berechnet. Durchläuft der Strahl größtenteils leeren Raum, also Positionen im Volumendatensatz die keine oder keine relevanten Daten enthalten, so wird dieser Raum dennoch abgetastet. Diese Abtastung kostet natürlich Rechenzeit. Empty-Space-Skipping versucht genau die Abtastung dieses leeren Raumes zu minimieren. So wird das im Volumen enthaltene Objekt möglichst genau durch eine “Hülle” aus primitiven (Dreiecken, Quads oder dergleichen) umschlossen. Nun muss nicht mehr der komplette “Würfel” des Volumendatensatzes abgetastet werden sondern nur das Stück des Strahls, das vom Eintrittspunkt des Strahls und der Hülle bis zu dessen Austrittspunkt reicht.
Octree
------
Ähnlich wie bei der Einteilung im Zweidimensionalen durch *Quadtrees*, bei denen ein 2D-Bild immer wieder in vier gleich große Stücke zerteilt wird, kann ein Objekt im dreidimensionalen Raum durch einen sogenannten *Octree* beschrieben werden. Hierzu wird ein Volumendatensatz in acht gleich große Blöcke zerteilt. Jeder Block, der relevante Daten enthält, wird in acht kleinere Blöcke zerteilt. Diese Blöcke wiederum werden mit ihrem Vaterknoten (dem sie umgebenden Block) verknüpft. Dieses Vorgehen wird rekursiv so lange wiederholt, bis die Blöcke eine bestimmte Mindestgröße oder die Größe eines einzelnen Voxels erreicht haben. Um für das Raycasting zu testen, welche Voxel nun von einem Abtaststrahl getroffen werden, testet man den Strahl mit dem Octree. Man fängt mit dem größten Block an und testet, ob der Strahl diesen schneidet. Wenn dies der Fall ist, testet man dessen Kinderblöcke und so weiter. Alle so gefunden Voxel werden sortiert in ein Ausgabearray geschrieben. Der Suchalgorithmus terminiert, wenn man entweder auf einen Block gestoßen ist, der keine Daten enthält oder bei den kleinsten Blöcken angekommen ist. In Abbildung \[label:Octree\] ist der Zusammenhang zwischen Baum und Blöcken grafisch dargestellt. Hier wird ein Block immer weiter zerteilt und parallel dazu die Baumstruktur angelegt. Octrees ermöglichen das deutlich schnellere finden von Schnittpunkten in einem Volumendatensatz und tragen zu einer deutlich höhreren Abastgeschwindkeit bei, da sie, ähnlich wie das Empty-Space-Skipping dazu führen, das nur “nicht-leerer Raum” abgetastet wird. Auf Grafikkarten sind Octrees jedoch relativ schwierig zu implementieren, da die Programmierung auf der Grafikkarte keine rekursiven Algorithmen erlaubt und der dafür benötigtet Stack selbst implementiert werden muss. Aus diesem Grund wurde in dem hier vorgestellten Raycaster bisher kein Octree implementiert.
![Aufbau eines Octrees[]{data-label="label:Octree"}](images/octree.pdf){width="1\columnwidth"}
Zusammenfassung und Ausblick
============================
Zusammenfassung
---------------
In dieser Ausarbeitung wurde das Volume Raycasting mit OpenCL dargestellt und erläutert. Es wurden zunächst die Grundlagen von Raytracing und Raycasting vorgestellt. Beim Raytracing werden Strahlen, die Rays, genutzt, um fotorealistische Bilder zu erzeugen. Hierzu werden Strahlen vom Auge des Betrachters in Richtung aller in der Szene befindlichen Objekte ausgesandt und mittels der Schnittpunkte Beleuchtung, Reflektion und Refraktion berechnet. Beim Raycasting werden ebenfalls Strahlen ausgesandt, jedoch werden diese auf sogenannte Volumendatensätze gerichtet. Reflektion und Refraktion werden beim Raycasting nicht betrachtet. Gradientenoperatoren dienen zur lokalen Oberflächenbestimmung innerhalb des Volumendatensatzes. Transmissions-Emmissionsmodelle werden genutzt, um z.B. in medizinischen Daten auch “unter die Haut” schauen zu können. Für die Farbbestimmung in medizinischen Daten wird die Hounsefield-Skala benutzt.
Der offene Standard OpenCL wurde vorgestellt und für diese Ausarbeitung genutzt, um einen Raycaster zu implementieren. OpenCL-Programme werden, im Gegensatz zu klassischen Computerprogrammen, direkt auf der Grafikkarte ausgeführt. Der Vorteil dieser Programme ist die massive Parallelität, die OpenCL bzw. Grafikkartenprogrammierung generelle, mit sich bringt. Aus diesem Grund bietet sich die direkte Grafikkartenprogrammierung auch an, um Raytracing- und Raycasting-Algorithmen auf diese zu portieren. Die Strahlenberechnungen sind voneinander unabhängig und können daher parallel ausgeführt werden. Hierbei ermöglichen Grafikkarten Rechengeschwindkeiten, die es dem Entwickler ermöglichen, Raytracing und Raycasting in Echtzeit ablaufen zu lassen.
Der innerhalb dieser Ausarbeitung entwickelte Raycaster und seine Funktionen wurden ebenfalls kurz vorgestellt. Anschließend zeigte eine Evalutation die Geschwindigkeiten (anhand Bildwiederholraten), die dieser relativ einfach konstruierte Raycaster, auch ohne besondere Optimierungen, liefern kann. Hierzu wurden im Evaluationskapitel zwei der implementierten Gradientenoperatoren gegenübergestellt.
Im Kapitel “Beschleunigungsmöglichkeiten für Raycasting” wurden mehrere Optimierungen vorgestellt, um das Raycasting “noch schneller” zu implementieren. Hierunter fielen Hitpoint-Refinement, Adaptives Sampling, Empty-Space-Skipping und der Octree. Allen diesen Verfahren ist gemein, dass sie unnötige Abtastungen im Volumendatensatz vermeiden bzw vermeidbar machen.
Ausblick
--------
Parallel zu dieser Ausarbeitung wurde ein eigener rudimentärer Raycaster geschrieben, der allerdings relativ akzeptable Ergebnisse liefert. Nichtsdestotrotz gibt es eine noch eine Menge Optimierungen, die den Raycaster “besser” machen könnten. Einige dieser Optimierung wurden im Kapitel “Beschleunigungsmöglichkeiten für Raycasting” kurz vorgestellt. So könnte die Implementierung eines Octrees den selbst entwickelten Raycaster um einiges schneller machen. Leider machen die mangelnden Debug-Möglichkeiten und ein teilweise sehr merkwürdiges Verhalten der selbstentwickelten OpenCL Programme (z.B. Speicherüberlauf oder einfach das Abstürzen des Grafikkartentreibers) eine Entwicklung von komplexeren Algorithmen auf der Grafikkarte deutlich schwieriger. Auch unterstützt die OpenCL-Implementierung von NVidia nicht alle Funktionen, die der offene Standard mit sich bringt. So wäre eine Portierung auf das von NVidia entwickelte und vermarktete CUDA eine Möglichkeit, einen Raycaster einfacher zu entwickeln und zu Debuggen, was vor der Auswahl des dem Raycaster zu Grunde liegenden Frameworks leider nicht bekannt war. So würde eine Portierung auf CUDA vermutlich einige der angesprochenen Probleme minimieren oder beseitigen, da NVidia CUDA deutlich mehr forciert als OpenCL. Durch die Portierung auf CUDA würde der Raycaster leider seine Kompatibilität zu ATI-Grafikkarten verlieren, da es nur auf NVidia Grafikkarten arbeitet.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A masses of a leptons deduced from a representation of a probability density vector by a spinors. A massive W and Z bosons and a massless A boson are obtained from a transformations for which a density vector is invariant.'
author:
- 'Gunn Alex Quznetsov[^1]'
title: A Piece of the Lepton Theory from a Probability
---
I use the following denotations:
$$1_2=\left[
\begin{array}{cc}
1 & 0 \\
0 & 1
\end{array}
\right] \mbox{, }0_2=\left[
\begin{array}{cc}
0 & 0 \\
0 & 0
\end{array}
\right]$$
and for $k\geq 2$:
$$1_{2k}=\left[
\begin{array}{cc}
1_k & 0_k \\
0_k & 1_k
\end{array}
\right] \mbox{, }0_{2k}=\left[
\begin{array}{cc}
0_k & 0_k \\
0_k & 0_k
\end{array}
\right] \mbox{.}$$
$$\sigma _1=\left[
\begin{array}{cc}
0 & 1 \\
1 & 0
\end{array}
\right] \mbox{, }\sigma _2=\left[
\begin{array}{cc}
0 & -i \\
i & 0
\end{array}
\right] \mbox{, }\sigma _3=\left[
\begin{array}{cc}
1 & 0 \\
0 & -1
\end{array}
\right]$$
are the Pauli matrices.
The Clifford pentad [@MD] $\stackrel{\circ }{\beta }$ is:
$$\begin{array}{c}
\beta _1=\left[
\begin{array}{cc}
\sigma _1 & 0_2 \\
0_2 & -\sigma _1
\end{array}
\right] ,\beta _2=\left[
\begin{array}{cc}
\sigma _2 & 0_2 \\
0_2 & -\sigma _2
\end{array}
\right] ,\beta _3=\left[
\begin{array}{cc}
\sigma _3 & 0_2 \\
0_2 & -\sigma _3
\end{array}
\right] , \\
\gamma _0=\left[
\begin{array}{cc}
0_2 & 1_2 \\
1_2 & 0_2
\end{array}
\right] =\beta _5,\beta _4=i\left[
\begin{array}{cc}
0_2 & 1_2 \\
-1_2 & 0_2
\end{array}
\right] \mbox{, }
\end{array}
\label{a0}$$
$$\beta _0=\left[
\begin{array}{cc}
1_2 & 0_2 \\
0_2 & 1_2
\end{array}
\right] \mbox{, }\gamma _5=\left[
\begin{array}{cc}
1_2 & 0_2 \\
0_2 & -1_2
\end{array}
\right] \mbox{.}$$
Masses
======
Let
$$\left\langle \rho \left( t,x,y,z\right) ,j_x\left( t,x,y,z\right) ,j_y\left(
t,x,y,z\right) ,j_z\left( t,x,y,z\right) \right\rangle$$
be a probability current 3+1 vector field [@PRB] and $\psi \left(
t,x,y,z\right) $ be any complex spinor field:
$$\psi =\left| \psi \right| \left[
\begin{array}{c}
\exp \left( i\stackrel{*}{\gamma }\right) \cos \left( \stackrel{*}{\beta }%
\right) \cos \left( \stackrel{*}{\alpha }\right) \\
\exp \left( i\stackrel{*}{\theta }\right) \sin \left( \stackrel{*}{\beta }%
\right) \cos \left( \stackrel{*}{\alpha }\right) \\
\exp \left( i\stackrel{*}{\varphi }\right) \cos \left( \stackrel{*}{\chi }%
\right) \sin \left( \stackrel{*}{\alpha }\right) \\
\exp \left( i\stackrel{*}{\upsilon }\right) \sin \left( \stackrel{*}{\chi }%
\right) \sin \left( \stackrel{*}{\alpha }\right)
\end{array}
\right] \mbox{.}$$
In this case the following system of equations:
$$\left\{
\begin{array}{c}
\psi ^{\dagger }\psi =\rho \mbox{,} \\
\psi ^{\dagger }\beta _1\psi =j_x\mbox{,} \\
\psi ^{\dagger }\beta _2\psi =j_y\mbox{,} \\
\psi ^{\dagger }\beta _3\psi =j_z
\end{array}
\right| \label{a1}$$
has got the following form:
$$\left\{
\begin{array}{c}
\psi ^{\dagger }\psi =\rho \mbox{,} \\
\left| \psi \right| ^2\left(
\begin{array}{c}
\cos ^2\left( \stackrel{*}{\alpha }\right) \sin \left( 2\stackrel{*}{\beta }%
\right) \cos \left( \stackrel{*}{\theta }-\stackrel{*}{\gamma }\right) - \\
-\sin ^2\left( \stackrel{*}{\alpha }\right) \sin \left( 2\stackrel{*}{\chi }%
\right) \cos \left( \stackrel{*}{\upsilon }-\stackrel{*}{\varphi }\right)
\end{array}
\right) =j_x\mbox{,} \\
\left| \psi \right| ^2\left(
\begin{array}{c}
\cos ^2\left( \stackrel{*}{\alpha }\right) \sin \left( 2\stackrel{*}{\beta }%
\right) \sin \left( \stackrel{*}{\theta }-\stackrel{*}{\gamma }\right) - \\
-\sin ^2\left( \stackrel{*}{\alpha }\right) \sin \left( 2\stackrel{*}{\chi }%
\right) \sin \left( \stackrel{*}{\upsilon }-\stackrel{*}{\varphi }\right)
\end{array}
\right) =j_y\mbox{,} \\
\left| \psi \right| ^2\left( \cos ^2\left( \stackrel{*}{\alpha }\right) \cos
\left( 2\stackrel{*}{\beta }\right) -\sin ^2\left( \stackrel{*}{\alpha }%
\right) \cos \left( 2\stackrel{*}{\chi }\right) \right) =j_z\mbox{.}
\end{array}
\right|$$
Hence for every probability current vector $\left\langle \rho
,j_x,j_y,j_z\right\rangle $: the spinor $\psi $, obeyed to this system, exists.
The operator $\widehat{U}\left( t,\triangle t\right) $, which acts in the set of these spinors, is denoted as the evolution operator for the spinor $%
\psi \left( t,x,y,z\right) $, if:
$$\psi \left( t+\triangle t,x,y,z\right) =\widehat{U}\left( t,\triangle
t\right) \psi \left( t,x,y,z\right) \mbox{.}$$
$\widehat{U}\left( t,\triangle t\right) $ is a linear operator.
The set of the spinors, for which $\widehat{U}\left( t,\triangle t\right) $ is the evolution operator, is denoted as the operator $\widehat{U}\left(
t,\triangle t\right) $ space.
The operator space is the linear space.
Let for an infinitesimal $\triangle t$:
$$\widehat{U}\left( t,\triangle t\right) =1+i\triangle t\widehat{H}\left(
t\right) \mbox{.}$$
Hence for an elements of the operator $\widehat{U}\left( t,\triangle
t\right) $ space:
$$i\widehat{H}=\partial _t\mbox{.}$$
If the functions $\rho $, $j_x$, $j_y$, $j_z$ fulfill to the continuity equation [@PRB]:
$$\partial _t\rho +\partial _xj_x+\partial _yj_y+\partial _zj_z=0$$
then:
$$\left( \left( \partial _t\psi ^{\dagger }\right) \beta _0+\left( \partial
_x\psi ^{\dagger }\right) \beta _1+\left( \partial _y\psi ^{\dagger }\right)
\beta _2+\left( \partial _z\psi ^{\dagger }\right) \beta _3\right) \psi =$$
$$=-\psi ^{\dagger }\left( \left( \beta _0\partial _t+\beta _1\partial
_x+\beta _2\partial _y+\beta _3\partial _z\right) \psi \right) \mbox{.}$$
Let:
$$\widehat{Q}=\left( i\widehat{H}+\beta _1\partial _x+\beta _2\partial
_y+\beta _3\partial _z\right) \mbox{.}$$
Hence:
$$\psi ^{\dagger }\widehat{Q}^{\dagger }\psi =-\psi ^{\dagger }\widehat{Q}\psi %
\mbox{.}$$
Hence $i\widehat{Q}\left( t,x,y,z\right) $ is the Hermitian for the matrix product operator.
Hence a real functions $\varphi _{i,j}\left( t,x,y,z\right) $ and $\varpi
_{i,j}\left( t,x,y,z\right) $ for which:
$$-i\widehat{Q}=$$
$$\left[
\begin{array}{cccc}
\varphi _{1,1} & \varphi _{1,2}+i\varpi _{1,2} & \varphi _{1,3}+i\varpi
_{1,3} & \varphi _{1,4}+i\varpi _{1,4} \\
\varphi _{1,2}-i\varpi _{1,2} & \varphi _{2,2} & \varphi _{2,3}+i\varpi
_{2,3} & \varphi _{2,4}+i\varpi _{2,4} \\
\varphi _{1,3}-i\varpi _{1,3} & \varphi _{2,3}-i\varpi _{2,3} & \varphi
_{3,3} & \varphi _{3,4}+i\varpi _{3,4} \\
\varphi _{1,4}-i\varpi _{1,4} & \varphi _{2,4}-i\varpi _{2,4} & \varphi
_{3,4}-i\varpi _{3,4} & \varphi _{4,4}
\end{array}
\right]$$
exist.
Let $G_t$, $G_z$, $K_t$ and $K_z$ are the solution of the following system of equations:
$$\left\{
\begin{array}{c}
G_t+G_z+K_t+K_z=\varphi _{1,1}\mbox{,} \\
G_t-G_z+K_t-K_z=\varphi _{2,2}\mbox{,} \\
G_t-G_z-K_t+K_z=\varphi _{3,3}\mbox{,} \\
G_t+G_z-K_t-K_z=\varphi _{4,4};
\end{array}
\right|$$
$G_x$ and $K_x$ are the solution of the following system of equations:
$$\left\{
\begin{array}{c}
G_x+K_x=\varphi _{1,2}\mbox{,} \\
-G_x+K_x=\varphi _{3,4}\mbox{;}
\end{array}
\right|$$
$G_y$ and $K_y$ are the solution of the following system of equations:
$$\left\{
\begin{array}{c}
-G_y-K_y=\varpi _{1,2}\mbox{,} \\
G_y-K_x=\varpi _{3,4}\mbox{.}
\end{array}
\right|$$
In this case:
$$\begin{array}{c}
-i\widehat{Q}= \\
=\left( G_t\beta _0+G_x\beta _1+G_y\beta _2+G_z\beta _3\right) + \\
+\left( K_t\beta _0+K_x\beta _1+K_y\beta _2+K_z\beta _3\right) \gamma _5+
\end{array}$$
$$+\left[
\begin{array}{cccc}
0 & 0 & \varphi _{1,3}+i\varpi _{1,3} & \varphi _{1,4}+i\varpi _{1,4} \\
0 & 0 & \varphi _{2,3}+i\varpi _{2,3} & \varphi _{2,4}+i\varpi _{2,4} \\
\varphi _{1,3}-i\varpi _{1,3} & \varphi _{2,3}-i\varpi _{2,3} & 0 & 0 \\
\varphi _{1,4}-i\varpi _{1,4} & \varphi _{2,4}-i\varpi _{2,4} & 0 & 0
\end{array}
\right] \mbox{.}$$
If
$$\left\{
\begin{array}{c}
\left( M_0+M_{z,0}\right) =\varphi _{1,3}\mbox{,} \\
\left( M_0-M_{z,0}\right) =\varphi _{2,4}\mbox{,}
\end{array}
\right|$$
$$\left\{
\begin{array}{c}
\left( M_4+M_{z,4}\right) =\varpi _{1,3}\mbox{,} \\
\left( M_4-M_{z,4}\right) =\varpi _{2,4}\mbox{,}
\end{array}
\right|$$
$$\left\{
\begin{array}{c}
\left( M_{x,0}+M_{y,4}\right) =\varphi _{1,4}\mbox{,} \\
\left( M_{x,0}-M_{y,4}\right) =\varphi _{2,3}\mbox{,}
\end{array}
\right|$$
$$\left\{
\begin{array}{c}
\left( M_{x,4}+M_{y,0}\right) =\varpi _{1,4}\mbox{,} \\
\left( M_{x,4}-M_{y,0}\right) =\varpi _{2,3}
\end{array}
\right|$$
then
$$\left[
\begin{array}{cccc}
0 & 0 & \varphi _{1,3}+i\varpi _{1,3} & \varphi _{1,4}+i\varpi _{1,4} \\
0 & 0 & \varphi _{2,3}+i\varpi _{2,3} & \varphi _{2,4}+i\varpi _{2,4} \\
\varphi _{1,3}-i\varpi _{1,3} & \varphi _{2,3}-i\varpi _{2,3} & 0 & 0 \\
\varphi _{1,4}-i\varpi _{1,4} & \varphi _{2,4}-i\varpi _{2,4} & 0 & 0
\end{array}
\right] =$$
$$=M_0\gamma _0+M_4\beta _4-M_{x,0}\gamma _\zeta ^0-M_{x,4}\zeta
^4+M_{y,0}\gamma _\eta ^0+M_{y,4}\eta ^4-M_{z,0}\gamma _\theta
^0-M_{z,4}\theta ^4\mbox{;}$$
here $\gamma _\zeta ^0$, $\zeta ^4$, $\gamma _\eta ^0$, $\eta ^4$, $\gamma
_\theta ^0$, $\theta ^4$ are the chromatic pentads [@QD], [@HC] members and $\gamma _0$ and $\beta _4$ is the light pentad $\stackrel{\circ
}{\beta }$ members. Since in this paper I will not consider a quarks then everywhere below: $$M_{x,0}=M_{x,4}=M_{y,0}=M_{y,4}=M_{z,0}=M_{z,4}=0$$
hence:
$$\begin{array}{c}
-i\widehat{Q}= \\
=\left( G_t\beta _0+G_x\beta _1+G_y\beta _2+G_z\beta _3\right) + \\
+\left( K_t\beta _0+K_x\beta _1+K_y\beta _2+K_z\beta _3\right) \gamma _5+ \\
+M_0\gamma _0+M_4\beta _4\mbox{.}
\end{array}$$
$\left\{ \beta _1,\beta _2,\beta _3,\beta _4,\gamma _0\right\} $ is the Clifford pentad.
If $j_x=\rho u_x$, $j_y=\rho u_y$, $j_z=\rho u_z$ then $u_x$, $u_y$, $u_z$ are the components of the average velocity. Hence $\beta _1$, $\beta _2$, $%
\beta _3$ define the components of the average velocity (\[a1\]).
If
$$j_{x_5}=\psi ^{\dagger }\gamma _0\psi \mbox{, }j_{x_4}=\psi ^{\dagger }\beta
_4\psi \mbox{, }j_{x_5}=\rho u_{x_5}\mbox{, }j_{x_4}=\rho u_{x_4}$$
then
$$\begin{array}{c}
u_{x_5}=\sin \left( 2\stackrel{*}{\alpha }\right) \left[
\begin{array}{c}
\cos \left( \stackrel{*}{\beta }\right) \cos \left( \stackrel{*}{\chi }%
\right) \cos \left( \stackrel{*}{\gamma }-\stackrel{*}{\varphi }\right) + \\
+\sin \left( \stackrel{*}{\beta }\right) \sin \left( \stackrel{*}{\chi }%
\right) \cos \left( \stackrel{*}{\theta }-\stackrel{*}{\upsilon }\right)
\end{array}
\right] \mbox{,} \\
u_{x_4}=\sin \left( 2\stackrel{*}{\alpha }\right) \left[
\begin{array}{c}
\cos \left( \stackrel{*}{\beta }\right) \cos \left( \stackrel{*}{\chi }%
\right) \sin \left( \stackrel{*}{\gamma }-\stackrel{*}{\varphi }\right) + \\
+\sin \left( \stackrel{*}{\beta }\right) \sin \left( \stackrel{*}{\chi }%
\right) \sin \left( \stackrel{*}{\theta }-\stackrel{*}{\upsilon }\right)
\end{array}
\right]
\end{array}$$
and if $\rho \neq 0$ then
$$u_x^2+u_y^2+u_z^2+u_{x_5}^2+u_{x_4}^2=1. \label{vel}$$
From [@PG] the maximal velocity of the information propagation in the space-time is 1.
Hence of only all five elements of the Clifford pentad lends the entire kit of the velocity components and, for the completeness, yet two ”space” coordinates $x_5$ and $x_4$ should be added to our three $x,y,z$.
Let
$$\begin{array}{c}
\mathbf{\Psi }\left( t,x,y,z,x_5,x_4\right) = \\
=\psi \left( t,x,y,z\right) \exp \left( -i\left( x_5M_0\left( t,x,y,z\right)
+x_4M_4\left( t,x,y,z\right) \right) \right) \mbox{.}
\end{array}$$
In this case the motion equation is the following:
$$\begin{array}{c}
\beta _0i\partial _t\mathbf{\Psi }+\beta _1i\partial _x\mathbf{\Psi }+\beta
_2i\partial _y\mathbf{\Psi }+\beta _3i\partial _z\mathbf{\Psi }+\gamma
_0i\partial _{x_5}\mathbf{\Psi }+\beta _4i\partial _{x_4}\mathbf{\Psi }+ \\
+\left( G_t\beta _0+G_x\beta _1+G_y\beta _2+G_z\beta _3\right) \mathbf{\Psi }%
+ \\
+\left( K_t\beta _0+K_x\beta _1+K_y\beta _2+K_z\beta _3\right) \gamma _5%
\mathbf{\Psi }=0
\end{array}
\label{gkk}$$
Let a evolution operator $\widehat{U}\left( t,\triangle t\right) $ be denoted as [*a Planck evolution operator*]{} if a tiny positive real number $h$ and a functions $N_\varphi \left( t,x,y,z\right) $ and $N_\varpi
\left( t,x,y,z\right) $, having a range of values in the set of the integer numbers, exist for which:
$$M_0=N_\varphi h\mbox{ and }M_4=N_\varpi h\mbox{.}$$
Let $-\frac \pi h\leq x_5\leq \frac \pi h$ , $-\frac \pi h\leq x_4\leq \frac
\pi h$,
$\mathbf{\Psi }\left( t,x,y,z,\pm \frac \pi h,x_4\right) =0$ and $\mathbf{%
\Psi }\left( t,x,y,z,x_5,\pm \frac \pi h\right) =0$.
In this case the Fourier series for $\mathbf{\Psi }$ is of the following form:
$$\begin{array}{c}
\mathbf{\Psi }\left( t,x,y,z,x_5,x_4\right) = \\
=\psi \left( t,x,y,z\right) \sum_{\nu ,\kappa }\delta _{-\nu ,N_\varphi
\left( t,x,y,z\right) }\delta _{-\kappa ,N_\varpi \left( t,x,y,z\right)
}\exp \left( -ih\left( \nu x_5+\kappa x_4\right) \right) \mbox{.}
\end{array}$$
Here:
$$\begin{aligned}
\delta _{-\nu ,N_\varphi } &=&\frac h{2\pi }\int_{-\frac \pi h}^{\frac \pi
h}\exp \left( ih\left( \nu x_5\right) \right) \exp \left( iN_\varphi
hx_5\right) dx_5=\frac{\sin \left( \pi \left( \nu +N_\varphi \right) \right)
}{\pi \left( \nu +N_\varphi \right) }\mbox{,} \\
\delta _{-\kappa ,N_\varpi } &=&\frac h{2\pi }\int_{-\frac \pi h}^{\frac \pi
h}\exp \left( ih\left( \kappa x_4\right) \right) \exp \left( iN_\varpi
hx_4\right) dx_4=\frac{\sin \left( \pi \left( \kappa +N_\varpi \right)
\right) }{\pi \left( \kappa +N_\varpi \right) }\mbox{.}\end{aligned}$$
If denote:
$$\phi \left( t,x,y,z,-\nu ,-\kappa \right) =\psi \left( t,x,y,z\right) \delta
_{\nu ,N_\varphi \left( t,x,y,z\right) }\delta _{\kappa ,N_\varpi \left(
t,x,y,z\right) }$$
then
$$\begin{array}{c}
\mathbf{\Psi }\left( t,x,y,z,x_5,x_4\right) = \\
=\sum_{\nu ,\kappa }\phi \left( t,x,y,z,\nu ,\kappa \right) \exp \left(
-ih\left( \nu x_5+\kappa x_4\right) \right) \mbox{.}
\end{array}$$
From the properties of $\delta $ in every point $\left\langle
t,x,y,z\right\rangle $: either
$$\mathbf{\Psi }\left( t,x,y,z,x_5,x_4\right) =0$$
or an integer numbers $\nu _0$ and $\kappa _0$ exist for which:
$$\mathbf{\Psi }\left( t,x,y,z,x_5,x_4\right) =\phi \left( t,x,y,z,\nu
_0,\kappa _0\right) \exp \left( -ih\left( \nu _0x_5+\kappa _0x_4\right)
\right) \mbox{.} \label{dlt}$$
That is for the every space-time point: either this point is empty or single mass is placed in this point.
Let on the space of these spinors the scalar product $\mathbf{\Phi }*\mathbf{%
\Psi }$ be denoted as the following:
$$\mathbf{\Phi }*\mathbf{\Psi }=\left( \frac h{2\pi }\right) ^2\int_{-\frac
\pi h}^{\frac \pi h}dx_5\int_{-\frac \pi h}^{\frac \pi h}dx_4\cdot \left(
\mathbf{\Phi }^{\dagger }\mathbf{\Psi }\right) \mbox{.}$$
In this case:
$$\mathbf{\Psi }*\beta _\mu \mathbf{\Psi }=\psi ^{\dagger }\beta _\mu \psi %
\mbox{.}$$
for $0\leq \mu \leq 3$
Hence from (\[a1\]):
$$\left\{
\begin{array}{c}
\mathbf{\Psi }^{\dagger }*\mathbf{\Psi }=\rho \mbox{,} \\
\mathbf{\Psi }^{\dagger }*\beta _1\mathbf{\Psi }=j_x\mbox{,} \\
\mathbf{\Psi }^{\dagger }*\beta _2\mathbf{\Psi }=j_y\mbox{,} \\
\mathbf{\Psi }^{\dagger }*\beta _3\mathbf{\Psi }=j_z\mbox{.}
\end{array}
\right|$$
Bi-zero-nonzero-mass state
--------------------------
Let
$$\begin{array}{c}
\mathbf{\Psi }\left( t,x,y,z,x_5,x_4\right) = \\
=\phi \left( t,x,y,z,0,0\right) +\phi \left( t,x,y,z,n,k\right) \exp \left(
ih\left( nx_5+kx_4\right) \right) \mbox{.}
\end{array}$$
Let $\epsilon _\mu $ ($1\leq k\leq 4$) be a basis in which pentad $\stackrel{%
\circ }{\beta }$ has got a form (\[a0\]) and let
$$\begin{array}{c}
\mathbf{\Psi }\left( x_5,x_4\right) = \\
=\sum_{r=1}^4\phi _r\left( 0,0\right) \epsilon _r+\exp \left( -ih\left(
nx_5+kx_4\right) \right) \sum_{k=1}^4\phi _k\left( n,k\right) \epsilon _k
\end{array}
\label{bir}$$
Hence in the basis
$$\left\langle \epsilon _r,\exp \left( -ih\left( nX+kY\right) \right) \epsilon
_k\right\rangle :$$
a 8-components bi-spinor:
$$\Psi =\left[
\begin{array}{c}
\phi _1\left( 0,0\right) \\
\phi _2\left( 0,0\right) \\
\phi _3\left( 0,0\right) \\
\phi _4\left( 0,0\right) \\
\phi _1\left( n,k\right) \\
\phi _2\left( n,k\right) \\
\phi _3\left( n,k\right) \\
\phi _4\left( n,k\right)
\end{array}
\right]$$
corresponds to $\mathbf{\Psi }$.
From (\[dlt\]): in every point $\left\langle t,x,y,z\right\rangle $:
$$\Psi =\left[
\begin{array}{c}
\phi _1\left( 0,0\right) \\
\phi _2\left( 0,0\right) \\
\phi _3\left( 0,0\right) \\
\phi _4\left( 0,0\right) \\
0 \\
0 \\
0 \\
0
\end{array}
\right] \mbox{ or }\Psi =\left[
\begin{array}{c}
0 \\
0 \\
0 \\
0 \\
\phi _1\left( n,k\right) \\
\phi _2\left( n,k\right) \\
\phi _3\left( n,k\right) \\
\phi _4\left( n,k\right)
\end{array}
\right] \label{dl2}$$
of $\delta $ characteristics.
Let us denote:
$$\phi _1\epsilon _1+\phi _2\epsilon _2=\phi _L\mbox {
and }\phi _3\epsilon _3+\phi _4\epsilon _4=\phi _R\mbox {.}$$
Hence from (\[bir\]):
$$\begin{array}{c}
\mathbf{\Psi }\left( x_5,x_4\right) =\phi _L\left( 0,0\right) +\phi _R\left(
0,0\right) + \\
+\exp \left( -ih\left( nx_5+kx_4\right) \right) \left( \phi _L\left(
n,k\right) +\phi _R\left( n,k\right) \right) \mbox{.}
\end{array}
\label{bir1}$$
If use denotation:
$$\underline{\vartheta }=\left[
\begin{array}{cc}
\vartheta & 0_4 \\
0_4 & \vartheta
\end{array}
\right] \mbox {,}$$
$$\gamma =\left[
\begin{array}{cc}
-\gamma _0 & 0_4 \\
0_4 & \gamma _0
\end{array}
\right] \mbox{, }\beta =\left[
\begin{array}{cc}
-\beta _4 & 0_4 \\
0_4 & \beta _4
\end{array}
\right]$$
and
$$\underline{n}=\left[
\begin{array}{cc}
0_4 & 0_4 \\
0_4 & n1_4
\end{array}
\right] \mbox {, }\underline{k}=\left[
\begin{array}{cc}
0_4 & 0_4 \\
0_4 & k1_4
\end{array}
\right]$$
then the motion equation is the following:
$$\begin{array}{c}
\underline{\beta _0}i\partial _t\Psi +\underline{\beta _1}i\partial _x\Psi +%
\underline{\beta _2}i\partial _y\Psi +\underline{\beta _3}i\partial _z\Psi +
\\
-h\underline{n}\gamma \Psi -h\underline{k}\beta \Psi + \\
+\left( G_t\underline{\beta _0}+G_x\underline{\beta _1}+G_y\underline{\beta
_2}+G_z\underline{\beta _3}\right) \Psi + \\
+\left( K_t\underline{\beta _0}+K_x\underline{\beta _1}+K_y\underline{\beta
_2}+K_z\underline{\beta _3}\right) \underline{\gamma _5}\Psi =0
\end{array}
\label{mtn}$$
and
$$\left\{
\begin{array}{c}
\Psi ^{\dagger }\Psi =\rho \mbox{,} \\
\Psi ^{\dagger }\underline{\beta _1}\Psi =j_x\mbox{,} \\
\Psi ^{\dagger }\underline{\beta _2}\Psi =j_y\mbox{,} \\
\Psi ^{\dagger }\underline{\beta _3}\Psi =j_z\mbox{.}
\end{array}
\right|$$
If use the following denotation: $t=x_0,x=x_1,y=x_2,z=x_3,\partial _\mu
=\frac \partial {\partial _{x_\mu }}$ then the lagrangian has got the following form:
$$\begin{array}{c}
\mathcal{L}_f=0.5i\left( \left( \sum_{\mu =0}^3\Psi ^{\dagger }\underline{%
\beta _\mu }\partial _\mu \Psi \right) -\left( \sum_{\mu =0}^3\partial _\mu
\Psi ^{\dagger }\underline{\beta _\mu }\Psi \right) \right) - \\
-\left( \Psi ^{\dagger }h\underline{n}\gamma \Psi +\Psi ^{\dagger }h%
\underline{k}\beta \Psi \right) + \\
+\Psi ^{\dagger }\left( \sum_{\mu =0}^3G_{x_\mu }\underline{\beta _\mu }%
\right) \Psi +\Psi ^{\dagger }\left( \sum_{\mu =0}^3K_{x_\mu }\underline{%
\beta _\mu }\right) \underline{\gamma _5}\Psi .
\end{array}
\label{lag}$$
This lagrangian is invariant for the rotation of $xOy,yOz,xOz$ and for the Lorentz transformation of $tOx,tOy,tOz$ and $G_{x_k}$ and $K_{x_k}$ behaves as the 4-vector fields [@WGA].
### Transformations
If $U$ is an 8$\times $8 complex matrix, $\Psi ^{\prime }=U\Psi $ and
$$\left\{
\begin{array}{c}
\Psi ^{\prime \dagger }\underline{\beta _1}\Psi ^{\prime }=j_x\mbox{,} \\
\Psi ^{\prime \dagger }\underline{\beta _2}\Psi ^{\prime }=j_y\mbox{,} \\
\Psi ^{\prime \dagger }\underline{\beta _3}\Psi ^{\prime }=j_z
\end{array}
\right| \label{uni}$$
then for $1\leq k\leq 3$: $U^{\dagger }\underline{\beta _k}U=$. In this case a real numbers
$a",b",c",g",u",v",k,s,a`,b`,c`,g`,u`,v`,k`,s`$ exist for which:
$$\underline{U}=\left[
\begin{array}{cccc}
\left( a"+b"i\right) 1_2 & 0_2 & \left( c"+ig"\right) 1_2 & 0_2 \\
0_2 & \left( a`+b`i\right) 1_2 & 0_2 & \left( c`+ig`\right) 1_2 \\
\left( u"+iv"\right) 1_2 & 0_2 & \left( k"+is"\right) 1_2 & 0_2 \\
0_2 & \left( u`+iv`\right) 1_2 & 0_2 & \left( k`+is`\right) 1_2
\end{array}
\right] .$$
If $\Psi ^{\prime \dagger }\Psi ^{\prime }=\rho $ then $U^{\dagger }U=1_8$ . Hence:
$$\begin{array}{c}
v"^2+b"^2+u"^2+a"^2=1, \\
c"^2+g"^2+k"^2+s"^2=1,
\end{array}$$
$$s"=-\frac{a"g"u"-u"b"c"+a"c"v"+b"g"v"}{u"^2+v"^2},$$
$$k"=\frac{-u"a"c"-u"b"g"+v"a"g"-b"c"v"}{u"^2+v"^2}.$$
$$\begin{array}{c}
v`^2+b`^2+u`^2+a`^2=1, \\
c`^2+g`^2+k`^2+s`^2=1,
\end{array}$$
$$s`=-\frac{a`g`u`-u`b`c`+a`c`v`+b`g`v`}{u`^2+v`^2},$$
$$k`=\frac{-u`a`c`-u`b`g`+v`a`g`-b`c`v`}{u`^2+v`^2}.$$
$\underline{U}$ has got 4 eigenvalues: $\exp \left( i\alpha _1\right) $, $%
\exp \left( i\alpha _2\right) $, $\exp \left( i\alpha _3\right) $, $\exp
\left( i\alpha _4\right) $ for 8 orthogonal eigenvectors:
$\mathbf{\varepsilon }_{1,1}$, $\mathbf{\varepsilon }_{1,2}$, $\mathbf{%
\varepsilon }_{2,1}$, $\mathbf{\varepsilon }_{2,2}$, $\mathbf{\varepsilon }%
_{3,1}$, $\mathbf{\varepsilon }_{3,2}$, $\mathbf{\varepsilon }_{4,1}$, $%
\mathbf{\varepsilon }_{4,2}$.
Let
$$K=\left[
\begin{array}{cccccccc}
\mathbf{\varepsilon }_{1,1} & \mathbf{\varepsilon }_{1,2} & \mathbf{%
\varepsilon }_{2,1} & \mathbf{\varepsilon }_{2,2} & \mathbf{\varepsilon }%
_{3,1} & \mathbf{\varepsilon }_{3,2} & \mathbf{\varepsilon }_{4,1} & \mathbf{%
\varepsilon }_{4,2}
\end{array}
\right] .$$
Let $\theta _1$, $\theta _2$, $\theta _3$, $\theta _4$ be the solution of the system of the equations:
$$\left\{
\begin{array}{c}
\theta _1+\theta _2+\theta _3+\theta _4=\alpha _1, \\
\theta _1+\theta _2-\theta _3-\theta _4=\alpha _1, \\
\theta _1-\theta _2+\theta _3-\theta _4=\alpha _1, \\
\theta _1-\theta _2-\theta _3+\theta _4=\alpha _1.
\end{array}
\right|$$
and
$$U_1=\exp \left( i\theta _1\right) 1_8,$$
$$U_2=K\left[
\begin{array}{cc}
\exp \left( i\theta _2\right) 1_4 & 0_4 \\
0_4 & \exp \left( -i\theta _2\right) 1_4
\end{array}
\right] K^{\dagger },$$
$$U_3=K\left[
\begin{array}{cccc}
\exp \left( i\theta _3\right) 1_2 & 0_2 & 0_2 & 0_2 \\
0_2 & \exp \left( -i\theta _3\right) 1_2 & 0_2 & 0_2 \\
0_2 & 0_2 & \exp \left( i\theta _3\right) 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & \exp \left( -i\theta _3\right) 1_2
\end{array}
\right] K^{\dagger },$$
$$U_4=K\left[
\begin{array}{cccc}
\exp \left( i\theta _4\right) 1_2 & 0_2 & 0_2 & 0_2 \\
0_2 & \exp \left( -i\theta _4\right) 1_2 & 0_2 & 0_2 \\
0_2 & 0_2 & \exp \left( -i\theta _4\right) 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & \exp \left( i\theta _4\right) 1_2
\end{array}
\right] K^{\dagger }.$$
In this case: $$U_1U_2U_3U_4=U$$
and
$$U_2=\left[
\begin{array}{cccc}
\exp \left( i\theta _2\right) 1_2 & 0_2 & 0_2 & 0_2 \\
0_2 & \exp \left( -i\theta _2\right) 1_2 & 0_2 & 0_2 \\
0_2 & 0_2 & \exp \left( i\theta _2\right) 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & \exp \left( -i\theta _2\right) 1_2
\end{array}
\right]$$
and a real number $a,b,c,g,u,v,k,s$ exist for which:
$$U_3U_4=\left[
\begin{array}{cccc}
\left( a+ib\right) 1_2 & 0_2 & \left( c+ig\right) 1_2 & 0_2 \\
0_2 & \left( u+iv\right) 1_2 & 0_2 & \left( k+is\right) 1_2 \\
\left( -c+ig\right) 1_2 & 0_2 & \left( a-ib\right) 1_2 & 0_2 \\
0_2 & \left( -k+is\right) 1_2 & 0_2 & \left( u-iv\right) 1_2
\end{array}
\right]$$
and
$$\begin{array}{c}
a^2+b^2+c^2+g^2=1, \\
u^2+v^2+r^2+s^2=1.
\end{array}$$
If
$$U^{\left( +\right) }=\left[
\begin{array}{cccc}
1_2 & 0_2 & 0_2 & 0_2 \\
0_2 & \left( u+iv\right) 1_2 & 0_2 & \left( k+is\right) 1_2 \\
0_2 & 0_2 & 1_2 & 0_2 \\
0_2 & \left( -k+is\right) 1_2 & 0_2 & \left( u-iv\right) 1_2
\end{array}
\right] \label{upls}$$
and
$$U^{\left( -\right) }=\left[
\begin{array}{cccc}
\left( a+ib\right) 1_2 & 0_2 & \left( c+ig\right) 1_2 & 0_2 \\
0_2 & 1_2 & 0_2 & 0_2 \\
\left( -c+ig\right) 1_2 & 0_2 & \left( a-ib\right) 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & 1_2
\end{array}
\right]$$
then
$$U_3U_4=U^{\left( -\right) }U^{\left( +\right) }=U^{\left( +\right)
}U^{\left( -\right) }.$$
#### B-boson
$$U_1U_2=\left[
\begin{array}{cccc}
e^{i\left( \theta _1+\theta _2\right) } & 0 & 0 & 0 \\
0 & e^{i\left( \theta _1-\theta _2\right) } & 0 & 0 \\
0 & 0 & e^{i\left( \theta _1+\theta _2\right) } & 0 \\
0 & 0 & 0 & e^{i\left( \theta _1-\theta _2\right) }
\end{array}
\right]$$
Let $\chi $ and $\varsigma $ be the solution of the following set of equations:
$$\left\{
\begin{array}{c}
0.5\chi +\varsigma =\theta _1+\theta _2, \\
\chi +\varsigma =\theta _1-\theta _2\mbox{,}
\end{array}
\right|$$
i.e.:
$$\begin{array}{c}
\chi =-4\theta _2\mbox{,} \\
\varsigma =\theta _1+3\theta _2\mbox{.}
\end{array}$$
Let
$$\overbrace{U}=\exp \left( i\varsigma \right) 1_8$$
and
$$\widetilde{U}=\left[
\begin{array}{cccc}
\exp \left( i\frac \chi 2\right) 1_2 & 0_2 & 0_2 & 0_2 \\
0_2 & \exp \left( i\chi \right) 1_2 & 0_2 & 0_2 \\
0_2 & 0_2 & \exp \left( i\frac \chi 2\right) 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & \exp \left( i\chi \right) 1_2
\end{array}
\right] \mbox{.}$$
In that case:
$$\widetilde{U}\overbrace{U}=U_1U_2\mbox{.}$$
Let $g_1$ be a positive real number and for $\mu \in \left\{ t,x,y,z\right\}
$: $F_\mu $ and $B_\mu $ be the solutions of the following systems of the equations:
$$\left\{
\begin{array}{c}
-0.5g_1B_\mu +F_\mu =G_\mu +K_\mu \\
-g_1B_\mu +F_\mu =G_\mu -K_\mu
\end{array}
\right|$$
i.e.:
$$\begin{aligned}
B_\mu &=&\frac 4{g_1}K_\mu \\
F_\mu &=&G_\mu +3K_\mu \mbox{.}\end{aligned}$$
Let [*the charge matrix*]{} be defined as the following:
$$\underline{Y}=-\left[
\begin{array}{cccc}
1_2 & 0_2 & 0_2 & 0_2 \\
0_2 & 2\cdot 1_2 & 0_2 & 0_2 \\
0_2 & 0_2 & 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & 2\cdot 1_2
\end{array}
\right]$$
In that case from (\[mtn\]):
$$\begin{array}{c}
\underline{\beta _0}i\partial _t\Psi +\underline{\beta _1}i\partial _x\Psi +%
\underline{\beta _2}i\partial _y\Psi +\underline{\beta _3}i\partial _z\Psi +
\\
-h\underline{n}\gamma \Psi -h\underline{k}\beta \Psi + \\
+\left( F_t\underline{\beta _0}+F_x\underline{\beta _1}+F_y\underline{\beta
_2}+F_z\underline{\beta _3}\right) \Psi + \\
+0.5g_1\underline{Y}\left( B_t\underline{\beta _0}+B_x\underline{\beta _1}%
+B_y\underline{\beta _2}+B_z\underline{\beta _3}\right) \Psi =0\mbox{.}
\end{array}
\label{umt1}$$
Let
$$\begin{array}{c}
\Psi \rightarrow \Psi `=\left( \widetilde{U}\Psi \right) \mbox{,} \\
n\rightarrow n`\mbox{,} \\
k\rightarrow k`\mbox{,} \\
F_\mu \rightarrow F_\mu `\mbox{,} \\
B_\mu \rightarrow B_\mu `
\end{array}$$
then:
$$\begin{array}{c}
\underline{\beta _0}i\partial _t\left( \widetilde{U}\Psi \right) +\underline{%
\beta _1}i\partial _x\left( \widetilde{U}\Psi \right) +\underline{\beta _2}%
i\partial _y\left( \widetilde{U}\Psi \right) +\underline{\beta _3}i\partial
_z\left( \widetilde{U}\Psi \right) + \\
-h\underline{n}`\gamma \widetilde{U}\Psi -h\underline{k}`\beta \widetilde{U}%
\Psi + \\
+\left( F_t\underline{`\beta _0}+F_x`\underline{\beta _1}+F_y`\underline{%
\beta _2}+F_z`\underline{\beta _3}\right) \widetilde{U}\Psi + \\
+0.5g_1\underline{Y}\left( B_t`\underline{\beta _0}+B_x`\underline{\beta _1}%
+B_y`\underline{\beta _2}+B_z`\underline{\beta _3}\right) \widetilde{U}\Psi
=0
\end{array}$$
hence:
$$\begin{array}{c}
\underline{\beta _0}i\left( \partial _t\widetilde{U}\right) \Psi +\underline{%
\beta _0}i\widetilde{U}\partial _t\Psi +\underline{\beta _1}i\left( \partial
_x\widetilde{U}\right) \Psi +\underline{\beta _1}i\widetilde{U}\partial
_x\Psi + \\
+\underline{\beta _2}i\left( \partial _y\widetilde{U}\right) \Psi +%
\underline{\beta _2}i\widetilde{U}\partial _y\Psi +\underline{\beta _3}%
i\left( \partial _z\widetilde{U}\right) \Psi +\underline{\beta _3}i%
\widetilde{U}\partial _z\Psi + \\
-h\underline{n}`\gamma \widetilde{U}\Psi -h\underline{k}`\beta \widetilde{U}%
\Psi + \\
+\left( F_t\underline{`\beta _0}+F_x`\underline{\beta _1}+F_y`\underline{%
\beta _2}+F_z`\underline{\beta _3}\right) \widetilde{U}\Psi + \\
+0.5g_1\underline{Y}\left( B_t`\underline{\beta _0}+B_x`\underline{\beta _1}%
+B_y`\underline{\beta _2}+B_z`\underline{\beta _3}\right) \widetilde{U}\Psi =
\\
=0\mbox{.}
\end{array}$$
Since
$$\partial _\mu \widetilde{U}=i\frac{\partial _\mu \chi }2\left[
\begin{array}{cccc}
\exp \left( i\frac \chi 2\right) & 0 & 0 & 0 \\
0 & 2\exp \left( i\chi \right) & 0 & 0 \\
0 & 0 & \exp \left( i\frac \chi 2\right) & 0 \\
0 & 0 & 0 & 2\exp \left( i\chi \right)
\end{array}
\right]$$
then
$$\partial _\mu \widetilde{U}=-i\frac{\partial _\mu \chi }2\underline{Y}%
\widetilde{U}\mbox{;}$$
Hence
$$\begin{array}{c}
\left( \underline{\beta _0}i\widetilde{U}\partial _t+\underline{\beta _1}i%
\widetilde{U}\partial _x+\underline{\beta _2}i\widetilde{U}\partial _y+%
\underline{\beta _3}i\widetilde{U}\partial _z\right) \Psi + \\
-h\left( \underline{n}`\gamma \widetilde{U}+\underline{k}`\beta \widetilde{U}%
\right) \Psi + \\
+\left( F_t\underline{`\beta _0}+F_x`\underline{\beta _1}+F_y`\underline{%
\beta _2}+F_z`\underline{\beta _3}\right) \widetilde{U}\Psi + \\
+0.5\left(
\begin{array}{c}
\left( g_1\underline{Y}\underline{\beta _0}B_t`+\underline{\beta _0}%
\underline{Y}\partial _t\chi \right) +\left( g_1\underline{Y}\underline{%
\beta _1}B_x`+\underline{\beta _1}\underline{Y}\partial _x\chi \right) + \\
+\left( g_1\underline{Y}\underline{\beta _2}B_y`+\underline{\beta _2}%
\underline{Y}\partial _y\chi \right) +\left( g_1\underline{Y}\underline{%
\beta _3}B_z`+\underline{\beta _3}\underline{Y}\partial _z\chi \right)
\end{array}
\right) \widetilde{U}\Psi = \\
=0
\end{array}$$
Since $\underline{Y}\underline{\beta _\mu }=\underline{\beta _\mu }%
\underline{Y}$ then
$$\begin{array}{c}
\left( \underline{\beta _0}i\widetilde{U}\partial _t+\underline{\beta _1}i%
\widetilde{U}\partial _x+\underline{\beta _2}i\widetilde{U}\partial _y+%
\underline{\beta _3}i\widetilde{U}\partial _z\right) \Psi + \\
-h\left( \underline{n}`\gamma \widetilde{U}+\underline{k}`\beta \widetilde{U}%
\right) \Psi + \\
+\left( F_t\underline{`\beta _0}+F_x`\underline{\beta _1}+F_y`\underline{%
\beta _2}+F_z`\underline{\beta _3}\right) \widetilde{U}\Psi + \\
+0.5\underline{Y}\left(
\begin{array}{c}
\underline{\beta _0}\left( g_1B_t`+\partial _t\chi \right) +\underline{\beta
_1}\left( g_1B_x`+\partial _x\chi \right) + \\
+\underline{\beta _2}\left( g_1B_y`+\partial _y\chi \right) +\underline{%
\beta _3}\left( g_1B_z`+\partial _z\chi \right)
\end{array}
\right) \widetilde{U}\Psi = \\
=0
\end{array}$$
hence:
$$\begin{array}{c}
\widetilde{U}^{\dagger }\left( \underline{\beta _0}i\widetilde{U}\partial _t+%
\underline{\beta _1}i\widetilde{U}\partial _x+\underline{\beta _2}i%
\widetilde{U}\partial _y+\underline{\beta _3}i\widetilde{U}\partial
_z\right) \Psi + \\
-h\widetilde{U}^{\dagger }\left( \underline{n}`\gamma \widetilde{U}+%
\underline{k}`\beta \widetilde{U}\right) \Psi + \\
+\widetilde{U}^{\dagger }\left( F_t\underline{`\beta _0}+F_x`\underline{%
\beta _1}+F_y`\underline{\beta _2}+F_z`\underline{\beta _3}\right)
\widetilde{U}\Psi + \\
+0.5\widetilde{U}^{\dagger }\underline{Y}\left(
\begin{array}{c}
\underline{\beta _0}\left( g_1B_t`+\partial _t\chi \right) +\underline{\beta
_1}\left( g_1B_x`+\partial _x\chi \right) + \\
+\underline{\beta _2}\left( g_1B_y`+\partial _y\chi \right) +\underline{%
\beta _3}\left( g_1B_z`+\partial _z\chi \right)
\end{array}
\right) \widetilde{U}\Psi = \\
=0\mbox{.}
\end{array}$$
Because:
$$\begin{array}{c}
\widetilde{U}^{\dagger }\gamma \widetilde{U}=\cos \left( \frac \chi 2\right)
\gamma -\sin \left( \frac \chi 2\right) \beta \mbox{,} \\
\widetilde{U}^{\dagger }\beta \widetilde{U}=\cos \left( \frac \chi 2\right)
\beta +\sin \left( \frac \chi 2\right) \gamma \mbox{,} \\
\widetilde{U}^{\dagger }\widetilde{U}=1_8\mbox{,} \\
\underline{\beta _\mu }\widetilde{U}=\widetilde{U}\underline{\beta _\mu }%
\mbox{,} \\
U^{\dagger }\underline{Y}U=\underline{Y}
\end{array}$$
then
$$\begin{array}{c}
\left( \underline{\beta _0}i\partial _t+\underline{\beta _1}i\partial _x+%
\underline{\beta _2}i\partial _y+\underline{\beta _3}i\partial _z\right)
\Psi + \\
-h\left( \underline{n}`\left( \cos \left( \frac \chi 2\right) \gamma -\sin
\left( \frac \chi 2\right) \beta \right) +\underline{k}`\left( \cos \left(
\frac \chi 2\right) \beta +\sin \left( \frac \chi 2\right) \gamma \right)
\right) \Psi + \\
+\left( F_t\underline{`\beta _0}+F_x`\underline{\beta _1}+F_y`\underline{%
\beta _2}+F_z`\underline{\beta _3}\right) \Psi + \\
+0.5\underline{Y}\left(
\begin{array}{c}
\underline{\beta _0}\left( g_1B_t`+\partial _t\chi \right) +\underline{\beta
_1}\left( g_1B_x`+\partial _x\chi \right) + \\
+\underline{\beta _2}\left( g_1B_y`+\partial _y\chi \right) +\underline{%
\beta _3}\left( g_1B_z`+\partial _z\chi \right)
\end{array}
\right) \Psi = \\
=0\mbox{.}
\end{array}$$
Therefore from (\[umt1\]):
$$\begin{array}{c}
F_x`=F_x\mbox{,} \\
B_\mu `=B_\mu -\frac 1{g_1}\partial _\mu \chi \mbox{,} \\
n`=-k\sin \frac \chi 2+n\cos \frac \chi 2\mbox{,} \\
k`=k\cos \frac \chi 2+n\sin \frac \chi 2\mbox{.}
\end{array}
\label{tt4}$$
But $k$ and $n$ are an integer numbers and $k`$ and $n`$ must be an integer numbers, too.
A triplet $\left\langle l,n,k\right\rangle $ of integer numbers is [*a Fermat triplet*]{} if
$$l^2=n^2+k^2\mbox{.}$$
Let $\varepsilon $ be any tiny positive real number. An integer number $l$ is [*a father number with a precise* ]{}$\varepsilon $ if for each real number $\chi $ and for every Fermat triplet $\left\langle l,n,k\right\rangle
$ a Fermat triplet $\left\langle l,n`,k`\right\rangle $ exists for which:
$$\begin{array}{c}
\left| -k\sin \frac \chi 2+n\cos \frac \chi 2-n`\right| <\varepsilon \mbox{,}
\\
\left| k\cos \frac \chi 2+n\sin \frac \chi 2-k`\right| <\varepsilon \mbox{.}
\end{array}$$
[*For every* ]{}$\varepsilon $[*: denumerable many of a father numbers with a precise* ]{}$\varepsilon $[* exist.*]{}
Excuse me, but I mean that a masses of the real members of the particles families are defined by a father numbers with a precise $h$. I.e. denumerable many of a families exist.
Therefore for the (\[tt4\]) transformation from (\[bir1\]):
$$\begin{array}{c}
\mathbf{\Psi }\left( x_5,x_4\right) =\sum_{r=1}^4\phi _r\left( 0,0\right)
\epsilon _r+\exp \left( -ih\left( nx_5+kx_4\right) \right) \sum_{k=1}^4\phi
_k\left( n,k\right) \epsilon _k= \\
=\phi _L\left( 0,0\right) +\phi _R\left( 0,0\right) +\exp \left( -ih\left(
nx_5+kx_4\right) \right) \left( \phi _L\left( n,k\right) +\phi _R\left(
n,k\right) \right) \rightarrow
\end{array}$$
$$\begin{array}{c}
\rightarrow \mathbf{\Psi }`\left( x_5,x_4\right) = \\
=\exp \left( i\frac \chi 2\right) \phi _L\left( 0,0\right) +\exp \left(
i\chi \right) \phi _R\left( 0,0\right) + \\
+\exp \left( -ih\left( \left( -k\sin \frac \chi 2+n\cos \frac \chi 2\right)
x_5+\left( k\cos \frac \chi 2+n\sin \frac \chi 2\right) x_4\right) \right)
\cdot \\
\cdot \left( \exp \left( i\frac \chi 2\right) \phi _L\left( n,k\right) +\exp
\left( i\chi \right) \phi _R\left( n,k\right) \right) \mbox {.}
\end{array}$$
#### $U^{\left( -\right) }$ transformation
$U^{\left( -\right) }$ has got the following eigenvalues and eigenvectors:
for the eigenvalue $1$: eigenvectors:
$$\underline{\iota }_1=\left[
\begin{array}{c}
0 \\
0 \\
1 \\
0 \\
0 \\
0 \\
0 \\
0
\end{array}
\right] ,\underline{\iota }_2=\left[
\begin{array}{c}
0 \\
0 \\
0 \\
1 \\
0 \\
0 \\
0 \\
0
\end{array}
\right] ,\underline{\iota }_5=\left[
\begin{array}{c}
0 \\
0 \\
0 \\
0 \\
0 \\
0 \\
1 \\
0
\end{array}
\right] ,\underline{\iota }_6=\left[
\begin{array}{c}
0 \\
0 \\
0 \\
0 \\
0 \\
0 \\
0 \\
1
\end{array}
\right] ; \label{si1}$$
for the eigenvalue $w=a+i\sqrt{1-a^2}$: eigenvectors:
$$\underline{\iota }_3=\frac 1{\sqrt{2}\sqrt{\sqrt{1-a^2}\left( b+\sqrt{1-a^2}%
\right) }}\left[
\begin{array}{c}
b+\sqrt{1-a^2} \\
0 \\
0 \\
0 \\
ic+g \\
0 \\
0 \\
0
\end{array}
\right] , \label{si2}$$
$$\underline{\iota }_4=\frac 1{\sqrt{2}\sqrt{\sqrt{1-a^2}\left( b+\sqrt{1-a^2}%
\right) }}\left[
\begin{array}{c}
0 \\
b+\sqrt{1-a^2} \\
0 \\
0 \\
0 \\
ic+g \\
0 \\
0
\end{array}
\right] \mbox{;} \label{si3}$$
for eigenvalue $w^{*}=a-i\sqrt{1-a^2}$: eigenvectors:
$$\underline{\iota }_7=\frac 1{\sqrt{2}\sqrt{\sqrt{1-a^2}\left( b+\sqrt{1-a^2}%
\right) }}\left[
\begin{array}{c}
ic-g \\
0 \\
0 \\
0 \\
b+\sqrt{1-a^2} \\
0 \\
0 \\
0
\end{array}
\right] \mbox{,} \label{si4}$$
$$\underline{\iota }_8=\frac 1{\sqrt{2}\sqrt{\sqrt{1-a^2}\left( b+\sqrt{1-a^2}%
\right) }}\left[
\begin{array}{c}
0 \\
ic-g \\
0 \\
0 \\
0 \\
b+\sqrt{1-a^2} \\
0 \\
0
\end{array}
\right] \mbox{.} \label{si5}$$
Hence the space of $U^{\left( -\right) }$ is divided on three orthogonal subspace:
the 4-dimensional $\mathcal{U}_1^{\left( -\right) }$on the basis $%
\left\langle \underline{\iota }_1,\underline{\iota }_2,\underline{\iota }_5,%
\underline{\iota }_6\right\rangle $ with eigenvalue $1$,
the 2-dimensional $\mathcal{U}_w^{\left( -\right) }$on the basis $%
\left\langle \underline{\iota }_3,\underline{\iota }_4\right\rangle $ with eigenvalue $w$ and
the 2-dimensional $\mathcal{U}_{w^{*}}^{\left( -\right) }$on the basis $%
\left\langle \underline{\iota }_7,\underline{\iota }_8\right\rangle $ with eigenvalue $w^{*}$.
Let
$$\widehat{\underline{\iota }}_k=\gamma \underline{\iota }_k\mbox{.}$$
In this case $\left\langle \widehat{\underline{\iota }}_3,\widehat{%
\underline{\iota }}_4,\widehat{\underline{\iota }}_7,\widehat{\underline{%
\iota }}_8\right\rangle $ is the orthonormal basis of $\mathcal{U}_1^{\left(
-\right) }$.
Let $\mathcal{U}_{\circ }^{\left( -\right) }$ be the space on the basis $%
\left\langle \widehat{\underline{\iota }}_3,\widehat{\underline{\iota }}_4,%
\underline{\iota }_3,\underline{\iota }_4\right\rangle $ and $\mathcal{U}%
_{*}^{\left( -\right) }$ be the space on the basis $\left\langle \widehat{%
\underline{\iota }}_7,\widehat{\underline{\iota }}_8,\underline{\iota }_7,%
\underline{\iota }_8\right\rangle $.
$$\begin{array}{c}
\Psi _{\circ }=\pi _{\circ }\Psi \mbox{, }\Psi _{*}=\pi _{*}\Psi \mbox{,} \\
\Psi _{\circ }\in \mathcal{U}_{\circ }^{\left( -\right) }\mbox{ and }\Psi
_{*}\in \mathcal{U}_{*}^{\left( -\right) }\mbox{.}
\end{array}
\label{ux}$$
In this case:
$$\pi _{\circ }=\frac 1{2\sqrt{1-a^2}}\left[
\begin{array}{cc}
\left( b+\sqrt{1-a^2}\right) 1_4 & \left( -ic+g\right) \gamma _5 \\
\left( ic+g\right) \gamma _5 & \left( \sqrt{1-a^2}-b\right) 1_4
\end{array}
\right] \mbox{,}$$
$$\pi _{*}=\frac 1{2\sqrt{1-a^2}}\left[
\begin{array}{cc}
\left( \sqrt{1-a^2}-b\right) 1_4 & \left( ic-g\right) \gamma _5 \\
\left( -g-ic\right) \gamma _5 & \left( b+\sqrt{1-a^2}\right) 1_4
\end{array}
\right] \mbox{.}$$
Hence
$$\begin{array}{c}
\mathbf{\Psi }_{\circ }\left( x_5,x_4\right) =\frac 1{2\sqrt{1-a^2}}\cdot \\
\cdot \left(
\begin{array}{c}
\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \phi _L\left( 0,0\right)
+\phi _R\left( 0,0\right) \right) - \\
-\left( ic-g\right) \left( \phi _L\left( n,k\right) -\phi _R\left(
n,k\right) \right) + \\
+\exp \left( -ih\left( nx_5+kx_4\right) \right) \cdot \\
\cdot \left(
\begin{array}{c}
\left( ic+g\right) \left( \phi _L\left( 0,0\right) -\phi _R\left( 0,0\right)
\right) + \\
+\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \phi _L\left( n,k\right)
+\phi _R\left( n,k\right) \right)
\end{array}
\right)
\end{array}
\right) \mbox{,}
\end{array}
\label{fr1}$$
$$\begin{array}{c}
\mathbf{\Psi }_{*}\left( x_5,x_4\right) =\frac 1{2\sqrt{1-a^2}}\cdot \\
\cdot \left(
\begin{array}{c}
\left( \sqrt{\left( 1-a^2\right) }-b\right) \left( \phi _L\left( 0,0\right)
+\phi _R\left( 0,0\right) \right) + \\
+\left( ic-g\right) \left( \phi _L\left( n,k\right) -\phi _R\left(
n,k\right) \right) + \\
+\exp \left( -ih\left( nx_5+kx_4\right) \right) \cdot \\
\cdot \left(
\begin{array}{c}
\left( g+ic\right) \left( -\phi _L\left( 0,0\right) +\phi _R\left(
0,0\right) \right) + \\
+\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \phi _L\left( n,k\right)
+\phi _R\left( n,k\right) \right)
\end{array}
\right)
\end{array}
\right) \mbox{.}
\end{array}
\label{fr2}$$
If $\lambda $ is the angle of the $U^{\left( -\right) }$ eigenvalue (i.e. $%
w=a+i\sqrt{1-a^2}$ and $\cos \lambda =a$ and $\sin \lambda =\sqrt{1-a^2}$ ) then
$$\begin{array}{c}
U^{\left( -\right) \dagger }\gamma U^{\left( -\right) }=\left( \gamma \cos
\lambda +\sin \lambda \left( \pi _{\circ }-\pi _{*}\right) \beta \right) %
\mbox{,} \\
U^{\left( -\right) \dagger }\beta U^{\left( -\right) }=\left( \beta \cos
\lambda -\sin \lambda \left( \pi _{\circ }-\pi _{*}\right) \gamma \right) %
\mbox{.}
\end{array}
\label{gb}$$
Let
$$\begin{array}{c}
\underline{n}\rightarrow \underline{n}`=\left( \underline{n}\cos \lambda +%
\underline{k}\sin \lambda \left( \pi _{\circ }-\pi _{*}\right) \right) %
\mbox{,} \\
\underline{k}\rightarrow \underline{k}`=\left( \underline{k}\cos \lambda -%
\underline{n}\sin \lambda \left( \pi _{\circ }-\pi _{*}\right) \right) %
\mbox{,} \\
\Psi \rightarrow \Psi `=U^{\left( -\right) }\Psi \mbox{,} \\
F_\mu \rightarrow F_\mu `=F_\mu \mbox{,} \\
B_\mu \rightarrow B_\mu `=B_\mu
\end{array}
\label{tt}$$
and the motion equation for $\Psi `$ be (\[umt1\]):
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }i\partial _\mu \Psi `-h\left(
\underline{n}`\gamma +\underline{k}`\beta \right) \Psi `+\sum_{\mu
=0}^3F_\mu `\underline{\beta _\mu }\Psi `+ \\
+0.5g_1\underline{Y}\sum_{\mu =0}^3B_\mu `\underline{\beta _\mu }\Psi
`=S\Psi `\mbox{.}
\end{array}
\label{um2}$$
Hence:
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu U^{\left(
-\right) }\right) \Psi +\sum_{\mu =0}^3\underline{\beta _\mu }iU^{\left(
-\right) }\left( \partial _\mu \Psi \right) - \\
-h\left( \underline{n}`\gamma +\underline{k}`\beta \right) \left( U^{\left(
-\right) }\Psi \right) +\sum_{\mu =0}^3F_\mu `\underline{\beta _\mu }\left(
U^{\left( -\right) }\Psi \right) + \\
+0.5g_1\underline{Y}\sum_{\mu =0}^3B_\mu `\underline{\beta _\mu }\left(
U^{\left( -\right) }\Psi \right) =S\left( U^{\left( -\right) }\Psi \right) %
\mbox{,}
\end{array}$$
then
$$\begin{array}{c}
U^{\left( -\right) \dagger }\sum_{\mu =0}^3\underline{\beta _\mu }i\left(
\partial _\mu U^{\left( -\right) }\right) \Psi +U^{\left( -\right) \dagger
}\sum_{\mu =0}^3\underline{\beta _\mu }iU^{\left( -\right) }\left( \partial
_\mu \Psi \right) - \\
-U^{\left( -\right) \dagger }h\left(
\begin{array}{c}
\left( \underline{n}\cos \lambda +\underline{k}\sin \lambda \left( \pi
_{\circ }-\pi _{*}\right) \right) \gamma + \\
+\left( \underline{k}\cos \lambda -\underline{n}\sin \lambda \left( \pi
_{\circ }-\pi _{*}\right) \right) \beta
\end{array}
\right) U^{\left( -\right) }\Psi + \\
+U^{\left( -\right) \dagger }\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }%
U^{\left( -\right) }\Psi ++U^{\left( -\right) \dagger }0.5g_1\underline{Y}%
\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }U^{\left( -\right) }\Psi = \\
=U^{\left( -\right) \dagger }SU^{\left( -\right) }\Psi \mbox{.}
\end{array}$$
Since
$$\begin{array}{c}
U^{\left( -\right) \dagger }\underline{\beta _\mu }=\underline{\beta _\mu }%
U^{\left( -\right) \dagger }\mbox{,} \\
\underline{Y}U^{\left( -\right) }=U^{\left( -\right) }\underline{Y}\mbox{,}
\\
U^{\left( -\right) \dagger }\left( \pi _{\circ }-\pi _{*}\right) =\left( \pi
_{\circ }-\pi _{*}\right) U^{\left( -\right) \dagger }
\end{array}$$
then
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }iU^{\left( -\right) \dagger }\left(
\partial _\mu U^{\left( -\right) }\right) \Psi +\sum_{\mu =0}^3\underline{%
\beta _\mu }i\partial _\mu \Psi - \\
-h\left(
\begin{array}{c}
\left( \underline{n}\cos \lambda +\underline{k}\sin \lambda \left( \pi
_{\circ }-\pi _{*}\right) \right) U^{\left( -\right) \dagger }\gamma
U^{\left( -\right) }+ \\
+\left( \underline{k}\cos \lambda -\underline{n}\sin \lambda \left( \pi
_{\circ }-\pi _{*}\right) \right) U^{\left( -\right) \dagger }\beta
U^{\left( -\right) }
\end{array}
\right) \Psi + \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\Psi +0.5g_1\underline{Y}%
\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\Psi = \\
=U^{\left( -\right) \dagger }SU^{\left( -\right) }\Psi \mbox{.}
\end{array}$$
From (\[gb\]):
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }iU^{\left( -\right) \dagger }\left(
\partial _\mu U^{\left( -\right) }\right) \Psi +\sum_{\mu =0}^3\underline{%
\beta _\mu }i\partial _\mu \Psi - \\
-h\left(
\begin{array}{c}
\left( \underline{n}\cos \lambda +\underline{k}\sin \lambda \left( \pi
_{\circ }-\pi _{*}\right) \right) \left( \gamma \cos \lambda +\sin \lambda
\left( \pi _{\circ }-\pi _{*}\right) \beta \right) + \\
+\left( \underline{k}\cos \lambda -\underline{n}\sin \lambda \left( \pi
_{\circ }-\pi _{*}\right) \right) \left( \beta \cos \lambda -\sin \lambda
\left( \pi _{\circ }-\pi _{*}\right) \gamma \right)
\end{array}
\right) \Psi + \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\Psi +0.5g_1\underline{Y}%
\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\Psi = \\
=U^{\left( -\right) \dagger }SU^{\left( -\right) }\Psi \mbox{.}
\end{array}$$
Since
$$\left( \pi _{\circ }-\pi _{*}\right) \left( \pi _{\circ }-\pi _{*}\right)
=1_8$$
then
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }iU^{\left( -\right) \dagger }\left(
\partial _\mu U^{\left( -\right) }\right) \Psi + \\
+\sum_{\mu =0}^3\underline{\beta _\mu }i\partial _\mu \Psi -h\left(
\underline{n}\gamma +\underline{k}\beta \right) \Psi +\sum_{\mu =0}^3F_\mu
\underline{\beta _\mu }\Psi +0.5g_1\underline{Y}\sum_{\mu =0}^3B_\mu
\underline{\beta _\mu }\Psi = \\
=U^{\left( -\right) \dagger }SU^{\left( -\right) }\Psi \mbox{.}
\end{array}$$
Hence from (\[umt1\]):
$$\sum_{\mu =0}^3\underline{\beta _\mu }iU^{\left( -\right) \dagger }\left(
\partial _\mu U^{\left( -\right) }\right) =U^{\left( -\right) \dagger
}SU^{\left( -\right) }$$
and
$$S=\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu U^{\left(
-\right) }\right) U^{\left( -\right) \dagger }$$
Therefore from (\[um2\]) the motion equation for the transformation (\[tt\]) is the following:
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }i\partial _\mu \Psi `-\sum_{\mu =0}^3%
\underline{\beta _\mu }i\left( \partial _\mu U^{\left( -\right) }\right)
U^{\left( -\right) \dagger }\Psi `- \\
-h\left( \underline{n}`\gamma +\underline{k}`\beta \right) \Psi `+ \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\Psi `+0.5g_1\underline{Y}%
\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\Psi `=0\mbox{.}
\end{array}
\label{um3}$$
##### $W$-bosons
Let $g_2$ be a positive real number.
If design ($a,b,c,g$ form $U^{\left( -\right) }$):
$$\begin{array}{c}
W_{0,\mu }=-2\frac 1{g_2g}\left(
\begin{array}{c}
g\left( \partial _\mu a\right) b-g\left( \partial _\mu b\right) a+\left(
\partial _\mu c\right) g^2+ \\
+a\left( \partial _\mu a\right) c+b\left( \partial _\mu b\right) c+c^2\left(
\partial _\mu c\right)
\end{array}
\right) \\
W_{1,\mu }=-2\frac 1{g_2g}\left(
\begin{array}{c}
\left( \partial _\mu a\right) a^2-bg\left( \partial _\mu c\right) +a\left(
\partial _\mu b\right) b+ \\
+a\left( \partial _\mu c\right) c+g^2\left( \partial _\mu a\right) +c\left(
\partial _\mu b\right) g
\end{array}
\right) \\
W_{2,\mu }=-2\frac 1{g_2g}\left(
\begin{array}{c}
g\left( \partial _\mu a\right) c-a\left( \partial _\mu a\right) b-b^2\left(
\partial _\mu b\right) - \\
-c\left( \partial _\mu c\right) b-\left( \partial _\mu b\right) g^2-\left(
\partial _\mu c\right) ga
\end{array}
\right)
\end{array}$$
and
$$W_\mu =\left[
\begin{array}{cccc}
W_{0,\mu }1_2 & 0_2 & \left( W_{1,\mu }-iW_{2,\mu }\right) 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & 0_2 \\
\left( W_{1,\mu }+iW_{2,\mu }\right) 1_2 & 0_2 & -W_{0,\mu }1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & 0_2
\end{array}
\right]$$
then
$$-i\left( \partial _\mu U^{\left( -\right) }\right) U^{\left( -\right)
\dagger }=\frac 12g_2W_\mu \label{w}$$
and from (\[um3\]):
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu -i\frac
12g_2W_\mu \right) \Psi `- \\
-h\left( \underline{n}`\gamma +\underline{k}`\beta \right) \Psi `+ \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\Psi `+0.5g_1\underline{Y}%
\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\Psi `=0\mbox{.}
\end{array}
\label{hW}$$
Let
$$\stackrel{,}{U}=\left[
\begin{array}{cccc}
\left( \stackrel{,}{a}+i\stackrel{,}{b}\right) 1_2 & 0_2 & \left( \stackrel{,%
}{c}+i\stackrel{,}{g}\right) 1_2 & 0_2 \\
0_2 & 1_2 & 0_2 & 0_2 \\
\left( -\stackrel{,}{c}+i\stackrel{,}{g}\right) 1_2 & 0_2 & \left( \stackrel{%
,}{a}-i\stackrel{,}{b}\right) 1_2 & 0_2 \\
0_2 & 0_2 & 0_2 & 1_2
\end{array}
\right] \mbox{,}$$
$$\stackrel{,}{\pi }_{\circ }=\frac 1{2\sqrt{1-\stackrel{,}{a}^2}}\left[
\begin{array}{cc}
\left( \stackrel{,}{b}+\sqrt{1-\stackrel{,}{a}^2}\right) 1_4 & \left( -i%
\stackrel{,}{c}+\stackrel{,}{g}\right) \gamma _5 \\
\left( i\stackrel{,}{c}+\stackrel{,}{g}\right) \gamma _5 & \left( \sqrt{1-%
\stackrel{,}{a}^2}-\stackrel{,}{b}\right) 1_4
\end{array}
\right] \mbox{,}$$
$$\stackrel{,}{\pi }_{*}=\frac 1{2\sqrt{1-\stackrel{,}{a}^2}}\left[
\begin{array}{cc}
\left( \sqrt{1-\stackrel{,}{a}^2}-\stackrel{,}{b}\right) 1_4 & \left( i%
\stackrel{,}{c}-\stackrel{,}{g}\right) \gamma _5 \\
\left( -\stackrel{,}{g}-i\stackrel{,}{c}\right) \gamma _5 & \left( \stackrel{%
,}{b}+\sqrt{1-\stackrel{,}{a}^2}\right) 1_4
\end{array}
\right] \mbox{.}$$
Let:
$$\cos \stackrel{,}{\lambda }=\stackrel{,}{a}\mbox{ and }\sin \stackrel{,}{%
\lambda }=\sqrt{1-\stackrel{,}{a}^2}$$
and
$$\begin{array}{c}
\Psi `\rightarrow \Psi ^{\prime }=\left( \stackrel{,}{U}\Psi `\right) %
\mbox{,} \\
\underline{n}`\rightarrow \underline{n}^{\prime }=\left( \underline{n}`\cos
\stackrel{,}{\lambda }+\underline{k}`\sin \stackrel{,}{\lambda }\left(
\stackrel{,}{\pi }_{\circ }-\stackrel{,}{\pi }_{*}\right) \right) \mbox{,}
\\
\underline{k}`\rightarrow \underline{k}^{\prime }=\left( \underline{k}`\cos
\stackrel{,}{\lambda }-\underline{n}`\sin \stackrel{,}{\lambda }\left(
\stackrel{,}{\pi }_{\circ }-\stackrel{,}{\pi }_{*}\right) \right) \mbox{,}
\\
F_\mu \rightarrow F_\mu ^{\prime }=F_\mu \mbox{,} \\
B_\mu \rightarrow B_\mu ^{\prime }=B_\mu \mbox{,}
\end{array}
\label{tt2}$$
and
$$W_\mu \rightarrow W_\mu ^{\prime }\mbox{.}$$
In that case from (\[w\]):
$$W_\mu ^{\prime }=-\frac{2i}{g_2}\left( \partial _\mu \left( \stackrel{,}{U}%
U^{\left( -\right) }\right) \right) \left( \stackrel{,}{U}U^{\left( -\right)
}\right) ^{\dagger };$$
Hence:
$$W_\mu ^{\prime }=-\frac{2i}{g_2}\left( \partial _\mu \stackrel{,}{U}\right)
\stackrel{,}{U}^{\dagger }-\frac{2i}{g_2}\stackrel{,}{U}\left( \partial _\mu
U^{\left( -\right) }\right) U^{\left( -\right) \dagger }\stackrel{,}{U}%
^{\dagger }\mbox{;}$$
i.e.:
$$W_\mu ^{\prime }=\stackrel{,}{U}W_\mu \stackrel{,}{U}^{\dagger }-\frac{2i}{%
g_2}\left( \partial _\mu \stackrel{,}{U}\right) \stackrel{,}{U}^{\dagger }%
\mbox{.}$$
If
$$F_{\mu ,\nu }=\left( \partial _\mu W_\nu -\partial _\nu W_\mu -i\frac{g_2}%
2\left( W_\mu W_\nu -W_\nu W_\mu \right) \right)$$
then
$$F_{\mu ,\nu }^{\prime }=\partial _\mu W_\nu ^{\prime }-\partial _\nu W_\mu
^{\prime }-i\frac{g_2}2\left( W_\mu ^{\prime }W_\nu ^{\prime }-W_\nu
^{\prime }W_\mu ^{\prime }\right) =UF_{\mu ,\nu }U^{\dagger }\mbox{.}$$
Therefore $F_{\mu ,\nu }$ is invariant for the transformation (\[tt2\]).
The Lagrangian for $F_{\mu ,\nu }$:
$$\mathcal{L}_F=\left( -\frac 14\sum_{\mu ,\nu }F^{\mu ,\nu }F_{\mu ,\nu
}\right) \mbox{.}$$
Hence the Euler-Lagrange equations for $W_\mu $ are the following:
$$\sum_\nu \partial ^\nu \left( \partial _\mu W_\nu -\partial _\nu W_\mu -i%
\frac{g_2}2\left[ W_\mu ,W_\nu \right] \right) =0\mbox{.}$$
For the components:
$$\begin{array}{c}
\sum_\nu \partial ^\nu \partial _\nu W_{0,\mu }=g_2\sum_\nu \partial ^\nu
\left( W_{1,\mu }W_{2,\nu }-W_{2,\mu }W_{1,\nu }\right) +\partial _\mu
\sum_\nu \partial ^\nu W_{0,\nu }\mbox{,} \\
\sum_\nu \partial ^\nu \partial _\nu W_{1,\mu }=g_2\sum_\nu \partial ^\nu
\left( W_{0,\nu }W_{2,\mu }-W_{0,\mu }W_{2,\nu }\right) +\partial _\mu
\sum_\nu \partial ^\nu W_{1,\nu }\mbox{,} \\
\sum_\nu \partial ^\nu \partial _\nu W_{2,\mu }=g_2\sum_\nu \partial ^\nu
\left( W_{0,\mu }W_{1,\nu }-W_{0,\nu }W_{1,\mu }\right) +\partial _\mu
\sum_\nu \partial ^\nu W_{2,\nu }\mbox{.}
\end{array}$$
Let:
$$\begin{array}{c}
\alpha _{0,\mu ,\nu }=\partial _\nu W_{0,\mu }-g_2\left( W_{1,\mu }W_{2,\nu
}-W_{2,\mu }W_{1,\nu }\right) \mbox{,} \\
\alpha _{1,\mu ,\nu }=\partial _\nu W_{1,\mu }-g_2\left( W_{0,\nu }W_{2,\mu
}-W_{0,\mu }W_{2,\nu }\right) \mbox{,} \\
\alpha _{2,\mu ,\nu }=\partial _\nu W_{2,\mu }-g_2\left( W_{0,\mu }W_{1,\nu
}-W_{0,\nu }W_{1,\mu }\right) \mbox{.}
\end{array}
\label{A1}$$
Hence if $\sum_\nu \partial ^\nu W_\nu =0$ then
$$\sum_\nu \partial ^\nu \alpha _{0,\mu ,\nu }=0,\sum_\nu \partial ^\nu \alpha
_{1,\mu ,\nu }=0,\sum_\nu \partial ^\nu \alpha _{2,\mu ,\nu }=0\mbox{.}$$
From (\[A1\]):
$$\partial _\nu W_{0,\mu }=\left( g_2\left( W_{1,\mu }W_{2,\nu }-W_{2,\mu
}W_{1,\nu }\right) +\alpha _{0,\mu ,\nu }\right) \mbox{,} \label{b1}$$
$$\partial _\nu W_{1,\mu }=\left( g_2\left( W_{0,\nu }W_{2,\mu }-W_{0,\mu
}W_{2,\nu }\right) +\alpha _{1,\mu ,\nu }\right) \mbox{,} \label{b2}$$
$$\partial _\nu W_{2,\mu }=\left( g_2\left( W_{0,\mu }W_{1,\nu }-W_{0,\nu
}W_{1,\mu }\right) +\alpha _{2,\mu ,\nu }\right) \mbox{;} \label{b3}$$
from (\[b1\]):
$$\begin{array}{c}
\partial _\nu \partial _\nu W_{0,\mu }=g_2\partial _\nu \left( W_{1,\mu
}W_{2,\nu }-W_{2,\mu }W_{1,\nu }\right) +\partial _\nu \alpha _{0,\mu ,\nu }=
\\
=g_2\left( \partial _\nu W_{1,\mu }W_{2,\nu }+W_{1,\mu }\partial _\nu
W_{2,\nu }-\partial _\nu W_{2,\mu }W_{1,\nu }-W_{2,\mu }\partial _\nu
W_{1,\nu }\right) +\partial _\nu \alpha _{0,\mu ,\nu };
\end{array}
\label{b4}$$
hence from (\[b4\]), (\[b2\]) and (\[b3\]):
$$\begin{aligned}
\partial _\nu \partial _\nu W_{0,\mu } &=&g_2\left(
\begin{array}{c}
\left( g_2\left( W_{0,\nu }W_{2,\mu }-W_{0,\mu }W_{2,\nu }\right) +\alpha
_{1,\mu ,\nu }\right) W_{2,\nu }- \\
-\left( g_2\left( W_{0,\mu }W_{1,\nu }-W_{0,\nu }W_{1,\mu }\right) +\alpha
_{2,\mu ,\nu }\right) W_{1,\nu }- \\
-W_{2,\mu }\partial _\nu W_{1,\nu }+W_{1,\mu }\partial _\nu W_{2,\nu }
\end{array}
\right) + \\
&&\ +\partial _\nu \alpha _{0,\mu ,\nu }\mbox{;}\end{aligned}$$
hence:
$$\begin{array}{c}
\partial _\nu \partial _\nu W_{0,\mu }= \\
=g_2\left(
\begin{array}{c}
g_2\left( -\left( W_{2,\nu }^2+W_{1,\nu }^2\right) W_{0,\mu }+\left(
W_{1,\mu }W_{1,\nu }+W_{2,\mu }W_{2,\nu }\right) W_{0,\nu }\right) + \\
+\alpha _{1,\mu ,\nu }W_{2,\nu }-\alpha _{2,\mu ,\nu }W_{1,\nu }+W_{1,\mu
}\partial _\nu W_{2,\nu }-W_{2,\mu }\partial _\nu W_{1,\nu }
\end{array}
\right) + \\
+\partial _\nu \alpha _{0,\mu ,\nu }\mbox{;}
\end{array}$$
and
$$\begin{array}{c}
\partial _\nu \partial _\nu W_{0,\mu }=-g_2^2\left( W_{2,\nu }^2+W_{1,\nu
}^2+W_{0,\nu }^2\right) W_{0,\mu }+ \\
+g_2^2\left( W_{0,\mu }W_{0,\nu }+W_{1,\mu }W_{1,\nu }+W_{2,\mu }W_{2,\nu
}\right) W_{0,\nu }+ \\
+g_2\left( \alpha _{1,\mu ,\nu }W_{2,\nu }-\alpha _{2,\mu ,\nu }W_{1,\nu
}+W_{1,\mu }\partial _\nu W_{2,\nu }-W_{2,\mu }\partial _\nu W_{1,\nu
}\right) +\partial _\nu \alpha _{0,\mu ,\nu }\mbox{;}
\end{array}$$
if $\sum_\nu \partial _\nu W_\nu =0$ then:
$$\begin{array}{c}
\sum_\nu \partial _\nu \partial _\nu W_{0,\mu }=-g_2^2W_{0,\mu }\sum_\nu
W_\nu ^2+ \\
+\frac{g_2^2}2\sum_\nu \left( W_\mu W_\nu +W_\nu W_\mu \right) W_{0,\nu
}+g_2\sum_\nu \left( \alpha _{1,\mu ,\nu }W_{2,\nu }-\alpha _{2,\mu ,\nu
}W_{1,\nu }\right) \mbox{,}
\end{array}
\label{d1}$$
$$\begin{array}{c}
\sum_\nu \partial _\nu \partial _\nu W_{1,\mu }=-g_2^2W_{1,\mu }\sum_\nu
W_\nu ^2+ \\
+\frac{g_2^2}2\sum_\nu \left( W_\nu W_\mu +W_\mu W_\nu \right) W_{1,\nu
}+g_2\sum_\nu \left( W_{0,\nu }\alpha _{2,\mu ,\nu }-\alpha _{0,\mu ,\nu
}W_{2,\nu }\right)
\end{array}$$
and
$$\begin{array}{c}
\sum_\nu \partial _\nu \partial _\nu W_{2,\mu }=-g_2^2W_{2,\mu }\sum_\nu
W_\nu ^2+ \\
+\frac{g_2^2}2\sum_\nu \left( W_\nu W_\mu +W_\mu W_\nu \right) W_{2,\nu
}+g_2\sum_\nu \left( \alpha _{0,\mu ,\nu }W_{1,\nu }-W_{0,\nu }\alpha
_{1,\mu ,\nu }\right) \mbox{.}
\end{array}$$
$$\alpha _{\mu ,\nu }=\left[
\begin{array}{cc}
\alpha _{0,\mu ,\nu } & \alpha _{1,\mu ,\nu }-i\alpha _{2,\mu ,\nu } \\
\alpha _{1,\mu ,\nu }+i\alpha _{2,\mu ,\nu } & -\alpha _{0,\mu ,\nu }
\end{array}
\right]$$
then
$$\begin{array}{c}
\sum_\nu \partial _\nu \partial _\nu W_\mu =-g_2^2W_\mu \sum_\nu W_\nu ^2+
\\
+\frac{g_2^2}2\sum_\nu \left( W_\nu W_\mu +W_\mu W_\nu \right) W_\nu -i\frac{%
g_2^2}2\sum_\nu \left( \alpha _{\mu ,\nu }W_\nu -W_\nu \alpha _{\mu ,\nu
}\right) \mbox{.}
\end{array}$$
It is the motion equation for the field $W_\mu $ which has got a less than unit 1 velocity. That is this field does not behave as a massless field.
Hence although $F_{\mu ,\nu }$ is a massless field but its components $W_\mu
$ do not behave like a massless fields.
If
$$\sum_\nu \left( W_\nu \frac{\partial W_\nu }{\partial W_\mu }+\frac{\partial
W_\nu }{\partial W_\mu }W_\nu \right) =0$$
then a real $\upsilon $ exists for which
$$\upsilon =\left( 2\sum_\nu W_\nu ^2\right) ^{\frac 12} \label{mw}$$
and
$$\partial _{W_\mu }\upsilon =0$$
then the Lagrangian of $W_\mu $ is:
$$\begin{array}{c}
\widehat{\mathcal{L}}=\sum_\nu \left( \partial _\nu W_\mu \right) \left(
\partial _\nu W_\mu \right) -g_2^2\frac{\upsilon ^2}2W_\mu ^2+ \\
+\frac{g_2^2}4\sum_\nu \left( W_\nu W_\mu +W_\mu W_\nu \right) ^2- \\
-i\frac{g_2^2}2\left( \left( \sum_\nu \left[ \alpha _{\mu ,\nu },W_\nu
\right] \right) W_\mu +W_\mu \left( \sum_\nu \left[ \alpha _{\mu ,\nu
},W_\nu \right] \right) \right) \mbox{.}
\end{array}$$
It is a lagrangian of a field with mass
$$M=g_2\frac \upsilon {\sqrt{2}}$$
and $M>0$.
##### $A$ and $Z$ bosons
Let $A_\mu $ and $Z_\mu $ are a fields for which [@KN]:
$$Z_\mu =\frac 1{\sqrt{g_1^2+g_2^2}}\left( g_2W_{0,\mu }-g_1B_\mu \right) %
\mbox{, }A_\mu =\frac 1{\sqrt{g_1^2+g_2^2}}\left( g_2B_\mu +g_1W_{0,\mu
}\right) \label{c2}$$
and
$$\sum_\nu \partial ^\nu \partial _\nu A_\mu =0\mbox{.} \label{c3}$$
Let denote:
$$\frac{g_2^2}2\sum_\nu \left( W_\mu W_\nu +W_\nu W_\mu \right) W_{0,\nu
}+g_2\sum_\nu \left( \alpha _{1,\mu ,\nu }W_{2,\nu }-\alpha _{2,\mu ,\nu
}W_{1,\nu }\right) =\Lambda \mbox{.}$$
Hence from (\[d1\]) and (\[mw\]):
$$\sum_\nu \partial _\nu \partial _\nu W_{0,\mu }=-g_2^2\frac{\upsilon ^2}%
2W_{0,\mu }+\Lambda \label{c4}$$
From (\[c2\]):
$$B_\mu =\frac 1{\sqrt{g_1^2+g_2^2}}\left( g_2A_\mu -g_1Z_\mu \right) \mbox{, }%
W_\mu ^0=\frac 1{\sqrt{g_1^2+g_2^2}}\left( g_1A_\mu +g_2Z_\mu \right) %
\mbox{.} \label{c1}$$
and
$$\sum_\nu \partial ^\nu \partial _\nu A_\mu =\frac 1{\sqrt{g_1^2+g_2^2}%
}\left( g_2\sum_\nu \partial ^\nu \partial _\nu B_\mu +g_1\sum_\nu \partial
^\nu \partial _\nu W_{0,\mu }\right) \mbox{,}$$
from (\[c4\]):
$$\sum_\nu \partial ^\nu \partial _\nu A_\mu =\frac 1{\sqrt{g_1^2+g_2^2}%
}\left(
\begin{array}{c}
g_2\left( \sum_\nu \partial ^\nu \partial _\nu B_\mu +g_1^2\frac{\upsilon ^2}%
2B_\mu -g_1^2\frac{\upsilon ^2}2B_\mu \right) + \\
+g_1\left( -g_2^2\frac{\upsilon ^2}2W_{0,\mu }+\Lambda \right)
\end{array}
\right) \mbox{,}$$
from (\[c1\])
$$\begin{aligned}
\sum_\nu \partial ^\nu \partial _\nu A_\mu &=&-\frac{\upsilon ^2}%
2g_1g_2\frac 1{g_1^2+g_2^2}\left( 2g_1g_2A_\mu +\left( g_2^2-g_1^2\right)
Z_\mu \right) + \\
&&\ \ +\frac 1{\sqrt{g_1^2+g_2^2}}\left( g_1\Lambda +g_2\left( \sum_\nu
\partial ^\nu \partial _\nu B_\mu +g_1^2\frac{\upsilon ^2}2B_\mu \right)
\right) \mbox{,}\end{aligned}$$
from (\[c3\]):
$$\begin{array}{c}
A_\mu =-\left( g_2^2-g_1^2\right) \frac 1{2g_1g_2}Z_\mu + \\
+\frac 1{\upsilon ^2\left( g_1g_2\right) ^2}\sqrt{g_1^2+g_2^2}\left(
g_1\Lambda +g_2\left( \sum_\nu \partial ^\nu \partial _\nu B_\mu +g_1^2\frac{%
\upsilon ^2}2B_\mu \right) \right)
\end{array}
\label{q1}$$
From (\[c2\]):
$$\sum_\nu \partial ^\nu \partial _\nu Z_\mu =\frac 1{\sqrt{g_1^2+g_2^2}%
}\left( g_2\sum_\nu \partial ^\nu \partial _\nu W_{0,\mu }-g_1\sum_\nu
\partial ^\nu \partial _\nu B_\mu \right) \mbox{,}$$
from (\[c4\]):
$$\sum_\nu \partial ^\nu \partial _\nu Z_\mu =\frac 1{\sqrt{g_1^2+g_2^2}%
}\left(
\begin{array}{c}
g_2\left( -g_2^2\frac{\upsilon ^2}2W_{0,\mu }+\Lambda \right) - \\
-g_1\left( \sum_\nu \partial ^\nu \partial _\nu B_\mu +g_1^2\frac{\upsilon ^2%
}2B_\mu -g_1^2\frac{\upsilon ^2}2B_\mu \right)
\end{array}
\right) \mbox{,}$$
from (\[c1\]):
$$\sum_\nu \partial ^\nu \partial _\nu Z_\mu =-\frac{\upsilon ^2}2\frac
1{g_1^2+g_2^2}\left( g_1^4+g_2^4\right) Z_\mu -g_1g_2\frac{\upsilon ^2}%
2\frac 1{g_1^2+g_2^2}\left( g_2^2-g_1^2\right) A_\mu +$$
$$+\frac 1{\sqrt{g_1^2+g_2^2}}\left( g_2\Lambda -g_1\left( \sum_\nu \partial
^\nu \partial _\nu B_\mu +g_1^2\frac{\upsilon ^2}2B_\mu \right) \right)$$
and from (\[q1\]):
$$\begin{array}{c}
\sum_\nu \partial ^\nu \partial _\nu Z_\mu =-\frac 12\frac{\upsilon ^2}%
2\left( g_1^2+g_2^2\right) Z_\mu + \\
+\frac 12\sqrt{g_1^2+g_2^2}\left( \frac 1{g_2}\Lambda -\frac 1{g_1}\left(
\sum_\nu \partial ^\nu \partial _\nu B_\mu +g_1^2\frac{\upsilon ^2}2B_\mu
\right) \right)
\end{array}$$
That is $Z_\mu $ has got the mass:
$$M_Z=\frac \upsilon 2\sqrt{g_1^2+g_2^2}\mbox{.}$$
#### Fragments
Since
$$\left( \pi _{\circ }+\pi _{*}\right) =1_8$$
then
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu -i\frac
12g_2W_\mu \right) \left( \pi _{\circ }+\pi _{*}\right) \Psi `- \\
-h\left(
\begin{array}{c}
\left( \underline{n}\cos \lambda \left( \pi _{\circ }+\pi _{*}\right) +%
\underline{k}\sin \lambda \left( \pi _{\circ }-\pi _{*}\right) \right)
\gamma + \\
+\left( \underline{k}\cos \lambda \left( \pi _{\circ }+\pi _{*}\right) -%
\underline{n}\sin \lambda \left( \pi _{\circ }-\pi _{*}\right) \right) \beta
\end{array}
\right) \Psi `+ \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\left( \pi _{\circ }+\pi
_{*}\right) \Psi `+ \\
+0.5g_1\underline{Y}\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\left( \pi
_{\circ }+\pi _{*}\right) \Psi `=0\mbox{.}
\end{array}$$
Because
$$\begin{array}{c}
\pi _{\circ }\beta =\beta \pi _{\circ }\mbox{, }\pi _{*}\beta =\beta \pi _{*}%
\mbox{,} \\
\pi _{\circ }\gamma =\gamma \pi _{\circ }\mbox{,}\pi _{*}\gamma =\gamma \pi
_{*}
\end{array}$$
then
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu -i\frac
12g_2W_\mu \right) \pi _{\circ }\Psi `+ \\
+\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu -i\frac
12g_2W_\mu \right) \pi _{*}\Psi `- \\
-h\left(
\begin{array}{c}
\underline{n}\cos \lambda \gamma \pi _{\circ }\Psi `+\underline{n}\cos
\lambda \gamma \pi _{*}\Psi `+ \\
+\underline{k}\sin \lambda \gamma \pi _{\circ }\Psi `-\underline{k}\sin
\lambda \gamma \pi _{*}\Psi `+ \\
+\underline{k}\cos \lambda \beta \pi _{\circ }\Psi `+\underline{k}\cos
\lambda \beta \pi _{*}\Psi `- \\
-\underline{n}\sin \lambda \beta \pi _{\circ }\Psi `+\underline{n}\sin
\lambda \beta \pi _{*}\Psi `
\end{array}
\right) + \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\pi _{\circ }\Psi `+\sum_{\mu
=0}^3F_\mu \underline{\beta _\mu }\pi _{*}\Psi `+ \\
+0.5g_1\underline{Y}\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\pi _{\circ
}\Psi `+ \\
+0.5g_1\underline{Y}\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\pi _{*}\Psi
`=0\mbox{.}
\end{array}$$
Let
$$\Psi _{\circ }`=\pi _{\circ }\Psi `\mbox{ and }\Psi _{*}`=\pi _{*}\Psi `%
\mbox{.}$$
In that case:
$$\begin{array}{c}
\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu -i\frac
12g_2W_\mu \right) \Psi _{\circ }`- \\
-h\left( \left( \underline{n}\cos \lambda +\underline{k}\sin \lambda \right)
\gamma +\left( \underline{k}\cos \lambda -\underline{n}\sin \lambda \right)
\beta \right) \Psi _{\circ }`+ \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\Psi _{\circ }`+0.5g_1%
\underline{Y}\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\Psi _{\circ }`+ \\
\sum_{\mu =0}^3\underline{\beta _\mu }i\left( \partial _\mu -i\frac
12g_2W_\mu \right) \Psi _{*}`- \\
-h\left( \left( \underline{n}\cos \lambda -\underline{k}\sin \lambda \right)
\gamma +\left( \underline{k}\cos \lambda +\underline{n}\sin \lambda \right)
\beta \right) \Psi _{*}`+ \\
+\sum_{\mu =0}^3F_\mu \underline{\beta _\mu }\Psi _{*}`+0.5g_1\underline{Y}%
\sum_{\mu =0}^3B_\mu \underline{\beta _\mu }\Psi _{*}`=0\mbox{.}
\end{array}
\label{um5}$$
Therefore for the (\[tt\]) transformation from (\[fr1\], \[fr2\]):
$$\mathbf{\Psi }\left( x_5,x_4\right) =\mathbf{\Psi }_{\circ }\left(
x_5,x_4\right) +\mathbf{\Psi }_{*}\left( x_5,x_4\right) \rightarrow$$
$$\rightarrow \mathbf{\Psi }`\left( x_5,x_4\right) =\mathbf{\Psi }_{\circ
}`\left( x_5,x_4\right) +\mathbf{\Psi }_{*}`\left( x_5,x_4\right) =$$
$$=\frac 1{2\sqrt{1-a^2}}\cdot$$
$$\cdot \left(
\begin{array}{c}
\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a+i\sqrt{1-a^2}%
\right) \phi _L\left( 0,0\right) +\phi _R\left( 0,0\right) \right) - \\
-\left( ic-g\right) \left( \left( a+i\sqrt{1-a^2}\right) \phi _L\left(
n,k\right) -\phi _R\left( n,k\right) \right) + \\
+\exp \left( -ih\left( \left( n\cos \lambda +k\sin \lambda \right)
x_5+\left( k\cos \lambda -n\sin \lambda \right) x_4\right) \right) \cdot \\
\cdot \left(
\begin{array}{c}
\left( ic+g\right) \left( \left( a+i\sqrt{1-a^2}\right) \phi _L\left(
0,0\right) -\phi _R\left( 0,0\right) \right) + \\
+\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a+i\sqrt{1-a^2}%
\right) \phi _L\left( n,k\right) +\phi _R\left( n,k\right) \right)
\end{array}
\right)
\end{array}
\right) +$$
$$+\frac 1{2\sqrt{1-a^2}}\cdot$$
$$\cdot \left(
\begin{array}{c}
\left( \sqrt{\left( 1-a^2\right) }-b\right) \left( \left( a-i\sqrt{1-a^2}%
\right) \phi _L\left( 0,0\right) +\phi _R\left( 0,0\right) \right) + \\
+\left( ic-g\right) \left( \left( a-i\sqrt{1-a^2}\right) \phi _L\left(
n,k\right) -\phi _R\left( n,k\right) \right) + \\
+\exp \left( -ih\left( \left( n\cos \lambda -k\sin \lambda \right)
x_5+\left( k\cos \lambda +n\sin \lambda \right) x_4\right) \right) \cdot \\
\cdot \left(
\begin{array}{c}
\left( g+ic\right) \left( -\left( a-i\sqrt{1-a^2}\right) \phi _L\left(
0,0\right) +\phi _R\left( 0,0\right) \right) + \\
+\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a-i\sqrt{1-a^2}%
\right) \phi _L\left( n,k\right) +\phi _R\left( n,k\right) \right)
\end{array}
\right)
\end{array}
\right) \mbox{.}$$
That is:
$$\mathbf{\Psi }\left( x_5,x_4\right) =\mathbf{\Psi }_{\circ }\left(
x_5,x_4\right) +\mathbf{\Psi }_{*}\left( x_5,x_4\right) \rightarrow$$
$$\rightarrow \mathbf{\Psi }`\left( x_5,x_4\right) =\mathbf{\Psi }_{\circ
}`\left( x_5,x_4\right) +\mathbf{\Psi }_{*}`\left( x_5,x_4\right) =$$
$$=\frac 1{2\sqrt{1-a^2}}\cdot$$
$$\cdot \left(
\begin{array}{c}
\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a+i\sqrt{1-a^2}%
\right) \phi _L\left( 0,0\right) +\phi _R\left( 0,0\right) \right) - \\
-\left( ic-g\right) \left( \left( a+i\sqrt{1-a^2}\right) \phi _L\left(
n,k\right) -\phi _R\left( n,k\right) \right) + \\
+\exp \left( -ih\left( \left( na+k\sqrt{1-a^2}\right) x_5+\left( ka-n\sqrt{%
1-a^2}\right) x_4\right) \right) \cdot \\
\cdot \left(
\begin{array}{c}
\left( ic+g\right) \left( \left( a+i\sqrt{1-a^2}\right) \phi _L\left(
0,0\right) -\phi _R\left( 0,0\right) \right) + \\
+\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a+i\sqrt{1-a^2}%
\right) \varphi _L\left( n,k\right) +\phi _R\left( n,k\right) \right)
\end{array}
\right)
\end{array}
\right) +$$
$$+\frac 1{2\sqrt{1-a^2}}\cdot$$
$$\cdot \left(
\begin{array}{c}
\left( \sqrt{\left( 1-a^2\right) }-b\right) \left( \left( a-i\sqrt{1-a^2}%
\right) \varphi _L\left( 0,0\right) +\phi _R\left( 0,0\right) \right) + \\
+\left( ic-g\right) \left( \left( a-i\sqrt{1-a^2}\right) \phi _L\left(
n,k\right) -\phi _R\left( n,k\right) \right) + \\
+\exp \left( -ih\left( \left( na-k\sqrt{1-a^2}\right) x_5+\left( ka+n\sqrt{%
1-a^2}\right) x_4\right) \right) \cdot \\
\cdot \left(
\begin{array}{c}
\left( g+ic\right) \left( -\left( a-i\sqrt{1-a^2}\right) \phi _L\left(
0,0\right) +\phi _R\left( 0,0\right) \right) + \\
+\left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a-i\sqrt{1-a^2}%
\right) \phi _L\left( n,k\right) +\phi _R\left( n,k\right) \right)
\end{array}
\right)
\end{array}
\right) \mbox{.}$$
Let in some point $\left\langle t,x,y,z\right\rangle $ $\phi _L\left(
n,k\right) \neq \mathbf{0}$ or/and $\phi _R\left( n,k\right) \neq \mathbf{0}$
In that case (\[dl2\]) in this point: $\phi _L\left( 0,0\right) =\mathbf{0}
$ and $\phi _R\left( 0,0\right) =\mathbf{0}$.
Hence:
$$\begin{array}{c}
\mathbf{\Psi }_{\circ }`\left( x_5,x_4\right) =\frac 1{2\sqrt{1-a^2}}\cdot
\\
\cdot \left(
\begin{array}{c}
-\left( ic-g\right) \left( \left( a+i\sqrt{1-a^2}\right) \phi _L\left(
n,k\right) -\phi _R\left( n,k\right) \right) + \\
+\exp \left( -ih\left( \left( na+k\sqrt{1-a^2}\right) x_5+\left( ka-n\sqrt{%
1-a^2}\right) x_4\right) \right) \cdot \\
\cdot \left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a+i\sqrt{%
1-a^2}\right) \phi _L\left( n,k\right) +\phi _R\left( n,k\right) \right)
\end{array}
\right) \mbox {,}
\end{array}
\label{ff1}$$
$$\begin{array}{c}
\mathbf{\Psi }_{*}`\left( x_5,x_4\right) =\frac 1{2\sqrt{1-a^2}}\cdot \\
\cdot \left(
\begin{array}{c}
\left( ic-g\right) \left( \left( a-i\sqrt{1-a^2}\right) \phi _L\left(
n,k\right) -\phi _R\left( n,k\right) \right) + \\
+\exp \left( -ih\left( \left( na-k\sqrt{1-a^2}\right) x_5+\left( ka+n\sqrt{%
1-a^2}\right) x_4\right) \right) \cdot \\
\cdot \left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a-i\sqrt{%
1-a^2}\right) \phi _L\left( n,k\right) +\phi _R\left( n,k\right) \right)
\end{array}
\right)
\end{array}
\label{ff2}$$
and
$$\begin{array}{c}
\mathbf{\Psi }`\left( x_5,x_4\right) = \\
=-i\left( ic-g\right) \phi _L\left( n,k\right) + \\
+\frac 1{2\sqrt{1-a^2}}\cdot \\
\cdot (\exp \left( -ih\left( \left( na+k\sqrt{1-a^2}\right) x_5+\left( ka-n%
\sqrt{1-a^2}\right) x_4\right) \right) \cdot \\
\cdot \left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a+i\sqrt{%
1-a^2}\right) \phi _L\left( n,k\right) +\phi _R\left( n,k\right) \right) +
\\
+\frac 1{2\sqrt{1-a^2}}\cdot \\
+\exp \left( -ih\left( \left( na-k\sqrt{1-a^2}\right) x_5+\left( ka+n\sqrt{%
1-a^2}\right) x_4\right) \right) \cdot \\
\cdot \left( b+\sqrt{\left( 1-a^2\right) }\right) \left( \left( a-i\sqrt{%
1-a^2}\right) \phi _L\left( n,k\right) +\phi _R\left( n,k\right) \right) %
\mbox {.}
\end{array}
\label{fff}$$
#### Local probabilities
Let:
$$\begin{array}{c}
\Psi _{\circ }^{\dagger }\Psi _{\circ }=\rho _{\circ }\mbox {, }\Psi _{\circ
}`^{\dagger }\Psi _{\circ }`=\rho _{\circ }`\mbox {,} \\
\Psi _{\circ }^{\dagger }\underline{\beta _1}\Psi _{\circ }=j_{\circ x}%
\mbox
{, }\Psi _{\circ }`^{\dagger }\underline{\beta _1}\Psi _{\circ }`=j_{\circ
x}`\mbox {,} \\
\Psi _{\circ }^{\dagger }\underline{\beta _2}\Psi _{\circ }=j_{\circ y}%
\mbox
{, }\Psi _{\circ }`^{\dagger }\underline{\beta _2}\Psi _{\circ }`=j_{\circ
y}`\mbox {,} \\
\Psi _{\circ }^{\dagger }\underline{\beta _3}\Psi _{\circ }=j_{\circ z}%
\mbox
{, }\Psi _{\circ }`^{\dagger }\underline{\beta _3}\Psi _{\circ }`=j_{\circ
z}`\mbox {,}
\end{array}$$
$$\begin{array}{c}
\Psi _{*}^{\dagger }\Psi _{*}=\rho _{*}\mbox {, }\Psi _{*}`^{\dagger }\Psi
_{*}`=\rho _{*}`\mbox {,} \\
\Psi _{*}^{\dagger }\underline{\beta _1}\Psi _{*}=j_{*x}\mbox {, }\Psi
_{*}`^{\dagger }\underline{\beta _1}\Psi _{*}`=j_{*x}`\mbox {,} \\
\Psi _{*}^{\dagger }\underline{\beta _2}\Psi _{*}=j_{*y}\mbox {, }\Psi
_{*}`^{\dagger }\underline{\beta _2}\Psi _{*}`=j_{*y}`\mbox {,} \\
\Psi _{*}^{\dagger }\underline{\beta _3}\Psi _{*}=j_{*z}\mbox {, }\Psi
_{*}`^{\dagger }\underline{\beta _3}\Psi _{*}`=j_{*z}`\mbox {,}
\end{array}$$
Because
$$\begin{array}{c}
\left( \Psi _{\circ }^{\dagger }\Psi _{\circ }\right) ^2-\left( \left( \Psi
_{\circ }^{\dagger }\underline{\beta _1}\Psi _{\circ }\right) ^2+\left( \Psi
_{\circ }^{\dagger }\underline{\beta _2}\Psi _{\circ }\right) ^2+\left( \Psi
_{\circ }^{\dagger }\underline{\beta _3}\Psi _{\circ }\right) ^2\right) = \\
=\left( \Psi _{\circ }^{\dagger }\gamma \Psi _{\circ }\right) ^2+\left( \Psi
_{\circ }^{\dagger }\beta \Psi _{\circ }\right) ^2\mbox {,}
\end{array}$$
$$\begin{array}{c}
\left( \Psi _{*}^{\dagger }\Psi _{*}\right) ^2-\left( \left( \Psi
_{*}^{\dagger }\underline{\beta _1}\Psi _{*}\right) ^2+\left( \Psi
_{*}^{\dagger }\underline{\beta _2}\Psi _{*}\right) ^2+\left( \Psi
_{*}^{\dagger }\underline{\beta _3}\Psi _{*}\right) ^2\right) = \\
=\left( \Psi _{*}^{\dagger }\gamma \Psi _{*}\right) ^2+\left( \Psi
_{*}^{\dagger }\beta \Psi _{*}\right) ^2
\end{array}$$
then the local densities are:
$$\begin{array}{c}
\rho _{\circ o}^2=\rho _{\circ }^2-\left( j_{\circ x}^2+j_{\circ
y}^2+j_{\circ z}^2\right) =\left( \Psi _{\circ }^{\dagger }\gamma \Psi
_{\circ }\right) ^2+\left( \Psi _{\circ }^{\dagger }\beta \Psi _{\circ
}\right) ^2 \\
\rho _{\circ o}`^2=\rho _{\circ }`^2-\left( j_{\circ x}`^2+j_{\circ
y}`^2+j_{\circ z}`^2\right) = \\
=\left( \Psi _{\circ }`^{\dagger }\gamma \Psi _{\circ }`\right) ^2+\left(
\Psi _{\circ }`^{\dagger }\beta \Psi _{\circ }`\right) ^2
\end{array}$$
$$\begin{array}{c}
\rho _{*o}^2=\rho _{*}^2-\left( j_{*x}^2+j_{*y}^2+j_{*z}^2\right) =\left(
\Psi _{*}^{\dagger }\gamma \Psi _{*}\right) ^2+\left( \Psi _{*}^{\dagger
}\beta \Psi _{*}\right) ^2 \\
\rho _{*o}`^2=\rho _{*}`^2-\left( j_{*x}`^2+j_{*y}`^2+j_{*z}`^2\right) = \\
=\left( \Psi _{*}`^{\dagger }\gamma \Psi _{*}`\right) ^2+\left( \Psi
_{*}`^{\dagger }\beta \Psi _{*}`\right) ^2
\end{array}$$
Let us design:
$$\begin{array}{c}
\gamma _{\circ }=\pi _{\circ }^{\dagger }\gamma \pi _{\circ }\mbox{, }\gamma
_{*}=\pi _{*}^{\dagger }\gamma \pi _{*}\mbox{,} \\
\beta _{\circ }=\pi _{\circ }^{\dagger }\beta \pi _{\circ }\mbox{, }\beta
_{*}=\pi _{*}^{\dagger }\beta \pi _{*}\mbox{,}
\end{array}$$
$$\begin{array}{c}
\gamma _{\circ }`=U^{\left( -\right) \dagger }\gamma _{\circ }U^{\left(
-\right) }\mbox{, }\gamma _{*}`=U^{\left( -\right) \dagger }\gamma
_{*}U^{\left( -\right) }\mbox{,} \\
\beta _{\circ }`=U^{\left( -\right) \dagger }\beta _{\circ }U^{\left(
-\right) }\mbox{, }\beta _{*}`=U^{\left( -\right) \dagger }\beta
_{*}U^{\left( -\right) }\mbox{.}
\end{array}$$
Since
$$\begin{array}{c}
\gamma _{\circ }`=a\gamma _{\circ }+\sqrt{1-a^2}\beta _{\circ }\mbox{, }%
\beta _{\circ }`=a\beta _{\circ }-\sqrt{1-a^2}\gamma _{\circ }\mbox{,} \\
\gamma _{*}`=a\gamma _{*}-\sqrt{1-a^2}\beta _{*}\mbox{, }\beta _{*}`=a\beta
_{*}+\sqrt{1-a^2}\gamma _{*}
\end{array}$$
then
$$\rho _{\circ o}^2=\rho _{\circ o}`^2\mbox { and }\rho _{*o}^2=\rho _{*o}`^2%
\mbox {.}$$
From (\[uni\]) since:
$$\rho =\rho `\mbox {, }j_x=j_x`\mbox {, }j_y=j_y`\mbox {, }j_z=j_z`$$
then the local densities:
$$\rho _o^2=\rho ^2-\left( j_x^2+j_y^2+j_z^2\right) =\rho _o`^2=\rho
`^2-\left( j_x`^2+j_y`^2+j_z`^2\right) \mbox {.}$$
Because
$$\begin{array}{c}
\left( \Psi ^{\dagger }\Psi \right) ^2-\left( \left( \Psi ^{\dagger }%
\underline{\beta _1}\Psi \right) ^2+\left( \Psi ^{\dagger }\underline{\beta
_2}\Psi \right) ^2+\left( \Psi ^{\dagger }\underline{\beta _3}\Psi \right)
^2\right) = \\
=\left( \Psi ^{\dagger }\gamma \Psi \right) ^2+\left( \Psi ^{\dagger }\beta
\Psi \right) ^2
\end{array}$$
and
$$\begin{array}{c}
\left( \Psi `^{\dagger }\Psi `\right) ^2-\left( \left( \Psi `^{\dagger }%
\underline{\beta _1}\Psi `\right) ^2+\left( \Psi `^{\dagger }\underline{%
\beta _2}\Psi `\right) ^2+\left( \Psi `^{\dagger }\underline{\beta _3}\Psi
`\right) ^2\right) = \\
=\left(
\begin{array}{c}
\sqrt{\left( \Psi _{\circ }`^{\dagger }\gamma \Psi _{\circ }`\right)
^2+\left( \Psi _{\circ }`^{\dagger }\beta \Psi _{\circ }`\right) ^2}+ \\
+\sqrt{\left( \Psi _{*}`^{\dagger }\gamma \Psi _{*}`\right) ^2+\left( \Psi
_{*}`^{\dagger }\beta \Psi _{*}`\right) ^2}
\end{array}
\right) ^2
\end{array}$$
but
$$\begin{array}{c}
\left( \Psi `^{\dagger }\Psi `\right) ^2-\left(
\begin{array}{c}
\left( \Psi `^{\dagger }\underline{\beta _1}\Psi `\right) ^2+ \\
+\left( \Psi `^{\dagger }\underline{\beta _2}\Psi `\right) ^2+\left( \Psi
`^{\dagger }\underline{\beta _3}\Psi `\right) ^2
\end{array}
\right) \neq \\
\neq \left( \Psi `^{\dagger }\gamma \Psi `\right) ^2+\left( \Psi `^{\dagger
}\beta \Psi `\right) ^2
\end{array}$$
then
$$\begin{array}{c}
\rho _o=\sqrt{\left( \Psi ^{\dagger }\gamma \Psi \right) ^2+\left( \Psi
^{\dagger }\beta \Psi \right) ^2}= \\
=\sqrt{\left( \Psi _{\circ }{}^{\dagger }\gamma \Psi _{\circ }\right)
^2+\left( \Psi _{\circ }{}^{\dagger }\beta \Psi _{\circ }\right) ^2}+ \\
+\sqrt{\left( \Psi _{*}{}^{\dagger }\gamma \Psi _{*}\right) ^2+\left( \Psi
_{*}{}^{\dagger }\beta \Psi _{*}\right) ^2}= \\
=\rho _{\circ o}+\rho _{*o}
\end{array}$$
and
$$\begin{array}{c}
\rho _o`= \\
=\sqrt{\left( \Psi _{\circ }`^{\dagger }\gamma \Psi _{\circ }`\right)
^2+\left( \Psi _{\circ }`^{\dagger }\beta \Psi _{\circ }`\right) ^2}+ \\
+\sqrt{\left( \Psi _{*}`^{\dagger }\gamma \Psi _{*}`\right) ^2+\left( \Psi
_{*}`^{\dagger }\beta \Psi _{*}`\right) ^2}= \\
=\rho _{\circ o}`+\rho _{*o}`
\end{array}$$
but
$$\rho _o`^2\neq \left( \Psi `^{\dagger }\gamma \Psi `\right) ^2+\left( \Psi
`^{\dagger }\beta \Psi `\right) ^2\mbox {.}$$
Therefore $\rho _o$ is a local probability density of a sum of two mutually exclusive events with a local densities $\rho _{\circ o}$ and $\rho _{*o}$.
Because:
$$\begin{array}{c}
\Psi _{\circ }^{\dagger }\gamma \Psi _{\circ }=\frac 12\left( 1-\frac b{%
\sqrt{1-a^2}}\right) \Psi ^{\dagger }\gamma \Psi \mbox {,} \\
\Psi _{\circ }^{\dagger }\beta \Psi _{\circ }=\frac 12\left( 1-\frac b{\sqrt{%
1-a^2}}\right) \Psi ^{\dagger }\beta \Psi \mbox {,} \\
\Psi _{*}^{\dagger }\gamma \Psi _{*}=\frac 12\left( 1+\frac b{\sqrt{1-a^2}%
}\right) \Psi ^{\dagger }\gamma \Psi \mbox {,} \\
\Psi _{*}^{\dagger }\beta \Psi _{*}=\frac 12\left( 1+\frac b{\sqrt{1-a^2}%
}\right) \Psi ^{\dagger }\beta \Psi
\end{array}$$
then $\rho _o$ and $\rho _o`$ do not depend from $U^{\left( -\right) }$.
For $U^{\left( +\right) }$ (\[upls\]):
$$\underline{\pi }_{\circ }=\frac 1{2\sqrt{1-u^2}}\left[
\begin{array}{cc}
\left( v+\sqrt{1-u^2}\right) 1_4 & \left( -s+ik\right) \gamma _5 \\
\left( -ik-s\right) \gamma _5 & \left( \sqrt{1-u^2}-v\right) 1_4
\end{array}
\right] \mbox{,}$$
$$\underline{\pi }_{*}=\frac 1{2\sqrt{1-u^2}}\left[
\begin{array}{cc}
\left( \sqrt{1-u^2}-v\right) 1_4 & \left( s-ik\right) \gamma _5 \\
\left( ik+s\right) \gamma _5 & \left( v+\sqrt{1-u^2}\right) 1_4
\end{array}
\right] \mbox{.}$$
Hence:
$$\begin{array}{c}
U^{\left( +\right) \dagger }\gamma U^{\left( +\right) }=u\gamma -\sqrt{1-u^2}%
\left( \underline{\pi }_{\circ }-\underline{\pi }_{*}\right) \beta \mbox{,}
\\
U^{\left( +\right) \dagger }\beta U^{\left( +\right) }=u\beta +\sqrt{1-u^2}%
\left( \underline{\pi }_{\circ }-\underline{\pi }_{*}\right) \gamma
\end{array}$$
and all rest for $U^{\left( +\right) }$ like to $U^{\left( -\right) }$.
[9]{} G. A. Quznetsov, The Poincare group deduced from the logic properties of the information in Photon : Old Problems in Light of New Ideas, V. Dvoeglazov (ed), Nova Science Publishers, NY, 2001 or http://arXiv.org/abs/physics/9901039
E.Madelung, [*Die Mathematischen Hilfsmittel des Physikers.*]{} (Springer Verlag, 1957)
Gunn A. Quznetsov, The probability in the relativistic m+1 space-time, http://arXiv.org/abs/physics/9803035
G. Quznetsov, The Probability Distribution to Leptons and Quarks, http://arXiv.org/abs/physics/9904024
Gunn Quznetsov, The lepton, quark and hadron\
currents, http://arXiv.org/abs/physics/9806007.
G. Quznetsov, Whence the Gauge Fields arise,\
http://arXiv.org/abs/physics/0006036
for instance, Gordon Kane, [*Modern Elementary Particle Physics.*]{}
(Addison-Wesley Publishing Company, Inc., 1987)
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
title: 'High Density and Non-volatile CRS-based CAM'
---
1.2em
[*Abstract*]{}-We present new computational building blocks based on memristive devices. These blocks, can be used to implement either supervised or unsupervised learning modules. This is achieved using a crosspoint architecture which is an efficient array implementation for nanoscale two-terminal memristive devices. Based on these blocks and an experimentally verified SPICE macromodel for the memristor, we demonstrate that firstly, the Spike-Timing-Dependent Plasticity (STDP) can be implemented by a single memristor device and secondly, a memristor-based competitive Hebbian learning through STDP using a $1\times 1000$ synaptic network. This is achieved by adjusting the memristor’s conductance values (weights) as a function of the timing difference between presynaptic and postsynaptic spikes. These implementations have a number of shortcomings due to the memristor’s characteristics such as memory decay, highly nonlinear switching behaviour as a function of applied voltage/current, and functional uniformity. These shortcomings can be addressed by utilising a mixed gates that can be used in conjunction with the analogue behaviour for biomimetic computation. The digital implementations in this paper use in-situ computational capability of the memristor.
Introduction
============
The classical von Neumann machine suffers from a large sequential (fetch-execute-store cycle) processing overload due to the existence of the data bus between memory and logic. Neuromorphic engineering introduces a more efficient (event driven) implementation but not necessarily low-power. Software techniques are power hungry and there traditionally has been was no low-power hardware device (switch) to provide tighter coupling between memory and logic, as in biological systems. The memristor is an emerging technology that combines (non-volatile) memory and in-situ computational characteristics in one device in the way that promises an entirely new computer architecture.
The mathematical foundation of the memristor, as the fourth fundamental passive element, has been expounded by Leon Chua [@memristor:chua:1971] and later extended to a broader class, known as memristive devices and systems [@memristor:chua:1976]. This broad classification today includes all resistance switching memory devices [@resistance:chua:2011]. Realisation of a solid-state memristor in 2008 [@memristor:strukov:2008] has generated a new wave of research in realization of both large memory arrays as well as new thinking in the neuromorphic engineering domain. Memristors (the term [*memristor*]{} is a portmanteau of [*memory*]{} and [*resistor*]{}) are capable of encoding information in two or more stable levels each with relatively long decay times. The decay can be long in human terms (e.g. days and weeks), which is a practical implementation of a non-volatile memory [@memristor:chua:1976; @memristor:strukov:2008; @fourth:kavehei:2010; @instar:snider:2011; @learning:hasegawa:2010]. It has also been experimentally proven -in small scale- that these two-terminal memristive devices are able to carry out logic operations [@memristive:borghetti:2010]. Therefore, memristor is a possible option for implementing a tighter coupling between memory and logic technologies.
There are many memristor-based applications. The obvious application of such a nanometer scale device is in implementing non-volatile, low-power, and dense memory arrays. Owing to the multi-stable state property and the relatively long term decay, memristors are also able to encode synaptic weights [@nanoscale:jo:2010]. Furthermore, several possibilities for neuromorphic engineering domain and learning have been also studied [@memory:pershin:2011; @instar:snider:2011; @cortical:snider:2008; @spike:zamarreno:2011]. In this paper we demonstrate very basic analogue and digital circuits that are implemented in memristor technology.
Contributions that this paper provides can be categorised as follow:
- Brief characterisation of memristor for neuromorphic purposes.
- Experimental results demonstrating the multi-stable state of a silver/titanium dioxide/indium tin oxide (Ag/TiO$_2$/ITO).
- Demonstration of the use of memristor as a synaptic connection that mimics the Spike-Timing Dependent Plasticity (STDP) rule.
- Show a memristor-based competitive Hebbian learning through STDP.
- Circuit for analogue multiplication and accumulation using fixed weights pattern.
- Experimental show that a sharp switching behaviour in a fabricated Ag/TiO$_2$/ITO and Pt/TiO$_2$/Pt (Pt: Platinum) memristors as well as state decay. Demonstrating a memristive-based analogue computing.
- CRS-based logic gates through material implication and PLA implementations.
Note that memristor and memristive device characteristics, modelling, materials, and underlying physics are not within the scope of this paper. The interested reader can find further details in [@memristor:chua:1971; @fourth:kavehei:2010; @memory:pershin:2011; @analytical:kavehei:2011; @fabrication:kavehei:2011] for further details. The simulations carried out in this work using SPICE macro-model implementation of presented model in [@analytical:kavehei:2011].
Memristor model {#sec:memmodel}
===============
Memristor device characteristics can be defined using a system of two equations, $$\begin{aligned}
\left\{ \begin{array}{ll}
I=g(w,V)\cdot V\\\label{equ:mem}
\frac{{\rm d}w}{{\rm d}t}=f(w,V)~,\end{array}\right.\\ \nonumber \end{aligned}$$ where $w$ is a physical variable indicating the internal memristor state that in theory is such that $0<w<L$, $L$ is the thickness of a thin-film metal-oxide (memristive) material sandwiched between two metallic electrodes, and $I$ and $V$ represent current and voltage, respectively. The $g(\cdot)$ function represents the memristor’s conductance. The state variable can be expressed using a normalised form $x=1-w/L$. In this case, $w\rightarrow 0$ or moving towards higher conductances can be expressed as $x\rightarrow 1$ and $w\rightarrow L$ or moving towards lower conductances can be shown as $x\rightarrow 0$. In this paper, $R_{\rm HRS}$ represents high resistance state and $R_{\rm LRS}$ shows low resistance state. Eq. (\[equ:mem\]) shows that the output of the system (here $I$), at a given time, depends on $w$ and $V$. State transition conditions are also explained by the function $f(\cdot)$. To measure this function several time-domain experiments for $I$ and $V$ are required. According to our measurements, a $\sinh(\cdot)$ like behaviour can explain dynamics of the device while an additional term is needed. The $\sinh(\cdot)$ term defines the dependency of velocity, $dw/dt$, to the effective applied electric field that has been described as an ionic crystal behaviour in an external electric field [@electronic:mott:1964]. The additional term highlights the dependency of conductance, $G_{t}$, to the previous conductance, $G_{t-1}$. Intuitively, we use an exponential form function $h(w)$ to define $dw/dt$ as a function of $w$ based on Fig. 3 in [@switching:pickett:2009]. The $h(w)$ function then should be multiplied by the $\sinh(\cdot)$. The conductance behaviour as a function of $w$ is also shown in Fig 2 of [@fabrication:kavehei:2011]. Due to the asymmetric behaviour of $w\rightarrow 0$ and $w\rightarrow L$ [@switching:pickett:2009], we have used two different $h(w)$ definition to address a more accurate switching properties [@switching:pickett:2009; @fabrication:kavehei:2011].
The state variable equation then can be defined as $$\begin{aligned}
\frac{dw}{dt}=h(w)\sum_{i}\upsilon_iV^i+d(w),\label{equ:newvelocity}\end{aligned}$$ where $\upsilon_i$ are coefficients for low and high electric fields. The index, $i$, is an positive odd integer so it is the expansion of $\sinh(\cdot)$. This demonstration help to easily extract linear approximation of the memristor model in [@memristor:strukov:2008] and also combine effects of Joule heating and $L-w$ (the effective electric field) in the coefficients [@analytical:kavehei:2011]. The function $d(w)$ represents the decay term which can be weeks, months, or more. The decay term appears to be very similar to synaptic weight update (learning) rule [@instar:snider:2011; @cortical:snider:2008]. The first term of Eq. (\[equ:newvelocity\]), represents a voltage dependent, highly nonlinear which makes high-speed digital computing possible. This property originated from the fact that resistance modulation inside the metal-oxide occurs via electron-ion interactions. This term creates a significant problem for learning applications in the current form.
To compensate this problem we have to take advantage of its high nonlinearity. This nonlinear behaviour produces a threshold-like region that voltages below that threshold does not change the conductance. Considering the fact that, memristor’s conductance, $G$, can be tuned by a series of voltage pulses with appropriate pulse widths and a voltage around the threshold, obviously, pulse time is the other parameter involved. Applying a voltage around the threshold slightly changes $x$ (or $w$) if it is maintained for a few $\mu$s. It is observed that such voltage cannot change the state if the duration is around a few ns. However, slightly increase in the applied voltage increases the speed by several orders of magnitude, which makes nanosecond (digital) switching possible. Therefore, a series of few $\mu$s pulses with an appropriate pulse shape can be used to mimic learning rule [@spike:zamarreno:2011].
Analogue Memory and Computing
=============================
Muti-stable state memory
------------------------
Here we demonstrate such behaviour in Ag/TiO$_2$/ITO experiment, which is an identification for existence of an ionic drift. Fig. \[fig:multistable\] illustrates the existence of the multi-stable memory levels. The experiments carried out using a Keithley 4200-SCS. Triangular input voltage was swept from $0~$V to $-0.9~$V and vice versa. Current compliance of $500~\mu$A was applied to avoid any damage to the device. At the end of each cycle device was disconnected from inputs.
The most critical limitation of analogue memristor is its state decay. Although many stable state can be observed, our measurements for five conductance levels showed decay distribution ranging from a few hours up to a few days. More measurements were not possible with our limited time.
![Memristor analogue behaviour. Experimental result from Ag/TiO$_2$/ITO memristor. Current values are normalised to their maximum value ($35~\mu A$). Inset shows a Device Under Test (DUT). The red and the green areas highlight a memristor device.[]{data-label="fig:multistable"}](./meas_analog.pdf){width="60.00000%"}
Memristive, plasticity, and learning
------------------------------------
The connection can be drawn between memristive devices and biological synaptic update rule, known as STDP, that has been observed in the brain [@nanoscale:jo:2010]. This can be achieved by collecting data from a memristive device based on the time difference, $\Delta t$, between two signals, so called pre- and post-synaptic signals. The results are shown in Fig. \[fig:stdp\] (a), which shows how the Device Under Test (DUT) weight (resistance) changes as a function of $\Delta t$. The intermediate states vanish after a certain decay duration whereas a significantly higher potentiation ($x\rightarrow 1$) will be kept as a long term memory. So, the existence of intermediate states decay helps in mimicking the long-term potentiation and short-term plasticity (LTP and LTP) behaviour [@short:ohno:2011].
{width="100.00000%"}
The collected information is then used as stimuli for a network of $1\times 1000$ memristors are connected to one neuron being implemented and pre- and post-synaptic spikes shape is the same as [@spike:zamarreno:2011], then this network implements the competitive Hebbian learning [@competitive:song:2000]. Initial states have been shown in Fig. \[fig:stdp\] (c) in red. Intentionally, a Gaussian distribution has been employed for the memristors’ initial state values. After running the simulation for $35$ minutes, the network results in a population distribution of weights similar to a previously published competitive Hebbian learning rules [@competitive:song:2000]. The additive and multiplicative features of a memristive network strictly depends on the device and its nonlinearity parameters. Fig. \[fig:stdp\] (b) demonstrates a Poissonian ISI distribution.
Programmable analogue circuits
------------------------------
Although plasticity plays an important role for adaptation and development, networks with fixed synaptic weight pattern should be also studied. Therefore, one of the challenges for this emerging technology is to integrate learning and unlearning hardware as part of a neural computational platform. Since memristors possess a threshold-like behaviour, usually low- or very low-voltage operations do not change the memristor’s initial state. This fact helps developing programmable analogue computing circuits [@practical:pershin:2010]. There is also a similar design in [@proposal:mouttet:2009]. The is no simulation or experimental result.
Here, we introduce the use of a memristive array for implementing a multiplication of inputs and the memristor’s internal state, $w$, which represents the memristor’s conductance. Fig. \[fig:proganal\] (a) illustrates a single row of the array and Fig. \[fig:proganal\] (b) shows its simulation results for two elements, M1 and M2, connected to two inputs, In$_{1}$ and In$_{2}$. In this case, we first applied a voltage pulse to M1 to read its conductance, then a pulse to M2 for the same reason. When two voltage pulses are simultaneously applied to M1 and M2, accumulation operation can be clearly observed.
![Multiply-accumulate module. (a) Shows a single row of multiply elements (memristors), In$_{i}\cdot w_{i}$. (b) Demonstrates simulation results for two memristors, M1 and M2. In this simulation, memristor M2 programmed at $x=0.5$, which is equivalent to $(R_{\rm HRS}+R_{\rm LRS})/2$. Then memristor M1 changes its resistance from $R_{\rm LRS}$ to $R_{\rm HRS}$ in three steps. Each step is a simulation that is shown with different colours. Blue for $R_{\rm M1}=R_{\rm LRS}$, green for $R_{\rm M1}$ close to $(R_{\rm HRS}+R_{\rm LRS})/2$, and red for $R_{\rm M1}$ close to $R_{\rm HRS}$. The summing amplifier can be replaced by any thresholding module for different applications.[]{data-label="fig:proganal"}](./promanalog.pdf){width="80.00000%"}
Existence of a threshold-like switching
---------------------------------------
In this part, we show the existence of a switching threshold in a TiO$_2$-based memristor. According to Chua’s definition [@memristor:chua:1971], memristor links electrical charge to flux, $\varphi$, and $\varphi=\int Vdt$. Therefore, the amount of flux passing through the device can be controlled by $V$ and/or time. So, low pulse widths should not change the conductance if the voltage is lower than a certain value and small voltages similarly do not change the conductance if the applied pulse width is not sufficient. The analysis started from the amorphous (RESET) state and a crystallisation window created above $0.8$ V and $100~\mu$s. Fig. \[fig:step3d\] illustrates the results from a Pt/TiO$_2$/Pt memristor. It is observed that the area of crystallisation window decreases as $R_{\rm HRS}$ increases in different devices [@fabrication:kavehei:2011].
![Existance of a switching threshold in the memristor material. The pulse widths are from $10~\mu$s to $1$ s.[]{data-label="fig:step3d"}](./step3d.pdf){width="100.00000%"}
Digital In-Situ Computing
=========================
The existence of the sharp switching threshold, functional uniformity, intermediate state initialisation, and most importantly state decay creates several problems that can be eliminated or compensated for by using the memristor device as a binary switch.
Complementary Resistive Switch (CRS)
------------------------------------
Although the memristor has introduced new possibilities and it is very well adapted in a crossbar architecture, the inherent interfering current paths between neighbouring cells of an addressed cell impose limitations on the array scalability [@analytical:kavehei:2011; @crs:linn:2010]. A possible solution is to build a diode or a transistor in series with a memristor. Using transistors adds other scalability issues due to the fact that transistors are not very well stackable and the application of diodes imposes a high drive current limiting the use of such array in an ultra-low-power applications.
Linn *et al.* [@crs:linn:2010] introduced a new paradigm by exploiting two anti-serially (with opposite polarities) connected memristors. The structure is similar to a [*memistor*]{} (note the missing “r”) [@memistor:widrow:1960; @solid:thakoor:1990; @memistor:xia:2011]. A (digital) CRS uses a combination of a High Resistance State (HRS) and a Low Resistance State (LRS) to encode logic “0" and logic “1". Consequently, the overall resistance of such device is always around HRS, resulting in significant reduction in the parasitic current paths through neighbouring devices. Fig. \[fig:crs\_states\] (a) summaries the CRS states. If $p$ and $q$ indicate resistances of the memristors A and B, respectively, four different states can be observed. For example, $p/q\leftarrow$L/H indicates that LRS is written in $p$ (memristor M1) and HRS in $q$ (memristor M2). Combinations L/H and H/L for $p$ and $q$ represent logic “1” and logic “0”, respectively. Note that the H/H state only occurs once in a “fresh” device. According to Fig. \[fig:crs\_states\] (c) any transition between the states occurs if the applied voltage exceed the SET thresholds, $V_{\rm th,S1}$ or $V_{\rm th,S2}$ and the device’s initial state supports the transition. Possible state transitions are shown in Table \[tab:crs\_states-trans\], where $p'/q'$ shows the next state, $p/q$ illustrates the initial state, and output is a current pulse or spike. These outputs enable us to have two different read-out mechanisms, logic$\rightarrow$ON or logic$\rightarrow$logic.
The transitions in Table \[tab:crs\_states-trans\] can be defined using [*material implication*]{} logic [@memristive:borghetti:2010; @crs:rosezin:2011]. It has been proven that implication and FALSE operation are a complete set for logical operations [@principia:whitehead:1912]. This logical operation results in change of $q$ depending on the state of $p$ (or vise versa), known as $p$ IMP $q$, ‘$p$ implies $q$’ or ‘if $p$ then $q$’. Therefore, if we represent $p$ NIMP $q$ it means ‘$p$ not implies $q$’, Table \[tab:crs\_states-trans\] (i), for example, represents $q\leftarrow$H and we say the conditions (initial $p/q$ and $\Delta V$) not implies $q$.
![CRS device structure and logical definition of each combination. (a) demonstrates all operational states, (b) illustrates the crossbar view, and (c) shows CRS functionality.[]{data-label="fig:crs_states"}](./crs_states.pdf){width="100.00000%"}
The destructive read-out should not be a problem for two reasons: (1) refreshing a digital memory is a normal task depends on the decay term and (2) there are no alternative available to combine the CRS properties and a non-destructive read-out.
[$p/q$]{} [$\Delta V=V_{{\rm I}_{\rm A}}-V_{{\rm I}_{\rm B}}$]{} [$p'/q'$]{} [Output]{}
---------- ----------- -------------------------------------------------------- ------------- ------------
[i)]{} [“1”]{} [$V_{\rm th,S1}<\Delta V<V_{\rm th,R1}$]{} [ON]{} [pulse]{}
[ii)]{} [“1”]{} [$V_{\rm th,R1}<\Delta V$]{} [“0”]{} [spike]{}
[iii)]{} [“0”]{} [$V_{\rm th,R2}<\Delta V<V_{\rm th,S2}$]{} [ON]{} [pulse]{}
[iv)]{} [“0”]{} [$\Delta V<V_{\rm th,R2}$]{} [“1”]{} [spike]{}
[v)]{} [ON]{} [$V_{\rm th,R1}<\Delta V$]{} [“0”]{} [–]{}
[vi)]{} [ON]{} [$\Delta V<V_{\rm th,R2}$]{} [“1”]{} [–]{}
: State transitions in a CRS[]{data-label="tab:crs_states-trans"}
CRS-based logical operations
----------------------------
Here, we introduce CRS-based logical operation and PLA (programmable logic array) that works with the two transitions, logic$\rightarrow$logic and logic$\rightarrow$ON, but we only present it with the later transition. The idea is to charge a bit-line in a crossbar array, and applying inputs to its word lines. The inherent implication property of the device causes a change under certain conditions that we have already discussed. In [@crs:rosezin:2011], AND and NOR operations are proposed using the logic$\rightarrow$logic transition and current spike read-out process. This method is very dependent on the current spike which occurs by a transient ON state between two logic states. In their implementation, two combinations have been evaluated out of two possible combinations for two CRS devices. Assume voltage, $\Delta V$, is applied across a CRS device that is exceeded its RESET threshold, in this situation this device changes its stored logic, $D$, if $D$ is a certain logic depends on the signature of $\Delta V$. Furthermore, if two CRS devices are connected together, that intermediate point can be connected to either ground or power supply to generate NOR/AND gate. That is the reason that no more possible state can be assumed using such approach.
![CRS-based logic gate structures. (a) $D$ represents stored data, $X$ is an input, and $R_{\rm pu}$ is pull-up resistor. The output is initially charged and it is discharged depends on $D$ and $In$. (b) Shows how a not implication, NIMP, can be implemented. Here $q'\leftarrow D$ NIMP $X$. (c) Two inputs NAND gate is implemented by storing one input as device state and another one as an actual input. Here complementary of signal $X$ is applied to the device. (d) Similar to NAND but complementary of $D$ stored in the CRS and $X$ is applied as an input. Therefore, an OR operation implemented, simply by a single CRS device. Obviously, operations are sequential and they requires one (or several) initialisation but this is a drawback for all of the available Boolean logic operations reported in [@crs:rosezin:2011; @memristive:borghetti:2010]. Pull-up (charge) voltage is enough to push a device to its ON state and not writing a logic, $V_{\rm th,S}<V_{\rm pu}<V_{\rm th,R}$. NOT function can be also implemented using a single CRS if $D$ stores (the data) $A$ and $X=0$, $F=\overline{A}$. (e) and (d) are PLA implementation of the two logic gates. Here we remove the outputs’ complementary signals, AND and NOR. The yellow highlights show the OR-plane and the rest are in the AND-plane.[]{data-label="fig:crscomp"}](./crscomp.pdf){width="100.00000%"}
Here two comprehensive forms of building logical gates are introduced. The first form, allows storing one or more inputs as device state and the second method does not. Fig. \[fig:crscomp\] illustrates how CRS works as an implementation of a not implication, NIMP, operation and how NAND and OR operations can be implemented using a single CRS device. Fig. \[fig:crscomp\] (a)-(d) are well explained in the figure’s caption and their operations is also described. Fig. \[fig:crscomp\] (e) and (f) follow similar phenomenon but in a form of a PLA. The idea is to have a logic$\rightarrow$ON transition in the OR-plane whenever an output product term is addressed. From the NIMP operation, we know that if the applied inputs are part of the output product terms, that bit-line does not discharged so there will be enough voltage across the output CRS device with stored logic “1” (greens) to turn to ON and conduct significantly more current to charge the output signal load.
In the case of using differential voltage pairs, $V_{\rm pu}=-V_{\rm pd}=1.4$ V was selected as $2.8$ V is the READ voltage (in Fig. \[fig:crs\_states\]), where $V_{\rm pu}$ and $V_{\rm pd}$ are pull-up and pull-down voltages. Here we applied $V_{\rm pu}=2.8$ V and $V_{\rm pd}=0$ V, so we used $0.25~\mu$m CMOS transistors in our CMOS domain. Therefore, equivalent input voltage for logics “1” is $2.8$ V and for logics “0” is $0$ V. The pull-up and pull-down resistors, $R_{\rm pu}$ and $R_{\rm pd}$, are both equal to $R_{\rm LRS}\sqrt{2(r+1)}$, where $r=R_{\rm HRS}/R_{LRS}$ [@analytical:kavehei:2011]. The used peripheral CMOS circuitries can be found in [@cmos:shakeel:2011]. The sense amplifier was designed for voltage sensitivity more than $100$ mV.
$D_1$ $D_2$ $X_1$ $X_2$ [Function]{}
---------------- ---------------- ---------------- ---------------- ------------------------
$\overline{A}$ $A$ $0$ $B$ $A\cdot B$ (AND)
$A$ $\overline{A}$ $0$ $\overline{B}$ $\overline{A+B}$ (NOR)
$A$ $\overline{A}$ $\overline{B}$ ${B}$ ${A\oplus B}$ (XOR)
$\overline{A}$ $A$ $\overline{B}$ ${B}$ ${A\odot B}$ (XNOR)
: CRS-based logic implementation with two inouts and two CRSs, $F=\overline{D_{1}}\cdot \overline{D_{2}}+\overline{D_{1}}\cdot X_{2}+\overline{D_{2}}\cdot X_{1}+X_{1}\cdot X_{2}$.[]{data-label="tab:crslogic"}
Assuming we have two inputs, $X_{1}$ and $X_{2}$, and two CRS devices, $D_{1}$ and $D_{2}$, connected to these inputs and a charged bit-line. A number of functions can be implemented by writing $\overline{F}=D_{1}\cdot \overline{X_{1}}+D_{2}\cdot \overline{X_{2}}$, hence, $F=\overline{D_{1}}\cdot \overline{D_{2}}+\overline{D_{1}}\cdot X_{2}+\overline{D_{2}}\cdot X_{1}+X_{1}\cdot X_{2}$. The first term, $\overline{D_{1}}\overline{D_{2}}$, indicates that if both CRSs store “0” TRUE ($F=1$) is implemented. Some other function that is implemented using this configuration are shown in Table \[tab:crslogic\]. In [@high:kavehei:2011] we demonstrates a CRS-based content addressable memory based on the XOR/XNOR function. Fig. \[fig:digitalcomp\] (a) illustrates simulation of a two input NAND function. The most significant advantage of this method is that the initialisation step (step 1) which is writing data into CRS arrays and not a simple refreshing cycle. While in a PLA structure, Fig. \[fig:digitalcomp\] (b), the initialisation is a refreshing cycle. Furthermore, in computer arithmetic operations signals arrive with relative delays, like SUM results and CARRY output, that can be used in parallel with the programming of CRS arrays.
![CRS-based logic gate simulations. (a) A 2-input NAND gate (Fig. \[fig:crscomp\] (c)) simulation. In this style, we are allowed to store one input as the CRS state. (b) A 3-input XOR (SUM) function, implemented in a PLA structure. In both cases, (a) and (b), dashed red line show worst-case low and high output voltages that are sent to sense amplifiers. Due to limited space, complementary output, XNOR, is not shown. Initialisation in (b) means, the array should be initialised before the next logical operation and this is the main reason that the first approach (in (a)) is a far more efficient implementation in terms of both hardware and number of steps. No initialisation is required in (a), because ’writing $D$’ effectively means writing one of the input’s logic into the device.[]{data-label="fig:digitalcomp"}](./digitalcomp.pdf){width="100.00000%"}
Conclusion
==========
This paper introduced basic functional blocks for analogue and digital computation based on memristive devices. It is difficult to have a fair comparison between emerging and the conventional devices as the emerging technologies are at their early stages. Moreover, architectural aspects for future computers seems to be dependent to the concept of universal memory and computational capability of one individual device or nano-system that is entirely different with the classical von Neumann computational framework. Therefore, introducing more compatible circuits and algorithms with these futuristic technologies could play an important role. This work presented the existence of ionic drift in the fabricated memristors. We have also illustrated how the memristor can be used to implement competitive Hebbian learning (additive STDP). An analogue multiply-accumulation circuit was introduced that is able to implement a low precision multiplication and addition. This circuit combines inherent non-volatile memory and dynamics of a memristor as a synapse. The problem of state decay then results in developing a digital version of such learning system which is out of the scope for this paper. However, the idea of digital computing using a more robust memristive device, CRS, was explained and two methods for implementing logical blocks were introduced.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work was supported by Australian Research Council (ARC) and grant No.R 33-2008-000-1040-0 from the World Class University project of MEST and KOSEF through Chungbuk National University.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
L. O. Chua, “Memristor – the missing circuit element,” *IEEE Transactions on Circuit Theory*, vol. 18, no. 5, pp. 507–519, 1971.
L. O. Chua and S. M. Kang, “[Memristive devices and systems]{},” *[Proceedings of the IEEE]{}*, vol. 64, no. 2, pp. 209–223, 1976.
L. Chua, “Resistance switching memories are memristors,” *Applied Physics A: Materials Science & Processing*, vol. 102, pp. 765–783, 2011.
D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, “The missing memristor found,” *Nature*, vol. 453, pp. 80–83, 2008.
O. Kavehei, A. Iqbal, Y. Kim, K. Eshraghian, S. Al-Sarawi, and D. Abbott, “The fourth element: characteristics, modelling and electromagnetic theory of the memristor,” *Proceedings of the Royal Society A: Mathematical, Physical and Engineering Science*, vol. 466, no. 2120, p. 2175, 2010.
G. S. Snider, “Instar and outstar learning with memristive nanodevices,” *Nanotechnology*, vol. 22, art no. 015201, 2011.
T. Hasegawa, T. Ohno, K. Terabe, T. Tsuruoka, T. Nakayama, J. Gimzewski, and M. Aono, “Learning abilities achieved by a single solid-state atomic switch,” *Advanced Materials*, vol. 22, no. 16, pp. 1831–1834, 2010.
J. Borghetti, G. Snider, P. Kuekes, J. Yang, D. Stewart, and R. Williams, “[‘Memristive’ switches enable ‘stateful’ logic operations via material implication]{},” *Nature*, vol. 464, no. 7290, pp. 873–876, 2010.
S. Jo, T. Chang, I. Ebong, B. Bhadviya, P. Mazumder, and W. Lu, “Nanoscale memristor device as synapse in neuromorphic systems,” *Nano Letters*, vol. 10, no. 4, pp. 1297–1301, 2010.
Y. Pershin and M. Di Ventra, “Memory effects in complex materials and nanoscale systems,” *Advances in Physics*, vol. 60, no. 2, pp. 145–227, 2011.
G. S. Snider, “Cortical computing with memristive nanodevices,” *SciDAC Review*, vol. 10, pp. 58–65, 2008.
C. Zamarre[ñ]{}o-Ramos, L. Camu[ñ]{}as-Mesa, J. P[é]{}rez-Carrasco, T. Masquelier, T. Serrano-Gotarredona, and B. Linares-Barranco, “On spike-timing-dependent-plasticity, memristive devices, and building a self-learning visual cortex,” *Frontiers in neuroscience*, vol. 5, 2011.
O. Kavehei, S. Al-Sarawi, K. Cho, K. Eshraghian, and D. Abbott, “An analytical approach for memristive nanoarchitectures,” *arXiv preprint:1106.2927*, 2011.
O. Kavehei, K. Cho, S. Lee, S. Kim, S. Al-Sarawi, D. Abbott, and K. Eshraghian, “Fabrication and modeling of [Ag/TiO$_{2}$/ITO]{} memristor,” in *54th IEEE International Midwest Symposium on Circuits and Systems (MWSCAS)*, 2011.
N. Mott and R. Gurney, *Electronic processes in ionic crystals*.1em plus 0.5em minus 0.4emDover, 1964, ch. 2.
M. D. Pickett, D. B. Strukov, J. L. Borghetti, J. J. Yang, G. S. Snider, D. R. Stewart, and R. S. Williams, “[Switching dynamics in titanium dioxide memristive devices]{},” *Journal of Applied Physics*, vol. 106, no. 7, art no. 074508, 2009.
T. Ohno, T. Hasegawa, T. Tsuruoka, K. Terabe, J. K. Gimzewski, and M. Aono, “Short-term plasticity and long-term potentiation mimicked in single inorganic synapses,” *Nature Materials*, vol. 10, no. 8, p. 591�595, 2011.
S. Song, K. Miller, and L. Abbott, “Competitive [Hebbian]{} learning through spike-timing-dependent synaptic plasticity,” *Nature Neuroscience*, vol. 3, pp. 919–926, 2000.
Y. Pershin and M. Di Ventra, “Practical approach to programmable analog circuits with memristors,” *IEEE Transactions on Circuits and Systems I: Regular Papers*, vol. 57, no. 8, pp. 1857–1864, 2010.
B. Mouttet, “Proposal for memristors in signal processing,” *Nano-Net*, pp. 11–13, 2009.
E. Linn, R. Rosezin, C. Kügeler, and R. Waser, “[Complementary resistive switches for passive nanocrossbar memories]{},” *Nature Materials*, vol. 9, no. 5, pp. 403–406, 2010.
B. Widrow, “An adaptive [‘ADALINE’]{} neuron using chemical “memistors”,” Stanford Electronics Laboratories Technical Report, Tech. Rep. TR-1553-2 23, Oct. 1960.
S. Thakoor, A. Moopenn, T. Daud, and A. Thakoor, “Solid-state thin-film memistor for electronic neural networks,” *Journal of Applied Physics*, vol. 67, no. 6, pp. 3132–3135, 1990.
Q. Xia, M. D. Pickett, J. J. Yang, X. Li, W. Wu, G. Medeiros-Ribeiro, and R. S. Williams, “Two-and three-terminal resistive switches: Nanometer-scale memristors and memistors,” *Advanced Functional Materials*, 2011 (in press).
R. Rosezin, E. Linn, C. Kügeler, R. Bruchhaus, and R. Waser, “Crossbar logic using bipolar and complementary resistive switches,” *IEEE Electron Device Letters*, vol. 32, no. 6, pp. 710–712, 2011.
A. Whitehead and B. Russell, *Principia mathematica*, 1912, vol. 2.
M. S. Qureshi, M. Pickett, F. Miao, and J. P. Strachan, “[CMOS]{} interface circuits for reading and writing memristor crossbar array,” in *IEEE International Symposium on Circuits and Systems (ISCAS)*, 2011, pp. 2954–2957.
O. Kavehei, S. Al-Sarawi, S. Sriram, M. Bhaskaran, and D. Abbott, “High density and non-volatile [CRS]{}-based [CAM]{},” *Arxiv preprint arXiv:1108.3716v1*, 2011.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Adaptive Cruise Control (ACC) is an Advanced Driver Assistance System (ADAS) that enables vehicle following with desired inter-vehicular distances. Cooperative Adaptive Cruise Control (CACC) is upgraded ACC that utilizes additional inter-vehicular wireless communication to share vehicle states such as acceleration to enable shorter gap following. Both ACC and CACC rely on range sensors such as radar to obtain the actual inter-vehicular distance for gap-keeping control. The range sensor may lose detection of the target, the preceding vehicle, on curvy roads or steep hills due to limited angle of view. Unfavourable weather conditions, target selection failure, or hardware issue may also result in target detection loss. During target detection loss, the vehicle following system usually falls back to Cruise Control (CC) wherein the follower vehicle maintains a constant speed. In this work, we propose an alternative way to obtain the inter-vehicular distance during target detection loss to continue vehicle following. The proposed algorithm integrates inter-vehicular communication, accurate vehicle localization, and a digital map with lane center information to approximate the inter-vehicular distance. In-lab robot following experiments demonstrated that the proposed algorithm provided desirable inter-vehicular distance approximation. Although the algorithm is intended for vehicle following application, it can also be used for other scenarios that demand vehicles’ relative distance approximation. The work also showcases our in-lab development effort of robotic emulation of traffic for connected and automated vehicles.'
author:
- 'Yuan Lin, and Azim Eskandarian, [^1] [^2]'
bibliography:
- 'references.bib'
title: 'Integrating Inter-vehicular Communication, Vehicle Localization, and a Digital Map for Cooperative Adaptive Cruise Control with Target Detection Loss'
---
[Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{}
Cooperative Adaptive Cruise Control, Inter-vehicular Communication, Vehicle Localization, Digital Map, Connected and Automated Vehicles.
Introduction
============
ACC systems are commercially available vehicle following systems that allow a vehicle to adjust its speed to maintain a desired distance or time gap between the vehicle itself and its preceding vehicle [@eskandarian2012]. The time gap setting is a popular choice as it allows the inter-vehicular distance to increase linearly with the follower vehicle’s speed, which abides with safety concerns. An ACC system usually includes two components: one is obtaining the inter-vehicular distance via sensors and their algorithms, and the other is gap-keeping feedback control [@liang1999optimal; @naranjo2003adaptive; @luo2010model; @ganji2014adaptive]. As inter-vehicular connectivity is introduced [@blum2004], CACC systems are developed such that vehicles follow one anther in a cooperative manner [@dey2016; @zheng2016stability]. An exemplary CACC system builds upon ACC feedback control and adds additional feedforward control which utilizes the preceding vehicle’s acceleration received wirelessly as the feedforward input [@naus2010]. It has been demonstrated that CACC systems maintain shorter gap following compared to ACC [@milanes2014cooperative; @lin2017].
Many automotive ACC systems utilize sensors such as radar, lidar, or stereo cameras to detect the preceding vehicle and obtain the inter-vehicular distance for gap-keeping control [@widmann2000comparison]. These sensors usually have a limited angle of view (except for 360$\degree$ view sensors) and may lose the target detection on curvy roads or steep hills [@ahmed2005object; @sudou2006adaptive; @engelman2001adaptive]. The target detection loss may also occur when the target selection algorithm of the sensor fails to differentiate the preceding vehicle from nearby vehicles in adjacent lanes. With unfavorable weather conditions such as fog, the sensors may also lose the target due to low reflectance [@austin1987relation; @difranco2004radar]. In addition, hardware problems might happen which could lead to sensor failure. During these target detection loss scenarios, ACC usually falls back to CC such that the follower vehicle keeps a constant speed until the sensor detects the preceding vehicle again [@winner2015handbook]. The ACC target detection loss is evident in the system development and experienced by users. Due to such limitations, ACC is an assistance instead of a safety system and requires drivers’ full attention at all time.
Vehicle-to-everything communication (V2X) which enables connectivity among vehicles, infrastructure, and pedestrians is a major trend of transportation revolution that will improve transportation mobility and safety. With inter-vehicular communication, vehicles can share acceleration for the CACC and share positions for blind spot warning. With vehicle-to-infrastructure communication, vehicles can perform communication-based highway merging [@rios2017automated], eco-routing [@elbery2015eco], and cooperative intersection control [@lee2012development]. Different countries may have different communication standards for V2X such as the Dedicated Short Range Communication (DSRC) in the US [@kenney2011dedicated] and Cellular-V2X in China [@chen2017vehicle]. 5G mobile network technology which is currently under development and deployment will provide much faster and more reliable wireless communication for V2X [@andrews2014will].
Automated driving is another trend of transportation revolution that reduces human drivers’ driving tasks. The Society of Automotive Engineers has characterized five levels of driving automation which ranges from “No Automation” to “Full Automation”. ACC systems fall into the second level which is “Driver Assistance”. The higher the automation level, the more sophisticated the technology. One of the technological areas for highly automated driving is vehicle localization [@bresson2017simultaneous]. A lot of automated driving tasks such as path following control and collision avoidance would require vehicle localization. The accuracy requirement for vehicle localization is at the centimeter scale due to performance and collision avoidance concerns [@vivacqua2018self]. Accurate vehicle localization is achieved through information fusion by fusing data from different sources which may include Global Positioning System (GPS), Inertial Measurement Unit (IMU), odometry, camera, Lidar, and a high-definition (HD) digital map [@levinson2007map; @levinson2010robust; @gu2016gnss]. Recent studies also investigate collaborative localization which adds communicated information in the data fusion to provide desirable localization results [@shen2017optimization].
The purpose of this work is to address the problem of vehicle following when the range sensor loses detection of the target preceding vehicle. We propose a method the can approximate the inter-vehicular distance using the essential functions of connected and automated vehicles which include inter-vehicular communication and vehicle localization. The proposed algorithm is an alternative way of obtaining inter-vehicular distances as opposed to directly using range sensors. As a proof of concept, we implement and validate the proposed algorithm on autonomous mobile robots for CACC robot following. In-lab experiments have demonstrated that the proposed algorithm provides desirable inter-robot distance approximation. We conclude that the proposed algorithm is a viable solution for vehicle following during target detection loss in real-world driving.
The rest of this paper is organized as follows: Section II introduces the CACC problem formulation and control system design. Section III details the proposed algorithm for inter-vehicular distance approximation. Section IV documents the setup of our robotic emulation of traffic for CACC robot following experiments. Section V shows the CACC robot following experiment results with the proposed inter-vehicular distance approximation algorithm. Section VI draws the conclusion of this work.
CACC Problem Formulation
========================
The problem that we are addressing is vehicle following with target detection loss. The vehicle following system can be either CACC or ACC. In this work, CACC is used since CACC has been proved in the literature to successfully enable shorter gap following as compared to ACC [@milanes2014cooperative; @lin2017]. In this section, we focus on the mathematical formulation and an exemplary control system design of CACC while also introducing the target detection loss issue.
Fig. \[fig:schematic\] shows a schematic of CACC vehicle following with target detection loss. The longitudinal distances traveled by the preceding $i-1$ and following $i$ vehicles (measured from the vehicle front bumper) are denoted as $l_{i-1}$ and $l_i$, respectively. Mathematically, the actual distance between these two vehicles is $l_{i-1} - l_i - b_{i-1}$ where $b_{i-1}$ is the body length of the preceding vehicle $i-1$. This actual inter-vehicular distance is obtained by range sensors such as radar in actual implementation. The desired distance is dictated by a constant time gap setting and is computed as $h \dot{l}_i + l_0$ where $h$ is the constant time gap, $\dot{l}_i$ is the follower vehicle’s velocity, and $l_0$ is a standstill safety distance. The objective of CACC vehicle following control is to minimize the gap-keeping error between the actual and desired distances, i.e., $e_i = l_{i-1} - l_i - b_{i-1} - (h \dot{l}_i + l_0)$.
In Fig. \[fig:schematic\], the symbols $x_{i-1}$ and $y_{i-1}$ are the horizontal and vertical positions of the preceding vehicle, respectively; the symbols $x_i$ and $y_i$ are the horizontal and vertical positions of the follower vehicle, respectively. These positions can be obtained through vehicle localization. These positions are cartesian coordinates that are used for the inter-vehicular distance approximation which is described in the next section.
![Schematic for CACC vehicle following with target detection loss.[]{data-label="fig:schematic"}](schematic){width="3.4in"}
Fig. \[fig:cacc\] shows the block diagram of an exemplary CACC system [@naus2010]. This CACC system consists of two portions: one is the traditional ACC feedback control that minimizes the error between the actual and desired distances; the other is the feedforward control that utilizes the acceleration of the preceding vehicle received through wireless communication. The feedforward control portion has a feedforward filter $F$ that filters the acceleration of the preceding vehicle. To derive the analytical expression of the feedforward filter $F$, the Laplace Transform of the error is obtained and set to zero $E_i = 0$ to obtain $F = 1/((1+sh)s^2 G)$. For the detailed derivation, readers can refer to [@naus2010] or our previous work [@lin2017experimental]. The feedforward input is supposed to be the instantaneous actual acceleration of the preceding vehicle. However, the actual acceleration that can be obtained through either IMU or wheel encoders can be noisy. Thus, we use the target acceleration of the preceding vehicle $u_{i-1}$ as the feedforward input.
![Block diagram of the CACC system.[]{data-label="fig:cacc"}](cacc){width="3in"}
The feedback control portion of the CACC system has a feedback controller $C$ that minimizes the gap-keeping error. The feedback controller that we use is a Proportional-Derivative (PD) controller [@dey2016]. It has been demonstrated in [@naus2010] that such CACC system design can realize vehicle following with string stability at a small time gap $h$ = 0.6 seconds.
Inter-vehicular Distance Approximation
======================================
This section illustrates the inter-vehicular distance approximation algorithm which is the core of this work. The algorithm requires accurate vehicle localization to obtain vehicle positions, inter-vehicular wireless communication to share those positions, and a digital map with lane center points. Fig. \[fig:distance\] shows the method of the inter-vehicular distance approximation. The main idea of this method is to obtain the inter-vehicular distance based on vehicles’ projected positions on a quadratic curve that fits the lane center points.
![Schematic for the inter-vehicular distance approximation algorithm.[]{data-label="fig:distance"}](distance){width="3.4in"}
Firstly, each vehicle needs accurate localization method to obtain global positions. As stated earlier in the Introduction, current localization methods provide vehicle positioning information with centimeter-level accuracy. The localization results may be GPS-type positions with latitude, longitude and height values based on the earth geodetic coordinate. The GPS-type positions for all vehicles need to be transformed into cartesian coordinates with the same coordinate origin to be used in the inter-vehicular distance approximation algorithm. The approximation algorithm neglects the height information and assumes two-dimensional flat road surface. With actual height changes of the road, the actual inter-vehicular distances are larger than the approximated values. This makes vehicle following safer since the vehicles actually keep larger gaps between them.
In Fig. \[fig:distance\], the triangles represent the localized positions given by the localization results, and the squares represent the vehicles’ true positions which are the positions of the GPS receivers on the vehicles. Via inter-vehicular wireless communication, the follower vehicle $i$ receives the localized position of its preceding vehicle ($x_{i-1}$, $y_{i-1}$). Together with the localized position of itself ($x_i$, $y_i$) and the digital map with lane center points, the follower vehicle obtains the approximated inter-vehicular distance using the following procedure.
Firstly, on the cartesian coordinate system, a bounding box is used to select a rectangular region that covers the localized positions of both vehicles. The bounding box is selected such that the smallest distances from the localized positions to the sides of the bounding box (shown by the dashed lines in Fig. \[fig:distance\]) are all the same. The bounding box covers a segment of the road with a number of lane center points. The lane center points covered by the bounding box are fitted with a quadratic curve function. The localized positions may not be exactly on the fitted quadratic curve of lane center points due to localization errors or that the vehicles may not follow the lane center exactly. Thus, the localized positions are projected on the quadratic curve. A projection point is obtained as the intersection between the quadratic curve and the straight line that passes through the localized position and is perpendicular to the two lane center points closest to the localized position. Two projection points are obtained for the localized positions of the two vehicles.
The approximated inter-vehicular distance is defined as the arc length between the two projection points on the quadratic curve, see the thicker black curve that overlaps some lane center points in Fig. \[fig:distance\]. Assuming the quadratic curve function as
$$\begin{split}
y = ax^2 + bx + c
\end{split}$$
where $a$, $b$, and $c$ are the coefficients that can be obtained through fitting the lane center points in the bounding box. The arc length $L$ between the two projection points, ($x^*_{i-1}$, $y^*_{i-1}$) for the preceding vehicle and ($x^*_i$, $y^*_i$) for the follower vehicle, on the quadratic curve can be computed as
$$\begin{split}
L = \int_{x^*_{i-1}}^{x^*_i} \sqrt{1+ y'^2} dx = \int_{x^*_{i-1}}^{x^*_i} \sqrt{4a^2x^2 + 4abx + b^2 + 1} dx
\end{split}$$
In the event that the lane center points are perfectly on a straight line, the coefficient $a$ of the quadratic function becomes zero. The quadratic function actually turns into a line function. However, the above methodology to obtain the approximated inter-vehicular distance still works by setting $a = 0$.
Robotic Emulation of Traffic
============================
We created the robotic emulation of traffic using mobile robots to evaluate the inter-vehicular approximation algorithm with robot following experiments. The robotic emulation of traffic is both computationally and financially affordable as compared to real-world connected and automated vehicles and allows us to proof-test algorithms in a faster manner. In the following, we introduce the mobile robot testbed preparation which include the robot hardware and software, system identification of robot longitudinal dynamics, robot wireless communication, robot lane keeping, in-lab emulated city and emulated GPS, and robot self-localization.
Robot Hardware and Software
---------------------------
The mobile robots are differential-drive skid-steering robots, Wifibot Lab V4, developed by Nexter Robotics, see the robots in Fig. \[fig:robots\]. Each robot has a mini-computer with Intel Core I5 CPU. The CPU operates a Linux System Ubuntu 14.04. Qt is installed on the Linux System as the C++ IDE. All the robot system capabilities are programmed in C++ for real-time application. The mini-computer connects to a forward-facing camera for lane detection, hosts a wifi card for inter-robot wireless communication, and interfaces with a low-level micro-controller through RS232 serial communication. The mini-computer executes high-level algorithm processing and sends velocity commands to the low-level micro-controller for motor motion control. The low-level micro-controller also obtains wheel encoder reading as robot velocity and Infrared (IR) sensor reading as inter-vehicular distance. Note that we use the IR sensor to emulate radar because the IR sensor is much more affordable and can provide desirable inter-vehicular distance measurements for the robot experiment purpose.
![Two robots used for CACC robot following experiments.[]{data-label="fig:robots"}](robots){width="3.4in"}
Robot Longitudinal Dynamics
---------------------------
The CACC system requires the transfer function of the mobile robot longitudinal dynamics $G$. Therefore, we send step inputs of desired velocity to the robot and obtain the actual velocity output using the wheel encoder, see Fig. \[fig:system\_id\]. Note that the robot takes in only velocity commands for motor control. We consider the transfer function from the desired velocity input to actual velocity ouput as a first-order system $\frac {1} {\tau s + 1} e^{\tau_d s}$ with $s$ as the Laplace Transform variable, $\tau$ as the time constant of the first order system, and $\tau_d$ as the time delay. Using MATLAB System Identification toolbox, we obtained $\tau$ = 0.0661 seconds and $\tau_d$ = 0.04 seconds for the experimental response data in Fig. \[fig:system\_id\]. We then use Simulink to obtain the simulated response of the obtained first-order system to the same desired step input. In Fig. \[fig:system\_id\], the simulated response matches the actual velocity output fairly well.
![System Identification to obtain the transfer function of robot longitudinal dynamics.[]{data-label="fig:system_id"}](system_id){width="3.4in"}
As the robot longitudinal dynamics transfer function $G$ in the CACC system requires the desired input as acceleration and output as traversed distance, we obtain $G$ as $$\begin{split}
G = \frac {1}{s^2(\tau s + 1)} e^{\tau_d s}
\end{split}$$
Wireless Communication
----------------------
The robots utilize wifi for inter-robot wireless communication. The mini-computer on each robot has a 802.11a/b/g Qualcomm Atheros AR93xx WLAN interface card and an antenna, see Fig. \[fig:robots\]. The wireless communication is realized through User Datagram Protocol (UDP) programmed in C++. The communication is decentralized ad hoc network with direct robot-to-robot communication. The communication spectrum is in the 5.9GHz frequency band with sending and receiving rate fixed at 20Hz. In-lab communication tests showed very little (less than 5%) packet loss.
In the CACC robot experiments, each robot has a unique IP address and pre-stores its predecessor and follower’s IP addresses. In this manner, a preceding robot sends its information exclusively to its follower robot using the follower robot’s IP address. We acknowledge that this approach may not be feasible in real-world driving. However, it’s possible to have a relative positioning algorithm to identify surrounding vehicles as long as they are engaged in the current vehicle following activity. This needs to be further investigated and implemented in future work.
Lane Keeping
------------
Lane keeping capability was developed for the mobile robots to run on in-lab artificial lanes autonomously. As the lane keeping method is documented in our previous work in details [@lin2018integrating], we provide a description of the method here. The lane keeping includes lane detection using computer vision and lane following using pure pursuit path following control. For lane detection, a forward-facing camera is used to obtain images of lane markers in front of the robot. The computer vision algorithm includes undistorting each image through camera calibration, extracting lane markers via edge detection, obtaining a top view of the lane markers through inverse perspective mapping, and removing outliers (erroneous lane markers) by checking lane width and lane slope. The middle point of the detected left and right lane marker points for each row within a defined region of interest is obtained to form the center of lane. The computer vision algorithm is programmed in C++ with OpenCV libraries.
Pure pursuit control is used for a mobile robot to follow the obtained center of lane. The pure pursuit control algorithm utilizes a constant look-ahead horizontal line to intersect the center of lane to obtain a look-ahead point. An instantaneous turning circle that connects the look-ahead point and the vehicle centroid, and whose circle center is on the line that is perpendicular to the robot body length direction can then be constructed. The radius of the instantaneous turning circle is the desired instantaneous turning radius for the robot. Given a target longitudinal velocity provided by the CACC control system, the left and right wheel velocities of the robot can be computed using robot kinematics to achieve the desired instantaneous turning radius. For the mathematical details, readers can refer to our previous work [@lin2018integrating]. Note that the pure pursuit control does not guarantee that the robot stays exactly on the center of lane since there is no lane deviation feedback control.
In-lab Track & Emulated GPS
---------------------------
We built an in-lab emulated city with artificial lanes to emulate road infrastructure. Fig. \[fig:localization\] in the CACC Experiments and Results section shows a top view of the in-lab emulated city from an overhead camera. The emulated city has an intersection in the middle and surrounding lane tracks on the outside. The robot following experiments were conducted on the very outside closed-loop track with solid lane markers. On the top-view image, we manually obtained lane center points for the outside track to create the digital map needed for the inter-vehicular distance approximation algorithm. The average distance between two lane center points is around 15 centimeters on the robot track.
A vision-based emulated GPS is developed using the overhead camera to provide robot positions in the emulated city. The main idea of the emulated GPS is to automatically identify and localize a LED light placed on top of each robot, see Fig. \[fig:robots\]. The corresponding computer vision algorithm includes undistorting the raw image obtained from the overhead camera, setting a brightness threshold on the raw image to obtain a black-and-white image whose white pixels represent the bright LED light, and obtaining the robot horizontal and vertical positions as the average of all the white pixels’ horizontal and vertical locations on the image, respectively. The obtained horizontal and vertical positions are cartesian coordinates with the bottom left corner of the undistorted image as the coordinate origin. The positions are originally in the unit of pixels but are converted to values in meters by comparing the same object length in pixels on the image and in meters of ground truth.
The emulated GPS is run on an independent laptop with a Linux system for real-time image processing since the overhead camera image size is large (5Mb). The independent laptop also has wifi which allows the position values to be sent to the corresponding robots instantaneously using the UDP introduced in the Wireless Communication session. The emulated GPS provides positions at 2Hz since the overhead camera frame rate is 2 frames per second. The emulated GPS also outputs heading angle value which is computed as the angle from a horizontal line to the line that connects two neighboring positions of the same robot obtained from two neighboring image frames.
Localization
------------
The localization provides localized positions of the robots to be used in the inter-vehicular distance approximation algorithm. The emulated GPS provides positions only at 2Hz, which is not sufficient for continuous operation of the approximation algorithm. Thus, we fuse the emulated GPS and IMU data using an extended Kalman filter to provide positioning information at the IMU frequency which is $f$ = 100Hz. We achknowledge that the state-of-the-art localization methods for self-driving cars fuse information from more sources such as lidar, camera, odometry, and/or a HD map to provide highly accurate positioning information. However, in our indoor lab setting, fusing just the emulated GPS and IMU data provides desirable accuracy for the approximation algorithm.
In the following, the extended Kalman filter method to fuse the emulated GPS and IMU data is explained. The state equations for the vehicle motion update are $$\begin{split}
x_{i,k+1} = x_{i,k} + \Delta t v_{i,k} \cos(\theta_{i,k} + \Delta t \dot{\theta}_{i,k}) \\
y_{i,k+1} = y_{i,k} + \Delta t v_{i,k} \sin(\theta_{i,k} + \Delta t \dot{\theta}_{i,k}) \\
\theta_{i,k+1} = \theta_{i,k} + \Delta t \dot{\theta}_{i,k}
\end{split}$$ where $x_{i,k+1}$, $y_{i,k+1}$, and $\theta_{i,k+1}$ are the horizontal position, vertical position, and heading angle of the robot, respectively, see Fig. \[fig:motion\]. These three variables are also called state variables. The $v_{i,k}$ is the longitudinal velocity magnitude of the robot obtained through wheel encoders. The $\dot{\theta}_{i,k}$ is the yaw rate (angular velocity) obtained through the IMU. The $\Delta t$ is the updating time step which is determined by the IMU frequency $\Delta t = 1/f$ = 0.01 seconds.
![Robot motion update from time step $k$ to $k+1$.[]{data-label="fig:motion"}](motion){width="3in"}
The measurement variables for the correction step in Kalman filtering are the same as the state variables and are provided by the emulated GPS. Using state space representation, we rewrite the state and measurement update equations in vector form as $$\begin{split}
X_{i,k+1} = f(X_{i,k},U_{i,k}) + W \\
Z_{i,k+1} = H X_{i,k+1} + V \\
\end{split}$$ where $X_{i,k+1} = [x_{i,k+1}, y_{i,k+1}, \theta_{i,k+1}]^\intercal$ is the vector of the state variables; $U_{i,k} = [v_{i,k}, \dot{\theta}_{i,k}]^\intercal$ is the vector of the inputs which include the longitudinal velocity $v_{i,k}$ obtained from wheel encoders and the yaw rate $\dot{\theta}_{i,k}$ obtained from the IMU; $Z_{i,k+1} = [x^{GPS}_{i,k+1}, y^{GPS}_{i,k+1}, \theta^{GPS}_{i,k+1}]^\intercal$ is the vector of the measurement variables. The $H$ matrix is $$H=
\begin{bmatrix}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{bmatrix}$$
The $W$ is the process noise covariance matrix that results from the inaccuracy of the state equations. In the state equations, the vehicle longitudinal velocity instead of the actual velocity is used which results in the inaccuracy. In other words, the actual velocity should have been used in the state equations to describe the correct vehicle motion update. The actual velocity consists of both longitudinal and lateral velocity components where the lateral velocity contributes to the vehicle turning behavior. As wheel encoder sensors measure just the longitudinal velocity, we use the longitudinal velocity in the state equations. Due to the inaccuracy, we define the constant $W$ as $$W=
\begin{bmatrix}
0.01 & 0 & 0 \\
0 & 0.01 & 0 \\
0 & 0 & 0.001
\end{bmatrix}$$
The $V$ is the measurement noise covariance matrix that results from the inaccuracy of the measurements. As our vision-based algorithm delivers highly accurate emulated GPS positions, we consider zero error for the position measurements. Thus, we define the constant $V$ as $$V=
\begin{bmatrix}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & 0.01
\end{bmatrix}$$
To use the extended Kalman filter, the state transition matrix $A$ is obtained as the Jacobian of the non-linear state equations. $$A=
\begin{bmatrix}
1 & 0 & -v_{i,k} \Delta t \sin(\theta_{i,k} + \Delta t \dot{\theta}_{i,k}) \\
0 & 1 & v_{i,k} \Delta t \cos(\theta_{i,k} + \Delta t \dot{\theta}_{i,k}) \\
0 & 0 & 1
\end{bmatrix}$$
The following shows the extended Kalman filter update steps. First, a prior error covariance estimate of the next time step $k+1$ is computed as $$\begin{split}
P^-_{k+1} = A P_k A^\intercal + Q
\end{split}$$ where $P_k$ is the posterior error covariance estimate for the current time step $k$; its initial value is defined as $$P_0=
\begin{bmatrix}
0.01 & 0 & 0 \\
0 & 0.01 & 0 \\
0 & 0 & 0.01
\end{bmatrix}$$
The Kalman filter gain is calculated as $$\begin{split}
K_{k+1} = \frac{P^-_{k+1} H^\intercal}{H P^-_{k+1} H^\intercal + R}
\end{split}$$
The posterior state estimate can then be computed as $$\begin{split}
X_{i,k+1} = f(X_{i,k},U_{i,k}) + K_{k+1} [Z_{i,k} - H f(X_{i,k},U_{i,k})]
\end{split}$$ These are the state variable outputs among which the robot horizontal and vertical positions are used in the inter-vehicular distance approximation algorithm.
Last, the posterior error covariance estimate $P_{k+1}$ is updated as $$\begin{split}
P_{k+1} = (I - K_{k+1} H) P^-_{k+1}
\end{split}$$ where I is a 3 by 3 identity matrix.
Note that the above extended Kalman filter update steps are only used when new emulated GPS measurements are available. When new emulated GPS measurements are not available, the state variable outputs are computed using just the robot motion update state equations $X_{i,k+1} = f(X_{i,k},U_{i,k})$. Thus, the robot positions for the time periods between the arrivals of two emulated GPS positions rely on the IMU data. The matrix calculation for the Kalman filtering was realized with Armadillo: a template-based C++ library for linear algebra [@sanderson2016armadillo; @sanderson2018user].
CACC Experiments and Results
============================
With the mobile robot testbed, we conducted two sets of robot following experiments to evaluate the proposed inter-vehicular distance approximation algorithm. For the first set of experiment, the inter-robot distance was provided solely by the proposed approximation algorithm. In other words, no range sensor (IR sensor) was used in the first set of experiment. For the second set of experiment, the range sensor (IR sensor) was used together with the proposed approximation algorithm to provide the inter-robot distance. In the second set of experiment, the IR sensor provided the inter-robot distance whenever possible and the proposed approximation algorithm was used only when the IR sensor was not able to provide the inter-robot distance around the curve. The second set of experiment is based on a switching mechanism that allows a vehicle to switch from using the range sensor to using the proposed approximation algorithm to obtain the inter-vehicular distance. The switching mechanism design can be potentially applied on the current commercial ACC systems so that it allows a vehicle to continue to follow during target detection loss. In the following, the two sets of robot experiments are explained and the corresponding results are shown.
CACC without range sensor
-------------------------
In the first set of robot following experiment, a follower robot obtained the inter-robot distance using solely the proposed inter-vehicular distance approximation algorithm during the entire experiment and the IR sensor was not used at all. Two robots were used with one being the leader robot and the other being the follower robot. The leader robot’s velocity profile was pre-defined and pre-programmed such that it traversed a complete circle of the outside track of the emulated city. We considered various speeds in one run to investigate the impact of speeds on the accuracy of the inter-vehicular distance approximation, see Fig. \[fig:v\]. A video of the robot following experiment can be seen on the website of the Autonomous Systems and Intelligent Machines Lab in the Mechanical Engineering Department at Virginia Tech.
Fig. \[fig:localization\] shows the localization results for both robots using the robot localization method introduced previously. Comparing the emulated GPS positions and the lane center points of the digital map, we conclude that the lane keeping method desirably restricted the robot to follow the center of lane. However, The robots were not exactly on the center of lane, especially around the curve where the robots seemed to slightly “over-steer” and stayed inside the circle of the center of lane. This could lead to a slight increase of the follower robot’s velocity when the leader robot was at the curve and a slight decrease of the follower robot’s velocity when the follower robot was at the curve, since the “over-steering” created shorter driving path compared to the lane center path. These can be validated in the velocity plots of Fig. \[fig:v\] and Fig. \[fig:v2\].
In Fig. \[fig:localization\], the distance between two emulated GPS positions increases as the robot speed increases. This should also be true to the localized positions obtained from localization. The localization method provides robot positions at a much higher frequency 100Hz compared to the emulated GPS frequency 2Hz. In the plots, the localization results are plotted using curves instead of discreet points due to visualization concerns. Since the latest localized positions are used to estimate the distance in the algorithm, a higher frequency for localization update is essential for smooth distance approximation. For the robot speeds considered, the present localization update frequency 100Hz is sufficient to obtain smooth inter-vehicular distance, see Fig. \[fig:d\].
In Fig. \[fig:tg\], the actual time gap values are very close to the desired time gap $h$ = 0.8 seconds despite the fluctuations mainly caused by the data measurement errors. The actual time gap values have an average of 0.8000 seconds and a standard deviation of 0.0264 seconds. This indicates that the CACC system leads to satisfying robot gap-keeping control.
CACC with switching mechanism
-----------------------------
In the second set of robot following experiment, the follower robot used the range sensor (IR sensor) as the priority option and switched to using the proposed approximation algorithm to obtain the inter-robot distance only when the IR sensor was not able to provide the distance values. Fig. \[fig:flow\_chart\] shows the flow chart for the switching mechanism. Note that the proposed approximation algorithm computes the inter-vehicular distance all the time in the background, while the computed distances are only used then range sensor is not able to provide the inter-vehicular distance.
![Flow chart for the switching mechanism to obtain inter-vehicular distance for CACC.[]{data-label="fig:flow_chart"}](flow_chart){width="3.4in"}
For the current automobile ACC systems, the sensors may lose detection of the preceding vehicle due to various reasons at various occasions. For our robot experiment, the occasion of driving around the curve is considered. Since the IR sensor on the robot provides just a scalar representing the distance, we define the switching condition as that the IR sensor returns erroneous distance values. Since the leader robot maintained a constant speed $v$ = 0.5m/s in this experiment (Fig. \[fig:v2\]), the IR sensor distance became erroneous when it was significantly larger or smaller than the constant desired distance $h \dot{l}_i + l_0$ = 0.8\*0.5 + 0.2 = 0.6m. In fact, we defined the erroneous values as $d < 0.55$ and $d > 0.65$ meters. In other words, the follower robot used the IR sensor to obtain the inter-robot distance when $0.55 \geq d \leq 0.65$ and used the proposed approximation algorithm otherwise. Note that the switching conditions for real cars are different since the existing automobile ACC systems self-determine target detection loss at various occasions. The goal of our robot experiment is to just introduce the switching method.
Fig. \[fig:v2\] shows the velocity results of both robots during the experiment. The follower robot experienced velocity fluctuations due to the “over-steering” around the curve explained previously. Fig. \[fig:d2\] shows the inter-robot distance. By observing the robot horizontal positions, we see that the follower robot utilized the IR sensor to obtain the inter-robot distance when both robots were on straight paths and utilized the proposed approximation algorithm otherwise (around the curve). When the IR sensor distance was used, the proposed approximation algorithm also computed the approximated distance in the background although it was not used. By comparing the IR sensor distance and the unused approximated distance, we conclude that the proposed approximation algorithm provided close approximation to the IR sensor distance. Fig. \[fig:tg2\] shows that the actual time gap values of the robot following. The actual time gap values are also close to the desired time gap $h$ = 0.8 seconds with an average of 0.8005 seconds and a standard deviation of 0.0359 seconds.
Conclusion
==========
In conclusion, we proposed an inter-vehicular distance approximation algorithm for vehicle following with target detection loss. The algorithm integrates inter-vehicular wireless communication, accurate vehicle localization, and a digital map with lane center information. More specifically, a follower vehicle receives the position of its preceding vehicle via wireless communication and utilizes its own position and road geometry from the map to mathematically compute the inter-vehicular distance. We have demonstrated through robot experiments that the proposed algorithm could be a complete replacement of range sensors for long-period vehicle following. We also designed the switching mechanism such that an ACC or CACC system can switch from using a range sensor to using the proposed algorithm to obtain the inter-vehicular distance for short-term target detection loss.
Through the robot following experiments, we discovered some factors that contribute to accurate and smooth inter-vehicular distance approximation using the proposed algorithm. A high vehicle localization update rate and a high wireless communication frequency are essential in the approximation since a follower vehicle requires instantaneous positions of both itself and its preceding vehicle to obtain the instantaneous inter-vehicular distance. The accuracy of the vehicle positions produced by the vehicle localization methods and the correctness of lane center locations of the digital map also directly impacts the approximation accuracy.
In the event of low localization update rates and communication frequencies or communication packet loss, an interpolation or prediction method may be needed to generate vehicle positions with high update rates. Future works also include implementing the proposed inter-vehicular approximation algorithm on real cars and conduct high-speed testing.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank Chaoxian Wu for providing the C++ code for the UDP wireless communication and Christopher Kappes for helping with the system identification to obtain the robot longitudinal dynamics.
[Yuan Lin]{} received the B.E. degree in Civil Engineering from Nanchang University, China, in 2011 and the Ph.D. degree in Engineering Mechanics from Virginia Tech, Blacksburg, VA, USA, in 2016. After obtaining his PhD, Yuan started to work as a Postdoctoral Research Associate of the Autonomous Systems and Intelligent Machines Lab in the Mechanical Engineering Department at Virginia Tech. His research interests include robotics, control, sensor fusion, and V2X.
[Azim Eskandarian]{} has been a Professor and Head of the Mechanical Engineering Department at Virginia Tech (VT) since August 2015. He became the Nicholas and Rebecca Des Champs chaired Professor in April 2018. He established the Autonomous Systems and Intelligent Machines laboratory at VT to conduct research in intelligent and autonomous vehicles, and mobile robotics. Prior to that, he was a Professor of Engineering and Applied Science at The George Washington University (GWU) and the founding Director of the Center for Intelligent Systems Research (1996-2015), the director of the “Transportation Safety and Security” University Area of Excellence (2002-2015), and the co-founder of the National Crash Analysis Center (1992) and its Director (1998-2002 & 5/2013-7/2015). Earlier, he was an Assistant Professor at Pennsylvania State University, York, PA (1989-92) and worked as an engineer/project manager in industry (1983-89). He was awarded the IEEE ITS Society’s Outstanding Researcher Award in 2017 and the GWU’s School of Engineering Outstanding Researcher Award in 2013. Dr. Eskandarian is a fellow of ASME, senior member of IEEE, and member of SAE professional societies. He received his BS, MS, and DSC degrees in Mechanical engineering from GWU, Virginia Tech, and GWU, respectively.
[^1]: Dr. Azim Eskandarian is a full professor, the Director of the Autonomous Systems and Intelligent Machines (ASIM) Lab, and the Department Head of Mechanical Engineering at Virginia Tech, Blacksburg, VA 24061, USA (e-mail: [email protected]).
[^2]: Dr. Yuan Lin was a postdoctoral research associate of the ASIM Lab in Mechanical Engineering Department at Virginia Tech when the research work was conducted (e-mail: [email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A powerful coronal mass ejection (CME) occurred on 2017 September 10 near the end of the declining phase of the historically weak solar cycle 24. We obtain new insights concerning the geometry and kinematics of CME-driven shocks in relation to their heliospheric impacts from the optimal, multi-spacecraft observations of the eruption. The shock, which together with the CME driver can be tracked from the early stage to the outer corona, shows a large oblate structure produced by the vast expansion of the ejecta. The expansion speeds of the shock along the radial and lateral directions are much larger than the translational speed of the shock center, all of which increase during the flare rise phase, peak slightly after the flare maximum and then decrease. The near simultaneous arrival of the CME-driven shock at the Earth and Mars, which are separated by 156.6$^{\circ}$ in longitude, is consistent with the dominance of expansion over translation observed near the Sun. The shock decayed and failed to reach STEREO A around the backward direction. Comparison between ENLIL MHD simulations and the multi-point in situ measurements indicates that the shock expansion near the Sun is crucial for determining the arrival or non-arrival and space weather impact at certain heliospheric locations. The large shock geometry and kinematics have to be taken into account and properly treated for accurate predictions of the arrival time and space weather impact of CMEs.'
author:
- 'Ying D. Liu, Bei Zhu, and Xiaowei Zhao'
title: 'Geometry, Kinematics and Heliospheric Impact of a Large CME-driven Shock in 2017 September'
---
Introduction
============
Coronal mass ejections (CMEs) are massive expulsions of plasma and magnetic flux from the solar corona [@webb12 and references therein]. Fast CMEs can drive shocks in the corona and interplanetary space, which are key accelerators of solar energetic particles [SEPs; @reames99]. Characterizing the shock three-dimensional (3D) geometry and kinematics and their connection with heliospheric impacts is of crucial importance for space weather research and forecasting.
CME-driven shocks have been tracked using type II radio bursts, with an effort to determine the shock distance and speed from the frequency drift of the associated type II burst [e.g., @reiner07; @liu08; @liu17b; @hu16; @zhao17; @zucca18]. Type II emissions, in general, cannot give details of shock geometry, although radio triangulation can provide some information of where in the shock the emission originates [e.g., @juan12; @magdalenic14]. In situ solar wind measurements have also been extensively used to study CME-driven shocks in interplanetary space [e.g., @burton92; @mostl12; @riley16; @lium18]. In particular, multiple spacecraft measurements at different locations in the heliosphere have indicated a large angular extent of CME-driven shocks [e.g., @reisenfeld03; @liu08; @delucas11]. For instance, a CME-driven shock from 2001 November was observed at both the Earth and Ulysses with a latitudinal separation of 73$^{\circ}$ and longitudinal separation of 64$^{\circ}$ [@reisenfeld03]. Another example in 2006 December was associated with an even larger angular separation between the Earth and Ulysses, i.e., 74$^{\circ}$ in latitude and 117$^{\circ}$ in longitude [@liu08]. In situ measurements at different spacecraft, however, are usually sporadically distributed in the heliosphere, and opportunities with well-aligned multiple spacecraft are rare for the investigation of CME-driven shocks.
Direct imaging of CME-driven shocks are now feasible with today’s coronagraphs and heliospheric imagers. They are usually observed as a faint edge ahead of the CME front in coronagraph images near the Sun [e.g., @vourlidas03; @vourlidas09; @liu08; @ontiveros09; @hess14]. When the shocks are far away from the Sun, they appear as a broad front in heliospheric imaging observations [@liu11; @liu12; @liu13; @maloney11; @volpes15]; the density within the CME driver decreases due to expansion in interplanetary space, so the shock and sheath become dominant in heliospheric imagers. These white-light imaging observations over a large distance range allow a possible prediction of the shock parameters at 1 AU [e.g., @liu11; @liu13; @volpes15]. Simultaneous observations from different vantage points enable the determination of the shock 3D geometry and kinematics with respect to the CME driver [@kwon14; @liu17a]. However, the 3D nature and kinematics of CME-driven shocks in relation to their heliospheric impacts, which requires coordinated remote-sensing and in situ observations from multiple spacecraft, are still not well understood.
The 2017 September 10 eruption, which was observed by a fleet of spacecraft at different vantage points, provides a great opportunity to study the 3D geometry, kinematics and heliospheric impacts of CME-driven shocks. The event has attracted significant attention owing to its unusual energetics and occurrence near the end of a weak solar activity cycle [e.g., @lee18; @luhmann18; @seaton18; @li18; @hu18; @yan18; @gopalswamy18 and references therein]. In this work we take advantage of the optimal, multi-spacecraft observations of the 2017 September 10 eruption to gain insights into CME-driven shocks. The view from the Earth as a limb event allows to track the flux rope and the CME-driven shock from their nascent stage continuously to the outer corona with projection effects minimized. In situ measurements from two well-separated locations also enable assessment of the heliospheric impacts in connection with remote-sensing observations.
Observations and Analysis
=========================
The eruption was associated with a long-duration X8.3 flare from NOAA AR 12673 (S09$^{\circ}$W91$^{\circ}$), which peaked at 16:06 UT on September 10. Figure 1 shows the eruption at its nascent stage viewed from GOES 16 and the well-developed CME from the Solar and Heliospheric Observatory [SOHO; @domingo95] and the Solar Terrestrial Relations Observatory A spacecraft [STEREO A; @kaiser08]. The combined EUV and coronagraph observations allow us to track the flux rope and shock separately from the early stage to the outer corona. The axis of the flux rope must be nearly parallel to the ecliptic plane, because otherwise the current sheet behind the flux rope would not be seen in the EUV observations. This interpretation is supported by the orientation of the post-eruption arcades (not shown here) and white-light CME reconstruction (see below). Note the wave and magnetic loop produced by the expansion of the flux rope in the EUV running-difference images. The EUV wave at the early time, which is the footprint of the CME-driven shock in the lower atmosphere [e.g., @patsourakos12; @kwon14 and references therein], and the top of the magnetic loop, which should be very close to the shock front in the upper atmosphere, can be used together to constrain the shock structure. The shock is also seen in the coronagraph images as a faint edge around the CME. It was expanding in all directions and appeared to enclose the whole Sun in both SOHO and STEREO A images at the times given in Figure 1. Similar shock geometries with 360$^{\circ}$ envelope around the Sun have also been found in other cases [e.g., @kwon15; @liu17a].
Based on the observations of the 2012 July 23 complex CME, @liu17a [hereinafter referred to as Paper 1] investigate the structure, propagation and expansion of the associated shock. They find that the shock can be modeled well by a simple spheroidal structure as in other studies [e.g., @hess14; @kwon14], which enables a separation between translation and expansion of the shock. Here we use an ellipsoidal model to simulate the shock [@kwon14; @kwon15], which can be applied to both the EUV and white-light observations. The cross section of the shock ellipsoid perpendicular to the propagation direction is assumed to be circular, which reduces a free parameter in the model. As for the CME, we employ a graduated cylindrical shell (GCS) method, which assumes a rope-like morphology with two ends anchored at the Sun, to estimate its propagation direction and distance [@thernisien06]. The GCS model is applied to the bright rim of the CME in coronagraph observations, for which the method is designed. Reasonable fits are generally obtained for both the CME and shock by adjusting the model parameters to visually match the two views from the Earth and STEREO A simultaneously (see Figure 1).
Figure 2 displays the modeled CME and shock projected onto the ecliptic. The average propagation longitude of the CME and shock is about 95$^{\circ}$ and 97$^{\circ}$ west of the Earth, respectively. The CME flux-rope has a small tilt angle (about 20$^{\circ}$), consistent with what the EUV images imply. The ellipsoidal structure of the shock is produced by the vast expansion of the ejecta in combination with its outward motion, as suggested in Paper 1. In agreement with this idea, the standoff distance between the shock and driver increases from the nose towards the flank. The large extent of the shock, which encircles the whole Sun, helps explain the early detection of energetic particles at Mars and STEREO A [@lee18; @luhmann18]. However, it is worth noting that at some point, especially near the wake, the structure in question could be just a wave without a non-linear steepening character (see Paper 1). Again, the simple structure of the shock enables an easy separation between the translational distance of the shock center and the expansion distances along and perpendicular to the propagation direction.
The kinematics of the CME and shock are shown in Figure 3. All the speeds increase during the flare rise phase, peak slightly after the flare maximum, and then decrease. (There is a bump in the X-ray flux before the maximum, which may indicate a small flare ahead of the major one.) The maximum radial nose velocity is about 3000 km s$^{-1}$ for the CME and about 3300 km s$^{-1}$ for the shock[^1]. The majority of the shock nose speed comes from the radial expansion speed with a peak value of about 2400 km s$^{-1}$. The radial expansion speed is always larger than the translational speed of the shock center, whose peak value is only of the order of 1000 km s$^{-1}$. The lateral expansion speed of the shock closely follows the radial expansion speed and is a little larger.
Figure 4 shows the in situ measurements at the Earth (1.01 AU) and Mars (1.66 AU). No ICME signatures were observed at the Earth, and only a forward shock arrived around 19:21 UT on September 12. A shock was also observed at Mars around 02:52 UT on September 13, possibly followed by a brief ICME interval [@lee18]. The shock arrival time at Mars is determined from the peak in the high-energy particles (see the bottom panel of Figure 4), known as energetic storm particles trapped around the shock. This arrival time agrees with the solar wind plasma measurements at Mars although the measurements are sparse [@halekas17]. The ICME and shock failed to arrive at STEREO A (not shown here). The arrival or non-arrival situation of the ICME/shock at the three locations is consistent with what the modeled CME and shock geometries suggest in Figure 2. The shock around the backward direction decayed before reaching the distance of STEREO A, like the case in Paper 1.
Of particular interest is the near simultaneous arrival of the shock at the Earth and Mars, despite a radial separation of about 0.65 AU between them. The active region produced several eruptions, but our analysis shows that the same CME on September 10 reaching both locations is the most likely scenario. Note that the shock speed at the Earth is about 570 km s$^{-1}$, with which it would take about 2 more days to reach the distance of Mars. This near simultaneous arrival tells the importance of the shock geometry and kinematics on its impact in the heliosphere. If the shock is assumed to be spherical, the locations of the Earth and Mars relative to the shock propagation direction (see Figure 2), together with a simultaneous arrival, would give a distance of about 0.9 AU from the Sun for the shock center and a radius of about 1.43 AU for the shock sphere. This is consistent with the dominance of expansion over translation as inferred from Figure 3.
To connect the imaging observations with the in situ measurements, we experiment with WSA-ENLIL MHD simulations [@arge04; @odstrcil04 hereinafter referred to as ENLIL]. The CME is inserted in the simulation as a plasma cloud through the inner boundary at 21.5 solar radii from the Sun with CME parameters determined from a cone model [@zhao02]. It has no magnetic field of its own and merely carries whatever field is present from the ambient solar wind model. This limitation will to some extent determine how the “CME" and shock evolve in the solar wind. Nevertheless it can do a reasonable job in some cases of predicting shock arrivals and jumps for example. The first run uses the measured CME half width from the cone model ($58^{\circ}$), and for the second run we increase the half width to $90^{\circ}$ to (partly) account for the large angular extent of the shock. The CME density is decreased by half for the second case, so the CME mass is roughly the same.
The simulation results are displayed in Figure 5, and the extracted time series of plasma and magnetic field parameters at the locations of the Earth and Mars are plotted in Figure 4 to compare with the actual measurements. For the first case using the measured CME width, the Earth is completely missed, and the shock arrival time at Mars is around 21:00 UT on September 13, which is about 18 hr later than observed (also see Figure 4). The simulation indicates that the Earth would only see a weak co-rotating interaction region (CIR). For the second case with the increased CME width, the shock reaches both the Earth and Mars with much better improved arrival times. The arrival time is around 16:00 UT on September 12 at the Earth, which is only about 3 hr earlier than observed, and around 10:00 UT on September 13 at Mars, which is only about 7 hr later than observed. In this case, the impact at Mars is much stronger, and the Earth sees the shock on top of the CIR. These results indicate that the shock geometry and kinematics must be properly simulated to have an accurate predication of the arrival and impact.
Although Mars was about $60^{\circ}$ away from the CME/shock nose, the event caused significant space weather effects at Mars [@lee18 and references therein]. The large expansion of the shock observed near the Sun must have enhanced the impact at Mars, as demonstrated above. The two preceding CMEs from September 9, which are included in the simulations, might also have helped by changing the plasma and magnetic field distributions in the ejecta. The simulations indeed show asymmetric distributions in the ejecta mass and speed with respect to the CME propagation direction, part of which can be attributed to the merging of the three CMEs.
Conclusions and Discussion
==========================
We have investigated the large CME-driven shock associated with the 2017 September 10 eruption, taking advantage of the coordinated remote-sensing and in situ observations from a fleet of well-aligned spacecraft. The results, which complement the findings of Paper 1 on the structure and evolution of CME-driven shocks, shed light on the 3D geometry and kinematics of CME-driven shocks in relation to their heliospheric impacts.
The shock and the CME driver can be tracked separately from the early stage to the outer corona using the multi-point EUV and white-light observations. The shock showed a large oblate structure and eventually enclosed the whole Sun, which can be attributed to the vast expansion of the ejecta in combination with its outward motion. The standoff distance between the shock and driver increases from the nose towards the flank and wake. The expansion speeds of the shock along the radial and lateral directions, which closely follow each other, are much larger than the translational speed of the shock center. The expansion and translational speeds increase during the flare rise phase, peak slightly after the flare maximum, and then decrease. Previous studies suggest a temporal relationship of CME acceleration with the associated flare [e.g., @zhang01]. Here our work indicates that a similar relationship also exists between shock kinematics and the flare. Our results also imply that the apparent acceleration of the CME leading edge, which is often used to investigate the acceleration mechanism of CMEs [e.g., @zhang01; @bein11], is largely dominated by the expansion due to the ejecta internal overpressure. The enormous lateral expansion of the shock has important implications for SEP production, because the shock lateral part can have a favorable geometry with respect to the ambient magnetic field for particle acceleration [e.g., @kozarev15; @zhu18]. This geometry, together with the large lateral expansion speed, may produce enhanced SEPs even for observers connected to the flank of the shock (i.e., not the nose). This helps explain the ground level enhancement observed at both the Earth and Mars in the current case [@lee18; @luhmann18; @gopalswamy18].
The CME-driven shock arrived at both the Earth and Mars near simultaneously, which agrees with the dominance of expansion over translation observed near the Sun. These two well-separated locations suggest a large longitudinal extent of the shock in interplanetary space (at least 156.6$^{\circ}$). This corresponds with previous multi-point in situ measurements [e.g., @reisenfeld03; @liu08; @delucas11] and here multi-spacecraft imaging observations of exceptionally wide shock structure. Note that, however, Äthe shock around the backward direction may have decayed before reaching the distance of STEREO A, since no shock signatures were observed at STEREO A. As we have found in Paper 1, at some point, especially near the wake, the structure in question could be just a wave without a non-linear steepening character; the shock would decay and disappear when the driver’s influence becomes weak. An important lesson we have learned by comparing ENLIL MHD simulations with the multi-point in situ measurements is that the shock expansion near the Sun is crucial for determining the arrival or non-arrival and space weather impact at certain heliospheric locations. To improve the predictions of the arrival time and space weather impact, the large shock geometry and kinematics have to be properly treated. Purely using the measured CME width in a simulation like ENLIL may not be sufficient for space weather prediction. Note that here we have compared two cases of simulations, which have the same solar wind background; the effects of the ambient solar wind on the shock propagation and impact, if any, are already taken into account in the simulations.
The research was supported by NSFC under grant 41774179 and the Specialized Research Fund for State Key Laboratories of China. We acknowledge the use of data from SOHO, STEREO, SDO, GOES, Wind and MAVEN, thank M. Leila Mays, Christina O. Lee and Janet G. Luhmann for their help and discussions on the work, and are grateful to R.-Y. Kwon for providing his ellipsoidal shock model. The ENLIL simulations were provided by CCMC through their public Runs on Request system (<http://ccmc.gsfc.nasa.gov>).
Arge, C. N., Luhmann, J. G., Odstrcil, D., Schrijver, C. J., & Li, Y. 2004, JASTP, 66, 1295
Bein, B. M., Berkebile-Stoiser, S., Veronig, A. M., et al. 2011, ApJ, 738, 191
Burton, M. E., Siscoe, G. L., & Smith, E. J. 1992, JGR, 97, 12283
de Lucas, A., Schwenn, R., dal Lago, A., Marsch, E., & Clúa de Gonzalez, A. L. 2011, JASTP, 73, 1281
Domingo, V., Fleck, B., & Poland, A. I. 1995, SoPh, 162, 1
Gopalswamy, N., Yashiro, S., Makela, P., et al. 2018, ApJL, 863, L39
Halekas, J. S., Ruhunusiri, S., Harada, Y., et al. 2017, JGR, 122, 547
Hess, P., & Zhang, J. 2014, ApJ, 792, 49
Hu, H., Liu, Y. D., Wang, R., Möstl, C., & Yang, Z. 2016, ApJ, 829, 97
Hu, H., Liu, Y. D., Zhu, B., et al. 2018, ApJL, submitted
Kaiser, M. L., Kucera, T. A., Davila, J. M., et al. 2008, SSRv, 136, 5
Kozarev, K. A., Raymond, J. C., Lobzin, V. V., & Hammer, M. 2015, ApJ, 799, 167
Kwon, R., Zhang, J., & Olmedo, O. 2014, ApJ, 794, 148
Kwon, R., Zhang, J., & Vourlidas, A. 2015, ApJL, 799, L29
Lee, C. O., Jakosky, B. M., Luhmann, J. G., et al. 2018, GeoRL, 45, 8871
Li, Y., Xue, J. C., Ding, M. D., et al. 2018, ApJL, 853, L15
Liu, M., Liu, Y. D., Yang, Z., Wilson, L. B., & Hu, H. 2018, ApJL, 859, L4
Liu, Y., Luhmann, J. G., Müller-Mellin, R., et al. 2008, ApJ, 689, 563
Liu, Y., Luhmann, J. G., Bale, S. D., & Lin, R. P. 2011, ApJ, 734, 84
Liu, Y. D., Luhmann, J. G., Möstl, C., et al. 2012, ApJL, 746, L15
Liu, Y. D., Luhmann, J. G., Lugaz, N., et al. 2013, ApJ, 769, 45
Liu, Y. D., Hu, H., Zhu, B., Luhmann, J. G., & Vourlidas, A. 2017a, ApJ, 834, 158
Liu, Y. D., Zhao, X., & Zhu, B. 2017b, ApJ, 849, 112
Lopez, R. E. 1987, JGR, 92, 11189
Luhmann, J. G., Mays, M. L., Li, Y., et al. 2018, SpWea, 16, 557
Magdalenić, J., Marqué, C., Krupar, V., et al. 2014, ApJ, 791, 115
Maloney, S. A., & Gallagher, P. T. 2011, ApJL, 736, L5
Martinez-Oliveros, J. C., Raftery, C. L., Bain, H. M., et al. 2012, ApJ, 748, 66
Möstl, C., Farrugia, C. J., Kilpua, E. K. J., et al. 2012, ApJ, 758, 10
Odstrčil, D., Riley, P., & Zhao, X. P. 2004, JGR, 109, A02116
Ontiveros, V., & Vourlidas, A. 2009, ApJ, 693, 267
Patsourakos, S., & Vourlidas, A. 2012, SoPh, 281, 187
Reames, D. V. 1999, SSRv, 90, 413
Reiner, M. J., Kaiser, M. L., & Bougeret, J.-L. 2007, ApJ, 663, 1369
Reisenfeld, D. B., Gosling, J. T., Forsyth, R. J., Riley, P., & St. Cyr, O. C. 2003, GeoRL, 30, 8031
Riley, P., Caplan, R. M., Giacalone, J., Lario, D., & Liu, Y. 2016, ApJ, 819, 57
Seaton, D. B., & Darnel, J. M. 2018, ApJL, 852, L9
Thernisien, A. F. R., Howard, R. A., & Vourlidas, A. 2006, ApJ, 652, 763
Volpes, L., & Bothmer, V. 2015, SoPh, 290, 3005
Vourlidas, A., Wu, S. T., Wang, A. H., Subramanian, P., & Howard, R. A. 2003, ApJ, 598, 1392
Vourlidas, A., & Ontiveros, V. 2009, in AIP Conf. Proc. 1183, Shock Waves in Space and Astrophysical Environments, ed. X. Ao, R. H. Burrows, & G. P. Zank (Melville, NY: AIP), 139
Webb, D. F., & Howard, T. A. 2012, LRSP, 9, 3
Yan, X. L., Yang, L. H., Xue, Z. K., et al. 2018, ApJL, 853, L18
Zhang, J., Dere, K. P., Howard, R. A., Kundu, M. R., & White, S. M. 2001, ApJ, 559, 452
Zhao, X., Liu, Y. D., Hu, H., & Wang, R. 2017, ApJ, 837, 4
Zhao, X. P., Plunkett, S. P., & Liu, W. 2002, JGR, 107, 1223
Zhu, B., Liu, Y. D., Kwon, R.-Y., & Wang, R. 2018, ApJ, 865, 138
Zucca, P., Morosan, D. E., Rouillard, A. P., et al. 2018, A&A, 615, A89
{width="18pc"}{width="18pc"}
[^1]: Note that there is a small convex-outward structure at the nose (see Figure 1), which is not covered by the present shock modeling. Its speed may be higher than given here. @gopalswamy18 obtain a larger speed (about 4000 km s$^{-1}$) using a spherical shock model to cover the convex-outward structure, but other parts of the shock around the nose are not modeled well.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe a method performing $w$-projection using the fast Gauss transform of @Strain91. We derive the theoretical performance, and simulate the actual performance for a range of $w$ for a canonical array. While our implementation is dominated by overheads, we argue that this approach could for the basis of a higher-performing algorithms with particular application to the Square Kilometre Array.'
author:
- |
K. W. Bannister$^{1,2,3}$[^1], T. J. Cornwell$^1$\
$^{1}$CSIRO Astronomy and Space Science, PO Box 76, Epping NSW 1710, Australia\
$^{2}$Bolton Fellow\
bibliography:
- 'Master.bib'
date: 'Accepted 2011 XXX XXX. Received XXX December XX; in original form XXX XXX XXX'
title: 'Memory-efficient $w$-projection with the fast Gauss transform'
---
\[firstpage\]
techniques: interferometric
Introduction
============
Interferometers are composed of an array of antennas arranged in a 3 dimensional volume. For long observations, even instantaneously planar arrays have baselines that are non-coplanar due to Earth rotation synthesis. Correcting non-coplanar baselines in interferometers require the use of a de-projection technique to compensate for the non-coplanarity. To date, $w$-projection [@Cornwell08] is the most computationally efficient algorithm known.
Interferometric images are usually formed by convolutional resampling of the measured visibilities on a regularly sampled, two dimensional $uv$ plane. Convolutional resampling operates by multiplying each visibility by a convolution function and adding the result to the $uv$-plane. In the simplest case, the convolution function is used for anti-aliasing purposes, although it can be used to compensate for primary beam affects, or in the case of $w$-projection, to compensate for non-coplanar baselines. $w$-projection uses a convolution function that is dependant on $w$, which is the convolution of an anti-aliasing function, and a Fresnel term.
While more efficient that other methods [@Cornwell08], $w$-projection does have some shortcomings. Firstly, the convolution functions take time to compute, and for a short observation, can dominate the time to compute an image. Secondly, the amount of memory to store the convolution functions can be large, and can exhaust the capabilities of a computing node. Finally, multiply-add operation is generally memory-bandwidth bound, so that the central processing unit (CPU) spends the majority of the time waiting for the data to arrive, and relatively little time doing the multiply-add operation.
The memory constraints of $w$-projection are of particular concern. The trend in computing has been for memory bandwidth to increase much more slowly than arithmetic capacity, which has roughly doubled every 18 months (Moore’s law), with this trend likely to continue for the foreseeable future. Therefore, the performance of memory-bandwidth bound algorithms such as $w$-projection will not improve as quickly as an algorithm which is bound by arithmetic capacity. Furthermore, modern high performance computers rely on having a very large number of small nodes, with each node having relatively limited memory. The requirement to store large convolution functions is, therefore, at odds with having nodes with small amounts of memory.
Memory efficient algorithms, therefore, will be an important step on the path to implementing wide-field imaging capabilities for the Square Kilometre Array (SKA). We have embarked on an program to research new approaches to interferometric imaging, with particular emphasis on algorithms that align more closely with the memory-bound computing that will be available in the time frame of the SKA.
In this paper we describe a different approach to $w$-projection. The key to our method is to represent the convolution function, including the $w$-projection term, and an anti-aliasing function, as a complex Gaussian. The gridding and degridding problems can then be solved using the variable-width fast Gauss transform (FGT; @Strain91). One advantage of this approach is that it does not require convolution functions to be computed and stored, and the theoretical memory bandwidth is less than standard gridding, in some situations.
In section \[sec:gauss\_wproj\] we describe Gaussian anti-aliasing functions used for $w$ projection and in section \[sec:wproj\] we apply these functions with the Fast Gauss Transform, to the gridding and degridding problems. In section \[sec:performance\] we derive the theoretical operations and memory requirements, and in section \[sec:perf\_comparison\] we compare the theoretical requirements for our method and the standard methods. In section \[sec:implementation\] we compare the theoretical results with a real-world implementation. We discuss our results in \[sec:discussion\] and we draw our conclusions in section \[sec:conclusions\].
$w$-projection with a Gaussian anti-aliasing function {#sec:gauss_wproj}
=====================================================
In this section we describe the gridding process, and use a Gaussian anti-aliasing function to obtain a closed form expression for the convolution function.
We acknowledge that the anti-aliasing properties of a Gaussian are not ideal. Typical interferometric imaging uses prolate spheroidal wavefunctions, which have a optimal out-of-band rejection [@Schwab80]. Nonetheless, Gaussian functions have simple analytical relationships (e.g. the convolutions and Fourier transforms of Gaussians are both Gaussians) that allow us to proceed.
Gridding
--------
The aim of convolutional gridding is to take visibilities sampled on arbitrary $u,v,w$ coordinates and interpolate them onto a regular grid so that a 2D fast Fourier transform (FFT) can be used to estimate the sky brightness distribution [@Taylor99 p. 134ff]. Mathematically, the grid is evaluated at a point $(u_c, v_c)$ by multiplying each visibility by a shifted convolution function according to:
$$F(u_c, v_c) = \sum_{k=1}^{M}{C(u_c - u_k, v_c - v_k, w_k) V(u_k, v_k, w_k)} \label{eq:gridsum}$$
where $M$ is the number of visibilities in to be gridded, $u$, $v$, and $w$ are the coordinates of the projected baseline (in wavelengths), $C(u, v, w)$ is a convolution function and $V(u, v, w)$ is the measured visibility weight. Typically, $C$ is assumed to be zero outside some region of support (6–8 pixels). If we take no account for the $w$ term, then the convolution function is chosen to be an anti-aliasing function, and is independent of $w$, i.e. $C(u,v,w) = A(u,v)$. If we account for the $w$ term using $w$-projection of @Cornwell08, then the convolution function is given by
$$C(u, v, w) = A(u, v) * G(u, v, w) \label{eq:convfunc}$$
where $*$ denotes convolution, $A(u, v)$ is an anti-aliasing function, and
$$G(u, v, w) = \frac{i}{w} \exp \left (-\pi i [u^2 + v^2]/w \right ) \label{eq:fresnel}$$
is the Fresnel diffraction term required by the $w$ projection algorithm.
Radially symmetric Gaussian anti-aliasing functions
---------------------------------------------------
For a fixed $w$, equation \[eq:convfunc\] is the convolution of the anti-aliasing function with a complex Gaussian. If we choose a Gaussian for the anti-aliasing function $A$, the resulting convolution function is also Gaussian. To simplify notation, we will consider the radially symmetric problem by introducing a new parameter $t^2 = u^2 + v^2$, and can now write the Fresnel term as:
$$\begin{aligned}
G(t, w) & = & \frac{i}{w} \exp\left (- \frac{\pi i t^2}{w} \right ) \\
& = & \frac{i}{w} \exp \left ( \frac{-t^2}{ i \delta_G} \right )\end{aligned}$$
with $\delta_G = - w/ \pi$. We can also write the Gaussian anti-aliasing function as
$$A(t) = \exp{\left ( \frac{-t^2}{\delta_A}\right )}.$$
The gridding function is given by the convolution $G(t) * A(t)$, which is another Gaussian, whose variance is equal to the sum of the variances of the original Gaussians, i.e., to within a scaling factor:
$$\begin{aligned}
C(t, w) & = & A(t) * G(t, w) \\
& = & \frac{i}{w} \exp \left ( - t^2\frac{1}{\delta_{A} + i \delta_{G}} \right ) \\
& = & \frac{i}{w} \exp \left ( - t^2 \frac{\delta_{A}}{\delta^2_{A} + \delta^2_{G}} \right )
\exp \left ( -t^2 \frac{ - i \delta_{G}}{\delta^2_{A} + \delta^2_{G}} \right ) \label{eq:prodconv}\\
& = & \frac{i}{w} \underbrace{\exp \left ( - \frac{t^2}{\delta_R} \right ) }_\textrm{real envelope}
\underbrace{ \exp \left ( - \frac{i t^2}{ \delta_I} \right )}_\textrm{complex chirp} \label{eq:conv} \\\end{aligned}$$
The width of the envelope is given by $$\delta_R = \frac{\delta_A^2 + \delta_G^2}{\delta_A}, \label{eq:deltar}$$
and width of the complex chirp is
$$\delta_I = \frac{\delta_A^2 + \delta_G^2}{\delta_G}.$$
The convolution function is, therefore, a complex Gaussian, which can be expressed as the product of real-valued Gaussian envelope (the $\delta_R$ term in equation \[eq:conv\]) with a complex chirp (the $\delta_I$ term). Clearly, as $|w|$ increases, the width of the Gaussian envelope $\delta_R$ increases.
For a given visibility indexed by $j$ the form of the convolution function can be calculated from its $w$ coordinate using: $$C(t, w_j) = \frac{i}{w_j} \exp{\left ( \frac{-t^2}{\delta_j} \right )} \label{eq:conv_ctw}$$
where $$\delta_j = \delta_A - i \frac{w_j}{\pi} \label{eq:conv_variance}$$
is the width of the relevant convolution function, and $w_j$ is the $w$ coordinate of the visibility.
Conversion from wavelengths to pixels
-------------------------------------
Up to this point, the $u,v,w$ co-ordinates and the Gaussian widths have been in units of wavelength. For the remainder of this paper, width will often be quoted in terms of pixels. We use the terms pixel and $uv$-cell interchangeably. To convert from wavelengths to pixels, a size of $uv$-cell is required. The size of the $uv$-cell, (in wavelengths) is set by the inverse of the desired field of view (in radians) i.e. $\phi = 1/\theta_{\rm f.o.v}$.
Therefore, a Gaussian width in pixels can be calculated from the width in wavelength by:
$$\delta (\rm pixels) = \frac{\delta (\rm wavelengths)}{\phi^2}$$
$w$ projection with the Fast Gauss Transform {#sec:wproj}
============================================
A brief introduction to the Fast Gauss Transform
------------------------------------------------
In this section we introduce the Fast Gauss Transform (FGT) with variable scales [@Strain91] (hereafter S91), which is an approximate technique for calculating the sum of Gaussians. We conform to the language of S91, which uses the term ‘source’ to describe a Gaussian function with a given position, amplitude and width. We will describe the application of the FGT onto the gridding problem in section \[sec:gridding\].
The FGT works by partitioning the evaluation region into square boxes, with each box containing a set of Taylor coefficients (Fig. \[fig:grid\]). For each Gaussian ‘source’, the Taylor coefficients for the boxes surrounding the source position are updated, with the size of the updated region being defined by the ‘width’ of the source, i.e. wider Gaussians update more boxes. When all sources have been applied, the Taylor coefficients are evaluated at any number of target locations. The algorithm is parameterised by three numbers: $r$, which sets the box size, $p$, which defines number of Taylor coefficients to consider, and $\delta$ which sets the smallest size of Gaussian that will be considered. Reducing the box size, or increasing the number of Taylor coefficients increases the accuracy of the sum.
Gridding with the Fast Gauss Transform {#sec:gridding}
---------------------------------------
Now that we have a Gaussian convolution function (Equation \[eq:conv\_ctw\]), we can solve the sum of Equation \[eq:gridsum\] the FGT with *source-dependant* scales. In the language of gridding visibilities, the evaluation region is the $uv$ plane. The ‘sources’ are the visibilities, with the source position being given by the coordinates of the visibility on the $uv$ plane. The width of the source is related to the $w$ coordinate of the visibility as in Equation \[eq:conv\_variance\], with the minimum width when $w=0$, being essentially the width of the anti-aliasing function. When all the visibilities have been applied to the Taylor coefficients, the Taylor coefficients are evaluated on a regular grid, for each $uv$ cell on the $uv$ plane.
![This figure illustrates the process of gridding or degridding a single visibility. The $uv$ plane of size $N_{\rm pix}$ is partitioned into boxes size $L$. The box size can be the same size as a $uv$ cell. Each box contains $(p - p_d+1)^2$ Taylor coefficients. To grid or degrid a visibility indexed by $j$, all boxes within $R_j$ of the visibility $uv$ position ($t_j$) are updated (circular blue shaded region). Within each box, $q^2$ Taylor coefficients are updated with no cheating. If cheating is enabled (i.e. $p_d > 0$), then only $q'^2$ of the $(p-p_d+1)^2$ Taylor coefficients (orange shaded region) are updated. $R_j$ has been exaggerated for clarity.[]{data-label="fig:grid"}](fig1-0.eps)
For a detailed description of the algorithm, we refer the reader to S91. For the purposes of explaining our implementation, we outline the algorithm in the following text, using similar notation to S91. This description applies to the one-dimensional case, but S91 describes the relationships to extend it to the multi-dimensional case (using multi-index notation), for which the 2D case is relevant here. Additionally, this description is only valid for real Gaussians. The extension to complex Gaussians will be considered in the next section.
The algorithm proceeds as follows:
1. Given values of $r$ and $p$ calculate:
$$\epsilon = \frac{r^p_p}{1 - r_p}, \label{eq:epsilon}$$
where $$\begin{aligned}
r_p & = & r\sqrt{\frac{e}{p+1}} < 1\end{aligned}$$
2. Partition the $uv$ plane into the set of boxes $B$ with centers $t_B$, and side length $L = r\sqrt{2 \delta}$ parallel to the axes, where $\delta = \min_j \delta_j = \delta_A$ is the smallest Gaussian width to be considered.
3. For each visibility indexed by $j$, calculate the range $R_j = \sqrt{- \delta_j \log \epsilon}$. Also calculate the order $q \le p$ satisfying
$$\frac{r^q_{q,j}}{1 - r_{q,j}} \le \epsilon \label{eq:qerror}$$
where $r_{q,j} = r\sqrt{e \delta/\delta_j (q + 1)}$.
4. For each box within $R_j$ of the visibility position $t_j$, update the Taylor coefficients according to:
$$C_\beta = \frac{1}{\beta !} \sum_{\delta_j \ge b} v_j \left (\frac{\delta}{\delta_j} \right )^{|\beta|/2} h_\beta \left( \frac{t_j - t_B}{\sqrt{\delta}} \right) \label{eq:cbeta}$$
where $\beta$ runs from $0$ to $q$, $v_j$ is the visibility weight, and $h_\beta$ is the Hermite function (see S91).
5. Once all the visibilities have been applied, evaluate the Taylor coefficients at each cell in the $uv$ plane, by finding the box in which the $uv$ cell is situated, and calculating the sum:
$$F(t) = \sum_{\beta \le p} C_\beta \left( \frac{t_C - t_B} {\sqrt{\delta}}\right) ^\beta$$
where $t_C$ is the position of the $uv$ cell, and $\beta$ runs from $0$ to $p$.
The error in the evaluation of the sum is bounded by: $$E_p \le V \frac{r^p_p}{1 - r_p},$$
where $$\begin{aligned}
V = \sum_{j=1}^{M} |v_j|.\end{aligned}$$
$r$ and $p$ are chosen to make $|E_p| \le V \epsilon$.
Extension to complex Gaussians
------------------------------
The formulation of S91 is explicitly for real Gaussians, i.e. with $\delta_j, v_j \in \Re$, but in our case $\delta_j, v_j \in \mathds{C}$. A full derivation of the algorithm for complex values is outside the scope of this work . The important questions are whether the series of Equation \[eq:cbeta\] converges (and for what values of a complex $\delta$) and what is a suitable range ($R_j$).
Empirically, we have determined that:
1. Using a range of $R_j = \sqrt{-\delta_{R,j} \log \epsilon}$ provides good results. This also makes sense because it is the amplitude of the envelope that determines the range.
2. For calculating the order $q$ (Equation \[eq:qerror\]) we used $$\delta_j = max(\delta_A, \delta_I) \label{eq:delta_error_complex}$$
Other possibilities (e.g. $\delta_j = \delta_R$) reduced $q$ too aggressively as $w$ increased, which resulted in very large errors.
In our case, the minimum value of $\delta$ occurs when $w=0$ which also coincides with $\Im(\delta) = 0$ so the box size is easily set as equal to $L = r \sqrt{2 \delta_A}$.
Optimisations {#sec:optimisations}
-------------
S91 describes a number of optimisations that can be used in the FGT with source-dependent scales:
1. S91 introduces a single set of Taylor coefficients, positioned at the center of the $uv$ plane, to capture Gaussians with the largest scales. Visibilities larger than an upper scale size $\delta_j > b$ can be added to this single set, rather than the sets associated with the boxes. S91 recommends the upper scale size be at least 10% of the size of the evaluation region, in order to allow for practical precision in evaluating the large number of Taylor coefficients. In practice, most array configurations would have relatively few baselines with such a large $w$ that their convolution functions would span $>$ 10% of the $uv$ plane, so this optimisation has limited impact.
2. S91 introduces a lower scale $a$ below which the Gaussian sum is directly evaluated. If there are only a few Gaussians with scales smaller than $a$, we can increase the box size, which reduces the memory and arithmetic capacity required, while incurring minimal penalty from direct evaluation. In practice, any interferometric observation is likely to be arranged such that the mean value of $w$ is approximately zero. Therefore, there are unlikely to be outlying, small Gaussians that would allow for increasing the box size. Once again, this optimisation has limited impact.
3. S91 proposes doing direct evaluation for boxes containing only a few evaluation points. As we are evaluating on a regular grid, we have the same number of evaluation points inside each box, so applying this optimisation would degenerate into standard gridding.
4. S91 proposes a method for avoiding needlessly accurate computation, i.e. for large scales (which are smoother than small ones) more boxes are updated, but smaller number of Taylor coefficients (i.e. the order $q$ described in Section \[sec:gridding\]) per box can be updated. This optimisation *is* useful and derives the most benefit for visibilities with large scales ($|w|$ large, see section \[sec:performance\].). We have incorporated this optimisation in Equation \[eq:qerror\].
5. S91 notes that the equation to calculate the order (Equation \[eq:qerror\]) overestimates the error by at sometimes two orders of magnitude. Memory bandwidth and operations can be saved by ‘cheating’ on the value of $q$ by decrementing $q$ by some value. This strategy is described in the section \[sec:cheating\].
6. The fact that each uv-cell contains a number of Taylor terms affords an additional axis available for parallelisation. The strategy is described in section \[sec:parallelisation\].
### Cheating {#sec:cheating}
There are two sources of error when a sample is gridded (see Figure \[fig:error\_pattern\]). The first is from truncating the support; that is, only boxes only within a finite range $R_j = \sqrt{- \delta_R \log \epsilon}$ are updated. Convolutional gridding suffers from the same type of truncation error. The second source of error is from truncation of the Taylor terms; that is, only $q < \infty$ Taylor terms are updated. S91 notes that the equation for calculating the error from truncation of Taylor terms (Equation \[eq:qerror\]) overestimates the error by several orders of magnitude. If this is the case, for a given value of $\epsilon$, the error will be dominated by the truncated support, and the contribution from the truncation of the Taylor terms will be negligible. In principle, therefore, one can reduce the number of Taylor terms without substantially increasing the overall error, as long as it remains less than the error from the truncation of support.
In order to balance the errors from both effects, we introduce an additional parameter to the algorithm $p_d$, which decrements value of $q$, so that the actual value of $q$ that is used is $q' = \rm{max}(q - p_d, 0)$. This reduces the CPU and memory requirements of updating a given box, at the expense of increasing the errors due to truncating the Taylor terms.
### Parallelisation {#sec:parallelisation}
One interesting property of our approach is that affords an additional axis of parallelisation (the Taylor terms) in addition to those available for standard gridding (typically, data parallelisation over frequency channels). We assume that the parallelism is achieved through message passing.
In section \[sec:performance\] we show that the memory bandwidth and operations count for the FGT is dominated by updating the Taylor terms. As each of the Taylor terms is independent, one can in principle store each of the $(p-p_d + 1)^2$ Taylor terms on (up to) as many nodes, thereby dividing the per-node storage requirement by the number of nodes. For the case where $q'^2$ is a multiple of the number of nodes, the work is spread equally among the nodes. Each node can update its Taylor terms for the boxes within range, and, the total memory bandwidth is increased by the number of nodes.
One penalty of parallelising in this way is that each node must calculate the same interim values (e.g. the values of the Hermite polynomials). This duplication can be reduced by partitioning the Taylor terms among fewer nodes, with each node being responsible for a particular row or column of the Taylor terms.
If $q'^2$ is less than the number of nodes, then some nodes will have no work to perform, as they are responsible for Taylor coefficients that are not being updated. One must be careful, therefore, to choose an algorithm parameterisation i.e. ($r$,$p$ and $p_d$) that guarantees never to give $q'^2$ less than the number of nodes.
The efficiency of this approach depends crucially on the properties of the computing nodes. It is useful if the nodes are storage or memory bound, but not if the cost of computing the interim values dominates the computing time. As this tradeoff requires intimate knowledge of the particular computing architecture, we will not pursue a detailed analysis of parallelisation in this paper.
Degridding with the Fast Gauss Transform
----------------------------------------
Degridding is used as part of the major cycle of Cotton-schwab clean [@Rau11]. Degridding involves taking the dot-product of the $uv$ plane with the convolution function shifted to the location of a visibility at an arbitrary location (i.e. not at the center of a $uv$ cell). It is the dual of gridding.
Degridding can be performed analogously to gridding using the fast Gauss transform with *target*-dependent scales (S91). All the same comments about optimisations for gridding (Section \[sec:optimisations\]) apply as for degridding.
Key differences between classical gridding and the FGT
------------------------------------------------------
The key conceptual differences between classical gridding and FGT gridding are described in Table \[tab:vs\]
Property Classical Gridding Fast Gauss Transform
--------------------------------------- --------------------------------------- -------------------------------------------------------------------------------------------------------------
Accuracy: Exact Approximate, but with controllable error.
Anti-aliasing function: Arbitrary. Gaussian only.
Fundamental scale: A $uv$-cell, set by imaging geometry. A box, which can be smaller, or larger than a $uv$-cell.
Gridding a single visibility updates: Cells on the $uv$-plane. Taylor coefficients in a set of boxes. The cells of the $uv$-plane is calculated as a post-processing step.
\[tab:vs\]
Theoretical performance {#sec:performance}
=======================
Here we derive the theoretical performance of the gridding and degridding processes using the fast Gauss transform, and traditional convolutional gridding.
Operations count of the fast Gauss transform
--------------------------------------------
### Gridding
To grid a single visibility of width $\delta_j$, the number of boxes that are within a circular region within the range $R_j = \sqrt{-\delta_R \log \epsilon}$ is approximately: $$N_B = \frac{\pi}{4}(2 \lceil R_j/L \rceil + 1)^2 .$$
The the order $q \le p$ is chosen to satisfy the error estimate:
$$\frac{r^q_{qj}}{1 - r_{qj}} \le \epsilon$$
where $r_qj = r\sqrt{e \delta/\delta_j(q + 1)}$.
To apply cheating (Section \[sec:cheating\]), we decrement the value of $q$, such that $q' = \rm{max}(q - p_d, 0)$.
Now we consider work required to update the Taylor coefficients in each box (Eq. \[eq:cbeta\]). The key is to compute and store the Hermite functions, and powers of $(\delta/\delta_j)^{|\beta|/2}$ separately, and perform the sum at the end, by taking advantage of the product form of the multi-index notation for multiple dimensions.
The Hermite function is a polynomial multiplied by a complex exponential. The argument for the complex exponential is the same for every order $q'$, so only one calculation is required per set of Hermite function evaluations. The way in which a complex exponential is computed by a CPU is highly implementation-dependant, so we simplify the analysis here and assume that it requires $N_{\rm cexp}$ floating point operations.
We observe that the Hermite function of order $q'$ has only $N_c=\lfloor q'/2 \rfloor + 1$ nonzero coefficients. Second, we note that we will evaluate the Hermite functions for order $0 \le \beta \le q'$ with the same argument. Therefore, we can calculate the powers of the argument initially with $q'$ operations. Evaluating each Hermite function requires only $N_c$ additional operations. Therefore, to compute all Hermite functions up to order $q'$ requires approximately: $$\begin{aligned}
N_{\rm ops, hermite} & = & N_{\rm cexp} + q' + 2 \sum_{\beta = 0}^{N_c} N_c \nonumber \\
& = & N_{\rm cexp} + q' + 2 \sum_{\beta = 0}^{N_c} \lfloor \beta/2 \rfloor + 1 \nonumber \\
& \simeq & N_{\rm cexp} + q' + N_C (N_C + 1)\end{aligned}$$
The multi-index powers of $t_\beta = (\delta/\delta_j)^{|\beta|/2}$ can be computed recursively, by computing $t_n = (\delta/\delta_j) t_{n-2}$ with $t_0=(\delta/\delta_j)^{1/2}$, and $t_1=(\delta/\delta_j)^1$. This requires $2(q'+1)$ operations $0 \le n \le 2q'$.
The weight $v_j$ can be folded into the power values also, with $2(q'+1)$ operations.
Finally, the full sum to update $C_\beta$ over two dimensions requires the multiplication of Hermite functions for 2 dimensions, and the powers, plus the accumulation, which requires $3 (q'+1)^2$ operations in total.
The total number of operations to update a single box is therefore the sum of two Hermite evaluations (one for each dimension), plus the powers of $\delta/\delta_j$, the weights and the final sum, i.e.:
$$N_{\rm grid} = 2 (N_{\rm cexp} + q' + N_C[N_C + 1]) + 4(q'+1) + 3(q'+1)^2$$
For small $q'$, the cost is dominated by the cost of evaluating the complex exponential. For large $q'$, it is dominated by the multiply-add step in the final sum.
### Degridding
The operations count for the degridding process is similar to the gridding case. The order $q'$ is evaluated identically, as are the Hermite functions and the powers of $\delta/\delta_j$. The only difference is that no weight is included, and the multiply-add stage includes the product of the two hermite functions, the power and the Taylor term, requiring $4(q'+1)^2$ operations.
The number of operations per visibility is therefore: $$N_{\rm degrid} = 2(N_{\rm cexp} + q' + N_C[N_C + 1]) + 2(q'+1) + 4(q'+1)^2.$$
Once again, for small $q'$, the cost is dominated by the cost of evaluating the complex exponential. For large $q'$, it is dominated by the multiply-add step in the final sum.
### Memory bandwidth
Calculating memory bandwidth is complicated by the issue of cache hierarchy. We assume the simplest case: i.e. no caching. In practice, this is a reasonable assumption, as visibilities are often stored in no particular order. Therefore, the desired grid (and convolution functions in the case of convolutional gridding) are essentially random, meaning that all memory accesses go to the main memory and bypass the caches.
For the fast Gauss transform, gridding a visibility requires the updating of $(q'+1)^2 N_B$ complex coefficients. The coefficients must be read into the processor, updated, and written back to memory, requiring $2 (q'+1)^2 N_B$ memory transactions per visibility.
For degridding, the coefficients do not need to be written back to memory, therefore the only $(q'+1)^2 N_B$ memory transactions are required.
### Storage
For an image size of $N_{\rm pix}^2$, the gridding operation requires storage for $(N_{\rm pix}/L)^2 (p- p_d+1)^2$ complex coefficients. Degridding requires the same storage.
Performance of convolutional gridding and degridding
----------------------------------------------------
In this section we derive the operations count, memory bandwidth and storage required to perform gridding and degridding with stored convolution funcitons. In order to put the two methods on equal footing in terms of imaging performance, we will use Gaussian anti-aliasing functions for the $w$-projection, rather than the commonly-used prolate spheroidal wavefunctions. This means that the imaging performance of the two approaches is essentially the same (except for the errors in truncating the Taylor series of the FGT) and the support size of the convolution functions is also easily computed.
We assume standard $w$ projection computes the convolution functions in advance and stores them in memory. We assume a convolution function of 1-D size for the $k$th $w$-plane $M_{\rm k} = k \Delta w$, where $k$ is an integer, and $\Delta w$ difference in size between different $w$ planes. Usually the cached convolution function is oversampled by a factor $\kappa$ of 4–8 in order to accurately grid visibilities whose coordinates are not exactly in the center of a $uv$ cell.
We assume that we truncate the convolution function when the real envelope reaches a value $\epsilon$, which is at a distance $t_\epsilon = \sqrt{- \delta_R \log \epsilon}$ from the centre of the convolution function. Therefore, the 1D support size in pixels, of the $k$th $w$-plane is (by substituting Equation \[eq:deltar\])
$$\begin{aligned}
M_{k} & = &2 t_\epsilon \nonumber \\
& = & 2 \sqrt{\log \epsilon \left (\delta_A + \frac{k^2 \Delta w^2}{\pi^2 \delta_A} \right )}\end{aligned}$$
### Operations count
Gridding a visibility requires two operations per point (weight times convolution function, add to grid), for a total of $2 M_{\rm k}^2$ operations per visibility. Degridding also requires two operations per point(convolution function times grid, add to result) and and therefore requires $2 M_{\rm k}^2$ operations per visibility.
### Memory bandwidth
Gridding requires the convolution function, and the grid position to be retrieved and written back to memory, requiring $3 M_{k}^2$ memory transactions per visibility. Degridding has no requirement to write back the result, (as the intermediate sum is stored in a register), so only $2 M_{\rm k}^2$ memory transactions are required.
### Storage
To compute the total memory required, we will determine the amount required to store $N_w$ $w$-planes for values of $w$ between $[0, +w_{\rm max}]$, uniformly sampled with width $\Delta w = w_{\rm max}/N_w$. In practice, we need to store $w$-planes for $[-w_{\rm max},0)$, so we will take the above result and multiply by two.
The amount of memory required to store the convolution functions is therefore equal to the sum of the oversampled convolution functions, i.e.:
$$\begin{aligned}
\frac{N_{\rm mem}}{2} & = & \sum_{k = 0}^{N_w}{(\kappa M_k)^2} \nonumber \\
& = & -4 \kappa^2 \log \epsilon \left (N_w \delta_A + \frac{\Delta w^2}{\pi^2 \delta_A} \frac{N_w}{6} [N_w + 1][2 N_w + 1] \right ) \\
& \simeq & -\frac{4}{3} \kappa^2 \log \epsilon \frac{w_{\rm max}^2}{\pi^2 \delta_A} N_w\end{aligned}$$
Theoretical Performance Comparison for Gridding {#sec:perf_comparison}
===============================================
We consider the gridding problem for two scenarios, $L=1$ and $L=2$. We assuming $\delta_A=1$ pixel, a required accuracy of $\epsilon=10^{-3}$, a 1 degree field of view and a 6 km maximum baseline at $\lambda = 0.2$ giving a $w_{\rm max}$ of $30 k \lambda$. The convolution function size at maximum $w$ of $M_{\rm wmax}=191$ pixels. We assume the cost of evaluating a complex exponential is $N_{\rm cexp} = 100$ flops. This estimate is also justified by measurements of our implementation (Fig. \[fig:procratio\_vs\_w\]).
The FGT method for $L=1$ and $L=2$ is outperformed by standard gridding both in terms of operations and memory bandwidth, if cheating is not enabled (Figure \[fig:fgt\_vs\_gridding\]). For the $L=1$ case in particular, the operations count is dominated by the complex exponential.
If cheating is enabled, then the situation is substantially improved. Once the $w$ value reaches above a certain threshold, the order $q' = 0$, and only one Taylor coefficient is updated per box. In the $L=1$ case, the memory bandwidth is reduced to 0.5 of the standard gridding case, because the FGT only requires only two memory transactions per pixel (read + write coefficient), while standard gridding requires three (read pixel + read convolution function + write pixel). For $L=2$ the improvement in memory bandwidth is even more with the FGT requiring 10 times less memory bandwidth than standard gridding. The FGT is more efficient as it only updates one coefficient for each box, which encapsulates 4 pixels. The FGT case is also improved because it only updates boxes within a circular region, while standard gridding updates all pixels within a square.
This encouraging result leads to a number of questions: is the FGT bound by memory bandwidth? Can these performance gains be realised in practice? Can cheating with $p_d=3$ give reasonable image errors? We will address these questions in the following section.
![Gridding with the fast Gauss transform can save memory bandwidth for large support sizes, at a cost of floating point operations. Plotted here is the ratio of operations (top panel) and memory bandwidth (bottom panel) for the fast Gauss transform compared with convolutional gridding (see Section \[sec:performance\]). The x-axis is the value of the $w$ coordinate, which is also a measure of the width of the convolution function (Equation \[eq:conv\_variance\]). The target error for each algorithm was set to $\epsilon = 10^{-3}$. The fast Gauss transform was configured with $\delta_A=1$ pixel and we have assumed $N_{\rm cexp}=100$. The step drops in the ratio occur when the FGT reduces the order $q'$ for the Taylor truncation. For $w>14 {\ensuremath{\, \mathrm{k\lambda}}}$, the FGT with $L=2$ requires 21.8 times more operations but only 0.1 times the memory bandwidth of classical gridding, but only if cheating is enabled with $p_d=3$.[]{data-label="fig:fgt_vs_gridding"}](fig2-0.eps){width="\linewidth"}
Implementation {#sec:implementation}
==============
We implemented the gridding and degridding algorithms as described in Section \[sec:wproj\] in C++. As with the theoretical simulation, we set the width of the anti-aliasing function equal to $\delta_A=1$ $uv$ cell. We aim explore the parameter space similar to the theoretical analysis, i.e. around $L=1$ and $L=2$ and ranges of errors around $10^{-3} < \epsilon < 10^{-2}$.
To compare the gridding errors, we compared the relative error of a visibility gridded with the FGT with the equivalent complex Gaussian (Equation \[eq:prodconv\]), calculated over $256^2$ pixels (which is larger than the 191 pixels we expect for $\epsilon = 10^{-3}$ and $w=w_{\rm max}$) . To compare the compute times, we compare the time to grid 100 visibilities of fixed $u,v,w$ coordinates with the FGT, with the same number of visibilities gridded with standard gridding. The support size of the gridding kernel was chosen to have with equivalent error to the FGT error, i.e. $M_k^2$ pixels. We chose 101 $w$ planes from $0$ to $w_{\rm max}$. The processing was performed on an Intel Core 2 Duo processor running at 2.66 GHz, with 32 kB of L1 cache, 3 MB of L2 cache and with 8 GB of 1057 MHz DDR3 RAM.
We consider here only the results for the gridding operation, as the degridding algorithm results essentially the same.
Error patterns
--------------
A typical error pattern is shown in Figure \[fig:error\_pattern\]. The absolute error in the gridded visibility contains two components. The circular component is due to the truncation to finite support, i.e. boxes outside the circle were not updated. The vertical and horizontal lines are due to truncation to a finite number of Taylor terms, with the largest error where the grid is evaluated near the boundary of adjacent boxes. The box structure of the gridding process can also be clearly seen in the FFT error, which also shows a box structure.
{height="14cm"}
Peak Image Errors vs $w$
------------------------
The peak error in the FFT of the visibility (i.e. the image plane), as a function of the $w$ for a range of algorithm parameterisations are shown in Figure \[fig:error\_vs\_w\]. To begin with, we consider the case for $L~\sim 1$. Somewhat surprisingly, Equation \[eq:epsilon\], which describes the error in gridding a single visibility on $uv$ plane, also predicts the error the image plane for many parameterisations and values of $w$. For $p_d=0$, and small $w$, the errors are substantially smaller than the predicted $\epsilon$. As $w$ increases, $q'$ is reduced, and the error generally matches the predicted $\epsilon$. For some parameterisations (e.g. $r=0.5$, $p=7$) there is a jump in error above predicted $\epsilon$ at a particular $w$.
For the $L~2$ case, for small $w$, the errors are at or below the theoretical limits in most cases, however in many cases the errors are considerably worse than for $L=1$. For example, for $w$ small in some cases (e.g. $L=2.5$, $p=18$, $p_d=0$) the error *increases* far beyond the theoretical limit. For large $w$, the errors are several orders of magnitude worse than predicted.
{width="\linewidth"} {width="\linewidth"}
Processing time
---------------
The measured processing time for the FGT vs. standard gridding is shown in Figure \[fig:procratio\_vs\_w\]. Our implementation operates substantially slower than standard gridding for all parameterisations and ranges of $w$, and the operations rate is well matched by the our theoretical model for the operations count.
{width="\linewidth"} {width="\linewidth"}
Discussion {#sec:discussion}
==========
FGT Implementation
------------------
The results from testing our implementation are somewhat conflicting. On the plus side, Figure \[fig:error\_vs\_w\] clearly shows that, for small values of $w$, Equation \[eq:qerror\] is overestimating the truncation error, and the value of $q'$ being used is too large. This was the original motivation for introducing cheating, and it is clear that the error is still overestimated even for $p_d=3$ in some cases.
Unfortunately, for larger values of $w$, the fact that the error jumps above the predicted $\epsilon$, even with no cheating, implies that our empirical extension to the FGT to complex Gaussians, (in particular Equation \[eq:qerror\], and Equation \[eq:delta\_error\_complex\]) is not correct. As a result of this incorrectness, our current cheating algorithm has limited value. Even with $p_d=1$, most parameterisations contain jumps in error of factors of few to 100 over the range of $w$. A different mapping between $w$ and $q'$ is required to maintain a constant error across $w$.
Increasing the box size above $L>2.3$ appears to have catastrophic affects on the error for small $w$. We suspect this is due either to numerical instability in evaluating high-order polynomials, or more likely, the Hermite expansions simply fail to converge when the Gaussian width is less than some function of the box size.
Our implementation’s processing time is clearly compute-bound as our computing times match our theoretical model for the operations count almost exactly in most instances. The theoretical model suggests the operations count is dominated by the cost of computing the complex exponential, and we expect that is the case, although other overheads may also be responsible. Sadly, we were not able to approach the memory-bound performance that originally motivated this work.
Future work
-----------
In spite of the disappointing performance of our implementation, we hope that some of the ideas from this work could be extended in future. In particular:
- The idea of having tuneable error is attractive. For arrays with high sidelobes in their synthesised beams (i.e. poor $uv$ coverage), large gridding errors can be made, as long as they remain below the clean threshold for the given major cycle.
- The required memory bandwidth of our algorithm is only weakly dependant on $|w|$, so for arrays with very long baselines, our algorithm may be suitable.
- The Hermite expansions were originally proposed by @Greengard91 for real Gaussians. It may be that faster-converging expansions can be found for complex Gaussians - particularly in the case for large $w$ where the real envelope is smooth and the complex chirp contains many peaks.
- The form of the convolution function is smooth at the centre and oscillates more rapidly as $t$ increases, with the fastest oscillations being damped at the edges by the envelope. Currently, our approach is to use the same order for all boxes, which leads to the largest errors being made away from the centre of the convolution function (Figure \[fig:error\_pattern\]) where the oscillations are the fastest and largest. There may, therefore, be some value in varying the order $q'$ as a function of the distance from the centre, thereby applying the majority of the computational effort where the errors are likely to be worst.
- We implemented the polynomial evaluations in a naive manner. More sophisticated methods exist [@TAOCP2 p. 485ff] that could reduce the operations count and improve numerical stability.
- To calculate the complex exponential, we used the [CEXP]{} function from the standard C library, which is accurate but slow. Faster, approximate methods are available[^2]. For an implementation whose run time is dominated by [CEXP]{} such as ours, these methods could improve speed by a factor of a few, albeit at the cost of some accuracy.
- Gaussian anti-aliasing functions are not well suited to imaging, as the alias rejection is not very good. The modern state of the art is to use prolate spheroidal wavefunctions [@Schwab80], that have better alias rejection. Unfortunately, the Hermite functions are not a suitable basis for Taylor expansion of the prolate spheroidal wavefunctions. If suitable analytic multipole expansions of the convolution of the complex chirp, and the prolate spheroidal wavefunctions can be found, then they can be applied with the fast multiple method [@Greengard87; @Greengard88] to obtain a similar result as our FGT, but with better anti-aliasing properties.
- The fast multipole method could also be used to compensate for the primary beam in the $uv$-plane (AW-projection, [@Bhatnagar08]). Once again, this would require analytic multipole expansions of the required convolution functions. While this may not be possible with an arbitrary primary beam, it may be more tractable if we assume Gaussian primary beams.
- For problems where the compute time of $w$-projection is dominated by the time to compute convolution functions (by Fourier transforming a phase screen in the image plane), using standing gridding with Gaussian anti-aliasing functions could be preferable, as they can be directly and efficiently computed in the $uv$ domain.
- Easily parallelised algorithms will clearly be important for large arrays such as SKA, where the computational work is likely to be spread over many thousands of nodes. Once the gridding problem has been distributed over the usual axes of frequency, pointing direction and polarisation, it is possible that problem may still be too large to be efficiently computed by a single node. For the gridding operation, the option of distributing and parallelising over Taylor coefficients adds an additional axis which an be exploited when distributing work among processors, over and above the traditional axes.
Conclusions {#sec:conclusions}
===========
We have described a procedure to perform $w$-projection with the fast Gauss transform with variable scales, where the anti-aliasing function is chosen to be a Gaussian. The gridding problem is solved by the FGT with source variable scales, and the degridding problem by the FGT with target variable scales. While the theoretical efficiency of our approach is encouraging, we were not able to approach the the theoretical performance gains with our implementation. Nonetheless, we find that $w$ projection with approximate algorithms such as the FGT or fast multipole methods may yet have promise, by having the attractive properties of tuneable error, an additional parallelisation axis, and no calculation and storage of convolution functions. The methods require additional research, to improve the practical implementation, find expansions with faster convergence, and find closed forms with better anti-aliasing properties.
Acknowledgements
================
The authors would like to thank Maxim Voronkov and Ben Humphries for invaluable help with C++ coding, Matthew Whiting for the idea of tuneable errors within major cycles, and Oleg Smirnov for his insightful referee comments.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: <http://gruntthepeon.free.fr/ssemath/> (e.g. [@Schroudolph98]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Based on two dissipative models, universal asymptotic behavior of flow equations for Hamiltonians is found and discussed. Universal asymptotic behavior only depends on fundamental bath properties but not on initial system parameters, and the integro-differential equations possess an universal attractor. The asymptotic flow of the Hamiltonian can be characterized by a non-local differential equation which only depends on one parameter - independent of the dissipative system or truncation scheme. Since the fixed point Hamiltonian is trivial, the physical information is completely transferred to the transformation of the observables. This yields a more stable flow which is crucial for the numerical evaluation of correlation functions. Furthermore, the low energy behavior of correlation functions is determined analytically. The presented procedure can also be applied if relevant perturbations are present as is demonstrated by evaluating dynamical correlation functions for sub-Ohmic environments. It can further be generalized to other dissipative systems.'
author:
- Tobias Stauber
title: Universal asymptotic behavior in flow equations of dissipative systems
---
Introduction
============
The flow equation method, introduced by G[ł]{}azek and Wilson [@Gla94] and independently by Wegner,[@Weg94] has become a powerful tool to analyze many-body systems in nuclear and solid-state physics.[@Weg01] The method is based on the continuous mapping of a given Hamiltonian on to its unitarily equivalent representations. This induces a flow of the system parameters which is governed by the so-called flow equations. The continuous transformations can be characterized by the anti-Hermitian generator $\eta$ and the flow equations of the Hamiltonian $H$ are then given by the differential form ${\partial_\ell}H=[\eta,H]$, $\ell$ denoting the flow parameter. The objective is to reach a simple fixed point Hamiltonian for $\ell\to\infty$.
The flow equations for many-body Hamiltonians are non-linear, integro-differential equations and in general analytically non-approachable. Yet, it is possible to analyze the asymptotic behavior as was done for the spin-boson model[@Keh96] as well as for an electron-phonon model.[@Len96] It was found that the asymptotic behavior of a given energy scale ${\Delta}$ generally leads to ${\Delta}\to{\Delta}^*+a\ell^{-1/2}$, with ${\Delta}^*\equiv{\Delta}(\ell=\infty)$ and $a$ denoting a dimensionless constant.[@FootEne] So far, $\eta$ has been chosen such that ${\Delta}^*$ was finite and the asymptotic fixed point Hamiltonian has thus been characterized by this renormalized energy scale.
In this paper we want to apply the flow equation technique to dissipative systems. The general Hamiltonian then consists of a system Hamiltonian $H_S$ with energy scale $\Delta_0$, a bath Hamiltonian $H_B$, and the interaction term $H_{SB}$. Instead of integrating out the bath degrees of freedom and thus reducing the effective configuration space, the flow equation method performs a sequence of infinitesimal unitary transformations with the objective that the interaction between the system and the bath becomes zero, i.e. $H_{SB}\to0$. This is done systematically by first “rotating away” interaction terms which connect states with high energy differences. If ${\Delta}^*$ is finite, the interaction between the system and the environment is thus transformed away in the following way: first decoupling the high-energy modes of the bath, then proceeding with decoupling the low-energy modes and in the end decoupling nearly degenerate states around the renormalized energy scale ${\Delta}_r={\Delta}^*\neq0$. Whereas in the first two regimes the coupling constants decrease exponentially as functions of $\ell$, the coupling constant belonging to the renormalized energy scale ${\Delta}^*$ follows an algebraic decay according to $\ell^{-1/2}$, see Ref. .
In order to evaluate spectral properties within the flow equation approach, observables need to be subjected to the same sequence of unitary transformations. This leads to flow equations for observables and the observable flow follows the systematic decoupling procedure of the Hamiltonian, i.e. the correlation function is first built up for high and low energies and the intermediate energy regime around ${\Delta}^*$ is determined at last.
Flow equations are more stable for $\ell$-values far away from degeneracies. This means that the regime which comprises most spectral weight is determined by flow equations which are the most unstable. In case of the spin-boson model, Kehrein and Mielke circumvent this problem by mapping the asymptotic flow of the spin-boson model onto the asymptotic flow of the exactly solvable dissipative harmonic oscillator.[@Keh97] They thus integrated the flow equations numerically up to some finite $\ell^*$ and then employed a conservation law which was derived for the exactly solvable model. Nevertheless, this mapping was restricted to a certain parameter regime[@FootKM] and might not be applied to other, more general dissipative models.
In this paper, we will set up flow equations with a different asymptotic behavior. The crucial point is that the fixed point Hamiltonian $H^*\equiv H(\ell=\infty)$ does not contain any characteristic renormalized energy scale, i.e. ${\Delta}^*=0$ and $H^*$ is thus just given by the non-interacting bath. As a consequence, the system is monotonously decoupled from the bath, starting from high-energy modes and finishing with the low-energy modes. But it also implies that the fixed point Hamiltonian does not contain any other information but the ground-state energy. If the physical renormalized energy scale is denoted by ${\Delta}_r$, the flow equations do not yield ${\Delta}_r={\Delta}^*$ as before. All physical information is transferred on to the observable flow.
The asymptotic behavior of the Hamiltonian is passed on to the observable flow as before. Since the asymptotic flow equations deal with the decoupling of the low-energy modes, the asymptotic behavior of the observable will determine the low-energy regime of correlation functions. The low-energy regime of dissipative systems is generally determined by universal power-laws. We thus find the connection between universal flow equations and universal spectral properties. We will further set up a conservation law, based on asymptotic, scale-invariant spectral functions, which will allow us to halt the numerical integration after some finite value $\ell^*$. Like this, we also obtain analytical results for the low energy behavior of correlation functions. The procedure should not be restricted to specific models as before.
There are other advantages of flow equations which lead to a trivial fixed point Hamiltonian only consisting of the free bath. First, one can easily include a logarithmic grid around $\omega=0$ to increase the numerical precision while decoupling almost degenerate states. Second, the spectrum of a dissipative system at low coupling[@FootBS] is generally that of the free bath, consisting of a ground-state and a gap-less continuum.[@Bac98] Unitary transformations do not alter the spectrum, but if the fixed point Hamiltonian consisted of the system with the renormalized energy scale ${\Delta}^*\neq0$ plus the free bath, the continuous spectrum of the bath would be superposed by the discrete spectrum of the system. Universal asymptotic behavior thus also guarantees the correct spectrum.
The paper is organized as follows. In Sec. II, we will treat the exactly solvable dissipative harmonic oscillator. We will first summarize the analytic solution of Kehrein and Mielke in a comprised form and then introduce and discuss universal asymptotic behavior. We explicitly show the connection between the conservation law obtained from the exact solution and an asymptotic, scale-invariant spectral function obtained from the universal behavior. In Sec. III, we will consider the spin-boson model for two different truncation schemes. In Sec. IIIA, we investigate flow equations stemming from a form-invariant truncation, i.e. no additional coupling terms are generated. This will allow a direct comparison with the results of the dissipative harmonic oscillator, but will yield rather poor results for dynamical correlation functions. In Sec. IIIB, we will therefore extend the truncation scheme and allow an additional coupling term to be generated. This approach will yield very good results for correlation functions within a large range of the parameter space including sub-Ohmic baths. The reader who is only interested in this non-trivial application, can directly go to Sec. IIIB, which is self-contained.
Dissipative Harmonic Oscillator {#SectionQuantumDissipativeHarmonicOscillator}
===============================
We will first discuss the harmonic oscillator coupled to an environment. Following the seminal work by Caldeira and Leggett,[@Cal83] we will model the bath as a set of non-interacting harmonic oscillators with a dense spectrum. We will also introduce the interaction induced renormalization of the potential so that the Hamiltonian is bounded from below. The dissipative harmonic oscillator shall thus be described by the following Hamiltonian: $$\begin{aligned}
\label{HOscillator}
H&=\frac{p^2}{2}+\kappa v q^2+
\sum_{\alpha}\Big(\frac{p_{\alpha}^2}{2}
+\frac{1}{2}
\omega_{\alpha}^2\big(x_{\alpha}-\frac{\lambda_{\alpha}}{\omega_{\alpha}^2}q\big)^2\Big)\end{aligned}$$ The parameter $\kappa$ can be decomposed as $\kappa=V_0/(mq_0^2)$ where $V_0$ shall denote the potential energy scale, $m$ the mass of the particle, and $q_0$ a length scale such that $v$ is dimensionless.[@FootScales] We identify the frequency of the harmonic oscillator as $\sqrt{2\kappa v/m}$. The operators obey the canonical commutation relations which read ($\hbar=1$) $$\begin{aligned}
\left[q,p\right]=i\quad,\quad\left[x_{\alpha},p_{\alpha^{\prime}}\right]=i
\delta_{\alpha,\alpha'}\quad.\end{aligned}$$
The system is exactly solvable and has been investigated by means of several techniques.[@Wei99] The exact solution of the model via the flow equation approach was first obtained by Kehrein and Mielke.[@Keh97] In Sec. \[AnaHarm\], we will recall the solution in a comprised form in order to introduce notations. In Sec. IIB, we will then discuss flow equations which exhibit universal asymptotic behavior. In Sec. \[NumHarm\], numerical results are presented.
Analytical Results {#AnaHarm}
------------------
### Hamiltonian flow
In order to solve the dissipative harmonic oscillator via flow equations,[@Weg94] the generator $\eta$ of the infinitesimal unitary transformation is chosen to be $$\begin{aligned}
\label{GeneralGenerator_HARM}
\eta&=i\Big(q\sum_{\alpha}\eta_{\alpha}^qp_{\alpha}
+p\sum_{\alpha}\eta_{\alpha}^px_{\alpha}+
\sum_{\alpha,\alpha'}\eta_{\alpha,\alpha'}x_{\alpha}p_{\alpha^{\prime}}
\Big)\\
&\equiv\eta^q+\eta^p+\eta_B\quad.\notag\end{aligned}$$ The first two terms follow from the canonical choice $\eta_c=[H,V]$ with the off-diagonal part $V=-q\sum{_{\alpha}}\lambda{_{\alpha}}x{_{\alpha}}$, but with generalized parameters $\eta{_{\alpha}}^q$ and $\eta{_{\alpha}}^p$. They will be determined later. The last term $\eta^B$ is needed to cancel new interaction terms between different bath modes which are generated by the flow equations. The solvability of the model is due to the fact that this cancellation is possible and that therefore the flow equations close exactly.
The constants of the generator are now chosen such that the Hamiltonian remains form-invariant during the flow, i. e. no new interaction terms are generated. This procedure is not unique; a possible parametrization is given by $$\begin{aligned}
&\eta_{\alpha}^p=\lambda_{\alpha}
f(\omega_{\alpha},\ell)\quad,\quad\eta_{\alpha}^q=-\lambda_{\alpha}
f(\omega_{\alpha},\ell)\quad,\notag\\
&\eta_{\alpha,\alpha'}=-\frac{\lambda_{\alpha}\lambda_{\alpha'}}{
\omega_{\alpha}^2-
\omega_{\alpha'}^2}\big(f(\omega_{\alpha},\ell)+f(\omega_{\alpha'},\ell)\big)\quad,\notag\end{aligned}$$ where $f(\omega{_{\alpha}},\ell)$ denotes an arbitrary function.
The flow of the system is characterized by the renormalized potential $$\begin{aligned}
\label{RenormPot}
\tilde{v}(\ell)\equiv v(\ell)+\kappa^{-1}\sum{_{\alpha}}\frac{\lambda{_{\alpha}}^2(\ell)}{2\omega{_{\alpha}}^2}\quad.\end{aligned}$$ Defining the effective frequency ${\Delta}^2(\ell)\equiv2\kappa\tilde v(\ell)$ and the spectral coupling function $$\begin{aligned}
J(\omega,\ell)=\frac{1}{{\Delta}(\ell)}\sum_{\alpha}\frac{\lambda_{\alpha}^2(\ell)}
{\omega_{\alpha}}\delta(\omega-\omega_{\alpha})\quad,\end{aligned}$$ the flow equations ${\partial_\ell}H=[\eta,H]$ are given by the following coupled integro-differential equations:[@FootZT] $$\begin{aligned}
\label{DiffvHarm}
&\partial_{\ell}{\Delta}(\ell)=\int_0^\infty d\omega
J(\omega,\ell)\omega f(\omega,\ell)\quad,\\
&\partial_{\ell}J(\omega,\ell)=
2J(\omega,\ell)\big(\omega^2-{\Delta}^2(\ell)\big)f(\omega,\ell)-J(\omega,\ell)\frac{\partial_{\ell}{\Delta}(\ell)}{{\Delta}(\ell)}\notag\\\label{DiffJHarm}
&-2{\Delta}(\ell)J(\omega,\ell)\int_0^\infty d\omega'
\frac{J(\omega',\ell)\omega'}{\omega^2-{\omega'}^2}F(\omega,\omega',\ell)\quad,\end{aligned}$$ where we defined $F(\omega,\omega',\ell)\equiv f(\omega,\ell)+f(\omega',\ell)$. Notice that the renormalization of the bath modes $\omega{_{\alpha}}$ was neglected since it vanishes in the thermodynamic limit which is a typical feature of dissipative systems.[@Keh97] Further, there is no mass renormalization in this approach, and we disregarded the renormalization of the ground-state energy.
To solve these equations, Kehrein and Mielke introduced the function $$\begin{aligned}
R(z,\ell)=\sum_{\alpha}\frac{\lambda_{\alpha}^2(\ell)}
{z-\omega_{\alpha}^2}\quad,\end{aligned}$$ for which the following differential equation holds: $$\begin{aligned}
\label{DiffR}
&\partial_{\ell}R(z,\ell)=-\partial_{\ell}{\Delta}^2(\ell)\\\notag
&+
2\big(z-{\Delta}^2(\ell)-R(z,\ell)\big)\sum_{\alpha}\frac{\lambda_{\alpha}^2(\ell)}
{z-\omega_{\alpha}^2}f(\omega_{\alpha},\ell)\end{aligned}$$ The algebraic equation $z-{\Delta}^2(\ell)-R(z,\ell)=0$ solves the above differential equation for any $z$, and apparently also for any $f(\omega{_{\alpha}},\ell)$. With $z^*={\Delta}^2(\ell=\infty)$ one imposes the boundary condition $R(z^*,\ell)\rightarrow0$ for $\ell\rightarrow\infty$ which guarantees that system and bath are decoupled for $\ell\rightarrow\infty$. This yields a self-consistent equation for ${\Delta}^*\equiv{\Delta}(\ell=\infty)$ which can be solved for $\ell=0$. Generally it reads $$\begin{aligned}
\label{FreqSelfCon}
{{\Delta}^*}^2={\Delta}^2(\ell)+{\Delta}(\ell)\int_0^\infty d\omega
\frac{J(\omega,\ell)\omega}{{{\Delta}^*}^2-\omega^2}\quad.\end{aligned}$$
We already want to stress that for the above argumentation ${\Delta}^*$ has to be finite. Otherwise, Eq. (\[FreqSelfCon\]) at $\ell=0$ and Eq. (\[RenormPot\]) cannot hold simultaneously unless $v(\ell=0)=0$. The case $v(\ell=0)=0$ describes the system of a dissipative free particle.
In the next subsection, we will choose $f(\omega{_{\alpha}},\ell)$ such that ${\Delta}^*=0$, and the system will still be decoupled from the bath for $\ell\to\infty$. The resulting flow equations thus belong to a different universality class since Eq. (\[FreqSelfCon\]) does not hold anymore. This is the key observation of this subsection.
### Observable flow
To determine correlation functions we have to apply the same sequence of infinitesimal transformations to the observables that led to the diagonalization of the Hamiltonian. Only then the diagonal structure of the Hamiltonian can be used to yield a simple time evolution of the operators in the Heisenberg picture. For the flow of the position operator we make the ansatz $$\begin{aligned}
q(\ell)=h(\ell)q+\sum_{\alpha}\chi_{\alpha}(\ell)x_{\alpha}\quad.\end{aligned}$$ The initial conditions are given by $h(\ell=0)=1$ and $\chi{_{\alpha}}(\ell=0)=0$. During the flow, the weight of the system operator will be transferred to the bath operators. In the language of the flow equation approach we speak of a dissipative system when the total weight of the system is being transferred to the bath during the flow, i.e. $h(\ell=\infty)=0$. The flow equations for the observable $\partial_{\ell}q=[\eta,q]$ close and with the spectral function $$\begin{aligned}
S(\omega,\ell)\equiv-\sum{_{\alpha}}\frac{\lambda{_{\alpha}}(\ell)\chi{_{\alpha}}(\ell)}{\omega{_{\alpha}}}
\delta(\omega-\omega{_{\alpha}})\end{aligned}$$ we obtain the following integro-differential equations: $$\begin{aligned}
\label{DiffhHarm}
{\partial_\ell}h(\ell)=\int_0^\infty d\omega \omega S(\omega,\ell)f(\omega,\ell)\quad,\end{aligned}$$
$$\begin{aligned}
\notag
{\partial_\ell}S(\omega,\ell)&=-{\Delta}(\ell)h(\ell)J(\omega,\ell)f(\omega,\ell)
-{\Delta}(\ell)J(\omega,\ell)\int_0^\infty d\omega' \frac{S(\omega',\ell)\omega'}
{\omega^2-{\omega'}^2}F(\omega,\omega',\ell)\\
&+(\omega^2-{\Delta}^2(\ell))S(\omega,\ell)f(\omega,\ell)
-{\Delta}(\ell)
S(\omega,\ell)\int_0^\infty d\omega' \frac{J(\omega',\ell)\omega'}
{\omega^2-{\omega'}^2}F(\omega,\omega',\ell)\quad.\label{DiffchiHarm}\end{aligned}$$
The flow equations for $p(\ell)=h(\ell)p+\sum_{\alpha}\chi_{\alpha}(\ell)p_{\alpha}$ are equivalent to Eqs. (\[DiffhHarm\]) and (\[DiffchiHarm\]) which merely demonstrates the fact that Hermite-city is conserved during the flow. From the above flow equations we obtain the following sum rule, which signifies that the commutation relation $[q(\ell),p(\ell)]=i$ holds for all $\ell$: $$\begin{aligned}
h^2(\ell)+\sum_{\alpha}\chi_{\alpha}^2(\ell)=1\end{aligned}$$ To solve the flow equations one introduces the functions $$\begin{aligned}
\begin{split}
S_1(z,\ell)&=\sum_{\alpha}\frac{\lambda_{\alpha}\chi_{\alpha}}
{z-\omega_{\alpha}^2}\quad,\quad\\
S_2(z,\ell)&=\sum_{\alpha}\frac{\chi_{\alpha}^2}
{z-\omega_{\alpha}^2}\quad.
\end{split}\end{aligned}$$ Kehrein and Mielke showed that the following quantity is conserved: $$\begin{aligned}
\label{ConservationHarm}
S_2(z,\ell)+\frac{\big(h(\ell)-S_1(z,\ell)\big)^2}{z-{\Delta}(\ell)-R(z,\ell)}=\text{const.}\end{aligned}$$ One can show that $h(\ell=\infty)=0$ for the initial conditions of interest such that the total weight of the system is indeed transferred to the bath. Correlation functions are obtained through the following identity: $$\begin{aligned}
\notag
K(\omega,\ell)&\equiv\sum_{\alpha}\chi_{\alpha}^2(\ell)\delta(\omega^2-\omega_{\alpha}^2)\\\label{K_Harm}
&=\frac{1}{\pi}\text{Im}S_2(\omega^2-i0,\ell)\end{aligned}$$ For example, the spectral function $\omega K(\omega)\equiv \omega K(\omega,\ell=\infty)$ is proportional to the Fourier transform of $\langle q(t)q\rangle$.
Choosing a Lorentzian spectral function $$\begin{aligned}
\label{Lorentzian_Harm}
J(\omega)\equiv J(\omega,\ell=0)=\frac{4\gamma^2\omega\alpha}{\omega^2+\gamma^2}\quad,\end{aligned}$$ the explicit solution is given by $(\omega_0\equiv{\Delta}(\ell=0))$[@Keh97]
$$\begin{aligned}
\label{K_of_Lorentzian_Harm}
K(\omega)=\frac{2\alpha\gamma^2\omega_0\omega(\gamma^2+\omega^2)}
{(\omega_0^2(\gamma^2+\omega^2)-2\pi\alpha\gamma^3\omega_0-\omega^2(\gamma^2+\omega^2))^2+4\pi^2\alpha^2\omega_0^2\gamma^4\omega^2}\quad. \end{aligned}$$
Universal Asymptotic Behavior {#UniversalAsymptotics_HARM}
-----------------------------
In the previous subsection, the generator of the flow equations, $\eta$, was chosen such that the initial Hamiltonian of Eq. (\[HOscillator\]) remained form-invariant. This condition defines the generator only up to an arbitrary function $f(\omega,\ell)$. It is often useful to explicitly specify $f(\omega,\ell)$ even though the final result must be independent of the particular choice. In Ref. , Kehrein and Mielke choose $f(\omega,\ell)=-(\omega-{\Delta}(\ell))/(\omega+{\Delta}(\ell))$. This leads to $J(\omega,\ell)\propto\exp[-2(\omega-{\Delta}(\ell))^2\ell]$ for $\ell\to\infty$ if one neglects the non-linear term in Eq. (\[DiffJHarm\]). The spectral function is therefore centered around the renormalized frequency ${\Delta}^*\neq0$. At $\omega={\Delta}^*$ the spectral function vanishes algebraically as $\ell^{-1/2}$. The asymptotic spectral function thus depends on the initial frequency through ${\Delta}^*$. We want to label this asymptotic behavior non-universal.
A different choice is $f(\omega,\ell)=-1$ which would have the consequence that the renormalized potential has to tend to zero so that the system is decoupled for $\ell\to\infty$, i.e. ${\Delta}^*=0$ in order that $J(\omega,\ell=\infty)=0$. It therefore belongs to a different universality class since Eq. (\[FreqSelfCon\]) does not hold anymore. The differential equation (\[DiffR\]) turns into a Ricatti equation which can formally be integrated.
That the choice $f(\omega,\ell)=-1$ really decouples the system from the bath is not clear from the beginning. One way to convince oneself is to study the asymptotic behavior of the flow equations (\[DiffvHarm\]) and (\[DiffJHarm\]). For this, we introduce an additional energy scale related to the coupling function, ${\Lambda}(\ell)\equiv\int_0^\infty d\omega J(\omega,\ell)/\omega$.
We now make the ansatz $$\begin{aligned}
\label{PartHarmAsymp}
{\Delta}(\ell)\rightarrow a\ell^{-1/2}\quad,\quad
{\Lambda}(\ell)\rightarrow b\ell^{-1/2}\end{aligned}$$ with constants $a$ and $b$ as $\ell\to\infty$. We further assume that[@FootDrop] $$\begin{aligned}
\label{SpectralHarmAsymp}
J(\omega,\ell)\rightarrow \omega^{s'}
\hat{J}(\ell)\bar{J}(\omega\sqrt\ell)\quad.\end{aligned}$$
The various contributions of $J(\omega,\ell)$ can be interpreted as follows: The energy dependence $\omega^{s'}$ will yield the (correct) algebraic low-energy behavior of the spectral function, $\bar{J}(\omega\sqrt\ell)$ represents the high-energy cutoff function with the $\ell$-dependent cutoff frequency $\omega_c^\ell=\ell^{-1/2}$, and the function $\hat{J}(\ell)$ is needed to assure units of energy.
The differential equations for ${\Lambda}(\ell)$ and $\hat{J}(\ell)$ can be obtained from Eq. (\[DiffJHarm\]) and the special feature of the exactly solvable dissipative harmonic oscillator is that the differential equation for ${\Lambda}(\ell)$ closes, i.e. only ${\Lambda}(\ell)$, ${\Delta}(\ell)$, and ${\partial_\ell}{\Delta}(\ell)$ are involved. One obtains $\hat{J}(\ell)\rightarrow \ell^{(s'-1)/2}$ and $$\begin{aligned}
a^2=1/2+s'/4\quad,\quad b=a/2\quad.\end{aligned}$$
We now define the dimensionless, scale-invariant variable $y\equiv\omega\sqrt{\ell}$ and set $\bar{J}(y)\rightarrow J_0$ for $y\rightarrow0$. We then obtain the following non-local differential equation for $J(y)\equiv y^{s'-1}\bar{J}(y)$: $$\begin{aligned}
\label{nonlocalJHarm}
\partial_y J(y)&=-4yJ(y)\Big(1-2a\int_0^{\infty}dy'\frac{J(y')}{y^2-{y'}^2}\Big)\\
\notag
&+(s'-1)\frac{J(y)}{y}\end{aligned}$$ The boundary condition is given by $a=2\int_0^\infty dyJ(y)$.
The above ansatz for the asymptotic behavior (Eq. (\[PartHarmAsymp\]) and (\[SpectralHarmAsymp\])) guarantees that the system will be decoupled from the bath since the support of the spectral function vanishes as $\ell^{-1/2}$ and $J(y)\rightarrow y^{4+(s'-1)}e^{-2y^2}$ for $y\to\infty$. There is thus a universal fixed point for all initial frequencies. They are all mapped onto the free particle plus bath.
To determine the asymptotic behavior of $h(\ell)$ and $S(\omega,\ell)$, we need the quantity $\Sigma(\ell)\equiv\int_0^\infty d\omega S(\omega,\ell)/\omega$. We make a similar ansatz as in the case of the spectral function, namely $$\begin{aligned}
h(\ell)\rightarrow c\ell^{-1/2-\xi}\quad,\quad
\Sigma(\ell)\rightarrow d\ell^{-1/2-\xi}\end{aligned}$$ with constants $c$ and $d$ as $\ell\to\infty$ and $$\begin{aligned}
S(\omega,\ell)\rightarrow \omega^{s'}
\hat{S}(\ell)\bar{S}(y)\quad.\end{aligned}$$ Notice that we introduced a further parameter $\xi$.
The differential equation for $\Sigma(\ell)$ and $\hat{S}(\ell)$ are obtained from (\[DiffchiHarm\]). One obtains $\hat{S}(\ell)\rightarrow \ell^{-\xi+(s'-1)/2}$ and $$\begin{aligned}
\xi=s'/4\quad,\quad c=2d\quad. \end{aligned}$$ Setting $\bar{S}(y)\rightarrow S_0$ for $y\rightarrow0$, we obtain the following non-local differential equation for $S(y)\equiv y^{s'-1}\bar{S}(y)$: $$\begin{aligned}
\label{nonlocalSHarm}
\partial_y S(y)&=-2yS(y)\Big(1-2a\int_0^{\infty}dy'\frac{J(y')}{y^2-{y'}^2}\Big)\\
&+4ayJ(y)\int_0^{\infty}dy'\frac{S(y')}{y^2-{y'}^2}\notag
+(s'-1)\frac{S(y)}{y}\end{aligned}$$ Eq. (\[nonlocalSHarm\]) is linear in $S(y)$. We can therefore not determine the constant $S_0$ from the asymptotic behavior. But we can extract the low-frequency behavior of the correlation function $K(\omega)=\lim_{\ell\to\infty}S^2(\omega,\ell)/(2{\Delta}(\ell)J(\omega,\ell))$: $$\begin{aligned}
K(\omega\to0)=\omega^{s'}\frac{S_0^2}{2aJ_0}\end{aligned}$$ The numerical results yield $s'=s$, where $s$ denotes the algebraic behavior of the initial spectral function at low energies, i.e. we recover the well-known result.
Finally, we will give an alternative derivation of the conservation law of Eq. (\[ConservationHarm\]) which is valid in the asymptotic regime. For this, we define the asymptotic function $K_a(\omega,\ell)$, following from the conservation law: $$\begin{aligned}
K_a(\omega,\ell)\equiv \frac{1}{\pi}\text{Im}\frac{(h(\ell)-S_1(\omega^2-i0,\ell))^2}{\omega^2-{\Delta}^2(\ell)-R(\omega^2-i0,\ell)}\end{aligned}$$ Taking $\ell^*$ to be in the asymptotic regime, we obtain $K_a(\omega,\ell)\equiv K(\omega)-K(\omega,\ell^*)=\omega^{s'}K_a(y^*)$ with $y^*=\omega\sqrt{\ell^*}$ and $$\begin{aligned}
\label{K_a_Asmptotic}
K_a(y^*)=4\int_{y^*}^{\infty}
dyy^{2-s'}S(y)\int_0^\infty dy'\frac{S(y')}{y^2-{y'}^2}\quad. \end{aligned}$$
One can thus derive the conservation law from the universal asymptotic behavior, given that $\ell^*$ is in the asymptotic regime. This is an important observation since conservation laws as Eq. (\[ConservationHarm\]) can only be derived for solvable models. But by analyzing the asymptotic behavior, we are able to integrate the flow equations up to a finite but asymptotic $\ell^*$ and than complete the spectral function by analyzing the non-local differential equations for the scale-invariant spectral functions. This procedure also works for non-trivial models.
Numerical Results {#NumHarm}
-----------------
We will now analyze the above flow equations numerically. Given the Lorentzian of Eq. (\[Lorentzian\_Harm\]) as initial coupling function, we will calculate $K(\omega)$ within the flow equation approach for two different choices of $f(\omega,\ell)$. To do so, we will employ the conservation law of Eq. (\[ConservationHarm\]), i.e. $K(\omega)=K(\omega,\ell)+K_a(\omega,\ell)$. This allows us to halt the integration routine after a finite $\ell^*$. Nevertheless, the final result must be independent of $\ell^*$ which is the case.
In Fig. \[Spectral\_Harm\], the results are shown for $\alpha=0.1$ and $\gamma/\omega_0=1$. The analytic solution superposes the solution obtained by flow equation approach and, obviously, $K(\omega)=K(\omega,\ell)+K_a(\omega,\ell)$ is independent of the two choices of $f(\omega,\ell)$. Nevertheless, the functions $K(\omega,\ell)$ and $K_a(\omega,\ell)$ are different. For $f(\omega,\ell)=-(\omega-{\Delta}(\ell))/(\omega+{\Delta}(\ell))$, $K_a(\omega,\ell)$ mostly contains the weight around intermediate time scales, whereas for $f(\omega,\ell)=-1$, $K_a(\omega,\ell)$ determines the low-frequency behavior.\
We will now investigate the asymptotic behavior of the flow equations for $f(\omega,\ell)=-1$. The differential equations (\[nonlocalJHarm\]) and (\[nonlocalSHarm\]) can be solved numerically via self-consistent iteration. Comparing these “analytic” results with the numerical results obtained via integrating the flow equations shall serve as a proof that we have found the correct asymptotic behavior.
In the following, the initial spectral coupling function shall be given by $J(\omega)=2\alpha K^{1-s}\omega^s\Theta(\omega_c-\omega)$ where $\alpha$ denotes the coupling strength and $\omega_c$ the high-frequency cutoff. The bath type is labeled by $s$ and $K$ denotes an additional energy scale present for $s\neq1$. One distinguishes the following three characteristic coupling types: super-Ohmic coupling ($s>1$), Ohmic coupling $(s=1)$, and sub-Ohmic coupling ($s<1$). The numerical data shows that $s'=s$.
The results for Ohmic coupling $s=1$ are shown in Fig. \[Asymp\_Harm\], where the asymptotic curves $J(y)$ and $S(y)$ obtained through direct integration and through self-consistent iteration are plotted for different coupling strengths $\alpha$. Indeed, $J(y)$ is universal for all coupling strengths, the functions $S(y)$ only differ by the constant $S_0$. This behavior also holds for $s\neq1$.
In Fig. \[Asymp\_s\], the scale-invariant functions $J(y)$ and $S(y)$ are shown at fixed coupling strength $\alpha=0.4$ for two different parameters $s$. In the super-Ohmic case $s=1.5$, $J(y)$ and $S(y)$ show a suppression whereas in the sub-Ohmic case $s=0.5$, $J(y)$ and $S(y)$ diverge as $y\to0$.
We close this section with a discussion on the asymptotic function $K_a(y)$ of of Eq. (\[K\_a\_Asmptotic\]) which is needed for the alternative conservation law. $K_a(y)$ is proportional to $S_0^2$ and thus depends on the coupling strength $\alpha$. Scaling out this factor, we obtain universal functions only depending on the bath type $(s'=s)$. In Fig. \[Abschluss\_s\], the results are shown for sub-Ohmic (lhs), Ohmic, and super-Ohmic baths (rhs). Notice that only for Ohmic dissipation $s=1$, $\partial_y K(y)\to0$ for $y\to0$.
Spin-Boson Model
================
The most prominent dissipative quantum system is the spin-boson model. It has been extensively studied over the past thirty years by several varying techniques - ranging from the Feynman-Vernon influence-functional formulation,[@Fey65] Liouville operator and projection methods,[@Fic90] to exact mathematical results.[@Bac95] Its popularity stems from the fact that it is the most simple non-trivial dissipative quantum system, consisting only of a two-state system and a bosonic bath that are linearly coupled in order to preserve the reflection symmetry of the system. Nevertheless, applications for the model are found in all fields of physics, starting from quantum optics and solid state physics to nuclear physics and chemistry.[@Wei99; @Leg87] In addition, from a mathematical and theoretical point of view, it attracts a lot of interest since it exhibits a Kosterlitz-Thouless transition for Ohmic coupling at $T=0$ and a crossover from coherent to incoherent tunneling at finite temperature for all coupling types. Moreover, it is related to other prominent models of theoretical physics - most strikingly to the anisotropic Kondo model.
The Kondo model as well as the spin-boson model belong to the so-called strong-coupling problems which make the use of renormalization group techniques almost indispensable. The low-energy behavior of the Kondo model was thus first obtained by Wilson employing his numerical renormalization group.[@Wil75] These ideas were also applied to the spin-boson model.[@Cos96] Flow equations results for the spin-boson model and Kondo model can be found in Refs. .
The Hamiltonian of the symmetric spin-boson model without bias is given by $$\begin{aligned}
\label{Hamiltonian_SpinBoson_Initial}
\begin{split}
H=-\frac{{\Delta}_0}{2}\sigma_x + \sum_{\alpha}
\omega_{\alpha}b_{\alpha}{^{\dagger}}b_{\alpha}
+\sigma_z\sum_{\alpha}\frac{\lambda_{\alpha}^0}{2}
(b_{\alpha}+b_{\alpha}{^{\dagger}}).
\end{split}\end{aligned}$$
The operators $b_{\alpha}^{(\dagger)}$ resemble the bath degrees of freedom and $\sigma_i$ with $i=x,y,z$ denote the Pauli spin matrices. They obey the canonical commutation relations and the spin-$1/2$ algebra, respectively. The coupling constants $\lambda{_{\alpha}}$ only enter via the spectral coupling function $$\begin{aligned}
J(\omega)=\sum_{\alpha}(\lambda_{\alpha}^0)^2
\delta(\omega-\omega_{\alpha}),\end{aligned}$$ which shall be parametrized as $$\begin{aligned}
J(\omega)=2\alpha K^{1-s}\omega^s\Theta(\omega_c-\omega).\end{aligned}$$ The dimensionless parameter $\alpha$ denotes the coupling strength and $\omega_c$ the high-frequency cutoff. The bath type is labeled by $s$ and $K$ denotes an additional energy scale present for $s\neq1$.
The Hamiltonian (\[Hamiltonian\_SpinBoson\_Initial\]) resembles an effective Hamiltonian where the high energy degrees of freedom of the bath were already integrated out - down to the bath cutoff $\omega_c$ by employing the Born-Oppenheimer approximation. The tunnel-matrix element ${\Delta}_0$ thus depends on this arbitrary energy scale and all physical results must be independent of the cutoff $\omega_c$. For Ohmic coupling one obtains $\Delta_0\propto(\omega_c)^\alpha$ and the only combination of $\Delta_0$ and $\omega_c$ that yields the units of an energy and is independent of $\omega_c$ is given by $\Delta_{r}\propto\Delta_0(\Delta_0/\omega_c)^{\alpha/(1-\alpha)}$. Physical observables should only depend on this effective tunnel-matrix element.
In the following, we will discuss the symmetrized equilibrium correlation function $$\begin{aligned}
C(t)=\frac{1}{2}{\langle}\{{\sigma_z}(t),{\sigma_z}\}{\rangle}=\int_0^\infty dt e^{i\omega t}C(\omega)\end{aligned}$$ at low coupling. In this regime, $C(\omega)$ has been evaluated for super-Ohmic baths and for Ohmic baths by Kehrein and Mielke employing flow equations which exhibited non-universal asymptotic behavior.[@Keh97] Here, we will set up and investigate flow equations which exhibit universal asymptotic behavior.
In Sec. \[FormInv\], we first discuss flow equations which preserve the form of the Hamiltonian of Eq. (\[Hamiltonian\_SpinBoson\_Initial\]). A direct comparison with the flow of the dissipative harmonic oscillator is then possible, but this truncation scheme will yield rather poor results for $C(\omega)$. In Sec. \[Extended\], we will therefore extend the truncation scheme and also allow additional coupling terms to be generated. This procedure will yield very good results for equilibrium correlation functions even for sub-Ohmic baths.
This extended scheme includes operators which we want to label “transitionally relevant”. Those operators neither appear in the initial nor in the fixed point Hamiltonian, but are obviously needed as transition link. This extents the classification of operators for the flow equation approach, initiated by Kehrein and Mielke in Ref. . We want to note that this “transitional relevance” can already be seen from a simpler model including only one bosonic mode.[@Sta03]
Form-invariant flow {#FormInv}
-------------------
### Flow equations
Applying flow equations to the spin-boson model we will first be guided by the premise that the flow shall not generate new interaction terms. This cannot be done exactly as in the case of the dissipative harmonic oscillator but one has to neglect normal ordered coupling terms of higher order. Then, one can follow the same procedure as outlined in the previous section. We therefore just cite the resulting equations. For details, we refer to Ref. and Ref. .
The flow equations ${\partial_\ell}H=[\eta,H]$ for the spin-boson model with form-invariant flow are given by the following coupled integro-differential equations ($T=0$): $$\begin{aligned}
{\partial_\ell}{\Delta}(\ell) &={\Delta}(\ell) \int_0^\infty d\omega
J(\omega,\ell)f(\omega,\ell)\\
{\partial_\ell}J(\omega,\ell)&=
2J(\omega,\ell)(\omega^2-{\Delta}^2(\ell))f(\omega,\ell)\\\notag
&-2{\Delta}(\ell) J(\omega,\ell)\int_0^\infty d\omega'
\frac{J(\omega',\ell)\omega'}{\omega^2-{\omega'}^2}F(\omega,\omega',\ell)\end{aligned}$$ We introduced the $\ell$-dependent spectral coupling function $J(\omega,\ell)\equiv\sum{_{\alpha}}\lambda{_{\alpha}}^2(\ell)\delta(\omega-\omega{_{\alpha}})$ and defined $F(\omega,\omega',\ell)=f(\omega,\ell)+f(\omega',\ell)$. Further, we disregarded the renormalization of the ground-state energy.
As in the previous section, the condition of form-invariance does not specify the flow completely and the flow equations still depend on an arbitrary function $f(\omega,\ell)$. But if one considers the spin-boson model with finite bias or starts from a unitarily equivalent representation of the initial Hamiltonian where the reflection symmetry is broken, e.g. $b{_{\alpha}}\to b{_{\alpha}}-\frac{\lambda{_{\alpha}}^0}{2\omega{_{\alpha}}}$, there is no open choice for $f(\omega,\ell)$ anymore and $f(\omega,\ell)=-1$ emerges automatically.[@StaD] Breaking the reflection symmetry of the system thus also breaks the “gauge-invariance”, expressed by the function $f(\omega,\ell)$.
The truncated flow of the observable shall be given by $$\begin{aligned}
\sigma_z(\ell)=h(\ell)\sigma_z+\sigma_x\sum{_{\alpha}}\chi{_{\alpha}}(\ell)(b{_{\alpha}}+b{_{\alpha}}{^{\dagger}})\quad.\end{aligned}$$ The flow equations for the observable $\partial_\ell\sigma_z=[\eta,\sigma_z]$ are approximated to yield $$\begin{aligned}
{\partial_\ell}h(\ell)&={\Delta}(\ell)\int_0^\infty d\omega S(\omega,\ell)f(\omega,\ell)\quad,\end{aligned}$$
$$\begin{aligned}
\label{diffS}
{\partial_\ell}S(\omega,\ell)&=-{\Delta}(\ell) h(\ell)J(\omega,\ell)f(\omega,\ell)
-{\Delta}(\ell)
J(\omega,\ell)\int_0^\infty d\omega' \frac{S(\omega',\ell)\omega'}
{\omega^2-{\omega'}^2}F(\omega,\omega',\ell)\\\notag
&+(\omega^2-{\Delta}^2(\ell))S(\omega,\ell)f(\omega,\ell)
-{\Delta}(\ell)
S(\omega,\ell)\int_0^\infty d\omega' \frac{J(\omega',\ell)\omega'}
{\omega^2-{\omega'}^2}F(\omega,\omega',\ell)\quad,\end{aligned}$$
where we defined $S(\omega,\ell)\equiv\sum{_{\alpha}}\lambda{_{\alpha}}(\ell)\chi{_{\alpha}}(\ell)\delta(\omega-\omega{_{\alpha}})$.
In Ref. , the flow equations are numerically integrated, choosing $f(\omega,\ell)=-(\omega-{\Delta}(\ell))/(\omega+{\Delta}(\ell))$. The asymptotic spectral function is then centered around the renormalized tunnel-matrix element $\Delta_r$ and one is able to map the asymptotic flow equations of the spin-boson model on to the asymptotic flow equations of the dissipative harmonic oscillator. One can then employ the exact solution of the dissipative harmonic oscillator outlined in Sec. IIA. The result is shown in Fig. \[FinalFinalFig\].
### Asymptotic behavior
In analogy to the dissipative harmonic oscillator we now investigate the flow equations with $f(\omega,\ell)=-1$ which will lead to universal asymptotic behavior. These flow equations show apparent deficiencies when calculating correlation functions. E.g., the sum rule $\langle{\sigma_z}^2(\ell)\rangle=1$ is not satisfied for $f(\omega,\ell)=-1$ which can be seen from the differential equation of $\langle{\sigma_z}^2(\ell)\rangle$, $$\begin{aligned}
{\partial_\ell}\langle{\sigma_z}^2(\ell)\rangle=-2{\Delta}(\ell)\int_0^\infty d\omega
d\omega'\frac{S(\omega,\ell)S(\omega',\ell)}{\omega+\omega'}\quad.\end{aligned}$$ Since $h(\ell)\to0$ for $\ell\to\infty$ and $S(\omega,\ell)\geq0$, the spectral function $C(\omega)=\sum{_{\alpha}}\chi{_{\alpha}}^2(\ell=\infty)\delta(\omega-\omega{_{\alpha}})$ cannot satisfy the spectral sum rule $\int_0^\infty d\omega C(\omega)=1$.
Analyzing the asymptotic behavior, we now show that the flow equations do not yield the right asymptotic behavior, either. We limit ourselves to Ohmic dissipation.
Following the discussion of the previous section, we make the ansatz $$\begin{aligned}
{\Delta}(\ell)\rightarrow a\ell^{-1/2}\quad,\quad
{\Lambda}(\ell)\rightarrow b\ell^{-1/2} \end{aligned}$$ with ${\Lambda}(\ell)\equiv\int_0^\infty d\omega J(\omega,\ell)/\omega$ as $\ell\to\infty$. We further assume that $J(\omega,\ell)\rightarrow \omega J(y)$, where $J(y)\to J_0$ for $y\to0$ and $y\equiv\omega\sqrt{\ell}$. We observe that $a=2b$ and thus obtain the following non-local differential equation: $$\begin{aligned}
\label{nonlocalJ}
\partial_y J(y)=-4yJ(y)\Big(1-2a\int_0^{\infty}dy'\frac{J(y')}{y^2-{y'}^2}\Big)\end{aligned}$$ The boundary condition is given by $1=2\int_0^\infty dyyJ(y)$.
The above equation is identical to Eq. (\[nonlocalJHarm\]) with the only difference that the parameter $a=2\int_0^\infty dyJ(y)$ takes on a different value and cannot be determined analytically, now. In the next subsection, we will see that the above equation also describes the asymptotic flow of the coupling function for an improved truncation scheme, but with $a=2$.
To determine the asymptotic behavior of $S(\omega,\ell)$ we make the ansatz $$\begin{aligned}
h(\ell)\rightarrow c\ell^{-1/2-\xi}\quad,\quad
\Sigma(\ell)\rightarrow d\ell^{-1/2-\xi}\end{aligned}$$ with $\Sigma(\ell)\equiv\int_0^\infty d\omega S(\omega,\ell)/\omega$ as $\ell\to\infty$. Further we assume that $S(\omega,\ell)\rightarrow \omega
\hat{S}(\ell)S(y)$, where $S(y)\rightarrow S_0$ for $y\rightarrow0$. This leads to the following non-local differential equation for $S(y)$: $$\begin{aligned}
\label{nonlocalS}
\partial_y S(y)&=-2yS(y)\Big(1-2a\int_0^{\infty}dy'\frac{J(y')}{y^2-{y'}^2}\Big)\\
+4ay&J(y)\int_0^{\infty}dy'\frac{S(y')}{y^2-{y'}^2}\notag
+2\xi\frac{S(y)-J(y)S_0/J_0}{y}\end{aligned}$$ There is one difference to the corresponding differential equation of the dissipative harmonic oscillator since the last term in Eq. (\[nonlocalS\]) is not present in Eq. (\[nonlocalSHarm\]).
To see the deficiency of the above asymptotic behavior, we now observe that $$\begin{aligned}
C(\omega,\ell)=\frac{S^2(\omega,\ell)}{J(\omega,\ell)}\to\omega\ell^{-2\xi}\frac{S_0^2}{J_0}\quad. \end{aligned}$$ We would only obtain the correct asymptotic behavior $C(\omega)\propto\omega$ for $\omega\to0$ if $\xi=0$. As we will see from the numerical results, this is not the case. We thus obtain the wrong asymptotic behavior which is an evidence for the general shortcoming of the form-invariant flow.
We want to close with a remark on non-universal asymptotic behavior. For Ohmic coupling it yields ${\Delta}(\ell)\to {\Delta}^*+\frac{1}{2}\ell^{-1/2}$ for coupling constants $\alpha<1$, with ${\Delta}^*$ denoting the renormalized tunnel-matrix element ${\Delta}^*\propto\Delta_0(\Delta_0/\omega_c)^{\alpha/(1-\alpha)}$, see Ref. . For $\alpha>1$ the asymptotic behavior is governed by $\Delta\to\ell^{-1/2}/\ln\ell$. The localization phenomena can thus be detected by the different asymptotic behavior of ${\Delta}(\ell)$. This is in contrast to the universal asymptotic behavior where for all coupling constants ${\Delta}(\ell)\to a\ell^{-1/2}$ holds. This indicates that our approach will only be valid for small coupling strength, away from the broken symmetry phase.
### Numerical results {#numerical-results}
The predictions of the above ansatz can be verified numerically by integrating the flow equations. In contrast to the exactly solvable dissipative harmonic oscillator, the constants $a$ and $\xi$ cannot be determined analytically. The self-consistent solution of the differential equations (\[nonlocalJ\]) and (\[nonlocalS\]) yields the following values: $$\begin{aligned}
a\approx1.21\quad,\quad\xi\approx 0.41\end{aligned}$$
In Fig. \[Spektralplus\_SPINBOSON\], the universal asymptotic functions $J(y)$ and $S(y)$ obtained from numerical integration of the flow equations are compared with the “analytic” solution following from the self-consistent evaluation of Eqs. (\[nonlocalJ\]) and (\[nonlocalS\]). The asymptotic results do only depend on the bath type (not shown here), but not on the coupling strength $\alpha$, the explicit form of the cutoff function, or on the value of $\omega_c$.
The symmetrized equilibrium correlation function $C(\omega)$ must of course depend on the initial conditions. Since the asymptotic functions $S(y)$ only differ by at most a constant for arbitrary initial conditions, this information must be contained in the primary and intermediate $\ell$-range of the flow. In Fig. \[SpectralUnivSpBo\], the flow equation results of $C(\omega)$ for a Ohmic bath with sharp cutoff at $\omega_c/{\Delta}_0=30$ at different coupling strengths $\alpha$ are shown.
The spectral function $C(\omega)$ is obtained by integrating the flow equations up to ${\Delta}_0^2\ell^*=10^3$ where one is already in the asymptotic regime for the coupling strengths considered. For higher coupling strengths one would have to integrate up to larger $\ell^*$, since ${\Delta}_r$ is smaller. Notice that we do not employ a conservation law to build up the spectral functions, nor is our approach limited to small coupling strengths. But one crucial shortcoming of the flow equations is that $C(\omega)$ does not fulfill the sum rule $\int_0^\infty d\omega C(\omega)=1$. Independent from the cutoff parameter, $\omega_c$ and bath type the sum rule only yields approximately $80\%$ for small coupling constants ($\alpha\lapp0.1$) and $83-85\%$ for larger coupling constants.
Nevertheless, $C(\omega)$ contains physical information. This is demonstrated in the inset of Fig. \[SpectralUnivSpBo\] where the maximum of $C(\omega)$ is plotted for different cutoff frequencies $\omega_c$ as a function of the coupling strength $\alpha$, according to the effective tunnel-matrix element $$\begin{aligned}
\Delta_{\text{eff}}\equiv\big(\cos(\pi\alpha)\Gamma(1-2\alpha)\big)^{-2(1-\alpha)}\Delta_0\Big(\frac{\Delta_0}{\omega_c}\Big)^{\frac{\alpha}{1-\alpha}},\end{aligned}$$ derived from the non-interacting blib approximation (NIBA).[@Leg87] The universal regime described by the NIBA is only reached for small coupling constants and for rather large $\omega_c$. Still, the flow equations capture the essential physics despite of the apparent shortcomings of the correlation function.
Extended flow equations {#Extended}
-----------------------
### Flow equations
Since the flow equations of the form-invariant flow did not yield satisfactory results, we will now extend the truncation scheme, i.e. we will choose a different generator and we will allow a new coupling term to be generated. But universal asymptotic behavior shall be retained.
The anti-Hermitian generator $\eta$ shall first be chosen canonically, i.e. $\eta_c=[H_B,H]$, with the diagonal Hamiltonian $H_B=\sum{_{\alpha}}\omega_{\alpha}b_{\alpha}{^{\dagger}}b{_{\alpha}}$ and $H$ given in Eq. (\[Hamiltonian\_SpinBoson\_Initial\]). This choice indicates that there will be universal asymptotic behavior since the diagonal Hamiltonian consists of the bath only and is thus $\ell$-independent.[@FootFS]
The commutator $[\eta_c,H]$ will now generate a new coupling term which is linear in the bosonic operators. Neglecting the shift in the ground-state energy, we thus arrive at the following truncated Hamiltonian: $$\begin{aligned}
H(\ell)&=-\frac{{\Delta}(\ell)}{2}\sigma_x+\sum{_{\alpha}}\omega{_{\alpha}}b{_{\alpha}}{^{\dagger}}b{_{\alpha}}\\\notag
&+i\sigma_y\sum{_{\alpha}}\frac{\lambda{_{\alpha}}^y(\ell)}{2}(b{_{\alpha}}-b{_{\alpha}}{^{\dagger}})
+\sigma_z\sum{_{\alpha}}\frac{\lambda{_{\alpha}}^z(\ell)}{2}(b{_{\alpha}}+b{_{\alpha}}{^{\dagger}})\end{aligned}$$ with ${\Delta}(\ell=0)={\Delta}_0$, $\lambda{_{\alpha}}^z(\ell=0)=\lambda{_{\alpha}}^0$, and $\lambda{_{\alpha}}^y(\ell=0)=0$.
The canonical generator $\eta_c$ also gives rise to coupling terms which are bilinear in the bosonic operators. These terms will not be included in the Hamiltonian flow explicitly. Still, they are adequately taken into account of by introducing an additional term to the generator which approximately cancels these bilinear contributions. This is in analogy to the procedure of the previous section. For details, see Ref. .
In terms of the spectral coupling functions $J^k(\omega,\ell)\equiv\sum{_{\alpha}}(\lambda{_{\alpha}}^k)^2(\ell)\delta(\omega-\omega{_{\alpha}})$, with $k=y,z$, we obtain the following flow equations: $$\begin{aligned}
{\partial_\ell}{\Delta}&=-2\int_0^\infty d\omega \omega J^m(\omega,\ell)\\
\label{FlowEquations_Extended_SpinBoson}
{\partial_\ell}J^k(\omega,\ell)&=-2\omega^2J^{\bar k}(\omega,\ell)+2{\Delta}\omega J^m(\omega,\ell)\\\notag
&+2J^k(\omega,\ell)\int_0^\infty d\omega'J^m(\omega',\ell)
\frac{\omega^2+{\omega'}^2}{\omega^2-{\omega'}^2}\\\notag
&+4J^m(\omega,\ell)\int_0^\infty d\omega'J^k(\omega',\ell)
\frac{\omega\omega'}{\omega^2-{\omega'}^2}\end{aligned}$$ We defined $\bar k=(z,y)$ for $k=(y,z)$ and $J^m(\omega,\ell)\equiv\sum{_{\alpha}}\lambda{_{\alpha}}^y\lambda{_{\alpha}}^z\delta(\omega-\omega{_{\alpha}})=(J^y(\omega,\ell)J^z(\omega,\ell))^{1/2}$.
The truncated flow of the observable shall be given by ${\sigma_z}(\ell)=h(\ell){\sigma_z}+{\sigma_x}\sum{_{\alpha}}\chi{_{\alpha}}(\ell)(b{_{\alpha}}+b{_{\alpha}}{^{\dagger}})$. The flow equations $\partial_\ell{\sigma_z}=[\eta,{\sigma_z}]$ are subjected to the same decoupling procedure as above. Defining the spectral functions $S^k(\omega,\ell)\equiv\sum{_{\alpha}}\lambda{_{\alpha}}^k(\ell)\chi{_{\alpha}}(\ell)\delta(\omega-\omega{_{\alpha}})$, we obtain the following flow equations: $$\begin{aligned}
\label{FlowEquations_Extended_SpinBoson_Observable}
{\partial_\ell}h(\ell) &=-\int_0^\infty d\omega \omega S^y(\omega,\ell)\quad,\end{aligned}$$
$$\begin{aligned}
\begin{split}
{\partial_\ell}S^y(\omega,\ell)&=h(\ell)\omega J^y(\omega,\ell)-\omega^2S^y(\omega,\ell)+{\Delta}(\ell)\omega S^z(\omega,\ell)\\
&+J^m(\omega,\ell)\int_0^\infty d\omega'S^y(\omega',\ell)
\frac{\omega^2+{\omega'}^2}{\omega^2-{\omega'}^2}
+2J^y(\omega,\ell)\int_0^\infty d\omega'S^z(\omega',\ell)
\frac{\omega\omega'}{\omega^2-{\omega'}^2}\\
&+S^z(\omega,\ell)\int_0^\infty d\omega'J^y(\omega',\ell)
\frac{\omega^2+{\omega'}^2}{\omega^2-{\omega'}^2}
+2S^y(\omega,\ell)\int_0^\infty d\omega'J^m(\omega',\ell)
\frac{\omega\omega'}{\omega^2-{\omega'}^2}
\end{split}\end{aligned}$$
Since $S^z(\omega,\ell)=S^y(\omega,\ell)J^z(\omega,\ell)/J^m(\omega,\ell)$, it suffices to define the integro-differential equation for $S^y(\omega,\ell)$.
One criterion for the assessment of the quality of the flow equations is again given by $\langle {\sigma_z}^2(\ell)\rangle$ which should be approximately equal to one for all $\ell$. It is governed by the differential equation $$\begin{aligned}
\partial_\ell\langle {\sigma_z}^2(\ell)\rangle=-\int_0^\infty d\omega d\omega'\frac{\omega-{\omega'}}{\omega+{\omega'}}S^y(\omega,\ell)S^z(\omega',\ell).\end{aligned}$$ In the asymptotic regime $\ell\to\infty$, the flow equations simplify considerably. i.e. $S^y(\omega,\ell)=S^z(\omega,\ell)$ which will be shown below. This property guarantees that the spectral weight of correlation functions is asymptotically conserved - in contrary to the form-invariant flow. Numerical calculations show that - throughout the flow - the spectral sum rule is fulfilled within less than $0.1\%$ relative error. This is a significant improvement w.r.t. the form-invariant flow.
### Asymptotic behavior
We now discuss the asymptotic behavior. As a first approximation, we take $\lambda{_{\alpha}}^y=\lambda{_{\alpha}}^z$ and thus drop the index $k$ of the spectral function. This assumption is consistent with the flow equations and justified later based on the numerical result. The asymptotic flow equations are then solved by the ansatz $$\begin{aligned}
\notag
{\Delta}(\ell)\to a\ell^{-1/2}\quad,\quad
J(\omega,\ell)\to\ell^{-1/2}J(\omega\sqrt{\ell})\end{aligned}$$ which leads to the following non-local differential equation for $J(y)$ with the scale-invariant variable $y\equiv\omega\sqrt{\ell}$: $$\begin{aligned}
\partial_y J(y)&=-4J(y)\Big((y-a)-2\int_0^\infty dy'\frac{J(y')}{y-y'}-\frac{t}{4y}\Big),\label{AsympJEx}\end{aligned}$$ where $t\equiv1-4\int_0^\infty dyJ(y)$. The spectral function depends on the parameter $a$ and the boundary condition is given by $a=4\int_0^\infty dyyJ(y)$.
With the similar ansatz for the observable flow, i.e. $$\begin{aligned}
\notag
h(\ell)=b\ell^{-1/4-s'/4}\;,\; S(\omega,\ell)\to\ell^{-1/4-s'/4}S(\omega\sqrt{\ell})\;,\end{aligned}$$ we obtain the following non-local differential equation for $S(y)$ with $y=\omega\sqrt{\ell}$: $$\begin{aligned}
\notag
\partial_y S(y)&=-2S(y)\Big((y-a)-2\int_0^\infty dy'\frac{J(y')}{y-y'}-\frac{t'}{2y}\Big)\notag\\\label{AsympSEx}
&+2J(y)\Big(b+2\int_0^\infty dy'\frac{S(y')}{y-y'}+\frac{t''}{y}\Big)\quad,\end{aligned}$$ where $t'\equiv(s'+t)/2$ and $t''\equiv\int_0^\infty dyS(y)$. The parameter $b$ is given by $b=-4\int_0^\infty dyyS(y)/(s'+1)$.
The above ansatz assumes that $a={\Delta}(\ell)\ell^{1/2}$ is constant. This is not the case, though - for the entire regime that is accessible by our numerical integration. We thus extend the asymptotic behavior of the tunnel-matrix element ${\Delta}(\ell)$ by setting ${\Delta}(\ell)=a(\ell)\ell^{-1/2}$ with $a(\ell)\to a$. The generalization induces a shift in the spectral function, i.e. $J(\omega,\ell)\to \ell^{-1/2}J(\omega\sqrt{\ell}+\hat{a}(\ell))$ with $[a-a(\ell)]/\hat a(\ell)=$ const.
The same extensions are also necessary in the case of the asymptotic observable flow, i.e. $h(\ell)\to b(\ell)\ell^{-1/4-s'/4}$ with $b(\ell)\to b$ and $S(\omega,\ell)\to\ell^{-1/4-s'/4}S(\omega\sqrt\ell+\hat a(\ell))$ with $[b-b(\ell)]/\hat a(\ell)=$ const. All limits hold for $\ell\to\infty$. A numerical fit for Ohmic dissipation with $\alpha=0.1$ and $\omega_c/{\Delta}_0=10$ - based on the resulting differential equation for $a(\ell)$ and $b(\ell)$ and on the fact that $[a-a(\ell)]/[b-b(\ell)]=$ const - suggests that $a=4.5$ and $b=2.0$.
But in the following, we want to demonstrate that the $\ell$-dependent shift in the spectral function does not essentially alter the functional dependence of $J(y)$ and $S(y)$. For this, we define $\widetilde J(y)\equiv J(y+a)$. Substituting $y\to y+a$ in Eq. (\[AsympJEx\]) and setting $a\to\infty$ on the lower bound of the integral, the boundary condition yields $t=0$ and we obtain $$\begin{aligned}
\label{nonlocalJSB}
\partial_y \widetilde J(y)&=-4y\widetilde J(y)\Big(1-4\int_0^{\infty}dy'\frac{\widetilde J(y')}{y^2-{y'}^2}\Big).\end{aligned}$$ This is the same non-local differential equation as in the case of the dissipative harmonic oscillator and the spin-boson model with form-invariant flow. But this time the parameter $a$ of Eq. (\[nonlocalJHarm\]) is given by $a=2$. The “analytic” result of the above equation is in good agreement with the flow equation result after a shift by the corresponding $a(\ell)$. This is demonstrated in Fig. \[AsympJExOhmic\], where on the lhs the flow equations were integrated numerically and on the rhs Eq. (\[nonlocalJSB\]) was used. The initial spectral function was taken to be $J(\omega)=2\alpha\omega\Theta(\omega_c-\omega)$ with $\alpha=0.1$ and $\omega_c/{\Delta}_0=10$ and the various values of $a(\ell)$ were obtained from the numerical integration of the flow equations.
The same transformation can be done for the asymptotic function of the observable. With $\widetilde S(y)\equiv S(y+a)$ this yields the following equation: $$\begin{aligned}
\notag
\partial_y \widetilde S(y)&=-2\widetilde S(y)\Big(y-2\int_{-\infty}^\infty dy'\frac{\widetilde J(y')}{y-y'}+\frac{s'}{4(y+a)}\Big)\\
&+2\widetilde J(y)\Big(b+2\int_{-\infty}^\infty dy'\frac{\widetilde S(y')}{y-y'}+\frac{t''}{y+a}\Big)\quad\label{Asymp_wwSEx}\end{aligned}$$ With $4t''=-s'b/a$ which follows from the boundary condition, the “analytic” solutions are again in good agreement with the flow equation solutions after a shift by the corresponding $a(\ell)$. This is demonstrated in Fig. \[AsympSExOhmic\], where on the lhs the flow equations were integrated numerically and on the rhs Eq. (\[Asymp\_wwSEx\]) was used. The various values of $b(\ell)$ were obtained from the numerical integration of the flow equations.
Keeping in mind that $J(y)$ is exponentially small for $y\to0$, we are allowed to neglect the inhomogeneous contribution in Eq. (\[AsympSEx\]) or (\[Asymp\_wwSEx\]) in this regime. The asymptotic behavior of $S(y)$ is then given by $S^2(y)\propto J(y)y^{s'}$ for $y\to0$. With $J(y)\to J_0$ and $S(y)\to S_0$ for $y\to0$, we thus obtain the $\ell$-independent correlation function for $\omega\to0$ $$\begin{aligned}
C(\omega\to0)=\omega^{s'}\frac{S_0^2}{J_0}\quad.\end{aligned}$$ For an initial spectral function $J(\omega)\propto \omega^s$, $s=s'$ is verified by the numerical data and we thus obtain the correct low-energy behavior $C(\omega)\propto\omega^s$.
We want to note that the equations for the asymptotic behavior hold independent of the initial coupling strength and of the bath type. The initial conditions only enter via the parameters $a$ and $b$ for which the corresponding asymptotic functions are obtained. But we also want to mention that for higher coupling strength $\alpha$ and for lower values of $s$ the asymptotic regime $\lambda{_{\alpha}}^y=\lambda{_{\alpha}}^z$ is only reached for rather large $\ell$. For an initial spectral function with $\alpha=0.1$, $\omega_c/{\Delta}_0=10$, $K/{\Delta}_0=1$ and $s=0.3$ we have e.g. ${\Delta}_0^2\ell^*\gapp10^{20}$, $\ell^*$ denoting the asymptotic regime were $\lambda{_{\alpha}}^y=\lambda{_{\alpha}}^z$ holds on all energy scales. The corresponding parameters at $\ell^*$ are given by $a=2.23$ and $b=1.43$.
Nevertheless, one can extend the range of the asymptotic regime if one drops the assumption $\lambda{_{\alpha}}^y=\lambda{_{\alpha}}^z$, i.e. one makes the ansatz $J^k(\omega,\ell)\to\ell^{-1/2}J^k(\omega\sqrt{\ell})$ and $S^k(\omega,\ell)\to\ell^{-1/4-s'/4}S^k(\omega\sqrt{\ell})$ for $k=y,z$ . The resulting set of non-local differential equations for the functions $J^k(y)$ recovers the flow equation results also for $\tilde\ell^*\lapp\ell^*$, if the second boundary condition is suitably chosen. In fact, $\tilde\ell^*$ can be chosen such that $\tilde\ell^*\ll\ell^*$.
To conclude the discussion, we now derive the asymptotic conservation law, in analogy to Eq. (\[K\_a\_Asmptotic\]). For this, we define the $\ell$-dependent correlation function $C(\omega,\ell)\equiv\sum_\alpha\chi{_{\alpha}}^2(\ell)\delta(\omega-\omega{_{\alpha}})$ and the asymptotic spectral function $C_a(\omega,\ell)\equiv C(\omega)-C(\omega,\ell)$. For $\tilde\ell^*$ being in the quasi-asymptotic regime, we obtain from the above flow equations $C_a(\omega,\tilde\ell^*)=\omega^{s'}I( \tilde y^*)$ where $I( \tilde y^*)$ is given by
$$\begin{aligned}
\label{FinalAsymptotics}
I( \tilde y^*)\equiv4\int_{ \tilde y^*}^\infty dy y^{-s'}\Big(bS^y(y)+S^z(y)\int_0^\infty dy'S^y(y')G^+(y,y')+S^y(y)\int_0^\infty dy'S^z(y')G^-(y,y')+t^{''}\frac{S^z(y)}{y}\Big).\end{aligned}$$
We defined $G^\pm(y,y')\equiv(y-y')^{-1}\pm(y+y')^{-1}$ and $ \tilde y^*=\omega\sqrt{\tilde\ell^*}$. The above expression simplifies considerably for $\ell^*$ being in the asymptotic regime.[@Sta02]
### Numerical results {#numerical-results-1}
We will now present the numerical results for the equilibrium correlation function $C(\omega)$.
On the left hand side of Fig. \[FinalFinalFig\], the spectral function $C(\omega)/\omega^s$ is shown for various bath types $s$. The initial coupling function was chosen to be $J(\omega)=2\alpha K^{1-s}\omega^s\Theta(\omega_c-\omega)$ with $\alpha=0.1$, $K/{\Delta}_0=1$, and $\omega_c/{\Delta}_0=10$. As can be seen, $C(\omega)/\omega^s$ is constant for low energies. But notice that $\partial_\omega C(\omega)/\omega^s\neq0$ for $\omega\to0$. This point was already mentioned in the discussion of the dissipative harmonic oscillator (see Fig. \[Abschluss\_s\]).
Apart from the sum rule $\langle \sigma_z(\ell)\rangle\approx1$ for all $\ell$, another criterion for the assessment of the quality of the flow equations is given by the generalized Shiba relation:[@Sas90] $$\begin{aligned}
\label{Shiba}
\lim_{\omega\to0}\frac{C(\omega)}{\omega^s}=2\alpha K^{1-s}\Big(\int_0^\infty d\omega\frac{C(\omega)}{\omega}\Big)^2\end{aligned}$$ The results are given in Table \[TableShiba\]. So far, to our knowledge, the generalized Shiba relation for sub-Ohmic dissipation can only be checked by the above procedure.
For Ohmic dissipation, the outlined approach can be contrasted with previous results. As initial spectral function we use $J(\omega)=2\alpha\omega\Theta(\omega_c-\omega)$ with $\alpha=0.1$ and $\omega_c/{\Delta}_0=10$. On the right hand side of Fig. \[FinalFinalFig\], the spectral function $C(\omega)/\omega$ is shown in comparison to the result obtained by integrating the flow equations of Sec. IIIA with $f(\omega,\ell)=-(\omega-{\Delta}(\ell))/(\omega+{\Delta}(\ell))$. It is obtained by integrating the flow equations up to ${\Delta}_0^2\ell^*=100$ and then using the conservation law of Eq. (\[ConservationHarm\]). For higher values of ${\Delta}_0^2\ell^*\gapp100$, the spectral function shows instabilities for energies close to the resonance $\omega\approx{\Delta}^*$. The situation becomes worse for higher values of $\alpha$ and generally for sub-Ohmic baths. For a detailed discussion, see Ref. .
In Table \[ContrastShiba\], the two flow equation approaches are contrasted with the help of the generalized Shiba relation for various weak Ohmic couplings. Throughout the parameter regime, flow equations with universal asymptotic behavior yield better results. For $\alpha=0.2$, the non-universal approach yields no stable result in order to check the generalized Shiba relation.
On the right hand side of Fig. \[FinalFinalFig\], $C(\omega)/\omega$ is also shown as obtained from the non-interacting blib approximation (NIBA) for suitable effective tunneling $\Delta_{\text{eff}}$.[@Leg87] The curve coincides with the one obtained from the universal approach for short and intermediate time scales. For long time scales, the graph diverges as $C(\omega)/\omega\propto\omega^{-2\alpha}$, which is incorrect.
The universal flow equation approach is applicable for not too small bath types $s\geq s^*$ and not too large coupling strengths $\alpha\leq\alpha^*$. As criterion we again use the generalized Shiba relation and define the ratio $R\equiv[\lim_{\omega\to0}C(\omega)/\omega^s]/[2\alpha K^{1-s}(\int_0^\infty d\omega C(\omega)/\omega)^2]$. We demand that $|1-R|\lapp0.3$. For Ohmic coupling and $\omega_c/{\Delta}_0=10$, e.g., we have $\alpha^*\lapp0.4$ and for a fixed coupling strength, $\alpha=0.1$ say, we have $s^*\gapp0.3$ ($K/{\Delta}_0=1$). These parameter regimes are further discussed in Ref. where also the case of finite bias is treated.
If one is only interested in the effective tunnel-matrix element or the half-width of the resonance, the flow equations with universal asymptotic behavior can yield the correct result even if the Shiba relation is not fulfilled satisfactorily.
$\lim_{\omega\to0}C(\omega)/\omega^s$ $2\alpha K^{1-s}(\int_0^\infty d\omega C(\omega)/\omega)^2$ Ratio
------- --------------------------------------- ------------------------------------------------------------- -------
s=1.0 0.364 0.391 0.93
s=0.8 0.356 0.390 0.91
s=0.5 0.384 0.432 0.89
: The generalized Shiba relation for various bath types $s$ with $\alpha=0.1$, $K/{\Delta}_0=1$, and $\omega_c/{\Delta}_0=10$.[]{data-label="TableShiba"}
R $\alpha=0.01$ $\alpha=0.05$ $\alpha=0.1$ $\alpha=0.2$
----------- --------------- --------------- -------------- --------------
Universal 1.00 0.97 0.93 0.86
Non-Uni. 1.03 1.19 1.38 -
: Contrasting the universal and non-universal flow equations with the help of the generalized Shiba relation, $R\equiv[\lim_{\omega\to0}C(\omega)/\omega]/[2\alpha (\int_0^\infty d\omega C(\omega)/\omega)^2]$, for a Ohmic bath with $\omega_c/{\Delta}_0=10$ and various coupling strengths $\alpha$.[]{data-label="ContrastShiba"}
Conclusions
===========
In this work, we applied the flow equation method to two dissipative systems, the exactly solvable dissipative harmonic oscillator and the spin-boson model. Investigating the former model, we have pointed that there are two basic universality classes for flow equations, i.e. either $\omega_{\infty}>0$ and Eq. (\[FreqSelfCon\]) holds or $\omega_{\infty}=0$. The former class was labeled non-universal since the asymptotic behavior depended on the initial conditions of the system. The latter class was labeled universal since the fixed point Hamiltonian was given by the non-interacting bath, only. The universal asymptotic behavior of the spectral coupling function as well as of the observable are characterized by scale-invariant, non-local differential equations from which the low-energy behavior of correlation functions can be deduced analytically.
These ideas were also applied to the non-trivial spin-boson model. It was mentioned that the form-invariant flow of the Hamiltonian necessarily displays universal asymptotic behavior when the reflection symmetry is broken. The resulting flow equations, though, did not yield the correct normalization condition nor the correct low-energy behavior of correlation functions. Nevertheless, an effective energy scale was recovered which was in good agreement with the predictions of the NIBA.
Due to the obvious shortcomings of the form-invariant flow, we set up extended flow equations which also displayed universal asymptotic behavior. The resulting flow equations yielded very good results for a wide range of the parameter regime where also sub-Ohmic baths could be included in the treatment. We further deduced the correct low-energy behavior of correlation functions from the universal asymptotic behavior and showed that the normalization condition and the Shiba relations were satisfied within numerical errors for a certain parameter regime.
Universal asymptotic behavior is crucial in order to determine correlation functions within the flow equation approach since only then one can assure that the dominant weight of the correlation functions is accounted for by the stable intermediate flow. Apart from this technical preference, the trivial fixed point Hamiltonian also accounts for the correct spectrum which is not superposed by the spectrum of the isolated system. We also want to mention that the asymptotic flow of the spectral function is described by the same non-local differential equation - independent of the model or the truncation scheme, but only differing by one parameter.
The presented procedure can easily be extended to other dissipative models by starting from the canonical generator $\eta_c=[H_B,H]$ where $H_B$ denotes the non-interacting bath.\
It is a pleasure to thank A. Mielke for innumerable discussions on the flow equation method. This work was funded by the Deutsche Forschungsgesellschaft.
[99]{} S. D. G[ł]{}azek and K.G. Wilson, Phys. Rev. D **48**, 5863 (1993); Phys. Rev. D **49**, 4214 (1994). F. Wegner, Ann. Phys. (Leipzig) **3**, 77 (1994). F. Wegner, Physics Reports [**348**]{}, 77 (2001). S.K. Kehrein, A. Mielke, and P. Neu, Z. Phys. B **99**, 269 (1996). P. Lenz and F. Wegner, Nucl. Phys. B **482** \[FS\], 693 (1996). The flow parameter $\ell$ has dimensions of inverse energy squared if one chooses the generator as $\eta=[H_0,H]$, with $H$ denoting the full and $H_0$ the diagonal Hamiltonian. S.K. Kehrein and A. Mielke, Ann. Phys. (Leipzig) **6**, 90 (1997). The mapping is only successful for super-Ohmic baths and for Ohmic baths at low coupling. We explicitly exclude the possibility of a quantum phase transition. Our approach will therefore only be valid for weak coupling. V. Bach, J. Fröhlich, and I.M. Sigal, Adv. in Math. **137**, 299 (1998). A.O. Caldeira and A.J. Leggett, Ann. Phys. (N.Y.) **149**, 374 (1983); erratum. Ann. Phys. (N.Y.) **153**, 445 (1984). We have chosen this parametrization to emphasize that the approach can easily be extended to more general, e.g. periodic, potentials, see Ref. . U. Weiss, [*Quantum Dissipative Systems*]{}. (Second Edition, Singapore World Scientific, 1999). Throughout this work, we consider the zero temperature limit, $T=0$. Notice that we dropped the index $s'$ on the functions $J(\omega,\ell)$, $\hat{J}(\ell)$, and $\bar{J}(y)$. R.P. Feynman and A.R. Hibbs, [*Quantum Mechanics and Path Integrals*]{} (Mc Graw-Hill, New York, 1996). E. Fick and G. Sauermann, [*The Quantum Statistics of Dynamic Processes*]{} (Springer Series in Solid-State Sciences, Vol. 88, Berlin 1990). H. Spohn, Commun. Math. Phys. **123**, 277 (1989); V. Bach, J. Fröhlich, and I.M. Sigal, Lett. Math. Phys. **34**, 183 (1995). A.J. Leggett, S. Chakravarty, A.T. Dorsey, M.P.A. Fisher, A. Grag, and W. Zwerger, Rev. Mod. Phys. **59**, 1 (1987). K.G. Wilson, Rev. Mod. Phys. **47**, 773 (1975). T.A. Costi and C. Kieffer, Phys. Rev. Lett. **76**, 1683 (1996). W. Hofstetter and S. Kehrein, Phys. Rev. B **63**, 140402(R) (2001). C. Slezak, S. Kehrein, Th. Pruschke, and M. Jarrell, Phys. Rev. B **67**, 184408 (2003). T. Stauber and A. Mielke, Phys. Lett. A **305**, 275 (2002). S.K. Kehrein and A. Mielke, J. Phys. A: Math. Gen. **27**, 4259 (1994). T. Stauber and A. Mielke, J. Phys. A: Math. Gen. **36**, 2707 (2003). T. Stauber, PhD thesis, University of Heidelberg (2002). Again, we neglect contributions of order $1/N$, $N$ being the number of modes. M. Sassetti and U. Weiss, Phys. Rev. Lett. **65**, 2262 (1990).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Consider a solution $u$ to $\Delta u +Vu=0$ on $\mathbb{R}^2$, where $V$ is real-valued, measurable and $|V|\leq 1$. If $|u(x)| \leq \exp(-C |x| \log^{1/2}|x|)$, $|x|>2$, where $C$ is a sufficiently large absolute constant, then $u\equiv 0$.'
address:
- 'Alexander Logunov: Department of Mathematics, Princeton University, Princeton, NJ, USA'
- 'Eugenia Malinnikova: Department of Mathematics, Stanford University, Stanford, CA, USA'
- 'Nicolai Nadirashvili: CNRS, Institut de Math'' ematique de Marseille, Marseille, France '
- 'Fedor Nazarov: Department of Mathematics, Kent State University, Kent, OH, USA'
author:
- 'A. Logunov'
- 'E. Malinnikova'
- 'N. Nadirashvili'
- 'F. Nazarov'
title: The Landis conjecture on exponential decay
---
The main result.
================
Let $u$ be a solution to $$\label{eq:schr}
\Delta u + V u =0$$ in $\mathbb{R}^n$, where $V$ is a measurable function with $|V | \leq 1 $ in the whole space. According to [@C15],[@KL88], in the late 1960s Landis conjectured that if $$|u(x)| \leq \exp(-C|x|),$$ where $C>0$ is a sufficiently large constant, then $u\equiv0$. The weaker statement, which was also conjectured by Landis according to [@C15], states that if $|u(x)|$ tends to $0$ faster than exponentially at $\infty$, i.e., $$|u(x)| \leq \exp(-|x|^{1+\varepsilon}),\varepsilon >0,$$ then $u\equiv 0$.
There are two versions of Landis’ conjectures: real and complex. Meshkov [@M92] constructed a counter-example to the complex version of Landis’ conjecture. He showed that there is a complex-valued potential $V$ with $|V|\leq 1$ and a non-zero solution $u$ to on $\mathbb{R}^2$ such that $|u(x)| \leq \exp(-c |x|^{4/3}).$ Meshkov also showed (in any dimension $n$) that if $$\sup_{\mathbb{R}^n} |u(x)| e^{-\tau |x|^{ 4/3}}< \infty \text{ for all } \tau>0,$$ then $u\equiv 0$. The question whether the Landis conjecture is true for real-valued $V$ is open. The main result of this article confirms the weak version of the Landis conjecture in dimension two.
\[main\] Suppose that $\Delta u +Vu=0$ on $\mathbb{R}^2$, where $u$ and $V$ are real-valued and $|V|\leq 1$. If $|u(x)| \leq \exp(-C |x| \log^{1/2}|x|)$, $|x|>2$, where $C$ is a sufficiently large absolute constant, then $u\equiv 0$.
A similar striking difference between the decay estimates for real and complex solutions has also been observed in [@C15], where a closely related equation $\Delta u+W\cdot\nabla u=0$ with a bounded vector field $W:\mathbb{R}^2 \to \mathbb{R}^2$ was studied.
There is a simple example of a solution to with bounded $V$ that decays exponentially. Define $u=e^{-|x|}$ in $\{ |x| > 1\}$ and extend it to a $C^2$ smooth positive function on the plane. Then $|\Delta u| \leq C|u|$ and by taking $u(\frac{1}{\sqrt C} \cdot)$ in place of $u$ one can make $|V|\leq 1$ in this example.
The assumption that $u$ is real-valued is redundant because in the case of real-valued $V$ the real and imaginary parts of $u$ also satisfy . But in the proof we will use that $u$ is real-valued. The proof of Theorem \[main\] combines the technique of quasiconformal mappings with two tricks. The tricks involve nodal sets (zero sets) of $u$ and holes that are made in nodal domains (connected components of the complement of the zero set). We describe the idea in Section \[Idea\]. Some two-dimensional tools are used in the proof and the Landis conjecture in higher dimensions is still open.
Our second result is a local version of Landis’ conjecture.
\[local 1\] Let $u$ be a real solution to $\Delta u+ Vu=0$ in $B(0,2R)\subset\mathbb{R}^2$, where $V$ is real-valued and $|V|\leq 1$. Suppose that $|u(0)|=\sup\limits_{B(0,2R)} |u|= 1 $. Then for any $x_0$ with $|x_0|=R/2>2$, we have $$\sup\limits_{B(x_0,1)}|u| \geq \exp(-C R \log^{3/2}R)$$ with some absolute constant $C>0$.
The previous best known bound $\sup\limits_{B(x_0,1)}|u| \geq \exp(-CR^{4/3}\log R)$, was obtained in any dimension by Bourgain and Kenig [@BK05] in their proof of Anderson localization for the Bernoulli model, see also [@C05].
Theorem \[local 1\] follows from the main local Theorem \[local main\], where we don’t assume that $|u(0)|=\sup\limits_{B(0,2R)} |u|= 1 $, and prove a version of the three balls inequality.
Landis’ conjecture was a subject to an extensive study. Under additional assumptions on $V$, some versions of Landis’ conjecture are known, see [@B12],[@D19], [@DKW19],[@EKPV],[@C05],[@C15],[@K98],[@Z16] and references therein. A related problem in a cylinder was studied in [@G81].
**Notation.** By $c,C,C',... >0$ we denote various constants. Typically small constants are denoted by small letters and we use capital letters for large constants. If a constant $C$ depends on a domain (or some other parameter), we say it. Sometimes we state theorems without reminding that the functions are assumed to be real-valued and $u$ is a solution to on $\mathbb{R}^2$. A ball with center at $x$ of radius $r$ is denoted by $B(x,r)$ and the two-dimensional Lebesgue measure is denoted by $m_2$.
**Acknowledgements.** The authors are grateful to Misha Sodin, Alexandru Ionescu, Charles Fefferman and Carlos Kenig for fruitful discussions.
This work was completed during the time A.L. served as a Clay Research Fellow and Packard Fellow. E.M. was partially supported by NSF grant DMS-1956294 and by Research Council of Norway, Project 275113. F.N. was partially supported by NSF grant DMS-1900008.
Strategy of the proof and local versions. {#Idea}
=========================================
The proof consists of three acts. First, we will explain the main ideas of each of them.
**Description of Act I.** We will use the following well-known fact about nodal sets, which is proved in the Appendix (Lemma \[le: diameter\]) for reader’s convenience. There is an absolute constant $r_0>0$ such that if $u$ is a solution to $\Delta u+ Vu$ in a neighborhood of a closed ball $\overline{B(z_0,r)}$ with $|V|\leq 1$, $u(z_0)=0$ and $0<r<r_0$, then the circle $ C(z_0,r)=\{ z:|z-z_0|=r\}$ is intersecting the zero set of $u$.
It is also true that the singular set $$S=\{ x: u(x)=0 \textup{ and } \nabla u(x)= 0\}$$ consists of isolated points and the nodal set $$F_0=\{x:u(x)=0\}$$ is a union of smooth curves, see [@CF85]. However the proof will not use it, but this structural result about nodal sets makes it easier to think about them.
Now, assume that $u$ is a solution to in $B(0,R)$, $R>1$. Take $\varepsilon>0$ (a small parameter to be chosen later) and add finitely many $C\varepsilon$ – separated closed disks of radius $\varepsilon$ to $F_0$ so that the distance from each disk to $F_0$ is $\geq C\varepsilon$ and $$F_0\cup \textup{ union of the disks } \cup \{ z:|z|\geq R\}$$ is a $3C\varepsilon$ – net on the plane (assume $C>2$). Let us denote by $F_1$ the union of the closed disks, see Figure \[puncture fig\].
![Puncturing nodal domains[]{data-label="puncture fig"}](puncture.png){width="80.00000%"}
It can be shown that $$\Omega=\{z:|z| < R, z \notin F_0\cup F_1 \}$$ is an open (possibly disconnected) set with the Poincare constant $\leq C'\varepsilon^2$, i.e., for every $u \in W_0^{1,2}(\Omega)$, we have $$\int_{\Omega} u^2 \leq C'\varepsilon^2 \int_{\Omega} |\nabla u|^2.$$
It allows one to construct a function $\varphi$ in $B(0,R)$ such that
- $\Delta \varphi + V \varphi=0$ in $\Omega$,
- $\varphi - 1 \in W_0^{1,2}(\Omega)$,
- $\| \varphi-1\|_{\infty} \leq C''\varepsilon^2$.
The details are given in Section \[sec:Poincare\].
**Description of Act II.** Consider $f= \frac{u}{\varphi}$. Then $f$ satisfies $${\rm{div}}(\varphi^2 \nabla f)=0$$ in $\Omega$. The set $\Omega$ is usually not connected and the functions $\varphi$ and $f$ may be not smooth across $F_0$. However due to the fact that $F_0$ is the zero set of $u$ it appears (after some work) that the equation ${\rm{div}}(\varphi^2 \nabla f)=0$ holds through $F_0$ in the whole $B(0,R)\setminus F_1$.
Here the theory of quasiconformal mappings joins the game. After noticing that $f \in W_{loc}^{1,2}$, we may use the Stoilow factorization theorem to make a $K$– quasiconformal change of variables $g$ mapping $0$ to $0$ and $B(0,R)$ onto $B(0,R)$ such that $$f =h \circ g$$ where $h $ is a harmonic function in $B(0,R)\setminus g(F_1)$. Moreover, $K$ is very close to $1$ when $\| \varphi - 1 \|_\infty$ is small: $$K \leq \frac{1+\left\|\frac{1-\varphi^2}{1+\varphi^2} \right\|_\infty}{1-\left \|\frac{1-\varphi^2}{1+\varphi^2} \right \|_\infty} \leq 1+C \varepsilon^2.$$ Mori’s theorem tells us how much the distances are distorted depending on $K$: $$\frac{1}{16}\left|\frac{z_1 - z_2}{R}\right|^{K} \leq \frac{|g(z_1) - g(z_2)|}{R}
\leq 16\left|\frac{z_1 - z_2}{R}\right|^{1/K}.$$ We choose $$\varepsilon \sim \frac{1}{\sqrt{\log R}}$$ so that the distortion on scales from $
\frac{1}{R}$ to $R$ is bounded and, moreover, the images of the disks in $F_1$ have size comparable to $\varepsilon$.
Then we get a harmonic function $h$ in $B(0,R)\setminus g(F_1)$, where $g(F_1)$ is the union of sets of diameter $\sim \varepsilon$ and each set (the image of a single disk) is surrounded by an annulus of width $\sim C\varepsilon$ in which $h$ does not change sign.
**Description of Act III.** By rescaling we get the following question: Let $h$ be harmonic in a punctured domain $B(0,R') \setminus \cup_j D_j$ where $R' \sim \frac{R}{\varepsilon}\sim R \sqrt{\log R}$ and $D_j$ are $1000$– separated unit disks. Assume also that $h$ does not change sign in $5D_j\setminus D_j$. What can be said about the decay of $|h|$?
\[exercise 1\] Under the above assumptions, we have
$$\sup_{B(0,R') \setminus \cup_j 3D_j}|h| \leq \exp(CR')
\sup_{ \{ z: R'/8<|z|<R' \} \setminus \cup_j 3D_j} |h| \quad \textup{for } R' > 2000$$ with some absolute constant $C>0$.
Theorem \[exercise 1\] is an immediate consequence of a more general Theorem \[local 3\]. The outcome is that $|u|$ cannot decay faster than $\exp(-CR\sqrt{\log R})$. A different proof of the estimate for harmonic functions in a punctured domain (with a slightly worse bound) is given in the Appendix. The second proof works in higher dimensions and uses the Carleman inequality with log linear weight.
**Local versions.** Local versions of Theorem \[main\] (on the two dimensional plane) are also true. Here is the main local Theorem \[local main\].
\[local main\] If $u$ is a solution to $\Delta u + V u =0$ in $B(0, R)$, $R>2$, $V$ is real-valued, $|V| \leq 1$, and $$\frac{\sup_{B(0,R)}|u|}{\sup_{B(0,R/2)}|u|} \leq e^N,$$ then $$\label{eq: Br}
\sup_{B(0,r)}|u| \geq (r/R)^{C (R \log^{1/2}R +N)}\sup_{B(0,R)}|u|$$ for any $r<R/4$, where $C$ is an absolute positive constant.
Theorem \[local main\] implies Theorems \[main\] and \[local 1\]. In order to deduce Theorem \[main\], we may assume that $|u|$ attains its global maximum at some point on the plane, otherwise $|u|$ does not tend to $0$ near infinity. Let $$|u(z_{max})|= \max_{\mathbb{R}^2} |u|=1.$$ Then for any $ R > 6|z_{max}|$ and any $x$ with $|x|=R/3$, we have $$\sup_{B(x,R)}|u| = \sup_{B(x,R/2)}|u|=1$$ and if additionally $R>2$, then by Theorem \[local main\] applied to $u(\cdot + x)$, we have $$\sup_{B(x,R/4)}|u| \geq e^{-C R \log^{1/2} R}$$ and therefore $$\sup_{|z|>R/12}|u| \geq e^{-C R \log^{1/2} R}.$$ In order to deduce Theorem \[local 1\] note that $$\sup_{B(x,R/2)}|u| = \sup_{B(x,R)}|u|=1$$ for any $x$ with $|x|=R/2$ because $|u(0)|=\max_{B(0,2R)} |u|$. Applying Theorem \[local main\] to $u(\cdot + x)$ we get $$\sup\limits_{B(x,1)}|u| \geq R^{-C R\log^{1/2} R }= e^{-C R\log^{3/2}R}.$$
\[local 2\] Let $A>4$. If $u$ is a solution to $\Delta u + V u =0$ in $B(0, 1)$, $V$ is real-valued, $|V| \leq A$, and $$\frac{\sup_{B(0,1)}|u|}{\sup_{B(0,1/2)}|u|} \leq \exp(N),$$ then $$\label{eq:B_r}
\sup_{B(0,r)}|u| \geq r^{C( \sqrt{A \log A}+N)}\sup_{B(0,1)}|u| \quad \text{ for } r\leq 1/4,$$ where $C$ is an absolute positive constant.
For the proof, consider $u(\frac{1}{\sqrt A}\cdot)$ in place of $u$. We obtain a solution to $\Delta u+ Vu=0$ in $B(0,\sqrt A)$ with $|V| \leq 1$ and $$\frac{\sup_{B(0,\sqrt A)}|u|}{\sup_{B(0,\sqrt A /2)}|u|} \leq e^{N}$$ and we can apply Theorem \[local main\] to the new $u$ and $R=\sqrt A$.
Inequality implies that the vanishing order of $u$ at $0$ is bounded by $C (\sqrt{A \log A}+N)$. This question was previously studied in [@B12],[@K98],[@Z16].
On any smooth two dimensional Riemannian manifold $(M,g)$ every equation $\Delta_g u+ Vu=0$ can be simplified in local isothermal coordinates to $\Delta u + V'u=0$ (with ordinary Euclidean Laplacian $\Delta$). Corollary \[local 2\] gives information on the distribution of solutions to Schrodinger equations on compact manifolds of dimension 2.
\[global\] Let $(M,g)$ be a smooth closed (compact and without boundary) Riemannian manifold of dimension $2$. Then for any function $u$ satisfying $\Delta_g u + V u =0$ on $M$ with $|V| \leq \lambda$, $\lambda >2$, we have $$\sup\limits_{B_r}|u| \geq r^{C \sqrt{\lambda \log \lambda}}\sup_{M}|u|$$ for any ball $B_r$ of radius $r<1/2$. The constant $ C$ depends on the manifold.
This result follows from Corollary \[local 2\] by iterations (see the argument in [@DF], page 162, after formula (1.5)).
A slightly better bound was obtained in [@DF] by Donnelly and Fefferman for Laplace eigenfunctions on closed Riemannian manifolds of any dimension. If $\Delta_g u+ \lambda u=0$ on $(M,g)$, then $$\sup_{B_r} |u| \geq c r^{C \sqrt \lambda } \sup_M |u|, \quad r \leq \frac 1 2.$$ So the vanishing order at any point is at most $C\sqrt \lambda$.
In Act I and Act II we will reduce (with a logarithmic loss) the main local Theorem \[local main\] to a general Theorem \[local 3\], which is a local statement about two dimensional harmonic functions.
Act I {#sec:Poincare}
=====
Poincare constant for porous domains.
-------------------------------------
\[Poincare\] Let $F$ be a closed set in $B(0,R)$, $R>1$, such that
1. For every $z_0 \in F$, $r\in (0,1]$, the circle $ C(z_0,r)=\{ z:|z-z_0|=r\}$ intersects $F\cup \partial B(0,R)$.
2. $ F \cup \partial B(0,R)$ is $C-$dense in $B(0,R)$, $C>1$.
Then the Poincare constant of $\Omega= B(0,R)\setminus F$ is bounded by some constant $\widetilde C$ that depends only on $C$.
Let $f \in C_0^\infty(\Omega)$. Extend $f$ by zero outside $\Omega$. First, we will show that if $z\in F\cup \partial B(0,R)$, then $$\int_{B(z,3C)} |f|^2 \lesssim \int_{B(z,3C)} |\nabla f|^2 .$$ Every circle $C_r=\partial B(z,r)$, $r\in(0,1)$, has a zero of $f$, whence $$\max\limits_{C_r}|f| \leq \int\limits_{C_r }|\nabla f|$$ and $$\int\limits_{B(z,1)}|f|^2 = \int_0^1 \left(\,\int\limits_{C_r}|f|^2\right) dr \leq \int_0^1 |C_r|\max\limits_{C_r}|f|^2dr \leq \int_0^1 |C_r|\left(\,\int\limits_{C_r }|\nabla f|\right)^2 dr \leq$$ $$\leq \int_0^1 |C_r|^2\left(\,\int\limits_{C_r }|\nabla f|^2\right) dr \leq (2\pi)^2 \int_0^1 \left(\,\int\limits_{C_r }|\nabla f|^2\right)dr = (2\pi)^2 \int\limits_{B(z,1)} |\nabla f|^2.$$ We therefore can find $r\in (1/2,1)$ such that $$\int\limits_{C_r }|f|^2 \leq C_1 \int\limits_{B(z,1)}|f|^2 \leq C_2 \int\limits_{B(z,1)} |\nabla f|^2.$$ Let $\Gamma_\psi$, $\psi \in[0,2\pi)$, be a segment starting at the point $$x_\psi:= z + re^{i\psi}$$ and ending at the point $z+3Ce^{i\psi}$. Note that $$\max_{\Gamma_\psi} |f|^2 \leq \left(|f(x_\psi)| + \int_{\Gamma_\psi}|\nabla f|\right)^2 \leq 2|f(x_\psi)|^2+2\left(\int_{\Gamma_\psi}|\nabla f|\right)^2 \leq$$ $$\leq 2 |f(x_\psi)|^2 + 2|\Gamma_\psi|\int_{\Gamma_\psi}|\nabla f|^2 \leq 2 |f(x_\psi)|^2 + 6C \int_{\Gamma_\psi}|\nabla f|^2$$ and therefore $$\int\limits_{ B(z,3C)\setminus B(z,1)} f^2 \leq (3C)^2 \int_0^{2\pi} \max_{\Gamma_\psi} |f|^2 d\psi \leq$$ $$\leq C_1(C) \left[ \quad \int\limits_{C_r }|f|^2 + \int_0^{2\pi} \left(\int_{\Gamma_\psi}|\nabla f|^2\right) d\psi \right] \leq C_2(C) \int_{B(z,3C)} |\nabla f|^2.$$ Thus $$\int_{B(z,3C)} |f|^2 \leq C_3(C) \int_{B(z,3C)} |\nabla f|^2 .$$
We can choose a finite collection $Z_*$ of points $z$ in $F \cup \partial B(0,R)$ such that the balls $B(z,3C)$ cover $B(0,R)$ and each point is covered a bounded number of times. Finally, we have $$\int\limits_{B(0,R)} f^2 \leq \sum\limits_{z\in Z_*} \int_{B(z,3C)} |f|^2 \leq C_3(C) \sum\limits_{z\in Z_*} \int_{B(z,3C)} |\nabla f|^2 \leq$$ $$\leq C_4(C) \int\limits_{B(0,R)} |\nabla f|^2.$$
We start proving Theorem \[local main\]. Recall that $\Delta u + Vu =0$ in the ball $B(0,R)$ (we may think that $R$ is a large number) and $F_0$ is the zero set of $u$. We will use the fact that $u\in C^1(B(0,R))$, which is proved in the Appendix, see Fact \[fact5\]. Now, consider the following setting:
\[contours\] {width="80.00000%"}
Take $\varepsilon>0$ (a small parameter to be chosen later). Choose finitely many $C\varepsilon$ – separated closed disks of radius $\varepsilon$, whose union will be denoted by $F_1$, so that the distance from each disk to $F_0$ and $\partial B(0,R)$ is $\geq C\varepsilon$ and $$F_0\cup F_1\cup \partial B(0,R)$$ is a $3C\varepsilon$ – net in $B(0,R)$ (we assume $C>2$).\
For instance, one can get $F_1$ by considering the maximal number of open non-intersecting disks of radius $(C+1)\varepsilon$ in $B(0,R)\setminus F_0 $. The centers of the disks are $(2C+2)\varepsilon$ – separated. There is no point $x$ in $B(0,R)\setminus F_0$ that is $(2C+2)\varepsilon$ far from the centers of the disks and from $F_0 \cup \partial B(0,R)$, otherwise we could add one more disk of radius $(C+1)\varepsilon$ with center at this point. So we may choose the disks of radius $\varepsilon>0$ with the same centers, they will be $C\varepsilon$ – separated and $F_0\cup F_1\cup \partial B(0,R)$ will be a $2(C+1)\varepsilon$ – net.
**Two points to avoid.** Now, let us remove from $F_1$ the disks that are $C\varepsilon$ close to $0$ or to the point $z_{\max} \in \overline{B(0,R/2)}$ such that $$|u(z_{\max})|= \sup_{B(0,R/2)}|u|.$$ The set $F_0\cup F_1\cup \partial B(0,R) $ will still be a $10C\varepsilon$ – net, but now all disks from $F_1$ are also $C\varepsilon$- separated from $0$ and $z_{\max}$. The detail about avoiding those two points will be used only in the end of Act II.
Recall that $F_0$ has the property that for any $z_0 \in F_0$, every circle $C(z_0,r)$ with $r<r_0$ intersects $F_0$ or $\partial B_R$. Taking $u(\varepsilon \cdot)$ in place of $u$ (so the assumptions of Lemma \[Poincare\] hold for $\varepsilon< r_0$) and applying Lemma \[Poincare\] we arrive to the following conclusion.\
**Outcome.** The domain $$\Omega= B(0,R)\setminus (F_0 \cup F_1)$$ has Poincare constant $\leq C'\varepsilon^2$ and $B(0,R)\setminus F_1$ contains $0$ and $z_{\max}$.
Solving $\Delta \varphi+ V\varphi=0$.
-------------------------------------
The goal of this section is to construct an auxiliary solution to in a domain with a small Poincare constant, so that the solution has boundary values $1$ and is uniformly close to $1$.
\[solving Schrodinger\] Let $\Omega$ be a bounded open set with the Poincare constant $k^2$. Let $V\in L^\infty(\Omega)$. Assume that
$$k^2 \|V\|_\infty \ll 1.$$ Then there exists $\varphi=1+\tilde \varphi$ with $$\tilde \varphi \in W_{0}^{1,2}(\Omega), \|\tilde \varphi\|_\infty \leq Ck^2\|V\|_\infty$$ such that $\varphi$ is a weak solution to $\Delta \varphi + V\varphi=0$ in $\Omega$, where $C$ is an absolute positive constant.
We will use the following fact, which is proved in the Appendix, Lemma \[lem: solving\] and Lemma \[lem: solving2\].
**Fact.** When the Poincare constant of $\Omega$ is $ 1$, $v\in L^{\infty}(\Omega)$, there is a solution $\varphi$ to $\Delta \varphi = v$ in $W_{0}^{1,2}(\Omega)$ with $$\| \varphi\|_\infty \leq C\| v\|_\infty$$ and $$\| \varphi \|_{W_0^{1,2}} \leq C \|v\|_2.$$
**Corollary** (follows by rescaling). If $\Omega$ has Poincare constant $k^2$, then we can find a solution $\varphi$ to $\Delta \varphi = v$ with $$\| \varphi \|_\infty \leq Ck^2\| v\|_\infty$$ and $$\| \varphi \|_{W_0^{1,2}} \leq C_1(k) \|v\|_2.$$
Now, let $\varphi_1$ solve $\Delta \varphi_1= -V$ and for $n\geq 2$ let $\varphi_n$ solve $$\Delta \varphi_n= -V\varphi_{n-1}.$$ Note that this sequence is well defined since on each step the right-hand side is in $L^\infty$. We have $$\|\varphi_n\|_\infty\leq Ck^2\|V\|_\infty\|\varphi_{n-1}\|_\infty,\quad n\geq 2,$$ and $\|\varphi_1\|_\infty \leq Ck^2 \|V\|_\infty.$ We are assuming that $Ck^2 \|V\|_\infty \leq 1/2$. Hence $\|\varphi_n\|_\infty \leq 2^{-n+1} Ck^2\|V\|_\infty $ and $$\|\varphi_n\|_{W_0^{1,2}} \leq C_1(k) \|\varphi_{n-1}\|_2 \leq C_2(k) \|\varphi_{n-1}\|_{\infty} \leq C_3(k) 2^{-n}.$$
Thus the series $$\tilde \varphi = \varphi_1+\varphi_2+\dots$$ converges both in $L^{\infty}$ and in $W_0^{1,2}(\Omega)$ with $$\| \tilde \varphi\|_{\infty} \leq C'k^2 \|V\|_\infty.$$ Also for any $h \in W_0^{1,2}(\Omega)$, we have $\int \nabla \varphi_n \nabla h = \int V \varphi_{n-1}h$ for $n \geq 2$ and $\int \nabla \varphi_1 \nabla h = \int Vh$. Thus $\Delta \tilde \varphi= - V(1+\tilde \varphi)$ and $$\Delta(1+\tilde \varphi) + V(1+\tilde \varphi) =0 \quad \text{ in } \Omega$$ as required.
**Outcome.** Since the Poincare constant of $\Omega= B(0,R) \setminus (F_0 \cup F_1)$ is $\leq \widetilde C\varepsilon^2$, using Lemma \[solving Schrodinger\], we can find $\varphi$ such that
- $\Delta \varphi + V \varphi=0$ in $\Omega$,
- $\varphi - 1 \in W_0^{1,2}(\Omega)$,
- $\| \varphi-1\|_{\infty} \leq C'\varepsilon^2$.
Act II.
=======
Reduction to a divergence type equation in a domain with holes.
---------------------------------------------------------------
Recall that $u$ is a solution to $\Delta u+ Vu=0$ in $B(0,R)$ and $F_0$ is the zero set of $u$. Extend the function $\varphi$ by 1 outside $$\Omega=B(0,R) \setminus (F_0\cup F_1).$$
\[div equation\]
The function $\frac{u}{\varphi} \in W^{1,2}_{loc}(B(0,R))$ and it is a solution to $$\textup{ div}(\varphi^2 \nabla (\frac{u}{\varphi})) = 0$$ in $B(0,R) \setminus F_1$ in the weak sense.
**Remark.** The lemma takes care of all “continuations through nodal lines" of $u$.
First, we would like to notice that the extended functions $\frac{1}{\varphi}, \varphi \in W^{1,2}_{loc}(\mathbb{R}^2)$ and $$\label{grad 1/phi}
\nabla \frac{1}{\varphi}= -\mathbbm{1}_{\Omega} \frac{\nabla \varphi}{\varphi^2} \quad \text{and} \quad \nabla \varphi = \mathbbm{1}_\Omega \nabla \varphi$$ in $\mathbb{R}^2$ in the sense of distributions: $$\int_{\mathbb{R}^2} \frac{1}{\varphi} \nabla \xi= \int_{\Omega}\frac{\nabla \varphi}{\varphi^2} \xi \quad \text{and} \quad \int_{\mathbb{R}^2} \varphi \nabla \xi= -\int_{\Omega}\nabla \varphi \xi$$ for any $\xi
\in C^\infty_0(\mathbb{R}^2) $. The formal check is performed in Fact \[fact6\] in the Appendix.
Now, we would like to verify that $ \frac{u}{\varphi} \in W^{1,2}_{loc}(B(0,R))$ and $$\nabla \frac{u}{\varphi} = \frac{\varphi\nabla u}{\varphi^2} - \frac{u \nabla \varphi }{\varphi^2} \mathbbm{1}_\Omega.$$
\[product\] Let $u,v \in W^{1,2}_{loc}(B(0,R))\cap L^{{}^{\scriptsize \infty}}_{loc}(B(0,R))$. Then $uv \in W^{1,2}_{loc}(
B(0,R)) $ and $\nabla(uv) = u \nabla v + v \nabla u$.
Fact \[product\] is proved in the Appendix.
Recall that $\varphi$ is extended by $1$ outside $\Omega$, $\frac{1}{\varphi} \in W^{1,2}_{loc}(\mathbb{R}^2)$ and $u$ is $C^1$-smooth in $B(0,R)$ by Fact \[fact5\]. By Fact \[product\] we know that $ \frac{u}{\varphi} \in W^{1,2}_{loc}(B(0,R))$ and, as expected, $$\nabla \frac{u}{\varphi} = \frac{\varphi \nabla u}{\varphi^2} - \frac{u \nabla \varphi}{\varphi^2} \mathbbm{1}_\Omega$$ in $B(0,R)$ in the sense of distributions. To establish the divergence-type equation for $\nabla \frac{u}{\varphi}$ we want to show that for every test function $h \in C_0^\infty(B(0,R)\setminus F_1)$, we have $$\int_{B(0,R)\setminus F_1} \varphi^2 \nabla (\frac{u}{\varphi}) \nabla h =0.$$ So we need to prove that $$\label{need}
\int_{B(0,R)\setminus F_1} (\varphi \nabla u - u \nabla \varphi \mathbbm{1}_\Omega) \cdot \nabla h =0.$$ Since $u$ is a solution to $\nabla u +Vu =0$ in $B(0,R)$, we have $$\label{have1}
\int_{B(0,R)\setminus F_1} \nabla u \cdot (\varphi \nabla h + h \nabla \varphi) = \int_{B(0,R)\setminus F_1} V\varphi u h$$ (we know the last equality under the assumption that $
\varphi$ is smooth, but it is also true for $\varphi \in W^{1,2}_{loc}(\mathbb{R}^2)$ by taking the norm limit). Consider a function $\xi \in C_0^\infty(B(0,R) \setminus F_0)$ that descends from $1$ to $0$ in the $\varepsilon$ – neighborhood of $F_0\cup \partial B(0,R)$ with $|\nabla \xi| < C/\varepsilon$.
Since $\Delta \varphi + V \varphi =0$ in $$\Omega= B(0,R) \setminus (F_0 \cup F_1)$$ and $uh\xi \in C_0^1(\Omega)$, we have
$$\label{have2}
\int_\Omega \nabla \varphi \cdot(h \nabla u \xi +u\nabla h \xi + uh \nabla \xi)= \int_\Omega V\varphi uh\xi.$$
Note that $\int_\Omega V\varphi uh\xi$ tends to $\int_\Omega V\varphi u h$ as $\varepsilon \to 0$ (the functions $V,\varphi,uh,\xi$ are uniformly bounded and the convergence holds pointwise in $\Omega$ because $\xi \to 1 $ in $B(0,R)\setminus F_0$). Note that $$h \nabla u \xi \to h\nabla u \quad \text{pointwise in } \Omega$$ and $$u\nabla h \xi \to u\nabla h \quad \text{pointwise in } \Omega$$ because $\xi \to 1$ in $\Omega$. Hence $$\label{have3} \int_\Omega \nabla \varphi \cdot(h \nabla u \xi +u\nabla h \xi) \to \int_\Omega \nabla \varphi (h\nabla u + u\nabla h) \quad \text {as } \varepsilon \to 0$$ by the Lebesgue dominated convergence theorem with the majorant $|\nabla \varphi| (|h||\nabla u| + |u||\nabla h|)$.
In order to prove we will show that $$\int_{\Omega} \nabla \varphi \cdot (uh\nabla \xi) \to 0.$$ And here is the main place where we use that $F_0$ is the zero set of $u$! Note that $uh \in C^1_0(B(0,R))$ and vanishes on $F_0$, so $|uh| \leq C_1(u,h)\varepsilon$ in the $\varepsilon$– neighborhood of the zero set of $u$. Thus $|uh\nabla \xi|$ is bounded by some constant $C(u,h)$ in $B(0,R)$. Also $m_2(\textup{supp} \nabla \xi)$ goes to 0. Hence
$$\int_\Omega \nabla \varphi \cdot (hu\nabla \xi) \leq C(u,h) \sqrt{m_2(\textup{supp} \nabla \xi)} \sqrt{\int_\Omega |\nabla \varphi|^2} \to 0.$$ By , we obtain $$\int_{\Omega} \nabla \varphi \cdot(h \nabla u +u\nabla h) = \int_{\Omega} V\varphi uh=\int_{B(0,R)\setminus F_1} V\varphi uh$$ (the second equality is due to the fact that $u=0$ on $F_0$). Using $\nabla \varphi = \nabla \varphi \mathbbm{1}_\Omega$ in the sense of distributions, we have $$\label{have4}
\int_\Omega \nabla \varphi (h\nabla u + u\nabla h) = \int_{B(0,R)\setminus F_1} \nabla \varphi (h\nabla u + u\nabla h).$$ Thus $$\int_{B(0,R)\setminus F_1} \nabla \varphi \cdot(h \nabla u +u\nabla h) = \int_{B(0,R)\setminus F_1} V\varphi uh$$ and, subtracting , we finish the proof of .
Quasiconformal change of variables.
-----------------------------------
We briefly describe some facts from the theory of quasiconformal mappings, which are used in the study of the solutions to equations in divergence form on the plane, and explain why the solutions behave like ordinary harmonic functions. We partially follow the exposition from [@NPS], where the quasiconformal mappings are applied to quasi-symmetry of Laplace eigenfunctions.
Let $B$ be a disk on the plane. Consider a real-valued function $f \in W^{1,2}_{loc}(B)$ satisfying $$\label{eq:div}
\textup{div}(\varphi^2 \nabla f) = 0$$ and assume that $ 0< c < \varphi(x) < C < +\infty$ in $B$. One can find a function $\tilde f \in W^{1,2}_{loc}(B)$ such that $$\varphi^2 f_x = \tilde f_y \textup{ and } \varphi^2 f_y = -\tilde f_x$$ (see Section \[sec: divergence\]) and $f$ appears to be the real part of $w=f+i\tilde f$. A direct computation shows that $w$ is a solution to the Beltrami equation: $$\label{eq:Beltrami}
\frac{\partial w}{\partial \overline z} = \mu \frac{\partial w}{\partial z}$$ with the Beltrami coefficient $$\label{eq:mu}
\mu= \frac{1-\varphi^2}{1+\varphi^2}\cdot \frac{f_x+i f_y}{f_x-if_y}.$$ When $\nabla f=0$, we put $\mu=0$.
We are going to apply the theory of quasiconformal mappings in a situation when $f=\frac{u}{\varphi} $ and the domain $$\Omega_1:= B(0,R)\setminus F_1$$ is not simply connected. In this case $w$ and $\tilde f$ can be defined only locally, but not in the whole $\Omega_1$. However the Beltrami coefficient $\mu$ is well defined by in $\Omega_1$ and $$|\mu| \leq \frac{1-\varphi^2}{1+\varphi^2} \leq C\varepsilon^2.$$ Let us extend $\mu$ by zero outside $\Omega_1$ to the whole complex plane. Now $\mu$ has a compact support.
The existence Theorem 5.3.2 [@AIM09] claims that there is a $K$-quasiconformal homeomorphism $\psi$ of the complex plane such that
- $\psi \in W^{1,2}_{loc}$,
- $ \frac{\partial \psi}{\partial \overline z} = \mu \frac{\partial \psi}{\partial z}$,
- $ K \leq \frac{1+\sup |\mu|}{1 - \sup |\mu|}.$
In our case $$K \leq 1+ C' \varepsilon^2.$$ **Claim.** The function $f \circ \psi^{-1}$ is harmonic in $\psi(\Omega_1)$.
Indeed, for any ball $B \subset \Omega_1$, we can define $w\in W^{1,2}_{loc}(B)$ such that $f=\Re w$ and $w$,$\psi$ solve the same Beltrami equation. Stoilow factorization theorem ([@AIM09], p.179, Theorem 5.5.1) claims that there is a holomorphic function $W$ such that $$w = W(\psi(z))$$ and therefore the harmonic function $\Re W$ satisfies $$f(z)= \Re W(\psi(z)).$$ Clearly, the local observation shows that $f\circ \psi^{-1}$ is a harmonic function in $\psi(\Omega_1)$.
Note that $\psi(B(0,R))$ is a simply connected domain (and not the whole plane). Using the Riemann uniformisation theorem we can find a conformal map that sends $\psi(B(0,R))$ back to $B(0,R)$ and $\psi(0)$ to $0$. The composition of this conformal map and the $K$-quasiconformal homeomorphism $\psi$ will be a $K$-quasiconformal homeomorphism $g$ of $B(0,R)$ onto itself with $g(0)=0$. Then the function $h=f \circ g^{-1}$ is harmonic in $g(\Omega_1)$.
**Distortion of quasiconformal mappings.** Mori’s theorem ([@Ahl66], Chapter III, Section C) tells us that distances are changed by $g$ in a controlled way: $$\label{eq:distortion}
\frac{1}{16}\left|\frac{z_1 - z_2}{R}\right|^{K} \leq \frac{|g(z_1) - g(z_2)|}{R} \leq 16\left|\frac{z_1 - z_2}{R}\right|^{1/K}.$$
We choose $$\varepsilon = \frac{c}{\sqrt{\log R}}$$ so that $$K\in [1,1+C c^2/ \log R), \quad R^K\asymp R \asymp R^{1/K}$$ and the distortion on scales from $
\frac{1}{R}$ to $R$ is bounded. Namely, we may choose $c$ so small that if $\frac{1}{R}\leq|z_1-z_2| \leq2R$, then $$\frac{1}{32}|z_1-z_2| \leq|g(z_1)-g(z_2)| \leq 32|z_1-z_2|.$$
Note that in the statement of Theorem \[local main\] one can safely assume that $R$ is sufficiently large ($R\gg 1$) by rescaling, which makes $\|V\|_\infty$ only smaller. It is needed to make $\varepsilon \geq 1/R$. Then we get a harmonic function $h$ in $B(0,R)\setminus g(F_1)$, where $g(F_1)$ is the union of sets of diameter $\sim \varepsilon$. The image of a single disk of radius $\varepsilon$ will be contained in a disk of radius $32\varepsilon$. Let us denote these disks of radius $32\varepsilon$ by $D_j$. The images of disks from $F_1$ are $ \frac{C}{32}\varepsilon$ – separated from each other and from the zero set of $h$. Hence $D_j$ are $ (\frac{C}{32}\varepsilon - 128\varepsilon)=C_132\varepsilon$ – separated from each other and from the zero set of $h$, and $h$ does not change sign in $C_1D_j \setminus D_j$. We have $$C_1 = \frac{C}{32^2} - 4 > 100$$ if $C= 10^6$.
We specifically asked that $0$ and $z_{\max}$ (the point where $\sup_{B(0,R/2)}|u|$ is attained) are $C\varepsilon$ – separated from the disks. Recall that $g(0)=0$, so the disks $C_1D_j$ do not contain $0$ and $g(z_{\max})$. The distortion estimate implies that $g(z_{\max}) \in \overline{B(0,R-R/64)}$. Since we had $$\frac{\sup_{B(0,R)}|u|}{\sup_{B(0,R/2)}|u|} \leq e^N,$$ we conclude that $$\frac{\sup_{B(0,R) \setminus \cup 3D_j}|h|}{\sup_{B(0,R-R/64)\setminus \cup 3D_j}|h|} \leq e^N.$$
If we make the rescaling by a factor of $32 \varepsilon$, then the disks $D_j$ become 100-separated unit disks and $R$ becomes $$R'=R\cdot32\varepsilon\sim R\sqrt{\log R}.$$
The goal of Theorem \[local main\] is to estimate $\sup_{B(0,r)}|u|$ from below. If $r<1/R$, the image of $B(0,r)$ may have radius significantly smaller than $r$. However $g(B(0,r))$ contains a disk with center at $0$ of radius $$\frac{R}{16}\left(\frac{r}{R}\right)^{K} \geq \frac{R}{16} \left(\frac{r}{R}\right)^2.$$ Let $\tilde g = \frac{1}{32\varepsilon} g$. Then $\tilde g(B(0,r))$ contains a ball $B(0,r')$, where $$r' \geq
\frac{R'}{16} \left(\frac{r}{R}\right)^2.$$ So $$\frac{R'}{r'} \leq 16 \frac{R^2}{r^2}.$$ In order to prove estimate , it is enough to show that $$\sup_{B(0,r')\setminus \cup 3D_j}|h| \geq c (r'/R')^{C (R' +N)}\sup_{B(0,R')\setminus \cup 3D_j}|h|.$$ It will be proved in Theorem \[local 3\].
Act III {#sec: toy}
=======
Before we formulate and prove the promised local Theorem \[local 3\] we will explain the main idea in the global case.
\[thm: toy\]
Let $\{D_j\}$ be a collection of $100$-separated disks with unit radius on the complex plane $\mathbb{C}$. Suppose that $u$ is a harmonic function in $\mathbb{C} \setminus \cup_j D_j$ which preserves sign in each annulus $5D_j\setminus D_j$. If $|u(z)| \leq e^{-L|z|}$ for all $z \in \mathbb{C} \setminus \cup_j D_j$ and $L$ is sufficiently large, then $u \equiv 0$.
We start with a simple observation.
Let $m_j= \min\limits_{\partial 3D_j}|u|$. Then for some absolute constant $A>0$, we have
1. $\max\limits_{\partial 3D_j}|u| \leq Am_j$,
2. $\max\limits_{\partial 3D_j}|\nabla u| \leq Am_j$.
By the Harnack inequality there exists a constant $A>0$ such that $$\sup\limits_{4D_j\setminus 2 D_j}|u| \leq A\inf\limits_{4D_j\setminus 2 D_j}|u| \leq Am_j,$$ which proves the first part of the claim. The second part follows from the Cauchy inequality.
Let $k\in (0,L)$ and consider the numbers $m_j e^{k \Re z_j}$, where $z_j$ is the rightmost point of $3D_j$.
{width="50.00000%"}
Since $$m_j \leq |u(z_j)| \leq e^{-L|z_j|},$$ there is $j_0$ such that $$m_{j_0} e^{k \Re z_{j_0}}= \max\limits_j m_je^{k \Re z_j}.$$
Now, consider the analytic in $\mathbb{C} \setminus \cup (3D_j)$ function $f=(u_x - i u_y) e^{kz}$. If $|u(z)| \leq e^{-L|z|}$ in $\mathbb{C} \setminus \cup (D_j)$, then $$|\nabla u(z)| \leq C \sup_{B(z,1)}| u| \leq Ce^{-L(|z|-1)} \quad \text{ for } z \in \mathbb{C} \setminus \cup (2D_j)$$ and $f(z) \to 0$ as $z \to \infty$, $z \in \mathbb{C} \setminus \cup (2D_j)$. So, by the maximum principle, there exists $j_1$ such that $$\max\limits_{\mathbb{C} \setminus \cup (3D_j)}|f|= \max\limits_{\partial 3D_{j_1}} |f| \leq Am_{j_1} e^{k\Re z_{j_1}} \leq Am_{j_0} e^{k\Re z_{j_0}},$$ whence $|\nabla u| \leq Am_{j_0} e^{-k\Re(z-z_{j_0})}$ in $\mathbb{C} \setminus \cup (3D_j)$. We may assume that $m_{j_0}\neq 0$, otherwise $u$ is constant and therefore zero.
Now, consider the ray $\{z_{j_0} + y: y \in (0,+\infty) \}$. There are two possibilities:
- The ray goes to $\infty$ without hitting any other disks $(3D_j)$. Then for any $y>0$, $$|u(z_{j_0}+y) - u(z_{j_0})| \leq \int_{0}^{\infty}|\nabla u(z_{j_0}+t)| dt \leq \int_{0}^{\infty} Am_{j_0} e^{-kt}dt = \frac A k m_{j_0}.$$ Since $|u(z_{j_0})| \geq m_{j_0}$, wee see that $|u|$ stays bounded from below by $(1- \frac A k)m_{j_0}$ on the ray. If $k>A$, this contradicts the decay assumption.
- The ray hits another disk $3D_j$
{width="90.00000%"} at some point $z_{j}'= z_{j_0}+y$. Then we still have $|u(z_j')|\geq (1-\frac A k )m_{j_0}$ and, due to the fact that the disks are separated, $$\Re (z_{j}' - z_{j_0} )=|z_{j}' - z_{j_0}| \geq 1.$$ Hence $$m_j e^{k\Re z_j } \geq m_j e^{k \Re z_j'} \geq \frac{|u(z_j')|}{A} e^{k(\Re z_{j_0} +1)} \geq$$ $$\geq \frac{1}{A}(1 - \frac{A}{k})e^k m_{j_0} e^{\Re z_{j_0}} > m_{j_0} e^{\Re z_{j_0}}$$ as soon as $k>2A$, which contradicts the choice of $j_0$.
This proves the theorem with any $L>2A$.
Now we formulate and prove the harmonic counterpart of the main local theorem.
\[local 3\] Let $D_j$ be a collection of 100 – separated unit disks on $\mathbb{R}^2=\mathbb{C}$ such that $0 \notin \cup 3D_j$. Let $R> 10^4$, $0<r\leq R/4$. Consider any harmonic function $u$ in $B(0,R) \setminus \cup D_j$ such that $u$ does not change sign in $(5D_j\setminus D_j) \cap B(0,R)$ for every $j$. Assume that $$\sup \limits _{B(0,R-R/64) \setminus
\cup 3D_j}|u| \geq e^{- N} \sup \limits_{B(0,R) \setminus
\cup 3D_j}|u|.$$ Then $$\label{eq:*}
\sup \limits _{B(0,r) \setminus
\cup 3D_j}|u| \geq \left(\frac r R \right)^{ C (R + N) } \sup \limits_{B(0,R) \setminus
\cup 3D_j}|u|,$$ with some absolute constant $C>0$.
WLOG, $\sup \limits _{B(0,R-R/64) \setminus
\cup 3D_j}|u|=1$. Fix $k=[C(N+R)]$ with sufficiently large $C>0$ and assume that $$\sup \limits _{B(0,r) \setminus
\cup 3D_j}|u| \leq \left(\frac r R \right)^{3k}.$$
Consider the domain $$\Omega:=\{r/2<|z|< R - 1 \} \setminus \cup (3D_j).$$ Let $W_1$ be the connected component of $\partial \Omega$ that intersects $\partial B(0,r/2)$. Note that each point of $W_1$ is either on $\partial B(0,r/2)$ or lies on some $\partial 3D_j$ that intersects $\partial B(0,r/2)$.\
**Estimate on $W_1$.** Recall that if $5D_j \subset B(0,R)$, we have
1. $\max\limits_{\partial3 D_j }|u| \leq A \min\limits_{\partial 3D_j}|u|$,
2. $\max\limits_{\partial 3D_j}|\nabla u| \leq A \min\limits_{\partial 3D_j}|u|$.
Hence on $W_1 \setminus \partial B(0,r/2) $, we have $$|u|,|\nabla u| \leq A \sup_{B(0,r)\setminus \cup (3D_j)} |u| \leq A \left(\frac r R \right)^{3k}.$$ If $x \in \partial B(0,r/2) \setminus \cup (3D_j)$, then either $x \in 4D_j$ for some $j$ or $B(x, min(1,r/2)) \subset B(0,r) \setminus \cup (3D_j) $. In the first case $u$ does not change sign in $B(x,1)$ and $$|\nabla u(x)| \leq A|u(x)| \leq A\sup_{B(0,r)\setminus \cup (3D_j)} |u|\leq A\left(\frac r R \right)^{3k}.$$ In the second case, we have $$|\nabla u(x)| \leq\frac{A}{\min(1,r/2)}\sup_{B(0,r)\setminus \cup (3D_j)}|u| \leq \frac{A}{\min(1,r/2)} \left(\frac r R \right)^{3k}.$$ Thus in all cases, if $C$ in the definition of $k$ is large enough, we have $$\max_{W_1}|u|, \max_{W_1}|\nabla u| \leq \frac{A}{\min(1,r/2)} \left(\frac r R \right)^{3k} \leq \left(\frac r R \right)^{2k} \quad$$ because $$\left( \frac R r \right)^{k} \geq 4^k > A$$ and $$\left( \frac R r \right)^{k} \geq 4^{k-1} \frac R r \geq \frac{2A}{r}.$$
Let $W_2$ be the connected component of $\partial \Omega$ that intersects $\partial B(0,R-1)$. Note that each point of $W_2$ is either on $\partial B(0,R-1)$ or lies on some $\partial 3D_j$ that intersects $\partial B(0,R-1)$.\
**Estimate on $W_2$.** Any point $x\in \overline{B(0,R-1)}\setminus \cup (3D_j)$ is either in $4D_j$ for some $j$ or $x\in \overline{B(0,R-1)}\setminus \cup (4D_j)$. In the first case $u$ does not change sign in $B(x,1)$ and therefore $$|\nabla u(x)|\leq A|u(x)| \leq A e^N.$$ In the second case $B(x,1) \subset B(0,R)\setminus \cup (3D_j)$ and $|\nabla u(x)| \leq A e^N$. Thus $$\max_{W_2}|u|, \max_{W_2}|\nabla u| \leq A e^N.$$ Note also that $$W_2 \subset \overline{B(0,R-1)}\setminus B(0,R-7).$$
Now, consider the analytic in $\Omega$ function $$f(z)=\frac{u_x-iu_y}{z^k}, \quad |f(z)|=\frac{|\nabla u(z)|}{|z|^k}.$$ Since $$\sup\limits_{B(0,R-R/64)\setminus \cup (3D_j)}|u|=1> \sup_{B(0,r/2)\setminus \cup 3D_j}|u|,$$ $$\max\limits_{W_1} |u| \leq \left( \frac r R \right)^{2k} < \frac{1}{2},$$ and since any point in $$\Omega_1=B(0,R-R/64)\setminus \cup (3D_j)$$ can be connected with $W_1$ by a curve of length at most $4R$ within $\Omega_1$, we must have $$\sup_\Omega|\nabla u| \geq \sup_{\Omega_1}|\nabla u| \geq \frac{1}{8R}$$ and $$\sup_{\Omega_1} |f|\geq \left( R-R/64\right)^{-k} \sup_{\Omega_1}|\nabla u| \geq \frac{1}{8R} \left( R-R/64 \right)^{-k}.$$ However $$\max_{W_1}|f|\leq \max_{W_1}|\nabla u |\left(\frac{2}{r} \right)^k \leq \left(\frac{r}{R} \right)^{2k}\left(\frac{2}{r} \right)^k= \left(\frac{2r}{R} \right)^{k} R^{-k} \leq$$ $$\leq 2^{-k} R^{-k} < \frac{1}{8R} R^{-k} < \sup_{\Omega_1} |f|\leq \sup_{\Omega} |f|$$ and $$\max_{W_2}|f|\leq A e^N \frac{1}{(R-7)^k} = A e^N \left( \frac{R-R/64}{R-7}\right)^k \left( R-R/64\right)^{-k}\leq$$ $$\leq A e^N \left( \frac{126}{127}\right)^k \left( R-R/64\right)^{-k} < \frac{1}{8R} \left( R-R/64\right)^{-k} \leq \sup_{\Omega} |f|$$ if $R-7> \frac{127}{128}R $ and $C$ in the definition of $k$ is large enough. By the maximum principle for holomorphic functions $\sup_\Omega |f|$ is achieved on $\partial 3D_j$ for some $3D_j\subset B(0,R-1)\setminus \overline{B(0,r/2)}$.
For every disk $D_j$ with $3D_j\subset B(0,R-1)\setminus \overline{B(0,r/2)}$, consider the point $z_j$ on $\partial3D_j$ closest to the origin. All $3D_j$ that are not in the annulus $B(0,R-1)\setminus \overline{B(0,r/2)}$ will not be considered further. Put $m_j=\min_{\partial3D_j}|u|$. Let $j_0$ be the index such that $$\frac{m_{j_0}}{|z_{j_0}|^k} = \max_j \frac{m_{j}}{|z_{j}|^k}.$$ If $\sup_\Omega|f|$ is achieved on $\partial3D_{j_1}$, then for $x\in \Omega$, $$\frac{|\nabla u(x)|}{|x|^k}\leq \frac{1}{|z_{j_1}|^k} \max_{\partial3D_{j_1}}|\nabla u|\leq A \frac{m_{j_1}}{|z_{j_1}|^k} \leq A \frac{m_{j_0}}{|z_{j_0}|^k}.$$ So we conclude that $$|\nabla u(x)| \leq \left( \frac{|x|}{|z_{j_0}|} \right)^k A m_{j_0}.$$ Recalling that $\frac{1}{8R} \leq \sup_\Omega|\nabla u|$, we get $$\frac{1}{8R} \leq \left(\frac{2R}{r}\right)^k Am_{j_0},$$ so $$|u(z_{j_0})| \geq m_{j_0} \geq \frac{1}{8AR} \left(\frac{r}{2R}\right)^k.$$
Now, let $(sz_{j_0},z_{j_0})$ be the longest subinterval of the radius $[0,z_{j_0}]$ starting at $z_{j_0}$ that is contained in $\Omega$. We have $$|u(sz_{j_0})| \geq |u(z_{j_0})| - |z_{j_0}| \int_{s}^{1}|\nabla u(t z_{j_0})| dt \geq |u(z_{j_0})| - |z_{j_0}| \int_{0}^{1} Am_{j_0} t^k dt \geq$$ $$\geq m_{j_0}(1-\frac{AR}{k+1}) \geq \frac{m_{j_0}}{2}$$ if the constant $C$ in the definition of $k$ is large enough. Note that $$\left(\frac{R}{r}\right)^{k} \geq 4^{k} > 16 AR \cdot 2^k.$$ Hence $$\frac{m_{j_0}}{2} \geq \frac{1}{16AR} \left(\frac{r}{2R}\right)^k > \left(\frac{r}{R}\right)^{2k}$$ and the point $sz_{j_0}$ cannot belong to $W_1$, whence it belongs to some $\partial 3D_j$ with $3D_j \subset B(0,R-1)\setminus \overline{B(0,r/2)}$. Then $$\frac{m_j}{|z_j|^k} \geq \frac{|u(sz_{j_0})|}{A|sz_{j_0}|^k} \geq \frac{1}{2As^k} \frac{m_{j_0}}{|z_{j_0}|^k}.$$ It remains to notice that, since the distance from $3D_j$ to $3D_{j_0}$ is at least $96$, we have $$s^k \leq (1-96/R)^k< \frac{1}{2A}$$ if the constant $C$ in the definition of $k$ is large enough. But then $\frac{m_j}{|z_j|^k} > \frac{m_{j_0}}{|z_{j_0}|^k}$, which contradicts the choice of $j_0$.
Appendix.
==========
The toy problem for harmonic functions in higher dimensions: a proof with extra logarithm.
------------------------------------------------------------------------------------------
Here we present another proof of a slightly worse bound for the toy problem for harmonic functions in a punctured domain. However this proof works in higher dimensions. We will denote by $B_R$ the ball in $\mathbb{R}^n$ with center at $0$ and of radius $R$.\
**Toy problem with extra logarithm.** Let $D_j$ be a collection of unit, $100$ – separated balls on the plane and let $R>100$. Then for any harmonic function $h$ in $B_R \setminus \cup D_j$ such that $h$ does not change sign in each $B_R\cap 5D_j\setminus D_j$, we have $$\int\limits_{B_{R}\setminus(B_{R/2} \bigcup (\cup 3D_j))} h^2 \geq \exp(-CR\log R) \int\limits_{B_{R/2}\setminus \cup 3D_j} h^2,$$ where $C$ is an absolute positive constant.
This inequality implies that Theorem \[thm: toy\] holds in higher dimensions if we assume that $|u(z)| \leq e^{-L|z|\log|z|}$ for sufficiently large $L$.
The proof is based on the Carleman inequality with log linear weight. Most of Carleman inequalities require strict log convexity-type properties of the weight. The next inequality is an exception: $$\label{eq: Carleman}
\int_{B_R} |{\Delta} u|^2 e^{kx_1} \geq \frac{ck^2}{R^2} \int_{B_R} u^2 e^{kx_1}$$ for any $u\in C^2_0(B_R)$. The inequality is not difficult to prove. Let $v=ue^{kx_1/2}$, then $$e^{kx_1/2} \Delta u = \Delta v - k v_{x_1} + \frac{k^2}{4} v$$ and $$\int_{B_R} |{\Delta} u|^2 e^{kx_1}= \int_{B_R} |\Delta v + \frac{k^2}{4} v|^2 + \int_{B_R} |k v_{x_1}|^2 - \int_{B_R} 2 (\Delta v + \frac{k^2}{4} v) kv_{x_1}.$$ Note that $$\int_{B_R} 2 v v_{x_1} =\int_{B_R} \frac{\partial}{\partial x_1}v^2=0.$$ Integrating by parts, we see that $$-\int_{B_R} \Delta v_{x_1} v=\int_{B_R} \Delta v v_{x_1} = \int_{B_R} v \Delta v_{x_1}$$ and therefore $$\int_{B_R} \Delta v v_{x_1} = 0.$$ Hence $$\int_{B_R} |{\Delta} u|^2 e^{kx_1}= \int_{B_R} |\Delta v + \frac{k^2}{4} v|^2 + \int_{B_R} |k v_{x_1}|^2 \geq \int_{B_R} |k v_{x_1}|^2 \geq$$ (by Poincare’s inequailty) $$\geq \frac{\pi^2}{4}\frac{k^2}{ R^2} \int_{B_R} v^2= c\frac{k^2}{ R^2} \int_{B_R} u^2 e^{kx_1}.$$ So we proved and would like to apply it for the harmonic function $h$. However $h$ is not in $C^2_0(B_{R})$ and inequality should be applied to $$u=h\eta,$$ where $\eta$ is a positive, $C^2$-smooth cut-off function:
- $\eta=0$ in each $2D_j$ and in $\{x: |x|>R-11\}$,
- $\eta=1$ in $B_{\frac{3}{4}R \setminus \cup 3D_j}$,
- the function $\eta$, as well as its first and second derivatives are bounded by a numerical constant.
We will choose the parameter $k$ later. For now we have $$\int_{B_{\frac{3}{4}R}\setminus\cup 3D_j} |{\Delta} h|^2 e^{kx_1} + \textup{``cut-off integrals"} \geq \frac{ck^2}{R^2} \int_{B_{\frac{3}{4}R}\setminus \cup 3D_j} h^2 e^{kx_1} =: \textup{RHS}.$$ It is good that $\Delta h=0$, so only the cut-off integrals are left on the left-hand side. There are two kinds of cut-off integrals: $$\textup{I}= \sum\limits_{5D_j \subset B_{\frac{3R}{4}}} \int_{3D_j\setminus 2D_j} \textup{``cut-off terms"}$$ and $$\textup{II}= \int_{B_{R-11}\setminus B_{(\frac{3R}{4}-10)}} \textup{``cut-off terms"}$$ where $$| \textup{``cut-off terms"}| \lesssim (h^2+|\nabla h|^2)e^{kx_1}$$ (recall that $\Delta h =0$). Note that $$\int_{3D_j\setminus 2D_j} e^{kx_1} \lesssim e^{-k/2}\int_{4D_j\setminus 3D_j} e^{kx_1}$$ because $4D_j\setminus 3D_j$ contains an open disk of radius $\frac{1}{4}$, where the function $e^{kx_1}$ is pointwise bigger than $e^{k/2} \cdot \sup_{3D_j\setminus 2D_j} e^{kx_1}$. Now, assuming $5D_j \subset B_R$ we will use the sign condition in $5D_j\setminus D_j$. By the Harnack inequality and the Cauchy estimate we know that there is a constant $a_j\geq 0$ such that $$|h| \asymp a_j \textup{ and } |\nabla h| \lesssim a_j \textup{ in } 4D_j\setminus 2D_j.$$ So $$\int_{3D_j\setminus 2D_j} \textup{``cut-off terms"} \lesssim
a_j^2 \int_{3D_j\setminus 2D_j} e^{kx_1} \lesssim a_j^2 e^{-k/2} \int_{4D_j\setminus 3D_j} e^{kx_1} \lesssim$$ $$\lesssim e^{-k/2} \int_{4D_j\setminus 3D_j} h^2 e^{kx_1}.$$ Hence $$\textup{I} \lesssim e^{-k/2}\int_{B_{\frac{3}{4}R}\setminus \cup 3D_j} h^2 e^{kx_1}.$$ Note that $$\textup{RHS}= \frac{ck^2}{R^2} \int_{B_{\frac{3}{4}R}\setminus \cup 3D_j} h^2 e^{kx_1} > 2 \textup{I}$$ if $$k^2/{R^2} \gg e^{-k/2}.$$ We make the choice $$k = C \log R$$ and it yields $$\sup_{B_{R}} e^{kx_1} \leq e^{CR\log R}.$$ Since $$\textup{I}+ \textup{II} \geq \textup{RHS} \quad \textup{and} \quad I\leq \frac{1}{2}\textup{RHS},$$ we have $$\int_{B_{R-11}\setminus B_{(\frac{3R}{4}-10)}} \textup{``cut-off terms"} = \textup{II} \geq \frac{1}{2}\textup{RHS} \asymp k^2/{R^2} \int_{B_{\frac{3}{4}R}\setminus \cup 3D_j} h^2 e^{kx_1} \geq$$ $$\geq \exp(-CR\log R)\int_{B_{\frac{1}{2}R}\setminus \cup 3D_j} h^2.$$ If $5D_j \subset B_R$, then $\int_{3D_j\setminus 2D_j} h^2 \asymp \int_{4D_j\setminus 3D_j} h^2$, whence $$\int\limits_{B_{R}\setminus(B_{R/2} \bigcup (\cup 3D_j))} h^2
\geq c_1 \int\limits_{B_{R-10}\setminus(B_{R/2} \bigcup(\cup 2D_j))} h^2 \geq$$ (by Cauchy estimate) $$\geq c_2 \int\limits_{B_{R-11}\setminus(B_{R/2} \bigcup(\cup 3D_j))} \! \! \! (h^2 +|\nabla h|^2)$$ and therefore $$\sup_{B_{R}} e^{kx_1} \int\limits_{B_{R}\setminus(B_{R/2} \bigcup (\cup 3D_j))} h^2 \geq c_2 \int\limits_{B_{R-11}\setminus(B_{R/2} \bigcup(\cup 3D_j))} (h^2 +|\nabla h|^2) e^{kx_1} \geq c_3 \textup{II}.$$ Thus $$\int\limits_{B_{R}\setminus(B_{R/2} \bigcup( \cup 3D_j))} h^2 \geq \exp(-C'R\log R) \int\limits_{B_{R/2}\setminus \cup 3D_j} h^2.$$
**Deduction of Theorem \[exercise 1\] from Theorem \[local 3\]**. We may assume that $$\sup\limits_{B(0,R')\setminus \cup 3D_j}|h|=\sup\limits_{B(0,R'/8)\setminus \cup 3D_j}|h|,$$ otherwise the statement is trivial. Consider any point $x$ on $\partial B(0,R'/4)\setminus \cup 3D_j$.
Note that $ B(0,R'/8) \subset B(x,3R'/8)$ and $B(x,3R'/4) \subset B(0,R')$. Hence $$\sup\limits_{B(x,3R'/8)\setminus \cup 3D_j}|h|=\sup\limits_{B(x,3R'/4)\setminus \cup 3D_j}|h|.$$ Applying Theorem \[local 3\] for the disk with center at $x$ (in place of $0$) of radius $R=3R'/4$ and $N=0$, we obtain the bound $$\sup\limits_{\{R'/8<|z|< R' \}\setminus \cup 3D_j}|h|\geq \sup\limits_{B(x,R'/8)\setminus \cup 3D_j}|h| \geq e^{-CR'}\sup\limits_{B(0,R')\setminus \cup 3D_j}|h|.$$
Sketches of general elliptic theory.
------------------------------------
\[fact1\] Denote by $E(z)=
\frac{1}{2\pi} \log|z|$ the fundamental solution of the Laplace operator on the plane in the sense that for every $C^\infty$ compactly supported function $h$, we have $$h = E * \Delta h.$$
\[fact2\] Let $\Omega$ be a bounded open set and $g \in L^1(\Omega)$. Put $f= E* g$. Then
1. $f\in L^p(\Omega)$ for all $p \geq 1$.
2. $\Delta f =g $ in the sense that for every $h \in C_0^\infty(\Omega)$, we have $$\int_{\Omega} f \Delta h = \int_{\Omega} gh.$$
**Agreement**. Writing $E*g$ we assume that $g$ is extended to $\mathbb{R}^2\setminus \Omega$ by zero.
1. Let $D=\textup{diam }\Omega$. Let $E_D= \mathbbm{1}_{B(0,D)} E$. Then in $\Omega$ we have $f=g* E_D$. Since $E_D \in L^p(\mathbb{R}^2)$ for all $p\geq 1$ and $g\in L^1$, the result follows from Young’s convolution inequality.
2. We have $$\int f\Delta h = \int (E*g)\Delta h =\int \limits_{\Omega \times \Omega} E(z-\zeta)g(\zeta) \Delta h(z)dm_2(z)dm_2(\zeta)=$$ $$= \int\limits_{\Omega} g(\zeta) \left[ \int\limits_{\Omega} E(z-\zeta)\Delta h(z) dm_2(z) \right] dm_2(\zeta)= \int\limits_\Omega gh.$$
\[fact3\] Let $V \in L^\infty(\Omega)$, $u\in L^1_{loc}(\Omega)$ and $\Delta u+ Vu=0$ in $\Omega$ in the sense that for every $h\in C^{\infty}_{0}(\Omega)$ we have $$\int\limits_{\Omega}u\Delta h + \int\limits_{\Omega}Vuh=0.$$ Then $u \in L^p_{loc}(\Omega)$ for every $ p\geq 1$.
Passing to a smaller bounded domain $\Omega'$, we may assume that $u\in L^1(\Omega)$, $\Omega$ is bounded. Consider $f= E * (Vu)$. By Fact \[fact2\], $f \in L^p(\Omega)$ for all $p \geq 1$. Note that $u-f \in L^1(\Omega)$ and $\Delta(u-f)=0$ in the sense of distributions. Hence, by Weyl’s lemma, $u-f$ is harmonic in $\Omega$, so $u-f \in L^p_{loc}(\Omega)$ and therefore $u \in L^p_{loc}(\Omega) $ too.
\[fact4\] Let $g \in L^p(\Omega)$ with $p>2$ and let $\Omega$ be bounded. Then $g * E \in C^1(\Omega)$.
$$E(z+t)-E(z)= \frac{1}{2\pi}(\log|z+t|-\log|z|)= \frac{1}{2\pi} \log\left|1+\frac{t}{z}\right| =$$
$$= \frac{1}{2\pi}
\Re \frac{t}{z} + O\left(
\begin{cases}
\frac{|t|^2}{|z|^2}, \quad \frac{1}{2} \geq \frac{|t|}{|z|} \\
\frac{|t|}{|z|}+|\log \bigl|1+\frac{t}{z}|\bigr|,\quad \frac{1}{2} \leq \frac{|t|}{|z|}
\end{cases}
\right).$$ Define $$W(\zeta) =\begin{cases}
\frac{1}{|\zeta|^2}, \quad \zeta > 2 \\
\frac{1}{|\zeta|}+\bigl|\log|1+\frac{1}{\zeta}|\bigr|,\quad \zeta \leq 2.
\end{cases}$$
Taking the convolution and applying Holder’s inequality, we have
$$(g*E)(z+t)- (g*E)(z) = \left[ g* \frac{1}{2\pi}
\Re \frac{t}{\cdot} \right](z) + O\left(\|g\|_p \|W(\cdot/t)\|_q\right) =$$ $$= \Re\left(t \left[g*\frac{1}{2\pi \cdot} \right] (z)\right) + O\left(\|W(\cdot/t)\|_q\right), \quad \text{where } 1/p+1/q=1.$$ Since $g*\frac{1}{2\pi \cdot} \in C(\Omega)$, the first term (as a function of $t \in \mathbb{C}=\mathbb{R}^2$) is a linear operator from $\mathbb{R}^2$ to $\mathbb{R}$, which depends continuously on $z$. It is enough to show that $\|W\|_q< \infty$ because $$\|W(\cdot/t)\|_q = \|W\|_q |t|^{2/q} = \|W\|_q \,o(t) \quad \text{ as } t \to 0$$ ($1<q<2$ if $p>2$). Indeed, $$\int|W|^q \lesssim \int_{|\zeta|>2} \frac{1}{|\zeta|^{2q}} + \int_{|\zeta|\leq 2} \left( \frac{1}{|\zeta|^q}+ \left|\log\Bigl|1+\frac{1}{\zeta}\Bigr|\right|^{q} \right)< \infty.$$
\[fact5\] Let $V \in L^\infty(\Omega)$. If $\Delta u + Vu=0$ in $\Omega$ in the sense of distributions and $u \in L^1_{loc}(\Omega)$, then $u \in C^1(\Omega)$.
By Fact \[fact3\], $u \in L^p_{loc}(\Omega)$ with $p > 2$. Again passing to a subdomain, if necessary, we may assume that $\Omega$ is bounded and $u \in L^p(\Omega)$. Consider $f= E*(Vu)$. Since $Vu \in L^p(\Omega)$, $f \in C^1(\Omega)$. However $u-f$ is harmonic. Hence $u \in C^1(\Omega)$.
\[lem: solving\] Let $\Omega$ be a bounded domain with Poincare constant smaller than 1. Then for any $v\in L^\infty(\Omega)$, we can find a solution $u \in W^{1,2}_0(\Omega)$ to $\Delta u = v$ in the sense of distributions such that $$\|u\|_{W^{1,2}_0(\Omega)} \leq 4 \|v\|_2.$$
Note that if $u \in W^{1,2}_{loc}(\Omega)$ and $h \in C^\infty_0(\Omega)$, then $$\int_\Omega \nabla u \nabla h = -\int_\Omega u \Delta h.$$ So $u \in W^{1,2}_{loc}(\Omega)$ is a solution to $\Delta u = v$ in the sense of distributions if and only if $$\int_\Omega \nabla u \nabla h = -\int_\Omega v h$$ for any $h \in C^\infty_0(\Omega)$.
Consider the functional $$\Phi(u)= \int |\nabla u|^2 + \int vu$$ for $u \in W_{0}^{1,2}(\Omega)$. Integrals in the next few lines will be over the domain $\Omega$. Notice that by Poincare’s inequality
$$\Phi(u) \geq \frac{1}{2} \int |\nabla u|^2 + \frac{1}{2} \int |u|^2 - \|v\|_2\|u\|_2 \geq$$ $$\geq \frac{1}{2} \|u\|^2_{W_{0}^{1,2}} - \|v\|_2\|u \|_{W_{0}^{1,2}} \geq -\frac{1}{2}\|v\|_2.$$
Thus $\Phi(u)$ is bounded from below. Note that $\Phi(u)>0=\Phi(0)$ as soon as $\|u\|_{W_{0}^{1,2}} > 2\|v\|_2.$ Let now $u_k\in C_0^{\infty}(\Omega)$ be any minimizing sequence for $\Phi$. Note that
$$\frac{\Phi(u')+ \Phi(u'')}{2} - \Phi\left(\frac{u'+u''}{2}\right) = \frac{1}{4} \int |\nabla(u'-u'')|^2 \geq \frac{1}{8} \|u'- u''\|^2_{W_0^{1,2}}.$$
Hence $u_k$ is a Cauchy sequence in $W_0^{1,2}(\Omega)$, so the limit $u=\lim u_k$ exists and minimizes $\Phi$. Now, take any test function $h \in C_0^\infty(\Omega)$ and consider $$\Phi(u+th)= \Phi(u) + t \left(2\int \nabla u \nabla h + \int vh \right)
+t^2 \int |\nabla h|^2.$$ Since $u$ is a minimizer, we must have $$2\int \nabla u \nabla h + \int vh=0,$$ i.e., $\Delta u = v/2$ in the sense of distributions. Taking $2u$ in place of $u$ we get a solution to $\Delta u = v$ with $$\|u\|_{W_0^{1,2} } \leq 4 \|v\|_2.$$
The next step is to show that $$\| u\|_\infty \leq C\| v\|_\infty$$ with some absolute constant $C>0$. WLOG, we will assume $|v|\leq 1$.
Uniform bound via Di Giorgi method
----------------------------------
Let $\Omega$ be any bounded open set in $\mathbb{R}^2$ with Poincare constant $k^{2} \leq k_{0}^{2},$ i.e., $$\int u^{2} \leq k^{2} \int|\nabla u|^{2} \text { for all } u \in C_{0}^{\infty}(\Omega).$$ **Claim I**: Let $k_0$ be sufficiently small and consider any smooth $\varepsilon-$minimizer of $$\Phi(u)=\int_{\Omega}|\nabla u|^{2}+\int_{\Omega} vu \quad\left(\|v\|_{\infty} \leq 1\right)
, \text{ i.e.},$$ $u\in C^\infty_0(\Omega)$ and for any $\tilde u\in C^\infty_0(\Omega)$, $\Phi(\tilde u) \geq \Phi(u) -\varepsilon$. Then $u$ satisfies $$\int_{B\cap \Omega} u^{2} \leq Ck^2(k^2+\varepsilon)$$ for every unit ball $B \subset \mathbb{R}^{2}$.
Let $\varphi(x)$ be a smooth positive radial function such that
- $\varphi(x)=1$ in $B(0,1)$,
- $\varphi(x) \in (0,1]$,
- $\varphi(x) \asymp e^{-|x|}$,
- $|\nabla \varphi| \leq \varphi$.
Let $ \psi=\varphi^{2}, \text { so }|\nabla \psi| \leqslant 2|\psi|$. Applying the Poincare inequality to $\varphi u$, we get $$\int \varphi^{2} u^{2}\leq k^{2} \int|\varphi \nabla u+u \nabla \varphi|^{2}
\leq 2 k^{2}\left(\int \varphi^{2}|\nabla u|^{2}+\int u^{2}|\nabla \varphi|^{2}\right) \leq$$ $$\leq 2k^{2} \int \varphi^{2}|\nabla u|^{2}+2 k^{2} \int u^{2} \varphi^{2}, \text { whence }$$ $$\int \varphi^2 u^2 \leq \frac{2 k^{2}}{1-2 k^{2}} \int \varphi^2|\nabla u|^{2}, \text{ i.e.},$$ $$\label{eq: *}
\int u^{2} \psi \leq \frac{2 k^{2}}{1-2 k^{2}} \int |\nabla u|^{2} \psi \leq 4 k^{2} \int |\nabla u|^{2} \psi \quad \text{ if } \quad k_{0} \leq \frac{1}{2}.$$
Now, consider the competitor $\tilde{u}=(1-\psi) u$. We have $$\quad \Phi(\tilde{u})=\int|(1-\psi) \nabla u-u \nabla \psi|^{2}+\int v(1-\psi) u\leq$$ $$\leq \int(1-\psi)^{2}|\nabla u|^{2}+2 \int|\nabla u| | u||\nabla \psi|+\int u^{2}|\nabla \psi|^{2}+\int v(1-\psi) u$$ $$\leq \int(\nabla u)^{2}+\int v u-\int |\nabla u|^{2}\psi+4 \int |\nabla u| | u| \psi +4 \int u^{2} \psi-\int v \psi u$$ (we used the inequalities $(1-\psi)^{2} \leq 1-\psi, |\nabla \psi| \leqslant 2 \psi, \psi^{2} \leq \psi$).
Since $\Phi(\tilde{u}) \geqslant \Phi(u)-\varepsilon,$ we must have $$\int|\nabla u|^{2} \psi \leq 4\left(\int|\nabla u| |u| \psi+\int u^{2} \psi\right)+\varepsilon+\int | v \psi u |.$$ However $$\int u^{2} \psi \leqslant 4 k^{2} \int |\nabla u|^{2} \psi \quad \text { by \eqref{eq: *}, }$$ and $$\int|\nabla u| |u| \psi \leq \sqrt{\int u^{2} \psi} \sqrt{\int | \nabla u|^{2} \psi} \leqslant 2 k \int|\nabla u|^{2} \psi.$$ So for sufficiently small $k_0$, $$4\left(\int|\nabla u| |u| \psi+\int u^{2} \psi\right) \leq \frac{1}{2} \int|\nabla u|^{2} \psi.$$ Hence $$\int|\nabla u|^{2} \psi\leq 2\varepsilon+ 2\int|v u \psi |$$ $$\leq 2\left(\varepsilon+\sqrt{\int \psi} \sqrt{\int u^{2} \psi} \, \, \right) \leqslant C\left(\varepsilon+ k \cdot \sqrt{\int |\nabla u|^{2} \psi} \, \,\right).$$ If the first term dominates, then $\int|\nabla u|^{2} \psi \leqslant C \varepsilon $. Otherwise $ \int|\nabla u|^{2} \psi \leqslant C k^{2}$.
By it follows that $$\int u^{2} \psi \leq C k^{2}\left(k^{2}+\varepsilon\right).$$
Note that we did not care in Claim I where the Poincare constant came from and what was special about the geometry of $\Omega$ that made it small. The next lemma gives a simple bound for the Poincare constant of “thin" domains.
\[le: thin\] Assume $\Omega$ is open and $m_2(\Omega \cap Q)\leq c < 1$ for all unit squares $Q\subset \mathbb{R}^2$. Then the Poincare constant of $\Omega$ is at most $2 + \frac{2}{1-c}$.
Let $f \in C_0^{\infty}(\Omega)$. Extend $f$ by zero outside $\Omega$. It is sufficient to show that if $Q$ is a unit square, then $$\int_Q | f|^2 \leq \left(2 + \frac{2}{1-c}\right)\int_Q | \nabla f|^2.$$ By tiling the plane with unit squares, it implies $$\int_\Omega | f|^2 \leq \left(2 + \frac{2}{1-c}\right) \int_\Omega | \nabla f|^2.$$ Let $Q=[0,1]^2$, $$I_x=((x,y): y\in [0,1]) \quad \text{ and } \quad I^y=((x,y): x\in [0,1]).$$ Let $X_0$ be the set of $x\in [0,1]$ such that $I_x$ contains a zero of $f$. Then for every $x\in X_0$, we have $$\max_{I_x} |f| \leq \int_{I_x} |\nabla f|$$ and $$\int_{I_x} f^2 \leq \max_{I_x} f^2 \leq \left(\int_{I_x} |\nabla f|\right)^2 \leq \int_{I_x} |\nabla f|^2.$$ Hence $$\int_{X_0 \times [0,1]} f^2 \leq \int_Q |\nabla f|^2.$$ The set $X_0$ has Lebesgue measure at least $1-c$, whence there is $x_0 \in X_0$ such that $$\int_{I_{x_0}} f^2 \leq \frac{1}{1-c} \int_Q |\nabla f|^2.$$ **Claim.** Let $I$ be a unit interval and let $z$ be any point in $I$. Then $$\int_I f^2 \leq 2|f(z)|^2 + 2 \int_I |\nabla f|^2.$$ Indeed, $$\int_I f^2 \leq \max_I f^2 \leq \left(|f(z)| + \int_I |\nabla f|\right)^2 \leq \left(|f(z)| + \sqrt{\int_I |\nabla f|^2}\,\,\right)^2 \leq$$ $$\leq 2|f(z)|^2 + 2\int_I |\nabla f|^2.$$ For every $y \in [0,1]$, it yields $$\int_{I^y} f^2 \leq 2|f(x_0,y)|^2 + 2 \int_{I^y} |\nabla f|^2.$$ Thus $$\int_Q f^2 = \int_0^1 \left(\int_{I^y} f^2dx\right)dy \leq 2\int_{I_{x_0}} f^2+2 \int_{Q} |\nabla f|^2 \leq \left(\frac{2}{1-c}+2\right)\int_{Q} |\nabla f|^2.$$
\[cor: thin\] If $$m_2(\Omega \cap Q)\leq k^2 \ll 1$$ for any unit square $Q$, then the Poincare consant of $\Omega$ is smaller than $Ck^2$.
For every square $Q_{2k}$ of size $2k$, we have $$m_2(\Omega \cap Q_{2k})\leq \frac{1}{4} m_2( Q_{2k}).$$ By $2k$ rescaling we reduce the problem to Lemma \[le: thin\].
Now we are almost ready to run the Di Georgi scheme. The only remaining preparatory part is smooth surgery. Let $u\in C_0^\infty(\Omega)$. Fix $t>0$ (level) and $\delta>0 $ (extremely small number). Let $\Theta$ be a $C^\infty$-smooth function on $\mathbb{R}$ described by Figure \[Theta\].
![ $y=\Theta(x)$[]{data-label="Theta"}](theta.pdf){width="70.00000%"}
The function $\Theta$ has the following properties:
- $ 0 \leq \Theta' \leq 1,$
- $ \Theta(0)= 0,$
- $ x-4\delta\leq \Theta(x) \leq x,$
- $\Theta(x) = t-2\delta \text{ on } (t -\delta, t+\delta).$
Let $\tilde u = \Theta \circ u$. Then $|u-\tilde u| \leq 4\delta$ and $|\nabla \tilde u| \leq |\nabla u|$ pointwise. Thus, if $u$ was an $\varepsilon$-minimizer, then $\tilde u$ is an $\varepsilon+ A\delta$-minimizer ($\delta$ is purely qualitative and $A=4\int_\Omega |v|$). Define $$\Theta_-(x)=
\begin{cases}
\Theta(x), \quad x\leq t+\delta\\
t-2\delta, \quad x\geq t+\delta
\end{cases} \quad \text{and} \quad \Theta_+=\Theta-\Theta_-.$$ The function $\tilde u$ naturally splits into two smooth compactly supported terms: $$\tilde u= \tilde u_-+ \tilde u_+, \text{ where } \tilde u_\pm = \Theta_\pm \circ u.$$ The function $\tilde u_+$ is compactly supported in $\{u>t\}$ ($\text{supp } \tilde u_+ \subset\{u \geq t+\delta \}$) and $\nabla \tilde u_+$ and $\nabla \tilde u_-$ have disjoint supports. We may then try to replace $\tilde u_+$ by some smooth competitor $w\in C_0^\infty(\{u>t\})$ and see if the functional can drop. Note that $$\Phi(\tilde u_-+w)= \int_\Omega|\nabla \tilde u_-|^2 + \int_{\{u>t\}}|\nabla w|^2 + \int_\Omega v\tilde u_- + \int_{\{u>t\}} v w,$$ so we just need to compare $$\int_\Omega|\nabla \tilde u_+|^2 + \int_\Omega v\tilde u_+ \quad \text{with} \quad \int_{\{u>t\}}|\nabla w|^2 + \int_{\{u>t\}} v w.$$ Hence $ \tilde u_+$ is an $(\varepsilon+A\delta)$-minimizer in the new domain $\{u> t \}$. We shall now fix the initial Poincare constant to be $k_0$ from Claim I. If $u$ is an $\varepsilon$-minimizer, then $ \int_B u^2 \leq Ck_0^2(k_0^2+\varepsilon)$ for every unit ball $B$, so $$m_2(\{u>t_0\} \cap B)\leq \frac{Ck_0^2(k_0^2+\varepsilon)}{t_0^2}.$$
Choose $t_{0}=C' \sqrt{k_{0}}$ with sufficiently large absolute constant $C'$. Then the domain $\Omega_{1}=\left\{u>t_{0}\right\}$ satisfies $m_2(\Omega_1\cap B) \leq \frac{C}{C'} k_{0}\left(k_{0}^{2}+\varepsilon\right)$ for any unit ball $B$ and, by Corollary \[cor: thin\], the Poincare constant of $\Omega_1$ is at most $k^{2}_1:=k_{0} \frac{k_{0}^{2}+\varepsilon}{2}$. Also $u_{1}=\tilde{u}_{+} \in C_{0}^{\infty}\left(\Omega_{1}\right)$ will be an $\varepsilon_{1}=\varepsilon+A\delta_{0}$-minimizer of $\Phi$ in $C_{0}^{\infty}\left(\Omega_{1}\right)$ where $\delta_{0}>0$ can be chosen arbitrary small. Finally, note that $u \leq t_{0}+u_{1}+4\delta_{0}$ everywhere in $\Omega$. We can now repeat this construction with $ u_{1}, \Omega_{1}, \varepsilon_1$ instead of $u, \Omega, \varepsilon$ to get $u_{2}, \Omega_{2}, \varepsilon_{2}$ and so on. We shall get a sequence of domains $\Omega_j$, functions $u_j \in C_0^{\infty}(\Omega_j)$ and numbers $k_j,t_j,\varepsilon_j, \delta_j>0$ such that $$\Omega_1 \supset \Omega_2 \supset ... \quad, \quad k_j^2= k_{j-1}\frac{k_{j-1}^2+\varepsilon_{j-1}}{2},$$ $$\varepsilon_j= \varepsilon_{j-1}+ A\delta_{j-1}, \quad t_j=C'\sqrt{k_j},$$ $$m_2(\Omega_j\cap B) \leq \frac{C}{C'}k_{j-1}(k_{j-1}^2 +\varepsilon_{j-1}) \quad \text{ for any unit ball } B,$$ and $$u \leq t_0+t_1+...+t_{l}+u_{l+1}+4\delta_0+4\delta_1+...+4\delta_{l} \quad \text{for any } l\geq 0.$$
In this construction, we can choose $\delta_j>0$ as small as we want, so putting $\delta_j=\frac{c\varepsilon}{2^j}$ with sufficiently small absolute constant $c>0$, we can guarantee that all $\varepsilon_j \leq 2 \varepsilon$.
Let $l$ be the first index for which $k_l^2<2\varepsilon$. Then (provided that $C' > 8C$, $\varepsilon <1/2$), we also have $m_2(\Omega_{l+1} \cap B)<\frac{\varepsilon}{2}$ for every unit ball $B$. For all $j\leq l$, we have $$k_j^2 \leq k_{j-1}^3,$$ so, if $k_0$ was chosen less than $\frac{1}{4}$, it implies that $k_j^2 \leq 2^{-j-2}$ for $j=0,1,...,l$, whence $t_j \leq C'2^{-\frac{j}{2}-1}$ and $$u \leq t_0+t_1+...+t_{l}+4\delta_0+4\delta_1+...+4\delta_{l} \leq C' \sum_{j\geq 0} 2^{-\frac{j}{2}-1}+\sum_{j\geq 0} 4\delta_j \leq$$ $$\leq 2C' +\varepsilon \leq 2C'+1=C_0$$ in $\Omega\setminus \Omega_{l+1}$ because $u_{l+1}=0$ in $\Omega\setminus \Omega_{l+1}$. Thus $$m_2(\{ u>C_0\}\cap B)<\varepsilon/2.$$ Considering $-u$ instead of $u$, we conclude that also $m_2(\{ u<-C_0\}\cap B)<\varepsilon/2$ and therefore $$m_2(\{ |u|>C_0\}\cap B)<\varepsilon$$ for every $\varepsilon$-minimizer $u$. Since the true minimizer $u$ is the limit of $\varepsilon$-minimizers in $L^2(\Omega)$, we get $$m_2(\{|u|>C_0\}\cap B)=0 \quad \text{ for any unit ball } B,$$ so $m_2(\{|u|>C_0\})=0$.
**Conclusion.** If $|v|\leq 1$ and the Poincare constant of $\Omega$ is not greater than $k_0^2\ll 1$, then the minimizer of $\int_\Omega |\nabla u|^2 + \int_\Omega vu$ in $W_0^{1,2}(\Omega)$ satisfies $$\|u\|_\infty\leq C_0.$$
If the Poincare constant of $\Omega$ is $k$, we put $\widetilde \Omega = \frac{k_0}{k}\Omega$, $\widetilde u = \frac{k_0^2}{k^2}u(\frac{k}{k_0}\cdot)$, $\widetilde v = v(\frac{k}{k_0}\cdot)$, so $\widetilde \Phi(\widetilde u)= \int_{\widetilde \Omega} |\nabla \widetilde u |^2+ \int_{\widetilde \Omega} \widetilde v \widetilde u = \frac{k_0^4}{k^4}\Phi(u)$. Applying the result that was just obtained, we get the final observation.
\[lem: solving2\] If the Poincare constant of $\Omega$ is $k>0$, then the minimizer of $\Phi(u)=\int_\Omega |\nabla u|^2 + \int_\Omega vu$ in $W_0^{1,2}(\Omega)$ (i.e., the solution to $\Delta u = v/2$) satisfies: $$\label{eq: uniform}
\|u\|_\infty\leq Ck^2\|v\|_\infty,$$ where $C$ is an absolute positive constant.
Other standard facts used in the proof.
---------------------------------------
\[trivialities 1\] Let $\Omega$ be a bounded open set, let $u \in W^{1,2}_{0}(\Omega) $ satisfy $\| u\|_\infty \leq 1$. Then there exists a sequence $u_k
\in C^\infty_0(\Omega)$ with $\|u_k\|_\infty \leq 2$ such that $ u_k \to u, \nabla u_k \to \nabla u $ in $L^2(\Omega)$ and almost everywhere in $\Omega$.
By the definition of $W^{1,2}_0(\Omega)$, we can find $\tilde u_k
\in C^\infty_0(\Omega) $ with $ \tilde u_k \to u, \nabla \tilde u_k \to \nabla u $ in $L^2$. Let $\Theta $ be defined by Figure \[Theta2\].
![The graph of $\Theta$[]{data-label="Theta2"}](theta1.pdf){width="70.00000%"}
The function $\Theta$ has the following properties.
- $\Theta(x)=x$ for $x\in [-1.5,1.5]$,
- $ \Theta $ is $C^\infty$-smooth and $|\Theta| \leq 1.75 $,
- $|\Theta'|\leq 1$ and $|\Theta(x)| \leq |x|$.
Put $u_k = \Theta(\tilde u_k)$. Note that $u_k= \tilde u_k$ and $\nabla u_k = \nabla \tilde u_k$ if $|\tilde u_k|\leq 1.5$ and we always have $|u_k|\leq |\tilde u_k|$, $|\nabla u_k|\leq |\nabla \tilde u_k|$. We need to show that $\|u_k-u\|_2$, $\|\nabla u_k- \nabla u\|_2 \to 0$. Since $\tilde u_k$ converge in $W^{1,2}_0(\Omega)$, the functions $|\tilde u_k|^2$ and $|\nabla \tilde u_k|^2$ are uniformly integrable, i.e., for any $\varepsilon>0$, there is $\delta>0$ such that if $m_2(E)<\delta$, then $$\int_E |\tilde u_k|^2, \int_E |\nabla \tilde u_k|^2 <
\varepsilon.$$ In the following computation $\int$ will denote the integral over $\Omega$: $$\int |u_k-u|^2 \leq 2 \left[ \int |u_k-\tilde u_k|^2 +\int |\tilde u_k - u|^2 \right]\leq$$
$$\leq 4 \!\!\!\!\!\!\int\limits_{\{|\tilde u_k|>1.5\}} (|u_k|^2+|\tilde u_k|^2) + 2\int |\tilde u_k - u|^2 \leq$$
$$\leq 8\!\!\!\!\!\!\int\limits_{\{|\tilde u_k|>1.5\}} |\tilde u_k|^2 + 2 \int|\tilde u_k - u|^2.$$ The second term tends to zero by the choice of $\tilde u_k$. Note that $$m_2(\{|\tilde u_k|>1.5\}) \leq m_2(|u-\tilde u_k|\geq 0.5) \leq 4 \|u-\tilde u_k\|_2^2 \to 0.$$ So the first term tends to zero by the uniform integrability property. In a similar way one can show that $\int|\nabla u_k - \nabla u|^2 \to 0 $.
Now we would like to choose a subsequence such that $ u_k \to u, \nabla u_k \to \nabla u $ almost everywhere in $\Omega$. It can be done by choosing any subsequence $u_{k_j}$ with $$\|u_{k_j}-u\|_2 \leq \frac{1}{4^j}, \quad
\|\nabla u_{k_j}- \nabla u\|_2 \leq \frac{1}{4^j}.$$ Let $E_j$ be the set of points $x\in \Omega$ such that $|u_{k_j}(x)-u(x)| \geq \frac{1}{2^{j}}$. Then $$\sqrt{m_2(E_j) \frac{1}{4^{j}} } \leq \|u_{k_j}-u\|_2 \leq \frac{1}{4^{j}}, \quad {\text{so}} \quad m_2(E_j) \leq \frac{1}{4^j}.$$ Note that if $x \notin \cup_{j=n}^\infty E_j$, then $u_{k_j}(x)$ converge to $u(x)$. However $$m_2(\cup_{j=n}^\infty E_j) \leq \sum_{j=n}^\infty m_2( E_j) \leq \frac{1}{2^{n}}.$$ Thus $u_{k_j}$ converge to $u$ almost everywhere in $\Omega$. In a similar way one can show that $\nabla u_{k_j}$ also converge to $\nabla u$ almost everywhere.
**Fact \[product\].** Let $\Omega$ be an open set. Assume that $u,v \in W^{1,2}_{loc}(
\Omega)\cap L^{{}^{\scriptsize \infty}}_{loc}(\Omega)$. then $uv \in W^{1,2}_{loc}(
\Omega) $ and $\nabla(uv) = u \nabla v + v \nabla u$.
Clearly, the fact is local. So we may assume $\Omega= B(0,r)$ and $u,v \in W^{1,2}(B(0,r)
)\cap L^\infty(B(0,r))$.
Let us fix a small $\delta>0$ and let $$K_\varepsilon(z)=\frac{1}{\varepsilon^2}\kappa(|z|/\varepsilon)$$ be a $C^\infty$-approximation to identity with $\text{supp} K_\varepsilon(z)\subset B(0,\varepsilon)$, $\varepsilon < \delta$. Then $K_\varepsilon* u$ and $K_\varepsilon*v$ converge to $u$ and $v$ in $L^2(B(0,r-\delta))$ and a.e. in $B(0,r-\delta)$ as $\varepsilon \to 0$. Consider any test function $\eta \in C_0^\infty(B(0,r-\delta))$ and extend it by $0$ outside $B(0,r-\delta)$. Then $\eta*K_\varepsilon \in C_0^\infty(B(0,r))$ and $K_\varepsilon * \nabla \eta = \nabla (K_\varepsilon*\eta) $. By Fubini’s theorem we have $$\int_{B(0,r-\delta)} (K_\varepsilon* u) \nabla \eta =\int_{B(0,r)} u (K_\varepsilon* \nabla \eta)=$$ $$=\int_{B(0,r)} u \nabla(K_\varepsilon* \eta)= -\int_{B(0,r)} \nabla u (K_\varepsilon* \eta)=$$ $$=- \int_{B(0,r)}\left(\int_{B(0,r-\delta)} \nabla u(x) K_\varepsilon(x-y)\eta(y) dy \right) dx=$$ $$= - \int_{B(0,r-\delta)}\left(\int_{B(0,r)} \nabla u(x) K_\varepsilon(y-x)\eta(y) dx \right) dy = -\int_{B(0,r-\delta)} (\nabla u *K_\varepsilon) \eta.$$
So $$u_\varepsilon:=K_\varepsilon* u \quad \textup{and} \quad v_\varepsilon:=\quad K_\varepsilon* v$$ are in $W^{1,2}(B(0,r-\delta))\cap C^\infty(B(0,r-\delta))$ with
1. $\nabla u_\varepsilon = \nabla u * K_\varepsilon$,$\nabla v_\varepsilon = \nabla v * K_\varepsilon$,
2. $\nabla u * K_\varepsilon \to \nabla u$, $\nabla v * K_\varepsilon \to \nabla v$ in $L^2(B(0,r-\delta))$ and a.e. in $B(0,r-\delta)$ as $\varepsilon \to 0$,
3. By Young’s inequality for convolutions, we have $|u_\varepsilon|< \|u\|_{L^\infty(B(0,r))}$, $|v_\varepsilon|< \|v\|_{L^\infty(B(0,r))}$ in $B(0,r-\delta)$.
We know that the convergence of $u_\varepsilon v_\varepsilon$ holds a.e. in $B(0,r-\delta)$ and $|u_\varepsilon v_\varepsilon|$ are bounded by $\|u\|_\infty\|v\|_\infty$ a.e. in $B(0,r-\delta)$ if $\varepsilon< \delta$. So by the Lebesgue dominated convergence theorem $u_\varepsilon v_\varepsilon \to uv$ in $L^2(B(0,r-\delta))$.
We want to show that $$\nabla(uv) = u \nabla v + v \nabla u$$ in the sense of $W^{1,2}(B(0,r-\delta))$. It is clear that $u \nabla v + v \nabla u$ is in $L^2(B(0,r-\delta))$ because $u,v$ are bounded and their gradients are in $L^2(B(0,r-\delta))$.
Consider again a test function $\eta \in C_0^\infty(B(0,r-\delta))$. We have $$\int uv \nabla \eta = \lim_{\varepsilon \to 0} \int u_\varepsilon v_\varepsilon \nabla \eta = - \lim_{\varepsilon \to 0} \int (\nabla u_\varepsilon v_\varepsilon + u_\varepsilon \nabla v_\varepsilon ) \eta=$$ $$=- \lim_{\varepsilon \to 0} \left[ \int \left((\nabla u_\varepsilon - \nabla u) v_\varepsilon + u_\varepsilon (\nabla v_\varepsilon - \nabla v) \right) \eta + \int ( \nabla u v_\varepsilon + u_\varepsilon \nabla v ) \eta \right].$$ Note that $\int (\nabla u_\varepsilon - \nabla u) v_\varepsilon \eta \to 0$ because $\nabla u_\varepsilon\to \nabla u$ in $L^2(B(0,r-\delta))$ and $|v_\varepsilon \eta | < \|v\|_\infty\| \eta\|_\infty$ in $B(0,r-\delta)$. Similarly, $ \int u_\varepsilon (\nabla v_\varepsilon - \nabla v) \eta \to 0$. Finally, by the Lebesgue dominated convergence theorem $$\int ( \nabla u v_\varepsilon + u_\varepsilon \nabla v ) \eta \to \int ( \nabla u v + u \nabla v ) \eta$$ because the convergence of the functions holds a.e. in $B(0,r-\delta)$ and there is the integrable majorant $(|\nabla u| \|v\|_\infty + \|u\|_\infty |\nabla v |) \|\eta\|_\infty$. Thus $\int uv \nabla \eta = - \int ( \nabla u v + u \nabla v ) \eta$.
\[fact6\] Let $\Omega$ be a bounded open set. Let $\varphi=1+\psi$, where $\psi \in W^{1,2}_0(\Omega)$, $\|\psi\|_\infty \leq \frac{1}{3}$. Then the functions $$\tilde \varphi =
\begin{cases}
\varphi \textup{ in } \Omega \\
1 \textup{ outside } \Omega
\end{cases}
\quad \textup{ and } \quad
\eta=
\begin{cases}
\frac 1 \varphi \textup{ in } \Omega \\
1 \textup{ outside } \Omega
\end{cases}$$ are in $W^{1,2}_{loc}(\mathbb{R}^2)$ and $$\nabla \tilde \varphi= \nabla \varphi \mathbbm{1}_\Omega, \quad \nabla \eta = - \frac{\nabla \varphi}{\varphi^2}\mathbbm{1}_\Omega.$$
Consider a sequence of functions $\psi_k \in C_0^\infty(\Omega)$ such that $\| \psi_k\|_\infty \leq \frac 2 3 $ and $\psi_k \to \psi$, $\nabla \psi_k \to \nabla \psi$ in $L^2(\Omega)$ and a.e. in $\Omega$. We can extend $\psi_k$ by zero outside $\Omega$ and get a sequence of $C_0^\infty(\mathbb{R}^2)$ functions, which we will still denote by $\psi_k$, such that $
\psi_k=0, \nabla \psi_k=0$ in $\mathbb{R}^2\setminus \Omega$ while $\psi_k \to \psi \mathbbm{1}_\Omega$ in $L^2(\mathbb{R}^2)$ and a.e., $\nabla \psi_k \to \nabla \psi \mathbbm{1}_\Omega$ in $L^2(\mathbb{R}^2)$. This immediately implies that $1+\psi_k \to \tilde \varphi$, $\nabla (1+\psi_k)=\nabla \psi_k \to \nabla \psi \mathbbm{1}_\Omega$ in $L^2_{loc}(\mathbb{R}^2)$, so $\tilde \varphi \in W^{1,2}_{loc}(\mathbb{R}^2)$ and $\nabla \tilde \varphi= \tilde \varphi \mathbbm{1}_\Omega.$ Note that $$\left| \frac{1}{1+\psi_k}- \frac{1}{1+\psi}\right|= \left| \frac{\psi_k-\psi}{(1+\psi_k)(1+\psi)}\right| \leq 9|\psi_k - \psi|$$ and $$\left| \nabla \frac{1}{1+\psi_k}+ \frac{\nabla \psi}{(1+\psi)^2}\right|= \left| \frac{\nabla \psi_k}{(1+\psi_k)^2} - \frac{\nabla \psi}{(1+\psi)^2}\right|\leq$$ $$\leq |\nabla \psi_k -\nabla \psi| \frac{1}{(1+\psi_k)^2}+|\nabla \psi|\left| \frac{1}{(1+\psi_k)^2} - \frac{1}{(1+\psi)^2}\right| \quad \textup{ in } \Omega.$$ Also $\frac{1}{1+\psi_k}=1$, $\nabla\frac{1}{1+\psi_k}=0$ in $\mathbb{R}^2\setminus \Omega$. Hence $\frac{1}{1+\psi_k} \to \eta$ in $L^2(\mathbb{R}^2)$ and we would like to show that $$\nabla\frac{1}{1+\psi_k} \to - \frac{\nabla \psi}{(1+\psi)^2} \mathbbm{1}_\Omega= - \frac{\nabla \varphi}{\varphi^2}\mathbbm{1}_\Omega \quad \textup{ in } L^2(\mathbb{R}^2).$$ To see the latter, note that $\frac{1}{(1+\psi_k)^2} \leq 9$, so $$\int_\Omega |\nabla \psi_k - \nabla \psi|^2 \frac{1}{(1+\psi_k)^4} \leq 81 \int_\Omega |\nabla \psi_k - \nabla \psi|^2 \to 0,$$ and the functions $|\nabla \psi|^2 \left[ \frac{1}{(1+\psi_k)^2} - \frac{1}{(1+\psi)^2} \right]^2$ have the integrable majorant $81|\nabla \psi|^2$ and tend to $0$ almost everywhere in $\Omega$. Thus $\eta \in W^{1,2}_{loc}(\mathbb{R}^2)$ and $\nabla \eta = - \frac{\nabla \varphi}{\varphi^2}\mathbbm{1}_\Omega$ as required.
\[le: W0\] Let $\Omega$ be a bounded open set and let a function $f\in C^1(\overline{\Omega})$ be zero on $\partial \Omega$. Then $f \in W^{1,2}_{0}(\Omega)$.
Let $\varepsilon>0$, denote by $\Omega_\varepsilon$ the set of points $x$ in $\Omega$ with distance to the boundary of $\Omega$ at least $\varepsilon$. Let $\eta$ be a function in $C_0^\infty(\Omega)$ with the following properties:
- $\eta(x)= 1$, if $x \in \Omega_\varepsilon$.
- $0 \leq \eta \leq 1$ and $|\nabla \eta| \leq \frac{C}{\varepsilon}$ in $\Omega$.
The function $f\eta$ is in $C_0^1(\Omega) \subset W_0^{1,2}(\Omega) $. We want to show that $f\eta$ converge to $f$ in $W_0^{1,2}(\Omega)$ norm as $\varepsilon \to 0$. Observe that $|\nabla f|$ is uniformly bounded in $\Omega$ by some constant $A=A(f)$, so $|f(x)| \leq A\varepsilon $ if the distance from $x$ to $\partial \Omega$ is smaller than $\varepsilon$ and therefore $$|\nabla(f\eta)|\leq |\nabla f||\eta|+ | f|| \nabla \eta| \leq A + AC \quad \text{ in } \Omega\setminus \Omega_\varepsilon.$$ Then $$\int\limits_\Omega |f -f \eta|^2 = \int\limits_{\Omega \setminus \Omega_\varepsilon } |f -f \eta|^2 \leq \int\limits_{\Omega \setminus \Omega_\varepsilon } |f|^2 \to 0$$ and $$\int\limits_\Omega |\nabla f - \nabla (f \eta)|^2 = \int\limits_{\Omega \setminus \Omega_\varepsilon } |\nabla f -\nabla(f \eta)|^2 \leq 2 \int\limits_{\Omega \setminus \Omega_\varepsilon } (|\nabla f|^2 +|\nabla(f \eta)|^2) \leq$$ $$\leq m_2(\Omega \setminus \Omega_\varepsilon) C_1 A^2.$$ Since $m_2(\Omega \setminus \Omega_\varepsilon) \to 0$ as $\varepsilon \to 0$, we have verified that $f \in W_0^{1,2}(\Omega)$.
\[le: diameter\] Let $u$ be a solution to $\Delta u + Vu=0$, $|V|\leq 1$, in a ball $B(x,r)$, where $r<r_0$ and $r_0$ is a sufficiently small universal constant. If $u$ is continuous up to $\partial B(x,r)$ and $u>0$ on $\partial B(x,r)$, then $u>0$ in $B(x,r)$.
We may assume that $u$ is larger than a positive constant $\delta$ on $\partial B(x,r)$. Consider the set $\Omega=\{ x \in B(x,r): u(x)< \frac \delta 2 \}$. This is an open set strictly inside $B(x,r)$ and if $u$ is not positive in $B(x,r)$, then $\Omega$ is not empty.
Since $u \in C^1(\overline{\Omega})$ by Fact \[fact5\] and $u=\frac \delta 2$ on $\partial \Omega$, we know by Lemma \[le: W0\] that $(u-\frac \delta 2) \in W_0^{1,2}(\Omega)$.
Note that $\overline{\Omega} \subset B(0,r) $, so $\Omega$ has a Poincare constant smaller than $Cr^2$. By Lemma \[solving Schrodinger\], if $r$ is sufficiently small, we can find $\varphi=1+\tilde \varphi$ with $\tilde \varphi \in W^{1,2}_0(\Omega), \|\tilde \varphi\|_\infty<\frac{1}{2}$ such that $\varphi$ is a solution to $\Delta \varphi + V \varphi=0$ in $\Omega$. Then $ (\frac{\delta}{2}\varphi - \frac{\delta}{2}) \in W^{1,2}_0(\Omega)$ and therefore the function $g=(\frac{\delta}{2}\varphi - u) \in W^{1,2}_0(\Omega) $. The function $g$ is also a solution to $\Delta g+ Vg=0$. For any $\eta\in C_0^\infty(\Omega)$, we have $\int_\Omega \nabla g \nabla \eta= \int_\Omega Vg\eta$ and taking the limit in $W^{1,2}_0(\Omega)$, we get $$\int_\Omega |\nabla g|^2 = \int_\Omega Vg^2 \leq \int_\Omega g^2.$$ However Poincare’s inequality implies $$\int_\Omega g^2 \leq Cr^2 \int_\Omega |\nabla g|^2.$$ If $r$ is sufficiently small, this could happen only if $g=0$ in $\Omega$. So $u= \frac{\delta}{2}\varphi$ in $\Omega$, but $\varphi> \frac 1 2$ in $\Omega$. So $u>\frac \delta 4$ in $\Omega$ and in $B(x,r)$.
Divergence free vector fields on the plane {#sec: divergence}
------------------------------------------
If $F=(F_1,F_2): B \to \mathbb{R}^2$ is a $C^1$- smooth vector field in a disk $B$ on the plane such that $F$ is divergence free: $\text{div} F=0$ in $B$, then there is a smooth function $u$ such that $$(F_1,F_2)= \nabla \times u := (u_{x_2},-u_{x_1}).$$
Sometimes people refer to the statement above as to Poincare’s lemma or the fundamental theorem of calculus, or the inverse gradient theorem. Here is the sketch of the standard proof. WLOG, $B=B(0,1)$. Consider any point $Q\in B$ and the rectangle $R \subset B$ with opposite vertices $0$ and $Q$, and sides parallel to $x_1$ and $x_2$ axes.
{width="50.00000%"}
Note that that the contour integral $$\int_{\partial R} (-F_2,F_1)\cdot dx= \int_{\partial R} F\cdot n(x) |dx| = \int_R \text{div } F$$ is zero. There are two simple paths that start at $0$, go along the sides of $R$ and end at $Q$. The integrals $\int(-F_2,F_1)dx$ over those two paths are the same and we define $u(Q)$ to be equal to both of them. The differentiation of $u(Q)$ in the horizontal and vertical directions shows that $(F_1,F_2)= (u_{x_2},-u_{x_1})$.
We need the version with less regularity assumptions on the divergence free vector field $F$: if $F\in L_{\text{loc}}^p(B(0,1))$, $1\leq p<\infty$, and $\int_{B(0,1)}F\nabla h=0$ for any $h \in C^\infty_0(B(0,1))$ (the divergence free condition), then there is a function $u\in W^{1,p}_{loc}(B(0,1))$ such that $(F_1,F_2)= \nabla \times u := (u_{x_2},-u_{x_1}).$
Indeed, let $$K_\varepsilon(z)=\frac{1}{\varepsilon^2}\kappa(|z|/\varepsilon)$$ be a $C^\infty$-approximation to identity with $\text{supp} K_\varepsilon(z)\subset B(0,\varepsilon)$. Define $$F_\varepsilon= F*K_\varepsilon= (F_1*K_\varepsilon, F_2*K_\varepsilon)$$ in the smaller ball $B(0,1-\varepsilon)$. Then $F_\varepsilon$ is divergence free in $B(0,1-\varepsilon)$. Indeed, if $f\in C_0^\infty(B(0,1-\varepsilon)$, then by Fubini’s theorem $$\int (F*K_\varepsilon) \nabla f= \int F (K_\varepsilon * \nabla f)=\int F \nabla (K_\varepsilon * f)=0.$$ So there is a $C^\infty$ function $u_\varepsilon$ such that $F_\varepsilon=\nabla \times u_\varepsilon$ in $B(0,1-\varepsilon)$. Fix $\delta \in (0,1)$. By the Lebesgue theory $F*K_\varepsilon$ converge to $F$ in $L^p(B(0,1-\delta))$. Thus $\nabla u_\varepsilon$ is a Cauchy sequence in $L^p(B(0,1-\delta))$. Let us add a constant to $u_\varepsilon$ so that $\int_{B(0,1-\delta)}u_\varepsilon=0$. By the Poincaré–Wirtinger inequality (see p.275, Theorem 1 in [@P]) $u_\varepsilon$ is a Cauchy sequence in $L^p(B(0,1-\delta))$. Thus we can find a function $\tilde u_\delta$ such that $u_\varepsilon$ converge to $\tilde u_\delta$ in $W^{1,p}(B(0,1-\delta))$ and $\nabla \times \tilde u_\delta = \lim_{\varepsilon \to 0} \nabla \times u_\varepsilon=(F_1,F_2)$. For any $\delta_1,\delta_2 \in (0,1)$, the gradients of $\tilde u_{\delta_1}$ and $\tilde u_{\delta_2}$ are the same in $B(0,1-\max(\delta_1,\delta_2))$ and therefore $\tilde u_{\delta_1} - \tilde u_{\delta_2}$ is constant almost everywhere in $B(0,1-\max(\delta_1,\delta_2))$. Finally, let us modify $\tilde u_\delta $ by subtracting a constant so that $ \int_{B(0,1/2)} \tilde u_\delta = 0 $ for all $\delta <1/2$. Then $u$ is well-defined by $u=\tilde u_\delta$ in $B(0,1-\delta)$.
[ZZZZ]{} K. Astala, T. Iwaniec, M. Gaven, Elliptic partial differential equations and quasiconformal mappings in the plane (PMS-48), Princeton; Oxford: Princeton University Press, 2009. L. Ahlfors, Lectures on quasiconformal mappings, Van Nostrand Co., Toronto, 1966. L. Bakri, Quantitative uniqueness for Schrödinger operator, Indiana University Mathematics Journal, 61(4), 2012, 1565–1580. J. Bourgain, C. E. Kenig, On localization in the continuous Anderson-Bernoulli model in higher dimension, Inventiones mathematicae, 161(2), 2005, 389–426. L. A. Caffarelli, A. Friedman, Partial regularity of the zero-set of solutions of linear and superlinear elliptic equations, Journal of Differential Equations, Volume 60, Issue 3, 1985, 420–433.
B. Davey, J. Zhu, Quantitative uniqueness of solutions to second-order elliptic equations with singular lower order terms, Communications in Partial Differential Equations, 44:11, 2019, 1217–1251. B. Davey, C. Kenig, J.-N. Wang, On Landis’ conjecture in the plane when the potential has an exponentially decaying negative part, Algebra i Analiz, 31:2, 2019, 204–226. H. Donnelly, C. Fefferman, Nodal sets of eigenfunctions on Riemannian manifolds, Invent. Math. 93, 1988, 161–183. L. Escauriaza, C. E. Kenig, G. Ponce, L. Vega, Uniqueness properties of solutions to Schrödinger equations. Bull. Amer. Math. Soc. (N.S.) 49, no. 3, 2012, 415–442. A. L. Gusarov, Liouville theorems for elliptic equations in a cylinder (Russian), Trudy Moskov. Mat. Obshch. 42, 1981, 254–266. L. C. Evans, Partial differential equations, Grad. Stud. Math., vol. 19, Amer. Math. Soc., Providence, RI, 1998. C. E. Kenig, Some recent quantitative unique continuation theorems, Séminaire Équations aux dérivées partielles, 2005–2006, 1–10. C. E. Kenig, L. Silvestre, J.-N. Wang, On Landis’ Conjecture in the Plane, Communications in Partial Differential Equations, 40:4, 2015, 766–789. V. A. Kondratiev, E. M. Landis, Qualitative theory of second order linear partial differential equations, Partial differential equations – 3, Itogi Nauki i Tekhniki, Ser. Sovrem. Probl. Mat. Fund. Napr., 32, VINITI, Moscow, 1988, 99–215. I. Kukavica, Quantitative uniqueness for second-order elliptic operators, Duke Math. J. 91(2), 1998, 225–240. E. M. Landis, Second order equations of elliptic and parabolic type, Translations of Mathematical Monographs, 171. American Mathematical Society, Providence, RI, 1998; translated from the 1971 Russian original by Tamara Rozhkovskaya; with a preface by Nina Uraltseva. V. Z. Meshkov, On the possible rate of decay at infinity of solutions of second order partial differential equations, Math. USSR Sbornik 72, 1992, 343–360.
F. Nazarov, L. Polterovich, M. Sodin, Sign and area in nodal geometry of Laplace eigenfunctions, Amer. J. Math., [ 127]{}, 2005, 879–910.
J. Zhu, Quantitative uniqueness of elliptic equations. Amer. J. Math., 138(3), 2016, 733–762.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that $d+1$-dimensional surface growth models can be mapped onto driven lattice gases of $d$-mers. The continuous surface growth corresponds to one dimensional drift of $d$-mers perpendicular to the $\left( d-1\right)$-dimensional “plane” spanned by the $d$-mers. This facilitates efficient, bit-coded algorithms with generalized Kawasaki dynamics of spins. Our simulations in $d=2,3,4,5$ dimensions provide scaling exponent estimates on much larger system sizes and simulations times published so far, where the effective growth exponent exhibits an increase. We provide evidence for the agreement with field theoretical predictions of the Kardar-Parisi-Zhang universality class and numerical results. We show that the $\left(2+1\right)$-dimensional exponents conciliate with the values suggested by Lässig within error margin, for the largest system sizes studied here, but we can’t support his predictions for $\left(3+1\right)d$ numerically.'
author:
- 'Géza Ódor (1), Bartosz Liedke (2) and Karl-Heinz Heinig (2)'
bibliography:
- 'ws-book9x6.bib'
title: 'Directed $d$-mer diffusion describing Kardar-Parisi-Zhang type of surface growth'
---
One of the simplest nonlinear stochastic differential equation set up by Kardar, Parisi and Zhang (KPZ) [@KPZeq] describes the dynamics of growth processes in the thermodynamic limit. It specifies the evolution of the height function $h(\mathbf{x},t)$ in the $d$ dimensional space $$\label{KPZ-e}
\partial_t h(\mathbf{x},t) = v + \sigma\nabla^2 h(\mathbf{x},t) +
\lambda(\nabla h(\mathbf{x},t))^2 + \eta(\mathbf{x},t) \ .$$ Here $v$ and $\lambda$ are the amplitudes of the mean and local growth velocity, $\sigma$ is a smoothing surface tension coefficient and $\eta$ roughens the surface by a zero-average Gaussian noise field exhibiting the variance $\langle\eta(\mathbf{x},t)\eta(\mathbf{x^{\prime}},t^{\prime})\rangle = 2 D
\delta^d (\mathbf{x-x^{\prime}})(t-t^{\prime})$. The notation $D$ is used for the noise amplitude and $\langle\rangle$ means the distribution average.
The KPZ equation was inspired in part by the the stochastic Burgers equation [@Burgers74], which belongs to the same universality class [@forster77], and it became the subject of many theoretical studies [@HZ95; @barabasi; @krug-rev]. Besides, it models other important physical phenomena such as directed polymers [@kardar85], randomly stirred fluid [@forster77], dissipative transport [@beijeren85; @janssen86], and the magnetic flux lines in superconductors [@hwa92]. The equation is solvable in $\left( 1+1\right) d$ [@kardar87], but in higher dimensions approximations are available only. As the result of the competition of roughening and smoothing terms, models described by the KPZ equation exhibit a roughening phase transition between a weak-coupling regime ($\lambda <\lambda _{c}$), governed by the Edwards-Wilkinson (EW) fixed point at $\lambda =0$ [@EWc], and a strong coupling phase. The strong coupling fixed point is inaccessible by perturbative renormalization group (RG) method. Therefore, the KPZ phase space has been the subject of controversies and the value of the upper critical dimension has been debated for a long time.
Using a directed polymer representation, the validity of a scaling hypothesis [@DK92] and the two-loop RG calculation for $d\ge 2$ [@FT94] was confirmed and extended to all orders in $d=(2+\epsilon)$ [@L95]. These results provided an argument for an upper critical dimension $d_{c}=4$ of the roughening transition, but the strong-coupling rough phase is not accessible by perturbation theory. Above $d=1$ the scaling behavior in the rough phase has been very controversial, diverse values for the scaling exponents were claimed [@H90; @Ste94; @L98]. In particular, assuming that height correlations exhibit no multiscaling and satisfy an operator product expansion, exact field-theoretic methods lead to rational number growth values in two and three dimensions [@L98]. Some theoretical approaches predict that $d_{c}=4$ is an upper critical dimension of the rough phase [@M95; @L97; @B98; @F05]. Recently, a non-perturbative RG study has been able to describe the strong coupling fixed point and has provided indications for a possible qualitative change of the critical behavior around $d=4$ [@CDDW09]. This is in contradiction with the numerical results [@AHK93; @ala99; @MPP; @MPPR02], which predict the lack of an upper critical dimension.
Mapping of surface growth onto reaction-diffusion system allow effective numerical simulations [@dimerlcikk; @Orev]. As a generalization of the $1+1$ dimensional roof-top model [@kpz-asepmap; @meakin] and the $2+1$ dimensional octahedron model [@asep2dcikk], we consider the deposition and removal processes of higher dimensional objects on $d\ge 2$ dimensional surfaces. We remind that in $1+1$ dimensions a continuous surface line having no overhangs can be approximated by 45 degree up/down slope elements (after appropriate length rescaling). A process with KPZ scaling can be realized by deposition at local minima or removal of local maxima (roof-top). If we associate the up slopes with ’particles’ and down slopes with ’holes’ of the base lattice (see Fig. \[map\].a), the adsorption/desorption corresponds to the asymmetric exclusion process (ASEP) [@Ligget]. This is a lattice gas [@KLS84; @S-Z], where particles can hop on adjacent sites with asymmetric rates and hard-core exclusion. Its behavior is well known, and variations of ASEP (disorder, interactions ... etc.) correspond to variations of $1+1$ dimensional KPZ growth models.
We have extended the roof-top construction to $\left(2+1\right)$ dimensions [@asep2dcikk] by the introduction of octahedra having four slopes. The up edges in the $x$ or $y$ directions can be represented by ’$+1$’, while the down ones by ’$-1$’, and a surface element update is a generalized (Kawasaki) exchange (Eq. (3) of [@asep2dcikk]). The translation of up edges to ’particles’ and the down ones to ’holes’ of the base lattice maps particle deposition/removal processes onto two simultaneous particle moves (one in the $x$ and one in the $y$ direction). One can also consider it as a dimer move in the bisectrix direction of $x$ and $y$ (see Fig. \[map\].b). Therefore, the $\left(2+1\right)$ surface dynamics can be mapped onto a “two-dimensional ASEP” of oriented dimers exhibiting hard-core exclusion. The asymmetric drift corresponds to an evolving surface exhibiting KPZ scaling, while the symmetric dimer diffusion is related to the EW behavior.
Now we proceed with this kind of construction, considering the discrete slope variables in higher dimensions and generalize the simultaneous $+1 \leftrightarrow -1$ (Kawasaki) exchange rule of them (Eq.(3) of [@asep2dcikk]) to $d$-dimensional updates $$\left(
\begin{array}{cc}
-1 & 1 \\
-1 & 1 \\
-1 & 1 \\
... &
\end{array}
\right) \overset{p}{\underset{q}{\rightleftharpoons }}\left(
\begin{array}{cc}
1 & -1 \\
1 & -1 \\
1 & -1 \\
... &
\end{array}
\right) \ , \label{rule}$$ with probability $p$ for attachment and probability $q$ for detachment (see Fig. \[map\].c for the 3d case). It is well known [@barabasi] that the surface evolution of the deterministic KPZ growth are described also by the Burgers equation [@Burgers74] for growth velocities **v**(**x**,t) in the surface normal obeying $$\label{BUR-e}
\partial_t \mathbf{v}(\mathbf{x},t) = \sigma\nabla^2 \mathbf{v}(\mathbf{x}%
,t) + \lambda \mathbf{v}(\mathbf{x},t) \nabla \mathbf{v}(\mathbf{x},t)$$ due to the transformation $\mathbf{v}(\mathbf{x},t) = \nabla h(\mathbf{x},t)$.
In the forthcoming part we will prove that our microscopic model for $d$-mers in the continuum limit can be mapped onto the anisotropic version of Eq. (\[BUR-e\]), similarly as shown in lower dimensions [@kpz-asepmap; @asep2dcikk]. The derivation is based on the formulation of the reduction of possible updates. Our surface model is represented by the discrete derivative elements: $\delta _{x}$, $\delta _{y}$, $\delta _{z}$ ... ($\in \pm 1$) at every lattice points. A generalized Kawasaki update (\[rule\]) is defined by a matrix $$\begin{aligned}
\label{up1}
\left(
\begin{array}{cc}
\delta _{x}(i-1,j,k,...) & \delta _{x}(i,j,k,...) \\
\delta _{y}(i,j-1,k,...) & \delta _{y}(i,j,k,...) \\
\delta _{z}(i,j,k-1,...) & \delta _{z}(i,j,k,...) \\
... &
\end{array}
\right) \ .\end{aligned}$$ In $d$ dimensions we define vectors of the slopes, the columns of (\[up1\]), analogously to one and two dimensions: $\overline{\sigma }_{i,j,k,..}=(\delta _{x}(i-1,j,k,...),\delta
_{y}(i,j-1,k,...),...)\ , \label{svect}$ around the lattice point, which we select for deposition/removal update and set up a microscopic master equation $$\begin{aligned}
\partial _{t}P(\{\overline{\sigma }\},t)
&=&\sum_{i,j,k,...}w_{i,j,k,...}^{\prime }(\{\overline{\sigma }\})P(\{%
\overline{\sigma }^{\prime }\},t) \notag \label{mastereq} \\
&-&\sum_{i,j,k,...}w_{i,j,k,...}(\{\overline{\sigma }\})
P(\{\overline{\sigma }\},t)\end{aligned}$$with the probability distribution $P(\{\overline{\sigma }\},t)$. Here the prime index denotes the state of $\overline{\sigma}$ following the update (\[rule\]). The transition probability of $\overline{\sigma}$-s can be expressed as $$\begin{aligned}
\label{wtr}
w_{i,j,k,...}(\{\overline{\sigma }\}) &=&A[1-\overline{\sigma }%
_{i+1,j+1,k+1,...}\overline{\sigma }_{i,j,k,...} \label{mast} \\
&+&\lambda (\overline{\sigma }_{i+1,j+1,k+1,...}-\overline{\sigma }%
_{i,j,k,...})]\ , \notag \end{aligned}$$ with $\lambda = 2 p/(p+q)-1$ parametrization, which formally looks like the Kawasaki exchange probability in $1$d, except the factor $A$, which is necessary to avoid surface discontinuity creation in higher dimensions. This means that we update the slope configurations only if the values of all coordinates of the vector $\overline{\sigma}$ are identical as shown by (\[rule\]). One can allow formally these updates via the expression $$\begin{aligned}
\label{filt}
A &=&1/2^{d+1}\det [(\overline{\sigma }_{i,j,k,..}+C\overline{\sigma }%
_{i,j,k,..}) \\
&&\times (\overline{\sigma }_{i+1,j+1,k+1,..}+C\overline{\sigma }%
_{i+1,j+1,k+1,..})I]\ , \notag\end{aligned}$$where $I$ and $C$ are the unity and the cyclic permutation matrices respectively. The matrix $C$ shifts each coordinate value to the next index value. Thus for $\overline{\sigma}$-s with mixed coordinate values, the vectors $(\overline{\sigma}_{i,j,k,..} + C\overline{\sigma}_{i,j,k,..} )=\overline{k}$ or $(\overline{\sigma}_{i+1,j+1,k+1,..} +C\overline{\sigma}_{i+1,j+1,k+1,..} )
=\overline{k'}$ possess zero elements. Therefore the determinant of $\overline{k} \overline{k'} I$, being the product of the diagonal elements, is zero in case of mixed coordinates and $A=1/2^{d+1}$ in case of equal coordinates.
For example a $d=3$ update is prohibited when the slope vector is $\overline{\sigma }=(1,1,-1)$, because $\overline{k}$ has one coordinate value of zero $$\overline{k}=\left(
\begin{array}{c}
1 \\
1 \\
-1%
\end{array}%
\right) + \left(
\begin{array}{ccc}
0 & 1 & 0 \\
0 & 0 & 1 \\
1 & 0 & 0%
\end{array}%
\right) \left(
\begin{array}{c}
1 \\
1 \\
-1%
\end{array}%
\right) =\left(
\begin{array}{c}
2 \\
0 \\
-2%
\end{array}
\right) \ .$$
By calling ’$+1$’-s as particles and the ’$-1$’-s as holes of the base lattice, their synchronous update can be considered to be a single step motion of an oriented $d$-mer in the bisectrix direction of the $x$, $y$, $z$, ... coordinate axes. Thus $d$-mers follow one-dimensional kinetics, described by Kawasaki exchanges (\[mast\]). To obtain a one-to-one mapping we update neighborhoods of the lattice points denoted by the green dots of Fig. \[map\].
To derive Eq. (\[BUR-e\]) first we have to average over the slope vectors $$\langle \overline{\sigma }\rangle =\sum_{\{\overline{\sigma }\}}
\overline{\sigma } P(\{\overline{\sigma }\},t) .$$ By calculating its time derivative using the master equation (\[mastereq\]) and the transition probabilities (\[wtr\]) $$\begin{aligned}
&&\partial _{t}\langle \overline{\sigma} \rangle =
\sum_{\{\overline{\sigma }\}} \Big[ \overline{\sigma}
\sum_{i,j,k,...}w_{i,j,k,...}^{\prime }(\{\overline{\sigma }\})
P(\{\overline{\sigma }^{\prime }\},t) \notag \\
&-&\overline{\sigma} \sum_{i,j,k,...}w_{i,j,k,...}
(\{\overline{\sigma }\})P(\{\overline{\sigma }\},t) \Big] \ ,\end{aligned}$$ in which we filter out vectors of non-equal coordinates (\[filt\]) (thus $w_{i,j,k,...}$ is nonzero only if $\overline{\sigma}_{i,j,...}\ne\overline{\sigma}_{i+1,j+1,...}$) we can obtain $$\begin{aligned}
&&2\partial _{t}\langle \overline{\sigma }_{i,j,k,...}\rangle =\langle
\overline{\sigma }_{i-1,j-1,...}\rangle -2\langle \overline{\sigma }%
_{i,j,...}\rangle \\
&+&\langle \overline{\sigma }_{i+1,j+1,...}\rangle +\lambda \langle
\overline{\sigma }_{i,j}(\overline{\sigma }_{i+1,j+1,...}-\overline{%
\sigma }_{i-1,j-1,...})\rangle \ , \notag\end{aligned}$$ analogously to one dimension [@kpz-asepmap]. Here one can see the discrete first and second differentials of $\overline{\sigma }_{i,j,k,...}$ corresponding to the operators of Eq. (\[BUR-e\]) in the bisectrix direction of the axes. These differentials are one-dimensional, because the $d$-mer dynamics is one-dimensional. In principle one could derive a set of coupled Burgers equations for the particles in each direction in an isotropic way in accordance with isotropic surface model, but the coordinated movements reduce the dimensionality and we can map onto an anisotropic equation of $d$-mers.
Making a continuum limit in each direction and taking into account the relation of height and slope variables ($\mathbf{v}(\mathbf{x},t) = \nabla h(\mathbf{x},t)$), we can arrive to the deterministic KPZ equation. The nonlinear term vanishes for $p=q$ ($\lambda =0$). The sign of the coefficient $\lambda $ of the nonlinear term can be interpreted as follows: For $p>q$ positive non-linearity (positive excess velocity) it is a consequence of growth with voids.
Since this derivation was applied just for the first one in the hierarchy of equations for correlation functions it does not prove the equivalence to the stochastic KPZ. Furthermore, the form of the noise term, which was not considered in our derivation, may also introduce differences. Although our surface model is spatially isotropic, we can map it onto a one-dimensional Burgers equation (of extended objects), therefore anisotropic scaling behavior might be expected. However, by going into the continuum description the hard-core exclusions necessary to provide continuous surfaces is lost, and the resulting equation looks trivial.
Here we investigate by numerical simulations this isotopic surface growth model via the one-dimensional directed migration of $d$-mers in the $d$-dimensional space. We have developed bit-coded algorithms for the updates (\[rule\]) and run it with $p=1$, such that randomness comes from the site selection only. Therefore it is important to use a very good random number generator, which provides uniformly distributed numbers with high resolution. Otherwise we would realize a process with quenched disorder, which for KPZ belongs to a different universality class (see [@Orev]). We used the latest Mersenne-Twister generator [@MT] in general, which has very good statistical properties and which is very fast, especially in the SSE2 registers. We tested our results using other random number generators as well. In practice each update site can be characterized by the $2^{d^2}$ different local slope configurations. However, due to the surface continuity we need only a few bits of a world (1 byte for $d=2,3,4$ and two bytes for $d=5$) for this purpose. This allows an efficient storage management in the computer memory and permits simulations of larger system sizes. The updates can be performed by logical operations, either on multiple samples at once, or on multiple (not overlapping) sites at once. Our bit-coded algorithm proved to be $\sim 40$ times faster than the conventional FORTRAN 90 code we started with. It is important to note that this stochastic cellular automaton like representation of the surface growth opens the possibility for an implementation on extremely fast graphic cards with massively parallel processors. Furthermore the construction permits the extension of the mapping for more complex surface reactions [@patscalcikk].
We performed dynamical simulations by starting from stripe ordered particle distributions. This corresponds to a flat surface with a small intrinsic width. The considered lattices gases had the maximum linear sizes $L_{max}=2^{15}, 2^{10}, 2^8, 2^6$ for $d=2,3,4,5$ dimensions, respectively and periodic boundary conditions were applied. A single step of the lattice gas algorithm comprises a random site selection and in case of an appropriate neighborhood configuration a $p=1$ Kawasaki $d$-mer update (\[rule\]). The time is incremented by $1/L^d$ in units of Monte Carlo steps (MCs). Throughout the paper we will use this unit of time.
We could exceed by magnitudes of order all previous numerical system sizes and simulation times. For example the largest five-dimensional simulations were done for $L=30$ and $t_{m}=230$ MCs [@AHK93]. Our $L=64$ simulations, where we have the good bulk/surface ratio: $5.4$, required 2GB memory size and a couple of weeks for a single realization up to $t_{max}=5000$ MCs. Similarly, the largest sized simulations in $d=2$ for $L=11520$ system could achieve $t_{max}=10^4$ MCs [@TFW92]. Our largest $L=32768$ sized simulations reached $t_{max}=44600$ MCs. The longest runs for $L=4096$ passed the saturation at $t\simeq 4\times10^5$ MCs and the samples were followed up to $t_{max}=10^6$ MCs.
We run the these lattice gas simulations for $10-1000$ independent realizations for each dimension and size considered, and calculated $h_{x,y,..}(t)$ and the second moment $$\label{width}
W(L,t) = \Bigl[ \frac{1}{L^{2d}} \, \sum_{x,y,..}^L
\,h^2_{x,y,...}(t) - \Bigl(\frac{1}{L^d} \, \sum_{x,y,...}^L \,h_{x,y,...}(t)
\Bigr)^2 \Bigr]^{1/2}$$ from the height differences at certain sampling times. The growth is expected to follow the Family-Vicsek scaling [@family] asymptotically, but due to the corrections it can be described by a power series $$\label{betacorr}
W(t,L\to\infty) = b t^{\beta}(1 + b_0 t^{\phi_0} + b_1 t^{\phi_1} ...) \ ,$$ with the surface growth exponent $\beta$. For finite system, when the correlation length exceeds $L$, the growth crosses over to a saturation with the scaling law $$\label{alphacorr}
W(t\to\infty,L) = a L^{\alpha}(1 + a_0 L^{\omega_0} + a_1 L^{\omega_1} ...) \ ,$$ characterized by the roughness exponent $\alpha$. In our case the intrinsic width of the initial state, which is represented by a zig-zag surface of width $1/2$ (see Fig.\[map\]), results in a constant correction term. Thus we have $b_0=1/2$, $\phi_0=-\beta$ and $a_0=1/2$, $\omega_0=-\alpha$. During our scaling analysis we dropped this contribution by subtracting $W^2(0)=1/4$ from the raw data and consider the next leading order correction as leading one. Furthermore we disregarded the initial time region $t < t_0 \simeq 50$, when basically an uncorrelated random deposition occurs. The dynamical exponent $z$ can be expressed by the ratio $z = \alpha/\beta$ and in case of the Galilean invariance of an isotropic KPZ equation the $z=2-\alpha$ relation should also hold.
Besides the extensive simulations we have performed careful correction to scaling analysis by calculating the local slopes of the exponents. The effective exponent of the surface growth can be estimated similarly as in case of other scaling laws [@Orev], as the discretized, logarithmic derivative of (\[width\]) $$\label{beff}
\alpha_{eff}(L) = \frac {\ln W(t\to\infty,L) - \ln W(t\to\infty,L')}
{\ln(L) - \ln(L^{\prime})} \ .$$ It was determined numerically for different discretizations: $t/t^{\prime}=2,3$, and we tried to fit it with the leading-order correction ansatz, which can easily be deduced from (\[betacorr\]) (see [@AA04] or [@Orev]) $$\label{betaslfit}
\beta_{eff}(t) = \beta + b_1\phi_1 t^{\phi_1} \ ,$$ for $t>t_0$ and before the saturation region. In other cases, such as ballistic deposition, which has a large unknown intrinsic width one can use another effective roughness exponent definition introduced in [@AA04].
We tested our method with the [*one-dimensional*]{}, exactly known case. Simulations were run on $L=5\times 10^5$ sized system up to $t_{max}=16666$ MCs for $40$ independent realizations. We determined the effective exponents $\beta_{eff}(t)$, which approaches $\beta=1/3$ from below, in a perfect agreement with the leading-order correction form (see Figure \[beta1\]). The fitting with (\[betaslfit\]) on the local slopes data resulted in $\beta=0.333(5)$ and $\phi_1=-0.53$.
Similarly to the time dependence we can analyze the size dependence following the saturation by determining the effective exponent of the roughness, which can be defined as the logarithmic derivative of (\[width\]) $$\label{aeff}
\alpha_{eff}(L) = \frac {\ln W(t\to\infty,L) - \ln W(t\to\infty,L')}
{\ln(L) - \ln(L^{\prime})} \ .$$ The finite size scaling was done for systems of linear sizes in between $L_{max}$ (discussed earlier) and $L_{min}$, which was $2^6$ for 2d, $2^5$ for 3d, $2^4$ for 4d and $2^3$ for 5d, respectively. To handle the boundary conditions effectively, system sizes of power of 2 were simulated. To get the asymptotic values we took into account all effective exponent points shown on Fig. \[alphas\] and applied a leading order, linear fitting. The error margins of exponents are estimated from the error-bars of Fig. \[alphas\]. This method gives a better estimate for the asymptotic values than just a least-square fitting on the data points, which completely disregards corrections to scaling. We also calculated rough, but independent estimates for $z$ by measuring the relaxation time, i.e. the time needed to reach 90% of the saturation value. The asymptotic value is extrapolated by a linear fitting: $z_{eff}(t) = z + c_1/L$.
In [*two dimensions*]{} we estimated the growth exponent in the largest system sizes considered ($L=2^{15}, 2^{14}, 2^{13}$) (see Fig. \[betas\]). Fitting in the $50 < t < 44600$ time window with the form (\[betaslfit\]) resulted in $b_1=0.83$ and $\beta=0.245(5)$, which is somewhat bigger than what was obtained by the largest known ($L=11520$) sized simulations: $\beta=0.240(1)$ [@TFW92], and all other previous numerical estimates including ours [@Ghai; @AA04; @asep2dcikk]. This value conciliates with the $\beta=1/4$ RG exponent of [@L98]. One can obtain this value by the late time behavior of effective exponent, which has not been seen before, because finite size effects have screened it. On the graph one can see strong oscillations for $L=2^{10}$ and intermediate times, which are damped before saturation. In the one-dimensional ASEP model such oscillations are shown to be the consequence of density fluctuations being transported through a finite system by kinematic waves [@GMGB07]. One can speculate that the slight final increase of $\beta_{eff}$ for the largest system sizes is just a fluctuation or oscillation effect, but we could not eliminate this overall tendency by increasing the statistics. Although the statistical fluctuations grow dramatically, as $t\to\infty$ the increase of the mean value is observable for each size $L > 2^{11}$. Our error-bar of $\beta$ reflects this uncertainty. The width saturation values have been investigated for $L=2^6, 2^7,...,2^{12}$. We took into account the leading order correction to-scaling by the following Ansatz $$\label{WFform}
\alpha_{eff}(L) = \alpha + a_1\omega_1 L^{\omega_1} \ ,$$ but due to the larger error-bars we restricted it to a linear approximation: $\omega_1=-1$. The local slopes of the steady state values $\alpha_{eff}(1/L)$ and of $z_{eff}(1/L)$ are shown on Fig. \[alphas\].
This provides $\alpha=0.395(5)$ and $a_1=2.02$ for the roughness and $z_{eff}=1.58(10)$, with the linear coefficient $c_1=1.83$ for the dynamical exponent. This roughness exponent is in agreement with RG value [@L98], and somewhat bigger than the existing figures $\alpha=0.393(3)$ [@MPP] for $L\le 1024$ and $\alpha=0.385(5)$ [@TFW92] for $L\le 128$.
In [*three dimensions*]{} the local slope analysis for $L=2^{10}$ results in $b_1 = 0.1$ and $\beta=0.184(5)$ agreeing with the numerical results from the literature: $\beta=0.180(2)$ [@TFW92; @AHK93], $\beta=0.186(1)$ [@MPP]. But our estimate is much higher than $\beta=0.168(3)$ [@Ghai] (based on $L < 200$ sized simulations) and $\beta=1/6$ predicted by RG [@L98].
For the saturation we obtained $a_1=1.40$ and $\alpha=0.29(1)$, matching $\alpha=0.29$ of [@Ghai] and in marginal agreement with $\alpha=0.3135(15)$ of [@MPP] and $\alpha=0.308(2)$ of [@AHK93]. The direct $z_{eff}$ measurement exhibits a strong correction to scaling: $z=1.60(1)$ ($c_1=1.10$) and one cannot differentiate it from the $2+1$ dimensional results within the error margins.
In [*four spatial dimensions*]{} our best fit for the growth exponent is $b_1=1.08$ and $\beta=0.15(1)$. In the literature $\beta=0.16(1)$ [@ala99] and $\beta=0.146(1)$ [@MPP] values are reported. For the width saturation values the linear fitting results in $\alpha=0.245(5)$ with $a_1=0.07$. This compares with the literature values $\alpha=0.245(1)$ [@AHK93] and $\alpha=0.255(5)$ [@MPP]. The $z_{eff}$ seems to converge to $z=1.91(10)$ (with $c_1=-0.64$) but the fluctuations are very strong and we could not reach saturation for sizes larger than $L=128$. Going further by a factor of two in system sizes would require simulations with 8GB memory and very long CPU times. Our results do not support the field theoretical prediction of $d_c=4$, because we don’t observe the disappearance of power-law growth.
In [*five dimensions*]{} the local slopes suggest $b_1=0.134$ and $\beta=0.115(5)$ in agreement with $\beta=0.11(1)$ [@AHK93] reported for smaller sizes. One can find strong oscillations before the saturation regime. Again these are due to kinematic transport waves in finite system. Initially for $L=64$ we saw a definite increase in $\beta_{eff}$ as $1/t < 0.005$ before the saturation, but this proved to be an artifact of the MT random number generator. When we used different, pseudo-random number generators: drand48 [^1] or random() of language C, the growth tendency for very late times was much weaker. We think that the site selection, the only source of randomness in case of $p=1$ might not be completely uniform among the $2^{30}$ possible places. To confirm this we repeated the $5d$ simulations using $p=0.9$ with the MT generator and found agreement with the results using drand48. For the saturation we estimate $\alpha=0.22(1)$ with $a_1=0.08$ and $z=1.95(15)$ with $c_1=-0.55$.
In conclusion we have shown that the mapping of a KPZ surface growth model onto driven lattice gases (DLG) can be extended to higher dimensions. Although the growth of the surfaces exhibits the spatial symmetry of the underlying lattice, one can map it onto an anisotropic DLG of more complex objects. The coarse grained, continuum description of these $d$-mers is an anisotropic Burgers equation. Still the DLG model is non-trivial, because it is just an oriented drift of $d$-mers with hard-core exclusions. The topological constraint is the consequence of the required surface continuity by the mapping. In two dimensions we confirmed [@patscalcikk] that the probability distribution $P(W^2)$ matches the universal scaling function determined for another KPZ model [@MPPR02]. We presented effective bit-coding simulations and high precision results for the exponents $\alpha$, $\beta$ and $z$ independently (see Table I.). The sensitive local slope analysis provides numerical agreement with former simulation results, but for larger sizes, which have not been investigated so far, we see a slight growing tendency in the $\beta_{eff}$ exponents in all dimensions. For $d=2$ our results marginally overlap with the $\beta=1/4$ value suggested some time ago by RG. The change towards a trivial behavior in higher dimensions in the DLG language would mean the disappearance of the topological constraints among the extended $d$-mer objects as they could follow a simple ASEP dynamics of point particles. This will be the target of further studies using massively parallel algorithms on graphic cards. We hope that we will be able to obtain a firm estimate for the upper critical dimension using extrapolation techniques.
$d$ $\alpha$ $\beta$ z
----- ---------- ---------- ----------
2 0.395(5) 0.245(5) 1.58(10)
3 0.29(1) 0.184(5) 1.60(10)
4 0.245(5) 0.15(1) 1.91(10)
5 0.22(1) 0.115(5) 1.95(15)
: Independent growth exponent estimates of the $d$-mer model in different dimensions
We thank Zoltán Rácz for the useful comments. Support from the Hungarian research fund OTKA (Grant No. T77629), the bilateral German-Hungarian exchange program DAAD-MÖB (Grant Nos. D/07/00302, 37-3/2008) is acknowledged. G. Ódor thanks for the access to the Clustergrid and the NIIF supercomputer.
[^1]: see for example: http://linux.die.net/man/3/drand48
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a thick $p$-brane embedded in an $n$-dimensional spacetime possessing radial symmetry in the directions orthogonal to the brane. We first consider a static brane, and find a general fine tuning relationship between the brane and bulk parameters required for the brane to be flat. We then consider the cosmology of a time dependent brane in a static bulk, and find the Friedmann equation for the brane scale factor $a(t)$. The singularities that would ordinarily arise when considering arbitrary codimensions are avoided by regularizing the brane, giving it a finite profile in the transverse dimensions. However, since we consider the brane to be a strictly local defect, we find that the transverse dimensions must have infinite volume, and hence gravity cannot be localized on the brane without resorting to some infra-red cutoff.'
author:
- |
Stephen A. Appleby and Richard A. Battye\
[*Jodrell Bank Observatory, Department of Physics and Astronomy,*]{}\
[*University of Manchester, Macclesfield, Cheshire, SK11 9DL*]{}
title: Regularized braneworlds of arbitrary codimension
---
Introduction
============
The idea that our Universe might be a four dimensional subspace embedded in a higher dimensional manifold has been exhaustively studied over the past decade; see for example [@br03; @shi1; @rg1; @ma04; @rd04; @cz100; @rs01; @rs02; @cl1; @ct102; @dl10] and references therein (and previously; see [@vr00], [@vr01] and [@bc00]). The most prominent model to have arisen since the inception of so called ‘braneworld’ theories is the Randall Sundrum (RS) model. In Refs. [@rs01], [@rs02], a thin, static, pure tension brane is embedded in five dimensional Anti-de-Sitter (AdS) space. It was found that a flat hypersurface could be embedded into the background space, but only if the brane tension was fine tuned to the bulk cosmological constant. Moreover, it was found that gravity could be localized in the vicinity of the four dimensional static brane. This important result meant that observers on a brane would feel standard four dimensional gravity at sufficiently low energies, where any massive bulk modes would be highly suppressed. In addition, this model offered a possible solution to the hierarchy problem.
Simple codimension one braneworld models such as the Randall Sundrum setup typically involve a static brane in a static higher dimensional space. However, the cosmological generalization to a time dependent brane has been considered by many authors since (see Ref. [@dl10] and references therein, for example). For codimension one objects, modifications to the standard four dimensional Friedmann equation have been found; specifically, one obtains a new radiation-like term that can be considered as a contribution to the brane evolution due to the bulk. In addition, one finds another modification to standard four dimensional cosmology which is quadratic in the brane energy density. This implies that for large energy densities on the brane, one would experience a completely different brane evolution than would be the case in conventional four dimensional gravity.
Progress in determining the cosmological evolution of codimension two branes has also been made [@n01; @cl00; @jv00; @pb00]. Initially seen as promising candidates for solving the cosmological constant problem, such models run into difficulties when one attempts to introduce matter on an otherwise pure tension brane. Specifically one obtains a singularity at the position of the brane, when considering more general equations of state for matter localized on the source. However, by treating the introduction of matter as a perturbation around a known static six dimensional braneworld model, it has been found in Ref. [@jv00] that if a thick brane is considered, then we can obtain standard late time FRW behaviour. Thickening the brane is one method of regularizing codimension two objects, but there are others (introducing Gauss-Bonnet terms for example [@pb00]).
Branes of arbitrary codimension have not been studied in general (although see for example Refs. [@is1; @cz20; @lb1]), for a number of reasons. In particular, one cannot simply embed a thin $(3+1)$-brane in an $n$-dimensional space, since one generically finds naked singularities in curvature invariants at the position of the brane. This problem can be resolved in a number of ways; for example one could thicken the brane, effectively smearing any singularities over a small region of the transverse space. This is the approach that we will take, that is we will replace the singular brane energy momentum tensor with some smoothed distribution. We then define effective four dimensional quantities by integrating the full field equations over the codimensions, averaged by the brane profile. Our approach follows Refs. [@ct102; @ct103; @ct104], where a similar regularization was considered.
This paper will be concerned with the cosmological evolution of a thick $(3+1)$-brane of arbitrary codimension $m$. We will show that four dimensional brane cosmology can be obtained for a large class of metrics where a thick $(3+1)$-brane is embedded in an $n$-dimensional background space. Our work is an extension of Refs. [@n01] and [@n02], where the cosmological equations for the brane scale factor $a(t)$ were obtained for codimension one and two branes. In Ref. [@n01] it was found that conventional late time cosmology can be obtained on a four dimensional brane in a six dimensional space. Our calculation closely follows these works, and we find a corresponding set of equations for a brane of arbitrary codimension. The method used in Ref. [@n01] to derive the cosmological equations will generalize for higher codimension, and we will show that we can recover standard late time cosmology regardless of the number of extra dimensions, subject to the assumptions we impose. We will concentrate on the case of a highly symmetric bulk spacetime, in which the brane is static. Some authors [@uz01; @sas01; @cc06] have considered dropping exact spherical symmetry in the bulk and studying the resulting brane field equations. This scenario is considerably more complicated than the models considered here, since the brane will generically experience a force when the bulk is not symmetric, and we will not consider such behaviour.
One important aspect of braneworld models is the requirement that conventional four dimensional gravity must be obtainable on the brane. For this to occur, we must have some mechanism that localizes gravity in the vicinity of the brane. One way in which this can be achieved is to have a transverse space with finite volume, as in the RS model. However, in this paper we will be considering local defects of arbitrary codimension, and it has been found in Ref. [@gh00] that for such models the transverse dimensions must have infinite volume. This means that gravity cannot be localized on the brane without introducing some large distance cutoff when we integrate over the transverse dimensions to obtain four dimensional gravity. This cutoff could arise by introducing a second brane, for example. We will consider this issue in section \[sec:100\].
The paper will proceed as follows. In the following section we review some important definitions and general braneworld results that we will need for the rest of the paper. In section $\ref{sec:st}$, we find the field equations for a particular class of static, spherically symmetric metrics. We then introduce a time-dependent metric in section $\ref{sec:1}$ and obtain an evolution equation for the brane scale factor, and show that our result correctly reduces to the codimension one and two examples as found in Refs. [@n01] and [@n02]. We write our result as a modified four dimensional Friedman equation, and doing so we find a non-standard cosmological equation with a dark radiation term present. The origin of this dark radiation term is explored in the appendix.
Formalism
=========
We begin with some important results that will be required in subsequent sections. In this paper, we will be considering codimension $m$, $p$-branes ($p$ being the number of spacetime dimensions of the brane) embedded in $n$-dimensional background spacetimes, hence $m=n-p$. Specifically, we will consider two different metrics. The first is a static, $n$-dimensional metric with spherical symmetry in the extra dimensions. It takes the form
$$\label{eq:1} ds^{2} = f(r)g_{AB}(x)dx^{A}dx^{B} -
dr^{2} - \alpha^{2}(r) \gamma_{ab}(y)dy^{a}dy^{b} ,$$
where $f(r)$ and $\alpha(r)$ are functions of the radial coordinate $r$ only, and $\gamma_{ab}$ is the metric of the unit $(m-1)$ sphere, so the extra dimensions are radially symmetric. The metric $g_{AB}(x)$ is the $(3+1)$ dimensional brane metric, and hence capital Latin indices $(A,B,..)$ run over the standard $(3+1)$ coordinates. The metric $\gamma_{ab}$ is that of the $(m-1)$-sphere (that is all of the brane orthogonal coordinates except the radial coordinate $r$), and hence lower case Latin indices $(a,b,..)$ run over $(m-1)$ of the codimensions. Finally, we will use the notation that Greek indices $(\mu,\nu,..)$ run over all $n$-dimensional coordinates. The full $n$-dimensional metric is $G_{\mu\nu}$, and in subsequent sections we will frequently need the measure $\sqrt{G}$, which is given by
$$\sqrt{G} = \lbrack f(r)\rbrack^{2} \alpha^{m-1}\sqrt{g\gamma} ,$$
where $g$ and $\gamma$ are the determinants of $g_{AB}$ and $\gamma_{ab}$ respectively.
In the next section, we will need the following decompositions of the $n$-dimensional Ricci tensor $R_{\mu\nu}$
$$\label{eq:2} R_{AB} = R^{(g)}_{AB} -
{1 \over 2}g_{AB} (\nabla_{a}\nabla^{a} + \nabla_{r}\nabla^{r}) f(r)
- {p -2 \over 4}g_{AB} f(r)^{-1} \nabla_{r}f(r) \nabla^{r}f(r)
,$$
$$\label{eq:3} R_{ab} = R^{(\gamma)}_{ab} - {p \over
2}f^{-1} \nabla_{a}\nabla_{b} f + {p \over 4}f^{-2}
\left(\nabla_{a}f\right) \left(\nabla_{b}f \right),$$
$$\label{eq:4} R_{rr} = R^{(\gamma)}_{rr} - {p \over
2}f^{-1} \nabla_{r}\nabla_{r} f + {p \over 4}f^{-2}
\left(\nabla_{r}f\right) \left(\nabla_{r}f \right),$$
where we will take $p=4$ in this paper, but is in general $p = g_{AB}g^{AB}$, and $R^{(g)}_{AB}$ is the Ricci tensor constructed from $g_{AB}(x)$. $R_{ab}^{(\gamma)}$ and $R_{rr}^{(\gamma)}$ are the components of the Ricci tensor constructed from the metric
$$ds^{2}_{[{\rm m}]} = -dr^{2} - \alpha(r)^{2}
\gamma_{ab}(y)dy^{a}dy^{b} .$$
In the next section, we will find that the Einstein equations for our metric ($\ref{eq:1}$) can be written in a very simple form in terms of the extrinsic curvatures of the subspaces with metrics $g_{AB}$ and $\gamma_{ab}$. For objects of codimension greater than one the extrinsic curvature is defined as [@c1]
$$\label{eq:extr1} K_{\mu\nu}{}^{\rho} \equiv \eta_{\nu}{}^{\sigma} \eta_{\mu}{}^{\beta} \nabla_{\beta}
\eta_{\sigma}{}^{\rho} ,$$
where $\nabla_{\beta}$ preserves the full $n$-dimensional background metric, and $\eta_{\nu}{}^{\sigma}$ projects tensors tangentially to the brane. For our metric, $\eta_{\nu}{}^{\sigma}$ is given by
$$\eta_{\mu}{}^{\nu} = \left( \begin{array}{cc}
g_{A}{}^{B} & 0 \\
0 & 0 \\
\end{array} \right) .$$
The first two indices of the extrinsic curvature tensor are tangential to the brane, and the last is orthogonal. The metrics that we are considering are highly symmetric, and as a result the extrinsic curvature tensor simplifies considerably. Using the metric ($\ref{eq:1}$), the only non-zero components of ($\ref{eq:extr1}$) are for $\rho = r$. We will use the extrinsic curvatures for our metric, $K_{A}{}^{B r}$ and $K_{a}{}^{b
r}$, which are given by
$$K_{A}{}^{B r} \equiv K_{A}{}^{B} = {f' \over
f}g_{A}{}^{B}, \qquad \qquad \qquad K_{a}{}^{b r} \equiv L_{a}{}^{b}
= 2{\alpha' \over \alpha}\gamma_{a}{}^{b} ,$$
where primes denote derivatives with respect to $r$. We have dropped the third index on the extrinsic curvature tensor, since the only non zero components of ($\ref{eq:extr1}$) are for $\rho =r$.
The second metric ansatz that we will consider is
$$\label{eq:197} ds^{2} = N^{2}(r,t) dt^{2} -
A^{2}(r,t)g_{ij}(x)dx^{i}dx^{j} - dr^{2} - \alpha^{2}(r,t)
\gamma_{ab}(y) dy^{a}dy^{b} ,$$
which is a more general version of ($\ref{eq:1}$), and will be used to model a time dependent brane. As before, we can split the line element ($\ref{eq:197}$) into brane tangential and brane orthogonal components. The ‘brane’ line element, $ds^{2}_{[b]}$, is given by
$$ds_{[{\rm b}]}^{2} = N^{2}(r,t) dt^{2} -
A^{2}(r,t)g_{ij}(x)dx^{i}dx^{j} ,$$
and for surfaces of constant $r$ is of the form of an FRW metric. We have introduced another set of indices $(i,j)$ in ($\ref{eq:197}$), which run over the standard three spatial dimensions. For the metric ($\ref{eq:197}$), the measure $\sqrt{G}$ is given by
$$\sqrt{G} = N A^{3} \alpha^{m-1} \sqrt{\gamma} .$$
The extrinsic curvatures $K_{A}{}^{B}$ and $L_{a}{}^{b}$ for this metric are
$$\label{eq:si1} K_{i}{}^{j} = 2\delta_{i}{}^{j} {A'\over A} , \qquad
\qquad \qquad K_{t}{}^{t} = 2 {N' \over N} , \qquad \qquad \qquad
L_{a}{}^{b} = 2 \delta_{a}{}^{b} {\alpha' \over \alpha},$$
and taking the trace gives
$$K = 2{N' \over N} + 6 {A' \over A}, \qquad \qquad \qquad L =
2(m-1){\alpha' \over \alpha} .$$
Regularization of the brane
---------------------------
Finally in this section, we explain our method of regularizing higher codimension branes, which follows Refs. [@ct102; @ct103; @ct104; @uz01]. For a thin brane, the total energy momentum tensor, $T_{\mu\nu}$, can be decomposed into distinct brane and bulk components. In the thin case, the standard definition of the $p$-brane energy momentum tensor $\hat{T}_{\mu\nu}$ is
$$\label{eq:397} \hat{T}_{\mu\nu} = \int \sqrt{g}
d^{p}\sigma
\bar{T}_{\mu\nu}\delta^{(n)}\left[x_{\alpha}-X_{\alpha}(\sigma^{A})\right],$$
where $\bar{T}_{\mu\nu}$ is the brane supported energy momentum tensor, $\sigma^{A}$ are the brane coordinates, and $x_{\alpha}$ the $n$-dimensional background coordinates. The brane is situated at $x_{\alpha} = X_{\alpha}(\sigma^{A})$.
In this paper we are considering branes of finite thickness. This means that we no longer treat the brane energy momentum tensor $\hat{T}_{\mu}{}^{\nu}$ as a singular source as in ($\ref{eq:397}$), but rather as some smoothed distribution. Explicitly, we consider the energy momentum tensor
$$\label{eq:fg1} \hat{T}_{\mu\nu} = \int \sqrt{g} d^{p}\sigma \bar{T}_{\mu\nu} D^{(n)}_{\epsilon}(x- X(\sigma)) ,$$
where we have replaced the $n$-dimensional $\delta$-function in ($\ref{eq:397}$) with the finite brane profile function $D^{(n)}_{\epsilon}(x- X(\sigma))$, which is peaked at $x=X(\sigma)$ and falls away sharply from the brane.
In this paper, we will consider the particular simple brane profile
$$\begin{aligned}
\label{eq:mn1} D_{\epsilon}^{(n)} = \quad &1 \quad r < \epsilon \\
&0 \quad r > \epsilon ,\end{aligned}$$
where $\epsilon$ can be considered as the brane thickness parameter. This profile has been considered previously, see for example Refs. [@dl10; @n02] (and a different brane profile was considered in [@uz01]). Although we use this particular profile, we expect that our results will be approximately valid for a large class of profile functions. Using ($\ref{eq:mn1}$), the brane supported energy momentum tensor $\tilde{T}_{\mu\nu}$ is given by
$$\tilde{T}_{\mu}{}^{\nu} = {1 \over \sqrt{g}|_{r=\epsilon}}
\int d^{m-1}y \int_{0}^{\epsilon} \sqrt{G} dr T_{\mu}{}^{\nu} .$$
We must also define the bulk energy momentum tensor, $T_{\mu\nu}^{\rm bulk}$. For simplicity, we will consider a cosmological constant only in the bulk, with no additional fields, so $T_{\mu\nu}|^{\rm bulk} = -\Lambda g_{\mu\nu}$, where $\Lambda$ is the $n$-dimensional cosmological constant.
Finally, we must discuss how to obtain four-dimensional equations for a thick brane. For thin branes, effective four dimensional equations are found by taking the full $n$-dimensional field equations and evaluating them on a surface of constant $y_{a}$ (the codimensions), at the position of the brane. This is equivalent to taking the full $n$-dimensional equations, and integrating them over the $m$ codimensions, weighted by the brane profile, which in this case is a delta function. For thick branes, we follow the same procedure; to obtain four-dimensional equations, we take the full $n$-dimensional equations, integrate them over the codimensions, weighted by the brane profile $D_{\epsilon}^{(n)}$, which is no longer singular. For the profile ($\ref{eq:mn1}$) used in this paper, our approach corresponds to integrating the field equations over the range $r=(0,\epsilon)$. In the next section, we will perform our method of regularization for a codimension $m$ static brane.
\[sec:st\] Static braneworld model
==================================
Let us first consider a braneworld model with a metric of the form ($\ref{eq:1}$). This is the simplest generalization of codimension one and two braneworlds that exist in the literature. Our approach follows the work of Refs. [@n01; @n02]; we will find that all of the field equations for the metric ansatz ($\ref{eq:1}$), except the $(r,r)$ equation, can be written approximately as total derivatives with respect to $r$. These equations can be integrated over the brane thickness to obtain a set of junction conditions. We then evaluate the remaining $(r,r)$ field equation at the brane-bulk boundary, that is at $r=\epsilon$, and use our junction conditions to write an equation relating the brane Ricci scalar $R^{(g)}$ to the brane energy momentum tensor $\tilde{T}_{\mu\nu}$. Before continuing we note that static metrics of the form ($\ref{eq:1}$) have been considered previously, see for example [@gh00].
To begin, we write the decompositions of the background Ricci tensor $R_{\mu}{}^{\nu}$, ($\ref{eq:2}$-$\ref{eq:4}$), in terms of the extrinsic curvature tensors $K_{A}{}^{B}$ and $L_{a}{}^{b}$,
$$\label{eq:b6} \sqrt{G} R_{A}{}^{B} = {\sqrt{G} \over
f}R^{(g)}{}_{A}{}^{B} + {1 \over 2}\left(\sqrt{G}K_{A}{}^{B}\right)'
= {\sqrt{G} \over M^{n-2}}\left( T_{A}{}^{B} - \delta_{A}{}^{B}{T
\over n-2}\right) ,$$
$$\label{eq:b7} \sqrt{G}R_{a}{}^{b} = {\sqrt{G} \over
\alpha^{2}} (m-2)\delta_{a}{}^{b} +{1 \over 2}
\left(\sqrt{G}L_{a}{}^{b}\right)' = {\sqrt{G} \over M^{n-2}}\left(
T_{a}{}^{b} - \delta_{a}{}^{b}{T \over n-2}\right) ,$$
$$\label{eq:b8} \sqrt{G} R_{r}{}^{r} =\sqrt{G} \left({1
\over 2} \left( L' + K'\right) + {1 \over
4}\left(K_{C}{}^{D}K_{D}{}^{C} + L_{a}{}^{b}L_{a}{}^{b}\right)
\right) = {\sqrt{G} \over M^{n-2}}\left( T_{r}{}^{r} - {T \over
n-2}\right) ,$$
where primes denote derivatives with respect to the radial coordinate $r$. $M$ is the $n$-dimensional fundamental mass scale. Note that equation ($\ref{eq:b7}$) does not exist for codimension one branes; we will return to this specific case shortly.
We now assume that derivatives tangential to the brane (that is derivatives with respect to the $x_{A}$ coordinates) will be negligible compared to derivatives in the brane-orthogonal directions. This implies that we can neglect the term $R^{(g)}{}_{A}{}^{B} / f$ in ($\ref{eq:b6}$), which then becomes
$$\label{eq:b90} \sqrt{G}R_{A}{}^{B} \approx {1 \over
2}\left(\sqrt{G}K_{A}{}^{B}\right)' = {\sqrt{G} \over M^{n-2}}\left(
T_{A}{}^{B} - \delta_{A}{}^{B}{T \over n-2}\right)$$
and we see that the ($A,B$) components of the Ricci tensor can be approximately written as a total derivative with respect to $r$. The next step is to integrate this expression over the brane thickness $0<r<\epsilon$. The integral of $(\sqrt{G}K_{A}{}^{B})'$ over the brane thickness will give us two terms, one at $r=0$ and one at $r= \epsilon$. We must choose our boundary conditions at $r=0$ carefully so that the metric is regular there, since a poor choice will generally create naked singularities in curvature invariants at the origin. For this reason we choose $\alpha(0) = 0$ and $\partial_{r} \alpha|_{r=0} = 1$, since this corresponds to the $m$ extra dimensions taking the form
$$ds^{2}_{[ m ]} = -dr^{2} - \alpha^{2}(r) \gamma_{ab}
dy^{a}dy^{b} = -dr^{2} - r^{2} d\Omega^{2}_{[m-1]},$$
at $r=0$. In other words, the extra dimensions are simply Minkowski space at the origin, and hence the metric is regular here. Integrating ($\ref{eq:b90}$) and using these boundary conditions, we obtain
$$\label{eq:b9} K_{A}{}^{B}|_{\epsilon} = {2M_{\rm b}^{m-1} \over
\Omega^{[m-1]}M^{n-2}}\left( \tilde{T}_{A}{}^{B} - \delta_{A}{}^{B}{\tilde{T}
\over n-2}\right) ,$$
where we have defined $M_{\rm b} = \alpha(\epsilon)^{-1}$. This result has been derived previously in Refs. [@n01], [@n02] for codimension one and two branes, and an explanation as to the relationship between these total derivatives and the underlying symmetries of the background space is given in Ref. [@rn00]. Note that if we consider the case $m=1$, and take the limit $\epsilon \to 0$ we recover the five dimensional junction conditions.
We now consider ($\ref{eq:b7}$). Unlike ($\ref{eq:b6}$), there is no total derivative in ($\ref{eq:b7}$), due to the presence of the $\sqrt{G}
R^{(\gamma)}{}_{a}{}^{b} / \alpha^{2}$ term. To proceed, we integrate ($\ref{eq:b7}$) over the region of the transverse space occupied by the brane,
$$\label{eq:b27} {1 \over 2}\Omega^{[m-1]} \int_{0}^{\epsilon}
(\sqrt{G}L_{a}{}^{b})' d r = {\Omega^{[m-1]} \over 2 M_{\rm
b}^{m-1}}\sqrt{g}|_{\epsilon} L_{a}{}^{b}|_{\epsilon} = {1 \over
M^{n-2}}\int d^{m-1}y dr \sqrt{G}\left(T_{a}{}^{b} -
\delta_{a}{}^{b}{T \over n-2}\right) - (m-2)\delta_{a}^{b}\int
d^{m-1}y dr \sqrt{G}{1 \over \alpha^{2}} .$$
which is valid for $m > 2$. The problematic term is the last one on the right hand side of ($\ref{eq:b27}$). Since we are only integrating over the small transverse region occupied by the brane (that is in the region $r<\epsilon $), we can find an approximate solution to the above equation. To do so, we note that we have chosen our boundary conditions at $r=0$ such that $\alpha(r)
= r$ for small $r$. In addition, we also impose the boundary condition ${d f \over dr}|_{r=0} = 0$, which is also required for a regular solution. We will assume that over the small brane region $(0,\epsilon)$, that $\alpha \approx r$ and $f(r) \approx {\rm
const}$. In making this approximation we find
$$\label{eq:nc1} L_{a}{}^{b}|_{r=\epsilon} \approx {2M_{\rm b}^{m-1}
\over \Omega^{[m-1]}M^{n-2}}\left( \tilde{T}_{a}{}^{b} -
\delta_{a}{}^{b}{\tilde{T} \over n-2} \right) - {2 \over
\epsilon}\delta_{a}{}^{b} ,$$
which is valid for $m > 2$. For the cases $m=1,2$, the last integral in ($\ref{eq:b27}$) is not defined; it diverges at $r=0$. However, this term is absent when we consider codimension one and two branes. We will consider the specific examples of $m=1,2$ shortly.
We now have approximate expressions for $K_{A}{}^{B}$ and $L_{a}{}^{b}$ at $r=\epsilon$, ($\ref{eq:b9}$) and ($\ref{eq:nc1}$). The next step is to consider the $(r,r)$ Einstein equation,
$$\label{eq:mm2} G_{r}{}^{r} = R_{r}{}^{r} - {1 \over 2}R
= {1 \over 8}\left(K_{C}{}^{D}K_{D}{}^{C} +
L_{a}{}^{b}L_{a}{}^{b}\right) - {1 \over 8}\left( L^{2} +
K^{2}\right) - {1 \over 4}KL - {1 \over 2}\left({R^{(g)} \over f} +
{R^{(\gamma)} \over \alpha^{2}}\right) = {T_{r}{}^{r}|_{{\rm bulk}}
\over M^{n-2}}.$$
To obtain an effective four dimensional field equation from ($\ref{eq:mm2}$), we should perform our regularization procedure and integrate it over the range $r=(0,\epsilon)$. However, for simplicity we will instead evaluate ($\ref{eq:mm2}$) at the surface $r = \epsilon$. We stress that this is only an approximation, and we should apply our regularization procedure to ($\ref{eq:mm2}$). However, given our particular profile function, we expect that evaluating ($\ref{eq:mm2}$) at $r=\epsilon$ is sufficient to obtain an approximate brane equation. The reason why we use this approximation is that by virtue of the ‘junction conditions’ ($\ref{eq:nc1}$) and ($\ref{eq:b9}$), we know $K_{A}{}^{B}$ and $L_{a}{}^{b}$ in terms of the brane energy momentum tensor $\tilde{T}_{\mu}{}^{\nu}$ at $r=\epsilon$. We consider the cases $m=1$, $m=2$ and $m>2$ separately.
Codimension one
---------------
For $m=1$ there are no $L_{a}{}^{b}$ components of the extrinsic curvature, and in addition there is no $R^{(\gamma)}/ \alpha^{2}$ term in ($\ref{eq:mm2}$). Therefore using ($\ref{eq:b9}$) and evaluating ($\ref{eq:mm2}$) at $r=\epsilon$, we find
$$\label{eq:np197} - {1 \over 2}{R^{(g)} \over f(\epsilon)} +
{A_{1}^{2} \over 8 (n-2)^{2}} \left[
\left((n-2)^{2}\tilde{T}_{A}{}^{B}\tilde{T}_{B}{}^{A} +
B_{1}(\tilde{T}_{A}{}^{A})^{2}\right) + \left(2B_{1}
\tilde{T}_{r}{}^{r} \tilde{T}_{A}{}^{A} +
C_{1}(\tilde{T}_{r}{}^{r})^{2}\right)\right] = -{\Lambda \over
M^{n-2}} ,$$
where the constants $A_{1}$, $B_{1}$ and $C_{1}$ are given by
$$A_{1}
= {1 \over M^{n-2}}, \qquad \qquad \qquad B_{1} = 1-p, \qquad
\qquad \qquad C_{1} = p-p^{2},$$
We note that the brane energy momentum tensor now has a non-zero component $\tilde{T}_{r}^{r}$ in the transverse direction. This component is zero for a thin brane.
Codimension two
---------------
The case $m=2$ is unique; to see why we return to equation ($\ref{eq:b7}$), and integrate over the region $r<\epsilon$. Evaluating this integral yields two terms, one at $r=0$ and the other at $r=\epsilon$,
$$\label{eq:bb1}{\Omega^{[m-1]} \over 2}
\int_{0}^{\epsilon} (\sqrt{G}L_{a}{}^{b})' dr = {\Omega^{[m-1]} \over
2}\left[ (\sqrt{G}L_{a}{}^{b})|_{r=\epsilon} -
(\sqrt{G}L_{a}{}^{b})|_{r=0}\right] .$$
When $m>2$ the term $(\sqrt{G}L_{a}{}^{b})|_{r=0}$ in ($\ref{eq:bb1}$) is zero by virtue of the boundary condition $\alpha(0)=0$. However, for $m=2$, $(\sqrt{G}L_{a}{}^{b})|_{r=0}
\neq 0$ and we have to include a boundary term at the origin. Hence, accounting for this extra term, equation ($\ref{eq:mm2}$), evaluated at $r=\epsilon$, is
$$\label{eq:np10} - {1 \over 2}{R^{(g)} \over
f(\epsilon)} + {A_{2}^{2} \over 8 (n-2)^{2}} \left[
\left((n-2)^{2}\tilde{T}_{A}{}^{B}\tilde{T}_{B}{}^{A} +
B_{2}(\tilde{T}_{A}{}^{A})^{2}\right) +
\left((n-2)^{2}\tilde{T}_{a}{}^{b}\tilde{T}_{b}{}^{a} +
B_{2}(\tilde{T}_{a}{}^{a})^{2}\right) \right]$$
$$+{A_{2}^{2} \over 8 (n-2)^{2}} \left[2B_{2}\left( \tilde{T}_{a}{}^{a} +
\tilde{T}_{r}{}^{r}\right) \tilde{T}_{A}{}^{A} + 2B_{2}
\tilde{T}_{a}{}^{a}\tilde{T}_{r}{}^{r} +
C_{2}(\tilde{T}_{r}{}^{r})^{2}\right]$$
$$+{1 \over 2}A_{2} M_{\rm
b}^{m-1}{\sqrt{g}|_{r=0} \over
\sqrt{g}|_{r=\epsilon}}\tilde{T}_{a}{}^{a} + {1 \over 2}A_{2}
M_{b}^{m-1}{\sqrt{g}|_{r=0} \over \sqrt{g}|_{r=\epsilon}}
\tilde{T}_{r}{}^{r} = -{\Lambda \over M^{n-2}},$$
where the constants $A_{2}$, $B_{2}$ and $C_{2}$ are given by
$$A_{2}
= { M_{{\rm b}} \over \pi M^{n-2}}, \qquad \qquad \qquad B_{2} =
-p, \qquad \qquad \qquad C_{2} = p-p^{2} -2p.$$
We have confirmed that equations ($\ref{eq:np197}$), ($\ref{eq:np10}$) agree with the results of Refs. [@n01] and [@n02].
Higher codimension
------------------
Now, the term $R^{(\gamma)}/\alpha^{2}$ is not zero, and we find the relation
$$\label{eq:np100} - {R^{(g)} \over
2 f(\epsilon)} + {A_{m}^{2} \over 8 (n-2)^{2}}
\left[\left((n-2)^{2}\tilde{T}_{A}{}^{B}\tilde{T}_{B}{}^{A} +
B_{m}(\tilde{T}_{A}{}^{A})^{2}\right) +
\left((n-2)^{2}\tilde{T}_{a}{}^{b}\tilde{T}_{b}{}^{a} +
B_{m}(\tilde{T}_{a}{}^{a})^{2}\right)\right]$$
$$+ {A_{m}^{2} \over 8 (n-2)^{2}}\left[ 2B_{m}\left( \tilde{T}_{a}{}^{a} +
\tilde{T}_{r}{}^{r}\right) \tilde{T}_{A}{}^{A} + 2B_{m}
\tilde{T}_{a}{}^{a}\tilde{T}_{r}{}^{r} +
C_{m}(\tilde{T}_{r}{}^{r})^{2} \right]$$
$$- {A_{m} \over
2\epsilon}\tilde{T}_{a}{}^{a} - {A_{m}(m-1) \over
2\epsilon}\tilde{T}_{r}{}^{r} - {(m-1)(m-2) \over \epsilon^{2}}=
-{\Lambda \over M^{n-2}}.$$
where $A_{m}$, $B_{m}$ and $C_{m}$ are given by
$$A_{m}
= {2 M_{{\rm b}}^{m-1} \over \Omega^{[m-1]} M^{n-2}}, \qquad \qquad
\qquad B_{m} = 2-m-p, \qquad \qquad \qquad C_{m} =
p-p^{2}+(m-1)-(m-1)^{2} -2p(m-1),$$
for $m>2$. The equation ($\ref{eq:np100}$) gives the brane Ricci scalar $R^{(g)}$ in terms of the bulk cosmological constant $\Lambda$ and the ’four dimensional’ energy momentum tensor $\tilde{T}_{\mu\nu}$. We will return to equations ($\ref{eq:np197}$), ($\ref{eq:np10}$) and ($\ref{eq:np100}$) shortly.
Next, we consider the $(A,a)$, $(A,r)$ components of $R_{\mu\nu}$ evaluated at $r=\epsilon$, which are simply given by $R_{A a} = 0$, and $ R_{A r} = 0$. For our particular metric, $R_{A r}=0$ tells us that $G_{A r} = 0$, which implies $T_{A r}|_{{\rm bulk}} = 0$. As this component of the energy momentum tensor describes the flow of energy from the brane to the bulk, it appears that for braneworld models with a metric given by ($\ref{eq:1}$) we will obtain no loss of energy into the bulk. However this conclusion is not necessarily true if the brane thickness is variable, as was discussed in Ref. [@n01].
Special cases of static braneworlds
===================================
The equations ($\ref{eq:np197}$), ($\ref{eq:np10}$) and ($\ref{eq:np100}$) that were derived in the previous section can act as generalized fine tuning relations for branes of codimension $m=1$, $m=2$ and $m>2$ respectively. To see this, we set $R^{(g)} = 0$. Taking the $m=1$ case as an example, we therefore see that the fine tuning relation
$$\label{eq:ft1}
\left[
\left((n-2)^{2}\tilde{T}_{A}{}^{B}\tilde{T}_{B}{}^{A} +
B_{1}(\tilde{T}_{A}{}^{A})^{2}\right) + \left(2B_{1} \tilde{T}_{r}{}^{r}
\tilde{T}_{A}{}^{A} + C_{1}(\tilde{T}_{r}{}^{r})^{2}\right)\right] = -{8
(n-2)^{2} \over A_{1}^{2} M^{n-2}} \Lambda$$
must hold to ensure the brane is flat. If we take the background space to be five dimensional, and consider a pure tension brane, then when we take the $\epsilon \to 0$ limit, we find $\tilde{T}_{r}{}^{r} \to 0$ and $\tilde{T}_{A}{}^{B} \to -{1 \over
2}\delta_{A}{}^{B} \lambda$, where $\lambda$ is the brane tension. Using these results in ($\ref{eq:ft1}$), we find the fine tuning relation
$$\label{eq:ft2} \Lambda = {\lambda^{2} \over 24
M^{3}},$$
which is equivalent to the standard Randall Sundrum fine tuning condition. There is a sign discrepancy between this fine tuning and the RS model, but this is because we are considering the metric,
$$ds^{2}_{[5]} = f(r) g_{\mu\nu}dx^{\mu}dx^{\nu} -
dr^{2} .$$
which is five dimensional de Sitter space, not AdS as in the RS model. If we consider an AdS space, we set $\Lambda \to -\Lambda$ and the correct fine tuning is recovered.
Whilst ($\ref{eq:ft2}$) is valid for a thin brane, for a thick brane we have the fine tuning relation ($\ref{eq:ft1}$), which we expect will depend on the brane profile. For a pure tension brane, we can write $\hat{T}_{A}^{B} = -{1 \over 2}\lambda \delta_{A}^{B}
D_{\epsilon}(y)$, where $D_{\epsilon}(y)$ is the brane profile. If we normalize the brane profile appropriately such that
$$\int_{-\epsilon}^{\epsilon} \sqrt{G}
D_{\epsilon}(y) dy = 1 ,$$
then we can write the effective four dimensional energy momentum tensor as $\tilde{T}_{A}^{B} = -{1 \over 2} \lambda
\delta_{A}^{B}$, as before. It follows that the fine tuning relation ($\ref{eq:ft1}$) has the form
$$\Lambda = {\lambda^{2} \over 24M^{3}} - {
\lambda \tilde{T}_{r}{}^{r} \over 6M^{3}} +
{(\tilde{T}_{r}{}^{r})^{2} \over 6M^{3}}.$$
We see for a thick brane the fine tuning relation is modified by the $\tilde{T}_{r}{}^{r}$ components of the brane energy momentum tensor.
We can continue in this manner and consider the codimension two case. We first consider a thin, pure tension codimension two brane for which we have $\tilde{T}_{a}{}^{b} = \tilde{T}_{r}{}^{r} =
0$. For pure tension branes, we can use the property $\tilde{T}_{A}{}^{B} = \delta_{A}{}^{B} \tilde{T}_{C}{}^{C}/4$. If we use this relation, and set $m=2$, $p=4$ and $n=6$ in ($\ref{eq:np10}$), we find that if we take $R^{(g)} = 0$, then all of the terms on the left hand side are zero. We are left simply with $\Lambda = 0$ at the position of the brane. If we consider instead a more general bulk energy momentum tensor $T_{r}{}^{r}|^{\rm bulk}$, then this result becomes $T_{r}{}^{r}|^{\rm bulk} = 0$. In [@sc1], a six dimensional model was considered, and it was found that the bulk energy momentum tensor had to be tuned in such a way that $T_{r}{}^{r}|^{\rm bulk}=0$ in order to have a flat brane. Hence our result is in agreement with [@sc1], and reflects that fact that to obtain a flat pure tension codimension two brane, the brane tension does not have to be fine tuned to the bulk energy momentum tensor.
Next, we consider a thick brane, so we cannot assume $\tilde{T}_{a}{}^{b}=\tilde{T}_{r}{}^{r} = 0$. We find that ($\ref{eq:np10}$) now reads
$$\label{eq:ff10} {A_{2}^{2} \over 8 (n-2)^{2}} \left[
\left((n-2)^{2}\tilde{T}_{a}{}^{b}\tilde{T}_{b}{}^{a} +
B_{2}(\tilde{T}_{a}{}^{a})^{2}\right)+ 2B_{2}\left(
\tilde{T}_{a}{}^{a} + \tilde{T}_{r}{}^{r}\right) \tilde{T}_{A}{}^{A}
+ 2B_{2} \tilde{T}_{a}{}^{a}\tilde{T}_{r}{}^{r} +
C_{2}(\tilde{T}_{r}{}^{r})^{2} \right] +$$
$$+{1 \over 2}A_{2} M_{\rm
b}^{m-1}{\sqrt{g}|_{r=0} \over
\sqrt{g}|_{r=\epsilon}}\tilde{T}_{a}{}^{a} + {(m-1) \over 2}A_{2}
M_{{\rm b}}^{m-1}{\sqrt{g}|_{r=0} \over \sqrt{g}|_{r=\epsilon}}
\tilde{T}_{r}{}^{r} = -{\Lambda \over M^{n-2}}.$$
Now if we assume that the $\tilde{T}_{a}{}^{b}$ and $\tilde{T}_{r}{}^{r}$ components of the brane energy momentum tensor are small, then we find the following approximate fine tuning relation that must be satisfied in order for a thick codimension two brane to be flat,
$$\label{eq:ff11} -{M_{\rm b}^{2} \over 16
\pi^{2}M^{8}}\left( \tilde{T}_{a}{}^{a} +
\tilde{T}_{r}{}^{r}\right) \tilde{T}_{A}{}^{A} + {M_{\rm b}^{2}
\over 2\pi M^{4}} {\sqrt{g}|_{r=0} \over \sqrt{g}|_{r=\epsilon}}
\left(\tilde{T}_{a}^{a} + \tilde{T}_{r}^{r}\right) = -{\Lambda \over
M^{4}}.$$
where we have taken $p=4$. We see that to obtain a flat brane in this case, we must tune $\tilde{T}_{A}{}^{A}$ to the bulk energy momentum tensor (unless we have a model in which $\tilde{T}_{a}{}^{a} = - \tilde{T}_{r}{}^{r}$, in which case we have $\Lambda \approx 0$, or more generally $T_{r}{}^{r}|^{\rm bulk} =
0$, as in the thin brane case.)
\[sec:1\]Cosmological Solution
==============================
Having considered metrics of the form ($\ref{eq:1}$) in the previous section, we now perform an analogous calculation for the time-dependent metric ($\ref{eq:197}$). We initially incorporate a time dependence in the extra dimensions by writing $\alpha(r,t)$ as a function of time. However, to perform the same calculation as above, we find that we must remove this time dependence. This amounts to assuming that the extra dimensions are static.
Calculating the Ricci tensor $R_{\mu}{}^{\nu}$ for the metric ($\ref{eq:197}$) as we did in the previous section gives
$$\label{eq:c1} R_{i}{}^{j} =
{1 \over 2}{\left(N A^{p-1} \alpha^{m-1} K_{i}{}^{j}\right)' \over N
A^{p-1}\alpha^{m-1}} + {R^{(g)}{}_{i}{}^{j} \over A^{2}} - {(m-1)
\over \alpha}\nabla^{j}\nabla_{i}\alpha = {1 \over
M^{n-2}}\left(T_{i}{}^{j} - \delta_{i}{}^{j}{T \over
n-2}\right),$$
$$\label{eq:c2}R_{t}{}^{t} =
{1 \over 2}{\left(N A^{p-1} \alpha^{m-1} K_{t}{}^{t}\right)' \over N
A^{p-1}\alpha^{m-1}} + {R^{(g)}{}_{t}{}^{t} \over A^{2}} - {(m-1)
\over \alpha}\nabla^{t}\nabla_{t}\alpha = {1 \over
M^{n-2}}\left(T_{t}{}^{t} - \delta_{t}{}^{t}{T \over
n-2}\right),$$
$$\label{eq:c3} R_{a}{}^{b} = {1 \over
2}{\left(N A^{p-1}\alpha^{m-1} L_{a}{}^{b}\right)' \over N
A^{p-1}\alpha^{m-1}} - {(m-2) \over \alpha^{2}}\delta_{a}{}^{b}
\nabla_{\mu} \alpha \nabla^{\mu} \alpha - {\delta_{a}{}^{b} \over
\alpha} \nabla^{\mu}\nabla_{\mu}\alpha + {(m-2) \over
\alpha^{2}}\delta_{a}{}^{b}$$
$$= {1 \over M^{n-2}}\left(T_{a}{}^{b} -
\delta_{a}{}^{b}{T \over n-2}\right),$$
$$R_{r}{}^{r} = {1 \over 2}\left( K' + L'\right)
+ {1 \over 4} \left(K_{A}{}^{B}K_{B}{}^{A} +
L_{a}{}^{b}L_{b}{}^{a}\right) .$$
where the extrinsic curvatures are defined in ($\ref{eq:si1}$). Primes denote derivatives with respect to the radial coordinate $r$, and dots derivatives with respect to time. We have defined the Ricci tensors of the metrics $g_{ij}$ and $\gamma_{ab}$ as $R^{(g)}{}_{i}{}^{j}$ and $R^{(\gamma)}{}_{a}{}^{b}$ respectively, and the covariant derivatives $\nabla_{\mu}$ preserve the metric $g_{\mu\nu}$.
The final equation that we will need is
$$\label{eq:b15} G_{A r} = R_{A r} = - {
(m-1)\partial_{A} \alpha' \over \alpha} + {(m-1) \partial_{B} \alpha
\over 2\alpha} K^{B}{}_{A} + {1 \over 2} \nabla^{B}(K_{AB} - g_{AB}K
)= {T_{A r}^{{\rm bulk}} \over M^{n-2}} .$$
which will give us a conservation equation for the brane energy momentum tensor $\tilde{T}_{\mu\nu}$.
We will now make a number of assumptions in order to obtain an equation for $A(t,r)$ at the surface of the brane. We begin as before, by assuming that brane tangential derivatives are small, in particular we will neglect the terms $\nabla^{j}\nabla_{i}\alpha$ and $\nabla_{t}\nabla^{t}\alpha$ in equations ($\ref{eq:c1}$), ($\ref{eq:c2}$) and ($\ref{eq:c3}$). With this assumption we have removed the time dependence of the extra dimensions, and so our solution will be valid only when the transverse dimensions are static or weakly time dependent.
We proceed as in the previous section. We neglect derivatives tangential to the brane, and integrate ($\ref{eq:c1}$-$\ref{eq:c3}$) over the brane thickness $0<r<\epsilon$, using the approximations $N(t,r=0) \approx N(t,r=\epsilon)=1$ and $\alpha(t,r<\epsilon)
\approx r$. We find
$$\label{eq:b17} K_{i}{}^{j}|_{r=\epsilon} = {2 M_{\rm b}^{m-1}
\over \Omega^{[m-1]}M^{n-2}}\left(\tilde{T}_{i}{}^{j} -
\delta_{i}{}^{j}{\tilde{T} \over n-2}\right) ,$$
$$\label{eq:b18} K_{t}{}^{t}|_{r=\epsilon} = {2 M_{\rm b}^{m-1}
\over \Omega^{[m-1]}M^{n-2}}\left(\tilde{T}_{t}{}^{t} -
\delta_{t}{}^{t}{\tilde{T} \over n-2}\right) ,$$
$$\label{eq:b19} L_{a}{}^{b}|_{r=\epsilon} = {2 M_{\rm b}^{m-1}
\over \Omega^{[m-1]}M^{n-2}}\left(\tilde{T}_{a}{}^{b} -
\delta_{a}{}^{b}{\tilde{T} \over n-2}\right) -{2 \over
\epsilon}\delta_{a}{}^{b},$$
which are valid for $m \ge 3$. As in the previous section, for codimension two objects we obtain an additional term in ($\ref{eq:b19}$), a boundary term at $r=0$ of the form $\left(\sqrt{G} L_{a}{}^{b}\right)|_{r=0}$. We will check throughout though that our results can give the results of [@n01] for a codimension two brane.
The next step is to evaluate the $(r,r)$ Einstein equation at $r=\epsilon$,
$$\label{eq:b20} G_{r}{}^{r}|_{r=\epsilon} = -{1 \over 2}R^{(g)}-{1
\over 2}{(m-1)(m-2) \over \alpha^{2}} + {1 \over 8}\left(
K_{C}{}^{D}K_{D}{}^{C} - K^{2}\right) + {1 \over 8}\left(
L_{a}{}^{b}L_{b}{}^{a} - L^{2} \right) - {1 \over 4}L K = -{\Lambda
\over M^{n-2}} .$$
We assume that $A(r=0) \approx A(r=\epsilon) = a(t)$ is the brane scale factor. We also set the energy momentum tensor on the brane $\tilde{T}_{\mu}{}^{\nu}$ as
$$\label{eq:1000} \tilde{T}_{\mu}{}^{\nu} = {\rm
diag}\left(\rho,-p_{\rm br},-p_{\rm br},-p_{\rm br},-p_{\rm r},-p_{\rm
bk},..,-p_{\rm bk}\right) ,$$
where $\rho$ is the energy density, $p_{\rm br}$ the brane pressure components, and $p_{\rm r}$ and $p_{\rm bk}$ the extra dimensional pressure components. There are $(m-1)$ $p_{\rm
bk}$ terms in ($\ref{eq:1000}$). Using ($\ref{eq:1000}$) in ($\ref{eq:b17}$-$\ref{eq:b19}$), the matching conditions become
$$\label{eq:b21} K_{i}{}^{j}|_{r=\epsilon} = -{2 M_{\rm b}^{m-1}
\over \Omega^{[m-1]}M^{n-2}(n-2)}\delta_{i}{}^{j}\left[
(n-1-p)p_{\rm br} +\rho - p_{\rm r} - (m-1)p_{\rm bk} \right] ,$$
$$\label{eq:b22} K_{t}{}^{t}|_{r=\epsilon} = {2 M_{\rm b}^{m-1}
\over \Omega^{[m-1]}M^{n-2}(n-2)}\left[ (n-3)\rho + (p-1)p_{\rm br}
+ p_{\rm r} + (m-1)p_{\rm bk} \right] ,$$
$$\label{eq:b23} L_{a}{}^{b}|_{r=\epsilon} = -{2 M_{\rm b}^{m-1}
\over \Omega^{[m-1]}M^{n-2}(n-2)}\delta_{a}{}^{b}\left[
(n-1-m)p_{\rm bk} + \rho - (p-1)p_{\rm br} -p_{\rm r} \right] -{2
\over \epsilon}\delta_{a}{}^{b} ,$$
and the $G_{r}^{r}$ Einstein equation is,
$$\label{eq:b25} G_{r}{}^{r}|_{r=\epsilon} =
3\left({\ddot{a}\over a}+ \left({\dot{a}
\over a}\right)^{2}\right)+ {A_{m}^{2} \over 8 (n-2)^{2}}
\left((n-2)^{2}\tilde{T}_{A}{}^{B}\tilde{T}_{B}{}^{A} +
B_{m}(\tilde{T}_{A}{}^{A})^{2} +
(n-2)^{2}\tilde{T}_{a}{}^{b}\tilde{T}_{b}{}^{a} +
B_{m}(\tilde{T}_{a}{}^{a})^{2}\right)$$
$$+ {A_{m}^{2} \over 8 (n-2)^{2}}\left(2B_{m}\left( \tilde{T}_{a}{}^{a} +
\tilde{T}_{r}{}^{r}\right) \tilde{T}_{A}{}^{A} + 2B_{m}
\tilde{T}_{a}{}^{a}\tilde{T}_{r}{}^{r} +
C_{m}(\tilde{T}_{r}{}^{r})^{2}\right) - {A_{m} \over 2
\epsilon}\tilde{T}_{a}{}^{a} - {(m-1)A_{m} \over 2\epsilon}
\tilde{T}^{r}{}_{r}- { (m-1)(m-2)\over \epsilon^{2}} = -{\Lambda
\over M^{n-2}}.$$
Now, using the brane energy momentum tensor ($\ref{eq:1000}$), we obtain the following equation for the brane scale factor
$$\label{eq:b30} 3\left({\ddot{a}\over a}+ \left({\dot{a}
\over a}\right)^{2}\right) = -{\Lambda \over M^{n-2}} - {A_{m}^{2}
\over 8(m+2)}\left[(1+m)(\rho + p_{{\rm br}} )^{2} + (2m-4) p_{{\rm
br}} ( p_{{\rm br}} - \rho) +3(m-1) p_{{\rm bk}}^{2} \right]+$$
$$-{A_{m}^{2} \over
8(m+2)} \left[ (2(m-1) p_{{\rm bk}} +2 p_{{\rm r}} )(\rho - 3
p_{{\rm br}} ) - 2(m-1) p_{{\rm bk}} p_{{\rm r}} - (m+3) p_{{\rm
r}}^{2} \right]$$
$$+ {(m-1)(m-2) \over \epsilon^{2}} -
{(m-1)A_{m} \over 2\epsilon}(p_{\rm r}+p_{\rm bk}) .$$
This equation describes the evolution of the scale factor $a(t)$ at the surface $r = \epsilon$ of a four dimensional, codimension $m$ brane, subject to the approximations and assumptions that we have made thus far. Equation ($\ref{eq:b30}$) is valid for $m>2$; as we have discussed in the previous section the codimension one and two cases are unique and should be considered separately. Before continuing with the general case, we briefly discuss the previously studied $m=1,2$ examples.
For the codimension one case, there is no $L_{a}{}^{b}$ field equation, and ($\ref{eq:b30}$) reads
$$\label{eq:bb30} 3\left({\ddot{a}\over a}+ \left({\dot{a}
\over a}\right)^{2}\right) = -{\Lambda \over M^{n-2}} - {1 \over
12M^{6}}\left[(\rho + p_{{\rm br}} )^{2} - p_{{\rm br}} ( p_{{\rm
br}} - \rho) + p_{{\rm r}} (\rho - 3 p_{{\rm br}} ) - 2 p_{{\rm
r}}^{2} \right],$$
which agrees with the results of Ref. [@n02], where a thick codimension one brane was considered. Next, we consider the codimension two case. Once again, a slight complication arises from a boundary term at $r=0$. We write the modified $(a,b)$ field equation as
$$\label{eq:md23} L_{a}{}^{b}|_{r=\epsilon} = -{ M_{{\rm b}}
\over 4\pi M^{4}}\delta_{a}{}^{b}\left[ 3p_{\rm bk} + \rho - 3p_{\rm
br} -p_{\rm r} \right] + {\sqrt{g}|_{r=0} \over
\sqrt{g}|_{r=\epsilon}}M_{{\rm b}},$$
and hence for $m=2$ the equation ($\ref{eq:b20}$) becomes
$$\label{eq:bbb30} 3\left({\ddot{a}\over a}+ \left({\dot{a}
\over a}\right)^{2}\right) = -{\Lambda \over M^{4}} - {M_{{\rm
b}}^{2} \over 32\pi^{2}M^{8}}\left[3(\rho + p_{{\rm br}} )^{2}
+3
p_{{\rm bk}}^{2} + 2(p_{{\rm bk}} +p_{{\rm r}})(\rho - 3 p_{{\rm
br}} ) - 2 p_{{\rm bk}} p_{{\rm r}} - 5 p_{{\rm r}}^{2} \right]$$
$$+
{M_{{\rm b}}^{2} \over 2\pi M^{4}}(p_{\rm r}+p_{\rm
bk}){\sqrt{g}|_{r=0} \over \sqrt{g}|_{r=\epsilon}}$$
Once again, this result agrees with the results in Ref. [@n01].
Having checked that our equation ($\ref{eq:b30}$) agrees with the well studied cases $m=1,2$, we can now consider the general case $m>2$. The brane energy momentum tensor contains the usual four dimensional energy density and pressure terms $\rho$ and $p_{{\rm
br}}$, and in addition non-zero pressure terms $p_{r}$ and $p_{{\rm
bk}}$ in the transverse directions. If we assume that $p_{{\rm r}}$ and $p_{{\rm bk}}$ are constant across the brane, we can write ($\ref{eq:b30}$) as
$$\label{eq:b31} 3\left({\ddot{a}\over a}+ \left({\dot{a}
\over
a}\right)^{2}\right) = \omega_{1} + \omega_{2}(\rho + p_{{\rm br}}
)^{2} + \omega_{3}p_{{\rm br}} ( p_{{\rm br}} - \rho) +
\omega_{4}(\rho - 3 p_{{\rm br}} )$$
where the constants $\omega_{1,2,3,4}$ are given by
$$\begin{aligned}
\omega_{1} &= -{\Lambda \over
M^{n-2}} - {A_{m}^{2} \over 8(m+2)} \left[ 3(m-1)p_{{\rm bk}}^{2} -
2(m-1)p_{{\rm bk}}p_{{\rm r}} - (m+3)p_{{\rm r}}^{2} \right] \\
&+ {(m-1)(m-2) \over \epsilon^{2}} -
{(m-1)A_{m} \over 2\epsilon}(p_{r}+p_{bk}), \\
\omega_{2} &= -{A_{m}^{2}(1+m) \over 8 (m+2)} , \\
\omega_{3} &= -{ A_{m}^{2}(m-2) \over 4 (m+2)}, \\
\omega_{4} &= -{ A_{m}^{2} \over 4
(m+2)}[(m-1)p_{{\rm bk}} + p_{{\rm r}}] .\end{aligned}$$
From ($\ref{eq:b30}$), we see that in addition to the standard $(\rho
- 3p_{{\rm br}})$ term, we also have the quadratic terms $(\rho
+ p_{{\rm br}})^{2}$ and $p_{{\rm br}}(p_{{\rm br}} -\rho)$. If we take $\rho = T + \rho_{m}$ and $p_{{\rm br}} = -T + w \rho_{m}$, where $\rho_{m}$ is a small energy component with arbitrary equation of state $p_{m}=w \rho_{m}$ and $T$ a constant brane tension, we can expand ($\ref{eq:b31}$) in powers of $\rho_{m}$. Doing so we obtain
$$\label{eq:b32} 3\left({\ddot{a}\over a}+ \left({\dot{a}
\over a}\right)^{2}\right) = (\omega_{1} +2T^{2}\omega_{3} +
4\omega_{4}T) + (\omega_{4} + T\omega_{3})(1 - 3w )\rho_{m} +
O(\rho_{m}^{2}),$$
which is the standard cosmological equation when considering late time cosmology (that is when $\rho_{m}$ is small). The cosmological constant problem in this particular model is why the constant $(\omega_{1} +2T^{2}\omega_{3} + 4\omega_{4}T)$ is either zero or very small. There is no reason to expect these terms to cancel one another, suggesting these models are not free from fine tuning. This is not a surprising result, since we had no reason to expect that self tuning behaviour exists in these models.
Thus we have found that late-time cosmology with quadratic corrections is a generic feature of certain braneworld models, regardless of the codimension, generalizing the arguments made in [@n01] and [@n02] for codimension one and two branes. We note that since we have assumed that $p_{r}$ and $p_{\rm bk}$ are constants, then we must necessarily get an expression of the form
$$\label{eq:ex1} 3\left({\ddot{a} \over a}+\left({\dot{a} \over a}\right)^{2}\right) =
\kappa_{1}(p_{r},p_{\rm bk}) + \kappa_{2}(p_{r},p_{\rm bk})(\rho -
3p_{\rm br}) + \kappa_{3}(\rho^{2},p_{br}^{2},\rho p_{br}) +{\Lambda
\over M^{n-2}}$$
where $\kappa_{1},\kappa_{2}$ are functions of $p_{r}$ and $p_{\rm bk}$ only, and hence are constants, and $\kappa_{3}$ is a term quadratic in the variables $\rho, p_{\rm br}$. It was remarked in [@n01] that if we obtained a cosmological equation like
$$\label{eq:ex2} 3\left({\ddot{a} \over a}+\left({\dot{a} \over a}\right)^{2}\right) =
F(\rho+p_{\rm br}) + G(p_{r},p_{\rm bk},\rho,p_{\rm br}) - {\Lambda
\over M^{n-2}}$$
where $F$ and $G$ are functions of the brane energy density, then if $F(\rho+p_{\rm br})$ was a function linear in $(\rho + p_{\rm br})$ then we would have a potential mechanism for self tuning. In our setup, the function $F(\rho +p_{\rm br})$ will always be quadratic in $(\rho+p_{\rm br})$.
The above arguments apply when the bulk energy momentum tensor $T_{r}^{r}|^{\rm bulk}$ is not a function of $\rho, p_{\rm br}$ and $p_{r}, p_{\rm bk}$ are constants. If we instead assume that $p_{r}$ and $p_{\rm bk}$ are related to $\rho$ by the equations of state $p_{r}
= w_{r}\rho$ and $p_{\rm bk} = w_{\rm bk}\rho$, then we would obtain a cosmological model that differs from ($\ref{eq:b32}$) [@n01].
Next, we derive a conservation equation for the brane energy momentum tensor $\tilde{T}_{\mu}^{\nu}$. We do so from equation ($\ref{eq:b15}$), evaluated at $r=\epsilon$. Using our matching conditions at the surface of the $n$-brane, we find that ($\ref{eq:b15}$) reads
$$\label{eq:sww1} {A_{m} \over 2}\nabla^{B}\tilde{T}_{AB} - {A_{m} \over 2(n-2)} (2-m-p)\nabla_{A}\tilde{T}_{r}^{r}
+{(m-1)A_{m} \partial_{B}M_{b} \over 2M_{b}} \left(\tilde{T}_{A}^{B}
- \delta_{A}^{B} {\tilde{T} \over n-2}\right) = {T_{A r}^{\rm
bulk}|_{r=\epsilon} \over M^{n-2}} .$$
We have assumed throughout this paper that the time dependence of the extra dimensions can be neglected to our level of approximation. This means that the $\partial_{A} M_{b}$ term in ($\ref{eq:sww1}$) can be neglected. In addition, we assumed that the $\tilde{T}_{r}^{r}$ component of the brane energy momentum tensor was approximately constant, which means that we can write ($\ref{eq:sww1}$) as
$$\label{eq:sww2} {A_{m} \over 2}( \dot{\rho} + 3{\dot{a}
\over a}(\rho + p_{{\rm br}}))
\approx {T_{t r}^{\rm bulk}|_{r=\epsilon} \over M^{n-2}},$$
and hence, if we have $T_{t r}^{\rm bulk}|_{r=\epsilon} =
0$ then we obtain the standard four dimensional conservation equation
$$\label{eq:sww3} \dot{\rho} + 3{\dot{a}
\over a}(\rho + p_{{\rm br}}) =0 ,$$
However, this is only approximate, subject to the assumption that $\partial_{t} M_{b} \approx 0$. We also note that if $T_{t r}^{\rm bulk}|_{r=\epsilon} \neq 0$, then the brane energy momentum tensor is not strictly conserved, and we can obtain a flow of energy into the bulk. However, if we assume that $T_{t r}^{\rm
bulk}|_{r=\epsilon} = 0$, then from ($\ref{eq:sww3}$) we can write
$$\label{eq:swf1} \rho = \rho_{c} a^{-3(1+w)} ,$$
where $\rho_{c}$ is a constant, and $w = p_{{\rm br}} /
\rho$. If we use ($\ref{eq:swf1}$) in ($\ref{eq:b31}$), then by multiplying the equation by $\dot{a}a^{3}$ we can write the left hand side as a total derivative,
$${3 \over 2}{d \over dt}
\left(\left(\dot{a}a\right)^{2}\right) =
\omega_{1}\dot{a}a^{3} + \left(\omega_{2}(1+w)^{2}+\omega_{3}
w(w-1)\right)\rho_{c}^{2} \dot{a} a^{-6(1+w)+3} +
\omega_{4}(1-3w)\rho_{c}\dot{a} a^{-3(1+w)+3} .$$
This equation can be integrated to give a Friedmann type equation for the brane scale factor,
$$\label{eq:fr1} H^{2} = \left({ \dot{a} \over
a}\right)^{2} = {\omega_{1} \over 6} + {2\omega_{4} \over 3} \rho +
{2\left[\omega_{2}(1+w)^{2}+\omega_{3} w(w-1)\right] \over
3[4-6(1+w)]}\rho^{2} + {C \over a^{4}},$$
where $C$ is an integration constant. The first two terms on the right hand side of ($\ref{eq:fr1}$) are what we would expect for standard four dimensional cosmology. The third term is quadratic in the brane energy momentum tensor, and the fourth term is a ‘dark radiation energy’ component. It is curious that we find a dark radiation contribution from the bulk; we might expect that in higher codimension additional terms would be present in the brane field equations. We will return to this point in the appendix.
\[sec:100\] Discussion
======================
To review our progress; we have written the full $n$-dimensional Ricci tensor and scalar in terms of $R^{(g)}$, $R^{(\gamma)}$ and the extrinsic curvatures $K_{A}{}^{B}$ and $L_{a}{}^{b}$ (see ($\ref{eq:b6}$),($\ref{eq:b7}$) and ($\ref{eq:b8}$)). We then decomposed the total energy momentum tensor into a brane component, strictly localized in the region $r<\epsilon$, and a bulk cosmological constant. Since the brane energy momentum tensor is assumed to localized in the region $r<\epsilon$, we are considering a local defect.
Our work is a particular example of the method outlined in Ref. [@c1], where brane quantities are defined as the projection of higher dimensional terms onto a four dimensional subspace. Doing so for the subspace $r=\epsilon$, we have found the approximate cosmological equation ($\ref{eq:b31}$). In our setup, we see that the four dimensional Planck mass $M_{{\rm pl}}^{2}$ is given by the coefficient of the $\rho$ term in ($\ref{eq:b32}$), that is
$$\label{eq:mp32} M_{{\rm pl}}^{2} = {1
\over 2\left(\omega_{4} + T \omega_{3}\right)}.$$
We note that our definition of the Planck mass differs from the effective action approach considered in (amongst others) Ref. [@gh00], where it is rather defined as
$$\label{eq:mp33} M_{\rm pl}^{2} = M^{m+2} \int \sqrt{G} d^{m}y ,$$
in other words an integral over the $m$ codimensional space, which for the metric considered in this paper is given by
$$M_{\rm pl}^{2} = M^{m+2} \Omega^{[m-1]} \int \alpha^{m-1} dr .$$
The difference between the definitions ($\ref{eq:mp32}$) and ($\ref{eq:mp33}$) reflects the two different philosophies adopted in the literature towards defining four dimensional quantities. In the effective action approach, four-dimensional quantities are obtained by integrating out the extra dimensions in the full $n$-dimensional action. However, in this paper we use a tensorial approach, and project $n$-dimensional tensors tangentially to the brane. For a discussion of the two approaches, see Ref. [@bt01].
As discussed in Refs. [@gh00], [@ben1], the codimension $m \geq 3$ case is different to the more commonly studied $m=1,2$ cases in the literature. To see the difference, we consider the $n$-dimensional Ricci scalar for our static metric ($\ref{eq:1}$),
$$\label{eq:ri1} R = K' + L' +{1 \over 4}\left(K_{A}{}^{B}K_{B}{}^{A} +
L_{a}{}^{b}L_{a}{}^{b}\right) + {1 \over 4} \left( K^{2} +
L^{2}\right) + {1 \over 2} KL + {R^{(g)} \over f} + {(m-1)(m-2)
\over \alpha^{2}}.$$
We see that the last term in ($\ref{eq:ri1}$) is present only when $m>2$, and represents the curvature of the bulk. We note that this is singular if $\alpha=0$ anywhere in the transverse space. Thus to avoid singularities in the curvature invariants, we must impose $\alpha \neq 0$ for all $r>0$. Note that we have imposed $\alpha = 0$ as a boundary condition at $r=0$, however this is simply a coordinate singularity.
As has been discussed in Ref. [@gh00], the last term in ($\ref{eq:ri1}$) is a problem when we consider localizing gravity on the brane. To see why, we consider the Einstein equations ($\ref{eq:b6}$), ($\ref{eq:b7}$) and ($\ref{eq:b8}$) for our metric ($\ref{eq:1}$). In the asymptotic limit $r \to \infty$, we require a solution to the Einstein equations that is Anti-de-Sitter, that is we require $R \sim
{\rm const}$. To obtain a solution of this form all terms on the right hand side of ($\ref{eq:ri1}$) must either asymptote to zero or a constant as $r \to \infty$. We are considering flat branes, so $R^{(g)} =0$. The $K_{A}{}^{B}$ terms asymptote to constants if the warp factor $f(r)$ asymptotes to an exponential, as we expect it to. The problematic term in ($\ref{eq:ri1}$) is the one of the form $(m-1)(m-2) / \alpha^{2}$, from which we deduce that the function $\alpha(r)$ must either asymptote to infinity or a constant in the limit $r \to \infty$. However, we find that no solution exists to the Einstein equations such that $\alpha \to {\rm const}$ as $r \to \infty$ and $f(r)$ is a real exponential function. Hence we are left only with the possibility $\alpha \to \infty$ as $r \to \infty$. This removes the problematic $\alpha^{-2}$ term in ($\ref{eq:b7}$), and we find that a valid solution exists in this case, where both $\alpha(r)$ and $f(r)$ are growing exponentials in $r$ (a result found in Ref. [@io1]). Since in this case both warp factors are increasing functions as $r \to \infty$, it follows that the transverse space has an infinite volume.
Since we are considering local defects of codimension $m>2$, the above analysis applies and we will generically obtain transverse spaces with infinite volume. This infinite volume is a problem, since we find that gravity cannot be localized on the brane in such a setup. The reason why we cannot localize gravity is that the zero mode in the graviton spectrum will not be normalizable. To avoid (but not solve) this problem, we could follow Ref. [@ben1] and introduce an infra-red cutoff when integrating the zero mode over the extra dimensions. This cutoff could arise as an interbrane separation, for example. In doing so, we would then obtain a finite integral over the transverse dimensions, and it would be possible to recover conventional four dimensional gravity. Of course, it would be preferable to obtain a model where gravity can be localized on the brane without the need to introduce a cutoff, and it appears that global defects are better suited to achieve this. Alternatively, we could introduce additional fields in the bulk in an attempt to remove this problematic behaviour. However, bulk fields have the highly undesirable property of inducing a non-trivial Weyl tensor contribution to the field equations on the brane, and a detailed analysis of the bulk would have to be performed in such a setup.
\[sec:2\] conclusion
====================
In this paper we have calculated the evolution equation for the scale factor $a(t)$ of a thick, codimension $m$, $3$-brane. By assuming radial symmetry in the bulk, we have found that many of the Einstein equations approximately admit a first integral, and we have integrated these equations over the brane thickness to obtain a set of approximate junction conditions at the surface of the brane $r=\epsilon$. We then used these junction conditions to write an equation for the evolution of the brane scale factor, $a(t)$, in terms of the brane energy momentum tensor. Since we considered a brane of arbitrary codimension, we were forced to make a large number of assumptions, which we review;
- Derivatives tangential to the brane can be neglected. Specifically, we assumed that the bulk metric has only a weak dependence on time, and is static to the level of approximation that we are working at.
- We assumed that in the core of the brane the metric may be approximately written as
$$\label{eq:ab1} ds^{2} = g_{\mu\nu}(x)dx^{\mu}dx^{\nu} -
dr^{2} -
r^{2} d\Omega^{2}_{[m-1]} .$$
The boundary conditions that we have used give us this form of the metric at the centre of the brane, that is at $r=0$, and we have assumed that the metric can be approximately written as ($\ref{eq:ab1}$) for $r<\epsilon$.
- We defined four dimensional quantities as $n$ dimensional quantities integrated over the brane thickness in the transverse dimensions. For example, the brane energy momentum tensor $\hat{T}_{A}{}^{B}$ is given by
$$\label{eq:cl1} \hat{T}_{A}{}^{B} \equiv \int \sqrt{\gamma}
d^{m-1}y \int_{0}^{\epsilon} \alpha^{m-1}N A^{n} T_{A}{}^{B} dr .$$
This is not an assumption, but rather a definition. However, there is some ambiguity in defining a four dimensional quantity when discussing thick braneworlds, and ($\ref{eq:cl1}$) is not unique.
- We have assumed that a solution to the full $n$-dimensional Einstein equations exists that respects the above assumptions. It is important to stress that we have not found a full solution to the field equations. We expect that a solution exists of the form postulated, subject to the above assumptions.
- We have assumed that the brane thickness is time independent, at least to first order. We expect that any matter on the brane will have a backreaction effect on the brane profile, and we have assumed that this effect is negligible. We have also assumed that this thickness is not as small as any fundamental lengthscale in the model. By this we mean that if the brane is too thin, then we could not use our classical arguments (we would not be able to resolve the thickness of the brane without appealing to quantum mechanics).
Based on these assumptions, we have found the standard cosmological equation plus quadratic terms in the brane energy density for a thick brane of arbitrary codimension. We have also found a general fine tuning condition required to make the effective four dimensional cosmological constant small in our model. It depends on the bulk energy momentum tensor, the Ricci scalar of the transverse dimensions, as well as the brane thickness, tension and transverse energy momentum components $p_{\rm r}$ and $p_{\rm bk}$. We assumed that the bulk energy momentum tensor in our model is simply a constant, and we might expect bulk fields to be present [@gh00]. Introducing new fields into the bulk will affect both the fine tuning relationship and the four dimensional Friedmann equation that we have found. Our final result is the conservation equation for the brane energy momentum tensor, ($\ref{eq:sww1}$). We find that the standard four dimensional conservation equation is obtained, but only if the component $T_{\mu r}^{\rm bulk}|_{r=\epsilon}$ of the bulk energy momentum tensor is zero, a well known result.
Acknowledgements {#acknowledgements .unnumbered}
================
S. A. would like to thank L. Gutteridge for many interesting discussions during the preparation of this manuscript. This work was supported by PPARC.
Appendix: Covariant braneworld equations {#appendix-covariant-braneworld-equations .unnumbered}
========================================
In the appendix, we discuss why we obtain a dark radiation term in the brane Friedmann equation, regardless of the codimension, and thereby produce an alternative derivation of the main results of this paper using tensorial notation. We then discuss the relationship between codimension one and codimension $m$ branes in our setup.
\[sec:f1\] Weyl tensor
----------------------
From our brane Friedmann equation ($\ref{eq:fr1}$), we see that a dark radiation term is present, which is a bulk effect contributing to $H$. We might expect extra terms to be introduced into the Friedmann equation, since we are considering more than one transverse dimension. We now consider the Weyl tensor in detail, to determine the origin of this dark radiation term.
The Weyl tensor $W_{\mu\nu}{}^{\alpha\beta}$ is defined as
$$\label{eq:wi37} W_{\mu\nu}{}^{\rho\sigma} = R_{\mu\nu}{}^{\rho\sigma} - {4 \over n-2}
g^{[\rho}{}_{[\mu}R^{\sigma]}{}_{\nu]} + {2 \over (n-1)(n-2)}R
g^{[\rho}{}_{[\mu}g^{\sigma]}{}_{\nu]} .$$
A certain contraction of this tensor determines the bulks effect on the brane. To see this, we consider the covariant braneworld equations of arbitrary codimension, given by Ref. [@c1]
$$\label{eq:ff200} R_{\mu\nu}^{(p)} = {p-2 \over
n-2}\eta_{\mu}{}^{\rho}\eta_{\nu}{}^{\sigma}R_{\rho\sigma} + {1 \over
n-2} \left( \eta^{\rho\sigma}R_{\rho\sigma} - {p-1 \over
n-1}R\right) \eta_{\mu\nu} + {p-1 \over p^{2}}
\bar{K}^{\sigma}\bar{K}_{\sigma}\eta_{\mu\nu} + {p-2 \over p}
\bar{C}_{\mu\nu}{}^{\sigma}\bar{K}_{\sigma}-
\bar{C}_{\mu}{}^{\rho\sigma}\bar{C}_{\nu\rho\sigma} + W_{\mu\nu},$$
where $\eta_{\mu\nu}$ is a tensor that projects other tensors tangentially to the brane. The Weyl tensor $W_{\mu\nu}$ is given by
$$\label{eq:w1} W_{\mu\nu} = \eta_{\mu}{}^{\sigma}\eta_{\nu}{}^{\kappa}
\eta_{\tau}{}^{\rho}W_{\rho\sigma}{}^{\tau}{}_{\kappa},$$
and $\bar{C}_{\mu\nu}{}^{\rho}$ by
$$\bar{C}_{\mu\nu}{}^{\rho} = \bar{K}_{\mu\nu}{}^{\rho} - {1
\over p} \eta_{\mu\nu} \bar{K}^{\rho} .$$
The extrinsic curvature $\bar{K}_{\mu\nu}{}^{\rho}$ in this notation is
$$\label{eq:ap1}\bar{K}_{\mu\nu}{}^{\rho} =
\eta_{\nu}{}^{\sigma}\eta_{\mu}{}^{\alpha}\nabla_{\alpha}
\eta_{\sigma}{}^{\rho}$$
In this section we use $\bar{K}_{\mu\nu}{}^{\rho}$ as the extrinsic curvature, as opposed to $K_{AB}$ which has been used in this paper. $K_{AB}$ is actually a particular example of the more general $\bar{K}_{\mu\nu}{}^{\rho}$ above. To see this, we write ($\ref{eq:ap1}$) as
$$\begin{aligned}
\label{eq:ex1} \bar{K}_{\mu\nu}{}^{\rho} &=
\eta_{\nu}{}^{\sigma}\eta_{\mu}{}^{\alpha}\perp^{\rho}{}_{\gamma}\nabla_{\alpha}
\eta_{\sigma}{}^{\gamma} \\ &= \eta_{\nu}{}^{\sigma}\eta_{\mu}{}^{\alpha}
\perp^{\rho}{}_{\gamma}\left(\partial_{\alpha}\eta_{\sigma}{}^{\gamma} +
\Gamma^{\gamma}{}_{\alpha\beta}\eta^{\beta}{}_{\sigma} -
\Gamma^{\beta}{}_{\alpha\sigma}\eta_{\beta}{}^{\gamma} \right).\end{aligned}$$
Next, we note that in our coordinate system, $\eta_{\mu\nu} = g_{AB}\delta^{A}{}_{\mu}\delta^{B}{}_{\nu}$, $\eta_{\mu}{}^{\nu} = \delta_{A}{}^{B}
\delta_{B}{}^{\nu}\delta_{\mu}{}^{A}$. Using this, as well as the relation $\eta_{\mu}{}^{\nu}\perp_{\alpha}{}^{\mu} = 0$, we find that the extrinsic curvature can be written as
$$\begin{aligned}
\bar{K}_{\mu\nu}{}^{\rho} &=
\delta_{\mu}{}^{A}\delta_{\nu}{}^{B} \perp^{\rho}{}_{\gamma}
\Gamma^{\gamma}{}_{AB} \\ &= -{1 \over 2}
\delta_{\mu}{}^{A}\delta_{\nu}{}^{B}
\perp^{\rho\gamma}\partial_{\gamma}g_{AB} .\end{aligned}$$
Now, we can use that fact that due to the symmetry imposed on our metric, the only non zero orthogonal derivative of $g_{AB}$ is in the radial direction, which means we can write
$$\bar{K}_{\mu\nu}{}^{\rho} = {1 \over 2}\delta_{r}{}^{\rho}
\delta_{\mu}{}^{A}\delta_{\nu}{}^{B}
\partial_{r}g_{AB}$$
where we used $\perp^{r \gamma} = g^{rr} = -1$. The only non-zero components of $\bar{K}_{\mu\nu}{}^{\rho}$ are $\rho
= r$, and hence we can drop this index, and write
$$\bar{K}_{\mu\nu} ={1 \over 2}
\delta_{\mu}{}^{A}\delta_{\nu}{}^{B}
\partial_{r}g_{AB} .$$
Comparing $\bar{K}_{\mu\nu}$ in this section with the $K_{AB}$ that we have been using;
$$K_{AB} = \partial_{r} g_{AB} ,$$
we see that they differ by a factor of ${1 \over 2}$, which we must account for in what follows. In addition, we note that $\bar{K}_{\mu\nu\rho}$ will be used. This is given by $\perp_{\rho\alpha}\bar{K}_{\mu\nu}{}^{\alpha}$. Since $g_{rr} = -1$, when we lower the third index on the extrinsic curvature, we must introduce a factor of $-1$,
$$\bar{K}_{\mu\nu\rho} = - \delta_{\rho}^{r}
\delta_{\mu}{}^{A}\delta_{\nu}{}^{B} K_{AB} .$$
Returning to our codimension one calculation, equation ($\ref{eq:ff200}$) can be used to calculate the brane evolution equation ($\ref{eq:fr1}$). Note the presence of the Weyl tensor in ($\ref{eq:ff200}$); for codimension one objects it is $W_{\mu\nu}$ which gives the dark radiation term.
In five dimensional, thin braneworld models, the Weyl tensor is singular at the position of the brane. For this reason, $W_{\mu\nu}$ in ($\ref{eq:ff200}$) is not evaluated on the brane. This is the approach that we will take; we evaluate the Weyl tensor at some $r = \epsilon + \delta \gtrsim \epsilon$ outside the core. We look for a solution to the Field equations $R_{\mu\nu} = \Lambda g_{\mu\nu}$, since we have a cosmological constant only in the bulk. A solution to these field equations is given by
$$\label{eq:l1001} ds^{2} = -h(a) dt^{2} + {da^{2} \over h(a)} + a^{2}
\left[d\chi^{2} + \chi^{2} \left(d\theta^{2} + \sin^{2}\theta d
\phi^{2}\right)\right] + \alpha^{2}_{0}
\gamma_{ab}dy^{a}dy^{b},$$
where
$$h(a) = {(n-1) \over 4L^{2}}a^{2} - {\alpha \over
a^{2}} ,$$
and the constant $\alpha^{2}_{0}$ is given by $\alpha^{2}_{0} = (m-2) L^{2} / (n-1)$.
From ($\ref{eq:l1001}$) we can now calculate the Weyl tensor ($\ref{eq:wi37}$). The relevant components that contribute to the brane evolution equations are $W_{DA}{}^{D B}$, where capital Latin indices run over the $(3+1)$ brane coordinates. To calculate the Weyl tensor contribution explicitly, we write $R_{DA}{}^{DB} =
R_{A}{}^{B} - R_{aA}{}^{aB} - R_{rA}{}^{rB}$, and use the fact that for the metric ($\ref{eq:l1001}$), we can set $R_{aA}{}^{aB} = 0$. Hence, using $R_{\mu\nu} = \Lambda_{\rm n} g_{\mu\nu}$ and $R = n
\Lambda_{\rm n}$, we can write the relevant Weyl tensor components as,
$$\label{eq:wkk1} W_{Di}{}^{Dj} = {(4-n) \over L^{2}}\delta_{i}{}^{j} +
{h' \over 2 a}\delta_{i}{}^{j} = {15-3n \over 4L^{2}}\delta_{i}{}^{j} +
{\alpha \over a^{4}}\delta_{i}{}^{j} ,$$
$$\label{eq:wkk2} W_{D t }{}^{D t} = {(4-n) \over L^{2}} + {h'' \over
2} = {15-3n \over 4L^{2}} - {3 \alpha \over a^{4}} ,$$
We note that the standard dark radiation term is present in ($\ref{eq:wkk1}$) and ($\ref{eq:wkk2}$).
Thus we have confirmed the presence of the dark radiation term in our setup. To understand why we obtain this term, we return to the metric ($\ref{eq:l1001}$). We see that we can split this metric into two components; a five dimensional part given by
$$\label{eq:l10010} ds^{2}_{[5]} = -h(a) dt^{2} + {da^{2} \over h(a)} + a^{2}
\left[d\chi^{2} + \chi^{2} \left(d\theta^{2} + \sin^{2}\theta d
\phi^{2}\right)\right] ,$$
which is the standard five dimensional metric considered in the literature [@ma04], that is the standard five dimensional Schwarzschild AdS line element, and a second component
$$\label{eq:rm1} ds^{2}_{[m-1]} = \alpha^{2}_{0}
\gamma_{ab}dy^{a}dy^{b},$$
which is simply pure AdS. This second component is the $(m-1)$ codimensions. From this split the origin of the dark radiation term becomes a little clearer; we might expect to obtain a dark radiation term from the five dimensional part ($\ref{eq:l10010}$) of our metric. The remaining $(m-1)$ codimensions in ($\ref{eq:rm1}$) are pure AdS only, and hence will contribute only terms like $\sim 1/L^{2}$ to $W_{\mu\nu}$. The fact that we obtain a dark radiation term in the brane Friedmann equation is a consequence of our choice of metric ansatz. If we dropped our assumption of spherical symmetry in the extra dimensions, or introduced additional bulk fields, then we would obtain more complicated bulk effects on the brane scale factor. In other words, in this particular model we have over-constrained the bulk, and assumed that it is simply pure AdS away from the brane.
Covariant braneworld equations
------------------------------
Finally, we verify that we can obtain our equations using the covariant braneworld equations given in, for example, [@c1]. We will see that we can consider our brane as either a codimension one or codimension $m$ object, and still obtain the same brane equation. We will explain why this is so at the end of the section.
We begin with the generalized Gauss equation for a $p$-brane of arbitrary codimension, as found in [@c1]. This equation relates the Ricci tensor of the brane, $R^{(p)}$, to the full $n$-dimensional Ricci tensor $R_{\mu\nu}$, the extrinsic curvature $\bar{K}_{\mu\nu}{}^{\rho}$ and the appropriately contracted Weyl tensor $W_{\mu\nu}$. We will show that we can obtain our brane Friedmann equation using two approaches. In the first method, we consider a codimension one object, with the $(m-1)$ ‘codimensions’ not as transverse dimensions but rather as brane parallel dimensions. In this approach the radial coordinate $r$ acts as the orthogonal coordinate, and we find our Friedmann equation arises from the Gauss equation.
In the second approach, we consider our brane as a codimension $m$ object, and calculate the Weyl tensor and background Ricci tensor in terms of the extrinsic curvatures $K_{A}^{B}$ and $L_{a}^{b}$. With this approach, we obtain the same Friedmann equation. We will show this, and then explain why we obtain the same result regardless of whether we consider the $(m-1)$ codimensions as brane parallel or brane orthogonal directions.
In [@c1], the Ricci tensor $R_{\mu\nu}^{(p)}$ and scalar $R^{(p)}$ of a $p$-brane embedded in an $n$-dimensional background space have been calculated, and are given by
$$\label{eq:ff2} R_{\mu\nu}^{(p)} = {p-2 \over
n-2}\eta_{\mu}{}^{\rho}\eta_{\nu}{}^{\sigma}R_{\rho\sigma} + {1 \over
n-2} \left( \eta^{\rho\sigma}R_{\rho\sigma} - {p-1 \over
n-1}R\right) \eta_{\mu\nu} + {p-1 \over p^{2}}
\bar{K}^{\sigma}\bar{K}_{\sigma}\eta_{\mu\nu} + {p-2 \over p}
\bar{C}_{\mu\nu}{}^{\sigma}\bar{K}_{\sigma}-
\bar{C}_{\mu}{}^{\rho\sigma}\bar{C}_{\nu\rho\sigma} + W_{\mu\nu},$$
where
$$\label{eq:w1} W_{\mu\nu} = \eta_{\mu}{}^{\sigma}\eta_{\nu}{}^{\kappa}
\eta_{\tau}{}^{\rho}W_{\rho\sigma}{}^{\tau}{}_{\kappa} ,$$
and
$$\label{eq:rp1} R^{(p)} = {p-1 \over n-2} \left( 2
\eta^{\rho\sigma}R_{\rho\sigma} - {p \over n-1}R \right) + {p-1
\over p} \bar{K}^{\sigma}\bar{K}_{\sigma} -
\bar{C}_{\lambda\mu}{}^{\nu} \bar{C}^{\lambda\mu}{}_{\nu} + W,$$
where
$$\bar{C}_{\mu\nu}{}^{\rho} = \bar{K}_{\mu\nu}{}^{\rho} - {1
\over p} \eta_{\mu\nu} \bar{K}^{\rho} .$$
To begin, we consider a codimension one object of dimension $p = m+3$. If we consider our analysis as describing a codimension one object, with $r$ being the codimension, then $R^{(p)} = R^{(g)} + R^{(\gamma)}$, and we can calculate our brane Friedmann equation from ($\ref{eq:rp1}$).
To do so, we will need to evaluate the projected Weyl tensor, specifically the trace of ($\ref{eq:w1}$),
$$\label{eq:w} W = \eta^{\kappa\sigma}
W_{\rho\sigma}{}^{\tau}{}_{\kappa}\eta^{\rho}{}_{\tau} .$$
Remembering that the brane tangential projection $\eta_{\mu\nu}$ now runs over the standard four dimensions as well as the $(m-1)$ spherically symmetric dimensions, the relevant components of ($\ref{eq:w}$) are given by
$$\label{eq:wi1} W_{Di}{}^{Dj} = {(4-n) \over L^{2}}\delta_{i}{}^{j} +
{h' \over 2a}\delta_{i}{}^{j} = -{15-3n \over
4(n-1)}\Lambda_{\rm n}\delta_{i}{}^{j} + {\alpha \over
a^{4}}\delta_{i}{}^{j},$$
$$\label{eq:wi2} W_{D t }{}^{D t} = {(4-n) \over L^{2}} + {h'' \over
2} = -{15-3n \over 4(n-1)}\Lambda_{\rm n} - {3 \alpha \over a^{4}},$$
$$\label{eq:wi3} W_{ab}{}^{ab} = {5 (m-1) \over
(m+3)} \Lambda_{\rm n} ,$$
$$\label{eq:wi4} W_{aB}{}^{aB} + W_{Ba}{}^{Ba} =
-{8(m-1) \over m+3}\Lambda_{\rm n} .$$
Now taking the trace of ($\ref{eq:wi1}$) and summing over all of the above Weyl tensor contributions to $W$ gives the result $W = 0$, as it should be; the trace of the Weyl tensor is zero for codimension one objects.
Using all of the above in ($\ref{eq:rp1}$), as well as the fact that $p = m+3$, $n = m+4$, gives us
$$\label{eq:dk1} R^{(g)} + R^{(\gamma)} = 2\left(R^{a}{}_{a} +
R^{A}{}_{A}\right) - R_{\alpha}{}^{\alpha} - {(m+2) \over
4(m+3)}\left(K^{2} + L^{2} + 2KL\right) + {1 \over
4}\left(K_{A}{}^{B}K_{B}{}^{A} + L_{a}{}^{b}L_{b}{}^{a}\right) - {1 \over
4(m+3)}\left( K^{2} + L^{2} + 2KL \right) .$$
Finally, we use the fact that
$$R^{(g)} = -6 \left( {\ddot{a} \over a} +
\left({\dot{a} \over a}\right)^{2} \right) ,$$
as well as
$$\label{eq:as1} 2R_{r}{}^{r} - R_{\alpha}{}^{\alpha} =
R_{\alpha}{}^{\alpha} - 2\left( R_{a}{}^{a} + R_{A}{}^{A} \right) = 2
{T_{r}{}^{r} \over M^{N-2}} ,$$
to write ($\ref{eq:dk1}$) as
$$\label{eq:jkb20} -6 \left( {\ddot{a} \over a} +
\left({\dot{a} \over a}\right)^{2} \right) = - 2{T_{r}{}^{r} \over
M^{n-2}} - R^{(\gamma)} - {1 \over 4} \left( K^{2} + L^{2} + 2
KL\right) + {1 \over 4} \left( K_{C}{}^{D}K_{D}{}^{C} +
L_{a}{}^{b}L_{b}{}^{a}\right) .$$
Next, we consider the field equations ’junction conditions’ for our thick brane approach. As before, we must integrate them over the brane orthogonal coordinates. However, since we are now simply considering a codimension one object, we only integrate over the $r$ coordinate. Doing so, we obtain
$$\label{eq:jkb17} K_{i}{}^{j}|_{r=\epsilon} = { M_{\rm b}^{m-1}
\over M^{n-2}}\left(\tilde{T}_{i}{}^{j} -
\delta_{i}{}^{j}{\tilde{T} \over n-2}\right) ,$$
$$\label{eq:jkb18} K_{t}{}^{t}|_{r=\epsilon} = { M_{\rm b}^{m-1}
\over M^{n-2}}\left(\tilde{T}_{t}{}^{t} -
\delta_{t}{}^{t}{\tilde{T} \over n-2}\right) ,$$
$$\label{eq:jkb19} L_{a}{}^{b}|_{r=\epsilon} = { M_{\rm b}^{m-1}
\over M^{n-2}}\left(\tilde{T}_{a}{}^{b} -
\delta_{a}{}^{b}{\tilde{T} \over n-2}\right) -{2 \over
\epsilon}\delta_{a}{}^{b},$$
The final step is to use ($\ref{eq:jkb17}$-$\ref{eq:jkb19}$) to write ($\ref{eq:jkb20}$) as
$$3\left({\ddot{a}\over a}+ \left({\dot{a}
\over a}\right)^{2}\right)+ {M_{\rm b}^{2m-2} \over 8 (n-2)^{2}M^{2n-4}}
\left((n-2)^{2}\tilde{T}_{A}{}^{B}\tilde{T}_{B}{}^{A} +
B_{m}(\tilde{T}_{A}{}^{A})^{2} +
(n-2)^{2}\tilde{T}_{a}{}^{b}\tilde{T}_{b}{}^{a} +
B_{m}(\tilde{T}_{a}{}^{a})^{2}\right)$$
$$+ {M_{\rm b}^{2m-2} \over 8 (n-2)^{2}M^{2n-4}}\left(2B_{m}\left( \tilde{T}_{a}{}^{a} +
\tilde{T}_{r}{}^{r}\right) \tilde{T}_{A}{}^{A} + 2B_{m}
\tilde{T}_{a}{}^{a}\tilde{T}_{r}{}^{r} +
C_{m}(\tilde{T}_{r}{}^{r})^{2}\right) - {M_{\rm b}^{m-1} \over 2
\epsilon M^{n-2}}\tilde{T}_{a}{}^{a}$$
$$- {(m-1)M_{\rm b}^{m-1} \over 2\epsilon M^{n-2}}
\tilde{T}^{r}{}_{r}- { (m-1)(m-2)\over \epsilon^{2}} = -{\Lambda
\over M^{n-2}}.$$
which is the same equation as ($\ref{eq:b25}$) for a codimension $m$ brane. Note that since we have assumed that our space contains a codimension one object, the trace of the Weyl tensor vanishes. However, $W_{\mu\nu}$ does not vanish; as we have shown above we obtain a dark radiation like term.
It may seem quite unnatural that we obtain the same Friedmann equation whether we consider a codimension one or $m$ object. The reason we do so is because we have assumed spherical symmetry. In our paper, we have considered surfaces of constant $r = \epsilon$ only; since the other codimensions are spherically symmetric we do not have to set them to a particular fixed value (since our final result will not depend on our choice). Hence the $(m-1)$ spherically symmetric dimensions could equally well be brane tangential or brane orthogonal coordinates. The only difference between the two would be the form of the $(m-1)$ components of the energy momentum tensor $\tilde{T}_{a}{}^{b}$, which would be small for a codimension $m$ brane, but potentially large for a codimension one brane.
We now proceed with our second approach, that is to consider our brane as a four-dimensional object of codimension $m$. Now, we will no longer have the trivial result $W=0$, however we find that we still obtain the same equation relating the brane scale factor $a(t)$ with $\tilde{T}_{A}{}^{B}$, $\tilde{T}_{r}{}^{r}$ and $\tilde{T}_{a}{}^{b}$.
To proceed, we consider ($\ref{eq:rp1}$) again. Now we have $p=4$ and $R^{(p)} = R^{(g)}$. We begin by calculating the projected Weyl tensor, using ($\ref{eq:wi37}$). The relevant components in our coordinate system are $W_{AB}{}^{AB}$, which are explicitly
$$\label{eq:wk1} W_{AB}{}^{AB} = R_{AB}{}^{AB} - {6 \over
(n-2)}R^{A}{}_{A} + {12 \over (n-1)(n-2)}R_{\alpha}{}^{\alpha}.$$
Using ($\ref{eq:wk1}$) in ($\ref{eq:rp1}$), we find that the brane Ricci tensor $R^{p}$ may be written as
$$\label{eq:as2} R^{(p)} = R_{AB}{}^{AB} - {3 \over 16} K^{2} + {1
\over 4} K^{A}_{B}K^{B}_{A} - {1 \over 16} K^{2} .$$
Next, we use the relation ($\ref{eq:as1}$), as well as
$$R_{AB}{}^{AB} = R_{A}{}^{A} - R_{aA}{}^{aA} -
R_{rA}{}^{rA}$$
to write ($\ref{eq:as2}$) as
$$\begin{aligned}
R^{(p)} &= R^{\alpha}{}_{\alpha} -
2R^{a}{}_{a} - R_{A}{}^{A} - R_{aA}{}^{aA} - R_{rA}{}^{rA} + {1 \over 4}
K^{A}{}_{B}K^{B}{}_{A} - {1 \over 4} K^{2} - 2{T_{r}{}^{r} \over M^{N-2}}
\\ \label{eq:as3} &= R_{r}{}^{r} - R_{a}{}^{a} - R_{aA}{}^{aA} - R_{rA}{}^{rA} + {1 \over 4}
K^{A}{}_{B}K^{B}{}_{A} - {1 \over 4} K^{2} - 2{T_{r}{}^{r} \over M^{n-2}}\end{aligned}$$
we now evaluate the terms in ($\ref{eq:as3}$). We find
$$R_{r}{}^{r} - R_{a}{}^{a} ={1 \over 2}
K' + {1 \over 4}\left( K_{A}{}^{B}K_{B}{}^{A} + L_{a}{}^{b}L_{b}{}^{a}
\right) - R^{(\gamma)} - {1 \over 4}L^{2} - {1 \over 4}KL ,$$
$$-R_{aA}{}^{aA} = -{1 \over 4}KL ,$$
$$- R_{rA}{}^{rA} = -{1 \over 2}K' - {1 \over 4}
K_{A}{}^{B}K_{B}{}^{A} .$$
Using these in ($\ref{eq:as3}$) gives us
$$\label{eq:as4} R^{(p)} = {1 \over
4}\left(K_{A}{}^{B}K_{B}{}^{A} + L_{a}{}^{b}L_{b}{}^{a}\right) - {1 \over 4}
\left(K^{2} + L^{2}\right) - {1 \over 2}KL - 2{T_{r}{}^{r} \over
M^{n-2}} - R^{(\gamma)} .$$
Once again, if the then proceed to write $K_{A}{}^{B}$ and $L_{a}{}^{b}$ in terms of $\tilde{T}_{\mu}{}^{\nu}$, we would find the same equation as we found in both the previous sections and above (when we considered the brane as a codimension one object). We stress that we obtain the same result because of the symmetry imposed on the $(m-1)$ codimensions. This means that our choice of the angular coordinates will not affect our Ricci scalar $R^{(p)}$, and hence we can consider the spherically symmetric coordinates either as being brane orthogonal or brane tangential.
Finally, we note that our result that codimension one and $m$ branes are equivalent for a metric such as ours is only valid for a thick brane of codimension one or $m$. If we consider thin branes, then in the codimension $m$ case we would obtain $\delta$-function singularities in $K_{A}{}^{B}$ and hence $R^{(p)}$, whereas for codimension one branes there would be no divergent behaviour. The singular behaviour is removed in the codimension $m$ case since we have smeared the brane energy momentum tensor over a finite region of space.
The brane Friedmann equation that we have derived is written in terms of the brane energy momentum tensor, which has the standard four dimensional components $\tilde{T}_{AB}$, and in addition non zero components in the $(a,b)$ and $(r,r)$ directions, $\tilde{T}_{ab}$ and $\tilde{T}_{rr}$. Above, we have shown that we can write down an evolution equation for $a(t)$ in terms of $\tilde{T}_{AB}$, $\tilde{T}_{ab}$ and $\tilde{T}_{rr}$, regardless of whether the object is codimension one or $m$. The only difference between the two setups will be the form of the $\tilde{T}_{ab}$ components of the brane energy momentum tensor. When we have $m$ codimensions, we expect that $\tilde{T}_{ab}$ will be some small quantity, which asymptotes to zero in the thin brane limit. For a codimension one brane however, the $\tilde{T}_{ab}$ will have the same form as the $\tilde{T}_{AB}$ components. This is the principle difference between the two treatments. However in this paper we do not explicitly specify the brane energy momentum tensor.
[99]{}
P Brax, C van de Bruck (2003) Class. Quant. Grav. 20, R201.
Roy Maartens (2004) Living Rev. Rel. 7, 7.
R Dick (2001) Class. Quant. Grav, Vol. 18, R1.
L. Randall, R. Sundrum (1999) Phys. Rev. Lett. 83, 3370.
L. Randall, R. Sundrum (1999) Phys. Rev. Lett. 83, 4690.
Tetsuya Shiromizu, Kei-ichi Maeda, Misao Sasaki (2000) Phys. Rev. D 62, 024012.
James M. Cline, Christophe Grojean, Geraldine Servant (1999) Phys. Rev. Lett. 83, 4245.
Peter Bowcock, Christos Charmousis, Ruth Gregory (2000) Class. Quant. Grav. 17, 4745-4764.
B. Carter, R. A. Battye, J.-P. Uzan (2003) Comm. Math. Phys. 235, 289-311.
D. Langlois (2002) Prog. Theor. Phys. Suppl. 148.
C Charmousis, R Zegers (2005) hep-th/0502170.
V. Rubakov, M. E. Shaposhnikov (1983) Phys. Lett. B 125, 136.
V. Rubakov, M. E. Shaposhnikov (1983) Phys. Lett. B 125, 139.
B. Carter (1992) Class. Quant. Grav. 9, 19-33.
I. Navarro, J. Santiago (2005), JHEP 0502, 007.
J.M. Cline, J. Descheneau, M. Giovannini, J. Vinet (2003) JHEP 0306, 048.
J. Vinet (2004) Int. J. Mod. Phys. A19, 5295.
P. Binetruy, C. Deffayet, D. Langlois (2000) Nucl. Phys. B565, 269.
I. Navarro K. Santiago (2004) hep-th/0402204.
Christos Charmousis, Robin Zegers (2005), JHEP 0508, 075.
Frederic Leblond, Robert C. Myers, David J. Winters (2001) JHEP 0107, 031.
R. A. Battye, B. Carter, and A. Mennim (2004) Phys. Rev. Lett. 92, 201305.
R. A. Battye, B. Carter, and A. Mennim (2005) Phys. Rev. D 71, 104026.
R. A. Battye, B. Carter, A. Mennim, and J. P. Uzan (2001) Phys. Rev. D 64, 124007.
D Yamauchi, M Sasaki (2007) arXiv: 0705.2443.
C Charmousis, R Gregory, A Padilla (2007) arXiv: 0706.0857.
P. Bostock, R. Gregory, I. Navarro, J. Santiago (2004) Phys. Rev. Lett. 92, 221601.
I. Navarro, J. Santiago (2006) JCAP 0603, 015.
A. L. Fitzpatrick, L. Randall (2006) JHEP 0601, 113.
S.M. Carroll, L. Mersini (2001) Phys. Rev. D 64, 124008.
S. M. Carroll, M. M. Guica (2003) hep-th/0302067.
T. Gherghetta, E. Roessl, M. Shaposhnikov (2000) Phys. Lett. B491, 353.
B. Carter (2001) Int. J. of Theor. Phys. 40, 2099.
K. Benson, I. Cho (2001) Phys. Rev. D 64, 065026.
I. Oda (2001), Phys. Rev. D 64, 026002.
P. Mounaix, D. Langlois (2002) Phys. Rev. D 65, 103523.
C. Csaki, M. Graesser, L. Randall, J. Terning (2000) Phys. Rev. D 62, 045015.
S. Mukohyama (2000) Phys.Lett. B473, 241.
Daisuke Ida (1999) arXiv:gr-qc/9912002.
R. A. Battye, A. Mennim (2001) Class. Quant. Grav. 18, 2171.
S. Kobayashi, K. Koyama, J. Soda (2002) Phys. Rev. D 65, 064014.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the growth of networks from a set of isolated ground nodes by the addition of one new node per time step and also of a fixed number of directed edges leading from the new node to randomly selected nodes already in the network. A fixed-width time window is used so that, in general, only nodes that entered the network within the latest window may receive new incoming edges. The resulting directed network is acyclic at all times and allows some of the ground nodes, then called sinks, to be reached from some of the non-ground nodes. We regard such networks as representative of abstract systems of partially ordered constituents, for example in some of the domains related to technological evolution. Two properties of interest are the number of sinks that can be reached from a randomly chosen non-ground node (its reach) and, for a fixed sink, the number of nonoverlapping directed paths through which the sink can be reached, at a given time, from some of the latest nodes to have entered the network. We demonstrate, by means of simulations and also of analytic characterizations, that reaches are distributed according to a power law and that the desired directed paths are expected to occur in very small numbers, perhaps indicating that recovering sinks late in the process of network growth is strongly sensitive to accidental path disruptions.'
author:
- 'Valmir C. Barbosa'
bibliography:
- 'dag.bib'
title: Reachability and recoverability of sink nodes in growing acyclic directed networks
---
Introduction
============
The study of large, essentially unstructured networks of interacting elements, also referred to as complex networks, has in the past several years received considerable attention. The main motivation behind so much interest has been the realization that networks occurring in many natural, technological, and social domains have common statistical properties that, though governed by strictly local interactions among the networks’ elements, relate globally to the networks’ structure or functionality. A comprehensive collection of papers spanning the main aspects of this emerging discipline, from origins to representative applications, can be found in [@bs03; @nbw06].
While it seems correct to say that most network models studied so far are undirected, reflecting the fact that the local interactions occur between pairs of interconnected elements in any of the two possible directions (this is the case, for example, of the networks that represent the Internet at some level), there are also several cases in which interactions are inherently unidirectional, as for example the WWW [@baj00], networks of bibliographic citations [@v01], and also networks that arise from certain flows of information in computer networks [@sb06a; @sb06b; @sb07]. Unidirectional interactions give rise to directed networks (that is, networks whose edges have directions), which in turn have been studied for both structural [@dms01; @bds03; @bds04; @bm06] and functional [@kkk02; @g04] properties.
The structure of directed networks is considerably more intricate than that of undirected networks, and this is due primarily to the existence of directed cycles, that is, node sequences in which it is possible to return to any node by following edges along their directions. The existence of such cycles in a directed network is strictly necessary for nontrivial strongly connected components to appear, so it comes as no surprise that many of the network’s properties depend on whether directed cycles exist, how large they are, and how they relate to other structures in the network. So, even though some attention has been given to network elements that lie outside directed cycles [@m03] or to how the network looks when directed cycles are broken [@c04], a fair appraisal seems to be that studying directed networks has so far concentrated primarily on properties that depend on the existence of directed cycles.
However, we find that a surprising number of systems are naturally representable by directed networks that are intrinsically acyclic, that is, contain no directed cycles (even though plenty of cycles exist if one ignores the edges’ directions). Such networks exist at much more abstract levels than the majority of the networks that have received attention from researchers, reflecting in general the partial order that is inherent to their nature or to the manner in which they are constructed. Important examples are: networks of immediate event precedence, both in history [@s96] and in the unfolding of distributed computations [@l78]; networks of object inheritance in object-oriented programs [@tcsh06]; the probabilistic graphical models, known as Bayesian networks, that represent the causal relationships among random variables in some artificial-intelligence systems [@p88]; networks that represent possible deductions in axiomatic systems of formal proof [@c02]; and networks of word etymology in large language groups [@babel-url].
Perhaps the reason why systems such as these have not yet been approached from a complex-network perspective is ultimately the elusiveness that they have about them. In some cases, data are simply not readily obtainable, as seems to be the case of the networks that reflect the innards of large software or artificial-intelligence systems. In others, as in the history and etymology systems, even defining the network’s elements depends on data that are no longer extant and thus requires extensive hypothesizing. Even so, it seems possible to postulate some prototypical growth model for acyclic directed networks and then use it in the study of properties that are expected to be of interest.
Our approach in this paper is to study the growth of acyclic directed networks from an initial set of ground nodes by the continual addition of new nodes and directed edges. At each time step, the growth is limited to the addition of one single node and a fixed number of edges outgoing from that node to randomly selected nodes already in the network. We impose a constraint on which are the nodes toward which new edges may be added: as a new node enters the network, the outgoing edges it acquires must necessarily lead to nodes inside a fixed-size window representing that time step’s immediate past. Both finite and infinite windows are considered, so we hope to be contemplating a wide variety of circumstances in regard to the previously mentioned networks as well as others.
Unlike most other studies of complex networks, in the present case the central entities to be observed are not node degrees (distributions are trivially obtainable for both in- and out-degrees, as we discuss shortly), but have to do instead with whether (and from which nodes) the ground nodes remain reachable as time elapses and, if they do, the nature of the directed paths that lead to them. What we have found is that ground-node reachability depends on how the number of ground nodes relates to window size, and also that the number of ground nodes that can be reached is at times distributed as a power law. As for recovering ground nodes from the latest nodes added to the network, this is expected to be achievable only through a very small number of nonoverlapping directed paths, thus indicating high susceptibility to failure should one such path be disrupted.
The model and basic properties
==============================
We study network evolution for discrete time $t\ge 1$ from an initial set of $n_0$ isolated ground nodes. One new node is added per time step, so the elapsing of time step $t$ causes the network to have $n_0+t$ nodes. We identify the ground nodes by the nonpositive integers $-n_0+1,\ldots,0$, thus imposing an arbitrary order on them, even though they are all assumed to be present when network growth begins. We also use $t$, interchangeably, to refer both to time step $t$ and to the node added at that time step. Upon entering the network, node $t$ acquires two outgoing edges leading to distinct nodes chosen randomly from the set $\{\max\{-n_0+1,t-w\},\ldots,t-1\}$ for some window $w\ge n_0$. If $t\le w+1$, then this set contains $$w_t=\min\{n_0,w-t+1\}$$ ground nodes; it contains no ground nodes otherwise. \[Note that the choice of $2$, as opposed to some other constant, as the number of outgoing edges per node added to the network is qualitatively irrelevant, so we make it for simplicity’s sake only. Similarly, we rule out the possibility of $w<n_0$, because this is qualitatively equivalent to using a number of ground nodes equal to $w$ (since it implies that $n_0-w$ ground nodes are guaranteed to remain isolated indefinitely).\]
Every non-ground node has an out-degree of exactly $2$. As for in-degrees, we may concentrate on some non-ground node $i$ and let $k\in\{0,\ldots,w\}$. The probability that $i$ has in-degree $k$ is clearly given by $${{w}\choose{k}}
\left(\frac{2}{w}\right)^k
\left(1-\frac{2}{w}\right)^{w-k}
\approx\frac{2^ke^{-2(1-k/w)}}{k!},$$ which approximates the probability that, at time $t\gg n_0$, a randomly chosen node has in-degree $k$. For $k\ll w$, it approaches the mean-$2$ Poisson distribution. (Note that, if we condition on ground nodes exclusively, the in-degree distribution becomes more concentrated at low degrees than the mean-$2$ Poisson, which implies a lower mean value.)
We henceforth refer to every non-isolated node having no outgoing edges as a sink, and to every non-isolated node having no incoming edge as a source. Clearly, every ground node becomes a sink when picked to be directed an edge at for the first time, and conversely only ground nodes may be sinks. Likewise, every non-ground node is a source upon entering the network, though it may cease being one afterward; conversely, no ground node may be a source.
Let $S_t$ denote the expected number of sinks just before the addition of node $t$ to the network. We have $S_1=0$ and, for $t\ge 1$, $$S_{t+1}=S_t+\Delta_t,
\label{eq:difference}$$ where $\Delta_t$ is the expected number of new sinks created when node $t$ is added. Of the $w_t$ ground nodes that may acquire a new incoming edge at time $t$, let those that are already sinks amount to an expected number $f_t$. Then $f_t=(w_t/n_0)S_t$ and $w_t-f_t=w_t(1-S_t/n_0)$.
The number of node pairs from which to choose at time $t$ is $(w_t+t-1)(w_t+t-2)/2$. Of these, $[w_t+t-1-(w_t-f_t)](w_t-f_t)$ are expected to lead to the creation of one new sink, while $(w_t-f_t)(w_t-f_t-1)/2$ others are expected to lead to the creation of two new sinks. We then obtain $$\begin{aligned}
\Delta_t
&=&\frac{2(w_t-f_t)(f_t+t-1)}{(w_t+t-1)(w_t+t-2)}\nonumber\\
&&\hspace{0.5in}\mbox{}+\frac{2(w_t-f_t)(w_t-f_t-1)}{(w_t+t-1)(w_t+t-2)}\\
&=&\frac{2w_t(1-S_t/n_0)}{w_t+t-1}.\end{aligned}$$ Approximating (\[eq:difference\]) by a differential equation yields two possibilities, depending on $t$. For $1\le t\le w+1-n_0$, $w_t=n_0$ and we get $$\frac{dS_t}{dt}+\frac{2S_t}{n_0+t-1}=\frac{2n_0}{n_0+t-1},$$ thence $$S_t=\frac{n_0(t-1)(2n_0+t-1)}{(n_0+t-1)^2}
\label{eq:solution1}$$ is obtained from $S_1=0$. For $w+1-n_0\le t\le w+1$, $w_t=w-t+1$ and we get $$\frac{dS_t}{dt}+\frac{2(w-t+1)S_t}{wn_0}=\frac{2(w-t+1)}{w},$$ thence $$S_t=n_0\left\{1-\left(\frac{n_0}{w}\right)^2
\exp\left[\left(\sqrt\frac{w}{n_0}-\frac{t-1}{\sqrt{wn_0}}\right)^2
-\frac{n_0}{w}\right]\right\}
\label{eq:solution2}$$ results from $S_{w+1-n_0}=n_0[1-(n_0/w)^2]$ \[cf. (\[eq:solution1\])\]. Notice that expressing $S_t/n_0$ as a function of $(t-1)/n_0$ in (\[eq:solution1\]), which is already independent of $w$, yields a constant with respect to $n_0$ as well. Doing the same in (\[eq:solution2\]) reveals an exclusive dependence on the ratio $n_0/w$.
Beginning at $t=w+1$, it is no longer possible for any sink to be created, so the expected number of sinks settles at the value, henceforth denoted by $S(n_0/w)$, given by $$S(n_0/w)=S_{w+1}=n_0\left[1-\left(\frac{n_0}{w}\right)^2e^{-n_0/w}\right],$$ following (\[eq:solution2\]). For $w=n_0$, this becomes $S(1)=n_0(1-e^{-1})$, which limits the expected number of sinks at about $63.21\%$ of the ground nodes. As $w$ grows, $S(n_0/w)$ approaches $n_0$ asymptotically.
Our study on the recoverability of sinks will be based on the nodes that, at time $t$, remain sources inside the latest window (i.e., the window comprising nodes $t-w+1,\ldots,t$). The probability that a node $i$ inside this window remains a source through time $t$ is $[(w-2)/w]^{t-i}$. The expected number of sources inside the latest window, denoted by $R$, is then $$R=\sum_{i=t-w+1}^t\left(\frac{w-2}{w}\right)^{t-i}
\approx w\left(\frac{1-e^{-2}}{2}\right),$$ amounting therefore to roughly $43.23\%$ of the nodes inside the window.
Reachability and recoverability of sinks
========================================
Reachability
------------
At time $t$, we say that a ground node is reachable from one of the $n_0+t$ nodes of the network when a directed path exists between them leading to the ground node. All ground nodes are reachable from themselves, but only sinks are reachable from non-ground nodes. The reach of a node is the number of ground nodes that are reachable from it. A node has unit reach if and only if it is a ground node, and the reach of a non-ground node refers to sinks exclusively.
Let $P_t(r)$ be the probability that, at time $t$, a randomly chosen node has reach $r$. Clearly, $$P_t(1)=\frac{n_0}{n_0+t}.$$ For $r>1$, however, we expect the number of sinks in the network to play a role in defining the value of $P_t(r)$.
As a node enters the network and connects out to two previously existing nodes, its reach has to account for every sink that is reachable from either of those two nodes. In the relatively early stages of network formation, and for sufficiently large $n_0$, it is likely that no sink is reachable from the two nodes concomitantly, and in this case the new node’s reach is simply the sum of their reaches. This becomes progressively less likely later on in the evolution of the network, thus making accurate predictions of $P_t(r)$ very difficult.
Our finds regarding $P_t(r)$ are summarized in Figure \[fig:wpl\], whose part (a) refers to $w=n_0$. In this case we see that, initially, non-unit reaches tend to be distributed exponentially. For $t=w=n_0$, in particular, the exponential character of the distribution is very clear \[cf. the inset in part (a) of the figure\] and may be expressed as $$P_{n_0}(r)
\approx\left(\frac{S(1)}{2n_0}\right)a^r
=\left(\frac{1-e^{-1}}{2}\right)a^r,
\label{eq:exp1}$$ for some constant $a$ such that $0<a<1$. Since the exponential seems to hold across all pertinent reach values, we can find $a$ by requiring $$P_{n_0}(1)+\sum_{r\ge 2}P_{n_0}(r)
=\frac{1}{2}+\left(\frac{1-e^{-1}}{2}\right)\sum_{r\ge 2}a^r
=1,$$ which leads to $a\approx 0.6958$. It also seems that an exponential approximation continues to hold for somewhat larger values of $t$. For $t\gg w$, though, we expect more and more nodes of reach around $S(1)$ to appear, owing to the finiteness of $w$. This is indeed what happens, but aside from this effect we have also found that the passage of time leads the initial exponential approximation to $P_t(r)$ to gradually become $$P_t(r)
\approx\left(\frac{S(1)}{n_0+t}\right)r^{-1}
=\left(\frac{n_0(1-e^{-1})}{n_0+t}\right)r^{-1},
\label{eq:pl1}$$ similar therefore to the power law known as Zipf’s law.
![(Color online) Reach distribution for $n_0=1\,000$, with $w=n_0$ (a), $w=2n_0$ (b), and $w=3n_0$ (c). Solid lines give the analytic predictions of (\[eq:exp1\]) and (\[eq:pl1\]) for part (a), of (\[eq:pl2\]) for parts (b) and (c). All simulation data are averages over $500$ independent runs.[]{data-label="fig:wpl"}](wpl1.eps "fig:")\
![(Color online) Reach distribution for $n_0=1\,000$, with $w=n_0$ (a), $w=2n_0$ (b), and $w=3n_0$ (c). Solid lines give the analytic predictions of (\[eq:exp1\]) and (\[eq:pl1\]) for part (a), of (\[eq:pl2\]) for parts (b) and (c). All simulation data are averages over $500$ independent runs.[]{data-label="fig:wpl"}](wpl2.eps "fig:")\
![(Color online) Reach distribution for $n_0=1\,000$, with $w=n_0$ (a), $w=2n_0$ (b), and $w=3n_0$ (c). Solid lines give the analytic predictions of (\[eq:exp1\]) and (\[eq:pl1\]) for part (a), of (\[eq:pl2\]) for parts (b) and (c). All simulation data are averages over $500$ independent runs.[]{data-label="fig:wpl"}](wpl3.eps "fig:")
As we increase $w$ beyond $n_0$ to $w=2n_0$ and $w=3n_0$, we obtain a similar evolution of $P_t(r)$ with respect to $t$, including the progressive probability accumulation around $r=S(1/2)$ or $r=S(1/3)$, depending on the case. This is illustrated, respectively, in parts (b) and (c) of Figure \[fig:wpl\], where we see that the power-law regime is established only for increasingly larger values of $t$. When this happens, a good approximation to $P_t(r)$ seems to be $$P_t(r)
\approx\left(\frac{S(1)}{n_0+t}\right)\frac{r^{-1}}{n_0/w}
=\left(\frac{w(1-e^{-1})}{n_0+t}\right)r^{-1},
\label{eq:pl2}$$ where, curiously, it is still $S(1)$ \[not $S(1/2)$ or $S(1/3)$, as we might expect\] that drives the distribution, after the simple scaling by $n_0/w$.
Recoverability
--------------
We now examine the network’s structure as it relates to the existence of directed paths from the sources in $\{t-w+1,\ldots,t\}$, at time $t$, to the sinks. While the average number of distinct paths over all such source-sink pairs is distributed quite widely, when we look at paths that are not merely distinct but edge-disjoint the situation is very different. For a given source and a given sink, a group of directed paths between them is edge-disjoint if no two paths in the group have any edges in common. The appropriate framework in which to compute the maximum number of edge-disjoint directed paths between two nodes is that of network flows.
Given a directed network with nonnegative numbers associated with the edges (the edges’ capacities), and assuming that it has at least one source and one sink, the maximum flow from a source to a sink is an assignment of numbers to the edges (their flows) such that: no edge flow exceeds the edge’s capacity; the total flow coming into any node equals that leaving the node (except for the source and the sink); and moreover no other assignment results in a greater net flow coming into the sink. By a well-known result from the theory of network flows (the max-flow min-cut theorem), the number of edge-disjoint directed paths from the source to the sink is precisely the maximum flow from the source to the sink under unit capacities [@amo93].
In our present context, the number of edge-disjoint directed paths from any given source to any given sink is at most the minimum between the source’s out-degree (equal to $2$) and the sink’s in-degree (distributed, as we have noted, such that the mean is less than $2$). So we know, beforehand, that the expected average number of such paths, taken over all source-sink pairs of interest, lies somewhere in the interval $[0,2]$. Computing this number is expected to require $RS(n_0/w)$ maximum-flow computations for each network. We have used the publicly available, efficient HIPR code of [@hipr-url] for $n_0=1\,000$ and three different values of $w$.
For $w=n_0$, we have found from $10$ independent runs that the expected average is $0.5024$ at $t=4\,000$, growing to the roughly stable value of $1.2402$ at $t=9\,000$. For $w=2n_0$ and $w=3n_0$, stabilization occurs later. For $t=4\,000$ and $t=19\,000$, the expected averages are, respectively, as follows: $0.0316$ and $1.4598$ for $w=2n_0$, $0.0122$ and $1.5069$ for $w=3n_0$. A small increase is then observed at stability as $w$ becomes larger.
Another pertinent indicator of the recoverability of sinks from sources in the latest window at time $t$ is the number of edge-disjoint directed paths from any of the sources to a given sink. Clearly, the expected average number of such paths, taken over all sinks, is some number in the interval $[0,2R]$, since the expected number of sources is $R$ and each has the potential of contributing two paths. However, the sink’s in-degree remains distributed with a less-than-$2$ mean, so it is very unlikely for an expected average significantly larger than $2$ to turn up. As for calculating the desired number of paths in a given network for a given sink, we note that, unlike the preceding case, a little artifice is needed before a maximum-flow computation can be performed (since it is unclear what the source is in such a computation). What we do is to add another source to the network and make capacity-$2$ directed edges outgo from it to all original sources. The combined number of edge-disjoint directed paths from the original sources to the sink is the maximum flow from the new source to the sink. For each network, we expect $S(n_0/w)$ maximum-flow computations to be needed.
Results for this second indicator are shown in Figure \[fig:npp\] for $w=n_0$ in the main plot set, $w=2n_0$ in the top inset, and $w=3n_0$ in the bottom inset. The resulting expected values are roughly stable at $t=9\,000$ and equal, respectively, $1.4036$, $1.8192$, and $2.2983$. It is clear from the figure that, for $w=n_0$, it is the distribution of the sinks’ in-degrees that exerts the greater influence on how the average number of edge-disjoint directed paths from all sources to one sink is distributed. For $w=3n_0$, it is the distribution of the non-sink nodes’ in-degrees (the mean-$2$ Poisson) that eventually does it.
![(Color online) Distribution of the average number of edge-disjoint directed paths from all sources to one sink for $n_0=1\,000$, with $w=n_0$, $w=2n_0$ (top inset), and $w=3n_0$ (bottom inset). Solid lines give the mean-$2$ Poisson distribution. All simulation data are averages over $500$ independent runs.[]{data-label="fig:npp"}](npp.eps)
The case of an infinite window
==============================
At time $t$, any value of $w$ surpassing $n_0+t-1$ has the effect of an infinite window; that is, any node in the network may be chosen to receive one of the two new edges as an incoming edge. When this is the case, none of our conclusions so far remains valid. Even though the case of infinite $w$ is of little general interest for modeling real systems (it is inherently dependent on global properties of the system as a new node comes in), we feel it is worth commenting on the resulting reach distribution, which differs strikingly from the finite case \[except when $r=1$, since $P_t(1)=n_0/(n_0+t)$ remains of course valid\].
Expressing $P_t(r)$ analytically seems infeasible for most values of $r>1$, but it can be done for $r=2$ and, interestingly, this leads directly to a good approximation for the general case, provided $t\lessapprox 9n_0$. Notice first that, for sufficiently large $n_0$, $$\begin{aligned}
P_t(2)
&\approx&\left(\frac{1}{n_0+t}\right)
\sum_{i=1}^t\left(\frac{n_0}{n_0+i-1}\right)^2\\
&=&P_t(1)n_0\zeta_t(2,n_0),\end{aligned}$$ where $$\zeta_t(2,n_0)=\sum_{u=0}^{t-1}\frac{1}{(n_0+u)^2}$$ is the truncation, to $t$ terms, of $\zeta(2,n_0)$, Riemann’s two-parameter zeta function [@gr00]. Our heuristic generalization for all values of $r$ is then simply the exponential $$P_t(r)\approx P_t(1)\left[n_0\zeta_t(2,n_0)\right]^{r-1}.
\label{eq:exp2}$$ Simulation results are shown in Figure \[fig:winf\], indicating that, for an infinite window, reach probabilities fall at least as fast as exponentially.
![(Color online) Reach distribution for $n_0=1\,000$ under an infinite window. Solid lines give the analytic predictions of (\[eq:exp2\]). All simulation data are averages over $500$ independent runs.[]{data-label="fig:winf"}](winf.eps)
Discussion and concluding remarks
=================================
We have considered directed networks that grow from a fixed set of ground nodes by the addition of one node per time step and of two edges directed from that node to previously existing, randomly chosen nodes inside a fixed-length sliding window. Networks thus constructed are devoid of directed cycles, and may be viewed as a prototypical representation of growing collections of partially ordered items, so long as some underlying time-like notion exists with respect to which the window mechanism makes sense. Laying down more than two edges per time step is expected to have no qualitatively significant effect (although it is unlikely for reaches of small even value to exist in the case of three edges, for example—a reach of $2$ is in fact impossible—and therefore reach distributions can be expected to undergo a sort of bifurcation as one moves from high reaches to lower).
Our study has been centered on the two notions that we deem especially relevant for the systems acyclic directed networks are purported to relate to. The first one is the property, here referred to as reachability, of nodes in the network to be able to reach ground nodes via directed paths. We found, by means of simulations and also through limited analytic predictions, that the number of ground nodes reachable from a randomly chosen non-ground node is distributed first exponentially, then as a power law as time elapses. The other notion on which we focused can be summarized as that of how to recover a specific ground node, in the sense of having edge-disjoint directed paths to get to it from some of the latest nodes to be added to the network. Our finds are that such paths are expected to occur in very small numbers on average (roughly somewhere near $2$), and therefore the recoverability of ground nodes may be severely affected by accidental path disruptions.
We believe this paper’s network model, along with its main observables, opens up new possibilities of investigation about abstract systems that are naturally representable as acyclic directed networks. Earlier we mentioned examples from fields related to computer software, artificial intelligence, mathematical logic, and also history. In addition to their being representable as networks such as the ones we studied, what these systems also have in common once viewed from the perspectives of ground-item reachability and recoverability is that many of them make reference, albeit indirectly, to the growing stack of digital technologies that currently separates “ground” pieces of information from their representations for end use. Concerns related to this issue are sometimes voiced in the media, referring, for example, to the digitization of documents [@unesco-url] or to a future in which, as some envisage, autonomous systems may become inscrutable regarding their internal organization [@itworld-url]. Even though such issues may seem like a far cry from the study we have pursued in this paper, carrying on with an eye on them may well prove worthwhile.
The author acknowledges partial support from CNPq, CAPES, and a FAPERJ BBP grant.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The two-type Richardson model describes the growth of two competing infection types on the two or higher dimensional integer lattice. For types that spread with the same intensity, it is known that there is a positive probability for infinite coexistence, while for types with different intensities, it is conjectured that infinite coexistence is not possible. In this paper we study the two-type Richardson model in the upper half-plane $\Z\times\Z_+$, and prove that coexistence of two types starting on the horizontal axis has positive probability if and only if the types have the same intensity.'
author:
- 'Daniel Ahlberg[^1]'
- 'Maria Deijfen[^2]'
- 'Christopher Hoffman [^3]'
date: August 2018
title: 'The two-type Richardson model in the half-plane'
---
Introduction
============
In 1998, Häggström and Pemantle [@HP_twotype] introduced a model for competing growth on $\Z^2$ known as the two-type Richardson model. Two competing entities, here referred to as type 1 and type 2 infection, initially occupy one site each of the $\Z^2$ nearest-neighbor lattice. As time evolves each uninfected site is occupied by type $i$ at rate $\lambda_i$ times the number of type $i$ neighbors. An infected site remains in its state forever, implying that the model indeed defines a competition scheme between the types.
Regardless of the values of the intensities, both types clearly have a positive probability of winning by surrounding the other type at an early stage. Attention hence focuses on the event $\coex$ that both types simultaneously grow to occupy infinitely many sites; this is referred to as *coexistence* of the two types. Deciding whether or not $\coex$ has positive probability is non-trivial since it cannot be achieved on any finite part of the lattice. By time-scaling and symmetry we may restrict to the case $\lambda_1=1$ and $\lambda_2=\lambda>1$. The conjecture, due to Häggström and Pemantle [@HP_twotype], then is that $\coex$ has positive probability if and only if $\lambda=1$. The *if*-direction of the conjecture was proved in [@HP_twotype], and extended to higher dimensions independently by Garet and Marchand [@GM_coex] and Hoffman [@Hoff_coex], using different methods. As for the *only if*-direction, Häggström and Pemantle [@HP_absence] showed in 2000 that coexistence is possible for at most countably many values of $\lambda$. Ruling out coexistence for *all* $\lambda>1$ remains a seemingly challenging open problem.
In this paper we study the analogous problem in the upper half-plane $\Z\times\Z_+=\{(x,y):y\ge0\}$ with $(0,0)$ initially occupied by type 1 and $(1,0)$ initially occupied by type 2, and show that coexistence has positive probability if and only if $\lambda=1$. That coexistence is possible for $\lambda=1$ follows from similar arguments as in the full plane, so the novelty lies in proving the *only if*-direction.
\[thm:main\] Consider the two-type Richardson model on $\Z\times\Z_+$ with $(0,0)$ and $(1,0)$ initially of type 1 and 2, respectively. Then we have that $\Pr(\coex)>0$ if and only if $\lambda=1$.
Some readers might suspect that the arguments used to prove this result could be adaptable to settle the Häggström-Pemantle conjecture in the full plane. This however is most likely not the case. It is known that, on the event of coexistence in the full plane, the speed of the growth is determined by the weaker type; see e.g. [@HP_absence Proposition 2.2]. This means that, in order not to grow too fast, the stronger type must survive by maintaining a meandering path surrounded by the weaker type. In fact, it can be shown that the fraction of the infected sites occupied by the stronger type is vanishing; see [@GM_invisible]. The crucial point in our half-plane argument is that infinite survival for the stronger type implies that it must occupy all sites along the positive horizontal axis. We use this to show that it will thereby grow fast enough to eventually surround the weaker type. Note that the role of the initial configuration is important for this argument. We have not been able to adapt the argument to rule out coexistence in the half-plane when the initial position of the stronger type is not connected to the horizontal axis. Indeed, working with general initial configurations seems to make the problem as hard as in the full plane. We remark that, in the full plane, it is shown in [@initial] that the initial configuration is irrelevant for the possibility of infinite coexistence, but that argument does not apply here.
One way of constructing the two-type process is by independently assigning a unit exponential random weight $\tau(e)$ to each nearest-neighbor edge $e$ of the lattice. The time required for type 1 to traverse an edge $e$ is then given by the associated weight $\tau(e)$, and the time for type 2 is $\lambda^{-1}\tau(e)$. Indeed, this construction provides a coupling of the two-type models for all $\lambda\ge1$ simultaneously. The curious partial result of [@HP_absence] is derived based on this coupling by showing that, in the probability measure underlying the coupling, there is almost surely at most one value of $\lambda$ for which coexistence may occur. That coexistence occurs with positive probability for at most countably many $\lambda\ge1$ is an easy consequence of this.
There are a number of proofs of coexistence for the case when the types have the same intensities, and (at least some of) these arguments can be adapted to prove the *if*-direction of Theorem \[thm:main\]. We shall however offer an alternative proof, since it is a simple by-product of the arguments required to prove the *only if*-direction of the theorem. To rule out coexistence for $\lambda>1$, we shall develop an argument inspired by the work of Blair-Stahn [@BS_thesis], and that incorporates elements of Busemann functions introduced by Hoffman [@Hoff_coex; @Hoff_geodesics]. Nevertheless, the proof will be a self-contained and elementary deduction from standard results in first-passage percolation.
The two-type Richardson model can be viewed as a two-type version of first passage percolation with exponential edge weights. One of the most fundamental results for first passage percolation is the shape theorem, asserting that the infected set at time $t$ converges on the scale $t^{-1}$ to a deterministic convex set $A$. In order to describe the structure of the proof of Theorem \[thm:main\], let $\theta$ denote the maximal angle between any supporting line of $A$ in the first coordinate direction and the vertical supporting line in the same coordinate direction; see Figure \[fig1\] (left picture). Then $\theta$ equals zero in case the shape is differentiable in the coordinate directions, and $\theta$ is at most $\pi/4$, which occurs if the shape is a diamond. Given $\eps>0$ and $n\in\Z$, we partition the upper half-plane $\Z\times\Z_+$ into two regions ${\mathcal{L}_\eps}(n)$ and ${\mathcal{R}_\eps}(n)$ as follows: Consider the semi-infinite line through $(n-1/2,0)$ with angle $\theta+\eps$ to the vertical line through the same point (see Figure \[fig1\], right picture), and write ${\mathcal{L}_\eps}(n)$ for the part of the upper half-plane to the left of this line, excluding points on the line, and ${\mathcal{R}_\eps}(n)$ for the part to the right of the line, including points on the line. Finally, define the strips $S_k:=\{(x,y)\in \Z^2:0\leq y\leq k\}$ and $S_k^+=\{(x,y)\in \Z^2: x\geq 0, 0\leq y\leq k\}$.
The proof of the *only if*-direction of Theorem \[thm:main\] can roughly be divided into three steps, where the first one may be considered the most fundamental:
Step (i)
: Show that, for every $\lambda\geq 1$ and $\eps>0$, if type 2 survives indefinitely, then almost surely type 2 reaches ${\mathcal{R}_\eps}(n)$ before type 1 for infinitely many $n\ge1$.
Step (ii)
: Show that, for every $\lambda>1$ there exists $\eps>0$ such that, if type 2 comes first to ${\mathcal{R}_\eps}(n)$, then for each each $k\ge1$ there is a positive probability (uniform in $n$) that type 2 occupies all vertices in $S_k\cap {\mathcal{R}_\eps}(n)$.
Step (iii)
: Show that, if type 2 conquers all but finitely many vertices in $S_k^+$ for $k$ large, then it will eventually almost surely defeat type 1.
Combining steps (i) and (ii) (or in fact a slight rephrasing of these claims) one obtains that, if type 2 survives indefinitely, then for all $k\geq 1$ it will almost surely conquer all but finitely many sites in the strip $S_k^+$ along the horizontal axis. According to step (iii), this means that type 1 will eventually become surrounded by type 2, ruling out coexistence.
(2,0) – ([4\*cos(45)-2]{},[4\*sin(45)]{}) – (-2,4) – ([4\*cos(135)-2]{},[4\*sin(135)]{}) – (-6,0) – cycle; (-7,0) – (2.5,0); (-2,-.5) – (-2,4.5); (2,0) – (2,4); (2,2) arc \[start angle=90, end angle=112.5, radius=2\]; (2,2) node\[anchor=south east\] [$\theta$]{}; (12,0) – ([12+4\*cos(67.5)/sin(67.5)]{},4) – (16,4) – (16,0) – cycle; (7,0) – (16.5,0); (8,-.5) – (8,4.5); (12,0) – (12,4); (12,2) arc \[start angle=90, end angle=67.5, radius=2\]; (12,0) – ([12+4\*cos(67.5)/sin(67.5)]{},4); (12,0) node\[anchor=north\] [$(n,0)$]{}; (13.25,2) node\[anchor=south east\] [$\scriptscriptstyle{\theta+\varepsilon}$]{}; (12,1.5) node\[anchor=south east\] [$\mathcal{L}_\varepsilon(n)$]{}; (16,1.5) node\[anchor=south east\] [$\mathcal{R}_\varepsilon(n)$]{};
The angle $\theta$ used to define the region ${\mathcal{R}_\eps}(n)$ can be motivated as follows: On one hand the claim in step (ii), which will be a consequence of the shape theorem, cannot hold for any angle larger than $\theta$. On the other hand, while the claim in step (i) certainly could be correct also for angles smaller than $\theta$ (assuming that $\theta>0$), proving such a thing would require detailed understanding of the structure of infinite one-sided geodesics in the half-plane setting. The information needed would go beyond our current understanding for the analogous objects in the full-plane. Of course, since we believe that the shape is differentiable (at least in coordinate directions) we consequently believe that $\theta=0$, and in this case we cannot do better that having ${\mathcal{R}_\eps}(n)$ defined as an $\varepsilon$-tilted vertical line.
The rest of the paper is organized so that relevant background on one-type first passage percolation is given in Section 2. In Section 3 we use Busemann functions to control the evolution of the one-type process to obtain a statement that will establish step (i). Section 4 is devoted to step (ii), which is essentially a consequence of the shape theorem. Finally, the proof of Theorem \[thm:main\] is completed in Section 5, where step (iii) is established by an adaption of an argument from [@HP_absence].
Preliminaries
=============
In standard first passage percolation each edge $e$ of some underlying graph is independently equipped with a non-negative random variable $\tau(e)$ from some common distribution. Throughout this paper, we shall assume that the underlying graph is the upper half-plane $\Z\times\Z_+$, equipped with edges between nearest-neighbors, and that the weights $\{\tau(e)\}$ are unit exponentials. Note that $\{\lambda^{-1}\tau(e)\}$ are then exponentials with parameter $\lambda$. Given a path $\Gamma$, we let $T_\lambda(\Gamma):=\sum_{e\in\Gamma}\lambda^{-1}\tau(e)$ and define the *passage time* between two sets $\Phi,\Psi\subset\Z\times\Z_+$ in the environment $\{\lambda^{-1}\tau(e)\}$ as $$T_\lambda(\Phi,\Psi):=\inf\big\{T_\lambda(\Gamma):\Gamma \mbox{ is a path in $\Z\times\Z_+$ connecting $\Phi$ to $\Psi$}\big\},\\$$ To simplify the notation, we write $T_1(\Gamma)=T(\Gamma)$, $T_1(\Phi,\Psi)=T(\Phi,\Psi)$, and $T_\lambda(x,y)$ for the passage time between $\{x\}$ and $\{y\}$ for $x,y\in\mathbb{Z}^2$. It is immediate from the construction that $T_\lambda(\Phi,\Psi)=\lambda^{-1} T(\Phi,\Psi)$ for all $\lambda\ge1$.
The above construction gives rise to a simultaneous coupling of the two-type processes for all $\lambda\geq 1$, where type 1 requires time $\tau(e)$ to traverse an edge $e$ while type 2 requires time $\lambda^{-1}\tau(e)$. The passage time $T(\mathbf{0},z)$ then denotes the time at which type 1 arrives at the site $z$, unless $z$ is already reached by type 2 by then, and $T_\lambda(\mathbf{1},z)$ similarly denotes the time it would take type 2 to reach $z$, unless impeded by type 1 along the way.[^4] In the case that $\lambda=1$, whether or not a site $z$ is eventually occupied by type 2 can be read out directly from $T$; it will in the case that $T(\mathbf{1},z)<T(\mathbf{0},z)$. Understanding the evolution in the two-type Richardson model thus leads us to recall some basic results for one-type first-passage percolation. Due to the relation $T_\lambda(x,y)=\lambda^{-1}T(x,y)$, we focus in the remainder of this section on the case $\lambda=1$; corresponding results for $\lambda>1$ are obtained by a simple scaling argument.
Although first passage percolation in half-planes has been studied before, e.g. in [@Ahl15; @ADH15; @WW98], the vast majority of the literature is concerned with the two and higher dimensional nearest-neighbor lattices. It will be convenient to survey some of the results here. In analogy with the notation in the half-plane, we shall denote by $\overline T(\Phi,\Psi)$ the passage time between the two sets $\Phi,\Psi\subset\Z^2$, where the infimum is now taken over paths in $\Z^2$ connecting $\Phi$ and $\Psi$.
A first crucial observation is that $\overline T$ defines a metric on $\Z^2$. In particular, it is subadditive in the sense that $$\overline T(x,y)\leq \overline T(x,z)+\overline T(z,y)\quad\mbox{for all }x,y,z\in\Z^2.$$ Using subadditive ergodic theory [@K68; @L85], one can establish the existence of a time constant $\mu\in(0,\infty)$ specifying the asymptotic inverse speed of the growth along the axes. Specifically, we have that $$\lim_{n\to\infty}\frac{\overline T(\org,\nbf)}{n}=\mu\quad \mbox{almost surely and in }L^1.$$ This can be extended to an arbitrary direction in the first octant, and hence by symmetry of $\Z^2$, to any arbitrary direction: For $\alpha\in[0,\pi/4]$, let $u_\alpha$ denote a unit vector with angle $\alpha$ to the $x$-axis, that is, $u_\alpha=(\cos\alpha,\sin\alpha)$. Also, for $x,y\in\R^2$, define $\overline T(x,y):=\overline T(x',y')$, where $x'$ and $y'$ are the points in $\Z^2$ closest to $x$ and $y$, respectively. Then there exists a directional time constant $\mu_\alpha\in(0,\infty)$ such that $$\label{dir_mu}
\lim_{n\to\infty}\frac{\overline T(\org,nu_\alpha)}{n}=\mu_{\alpha}\quad \mbox{almost surely and in }L^1.$$ By definition, we have $\mu_0=\mu$. We remark that passage times to lines rather than single points obey the same asymptotics. For instance, with $\bar{\ell}_\alpha(n)$ denoting the straight line through $nu_\alpha$ with normal vector $u_\alpha$, we have that $\frac1n\overline T(\org,\bar{\ell}_\alpha(n))$ converges to $\mu_\alpha$ almost surely. This can be seen as a consequence of the fundamental shape theorem, which in its first version dates back to the work of Richardson [@R73].
Since $\overline T$ defines a (random) metric on $\Z^2$ it is natural to investigate the shape of large balls in this metric. The shape theorem [@Kes73; @R73] states that the set of sites that can be reached from the origin within time $t$ converges almost surely on the scale $t^{-1}$ to a deterministic shape $A$, that is, with probability one, we have for every $\eps>0$ that $W(t):=\{x\in\R^2:\overline T(\org,x)\leq t\}$ satisfies $$(1-\eps)A\subset \frac{W(t)}{t}\subset (1+\eps)A \quad\mbox{for all large }t.$$ The asymptotic shape $A$ can be characterized as the unit ball in the norm defined by $\mu(x)=\lim_{n\to\infty}\frac1n\overline T(\org,nx)$ for $x\in\R^2$. It is thus known to be compact and convex, with non-empty interior, and it inherits all symmetries of $\Z^2$. Apart from this, very little is known about the properties of the shape. It has been studied by aid of simulations in [@AlmDeijfen15], where the results indicate that it is close to, but not identical to, a Euclidean disk. We remark that there is no theoretical support for $A$ being a Euclidean disk, and in large dimension it is known not to be a Euclidean ball.
When restricting the growth to a strip $S_k:=\{(x,y)\in \Z^2:0\leq y\leq k\}$ for some $k\ge1$, the speed of progression decreases. However, the thicker the strip, the smaller is the effect. To be precise, let ${T^{{\scriptscriptstyle}(k)}}(\Phi,\Psi)$ denote the passage time between $\Phi\subset S_k$ and $\Psi\subset S_k$, where the infimum is taken over paths $\Gamma\subset S_k$ connecting $\Phi$ and $\Psi$. Again, the subadditive ergodic theorem shows that $\frac1n{T^{{\scriptscriptstyle}(k)}}(\org,\nbf)$ converges (almost surely and in $L^1$) to some constant ${\mu^{{\scriptscriptstyle}(k)}}\in(0,\infty)$. Moreover, $$\label{mu_strip}
{\mu^{{\scriptscriptstyle}(k)}}\searrow\mu\quad\mbox{as }k\to\infty;$$ see e.g. [@Ahl15 Proposition 8]. A similar statement holds for directions other than the axes directions. As a consequence, one can show that a shape theorem holds also for first passage percolation in the upper half-plane $\Z\times\Z_+$, and that the asymptotic shape in this case is the half-plane restriction of the shape $A$ arising in the full-plane growth; see [@Ahl15 Theorem 1]. We shall occasionally need the following stronger form of this half-plane shape theorem, which is a consequence e.g. of [@Ahl15 Proposition 15]: For every $\eps>0$ we have, almost surely, for all $y$ and all but finitely many $z$ in $\Z\times\Z_+$ that $$\label{eq:better_shape}
\big|T(y,z)-\mu(z-y)\big|<\eps\max\{|z|,|z-y|\},$$ where $\mu$ is the time constant as determined by $\overline T$.
A one-type lemma
================
The aim of this section is to take the first and most fundamental step towards a proof of our main theorem. It will be crucial for ruling out coexistence in the case when $\lambda>1$, but we will use it also to give a short proof of coexistence in the case when $\lambda=1$. The result is a statement for the one-type process on $\Z\times\Z_+$.
\[lma:main\] For every $\eps>0$ there exists $\gamma>0$ such that $$\Pr\Big(T(-\nbf,\org)<T(-\nbf,{\mathcal{R}_\eps}(0)\backslash \{\org\})\text{ for all }n\ge1\Big)>\gamma.$$
Key to the proof of the lemma will be the notion of Busemann functions. Define, for all $n\ge1$ and sites $u,v$ in the half-plane, the Busemann-like function $$B_n(u,v):=T(-\nbf,u)-T(-\nbf,v).$$ Lemma \[lma:main\] can be rephrased to say that with positive probability $B_n(\org,v)<0$ for all $v\in{\mathcal{R}_\eps}(0)\setminus\{\org\}$ and $n\ge1$. We shall first show that, almost surely, $B_n(\org,v)<0$ may fail for some $n$ for at most finitely many $v$ (Lemma \[lma:negative\]). A local modification argument will then show that with positive probability it does not.
A key observation is that, for fixed $m\ge1$, the sequence $\{B_n(\org,\mbf)\}_{n\ge1}$ is almost surely increasing. The limit $$B(\org,\mbf):=\lim_{n\to\infty}B_n(\org,\mbf)$$ hence exists almost surely. Indeed, this turns out to be true for all $u$ and $v$, see [@ADH15], but we shall not need this fact. Instead, we shall make use of the following asymptotic property.
\[lma:Busemann\] For all $m\ge1$, we have that $\operatorname{{\mathbb E}}[B(\org,\mbf)]=-\mu\cdot m$, and $$\lim_{m\to\infty}\frac1mB(\org,{\bf m})=-\mu\quad\text{almost surely}.$$
A useful property of $B_n$ is that it is additive. The additivity carries over in the limit as $n\to\infty$ and for $B$ this implies that $$\label{eq:B}
\frac1mB(\org,\mbf)=\frac1m\sum_{j=0}^{m-1}B({\bf j},{\bf j+1}),
$$ where $B({\bf j},{\bf j+1}):=\lim_{n\to\infty}B_n({\bf j},{\bf j+1})$. Due to invariance with respect to horizontal shifts, sending $m$ to infinity in , the ergodic theorem yields the almost sure limit $\operatorname{{\mathbb E}}[B(\org,{\mathbf1})]$. By additivity, it only remains to identify $\operatorname{{\mathbb E}}[B(\org,{\mathbf1})]$ with $-\mu$.
To this end, we rephrase $B(\org,{\mathbf1})$ as a limit of partial averages, and obtain $$\operatorname{{\mathbb E}}[B(\org,{\mathbf1})] = \operatorname{{\mathbb E}}\Big[\lim_{n\to\infty}\frac1n\sum_{j=0}^{n-1}B_j(\org,{\mathbf1})\Big]=\lim_{n\to\infty}\frac1n\sum_{j=0}^{n-1}\operatorname{{\mathbb E}}[B_j(\org,{\mathbf1})],$$ where extraction of the limit is allowed by dominated convergence, since $|B_j(\org,{\mathbf1})|\le T(\org,{\mathbf1})$. Due to invariance with respect to horizontal shifts, we have further that $$\operatorname{{\mathbb E}}[B(\org,{\mathbf1})]=\lim_{n\to\infty}\frac1n\sum_{j=0}^{n-1}\operatorname{{\mathbb E}}[T(\org,{\bf j})-T(\org,{\bf j+1})]=\lim_{n\to\infty}\frac1n\operatorname{{\mathbb E}}[-T(\org,{\bf n})]=-\mu,$$ as required.
Let $\partial {\mathcal{R}_\eps}(n)$ denote the set of sites in ${\mathcal{R}_\eps}(n)$ that have at least one neighbor in ${\mathcal{L}_\eps}(n)$.
\[lma:negative\] There exists $\delta>0$ such that, with probability one, for all $n\ge1$ and all but finitely many $v\in\partial{\mathcal{R}_\eps}(0)$, we have that $$B_n(\org,v)<-\delta|v|\mu<0.$$
Note that, by convexity of the shape and the definition of $\theta$, there exists $\delta>0$ such that for each $v\in\partial {\mathcal{R}_\eps}(0)$ there is $m=m(v)$ such that $$\mu(v-\mbf)\le(1-\delta)\mu(\mbf);$$ see Figure \[fig2\] (left picture).
(-10,0) – (6,0); (0,-.5) – (0,4); (0,0) – (1.5,4); (14,0) – (13.5,0) – (14.625,3) – (15.125,3) – cycle; (11,0) – (16,0); (14,-.5) – (14,4); (14,0) – (15.5,4); (-9,0) to \[curve through=[(-8.5,.4) (-8,.6) (-7.5,.5) (-7,.65) (-6,.75) (-5,1) (-4,1.5) (-3,1.75) (-2,1.75) (-1,1.65) (0,1.8)]{}\] (.75,2); (-6,.75) to \[curve through=[(-5,.65) (-4,.45) (-3,.75) (-2,.45) (-1,.35)]{}\] (0,0); (14,0) to \[curve through=[(13.5,.25) (13.2,.2) (13.4,.75) (13,1.25) (13.5,1.75) (14,2.3) (14.25,2.2)]{}\] (14.75,2); (-9.3,0) node\[anchor=north\] [$-\mathbf{n}$]{}; (4,0) node\[anchor=north\] [$\mathbf{m}$]{}; (12,1.5) node\[anchor=west\] [$\Gamma$]{}; (.75,2) node\[anchor=west\] [$v$]{}; (2.5,2.7) node\[anchor=west\] [${\mathcal{R}_\eps}(0)$]{}; (14.75,2) node\[anchor=west\] [$\ell$]{}; (15.125,3) node\[anchor=west\] [$\ell'$]{}; (0,0) circle (.125); (.75,2) circle (.125); (-9,0) circle (.125); (4,0) circle (.125); (14.625,2) – (14.825,2); (15,3) – (15.25,3);
Indeed, $m$ can be chosen to roughly equal $c|v|$ for some $c>0$. Together with the strong version of the shape theorem stated in , it follows that almost surely for all but finitely many $v\in\partial {\mathcal{R}_\eps}(0)$ we have that $$B_n(\mbf,v)\le T(\mbf,v)\le(1+\delta)\mu(v-\mbf)\le(1-\delta^2)\mu(\mbf).$$ Moreover, by monotonicity of $B_n$ and Lemma \[lma:Busemann\], we have almost surely for all $n\ge1$ and large $m$ that $$B_n(\org,\mbf)\le B(\org,\mbf)\le-(1-\delta^2/2)\mu(\mbf).$$ Combining the two estimates we conclude that almost surely, for all but finitely many $v\in \partial {\mathcal{R}_\eps}(0)$, we have for all $n\ge1$ that $$B_n(\org,v)=B_n(\org,\mbf)+B_n(\mbf,v)\le-(\delta^2/2)\mu(\mbf)<0.$$ Since $m$ is roughly $c|v|$ for some $c>0$, the lemma follows.
Let $(x_n,y_n)$ be the point in ${\mathcal{R}_\eps}(0)$ with the smallest passage time to $-\nbf$. By Lemma \[lma:negative\], the sequence $(y_n)_{n\ge1}$ is almost surely bounded. Fix $\ell$ large so that, with probability at least $3/4$, we have $y_n\leq \ell$ for all $n$. Then pick some finite path $\Gamma$, connecting the origin to a point in $\partial{\mathcal{R}_\eps}(0)$ of the form $(x,\ell+1)$, which except for its endpoints is contained in ${\mathcal{L}_\eps}(0)$; see Figure \[fig2\] (right picture). Next, take $t$ large so that, with probability at least $3/4$, the total passage time $T(\Gamma)$ is at most $t$. Note that, since any path from $-\nbf$ to $(x,y)\in\partial{\mathcal{R}_\eps}(0)$ with $1\leq y\leq \ell$ must hit $\Gamma$ before hitting $(x,y)$, we have for all $n\geq 1$ that, on the intersection of the above two events, $$T(-\nbf,\org)\leq T(-\nbf,{\mathcal{R}_\eps}(0))+T(\Gamma)\le T(-\nbf,{\mathcal{R}_\eps}(0))+t.$$ Write $U_{\ell'}$ for the set of sites $(x,y)\in \partial{\mathcal{R}_\eps}(0)$ with $y\geq \ell'$. Due to Lemma \[lma:negative\], we may pick $\ell'\geq \ell$ such that $T(-\nbf,U_{\ell'})\geq T(-\nbf,\org)+2t$ for all $n\ge1$ with probability at least $3/4$. Define $C$ to be the intersection of all three events above. That is, let $$C:=\big\{y_n\leq \ell\text{ for all }n\big\}\cap \big\{T(\Gamma)\leq t\big\}\cap \big\{T(-\nbf,U_{\ell'})\geq T(-\nbf,\org)+2t\text{ for all }n\big\},$$ and note that $\Pr(C)\ge1/4$.
Let $\Lambda_{\ell'}$ denote the set of edges connecting sites $(x,y)\in\partial{\mathcal{R}_\eps}(0)\setminus\{\org\}$ with $y\leq \ell'$ to sites in ${\mathcal{L}_\eps}(0)$; see Figure \[fig2\] (shaded area in the right picture). We complete the proof by arguing that, on the event $C$, a configuration where the origin is the closest point in $\partial{\mathcal{R}_\eps}(0)$ to $-\nbf$ for all $n\geq 1$ is obtained by increasing the weight of all edges in $\Lambda_{\ell'}$ to $2t$. Indeed, the time minimizing path from $-\nbf$ to ${\mathcal{R}_\eps}(0)$ will then not hit a point $(x,y)\in{\mathcal{R}_\eps}(0)$ for $y=1,\ldots,\ell'$, since it would have reached the origin via $\Gamma$ before the last edge is traversed. It will also not hit ${\mathcal{R}_\eps}(0)$ for $y\geq \ell'$, since it will take at least time $2t$ from the moment when $\Gamma$ is hit to reach that level.
To formalize this, we define another i.i.d. family of edge weights $\{\hat\tau(e)\}$, where $\hat\tau(e)=\tau(e)$ for $e\not\in\Lambda_{\ell'}$ and where $\hat\tau(e)$ is sampled independently of $\tau(e)$ for $e\in\Lambda_{\ell'}$. Denoting by $Q$ the event $\{\hat{\tau}(e)>t\mbox{ for all }e\in\Lambda_{\ell'}\}$, and distances with respect to $\{\hat\tau(e)\}$ by $\hat T$, the above reasoning gives that $$\Pr\left(\hat{T}(-\nbf,\org)<\hat{T}(-\nbf,{\mathcal{R}_\eps}(0)\backslash \{\org\})\text{ for all }n\ge1\right)
\geq \Pr\big(C\cap Q\big)=\Pr(C)\Pr(Q)>0,$$ due to independence of the two configurations on $\Lambda_{\ell'}$. Since the two configurations are equal in distribution, the lemma follows.
A two-type lemma
================
The next lemma concerns the two-type process with an unbounded initial configuration. It applies when type 2 is strictly stronger than type 1, and is derived as a geometric consequence of the shape theorem. Recall that $S_k^+=\{(x,y)\in \Z^2: x\geq 0, 0\leq y\leq k\}$. Note also that for small enough values of $\eps>0$ the origin is the only site on the horizontal axis contained in $\partial{\mathcal{R}_\eps}(0)$.
\[lma:chance\] For every $\lambda>1$ there is $\eps>0$ such that if initially $\org$ is occupied by type 2 and all sites in $\partial{\mathcal{R}_\eps}(0)\backslash\{\org\}$ are occupied by type 1, then, for every $k\ge1$, there is a positive probability that type 2 occupies all initially uninfected sites in the half-strip $S_k^+$.
Fix $\lambda>1$. Note that it suffices to prove the lemma for large $k$, since if type 2 occupies all uninfected sites in $S_k^+$, then this is trivially the case also for all $k'\leq k$. By we thus pick $k$ large so that $$\lambda^{-1}\mu\le\lambda^{-1}{\mu^{{\scriptscriptstyle}(k)}}<\mu.$$ Let $\delta=(\mu-\lambda^{-1}{\mu^{{\scriptscriptstyle}(k)}})/4$ and set $\rho=\lambda^{-1}{\mu^{{\scriptscriptstyle}(k)}}+2\delta_0$.
It follows from the half-plane shape theorem (the version stated in ), convexity of the shape and the definition of ${\mathcal{R}_\eps}(0)$ that, almost surely, for large $n$ we have that $$T(\partial{\mathcal{R}_\eps}(0),(n,k))>(\mu-\delta')n$$ for some $\delta'=\delta'(\eps)>0$, with $\delta'\to 0$ as $\eps\to 0$. Hence, for $\eps>0$ small, $\mu-\delta'\geq\rho+\delta$. Moreover, almost surely, we have that $${T^{{\scriptscriptstyle}(k)}}_\lambda(\org,(n,k))<(\rho-\delta)n$$ for all large $n$. Finally, write $\tilde{T}^{{\scriptscriptstyle}(k)}_\lambda(\org,(n,k))$ for the above passage time in the process based on $\{\lambda^{-1}\tau(e)\}$, when sites in $\partial{\mathcal{R}_\eps}(0)$ cannot be used, and note that this clearly obeys the same asymptotics. (We assume here and in what follows that $\eps>0$ is small, so that the origin is the only site on the horizontal axis contained in $\partial{\mathcal{R}_\eps}(0)$.)
For $m\ge1$ now define $$\begin{aligned}
D_m&:=\{T(\partial{\mathcal{R}_\eps}(0),(n,k))>(\rho+\delta)n\mbox{ for all }n\geq m\},\\
D'_m&:=\{\tilde{T}^{{\scriptscriptstyle}(k)}_\lambda(\org,(n,k))<(\rho-\delta)n\mbox{ for all }n\geq m\},
\end{aligned}$$ and pick $m$ large so that $\Pr(D_m\cap D_m')>3/4$. Let $\Omega_m$ denote the set of edges consisting of all edges connecting an initially type 1 infected site to a neighbor in $S_k^+$, and all vertical edges connecting a site $(j,k+1)$ in ${\mathcal{R}_\eps}(0)$ with $j\le m$ to $(j,k)$. Hence $\Omega_m$ consists of all edges up to the level $x=m$ through which type 1 can enter the strip; see Figure \[fig3\].
(-1,0) – (9,0); (0,-.5) – (0,4); (0,2) node\[anchor=east\] [$k$]{}; (7,0) node\[anchor=north\] [$m$]{}; (0,0) – (0.86,2.3) – (7,2.3) –(7,2) – (1.1,2) – (0.3,0) – cycle; (0,0) – (1.5,4); (0,2) – (9,2); (7,0) – (7,4); (3,1) node\[anchor=west\] [$\Omega'_m$]{}; (4,3) node\[anchor=west\] [$\Omega_m$]{}; (4,3) – (2.8,2.15);
Also, let $\Omega_m'$ denote the set of edges connecting initially uninfected sites in $S_k^+$ up to level $x=m$, and note that $\Omega_m$ and $\Omega_m'$ are disjoint.
Next, let $$\begin{aligned}
E_{m,t}&:=\{\mbox{$\tau(e)>tkm$ for all $e\in\Omega_m$}\},\\
E_{m,t}'&:=\{\mbox{$\lambda^{-1}\tau(e)<t$ for all $e\in\Omega_m'$}\}.
\end{aligned}$$ Since $\Pr(E_{m,t}')\to 1$ as $t\to\infty$, we can pick $t$ large so that $\Pr(D_m\cap D_m'\cap E_{m,t}')>1/2$. We claim that, on $D_m\cap D_m'\cap E_{m,t}\cap E_{m,t}'$, type 2 occupies all initially uninfected sites in $S_k^+$. To see this, note that $E_{m,t}\cap E_{m,t}'$ ensures that type 1 cannot enter the strip at a site $(j,k)$ with $j<m$, since any such site can be reached from the origin by a path in $\Omega_m'$ with weight at most $mkt$. The event $D_m\cap D_m'$ then guarantees that type 1 cannot enter the strip at a site $(j,k)$ with $j\geq m$, since type 2 is faster to all such sites once it has access to the initial piece of the strip.
It remains to prove that $\Pr(D_m\cap D_m'\cap E_{m,t}\cap E_{m,t}')>0$. To this end, write $$\Pr(D_m\cap D_m'\cap E_{m,t}\cap E_{m,t}')=\Pr(D_m\cap D_m'\cap E_{m,t}'|E_{m,t})\Pr(E_{m,t}).$$ The events $D_m'$ and $E_{m,t}'$ involve only edges in $\Omega_m'$ while $E_{m,t}$ involves only edges in $\Omega_m$. Hence, since $\Omega_m'$ and $\Omega_m$ are disjoint, the conditioning on $E_{m,t}$ does not affect $D_m'$ and $E_{m,t}'$. As for $D_m$, the event $E_{m,t}$ stipulates that the passage times on certain edges are large. This clearly increases the probability of $D_m$ so that, in summary, $\Pr(D_m\cap D_m'\cap E_{m,t}'|E_{m,t})\geq \Pr(D_m\cap D_m'\cap E_{m,t}')$. The desired conclusion follows by noting that $\Pr(E_{m,t})>0$ since $\Omega_m$ is finite and $t$ fixed.
Proof of Theorem \[thm:main\]
=============================
In this section we prove Theorem \[thm:main\]. As mentioned, there are a number of proofs in the literature showing that coexistence is possible on $\Z^2$ when $\lambda=1$, and some of these are easily adapted to show the same statement in the half-plane. However, this can also be obtained by a short argument based on Lemma \[lma:main\].
Take $\lambda=1$. Let $F$ denote the event in Lemma \[lma:main\], and let $\bar F$ denote its reflection in the vertical axis. Let further $\bar F_m$ denote the translate of $\bar F$ along the vector $\mbf$. We observe that, on $F$, type 1 will be first to all sites along the negative horizontal axis. Similarly, on $\bar F_1$, type 2 will be first to all sites along the positive horizontal axis. Although there is no guarantee that the intersection of the two events occurs with positive probability, since $\bar F_m$ occurs with a density (due to the ergodic theorem), we may fix $m\ge1$ so that $\Pr(F\cap\bar F_m)>0$. To guarantee coexistence, it then remains to show that, on $F\cap\bar F_m$, there is positive probability for type 2 to reach $(m,0)$ before type 1 reaches ${\mathcal{R}_\eps}(m)$.
Let $O$ denote the event that each edge adjacent to the origin has weight at least $\delta$, and note that $\Pr(F\cap\bar F_m\cap O)>0$ for small $\delta>0$. Let $O'$ denote the event that the sum of the weights on the edges along the axis connecting ${\mathbf1}$ to $\mbf$ is at most $\delta/2$. Note that, on $O\cap O'$, type 2 will reach $\mbf$ before type 1 takes its first step. Since $F$, $\bar F_m$ and $O$ are independent of the state of the edges defining $O'$, it follows that $$\Pr(\coex)\geq\Pr(F\cap\bar F_m\cap O\cap O')=\Pr(F\cap\bar F_m\cap O)\Pr(O')>0,$$ as required.
We proceed with the *only if*-direction, and start by combining Lemmas \[lma:main\] and \[lma:chance\] into a statement for the two-type process.
\[lma:strip\] For every $\lambda>1$ and $k\geq 1$, if type 2 occupies infinitely many sites in the two-type model on $\Z\times \Z^+$, then type 2 will almost surely occupy all but finitely many vertices in $S_k^+$.
Fix $\lambda>1$ and $k\ge1$. Write $F$ for the event in Lemma \[lma:main\], and let $F_m$ denote the translate of $F$ along the vector $\mbf$. Also, let $G$ denote the event in Lemma \[lma:chance\], and let $G_m$ denote the translate of $G$ along the vector $(m,0)$. Each of the two events $F$ and $G$ occur with positive probability. Moreover, $F$ is determined by edges between sites in ${\mathcal{L}_\eps}(0)\cup \partial{\mathcal{R}_\eps}(0)$ involving at least one site in ${\mathcal{L}_\eps}(0)$, while $G$ is determined by edges between pairs of sites in ${\mathcal{R}_\eps}(0)$. Hence, the two events are independent and $\Pr(F\cap G)>0$. By the ergodic theorem, $F_m\cap G_m$ will occur for infinitely many $m\ge1$, almost surely.
It remains to prove that, on the event $F_m\cap G_m\cap \{\text{type 2 survives}\}$, where $m\ge1$, type 2 occupies all but finitely many vertices in $S_k^+$. For this, clearly it suffices to see that, if type 2 survives indefinitely, then $F_m$ implies that type 2 reaches $\mbf$ before any other site in $\partial{\mathcal{R}_\eps}(m)$ is reached by type 1. To this end, let $\Gamma$ denote the time minimizing path from the origin to $\mbf$. Note that, if type 2 survives indefinitely, then $\mbf$ must be occupied by type 2 in the two-type process. Let $v$ denote the first (in time) point on the path $\Gamma$ that is occupied by type 2 in the two-type process. The fastest way to get from $v$ to ${\mathcal{R}_\eps}(m)$ is to follow $\Gamma$ and, doing this, type 2 will arrive at $\mbf$ before any other site in ${\mathcal{R}_\eps}(m)$ is infected, as desired.
The last ingredient we need in order to prove the *only if*-direction of Theorem \[thm:main\] is a half-plane version of a result from [@HP_absence Proposition 2.2]. More precisely, we need to show that, if type 2 conquers a wide half-strip, then type 2 will end up surrounding type 1. The argument will be similar to that of [@HP_absence], but the geometric construction is easier in our case and the proof consists of applying the ideas in Lemmas \[lma:chance\] and \[lma:strip\] in non-axis directions. We shall therefore be brief.
\[lma:speed\] For every $\lambda>1$, there is $k\ge1$ such that, if type 2 occupies all but finitely many sites in $S_k^+$, then almost surely type 1 will occupy only finitely many sites.
If type 2 occupies all but finitely many sites in the half-strip $S_k^+$ for $k$ sufficiently large, then the type 2 speed along the axis in $S_k^+$ will be strictly larger than the speed of type 1. As we shall see, type 2 will then be strictly faster than type 1 also in direction $\alpha$, for some small $\alpha>0$. This can be used to show that type 2 occupies all but finitely many vertices in an $\alpha$-cone. By repeating the argument we then show that type 2 will also occupy almost all sites in a $2\alpha$-cone, etc.
Recall the definition, in , of the time constant $\mu_\alpha$ in direction $\alpha$ based on unit rate exponential edge weights. The time constant in direction $\alpha$ based on exponential edge weights with parameter $\lambda$ is then given by $\lambda^{-1}\mu_\alpha$. As is well-known, the directional time constant $\mu_\alpha$ is Lipschitz continuous, since $\mu$ defines a norm. In particular, there exists a constant $c>0$ such that, for any $\alpha_0,\alpha\in[0,2\pi]$, we have that $$\mu_{\alpha_0+\alpha}\leq \mu_{\alpha_0}(1+c\alpha).$$ It follows that, uniformly in the choice of $\alpha_0$, we have $\lambda^{-1}\mu_{\alpha_0+\alpha}\leq \mu_{\alpha_0}$ if $\alpha$ is sufficiently small. By picking $\alpha$ even smaller, we further obtain that $\lambda^{-1}\mu_{\alpha_0+\alpha}<\mu_{\alpha_0}\cos\alpha$. For the remainder of this proof we fix $\alpha>0$ so that for all $\alpha_0\in[0,2\pi]$ we have $$\label{eq:alpha_choice}
\lambda^{-1}\mu_{\alpha_0+2\alpha}<\mu_{\alpha_0}\cos(2\alpha).$$
Let $\ell_\alpha(0)$ denote the semi-infinite line starting at the origin with angle $\alpha$ to the horizontal axis. In a first step, we argue that if type 2 occupies all sites in a thick strip, then type 2 will almost surely occupy all but finitely many sites below the line $\ell_\alpha(0)$. Pick $k$ large so that $\lambda^{-1}{\mu^{{\scriptscriptstyle}(k)}}<\mu$, which is possible by . Let $H_m$ be the event that type 2 eventually occupies the site $(m,k)$, and that at the time at which this occurs type 1 has not yet reached the vertical line $L(m)=\{(x,y)\in\mathbb{Z}^2:x=m\}$. The choice of $k$ assures that, given that type 2 captures all but finitely many sites in the strip $S_k^+$, the probability of $H_m$ tends to one as $m$ tends to infinity.
Write $\ell_{2\alpha}(m,k)$ for the semi-infinite line starting at the point $(m,k)$ with angel $2\alpha$ to the horizontal line through $(m,k)$. By , we have that $\lambda^{-1}\mu_{2\alpha}< \mu\cos(2\alpha)$, and hence that the asymptotic type 2 time from $(m,k)$ to a point on $\ell_{2\alpha}(m,k)$ far from $(m,k)$ is strictly smaller than the type 1 passage time from $L(m)$ to the same point; see Figure \[fig4\].
(-5,-1) – (9,-1); (-4,-2) – (-4,5); (0,-1) – (0,5); (-4,0) – (9,0); (-5,0) node\[anchor=west\] [$k$]{}; (0,-1) node\[anchor=north\] [$m$]{}; (0,0) circle (.125); (5.6,3) circle (.125); (0,0) – (7.5,4); (7.5,4) node\[anchor=north west\] [$\ell_{2\alpha}(m,k)$]{}; (1,0) arc \[start angle=0, end angle=27, radius=1\]; (1.1,0.4) node\[anchor=west\] [$2\alpha$]{}; (0,3) – (5.6,3); (2.2,1.8) node\[anchor=west\] [$x$]{}; (1,3.3) node\[anchor=west\] [$x\cos(2\alpha)$]{};
(Here, we say that a point $z\in\R^2$ is infected when the closest point in $\Z^2$ is infected.) Let $G_m^{{\scriptscriptstyle}2\alpha}$ denote the event that, starting from a configuration in which $(m,k)$ is of type 2 and the rest of the line $L(m)$ is of type 1, every point along the line $\ell_{2\alpha}(m,k)$ is eventually captured by type 2. A similar argument as that used to prove Lemma \[lma:chance\] then shows that $G_m^{{\scriptscriptstyle}2\alpha}$ occurs with positive probability. The ergodic theorem implies that $G_m^{{\scriptscriptstyle}2\alpha}$ occur for a positive density of all $m\ge1$, almost surely, and since the conditional probability that $H_m$ occurs, given that type 2 takes the strip, tends to one, their intersection will occur for some (large) value of $m$ almost surely. The occurrence of $H_m\cap G_m^{{\scriptscriptstyle}2\alpha}$ guarantees that type 2 captures the whole line $\ell_{2\alpha}(m,k)$, and consequently that the whole area below the line $\ell_{2\alpha}(m,k)$ is captured by type 2. Since $\ell_\alpha(0)$ eventually enters this region, we conclude that if type 2 captures all but finitely many sites in $S_k^+$ (and $k$ is large), then almost surely type 2 captures all but finitely many sites in the cone below the line $\ell_\alpha(0)$.
In a second step we show that for any $\alpha_0>0$, if type 2 occupies all but finitely many vertices in the $\alpha_0$-cone below the line $\ell_{\alpha_0}(0)$, then the same is true for the $(\alpha_0+\alpha)$-cone below the line $\ell_{\alpha_0+\alpha}(0)$, almost surely. Since $\alpha_0$ is arbitrary, this will complete the proof of the lemma. We repeat the argument above, and let $v_m$ denote the point on $\ell_{\alpha_0}(0)$ at distance $m$ from the origin, write $\ell_{2\alpha}^{\alpha_0}(m)$ for the semi-infinite line starting at $v_m$ with angel $2\alpha$ to $\ell_{\alpha_0}(0)$, and let $\bar{\ell}_{\alpha_0}(m)$ be the line through $v_m$ that is orthogonal to $\ell_{\alpha_0}(0)$; see Figure \[fig5\].
(-1,0) – (12,0); (0,-1) – (0,6); (0,0) – (10,2.5); (9,1.6) node\[anchor=west\] [$\ell_{\alpha_0}(0)$]{}; (4.8,0) – (3.53,5); (4.2,5.8) node\[anchor=east\] [$\bar{\ell}_{\alpha_0}(m)$]{}; (1.5,0) arc \[start angle=0, end angle=20, radius=1\]; (1.5,0.22) node\[anchor=west\] [$\scriptstyle\alpha_0$]{}; (4.5,1.125) – (6.5,5.5); (5.5,6.2) node\[anchor=west\] [$\ell_{2\alpha}^{\alpha_0}(m)$]{}; (5.5,1.375) arc \[start angle=0, end angle=60, radius=1\]; (5.2,2.1) node\[anchor=west\] [$\scriptstyle2\alpha$]{}; (4.5,1.125) circle (.125);
Now, if type 2 occupies all but finitely many vertices in the $\alpha_0$-cone, then its asymptotic speed in direction $\alpha_0$ is determined by $\lambda^{-1}\mu_{\alpha_0}$. Hence, if type 2 occupies all but finitely many vertices in the $\alpha_0$-cone, then the event $H_m^{\alpha_0}$ that type 1 has not yet reached $\bar{\ell}_{\alpha_0}(m)$ when type 2 reaches $v_m$ has probability tending to one as $m\to\infty$.
Furthermore, by it again follows that the type 2 time from $v_m$ to a point far along the line $\ell_{2\alpha}^{\alpha_0}(m)$ is with high probability strictly smaller than the type 1 passage time from $\bar{\ell}_{\alpha_0}(m)$ to the same point. Again repeating the argument in the proof of Lemma \[lma:chance\], we may show that the event $G_m^{{\scriptscriptstyle}\alpha_0,2\alpha}$ that the whole line $\ell_{2\alpha}^{\alpha_0}(m)$ is captured by type 2, when starting from a configuration where $v_m$ is of type 2 and the rest of the sites on or to the left of the line $\bar{\ell}_{\alpha_0}(m)$ is of type 1, occurs with positive probability. Appealing to the ergodic theorem we again find that, given that type 2 takes all but finitely many sites in the $\alpha_0$-cone, the event $H_m^{\alpha_0}\cap G_m^{\alpha_0,2\alpha}$ will occur for some (large) $m$, almost surely, and so type 2 will occupy all but finitely many sites in the $(\alpha_0+\alpha)$-cone below the line $\ell_{\alpha_0+\alpha}(0)$. Since $\alpha_0$ was arbitrary, this completes the proof.
The *only if*-direction of Theorem \[thm:main\] is an immediate consequence of Lemmas \[lma:strip\] and \[lma:speed\].
[99]{}
Ahlberg, D. (2015): Convergence towards an asymptotic shape in first-passage percolation on cone-like subgraphs of the integer lattice, *J. Theoret. Probab.* [**28**]{}, 198–222.
Auffinger, A., Damron, M. and Hanson, J. (2015): Limiting geodesics for first-passage percolation on subsets of $\mathbb{Z}^2$, *Ann. Appl. Probab.* [**25**]{}, 373-405.
Alm, S.E. and Deijfen, M. (2015): First passage percolation on $\mathbb{Z}^2$ – a simulation study, *J. Stat. Phys.* [**161**]{}, 657-678.
Blair-Stahn, N. (2012): *A geometric perspective on first-passage percolation*, Ph.D. Dissertation in Mathematics, University of Washington, arXiv:1212.6254.
Deijfen, M. and Häggström, O. (2006): The initial configuration is irrelevant for the possibility of mutual unbounded growth in the two-type Richardson model, [*Comb. Probab. Computing*]{} **15**, 345-353.
Garet, O. and Marchand, R. (2005): Coexistence in two-type first-passage percolation models, [*Ann. Appl. Probab.*]{} [**15**]{}, 298-330.
Garet, O. and Marchand, R. (2007): First-passage competition with different speeds: positive density for both species is impossible, *Electron. J. Probab.* [**13**]{}, 2118-2159.
Häggström, O. and Pemantle, R. (1998): First passage percolation and a model for competing spatial growth, *J. Appl. Probab.* **35**, 683-692.
Häggström, O. and Pemantle, R. (2000): Absence of mutual unbounded growth for almost all parameter values in the two-type Richardson model, *Stoch. Proc. Appl.* **90**, 207-222.
Hoffman, C. (2005): Coexistence for Richardson type competing spatial growth models, [*Ann. Appl. Probab.*]{} [**15**]{}, 739-747.
Hoffman, C. (2008): Geodesics in first passage percolation, *Ann. Appl. Probab.* [**18**]{}, 1944-1969.
Kesten, H. (1973): Discussion contribuiton, *Ann. Probab.* [**1**]{}, 903.
Kingman, J.F.C. (1968): The ergodic theory of subadditive stochastic processes, *J. Roy. Statist. Soc. Ser. B* **30**, 499-510.
Liggett, T.M. (1985): An improved subadditive ergodic theorem, *Ann. Probab.* **13**, 1279-1285.
Richardson, D. (1973): Random growth in a tessellation, *Proc. Cambridge. Philos. Soc.* [**74**]{}, 515-528.
Wehr, J. and Woo, J. (1998): Absence of geodesics in first-passage percolation on a half-plane, *Ann. Probab.* **26**, 358-367.
[^1]: Department of Mathematics, Stockholm University; [[email protected]]{}
[^2]: Department of Mathematics, Stockholm University; [[email protected]]{}
[^3]: Department of Mathematics, University of Washington; [[email protected]]{}
[^4]: Throughout the paper, we shall let bold letters like $\nbf$ be short for the horizontal vectors $(n,0)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the problem of varying the security of blockchain transactions according to their importance. This adaptive security is achieved by using variable size consensus committees. To improve performance, such committees function concurrently. We present two algorithms that allow adaptive security by forming concurrent variable size consensus committees on demand. One is based on a single joint blockchain, the other is based on separate sharded blockchains. For in-committee consensus, our algorithms may use various available byzantine-robust fault tolerant algorithms (BFT). We implement synchronous BFT, asynchronous BFT and proof-of-work consensus. We thoroughly evaluate the performance of our adaptive security algorithms.'
author:
- 'Shishir Rai, Kendric Hood, Mikhail Nesterenko, and Gokarna Sharma\'
- 'Shishir Rai, Kendric Hood, Mikhail Nesterenko, and Gokarna Sharma'
bibliography:
- 'blockguard.bib'
title: 'Blockguard: Adaptive Blockchain Security'
---
Introduction
============
**Blockchain.** A secure distributed ledger, *blockchain* allows a decentralized network of peers to register a sequence of transactions despite potentially malicious actions of a minority of peers. Blockchain technology is poised to revolutionize a variety of fields: from currency and payment systems that are impervious to state and corporate manipulation [@nakamoto; @ethereum], to automatically enforced contracts [@Abdellatif2018; @ConcurrencySmartContracts], to internet-of-things massive data recording [@iotchain].
Peers acting maliciously are *Byzantine* [@byzantine]: a Byzantine peer may exhibit arbitrary behavior. Such malicious peers are controlled by *adversary* that uses them to compromise the blockchain. To overcome the adversary, peers coordinate their collective decision. This decision may be achieved using classic byzantine-robust coordinated consensus [@pbft; @paxos; @byzantine] or novel proof-of-work [@nakamoto; @ethereum].
In a coordinated consensus algorithm, the peers exchange messages to arrive at a uniform conclusion. Such an algorithm usually requires every peer to know the identities of all other peers. That is, a completely connected network topology is necessary. A coordinated consensus algorithm operates correctly so long as the number of Byzantine peers is less than its resiliency threshold. This threshold is a fraction of the committee size. In a fast changing peer-to-peer network, such strict membership requirement is problematic. In a proof-of-work algorithm, the peers compete to solve a computationally intensive problem. All peers accept the solution of the peer who solves the problem first. Such algorithms do not require membership or network topology maintenance. The algorithm operates correctly provided that the computation power of the Byzantine peers does not exceed the power of the correct peers. That is, the resiliency threshold of a proof-of-work algorithm is expressed in terms of the computation power. However, such an algorithm needs extensive computing resources and has poorer throughput and latency. This limits the scalability of proof-of-work networks.
There are plenty of studies enhancing the scalability of proof-of-work networks [@solida; @decker2016bitcoin; @bitcoinng; @algorand; @byzcoin; @honeybadger; @thunderella; @phantomGhostTAG] , optimizing membership maintenance of the coordinated consensus schemes or joining the two approaches [@hybrid]. Ultimately, the scalability of both coordinated consensus and proof-of-work networks is limited by the need to broadcast the information to all peers. To avoid this broadcast, the network needs to be logically split up or sharded. In *sharding*, the network is separated into independent committees that process transactions concurrently. Sharded committees may use either coordinated or competitive consensus.
Sharding may potentially resolve performance issues. However, the separate committees do not provide the same protection as the complete network. Indeed, to compromise transaction processing, the adversary needs to exceed the resiliency threshold of the consensus algorithm for the single committee. Sharding also introduces the problem of cross-shard transactions where transaction inputs and outputs span shards [@crossChain; @omniledger; @elastico; @rapidchain].
\
**Adaptive security.** Thus, performance and security are cross-purposes of blockchain design. We propose to mitigate this problem with *adaptive security*. Specifically, the client may request the security level for the submitted transaction on the basis of its importance. For example, the security level of purchasing a cup of coffee, may be lower than that of buying a car. The fees for registering a higher security level transaction in the blockchain may be higher. The greater security level is achieved by composing a larger committee which is less vulnerable to adversarial attack since it has a higher resiliency threshold.
A stream of various security level transactions requires an *adaptive security algorithm* that provides an appropriate size committee for each transaction. A naive solution would sequentially process transactions at the highest security level. However, employing an excessively large committee would waste network resources while sequential transaction processing may result in low throughput and high transaction waiting time. Instead, it is more efficient to assemble the consensus committees on demand and allow concurrency in transaction processing.
In this paper, we propose two such security algorithms: Composite Blockguard and Dynamic Blockguard. Composite Blockguard assembles a committee from groups of peers that maintain independent ledgers. Dynamic Blockguard selects a committee from the processes that most recently wrote to the shared ledger. Both algorithms may operate with many consensus algorithms.
\
**Our contribution.** We state the problem of adaptive security and propose two efficient algorithms that solve it. We evaluate their performance with major consensus algorithms: PBFT, SBFT and proof-of-work. We measure their throughput, transaction waiting time and resistance to Byzantine peer corruption. Our results suggest that the adaptive security provides an effective trade-off between network performance and security without significant increase in network complexity or major architectural modifications. Thus, it should be adopted by current blockchain networks.
\
**Related Work.** In a classic Byzantine consensus algorithm, a leader is elected and replaced if it is found faulty. The algorithm by Castro and Liskov [@pbft], known as PBFT, requires weak network synchronization for leader change and tolerates up to $f < n/3$ Byzantine peers. Synchronous network provides more information. A peer can determine whether its neighbor is faulty if the message is not received. An efficient synchronous algorithm by Abraham et al. [@sbft], that we call SBFT, tolerates up to $f < n/2$ faults and achieves consensus in 4 expected rounds of message exchanges. Honeybadger [@honeybadger] and Thunderella [@thunderella] are newer approaches to cooperative consensus that try to improve its performance.
There is a number of blockchain sharding algorithms [@rscoin; @byzcoin; @omniledger; @elastico; @rapidchain]. RSCoin [@rscoin] uses central bank authority to regulate data and sharding distribution while using a peer-to-peer network for transaction registration. In Elastico [@elastico], only the agreement algorithm is sharded, the shared blockchain is replicated at every peer. Each committee runs PBFT. The agreed value is sent to the reference committee. This reference committee then broadcasts this value to the whole network. The cross-shard transactions have to lock multiple committees. OmniLedger [@omniledger] shards both data and agreement portions of the network thus improving on the scalability of Elastico. It improves the cross-transaction locking mechanism and eliminates committee assignment security issues present in Elastico. RapidChain [@rapidchain] further improves the sharded design. RapidChain optimizes overall message communication. It improves on the cross-shard transaction locking mechanism of Elastico and OmniLedger by moving the locking of input transactions from the client to the committee. RapidChain also improves on cross committee communication by introducing a committee routing mechanism and optimizes peer churn handling. None of the above sharding algorithms consider adaptive security.
Definitions and Consensus Algorithms
====================================
Definitions
-----------
A set of $n$ *peer processes* (or *peers*) forms a network to maintain the blockchain. The *blockchain* is a sequence of blocks or transactions. We use the terms interchangeably, i.e. we assume that a block contains a single transaction. A *transaction* is a unit of blockchain recording. Each subsequent transaction is cryptographically linked to the previous one. The first transaction in the blockchain is the *genesis* transaction.
Each transaction has a unique identifier. The payload (content) of a transaction is immaterial. Any peer may generate a new transaction. Such peer is *generating*. Peers do not share memory. Peer communication is through messages. One peer may communicate with any other peer. This communication ability is always bi-directional. A peer *broadcasts* a message if it sends it to all other peers. Message delivery is FIFO. There is no message loss. Messages cannot be forged. Specifically, every peer signs its message and all other peers have ways of verifying this signature.
Peers are either *honest* or *Byzantine*. A set of peers that cooperate to approve a transaction despite actions of Byzantine peers is a *consensus committee*. \
**Adversary**. All Byzantine peers behave as if controlled by a single *adversary* aiming to cause maximum amount of damage to the network. The consensus algorithm may discover the activity of a Byzantine peer, detect its identity and exclude it from further operation. However, the adversary may adapt by compromising honest peers. So as not to exceed the maximum number of allowed Byzantine peers $f$, once the adversary claims another Byzantine peer, it needs to allow one of the already corrupted peers to become honest. This is *peer shuffle*. We assume that the peer shuffle happens only when peers are idle. That is, during consensus, a peer is either honest or Byzantine. \
**Sharding.** A *(recording) group* is a set of processes that maintain a single blockchain. There are as many groups as there are separate blockchains. In case of sharding, a peer in the consensus committee that approves a certain transaction in a blockchain does not necessarily belong to the group that records it. However, a peer may belong to only one recording group and only one consensus committee at a time.
Consensus Algorithms
--------------------
**PBFT.** The committee of peers elect the *leader*. The leader is unambiguously determined by the identities of the peers of the committee. A peer that generates a transaction sends it to the leader. The leader runs consensus on every arriving transaction consecutively. Once a transaction reaches the leader, it sends a *pre-prepare* message with this transaction to all the committee peers. After a non-leader peer receives the pre-prepare message, the peer broadcasts a prepare message to the committee. Once a peer receives $2f+1$ prepare messages, it broadcasts the commit message. After the peer receives $2f+1$ commit messages, it locally confirms the transaction. Since the peer has $2f+1$ commit messages, at least one honest voted for the transaction and no other can receive $2f+1$ different commit messages. This is also true for the prepare messages. A non-leader Byzantine peer may delay messages or send incorrect messages. However, if the fraction $f$ of Byzantine peers is small, the honest peers are guaranteed to receive sufficient number of correct massages and then commit. That is, the actions of non-leader Byzantine peer may only delay the consensus. A Byzantine leader may temporarily block the consensus by sending different messages to different peers or not sending messages altogether. In either case, the honest peers discover the Byzantine leader and replace it by forcing a *view change*. PBFT is guaranteed to withstand up to $f < n/3$ Byzantine peers regardless of the message propagation delay. \
**SBFT.** Leader election is similar to PBFT. The algorithm works in four rounds. (i) The generating peer sends its transaction to the leader. (ii) The leader sends the proposal message to the peers. (iii) Once a peer receives the proposal message, it commits the transaction and sends the commit message to all other peers. If a peer receives $f+1$ valid commit messages, it confirms the transaction. A Byzantine leader may be able to prevent some or all peers from committing by either sending different proposal messages to different peers or not sending messages at all. However, the honest peers discover such behavior and elect the new leader. (iv) The peer that confirms the transaction sends a notification message about the confirmation. At the end of the fourth round, if there are peers that have not confirmed the transaction and terminated, the new leader is elected and the algorithm is repeated.
Similar to PBFT, this algorithm relies on at least one honest peer confirming the transaction. However, it assumes that there is a bound on communication delay between honest peers. If a message is not received after a certain delay, it is guaranteed never to arrive. On the other hand, the algorithm has to delay to ascertain this lack of message receipt. In practice this may make SBFT slower. However, it has higher resilience threshold. It can tolerate up to $f < n/2$ Byzantine peers.
1234123451234512345123451234512345123451234512345=$\textbf{constants}$\
$gsize$ // group size\
\
$\textbf{variables}$\
$freeGroups$ // list of all the groups currently idle\
$waitingTrans$ // transaction waiting queue\
$activeComs$ // active committees\
\
$\textbf{commands}$\
new transaction $t$ generated $\longrightarrow$\
add $t(ts)$ to tail of $transWaiting$\
$evaluate()$\
\
consensus committee $c(gl)$ done $\longrightarrow$ // $gl$ is the list of groups in the committee\
remove $c$ from $activeComs$\
add $gl$ to $freeGroups$\
$evaluate()$\
\
$\textbf{function}$\
$evaluate()$\
let $t(ts)$ be at head of $waitingTrans$ // $ts$ is security level\
$\textbf{while}(ts*gsize \leq size(freeGroups)\ \textbf{and}\ \textbf{not}\ empty(waitingTrans))$\
remove $t$ from head of $waitingTrans$\
form committee $c$ from first $ts$ groups in $freeGroups$\
add $c$ to $activeComs$\
remove $ts$ groups from $freeGroups$\
let $t(ts)$ be at head of $waitingTrans$\
run consensus in committee $c$\
\
**PoW.** Adding a transaction to the blockchain requires solving a computationally intensive problem involving the data from the new transaction. The new transaction is then cryptographically linked to last transaction in the blockchain. This task is *mining* the new transaction. Once mined, the integrity of the transaction is easily verified.
The generating peer broadcasts the transaction to the network. Once a peer receives a new transaction, it attempts to mine it. After some peer mines the transaction, it broadcasts the mined transaction. Once a peer receives a mined transaction, it verifies it, attaches it to the blockchain and starts mining a new transaction on top of it. Several peers may mine transactions concurrently. This is a *fork* in the blockchain. A branch of a fork may be extended by addition transactions mined on top of the current block. The shorter branch is discarded.
PoW consensus works correctly provided that the computational power of honest peers exceeds that of Byzantine peers. If peers have the same computational power, PoW consensus tolerates up to $f < n/2$ Byzantine peers.
The Adaptive Security Problem and Solutions
===========================================
**The Adaptive Security Problem.** Consider a sequence of transactions that arrive over time. They need to be recorded into the blockchain. Each transaction has a security level. This security level is fixed and a solution can neither modify nor anticipate it. Consider a *committee consensus algorithm* that, given a fixed number of peers and a transaction, allows peers to agree to this transaction in a fixed number of steps. This agreement is effected despite the actions of Byzantine peers so long as the number of them is less than a certain resiliency threshold $f$. A peer may be in at most one committee at a time. For a transaction at security level $i$, the committee consensus algorithm needs to contain more peers than at level $i-1$. For example, a consensus algorithm may contain $2^i$ peers, i.e. the consensus committee size may grow exponentially with security levels. In this case, the number of security levels is $\log n$.
[*The *Adaptive Security Problem* requires the solution, an adaptive security algorithm, to assign committees to the transactions such that each committee satisfies the transaction security level.*]{}
A trivial solution assigns all peers to every transaction. In other words, every transaction is processed at the highest security level. However, such solution is inefficient as communication resources are wasted with large committee agreeing on a low security level transaction. Another solution forms committees of appropriate size for each transaction but processes them sequentially. This solution uses the resources efficiently. However, it has low throughput since transactions are processed sequentially. Hence, we are led to consider an adaptive security algorithm that selects appropriate size committees and processes transactions with as much parallelism as possible. We present two such algorithms: *Composite Blockguard* and *Dynamic Blockguard*.
\
**Common features of the security algorithms.** We first discuss the features that are common to both Composite Blockguard and Dynamic Blockguard. There are two committee types: reference and consensus. *Reference committee* schedules transactions for verification. We assume its existence and do not discuss its formation and maintenance. Several papers discuss reference committee maintenance [@rapidchain; @elastico].
Once the new transaction is generated, it is *pending*. The reference committee maintains a queue of pending transactions. Transactions are processed in FIFO order. If appropriate consensus committee is available, the pending transaction at the head of the queue is removed and dispatched to the committee for processing. The transaction is then *tentative*. If the committee approves the tentative transaction, it is added to the blockchain and becomes *recorded*. Once a sufficient size committee is available, the transaction is dispatched and the next transaction is considered. If enough peers are available, multiple transactions are processed concurrently.
1234123451234512345123451234512345123451234512345=$\textbf{constants}$\
$winSize$ // window size\
$secMult$ // security level multiplier\
\
$\textbf{variables}$\
$bc$ // blockchain DAG\
$waitingTrans$ // transaction waiting queue\
$freePeers$ // list of peers not in committees\
$activeComs$ // active committees\
\
$\textbf{commands}$\
new transaction $t$ generated $\longrightarrow$\
add $t$ to tail of $waitingTrans$\
\
end of recording stage $\longrightarrow$\
let $t(ts)$ be at head of $waitingTrans$ // $ts$ is security level\
let $peerlist$ be the first $winSize$ peers of totally ordered $bc$\
$\textbf{while} (size(peerlist) > ts * secMult\
\textbf{and}\ \textbf{not}\ empty(waitingTrans))$\
form committee $c$ by randomly selecting\
$ts*secMult$ unassigned peers of $peerlist$\
add $c$ to $activeComs$\
start consensus stage // run consensus algorithms in all committees of $activeComs$\
\
end of consensus stage $\longrightarrow$\
start recording stage // each committee records accepted transaction into $bc$\
\
**Composite Blockguard adaptive security algorithm.** The algorithm is shown in Figure \[figBC\]. In this algorithm, peers are divided into storage groups maintaining independent blockchains. The algorithm maintains a list of idle groups $freegroups$ and stores pending transactions in $waitingTrans$. Once a new transaction arrives or a consensus committee is done, Composite Blockguard finds appropriate number of available groups, forms a consensus committee to process the next transaction in $watingTrans$ and dispatches the transaction. If not enough idle groups are available, the pending transactions wait.
\
**Dynamic Blockguard adaptive security algorithm.** The algorithm is shown in Figure \[figDB\]. This algorithm has a single blockchain and thus a single recording group. A consensus committee is selected out of this group of peers. Multiple consensus committees may operate concurrently if their members do not intersect. This means that the committees have to concurrently write to the same blockchain. To ensure the integrity of the blockchain, the computation proceeds by alternating two stages: consensus stage and recording stage. In the *consensus* stage, committees agree on blocks to be written to the blockchain. Every committee must reach consensus before any committee may proceed to the next stage. In the *recording* stage, each committee broadcasts the transaction to the group maintaining the blockchain. That is, they broadcast it to the whole network. Each written transaction is cryptographically linked to all the written transaction in the previous recording stage. This way, the resultant blockchain is a series-parallel graph. \
*Committee selection window* is the set of unique peers that published in the blockchain most recently. Committee peers are picked at random from the committee selection window. For example, to select $64$-peer committee, a $512$ selection window may be chosen. Then the individual $64$ members are selected from the $512$ members of the window.
\
**Algorithm analysis and comparison.** In both Composite and Dynamic Blockguard, changing the security of each level or adding security levels is relatively easy. In Dynamic, it is just a matter of adjusting Committee Selection Window size or committee sizes. In Composite, the security levels can be changed by modifying the number of groups being merged into a committee
Composite Blockguard is simpler to implement. Since the groups do not overlap, the parallelism is potentially greater. On the other hand, Dynamic Blockguard automatically prefers the most active and, potentially, more reliable peers. However, Composite Blockguard has fewer synchronization issues between parallel committees as they are writing their results in separate ledgers. Composite Blockguard, though, has the added complexity of *cross-shard transactions* where a transaction affects more than a single blockchain. Dynamic Blockguard does not have this complication since it uses just one blockchain.
-- --
-- --
Performance Evaluation
======================
Preliminaries
-------------
**Setup.** We evaluate the performance of Composite and Dynamic Blockguard using abstract simulation. The code for our simulation is available on GitHub [@github]. The behavior of each algorithm is represented as a computation and the performance of such computations is evaluated. Individual computation consists of a sequence of rounds. In every round, each peer may receive a single new message, do local computation and send messages to other peers.
Message propagation may take several rounds. The message propagation delay is uniformly distributed between one round and the maximum. If some peer $p_s$ sends several concurrent messages to the same peer $p_r$, message propagation delay is implemented as follows. Once $p_r$ receives the first message, the next message is delayed between one and the maximum number of rounds. Once this one is received by $p_r$, the next message delay is computed and so on. Message delivery is FIFO. In a single round the recipient may process only a single message from the same sender. However, if multiple messages from different senders are available for delivery by a single peer $p_r$, $p_r$ processes them all in a single round.
Unless specified otherwise, the fraction of Byzantine faults in the network is $n/10$. Byzantine behavior is difficult to simulate in its full complexity. We implement simplified Byzantine peer behavior as follows. A committee is *reliable* if the number of Byzantine peers in it does not exceed its tolerance threshold. The committee is *defeated* otherwise. For example, a committee running PBFT is reliable if the number of Byzantine peers is less than $1/3$ of the total number of its members. The network may not recognize a defeated committee. A defeated committee proceeds operating as normal and writes its transaction to the blockchain. This transaction is counted as defeated. In PoW, if an honest peer mines a transaction in a defeated committee, the transaction is counted as reliable.
Byzantine peer behavior affects reliable committees as well. A committee running PBFT or SBFT may elect a Byzantine peer as a leader. In our simulation, an election of Byzantine leader forces a view change and, in effect, slows down the consensus. A correct transaction is, eventually, recorded in the blockchain. Let us consider the operation of reliable PoW. If an honest peer mines a transaction first, it is recorded in the blockchain. If a Byzantine peer mines a transaction first, the transaction is discarded and mining re-starts. That is, similar to PBFT and SBFT, the presence of Byzantine peers slows down the operation of the consensus algorithm.
In SBFT and PBFT, non-leader Byzantine peers have little influence over the performance of the algorithm, while Byzantine leaders are detected by the consensus algorithm. It is therefore possible that all Byzantine peers may eventually be detected and removed from the network. To counter this, our adversary may shuffle Byzantine peers. That is, a peer may start honest, become Byzantine and then become honest again over the course of our experiment. The adversary may never have more than $f$ Byzantine peers in the network at any given time. Let $s$ rounds be a shuffle period. Every $s$ rounds, a random number of Byzantine peers become honest and an equal number of honest peers become Byzantine. This maintains the ratio of Byzantine to honest peers in the network. Peers assigned to a committee may not be shuffled. In Dynamic Blockguard, Byzantine peers are shuffled after the recording stage. In Composite Blockguard, shuffling happens every round but only affects non-assigned peers.
We use geometric distribution to select the security level of newly generated transaction. The selection probability is $50\%$. That is, for security level $k$, the probability of selection is $0.5^k$. For the highest security level, the probability is $0.5^{k-1}$. To put another way, half of the transactions are at the lowest security level $1$, then $25\%$ of transactions is at the next security level up and so on.
We do not take into account the performance of the reference committee in our experiments. However, we assume that the reference committee carries out the following tasks: it allocates peers to consensus committees, it conducts synchronous stages in Dynamic Blockguard.
-- --
-- --
\
**Experiment parameters and evaluation metrics.** Unless stated otherwise, in the below experiments, the parameters are set as follows. The number of peers in the network is $1024$. The number rounds in a computation is $1000$. For each data point, we carry out $10$ computations and compute the average of the evaluated metric.
A new transaction is generated in every two rounds. This transaction generation rate slightly exceeds the maximum throughput of all consensus algorithms. The new transaction is generated by a randomly selected peer.
We have $5$ security levels. The highest security level is the 5-th level which contains the whole network. That is, the level-5 committee contains $1024$ peers. Each lower level contains half of the peers of the higher level. The lowest security level contains $64$ peers.
In PoW, we use binomial distribution to determine the number of rounds it takes the peers to mine a transaction. The mode, i.e. most frequently occurring value, is $5$ and variance $2.5$.
We vary maximum message delay and the fraction of Byzantine peers in the network. We compute the following metrics. *Throughput* is the number of consensuses per round. We compute it as the number of successful consensuses divided by the length of the computation. Consensuses of defeated committees are not counted. *(Transaction) waiting time* is computed as follows. For coordinated consensus algorithms, i.e. PBFT and SBFT, it is the number of rounds from the moment the transaction is generated till the first peer determines that the transaction is committed. For PoW, it is the time from this transaction. The waiting time for transactions of defeated committees is counted.
-- --
-- --
Results and Analysis
--------------------
**Motivation experiments.** The results of the first series of experiments are shown in Figure \[figFixed\]. The results demonstrate the need for adaptive security. We show that there is a trade-off between the performance and the security of the network. The security level of every transaction is ignored and all transactions are approved by committees of a specific size. We vary this single committee size. The committees proceed with maximum possible concurrency.
Figures \[figFixedThroughputComposite\] and \[figFixedThroughputDynamic\] show throughput for Composite and Dynamic Blockguard respectively. Similarly, Figures \[figFixedWaitingTimeComposite\] and \[figFixedWaitingTimeDynamic\] show waiting time for the two algorithms. The results indicate that as the committee size increases, the throughput declines and transaction waiting time increases. The throughput decline is more pronounced for SBFT since this is a synchronous algorithm. It has to wait for the maximum delay time to ascertain the lack of message receipt. Similarly, wating time is greater for SBFT.
Conversely, Figures \[figFixedDefeatedCommitteesComposite\] and \[figFixedDefeatedCommitteesDynamic\] shows the ratio of defeated committees for particular fraction of Byzantine peers. We show the results for PBFT only and for the lower three security levels. The results for other consensus committee algorithms and security levels are similar. The results indicate that smaller size committees are defeated with greater ease.
\
**Algorithm performance experiments.** The results of the performance evaluation of the adaptive security algorithms are shown in Figures \[figThroughput\], \[figWaitingTime\], and \[figDefeated\]. Let us first discuss the results in Figure \[figThroughput\]. Figures \[figThroughputDelayComposite\] and \[figThroughputDelayDynamic\] demonstrate how throughput depends on the network delay for Composite and Dynamic Blockguard respectively. As network delay increases, the throughput declines. However, different consensus committees react to this increase differently. PBFT has the best performance and lowest decline since the committees just wait for the actual messages to arrive. SBFT exhibits the most sensitivity to the network delay. The reason is that SBFT has to wait for the maximum delay to determine that the message is not coming.
Let us discuss Figures \[figThroughputFractionComposite\] and \[figThroughputFractionDynamic\]. It shows that the performance of Composite and Dynamic Blockguard decreases as the fraction of Byzantine peers in the network increase. This is due to Byzantine peers slowing down the consensus algorithms. PBFT suffers the most since its tolerance threshold is only a third of the peers.
Let us address the results in Figure \[figWaitingTime\]. Figures \[figWaitDelayComposite\] and \[figWaitDelayDynamic\] show the dependency of transaction waiting time on network delay. As expected, the waiting time increases with delay. SBFT is the most vulnerable to this increase since it has to wait for maximum delay time. Figures \[figWaitFractionComposite\] and \[figWaitFractionDynamic\] show how waiting time varies with the fraction of Byzantine peers. Let us explain the trends in the data. As the consensus committee approaches its resiliency threshold, the number of view changes or repeated transaction mining increases which increases the transaction waiting time. If the fraction is away from this threshold, the committees are either reliable or defeated. In either case the waiting time is relatively low. Thus, there is a peak near $n/3$ for PBFT and near $n/5$ for SBFT and PoW. This trend is less pronounced in Dynamic Blockguard since it is masked by synchronization across consensus committees in the same stage.
-- --
-- --
Let us now focus on the results in Figure \[figDefeated\]. The number of defeated committees increases with the fraction of Byzantine peers. It increases fastest for PBFT since it has the lowest tolerance threshold. It increases slowest for PoW since honest miners may still record a reliable transaction in a defeated committee.
\
The results of our experiments indicate that both Composite and Dynamic blackguard algorithm provide adaptive security with a trade-off between performance and security parameters. Composite and Dynamic Blockguard operate adequately regardless of the specific consensus algorithm that they use.
Conclusions and Future Work
===========================
In this paper, we defined the Adaptive Security Problem and showed two efficient solutions for it: Composite and Dynamic Blockguard algorithms. In conclusion, we would like to list further algorithm improvements and possible research directions. Composite and Dynamic Blockguard may be combined to further increase network efficiency. In both algorithms, rather than processing transactions FIFO, the reference committee may re-order transactions to better utlize available peers. Both algorithms have to be able to handle churn of peers. As old peers leave and new peers arrive, the algorithms have to be able to add them to the committees. To further demonstrate their practicality, both algorithms may be implemented and tested in a realistic blockchain system.
Appendix {#appendix .unnumbered}
========
Parameter Experiments {#parameter-experiments .unnumbered}
---------------------
-- --
-- --
Let us discuss the results in Figure \[figTimeline\]. Each metric is taken as a rolling average where only transactions that have been confirmed in the last 200 rounds of computation are considered. For example, throughput is calculated by taking the total number of transactions over the past 200 rounds that have been confirmed over the total number of transactions submitted to the network in that time. In Figures \[figTimelineThroughputComposite\] and \[figTimelineThroughputDynamic\], we see that while the throughput varies it is consistently confirming one transaction in two rounds for PBFT and PoW. In the case of SBFT, this throughput is lower: one transaction per 3 to 4 rounds. This is due to to the need for SBFT to wait for the maximum network delay. We note that same trends in Figure \[figTimelineWaitingComposite\] for Composite Blockguard. In Figure \[figTimelineWaitingDynamic\]. the waiting time gradually increases for Dynamic Blockguard. This indicates that transactions are being submitted faster then the throughput of the network. This is not apparent in Figure \[figTimelineThroughputDynamic\] because throughput is measured as a rolling average. That is, the number of transactions confirmed by the network is constant but the queue of waiting transactions grows. The results show that Dynamic Blockguard can not handle as high of a transaction rate as Composite Blockguard. However, in order to compare the two, the same transaction rate should be used. The stability and consistency of the throughput shows that 1000 rounds is sufficient to take measurements on how the network will behave over a much longer duration.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We perform a study of 62 solar analog stars to compute their effective temperatures (${T_\mathrm{eff}}$) using the Balmer line wing fitting procedure and compare them with ${T_\mathrm{eff}}$ values obtained using other commonly employed methods. We use observed H$\alpha$ spectral lines and a fine grid of theoretical LTE model spectra calculated with the best available atomic data and most recent quantum theory. Our spectroscopic data are of very high quality and have been carefully normalized to recover the proper shape of the H$\alpha$ line profile. We obtain ${T_\mathrm{eff}}$ values with internal errors of about 25K. Comparison of our results with those from other methods shows reasonably good agreement. Then, combining ${T_\mathrm{eff}}$ values obtained from four independent techniques, we are able to determine final ${T_\mathrm{eff}}$ values with errors of about 10K.'
---
Introduction
============
The effective temperature (${T_\mathrm{eff}}$) is one of the most important parameters in the study of stars. For example, precise and accurate ${T_\mathrm{eff}}$ values allow us to reliably measure the chemical compositions of stars. Other important stellar parameters such as luminosity, radius, etc., can only be obtained once ${T_\mathrm{eff}}$ is known. A number of techniques have been devised to derive ${T_\mathrm{eff}}$. In this work, we use the relative flux level in the wings of H$\alpha$ line profile as an indicator of the star’s effective temperature (e.g., Gehren 1981, Barklem et al. 2002).
In studies of stars like the Sun, systematic errors can be minimized if the data are carefully treated with a differential analysis. Thus, very precise ${T_\mathrm{eff}}$ values can in principle be derived using high quality data of solar analog stars. The aim of this work is to derive ${T_\mathrm{eff}}$ values using model fits to the H$\alpha$ line wings of 62 solar analogs and to compare the results with the ${T_\mathrm{eff}}$ values derived using three other methods.
Determination of the effective temperature using H$\alpha$
==========================================================
The method we use consists of finding the best match to an observed H$\alpha$ line profile from a theoretical grid (see Fig. 1). Spectroscopic data acquired with the R. G. Tull coudé spectrograph on the 2.7m Telescope at McDonald Observatory, properly normalized, are employed. The spectral resolution is $R=60,000$ and the average signal-to-noise ratio is 300. Spectral windows free from telluric lines in our solar spectrum (asteroid reflected sunlight) are identified first and later used for the entire sample. The model grid was calculated as in Barklem et al. (2002) and it has a fine spacing of 10K in ${T_\mathrm{eff}}$, 0.05dex in ${\log g}$, and 0.05dex in ${\mathrm{[Fe/H]}}$. The ${T_\mathrm{eff}}$ and its error are derived using least squares minimization. We find ${T_\mathrm{eff}}=5752\pm16$K for the Sun (error bar corresponds to observational noise only). We applied zero point corrections to our solar ${T_\mathrm{eff}}$’s based on solar spectrum adding the difference in temperature that forces the solar ${T_\mathrm{eff}}$ to be equal to 5777K, adding the same temperature difference on the whole sample. Internal errors in our derived ${T_\mathrm{eff}}$ values are about 25K. Note, however, that Barklem et al. (2002) point out that systematic errors can be as large as 80K. Nevertheless, in our differential analysis of solar analog stars, we expect the systematic errors to have a small impact.
![Our observed solar spectrum is superposed on a theoretical grid of H$\alpha$ line profiles.[]{data-label="Fig1"}](vesta.ps){width="9.00cm"}
Comparison with other methods
=============================
We compared our ${T_\mathrm{eff}}$(H$\alpha$) with the ${T_\mathrm{eff}}$’s obtained from the method of the spectral line-depth ratios (Ldr; e.g., Gray & Johansson 1991, Gray 1994), using the calibration formulae by Kovtyukh et al. (2003). We also compared our temperatures with those from the infrared flux method (IRFM; e.g., Ramírez et al. 2005) ${T_\mathrm{eff}}$ scale, using the color calibrations by Casagrande et al. (2010). Finally, we also made a comparison with the values of ${T_\mathrm{eff}}$ obtained from the excitation equilibrium of [Fe<span style="font-variant:small-caps;">i</span>]{} lines, as derived by Ramírez et al. (2009, hereafter R09). Our ${T_\mathrm{eff}}$(H$\alpha$) values are in reasonably good agreement with those from the Ldr, IRFM, and R09 methods, as shown in Fig. 2.
Careful inspection of the residuals of the ${T_\mathrm{eff}}$ value comparisons revealed small offsets and trends with stellar parameters. For example, Ldr–H$\alpha$ showed a clear ${\mathrm{[Fe/H]}}$ dependency while R09–H$\alpha$ revealed an offset of about 40K, as shown in Fig. 3. The former could be due to the fact that Ldr calibration formulae do not take ${\mathrm{[Fe/H]}}$ into account while the latter may be related to the degeneracy between ${T_\mathrm{eff}}$ and ${\log g}$ derived only from an iron line analysis (i.e., forcing excitation/ionization balance). We re-calculated the values of ${T_\mathrm{eff}}$(R09) and ${T_\mathrm{eff}}$(Ldr), thus eliminating the small trends and offsets with linear corrections. In this way, residuals of the ${T_\mathrm{eff}}$ comparisons are dominated by measurement errors (Fig. 4).
![${T_\mathrm{eff}}$(Ldr) – ${T_\mathrm{eff}}$(H$\alpha$) residuals vs. ${\mathrm{[Fe/H]}}$ (upper panel) and ${T_\mathrm{eff}}$(R09) – ${T_\mathrm{eff}}$(H$\alpha$) residuals vs. ${\mathrm{[Fe/H]}}$ (lower panel). Solid lines are linear fits to the residuals.[]{data-label="fig3"}](metall-tren.ps){width="10.00cm"}
![Residuals of ${T_\mathrm{eff}}$ value comparison from different methods after removing small trends and offsets.[]{data-label="fig4"}](without_trend.ps){width="10.00cm"}
Conclusions
===========
Effective temperatures have been determined using the method of Balmer line fitting for a sample of 62 solar analog stars, with internal errors of about 25K. The other methods discussed in this work have internal errors of about 30 - 50K. The high precision of our ${T_\mathrm{eff}}$ values is useful to find small residual trends in the comparison with other methods. We find reasonably good agreement with the ${T_\mathrm{eff}}$’s obtained with the Ldr, IRFM, and R09 methods, but small trends and offsets for the residuals are detected and removed with linear corrections. We argue that high accuracy effective temperatures, with errors of order 10K, are possible to achieve for solar analog stars if several independent measurement are combined, mainly because the impact of errors is very small and can be understood and removed empirically.
Acknowledgements
================
I. R.’s work was performed under contract with the California Institute of Technology (Caltech) funded by NASA through the Sagan Fellowship Program. D. C. thanks the Organizing Commitee of the event for the financial support, J. F. Valle of the direction of Astrophysics of CONIDA - Space Agency of Perú, for his suggestions and the CONIDA for their support of this work. P. S. B is a Royal Swedish Academy of Sciences Research Fellow supported by a grant from the Knut and Alice Wallenberg Foundation.\
2002, A&A, 385, 951
2010, A&A, 512, 54
1981, A&A, 100, 97
1991, PASP, 103, 439
1994, PASP, 106, 1248
2003, A&A, 559, 564
2005, ApJ, 626, 465
2009, A&A, 508, L17
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Rafit Izhak-Ratzin \*&& Hyunggon Park \*&& Mihaela van der Schaar\
PaloAlto && Electrical Engineering Institute && Electrical Engineering Department\
Networks && Ewha Womans University && University of California\
Sunnyvale, CA && Seoul, Korea && Los Angeles, CA
bibliography:
- 'IEEEabrv.bib'
- 'RL\_P2P.bib'
title: Reinforcement Learning in BitTorrent Systems
---
[^1]
[^1]: \* This work was done while Dr. Izhak-Ratzin and Dr.Park were at UCLA
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Micron-sized objects confined in thin liquid films interact through forces mediated by the deformed liquid-air interface. This capillary interactions provide a powerful driving mechanism for the self-assembly of ordered structures such as photonic materials or protein crystals. Direct probing of capillary interactions requires a controlled force field to independently manipulate small objects while avoiding any physical contact with the interface. We demonstrate how optical micro-manipulation allows the direct measurement of capillary interactions between two micron sized spheres in a free standing liquid film. The force falls off as an inverse power law in particles separation. We derive and validate an explicit expression for this exponent whose magnitude is mainly governed by particles size. For micron-sized objects we found an exponent close to, but smaller than one, making capillary interactions a unique example of strong and very long ranged forces in the mesoscopic world.'
author:
- 'R. Di Leonardo$^1$, F. Saglimbeni$^2$ & G. Ruocco$^2$'
title: The very long range nature of capillary interactions in liquid films
---
It is a well known fact that small objects floating on a liquid surface cluster together. Bubbles on the surface of a soap solution[@bragg] or cereals in a bowl of milk[@cheerio] attract each other with long range forces arising from the interface deformation under particles weight (or buoyancy for bubbles). Close packed configurations for such macroscopic objects are found to minimize gravitational potential energy. Shrinking lengths to the mesoscopic scale, particles weight soon become too weak to produce any significant deformation and hence attraction. However if the particles are confined in thin liquid films a deformation of the interface is unavoidable. This is the case, for example, of a colloidal suspension drying on a solid substrate, or dispersed in a free standing thin film. When the thickness of the liquid film becomes smaller than the bead diameter, the interface has to deform with an increase in surface energy. The liquid interface will then react on the particles with forces aiming to reach a minimum surface (energy) configuration, that usually corresponds to close packed two dimensional crystals[@2d]. Such phenomena, already observed by Perrin in 1909[@perrin], have attracted considerable interest in recent times due to their relevance for the engineering of photonic materials[@vlasov] and protein crystallography[@yoshimura]. Consequently, a strong effort has been devoted to the theoretical analysis of the involved forces, resulting in a long series of papers reviewed in[@review]. Prediction for macroscopic objects have been confirmed by experiments on immersed cylinders [@velev] or particles attached to holders[@dushkin].
However no experiment so far has been able to directly measure the strong capillary force acting between an isolated pair of mesoscopic objects, despite the fact that it is in the mesoscopic and nanoscopic realm that this effect finds the most interesting applications. Any physical contact with the particles would inevitably produce a significant deformation of the liquid-air interface and dramatically affect the interaction. On the other hand, due to long ranged hydrodynamic interactions in 2D, particles mobility are very sensitive to interparticle distances and force measurements are difficult to deduce from particles trajectories. A static, highly non-invasive method is required for a direct and reliable measurement of these interactions.
In this Letter we demonstrate how optical micro-manipulation [@ashkin] allows the precise measurement of capillary interactions between two micron sized spheres confined in a free standing thin liquid film. Holographic optical tweezers [@hot] allow to isolate a single pair of particles and scan their relative distance from close contact to several tens of diameters. Capillary forces will tend to push the particles towards each other and out of the optical traps until the restoring trap forces will balance the attractive force. The intensity of capillary force can be then deduced by measuring particles displacements from trap centers, after trap calibration. Working in a free standing liquid film is essential for accurate capillary force determinations since no particle-substrate interactions have to be taken into account.
{width=".45\textwidth"} \[fig1\]
The force law can be anticipated by calculating the surface tension forces acting on a pair of spherical objects confined in a thin liquid film. Kralchewsky et al.[@kral] derived the shape of the meniscus around the two particles using the method of matched asymptotic expansions. An implicit expression for the force law was obtained, whose evaluation required the numerical solution of a system of nonlinear equations in the accessible system parameters. Confining here to mesoscopic objects, we show that Kralchewsky solution for the force can be very well approximated by an inverse power law whose parameters have an explicit expression in terms of the system’s physical properties. We assume complete wetting (zero contact angle) with the liquid surface departing from a contact ring of radius $r_c$ with continuous slope angle $\psi_c$ (Fig. 1a). Surface tension will exert a force on the contact ring whose resultant is orthogonal to the ring plane and has a modulus:
$$\label{F} F=\gamma\; 2\pi r_c\ \sin\psi_c$$
where $\gamma$ is the liquid-air surface tension ($\gamma\simeq35$ mN/m in our experiment). For small gradients $r_c=a \sin\psi_c\sim a \psi_c$, $z_c=a(1-\cos\psi_c)\simeq a\psi_c^2/2$, and the force $F\simeq 4\pi\gamma z_c$ is then proportional to $z_c$ with a strength of order 1 nN/nm. The order of magnitude of other forces into play is 1 pN for particle’s weight and 100 pN for the maximum optical force exerted by our trap. The interface is supposed to be flat and horizontal far from the particle and we indicate with $\zeta(r)$ the local vertical displacement of the interface from this reference surface. In the small gradient approximation, Laplace equation for the pressure drop across free standing portion of the upper surface ($r>r_c$) reads[@landau]:
$$\label{laplace}
\mathbf \gamma\nabla^2\zeta(r)=\rho g \zeta(r)$$
where $\rho$ is the liquid density and $g$ the acceleration due to gravity. The only axisymmetrical solution to (\[laplace\]) vanishing at infinity takes the form[@nico]:
$$\zeta(r)=-\frac{\tan\psi_c}{q}\frac{K_0(q r)}{K_1(q r_c)}$$
where $K_i(x)$ is the modified Bessel function of ith order[@abramowitz] and $q^{-1}=\sqrt{\gamma/\rho g}$ is the capillary length. This is the length scale below which gravitational forces play no role in determining the interface shape and it’s about $2$ mm in typical solvents. Therefore we can safely replace $K_0(x)$ and $K_1(x)$ with their small argument expansion and write for small gradients:
$$\label{profiloth}
\zeta(r)=-2 z_c \left[\log(q r/2)+\gamma_e \right]$$
where $\gamma_e$ is the Euler-Mascheroni constant. At the micron scale gravitational energy will be negligible and the bottom surface will have a symmetrical shape to the top one. Therefore two equally strong capillary forces will act on the top and bottom contact rings of an isolated particles but they will cancel out giving no net force. When a second particle is introduced in the film at a distance $s$, it will in turn contibute to the interfaces deformation producing a tilt of the contact lines around the first particle (Fig. 1b). For small deformations we can still think of contact lines as circles but this time slightly tilted. Within the superposition approximation by Nicolson[@nico], the amount of tilt would simply be given by the gradient of the surface deformation produced by an isolated particle at the location of the second one. According to (\[profiloth\]) the interface height around an isolated particle is expected to decay logarithmically with the distance producing a tilt of the contact lines that falls off as the inverse interparticle distance. Capillary forces $F$ acting on the contact rings will not be balanced anymore but a net attractive force would appear whose intensity is given by the projections of $F$ on the film reference plane:
$$\label{force}
f(s)=2 F(s) \left.\frac{\partial \zeta(r)}{\partial r}\right|_{r=s}=
16\pi\gamma z_c^2(s) \frac{1}{s}$$
where $s$ is the interparticle distance. The resulting force is than proportional to the square of $z_c$ and inversely proportional to the distance separating the particles. In analogy with Coulomb electrostatic force in 2D, the quantity $z_c$ is usually called the “capillary charge” of the particle[@charge]. However, the capillary charge is not an intrinsic constant property but a slowly varying function of particle separation. This function can be calculated by imposing the continuity and differentiability of the interface across the contact ring.
$$\begin{aligned}
\label{boundary1}
&&\zeta(r_c)+\zeta(s)+h=a-z_c\\
\label{boundary2}
&&r_c=a \sin\psi_c\sim\sqrt{2 a z_c} \end{aligned}$$
where we rely on the superposition approximation [@nico] to express the interface vertical displacement field as the sum of two single particle displacement fields (\[profiloth\]). We are also assuming in (\[boundary1\]) that the deformation field $\zeta(s)$ produced by one particle is constant over the contact line of the other. Equations \[boundary1\],\[boundary2\] can be solved analytically giving:
$$\label{zc}
z_c(s)=\frac{a-h}{-W\left[-q^4 s^2 a(a-h) \exp[4\gamma_e-1]/8\right]}$$
Where $W$ is the Lambert-W function [@lambert]. The above expression for $z_c$ is a slowly varying function of $s$ in mesoscopic systems ($a<s\ll
q^{-1}$) and it can be well approximated by its logarithmic expansion about $s=2a$:
$$\begin{aligned}
&&z_c(s)\sim z_{0} (s/2a)^{\alpha}\\[.2cm]
&&z_0=z_c(2 a)\\[.1cm]
&&\alpha=\left.\frac{d\log z_c}{d\log s}\right|_{s=2 a}=\frac{2 z_{0}}{a-h-z_{0}}\end{aligned}$$
Accounting for changes in $z_c$ the attractive force (\[force\]) will still display a power law behavior but with an exponent smaller than one:
$$\label{force0}
f(s)=\frac{16\pi\gamma z_{0}^2}{2a} \left(\frac{2a}{s}\right)^{1-2\alpha}$$
where $\alpha$ is an explicit function of the three lengths: particle size, film thickness, capillary length of the solvent. The exponent $1-2\alpha$ is equal to $1$ for $h\rightarrow a$ but then quickly drops to a fairly constant value as soon as $h$ is small enough to produce significative forces ($\sim$ 1 pN $\sim 10^3 K_B T/a$). This constant value depends practically only on the particle size and varies very little when using different typical solvent properties. Changing particle size from 10 nm to 10 $\mu$m produces a corresponding exponent variation in the range 0.92 to 0.82. For our particle size we predict an exponent of 0.86.
{width=".35\textwidth"} \[fig2\]
{width=".45\textwidth"} \[fig3\]
A schematic view of the experimental setup is reported in Fig. 2a.
Latex beads (5.7 $\mu m$ diameter) are dispersed in a 2/3 water - 1/3 glycerol mixture with added 0.2% wt surfactant (SDS). A thin film is obtained by sweeping the solution on a square frame (10 mm side) of nylon wires (60 $\mu m$ thickness). The free standing liquid film is enclosed in a humidity chamber and placed over the 40x objective of an inverted optical microscope (Nikon TE2000-U). The same objective is used to focus the laser beam ($\lambda$=532 nm) reflected off a spatial light modulator (Holoeye LCR-2500) into two, dynamically reconfigurable, optical traps [@hot]. Axial confinement is guaranteed by the normal components of capillary forces. We can also access the thickness variations around an isolated particle by viewing the film under reflected monochromatic light. To this end a red diode laser beam, overlapped to the trapping green beam, is focused by the same microscope objective far from the film surface. The observed portion of the red beam wavefront is approximately plane and the film reflectance than varies with $\cos[4 \pi
\zeta(r)n/\lambda]$ giving interference rings. $n=1.37$ is the refractive index of the liquid mixture and $\lambda=657$ nm the red laser wavelength. The distorted film will then show ring shaped interference fringes centered on the trapped particle (Fig. 3b). The thickness profile around an isolated trapped particle can be extracted (within an additive constant) and is reported here in Fig. 3a as open circles. A clean logarithmic shape is found up to 200 $\mu$m, in perfect agreement to prediction in (\[profiloth\]), as shown by the logarithmic fit curve (red line).
To extract the force law one of the two traps is held fixed while the other is continously scanned through different distances with a step of 2 $\mu$m. An image of the two particle is digitally recorded (Fig. 2b) for every scan step and subsequently used to extract particle positions. Due to liquid drainage, the film is slowly but constantly thinning. When the thickness is so small that optical forces cannot balance the capillary interaction, one of the two particles jumps out of the trap and collapses onto the other. Until that time we can extract the attractive capillary force from the interparticle distance. Calling $k_1$ and $k_2$ the two trap elastic strengths, each particle will be displaced towards the other by a distance $\Delta x_i=f/k_i$. The observed interparticle separation will be then smaller than trap separation by an amount: $\Delta s=(\Delta x_1+\Delta x_2)=f(1/k_1+1/k_2)=f/k'$. Particle distance will actually fluctuate due to Brownian motion with a mean squared value given by [@meiners]: $\langle\Delta d^2\rangle=K_B T/k'$ which can be used to experimentally determine $k'=29$ pN/$\mu$m.
![ (Color online) Intensity of the capillary interaction between two beads confined in a thin film, as a function of their relative distance. Open circles are experimental determinations. The black solid line represents a fit to the predicted power law (\[force0\]) having the film thickness $h$ as the only free parameter. Red dashed line is the full Kralchewsky theory[@kral] for the same $h$ parameter. The gray region marks the unaccessible range of separations $s$ smaller than close contact distance $2a$. Double log plot insert more immediately shows that the data are well represented by a power law. A 1/r law is also reported there, for reference, as a black dashed line.](fig4.eps "fig:"){width=".45\textwidth"} \[fig4\]
As already discussed, the logarithmic film shape in Fig. 3 would lead, in the Nicolson approximation, to an attractive force decaying roughly as the inverse of interparticle distance. More precisely we would expect to find the power law in (\[force0\]). Indeed a clean power law is found experimentally as evidenced by the double log plot reported in Fig. 4b. The overall behavior is very well fitted by the power law in (\[force0\]) leaving the film thickness at large distance $h$ as the only free parameter. We find $h=2.2\mu\textrm{m}=0.8 a$ that makes an average gradient $\psi_c=0.12$ confirming that the small gradient approximation is justified for our system. For this fitted $h$ value we get a power law exponent $1-2\alpha=0.86$. The full Kralchewsky prediction for the same parameter values, also reported in Fig. 4b, is almost completely overlapped to our power law expansion, confirming the goodness of our approximations.
We have shown how optical micromanipulation provides an unprecedented tool for investigating capillary forces which govern aggregation and self-assembly of colloids in liquid films. We provide a static force measurement of the capillary attraction between an isolated colloidal pair and perform a direct test of the theoretically predicted power law. The exponent of the power law is found to be close to, but smaller than one, making capillary forces a quite unique example of very long ranged interactions in the mesoscopic world. The experiment opens the way to a variety of further developments addressing the role of many body-effects, membrane elasticity, wetting properties, surfactant dynamics, hydrodynamic interactions in 2D. A deep insight into the nature of interface-mediated forces at the mesoscopic scale could suggest new routes to self-assembly of meso- and nano-structures[@whitesides]. Optical trapping of colloidal particles bound to lipid membranes [@koltover] could also provide new insights in the dynamics of biomembrane inclusions[@reynwar].
[1]{}
W.L. Bragg, J. Sci. Instrum. [**19**]{}, 148 (1942).
D. Vella, L. Mahadevan, [*Am. J. Phys.*]{} [**73**]{}, 817 (2005) .
N.D. Denkov, O.D. Velev, P.A. Kralchevsky, I.B. Ivanov, H. Yoshimura, K. & Nagayama, Nature [**361**]{}, 26 (1993).
J. Perrin, Ann. Chim. Phys. [**18**]{}, 1 (1909).
Y. Vlasov, X.Z. Bo, J.C. Sturm, D.J. Norris, Nature [**414**]{}, 289 (2001).
H. Yoshimura, M. Matsumoto, S. Endo, K. Nagayama, Ultramicroscopy [**32**]{}, 265 (1990).
P.A. Kralchevsky, K. Nagayama, Adv. Colloid Interface Sci. [**85**]{}, 145 (2000).
O.D. Velev, N.D. Denkov, V.N. Paunov, P.A. Kralchevsky, K. Nagayama, Langmuir [**9**]{}, 3702 (1993).
C.D. Dushkin, P.A. Kralchevsky, H. Yoshimura, K. Nagayama, Phys. Rev. Lett. [**75**]{}, 3454 (1995).
A. Ashkin, J.M. Dziedzic, J.E. Bjorkholm, S. Chu, Opt. Lett. [**11**]{}, 288 (1986).
J.E. Curtis, B.A. Koss, D.G. Grier, Opt. Commun. [**207**]{}, 169 (2002).
P.A. Kralchewsky, V.N. Paunov, I.B. Ivanov, K. Nagayama, J. Colloid Interface Sci. [**151**]{}, 79 (1992).
L.D. Landau, E.M. Lifshitz, Fluid Mechanics, Pergamon/Addison Wesley, London, (1959).
N.N. Nicolson, Proc. Cambridge Philos. Soc. [**45**]{}, 288 (1949).
M. Abramowitz, I.A. Stegun, Handbook of Mathematical functions, Dover, New York (1965).
V.N. Paunov, P.A. Kralchewsky, N.D. Denkov, K. Nagayama, J. Colloid Interface Sci. [**157**]{}, 100 (1993).
R.M. Corless, G.H. Gonnet, D.E.G. Hare, D.J. Jeffrey, D.E. Knuth, Adv. Comput. Math. [**5**]{}, 329 (1996).
J.C. Meiners, S.R. Quake, Phys. Rev. Lett. [**82**]{}, 2211 (1999).
G.M. Whitesides, B. Grzybowski, Science [**295**]{}, 2418 (2002).
I. Koltover, J.O. Rädler, C.R. Safinya, Phys. Rev. Lett. [**82**]{}, 1991 (1999).
B.J. Reynwar, G. Illya, V.A. Harmandaris, M.M. Müller, K. Kremer, M. Deserno, Nature [**447**]{}, 461 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study order parameters for the confinement-deconfinement phase transition related to the Polyakov-loop variable. The functional renormalisation group is used to compute these order parameters in a unified, non-perturbative continuum approach. Our result for the expectation value of the traced Polyakov loop agrees quantitatively with the lattice result. Furthermore, we discuss how this order parameter differs from the standard continuum Polyakov loop. For temperatures close to the phase transition temperature there are significant deviations. We argue that these deviations are of crucial importance for QCD effective models, which usually implicitly rely on a Gaußian approximation neglecting this difference.'
author:
- Tina Katharina Herbst
- Jan Luecker
- 'Jan M. Pawlowski'
bibliography:
- '../refs.bib'
title: Confinement order parameters and fluctuations
---
Introduction {#sec:intro}
============
The physics of low energy QCD is mainly governed by two phenomena: spontaneous chiral symmetry breaking and confinement. Order parameters for the former are local observables that are not invariant under chiral transformations, most prominently the chiral condensate. In turn, a smoking gun for confinement is the linear rise of the static quark-potential at infinite distance in pure $SU(N)$ Yang-Mills theory. The underlying symmetry is the $Z_N$-center symmetry, which is only accessible via non-local observables such as the traced Polyakov loop. In QCD at the physical point, none of these symmetries is fully realised and the respective phase transitions turn into crossovers: chiral symmetry is explicitly violated by the current quark masses and the chiral anomaly, and center symmetry is violated explicitly by dynamical quarks in the fundamental representation. Nonetheless, the respective order parameters are interesting observables in full QCD as they give access to the amount of explicit symmetry breaking, and hence the proximity of QCD at finite temperature and density to a potential critical end-point in the phase diagram.
The confinement-deconfinement phase transition, or more precisely the transition from the quark into the hadronic phase, is also accessible with baryonic observables such as baryonic fluctuations. The latter are directly accessible within experiments and are hence especially interesting. In this context, observables based on the Polyakov-loop variable are of specific interest as the related gluonic backgrounds play an important r$\hat{\rm o}$le in the computation of the baryonic fluctuations, see [@Fu:2015naa].
More generally, in functional approaches to QCD the temporal gluonic background $\langle A_0\rangle$ relates to the Polyakov loop. It provides a natural expansion point for systematic expansions of first-principle QCD, as well as for QCD-enhanced low-energy effective models, see e.g. [@Braun:2009gm; @Pawlowski:2010ht; @Haas:2013qwp; @Herbst:2013ufa; @Mitter:2014wpa; @Braun:2014ata] and the recent survey [@Pawlowski:2014aha]. This approach has been established in [@Braun:2007bx; @Marhauser:2008fz; @Braun:2009gm; @Braun:2010cy; @Fister:2013bh]. There it has been worked out how to define and compute the gauge invariant non-perturbative glue potential $V[A_0]$ in Yang-Mills theory and QCD within the functional renormalisation group, as well as general functional methods. It has also been shown that the expectation value $\langle A_0\rangle$ of the gluonic background, defined as the minimum of the glue potential $V[A_0]$, serves as a gauge-invariant order parameter for the confinement-deconfinement phase transition. It is indeed directly linked to the gauge-invariant eigenvalues of the untraced Polyakov loop. Moreover, the setting enabled us to link the confinement-deconfinement phase transition algebraically to a mass gap in the gluon propagator, [@Braun:2007bx; @Fister:2013bh]. In the latter work [@Fister:2013bh], the framework has been extended to general functional approaches including the Dyson-Schwinger equations and the 2PI formalism, for applications to QCD see [@Fischer:2013eca; @Fischer:2014vxa; @Fischer:2014ata]. In [@Reinhardt:2012qe; @Reinhardt:2013iia; @Heffner:2015zna] the approach has been extended to the Coulomb gauge in the Hamiltonian formulation. It has also been picked up in more phenomenological applications to QCD, see e.g. [@Kondo:2010ts; @Fukushima:2012qa; @Reinosa:2014ooa; @Reinosa:2014zta; @Kondo:2015noa].
In the present work we study the relation between the expectation value of the traced Polyakov loop and the expectation value $\langle
A_0\rangle$, the temporal gluonic background of the theory, including the gauge invariant definition of the latter. In Secs. \[sec:OrderParams\] and \[sec:OrderParamsgauge\] of this work we show how these order parameters for the confinement-deconfinement phase transition are derived from the Polyakov loop, and discuss their properties and the underlying symmetries. Subsequently we set up the functional renormalisation group (FRG) as a framework to study these objects in a unified approach, see . Notably, we present the first non-perturbative continuum calculation of the expectation value of the traced Polyakov loop, which is a standard observable on the lattice. We discuss the renormalisation of this object in and demonstrate the quantitative agreement with the lattice results in .
Order parameters and the Polyakov loop {#sec:OrderParams}
======================================
In the present work we compute and compare different order parameters for the confinement-deconfinement phase transition that can be derived from the Polyakov loop. The properties of the Polyakov loop itself and related constructions of order parameters have been discussed at length in the literature. This includes the discussion of simple representations in physical gauges such as the Polyakov gauge, maximal Abelian gauges and axial gauges The latter gauges, and in particular the Polyakov gauge, also facilitate the access to topological excitations or defects and their r$\hat{\rm o}$le for the confinement-deconfinement phase transition, see e.g.[@Ford:1998bt; @vanBaal:2000zc] and references therein. Here we briefly review some important properties.
The Polyakov loop, the Wilson loop in the temporal direction, is defined as $$\label{eq:Poloop}
P(\vec x)= {\cal P} \exp\left\{ i g \int_0^{\beta} dx_0\, A_0(x_0,\vec x)
\right\} \,,\quad{\rm with}\quad \beta=\0{1}{T}\,,$$ where $\cal P$ denotes path ordering, and $g$ the gauge coupling. Due to the periodicity in the temporal direction, $x_0 \to x_0+\beta$, gauge fields have to be periodic up to gauge transformations, the temporal transition functions $t_0(x)$. The latter can be chosen periodic, $t_0(x_0+\beta,\vec x)= t_0(x_0,\vec x)$. Under general gauge transformations $U(x_0,\vec x)$, the Polyakov loop transforms as $$P(\vec x)\to U^{-1}(0,\vec x)P(\vec x) U(\beta,\vec x)\,.$$ Notably, the traced Polyakov loop, ${\ensuremath{\operatorname{Tr}}}\, P(\vec x)$, is only invariant under periodic gauge transformations with $U(x_0+\beta,\vec
x)= U(x_0,\vec x)$, that also guarantee the periodicity of the transition function $t_0$. Under these periodic gauge transformations transforms as a tensor $$\label{eq:Utrafo}
P(\vec x)\to U^{-1}(\vec x)P(\vec x) U(\vec x)\,,\ \ \text{with}\ \
U(\vec x)= U(0,\vec x)\,,$$ and the traced Polyakov loop is invariant. In turn, under non-periodic gauge transformations the periodicity of the temporal transition functions changes and $P(\vec x)$ has to be augmented by transition functions in order to provide a tensor under gauge transformations. This generalisation is necessary e.g. if a temporal axial gauge is considered with $A_0=0$. Clearly $ P(\vec
x)=\id$ in such a gauge and carries no physics information. Then, the information about the phase transition is comprised entirely in the transition functions, see e.g. [@Ford:1998bt] for more details.
Here, however, we resort to the periodic setting. Then one well-known standard order parameter is given by the expectation value of the trace of the Polyakov loop. The normalised trace reads $$\label{eq:tracedPL}
L[A_0] = \frac{1}{N} {\ensuremath{\operatorname{Tr}}}_f P(\vec x)\,,$$ which is gauge invariant under periodic gauge transformations. The expectation value of , $\lLA$, relates to the free energy, $F_{q \bar q}$, of a static quark–anti-quark pair at infinite distance, $$\label{eq:freen}
\lLA \sim e^{-\012 \beta F_{q\bar q}}\,.$$ Since the free energy of a quark–anti-quark pair is finite in the deconfined phase and diverges in the confined phase, one can deduce that $\lLA$ indeed is an order parameter for the confinement-deconfinement transition. More precisely, it is an order parameter for center symmetry breaking: transformations with $z\in Z_N$ transform the Polyakov loop into $$\label{eq:ztrafo}
P(\vec x)\to z P(\vec x)\quad \Rightarrow \quad L[A_0]\to z L[A_0] \,,$$ where $z$ in is in the fundamental representation. Hence, in the center-symmetric phase the expectation value of $L$ has to vanish. In turn, the ground state in the perturbative high-temperature phase is given by vanishing gauge fields which break center symmetry maximally. Then we have $\lLA>0$, if we single out the group direction that gives real and positive values for the Polyakov loop. This can either be done by a small explicit symmetry breaking in the effective Polyakov loop potential or by appropriate temporal boundary conditions, both are common procedures for computing the order parameters of spontaneous symmetry breaking. We are finally led to $$\label{eq:exptrL}
\lLA \ \left\{\begin{array}{rcl} = 0 & \quad {\rm for}\quad & T<T_c\\[2ex]
> 0 & \quad {\rm for}\quad & T>T_c\end{array}\right. \,.$$ As the underlying symmetry is the discrete $Z_N$-symmetry, we expect a second-order phase transition for $SU(2)$ and a first-order transition for $SU(N>2)$. This has indeed been verified on the lattice as well as in the continuum. Moreover, the temperature-dependence of the order parameter shows rapid convergences towards the large-$N$ limit, for results in the present continuum QCD setting and on the lattice see e.g. [@Braun:2010cy; @Mykkanen:2012ri]. Despite this successful classification of the confinement-deconfinement phase transition in semi-simple Lie groups there are strong indications that it is the number of degrees of freedom that triggers the order of the confinement-deconfinement phase transition, rather than the center. For example, for exceptional Lie groups one sees first-order phase transitions, e.g. [@Braun:2010cy], while for $SO(3)$ a second order phase transition is observed, e.g. [@deForcrand:2002vs].
{height=".26\textheight"} {height=".26\textheight"}
The expectation value of the traced Polyakov loop is a simple order parameter for the confinement-deconfinement phase transition which is easily accessible on the lattice. In terms of the gauge field, however, it requires the computation of infinite-order correlation functions of the temporal gauge field. Hence, within continuum computations based on a description in gauge fields this asks for the definition of an order parameter which is more easily accessible. To this end we note that the Polyakov loop can also be written as the exponential of an algebra-valued field $\varphi$, $$\label{eq:varphiP}
P(\vec x)= e^{ 2 \pi i\, \varphi(\vec x)}\,,\quad {\rm with}\quad
\varphi(\vec x)\to U^{-1}(\vec x)\, \varphi(\vec x)\, U(\vec x)\,.$$ The transformation property of $\varphi$ in follows straightforwardly from that of the Polyakov loop . Hence, $\varphi$ also transforms as a tensor under gauge transformations. This entails that the eigenvalues $\nu_n$ of $\varphi$ are gauge invariant, as gauge transformations are unitary rotations that do not change the eigenvalues of a matrix. We utilise this freedom and rotate $\varphi$ into the Cartan subalgebra. We shall briefly discuss below that this amounts to taking the Polyakov gauge. Then, the gauge-invariant eigenvalues are given by the relation $$\label{eq:nun}
\varphi(\vec x) |\phi_n\rangle = \nu_n(\vec x) |\phi_n \rangle \,,$$ with eigenvectors $|\phi_n\rangle$ that span the Cartan subalgebra. This allows us to define a constant background $$\label{eq:barvarphi}
\bar\varphi= \sum_n \langle \nu_n \rangle |\phi_n
\rangle\langle \phi_n |\,,$$ which carries the gauge invariant information about the eigenvalues of the Polyakov loop, $\exp \{2 \pi i\, \nu_n\}$, and hence about the confinement-deconfinement phase transition. Note that for the expectation values $\langle \nu_n\rangle$ a preferred group direction has to be singled out, similarly to the computation of the expectation value of the traced Polyakov loop.
Indeed, $\bar\varphi$ itself is an order parameter, which can be deduced from the transformation properties of $\varphi$ under center transformations. Before this is detailed, let us remark that in the Polyakov gauge there is a simple relation between the algebra-valued field $\varphi$ and the gauge field. The Polyakov gauge reads $$\label{eq:polgauge}
A_0(x) = A_0^c(\vec x)\,,$$ where $A_0$ depends on the spatial coordinates $\vec x$ only and is rotated into the Cartan subalgebra indicated by the superscript $c$ in . Then we have $$\label{eq:varphiPolgauge}
2\pi\varphi(\vec x)= g\beta A_0^c(\vec x)\,.$$ The relation entails that the eigenvalues of $
g\beta A_0^c(\vec x)$ in Polyakov gauge are the gauge-invariant eigenvalues of $\varphi$. Hence, despite working in a gauge-fixed setting, we can directly extract gauge-invariant information from the expectation value of the gauge field. With the traced Polyakov loop then takes the simple form $$\label{eq:LA0}
L[A_0]=\01N {\ensuremath{\operatorname{Tr}}}_f e^{ig\beta A_0^c} =\01N {\ensuremath{\operatorname{Tr}}}_f e^{2\pi i \varphi} =:L(\varphi)\,.$$ Note also that holds for any constant gauge field as constant fields can be rotated in the Cartan subalgebra. For these fields in $SU(2)$ we have $\beta g A_0= 2 \pi \varphi_3\, \sigma^3/2$ with the generator $\sigma^3/2$ of the Cartan subalgebra. For $SU(3)$ we write $$\label{eq:varphi}
\beta g A_0 = 2\pi\left(\varphi_3t^3 + \varphi_8t^8\right) =: 2\pi
\varphi\,,$$ with $t^{3},\,t^8$ the generators of the Cartan subalgebra of $SU(3)$. With the traced Polyakov loop in $SU(3)$ Yang-Mills theory takes the form $$\label{eq:Lvarphi}
L(\varphi_3,\varphi_8) = \013 \left( e^{- \0{2 \pi i \varphi_8}{\sqrt{3}} } +2\cos(
\pi\varphi_3)\, e^{\0{\pi i \varphi_8}{\sqrt{3}} } \right)\,.$$ The above entails that the corresponding effective potential $V(\varphi)$ simply is that of constant temporal gauge fields $V[A_0]$. The expectation value $\bar\varphi$ can be determined from the minimum of the effective potential $V(\varphi)$. As for $\lLA$, we single out the minimum with $L(\bar\varphi)\in \mathbb{R}$ and positive, leading to $\varphi_8=0$, see . Contour plots of the effective potential both in the deconfined and in the confined phase are shown in , and details about its continuum definition are provided in the next section. In the confined phase we have $\bar\varphi_3=2/3$ and $L[\bar\varphi]=0$ while in the deconfined phase we have $0\leq\bar\varphi_3<2/3$, $L[\bar\varphi]\neq
0$, with $\bar\varphi_8=0$ in both cases. This already indicates that $\bar\varphi$ as well as $L[\bar\varphi]$ are order parameters for center symmetry breaking.
Now we turn to the properties of $\varphi$ and the expectation value $\bar\varphi$ under symmetry transformations, and in particular center transformations. On the level of the algebra field $\varphi$, center symmetry is realised as a shift $$\label{eq:thetaz}
\varphi\to \varphi+ \theta_z\,,\quad {\rm with} \quad
z=e^{2 \pi i \theta_z}\,.$$ The center transformation is a symmetry transformation of the effective potential $V(\varphi)$, as can be seen from the contour plots in . For $SU(2)$ the center elements are given by $z=\id, -\id$. The corresponding algebra elements of $z$ are $\theta_z = 0, \sigma^3/2$ respectively, with the Pauli matrix $\sigma_3$. For $SU(3)$ the center elements $z$ and the corresponding algebra elements $\theta_z$ are given by $$\label{eq:thetazSU3}
z=\id\,,\ \id e^{\0{2 \pi}{3} i}\,, \ \id e^{\0{4 \pi}{3} i }\,,\qquad \theta_z =
0\,, \ \0{2}{\sqrt{3}}\, t^8 \,,\ t^3 - \0{1}{\sqrt{3}}\,t^8\,,$$ respectively with the Cartan generators $t^3,t^8$ of $SU(3)$ being half of the Gell-Mann matrices $\lambda^3,\lambda^8$.
The transformation can be combined with a Weyl reflection, i.e. a reflection about the edges of the Weyl chamber. Weyl reflections are an isometry of the roots of the gauge group and are generated by specific constant gauge transformations. Hence they are a symmetry of the theory. In $SU(2)$ they are simply given by $\varphi\to -\varphi$, while in $SU(3)$ the Weyl reflections can be read off from . Clearly this is a symmetry of the potential. Restricting ourselves to the fundamental chamber, they map the chamber onto itself, and the centers of the Weyl chambers are fixed-points of the combined symmetry transformations, see e.g.[@vanBaal:2000zc] for a more detailed discussion. Hence, in the center-symmetric phase $\bar\varphi$ has to settle at these points. We conclude that $\bar\varphi$ and the Polyakov loop $L[\bar\varphi]$ are indeed order parameters for the confinement-deconfinement phase transition, $$\label{eq:barL}
L[\bar\varphi]\ \left\{\begin{array}{rcl} = 0 & \quad {\rm for}\quad & T<T_c\\[2ex]
> 0 & \quad {\rm for}\quad & T>T_c\end{array}\right. \,,$$ where we have singled out the positive semi-definite minimum on the real axis, i.e. $\bar\varphi^8=0$.
Order parameters and gauge fields {#sec:OrderParamsgauge}
=================================
In the following we compute the order parameters, and , within a non-perturbative approach to continuum Yang-Mills theory formulated in terms of the gauge fields $A_\mu$. Both order parameters can be expressed in terms of the temporal gauge field $A_0$ and expectation values of correlations of $A_0$. We have already discussed that the computation of $\lLA$ within an expansion in the gauge field requires the computation of infinite-order correlation functions of the gauge field. On the lattice, on the other hand, it is straightforward to compute $\lLA$, see e.g. [@Boyd:1996bx; @Kaczmarek:2002mc; @Zantow:2003uh; @Gupta:2006qm; @Gupta:2007ax; @Dumitru:2003hp; @Gavai:2010qd]. In this manner, the first-order nature of the phase transition in $SU(3)$ with critical temperature $T_c/\sqrt{\sigma}=0.646$ in units of the string tension $\sigma$ has been established. We shall see in that within the functional renormalisation group the task of computing $\lLA$ in the continuum is also tractable. It can be reduced to solving a differential equation that is linear in the scale-derivative of $L$ and its second $\varphi$-derivatives.
In turn, the computation of the order parameter only requires the computation of the non-perturbative expectation value of the constant temporal background gauge field in a background-field approach, cf. [@Braun:2007bx; @Fister:2013bh], or in the Polyakov gauge [@Marhauser:2008fz], $$\label{eq:contPL}
L[\lA0] = \frac{1}{N}{\ensuremath{\operatorname{Tr}}}_f e^{i g\beta \lA0 }\,.$$ The expectation value $\lA0$ is determined as the minimum of the non-perturbative effective glue background potential $V[\bar A_0]$ in the background-field approach, e.g. [@Abbott1981189]. To this end we introduce a background field $\bar A_\mu$ by splitting the full gauge field linearly into $$\label{eq:backsplit}
A_\mu=\bar A_\mu+a_\mu\,,$$ with $\langle a_\mu\rangle =0$. The background field enters the background gauge-fixing condition $$\label{eq:backgauge}
\bar D_\mu\, a_\mu =0\,,\quad {\rm with}\quad \bar D_\mu
=\partial_\mu + i g \bar A_\mu \,.$$ The background-field effective action $\Gamma_k[\bar A,\phi]$ depends on the background field $\bar A_\mu$ and the fluctuation field $\phi=(a_\mu,\, c,\,\bar c,\, ...)$. This setting allows for the definition of a gauge invariant effective action $$\label{eq:Gamgauge}
\Gamma_k[A_\mu]=\Gamma_k[A_\mu,0]\,,$$ that is invariant under the gauge transformation $$\label{eq:gaugetrafo}
A_\mu \to -\0{i}{g} ( U D_\mu U^{-1})\,,\quad {\rm with} \quad D_\mu
= \partial_\mu + i g A_\mu\,.$$ For constant fields this reduces to $A_\mu\to U A_\mu U^{-1}$ as for $\varphi$ in . The related effective background potential is given by $$\label{eq:Vglue}
V[A_0]= \01{\beta {\cal V}}\, \Gamma[A_0]\,,$$ where $\cal{V}$ the spatial volume. It is evident from the discussion in the last section that simply is the potential for constant $\varphi$ with $V(\varphi) = V[A_0(\varphi)]$ with the relation for $SU(3)$.
In perturbation theory the effective potential $V[A_0]$ has first been computed in [@Gross:1980br; @Weiss:1980rj]. It features only the deconfining minimum at $A_0=0$. A non-perturbative approach for the computation of $ V[A_0]$ has been established in [@Braun:2007bx; @Marhauser:2008fz; @Fister:2013bh]. By now computations of the glue potential from a variety of methods have been put forward, including the FRG and Dyson-Schwinger equations, 2PI-schemes, perturbative approaches and the lattice [@Braun:2007bx; @Braun:2009gm; @Braun:2010cy; @Marhauser:2008fz; @Heffner:2015zna; @Reinhardt:2012qe; @Fister:2013bh; @Fischer:2013eca; @Reinhardt:2013iia; @Fukushima:2012qa; @Reinosa:2014ooa; @Reinosa:2014zta; @Diakonov:2013lja; @Greensite:2012dy; @Langfeld:2013xbf].
The expectation value $\bar\varphi= g\beta \langle A_0\rangle/(2 \pi)$ is then given by the minimum of the background effective potential, i.e. it is defined via $$\label{eq:<A>}
\left. \0{\partial V[A_0]}{\partial A_0}\right|_{A_0=\langle
A_0\rangle}=0\,,\quad {\rm with}\quad L[\langle A_0\rangle]\geq 0\,.$$ As can be seen from , the minimum is not unique, and we choose the one which gives a real, positive semi-definite traced Polyakov loop via , cf. our discussion in . The resulting critical temperature for $SU(3)$ is in quantitative agreement with the lattice result with a critical temperature $T_c/\sqrt{\sigma}=0.655$ in units of the string tension $\sigma$, see [@Fister:2013bh]. Here, the biggest systematic uncertainty concerns the relative scale setting in the continuum and on the lattice, leading to a systematic error estimate of about $10\%$ [@Fister:2013bh].
Note that the two order parameters derived from the Polyakov loop, $\lLA$ and $L[\lA0]$, are indeed closely related: Jensen’s inequality yields the relation [@Braun:2007bx] $$\langle L[A_0]\rangle \leq L[\lA0]\,,
\label{eq:Jensen}$$ in the region where $L[A_0]$ is convex. Furthermore we have, cf. [@Marhauser:2008fz; @Fister:2013bh], $$L[\lA0]=0\Leftrightarrow \lLA=0\,.
\label{eq:PLTc}$$
Both these observables have been studied extensively and naturally reproduce the order of the phase transition as well as the critical temperature. However, the computations also reveal that the inequality is far from being saturated. This statement holds for $T> T_c$, i.e. in the center broken phase in pure Yang-Mills theory, and for all temperatures in QCD.
For QCD effective models, e.g. [@Fukushima:2003fw; @Megias:2004hj; @Ratti:2005jh; @Mukherjee:2006hq; @Roessner:2006xn; @Schaefer:2007pw; @Fukushima:2008wg; @Herbst:2010rf; @Skokov:2010uh; @Skokov:2010wb; @Schaefer:2011ex; @Herbst:2013ufa; @Herbst:2013ail; @Haas:2013qwp] this is particularly important. There, one usually relies on input from other methods to fix the gauge sector. The common procedure is to introduce a Polyakov-loop potential $V_L(L,\bar L)$, with $\bar L$ being the conjugate Polyakov loop. The coefficients of $V_L$ are then fitted to lattice Yang-Mills data, i.e. the potential is based on $\lL$.
This has to be contrasted with the fact that most low-energy effective models can be derived from continuum QCD in a standard formulation with gauge fields. Hence, they should be formulated in terms of $\langle A_\mu\rangle$. Then, also the ultraviolet parameters of these models at their initial scale, $\Lambda\approx 1$ GeV, can be determined from continuum QCD at this scale, and the gauge sector is related to the glue potential, $V[A_0]$.
Alternatively, a formulation in terms of $\lL$ is possible. Then, however, the matter part of the models is subject to an inherent Gaußian approximation $$\label{eq:Gauss}
\langle L_1 \cdots L_n \bar L_1 \cdots \bar L_m\rangle =
\prod_{i=1}^n \langle
L_i\rangle \prod_{j=i}^m \langle \bar L_i\rangle\,,$$ with the shorthand notation $L_i=L(\vec x_i)$ and similarly for $\bar
L_j$. Further reductions such as $$\label{eq:Gausstr}
\0{1}{N^2} \langle{\ensuremath{\operatorname{Tr}}}P_1 P_2 \rangle = \langle L_1
\rangle \langle L_2
\rangle \,,$$ with $P_i=P(\vec x_i)$ are used. neglects spatial fluctuations of the Polyakov-loop variable $P(\vec x)$. Note that in the Gaußian approximation underlying and , the inequality saturates, which is most easily seen in the Polyakov gauge . In other words, the difference between the observables $\lL$ and $L[\lA0]$ encodes the non-Gaußianity of the Polyakov loop, and the explicit results show the failure of the Gaußian approximation. Thus, to achieve a consistent treatment of the gauge sector in QCD effective models, a reliable resolution of the deviation between these two quantities is of utmost importance. In the present work we fill this gap and compute both definitions of the Polyakov loop in a unified approach from the FRG.
Apart from its relevance for effective models, the calculation of $\lL$ from the FRG represents one more observable that is available both, in the continuum and on the lattice, and allows us to demonstrate that the relevant dynamics of the system is properly accounted for in the functional continuum approach.
Continuum approach to confinement order parameters {#sec:PLFlows}
==================================================
As already emphasised, the Polyakov loop $\lLA$ is a non-local and all-order correlation function of the temporal gauge field. First-principle continuum approaches to QCD, however, are based on the computation of local correlation functions of the fundamental fields, and in particular $A_0$. It is this difference that makes the continuum computation of $\lLA$ so difficult. At large momentum scales $k\gg \Lambda_{\text{\tiny QCD}}$, this problem is resolved within the perturbative expansion that allows us to drop the higher-order correlation functions due to the small gauge coupling. An extension of such a study into the non-perturbative domain is made feasible by functional methods. In the present work we resort to the functional renormalisation group (FRG).
![Diagrammatic representation of the FRG flow for YM theory. Curly lines denote gluonic degrees of freedom, while dashed lines represent the ghost and anti-ghost. The black dot indicates the full propagator and the crossed circle symbolises the regulator insertion.[]{data-label="fig:YMFlow"}](Gammakflow){height=".085\textheight"}
Functional renormalisation group
--------------------------------
The FRG allows us to systematically compute the scale dependence of correlation functions. To this end, an infrared regulator function, $R_k$, is introduced into the propagators of the theory. Such a regulator suppresses the propagation of quantum and thermal fluctuations below the infrared cutoff scale $k$. Lowering the cutoff scale $k$ implements the Wilsonian idea of integrating out fluctuations momentum shell by momentum shell. In the presence of the scale $k$ the scale-dependent effective action, $\Gamma_k$, only carries the quantum and thermal physics of momentum fluctuations above the cutoff scale. In turn, all quantum and thermal fluctuations below this scale are suppressed. Hence $\Gamma_k$ interpolates between the bare action at an ultraviolet scale $\Lambda$, and the full quantum effective action for $k=0$. Its evolution is described by the Wetterich equation [@Wetterich:1992yh], $$\partial_t\Gamma_k[\phi] = \012 {{\ensuremath{\operatorname{Tr}}}}\, G_k[\phi] \dot R_k\,,\ \
{\rm with}\ \ G_k[\phi]=\0{1}{\Gamma^{(2)}_k[\phi] + R_k}\,,
\label{eq:FRG}$$ and the trace sums over momenta, internal indices and species of fields, including minus signs for fermionic loops. In , $\dot{R}_k = \partial_t R_k=k\partial_k R_k$ denotes the scale derivative of the regulator w.r.t. the FRG time, $t=\log(k/\Lambda) $, and $G_k$ is the full propagator at the scale $k$. A pictorial representation of this equation is given in for the special case of Yang-Mills (YM) theory used in this work. The curly line denotes the scale-dependent gluon propagators, while the dashed line represents the ghost propagator. The crossed circle symbolises the regulator insertion, $\dot{R}_k$. By virtue of the regulator function, is ultraviolet and infrared finite. It has a one-loop structure, but is fully non-perturbative. The solution of this equation relies on the propagators, which themselves obey flow equations derived from , see [@Fister:2011uw; @Fister:2013bh].
As can be seen from relation , the flow of the glue potential, $V_k[A_0]$, is directly related to that of the effective action, $$\partial_tV_k[A_0] = \frac{1}{\beta\mathcal{V}}\,\partial_t\Gamma_k[A_0]\,.
\label{eq:FlowVk}$$
Furthermore, in [@Pawlowski:2005xe] a flow equation for general observables $I_k= \langle \hat I_k[J,\hat\phi] \rangle$ has been derived, where $\hat\phi$ is the quantum field with $\phi=\langle
\hat\phi\rangle$. The flow equation for $I_k$ reads $$\begin{aligned}
\partial_t I_k[\phi] = -\frac12 {\rm Tr}\left\lbrace \left( G_k\dot
R_kG_k\right)_{\bf ab}
\frac{\delta}{\delta\phi_{\bf b}}\frac{\delta}{\delta\phi_{\bf a}}
I_k[\phi]\right\rbrace\,.
\label{eq:MasterFlowEq}\end{aligned}$$ The bold letter indices ${\bf a},{\bf b}$ collect species of fields, Lorentz and internal indices and again, $G_k$ denotes the full propagator at RG-scale $k$. We refer the reader to [@Pawlowski:2005xe] for more details and the derivation of , and to [@Litim:1998nf; @Berges:2000ew; @Polonyi:2001se; @Pawlowski:2005xe; @Gies:2006wv; @Schaefer:2006sr; @Rosten:2010vm; @Braun:2011pp; @vonSmekal:2012vx] for QCD-related reviews of the FRG.
![Diagrammatic representation of the FRG flow for general operators $I_k[\phi]$ in YM theory. The boxes indicate the second functional derivative of the operator w.r.t. the gluon and ghost fields, respectively.[]{data-label="fig:IkFlow"}](./Ikflow){height=".085\textheight"}
shows the diagrammatic representation of the flow , again for YM theory. Its structure is similar to , except for the presence of second derivatives of the operator $I_k$ w.r.t. the fields, indicated by the boxes. Once more the propagators of the theory serve as input for the solution of this flow equation.
The set of operators $I_k$ for which is valid, includes, e.g., the (connected and disconnected) $n$-point correlation functions $ \lbrace \langle
\phi_1\phi_2\dots\phi_n\rangle\,|\,n\in\mathbb N\rbrace$ as well as $\tfrac{\delta\Gamma}{\delta\phi}$, cf. [@Pawlowski:2005xe]. In particular, the observable $I_k~=~\lLA$ falls into this class of operators.
Flow equation for the Polyakov loop {#sec:flowPol}
-----------------------------------
From now on we restrict our discussion to $SU(3)$ Yang-Mills theory, full QCD will be discussed elsewhere. We use the covariant background-field formalism [@Abbott1981189] with the background gauge-fixing condition in the Landau-deWitt gauge. The vertices $\Gamma^{(n)}$ of the effective action $\Gamma_k[A_\mu]$ are directly related to S-matrix elements, and hence directly carries physics information. Its flow equation is simply that of $\Gamma_k[A_\mu,0]$ in , $$\label{eq:FRGback}
\partial_t\Gamma_k[A_0] = \012 {{\ensuremath{\operatorname{Tr}}}}\, G_k[A_0] \dot R_k\,,$$ with $$\label{eq:FRGbackG}
G_k[A_0] = \0{1}{\0{\delta^2 \Gamma_k}{\delta\phi^2} +
R_k}[A_0,\phi=0]\,.$$ Eqs. and entail that the propagators of the fluctuation fields evaluated at vanishing fluctuation field, $G_k[A_0]$, govern the flow of the gauge-invariant effective action. This is the peculiar situation that the flow of a gauge invariant quantity is governed by a gauge variant quantity, the fluctuation field propagator. The latter satisfies a complicated Slavnov-Taylor identity, see e.g. [@Pawlowski:2005xe]. However, for vanishing fluctuation field, $\phi=0$, the propagator, , is gauge covariant and can be expanded in gauge covariant tensors. It is this property that enables us to use the results of the Landau gauge, corresponding to the background Landau-deWitt gauge , [@Braun:2007bx; @Braun:2010cy; @Fister:2011uw; @Fister:2013bh]. The gauge covariance allows us to expand the two point correlator of the fluctuation field, $\delta^2/\delta\phi^2\Gamma[A_0,0](p^2)$ in Landau-deWitt gauge about the Landau gauge propagator [@Braun:2007bx; @Fister:2013bh], $$\label{eq:expandLandau}
\0{\delta^2\Gamma[A_0,0]}{\delta\phi^2}(p) =
\0{\delta^2\Gamma_{\text{\tiny{Lan}}}}{\delta\phi^2}(D) + f(D,F,A_0) \,,$$ where the covariant tensor $f$ satisfies $f(D,0,0)\equiv 0$ and the $A_0$-dependence of $f(D,F,A_0)$ indicates the dependence on the covariant tensor $P(\vec x)$. The subscript ’Lan’ indicates the standard Landau gauge with $\partial_\mu A_\mu=0$. The extension to covariant momenta is done such that $ D\cdot \delta^2/\delta A_0^2
\Gamma_{\text{\tiny{Lan}}}(D) =0$, up to the gauge fixing term. For constant fields the field-strength tensor vanishes, $F=0$, and the covariant tensor reduces to $f(D,A_0)$.
allows us to use the results of the finite temperature ghost and gluon propagators of [@Fister:2011uw] for the first term on the right-hand side of . The computation of $f(D,F,A_0)$ will be discussed in . The Landau gauge two point functions at finite temperature are conveniently parametrised as $$\label{eq:G2}
\Gamma_{L/T,k}^{(2)}(p) = Z_{L/T,k}(p)\, p^2 P^{L/T} \,, \quad
\Gamma_{\bar c c,k}^{(2)}(p) = Z_{c,k}(p)\, p^2 \,,$$ where $P^{L}$ and $P^T$ are the projection operators onto the chromo-electric and chromo-magnetic gluon modes, and the identity matrix in colour space is implied. For more details see [@Fister:2011uw]. The following computation of Polyakov loop expectation values only implicitly depends on $Z_T$, which relates to the spatial components of the gauge field. For the sake of simplicity we hence denote $$\label{eq:Za}
Z_{a,k}(p) = Z_{L,k}(p)\,, \quad \quad Z_{\bar A,k}(p) = (Z_{\bar A})_{L,k}(p)\,.$$ Both, the fluctuating, $a_0$, and the background, $\bar A_0$, wave-function renormalisations are the chromo-electric ones.
With this choice of gauge, the background temporal gauge field, $\bA0$, is in the Polyakov gauge. For vanishing fluctuation field, $a_\mu=0$, we define, cf. Eqs. and , $$\label{eq:varphiold}
\beta \bar g_k\, Z_{\bar A,k}^{1/2}\, \bar A_0 = 2\pi\left(\bar \varphi_3t^3 +
\bar \varphi_8t^8\right) =: 2\pi \bar\varphi\,,$$ where we have made the wave-function renormalisation $Z_{\bar
A,k}=Z_{\bar A,k}(p_0=0,\vec p^2=k^2)$ of the background field explicit. The wave-function renormalisation is evaluated at $\vec p\,^2
=k^2$, which is the important momentum scale in the flow. The running coupling $\bar g_k$ is the background running coupling, evaluated at the symmetric point $p^2 =k^2$. It tends toward a finite value in the infrared and shows an infrared plateau, see e.g.[@Eichhorn:2010zc].
{height=".26\textheight"} {height=".26\textheight"}
Within our approach, the dependence on spatial gauge fields and ghosts originates solely from $A_0-\vec A$ and $A_0$-ghost vertices. Such vertices entail that all quantities involved are fully dressed, which is accounted for by using the running coupling, $g=\bar g_k$, in . Furthermore, the coupling $\bar g_k$ in is the renormalisation group (RG) invariant background running coupling. The combination $Z_{\bar A,k}^{1/2}\,
\bA0$ is RG-invariant, too, and hence $\bar\varphi$ is RG-invariant. Moreover, we also have $\partial_t (g_k\, Z_{\bar
A,k}^{1/2}\, \bA0)=0$, and hence $\partial_t \bar\varphi=0$.
Now we use that all background correlations can be expressed in terms of those of the fluctuations with the help of Nielsen identities. In this is particularly simple for the cutoff-independent combination $\bar g_k Z_{\bar A,k}$. We rewrite it as $$\label{eq:barAAmap}
\bar g_k Z^{1/2}_{\bar A,k} = g_k Z^{1/2}_{A,k} \,, \quad Z_{A,k} =
\0{1}{Z_{a,k} Z_{c,k}^2}\,,\quad
g_k = \0{\bar g}{Z_{a,k}^{1/2} Z_{c,k}}\,,$$ where $g_k$ and $Z_{A,k}$ are defined via fluctuating wave-function renormalisations, $Z_{a/c,k}=Z_{a/c,k}(p_0=0,\vec p\,^2=k^2)$. Moreover, $\bar g$ is the background coupling at $k$ with $Z_{a,k}^{1/2}
Z_{c,k}=1$. We emphasise that the splitting is not unique, but rather a convenient definition of $Z_{A,k}$, enabled by the cutoff independence of the combination $\bar g_k Z_{\bar A,k}$.
For non-vanishing temporal fluctuation, $a_0\neq 0$, the parametrisation generalises to $$\label{eq:varphia}
2 \pi \varphi = \beta g_k\, Z_{A,k}^{1/2}\, \left(\bar A_0
+ \tfrac{Z_{a,k}^{1/2}}{Z_{ A,k}^{1/2}}\, a_0\right) \,.$$ Note that the combination is RG-invariant, but not $t$-independent for $a_0\neq 0$. Moreover, it is the choice of the splitting and the relation of $Z_{A,k}$ to the fluctuation wave-function renormalisations that leads to the simple ratio $Z_{a,k}^{1/2}/Z_{ A,k}^{1/2}$ in : taking derivatives w.r.t. the fluctuation fields naturally leads to powers of the wave-function renormalisation and the fluctuation coupling $g_{k} Z^{1/2}_a$, see the vertex parametrisation [@Fischer2009b] It has been shown in [@Mitter:2014wpa] that this approximation of the vertices is quantitatively reliable. The ratio is then given by $$\label{eq:barZ}
\tfrac{Z_{a,k}^{1/2}}{Z_{\bar A,k}^{1/2}} = \0{1}{Z_{c,k}}\,.$$ With these prerequisites we can now apply the flow to $\lLA$. To this end we employ the parametrisation $$\lLA = Z_{L,k}[\bar A,\phi]\cdot L(\varphi)\,,
\label{eq:lLpara}$$ where the relative factor, $Z_{L,k}[\bar A,\phi]$, depends on all fluctuation fields $\phi$ and on the background field $\bar
A$. Furthermore we have $\phi = \phi(\varphi)$ via . It is the renormalisation factor of the composite operator $L$, in analogy to the wave-function renormalisation $Z_\phi$ of the field operators $\phi$.
The $t$-derivative on the left-hand side of is taken at fixed background $\bar A_0$. As we have seen above, this corresponds to $\partial_t \bar\varphi=0$, which implies $\partial_t
L(\bar\varphi)=0$ and leads to the flow equation $$\begin{aligned}
\partial_t \lL = & L \,\partial_t Z_{L,k} \nonumber\\
= & -\frac{g^2_k\, Z_{a,k}\,\beta}{8 \pi^2 } \int \0{d^3p}{(2 \pi)^3}
\left[ G_k\dot R_kG_k
\right]_{ab} (\omega_0=0,\vec p)\, \nonumber\\[2ex]
& \times \left( Z_{L,k}\, L^{ba} + Z_{L,k}^b \,L^a +Z_{L,k}^a\,L^b +
Z_{L,k}^{ba}\, L\right)\,,
\label{eq:flowL}\end{aligned}$$ with gauge group indices $a,b=1,...,N^2-1$ in $SU(N)$. For the sake of brevity we have dropped the $\varphi$-arguments on the right-hand side and have introduced the notation $$\label{eq:diffnot}
X^a = \0{\partial X}{\partial {\varphi_a}} \,,\qquad X^{ab} =
\0{ \partial^2 X}{\partial {\varphi_a}\partial\varphi_b} \,,$$ with $X=L, Z_{L}$. The $x_0$-independence of $\lA0$, due to the Polyakov gauge for the background field, entails that only the lowest Matsubara frequency, $\omega_0=0$, is present in this equation. Furthermore, in this gauge only the derivatives with $a,b=3,8$ are non-vanishing in . Upon taking the trace, only the components of the propagators that carry no explicit dependence on the background field remain. Hence, the only field dependence on the right-hand side is that of $L(\varphi)$ and $Z_{L,k}$. As the field-dependence of the latter is only generated by the flow, we conclude that the whole field dependence originates from $L(\varphi)$. Note also that the above analysis also holds for general constant backgrounds without gauge fixing. In summary, the symmetry properties of $L(\varphi)$ under shifts and center transformations are carried over to $\lLA$ via the flow .
effectively is a flow equation for the ratio $Z_{L,k}$ of the two order parameters, $L(\varphi)=L[\lA0(\varphi)]$ and $\lL(\varphi)$ for general background $\varphi$. Eventually we evaluate all observables at the minimum $\bar\varphi$ defined by . $Z_{L,k}$ provides a measure for the different impact of quantum and thermal fluctuation on the two definitions of the Polyakov loop. It hence is a measure of the non-Gaußianity of Polyakov loop correlations as well as of gauge field correlations that is relevant for low energy effective models. The seeming parametric singularity of when $L(\varphi)=0$ is lifted by our parametrisation , which guarantees that in this case also $\lL=0$, cf. .
It is also worth noting that the off-diagonal entries in the last line of in general are non-vanishing and purely imaginary. They couple to the off-diagonal terms of the propagator, $G_k^{38},\, G_k^{83}$, which are only present at finite quark chemical potential. This property facilitates the emergence of different, real values for the Polyakov loop and its conjugate, $\lL$ and $\lLc$, at finite density. This is expected since these observables relate to the free energies of quarks and anti-quarks, respectively.
Additionally, knowledge of $Z_{L,k}$ enables us to translate between the glue effective potential, $V(\varphi)$, cf. , as computed in [@Braun:2007bx; @Braun:2009gm; @Braun:2010cy; @Marhauser:2008fz; @Heffner:2015zna; @Reinhardt:2012qe; @Fister:2013bh; @Fischer:2013eca; @Reinhardt:2013iia; @Fukushima:2012qa; @Reinosa:2014ooa; @Reinosa:2014zta], and the Polyakov loop potential used in models, $V_{L}(L=\lL)$ by the above relation $\lL(\varphi) =
Z_L(\varphi) L[\varphi]$. This can, for example, be used to evaluate and improve model potentials.
Input {#sec:input}
-----
In order to solve the flow , the scale-dependent YM propagators and the glue potential, $V_k(\varphi)$, are the only inputs needed. The covariant tensor $f(D,A_0)$ is directly related to the background potential via , to wit $$\label{eq:fV}
f(D,A_0)= Z_{a,k}\, \beta^2\0{g_k^2 }{4 \pi^2} V_k^{ab}(\varphi)\,.$$ Here the prefactor $Z_{a,k}$ carries the correct RG-scaling of a gluonic two-point correlation function of the fluctuation field, and the rest is RG-invariant. FRG calculations for the YM propagators have previously been put forward in [@Fister:2011uw], and an FRG approach to the glue potential has been presented in, e.g., [@Fister:2013bh]. Here we build upon these results to achieve a consistent description of fluctuation effects on the Polyakov loop.
First, we briefly recapitulate the calculation of the glue potential, whose flow equation is depicted diagrammatically in . To solve this equation we make use of the temperature- and scale-dependent results for the ghost and gluon propagators from [@Fister:2011uw]. There, results have been obtained with the ghost and gluon regulators $$\label{eq:regs}
R_{a/c,k}(p^2) = \bar Z_{a/c,k}\, p^2\, r(p^2/k^2)\,, \quad \quad r(x)=
\frac{x}{ e^{x^2}-1}\,,$$ which ensures the necessary momentum locality of the flow, as well as an exponential thermal decay proportional to $k/T$ for large cutoff scales. The wave-function renormalisations $\bar Z_{a/c}$ are related to $Z_{a/c}$, for details see [@Fister:2011uw]. It has been shown in [@Fister:2013bh] that the corresponding glue potential correctly reproduces the order and critical temperature of the $SU(3)$ deconfinement phase transition. The resulting glue potential, $V_{k\to0}(\varphi_3,\varphi_8)$, is shown in in the confined (left) and deconfined phase (right). The symmetries discussed in are clearly visible and we have indicated the line $\varphi_8=0, \varphi_3\geq0$, to which we restrict, by the black, dashed line. The corresponding potential along this line is then shown in for several temperatures below and above the critical temperature, $T_c=264\pm26$ MeV. As also discussed in [@Fister:2013bh], the absolute scale in this computation is set by a comparison of the peak position in the propagators with the lattice [@Maas:2011se; @Fischer:2010fx; @Maas:2011ez], which results in an error in $T_c$ of about $10~\%$.
The physical gauge field, $\lA0\sim\bar\varphi$, is determined as the global minimum of the infrared glue potential, $V_{k\to0}$, see . The order parameter $L(\bar\varphi)$ is calculated via . FRG data for the YM propagators are available up to $T\lesssim1.1$ GeV [@Fister:2013bh], which allows us to compute the glue potential fully non-perturbatively up to this scale. For all temperatures the glue potential $V(\varphi)$ can be fitted well via the perturbative Weiss potential as $$V(\varphi) = a(T)\, V_W^{\rm SU(3)}(\varphi)
+ b(T)\, V_W^{\rm SU(3)}(\varphi)^2\,,$$ where the Weiss potential for $SU(2)$ is given by $V_W^{\rm
SU(2)}(\varphi)=(1-(1-2\varphi)^2)^2$ and its $SU(3)$-counterpart can be constructed in the standard way as $$V_W^{\rm SU(N)}(\varphi) = \012\sum_{n=1}^{N^2-1}
V_W^{\rm SU(2)}(x_n(\varphi))\,,$$ with $x_n=\lbrace 0,
\pm\varphi_3,\pm\tfrac{\varphi_3\pm\sqrt{3}\varphi_8}{2} \rbrace$ in $SU(3)$ see e.g. [@Braun:2010cy]. For high temperatures, $T\gtrsim
1$ GeV we find that the coefficient $b(T)$ is rather small and the potential is well approximated by the perturbative form. We exploit this observation in the following and replace the glue potential in by the Weiss potential at $T\gtrsim1$ GeV.
Furthermore, it turns out that the thermal propagators can be approximated by the $T=0$ propagators plus a Debye mass which is fitted to the data. At large temperatures, $T\gtrsim1$ GeV, where no FRG data for the propagators is available yet, we use this construction. We have checked that the detailed form of the mass term has little impact on the resulting Polyakov loop. In the plots shown in the next section, we indicate the scale where we switch from first-principles data to the fit for the Debye mass and Weiss potential by a dashed, vertical line.
To finally solve the flow for $\lLA$, the last remaining ingredient is the running coupling, $g_k$. Within our approach, it is related to the strength of the ghost-gluon vertex, $g_k^2 \sim
Z_{a,k}^{-1} Z_{c,k}^{-2}$, and can also be deduced from the propagators, cf. [@Fister:2011uw].
![FRG result for the $SU(3)$ glue potential along the line $\varphi_3\geq0, \varphi_8=0$ for various temperatures. The critical temperature, associated with the first-order transition is given by $T_c=264\pm26$ MeV.[]{data-label="fig:VkSU3"}](./VkSU3){width="\columnwidth"}
Fluctuations and renormalisation {#sec:Renormalization}
================================
Following the procedure outlined in the last section we are now in the position to compute both Polyakov loops, $L(\bar\varphi)$ and $\lL =
Z_{L,k=0}(\bar\varphi) L(\bar\varphi)$. For the computation of the latter we first remark that the composite operator $\lL$ is UV-relevant and has to be renormalised. In the present functional renormalisation group approach the renormalisation procedure is easily accessible. It simply entails that $$\label{eq:RenPL}
\Lambda \partial_\Lambda \lL_{k=0} \stackrel{!}{=} 0\,,$$ for the renormalised traced Polyakov loop: the expectation value $\lLA$ at vanishing cutoff scale, $k=0$, does not depend on the initial cutoff scale, $k=\Lambda$. Hence, this amounts to adjusting the $\Lambda$-dependence of $\lLA_\Lambda$. For to hold, $\lLA_\Lambda$ has to satisfy the flow equation , see e.g. [@Pawlowski:2005xe]. Within the parametrisation this translates into the requirement that $Z_{L,\Lambda}$ has to satisfy . In turn, the unrenormalised traced Polyakov loop is defined by using the trivial initial condition $Z_{L,\Lambda}=1$, which violates . Thus, the multiplicative factor $Z_{L,\Lambda}$ carries the renormalisation of the Polyakov loop $\lLA$, and is related to the multiplicative renormalisation factor $Z_{\text{\tiny{lat}}}$ used on the lattice.
For asymptotically large scales, that is $\Lambda/\Lambda_{\text{\tiny
QCD}}\to \infty$ and $\Lambda/T_{\text{\tiny{max}}}\to \infty$, it is sufficient to solve up to sub-leading terms that vanish in this limit. Asymptotically large $
\Lambda/\Lambda_{\text{\tiny QCD}}$ guarantees the perturbative limit that facilitates the computation of the initial $Z_{L}$. The second requirement, $ \Lambda/T_{\text{\tiny{max}}}\to\infty$ guarantees the temperature-independence of the initial condition.
In the above asymptotic limit, the temperature dependence of the gluon propagators, including the term proportional to the potential curvature $V^{ab}$, and of the gluonic wave-function renormalisation $Z_a$ decays exponentially with $\exp \{-c(r) k/T\}$, see [@Fister:2011uw; @Fister:2013bh; @Fister:2015eca]. The prefactor $c(r)$ depends on the shape function $r$ in , and vanishes for non-analytic regulators, see [@Fister:2015eca]. In such a case the thermal suppression is only polynomial in $T/k$. For the present regulators, , we have an exponential decay. The same exponential decay holds for general correlation functions, and hence also applies to the coupling $\alpha_s$. Accordingly, the temperature dependence of the flow decays exponentially, reflecting the temperature independence of the UV renormalisation. Moreover, the flow also becomes $\varphi$-independent exponentially fast as the sum of the $L^{ab}$-terms is proportional to $L$ for vacuum propagators which are proportional to $\delta^{ab}$. Hence, all $Z_L^{a},Z^{ab}_L$ terms on the right-hand side can be dropped, and we are left with the asymptotic flow $$\label{eq:flowLas}
\0{\partial_t Z_{L,k}}{Z_{L,k}} = \frac{g^2_k\, Z_{a,k}\,\beta}{6
} \int \0{d^3p}{(2 \pi)^3} \left[ G_k\dot R_kG_k
\right]_{bb} (0,\vec p)\,,$$ for the renormalisation factor. The prefactor has a logarithmic scaling with $k$ from both, the running coupling $g_k^2$ and the gluon wave-function renormalisation $Z_{a,k}$. The scaling of the latter is cancelled by the logarithmic scaling of the propagators and regulator. The remaining spatial momentum integral scales linearly with the cutoff scale. Hence we arrive at the total scaling $$\label{eq:Zren}
\frac{\partial_t Z_{L,k}}{Z_{L,k}} = z_L\,\alpha_{s,k}\,\0{k}{T}\,,$$ with the strong running coupling $\alpha_{s,k}=g^2_k/(4\pi)$ at $T=0$ up to exponentially suppressed terms. The factor $z_L$ is given by $$\label{eq:a}
z_L = \frac{2 \pi}{ 3} \int \0{d^3p}{(2 \pi)^3} \0{1}{k}\left[
Z_{a,k}\, G_k\dot R_kG_k \right]_{bb} (0,\vec p)\,,$$ with the propagators $G_k$ and $Z_{a,k}$ at vanishing temperature, up to exponentially suppressed terms. Moreover, $z_L$ tends towards a constant for large cutoff scales, $z_{L,k\to\infty}= 0.0177$.
Note also that the exponential thermal decay does not hold in momentum space at vanishing cutoff scale. All correlation functions, including the propagator and the running coupling, $\alpha_s(p,T)$, only show a polynomial decay of the thermal contributions with large momenta, that is with powers of $T/p$, similarly to the polynomial decay with $T/k$ for non-analytic cutoffs. However, contrary to the latter, the polynomial decay with $T/p$ relates to physics, for example that of the propagator relates to the Tan relations in many-body physics, for a discussion in Yang-Mills theory see [@Fister:2011uw]. In turn this entails that the choice of temperature-independent renormalisation conditions in a non-perturbative approach based on the running coupling at asymptotic momentum scales is intricate. Hence, in particular the identification of the finite temperature running coupling at large momenta with the $T=0$ one, $\alpha_s(p\gg
T,T)=\alpha_s(p\gg T, 0)$ introduces a temperature-dependent renormalisation scheme due to the sub-leading powers of $T/p$.
We emphasise that for the present renormalisation scheme with the regulator this is avoided: at large cutoff scales $k/T
\gg 1$ the difference of the couplings $\alpha_{s,k}(p,T)$ and $\alpha_{s,k}(p,T=0)$ is exponentially suppressed with $f(T/k,T/p)\exp\{ - c_T k/T\}$. Here, $f(x,y)$ has power law suppressions for both, $k/T\to\infty$ and $p/T\to\infty$. At $k=0$ the exponential suppression disappears. Hence the integrated flow leads to the sub-leading momentum-dependent terms in the coupling in $T/p$ that are only power law suppressed at large momentum scales. This has been shown in [@Fister:2011uw].
Nonetheless, a temperature-dependent renormalisation scheme is a consistent choice, and such a general setting is taken into account in the present setting by allowing for thermal shifts in the cutoff scale. This amounts to a sub-leading change of the initial condition for the flow of $Z_L$ proportional to $T/\Lambda$, and hence a sub-leading contribution to the flow proportional to $T/\Lambda$. The flow is peaked at about the temperatures scale, and we are led to the final expression for the renormalisation factor $Z_{L,\Lambda}$ at asymptotically large cutoff scales $\Lambda$ $$\label{eq:ZLlambda}
Z_{L,\Lambda}(T) = e^{a_{\text{\tiny{cont}}} \Lambda/T}\,, \quad
a_{\text{\tiny{cont}}} =a_0 + a_1 \frac{T}{\Lambda}\,\alpha_s( c_T
T) + \dots \,,$$ where the leading term $a_0 =1/\Lambda \int^\Lambda d k\,
z_L\,\alpha_{s,k}$ is proportional to $\alpha_{s,\Lambda}$ for large cutoff scales. However, it is very slowly varying due to its only logarithmic decay, and for the chosen initial cutoffs $\Lambda=10, 15,
18$ GeV in it is given by $a_0=0.0175$. More importantly, demonstrates the renormalisation group invariance of the traced Polyakov loop observable $\lL$. The higher-order term, with free coefficient $a_1$, acts as a constant shift in the Polyakov loop at low temperatures, and decreases at large $T$, respecting the correct high-temperature limit. It implements the subleading temperature-dependent renormalisation discussed above, and the parameters $a_1, c_T$ specify the scheme. Note also that for small temperatures above, but close to the phase transition the flow for $Z_L$ is peaked at the mass gap of the gluons and the scale $c_T T$ in $\alpha_s( c_T T) $ has to freeze at about the mass gap. This entails that the sub-leading term related to the potential temperature-dependent renormalisation provides a constant off-set in the infrared, and the ratio for expectation values of the traced Polyakov loop is independent of the temperature-dependent part of the renormalisation. In summary, the potential temperature-dependent RG scheme introduces in $Z_L$ a temperature-dependent, but cutoff-independent factor for asymptotically large temperatures, while it is constant for low temperatures due to the plateau in the background coupling. In we have used $a_1=0$, since it does not explicitly depend on $\Lambda$. This subtlety is discussed further in the next section.
![Demonstration of the cutoff-independence of the Polyakov loop, $\lL$ with $a_1=0$ in , and the comparison to $L[\lA0]$. The high-T limit in both cases is unity.[]{data-label="fig:PL_ren"}](./PLSU3_renLambda){width="\columnwidth"}
The above renormalisation procedure in the continuum is to be compared to the lattice renormalisation. There, it is known that taking the infinite volume limit at a fixed temperature leads to a vanishing Polyakov loop and an appropriate renormalisation procedure is needed [@Kaczmarek:2002mc; @Zantow:2003uh; @Gupta:2006qm; @Gupta:2007ax]. It was suggested already by Polyakov [@Polyakov1980171], that the linear divergences appearing in $\lL$ at any order in perturbation theory can be combined into an exponential factor. This leads to the definition of the renormalised Polyakov loop $$\lL_{\text{\tiny{lat}}} = Z_{\text{\tiny{lat}}}(T) \lL_{\text{\tiny{bare}}}\,,\quad
\text{with}\quad Z_{\text{\tiny{lat}}}(T) = e^{a_{\rm\tiny{lat}}\Lambda/T}\,,
\label{eq:renPL}$$ where suggests $a_{\text{\tiny{cont}}}=a_{\text{\tiny{lat}}}$ up to differences in the renormalisation condition. On the lattice, the constant $a_{\text{\tiny{lat}}}$ can be fixed via the zero-temperature heavy-quark potential [@Kaczmarek:2002mc; @Zantow:2003uh; @Gupta:2006qm], for other renormalisation schemes see e.g. [@Gupta:2007ax; @Dumitru:2003hp; @Gavai:2010qd; @Mykkanen:2012ri]. Such a renormalisation procedure is unique only up to a constant, i.e. an overall multiplicative renormalisation remains, cf. [@Gupta:2006qm; @Gupta:2007ax]. The different normalisations used on the lattice and in the continuum may also introduce a temperature-dependent part of the renormalisation as discussed above. This has to be taken into account in a comparison of the different schemes, which is deferred to the next section.
While both continuum observables, $L[\lA0]$ and $\lL$, serve as order parameters for the confinement-deconfinement transition, reveals clear differences between the two above $T_c$. It had been found previously, and is also confirmed by our data, that the overall shape of $L[\lA0]$ is much steeper than that of $\lL$. The difference between these two observables is quantified by $Z_{L,k\to0}(T)$ and is due to the different impact of thermal and quantum fluctuations on both observables.
First we note that $L[\lA0]$ saturates at its high-temperature limit, $L[\lA0]\stackrel{T\to\infty}{\longrightarrow}1$, already at $T\gtrsim
0.4$ GeV. In turn, $\lL$ rises much more gradually. It eventually overshoots unity at $T\approx 1.5$ GeV and approaches its high-T limit of one from above, in agreement with the expectation from perturbation theory [@Gava1981285].
The non-Gaußianity of the Polyakov loop in terms of correlations of the gauge field is quantified by the difference between $L[\lA0]$ and $\lL$ as discussed in . It is clearly visible in , and originates in two qualitatively different mechanisms: The first one simply encodes the renormalisation of the gauge field correlation functions that are in line with (resummed) thermal perturbation theory. These deviations are similar to those observed in the pressure, which are known to be significant even for temperatures $T\gg T_c$ as the approach to the Stefan-Boltzmann limit is rather slow. Still, they are captured well by HTL or similar resummation schemes such as (perturbative) two-particle-irreducible computations. This is visible in $Z_{L,k}$ which is approximately linear in the cutoff scale up to the thermal gap for temperatures $T\gtrsim (2-3)T_c$. In turn, for temperatures $T\lesssim (2-3)T_c$ the factor $Z_{L,k}$ shows an additional significant drop at about the non-perturbative mass gap of Yang-Mills and freezes below this scale.
In summary, the non-perturbative renormalisation factor $Z_L(T)$ carries much of the information about the perturbative and non-perturbative thermal and quantum fluctuations. The deviations are particularly large for the low temperature regime $T\leq (2-3)
T_c$. In this regime we also see a significant non-trivial dynamics related to the confinement-deconfinement phase transition. It is expected that these differences have a significant impact on model calculations if the standard Gaußian approximation is lifted.
Continuum and lattice results for $\lLA$ and renormalisation {#sec:Comparison}
============================================================
Now we turn to a comparison of $\lL$ in the continuum and on the lattice. This is done as a function of the reduced temperature, $T/T_c$, to accommodate for the differences in the relative scale setting procedure mentioned before. Using the renormalisation procedure discussed above with only $a_0$, the ratio of the continuum and lattice Polyakov loop expectation values is given by $\lL_{\text{\tiny{cont}}}/\lL_{\text{\tiny{lat}}}= 0.92$ for temperatures $T/T_c \lesssim 10$, that is up to about $3$ GeV. It has been discussed at length in the last section that such a constant low temperature off-set is introduced by a relative temperature-dependent renormalisation. Hence, this result confirms the quantitative agreement of both computations. Above this temperature we observe a change of this ratio towards unity triggered by the normalisation of both Polyakov loops at infinite temperature. As discussed in , the subleading term in the continuum renormalisation, that signals a temperature-dependent renormalisation scheme, on the other hand, is approximately constant at low $T\lesssim
10 T_c$. With the choice $a_1=0.47$ it accounts for the factor $1.088
=1/0.92$, see . Note that this term also accommodates possible sub-leading corrections of the continuum computation in the presence of full gluon propagators for asymptotically large temperatures. In the present work we have tested a large, but not complete set of sub-leading temperature corrections to the propagators, and did not observe any impact. In summary this strongly suggests a relative temperature-dependent renormalisation. To highlight the quantitative agreement up to asymptotically high temperatures, the inset in shows the continuum Polyakov loop with full renormalisation , and the lattice Polyakov loop from [@Gupta:2007ax] vs. $T/T_c$. A full investigation of this large temperature intricacy for $T/T_c \gg 1$ will be discussed elsewhere.
![Polyakov loop, $\lL$, compared to the renormalised lattice result, [@Kaczmarek:2002mc; @Gupta:2007ax]. The inset shows the high-T behaviour vs $T/T_c$.[]{data-label="fig:PLlat_T"}](./PLSU3_ren_T_small){width="\columnwidth"}
Conclusion and Outlook {#sec:conclusion}
======================
In this work we have discussed different order parameters for the confinement-deconfinement transition in Yang-Mills theory, based on the Polyakov loop variable $P(\vec x)= \exp(2 \pi i \varphi)$. The common order parameter is the expectation value of the traced Polyakov loop, $\lLA$, that is related to the free energy of a quark–anti-quark pair. It has been computed on the lattice for both, Yang-Mills theory and full QCD at finite temperature, but quantitative continuum computations had been missing so far. A further order parameter is related to the expectation value $\bar\varphi$ of the algebra-valued field $\varphi$, built from the expectation value of the gauge invariant eigenvalues, see the discussion around . With $\bar\varphi$ we can also compute the traced Polyakov loop $L(\bar\varphi)$, an order parameter that bears similarities to $\lLA$. The field $\varphi$ is directly related to the gauge field and its expectation values are $\langle A_0\rangle$. It has first been computed non-perturbatively within continuum approaches to Yang-Mills theory and QCD in [@Braun:2007bx; @Marhauser:2008fz; @Braun:2009gm; @Braun:2010cy; @Fister:2013bh], first lattice results have been presented e.g. in [@Langfeld:2013xbf; @Diakonov:2013lja]. It has also been argued that Polyakov-loop enhanced low-energy effective models should rather be based on the latter observables, $\bar\varphi$ and $L(\bar\varphi)$, than on the former, $\lLA$, see e.g. [@Braun:2009gm; @Pawlowski:2010ht; @Herbst:2013ail; @Haas:2013qwp; @Herbst:2013ufa; @Pawlowski:2014aha].
We have argued that the order parameters $\lLA$ and $L(\bar\varphi)$ agree within Gaußian approximations for correlations of the temporal gauge field $A_0$ or $\varphi$, and consequently within such an approximation for the correlations of $L$, see the discussion around Eqs. to on page 5. Low-energy effective models built on $\lLA$ rest on these Gaußian approximations for the computation of the quark-gauge field or rather quark-Polyakov loop fluctuations. In contradistinction, models built on $\bar\varphi$ are not subject to these Gaußian approximations as they can be derived directly from QCD. The difference between $\lLA$ and $L(\bar\varphi)$ is a direct measure for the non-Gaußianity of these fluctuations, and hence a measure for the quantitative reliability of $\lLA$-based models.
In the present work we have discussed the differences and similarities of these observables. We have also provided a review of the underlying symmetries and in particular of the connection of the algebra-valued field $\varphi$ to the gauge field, which is the basis of continuum formulations of Yang-Mills theories. The background glue potential in Yang-Mills theory has been computed within the functional renormalisation group, similarly to [@Braun:2007bx; @Braun:2010cy; @Fister:2013bh]. Using the flow equation for general composite operators put forward in [@Pawlowski:2005xe] we have furthermore derived a flow equation for $\lLA$. This flow equation is fully non-perturbative and also utilises the expectation value $\bar\varphi$. The difficult task of computing infinite-order correlation functions of the gauge field is resolved within the successive integration of momentum fluctuations: each iterative infinitesimal momentum-shell integration increases the order of gauge field fluctuations taken into account.
We have shown in that the present continuum computation of $\lLA$ is renormalisation group invariant. Moreover, the present functional renormalisation group approach provides a direct and simple access to the renormalisation procedure. This also allows us to discuss the relative renormalisation between the lattice and present continuum computations for $\lLA$. It has been argued that it amounts to a temperature-dependent renormalisation at large temperatures, see .
In we have shown that the results for $\lLA$ within the present non-perturbative continuum approach agree quantitatively with the lattice results, see . As the present continuum computation of $\lLA$ is based on the input $\bar\varphi$ or rather $\lA0$, it also is a further non-trivial support for the quantitative reliability of the results for $\lA0$ in [@Braun:2007bx; @Marhauser:2008fz; @Braun:2009gm; @Braun:2010cy; @Fister:2013bh], and related applications in QCD-enhanced low energy effective models [@Herbst:2013ail; @Haas:2013qwp; @Herbst:2013ufa].
We have shown that the non-Gaußianity of the fluctuations grows strong in the regime $T\lesssim 3 T_c$ which is the regime of interest for the low energy effective models, see . This suggests that, in particular for an analysis of fluctuations in QCD, it is important to take into account the Polyakov loop fluctuations if one aims at quantitative precision. This can either be done by using models based on $\bar\varphi$ or $\lA0$, or by amending the $\lLA$-based effective models by a fluctuation analysis of the Polyakov loop.
In turn, both order parameters are well described by (resummed) perturbation theory for temperatures $T\gtrsim 3 T_c$. Such a behaviour is already well-known from the trace anomaly. In the present continuum approach in the Landau-deWitt gauge it can be traced back to the non-perturbative mass-gap in the gluon propagator. Even though the propagator is a gauge-dependent quantity, this mass-gap reflects the mass-gap in Yang-Mills theory, and is directly related to confinement.
Finally, the extension of the presented approach to other gauge groups, such as $SU(2)$ or $SU(N)$ with $N>3$ and exceptional Lie groups, as well full QCD is straightforward and relies solely on the knowledge of the corresponding propagators. Here, the case of QCD with $N_f$ quark flavours is of particular interest: the corresponding unquenched propagators and unquenching of the glue potential have been put forward in [@Braun:2009gm; @Fischer:2013eca]. The structure of the flow equation then suggests that a splitting between $\lL$ and $\lLc$ will arise naturally, accounting for the fact that these objects are related to the free energies of quarks and anti-quarks, respectively. Furthermore, since light quarks break the center symmetry explicitly, the confinement-deconfinement transition is a crossover in this case. The temperature-dependent differences between $L[\lA0]$ and $\lL$, as observed here, then have a drastic impact on the pseudo-critical temperature, usually deduced from the inflection point of these observables. Indeed, the order parameter $L(\lA0)$ has been computed with continuum methods in $N_f=2$ flavour QCD in [@Braun:2009gm] for finite temperature and imaginary chemical potential, and in $N_f=2+1$ flavour QCD in [@Fischer:2013eca] for finite temperature and real chemical potential. As expected, both show a significantly stepper thermal rise in comparison to the lattice results for $\lL$, see e.g.[@Borsanyi:2010bp] for $N_f =2+1$ flavour results. Moreover, the pseudo-critical temperature of $L[\lA0]$ for the confinement-deconfinement phase transition agrees well with the chiral pseudo-critical temperature, see [@Braun:2009gm; @Fischer:2013eca]. From the present work it is clear that this originates in the different treatment of non-Gaußian fluctuations in the order parameters. The close chiral and confinement-deconfinement transition temperatures also suggests that the observable $L[\lA0]$ shows a closer resemblance to baryonic fluctuation observables that have a direct physics interpretation in dynamical QCD.
In summary, this work lays the foundation for a systematic study of the confinement-deconfinement transition in the phase diagram of QCD, both in model approaches and from first principles.
Acknowledgements {#acknowledgements .unnumbered}
----------------
We thank the fQCD collaboration, in particular L. Fister, for discussions and collaboration on related topics. This work is supported by the Helmholtz Alliance HA216/EMMI, by ERC-AdG-290623 and by the BMBF grant OSPL2VHCTG.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Radio-frequency induced spin transitions of one individual proton are observed for the first time. The spin quantum jumps are detected via the continuous Stern-Gerlach effect, which is used in an experiment with a single proton stored in a cryogenic Penning trap. This is an important milestone towards a direct high-precision measurement of the magnetic moment of the proton and a new test of the matter-antimatter symmetry in the baryon sector.'
author:
- 'S. Ulmer$^{1,2,3}$, C.C. Rodegheri$^{1,2}$, K. Blaum$^{1,3}$, H. Kracke$^{2,4}$, A. Mooser$^{2,4}$, W. Quint$^{3,5}$, J. Walz$^{2,4}$'
title: Observation of Spin Flips with a Single Trapped Proton
---
The challenge to understand the structure of the proton and to measure its properties inspires very different branches of physics. A very important property of the particle is its magnetic moment $\mu_\text{p} = (g_\text{p} / 2) \mu_N$, where $g_\text{p}$ is the Landé *g*-factor, $\mu_\text{N} = e \hbar / (2 m_\text{p})$ is the nuclear magneton, and $e/m_\text{p}$ is the charge-to-mass ratio. The most precise value for the *g*-factor of the free proton $g_\text{p}=5.585\,694 \,706(56)$ is extracted theoretically from a measurement of the magnetic field dependence of the hyperfine splitting in hydrogen by means of a maser [@winkler1972magnetic; @KB]. The precision achieved by that experiment is 1 part in $10^8$, limited by the “wall shift” due to the interaction of the hydrogen atoms with the maser cavity.\
Our experiment aims at a direct measurement of the magnetic moment of one individual proton stored in a Penning trap [@CCR; @Klaus]. The particle is confined to a volume below 100$\,\mu$m$^3$ by means of electrostatic and magnetostatic fields. Thus, the experiment is exceptionally “clean” in the sense that interactions with the apparatus can be neglected, no electron is involved, and thus, no theoretical corrections are required. It should be possible to improve the accuracy of the proton *g*-factor by one order of magnitude, at least.\
This opens up a highly exciting perspective; the same techniques can also be used to measure the magnetic moment of the antiproton [@quint1993mma], which is presently known at a level of $10^{-3}$, only [@Pask]. Thus, an improvement by six orders of magnitude is possible, which will represent a very stringent test of the matter-antimatter symmetry (*CPT*-symmetry) in the baryon sector [@Bluhm].\
The *g*-factor determination reduces to the measurement of two frequencies, $g_\text{p}=2\nu_\text{L}/\nu_\text{c}$. Both frequencies, the free cyclotron frequency $\nu_c=e B/(2\pi m_\text{p})$, and the Larmor frequency $\nu_L=g_\text{p} e B/(4\pi m_\text{p})$, are defined by the charge-to-mass ratio of the proton, and the magnetic field $B$ of the Penning trap. The free cyclotron frequency is obtained via the relation $\nu_\text{c}^2={\nu_+^2+\nu_-^2+\nu_z^2}$, where $\nu_+, \nu_-$ and $\nu_z$ are the respective eigenfrequencies of the single proton stored in the Penning trap [@brown1986gtp]. The measurement of the Larmor frequency $\nu_\text{L}$ is based on the continuous Stern-Gerlach effect [@dehmelt]. This elegant scheme for the non-destructive determination of the spin direction has been used with great success in measurements of the *g*-factor of the electron [@Hanneke], the positron [@vandyckjr1987nhp] and of bound electrons [@Hermanspahn; @Verdu]. All these experiments dealt with magnetic moments on the level of the [*[Bohr magneton]{}*]{}. A measurement of the magnetic moment of the proton is much more challenging, because it is on the level of the [*[nuclear magneton]{}*]{}, which is about three orders of magnitude smaller.\
In this Letter we demonstrate the detection of spin flips of a single proton stored in a Penning trap, which is the most important milestone for our experiment.\
A double Penning trap [@CCR]
![Schematic of the proton *g*-factor experiment. The system consists of two Penning traps which are connected by transport electrodes. The central ring electrode of the analysis trap is made of ferromagnetic material. The lower graph shows the magnitude of the magnetic field along the *z*-axis. For further details see text.[]{data-label="fig:TrapDetect"}](Fig1.eps){width="9cm"}
consisting of a precision trap (PT) and an analysis trap (AT) is mounted in the horizontal bore of a superconducting magnet with a magnetic field of $1.89\,$T. Both traps are of cylindrical and compensated design [@gabrielse1989oep]. The precision trap has an inner radius of $3.5\,$mm and is located in the homogeneous center of the magnetic field. The ring electrode of the analysis trap has an inner radius of $1.8\,$mm and is made out of ferromagnetic Co/Fe-material which distorts the homogeneous magnetic field. Near the center of the trap the magnetic field is approximately described by $$\begin{aligned}
\vec{B}=B_0 \vec{e}_z+B_2\left((z^2-\rho^2/2)\vec{e}_z- z\rho\vec{e}_\rho\right)~.
\label{eq:Bottle}
\end{aligned}$$ The “magnetic bottle” coefficient $B_2=3.00(10)\cdot10^5\,$T$\cdot$m$^{-2}$ has been determined by shifting a single proton through the trap and measuring its cyclotron frequency. $B_2$ is $3.8$ times larger than that used in a competing experiment [@Nick] and is the highest magnetic field inhomogeneity ever superimposed to a Penning trap.\
The two traps are connected by transport electrodes. The inner radius reduces along these electrodes to match the different trap radii, as shown in Fig. \[fig:TrapDetect\]. Voltage ramps are applied to the electrodes of the transport section to move the proton adiabatically between both Penning traps. The whole electrode stack is mounted in a sealed vacuum chamber and cooled to $3.8\,$K with a pulse tube cooler. Due to cryo-pumping the background pressure in such systems is $<10^{-14}\,$Pa. We can store a proton in our trap for arbitrarily long times (months).\
Protons are produced in the precision trap by electron bombardment of a polyethylene-target and subsequent electron impact ionization. Impurity ions are selectively heated out of the trap using broadband noise and resonant radio frequency (rf) drives. Particles are then removed from the clean proton cloud until just one single particle remains. The trapped proton oscillates at three different eigenmotions [@brown1986gtp], the modified cyclotron motion at $\nu_\text{+,PT}\approx28.97\,$MHz, the axial motion at $\nu_\text{z}\approx674\,$kHz, and the magnetron motion at $\nu_\text{-,PT}\approx 8\,$kHz. The oscillation of the particle induces image currents in the trap electrodes which are detected as rf voltages across a resistance $R$. In practice a coil is used instead of a resistor. The inductance together with the stray capacitance of the trap system forms a parallel tuned circuit with a quality factor *Q*. The axial frequency of the stored proton in the precision trap and in the analysis trap is measured with a superconducting Nb/Ti coil connected to both traps. The quality factor is about $5700$ and the effective on-resonance resistance is $36\,$M$\Omega$ [@ulmer2009quality]. The detection system for the measurement of the modified cyclotron frequency in the precision trap (not shown in Fig. \[fig:TrapDetect\]) is made out of copper. It has a quality factor of $1250$ and an on-resonance resistance of $380\,$k$\Omega$. The voltage signals induced by the proton are amplified with cryogenic low-noise amplifiers and analyzed with an FFT spectrum analyzer. The magnetron frequency is measured by sideband coupling [@Cornell] to the axial eigenmotion. The proton interacts with the tank circuits and thus cools resistively to ambient cryogenic temperatures. The magnetic field in the center of the analysis trap is $1.17\,$T due to the field distortion by the ferromagnetic ring electrode. The cyclotron frequency of the proton is measured as described in [@Hermanspahn] and is $\nu_\text{+,AT}\approx17.91\,$MHz. The magnetron frequency in the analysis trap is $\nu_\text{-,AT}\approx13\,$kHz.\
In an ideal Penning trap the axial frequency $\nu_z$ of a trapped proton is $$\begin{aligned}
\nu_{z,0}=\frac{1}{2\pi}\sqrt{\frac{e}{m_p}\frac{d}{dz}\left(-\vec\nabla\Phi_\text{E}|\vec{e}_z\right)}=\frac{1}{2\pi}\sqrt{\frac{e V_0}{m_pd^2}}~,
\end{aligned}$$ where $d$ is a characteristic length of the trap, $V_0$ is the trapping voltage and $e\Phi_\text{E}$ is the electrostatic energy. In the analysis trap with its strong magnetic bottle the magnetostatic energy $\Phi_\text{M}=- \vec\mu \cdot \vec B$ contributes to the axial frequency. Here $\vec \mu$ is the sum of the Landau-magnetic moment, which is due to the angular momenta of the cyclotron and the magnetron motion, respectively, and the magnetic moment of the proton spin $\mu_s=(g_\text{p}e\hbar m_s)/(2m_\text{p})$, where $m_s$ is the spin quantum number. The energy $\Phi_\text{M}$ contains a term proportional to $z^2$ where $z$ is the axial distance to the center of the trap. The axial frequency thus becomes $\nu_\text{z}=\nu_\text{z,0}+\Delta\nu_\text{z}(n_+,n_-,m_s)$, where $n_\pm$ denotes the number of quanta in the radial modes, and $$\begin{aligned}
\Delta\nu_z(n_+,n_-,m_s)&=&\nonumber\\
\frac{h \nu_+}{4\pi^2 m_\text{p}\nu_z}\frac{B_2}{B_0}&\cdot&\left(n_++\frac{1}{2}+\frac{\nu_-}{\nu_+}\left(n_-+\frac{1}{2}\right)+\frac{g_\text{p}m_s}{2}\right)~.
\label{eq:Bottleshift}
\end{aligned}$$ A cyclotron quantum jump $\Delta n_+=\pm 1$ corresponds to a radial energy change of $\Delta E_+=\pm74\,$neV and causes an axial frequency shift of $\Delta\nu_z=\pm68\,$mHz. A transition of the magnetron quantum number $\Delta n_-=\pm 1$ leads to $\Delta\nu_z=\pm49\,\mu$Hz. Changes of both radial quantum numbers $\Delta n_\pm=\pm 1$ are due to electric dipole transitions. A magnetic-dipole spin-flip transition $\Delta m_s = \pm 1$ causes a jump of the axial frequency $\Delta\nu_\text{z,SF}=\pm190\,$mHz. This is most important because the axial frequency can thus be used to detect the spin direction of the proton.\
Spin-flip transitions are driven using a disc coil mounted close to the electrode stack as shown in Fig. \[fig:TrapDetect\]. The coil generates a transverse magnetic rf-field $\vec{b}_\text{rf}$ with frequency $\nu_\text{rf}$. To drive spin flips $\vec{b}_\text{rf}$ is tuned near to the Larmor frequency $\nu_\text{L}=g_\text{p}\nu_c/2$. The field penetrates into the trap through slits between the electrodes. This rf field causes a precession of the proton spin around the $\vec{b}_\text{rf}$-axis at the Rabi frequency $\Omega_R/2\pi=\nu_\text{L} b_\text{rf}/B_0$. In the presence of a strong magnetic bottle $B_2$ the spin-flip probability is given by [@Brown2] $$\begin{aligned}
P_\text{SF}=\frac{1}{2}\left(1-\exp\left(-\frac{1}{2}\Omega_R^2 t_0~\chi(2\pi\nu_\text{rf},B_2,T_z)\right)\right)~,
\label{eq:SFProb}
\end{aligned}$$ where $t_0$ is the irradiation time, $\chi(2\pi\nu_\text{rf},B_2,T_z)$ the transition lineshape and $T_z$ the axial temperature.\
The axial frequency $\nu_z$ has a strong dependence on the radial energy $E_\rho=|E_+|+|E_-|$ with $(\Delta\nu_z)/(\Delta E_\rho) \approx0.92\,$Hz/$\mu$eV, where $|E_\pm|$ is the energy of the cyclotron mode and the magnetron mode, respectively. It is thus essential to avoid induced spurious changes of the radial energy of the proton. Care is taken to shield the apparatus from perturbing rf-signals and noise. In the lines for the spin-flip drive a chain of band-pass filters attenuates spurious rf-fields at the cyclotron frequency by $100\,$dB. The line is shorted to ground with an rf-relay whenever possible.\
![Spectrum of the axial detector. The peak in the upper graph is the thermal noise spectrum of the cryogenic tank circuit. Its width is $\nu_z/Q\approx 120\,$Hz. The frequency axis is stretched in the lower graph to show the narrow dip, which is due to the trapped proton shorting the noise of the detector.[]{data-label="fig:CauseRes"}](Fig2.eps){width="6.5cm"}
The frequency $\nu_z$ is measured by observing the thermal noise spectrum of the cryogenic axial tank circuit, see Fig. \[fig:CauseRes\]. The particle acts like a series tuned circuit, shorting the thermal voltage-noise $e_\text{th}=\sqrt{4k_\text{B} T_z R\Delta F}$ of the detector, where $\Delta F$ is the FFT spectrum analyzer bandwidth. The linewidth of the dip is $1/(2\pi\tau_z)$, where $\tau_z=(m_\text{p}/R)\cdot(D^2/q^2)$ is the cooling time constant and $D$ is a characteristic length of the trap. In the precision trap the signal linewidth is $1.25\,$Hz and in the analysis trap $3.5\,$Hz due to the smaller trap-size. The detection system has a signal-to-noise ratio of 16$\,$dB.\
To determine the spin direction of the proton, the axial frequency $\nu_z$ has to be measured with high resolution. The axial frequency resolution achieved by noise-dip detection increases with measuring time $\Delta t$ due to spectrum analyzer averaging. On the other hand, tiny perturbing effects can drive the radial modes and cause small changes of the axial frequency which increase with time. It was found in a series of experiments that fluctuations of the cyclotron energy at an average rate of 0.045 quantum jumps per second are responsible for the residual instability of the axial motion. Details of these experiments will be discussed in a forthcoming publication.\
In the following discussion the change of the axial frequency between time $t$ and $t+\Delta t$ is considered and the standard deviation of $\alpha=\nu_z(t)-\nu_z(t+\Delta t)$ is considered. We quantify the frequency fluctuation by using the standard deviation of $\alpha$, namely $\Xi=\left((N-1)^{-1}\sum(\alpha_i-\bar\alpha)^2\right)^{1/2}$. In our experiment at an FFT averaging time of $\Delta t_\text{opt}=80\,$s the frequency fluctuation is minimal, $\Xi_\text{min}\approx150\,$mHz. This value is not yet sufficiently stable to detect spin flips directly in one measurement sequence as described in [@Hartmut]. However, averaging can be used to detect proton spin flips.\
For such a statistical measurement the axial frequency $\nu_z$ is determined in sequences of three measurements $\nu_{z,1}$, $\nu_{z,2}$, and $\nu_{z,3}$. Between the first and the second measurement a spin-flip drive is turned on near the Larmor frequency $\nu_\text{L}\approx\nu_\text{rf}=50.102\,$MHz. Between the second and the third measurement the rf-synthesizer is tuned to a reference frequency 100$\,$kHz below $\nu_\text{L}$ which corresponds to approximately one linewidth. Between the third measurement and the first measurement of the following cycle $\nu_{z,1'}$ no rf-signal is applied to the trap. This sequence is repeated several hundred times and the standard deviations of the frequency differences $\Xi_\text{SF}=\nu_{z,2}-\nu_{z,1}$, $\Xi_\text{ref}=\nu_{z,3}-\nu_{z,2}$ and $\Xi_\text{back}=\nu_{z,1'}-\nu_{z,3}$ are computed. If spin flips are driven the corresponding axial frequency shifts add up in a statistical way. If the spin of the proton flips in *M* cycles out of *N*, the total fluctuation is $$\begin{aligned}
\Xi_\text{SF}&=&\sqrt{\sum_i^M\frac{\left(\alpha_i\pm\Delta\nu_{z,SF}-\bar\alpha\right)^2}{N-1}
+\sum_{i=M}^N\frac{\left(\alpha_i-\bar\alpha\right)^2}{N-1}}\nonumber\\
&\approx&\sqrt{\Xi_{back}^2+P_{SF}\Delta\nu_{z,SF}^2}~,
\label{eq:SFBack}
\end{aligned}$$ where $P_\text{SF}$ is given by Eq. (\[eq:SFProb\]). The comparison of $\Xi_\text{ref}$ and $\Xi_\text{back}$ is a means to test for spurious heating of the cyclotron motion due to rf-synthesizer noise, which might increase the axial frequency fluctuation. If both quantities are the same within error bars this indicates that the cyclotron motion is not affected and that an increase of $\Xi$ is due to proton spin flips. A resolution limit for the detection of spin flips by this method is reached if the error $\sigma_{\Delta\Xi}=((\Xi_\text{ref}/\sqrt{2N-2})^2+(\Xi_\text{SF}/\sqrt{2N-2})^2)^{0.5}$ is of the same size as the increase of the frequency fluctuation $\Delta\Xi=\Xi_\text{SF}-\Xi_\text{ref}$. With the conditions of our experiment for $P_\text{SF}=50\,\%$, $\Delta\Xi\approx50\,$mHz are expected. This shows that it is possible to statistically detect spin flips in a series of only fifty measurement cycles.\
![Proton spin flip data. In a.) the evolution of the frequency fluctuations $\Xi_\text{SF}$ and $\Xi_\text{ref}$ as a function of measurement cycles is shown. The clear separation of the two lines shows that spin flips are detected. An alternative analysis is shown in b.). The entire data sequence was split into subsets and the values of $\Xi_\text{SF}$ and $\Xi_\text{ref}$ were binned to two histograms.[]{data-label="fig:Result"}](Fig3.eps){width="8.0cm"}
In the experiment the measurement sequence is repeated several hundred times and the frequency fluctuations $\Xi_\text{SF}$ and $\Xi_\text{ref}$ are determined. The evolution of both quantities as a function of measurement cycles is shown in Fig. \[fig:Result\] a.) along with a confidence band given by $\sigma_{\Delta\Xi}$. Both fluctuations converge for large measuring times to nearly constant values. A significant shift of $\Delta\Xi=47\,$mHz between the reference measurement $\Xi_\text{ref}$ and the resonant measurement $\Xi_\text{SF}$ is observed. This means that spin flips are detected in the experiment. Using Eq. (\[eq:SFBack\]) we find an approximately saturated spin-flip probability of $P_\text{SF}=47\pm7\,\%$. The data set has also been analyzed in a different way. The whole sequence of 800 data points is split into subsequences of 50 measurements. For every subsequence $\Xi_\text{SF}$ and $\Xi_\text{ref}$ are determined and binned into histograms. Figure \[fig:Result\] b.) shows a distinct splitting between both distributions. This clearly shows, again, that proton spin flips are detected.\
A spin flip resonance of the proton in the analysis trap has been measured by tuning $\nu_\text{rf}$ in steps across the Larmor frequency $\nu_\text{L}$. For every frequency point the data are analyzed as above and the corresponding spin-flip probability $P_\text{SF}$ is determined. In the magnetic bottle the Larmor frequency $\nu_\text{L}$ is a function of the particle position $z$. Due to the axial oscillation $z(t)$ the spin-flip resonance curve has a broad width. It is asymmetric due to the Boltzmann distribution of the axial energy. Figure \[fig:SFProb\] presents the spin-flip resonance. The solid line is the best fit of Eq. (\[eq:SFProb\]) to the data. Fixed parameters are the irradiation time $t_0=10\,$s, the magnetic bottle strength $B_2=3.00(10)\cdot10^5\,$T$\cdot$m$^{-2}$, and the axial temperature of $T_z=9.5\,$K. $B_2$ and $T_z$ have been measured independently. Free fit parameters are the amplitude $b_\text{rf}$ of the magnetic rf-field, which comes out as $2.5\,\mu$T, and the Larmor frequency $\nu_\text{L}$. From the fit the Larmor frequency can be determined with a precision of 2$\cdot$10$^{-4}$. The result is consistent with the independent measurement of the magnetic field in the analysis trap using the proton’s cyclotron frequency. The agreement between the data and the fit for the variation of the spin-flip probability with the drive frequency provides convincing evidence that proton spin flips were detected.\
![Proton spin flip resonance in the analysis trap with its inhomogeneous magnetic field.[]{data-label="fig:SFProb"}](Fig4.eps){width="6.5cm"}
In the future we plan to drive the spin flip in the precision trap and to transport the particle into the analysis trap to analyze its spin direction [@Hartmut]. In the precision trap the magnetic field is four orders of magnitude more homogeneous than in the analysis trap and a narrow spin-flip resonance is expected. Together with a measurement of the cyclotron frequency this will be a direct high-precision measurement of the *g*-factor of the proton.\
In conclusion, spin flips of a single trapped proton have been observed for the first time. A strong magnetic bottle $B_2$ has been used to couple the proton spin to its axial motion in a Penning trap. The observation of tiny changes of the axial frequency clearly indicates proton spin flips. A Larmor resonance was clearly observed by averaging. This is an important milestone towards a direct high-precision measurement of the magnetic moment of the proton and the antiproton.\
We acknowledge the support of the Max-Planck Society, the BMBF, the DFG (QU-122-3), the Helmholtz-Gemeinschaft, HGS-HIRE, Al$\beta$an (E06D101305BR) and the IMPRS-QD.
[99]{} P. F Winkler, D. Kleppner, T. Myint, and F. G. Walther, Phys. Rev. A **5**, 83 (1972). S. G. Karshenboim and V. G. Ivanov, Phys. Lett. B **566**, 27 (2003). C. C. Rodegheri *et al.*, Hyperfine Interact. **194**, 93 (2009). K. Blaum *et al.*, J. Phys. B: At. Mol. Opt. Phys. **42**, 154021 (2009). W.Quint and G. Gabrielse, Hyperfine Interact. **76**, 379 (1993). T. Pask *et al.*, J. Phys. B **41**, 081008 (2008). R. Bluhm, V. A. Kostelecky, and N. Russell, Phys. Rev. D **57**, 3932 (1998). L.S. Brown and G. Gabrielse, Rev. Mod. Phys. **58**, 233 (1986). H. Dehmelt and P. Ekstrom, Bull. Am. Phys. Soc. **18**, 72 (1973). D. Hanneke, S. Fogwell, and G. Gabrielse, Phys. Rev. Lett. **100**, 120801 (2008). R. S. Van Dyck, Jr., P. B. Schwinberg and H. G. Dehmelt, Phy. Rev. Lett. **59**, 26 (1987). N. Hermanspahn *et al.*, Phys. Rev. Lett. **84**, 427 (2000). J. Verdu *et al.*, Phys. Rev. Lett. **92**, 093002 (2004). G. Gabrielse, L. Haarsma, and S. L. Rolston, Int. J. Mass Spec. **88**, 319 (1989). N. Guise, J. DiSciacca, and G. Gabrielse, Phys. Rev. Lett. **104**, 143001 (2010). S. Ulmer *et al.*, Rev. Sci. Inst. **80**, 123302 (2009). E. A. Cornell, R. M.Weisskoff, K. R. Boyce, and D. E.Pritchard, Phys. Rev. A, **41**, 312 (1990). L. S. Brown, Ann. Phys. **159**, 62 (1985) H. Häffner *et al.*, Phys. Rev. Lett. **85**, 5308 (2000).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Over an algebraically closed field, various finiteness results are known regarding the automorphism group of a K3 surface and the action of the automorphisms on the Picard lattice. We formulate and prove versions of these results over arbitrary base fields, and give examples illustrating how behaviour can differ from the algebraically closed case.'
address:
- |
Mathematisch Instituut\
Niels Bohrweg 1\
2333 CA Leiden\
Netherlands
- 'The Tutte Institute for Mathematics and Computation, P.O. Box 9703, Terminal, Ottawa, ON K1G 3Z4, Canada'
- 'School of Mathematics and Statistics, 4302 Herzberg Laboratories, 1125 Colonel By Drive, Carleton University, Ottawa, ON K1S 5B6, Canada'
- |
Mathematisch Instituut\
Niels Bohrweg 1\
2333 CA Leiden\
Netherlands
author:
- Martin Bright
- Adam Logan
- Ronald van Luijk
bibliography:
- 'k3q.bib'
title: Finiteness results for K3 surfaces over arbitrary fields
---
Introduction
============
The geometry of K3 surfaces over the complex numbers has a long history, with many results known about the cohomology, the Picard group, and the automorphism group of an algebraic K3 surface, and how these objects interact. Such results over the complex numbers carry over to other algebraically closed fields of characteristic zero, and similar results are also known over algebraically closed fields of other characteristics. For a comprehensive treatment of the geometry of K3 surfaces, we refer the reader to the lecture notes of Huybrechts [@huybrechts]. Much of the theory we use was originally developed by Nikulin [@nikulin80].
K3 surfaces are also interesting from an arithmetic point of view, with much recent work on understanding the rational points, curves, Brauer groups and other invariants of K3 surfaces over number fields. In this article, we investigate the extent to which some standard finiteness results for K3 surfaces over algebraically closed fields remain true over more general base fields. In particular, we show how to define the correct analogue of the Weyl group, and give an explicit description of it. This allows us to formulate and prove finiteness theorems over arbitrary fields, modelled on those already known over algebraically closed fields. The tools we use include representability of the Picard and automorphism schemes, classification of transitive group actions on Coxeter–Dynkin diagrams, and an explicit description of the walls of the ample cone. We follow these theoretical results with several detailed examples, showing how the relationship between the Picard group and the automorphism group can be different from the geometric case. We end by proving that a surface over ${\mathbb{Q}}$ of the form $x^4-y^4 = c(z^4-w^4)$ has finite automorphism group for $c \in {\mathbb{Q}}^*$ that is not in the subgroup generated by squares together with $-1, 2$.
The specific finiteness results we address go back to Sterk [@sterk]. To state them, we need some definitions; we follow the notation of [@huybrechts]. In this article, by a K3 surface we will always mean an algebraic K3 surface, which is therefore projective.
Let $k$ be a field, and let $X$ be a K3 surface over $k$. Denote the group of isometries of $\operatorname{Pic}X$ by ${{\mathrm{O}}}(\operatorname{Pic}X)$. Reflection in any $(-2)$-class in $\operatorname{Pic}X$ defines an isometry of $\operatorname{Pic}X$, and we define the *Weyl group* $W(\operatorname{Pic}X) \subset {{\mathrm{O}}}(\operatorname{Pic}X)$ to be the subgroup generated by these reflections.
We recall the definitions of the positive, ample and nef cones associated to the K3 surface $X$; see also [@huybrechts Chapter 8]. Let $(\operatorname{Pic}X)_{\mathbb{R}}$ denote the real vector space $(\operatorname{Pic}X) \otimes_{\mathbb{Z}}{\mathbb{R}}$. By the Hodge index theorem, the intersection product on $(\operatorname{Pic}X)_{\mathbb{R}}$ has signature $(1, \rho-1)$; so the set $\{ \alpha \in (\operatorname{Pic}X)_{\mathbb{R}}\mid \alpha^2 > 0 \}$ consists of two connected components. The *positive cone* ${{\mathcal{C}}}_X \subset (\operatorname{Pic}X)_{\mathbb{R}}$ is the connected component containing all the ample classes.
The *ample cone* $\operatorname{Amp}(X)$ is the cone in $(\operatorname{Pic}X)_{\mathbb{R}}$ generated by all classes of ample line bundles. The *nef cone* $\operatorname{Nef}(X)$ is defined as $$\operatorname{Nef}(X) = \{ \alpha \in (\operatorname{Pic}X)_{\mathbb{R}}\mid \alpha \cdot C \ge 0 \text{ for all curves } C \subset X \}.$$ Finally, we define $\operatorname{Nef}^e(X)$ to be the real convex hull of $\operatorname{Nef}(X) \cap \operatorname{Pic}X$. An application of the criterion of Nakai–Moishezon–Kleiman shows that $\operatorname{Amp}(X)$ is the interior of $\operatorname{Nef}(X)$ and $\operatorname{Nef}(X)$ is the closure of $\operatorname{Amp}(X)$: see [@huybrechts Corollary 8.1.4].
The following finiteness theorems are due to Sterk [@sterk] for $k={\mathbb{C}}$ and to Lieblich and Maulik [@lm] when $k$ has positive characteristic not equal to $2$. As in Huybrechts [@huybrechts Chapter 8], a *fundamental domain* for the action of a discrete group $G$ acting continuously on a topological manifold $M$ is defined as the closure $\overline{U}$ of an open subset $U\subset M$ such that $M = \cup_{g \in G} g \overline{U}$ and such that for $g \neq h \in G$ the intersection $g\overline{U} \cap h\overline{U}$ does not contain interior points of $gU$ or $hU$.
\[thm:old\] Let $k = {\bar{k}}$ be an algebraically closed field of characteristic not equal to $2$, and let $X$ be a K3 surface over $k$.
1. [@huybrechts Corollary 8.2.11] The cone $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$ is a fundamental domain for the action of $W(\operatorname{Pic}X) \subset {{\mathrm{O}}}(\operatorname{Pic}X)$ on the positive cone ${{\mathcal{C}}}_X$.
2. \[it2\] [@huybrechts Theorem 15.2.6], [@lm Proposition 5.2] The subgroup $W(\operatorname{Pic}X)$ is normal in ${{\mathrm{O}}}(\operatorname{Pic}X)$; the natural map $\operatorname{Aut}X \to {{\mathrm{O}}}(\operatorname{Pic}X)/W(\operatorname{Pic}X)$ has finite kernel and image of finite index.
3. [@huybrechts Theorem 8.4.2] The action of $\operatorname{Aut}X$ on $\operatorname{Nef}^e X$ admits a rational polyhedral fundamental domain.
4. [@huybrechts Corollary 8.4.6] The set of orbits under $\operatorname{Aut}X$ of $(-2)$-curves on $X$ is finite. More generally, for any $d$ there are only finitely many orbits under $\operatorname{Aut}X$ of classes of irreducible curves of self-intersection $2d$.
In Section \[sec:ft\] we will prove analogues of the various statements of Theorem \[thm:old\] when $k$ is replaced by an arbitrary base field of characteristic different from $2$.
A consequence of Theorem \[thm:old\](\[it2\]) is that the finiteness of $\operatorname{Aut}X$ depends only on the lattice $\operatorname{Pic}X$. Those possible Picard lattices for which ${{\mathrm{O}}}(\operatorname{Pic}X)/W(\operatorname{Pic}X)$ is finite have been classified [@nikulin]. Over an arbitrary base field we will see that, instead of using the Weyl group $W(\operatorname{Pic}X)$, we must use the Galois-invariant part of the geometric Weyl group. This means that the finiteness of $\operatorname{Aut}X$ is no longer determined purely by the Picard lattice $\operatorname{Pic}X$; rather, it depends on the geometric Picard lattice together with the Galois action. In Section \[sec:eg\], we give several examples that illustrate this difference to the classical case.
We thank the referee, whose thorough report helped us improve the exposition and correct a number of errors in the paper. The second author would like to thank the Tutte Institute for Mathematics and Computation for its partial support for a visit to the University of Leiden during which much of this research was done.
Lemmas on lattices
==================
In this section we will study lattices with the action of a group. Given a lattice ${\Lambda}$ with the action of a finite group $H$, we consider the group of automorphisms that commute with $H$, and the group of automorphisms that preserve the sublattice fixed by $H$.
A *lattice* ${\Lambda}$ is a free abelian group of finite rank with a nondegenerate integer-valued symmetric bilinear form. If the form is (positive or negative) definite, we likewise refer to ${\Lambda}$ as *definite*. The group of automorphisms of ${\Lambda}$ preserving the form is denoted ${{\mathrm{O}}}({\Lambda})$. A *sublattice* of ${\Lambda}$ is a subgroup on which the restriction of the form is nondegenerate. Given a sublattice $M \subseteq {\Lambda}$, we use ${{\mathrm{O}}}({\Lambda},M)$ for the subgroup of ${{\mathrm{O}}}({\Lambda})$ fixing $M$ as a set. For a subgroup $H \subseteq {{\mathrm{O}}}({\Lambda})$, let ${\Lambda}^H$ be the subgroup of ${\Lambda}$ consisting of the elements fixed by every element of $H$ (note that according to our conventions ${\Lambda}^H$ may not be a lattice, because the quadratic form on ${\Lambda}$ may be degenerate when restricted to ${\Lambda}^H$). The vector space ${\Lambda}\otimes_{\mathbb{Z}}{\mathbb{Q}}$ will be denoted $\Lambda_{\mathbb{Q}}$.
Our goal is to prove the following proposition.
\[lattice-prop\] Let ${\Lambda}$ be a lattice and $H \subseteq {{\mathrm{O}}}({\Lambda})$ a subgroup such that $M = {\Lambda}^H$ is a lattice. Then the following hold:
1. the natural map ${{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M)$ has image of finite index;
2. suppose that $M^\perp$ is definite. Then ${{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M)$ has finite kernel, and the centralizer $Z_{{{\mathrm{O}}}({\Lambda})} H$ is a finite-index subgroup of ${{\mathrm{O}}}({\Lambda},M)$.
Our interest in this situation arises from the geometry of K3 surfaces. Let $X$ be a projective K3 surface defined over a field $F$, and let $K/F$ be a Galois extension. Let ${\Lambda}= \operatorname{Pic}X_K$ with the intersection pairing, and let $H$ be the image of $\operatorname{Gal}(K/F)$ in ${{\mathrm{O}}}({\Lambda})$. If $X$ has a rational point over $F$, we have $\operatorname{Pic}X_F = {\Lambda}^H$. (See Section \[sec:ft\] for more details. This statement holds slightly more generally: for example, if $F$ is a number field and $X$ has points everywhere locally over $F$.) The Hodge index theorem states that ${\Lambda}$ has signature $(1,n)$ and ${\Lambda}^H$ has signature $(1,m)$: therefore $M^\perp$ is definite.
Before giving the proof we first collect a few helpful statements, which are probably well known.
\[sublemma\] Let ${\Lambda}$ be a lattice and $M$ a sublattice. Let $H$ be a subgroup of ${{\mathrm{O}}}({\Lambda})$ such that ${\Lambda}^H$ is a lattice. For groups $G' \subset G$, let $Z_G(G')$ denote the centralizer of $G'$ in $G$.
1. \[Mperp\] $M^\perp$ is a sublattice of ${\Lambda}$.
2. \[Operp\] There is a natural injection $d: {{\mathrm{O}}}({\Lambda},M) \hookrightarrow {{\mathrm{O}}}(M) \oplus {{\mathrm{O}}}(M^\perp)$ with image of finite index.
3. \[cent1\] $Z_{{{\mathrm{O}}}({\Lambda})} H$ is contained in ${{\mathrm{O}}}({\Lambda},{\Lambda}^H)$.
4. \[cent2\] The kernel of the map ${{\mathrm{O}}}({\Lambda},{\Lambda}^H) \to {{\mathrm{O}}}({{\Lambda}^H}^\perp)$ is contained in $Z_{{{\mathrm{O}}}({\Lambda})} H$.
To prove (\[Mperp\]), we just have to prove that the pairing on the subspace $M^\perp_{\mathbb{Q}}\subset \Lambda_{\mathbb{Q}}$ is non-degenerate; this is [@Eichler Satz 1.2].
Let $\phi \in {{\mathrm{O}}}({\Lambda},M)$. By definition $\phi$ restricts to an endomorphism of $M$. Let $S_M$ be the saturation of $M$ in ${\Lambda}$. Clearly $\phi(S_M) \subseteq S_M$. Now, $\phi^{-1}(S_M)$ has the same rank as $S_M$ and contains $S_M$, so it is equal to $S_M$. So if $\phi(S_M) \ne S_M$, then the image of $\phi$ does not contain $S_M$, contradicting the hypothesis that $\phi$ is an automorphism of ${\Lambda}$. Since $M$ is a subgroup of finite index of $S_M$, this implies that $\#(S_M/M) =
\#(S_M/\phi(M))$. But $\phi(M) \subseteq M$, so it follows that $\phi(M) = M$. Thus there is a map ${{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M)$. Now, if $\phi \in {{\mathrm{O}}}({\Lambda},M)$ and $y \in M^\perp$, then $m \cdot \phi(y) =
\phi^{-1}(m) \cdot y = 0$ for all $m \in M$, so $\phi(y) \in M^\perp$, and we get a map ${{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M^\perp)$ in the same way. Combining these two maps gives a map $d: {{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M) \oplus {{\mathrm{O}}}(M^\perp)$.
If $d(\phi) = 1$, then $d$ is the identity on $M \oplus M^\perp$, which is a subgroup of ${\Lambda}$ of finite index. Because ${\Lambda}$ is torsion-free, this forces $\phi$ to be the identity. To show that $\operatorname{im}d$ has finite image in $(O(M) \oplus O(M^\perp))$, et $n$ be the smallest positive integer such that $n{\Lambda}\subseteq M \oplus M^\perp$, and let $k = [M \oplus M^\perp:n{\Lambda}]$. Every element of ${{\mathrm{O}}}(M) \oplus {{\mathrm{O}}}(M^\perp)$ that fixes $n{\Lambda}$ as a set is in the image of $d$, because the induced automorphism of $n{\Lambda}$ extends to an automorphism of ${\Lambda}$ with the same action on $M \oplus M^\perp$. Since there are only finitely many subgroups of index $k$ in $M \oplus M^\perp$, the stabilizer of $n{\Lambda}$ is of finite index, and we have proved (\[Operp\]).
To prove (\[cent1\]), let $\phi \in Z_{{{\mathrm{O}}}({\Lambda})} H$, and let $m \in {\Lambda}^H$ and $h \in H$. Then $h(m) = m$ and $\phi
\circ h = h \circ \phi$. So $h(\phi(m)) = \phi(h(m)) = \phi(m)$, establishing that $\phi(m) \in {\Lambda}^H$.
Finally we prove (\[cent2\]). Choose $\phi$ in the kernel and $h \in H$, and let $x \in {\Lambda}$. We will view $\phi$ and $h$ as automorphisms of ${\Lambda}_{\mathbb{Q}}$. In ${\Lambda}_{\mathbb{Q}}$ we may write $x = x_1 + x_2$, where $x_1 \in {\Lambda}^H_{\mathbb{Q}}$ and $x_2 \in ({{\Lambda}^H_{\mathbb{Q}}})^\perp$. Then $h(\phi(x)) = h(\phi(x_1 + x_2)) = h(\phi(x_1)) + h(\phi(x_2)).$ However, $\phi(x_1) \in {\Lambda}^H_{\mathbb{Q}}$, so $h(\phi(x_1)) = \phi(x_1)$, and $\phi$ is in the kernel of the map to ${{\mathrm{O}}}({{\Lambda}^H}^\perp)$, so $\phi(x_2)
= x_2$. It follows that $h(\phi(x)) = \phi(x_1) + h(x_2)$. Similarly, $\phi(h(x)) = \phi(h(x_1+x_2)) = \phi(h(x_1)) +
\phi(h(x_2)) = \phi(x_1) + h(x_2) = h(\phi(x))$, establishing that $\phi$ commutes with $h$.
The map ${{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M)$ is a composition ${{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M) \oplus {{\mathrm{O}}}(M^\perp) \to {{\mathrm{O}}}(M)$. In part (2) of the lemma just proved we showed that the first map has image of finite index. The second map is surjective, so the composition has image of finite index as well.
We now suppose that $M^\perp$ is definite to prove the second statement. Then ${{\mathrm{O}}}(M^\perp)$ is finite, so ${{\mathrm{O}}}({\Lambda},M) \to {{\mathrm{O}}}(M)$ is a composition of an injective map with a map with finite kernel and so its kernel is finite. Let $K = \ker({{\mathrm{O}}}({\Lambda},M)
\to {{\mathrm{O}}}(M^\perp))$. Then $Z_{{{\mathrm{O}}}({\Lambda})} H/K$ has finite index in ${{\mathrm{O}}}({\Lambda},M)/K$, because both inject into the finite group ${{\mathrm{O}}}(M^\perp)$. Therefore $Z_{{{\mathrm{O}}}({\Lambda})} H$ has finite index in ${{\mathrm{O}}}({\Lambda},M)$ too.
Finiteness results for K3 surfaces {#sec:ft}
==================================
In this section we formulate and prove analogues of the statements of Theorem \[thm:old\] when $k$ is an arbitrary field. We first look at the case of $k$ separably closed, which is straightforward.
\[lem:sep\] Let $k = {k^s}$ be a separably closed field, and let ${\bar{k}}$ be an algebraic closure of $k$. Let $X$ be a K3 surface over $k$, and let ${{\bar{X}}}$ be the base change of $X$ to ${\bar{k}}$. Then the natural maps $\operatorname{Pic}X \to \operatorname{Pic}{{\bar{X}}}$ and $\operatorname{Aut}X \to \operatorname{Aut}{{\bar{X}}}$ are isomorphisms.
As $X$ is projective, the Picard scheme ${\mathbf{Pic}}_{X/k}$ exists, is separated and locally of finite type over $k$, and represents the sheaf $\operatorname{Pic}_{(X/k)({\textrm{{\'e}t}})}$, which is defined to be the sheafification on the big étale site over $k$ of the presheaf $$T \mapsto \operatorname{Pic}(X \times_k T) / \operatorname{Pic}T$$ (see [@FAG Theorem 9.4.8]). In particular, because $k$ and $\bar{k}$ are both separably closed, we have ${\mathbf{Pic}}_{X/k}(k) = \operatorname{Pic}X$ and ${\mathbf{Pic}}_{X/k}({\bar{k}}) = \operatorname{Pic}{{\bar{X}}}$. From ${\mathrm{H}}^1(X,{{\mathcal{O}}}_X)=0$ it follows from [@FAG Theorem 9.5.11] that ${\mathbf{Pic}}_{X/k}$ is étale over $k$. So every ${\bar{k}}$-point of ${\mathbf{Pic}}_{X/k}$ is defined over $k$, and $\operatorname{Pic}X \to \operatorname{Pic}{{\bar{X}}}$ is an isomorphism.
The functor taking a $k$-scheme $S$ to the group $\operatorname{Aut}(X \times_k S)$ is represented by a scheme ${\mathbf{Aut}}_{X/k}$: see [@FAG Theorem 5.23]. A standard argument in deformation theory shows that the tangent space at the identity element is isomorphic to ${\mathrm{H}}^0(X, T_X)$, where $T_X$ denotes the tangent sheaf on $X$. Indeed, an element of the tangent space is given by a morphism $S=\operatorname{Spec}k[\epsilon]/(\epsilon^2) \to {\mathbf{Aut}}_{X/k}$ extending the morphism sending $\operatorname{Spec}k$ to the identity automorphism. Such a morphism corresponds to an automorphism of $X \times_k S$ restricting to the identity on the central fibre. By [@FAG Theorem 8.5.9], the set of these morphisms forms an affine space under ${\mathrm{H}}^0(X,T_X)$. In our case, the group ${\mathrm{H}}^0(X, T_X)$ is zero [@huybrechts Theorem 9.5.1], so the scheme ${\mathbf{Aut}}_{X/k}$ is étale over $k$, and $\operatorname{Aut}X \to \operatorname{Aut}{{\bar{X}}}$ is an isomorphism.
\[cor:sep-2\] In the situation of Lemma \[lem:sep\], every $(-2)$-curve on ${{\bar{X}}}$ is defined over $k$.
Let $\bar{C}$ be a $(-2)$-curve on ${{\bar{X}}}$. Then Lemma \[lem:sep\] shows that there is a line bundle $L$ on $X$ whose base change to ${{\bar{X}}}$ is isomorphic to ${{\mathcal{O}}}_{{{\bar{X}}}}(\bar{C})$. The Riemann–Roch theorem and flat base change give $h^0(X,L) = h^0({{\bar{X}}},{{\mathcal{O}}}_{{{\bar{X}}}}(\bar{C})) = 1$. So all nonzero sections of $L$ cut out the same divisor $C \subset X$ and the base change of $C$ to ${{\bar{X}}}$ must coincide with $\bar{C}$. In other words, $\bar{C}$ is defined over $k$.
We now pass to the case of a general field. Let $k$ be a field; fix an algebraic closure ${\bar{k}}$ of $k$, and let ${k^s}$ be the separable closure of $k$ in ${\bar{k}}$. Let $X$ be a K3 surface over $k$, and let ${X^s}$ and ${{\bar{X}}}$ denote the base changes of $X$ to ${k^s}$ and ${\bar{k}}$, respectively. Write ${{\Gamma_k}}= \operatorname{Gal}({k^s}/k)$.
The group ${{\Gamma_k}}$ acts on $\operatorname{Pic}{X^s}$ preserving intersection numbers, giving a representation ${{\Gamma_k}}\to {{\mathrm{O}}}(\operatorname{Pic}{X^s})$. Let ${{\Gamma_k}}$ act on ${{\mathrm{O}}}(\operatorname{Pic}{X^s})$ by conjugation, that is, such that $({{{}^{\sigma} f}})(x) = \sigma(f(\sigma^{-1} x))$ for all $x \in \operatorname{Pic}{X^s}$. For a $(-2)$-class $\alpha \in \operatorname{Pic}{X^s}$, denote the reflection in $\alpha$ by $r_\alpha$; then we have $({{{}^{\sigma} r_\alpha}}) = r_{\sigma\alpha}$. So the action of ${{\Gamma_k}}$ on ${{\mathrm{O}}}(\operatorname{Pic}{X^s})$ restricts to an action on $W(\operatorname{Pic}{X^s})$.
\[defn:rx\] Define $R_X$ to be the group $W(\operatorname{Pic}{X^s})^{{\Gamma_k}}$.
Recall that $\operatorname{Pic}X$ is contained in, but not necessarily equal to, the fixed subgroup $(\operatorname{Pic}{X^s})^{{\Gamma_k}}$. The Hochschild–Serre spectral sequence gives rise to an exact sequence $$0 \to \operatorname{Pic}X \to (\operatorname{Pic}{X^s})^{{\Gamma_k}}\to \operatorname{Br}k \to \operatorname{Br}X.$$ If $X$ has a $k$-point, then evaluation at that point gives a left inverse to $\operatorname{Br}k \to \operatorname{Br}X$, showing that $\operatorname{Pic}X \to (\operatorname{Pic}{X^s})^{{\Gamma_k}}$ is an isomorphism. In general this does not have to be true. However, because $\operatorname{Br}k$ is torsion and $\operatorname{Pic}{X^s}$ is finitely generated, the above sequence shows that $\operatorname{Pic}X$ is of finite index in $(\operatorname{Pic}{X^s})^{{\Gamma_k}}$.
It is easy to see that the action of $R_X$ on $\operatorname{Pic}{X^s}$ preserves $(\operatorname{Pic}{X^s})^{{\Gamma_k}}$, but it is not immediately obvious that this action preserves $\operatorname{Pic}X$. To show that this is the case, we use an explicit description of $R_X$ provided by a theorem of Hée and Lusztig, for which Geck and Iancu gave a simple proof. Before stating their theorem, we establish some notation and conventions for Coxeter systems.
\[def:coxeter\] Let $W$ be a group generated by a set $T \subset W$ of elements of order $2$. For $t_i, t_j \in T$, let $n_{i,j} = n_{j,i}$ be the order of $t_i t_j$ if $t_i t_j$ has finite order and $0$ otherwise. Suppose that the relations $t_i^2 = 1, (t_i t_j)^{n_{i,j}} = 1$ for $i,j$ with $n_{i,j} \ne 0$ are a presentation of $W$. Then $(W,T)$ is a *Coxeter system.*
Let $G$ be a graph with vertices $T$ and such that $t_i, t_j$ are adjacent in $G$ if and only if $t_i$ does not commute with $t_j$; in this case, label the edge joining $t_i$ to $t_j$ with $n_{i,j} - 2$ for $n_{i,j} > 0$ and $0$ otherwise. We refer to $G$ as the *Coxeter–Dynkin diagram* of $(W,T)$. The Coxeter system $(W,T)$ is said to be *irreducible* if its Coxeter–Dynkin diagram is connected.
Let the [*length*]{} $\ell(w)$ of an element $w \in W$ be the length of a shortest word in the $t_i$ that represents it. If $W$ is finite, there is $w_0 \in W$ such that $\ell(w_0) > \ell(w)$ for all $w \ne w_0 \in W$ (see [@bjorner-brenti Proposition 2.3.1]). We refer to $w_0$ as the *longest element* of $W$.
Let $\sigma$ be a permutation of $T$. Then there is at most one way to extend $\sigma$ to a homomorphism $W \to W$, because $T$ generates $W$. If there is such an extension, it is an automorphism, because $\sigma^{-1}$ extends to its inverse, and we speak of it as the automorphism [*induced by $\sigma$*]{}.
If $(W,T)$ is a Coxeter system, and $I$ is a subset of $T$, let $W_I$ denote the subgroup of $W$ generated by the elements of $I$. Then $(W_I,I)$ is a Coxeter system: see [@bjorner-brenti Proposition 2.4.1 (i)].
\[thm:gi\] Let $(W,T)$ be a Coxeter system. Let $G$ be a group of permutations of $T$ that induce automorphisms of $W$. Let $F$ be the set of orbits $I \subset T$ for which $W_I$ is finite, and for $I \in F$ let $w_{I,0}$ be the longest element of $(W_I,I)$. Then $(W^G,\{w_{I,0}: I \in F\})$ is a Coxeter system.
We will apply this theorem with $W=W(\operatorname{Pic}{X^s})$ and $T$ being the set of reflections in $(-2)$-curves on ${X^s}$.
\[rx\] Let $F$ be the set of Galois orbits $I$ of $(-2)$-curves on ${X^s}$ of the following two types:
(i) $I$ consists of disjoint $(-2)$-curves;
(ii) $I$ consists of disjoint pairs of $(-2)$-curves, each pair having intersection number 1.
Then the following statements hold.
1. \[rx1\] For each $I \in F$, let $W_I$ be the subgroup of $W(\operatorname{Pic}{X^s})$ generated by reflections in the classes of curves in $I$, and let $r_I$ be the longest element of the Coxeter system $(W_I,I)$. Then $(R_X, \{r_I: I \in F\})$ is a Coxeter system.
2. \[rx2\] For each $I \in F$, let $C_I \in (\operatorname{Pic}{X^s})^{{\Gamma_k}}$ be the sum of the classes in $I$. Then $r_I$ acts on $(\operatorname{Pic}{X^s})^{{\Gamma_k}}$ as reflection in the class $C_I$.
3. \[rx3\] The action of $R_X$ on $\operatorname{Pic}{X^s}$ preserves $\operatorname{Pic}X$.
Let $I$ be a Galois orbit of $(-2)$-curves, and suppose that the subgroup $W_I$ is finite. We will show that $I$ is of one of the two types described. Firstly, no two $(-2)$-curves in $I$ have intersection number greater than $1$, for then the corresponding reflections would generate an infinite dihedral subgroup of $W_I$. Since $W_I$ is finite, its Coxeter–Dynkin diagram is a finite union of trees [@bjorner-brenti Exercise 1.4]. In particular, it contains a vertex of degree $\le 1$. However, the Galois group ${{\Gamma_k}}$ acts transitively on the diagram, so we conclude that either every vertex has degree $0$, or every vertex has degree $1$. These two possibilities correspond to the two types of orbits described. Now (\[rx1\]) follows from Theorem \[thm:gi\].
We prove (\[rx2\]) separately for the two types of orbits. In the first case we have $I = \{ E_1, \dotsc, E_r \}$. The reflections in the $E_i$ all commute, so $W_I$ is isomorphic to the Coxeter group $A_1^r$. The longest element is $r_I = r_{E_1} \circ \dotsb \circ r_{E_r}$. For $D \in (\operatorname{Pic}{X^s})^{{\Gamma_k}}$, the intersection numbers $D \cdot E_i$ are all equal, and one calculates $$r_I(D) = D + (D \cdot E_1)(E_1 + \dotsb + E_r),$$ that is, $r_I$ coincides on $(\operatorname{Pic}{X^s})^{{\Gamma_k}}$ with reflection in the class $C_I = E_1 + \dotsb E_r$, of self-intersection $-2r$.
In the second case, write $I = \{ E_1, E'_1, \dotsc, E_r, E'_r \}$, where $E_i \cdot E'_i = 1$ and the other intersection numbers are all zero. The two reflections $r_{E_i}$ and $r_{E'_i}$ generate a subgroup isomorphic to the Coxeter group $A_2$, in which the longest element is $$r_I = r_{E_i} \circ r_{E'_i} \circ r_{E_i} = r_{E'_i} \circ r_{E_i} \circ r_{E'_i} = r_{E_i + E'_i}.$$ Thus we have $W_I \cong A_2^r$ and the longest element in $W_I$ is the product of the longest elements in the factors $A_2$, that is, it equals $r_I = r_{E_1 + E'_1} \circ \dotsb \circ r_{E_r + E'_r}$. For $D \in (\operatorname{Pic}{X^s})^{{\Gamma_k}}$, all the $D \cdot E_i$ and $D \cdot E'_i$ are equal, and we have $$r_I(D) = D + 2(D \cdot E_1)(E_1 + E'_1 + \dotsb + E_r + E'_r),$$ which coincides with reflection of $D$ in the class $C_I = E_1 + E'_1 + \dotsb + E_r + E'_r$, of self-intersection $-2r$.
Finally, each class $C_I$ is, by construction, the class of a Galois-fixed divisor on ${X^s}$, so lies in $\operatorname{Pic}X$. So in both cases the formula for $r_I$ given above shows that reflection in $C_I$ preserves $\operatorname{Pic}X$, and therefore the action of $R_X$ preserves $\operatorname{Pic}X$, proving (\[rx3\]).
We now turn to the ample and nef cones. As ampleness is a geometric property, and the nef cone is the closure of the ample cone over any base field, it follows that $$\operatorname{Amp}(X) = \operatorname{Amp}({{\bar{X}}}) \cap (\operatorname{Pic}X)_{\mathbb{R}}\text{ and } \operatorname{Nef}(X) = \operatorname{Nef}({{\bar{X}}}) \cap (\operatorname{Pic}X)_{\mathbb{R}},$$ the intersections taking place inside $(\operatorname{Pic}{{\bar{X}}})_{\mathbb{R}}$. The following result is well known when $k$ is algebraically closed (see [@huybrechts Corollary 8.2.11]), and descends easily to arbitrary $k$.
\[prop:fundamental-domain\] Let $X$ be a K3 surface over $k$. The cone $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$ is a fundamental domain for the action of $R_X$ on the positive cone ${{\mathcal{C}}}_X$, and this action is faithful.
We will prove two things: first, that every class in ${{\mathcal{C}}}_X$ is $R_X$-equivalent to an element of $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$; and second, that the translates of $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$ by two distinct elements of $R_X$ meet only along their boundaries. The second of these shows in particular that the action is faithful. (When we refer to the boundary of $\operatorname{Nef}(X)$ or one of its translates, we mean the boundary within $(\operatorname{Pic}X)_{\mathbb{R}}$. The boundary of $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$ in ${{\mathcal{C}}}_X$ is just the boundary of $\operatorname{Nef}(X)$ intersected with ${{\mathcal{C}}}_X$, so that distinction is not so important.)
To prove the first statement, let $D \in {{\mathcal{C}}}_X$. Suppose first that $D$ has trivial stabilizer in $W(\operatorname{Pic}{{\bar{X}}})$. Then, by [@huybrechts Corollary 8.2.11], there exists a unique $g \in W(\operatorname{Pic}{{\bar{X}}}) = W(\operatorname{Pic}{X^s})$ such that $gD$ lies in the interior of $\operatorname{Nef}({{\bar{X}}}) \cap {{\mathcal{C}}}_{{\bar{X}}}$. We claim that $g$ lies in $R_X$. For any $\sigma \in {{\Gamma_k}}$, we have $$({{{}^{\sigma} g}}) D = \sigma(g(\sigma^{-1} D)) = \sigma (g(D)) \in \operatorname{Nef}({{\bar{X}}}) \cap {{\mathcal{C}}}_{{\bar{X}}}$$ since the Galois action preserves the properties of being nef and positive. By uniqueness of $g$, we conclude that $g = {{{}^{\sigma} g}}$, that is, $g$ lies in $R_X$. It then follows that $gD$ lies in $((\operatorname{Pic}{X^s})_{\mathbb{R}})^{{\Gamma_k}}= (\operatorname{Pic}X)_{\mathbb{R}}$ and therefore in $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$.
Now suppose that $D \in {{\mathcal{C}}}_X$ has non-trivial stabilizer. Then $D$ lies on at least one of the walls defined by the action of $W(\operatorname{Pic}{{\bar{X}}})$ on $(\operatorname{Pic}{{\bar{X}}})_{\mathbb{R}}$; see [@huybrechts Section 8.2]. By [@huybrechts Proposition 8.2.4], the chamber structure of this group action is locally polyhedral within ${{\mathcal{C}}}_{{\bar{X}}}$, so a small enough neighbourhood of $D$ meets only finitely many chambers. Also note that $\operatorname{Pic}X$ is not contained in any of the walls, because $X$ admits an ample divisor. This allows us to construct a sequence $(D_i)_{i=1}^\infty$ of elements of ${{\mathcal{C}}}_X$, tending to $D$ and all lying in the interior of the same chamber of ${{\mathcal{C}}}_{\bar{X}}$. As in the previous paragraph, there is a unique $g \in R_X$ satisfying $g D_i \in \operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$ for all $i$. By continuity, $gD$ also lies in $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$.
We now prove the second statement, that $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$ intersects the translate by any non-trivial element of $R_X$ only in its boundary. Suppose that $x \in {{\mathcal{C}}}_X$ lies in the intersection $\operatorname{Nef}(X) \cap g \operatorname{Nef}(X)$, for some non-trivial $g \in R_X$. By [@huybrechts Corollary 8.2.11], we see that $x$ lies in the boundary of $\operatorname{Nef}({{\bar{X}}})$. The following lemma shows that $x$ lies in the boundary of $\operatorname{Nef}(X)$.
\[lem:boundary\] Let $V$ be a real vector space, let $C \subset V$ be a closed convex cone, and let $S \subset V$ be a subspace having non-empty intersection with the interior of $C$. Then we have $$\partial_S(C \cap S) = \partial_V(C) \cap S.$$
The dual cone $C^* \subset V^*$ is defined by $$C^* = \{ \phi \in V^* \mid \phi(x) \ge 0 \text{ for all } x \in C \}.$$ By the supporting hyperplane theorem, $C^*$ has the property that a point $x \in C$ lies in $\partial_V(C)$ if and only if there exists a non-zero $\phi \in C^*$ satisfying $\phi(x)=0$. (The hyperplane $\phi=0$ is called a *supporting hyperplane* of $C$ at $x$.) Let $f \colon V^* \to S^*$ be the natural restriction map; then the dual $(C \cap S)^*$ is equal to $f(C^*)$ (see [@rockafellar Corollary 16.3.2]).
Let $x$ be a point of $\partial_V(C) \cap S$. Then there is a supporting hyperplane to $C$ at $x$, that is, there exists a non-zero $\phi \in C^*$ satisfying $\phi(x)=0$. The condition that $S$ meet the interior of $C$ implies that $\phi$ does not vanish identically on $S$, so $f(\phi)$ is non-zero. Thus $f(\phi)$ is a non-zero element of $(C \cap S)^*$ vanishing at $x$, so $x$ lies in $\partial_S(C \cap S)$.
Conversely, suppose that $x$ lies in $\partial_S(C \cap S)$. Then there is a supporting hyperplane to $C \cap S$ at $x$, that is, there exists a non-zero $\psi \in (C \cap S)^*$ satisfying $\psi(x)=0$. Let $\phi \in C^*$ satisfy $f(\phi)=\psi$; then we have $\phi(x)=0$ and so $x \in \partial_V(C)$.
\[walls\] We can also give an explicit description of the walls of $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$. According to [@huybrechts Corollary 8.1.6], a class in ${{\mathcal{C}}}_X$ lies in $\operatorname{Nef}(X)$ if and only if it has non-negative intersection number with every $(-2)$-curve on ${X^s}$, or, equivalently, with every Galois orbit of $(-2)$-curves. The question is to determine which Galois orbits are superfluous, and which actually define walls of $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$.
Let $I$ be a Galois orbit of $(-2)$-curves on ${X^s}$, and suppose that the subgroup $W_I \subset W(\operatorname{Pic}{X^s})$ generated by reflections in the elements of $I$ is finite, that is, $I$ is as described in Proposition \[rx\]. The longest element of $W_I$ acts on $\operatorname{Pic}X$ by reflection in the class $C_I = \sum_{E \in I} E$, which has negative self-intersection. The hyperplane orthogonal to $C_I$ is a wall of $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$: by the same argument as in [@sterk], given a class $D\in \operatorname{Amp}(X)$, the class $$D - \frac{D \cdot C_I}{C_I \cdot C_I} C_I$$ is orthogonal to $C_I$ but has positive intersection number with all $(-2)$-curves outside $I$.
On the other hand, let $I$ be a Galois orbit of $(-2)$-curves such that the subgroup $W_I$ is infinite. Then, for a $(-2)$-curve $C \in I$, the hyperplane orthogonal to $C$ in $\operatorname{Pic}({X^s})_{\mathbb{R}}$ does not meet $\operatorname{Nef}(X) \cap {{\mathcal{C}}}_X$, as the following argument shows. Let $x$ be a class in ${{\mathcal{C}}}_X$ orthogonal to $C$; then $x$ is also orthogonal to all the other curves in $I$, and so is fixed by $W_I$. Therefore $x$ is also orthogonal to the infinitely many images of $C$ under the action of $W_I$, contradicting the fact that the chamber structure induced by $W(\operatorname{Pic}{X^s})$ on ${{\mathcal{C}}}_{X^s}$ is locally polyhedral.
Next we would like to prove an analogue of Theorem \[thm:old\] (2). However, while the authors do not have an explicit counterexample, there seems to be no reason for the image of $R_X$ in ${{\mathrm{O}}}(\operatorname{Pic}X)$ to be normal. Instead, we will see that there is a natural homomorphism from a semidirect product $\operatorname{Aut}X \ltimes R_X$ to ${{\mathrm{O}}}(\operatorname{Pic}X)$ having finite kernel and image of finite index.
Note that the natural action of $\operatorname{Aut}X \subset \operatorname{Aut}{{\bar{X}}}$ on ${{\mathrm{O}}}(\operatorname{Pic}{{\bar{X}}})$ by conjugation fixes $W(\operatorname{Pic}{{\bar{X}}})$ and commutes with the Galois action, so fixes $R_X$. This gives an action of $\operatorname{Aut}X$ on $R_X$, and a homomorphism from the associated semidirect product $\operatorname{Aut}X \ltimes R_X$ to ${{\mathrm{O}}}(\operatorname{Pic}{{\bar{X}}})$. Since $\operatorname{Aut}X$ and $R_X$ both fix $\operatorname{Pic}X$, we also obtain a homomorphism $\operatorname{Aut}X \ltimes R_X \to {{\mathrm{O}}}(\operatorname{Pic}X)$.
\[main-result\] Let $X$ be a K3 surface over a field $k$ of characteristic different from $2$. Then the natural map $$\operatorname{Aut}X \ltimes R_X \to {{\mathrm{O}}}(\operatorname{Pic}X)$$ has finite kernel and image of finite index.
In the literature, this statement appears in various different but equivalent forms.
- Because the action of $\operatorname{Aut}X$ fixes the ample cone, the image of $\operatorname{Aut}X$ in ${{\mathrm{O}}}(\operatorname{Pic}X)$ meets $R_X$ only in the identity element. Therefore the finiteness of the kernel in Proposition \[main-result\] is equivalent to the finiteness of the kernel of $\operatorname{Aut}X \to {{\mathrm{O}}}(\operatorname{Pic}X)$.
- Let $\operatorname{Aut}_s X \subset \operatorname{Aut}X$ be the subgroup of symplectic automorphisms [@huybrechts Definition 15.1.1]. Over an algebraically closed field $k$ of characteristic zero, the induced map from $\operatorname{Aut}_s X$ to ${{\mathrm{O}}}(\operatorname{Pic}X)$ is injective, so that $\operatorname{Aut}_s X \ltimes W(\operatorname{Pic}X)$ can be viewed as a subgroup of ${{\mathrm{O}}}(\operatorname{Pic}X)$. Instead of our Proposition \[main-result\], Huybrechts [@huybrechts Theorem 15.2.6] makes the statement that $\operatorname{Aut}_s X \ltimes W(\operatorname{Pic}X)$ has finite index in ${{\mathrm{O}}}(\operatorname{Pic}X)$. This is equivalent to our formulation, because $\operatorname{Aut}_s X$ is of finite index in $\operatorname{Aut}X$. However, when $k$ is not algebraically closed there is no reason for $\operatorname{Aut}_s X \to {{\mathrm{O}}}(\operatorname{Pic}X)$ to be injective, so there is no advantage to stating the result in terms of $\operatorname{Aut}_s X$.
- Lieblich and Maulik [@lm Theorem 6.1] define $\Gamma \subset {{\mathrm{O}}}(\operatorname{Pic}X)$ to be the subgroup of elements preserving the ample cone, and then show that $\operatorname{Aut}X \to \Gamma$ has finite kernel and cokernel. Over an algebraically closed base field $k$, the subgroups $\Gamma$ and $W(\operatorname{Pic}X)$ generate ${{\mathrm{O}}}(\operatorname{Pic}X)$: any element of ${{\mathrm{O}}}(\operatorname{Pic}X)$ permutes the $(-2)$-classes in $\operatorname{Pic}X$ and therefore takes the ample cone to one of its translates under $W(\operatorname{Pic}X)$; so composing with an element of $W(\operatorname{Pic}X)$ gives an element of $\Gamma$. Therefore that condition is also equivalent to the condition of Proposition \[main-result\].
Before proving Proposition \[main-result\], we first state two lemmas which are well known in the case of abelian groups.
\[lem:kercoker\] Let $G$ be a finite group, and $f \colon A \to B$ a homomorphism of (possibly non-commutative) $G$-modules having finite kernel and image of finite index. Then the induced homomorphism $f^G \colon A^G \to B^G$ also has finite kernel and image of finite index.
The kernel of $f^G$ is contained in the kernel of $f$, so is finite. For the statement about the image, consider the short exact sequence $$0 \to \ker f \to A \xrightarrow{f} \operatorname{im}f \to 0.$$ The associated exact sequence of cohomology gives $$0 \to (\ker f)^G \to A^G \xrightarrow{f^G} (\operatorname{im}f)^G \to {\mathrm{H}}^1(G, \ker f)$$ and ${\mathrm{H}}^1(G,\ker f)$ is a finite set, showing that $\operatorname{im}(f^G)$ is of finite index in $(\operatorname{im}f)^G$. On the other hand, we claim that $(\operatorname{im}f)^G$ is of finite index in $B^G$. Indeed, by hypothesis $\operatorname{im}(f)$ is of finite index in $B$, and this property is preserved on intersecting with the subgroup $B^G$. Thus $\operatorname{im}(f^G)$ is of finite index in $B^G$.
The following lemma is standard and easy to prove; we state it for reference.
\[lem:comp\] Let $f \colon A \to B$ and $g \colon B \to C$ be two homomorphisms of groups.
1. If $f$ and $g$ both have finite kernel and image of finite index, then so does the composition $g \circ f$.
2. If $g \circ f$ has finite kernel and image of finite index, and $g$ has finite kernel, then $f$ has finite kernel and image of finite index.
3. If $g \circ f$ has finite kernel and image of finite index, and $f$ has image of finite index, then $g$ has finite kernel and image of finite index.
We leave this as an exercise for the reader.
By [@huybrechts Theorem 15.2.6] in characteristic zero, or [@lm] in characteristic $p > 2$, the natural map $$\operatorname{Aut}{{\bar{X}}}\to {{\mathrm{O}}}(\operatorname{Pic}{{\bar{X}}}) / W(\operatorname{Pic}{{\bar{X}}})$$ has finite kernel and image of finite index. Lemma \[lem:sep\] shows that the same is true for $\operatorname{Aut}{X^s}\to {{\mathrm{O}}}(\operatorname{Pic}{X^s}) / W(\operatorname{Pic}{X^s})$. The action of ${{\Gamma_k}}$ on all of these groups factors through a finite quotient, so Lemma \[lem:kercoker\] shows that the induced homomorphism $$\label{eq:hom}
(\operatorname{Aut}{X^s})^{{\Gamma_k}}\to ({{\mathrm{O}}}(\operatorname{Pic}{X^s}) / W(\operatorname{Pic}{X^s}))^{{\Gamma_k}}$$ also has finite kernel and image of finite index. Because the automorphism group functor is representable, we have $(\operatorname{Aut}{X^s})^{{\Gamma_k}}= \operatorname{Aut}X$.
There is an exact sequence $$1 \to R_X \to {{\mathrm{O}}}(\operatorname{Pic}{X^s})^{{\Gamma_k}}\to ({{\mathrm{O}}}(\operatorname{Pic}{X^s}) / W(\operatorname{Pic}{X^s}))^{{\Gamma_k}}$$ and the homomorphism factors through ${{\mathrm{O}}}(\operatorname{Pic}{X^s})^{{\Gamma_k}}$, and hence through the injective map $${{\mathrm{O}}}(\operatorname{Pic}{X^s})^{{\Gamma_k}}/R_X \to ({{\mathrm{O}}}(\operatorname{Pic}{X^s}) / W(\operatorname{Pic}{X^s}))^{{\Gamma_k}}.$$ Therefore, by Lemma \[lem:comp\], the map $$\operatorname{Aut}X \to {{\mathrm{O}}}(\operatorname{Pic}{X^s})^{{\Gamma_k}}/ R_X$$ has finite kernel and image of finite index. Since the image of $\operatorname{Aut}X$ in ${{\mathrm{O}}}(\operatorname{Pic}X^s)^{{\Gamma_k}}$ meets $R_X$ only in the identity element, this shows that the natural map $$\label{eq:hom2}
\operatorname{Aut}X \ltimes R_X \to {{\mathrm{O}}}(\operatorname{Pic}X^s)^{{\Gamma_k}}$$ also has finite kernel and image of finite index.
We apply Proposition \[lattice-prop\] with $\Lambda = \operatorname{Pic}{X^s}$ and $H$ being the image of ${{\Gamma_k}}$ in ${{\mathrm{O}}}(\operatorname{Pic}{X^s}))$. The centralizer $Z_{{{\mathrm{O}}}(\Lambda)}(H)$ is then ${{\mathrm{O}}}(\operatorname{Pic}{X^s})^{{\Gamma_k}}$, so part (2) of Proposition \[lattice-prop\] shows that ${{\mathrm{O}}}(\operatorname{Pic}{X^s})^{{\Gamma_k}}$ is of finite index in ${{\mathrm{O}}}(\operatorname{Pic}{X^s},(\operatorname{Pic}{X^s})^{{\Gamma_k}})$. Parts (1) and (2) of Proposition \[lattice-prop\] combined show that the map $${{\mathrm{O}}}(\operatorname{Pic}{X^s},(\operatorname{Pic}{X^s})^{{\Gamma_k}}) \to {{\mathrm{O}}}((\operatorname{Pic}{X^s})^{{\Gamma_k}})$$ has finite kernel and image of finite index; by Lemma \[lem:comp\] so does the map ${{\mathrm{O}}}(\operatorname{Pic}{X^s})^{{\Gamma_k}}\to {{\mathrm{O}}}((\operatorname{Pic}{X^s})^{{\Gamma_k}})$.
Composing with and applying Lemma \[lem:comp\] shows that $$\operatorname{Aut}X \ltimes R_X \to {{\mathrm{O}}}((\operatorname{Pic}{X^s})^{{\Gamma_k}})$$ has finite kernel and image of finite index. Since the actions of both $\operatorname{Aut}X$ and $R_X$ on $(\operatorname{Pic}{X^s})^{{\Gamma_k}}$ preserve $\operatorname{Pic}X$, this last map factors as $$\label{eq:hom4}
\operatorname{Aut}(X) \ltimes R_X \to {{\mathrm{O}}}((\operatorname{Pic}{X^s})^{{\Gamma_k}},\operatorname{Pic}X) \to {{\mathrm{O}}}((\operatorname{Pic}{X^s})^{{\Gamma_k}}).$$ As the second map in this composition is clearly injective, Lemma \[lem:comp\] shows that the first one has finite kernel and image of finite index.
Finally, $\operatorname{Pic}X$ is of finite index in $(\operatorname{Pic}{X^s})^{{\Gamma_k}}$, so its orthogonal complement in the latter is trivial. Lemma \[sublemma\](\[Operp\]) shows that the natural map $${{\mathrm{O}}}((\operatorname{Pic}{X^s})^{{\Gamma_k}}, \operatorname{Pic}X) \to {{\mathrm{O}}}(\operatorname{Pic}X)$$ is injective, and its image has finite index. Composing with the first map of and applying Lemma \[lem:comp\] again, we deduce that $\operatorname{Aut}X \ltimes R_X \to {{\mathrm{O}}}(\operatorname{Pic}X)$ has finite kernel and image of finite index.
The finiteness of the kernel can also be proved directly, by the same proof as over an algebraically closed field.
Having proved Proposition \[main-result\], we can deduce the remaining results exactly as in the classical case. Define the cone $\operatorname{Nef}^e(X)$ to be the real convex hull of $\operatorname{Nef}(X) \cap \operatorname{Pic}X$.
\[rat-poly\] The action of $\operatorname{Aut}X$ on $\operatorname{Nef}^e(X)$ admits a rational polyhedral fundamental domain.
This is as in the case of $k={\mathbb{C}}$; we briefly recall the argument of Sterk [@sterk], making the necessary adjustments.
Let $\Lambda$ be a lattice of signature $(1,\rho-1)$ and let $\Gamma \subset {{\mathrm{O}}}(\Lambda_{\mathbb{R}})$ be an arithmetic subgroup (for example, a subgroup of finite index in ${{\mathrm{O}}}(\Lambda)$). Let $C \subset \Lambda_{\mathbb{R}}$ be one of the two components of $\{ x \in \Lambda_{\mathbb{R}}\mid (x \cdot x) < 0 \}$; this is a self-adjoint homogeneous cone [@ash Remark 1.11]. Let $C_+$ be the convex hull of $\overline{C} \cap \Lambda_{\mathbb{Q}}$. Pick any $y \in C \cap \Lambda$; the argument of [@sterk p. 511] shows that the set $$\Pi = \{ x \in C_+ \mid ( \gamma x \cdot y ) \ge ( x \cdot y ) \text{ for all } \gamma \in \Gamma \}$$ is rational polyhedral (for this, the reference to Proposition 11 of the first edition of [@ash] has become Proposition 5.22 in the second edition) and is a fundamental domain for the action of $\Gamma$ on $C_+$. (Sterk does not explicitly prove that two translates of $\Pi$ intersect only in their boundaries, but this is easy to show from the description above.) In our case, applying this with $\Lambda = \operatorname{Pic}X$ and $\Gamma$ being the image of $\operatorname{Aut}X \ltimes R_X \to {{\mathrm{O}}}(\operatorname{Pic}X)$ gives a rational polyhedral fundamental domain $\Pi$ for the action of $\operatorname{Aut}X \ltimes R_X$ on $({{\mathcal{C}}}_X)_+$.
If we choose $y$ to be an ample class in $\operatorname{Pic}X$, then the resulting $\Pi$ is contained in $\operatorname{Nef}(X)$, as we now show. Let $I$ be a Galois orbit of $(-2)$-curves on ${X^s}$ such that the corresponding group $W_I \subset W(\operatorname{Pic}{X^s})$ is finite. Proposition \[rx\] states that the longest element $w$ of $W_I$ acts on $\operatorname{Pic}X$ as reflection in the class $C_I = \sum_{E \in I} E$, and that these elements generate $R_X$. Taking $\gamma=w$ in the definition of $\Pi$ shows that $\Pi$ is contained in the half-space $\{x \mid x.C_I \ge 0 \}$. As this holds for all such $I$, Remark \[walls\] shows that $\Pi$ is contained in $\operatorname{Nef}(X)$.
We conclude as in [@sterk]. If $x$ is a class in $\operatorname{Nef}^e(X)$ then, since $\Pi$ is a fundamental domain for the action of $\operatorname{Aut}X \ltimes R_X$ on $({{\mathcal{C}}}_X)_+$, we can find $\phi \in \operatorname{Aut}X$ and $r \in R_X$ such that $r \phi(x)$ lies in $\Pi$. But now $\phi(x)$ and $r \phi(x)$ both lie in $\operatorname{Nef}^e(X)$, so they are equal and lie in $\Pi$. This shows that $\Pi$ is a fundamental domain for the action of $\operatorname{Aut}X$ on $\operatorname{Nef}^e(X)$.
\[finite-class-orbits\]
1. \[g0\] There are only finitely many $\operatorname{Aut}X$-orbits of $k$-rational $(-2)$-curves on $X$.
2. \[g1\] There are only finitely many $\operatorname{Aut}X$-orbits of primitive Picard classes of irreducible curves on $X$ of arithmetic genus $1$.
3. \[g2\] For $g \ge 2$, there are only finitely many $\operatorname{Aut}X$-orbits of Picard classes of irreducible curves on $X$ of arithmetic genus $g$.
Let $\Pi$ be a rational polyhedral fundamental domain for the action of $\operatorname{Aut}X$ on $\operatorname{Nef}^e(X)$, as in Corollary \[rat-poly\]. Every $k$-rational $(-2)$-curve on $X$ defines a wall of $\operatorname{Nef}(X)$, by Remark \[walls\]. Since $\Pi$ meets only finitely many walls of $\operatorname{Nef}(X)$, this proves (\[g0\]).
Gordan’s Lemma states that the integral points of the dual cone of a rational polyhedral convex cone form a finitely generated monoid. Applying this to the dual cone of $\Pi$, let $D_1, \dotsc, D_r$ be a minimal set of generators for $\Pi \cap \operatorname{Pic}X$. Since these all lie in $\operatorname{Nef}(X)$, we have $D_i \cdot D_i \ge 0$ for all $i$, and $D_i \cdot D_j > 0$ for $i \neq j$. As observed in [@sterk], this implies that, for any $n > 0$, there are only finitely many classes in $\Pi \cap \operatorname{Pic}X$ of self-intersection $n$; and there are only finitely many primitive classes in $\Pi \cap \operatorname{Pic}X$ of self-intersection zero. The class of an irreducible curve of arithmetic genus $g \ge 1$ has self-intersection $2g-2$ and therefore lies in $\operatorname{Nef}(X)$, so this proves (\[g1\]) and (\[g2\]).
It is not true that every irreducible curve on $X$ of arithmetic genus $0$ is a $k$-rational $(-2)$-curve. However, there are not many possibilities, as we now show. Let $C$ be such a curve, and let $C_1, \dotsc, C_r$ be the geometric components of $C$; the Galois group ${{\Gamma_k}}$ acts transitively on them. In order to achieve $C^2=-2$, we must have $C_i^2<0$ for all $i$, so each $C_i$ is a $(-2)$-curve. Consider the intersection matrix $(C_i \cdot C_j)$: the sum of the entries in the matrix is $C^2=-2$, and the Galois action shows that every row sum is the same. Therefore the number $r$ of rows divides $2$, and there are only two options: $r=1$, so that $C$ is a rational $(-2)$-curve; or $r=2$, and $C=C_1 \cup C_2$ is the union of two conjugate $(-2)$-curves meeting transversely in a single point. By Corollary \[rat-poly\], both types of curves define walls of $\operatorname{Nef}(X)$, so in fact both types fall into finitely many orbits under the action of $\operatorname{Aut}(X)$.
Both types occur on the surfaces considered in Section \[finite-diagonal\] below. Indeed, the line $x = y, z = w$ on the surface $x^4 - y^4 = c(z^4 - w^4)$ is a rational $(-2)$-curve, while the line $x = y, z = iw$ meets its conjugate transversely in a single point.
In the case $g=1$, the condition that the class be primitive cannot be omitted, for the following reason. Take for example $k={\mathbb{Q}}$, and suppose that $X$ admits an elliptic fibration $\pi \colon X \to {\mathbb{P}}^1$. A general fibre of such a fibration is a smooth, geometrically irreducible curve $E$ of genus $1$, whose class in $\operatorname{Pic}X$ has self-intersection $0$ and is primitive. (In fact, the class is even primitive in $\operatorname{Pic}{{\bar{X}}}$.) If $s \in {\mathbb{P}}^1$ is a point of degree $m>1$, then in general the fibre $\pi^{-1}(s)$ will be an irreducible curve on $X$ of arithmetic genus $1$, linearly equivalent to $mE$. As $m$ varies, this construction gives infinitely many such classes that are clearly not $\operatorname{Aut}X$-equivalent.
Examples {#sec:eg}
========
In this section we give three examples that illustrate some of the theory developed up to this point. First, we will give a K3 surface $X$ with finite automorphism group, even though all K3 surfaces $V$ over $\bar {\mathbb{Q}}$ with $\operatorname{Pic}V \cong \operatorname{Pic}X$ have infinite automorphism group. Second, we will construct a surface $Y$ with finite automorphism group, even though, for all extensions $k/{\mathbb{Q}}$, all K3 surfaces $V$ over $\bar {\mathbb{Q}}$ with $\operatorname{Pic}V \cong \operatorname{Pic}Y_k$ have infinite automorphism group. For a third example, we will prove that the quartic surface $Z$ in ${\mathbb{P}}^3$ defined by $x^4 - y^4 = c(z^4 - w^4)$ over a field $k$ of characteristic $0$ has finite automorphism group when $c \in k$ is such that the Galois group of the field of definition of the Picard group has degree $16$, the largest possible.
First example {#finite-from-overlattice}
-------------
We construct a surface $X$ such that $\operatorname{Aut}\bar{X}$ is finite, and thus so is $\operatorname{Aut}X$, while any K3 surface over $\bar{{\mathbb{Q}}}$ having the same Picard lattice as $X$ has infinite automorphism group. In contrast, over an algebraically closed field of characteristic not equal to $2$, finiteness of the automorphism group depends only on the isomorphism type of the Picard lattice. This is an immediate consequence of statement $2$ of Theorem \[thm:old\].
Let $M$ and $N$ be the block diagonal matrices $$M = \begin{pmatrix} 0&1&0\\1&0&0\\0&0&-8 \end{pmatrix}, \qquad
N = \left( \begin{array}{cc|c} 0&1\\1&0\\ \hline &&-2I_4 \end{array}\right),$$ where $I_4$ is the $4\times 4$ identity matrix. Let $L_N$ be a lattice with basis $( e_1, \dotsc, e_6 )$ and Gram matrix $N$ with respect to that basis. Let $L_M \subset L_N$ be the sublattice generated by $e_1$, $e_2$ and $e_3+e_4+e_5+e_6$. The Gram matrix for $L_M$ with respect to this basis is $M$. The surface $X$ that we will construct will have compatible isomorphisms $\operatorname{Pic}\bar{X} \cong L_N$ and $\operatorname{Pic}X \cong L_M$.
\[prop:aut-x-inf\] Let $V$ be a K3 surface over $\bar {\mathbb{Q}}$ with $\operatorname{Pic}V \cong L_M$. Then $\operatorname{Aut}V$ is infinite.
There is an obvious embedding of the hyperbolic lattice $U$ with Gram matrix $\left( \begin{smallmatrix}0&1\\1&0\end{smallmatrix} \right)$ into $L_M \cong \operatorname{Pic}V$. By [@huybrechts Remark 11.1.4], this implies that there is an elliptic fibration $\pi \colon V \to {\mathbb{P}}^1$ with a section. Let $E$ be the class of a fibre of $\pi$ and $O$ the class of a section; then $E$ and $O$ generate a sublattice of $\operatorname{Pic}V$ isomorphic to $U$ (though not necessarily the obvious one). If $\pi$ were to have a reducible fibre, then any component of that fibre other than the one meeting $O$ would lie in ${\langle E,O \rangle}^\perp$ and have self-intersection $-2$. However, ${\langle E,O \rangle}$ is a lattice of rank $2$ and determinant $-1$, so the determinant of ${\langle E,O \rangle}^\perp$ is the negative of that of $\operatorname{Pic}V$ and its rank is $2$ less. Thus it is generated by a single vector of norm $-8$ and so there are no reducible fibres. It follows by the Shioda–Tate formula [@huybrechts Corollary 11.3.4] that the Mordell–Weil group of the fibration has rank $1$. Translation by a non-torsion section gives an automorphism of $V$ of infinite order.
In fact, Shimada [@shimada Remark 9.3] has proved that this translation, and negation in the Mordell–Weil group, generate the whole of $\operatorname{Aut}V$.
In contrast with Proposition \[prop:aut-x-inf\], we will see that there exist K3 surfaces over ${\mathbb{Q}}$ with Picard lattice isomorphic to $L_M$ for which the automorphism group is finite. We use the standard Kodaira symbols for reducible fibres of elliptic fibrations, as in [@silverman IV.9, Table 4.1].
\[prop:autxfinite\] Let $V$ be a K3 surface over an algebraically closed field with an elliptic fibration $\pi \colon V \to {\mathbb{P}}^1$ that has a section. Suppose that $\pi$ has four fibres each of type either $\textit{I}_2$ or $\textit{III}$. Then the following statements are equivalent.
1. \[pic6\] The rank of $\operatorname{Pic}V$ is at most $6$.
2. \[picLN\] The Picard lattice $\operatorname{Pic}V$ is isomorphic to $L_N$.
3. \[mw\] The Mordell–Weil group associated to $\pi$ is trivial, and there are only four reducible fibres.
If these equivalent conditions hold, then $\operatorname{Aut}V$ is finite.
Before proving the proposition, we state and prove a lemma.
\[lem:cant-divide-four\] Let $V$ be a K3 surface containing four pairwise disjoint smooth rational curves $C_1, C_2, C_3, C_4$. Then the class $\sum_{i=1}^4 [C_i]$ cannot be divided by $2$ in $\operatorname{Pic}V$.
Suppose otherwise, and let $D = (\sum_{i=1}^4 [C_i])/2 \in \operatorname{Pic}V$. Then $D^2 = -2$, so either $D$ or $-D$ is effective; since $2D$ is effective, we deduce that $D$ is effective. Now $(D,[C_i]) = -1$ for $1 \le i \le 4$, so the $C_i$ are base components of the linear system $|D|$ and $D - \sum_{i=1}^4 [C_i] = -D$ is also effective, giving a contradiction.
The equivalence of (\[picLN\]) and (\[mw\]) is proved as follows. By [@huybrechts Proposition 11.3.2], there is an exact sequence $$0 \to A \to \operatorname{Pic}V \to G \to 0,$$ where $A \subset \operatorname{Pic}V$ is the subgroup generated by vertical divisors and a chosen section, and $G$ is the Mordell–Weil group of the elliptic fibration. Let $E$ be a fibre of $\pi$, let $O$ be a chosen section, and let $W_1, \dotsc, W_4$ be the components of the four given fibres that do not meet $O$. The classes $[E], [E]+[O], [W_1], \dotsc, [W_4]$ lie in $A$ and have intersection matrix equal to $N$, so they generate a sublattice of $A$ isomorphic to $L_N$. If $G$ is trivial and there are no other reducible fibres, then these six classes generate $\operatorname{Pic}V$ and so we have $\operatorname{Pic}V \cong L_N$. Conversely, if $\operatorname{Pic}V$ is isomorphic to $L_N$ then these six classes must generate $\operatorname{Pic}V$, so $G$ is trivial. Also, there can be no further reducible fibres, for the class of a curve in such a fibre would be independent of the given generators of $\operatorname{Pic}V$.
The implication (\[picLN\])$\Rightarrow$(\[pic6\]) is trivial; we now prove (\[pic6\])$\Rightarrow$(\[picLN\]). As above, we have an embedding $L_N \hookrightarrow \operatorname{Pic}V$, so $\operatorname{Pic}V$ must have rank exactly $6$. Since $\operatorname{Pic}V$ has rank $6$ this embedding has finite index, and we must prove it to be an isomorphism. The determinant of $N$ is $-16$; the square of the index $[\operatorname{Pic}V:L_N]$ must divide this, so the index is $1, 2,$ or $4$. If it is not $1$, there is some element of $L_N$ that can be divided by $2$ in $\operatorname{Pic}V$ but not in $L_N$. We take this element to be of the form $a[E] + b[O] + \sum_{i=1}^4 c_i [W_i]$, where all the coefficients are $0$ or $1$ and not all are $0$. Then $a = b = 0$, for otherwise the intersection number with $[O]$ or $[E]$ would be odd; and all the $c_i$ must be equal, because the self-intersection of any divisor on a K3 surface is even and hence that of any divisor that can be divided by $2$ is a multiple of $8$. Thus all $c_i$ are equal to $1$. However, Lemma \[lem:cant-divide-four\] shows that $\sum_{i=1}^4 [W_i]$ is not divisible by $2$. This proves $\operatorname{Pic}V \cong L_N$.
Finally, Nikulin [@nikulin Theorem 3.1] has listed the finitely many possibilities for $\operatorname{Pic}V$ of rank $\ge 6$ that give rise to finite automorphism groups. The lattice $L_N \cong U \oplus 4A_1$ is in the list, showing that $\operatorname{Aut}V$ is finite.
\[prop:auttfinite\] Let $X$ be a K3 surface over a field $k$ with an elliptic fibration $\pi \colon X \to {\mathbb{P}}^1$ that has a section. Suppose that $\pi$ has four conjugate fibres of type $\textit{I}_2$ or $\textit{III}$ and that the rank of $\operatorname{Pic}\bar{X}$ is at most $6$. Then there are compatible isomorphisms $\operatorname{Pic}\bar{X} \cong L_N$ and $\operatorname{Pic}X \cong L_M$, and $\operatorname{Aut}X$ is finite.
Applying Proposition \[prop:autxfinite\] to the surface $\bar{X}$ shows that $\operatorname{Pic}\bar{X}$ is isomorphic to $L_N$. More precisely, the proof shows that there is an isomorphism $\operatorname{Pic}\bar{X} \cong L_N$ that identifies the basis $(e_1, \dotsc, e_6)$ of $L_N$ with the basis $([E], [E]+[O], [W_1], \dotsc, [W_4])$ of $\operatorname{Pic}\bar{X}$, where $E$ is a fibre of $\pi$ and $O$ is a section, and $W_1, \dotsc, W_4$ are the components of the four reducible fibres that do not meet $O$. Since by Proposition \[prop:autxfinite\] there are no other reducible fibres, the Galois action permutes $W_1, \dotsc, W_4$ transitively and so the Galois-invariant subgroup is identified with $L_M$. As $X$ contains a $k$-rational curve $O$ of genus $0$, it has rational points over $k$ and hence $\operatorname{Pic}X = (\operatorname{Pic}\bar{X})^{\Gamma_{\mathbb{Q}}} = L_M$.
Proposition \[prop:autxfinite\] also shows that $\operatorname{Aut}\bar{X}$ is finite, and *a fortiori* that $\operatorname{Aut}X$ is finite.
We now construct a K3 surface $X$ over ${\mathbb{Q}}$ satisfying the conditions of Proposition \[prop:auttfinite\], as the Jacobian of a genus-$1$ fibration on a quartic $U$ in ${\mathbb{P}}^3$. See [@huybrechts Section 11.4] for the properties of the Jacobian of a genus-$1$ fibration on a K3 surface.
Let $U$ be a smooth quartic surface in ${\mathbb{P}}^3$ over ${\mathbb{Q}}$ containing a line $L$. Projection away from $L$ induces a morphism $\pi_L \colon U \to {\mathbb{P}}^1$ whose fibres are the residual intersections with $U$ of planes containing $L$. The generic fibre is a smooth curve of genus $1$, and the induced morphism $L \to {\mathbb{P}}^1$ has degree $3$.
\[prop:four-a1\] Let $U$ be a smooth quartic surface in ${\mathbb{P}}^3_{\mathbb{Q}}$ containing a rational line $L$ and a Galois orbit $\mathcal{L}$ of four lines that meet $L$. Suppose in addition that each of the four planes containing $L$ and a line in $\mathcal{L}$ meets $U$ in one further component, which is a smooth conic. Then $\pi_L$ has four conjugate fibres of type $\textit{I}_2$ or $\textit{III}$.
Let $X \to {\mathbb{P}}^1$ be the relative Jacobian of $\pi_L$, and suppose in addition that $\operatorname{Pic}(\bar{U})$ has rank at most $6$. Then there are compatible isomorphisms $\operatorname{Pic}\bar{X} \cong L_N$ and $\operatorname{Pic}X \cong L_M$, and $\operatorname{Aut}X$ is finite.
Let $H$ be one of the four conjugate planes containing $L$ and a line $L' \in \mathcal{L}$, and let $C$ denote the residual smooth conic. The union $L' \cup C$ is a fibre of $\pi_L$. We have $L' \cdot C = 2$, because $L', C$ are a line and a conic in the plane $H$. Either $L'$ is tangent to $C$, in which case we have a fibre of type $\textit{III}$, or they intersect in two distinct points, and the fibre is of type $\textit{I}_2$. The same description holds for the other three planes that are Galois conjugates of $H$.
The relative Jacobian $X$ is a K3 surface [@huybrechts Proposition 11.4.5] that has the same geometric Picard number as $U$ [@huybrechts Corollary 11.4.7 and the discussion following it], and $X \to {\mathbb{P}}^1$ has the same geometric fibres as $\pi_L$ [@huybrechts Chapter 11, equation (4.1)]. Now apply Proposition \[prop:auttfinite\].
A very general surface $U$ constructed according to this proposition has Picard group generated by the classes of $L$, the lines in $\mathcal{L}$ and a fibre of $\pi_L$. This does not imply that such a surface exists over ${\mathbb{Q}}$, but we will see that it is not difficult to find an example.
\[ex:construct-t\] We claim that the surface $U \subset {\mathbb{P}}^3_{\mathbb{Q}}$ given by the equation $$\begin{gathered}
\label{eq:R}
-2x^3z - 3x^2yz - 3y^3z + x^2z^2 - 3xyz^2 + 2y^2z^2 + xz^3 + yz^3 - 13x^3w \\ + 24x^2yw
- 13xy^2w + 8y^3w - x^2zw + 51xz^2w - 37x^2w^2 + 47xyw^2 - 16y^2w^2 \\ + 111xzw^2
- 38yzw^2 - 57z^2w^2 - 227xw^3 + 24yw^3 - 94zw^3 + 303w^4 = 0\end{gathered}$$ satisfies the conditions of Proposition \[prop:four-a1\]. Indeed, $U$ contains the line $L$ defined by $w=z=0$ and, for any $\alpha$ satisfying $\alpha^4+\alpha-1 = 0$, $U$ contains the line $L_\alpha$ through the point $(-\alpha^3 + \alpha^2 - \alpha + 1 : 1 : 0 : 0)$ on $L$ and the point $(-\alpha^3 - \alpha + 1 : -\alpha^3 - \alpha^2 + 1 : \alpha^3 - \alpha^2 + \alpha + 1 : 1)$. The plane $H_\alpha$ containing $L$ and $L_\alpha$ is given by $z=(\alpha^3-\alpha^2+\alpha+1)w$, and one verifies that $U \cap H_\alpha$ consists of $L$, $L_\alpha$ and a smooth conic.
Let $U_3 \subset {\mathbb{P}}^3_{{\mathbb{F}}_3}$ be the surface defined by the equation , which is smooth. Let $\bar{U}_3$ be the base change of $U_3$ to an algebraic closure of ${\mathbb{F}}_3$, and $F \colon \bar{U}_3 \to \bar{U}_3$ the geometric Frobenius morphism, defined by $(x,y,z,w) \mapsto (x^3,y^3,z^3,w^3)$. Choose a prime $\ell \neq 3$ and let $F^*$ be the endomorphism of ${\mathrm{H}}^2(\bar{U}_3, {\mathbb{Q}}_\ell(1))$ induced by $F$. By [@rvl Proposition 6.2] the Picard rank of $\bar{U}$ is bounded above by that of $\bar{U}_3$, which in turn is at most the number of eigenvalues of $F^*$ that are roots of unity by [@rvl Corollary 6.4]. As in [@rvl], we find the characteristic polynomial of $F^*$ by counting points on $U_3$. The results are shown in Table \[counts\].
$n$ $1$ $2$ $3$ $4$ $5$ $6$ $7$ $8$
----------------------------- ------ ------ ------- -------- --------- ---------- ----------- ------------
$\#U_3({\mathbb{F}}_{3^n})$ $16$ $94$ $676$ $7066$ $60076$ $533818$ $4785076$ $43101802$
: Point counts on the surface $U_3$[]{data-label="counts"}
From the Lefschetz fixed point formula we find that the trace of the $n$th power of Frobenius acting on ${\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{U}_3,{\mathbb{Q}}_\ell)$ is equal to $\#U_3({\mathbb{F}}_{3^n}) - 3^{2n} - 1$. The trace on the Tate twist ${\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{U}_3,{\mathbb{Q}}_\ell(1))$ is obtained by dividing by $3^n$, while on the subspace $V \subset {\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{U}_3,{\mathbb{Q}}_\ell(1))$ generated by $H, L$, and the four lines $L_\alpha$, the trace $t_n$ is equal to $6$ if $n$ is a multiple of $4$, and equal to $2$ if $n$ is not a multiple of $4$; hence, on the $16$-dimensional quotient $Q={\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{U}_3,{\mathbb{Q}}_\ell(1))/V$, the trace equals $\#U_3({\mathbb{F}}_{3^n})/3^n - 3^{n} - 3^{-n} - t_n$. These traces are sums of powers of eigenvalues, and we use the Newton identities to compute the elementary symmetric polynomials in these eigenvalues, which are the coefficients of the characteristic polynomial $f$ of Frobenius acting on $Q$. This yields the first half of the coefficients of $f$, including the middle coefficient, which turns out to be nonzero. This implies that the sign in the functional equation $t^{16}f(1/t) = \pm f(t)$ is $+1$, so this functional equation determines $f$, which we calculate to be $$f = \tfrac{1}{3}\big(3t^{16} + t^{14} + 4t^{13} + 2t^{10} - 2t^8 + 2t^6 + 4t^3 + t^2 + 3\big).$$ As the characteristic polynomial of Frobenius acting on $V$ is $(t-1)^2(t^4-1)$, we find that the characteristic polynomial of Frobenius acting on ${\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{U}_3,{\mathbb{Q}}_\ell(1))$ is $(t-1)^2(t^4-1)f$. The polynomial $3f \in {\mathbb{Z}}[t]$ is irreducible, primitive and not monic, so its roots are not roots of unity. Thus we obtain an upper bound of $6$ for the Picard rank of $\bar{U}$.
We found $U$ by first fixing the lines $L$ and $L'$. The space of rational quartic polynomials vanishing on $L$ and on the conjugates of $L'$ is easily checked to have dimension $14$. (This is the expected dimension. The space of quartic polynomials has dimension $35$ and vanishing on a line is equivalent to vanishing on $5$ points of the line and so imposes $5$ conditions. However, if two lines meet in a point, that point gives the same condition for both lines. With four pairs of intersecting lines, we therefore expect the dimension to be $35 - 5 \times 5 + 4 = 14$.) We randomly chose elements of this space until we obtained one defining a smooth surface with good reduction at $3$ and suitable characteristic polynomial of Frobenius.
\[nine-curves\] The surface $\bar X$ has exactly $9$ smooth rational curves.
Let the $W_i$ be the components of the reducible fibres that meet $O$. In our basis, the known rational curves have classes $[O], [W_i], [E]-[W_i]$ for $1 \le i \le 4$. Suppose that there is another rational curve, of class $C = a[E] + b[O] + \sum_{i=1}^4 c_i [W_i]$. It must have non-negative intersection with the known curves, which implies the inequalities $$a \ge b \ge 0, \quad c_i \le 0, \quad b + c_i \ge 0 \qquad (1 \le i \le 4).$$ Let $m = \min(\{c_i\})$: then $ab \ge b^2 \ge 4m^2$. We thus find $$C^2 = 2ab - 2\sum_{i=1}^4 c_i^2 \ge 2b^2 - 8m^2 \ge 0,$$ which contradicts the fact that $C^2 = -2$.
\[more-curves-on-r\] The surface $\bar U$, on the other hand, has infinitely many smooth rational curves. Indeed, by [@huybrechts Chapter 11, equation (4.5)], the Picard lattice of $\bar U$ has discriminant $-144$. However, the list of Picard lattices of rank $\ge 6$ giving finite automorphism group in [@nikulin Theorem 3.1] does not include any lattices of rank $6$ and discriminant $-144$. Hence $\operatorname{Aut}\bar U$ is infinite. Let $\mathcal C$ be the union of $\{L\}$ and the set of components of the reducible fibres of $\pi_L$. By construction $\mathcal C$ spans a subgroup of $\operatorname{Pic}\bar U$ of finite index (in fact, by computing the discriminant one checks that $\mathcal C$ generates $\operatorname{Pic}\bar U$). It follows that the stabilizer of $\mathcal C$ in $\operatorname{Aut}\bar U$ is finite and hence that the orbit is infinite.
\[prop:min-deg\] We now explain why we do not construct the Jacobian of $U$ directly as a smooth surface in a projective space.
If $H$ is an ample divisor class on a K3 surface $X$ with $\operatorname{Pic}X \cong L_N$, then $H^2 \ge 16$. To see this, let $H = a[E] + b[O] + \sum_{i=1}^4 c_i [W_i]$ be such a class. Since $H$ is ample, all of $H \cdot [O], H \cdot [W_i], H \cdot ([E] - [W_i])$ must be positive: that is, $$a - 2b + \sum_{i=1}^4 c_i \ge 1, \quad b - 2c_i \ge 1, \quad c_i \ge 1.$$ So $b \ge 3$ and $b+2c_i \ge 5$. We thus find $$\begin{aligned}
H^2 &= 2ab - 2b^2 + 2 \sum_{i=1}^4 bc_i - 2 \sum_{i=1}^4 c_i^2 \cr
&\ge 2b(2b - \sum_{i=1}^4 c_i + 1) - 2b^2 + 2 \sum_{i=1}^4 bc_i - 2 \sum_{i=1}^4 c_i^2 \cr
&= 2b^2 + 2b - 2 \sum_{i=1}^4 c_i^2 \cr
&= 2b + \frac{\sum_{i=1}^4 (b-2c_i)(b+2c_i)}2\cr
&\ge 6 + 4\cdot 1 \cdot 5/2 = 16.\cr
\end{aligned}$$
Conversely, such a surface has an ample divisor class of self-intersection $16$. To see this, note that equality is attained in the above with $a = 3, b = 3, c_i = 1$, and this gives an ample divisor class by [@huybrechts Corollary 8.1.7].
Thus an ample divisor class cannot give an embedding into ${\mathbb{P}}^n$ for $n < 9$.
Second example {#finite-by-reflections}
--------------
Now we give an example that is perhaps more surprising: a K3 surface $Y/{\mathbb{Q}}$ for which $\operatorname{Aut}Y$ is finite even though, for all field extensions $L/{\mathbb{Q}}$, a K3 surface over $\bar {\mathbb{Q}}$ whose Picard lattice is isomorphic to $\operatorname{Pic}Y_L$ would have infinite automorphism group.
Let $$M = \begin{pmatrix}10&0\cr 0&-4\cr \end{pmatrix}, \quad
N = \begin{pmatrix}10&0&0\cr 0&-2&0\cr 0&0&-2\cr\end{pmatrix},$$ and let $L_M, L_N$ be the lattices with Gram matrices $M, N$ respectively.
We will choose $Y$ to be a K3 surface whose Picard lattice over ${\mathbb{Q}}$ is isomorphic to $L_M$, while over $\bar {\mathbb{Q}}$, and indeed over a certain quadratic extension $K_Y$, the Picard lattice is isomorphic to $L_N$. The Galois group will act through the quotient ${\mathbb{Z}}/2{\mathbb{Z}}$ by exchanging the second and third generators.
\[prop:describe-u\] Let $Y$ be a K3 surface in ${\mathbb{P}}^4_{\mathbb{Q}}$ given as the intersection of a quadric and a cubic, containing two disjoint Galois-conjugate conics $C_1, C_2$ defined over a quadratic field $K_Y$, and having Picard number $3$ over $\bar {\mathbb{Q}}$. Then $\operatorname{Pic}Y \cong L_M$ and $\operatorname{Pic}Y_{\bar {\mathbb{Q}}} \cong L_N$.
Let $H$ be a hyperplane section. Then the divisors $H + C_1 + C_2, C_1, C_2$ have the intersection matrix $N$, whose determinant is $-40$. Thus we have an embedding $L_N \hookrightarrow \operatorname{Pic}Y_{\bar {\mathbb{Q}}}$ whose image has index $1$ or $2$. If it is $2$, then either $[H] + [C_1]$ or $[H] + [C_2]$ can be divided by $2$. If $[H] + [C_1] \sim 2D$, then, letting $\sigma$ be an extension of the nontrivial automorphism of the field of definition of $C_1$ to that of $D$, we have $[H] + [C_2] \sim 2D^\sigma$, and so both classes can be divided by $2$ and the index is a multiple of $4$, a contradiction. Similarly, $[H] + [C_2]$ cannot be divided by $2$. We conclude that the index is $1$: that is, $\operatorname{Pic}Y_{\bar {\mathbb{Q}}} \cong L_N$. There are Galois-invariant divisors of the classes $[H] + [C_1] + [C_2], [C_1] + [C_2]$, and these span the invariant subspace, so they generate $\operatorname{Pic}Y$. Their intersection matrix is $M$.
\[rem:bdry-ample\] The divisor class $[H]$ is very ample, because it is the hyperplane class on the smooth projective surface $Y$. The divisor class $D = [H]+[C_1]+[C_2]$ is not ample, but we will show that it is nef. Indeed, for any irreducible curve $C$ other than $C_1$ and $C_2$, the intersection numbers $H \cdot C$, $C_1 \cdot C$, $C_2 \cdot C$ are all non-negative, while $D \cdot [C_1]$ and $D \cdot [C_2]$ are both zero. In fact, one can show that $D$ is the hyperplane class for a model of $Y$ as a surface of degree $10$ in ${\mathbb{P}}^6$ with two ordinary double points ($A_1$ singularities).
\[prop:aut-y-z-infinite\]
1. \[itemLM\] The group ${{\mathrm{O}}}(L_M)$ is infinite. If $\alpha \in {{\mathrm{O}}}(L_M)$ has infinite order and $A$ is a normal subgroup of ${{\mathrm{O}}}(L_M)$ containing $\alpha$ then ${{\mathrm{O}}}(L_M)/A$ is finite.
2. \[item:Y\] Let $V$ be a K3 surface over an algebraically closed field having Picard lattice isomorphic to $L_M$. Then $\operatorname{Aut}V$ is infinite.
3. \[item:Z\] Let $W$ be a K3 surface over an algebraically closed field having Picard lattice isomorphic to $L_N$. Then $\operatorname{Aut}W$ is infinite.
We first prove (\[itemLM\]). To find ${{\mathrm{O}}}(L_M)$, we consider the quadratic form $10x^2 - 4y^2 = -N(2y + \sqrt{10} x)$ associated to $M$. Its automorphism group is generated by the sign changes $(x,y) \to (x,-y)$ and $(x,y) \to (-x,y)$ and by multiplication by a generator of the group of totally positive units of ${{\mathrm{O}}}_{{\mathbb{Q}}(\sqrt{10})}$. (This generator is $19 + 6 \sqrt{10}$ and it takes $(x,y)$ to $(19x+12y,30x+19y)$.) Thus ${{\mathrm{O}}}(L_M)$ has a subgroup of finite index isomorphic to ${\mathbb{Z}}$, so the quotient by any infinite normal subgroup is finite.
We continue with (\[item:Y\]). Working mod $5$ we see that $\operatorname{Pic}V$ has no vectors of norm $-2$, whence $V$ has no rational curves. Thus $\operatorname{Aut}V$ coincides up to finite index with the infinite group ${{\mathrm{O}}}(\operatorname{Pic}V) = {{\mathrm{O}}}(L_M)$: this follows from Theorem \[thm:old\] (2), because $W(\operatorname{Pic}V)$ is trivial.
We now turn to (\[item:Z\]). Let $D_1, D_2, D_3$ be divisors on $W$ whose intersection matrix is $N$, and let $D = D_1-2D_2-D_3$. Since $D^2 = 0$ and $D$ is primitive, for some $\alpha \in {{\mathrm{O}}}(L_N)$ there is a genus-$1$ fibration $\pi$ with fibres of class $\alpha(D)$. There is no section of $\pi$ (indeed, no two curves on $W$ have odd intersection), but there is a $2$-section. The Jacobian $J$ of $\pi$ is a K3 surface of Picard number $3$ [@huybrechts Corollary 11.4.7] on which the determinant of the intersection pairing is $(\det N)/2^2 = -10$ [@huybrechts Equation (11.4.5)]. As in Proposition \[prop:aut-x-inf\], this shows that $J$ has no reducible fibres: the non-identity component of a reducible fibre would be orthogonal to the classes of both a fibre and the zero-section, and have self-intersection $-2$, which is incompatible with the required determinant. The Shioda–Tate formula then shows that the Mordell–Weil group of $J$ has rank $1$. This Mordell–Weil group acts faithfully on $W$, so it follows as in Proposition \[prop:aut-x-inf\] that $\operatorname{Aut}W$ is infinite.
Proposition \[prop:aut-y-z-infinite\] states that any K3 surface over an algebraically closed field having Picard lattice isomorphic to either $L_M$ or $L_N$ has infinite automorphism group. In contrast, the following proposition shows that a K3 surface over ${\mathbb{Q}}$ having Picard lattice isomorphic to $L_M$ or $L_N$ over any algebraic extension of ${\mathbb{Q}}$ may have finite automorphism group. Indeed, the last part of this section will be devoted to finding an example of such a surface.
\[prop:auts-finite\] Let $Y$ be a K3 surface as in Proposition \[prop:describe-u\]. Then $\operatorname{Aut}Y$ is finite.
Let $D_1 = [H] + [C_1] + [C_2]$ and $D_2 = [C_1] + [C_2]$, so that $D_1, D_2$ have the intersection matrix $M$. Since $C_1, C_2$ are disjoint conjugate rational curves, the product $r_1$ of the reflections in their classes is in our reflection group $R_Y$ (Definition \[defn:rx\]). In the given basis, $r_1$ acts on $\operatorname{Pic}Y$ with matrix $$A_1 = \begin{pmatrix}1&0\cr 0&-1\cr\end{pmatrix}.$$
We now show that there is a smooth rational curve in the class $F = 6D_1 - 9[C_1] - 10[C_2]$, defined over the same field $K_Y$ as the $C_i$. The class $F$ is effective, since it has self-intersection $-2$ and positive intersection with the very ample divisor class $[H]$. An irreducible curve of self-intersection $-2$ must be rational and smooth, so it suffices to show that $F$ has no nontrivial expression as a sum of effective classes.
Let $E$ be an effective divisor with $[E] = aD_1 - b_1 [C_1] - b_2 [C_2]$. By Remark \[rem:bdry-ample\], $D_1$ is nef, so $0 \le [E] \cdot D_1 = 10a$, and therefore $a \ge 0$. Moreover, if $E$ is irreducible, then the inequality $E^2 \geq -2$ yields $10a^2\geq 2b_1^2+2b_2^2-2$, and if furthermore we have $a>0$, then the inequality $C_i \cdot E \geq 0$ gives $b_i \geq 0$.
We write the class $F$ as a sum $F = \sum_{i=1}^s [E_i]$ of classes of irreducible effective divisors $E_1, \ldots, E_s$, and we write $[E_i] = a_iD_i - b_{i,1}[C_1] - b_{i,2}[C_2]$. We saw above that $a_i \geq 0$ for each $i$, so from $\sum_i a_i=6$, we find $a_i \leq 6$ for each $i$. Furthermore, from $\sum_i (b_{i,1}+b_{i,2}-3a_i) = 9+10 - 3\cdot 6 = 1 >0$, we conclude that there is an $i$ with $b_{i,1}+b_{i,2} > 3a_i$, and hence $b_{i,1}+b_{i,2} \geq 3a_i+1$. Again from the above, we find for such $i$ that $$10a_i^2 \geq 2b_{i,1}^2+2b_{i,2}^2-2 \geq (b_{i,1}+b_{i,2})^2 - 2 \geq (3a_i+1)^2-2,$$ which implies $a_i \geq 6$, so $a_i = 6$. The left- and right-hand sides of the sequence of inequalities above then differ by only $1 = 360-359$, so we also find $b_{i,1}+b_{i,2} = 3a_i+1=19$ (not $-19$ because $b_{i,j} \ge 0$) and $(2b_{i,1}^2+2b_{i,2}^2) - (b_{i,1}+b_{i,2})^2 \leq 1$. Equivalently, $(b_{i,1}-b_{i,2})^2 \leq 1$, which yields $[E_i] = F$ or $[E_i] = F' = 6D_1 - 10[C_1]-9[C_2]$. But $F - [E_i]$ is effective by definition, and $F - F' = [C_1] - [C_2]$ is not effective (the intersection with the ample divisor $H$ is $0$), so $[E_i] = F$ and $F$ is indeed the class of an irreducible curve, say $R$.
Let $\sigma$ be an automorphism of $\bar {\mathbb{Q}}$ that exchanges $C_1$ and $C_2$. Then the class $F^\sigma=[R^\sigma]$ is $6D_1 - 10[C_1] - 9[C_2]$, and one checks immediately that $F \cdot F^\sigma = 0$. The product $r_2$ of the reflections in $F$ and $F^\sigma$ has matrix $$A_2 = \begin{pmatrix}721&456 \cr -1140&-721\cr \end{pmatrix}$$ with respect to the basis $(D_1,D_2)$ of $\operatorname{Pic}Y$. Now, $r_1 r_2$ is an element of $R_Y$ of infinite order. From Proposition \[prop:aut-y-z-infinite\] (\[itemLM\]), it follows that the quotient ${{\mathrm{O}}}(\operatorname{Pic}Y)/R_Y$ is finite and hence that $\operatorname{Aut}Y$ is finite.
Given equations for $Y$, it is quite practical to make the curve $R$ used in the proof explicit. For any variety $V \subset {\mathbb{P}}^4$, let $I(V)$ be the ideal in the homogeneous coordinate ring of ${\mathbb{P}}^4$ of polynomials that vanish on $V$. Then the dimension of the degree-$6$ part of $((I(Y)+I(C_1)^3) \cap (I(Y)+I(C_2)^4))/I(Y)$ is $1$; let $f_6$ be a generator, so that the divisor cut out on $Y$ by $f_6$ is of the form $C + 3C_1 + 4C_2$. Using Magma one checks that $C$ is indeed an irreducible curve of arithmetic genus $0$.
The numerical properties of $F$ follow from the fact that $(19,6)$ is a solution to the Pell equation $x^2 - 10y^2 = 1$. More generally, let $k$ be even and suppose that $D_1, [C_1], [C_2]$ are classes of pairwise disjoint divisors on a surface with $D_1^2 = k$ and $C_1^2 = C_2^2$, where $C_1$ and $C_2$ constitute a Galois orbit. Suppose further that $m^2 - kn^2 = 1$: clearly $m$ is odd. Let $D$ be the divisor class $nD_1 - \lfloor \frac{m}{2} \rfloor [C_1] - \lceil \frac{m}{2} \rceil [C_2]$ and let $D^\sigma$ be its Galois conjugate. Then $D^2 = kn^2 - 2 \left(\frac{m-1}{2}\right)^2 - 2 \left(\frac{m+1}{2}\right)^2 = -2$ and $D \cdot D^\sigma = kn^2 - 4\frac{(m-1)(m+1)}{4} = 0$.
The remainder of this section will be devoted to constructing a K3 surface $Y$ satisfying the conditions of Proposition \[prop:describe-u\]. The verification will be more complicated than that of Example \[ex:construct-t\], because the Picard number over $\bar {\mathbb{Q}}$ in this example is odd, so that the reduction map cannot induce an isomorphism of $\operatorname{Pic}Y_{\bar {\mathbb{Q}}}$ to $\operatorname{Pic}Y_{\bar {\mathbb{F}}_p}$. We use the following proposition.
\[prop:how-to-find-u\] Let $Y$ be a smooth intersection of a quadric and a cubic in ${\mathbb{P}}^4({\mathbb{Q}})$ containing two disjoint conics $C_1, C_2$ that are defined and conjugate over a quadratic field $K_Y$. Let $p \ne 2, q$ be primes such that:
1. $Y$ has good reduction at $p$ and $q$;
2. the reduction $Y_{{\mathbb{F}}_p}$ contains a line $L$ disjoint from the reductions of $C_1, C_2$, and its Picard group over $\bar{\mathbb{F}}_p$ has rank $4$;
3. the reduction $Y_{{\mathbb{F}}_q}$ base changed to $\bar {\mathbb{F}}_q$ contains no lines disjoint from the reductions of $C_1$ and $C_2$.
Then $Y_{\bar {\mathbb{Q}}}$ has Picard number $3$ and hence $Y$ satisfies the conditions of Proposition \[prop:describe-u\].
To speak of $Y_{{\mathbb{F}}_p}$ we first need to choose a model of $Y$ over ${\mathbb{Z}}$. When we give equations with integral coefficients for $Y$ over ${\mathbb{Q}}$, it is understood that the model over ${\mathbb{Z}}$ is defined by the same equations. In each case it may be checked that the ideal defining this model is generated by the intersection of the ideal defining $Y$ over ${\mathbb{Q}}$ with the ring of polynomials in the same variables with integral coefficients.
Let $s_p$ be the specialization homomorphism $\operatorname{Pic}Y_{\bar {\mathbb{Q}}} \to \operatorname{Pic}Y_{\bar {\mathbb{F}}_p}$, which is injective by [@rvl Proposition 6.2]. Then by [@elsenhans-jahnel Theorem 1.4] the cokernel of $s_p$ is torsion-free. Note that specialization preserves intersection numbers [@fulton Section 20.3]. If $s_p$ is surjective, the class of $L$ is in the image. The class $L_0$ with $s_p(L_0) = [L]$ has positive intersection with the hyperplane class and self-intersection $-2$, so it is effective; its intersection with the hyperplane class is $1$, so it is the class of a line on $Y_{\bar {\mathbb{Q}}}$. We have $[L] \cdot s_p(C_i) = 0$, so $L_0 \cdot C_i = 0$ and $s_q(L_0) \cdot s_q(C_i) = 0$ as well. This contradicts hypothesis (3). It follows that $\operatorname{rk}\operatorname{Pic}Y_{\bar {\mathbb{Q}}} < \operatorname{rk}\operatorname{Pic}Y_{\bar {\mathbb{F}}_p} = 4$. On the other hand, the hyperplane class in ${\mathbb{P}}^4$ and the classes of $C_1, C_2$ are independent, so $\operatorname{rk}\operatorname{Pic}Y_{\bar {\mathbb{Q}}} \ge 3$.
In our construction, we will use $p = 3, q = 5$. We will take the conics $C_1, C_2$ to be defined over ${\mathbb{Q}}(\sqrt{19})$, the smallest quadratic field in which both $p$ and $q$ split. Let $\rho = \sqrt{19}$, and denote the coordinates on ${\mathbb{P}}^4$ by $x_0, \dots, x_4$. We choose the conic $C_1$ defined by $$\begin{aligned}
&(-2\rho + 2)x_0 + x_2 + (\rho - 2)x_3 + (-\rho - 2)x_4 = \\
&(2\rho + 1)x_0 + (\rho + 2)x_1 + 2x_2 + (\rho + 2)x_3 + (-\rho + 2)x_4 = \\
& \quad (-\rho - 2)x_0^2 + (-2\rho + 1)x_0x_1 + (-2\rho - 1)x_1^2 + (-\rho + 2)x_0x_2 + (-\rho + 2)x_1x_2 \\
& \qquad + (\rho - 2)x_2^2 + (2\rho + 1)x_0x_3 + (2\rho - 2)x_1x_3 + (2\rho + 2)x_2x_3 - 2x_3^2 \\
& \qquad + (\rho + 2)x_0x_4 + (-\rho + 2)x_1x_4 + (-\rho + 1)x_2x_4 - x_3x_4 + 2x_4^2 = 0
\end{aligned}$$ and let $C_2$ be its Galois conjugate. Next, let $L$ be the line in ${\mathbb{P}}^4({\mathbb{F}}_3)$ through $(1:1:2:0:1),(2:1:1:1:0)$ and let $C_{3,1}, C_{3,2}$ be the reductions of $C_1$ and $C_2$, respectively, at one of the two primes above $3$.
\[prop:char-3\] The surface $Y_3$ in ${\mathbb{P}}^4({\mathbb{F}}_3)$ defined by $F_3=G_3=0$, where $$\begin{aligned}
F_3 &= 2x_0x_1 + x_1x_2 + x_2^2 + 2x_0x_3 + x_1x_3 + x_3^2 + 2x_0x_4 + x_3x_4
+ 2x_4^2, \\
G_3 &= 2x_0x_1^2 + 2x_0^2x_2 + 2x_0x_1x_2 + 2x_1^2x_2 + 2x_0x_2^2 + 2x_1x_2^2 + 2x_2^3 + 2x_0x_1x_3 + x_1^2x_3 \\
& \quad + 2x_0x_2x_3 + 2x_1x_2x_3 + 2x_0x_3^2 + 2x_2x_3^2 + 2x_3^3 + x_0^2x_4 +
2x_1^2x_4 + 2x_0x_2x_4 \\
& \quad + x_2^2x_4 + 2x_0x_3x_4 + 2x_2x_3x_4 +
x_3^2x_4 + 2x_0x_4^2 + x_1x_4^2 + x_2x_4^2 + x_4^3,
\end{aligned}$$ is smooth, contains $L, C_{3,1}$, and $C_{3,2}$, and has Picard number $4$.
The verification that $Y_3$ is smooth and contains the given curves is routine. To see that $Y_3$ has Picard number $4$, we proceed as in Example \[ex:construct-t\]. In order to count the number of points on $Y_3$ over ${\mathbb{F}}_{3^n}$ for $1\leq n \leq 9$, we write $Y_3$ as an elliptic surface using the fibration of Picard class $H - C_{3,1}$, for which $L$ is a section, and we sum the number of points of the fibres. We obtain point counts as in Table \[counts2\].
$n$ $\#Y_3({\mathbb{F}}_{3^n})$
----- -----------------------------
$1$ $18$
$2$ $104$
$3$ $846$
$4$ $6776$
$5$ $59658$
$6$ $532694$
$7$ $4790811$
$8$ $43068056$
$9$ $387398079$
: Point counts on the surface $Y_3$[]{data-label="counts2"}
As in Example \[ex:construct-t\], we find that the trace of the $n$th power of Frobenius acting on ${\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{Y}_3,{\mathbb{Q}}_\ell(1))$ is equal to $\#Y_3({\mathbb{F}}_{3^n})/3^n - 3^{n} - 3^{-n}$, while on the subspace $V \subset {\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{Y}_3,{\mathbb{Q}}_\ell(1))$ generated by $H, C_{3,1}, C_{3,2}$, and $L$, the trace is equal to $4$; hence, on the $18$-dimensional quotient $Q={\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{Y}_3,{\mathbb{Q}}_\ell(1))/V$, the trace equals $\#Y_3({\mathbb{F}}_{3^n})/3^n - 3^{n} - 3^{-n} - 4$. Again, these traces, together with the Newton identities, determine the coefficients of the characteristic polynomial $f$ of Frobenius acting on $Q$. The point counts in Table \[counts2\] yield the first half of the coefficients of $f$, including the middle coefficient, which turns out to be nonzero. This implies that the sign in the functional equation $t^{18}f(1/t) = \pm f(t)$ is $+1$, so this functional equation determines $f$, which we calculate to be $$\begin{aligned}
f = \tfrac{1}{3} \big( 3t^{18} &- 4t^{17} + 5t^{16} - 4t^{15} + 4t^{14} - 4t^{13} + 5t^{12} - 5t^{11} + 5t^{10} \\
&- 6t^9 + 5t^8 - 5t^7 + 5t^6 - 4t^5 + 4t^4 - 4t^3 + 5t^2 - 4t + 3 \big). \end{aligned}$$ As the characteristic polynomial of Frobenius acting on $V$ is $(t-1)^4$, we find that the characteristic polynomial of Frobenius acting on ${\mathrm{H}}^2_{\textrm{{\'e}t}}(\bar{Y}_3,{\mathbb{Q}}_\ell(1))$ is $(t-1)^4f$. The polynomial $3f$ is irreducible, primitive and not monic, so its roots are not roots of unity. Thus we obtain an upper bound of $4$ for the Picard rank of $\bar{Y}_3$. The rank is equal to $4$ as the Picard group contains the linearly independent classes of $H$, $L$, $C_{3,1}$, and $C_{3,2}$.
Now let $C_{5,1}, C_{5,2}$ be the reductions of $C_1$ and $C_2$, respectively, at one of the two primes above $5$.
\[prop:char-5\] Let $Y_5$ be the surface in ${\mathbb{P}}^4_{{\mathbb{F}}_5}$ defined by $F_5=G_5=0$, where $$\begin{aligned}
F_5 &= 2x_0^2 + x_0x_1 + 3x_1^2 + 2x_0x_2 + 2x_2^2 + 2x_1x_3 + 2x_2x_3 +
3x_3^2 + 3x_0x_4 + 2x_2x_4, \\
G_5 &= x_0^2x_1 + 3x_0x_1^2 + 2x_1^3 + 4x_0x_1x_2 + x_1^2x_2 + 2x_1x_2^2 +
3x_2^3 + 2x_0^2x_3 + 4x_1^2x_3 \\
& \quad + 2x_0x_3^2 + 2x_1x_3^2 +
2x_2x_3^2 + 2x_3^3 + x_0x_1x_4 + 4x_1^2x_4 + x_0x_2x_4 +
2x_2^2x_4 \\
& \quad + 3x_1x_3x_4 + 2x_2x_3x_4 + 3x_3^2x_4 + 3x_0x_4^2 +
2x_1x_4^2 + 3x_2x_4^2 + x_3x_4^2.
\end{aligned}$$ Then $Y_5$ is smooth. It contains $C_{5,1}$ and $C_{5,2}$, but no lines over $\bar {\mathbb{F}}_5$ that are disjoint from $C_{5,1}$.
Again, it is easy to check that $Y_5$ is smooth and that the $C_{5,i}$ are on $Y_5$. It is also easy to verify directly that $Y_5$ contains no lines defined over ${\mathbb{F}}_5$ by checking that no line through two distinct ${\mathbb{F}}_5$-points of $Y_5$ is contained in $Y_5$; however, this is not sufficient. Suppose that $L$ is a line contained in $Y_5$ and disjoint from $C_{5,1}$. We embed $Y_5$ in ${\mathbb{P}}^5_{{\mathbb{F}}_5}$ by the divisor class $H + [C_{5,1}]$, where $H$ is the hyperplane class in ${\mathbb{P}}^4_{{\mathbb{F}}_5}$. In this embedding, $L$ is still a line, because $L \cdot (H+[C_{5,1}]) = 1$.
The image is a surface $T_5$ defined by three quadrics $Q_1, Q_2, Q_3$. For each $Q_i$, let $Z_i$ be the subscheme of ${\mathbb{P}}^5 \times {\mathbb{P}}^5$ consisting of pairs of points $P, P'$ for which $P = P'$ or the line joining $P$ to $P'$ is on the variety defined by $Q_i = 0$. The image of $Z_i$ in ${\mathbb{P}}^{14}$ under the Pl" ucker embedding is isomorphic to the Fano scheme of lines on $Q_i$. It is easy to check in Magma that the intersection of the images of the $Z_i$ in ${\mathbb{P}}^{14}$ is empty, which implies that there are no lines on $T_5$ even over $\bar {\mathbb{F}}_5$.
In fact, we claim that $Y_5$ contains no lines at all. Indeed, a calculation similar to the one in the proof of Proposition \[prop:char-5\] shows that there are also no lines on $Y_5$ disjoint from $C_{5,2}$. To finish the proof of the claim, it suffices to show that there are no lines that meet both $C_{5,1}$ and $C_{5,2}$. Let $Z_2$ be the closure of the subset of ${\mathbb{P}}^4_{{\mathbb{F}}_5} \times {\mathbb{P}}^4_{{\mathbb{F}}_5}$ consisting of all pairs of points $(x,y)$ such that $x \neq y$ and the line through $x$ and $y$ lies on the hypersurface defined by $F_5$. Define $Z_3$ similarly for $G_5$. Since $C_{5,1}$ and $C_{5,2}$ are disjoint, the intersection $Z_2 \cap Z_3 \cap (C_{5,1} \times C_{5,2})$ consists of all pairs $(x,y)$ with $x \in C_{5,1}$ and $y\in C_{5,2}$ for which there is a line on $Y_5$ through $x$ and $y$. We then check in Magma that this intersection is empty. This is done in our Magma scripts [@script].
In principle one can prove directly that $Y_5$ has no lines by checking that the images of $Z_2$ and $Z_3$ under the Pl" ucker embedding in ${\mathbb{P}}^9$ are disjoint, but this calculation is very slow.
It remains to show that there exists a surface $Y \subset {\mathbb{P}}^4_{\mathbb{Q}}$ containing the conics $C_1,C_2$ and lifting the surfaces $Y_3,Y_5$ described above. Denote by $V_2(C_1 \cup C_2, {\mathbb{Q}})$ the vector space of homogeneous forms of degree $2$, with coefficients in ${\mathbb{Q}}$, vanishing on $C_1 \cup C_2$. One easily computes that $V_2(C_1 \cup C_2, {\mathbb{Q}})$ has dimension $5$; let $q_1, \dotsc, q_5$ be a ${\mathbb{Z}}$-basis for the free abelian group $V_2(C_1 \cup C_2, {\mathbb{Z}})$ of forms having integer coefficients. One also easily computes that the analogous spaces $V_2(C_{3,1} \cup C_{3,2}, {\mathbb{F}}_3)$ and $V_2(C_{5,1} \cup C_{5,2}, {\mathbb{F}}_5)$ both have dimension $5$ over ${\mathbb{F}}_3$ and ${\mathbb{F}}_5$ respectively, which implies that the images of $q_1, \dotsc, q_5$ span these spaces. Now the Chinese remainder theorem guarantees the existence of a quadratic form $F \in V_2(C_1 \cup C_2, {\mathbb{Z}})$ satisfying both $F \equiv F_3 \pmod 3$ and $F \equiv F_5 \pmod 5$. Similarly, the analogous spaces of cubic forms vanishing on $C_1 \cup C_2$ are all of dimension $21$, which implies the existence of a cubic form $G$ satisfying $G \equiv G_3 \pmod 3$ and $G \equiv G_5 \pmod 5$. Remark \[syzygies\] below shows that one can avoid explicit computations to calculate these dimensions. The surface $Y$ given by $F=G=0$ reduces modulo $3$ and $5$ to $Y_3$ and $Y_5$, respectively.
\[syzygies\] Let $k$ be any field. Let $C_1,C_2 \subset {\mathbb{P}}^4_k$ be two conics; suppose that the two planes spanned by $C_1,C_2$ respectively meet only in one point $P$, and suppose that $P$ does not lie on either $C_1$ or $C_2$. Then the space of forms of degree $d$ vanishing on $C_1$ and $C_2$ has dimension $$4\binom{d+2}{4} - 4\binom{d+1}{4} + \binom{d}{4} + 2\binom{d}{2} - 1 =
\tfrac{1}{24}d(d-1)(d^2+11d+46) - 1.$$
Finite automorphism group for a family of diagonal surfaces {#finite-diagonal}
-----------------------------------------------------------
In this example, for which we summarize the calculations in a Magma script [@script], we consider a quartic surface $Z$ over a field $k$ of characteristic $0$ defined by $$x^4 - y^4 = c(z^4 - w^4).$$ for some nonzero $c \in k$. Conjecture 2.5 of [@vluijk] states that if $k$ is the field ${\mathbb{Q}}$ of rational numbers, then for every nonzero $c \in {\mathbb{Q}}$, the set of rational points on the associated surface $Z$ is dense. The $48$ lines on a surface in this family are defined over the field $k(\zeta_8, \root 4 \of c)$. We assume that $k, c$ are such that this is a Galois extension of $k$ of degree $16$: with $k = {\mathbb{Q}}$, this is equivalent to $|c|$ being neither a square nor twice a square. (This Galois action is case A45 of [@bright Appendix A].) We will prove under this assumption that $\operatorname{Aut}Z$ is finite. The argument will proceed by the following steps.
1. Determination of $\operatorname{Pic}Z$, and comparison to the Picard lattice of a K3 surface $V$ with finite automorphism group over $\bar k$.
2. Determination of ${{\mathrm{O}}}(\operatorname{Pic}V)$ and ${{\mathrm{O}}}(\operatorname{Pic}Z)$ up to finite index.
3. Construction of elements of $R_Z$ and verification that ${{\mathrm{O}}}(\operatorname{Pic}Z)/R_Z$ is finite.
If the characteristic of $k$ is congruent to $1 \bmod 4$, then $k$ contains the $4$th roots of $1$ and this Galois action is not possible. If it is $3 \bmod 4$, then $Z$ is a supersingular surface and $Z_{\bar k}$ has Picard number $22$, so the argument given here does not apply. If $k$ has characteristic $2$, then of course $Z$ is not a smooth surface.
Let $L_N$ be the lattice with Gram matrix $$N = \left( \begin{array}{cc|c} 0&1\\1&0\\ \hline &&-2I_4 \end{array}\right),$$ as in Section \[finite-from-overlattice\].
\[lem:pic-diagonal\] There is a sublattice of $L_N$ of index $4$ that is isomorphic to $\operatorname{Pic}Z$.
The Picard group of $\bar{Z}$ is generated by the classes of the $48$ lines. (See [@ssvl Section 6.1] for a history of proofs of this fact.) One easily verifies that the subspace fixed by the action of Galois is of rank $6$ and that the intersection form has discriminant $-256$. Since the surface has rational points such as $(1,1,0,0)$, the fixed subspace of the Picard group of $\bar{Z}$ is in fact the Picard group of $Z$.
Let $\Lambda$ be the lattice obtained from $\operatorname{Pic}Z$ by adjoining $v/2$ for each $v$ in $$\{\,\, v \in \operatorname{Pic}Z \,\, : \,\, 8\mid (v,v) \mbox{ and } 2\mid (v,w) \mbox{ for all } w \in \operatorname{Pic}Z \,\, \}.$$ Then $\Lambda$ contains $\operatorname{Pic}Z$ with index $4$ and hence has discriminant $-16$. It is verified in [@script] that $L_N$ and $\Lambda$ are isomorphic.
We now proceed to step $2$, determining ${{\mathrm{O}}}(L_N)$. Let $V$ be a K3 surface over $\bar k$ with Picard lattice $L_N$, such as those constructed in Proposition \[prop:four-a1\]. We choose a basis $e_1, \dots, e_6$ for $L_N \cong \operatorname{Pic}V$ to consist of the classes $E, E+O, C_1, C_2, C_3, C_4$, where $E$ is the class of an elliptic fibration with zero section $O$ and the $C_i$ are the components of the reducible fibres that do not meet $O$. In this basis, the Gram matrix of $\operatorname{Pic}V$ is as above. Let $G \subset {{\mathrm{O}}}(L_N)$ be the image of the symmetric group ${{\mathcal{S}}}_4$ by the homomorphism taking a permutation $\pi$ to the automorphism fixing $E$ and $E+O$ and taking $C_i$ to $C_{\pi(i)}$. Let $A_{L_N}$ be the discriminant group of $L_N$ with its discriminant form, a quadratic form with values in ${\mathbb{Q}}/2{\mathbb{Z}}$ [@huybrechts Chapter 14].
\[outer-auts\] $A_{L_N}$ is generated by the classes of $e_i/2$ for $3 \le i \le 6$. The discriminant form takes $\sum_{i=3}^6 a_ie_i/2$ to $-\sum_{i=3}^6 a_i^2/2 \bmod 2$, and the natural map $G \to \operatorname{Aut}A_{L_N}$ is an isomorphism.
The $e_i/2$ for $3 \le i \le 6$ belong to $L_N^*$ and they generate a group that contains $L_N$ with index $16$. Since $L_N$ has discriminant $16$, they generate $L_N^*$ and their classes generate $A_{L_N}$. The quadratic form takes $\sum_{i=3}^6 a_ie_i/2$ to $\sum_{i=3}^6 a_i^2 e_i^2/4 \bmod 2 = -\sum_{i=3}^6 a_i^2/2$. It follows that the only elements of $A_{L_N}$ taken to $-1/2$ by the quadratic form are the basis vectors, so every automorphism of $A_{L_N}$ must permute these. Conversely it is clear that every permutation of the basis vectors extends to an automorphism of $A_{L_N}$ and that this automorphism is the image of a unique element of $G$.
Recall [@huybrechts Definition 15.1.1] that an automorphism of $V$ is *symplectic* if it acts trivially on ${\mathrm{H}}^0(V,\Omega^2_V)$.
The group $\operatorname{Aut}V$ is finite cyclic and the subgroup $\operatorname{Aut}_s V$ of symplectic automorphisms of $V$ is trivial.
By Lemma \[nine-curves\], there are exactly nine smooth rational curves on $V$. Every automorphism of $V$ must preserve the configuration of these curves. Exactly one of them, the section of the elliptic fibration, intersects four of the others. Hence, the only automorphisms of the configuration are those that permute the reducible fibres. Since the nine curves generate $\operatorname{Pic}V$, we conclude that every automorphism of $V$ acts on $\operatorname{Pic}V$ by elements of $G$. Now, a symplectic automorphism $\alpha$ of a K3 surface $V$ in characteristic $0$ acts as the identity on the transcendental lattice in ${\mathrm{H}}^2(V,{\mathbb{Z}})$ [@huybrechts Remark 15.1.2] and hence on its discriminant group, so in addition it acts trivially on the discriminant group of $\operatorname{Pic}V$ by [@huybrechts Lemma 14.2.5]. Since $\alpha$ acts on $\operatorname{Pic}V$ through $G$, combining this with the last lemma shows that $\alpha$ acts trivially on $\operatorname{Pic}V$. It follows that $\alpha$ acts trivially on ${\mathrm{H}}^2(V,{\mathbb{Z}})$; it is therefore the identity by [@huybrechts Corollary 15.1.6].
Since we are in characteristic $0$, there is an exact sequence [@huybrechts (15.1.3)] $$1 \to \operatorname{Aut}_s V \to \operatorname{Aut}V \to \mu_m \to 1$$ for some $m$, where $\operatorname{Aut}_s V$ is the group of symplectic automorphisms that we have just shown to be trivial. It follows that $\operatorname{Aut}V$ is finite cyclic.
There is always an automorphism of order $2$ of $V$ given by negation of the elliptic fibration with respect to the unique section. For very general $V$ the order of $\operatorname{Aut}V$ is $2$, but it might be possible for $\operatorname{Aut}V$ to be of order $m = 4, 6$, or $8$ if there is an automorphism that permutes the reducible fibres of the fibration nontrivially. In this case the field of definition of $\operatorname{Pic}V$ would contain $\mu_m$, which does not happen generically.
By [@dolgachev Proposition 5.10], we know that ${{\mathrm{O}}}(L_N)$ is generated by the automorphisms of the ample cone and the reflections in the classes of rational curves, together with $-1$; the automorphisms of the ample cone are just the $G$ of Lemma \[outer-auts\]. Since $G$ is generated by $2$ elements and has $3$ orbits on the set of rational curves, we have a set of $6$ generators for ${{\mathrm{O}}}(L_N)$.
To pass to ${{\mathrm{O}}}(\operatorname{Pic}Z)$, we use a simple lemma.
\[commensurable\] Let ${\Lambda}$ be a lattice with sublattices ${\Lambda}_1, {\Lambda}_2$ with $[{\Lambda}_1:{\Lambda}_1 \cap {\Lambda}_2]$ finite. Then the stabilizer of ${\Lambda}_1 \cap {\Lambda}_2$ in ${{\mathrm{O}}}({\Lambda}_1)$ has finite index. If in addition ${\Lambda}_1, {\Lambda}_2$ have equal rank, then ${{\mathrm{O}}}({\Lambda}_1) \cap {{\mathrm{O}}}({\Lambda}_2)$ has finite index in ${{\mathrm{O}}}({\Lambda}_1)$ and ${{\mathrm{O}}}({\Lambda}_2)$, where the ${{\mathrm{O}}}({\Lambda}_i)$ are regarded as subgroups of ${{\mathrm{O}}}({\Lambda}_1 \otimes {\mathbb{Q}})$.
There are only finitely many sublattices of ${\Lambda}_1$ of index $[{\Lambda}_1:{\Lambda}_1 \cap {\Lambda}_2]$, so the stabilizer is of finite index. If $[{\Lambda}_1:{\Lambda}_1 \cap {\Lambda}_2]$ is finite and ${\Lambda}_1, {\Lambda}_2$ have equal rank then $[{\Lambda}_2:{\Lambda}_1 \cap {\Lambda}_2]$ is finite as well and ${{\mathrm{O}}}({\Lambda}_1) \cap {{\mathrm{O}}}({\Lambda}_2)$ is the stabilizer of ${\Lambda}_1 \cap {\Lambda}_2$ in ${{\mathrm{O}}}({\Lambda}_i)$ for $i \in \{1,2\}$.
The group generated by ${{\mathrm{O}}}({\Lambda}_1)$ and ${{\mathrm{O}}}({\Lambda}_2)$ usually does not contain ${{\mathrm{O}}}({\Lambda}_1)$ with finite index.
Using Magma [@magma] to compute the permutation representation of ${{\mathrm{O}}}(L_N)$ on sublattices with quotient $({\mathbb{Z}}/2{\mathbb{Z}})^2$, we find a set of $30$ generators for ${{\mathrm{O}}}(L_N) \cap {{\mathrm{O}}}(\operatorname{Pic}Z)$. However, some of these are products of others, so we easily reduce to a set of $9$ generators. We now perform step $3$, the construction of elements of $R_Z$.
\[prop:finite-orbits\] There are $14$ Galois orbits of lines and $8$ of conics on $\bar{Z}$ that give finite Coxeter groups.
We list the conics by observing that their Picard classes are of the form $H/2 + M$, where $M \in \frac{1}{2} \operatorname{Pic}\bar{Z}$ is such that $H \cdot M = 0$ and $M^2 = -3$. Since $H^\perp \subset \frac{1}{2} \operatorname{Pic}\bar{Z}$ is a definite lattice, the set of $M$ with these properties is finite. Such a class is represented by a conic if and only if it is integral and not the sum of two classes of lines, both of which are easily checked.
We finish the proof by referring to Proposition \[rx\] for the configurations of curves that give finite Coxeter groups.
It appears that there are no further generators of $R_Z$.
$\operatorname{Aut}Z$ is finite.
We use Proposition \[main-result\]. To do so, we search for relations among the generators of ${{\mathrm{O}}}(L_N) \cap {{\mathrm{O}}}(\operatorname{Pic}Z)$ and the $22$ elements of $R_Z$ that we have found, simply by checking which products of up to $5$ generators of ${{\mathrm{O}}}(L_N) \cap {{\mathrm{O}}}(\operatorname{Pic}Z)$ and up to $4$ of these together with the generators of $R_Z$ give the identity matrix. Magma quickly verifies that these relations are sufficient to show that $\langle {{\mathrm{O}}}(L_N) \cap {{\mathrm{O}}}(\operatorname{Pic}Z),R_Z\rangle/R_Z$ is finite, from which we conclude by Proposition \[main-result\] that $\operatorname{Aut}Z$ is finite.
There are finitely many classes of geometrically irreducible curves on $Z$ of each genus.
This follows by combining the theorem with Corollary \[finite-class-orbits\].
\[hard-to-prove-density\] We conclude that in order to prove that the rational points of $Z$ are Zariski-dense, we do not have an infinite automorphism group $\operatorname{Aut}Z$ to our disposal. In particular, we will not be able to find an elliptic fibration on $Z$ with a section of infinite order, as translation by such a section would be an automorphism of infinite order.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Miodrag Cristian Iovanov\
Department of Algebra, Faculty of Mathematics, University of Bucharest\
Academiei 14, Bucharest, Romania
title: '**Characterisation of PF rings by the Finite Topology on duals of $R$ Modules**'
---
\[section\] \[th\][**Definition**]{}\[th\][**Proposition**]{}\[th\][**Corollary**]{}\[th\][**Lemma**]{}\[th\][**Remark**]{}\[th\][**Example**]{}
-24mm-20mm
> [. In this paper we study the properties of the finite topology on the dual of a module over an arbitrary ring. We aim to give conditions when certain properties of the field case are can be still found here. Investigating the correspondence between the closed submodules of the dual $M^{*}$ of a module $M$ and the submodules of $M$, we prove some characterisations of PF rings: the up stated correspondence is an anti isomorphism of lattices iff $R$ is a PF ring.]{}
[Introduction and preliminaries]{} Let $R$ be an arbitrary (non commutative) ring. We will use the notations ${\rm Hom}_{R}(M,N)$ for the set of $R$ module morphisms from $M$ to $N$ for right modules $M,N$ and ${}_{R}{\rm Hom}(M,N)$ respectively for left modules $M,N$. Also we use $M^{*}={\rm Hom}_{R}(M,R)$ for any right module $M$ and ${}^{*}M={}_{R}{\rm Hom}(M,R)$ for a left module $M$.\
Given two right $R$ modules $M$ and $N$, recall that the finite topology on ${\rm Hom}_{R}(M,N)$ is the linear topology for which a basis of open neighborhoods for $0$ is given by the sets $\{f\in {\rm Hom}_{R}(M,N)\mid f(x_{i})=0,\,\forall\,i\in \{1,\dots,n\}\}$, for all finite sets $\{x_{1},\dots,x_{n}\}\subseteq M$. This is actually the topology induced on ${\rm Hom}_{R}(M,N)$ from ${\rm Hom}_{Set}(M,N)=N^{M}$ which is a product of topological spaces, where $N$ is the topological discrete space on the set $N$. For an arbitrary set $X\subseteq M$ we denote by $X^{\perp}=\{f\in {\rm Hom}_{R}(M,N)\mid f\vert_{X}=0\}$. Denoting by $<X>_{R}$ the $R$ submodule generated by $X$, we obviously have $(<X>_{R})^{\perp}=X^{\perp}$, so we will work with finitely generated submodules $F\leq M$ and the basis of open neighborhoods $\{F^{\perp}\mid F\leq M\,{\rm finitely\,generated}\}$. Also for left $R$ modules $X$ and $Y$ and $U\leq X$ a submodule of $X$ we will denote $U^{\perp}_{{}_{R}{\rm Hom}(M,N)}$ or simply $U^{\perp}=\{g\in {}_{R}{\rm Hom}(X,Y)\mid g\vert_{X}=0\}$ when there is no danger of confusion. If $W\leq {\rm Hom}_{R}(M,N)$ is a subgroup with $M$ and $N$ left $R$ modules we denote $W^{\perp}=\{x\in N\mid f(x)=0,\,\forall\,f\in W\}$. If $N$ is an $R$ bimodule then we consider the left $R$ module structure on ${\rm Hom}_{R}(M,N)$ given by $(r\cdot f)(x)=rf(x)$, for all $x\in M,\,f\in{\rm Hom}_{R}(M,N),\,r\in R$. If $W$ is a (left) submodule in ${\rm Hom}_{R}(M,N)$, then $W^{\perp}$ is a (right) submodule of M.\
For any right module $M$ we denote by $\Phi_{M}$ the right $R$ modules morphism $$M\stackrel{\Phi_{M}}{\longrightarrow}{}^{*}(M^{*})$$ defined by $\Phi_{M}(m)(f)=f(m)$, for all $f\in M^{*}$ and all $m\in M$. Then $\Phi$ is a functorial morphism from $id_{{\cal M}_{R}}$ to the functor ${}^{*}((-)^{*})$.\
Over a field, there is a series of properties involving the orthogonal $F^{\perp}$ for a vector space $V$ and its dual $V^{*}$ which we will state in a more general setting.
\[1\] Let $M,N$ be $R$ modules.\
(i) If $X\subseteq Y$ are submodules of $M$ then $Y^{\perp}\leq X^{\perp}$.\
(ii) If $U\subseteq V$ are subgroups of ${\rm Hom}_{R}(M,N)$ then $V^{\perp}\leq U^{\perp}$.
\[2\] For $M,N$ right $R$ modules we have:\
(i) If $X\leq M$ is a submodule of $M$ then $(X^{\perp})^{\perp}\supseteq X$ and if we denote $\overline{0}$ the class of $0$ in $M/X$ then we have $(\{\overline{0}\}^{\perp})^{\perp}=(X^{\perp})^{\perp}/X$. If $N$ is an injective cogenerator of ${\cal M}_{R}$ then the equality $(X^{\perp})^{\perp}=X$ holds.\
(ii) If $Y\leq {\rm Hom}_{R}(M,N)$ is a (left) submodule of ${\rm Hom}_{R}(M,N)$ then $(Y^{\perp})^{\perp}\supseteq \overline{Y}$ ($\overline{Y}$ is the closure of $Y$ in ${\rm Hom}_{R}(M,N)$). If $N=R$ and $R$ is a left PF ring (${}_{R}R$ is injective and a cogenerator of ${}_{R}{\cal M}$) then the equality $(Y^{\perp})^{\perp}=\overline{Y}$ holds for all modules $M$ and (left) submodules $Y\leq M^{*}$.
\(i) If $x\in X$ then take $f\in X^{\perp}$; then $f(x)=0$ as $f\vert_{X}=0$. We get that $f(x)=0,\,\forall\, f\in X^{\perp}$ so $x\in (X^{\perp})^{\perp}$. Moreover, $\overline{x}\in (\{\overline{0}\}^{\perp})^{\perp}$ if and only if $\tilde{h}(\overline{x})=0,\,\forall\, \tilde{h}:M/X\longrightarrow N$, equivalent to $h(x)=0,\,\forall h\in X^{\perp}$, i.e. $x\in (X^{\perp})^{\perp}$.\
Suppose now $N$ is an injective cogenerator of ${\cal M}_{R}$ and take $x\in (X^{\perp})^{\perp}$. If $x\notin X$ then there is $f:M/X\longrightarrow N$ such that $f(\hat{x})\neq 0$ ($\hat{x}$ is the image of $x$ in $M/X$ via the canonic morphism $\pi: M\longrightarrow M/X$). Then there is $g=f\circ \pi,\, g\in {\rm Hom}_{R}(M,N)$ such that $g\vert_{X}=0$ ($g\in X^{\perp}$) and $g(x)\neq 0$, showing that $x\notin (X^{\perp})^{\perp}$, a contradiction.\
(ii) Let $f\in \overline{Y}$ and take $x\in Y^{\perp}$. Then there is $g\in Y$ such that $f(x)=g(x)$. But $g(x)=0$ because $x\in Y^{\perp}$ so $f(x)=0$. Thus $f\vert_{Y^{\perp}}=0$ and $f\in (Y^{\perp})^{\perp}$.\
For the converse, first we see that ${}_{R}R$ injective implies that for all finitely generated right $R$ modules $F$ we have that $F\stackrel{\Phi_{F}}{\longrightarrow}{{}^{*}(F^{*})}$ is an epimorphism. Take $\pi:P=R^{n}\longrightarrow F$ an epimorphism in ${\cal M}_{R}$. Then we have a monomorphism $0\longrightarrow P^{*}\longrightarrow F^{*}$ in ${}_{R}{\cal M}$, and as ${}_{R}R$ is injective we obtain an epimorphism of right modules ${}^{*}(P^{*})\stackrel{{}^{*}(p^{*})}{\longrightarrow}{}^{*}(F^{*})\longrightarrow 0$. Because $\Phi$ is a functorial morphism then we have the commutative diagram $$\xymatrix{
P \ar[r]^\pi\ar[d]_{\Phi_P} & F \ar[r]\ar[d]^{\Phi_F} & 0\\
{}^{*}(P^{*})\ar[r]_{{}^{*}(\pi^{*})} & {}^{*}(F^{*}) \ar[r] & 0
}$$ showing that $\Phi_{F}$ is surjective, as $\Phi_{P}=\Phi_{R^{n}}$ is an isomorphism. Now to prove the desired equality, take $f\in (Y^{\perp})^{\perp}$, $(f_{i})_{i\in I}$ a family of generators of the left $R$ module $Y$, and $F<M$ a finitely generated submodule of $M$. Then $f_{i}\vert M\in F^{*}$ and if $f\vert F\notin {}_{R}<f_{i}\vert_{F}\mid i\in I>$ then as ${}_{R}R$ is an injective cogenerator of ${}_{R}{\cal M}$ we can find a morphism of left $R$ modules $\phi: F^{*}\longrightarrow R$ such that $\phi(f_{i})=0,\,\forall i\in I$ and $\phi(f)\neq 0$. But as $\Phi_{F}$ is surjective, we can then find $x\in F$ such that $\phi=\Phi(x)$ and then $f_{i}(x)\Phi(x)(f_{i})=\phi(f_{i})=0,\,\forall i\in I$, showing that $x\in Y^{\perp}$ and $f(x)=\Phi(x)(f)=\phi(f)\neq 0$ which contradicts the fact that $f$ belongs to $(Y^{\perp})^{\perp}$. Thus we must have $f\vert_{F}\in {}_{R}<f_{i}\vert_{F}\mid i\in I>$ so there is $(r_{i})_{i\in I}$ a family of finite support such that $f\vert_{F}=\sum\limits_{i\in I}r_{i}(f_{i}\vert_{F})=(\sum\limits_{i\in I}r_{i}f_{i})\vert_{F}$. This last relation shows that $f\in \overline{Y}$.
If $R$ is a PF ring (left and right) then for any right (or left) $R$ module $M$ and $Y<M^{*}$ we have that $Y$ is dense in $M^{*}$ if and only if $Y^{\perp}=0$.
Let $M$ be a right $R$ module.\
(i) If $X\leq M$ then we have $((X^{\perp})^{\perp})^{\perp}=X^{\perp}$ and $X^{\perp}$ is closed.\
(ii) If $Y\leq {\rm Hom}_{R}(M,N)$ then $((Y^{\perp})^{\perp})^{\perp}=Y^{\perp}$.
“$\subseteq$” from (i) and (ii) follow from Proposition \[1\] and Lemma \[2\].\
(i) “$\supseteq$” Let $f\in X^{\perp}$. Take $x\in (X^{\perp})^{\perp}$; then $f(x)=0$ so $f\in ((X^{\perp})^{\perp})^{\perp}$. To show that $X^{\perp}$ is closed take $f\in \overline{X^{\perp}}$ and $x\in X$. Then there is $g\in X^{\perp}$ such that $g(x)=f(x)$ so $f(x)=0$ ($x\in X$). We obtain that $f\vert_{X}=0$ so $f\in X^{\perp}$.\
(ii) “$\supseteq$” Let $x\in Y^{\perp}$. If $f\in (Y^{\perp})^{\perp}$ then $f\vert_{Y^{\perp}}=0$ so $f(x)=0$ showing that $x\in ((Y^{\perp})^{\perp})^{\perp}$.
\[3\] Let $M,N$ be right $R$ modules and $(X_{i})_{i\in I}$ a family of submodules of $M$. Then\
(i) $(\sum\limits_{i\in I}X_{i})^{\perp}=\bigcap\limits_{i\in I}X_{i}^{\perp}$.\
(ii) $(\bigcap\limits_{i\in I}X_{i})^{\perp}\supseteq \sum\limits_{i\in I}X_{i}^{\perp}$. If $I$ is finite and $N$ is injective then equality holds.
\(i) $f\in (\sum\limits_{i\in I}X_{i})^{\perp}\Leftrightarrow f\vert_{\sum\limits_{i\in I}X_{i}}=0\Leftrightarrow f\vert_{X_{i}}=0,\,\forall i\in I\Leftrightarrow f\in X_{i}^{\perp},\,\forall i\in I\Leftrightarrow f\in \bigcap\limits_{i\in I}X_{i}^{\perp}$.\
(ii) “$\supseteq$” is obvious, for Proposition \[1\] shows that $X_{i}^{\perp}\subseteq {\bigcap\limits_{j\in I}X_{j}}^{\perp},\,\forall i\in I$. For the converse it is enough to prove the equality for two submodules $X,Y$ of $M$. Denote $\pi:M\longrightarrow M/X\cap Y$, $p:M\longrightarrow M/X$, $q:M\longrightarrow M/Y$ the canonical morphisms. If $f\in {\rm Hom}_{R}(M,N)$ such that $f\vert _{X\cap Y}=0$ then denote $\overline{f}:M/X\cap Y\longrightarrow N$ the factorisation of $f$ ($f=\overline{f}\circ \pi$) and $i:M/X\cap Y\longrightarrow M/X\oplus M/Y$ the injection $i(\pi(x))=(p(x),q(x)),\,\forall x\in M$. Then the diagram $$\xymatrix{
0 \ar[r] & \frac{M}{X\cap Y} \ar[r]^{i}\ar[d]_{\overline{f}} & \frac{M}{X}\oplus\frac{M}{Y} \ar[dl]^{h=\overline{u}\oplus\overline{v}}\\
& N &
}$$ is completed commutatively by $h$. Then $h=\overline{u}\oplus\overline{v}$, with $\overline{u}\in {\rm Hom}_{R}(M/X,N)$ and ${\rm Hom}_{R}(M/Y,N)$, such that $h(p(x),q(x))=\overline{u}(p(x))+\overline{v}(q(x))$. Taking $u=\overline{u}\circ p$ and $v=\overline{v}\circ q$ we have $u\in X^{\perp},\,v\in Y^{\perp}$ and $f(x)=\overline{f}(\pi(x))=h(i(\pi(x)))=h(p(x),q(x))=\overline{u}(p(x))+\overline{v}(q(x))=u(x)+v(x),\,\forall x\in M$, so $f\in X^{\perp}+Y^{\perp}$.
\[4\] Let $M,N$ be right $R$ modules and $(Y_{i})_{i\in I}$ a family of submodules of ${\rm Hom}_{R}(M,N)$. Then:\
(i) $(\sum\limits_{i\in I}Y_{i})^{\perp}=\bigcap\limits_{i\in I}Y_{i}^{\perp}$.\
(ii) $(\bigcap\limits_{i\in I}Y_{i})^{\perp}\supseteq\sum\limits_{i\in I}Y_{i}^{\perp}$. If $N=R$ and $R$ is a PF ring (both left and right PF) and $Y_{i}$ are closed subsets of $M^{*}={\rm Hom}_{R}(M,R)$ then the equality holds: $(\bigcap\limits_{i\in I}Y_{i})^{\perp}=\sum\limits_{i\in I}Y_{i}^{\perp}$.
\(i) Obvious.\
(ii) “$\supseteq$” similar to (ii)“$\supseteq$” of the previous proposition. For the converse inclusion, take $(Y_{i})_{i\in I}$ a family of submodules of $M^{*}$. Then $$\begin{aligned}
\sum\limits_{i\in I}Y_{i}^{\perp} & = & ((\sum\limits_{i\in I}Y_{i}^{\perp})^{\perp})^{\perp} \;\;\;({\rm from\,Lemma\,}\ref{2}:R {\rm \,is\,right\,PF} )\\
& = & (\bigcap\limits_{i\in I}(Y_{i}^{\perp})^{\perp})^{\perp} \;\;\;({\rm from\,Proposition\,}\ref{3})\\
& = & (\bigcap\limits_{i\in I}Y_{i})^{\perp}\;\;\;({\rm Lemma\,}\ref{2}\,:Y_{i}{\rm \,are\,closed\,and\,}R{\rm\,is\,left\,PF})\end{aligned}$$
\(i) We show that the equality in Proposition \[3\] does not hold for infinite sets. Let $V$ be an infinite dimensional space with a countable basis indexed by the set of natural numbers: $(e_{n})_{n\in\mathbf{N}}$. Put $V_{n}=<e_{k}\mid k\geq n>$. Then we can easily see that $\bigcap\limits_{n\in\mathbf{N}} V_{n}=0$ so $(\bigcap\limits_{n\in\mathbf{N}} V_{n}=0)^{\perp}=V^{*}$. Let $f\in V^{*}$ be the function equal to $1$ on all the $e_{n}$-s. Then as $V_{n}^{\perp}<V_{m}^{\perp},\,\forall\, n<m$, we have that $f\in \sum\limits_{n\in\mathbf{N}}V_{n}^{\perp}\Leftrightarrow\exists\,n\in{\mathbf N}\, such\, that \,f\in V_{n}^{\perp}$ which is impossible as $f(e_{n})=0,\,\forall\, n$. We obtain $\bigcap\limits_{n\in\mathbf{N}} V_{n}\supset\sum\limits_{n\in\mathbf{N}}V_{n}^{\perp}$ a strict inclusion.
\(ii) We show now that the equality in Proposition \[4\] does not hold for non-closed sets. Let again $V$ be a vector space with a countable basis $B=(e_{n})_{n\in\mathbf{N}}$. Denote by $e_{n}^{*}$ the linear map equal to $1$ on $e_{n}$ and $0$ on the other elements of the basis $B$ and by $f^{*}$ the linear map equal to $1$ on all the $e_{n}$-s. Take $H=<e_{n}^{*}\mid n\in\mathbf{N}>$ and $L=<f^{*},e_{n}^{*}\mid n\in\mathbf{N}^{*}>$. Then we can easily see that $H^{\perp}=0$, $L^{\perp}=0$ and $H\cap L=<e_{n}^{*}\mid n\in \mathbf{N}^{*}>$, so $H^{\perp}+L^{\perp}=0$, but $(H\cap L)^{\perp}=<f^{*},e_{n}^{*}\mid n\in {\mathbf N}>^{\perp}=<e_{0}>$, thus $H^{\perp}+L^{\perp}\neq(H\cap L)^{\perp}$.
\(iii) Given the same vector space, we give an example of a family of dense subspaces of $V^{*}$ whose intersection is 0. For $p\in {\mathbf N}$ let $H_{p}=<e_{n}^{*}+e_{n+1}^{*}+\dots+e_{n+p}^{*}\mid n\in {\mathbf N}>$. Then a short computation shows that $H_{n}^{\perp}=0$ showing that $H_{n}$ is closed in $V^{*}$. But $\bigcap\limits_{n\in {\mathbf N}}H_{n}=0$, because if $f=\sum\limits_{i=1}^{m}\lambda_{i}e_{i}^{*}\in\bigcap\limits_{n\in {\mathbf N}}H_{n}\subset H_{0}$, then $f\in H_{m+1}$ which shows that if $f\neq 0$, than it can be written as a linear combination of $e_{i}^{*}$ in which at least one of the $e_{i}^{*}$ has $i>m$. This is impossible as the $e_{n}^{*}$-s are independent.
[The Finite Topology vs PF Rings]{} If $R$ is a ring then we have $(R^{n})^{*}={\rm Hom}_{R}(R,R)\simeq {}_{R}R^{n}$. So we can identify $R$ submodules of the right dual of $R^{n}$ with left submodules of ${}_{R}R$ and vice versa. For all $x=(x_{1},\dots,x_{n})\in R^{n}$ we denote by $\varphi_{x}:R^{n}\longrightarrow R$ the morphism of right $R$ modules $\varphi_{x}(r_{1},\dots,r_{n})=\sum\limits_{i=1}^{n}x_{i}r_{i}$ and by $\psi_{x}$ the morphism of left modules defined by $\psi_{x}(r_{1},\dots,r_{n})=\sum\limits_{i=1}^{n}r_{i}x_{i},\,\forall \, (r_{1},\dots,r_{n})\in R^{n}$. Also because of the isomorphism $(R^{n})^{*}\simeq {}_{R}R^{n}, \, x\mapsto \varphi_{x}$, we will denote by $I^{\perp}=\{x\in R^{n}\mid \varphi_{x}(r)=0,\,\forall \, r\in I\}$ if $I$ is a right submodule of $R^{n}$ and similarly for left submodules $X$ of $R^{n}$, $X^{\perp}=\{x\in R^{n}\mid \psi_{x}(r)=0,\,\forall \, r\in X\}$.\
Over a vector space $V$ there is an anti isomorphism of lattices between the lattice of closed subspaces of $V^{*}$ and the subspaces of $V$ given by $X\mapsto X^{\perp},\,\forall\,X\leq V$. We have the obvious
\[f.1\] For a right module $M$ the following are equivalent:\
(i) The applications $M\geq X\mapsto X^{\perp}\leq M^{*}$ and $M^{*}\geq Y\mapsto Y^{\perp}\leq M$ between the lattice of the submodules of $M$ and the lattice of the closed submodules of $M^{*}$ are inverse anti isomorphism of lattices.\
(ii) $(X^{\perp})^{\perp}=X,\,\forall\,X\leq M$ and $(Y^{\perp})^{\perp}=\overline{Y},\,\forall\,Y\leq M^{*}$.\
(iii) $(X^{\perp})^{\perp}=X,\,\forall\,X\leq M$ and $(Y^{\perp})^{\perp}=Y,\,\forall\,Y\leq M^{*}$, $Y$ closed.\
(iv) The applications of (i) are inverse to each other.
If $F$ is a finitely generated right $R$ module then every submodule of $F^{*}$ is closed, as if $Y$ is a left submodule of $F^{*}$ and $f\in \overline{Y}$, taking $\{x_{1},\dots,x_{n}\}$ the a system of generators of $F$, there is $g\in Y$ such that $g(x_{i})=f(x_{i})$, for all $i$, so $f=g\in Y$. Also it is easy to see that $R^{n}$ has orthogonal equivalence as right module if and only if it has orthogonal equivalence as left module, and this is equivalent to $(I^{\perp})^{\perp}=I,\,\forall\,I\leq R^{n}_{R}$ and $(X^{\perp})^{\perp}=X,\,\forall\,X\leq {}_{R}R^{n}$.
We will say that a right $R$ module $M$ has orthogonal equivalence (or orthogonal isomorphism, or shortly $M$ has $\perp$ equivalence) if the equivalent statements of Proposition \[f.1\] hold. The ring $R$ will be called with $\perp$ equivalence if $R_{R}$ (or equivalently ${}_{R}R$) is a module with orthogonal equivalence.
\[f.3\] Let $M$ be a right $R$ module and $X$ a submodule of $M$. Then we have the exact sequence $$0\longrightarrow (0^{\perp})^{\perp}\longrightarrow M\stackrel{\Phi_{M}}{\longrightarrow}{}^{*}(M^{*})$$
For $x\in M$ we have $\Phi_{M}(x)=0\Leftrightarrow f(x)=0,\,\forall\,f\in M^{*}$ and this equivalent to $x\in (M^{*})^{\perp}=(0^{\perp})^{\perp}$, thus ${\rm ker }\,\Phi_{M}=(0^{\perp})^{\perp}$.
\[f.4\] (i) For an $R$ module $M$ we have $(0^{\perp})^{\perp}=0$ if and only if $M$ is $R$ cogenerated, i.e. there is a monomorphism $M\hookrightarrow R^{I}$ for some set $I$.\
(ii) If $\cal C$ is a class of right $R$ modules which is closed under quotients then the following are equivalent:\
(a) $(X^{\perp})^{\perp}=X$ for all $M$ in ${\cal C}$, $X<M$.\
(b) $(0^{\perp})^{\perp}=0$ for all $M$ in $\cal C$.\
(c) Any $M\in {\cal C}$ is cogenerated by $R$.\
(d) $\Phi_{M}$ is a monomorphism for every $M$ in $\cal C$.
\(i) If $(0^{\perp})^{\perp}=0$ then take $I=M^{*}$ and $M\stackrel{i}{\longrightarrow}R^{I},\,i(x)=(f(x))_{f\in I}$; then of course $i$ is a monomorphism as $i(x)=0$ if and only if $f(x)=0,\,\forall f\in I=M^{*}$ i.e. $x\in (0^{\perp})^{\perp}=0$. Conversely, given a monomorphism $M\stackrel{i}{\hookrightarrow}R^{I}$, taking $\pi_{j}$ the canonical projections for all $j\in I$, we obtain the morphisms $f_{j}=\pi_{j}\circ i\in M^{*}$ and then $x\in (0^{\perp})^{\perp}=(M^{*})^{\perp}$ implies $f_{j}(x)=0,\,\forall\,j\in I$, i.e. $i(x)=0$ so $x=0$, as $i$ is injective. Thus $(0^{\perp})^{\perp}=0$.\
(ii) (b) $\Leftrightarrow$ (c) by (i). (a) $\Leftrightarrow$ (b) follows as $\cal C$ is closed under quotient objects and denoting $\overline{0}$ the zero element of $M/X\in {\cal C}$ we have $(\{\overline{0}\}^{\perp})^{\perp}=(X^{\perp})^{\perp}$ from Lemma \[2\]. Equivalence with (d) follows from Proposition \[f.3\]
\[f.5\] Suppose $R_{R}$ is a module with $\perp$ equivalence. Then $R$ contains all left simple modules and all right simple modules (up to an isomorphism; this is called a right - and left- Kasch ring).
It is easy to see that for every right ideal $I$ of $R$ we have the isomorphism of left $R$ modules $(\frac{R}{I})^{*}\simeq I^{\perp}$, given by $I^{\perp}\ni f\mapsto f\circ \pi\in (\frac{R}{I})^{*}$, with $\pi:R\longrightarrow R/I$ the canonical projection. Then if $S$ is simple right module there is a maximal right ideal $M<R$ and an isomorphism $S\simeq \frac{R}{M}$. Then $S^{*}\simeq (\frac{R}{M})^{*}\simeq M^{\perp}\neq 0$ because if $M^{\perp}=0$ then $M=(M^{\perp})^{\perp}=0^{\perp}=R$, which contradicts the maximality of $M$. In a similar way one can see that $R$ contains all the isomorphism types of left $R$ modules.
We shall say a right (or left) $R$ module is $n$ generated if it has a system of $n$ generators.
\[f.6\] Let $X$ be a right $R$ module such that every monomorphism $i:X\hookrightarrow M$ with the property that $M/{\rm Im}\,i$ is 1-generated splits. Then $X$ is an injective module.
Let $M$ be a right $R$ module such that $X<M$ (we identify $X$ with its image in $M$) and suppose $X\neq M$. Let ${\cal L}=\{Y<M\mid Y\neq 0 \,{\rm and}\,X\cap Y=0\}$. Then ${\cal L}\neq \emptyset$, because if $x\in M\setminus X$ then as $(X+xR)/X\neq 0$ is finitely generated then the hypothesis shows that there is $Y<X+xR$ such that $X\stackrel{.}{+}Y=X+xR$ and then $Y\neq 0$ as $x\notin X$, so $Y\in {\cal L}$. We can easily see that ${\cal L}$ is inductive, because if $(Y_{i})_{i\in I}$ is a totally ordered family of elements of ${\cal L}$ then $\bigcup\limits_{i\in I}Y_{i}$ is its majorant in ${\cal L}$. Take $N$ a maximal element of ${\cal L}$ and suppose $X+N\neq M$. Then there is $x\in M\setminus (X+N)$ and as $(X+N+xR)/(X+N)$ is finitely generated, by the hypothesis we can find $Y<M$ such that $X+N+Y=X+N+xR$ and $(X+N)\cap Y=0$. An easy computation shows now that $(N+Y)\cap X=0$ and so $N+Y=N$ by the maximality of $N$. Thus we obtain $X+N+Y=X+N=X+N+xR$ which is a contradiction, because $x\notin X+N$. We find that $X$ is a direct summand in $M$ for every module $M$ such that $X\hookrightarrow M$, so $X$ is injective in ${\cal M}_{R}$.
\[f.7\] Let $R$ be a ring with $\perp$ equivalence. If $R\stackrel{j}{\hookrightarrow} X$ is a monomorphism of right (left) $R$ modules and $X$ is $R$ cogenerated then $j$ splits.
Consider $X\stackrel{\sigma}{\hookrightarrow}R^{I}$ a monomorphism and let $(x_{i})_{i\in I}=\sigma(j(1))$. Then we have $(x_{i}r)_{i\in I}=\sigma(j(1))r=\sigma(j(r))$ and as $j,\sigma$ are injective we see that $x_{i}r=0,\,\forall\,i\in I$ if and only if $r=0$. This shows that $\bigcap\limits_{i\in I}Rx_{i}^{\perp}=0$. Then we have $0=\bigcap\limits_{i\in I}Rx_{i}^{\perp}=(\sum\limits_{i\in I}Rx_{i})^{\perp}$ (by Proposition \[3\]), so $\sum\limits_{i\in I}Rx_{i}=((\sum\limits_{i\in I}Rx_{i})^{\perp})^{\perp}=0^{\perp}=R$. Then we find that there is $F$ a finite subset of $I$ such that $\sum\limits_{i\in F}Rx_{i}=R$, thus there are $(y_{i})_{i\in F}\in R$ such that $\sum\limits_{i\in F}y_{i}x_{i}=1$. Now if we denote by $\pi_{F}$ the projection of $R^{I}$ on $R^{F}$, $\pi_{F}((r_{i})_{i\in I})=(r_{i})_{i\in F}$ and by $y=(y_{i})_{i\in F}\in R^{F}=R^{(F)}$, then $\varphi_{y}(\pi_{F}(\sigma(j(r))))=\varphi_{y}(\pi_{F}((x_{i}r)_{i\in I}))=\varphi_{y}((x_{i}r)_{i\in F})=\sum\limits_{i\in F}y_{i}x_{i}r=r$, so $\varphi_{y}\circ\pi_{F}\circ\sigma\circ j=id_{R}$, showing that the morphism of right modules $\varphi_{y}\circ\pi_{F}\circ\sigma:X\longrightarrow R$ is a split for $j$.
\[f.8\] $R^{n}$ has orthogonal equivalence (as left or right $R$ module) if and only if every $n$ generated right (or left) module has orthogonal equivalence.
Suppose $R^{n}$ has $\perp$ equivalence. Let $F=R^{n}/X$ be a right $n$ generated $R$ modules and $\pi:R^{n}\longrightarrow F$ the canonical projection. For each $g\in X^{\perp}$ ($X<R^{n}$) we denote by $\overline{g}\in F^{*}$ the (unique) morphism for which $\overline{g}\circ\pi=g$ and with $\hat{x}=\pi(x)$ - the class of an element $x\in R^{n}$. Now we see that if $Y<F^{*}$ and $Z=\{\alpha\circ\pi\mid\alpha\in Y\}$, then $Y=\{\overline{g}\mid g\in Z\}$, $Y^{\perp}=\{\hat{x}\mid \overline{g}(\hat{x})=0,\,\forall\,g\in Z\}=Z^{\perp}/X$ ($Z\subseteq X^{\perp}$ so $Z^{\perp}\supseteq (X^{\perp})^{\perp}=X$) and $(Y^{\perp})^{\perp}=\{\overline{g}\mid\overline{g}(\hat{x})=0,\,\forall\,\hat{x}\in Z^{\perp}/X\}=\{\overline{g}\mid g(x)=0,\,\forall\,x\in Z^{\perp}\}=\{\overline{g}\mid g\in (Z^{\perp})^{\perp}=Z\}=Y$.\
Now if $Y<F$ and $Z=\pi^{-1}(Y)$ then $Y^{\perp}=\{\overline{g}\mid\overline{g}(\hat{x})=0,\,\forall\,\hat{x}\in Y\}=\{\overline{g}\mid g(x)=0,\,\forall\,x\in Z\}=\{\overline{g}\mid g\in Z^{\perp}\}$ and $(Y^{\perp})^{\perp}=\{\hat{x}\mid\overline{g}(\hat{x})=g(x)=0,\,\forall\,g\in Z^{\perp}\}=\{\hat{x}\mid x\in (Z^{\perp})^{\perp}=Z\}=Y$.
The following assertions are equivalent:\
(i) Every right $R$ module has $\perp$ equivalence.\
(ii) Every finitely generated module has $\perp$ equivalence.\
(iii) Every left $R$ module has $\perp$ equivalence.\
(iv) Every finitely generated module has $\perp$ equivalence.\
(v) $R$ is a PF ring (both left and right).\
(vi) $(X^{\perp})^{\perp}=X$ for all $X<M$ in ${\cal M}_{R}$ or in ${}_{R}{\cal M}$.\
(vii) $R^{2}$ has $\perp$ equivalence.
$\bullet$ (v) $\Rightarrow$ (i) and (v) $\Rightarrow$ (vi) follow from Lemma \[2\] so we have the implications (v) $\Rightarrow$ (i) $\Rightarrow$ (ii) $\Rightarrow$ (vii) and (v) $\Rightarrow$ (vi) $\Rightarrow$ (vii)\
$\bullet$ (v) $\Rightarrow$ (iii) $\Rightarrow$ (iv) $\Rightarrow$ (vii) is the left symmetric of (v) $\Rightarrow$ (i) $\Rightarrow$ (ii) $\Rightarrow$ (vii).\
$\bullet$ (vii) $\Rightarrow$ (v) If $R^{2}$ has $\perp$ equivalence, then by Lemma \[f.8\] we have that any 2 generated right (and any left) module has $\perp$ equivalence, in particular $R$ has orthogonal equivalence. Now let $R\stackrel{i}{\hookrightarrow}X$ be a monomorphism in ${\cal M}_{R}$ such that $X/i(R)$ is 1 generated. Then as $X$ has $\perp$ equivalence, Proposition \[f.4\] shows that $X$ is $R$ cogenerated as right $R$ module. Now by Proposition \[f.7\] $i$ splits, as $X$ is $R$ cogenerated and $R$ has $\perp$ equivalence. Then we can apply Lemma \[f.6\] and obtain that $R_{R}$ is injective. Because $R$ has $\perp$ equivalence, by Proposition \[f.5\] we obtain that $R_{R}$ contains all isomorphism types of simple right modules, and as $R_{R}$ is injective, we obtain that $R_{R}$ is an injective cogenerator of ${\cal M}_{R}$, i.e. a right RF ring. Similarly we can show that $R$ is also a left PF ring.
If $R$ is a PF ring, then $F\simeq {}^{*}(F^{*})$ by $\Phi_{F}$ for every finitely generated left module (the analogue holds for right modules).
Proposition \[f.3\] shows that $\Phi_{F}$ is injective. By the same argument as in the proof of Lemma \[2\] we have that ${}_{R}R$ injective implies that $\Phi_{F}$ is an epimorphism and the conclusion is proved.
$R$ is a PF ring if and only if for every finitely generated right (or left) $R$ module $F$, the lattice of the submodules of $F$ is anti isomorphic to the lattice of the submodules of $F^{*}$ via the $\perp$ applications of Proposition \[f.1\], equivalently, the dual lattice of the submodules of any finitely generated right module is isomorphic (via $\perp$ applications) to the lattice of the submodules of the dual of that module.
[99]{}
R. B. Ash, [*Abstract Algebra* ]{}, University of Illinois, 2000.
X. Chen and Q. Xu, Toeplitz operators on discrete abelian groups, [*Mathematical Proceeding of the Royal Irish Academy*]{}, [**100A**]{}(2), (2000), 139-148.
E. H. Connel, [*Elements of Abstract and Linear Algebra*]{}, University of Miami, 2002.
I. D. Ion, [*Algebra*]{}, Bucharest, 1974.
I. M. Isaac, [*Algebra*]{}, Belmont, California, 1994.
L. Lady, [*Finite Rank Torsion Free Modules over Dedekind Domains* ]{}, University Hawaii, 1998.
J. D. Lewis, [*Commutative ring theory*]{}, Alberta University, 2000.
J. Rada and M. Saorin, On semiregular rings whose finitely generated modules embed in free modules, [*Canad. Math. Bull.* ]{}, Vol. [**40**]{}(2), (1997).
T. M. Viswanathan, Ordered modules of fractions, [*J. Reine Sngew. Math.*]{}, (1969).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Supporting convolutional neural network (CNN) inference on resource-constrained IoT (Internet of Things) devices in a timely manner has been an outstanding challenge for emerging smart systems. To mitigate the burden on IoT devices, one prevalent solution is to offload the CNN inference task, which is usually composed of billions of operations, to public cloud. However, the “offloading-to-cloud” solution may cause privacy breach while moving sensitive data to cloud. For privacy protection, the research community has resorted to advanced cryptographic primitives (e.g., homomorphic encryption) and approximation techniques to support CNN inference on encrypted data. Consequently, these attempts cause impractical computational overhead on IoT devices and degrade the performance of CNNs. Another concern of “offloading-to-cloud” besides the privacy issue is the integrity of data. In order to avoid the heavy computation resulted by the offloaded tasks, public cloud can dismiss the inference request by sending back random results to IoT devices. Moreover, relying on the remote cloud can cause additional network latency and even make the system dysfunction when network connection is off.
To address the challenge, we proposes an extremely ightweight dge device assisted rivate CNN inference solution for IoT devices, namely LEP-CNN. The main design of LEP-CNN is based on a novel online/offline encryption scheme. The decryption of LEP-CNN is pre-computed offline via utilizing the linear property of the most time-consuming operations of CNNs. As a result, LEP-CNN allows IoT devices to securely offload over 99% CNN operations, and edge devices to execute CNN inference on encrypted data as efficient as on plaintext. To prevent edge devices from returning an incorrect result, LEP-CNN provides an integrity check option to enable the IoT devices to detect incorrect results with a successful rate over 99%. Experiments on AlexNet show that our scheme can speed up the CNN inference for more than $35\times$ for IoT devices with comparable computing power of Raspberry Pi. We also implemented a homomorphic encryption based privacy preserving AlexNet using the well-known CryptoNets scheme and compared it with LEP-CNN. We demonstrated that LEP-CNN has a better performance than homomorphic encryption based privacy preserving neural networks under time-sensitive scenarios.
author:
- '[email protected], [email protected] *Embry-Riddle Aeronautical University, USA*'
- '[email protected] *Stevens Institute of Technology, USA*'
- '[email protected] *Boise State University, USA*'
bibliography:
- 'CNS-19-LEPCNN-Journal.bib'
title: 'LEP-CNN: A Lightweight Edge Device Assisted Privacy-preserving CNN Inference Solution for IoT'
---
Internet of Things, Deep Neural Networks, Convolutional Neural Network, Privacy
Introduction
============
With the recent advances in artificial intelligence, the integration of deep neural networks and IoT is receiving increasing attention from both academia and industry [@dl-iot-1; @dl-iot-2; @dl-iot-5]. As the representative of deep neural networks, convolutional neural network (CNN) has been identified as an emerging technique to enable a spectrum of intelligent IoT applications [@dl-iot-survey], including visual detection, smart security, audio analytics, health monitoring, infrastructure inspection, etc. However, due to the high computational cost introduced by CNNs, their deployment on resource-constrained IoT devices for time-sensitive services becomes very challenging. For example, popular CNN architectures (e.g., AlexNet [@AlexNet], FaceNet [@FaceNet], and ResNet [@ResNet]) for visual detection require billions of operations for the execution of a single inference task. Our evaluation results show that an inference task using AlexNet can cost more than two minutes on an IoT device with comparable computing capability as a Raspberry Pi (Model A). To mitigate such a burden for IoT devices, offloading CNN tasks to public cloud has become a popular choice in the literature. However, this type of “cloud-backed" system may raise privacy concerns [@iot-cloud-privacy] by sending sensitive data to remote cloud. Also, the integrity of the returned results cannot be guaranteed when the cloud is “curious-and-dishonest”. Moreover, connecting to cloud can cause additional latency to the system under network congestion and even make the system dysfunction when network is off [@edge-computing].
In time-sensitive CNN-driven IoT applications, the inference stage of a trained CNN is executed for deep data analytics. To address the privacy concern with offloading CNNs, researchers have attempted to execute the inference stage over encrypted data [@CryptoNets; @IACR-PPDL; @CryptoDL]. These schemes usually first use approximation strategies to convert non-linear layers in a CNN to linear operations. Then, homomorphic encryption is utilized to enable privacy-preserving execution of these converted operations and other layers in the CNN using cloud computing. Nevertheless, the adoption of homomorphic encryption introduces extremely high encryption cost and/or communication load to the local IoT devices. For example, a quad-core Raspberry Pi, which outperforms most IoT devices in terms of computational capability, can perform only four Paillier homomorphic encryption per second [@paillier-benchmark]. Given a single input of AlexNet that has $227\times227\times 3$ elements, it requires more than 10 hours to complete the encryption, which is impractical for most applications in terms of both time delay and energy consumption. (As a comparison, to execute an inference task with the same AlexNet, a single-core Raspberry Pi with our solution can finish all local encryption and decryption within 0.3 second as shown in Section \[s:evaluation\].) Besides the high computational cost, the utilization of approximation in existing research also results in accuracy loss to some extent. Furthermore, these research adopt batch processing to improve their performance, which is more suitable for the “Data Collection and Post-Processing” routine. Differently, on-the-fly processing is desired for IoT devices to fulfill time-sensitive tasks. Another line of related research utilizes differential privacy to achieve privacy-preserving offloading of the training stage of CNNs [@CCS16; @ICDM17-PHAN]. These research control the amount of information leaked from each individual record in the training dataset. However, differential privacy becomes unsuitable for the inference stage, because only a single input is available at this stage.
The other limitation with offloading time-sensitive IoT tasks to cloud is the reliance on the availability of cloud and its network condition. To overcome this limitation, one prevalent solution is to utilize the computing resources at the edge of network. Compared with cloud computing, edge devices are geographically closer to IoT devices and usually within one-hop communication range. Such physical proximity can effectively ameliorate the network latency and availability issue. Most current research [@edge-iot-2; @edge-iot-3] on edge computing mainly emphasizes on fundamental issues such as resource allocation but assumes the edge devices are fully trusted. Such an assumption, while is probably true for many applications, can be problematic for other applications where IoT devices are mobile or edge devices are shared by ad hoc multiple parties that do not share mutual trust. Examples of such applications include vehicular networks, drones, and mobile edge computing (MEC) [@5G-MEC] in general, to name a few. Some research [@iot-survey] does address data security and privacy issues in edge computing. However, efficient and private offloading of CNN inference to edge devices remains an open challenge.
Another issue is that the IoT devices are not able to judge the correctness of the returned results from the cloud. Since the offloaded computation is resource consuming, public clouds may not be willing to allocate expensive computational resources and may tend to cheat the IoT devices by returning random data with the same size of the desired data. According to [@kandukuri2009cloud], in some cases, dishonest cloud service provides may even discard the data to save resources. How to effectively detect such dishonest behaviors while maintaining the lightweight computation on IoT devices and overall performance in time-sensitive CNN inference tasks is another essential challenge to be solved.
This paper addresses such challenges and proposes an extremely lightweight edge device assisted private CNN inference solution for IoT devices, namely LEP-CNN. LEP-CNN enables IoT devices to securely offload CNN inference tasks to local edge devices. To significantly speed up the offloading process, LEP-CNN adopts a novel online/offline encryption. Specifically, since CNN linear operation over input data and random noise are separable, encryption and decryption can be efficiently computed offline. In practical CNN architectures such as AlexNet and FaceNet, linear operations are dominant due to their vast number. Therefore, it is rewarding to trade offline computation and storage (of random noise) for online computation. As a result, our online/offline encryption allows IoT devices to securely offload over 99% CNN operations to edge devices. And edge devices are able to execute CNN inference on encrypted data as efficiently as on unencrypted data. In addition, LEP-CNN does not introduce any accuracy loss as compared to CNN inference over unencrypted data. Compared with homomorphic encryption based privacy preserving neural networks [@CryptoNets; @CryptoDL; @IACR-PPDL], LEP-CNN achieves a better performance under time-sensitive scenarios due to its lightweight property. Furthermore, to detect dishonest behaviors from “curious-and-dishonest” edge devices, our scheme provides an integrity check mechanism which helps the IoT devices detect incorrect returned results from edge devices with an over 99% success rate. Minor computation overhead (1.1% drop of offload percentage in worst case) is introduced when this integrity check is on. This integrity check can be optionally turned off in such scenario that the cloud is considered “curious-and-honest” to save local computational resources. LEP-CNN can be customized to support flexible CNN architectures that fulfill the requirements of different applications.
Extensive experimental evaluation shows the efficiency, scalability, accuracy and integrity of LEP-CNN. We implemented a prototype over well-known ImageNet [@imagenet_cvpr09] dataset using an uncompressed AlexNet architecture, which involves 2.27 billion operations for each inference result and has comparable complexity with these widely adopted architectures, e.g., FaceNet (1.6 billion operations) and Results (3.6 billion operations). The experimental results show that LEP-CNN can securely offload $99.95\%$ computation from the IoT devices for AlexNet. As a result, we are able to speed up $35.63\times$ for the execution of CNN-driven IoT AlexNet tasks using a single laptop as the edge device. Meanwhile, LEP-CNN saves over $95.56\%$ energy consumption compared with fully executing an AlexNet request on the IoT device. We also deployed LEP-CNN in a “curious-and-dishonest” scenario. With the integrity check feature being turned on, our scheme can still maintain a high computation offloading rate of $99.33\%$ from the IoT devices and a $30.00\times$ speedup. In addition, our scheme keeps the high speedup rate as the complexity of CNN layers increases. Therefore, it is promising to be scaled up for more complex CNN architectures.
The rest parts of this paper are organized as follows: In Section \[s:cnn\], we introduce the background of CNN. Section \[s:construction\] presents the detailed construction of our scheme. We state security analysis and numerical analysis in Section \[s:analysis\]. We further evaluate the practical performance of our scheme with a prototype evaluation in Section \[s:evaluation\]. We review and discuss related works in Section \[s:related-work\] and conclude this paper in Section \[s:conclusion\].
Background of Convolutional Neural Network {#s:cnn}
==========================================
A CNN contains a stack of layers that transform input data to outputs with label scores. There are four types of most common layers in CNN architectures, including: *Convolutional Layers, Pooling Layers, Activation Layers*, and *Fully-connected Layers*.
Convolutional layers extract features from input data. Fig.\[f:con-layer\] depicts an example of convolutional layer that has an input data of size $n\times n\times D$ and $H$ kernels, each of size $k\times k \times D$. The input will be processed into all $H$ kernels independently to extract $H$ different features. Considering the input and each kernel as $D$ levels, where each level of the input and kernel are a $n\times n$ matrix and a $k\times k$ matrix respectively. Each level of a kernel starts scanning the corresponding input level from top-left corner, and then moves from left to right with $s$ elements, where $s$ is the stride of the convolutional layer. Once the top-right corner is reached, the kernel moves $s$ elements downward and scans from left to right again. This convolution process is repeated until the kernel reaches the bottom-right corner of this input level. For each scan, an output is computed using the dot product between the scanned window of input and the kernel as an example shown in Fig.\[f:con-layer\]. For each kernel, the output for all $D$ levels will be summed together.
![Examples of a Convolutional Layer and a Fully-connected Layer[]{data-label="f:con-layer"}](conv-layer){width="8cm"}
Pooling layers and activation layers are usually non-linear layers. A pooling layer is periodically inserted between convolutional layers. Pooling layers progressively reduce the spatial size of outputs from convolutional layers, and thus to make data robust against noise and control overfitting. An activation layer utilizes element-wise activation functions to signal distinct identification of their input data. There are a number of popular pooling strategies (e.g., max-pooling and average-pooling) and activation functions (e.g., rectified linear units (ReLUs) and continuous trigger functions), which are extremely computational efficient compared with convolutional layers and fully-connected layers. In our scheme, these two efficient layers will be directly handled on the IoT devices.
Fully-connected layers are usually the final layers of a CNN to output the final results of the network. In case of a fully-connected layer, all neurons in it have full connections to all outputs from the previous layer. As an example shown in Fig.\[f:con-layer\], the connection between each neuron and input element has a weight. To obtain the output of a neuron, elements connected to it will be multiplied with their weights and then accumulated.
More details about CNN can be found in ref [@cnn-wiki].
Detailed Construction of LEP-CNN {#s:construction}
================================
***System Model***: LEP-CNN consists of three major entities: *IoT Devices*, the *Owner of Devices*, and *Edge Devices*. To deploy IoT devices for CNN-driven tasks, the owner first performs the offline phase to generate encryption and decryption keys and stores them into the IoT devices. As each set of encryption and decryption keys will only be used for one CNN request for security purpose, the owner needs to pre-load multiple sets of keys into the IoT devices. In Section \[ss:key-update\], we discuss the storage and remote update of keys to ensure the performance of different IoT application scenarios. For trained CNN architectures, the owner stores them using cloud storage or other accessible platforms, which can be retrieved by the edge devices when the IoT device enters the their coverage or pre-requested by the owner. In the online phase, once the IoT device has a piece of data needs to be processed by the CNN, it interacts with the edge device to perform privacy-preserving execution. During this process, the IoT device efficiently offloads expensive convolutional layers and fully-connected layers to the edge device, and only keeps the compute-efficient layers at local. This is motivated by the fact that convolutional and fully-connected layers occupy majority of computation and parameters storage in typical CNNs [@Cong2014].
***Threat Model***: In our system, we consider the edge devices to be “curious-but-honest”, i.e., the edge devices follow our scheme to correctly conduct the storage, communication, and computational tasks, but try to learn sensitive information in IoT devices’ input for the CNN. With our integrity check feature turned on, our system can also detect “curious-and-dishonest” edge devices at a high successful rate with minor efficiency trade-off. We assume the IoT devices are fully trusted and will not be compromised. The research on protecting IoT devices is orthogonal to this work. The edge device has access to the trained CNNs and all ciphertexts of inputs and outputs for the offloaded convolutional layers and fully-connected layers in the CNN. This assumption is consistent with majority of existing works that focus on privacy-preserving computation offloading [@CryptoNets; @IACR-PPDL; @CryptoDL]. LEP-CNN focuses on preventing the edge devices from learning the IoT devices’ inputs and outputs of each layer in the offloaded CNN.
There is also a line of research that proposes to infer the inputs of CNN layers from their outputs [@output-inference1]. This kind of inference is not applicable to LEP-CNN, since the edge device only has access to encrypted outputs from each layer.
We now present the detailed construction of LEP-CNN. Important notation is summarized in Table \[t:notations\].
--------------------- -----------------------------------------------------------------------
the size of each level of a convolutional layer’s
input
the depth of the input of a convolutional layer
and kernel
$k\times k$ the size of a convolutional layer’s kernel matrix
$H$ the number of kernels of a convolutional layer
$s$ the size of stride used for a convolutional layer
$p$ the size of padding used for a convolutional layer
$\mathcal{R}_{c,d}$ $n\times n$ random matrices to encrypt the input of a
$1\leq d\leq D$ convolutional layer
$\alpha_i$ $(\frac{n-k+2p}{s}+1)\times (\frac{n-k+2p}{s}+1)$ matrices to decrypt
$1\leq i\leq H$ a convolutional layer’s output from $H$ kernels
$m$ the size of a fully-connected layer input vector
$T$ the number of neurons of a fully-connected layer
a $m$-dimensional random vector to encrypt the
input of a fully-connected layer
a $T$-dimensional decryption vector to decrypt
fully-connected layer outputs
$r$ the sample rate of returned data of a convolutional layer
$\theta$ the error rate of returned data of a convolutional layer
--------------------- -----------------------------------------------------------------------
: Summary of Notations[]{data-label="t:notations"}
Offline Phase
-------------
In the offline phase, the owner generates encryption and decryption keys for all convolutional layers and fully-connected layers in a trained CNN. In LEP-CNN, we consider each element in the input data of convolutional layers and fully-connected layers is $\gamma$-bit long, and $\lambda$ is the security parameter. To ensure the security $\frac{1}{2^{\lambda-\gamma-1}}$ shall be a negligible value in terms of computational secrecy [@Crpto-book-cp3.3], e.g., $<\frac{1}{2^{128}}$. Detailed selection of security parameter is discussed in Section \[ss:security-analysis\].
As described in Algorithm.\[a:off-covl\], given a convolutional layer with a $n\times n\times D$ input, stride as $s$, padding as $p$, and $H$ kernels ($k\times k$ matrices), the owner generates $\{\mathcal{R}_{c,d},1\leq d\leq D\}$ as the encryption keys and $\{\alpha_{i},1\leq i\leq H\}$ as its decryption keys, where $\mathcal{R}_{c,d}$ is a $n\times n$ random matrix and $\alpha_{i}$ is a $(\frac{n-k+2p}{s}+1)\times (\frac{n-k+2p}{s}+1)$ matrix. For expression simplicity, we use $\textbf{Conv}(\mathcal{R}_{c,d},i_{th})$ to denote the convolution operation for the $i_{th}$ kernel with $\mathcal{R}_{c,d}$ as input.
Given a fully-connected layer with a $m$-dimensional vector as input and $T$ neurons, the owner first generates a $m$-dimensional random vector $\mathcal{R}_{f}$. Then, the owner takes $\mathcal{R}_{f}$ as the input of the fully-connected layer to output a $T$-dimensional vector $\beta$. $\mathcal{R}_f$ and $ \beta$ are set as the encryption key and decryption key respectively for this layer.
For a CNN with $x$ convolutional layers and $y$ fully-connected layers, $x$ sets of $\{\mathcal{R}_{c,d},\alpha_{i}\}_{1\leq d\leq D_x, 1\leq i\leq H_x}$ and $y$ sets of $\{\mathcal{R}_f, \beta\}$ are generated by the owner as a final set of keys $\{Enc_{key},Dec_{key}\}$. ***Note that***, *each set of keys is only valid for one CNN request in the later online phase*. Thus, the owner will generate multiple sets of keys according to the necessity of application scenarios as discussed in Section \[ss:key-update\].
Generate random $n\times n$ matrices $\mathcal{R}_{c,d},1\leq d\leq D$
**Set** the Layer Input $\mathcal{M}$ = Input Data **Set** Layers = the collection of all Convolutional Layers and Fully-connected Layers in CNN **Set** Layer = the first Layer from Layers $\mathcal{M}$ as result
Online Phase
------------
During the online phase, the IoT device can efficiently interact with the edge device to process data using CNN in a privacy-preserving manner. The overall process of our online phase is depicted in Algorithm.\[a:online-cnn\]. Specifically, the IoT device offloads encrypted data to the edge device for performing compute-intensive convolutional layers and fully-connected layers. Intermediated results are returned back to the IoT device for decryption. Then, these decrypted results are processed with the follow up activation layer and pooling layer (if exist). Outputs are encrypted and offloaded again if the next layer is a convolutional layer or a fully-connected layer. This procedure is conducted iteratively until all CNN layers are executed.
To fulfill these tasks, we designed two privacy-preserving schemes **PPCL** and **PPFL** for convolutional layers and fully-connected layers respectively.
### PPCL: Privacy-preserving Convolutional Layer {#sss:ppcl}
In *PPCL*, we consider a general convolutional layer with a $n\times n\times D$ input, stride as $s$, padding as $p$, and $H$ kernels with each size of $k\times k\times D$. The $d_{th}$ level of the input is denoted as a $m\times m$ matrix $\mathcal{I}_d$.
**Input Encryption**: The IoT device encrypts the input using the pre-stored keys $\{\mathcal{R}_{c,d}\}$ for this convolutional layer as $$\begin{aligned}
\label{e:ppcl-enc}
Enc(\mathcal{I}_d)=\mathcal{I}_d+\mathcal{R}_{c,d}\end{aligned}$$ where $\{Enc(\mathcal{I}_d)\},1\leq d\leq D$ are sent to the edge device.
**Privacy-preserving Execution**: The edge device takes each $Enc(\mathcal{I}_d),1\leq d\leq D$ as the input of kernels to perform the convolution process. For the $i_{th}$ kernel, the edge device outputs $$\begin{aligned}
&&\sum_{d=1}^D \textbf{Conv}(Enc(\mathcal{I}_d),i_{th})\\
&&=\sum_{d=1}^D \textbf{Conv}(\mathcal{I}_d,i_{th})+\sum_{d=1}^D \textbf{Conv}(\mathcal{R}_{c,d},i_{th})\nonumber\end{aligned}$$ $\sum_{d=1}^D \textbf{Conv}(Enc(\mathcal{I}_d),i_{th}),1\leq i\leq H$ are returned back to the IoT device as intermediate results.
**Decryption and Preparation for the Next Layer**: Given the returned $\sum_{d=1}^D \textbf{Conv}(Enc(\mathcal{I}_d),i_{th}),1\leq i\leq H$, the IoT device quickly decrypts them as $$\begin{aligned}
\sum_{d=1}^D \textbf{Conv}(Enc(\mathcal{I}_d),i_{th})-\alpha_i=\sum_{d=1}^D \textbf{Conv}(\mathcal{I}_d,i_{th})\end{aligned}$$ where $\{\alpha_i=\sum_{d=1}^D \textbf{Conv}(\mathcal{R}_{c,d},i_{th})\}, 1\leq i\leq H$ are the pre-stored decryption keys for this layer. Afterwards, the IoT device performs the activation layer and pooling layer directly over convolutional output, which are extremely compute-efficient. For example, one of the most popular activation layer ReLU only requires translating negative values in the input to 0. The popular max-pooling (or average-pooling) layer simply shrinks the data by taking the max value (or average value respectively) every few values. The output will be encrypted and sent to the edge device using *PPCL* for the next convolutional layer (or *PPFL* respectively for a fully-connected layer).
### PPFL: Privacy-preserving Fully-connected Layer
In *PPFL*, we consider a general fully-connected layer with $T$ neurons and takes a $m$-dimensional vector $\mathcal{V}$ as input.
**Input Encryption**: Given the input, the IoT device encrypts it using the pre-stored encryption key $\mathcal{R}_f$ for this layer as $$\begin{aligned}
\label{e:ppfl-enc}
Enc(\mathcal{V})=\mathcal{V}+\mathcal{R}_f\end{aligned}$$ $Enc(\mathcal{V})$ is then sent to the edge device.
**Privacy-preserving Execution**: On receiving $Enc(\mathcal{V})$, the edge device takes $Enc(\mathcal{V})$ as the input of the fully-connected layer. Specifically, the encrypted outcome $Enc(\mathcal{O}[j]),1\leq j\leq T$ of each neuron is computed as $$\begin{aligned}
Enc(\mathcal{O}[j])&=&\sum_{i=1}^m Enc(\mathcal{V})[i]\times w_{i,j}=\mathcal{O}[j]+\beta[j]
%&=&\sum_{i=1}^m (\mathcal{V}[i]*w_{i,j}+\mathcal{R}_{f}[i]*w_{i,j}) \nonumber \\
%&=&\mathcal{O}[j]+\beta[j] \nonumber \end{aligned}$$ where $w_{i,j}$ is the weight between the $i_{th}$ element of input vector and the $j_{th}$ neuron. $Enc(\mathcal{O})=\{Enc(\mathcal{O}[1]),Enc(\mathcal{O}[2]),\cdots, Enc(\mathcal{O}[T])\}$ is sent back to the IoT device as intermediate results.
**Decryption and Preparation for the Next Layer**: Given the returned $Enc(\mathcal{O})$, the IoT device decrypts each $Enc(\mathcal{O})$ with the pre-stored decryption key $\beta$ of this layer as $$\begin{aligned}
\label{e:ppfl-dec}
\mathcal{O} = Enc(\mathcal{O}) - \beta\end{aligned}$$ Then, the IoT device executes the activation layer with $\mathcal{O}$ as input. The output from the activation layer will be encrypted and sent to edge device using *PPFL* if there are any additional fully-connected layers in the CNN.
To this end, the IoT device is able to efficiently handle each layer in a CNN. Compute-intensive convolutional and fully-connected layers are securely offloaded to the edge using *PPCL* and *PPFL*. These compute-efficient layers are directly handled by the IoT device. Since we develop *PPCL* and *PPFL* as independent modules, they can be customized and recursively plugged into any CNN no matter how many different convolutional layers and fully-connected layers it contains.
Discussion - Storage and Update of Pre-computed Keys {#ss:key-update}
----------------------------------------------------
LEP-CNN considers two major types of resource-constrained IoT devices that run CNN-driven applications.
- Type-1: Mobile IoT devices with limited battery life and computational capability, such as drones.
- Type-2: Static devices with power supply but has limited computational capability, such as security cameras.
The type-1 devices are usually deployed to perform tasks for a period time. Therefore, before each deployment, the device owner can pre-load enough keys to support its CNN tasks. With regards to the type-2 devices, the owner can perform an initial key pre-loading and then use remote update to securely add new offline keys as described in Fig.\[f:key-update\].
![Key Update for Power Connected Devices[]{data-label="f:key-update"}](key-loading.pdf){width="8cm"}
LEP-CNN proposes to ensure the timely processing of CNN requests when they are needed on IoT devices. Instead of performing real-time CNN requests on every piece of data collected, resource-constrained IoT devices usually require in-depth analytics using CNN when specific signals are detected. Taking real-time search and monitoring using drones as an example application for type-1 devices, fast local processing will be first performed for data collected to get estimated results [@uav-cloud]. Once suspicious signs are detected in estimated results, CNN based analytics are further conducted for a small set of data (e.g., video frames with the detected suspicious object). Given the high efficiency of LEP-CNN, the performance of such CNN requests will be timely supported when enough pre-computed keys are available. For example, when the average frequency of CNN requests is every one per ten seconds for a drone, only 360 sets of pre-computed keys are needed for one-hour deployment, which is longer than most current drones’ battery life [@uav-battery]. Security camera is an example of type-2 devices, which requires CNN-based analytics to extract detailed information only when alarm is triggered by motion or audio sensors of the camera. Similar to the drone case, LEP-CNN can timely support the peak CNN requests when suspicious signs are detected.
Assuming the average frequency of CNN-required alarm in a security camera is one per 10 minutes, and each alarm requires 5 CNN requests, 720 sets of pre-computed keys are needed for one-day usage. As evaluated in Section \[ss:storage\], an IoT device with a 32GB SD card is able to store keys to support 1600 requests for AlexNet. Such a result indicates 4.4 deployments and a 2.22-day support for type-1 and type-2 devices respectively when using AlexNet.
Note that, LEP-CNN is designed for low-cost resource-constrained devices that require timely processing of moderate (or low) frequent CNN requests. For application scenarios that involve a large number of constant CNN requests, e.g., security critical surveillance systems, computational powerful devices are suggested to handle CNN requests directly at local.
Discussion - Integrity Check on Returned Data {#ss:integrity-check}
---------------------------------------------
In a scenario where the edge devices are “curious-and-dishonest”, those edge devices may perform dishonest behaviors so that they can save their resource utilities. After receiving inference requests from the IoT devices, the edge devices may cheat the IoT devices by skipping the heavy convolutional operations and sending back random, apparently not correct, results to the IoT devices. These incorrect results can badly effect or even completely ruin the final result of the entire CNN inference. In order to ensure the integrity of returned data from edge devices, LEP-CNN also provides an optional integrity check functionality with only a minor efficiency cost. By enabling the integrity check, the IoT device can achieve an error detection rate of $99\%$ while only losing $1.1\%$ in offload percentage in the worst case compared with LEP-CNN disabling integrity check. The users can decide whether to turn on this functionality based on the actual deployment scenario and the trustworthiness of the edge devices.
The basic strategy of integrity check is to first sample a small portion of elements from returned data in each layer and then check whether there is an incorrect result occurring in the selected elements. To validate the correctness of a single element, the IoT device needs to go through the corresponded convolution operations locally. Although resource consuming, with high probability, this validation process can block IoT devices from taking incorrect results into next layer. The error detection rate $Pr(ED)$ can be calculated as below: $$\begin{aligned}
Pr(ED) = 1 - \frac{{(1-\theta)N \choose rN}}{{N \choose rN}}\end{aligned}$$
where $N$, $r$ and $\theta$ is the size, sample rate and error rate of returned data in a convolutional layer. ${(1-\theta)N \choose rN}$ is the combination operation for selecting $r \times N$ elements from $(1-\theta) \times N$ elements. In order to increase the error detection rate while lowering the additional validation computation on IoT device, we provide detailed evaluation by performing numerical analysis and practical experiments in Section \[s:analysis\] and Section \[s:evaluation\] respectively. Based on our experiment results on AlexNet, LEP-CNN with integrity check turned on could achieve over 99% error detection rate while maintaining a similar computation offload rate compared with LEP-CNN with integrity check turned off.
Analysis of LEP-CNN {#s:analysis}
===================
Security Analysis {#ss:security-analysis}
-----------------
\[th-1\] Given the ciphertext $\mathcal{C}$ of a $\gamma$-bit random message $\mathcal{M}$ generated using $\textit{PPCL}$ or $\textit{PPFL}$ in LEP-CNN, the probability for a probabilistic polynomial time (PPT) adversary $\mathcal{A}$ to output a correct guess for $\mathcal{M}$ shall have $$\begin{aligned}
Pr[(\mathcal{M}^*=\mathcal{M})|\mathcal{C}] - Pr[\mathcal{M}^*=\mathcal{M}] \leq \epsilon\end{aligned}$$ where $\epsilon$ is a negligible value in terms of computational secrecy [@Crpto-book-cp3.3], $\mathcal{M}^*$ is $\mathcal{A}$’s guess for $\mathcal{M}$, and $Pr[\mathcal{M}^*=\mathcal{M}]$ is the probability $\mathcal{A}$ makes a correct without ciphertext. Specifically, the corresponding ciphertext generated using $\textit{PPCL}$ or $\textit{PPFL}$ only introduces negligible additional advantages to $\mathcal{A}$ for making a correct guess of $\mathcal{M}$.
As shown in Eq.\[e:ppcl-enc\], each level of the input data $\mathcal{I}_d$ in *PPCL* is encrypted by adding an *one-time* random matrix $\mathcal{R}_{c,d}$. With regards to each $\gamma$-bit element in $\mathcal{I}_d$, it is encrypted by adding a $\lambda$-bit random number $\mathcal{R}_{c,d}[e]$, i.e., $Enc(\mathcal{I}_d[e])=\mathcal{I}_d[e]+\mathcal{R}_{c,d}[e]$. Similarly, each element $\mathcal{V}[e]$ in the input of *PPFL* is encrypted with a random number $\mathcal{R}_{f}[e]$ as $Enc(\mathcal{V}[e])=\mathcal{V}[e]+\mathcal{R}_{f}[e]$. For expression simplicity, we use $\mathcal{M}$ to denote a $\gamma$-bit input element for *PPCL* or *PPFL*, $R$ is the random number to encrypt $\mathcal{M}$ as $\mathcal{C}=\mathcal{M}+R$. Note that, $R$ will be re-generated for each encryption using *PPCL* or *PPFL*.
To make a correct guess of $\mathcal{M}$ without the ciphertext, the adversary $\mathcal{A}$ has $Pr[\mathcal{M}^*=\mathcal{M}]=\frac{1}{2^\gamma}$, where $\mathcal{M}^*$ is $\mathcal{A}$’s guess for $\mathcal{M}$.
By given a ciphertext $\mathcal{C}$, there are $2^\gamma$ possible values for its plaintexts $\mathcal{M}$ if $2^\gamma\leq \mathcal{C} \leq 2^\lambda-2^\gamma$, because $\mathcal{C}$ has the same distribution as the random number $R$ [@Crpto-book-cp11]. Now, if the random number $R$ used for encryption is the range of $[2^\gamma,2^\lambda-2^\gamma]$, we have $Pr[(\mathcal{M}^*=\mathcal{M})|\mathcal{C}] = Pr[\mathcal{M}^*=\mathcal{M}] = \frac{1}{2^\gamma}$. When $\mathcal{C}<2^\gamma$ or $\mathcal{C}>2^\lambda$, we have $Pr[(\mathcal{M}^*=\mathcal{M})|\mathcal{C}]>1/2^\gamma$. This is because the total possible inputs are reduced to $\mathcal{C}$ or $\mathcal{C}-2^\lambda$ respectively. Fortunately, the probability $Pr[\mathcal{C}<2^\gamma]$ or $Pr[\mathcal{C}>2^\lambda]$ in LEP-CNN is negligible when appropriate security parameter is selected. To be specific, $\mathcal{C}<2^\gamma$ or $\mathcal{C}>2^\lambda$ can appear when $R<2^\gamma$ or $R>2^\lambda-2^{\gamma}$. As $Pr[R<2^\gamma]=Pr[R>2^\lambda-2^{\gamma}]=\frac{2^\gamma}{2^\lambda}$, we have $$\begin{aligned}
&&Pr[R<2^\gamma~or~R>2^\lambda-2^{\gamma}] \nonumber\\
&&=Pr[R<2^\gamma]+Pr[R>2^\lambda-2^{\gamma}]=\frac{1}{2^{\lambda-\gamma-1}} \nonumber\end{aligned}$$ Thus, to guarantee $\frac{1}{2^{\lambda-\gamma-1}}$ is a negligible probability, such as $\frac{1}{2^{128}}$, LEP-CNN can set the security parameter $\lambda$ according to size of input message, i.e., $\lambda-\gamma-1>128$.
We now use $\epsilon=\frac{1}{2^{\lambda-\gamma-1}}$ to denote the negligible probability, and get the probability $Pr[(\mathcal{M}^*=\mathcal{M})|\mathcal{C}]$ as $$\begin{aligned}
Pr[(\mathcal{M}^*=\mathcal{M})|\mathcal{C}]\leq \frac{1}{2^\gamma}*(1-\epsilon)+1*\epsilon = \frac{1}{2^\gamma}+(1-\frac{1}{2^\gamma})\epsilon \nonumber\end{aligned}$$ where $\frac{1}{2^\gamma}*(1-\epsilon)$ is the probability for a correct guess for $2^\gamma\leq R\leq 2^\lambda-2^\gamma$, and the “1” in $1*\epsilon$ is best probability for a correct guess $\mathcal{A}$ can have when $[R<2^\gamma~or~R>2^\lambda-2^{\gamma}]$. As a result, we get $$\begin{aligned}
Pr[(\mathcal{M}^*=\mathcal{M})|\mathcal{C}] - Pr[\mathcal{M}^*=\mathcal{M}]\leq(1-\frac{1}{2^\gamma})\epsilon<\epsilon \nonumber\end{aligned}$$ Since $\epsilon$ is negligible value, Theorem \[th-1\] is proved.
[ | c | c | c | c | c | c | c |]{}\
& & & **Offloaded Cost**& **Communication**&\
& & **Input** & **Results** &**to the Edge**&**Cost**&\
& & **Encryption** & **Decryption** &**(FLOPs)**&(**Elements**)&**(Elements)**\
**Convolutional** & $n\times n \times D$ &$Dn^2$ &$H(\frac{n-k+2p}{s}+1)^2$&$2DHk^2(\frac{n-k+2p}{s}+1)^2$ &$Dn^2+H(\frac{n-k+2p}{s}+1)^2$&$Dn^2+$\
& & &&&&$H(\frac{n-k+2p}{s}+1)^2$\
**Fully-connected**& $m$ & $m$ & $T$ & $2mT$ &$m+T$ &$m+T$\
\
**Convolutional** & $n\times n \times D$ &N/A &N/A&$2DHk^2(\frac{n-k+2p}{s}+1)^2$ &$Dn^2+H(\frac{n-k+2p}{s}+1)^2$&N/A\
**Fully-connected**& $m$ & N/A&N/A & $2mT$ &$m+T$ &N/A\
In this table: $s$ is the stride, $p$ is the size of padding, $H$ is the number of kernels, $k\times k$ is the size of kernels of a convolutional layer; $T$ is the number of neurons of a fully-connected layer. Each element is 20 Bytes.
[ | c | c | c | c | c | c | c |]{}\
& & **Offloaded Cost**& **Communication**&\
& **Input** & **Results** & **Results** &**to the Edge**&**Cost**&\
& **Encryption** & **Decryption** & **Validation** &**(FLOPs)**&(**Elements**)&**(Elements)**\
**Integrity** & &$H(\frac{n-k+2p}{s}+1)^2$ &$2Dk^2\times$&$2DHk^2\times$ &$Dn^2+$&$Dn^2+Hk^2+$\
**Check**& & &$\lceil rH(\frac{n-k+2p}{s}+1)^2 \rceil$&$(\frac{n-k+2p}{s}+1)^2$&$H(\frac{n-k+2p}{s}+1)^2$&$H(\frac{n-k+2p}{s}+1)^2$\
**No Integrity**& & $H(\frac{n-k+2p}{s}+1)^2$ & $0$ & $2DHk^2\times$ &$Dn^2+$ &$Dn^2+$\
**Check**&&&&$(\frac{n-k+2p}{s}+1)^2$&$H(\frac{n-k+2p}{s}+1)^2$&$H(\frac{n-k+2p}{s}+1)^2$\
In this table: $s$ is the stride, $p$ is the size of padding, $H$ is the number of kernels, $k\times k$ is the size of kernels of a convolutional layer; $\theta$ is the error rate of the returned data; $r$ is the sample rate of the returned data. Each element is 20 Bytes.
[ |c |c |c | c |c | c |c |c |]{} &**Parameters**&**Input Size**&**Computation**&**Offloaded Cost**&**Offloaded**&**Communication**&**Storage**\
&&&**of the IoT**&&**Percentage**&**Cost**&**Overhead**\
&n=227, H=96 & & & & & &\
&k=11, s=4 & $227\times227\times 3$ & 444,987 FLOPs & 210,830,400 FLOPs & 99.79% & 8691.15 KB & 8691.15 KB\
&n=27, H=256 & & & & & &\
&k=5, s=1 & $27\times27\times 96$ & 256,608 FLOPs & 895,795,200 FLOPs & 99.97% & 5011.88 KB & 5011.88 KB\
&n=13, H=384 & & & & & &\
&k=3, s=1 & $13\times13\times 256$ & 108,160 FLOPs & 299,040,768 FLOPs & 99.96% & 2112.50 KB & 2112.50 KB\
&n=13, H=384 & & & & & &\
&k=3, s=1 & $13\times13\times 384$ & 129,792 FLOPs & 448,561,152 FLOPs & 99.97% & 2535.00 KB & 2535.00 KB\
&n=13, H=256 & & & & & &\
&k=3, s=1 & $13\times13\times 384$ & 108,160 FLOPs & 299,040,768 FLOPs & 99.96% & 2112.50 KB & 2112.50 KB\
**FC-1** &m=9216,T=4096& 9216 & 13,312 FLOPs & 75,497,472 FLOPs & 99.98% & 260.00 KB & 260.00 KB\
**FC-2** &m=4096,T=4096& 4096 & 8,192 FLOPs & 33,554,432 FLOPs & 99.98% & 160.00 KB & 160.00 KB\
**FC-3** &m=4096,T=1000& 4096 & 5,096 FLOPs & 8,192,000 FLOPs & 99.94% & 99.53 KB & 99.53 KB\
**Total Cost** &N/A &N/A & 1,074,307 FLOPs & 2,270,512,192 FLOPs & 99.95% & 20.49 MB & 20.49 MB\
\
------------ ---------------------- ----------------- ------------------ --------------- ------------------ --------------- ------------------ ---------------
[**No Integrity**]{} **Integrity** **No Integrity** **Integrity** **No Integrity** **Integrity** **No Integrity** **Integrity**
**Check** **Check** **Check** **Check** **Check** **Check** **Check** **Check**
**Conv-1** 444,987 FLOPs 866,793 FLOPs 99.79% 99.59% 8691.15 KB 8691.27 KB 8691.15 KB 8918.03 KB
**Conv-2** 256,608 FLOPs 2,944,608 FLOPs 99.97% 99.67% 5011.88 KB 5011.98 KB 5011.88 KB 5136.88 KB
**Conv-3** 108,160 FLOPs 2,504,320 FLOPs 99.96% 99.16% 2112.50 KB 2112.60 KB 2112.50 KB 2180.00 KB
**Conv-4** 129,792 FLOPs 3,724,032 FLOPs 99.97% 99.17% 2535.00 KB 2535.10 KB 2535.00 KB 2602.50 KB
**Conv-5** 108,160 FLOPs 3,398,272 FLOPs 99.96% 98.86% 2112.50 KB 2112.59 KB 2112.50 KB 2157.50 KB
------------ ---------------------- ----------------- ------------------ --------------- ------------------ --------------- ------------------ ---------------
Numerical Analysis
------------------
The numerical analysis of LEP-CNN is summarized in Table \[t:na-summ\]. For expression simplicity, we use one floating point operation ***FLOP*** to denote an addition or a multiplication. For a general convolutional layer, $n\times n\times D$ is the size of input, $s$ is the stride, $p$ is the size of padding, $H$ is the number of kernels, and $k\times k$ is the size of kernel matrix. For a general fully-connected layer, $m$ is the dimension of the input vector, $T$ is the number of neurons. For a pooling layer, $q\times q$ is the size of pooling regions. We use an uncompressed AlexNet [@AlexNet] architecture as the study case for analysis, which is a complex CNN architecture that requires 2.27 billion FLOPs for each inference request, which has comparable computing loads as the current prevalent FaceNet (1.6 billion FLOPs) [@FaceNet] and ResNet (3.6 billion FLOPs) [@ResNet].
### Computational Cost {#sss:comp-cost}
In the *Online* phase of LEP-CNN, the IoT device offloads compute-intensive convolutional layers and fully-connected layers to the edge devices. Given a general convolutional layer, the IoT device only needs to perform $D$ matrix addition with $Dn^2$ FLOPs for encryption and $H(\frac{n-k+2p}{s}+1)^2$ FLOPs for decryption respectively. Compared with executing the same convolutional layer fully on the IoT device, which takes $2DHk^2(\frac{n-k+2p}{s}+1)^2$ FLOPs, LEP-CNN significantly reduces real-time computation on the IoT device. It is worth to note that the stride $s$ in a convolutional layer is typically a small value (e.g., 1 or 2). For a general fully-connected layer, the IoT device needs to perform $m$ FLOPs for encryption and $T$ FLOPs for decryption as shown in Eq.\[e:ppfl-enc\] and Eq.\[e:ppfl-dec\] respectively. Differently, if the IoT device executes such a fully-connected layer at local, $2mT$ FLOPs are needed.
Besides the offloading of convolutional layers and fully-connected layers, the IoT device also needs to process non-linear layers at local. Fortunately, these non-linear layers are extremely compute-efficient. Taking the widely adopted activation layer - ReLU as an example, it only requires $\frac{1}{2Dk^2}$ of its previous convolutional layer’s cost, and $\frac{1}{2m}$ of the cost if the previous layer is a fully-connected layer.
We now discuss the computational cost of LEP-CNN using AlexNet. As shown in Table \[t:AlexNet\], LEP-CNN can offload over 99.9% computational cost for convolutional layers and fully-connected layers, and only leaves lightweight encryption and decryption on the IoT device. Compared with the offloaded convolutional layers and fully-connected layers, the local execution of all non-linear layers only requires $0.08\%$ operations for AlexNet. This result further affirms our motivation to offload convolutional layers and fully-connected layers.
With regards to the encrypted execution on the edge device, LEP-CNN achieves the same computational cost as that directly using unencrypted data as shown in Table \[t:na-summ\]. This is because our encryption (Eq.\[e:ppcl-enc\] and Eq.\[e:ppfl-enc\]) in *PPCL* and *PPFL* schemes make the ciphertexts $Enc(\mathcal{I}_d)$ and $Enc(\mathcal{V})$ remain the same dimension as their plaintexts $\mathcal{I}_d$ and $\mathcal{V}$. Such a decent property guarantees real-time computational performance on the edge device.
In the *Offline* phase, the IoT device owner first prepares encryption keys by choosing random matrices for convolutional layers and fully-connected layers that will be offloaded. Meanwhile, the owner will take these encryption keys as inputs for their corresponding convolutional layers or fully-connected layers to obtain results as the decryption keys. In Section \[s:evaluation\], we show that the offline phase can be efficiently executed using a regular laptop.
### Communication Cost
The communication cost of LEP-CNN comes from the transmission of encrypted inputs and outputs of convolutional layers and fully-connected layers. In our implementation, we use 160-bit random numbers (i.e., $\lambda=160$) during all encryption processes in Eq.\[e:ppcl-enc\] and Eq.\[e:ppfl-enc\]. Thus, each element in the ciphertext (a matrix or a vector) is 20-Byte long. To offload a convolutional layer with a $n\times n \times D$ input, the IoT device first sends its corresponding ciphertext contains $D$ encrypted matrices with $Dn^2$ elements in total. Then, $H$ encrypted result matrices are received from the edge device with each size of $(\frac{n-k+2p}{s}+1)\times (\frac{n-k+2p}{s}+1)$. With regards to the offloading of a fully-connected layer that takes a $m$-dimensional vector as input, the IoT device needs to send a $m$-dimensional vector as encrypted input and receive a $T$-dimensional vector as encrypted output from the edge device.
As shown in Table \[t:AlexNet\], the communication cost for an offloading of the AlexNet is 20.49MB, which can be efficiently handled under the edge computing environment [@5G-MEC].
---------------- ----------------- --------------------- ---------------------- ------------------- -------------- -------------------
**IoT without**
**Offloading** **IoT Computation** **Edge Computation** **Communication** **Total**
**(second)** **(second)** **(second)** **(second)** **(second)**
**Conv-1** 10.01 0.037 [0.0103]{} 0.849 [0.896]{} 11.17$\times$
**Conv-2** 40.68 0.0405 0.0435 0.489 [0.573]{} [70.99]{}$\times$
**Conv-3** 19.93 0.0437 0.013 0.206 [0.263]{} [75.78]{}$\times$
**Conv-4** 29.78 0.0498 0.0184 0.248 [0.316]{} [94.24]{}$\times$
**Conv-5** 19.88 0.0420 [0.0127]{} 0.206 [0.261]{} [76.17]{}$\times$
**FC-1** 2.22 0.0013 0.0043 0.025 [0.031]{} [71.61]{}$\times$
**FC-2** 1.08 0.001 [0.0025]{} 0.016 [0.019]{} [56.84]{}$\times$
**FC-3** 0.27 0.0008 [0.0009]{} 0.01 [0.012]{} [22.5]{}$\times$
**Non-linear** 1.137 1.137 N/A N/A 1.137 N/A
**Total Cost** 124.99 1.353 [0.106]{} 2.049 [3.508]{} [35.63]{}$\times$
---------------- ----------------- --------------------- ---------------------- ------------------- -------------- -------------------
---------------- ----------------- --------------------- --------------- ------------------------ -------------------
**IoT without**
**Offloading** **Integrity Check** **No Integrity Check**
**(second)** **(second)** **(second)**
**Conv-1** 10.01 0.916 10.93$\times$ [0.896]{} 11.17$\times$
**Conv-2** 40.68 0.655 62.11$\times$ [0.573]{} [70.99]{}$\times$
**Conv-3** 19.93 0.402 49.58$\times$ [0.263]{} [75.78]{}$\times$
**Conv-4** 29.78 0.524 56.83$\times$ [0.316]{} [94.24]{}$\times$
**Conv-5** 19.88 0.470 42.30$\times$ [0.261]{} [76.17]{}$\times$
**FC-1** 2.22 0.031 71.61$\times$ [0.031]{} [71.61]{}$\times$
**FC-2** 1.08 0.019 56.84$\times$ [0.019]{} [56.84]{}$\times$
**FC-3** 0.27 0.012 22.5 $\times$ [0.012]{} [22.5]{}$\times$
**Non-linear** 1.137 1.137 N/A 1.137 N/A
**Total Cost** 124.99 4.166 30.00$\times$ [3.508]{} [35.63]{}$\times$
---------------- ----------------- --------------------- --------------- ------------------------ -------------------
### Storage Overhead {#ss:storage}
For the offloading of a convolutional layer with a $n\times n \times D$ input, the IoT device needs to store $D$ random matrices with $n^2$ elements each as the encryption keys, and $H$ matrices with size of $(\frac{n-k+2p}{s}+1)\times (\frac{n-k+2p}{s}+1)$ as the decryption keys. To offload a fully connected layer with a $m$-dimensional vector as input, a $m$-dimensional vector and a $T$-dimensional vector need to be pre-stored as the encryption key and decryption key respectively.
Table \[t:AlexNet\] shows the offloading of an AlexNet request needs 20.49MB storage overhead. With the rise of IoT devices, low-power-consumption SD memory card has become an excellent fit to economically extend the storage of IoT devices [@SD-IoT], which usually have more than 32GB capacity.
### Additional Resource Consumption of Integrity Check {#ss:analysis-of-integrity-check}
Turning on the integrity check leads to additional resource consumption to local IoT device. As shown in Table \[t:Numerical-Analysis-Integrity-Check\], given a returned matrix of size $H(\frac{n-k+2p}{s}+1)^2$ and a sample rate of $r$, the validation process in Section \[ss:integrity-check\] brings $\lceil rH(\frac{n-k+2p}{s}+1)^2 \rceil$ additional computation and makes the total computational cost of IoT devices rise to $2Dk^2\lceil rH(\frac{n-k+2p}{s}+1)^2 \rceil$. Since any convolutional result in the entire set of response map can be incorrect, IoT devices need to store all kernel parameters of each convolutional layer locally, which adds on $Hk^2$ storage overhead and makes the total IoT storage overhead to be $Dn^2 + H(\frac{n-k+2p}{s}+1)^2 + Hk^2$.
Table \[t:comparison-integrity-check\] shows the resource consumption comparison between LEP-CNN with integrity check turned on and turned off. The results are calculated when error rate $\theta = 1\%$ and sample rate $r = 0.2\%, 0.3\%, 0.8\%, 0.8\%, 1.1\%$ for Conv-1 - Conv-5 respectively. Under this setting, IoT device can achieve $99\%+$ error detection rate in each convolutional layer. Since all the multiplication results of $r\theta$ are less equal to $1.1\times10^{-4}$, the additional communication costs resulted from integrity check are tiny. As a result, the communication increments are less than $4.74\times10^{-3}\%$ of the original communication costs. Compared with the heavy parameters in fully-connected layers, the parameters in convolutional layers only stand for a minor portion of the entire neural network model. Thus, even the highest additional storage overhead is only 227 KB while the lowest increment can be as low as 45 KB.
Prototype Evaluation {#s:evaluation}
====================
We implemented a prototype of LEP-CNN using Python 2.7. In our implementation, TensorFlow and Keras libraries are adopted to support CNNs. The resource-constrained IoT device is a Raspberry Pi (Model A) with Raspbian Debian 7, which has 700 MHz single-core processor, and 256MB memory, and 32GB SD card storage. The edge device and the IoT device owner is a Macbook Pro laptop with OS X 10.13.3, 3.1 GHz Intel Core i7 processor, 16GB memory, and 512GB SSD. The IoT device and the edge device are connected using WiFi in the same subnet. We use the well-known ImageNet [@imagenet_cvpr09] as the dataset for the evaluation of AlexNet. The security parameter $\lambda$ is set as 160 in our implementation. We also implemented an AlexNet-structured CryptoNets [@CryptoNets] as an example to compare our scheme with homomorphic encryption based privacy preserving neural networks [@CryptoNets; @CryptoDL; @IACR-PPDL].
Evaluation Results
------------------
***Efficiency***: In this section, we first evaluate the efficiency of offline phase of LEP-CNN, and then discuss the online phase for CNN execution efficiency.
To generate the encryption and decryption keys for the execution of one AlexNet request, LEP-CNN only requires 114ms for the IoT device owner. While each set of keys will only be used for one request, the owner can efficiently compute more than 2600 sets of keys for AlexNet using 5 minutes.
In the online phase, the IoT device in LEP-CNN efficiently executes a CNN request with privacy-preserving offloading to the edge device. Table \[t:AlexNet-IoT-Comparison\] summarizes the evaluation results of LEP-CNN on AlexNet. By applying LEP-CNN, the required computational time on the IoT device is reduced to about $\frac{1}{92}$ for AlexNet. With such a high computational reduction, we not only overcome the challenges from limited computational resources of IoT devices, but also save energy consumption for them to achieve longer battery life. The other part of computational cost of LEP-CNN is from the privacy-preserving execution of convolutional layers and fully-connected layers on the edge device. As shown in the fourth column of Table \[t:AlexNet-IoT-Comparison\], the edge device served by a laptop can efficiently handle these operations using encrypted data. In practice, the selection of layers to offload in CNNs can be customized according to their complexity, since our *PPCL* and *PPFL* schemes are designed as independent modules for flexible combination. We also compare the privacy-preserving execution of CNN layers on the edge device with that using unencrypted data. Table \[t:comp-no-privacy\] shows that the encryption execution on the edge device using our scheme spends almost the same time as executing these layers without privacy protection. This is also consistent with our numerical analysis in Section \[sss:comp-cost\], since our ciphertext has the same dimensions as its corresponding plaintext.
As our scheme requires the interaction between the IoT device and the edge device during the execution of a CNN request, another part of cost of LEP-CNN is the communication cost. In our implementation, we use a wireless network with 10MB/s transmission speed between the IoT device and the edge device. In real-world scenario, the devices are likely to be connected via wired or cellular connection, which allows a higher transmission speed than our experimental environment. As presented in the fifth column of Table \[t:AlexNet-IoT-Comparison\], the total communication time for AlexNet in our network environment is only about $\frac{1}{61}$ compared with processing the entire AlexNet on the IoT device without LEP-CNN. Moreover, the upcoming 5G era for MEC environment will significantly empower the transmission speed [@5G-MEC] and further optimize the communication performance our scheme.
We now compare the total cost of LEP-CNN with directly executing CNN on the IoT device. As shown in Table \[t:AlexNet-IoT-Comparison\], LEP-CNN can speed up the execution of an AlexNet request for 35.63$\times$. Among convolutional layers and fully-connected layers in AlexNet, LEP-CNN can speed up the execution for over 90$\times$. In Table \[t:experiment-integrity-check\], when the integrity check is turned on, LEP-CNN can still achieve a high speedup rate of $30.00\times$ compared with AlexNet local execution. These results also validate the scalability of LEP-CNN. More to mention, with increasing complexity of convolutional layers and fully-connected layers, LEP-CNN retains or increases the high speedup rate as shown in the last column of Table \[t:AlexNet-IoT-Comparison\] and \[t:experiment-integrity-check\]. Taking AlexNet as example, the highest speedup rates for them are all achieved with these more complex layers. Therefore, LEP-CNN is promising to be scaled to support more complex CNN architectures according to practical requirements.
---------------- -------------- ---------------------------
**LEP-CNN** **No Privacy Protection**
**(Second)** **(Second)**
**Conv-1** 0.014 0.012
**Conv-2** 0.0435 0.041
**Conv-3** 0.013 0.013
**Conv-4** 0.0184 0.016
**Conv-5** 0.012 0.012
**FC-1** 0.0043 0.004
**FC-2** 0.0065 0.0063
**FC-3** 0.0022 0.002
**Total Cost** 0.123 0.106
---------------- -------------- ---------------------------
: Executing Each Layer of AlexNet using LEP-CNN and Non-privacy-preserving Approach on the Edge[]{data-label="t:comp-no-privacy"}
***Energy Consumption***: Compared with fully executing AlexNet inference task on the IoT device with high energy consumption, LEP-CNN significantly saves the energy consumption for computation of the IoT device while introducing slight extra energy consumption for communication. In our evaluation, the IoT device (Raspberry Pi Model A) is powered by a 5V micro-USB adapter. The voltage and current is measured using a Powerjive USB multimeter [@powerjive]. Table \[t:energy-consumption\] shows the average IoT power consumption under different IoT device status. We observe that the network connection is a major power cost in IoT device. An idle IoT device with network connection can have a comparable power cost as executing AlexNet locally without network connection. In our measurement, the average active current consumption for the IoT device is $162mA$, which indicates at least 101.24J energy consumption when fully executing an inference task on the IoT device with 124.99 seconds as stated in Table \[t:energy-consumption\]. Differently, LEP-CNN reduces the computation on the IoT device to 1.353 seconds (1.59J energy consumption) with 2.049 seconds extra communication (2.90J energy consumption). Therefore, LEP-CNN can save IoT energy consumption by $\frac{101.19-(1.59+2.90)}{101.19}=95.56\%$.
---------------- ------------------------ ------------------------ ----------------- -------------------
**IoT without**
**Offloading &** **Idle IoT with** **IoT** **IoT**
**Network Connection** **Network Connection** **Computation** **Communication**
**Power (W)** 0.81 0.78 1.17 1.42
**Energy (J)** 101.19 N/A 1.59 2.90
---------------- ------------------------ ------------------------ ----------------- -------------------
***Accuracy***: To validate that there is no accuracy loss in LEP-CNN, we also implemented original AlexNet without any encryption. By using the same parameters, LEP-CNN achieves the exact same accuracy ($80.1\%$) as that obtained using original AlexNet [@AlexNet] without any encryption, because there is no approximation design in LEP-CNN.
***Evaluation of Sample Rate $r$***: In order to achieve a high error detection rate, different sample rate $r$ needs to be calculated based on different settings in each convolutional layer. As shown in Figure \[f:sample-rate-error-detection-rate\], to make the error detection rate to surpass $99\%$, Conv-1 - Conv-5 need to set $r$ to be $0.2\%, 0.3\%, 0.8\%, 0.8\%, 1.1\%$ respectively. Figure \[f:sample-rate-returned-data-size\] shows that as the size of the returned data rises, the sample rate $r$ required to reach $99\%+$ error detection rate drops correspondingly. From this observation combined with Figure \[f:sample-rate-offloaded-computation-percentage\], the scalability of the integrity check feature is ensured and the additional resource consumption of a larger, more complex CNN is always minor compared with its original costs.
![Evaluation of Sample Rate $r$ and Error Detection Rate[]{data-label="f:sample-rate-error-detection-rate"}](r-pred){width="8cm"}
![Evaluation of Sample Rate $r$ and Returned Data Size[]{data-label="f:sample-rate-returned-data-size"}](r-size){width="8cm"}
![Evaluation of Convolutional Layers and Offloaded Computation Percentage[]{data-label="f:sample-rate-offloaded-computation-percentage"}](r-computation){width="8cm"}
----------------- --------------- ------------------------------
**LEP-CNN** **AlexNet Scale CryptoNets**
**(seconds)** **(seconds)**
**Encryption** 0.012 0.360
**Convolution** 0.0103 625.856
**Decryption** 0.025 0.271
----------------- --------------- ------------------------------
: Comparison Between LEP-CNN and CryptonNets in First Convolutional Layer of AlexNet[]{data-label="t:cryptonets-comparison"}
***Comparison with CryptoNets***: To compare LEP-CNN with homomorphic encryption based privacy preserving neural networks [@CryptoNets; @CryptoDL; @IACR-PPDL] under the same scale, we implemented AlexNet in CryptoNets version, denoted as A-CryptoNets, using the same network settings as in [@AlexNet] and same YASHE cryptosystem [@YASHE] as in [@CryptoNets]. Table \[t:cryptonets-comparison\] shows the efficiency of the first convolutional layer in LEP-CNN and A-CryptoNets. Due to the large input size of AlexNet, the polynomials in A-CryptoNets needs to be big enough to prevent multiplication overflow. These big polynomials make the first convolutional operation cost A-CryptoNets over 10 minutes, which is even more than the total time cost of walking through each layer in LEP-CNN. Although processing requests in batches can help relieve the high computation cost of each inference request in A-CryptoNets, additional time cost is introduced to collect and form single requests into a batch. Thus, in time-sensitive scenarios, LEP-CNN has a better performance than homomorphic encryption based solutions.
Related Work {#s:related-work}
============
To enable the offloading of the CNN in a privacy-preserving manner, Gilad-Bachrach et al. [@CryptoNets] propose a CryptoNets using “YASHE” homomorphic encryption. CryptoNets allows cloud computing servers to perform the inference stage of a CNN using encrypted data only. Meanwhile, the activation function in the activation layers is replaced with the squared function to integrate homomorphic encryption. However, high computational and communication cost are introduced in CryptoNets due to the utilization of homomorphic encryption, and thus making them inefficient for time-sensitive application. In addition, the encryption cost for each request using homomorphic encryption is also expensive for resource-constrained IoT devices. Another limitation of CryptoNets is that its effectiveness can only be guaranteed for small number of activation layers as indicated in ref [@IACR-PPDL]. As a result, only small scale CNN architectures can be supported. To improve CryptoNets, Chabanne et al. [@IACR-PPDL] utilize low-degree polynomials to approximate activation layers. In addition, a normalize layer is added before the non-linear layer with batch normalization, with which polynomial approximation only needs to be accurate on a small and fixed interval. Nevertheless, ref [@IACR-PPDL] follows the same homomorphic encryption-based design for privacy protection as that in CryptoNets, and also suffers from the high computational and communication cost. Recently, ref [@CryptoDL] introduces CryptoDL that enhances CryptoNets in terms of efficiency and accuracy. In CryptoDL, low-degree polynomial-based approximation designs are proposed to support commonly used activation functions (i.e. ReLU, Sigmoid, and Tanh) in activation layers. While CryptoDL reduces about $50\%$ run time compared with CryptoNets, it still requires high local encryption cost on the IoT device. Furthermore, time-sensitive tasks require IoT devices to process data on-the-fly. Unfortunately, these existing research [@CryptoNets; @IACR-PPDL; @CryptoDL] are more suitable for the “Data Collection and Post-Processing” routine, since they require the batch processing of a large number of requests to improve efficiency.
Conclusion {#s:conclusion}
==========
In this paper, we proposed LEP-CNN that enables resource constrained IoT devices to efficiently execute CNN requests with privacy protection. LEP-CNN uniquely designs a lightweight online/offline encryption scheme. By discovering the fact that linear operations in CNNs over input and random noise can be separated, LEP-CNN can pre-compute decryption keys to remove random noise and thus boosting the performance of real-time CNN requests. By integrating local edge devices, LEP-CNN ameliorates the network latency and service availability issue. LEP-CNN also makes the privacy-preserving operation on the edge device as efficient as that on unencrypted data. Moreover, the privacy protection in LEP-CNN does not introduce any accuracy loss to the CNN inference. LEP-CNN also provides optional integrity check functionality to help IoT devices detect erroneous results from dishonest edge devices. Thorough security analysis is provided to show that LEP-CNN is secure in the defined threat model. Extensive numerical analysis as well as prototype implementation over the well-known CNN architectures and datasets demonstrate the practical performance of LEP-CNN. Our experimental results also depict that LEP-CNN prevails in terms of accuracy and efficiency under time-sensitive scenarios compared with homomorphic encryption based offloading solutions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the relation between stellar specific angular momentum $j_*$, stellar mass $M_*$, and bulge-to-total light ratio $\beta$ for THINGS, CALIFA and Romanowsky & Fall datasets, exploring the existence of a fundamental plane between these parameters as first suggested by Obreschkow & Glazebrook. Our best-fit $M_*-j_*$ relation yields a slope of $\alpha = 1.03 \pm 0.11$ [with a trivariate fit including $\beta$]{}. When ignoring the effect of $\beta$, the exponent $\alpha = 0.56 \pm 0.06$ is consistent with $\alpha = 2/3$ predicted for dark matter halos. There is a linear $\beta - j_*/M_*$ relation for $\beta \lesssim 0.4$, exhibiting a general trend of increasing $\beta$ with decreasing $j_*/M_*$. Galaxies with $\beta \gtrsim 0.4$ have higher $j_*$ than predicted by the relation. Pseudobulge galaxies have preferentially lower $\beta$ for a given $j_*/M_*$ than galaxies that contain classical bulges. Pseudobulge galaxies follow a well-defined track in $\beta - j_*/M_*$ space, consistent with Obreschkow & Glazebrook, while galaxies with classical bulges do not. These results are consistent with the hypothesis that while growth in either bulge type is linked to a decrease in $j_*/M_*$, the mechanisms that build pseudobulges seem to be less efficient at increasing bulge mass per decrease in specific angular momentum than those that build classical bulges.'
author:
- 'Sarah M. Sweet'
- David Fisher
- Karl Glazebrook
- Danail Obreschkow
- Claudia Lagos
- Liang Wang
bibliography:
- 'references.bib'
title: 'Revisiting the stellar mass – angular momentum – morphology relation: extension to higher bulge fraction, and the effect of bulge type.'
---
Introduction
============
Galaxy stellar mass $M_*$ and angular momentum $J$ are fundamental properties of galaxies: they have been shown to correlate strongly with galaxy size and density [@Mo+1998], disk thickness and colour [@Hernandez+2006], and morphology [@RF12; @OG14; @Cortese+2016 hereafter RF12, OG14, [C16]{}].
$M$ and $J$ are not independent in that $J$ is scaled by mass, so the standard method for studying their relationship is to remove the mass dependence to obtain specific angular momentum $j = J/M$. Specific angular momentum of baryons in the galaxy $j_{baryons}$ is empirically similar to that of the dark matter halo $j_h$ [@Fall1983]. This similarity is expected for baryons in the halo, since the same tidal forces are experienced during spin-up [@Catelan+1996a; @Catelan+1996b; @vandenBosch+2001; @Barnes+1987], but has been historically difficult to reconcile for baryons in the disk, *viz.* the “angular momentum catastrophe” [@Governato+2010; @Agertz+2011]. $j_{baryons}$ is typically further resolved into analogous specific angular momenta for stars, H$\alpha$, HI and H$_2$ ($j_*$, $j_{H\alpha}$, $j_{HI}$, $j_{H_2}$ respectively) depending on the observed kinematics and mass profiles[^1] available to be studied. In this work we focus on stellar specific angular momentum $j_* = J_*/M_*$.
The observational $M_*-j_*$ plane was first studied by @Fall1983, who found $j_* \propto qM_*^\alpha$, with [parallel tracks defined by late-type and early-type galaxies. These tracks have exponent $\alpha = 2/3$, in agreement with the natural scaling of CDM halos in a hierarchical universe. @RF12 [RF12] later studied the relation between $M_*$, $j_*$ and $\beta$, confirming the earlier result of $j \propto M_*^{2/3}$, with factor $q$ depending on whether disks or bulges are considered.]{} @OG14 [OG14] used high-quality observations of 16 galaxies from The HI Nearby Galaxy Survey [THINGS, @Leroy+2008; @Walter+2008] to further investigate the $M_*-j_*-\beta$ relation, finding that $\alpha = 2/3$ for $0 \leqslant \beta \leqslant 0.32$, but that $\alpha \sim 1$ [when $\beta$ is treated as a free parameter.]{} [More recently, C16 analysed a subset of galaxies from the Sydney-AAO Multi-object Integral field [SAMI, @Croom+2012] Galaxy Survey [@Bryant+2015; @Allen+2015; @Sharp+2015], and similarly found, when considering $j$ at one effective radius $r_e$, that $\alpha$ is consistent with 2/3 for the whole range of morphologies, but higher when a single morphology class is considered, and approaching $\alpha = 1$ for late-type galaxies.]{}
There is a known dichotomy in the properties of pseudo- vs. classical bulges [e.g. @kormendy2004; @fisher2016]; classical bulges are pressure-supported components thought to form [by minor mergers [@Toomre1977; @Schweizer1990] or disk instabilities [@Toomre1964]]{}, while pseudobulges are rotationally-supported components formed during secular evolution of the disk, so naturally have higher $j_*$ [@kormendy2004; @Wyse+1997]. Classical bulges generally contribute a larger $\beta$ than pseudobulges [@fisher2016]. These properties are intimately related to galaxy angular momentum and morphology. However, previous studies of the $M_*-j_*-\beta$ relation have not analysed [galaxies that contain]{} classical bulges separately from [those that contain]{} pseudobulges, so the effect of bulge type is unknown.
In this work we investigate the effect of bulge type on the relationship between stellar mass, specific angular momentum and morphology across a large range of $\beta$, by combining OG14 with high-quality subsets of the sample in RF12 and the Calar Alto Legacy Integral Field Area Survey [CALIFA, @Sanchez+2012; @Husemann+2013; @Walcher+2014; @Sanchez+2016]. In Section \[sec:methods\] we describe our methods for measuring bulge properties and $j_*$, and introduce the datasets. In Section \[sec:results\] we present the $M_* - j_* - \beta$ relation as it relates to bulge type, given the known dichotomy in the properties of pseudo- vs. classical bulges [and the galaxies that host them]{} [e.g. @kormendy2004; @fisher2016]. Section \[sec:conclusion\] concludes this letter with a discussion of the significance of these results.
Sample & Methods\[sec:methods\]
===============================
We combine observations from THINGS, RF12 and CALIFA datasets to trace the fundamental relation between $M_*$, $j_*$ and $\beta$ over a wide range of $\beta$. The three samples are complementary. We have high-quality $j_*$, $\beta$ and bulge classifications for THINGS, but the sample is limited to low to moderate $\beta$, with few [galaxies that contain]{} classical bulges. We thus employ the RF12 galaxies for which we have high-quality $\beta$ and bulge classification to extend our sample to higher $\beta$ and increase the number of [galaxies with]{} classical bulges. Similarly, we also include a subset of the CALIFA sample for which we measure high-quality $j_*$, and a greater range of $\beta$. Below we present our methods for determining bulge properties and $j_*$, before giving specific details for each of our samples.
Bulge-to-total mass ratio and type
----------------------------------
We obtain bulge properties by cross-correlating the OG14 and RF12 samples with the combined data set of @fisher2010 [@fisher2011], @fabricius2012 and @fisher2013. These samples use the same method, software and wavelength range to conduct 2D bulge-disk decompositions [described in @fisher2008 FD08]. The method combines high-resolution HST imaging with wide-field ground based imaging to reduce uncertainties and degeneracies inherent to bulge-disk decompositions. It also accounts for the different mass-to-light ratio of the bulge and disk. Conversely, OG14 simply measured the bulge as the excess light over fitted exponential disk, while RF12 fitted two elliptical isophotes in projection [@Kent1986]. Figure \[fig:betabeta\] compares FD08 $\beta$ with OG14 and RF12, illustrating that OG14 and RF12 present $\beta$ that are mutually inconsistent and systematically underestimated with respect to FD08. Importantly, our consistent method allows for an accurate comparison of bulge properties between OG14 and RF12 [, with an uncertainty of $\Delta \beta_{FD08} = \pm 0.05$.]{}
We use the well-known correlation between bulge Sérsic index $n_b$ [@Sersic1963] and bulge type [@fisher2008; @fisher2016] to classify [galaxies that contain]{} pseudobulges and classical bulges, such that $n_b<2$ implies a [galaxy with a]{} pseudobulge and $n_b>2$ implies a [galaxy with a]{} classical bulge. The exception is NGC3593 in RF12, with a low $n_b\approx1.2$, which would ordinarily imply [that it contains]{} a pseudobulge. However, its bulge is “not classifiable” [@fabricius2012], for the following reasons: 1) NGC3593 is an extreme example of counter-rotating kinematics [@fabricius2012; @bertola1996], suggesting a recent merger [e.g. @bassett2017], and the empirical methods of bulge classification fail for most galaxies that are experiencing interactions [@fisher2016]; 2) the surface photometry cannot be reliably fit due to chaotic dust profile [@ravindranath2001]; 3) the galaxy has a “peculiar” global morphology [e.g. @sandage1994].
Specific angular momentum
-------------------------
We adopt the same method as @Obreschkow+2015 for calculating $j_*$ from integral field spectroscopic observations (IFS). Compared with slit spectroscopy, this method significantly increases accuracy in tracing the kinematic field, since kinematic and photometric major axes may be misaligned [e.g. @Sweet+2016], and the velocity fields of many galaxies are not well described by simple 1D rotation curves.
[We use a combination of the observed deprojected angular momentum where available, together with [a model-informed estimate of the]{} deprojected angular momentum in the spaxels where observations are not of sufficient quality, e.g. due to low signal-to-noise ratios [**$< 3$**]{} in either the stellar surface density maps or the kinematic maps. ]{}
1\) The observed [deprojected]{} angular momentum $J_i = {\bf r}_i {\bf v}_i m_i$ in every spaxel $i$ at deprojected radius ${\bf r}$ whose circular velocity ${\bf v}$ is derived from kinematic maps and mass $m$ from stellar surface density maps. [The deprojection is based on inclination and position angle derived from a fit to the stellar surface density maps; inclination and position angle are assumed to be constant with radius. Non-circular motions are neglected in this work, but see Sweet et al. (in prep.) for a treatment of the contribution of non-circular motions to total and spatially-resolved $j_*$.]{}
2\) The model ${J}_i$ at each spaxel is computed by fitting an exponential profile to the disk, in order to reach the total angular momentum, traced by the flat part of the rotation curve. [The surface mass density is characterised by $\Sigma({\bf r}_i) \propto \rm{exp}(-{\bf x}_i)$, where ${\bf x}_i = {\bf r}_i/r_{flat}$ and [the exponential scale length]{} $r_{flat}$ is [**assumed to be**]{} the radius at which the velocity reaches the converged velocity $v_{flat}$. [While not the case in general, this simplifying assumption is made in order to keep the number of free parameters at a minimum.]{}]{} [The exponential fit is given by equation 7 of OG14; ${\bf \tilde{v}}_i \approx v_{flat} \left(1 - \rm{exp}(-{\bf x}_i)\right)$. Following OG14 equation 8, the model ${J}_i = 2((1+{\bf x}_i)^3-1)/(1+{\bf x}_i)^3 {\bf r}_i v_{flat} m_i$.]{} [The model is on average consistent with the observed $J$ to the 5% level, if both are summed over the same high signal-to-noise spaxels. However, note that the purpose of the model is only to serve as an estimate of $J_i$ in the low signal-to-noise spaxels. ]{}
3\) The total $j_*$ is then given by $J/M_*$, where $J = \lvert\sum^i {J}_i\rvert$ is the norm of the sum over the observed $J_i$ where defined, and [estimated]{} $J_i$ in other spaxels, [integrated to ${\bf r}_i=\infty$]{}. [Including the estimated $J_i$ in the spaxels where data is missing comprises an average of 20% of the total $j_*$. ]{}
[The uncertainty in this method is typically $\Delta j_* / j_* = \pm$ 5-10%, predominantly contributed by [the uncertainty in extrapolating $j_*(r)$ beyond where it is converged]{} (that is, [the assumption]{} that the observations reach the flat part of the rotation curve), as well as the assumption of pure circular motion, and the uncertainty in inclination (see OG14 for further details).]{}
THINGS
------
THINGS [@Walter+2008] is a survey of 34 nearby galaxies observed to high multiples of effective radius $r_e$. OG14 presented $j_*$ measured with HI kinematics for the sixteen spiral THINGS galaxies that have stellar surface density maps published by @Leroy+2008. [Stellar masses and uncertainty $\Delta M_*/M_* = \pm 0.11$ dex are also taken from OG14.]{} This sample contains 13 [galaxies with]{} pseudobulges and 3 [with]{} classical bulges, with bulge-to-total mass ratios $0 \leqslant \beta \leqslant 0.41$.
Romanowsky & Fall
-----------------
RF12 presented $j_*$ for a sample of spiral and elliptical galaxies, calculated using stellar kinematics from slit spectroscopy of starlight and ionized gas. OG14 found these to systematically [vary]{} with respect to their own IFS observations, so we rescale RF12 $j_*$ [using equation 6 of OG14: $$\label{eqn:OG14-6}
\left(\frac{j_*}{10^3\ {\rm kpc\ km\ s^{-1}}}\right) \approx
1.01\left(\frac{\widetilde{j_*}}{10^3\ {\rm kpc\ km\ s^{-1}}}\right)^{1.3}$$]{} [The relative uncertainty is $\Delta j_* / j_* = \pm 32\%$, given by the quadrature sum of the uncertainty in RF12 $j_*$ (10%) and the RMS scatter of the calibrating relation (30%).]{}
[RF12 derived $M_*$ using @Bell+2003 colours and a diet Salpeter IMF located between @Kroupa2001 and @Salpeter1955, which translates to an assumed $K-$band mass-to light ratio $M/L{_K} = 1M_\odot/L_{\odot,K}$. This differs to @Leroy+2008, who assumed a @Kroupa2001 IMF and consequently $M/L{_K} = 0.5M_\odot/L_{\odot,K}$. We therefore scale the RF12 $M_*$ by 0.5 to achieve consistency with our THINGS sample. The uncertainty in RF12 $M_*$ of $\Delta M_*/M_* = \pm 0.2$ dex is taken from OG14.]{}
Motivated by the desire to calculate $\beta$ and classify bulge type in the same manner as for our THINGS sample, we select the 25 galaxies from RF12 for which we have existing high-quality bulge-disk decompositions. The RF12 sample thus contains 12 [galaxies that contain]{} pseudobulges, 12 [that contain]{} classical bulges and one [whose bulge is]{} unclassifiable (NGC3593), with $0 \leqslant \beta \leqslant 0.53$.
CALIFA
------
The CALIFA survey has made available stellar kinematic and surface density maps for 300 nearby galaxies [@Falcon-Barroso+2017]. We use the OG14 method described above to measure $j_*$ for these galaxies, and note that $j_*(<r)$ converges to $>0.99j_*$ at a radius $r \sim 3r_e$. This motivates us to select the subset observed to at least that radius.
We take stellar masses from @Falcon-Barroso+2017[, who used the methods outlined in @Walcher+2014; namely, assuming @Bruzual+2003 stellar populations and a @Chabrier2003 IMF. The RMS scatter between their two implementations gives the uncertainty $\Delta M_*/M_* = \pm 0.15$ dex. The @Bruzual+2003 stellar populations were shown by @Sanchez+2013 to give stellar masses consistent with those derived from @Bell+2001 colours (which themselves are consistent with @Bell+2003, as used by @Leroy+2008). $K$-band $M/L{_K}$ ratios based on the @Chabrier2003 IMF differ from those assuming a @Kroupa2001 IMF [as in @Leroy+2008] by only 10% [@Longhetti+2009 table 2], well within the scatter. Hence, we are comfortable that the CALIFA stellar masses are comparable with those of our THINGS sample.]{}
Hubble types are taken from @Falcon-Barroso+2017 and $\beta$ from the bulge-disk decompositions presented in @Mendez-Abreu+2017. @Mendez-Abreu+2017 includes only one FD08 galaxy; $\beta_{FD08} = 0.5$ *cf.* $\beta_{CALIFA} = 0.6$. [Since the @Mendez-Abreu+2017 decompositions were based on SDSS imaging, we estimate the uncertainty as $\Delta \beta_{CALIFA} = \pm 0.1$]{}. We remove five pure elliptical galaxies, since $j_*$ as measured here is strictly applicable to systems that contain a disk. There is a lack of imaging of sufficient resolution to reliably recover the bulge Sérsic index, so we do not categorise CALIFA [galaxies]{} into [those with]{} classical [or]{} pseudobulges. Our high-quality CALIFA subset comprises 35 spiral and 15 elliptical/lenticular galaxies, and spans $0 \leqslant \beta \leqslant 0.73$.\
The properties of the resulting samples are given in Table \[tab:data\].
--------------- -------- ------ -------------------- ------------------ --------- ---------------- --------- ------------ ----------------- --------------------- --------------------- ------------- -- --
Name Survey Type $M_*$ $\Delta M_*/M_*$ $\beta$ $\Delta \beta$ $r_d$ $r_{flat}$ $v_{flat}$ $j_*$ $\Delta j_*$ $n_{bulge}$
\[log($M_\odot$)\] \[dex\] \[kpc\] \[kpc\] \[km s$^{-1}$\] \[kpc km s$^{-1}$\] \[kpc km s$^{-1}$\]
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)
NGC0628 THINGS Sc 10.10 0.11 0.17 0.05 2.3 0.8 217 955 95 1.53
NGC0925 THINGS SBcd 9.90 0.11 0.06 0.05 4.1 6.5 136 871 87 0.90
NGC2403 THINGS SBc 9.70 0.11 0.06 0.05 1.6 1.7 134 417 42 0.80
... ... ... ... ... ... ... ... ... ... ... ... ...
NGC0224 RF12 Sb 10.76 0.20 0.48 0.05 5.9 234 2967 938 2.13
NGC0247 RF12 Sd 9.54 0.20 0.00 0.05 4.1 92 749 237 0.00
NGC0300 RF12 Sd 8.93 0.20 0.00 0.05 1.6 60 173 55 1.64
... ... ... ... ... ... ... ... ... ... ... ... ...
IC1151 CALIFA Scd 9.85 0.15 0.02 0.10 1.9 4.0 113 1122 154
MCG-02-02-030 CALIFA Sb 10.37 0.15 0.08 0.10 3.1 8.8 177 2698 340
NGC0001 CALIFA Sbc 10.80 0.15 0.46 0.10 1.8 3.4 169 1564 117
... ... ... ... ... ... ... ... ... ... ... ... ...
--------------- -------- ------ -------------------- ------------------ --------- ---------------- --------- ------------ ----------------- --------------------- --------------------- ------------- -- --
\
The relation between stellar mass, specific angular momentum and morphology\[sec:results\]
==========================================================================================
We present our $M_*-j_*-\beta$ relation for THINGS, RF12 and CALIFA, fitting the data with the log-linear three-parameter model given in eq. (9) of OG14, $$\label{eq:master}
\beta = p_1\log_{10}M_*+p_2\log_{10}j_*+p_3,$$ where $\beta$ is the bulge-to-total stellar mass ratio, $M_*$ is the stellar mass in units of $10^{10}M_\odot$ and $j_*$ is the stellar specific angular momentum in units of $\rm 10^3 kpc~km~s^{-1}$. The maximum likelihood solution is easily computed using the [`hyper.fit`]{} algorithm of @Robotham+2015. This yields $p_1=0.39\pm0.04$, $p_2=-0.38\pm0.06$ and $p_3=0.06\pm0.02$ with an intrinsic scatter of standard deviation $\sigma=0.07\pm0.02$[^2]. The uncertainties are standard deviations, i.e. the square-roots of the diagonal elements of the covariance matrix, approximated as the negative inverse Hessian matrix of the likelihood at its maximum (Laplace approximation). The parameters are in agreement with OG14, while the scatter is increased due to revised $\beta$.
In Figure \[fig:m\_j\] we show the $M_*-j_*$ projection, overlaid with lines of constant $\beta$. These best-fitting lines take the form of eq. (10) of OG14, $$\label{eqn:OG14-10}
\frac{j_*}{10^3\ {\rm kpc\ km\ s^{-1}}} =
k e^{(-g\beta)} \left(\frac{M_*}{10^{10}M_\odot}\right)^\alpha,$$ where coefficients $k$ and $g$ together are instructed by baryonic physics, $g$ modifies the bulge-dependent scale, and the exponent $\alpha$ is predicted by CDM to be $\alpha = 2/3$ for DM halos. [Equation \[eqn:OG14-10\] is simply obtained by potentiating Equation \[eq:master\]. Since our observational uncertainties tend to be normal in log($M_*$), log($j_*$) and $\beta$ (with the exception that $0\leq \beta \leq 1$), rather than in $M_*$, $j_*$ and $exp(\beta)$, it is sensible to fit the parameters ${p} = \{p_1,p_2,p_3\}$ of Equation \[eq:master\] and propagate them to the parameters ${q} = \{k=10^{(-p_3/p_2)},g=-ln(10)/p_2,\alpha=-p_1/p_2\}$ of Equation \[eqn:OG14-10\]. The covariance matrix $C_{q}$ of the new parameters ${q}$ can then be estimated by linearly propagating the covariance matrix $C_{p}$ of the parameters ${p}$, i.e. $C_{q} = JC_{p} J^\dag$, where the Jacobian $J$ is defined as $J_{\rm ij}=\partial{q_{\rm i}}/\partial{p_{\rm j}}$.]{} [This method is [consistent with]{} fitting Equation \[eqn:OG14-10\] assuming log-normal uncertainties in $M_*$ and $j_*$ and normal in $\beta$. Note that the orthogonal scatter is [not fit here but is ]{}[also]{} minimized [see @Robotham+2015 for details]. ]{}
We give the resulting coefficients [and their uncertainties]{} in Table \[tab:fits1\]. We find $\alpha = 1.03 \pm 0.11$, consistent with OG14. [The importance of correctly accounting for measurement uncertainties in all variables as well as intrinsic scatter is demonstrated by the fact that a simple, linear least-squares fit to Equation \[eqn:OG14-10\] that only accounts for measurement uncertainties in $j_*$ and not intrinsic scatter yields a significantly different exponent $\alpha = 0.66 \pm 0.06$. The]{} bulge dependence $g$ and prefactor $k$ [from our fit to Equation \[eq:master\] are also]{} consistent [with OG14]{} at the 3-$\sigma$ level. The fits for THINGS or RF12 alone are consistent with the main sample; CALIFA is significantly different, but this is likely due to their angular size selection function, which results in a lack of low-mass, low-$j_*$ galaxies. Galaxies certain to contain a pseudobulge (or no bulge) are consistent with the main sample. Conversely, the galaxies in our sample that contain classical bulges [do not have sufficient dynamic range in $M_*$ to measure the presence or absence of a relation.]{}
The corresponding two-dimensional fit to $j_* \propto M_*^\alpha$ (that is, ignoring the effect of $\beta$) [can be found in a similar manner by fitting ${\rm log}j_* = \alpha {\rm log}(M_*) + a$; this ]{}gives $\alpha = 0.56 \pm 0.06$, consistent with the CDM prediction for halos of $\alpha = 2/3$.
[lLLL]{} This work:\
All & 1.47 0.24 & 6.13 0.95 & 1.03 0.11\
THINGS & 1.38 0.52 & 8.65 3.07 & 1.27 0.27\
RF12 & 1.75 0.82 & 6.60 2.31 & 1.07 0.21\
CALIFA & 1.76 0.25 & 4.25 0.79& 0.63 0.12\
Pseudobulge ($n_b < 2$) & 1.21 0.30 & 6.94 1.82 & 0.86 0.12\
OG14:\
fixed $\alpha$ & 0.91 0.09 & 7.59 0.79 & 1.00\
free $\alpha$ & 0.89 0.11 & 7.03 1.35 & 0.94 0.07\
& &\
This work:\
All & -0.35 0.04 & 0.08 0.02\
Pseudobulge, all & -0.26 0.06 & 0.05 0.03\
Pseudobulge, THINGS & -0.23 0.06 & 0.04 0.02\
OG14 & -0.30 0.03 & -0.01 0.01\
In Figure \[fig:bt\_jm\] we fix $\alpha=1$ and show the $\beta-j_*/M_*$ relation for pseudobulges and for all bulge types. The best-fitting lines take the form of equation 11 of OG14: $$\label{eqn:OG14-11}
\beta = k_1{\rm log}\left(\frac{j_*\ M_*^{-1}}{10^{-7}\ {\rm kpc\ km\ s^{-1}}\ M_\odot^{-1}}\right)\ +\ k_2,$$ with coefficients $k_1$ and $k_2$ given in Table \[tab:fits1\]. [This is obtained by imposing $p_2 = -p_1$ in Equation \[eq:master\], and propagating parameters ${p} = \{p_1,-p_1,p_3\}$ to ${q} = \{k_1=-p_1,k_2=p_3,\alpha=-p_1/p_2=1\}$.]{} The fit to all bulge types has slope $k_1 = -0.35 \pm 0.04$. Galaxies with high $\beta \gtrsim 0.4$ (where most are dominated by a classical bulge) all lie above this relation, indicating a high [$\beta$ for their $j_*$ and stellar mass]{}. The [sample of galaxies that host pseudobulges]{} follows a [shallower]{} relation to the fit to [galaxies with]{} all bulge types, with $k_1 = -0.26 \pm 0.06$, exhibiting a lower $\beta$ for a given $j_*/M_*$ than those that contain classical bulges. [Galaxies that contain classical bulges have a small range of $M_*$, so we cannot determine whether or not there is a corresponding relation for that sample. ]{} Our sample of THINGS [galaxies with]{} pseudobulges is marginally consistent with [and follows a shallower relation than]{} OG14 (which is predominantly [comprised of galaxies that host]{} pseudobulges); the [main]{} difference is a direct result of our revised $\beta$.
Discussion & conclusion\[sec:conclusion\]
=========================================
In Section \[sec:results\] we presented the relation between specific angular momentum, stellar mass and bulge-to-total mass ratio. [Galaxies with]{} pseudobulges have lower $\beta$ per $j_*/M_*$ than [those with classical bulges]{}, and exhibit a well-defined track in $\beta-j_*/M_*$ space.
We investigate over what range in $\beta$ the $\beta-j_*/M_*$ relation applies. Figure \[fig:bt\_jm\] illustrates that the relation for all galaxies extends to $\beta$ $\sim 0.4$, albeit with some scatter, partly arising from the difficulty in measuring $\beta$. As discussed in Section \[sec:methods\], we employ bulge-disk decompositions for the three datasets presented here. The methods for RF12 and OG14 are identical, but the CALIFA decompositions may subtly differ due to the different imaging and software used. There may be also systematic differences between [samples owing]{} to the type of kinematics used in this study: [stellar for CALIFA and RF12, and HI for THINGS]{}. Firstly, we have assumed (as in OG14) that the HI and stars co-rotate, but asymmetric drift may be appreciable, contributing to [$j_{gas}$ being *higher* than $j_{*}$]{} by 0.1 dex [@Cortese+2016]. Those $j_{gas}$ were measured from H$\alpha$ kinematics, not HI, so we cannot apply the magnitude of that correction, but the direction of the effect will likewise be in the opposite direction to that required. Secondly, stellar kinematics trace both disk and bulge, but HI kinematics trace only the disk. [Inclusion of the bulge component in the THINGS measurements would serve to increase $j_*$ towards the CALIFA and RF12 values]{}, but only at an average of 0.3% and maximum 1.3% (OG14). Neither effect [explains why]{} many of the CALIFA and RF12 galaxies [with $\beta \gtrsim 0.4$ ]{} lie significantly above the relation defined by all three samples. The apparent upturn may seem to suggest that the true relation [takes a different functional form than that assumed here.]{}
In general, mechanisms that increase bulge mass appear to decrease the ratio of $j_*/M_*$. The relation first presented in OG14 for $0 \leqslant \beta \leqslant 0.32$ is confirmed here for $0 \leqslant \beta \lesssim 0.4$, but breaks down for $\beta \gtrsim 0.4$. [At fixed $j_*/M_*$, galaxies hosting a classical bulge exhibit a range of $\beta$, extending upwards from the relation defined by galaxies that host a pseudobulge. A large $\beta$ for classical galaxies is well explained by noting from Figure \[fig:m\_j\] that galaxies that host classical bulges are generally more massive than those that contain pseudobulges [and see @fisher2016], and that more massive galaxies typically have larger $\beta$ [@Koda+2009]. ]{} The same general trend is also seen in the Evolution and Assembly of GaLaxies and their Environments simulations [EAGLE, @Schaye+2015; @Lagos+2017], [where the most bulge-dominated galaxies lie above the best-fitting line. This is interpreted as those galaxies having higher $j_*$ than predicted by the relation, which points to an absence of gas-poor mergers [@Lagos+2017].]{}
Interestingly, $ j_* \propto M_*^{\alpha}$ is well fit by $\alpha = 1$ for fixed $\beta$ over all bulge types, in line with the finding of OG14 for their sample of predominantly [galaxies with]{} small pseudobulges. [OG14 outlined a theoretical argument for a physical motivation to an exponent $\alpha = 1$ for a given morphology, whereby]{} $j_*/M_*$ traces inverse surface density, which is inversely related to the @Toomre1964 Q parameter. Q quantifies instability against rotation, required for pseudobulge formation, so decreases while $\beta$ increases. [Testing this interpretation is outside of the scope of this paper but would make interesting future work.]{}
When ignoring the $\beta$ dimension we find $\alpha \sim 2/3$, as seen by @Fall1983, OG14 and @Cortese+2016. [This confirmation is notwithstanding several important distinctions, namely: a different bulge decomposition method to @Fall1983, OG14 and @Cortese+2016, the non-linear correction applied to the @Fall1983 $j_*$ to approximate our IFU data, the use of data out to 3$r_e$ *cf.* 1$r_e$ in @Cortese+2016, and the extended range of morphology with respect to OG14.]{} This exponent is consistent with the CDM prediction for halos $ j_{h} \propto M_{h}^{2/3}$. Connecting that prediction with our observed relation for stars implies that $M_*$ and $j_*$ depend respectively on $M_{h}$ and $j_{h}$ with the same functional form. The $M_*-M_{h}$ relation is shown to be complex [@Guo+2010]; future large IFS surveys [e.g. Hector, @Lawrence+2012] are required to test whether the $j_*-j_h$ relation takes a similar form.
We investigate the possibility of two $\beta-j_*/M_*$ tracks: one for galaxies with classical bulges, thought to form by merging; and the other for galaxies with pseudobulges, formed by secular evolution [@kormendy2004; @Wyse+1997]. Secular evolution refers to [angular momentum transport causing some disk material to fall towards the galaxy centre, contributing to the pseudobulge with a small increase in $\beta$. The same process feeds star formation in the pseudobulge, which causes a small amount of $j_*$ to be lost in outflows due to stellar winds. There is a corresponding small]{} change in $M_*$, so the galaxy moves along a well-defined track in $\beta-j_*/M_*$. This is consistent with the distinct relation we find for pseudobulges. The lower $\beta$ implies that the processes that form pseudobulges are less efficient at rearranging $j_*$ and $M_*$ while forming bulges than those responsible for classical bulges.
Conversely, mergers can significantly increase both $M_*$ and $j_*$, though some $j_*$ will cancel due to misalignment of the galaxy spin axes [@RF12; @Lagos+2017]. There is a correspondingly large increase in $\beta$, so mergers move a galaxy *above* the pseudobulge $\beta-j_*/M_*$ relation, while forming a classical bulge. [We cannot include EAGLE as a control sample on Figure \[fig:bt\_jm\], since the EAGLE $\beta$ are measured from kinematic bulge-disk decompositions instead of photometric methods. As a result, they are expected to be on average larger than our $\beta$ by $\sim 0.5$ [@Obreja+2016]. In addition to this systematic offset, there is considerable scatter, so that one cannot apply a simple correction factor. In future work (Lagos et al., in prep.) we will present ‘photometric’ bulge-disk decompositions of the synthetic images of galaxies to obtain $\beta$ measurements that are directly comparable to observations. ]{} While EAGLE $\beta$ are not directly comparable with our $\beta$, we can use those simulations to make quantitative predictions of movement along the $j_*/M_*$ axis [, since mergers, which build classical bulges, appear to do so while moving the galaxy above the pseudobulge relation]{}. We see in Figure \[fig:bt\_jm\] that [galaxies that host]{} classical bulges lie on or above the pseudobulge relation, so we use that relation to calculate a lower limit to $\Delta\beta$ for mergers that build classical bulges. @Lagos+2017 predict that a typical dry (wet) minor merger[^3] that increases $M_*$ by $\Delta M_* = 0.15$ dex will decrease (increase) $j_*$ by $\Delta j_* = - 0.15 (0.04)$ dex. Combining the EAGLE predictions with our observed [relation for galaxies that contain]{} pseudobulges, we expect that minor mergers of $\Delta M_* = 0.15$ dex will increase $\beta$ by [more than]{} $\Delta\beta \geqslant 0.08 (0.03)$. Assuming that a bulgeless progenitor of a [galaxy with a]{} classical bulge begins at ${\rm log} (j_*/M_*) \sim -6.75$, then several mergers of this magnitude would be required to achieve $-7.7 \lesssim {\rm log} (j_*/M_*) \lesssim -7.2$ as we observe. Alternatively, a galaxy that already hosts a pseudobulge and lies in that $j_*/M_*$ range may only need to experience one such merger to form a classical bulge with those properties. However, it is not known where the progenitors of classical bulge galaxies lie in $\beta-j_*/M_*$ space. In reality there is a range of possible merger ratios, which increases the range of expected $\Delta \beta$, and may serve to explain the observed dispersion in $\beta-j_*/M_*$. The apparent failure of the relation for high $\beta$ may then simply reflect the difference between the pseudobulge and classical bulge regimes. We note that other physical processes such as outflows also modify $j_*$; these will be discussed further in Sweet et al. (in prep), when we present the internal distribution of $j_*$.
The dependence on bulge type that we see in the $\beta-j_*/M_*$ relation is reminiscent of the black hole mass – galaxy velocity dispersion $M_{BH} - \sigma$ relation [@Ferrarese+2000; @Gebhardt+2000], where [galaxies with]{} classical bulges follow a well-defined relation in $M_{BH} - \sigma$ space but [those that contain]{} pseudobulges show no such relation, as suggested by @Kormendy+2001 and shown by @Hu2008 and @Saglia+2016. In this letter we see the opposite effect for $\beta-j_*/M_*$, where it is instead [galaxies with]{} pseudobulges that show the well-defined relation, and [those with]{} classical bulges that do not. This [is consistent with]{} earlier suggestions that classical bulges are sensitive to black hole evolution, while the evolution of pseudobulges is linked to the disk, which dominates the $j_*$ budget [@Kormendy+Ho2013].
In conclusion, we have presented high-quality integrated specific angular momenta for a subset of CALIFA galaxies, and revisited the stellar mass – specific angular momentum – morphology relation for $0 \leqslant \beta \leqslant 0.73$, using galaxies from THINGS, RF12 and CALIFA. We confirm the OG14 $\beta-j_*/M_*$ relation for [galaxies that host]{} pseudobulges, albeit with increased scatter. The relation does not describe [galaxies with classical bulges]{}, in line with separate evolutionary channels for the formation of the two major bulge types. Future work will employ a large, homogeneous sample with high-quality measurements of $\beta$ and $j_*$ to mitigate selection biases and confirm the strength of this relation. The next critical stage is to understand the place of progenitors of galaxies that contain classical bulges, with a detailed study of specific angular momentum in main-sequence high-redshift galaxies.
We thank the anonymous referee for thoughtful comments, which helped to improve the paper.
This study uses data provided by the Calar Alto Legacy Integral Field Area (CALIFA) survey (http://califa.caha.es/). Based on observations collected at the Centro Astron[ó]{}mico Hispano Alem[á]{}n (CAHA) at Calar Alto, operated jointly by the Max-Planck-Institut f[ü]{}r Astronomie and the Instituto de Astrof[í]{}sica de Andaluc[í]{}a (CSIC).
Funding for the Sloan Digital Sky Survey IV has been provided by the Alfred P. Sloan Foundation, the U.S. Department of Energy Office of Science, and the Participating Institutions. SDSS-IV acknowledges support and resources from the Center for High-Performance Computing at the University of Utah. The SDSS web site is www.sdss.org.
SDSS-IV is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS Collaboration including the Brazilian Participation Group, the Carnegie Institution for Science, Carnegie Mellon University, the Chilean Participation Group, the French Participation Group, Harvard-Smithsonian Center for Astrophysics, Instituto de Astrofísica de Canarias, The Johns Hopkins University, Kavli Institute for the Physics and Mathematics of the Universe (IPMU) / University of Tokyo, Lawrence Berkeley National Laboratory, Leibniz Institut für Astrophysik Potsdam (AIP), Max-Planck-Institut für Astronomie (MPIA Heidelberg), Max-Planck-Institut für Astrophysik (MPA Garching), Max-Planck-Institut für Extraterrestrische Physik (MPE), National Astronomical Observatories of China, New Mexico State University, New York University, University of Notre Dame, Observatário Nacional / MCTI, The Ohio State University, Pennsylvania State University, Shanghai Astronomical Observatory, United Kingdom Participation Group, Universidad Nacional Autónoma de México, University of Arizona, University of Colorado Boulder, University of Oxford, University of Portsmouth, University of Utah, University of Virginia, University of Washington, University of Wisconsin, Vanderbilt University, and Yale University.
[^1]: Surface density is typically used as a proxy; [$j$ does not have the mass scaling of $J$]{}, but the mass profile is used as a weighting factor.
[^2]: The intrinsic Gaussian scatter of Equation \[eq:master\] is defined along the $\beta$-axis, but can be propagated to another set of axes appropriate to the chosen independent variable, with [consistent results]{}. [For example, writing Equation \[eq:master\] as $\log_{10}(j_*)=q_1\log_{10}(M*)+q_2\beta+q_3$ gives intrinsic scatter of $\sigma=0.20\pm0.05$ in log($j_*$), and parameters $q_1 = 1.03\pm0.11$, $q_2 = -2.66\pm0.41$, $q_3 = 0.17\pm0.07$, which propagate identically back to $p$.]{}
[^3]: mass ratio $<1:3$
| {
"pile_set_name": "ArXiv"
} |
---
title: 'FACT - How stable are the silicon photon detectors?'
---
Introduction
============
The First G-APD Cherenkov Telscope (FACT, [@bib:design]) is the first installation of a complete focal plane using silicon photo sensors. It is in operation since Oct. 2011 and its main goals are to prove the applicability of Geiger-mode avalanche photo diodes (G-APD) for focal planes with changing environmental conditions and the long-term monitoring of the brightest TeV blazars.
While G-APDs are very robust and easy in their application taken their relatively low bias voltage of usually less than 100V into account, their operation properties are also very sensitive to their temperature and the applied voltage.
Since active cooling and the achievement of a homogeneous temperature on a large surface exposed to the environment is difficult, the camera only features a passive cooling and thermal insulation to ensure that the waste heat of the electronics does not significantly heat the sensors.
By thermal design, the temperature gradient in the focal plane is small enough to allow powering of four and five sensors at the same time with the same bias voltage. They have been sorted accordingly. The bias voltage supply system comprises 320 channels in total and allows to adapt the voltage with a precision of about 22mV. Simultaneously, it enables the readout of the provided current with a precision of 1.2$\mu A$.
The feedback system
===================
To keep the gain, optical crosstalk probability and afterpulse probability of G-APDs stable, a constant overvoltage must be applied. The applied overvoltage is the voltage difference between the applied absolute voltage and the breakdown voltage, which is a function of temperature. The dependence of the breakdown voltage on temperature is linear, well defined and identical for all channels and can thus be corrected adapting the applied voltage according to the measured temperature.
To correct for the change in overvoltage, induced by the change in current and temperature, a feedback system is applied. As feedback values, 31 temperature sensors in the sensor compartment are available, as well as the current readout of each bias voltage channel.
The bias voltage is distributed to the G-APDs using a passive filter network. Its serial resistance induces a voltage drop depending on the current flowing. Since bright ambient light condition as moon lit nights can induce count rates up to 1000 times higher than during dark night conditions, every change in ambient light level would directly influence the applied overvoltage. The voltage drop at the G-APD can be calculated from the measured current and resistances.
For the correction of both effects, a feedback loop has been implemented in the slow control software, which calculates the correction offset according to the feedback values. Since temperature changes are assumed to be slow, the temperature is evaluated every 15s. Changes in current can be comparably fast due to bright stars moving over the field-of-view, therefore, the voltage drop is calculated once every second.
Measuring the stability
=======================
To measure the performance of the feedback system, the G-APD properties are evaluated. Different methods to determine the dependence on the applied bias supply voltage, the temperature and the ambient light condition, have been applied and are discussed hereafter.
Light pulser measurements
-------------------------
As an indirect measurement of the gain, the measured amplitude of a light pulse is used. The light pulses are emitted by a light pulser installed in the reflector dish. Since its light yield is temperature stabilized, the measured signal amplitude is an indirect measurement of the gain of all channels.
For each measurement, the light-pulser is flashed one thousand times and the readout is self-triggered by the camera’s trigger system. The signal is extracted either by searching the peak or integrating the pulse over a certain range around the peak, both done by a order spline interpolation. For each channel, the average amplitude is calculated and the median of all channel is taken, hereafter, just called [*light-pulser amplitude*]{}.
From random triggered events, a measure for the ambient light level can be derived calculating the fluctuations of the recorded signal. For this, the signal-extraction algorithms is applied in each event at a random position. The strength of the fluctuation corresponds to the background photon rate. To be able to correlate the measured values with temperature, an average of all temperature sensors is calculated during these five minutes.
The measured light-pulser amplitude is shown in Fig. \[fig:AmplVsPed\] versus average temperature with only the temperature-feedback switched on (blue) and with the whole feedback system enabled (black). With just the temperature correction, a clear dependence of the light-pulser amplitude on the ambient light level is visible, as expected. Applying the correction based on the measured current fully corrects this dependency.
![Distribution of light-pulser amplitudes measured at average compartment temperatures between 5C and 25C fitted with a Gaussian.[]{data-label="fig:TempDist"}](icrc2013-683-02){width="48.00000%"}
{width="48.00000%"} {width="48.00000%"}
{width="18.00000%"} {width="38.00000%"} {width="38.00000%"}
![An estimate of the dark-count rate obtained from the integration of the normalized sum-spectra versus average compartment temperature, extraction efficiency and dead-time are neglected. The correlation fits very well with the data-sheet.[]{data-label="fig:DarkCounts"}](icrc2013-683-08){width="31.00000%"}
In Fig. \[fig:TempDist\], the distribution of the light-pulser amplitude is shown. The sigma of the Gauss fit yields $\pm 6\%$ of the mean value.
Direct gain measurement
-----------------------
Due to the high precision of the charge released in breakdowns of single G-APD cells, they feature excellent resolution for single signals (single-[p.e.]{}). This allows the extraction of their properties from their dark count spectrum. For this, randomly triggered data is recorded with closed lid. The comparably low rate ensures that the probability for two coinciding dark count signals is reasonably low simplifying their identification and analysis.
#### Gain extraction
The quality of the gain measurement relies entirely on the quality of the single-[p.e.]{} spectrum, i.e. the recognition and extraction of pulses induced from a single primary breakdown. The gain is determined from the distance of higher order peaks to avoid a bias from a possible baseline shift, although the baseline for each channel is determined and corrected. For this study, the dark count spectrum of all 1440 channels has been investigated individually. For each measurement a run with 3000 randomly triggered events was taken.
To extract the pulses, a sliding average of ten samples is computed from the data. This improves the signal-to-noise ratio and eliminates some occasional noise with small amplitudes and a fixed frequency. To identify pulses, the remaining samples are scanned for a leading edge defined by a threshold-crossing and the constrain that four samples before and four samples after that point the signal must still be below and above this threshold, respectively. From this point, the local maximum is determined within the following 30 samples and the point at which the leading edge reaches 50% of the local maximum, called [*arrival time of the pulse*]{} hereafter. If the distance between the maximum and the arrival time exceeds 7ns the pulse is discarded. From the arrival time position onwards, 30 samples of the raw-signal are integrated. Since the arrival time is the point of the steepest slope, the integration starting at this point yields the most stable results, which was confirmed by simple cross-checks.
#### Fit function
The resulting distribution is a superposition of the distribution of the single-[p.e.]{} peak and higher order peaks resulting from optical crosstalk. Each distribution is believed to be Gaussian with a width $\sigma_{n}$ being compiled from a constant noise component $\sigma_{el}$, and a noise originating from the fluctuations of the amplitude of a single avalanche, $\sigma_{pe}$. The ratio between the number of events represented by two consecutive peaks is described by the probability $p$, the probability that an avalanche induces exactly one other avalanche in another micro-cell.
In the obtained spectra, a deviation from this strict exponential behavior is observed. Since this study does not aim at the precise determination of other parameters than the gain, this effect is not discussed in details.
#### Fit procedure
To determine reasonable start values for fitting the spectra of all channels individually, a combined spectrum of all pixels is filled and fitted for each run. With the baseline and gain value determined for each channel, the individual spectra are normalized and compiled into a sum-spectrum. Fig. \[fig:spectrum\] (right) shows the sum of these spectra of all available data and a fit with the function given above. The very good agreement of the individual spectra after normalization is evident.
#### Result
From the primary sum-spectrum, the average gain of all channels is determined. The dependence of the extracted gain versus the average sensor temperature is shown in Fig. \[fig:gain\] (left). A small remaining dependence of the gain on temperature is visible which is of the order of 5% per 20K. This is an effect of a non-ideal temperature adaption coefficient in the control software which has intentionally not been touched since the deployment of the camera to ensure consistency of the obtained data during a reasonably long period. The center plot shows the distribution of the measured gain values. The fitted Gaussian yields a width of the distribution of about $\pm 3\%$ which is well within the specification requiring a gain stability of $5\%$. The right distribution corresponds to the average root-mean square of the difference between the gain obtained from the single pixel fit and the average gain obtained from the fit of the combined spectrum of all pixels. It can be interpreted as the typical gain spread of pixels within the camera. The fitted Gaussian yields a mean of slightly more than 4%.
As a cross-check, the dark count rate can be estimated from the spectra and plotted versus temperature, see Fig. \[fig:DarkCounts\]. This measurement is not a precise measurement of the dark count rate because dead-time and efficiency effects are neglected. Nevertheless, the obtained dark count rate fits very well the data-sheet values.
Ratescans
---------
Another method to access the gain are so called ratescans. Ratscans use the cosmic ray spectrum for an indirect measurement. They measure the dependence of the total trigger rate of the system on the trigger threshold settings.
The trigger of the camera comprises two stages. In the first stage, the signals of nine channels are summed. This sum-signal is clipped to suppress effects of baseline fluctuations in case of high noise conditions. A comparator converts this signal into a a digital signal. Four of these signals are again summed and discriminated slightly below the threshold for a single input signal. This second step mainly reduces fake triggers from noise of the electronics itself by suppressing too short trigger signals.
Changing the comparator thresholds, the dependence of the total trigger rate on the threshold settings can be investigated. For low thresholds the rate is dominated by noise and photons from the diffuse night-sky background. For high thresholds, random triggers by artificial coincidences are suppressed, and only simultaneously arriving photons from cosmic-ray induced showers will trigger the system.
#### Results
For rates above 10MHz, the system saturates. Below, the trigger rate is dominated by random coincidences of night-sky background photons turning into triggers caused by coincident photons from cosmic-ray induced showers.
To prove the stability of the gain over a longer time-period and its independence of the ambient light condition, several ratescans have been taken between March and July 2012 with changing light conditions. Fig. \[fig:Ratescans0V\] shows 26 ratescans taken with light conditions ranging from dark night to almost full moon ($\approx 90\%$). It is visible that the shower induced part of the curves is independent of the ambient light conditions corresponding to a stable gain.
More on measured current, light condition, trigger threshold and rate can be found in [@bib:threshold; @bib:ratescans; @bib:moon].
Conclusion
==========
The camera of the FACT telescope is now operated since 20 months and has proven the applicability of silicon photo sensors in real-life focal plane installations, in particular, in Cherenkov telescopes.
To achieve a stable operation, of the applied Geiger-mode avalanche photo diodes, a feedback system adapting the applied voltage according to temperature readings and the measured current has been developed.
Three different methods to measure the gain of the system directly or indirectly were applied. They show consistent results on gain-stability. The presented results are limited by the calibration procedure of the bias voltage system, which is currently improved. The most precise and direct method, the extraction of the dark count spectrum, has shown a long-term stability over several months under changing temperature conditions at the few percent level. The measurement of the amplitude of an external light-pulser signal has proven the stability independent of the background light level also at the few percent level.
The implemented feedback system renders the need for an external calibration device obsolete, which is a big advantage for further Cherenkov telescope projects saving a lot of development time and costs.
A detailed description of the telescope and camera hardware and software can be found in [@bib:status; @bib:design]. A more detailed discussion of the stability will be available soon in [@bib:feedback].
T. Bretz et al. (FACT Collaboration), these proc., ID682.
H. Anderhub et al. (FACT Collaboration), 2013, JINST [**8**]{} P06008 \[arXiv:1304.1710\].
FACT Collaboration, [*in prep.*]{}
M. Koetig et al. (FACT Collaboration), these proc., ID695.
T. Bretz et al. (FACT Collaboration), these proc., ID720.
D. Hildebrand et al. (FACT Collaboration), these proc., ID709.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Aislynn Wallach
- 'Brett M. Morris'
- Doug Branton
- 'Teagan O’Reilly'
- Brittany Platt
- Ada Beale
- Andrew Yetter
- Katie Reil
- Kristen Garofali
- Eric Agol
title: 'Pre-MAP Search for Transiting Objects Orbiting White Dwarfs'
---
Introduction
============
Metal pollution in white dwarf (WD) atmospheres may be the accreted remnants of planetary objects [@Jura2009; @Gansicke2012; @Farihi2013; @Veras2015]. After the discovery of disintegrating planetary objects transiting WD 1145+017 [@Vanderburg2015], undergraduates in the University of Washington’s Pre-Major in Astronomy Program (Pre-MAP) were inspired to collect photometry of the brightest white dwarfs to hunt for similar transiting objects around other metal-polluted WDs. Prior surveys have yet to make a detection of a transiting planet orbiting a WD [@Faedi2011; @Fulton2014; @Maoz2015; @Belardi2016; @Sandhaus2016; @vanSluijs2018], yet WDs are still an attractive target for searches of small, rocky planetary material. Since a typical white dwarf is Earth-sized, transits of Earth-sized planets could have depths $>50\%$, so even low S/N photometry has a chance at discovering transiting material.
Rocky planetary material near the Roche limit of a white dwarf host will become tidally disrupted at orbital periods of only 3-10 hours [@Veras2014], enabling the discovery of such transiting objects and their periods in a single night of observations (for a fortuitous observer). The probability of transit for a Moon-sized object at these periods is of order $p_\mathrm{trans} = 0.01$ [@Agol2011].
Observations
============
We identified bright, northern, metal-polluted white dwarfs in the SDSS DR10 white dwarf catalog from [@Kepler2015] for photometric monitoring. We identified observable WDs from the Astrophysical Research Consortium Small Aperture Telescope (ARCSAT) 0.5-meter telescope at Apache Point Observatory with the observation planning package `astroplan` [@astroplan]. We observed each WD with typical exposure times of 60 seconds in the SDSS g filter with 3x3 binning to maximize observing efficiency. We present here five white dwarf light curves at 60 second cadence: G226-29, Wolf 1516, Wolf 28, SDSS J160401.31+083109.0, and known eclipsing white dwarf binary system SDSS J1152+0248 [@Hallakoun2016].
Results
=======
Figure \[fig:1\] shows photometry of four bright, metal-polluted WDs. Four of these light curves present no obvious periodic dimming events consistent with transiting planets or debris on short orbital periods. The fifth was a control target, SDSS J1152+0248, a known eclipsing white dwarf binary.
After visually inspecting the light curves for transient events, we search for periodic transit events using Lomb-Scargle and box-least squares (BLS) periodograms [@gatspy; @kovacs2002]. All WDs except the control eclipsing system exhibit negligible flux periodicity, as one might expect from the small transit probabilities predicted in @Agol2011. For SDSS J1152+0248, BLS detects the correct orbital period of 2.4 h, consistent with other observations [@Hallakoun2016].
We estimate the minimum potential planet radius observable under our conditions for each star by assuming that we could detect any transit with depth $\delta \approx (R_p/R_\star)^2 >5\times$ the rms photometric scatter (Figure \[fig:1\], right panel) with an assumed transit duration of one minute and $R_\star = 1.4R_\oplus$. The ARCSAT photometry has sufficient precision to detect Moon-sized objects or larger at short orbital periods, though no such planets were detected for these targets.
Though we have not detected any transiting debris, we look forward to surveys which may find planets orbiting white dwarfs, such as NASA’s TESS, ESA’s PLATO, and the Evryscope [@Ricker2014; @plato; @Law2015].
Based on observations obtained with the Apache Point Observatory 3.5 m and 0.5-meter telescopes, which are owned and operated by the Astrophysical Research Consortium.
natexlab\#1[\#1]{}\[1\][[\#1](#1)]{}
, E. 2011, , 731, L31
, C., [Kilic]{}, M., [Munn]{}, J. A., [et al.]{} 2016, , 462, 2506
, F., [West]{}, R. G., [Burleigh]{}, M. R., [Goad]{}, M. R., & [Hebb]{}, L. 2011, , 410, 899
, J., [G[ä]{}nsicke]{}, B. T., & [Koester]{}, D. 2013, Science, 342, 218
, B. J., [Tonry]{}, J. L., [Flewelling]{}, H., [et al.]{} 2014, , 796, 114
, B. T., [Koester]{}, D., [Farihi]{}, J., [et al.]{} 2012, , 424, 333
, N., [Maoz]{}, D., [Kilic]{}, M., [et al.]{} 2016, , 458, 845
, M., [Farihi]{}, J., & [Zuckerman]{}, B. 2009, , 137, 3191
, S. O., [Pelisoli]{}, I., [Koester]{}, D., [et al.]{} 2015, , 446, 4078
, G., [Zucker]{}, S., & [Mazeh]{}, T. 2002, , 391, 369
, N. M., [Fors]{}, O., [Ratzloff]{}, J., [et al.]{} 2015, , 127, 234
, D., [Mazeh]{}, T., & [McQuillan]{}, A. 2015, , 447, 1749
Morris, B. M., Tollerud, E., Sipocz, B., [et al.]{} 2018, The Astronomical Journal, 155, 128. <http://stacks.iop.org/1538-3881/155/i=3/a=128>
, H., [Catala]{}, C., [Aerts]{}, C., [et al.]{} 2014, Experimental Astronomy, 38, 249
, G. R., [Winn]{}, J. N., [Vanderspek]{}, R., [et al.]{} 2014, in , Vol. 9143, Space Telescopes and Instrumentation 2014: Optical, Infrared, and Millimeter Wave, 914320
, P. H., [Debes]{}, J. H., [Ely]{}, J., [Hines]{}, D. C., & [Bourque]{}, M. 2016, , 823, 49
, L., & [Van Eylen]{}, V. 2018, , 474, 4603
, A., [Johnson]{}, J. A., [Rappaport]{}, S., [et al.]{} 2015, , 526, 546
Vanderplas, J. 2015, [gatspy: General tools for Astronomical Time Series in Python]{}, , , doi:10.5281/zenodo.14833. <http://dx.doi.org/10.5281/zenodo.14833>
, D., & [G[ä]{}nsicke]{}, B. T. 2015, , 447, 1049
, D., [Leinhardt]{}, Z. M., [Bonsor]{}, A., & [G[ä]{}nsicke]{}, B. T. 2014, , 445, 2244
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The evolution of local spin polarization in finite systems involves interference phenomena that give rise to [**quantum dynamical echoes** ]{}and non-ergodic behavior. We predict the conditions to observe these echoes by exploiting the NMR sequences devised by Zhang et al. \[Phys. Rev. Lett. [**69**]{}, 2149 (1992)\], which uses a rare $^{13}$C as [**local probe** ]{}for a dipolar coupled $^1$H spin system. The non-ideality of this probe when testing mesoscopic systems is carefully analyzed revealing the origin of various striking experimental features.'
address: |
Facultad de Matemática, Astronomía y Física. Universidad\
Nacional de Córdoba.\
Ciudad Universitaria. 5000 Córdoba-ARGENTINA
author:
- 'Horacio M. Pastawski, Patricia R. Levstein and Gonzalo Usaj'
date: received 31 May 1995
title: 'Quantum Dynamical Echoes in the Spin ’Diffusion’ in Mesoscopic Systems. '
---
In recent years interest has aroused on transport properties of mesoscopic systems [@Lee]. Examples range from photons propagating through a colloidal solution to electrons in nanostructures. In most of the systems the quantum phenomena manifest as transport properties which are not linear with the size of the system. Since the Boltzmann equation does not retain interferences after collisions, it is necessary the use of the Schrödinger equation (e.g. in Liouville [@Liouville] or Keldysh’s form [@Pastawski]). Thus, a localized excitation, after diffusing away and reaching the boundaries, might return as [**quantum**]{} [**dynamical echoes** ]{}[@Prigodin][** **]{} resembling the familiar sound echoes. A trivial case occurs when a particle is placed in one side of a symmetrical double well potential and oscillates among them [@Levstein]. While in these examples the high group velocities of the excitations restrict the observation of interference phenomena in the time domain, the local excitations in a system of nuclear spins with magnetic dipolar interactions evolve with characteristic times on the hundreds of micro-seconds and can be tested by using an NMR technique devised recently by Zhang, Meier and Ernst [@ZME] (ZME). While they addressed a realization of the Lochsmidt daemon [@Rhim], they showed that, by applying an appropriate radio-frequency pulse sequence involving cross-polarization [@HH], the spin $S$ of a rare $^{13}$C$\,\,\,\,$(1.1% abundance) bonded to a $^1$H$\,\,$with spin $I_1$ can be used as a [**local probe**]{} that injects magnetization in the proton and captures it at a later time. The set of magnetic dipolar coupled $^1$H within a molecule constitutes the mesoscopic system where the spin dynamics can be monitored. The rest of the crystal constitutes a weakly interacting thermal bath.
The experimental [@ZME] time evolution of the polarization of a proton spin in a crystal of ferrocene, (C$_5$H$_5$)$_2$Fe$,$ shows a nearly exponential decay of the local magnetization toward an asymptotic value of $%
0.2$ of the initial polarization. This was interpreted as an ”unhindered” spin diffusion among the five protons of a single ring. This is based on the assumption that the $^{13}$C is an [**ideal local probe**]{}, i.e. the polarization transfer between the $^1$H and $^{13}$C is complete and it takes a short time as compared with the characteristic time of the $^1$H-$^1$H interactions [@Muller]. However, there are two aspects in the results of this ingenious experiment which do not fit with our current understanding of mesoscopic quantum dynamics: One is the linear decay of the magnetization observed for short times which is obviously inconsistent with the expected quadratic decay. The other is the monothonic decay which does not reveal any of the interference features characteristic of quantum evolution in systems with few degrees of freedom. The search for the conditions to actually observe quantum dynamical echoe and answer to these questions lead us to solve the Schrödinger equation for the model system proposed by ZME.
Let us consider an ideal situation in which [*there is*]{} a perfect local probe which is able to inject magnetization in a single $^1$H nucleus and to measure it later. The system consists of $N$ $I$-spins governed by the magnetic dipolar ($\alpha =1$) Hamiltonian truncated with respect to the dominant Zeeman interaction:
$${\cal H}_{II}=\sum_{j>k}\sum_k^{\,\,}d_{jk}\,\left[ \alpha 2I_j^zI_k^z-\frac
12\left( I_j^{+}I_k^{-}+I_j^{-}I_k^{+}\,\right) \right] . \label{Hii}$$
Site $j$ spin operators $I_j$ are dimensionless. Subscripts start sequentially from $j=1.$ The interaction parameters are $$d_{jk}^{}=-\frac{\mu _{0\,\,}\gamma _I^2\,\,\,\,\hbar ^2}{4\pi r_{jk}^3}\,\,%
\frac 12\left\langle 3\cos ^2\theta _{jk}-1\right\rangle , \label{d}$$ where $\gamma _I^{}$ is the gyromagnetic factor of the $I$-spins, $r_{jk}$ are internuclear distances and $\theta _{jk\text{ }}$ their angles with the static magnetic field. A simplifying fact about ferrocene is that the ring performs fast rotations around its five-fold symmetry axis with a very short correlation time[@Seiler] ($\approx 10^{-12}{\rm s}$) and therefore the interaction parameter is time averaged. The new constants depend [@Slichter] only on the angle $\theta $ between the molecular axis and the magnetic field and the angles $\gamma _{j,k}$ between the internuclear vectors and the rotating axis. The factor in brackets in Eq.(\[d\]) becomes $\left\langle \frac 12(3\cos ^2\theta -1)(3\cos ^2\gamma
_{j,k}-1)\right\rangle =1/2$ for a magnetic field normal to the axis. All sites in a molecule are then equivalent and the five fold symmetry is recovered, giving $d_1=1576Hz\times 2\pi \hbar $ for nearest neighbors. A locally polarized initial state $\left| i\right\rangle $ evolves toward a final state $\left| f\right\rangle $ with probability$:$
$$P_{fi}(t_1)=\left| \langle f\mid \exp [-\frac{{\rm i}}\hbar {\cal H}%
_{II}{}t_1]\mid i\rangle \right| ^2. \label{Pt}$$
In the high temperature limit all the $N_i\equiv N_f$ configurations compatible with an up projection of the spin $I_1$ are equally probable. Then the magnetization will be:
$$M(t)=2\left[ \sum_f^{N_f}\sum_i^{N_i}\frac 1{N_i}P_{fi}(t_1)-\frac 12\right]
. \label{Mt}$$
For an [*ideal*]{} system of five spins arranged in a ring configuration, there is $1$ configuration with total projection $5/2$, and there are $5$, $%
10$, $10$, $5$ and $1$ configurations with projections $3/2$, $1/2$, $-1/2$, $-3/2$ and $-5/2$ respectively. On each of these subspaces there are $1$, $4$, $6$, $4$, $1$ and $0$ possible initial and final states with spin $I_1$ polarized summing up $N_i=16$. The evolution under Hamiltonian (\[Hii\]) is called ’diffusion’ in the [*laboratory frame*]{}. For the results of local polarization presented in Fig. 1 we multiply by $-[\frac 12]$ the above interaction constants, which correspond to the experimental case of ’diffusion’ in the [*rotating frame*]{}. The dashed line represents the evolution when only nearest neighbor interactions are taken into account, while the full line considers all the interactions. At [*intermediate times*]{}, the first shows a clear quantum dynamical echo at around $1400\mu
{\rm s}\approx 5\times \sqrt{2}\times \hbar /([\frac 12]d_1)$, which can be interpreted as the time the excitation needs to wind around a ring of length $L=5a$ at an average speed $v_M/\sqrt{2}$, with maximum group velocity $%
v_M=a[\frac 12]d_1/\hbar $. This is confirmed by the study of correlation functions at different sites: the polarization at site $2$ and $5$ has the first maximum at $280\mu {\rm s}$ and at sites $3$ and $4$ around $560\mu
{\rm s}$. Including second neighbors interactions, maxima are split and weakened, this can be assigned to the effect of short-cutting the path around. Although dynamical echoes are still visible, one may conclude that they are weaker when more degrees of freedom become effectively mixed. In the [*long time*]{} regime a striking effect is noticeable in the inset of Fig. 1: local polarization does not distribute around the value $1/5$ but rather oscillates around $\overline{M}$=$\,\lim_{t\rightarrow \infty
}\,t^{-1}\int_0^tM(t^{\prime }){\rm d}t^{\prime }\approx 0.32.$ This non-ergodic behavior is a consequence of the degeneracy of opposite quasi-momenta in perfect rings. We have observed this for single rings as well as for pairs of coaxial rings with $N\leq 10$ spins interacting according to dipolar ($\alpha =1$), Ising like ($\alpha \gg 1$), Heisenberg ($\alpha =-\frac 12$) and XY ($\alpha =0$) interactions by studying $%
\overline{M}.$ In the last situation we obtained the analytical high temperature solution of Eq. ($\ref{Mt})$, which for odd rings gives: $%
\overline{M}_{{\rm i}}=1/N+(N-1)/N^2,$ where the second term accounts for the degenerate quasi-momentum states. The [*complete ferrocene molecule*]{} gives $\overline{M}_{{\rm i}}\approx 0.2.$ Here we distinguish rings rotating independenly or as a whole, keeping an eclipsed or staggered conformation. While the first present a weaker version of the dynamical echoes of the single ring, in the other two the echoes and valleys merge in a sort of plateau for intermediate times. For very [*short times*]{} the magnetization decreases quadratically with $t_1^{}.$ The coefficient is the configuration average of the sum of squares of the coupling constants effective to the flip-flop process in each configuration$.$
These results clearly do not represent the experimental data in ZME [@ZME]. Although quantum dynamical echoes are indeed attenuated by including the interactions with other rings and molecules, the linear decay of the magnetization suggests that some quantum evolution has already occurred for the experimental time $t_1=0$. Therefore, it is important to consider that magnetization is injected and monitored using the spin $S$, which constitutes a [**non-ideal probe**]{}. The complete pulse sequence is schematized in Fig. 2. The most relevant part is as follows: [**A)** ]{}The magnetization from an initially $y$ polarized $S$ spin is transferred to the $y$-axis of the $I_1^{}$ spin (and in lesser degree to its neighbors) through a cross polarization pulse of duration $t_d$, when both the abundant and rare spins are irradiated at their respective resonant frequencies with field strengths fulfilling the Hartmann-Hahn [@HH] condition $\gamma
_IB_{1I}^y=\gamma _SB_{1S}^y.$ The time $t_d$ is the shortest of those that maximize the polarization transfer [@Muller]. [**B)** ]{}The $I$-spins evolve in presence of $B_{1I}^y$ during a time $t_1$. Thus, the relevant evolution of the spin-diffusion sequence occurs in a rotating frame.$\,\,$[**C)**]{} Another cross-polarization pulse of length $t_p=t_d$ is applied to transfer back the polarization to the $x$-axis of $S$. [**D)** ]{}The $S$ polarization is detected while the $I$-system is kept irradiated (high-resolution condition). Therefore a better simulation of the experiment includes the $I$’s spins system, the non ideal probe $S$ and their interaction. The $I$-$S$ dipolar interaction extends the sum in (\[Hii\]) to the $S$ spin. The interaction coefficient is $b_k^{}$ defined as ($\ref{d}%
)$ with a $\gamma _I$ replaced by a $\gamma _S$ and distances and angles modified accordingly. Given the difference in gyromagnetic factors the dipolar $I$-$S$ interaction is truncated dropping terms that mix different Zeeman subspaces, hence: $${\cal H}_{IS}=\sum_kb_k\,\,2I_k^zS_{}^z, \label{His}$$
When two r.f. fields are applied to both $I$ and $S$ spins, these are quantized in their corresponding rotating frames and the total Hamiltonian is:
$${\cal H}_{}^{y(x)}={\cal H}_{II}^{}+{\cal H}_I^{}+{\cal H}_S^{y(x)}+{\cal H}%
_{IS}^{} \label{Hyx}$$
the superscript indicates the direction of $B_{1S}^{}$.
The [**isolated I-spins system** ]{}is described by the first two terms, where ${\cal H}_I=-\hbar \omega _{1I}\sum_kI_k^y$ with $\hbar \omega _{1I}=\gamma
_IB_{1I}^y$ is the effect of the r.f.. In the case of strong irradiation, $%
\mid \hbar \omega _{1I}\mid \gg \mid d_{jk}\mid $ , we can approximate (\[Hii\]) in a frame tilted to the rotating $y$ axis by $${\cal H}_{II}^{^{\prime }}=-[%
%TCIMACRO{\tfrac 12}
%BeginExpansion
{\textstyle {1 \over 2}}
%EndExpansion
]\sum_{j>k}\sum_k^{\,\,}d_{jk}\,\left[ 2I_j^yI_k^y-\frac 12\left(
I_j^{+}I_k^{-}+I_j^{-}I_k^{+}\,\right) \right] . \label{Hii'}$$ As usual, we [*neglected*]{} non-secular terms of the form $\left(
I_j^{+}I_k^{+}+I_j^{-}I_k^{-}\,\right) .$ It is the factor $-[\frac 12]$ relating (\[Hii\]) and (\[Hii’\]) what allows the realization of the Lochsmidt daemon [@Rhim] by experimentally switching the evolution from one situation to the other which leads to a [*polarization echo*]{} [@ZME], not to be confused with the quantum dynamical echoes described above. The [**isolated probe**]{} in the presence of a r.f. field along $y$ $(x)$ direction is described by ${\cal H}_S^{y(x)}=-\hbar \omega
_{1S}S_{}^{\,\,y(x)}$ with $\hbar \omega _{1S}=\gamma _SB_{1S}^{y(x)}$. The [**system-probe interaction**]{} with an additional truncation in the doubly rotating frame is:
$${\cal H}_{IS}^{}\approx \sum_kb_k\frac{({\rm i})}2\,[I_k^{+}S_{}^{-}+(-1)\,%
\,\,I_k^{-}S_{}^{+}], \label{His'}$$
where factors in parenthesis correspond to $B_{1S}^x$. In Fig. 2, the evolution starting with the polarized $S$ spin involves two Hamiltonians, during the times $t_d$ and $t_p$ the evolution is given by the total Hamiltonian ${\cal H}_{}^{y(x)},$ but during $t_1$ as well as during the acquisition, the irradiation makes the action of ${\cal H}_{IS}^{}$ negligible (high resolution condition) and evolves with ${\cal H}%
_{II}^{\prime }+{\cal H}_I^{}$. The probability of having an initial state $%
\left| i_y\right\rangle $ with the probe polarized along $y$ direction and a final state $\left| f_x\right\rangle $ with the probe polarized along $x$ is: $$P_{f,i}(t_1)=\mid \langle f_x\mid \exp [-%
%TCIMACRO{\tfrac{{\rm i}}\hbar }
%BeginExpansion
{\textstyle {{\rm i} \over \hbar}}
%EndExpansion
{\cal H}^xt_p]\exp [-%
%TCIMACRO{\tfrac{{\rm i}}\hbar }
%BeginExpansion
{\textstyle {{\rm i} \over \hbar}}
%EndExpansion
({\cal H}_I+{\cal H}_{II}^{\prime })t_1]\exp [-%
%TCIMACRO{\tfrac{{\rm i}}\hbar }
%BeginExpansion
{\textstyle {{\rm i} \over \hbar}}
%EndExpansion
{\cal H}^yt_d]\mid i_y\rangle \mid ^2$$
The resulting magnetization is shown in Fig. 3 with open circles. For short $%
t_1$ the evolution is linear with time and the dynamical echoes are still noticeable but appear shifted to shorter times. Both effects are consequence of the evolution of the magnetization in the $I$-$I$ system during the two cross polarization periods: by comparing with Fig.1, it is clear that the overall effect of them is to shift the curve in a time shorter than their sum. This is verified by superposing the ideal evolution from Fig. 1, shifted in $t_{{\rm shift}}=126\mu s<(t_d+t_p)$ and renormalized with a factor of $a=1.22$ obtained from the short time behavior. This explains why the asymptotic value of the magnetization in the complete sequence has been raised to $\overline{M}_{{\rm n.i}.}=a\overline{M}_{{\rm i}}\approx 0.4$ from the $\overline{M}_{{\rm i}}\approx 0.32$ around which the ideal polarization of Fig.1 oscillates. Then $\overline{M}_{{\rm n.i}.}$ overestimates the fraction of magnetization remaining in the proton. The fact that Figs. 1 and 3 can be superposed indicates that $S$-$I$ interactions do not break the five-fold symmetry appreciably within the relevant experimental times, this is a consequence of the high resolution condition.
These results indicate that the experimental [@ZME] asymptotic value $%
\overline{M}_{\exp }=0.2=\overline{M}_{{\rm n.i}.}$ originates on the evolution that has already occurred at $t_1=0,$when polarization starts to be recorded and is normalized. A more proper normalization would give an ideal asymptotic value $\overline{M}_{{\rm i}}$ $\approx \overline{M}_{\exp
}/a=$ $0.16,$ lower than $\overline{M}_{{\rm i}}$ for the complete molecule (two coupled rings). Since, as discussed above, the quantum evolution [**does not give an equi-distribution of polarization,**]{} to achieve the experimental $0.2$, the evolution must involve a system with $N\gg 10.$ With the participation of an important number of spins other than those belonging to one molecule, the boundary of the system becomes fuzzy, weakening the interferences that originate the quantum dynamical echoes. Then, it becomes clear why these do not show up in the reported experiments.
An interesting phenomenon shown up by the non-ideal evolution in Fig 3 is the small high frequency ($\approx \omega _{1I}$) oscillation around the superposed ideal curve. This effect is due to an incomplete $S$-$I$ transfer which allows interference of phases accumulated during evolution $t_1$ by sub-spaces with different total spin $\sum I_k^{\,\,\,y}$ projection; i.e. the interference of the dashed and dotted paths of polarization amplitudes in Fig. 2. They may originate the small ’dispersion’ observed in the top of the polarization echo in the experimental curve.
In summary, we have solved the relevant part of a non-ideal spin ’diffusion’ experiment for a simple five-spins system resorting to the usual assumptions to generate the system, probe and system-probe Hamiltonians: truncation of the dipolar Hamiltonian and total decoupling under high resolution conditions (during the time $t_1$)$.$ We have shown that while $^{13}$C behaves as a [**non-ideal local probe**]{}, its effect can be reasonably quantified. As occurs in all mesoscopic physics the probes are actually part of the system [@Buttiker] and therefore have a relevant role in the interpretation of the results. This undesired effect can be experimentally overcome in part, by studying the diffusion in the laboratory frame which tends to compensate the non-ideality of the probe and shrinks in a factor $%
\left[ \frac 12\right] $ the time scale, ’cleaning’ the evolution from interactions different than the [*secular*]{} dipolar one. We predict that a spin ’diffusion’ experiment in molecular crystals containing few interacting spins per molecule should present [**quantum dynamical echoes**]{}. Possible candidates are molecules with single rings of the form (C$_n^{}$H$_n$)$.$ Molecules with linear topology in the interaction network present less important second neighbors interaction, making the interference effects even stronger.
This work was performed at LANAIS de RMN (UNC-CONICET) with financial support from Fundación Antorchas, CONICOR and SeCyT-UNC.
edited by[* *]{}B. L. Altshuler, P. A. Lee and R. A. Webb (North Holland, Amsterdam, 1992).
W. R. Frensley, Rev. Mod. Phys. [**62**]{},754 (1990).
H. M. Pastawski, Phys. Rev. B. [**44**]{} 6329 (1991); [**46,**]{} 4053 (1992).
V. N. Prigodin, B. L. Altshuler, K. B. Efetov and S. Iida, Phys. Rev. Lett. [**72**]{}, 546 (1994)
P. R. Levstein, H. M. Pastawski and J. L. D’Amato, J. Phys. Condens. Matt. [**2**]{}, 1781 (1990)
S. Zhang, B. H. Meier, and R. R. Ernst, Phys. Rev. Lett. [**69**]{}, 2149 (1992)
W. -K. Rhim, A. Pines and J. S. Waugh, Phys. Rev. Lett., [**25**]{} , 218 (1970)
S. R. Hartmann and E. L. Hahn, Phys. Rev. [**128**]{}, 2042 (1962); M. H. Levitt, D. Suter, and R. R. Ernst, J. Chem. Phys. [**84**]{}, 4243 (1986)
L. Müller, A. Kumar, T. Baumann and R. R. Ernst, Phys. Rev. Lett. [**32**]{}, 1402 (1974)
P. Seiler and J. D. Dunitz, Acta Cristallog. Sect. B [**35**]{}, 1068 (1979)
C. P. Slichter, [*Principles of Magnetic Resonance*]{} (Springer, New York, 1992)
M. Büttiker, Phys. Rev. Lett. [**57**]{}, 1761 (1986)
[**Figure 1:** ]{} Ideal evolution of a $^1$H-spin polarization as a function of $t_1$ for a system of $5$ spins in a ring. Dashed line considers only nearest neighbor interaction $d_1$. Full line is the exact dipolar evolution. The inset shows fluctuations around the value $\overline{M}_{{\rm %
i}}=0.32$.
[**Figure 2:**]{} Pulse sequence for measurement of proton spin ’diffusion’ in the rotating frame. A $\pi /2$ pulse on the abundant $^1$H spins system creates a polarization that is transferred during $t_C$ to a rare $^{13}$C. After the decay of the proton spin coherence during $t_S$: ([**A**]{}) the $%
^{13}$C polarization is locally injected to the bonded proton, ([**B**]{}) ’diffusion’ in the protons system is allowed and ([**C**]{}) it is captured again in the $^{13}$C where ([**D**]{}) it is recorded.
[**Figure 3:**]{} Non-ideal evolution of a $^1$H spin polarization in a ring of ferrocene (open circles) as detected in a strongly coupled $^{13}$C spin. Calculation uses the pulse sequence in Fig. 2 with $t_d=t_p=85\mu {\rm s}$ with the molecular symmetry axis at an angle of $\pi /2$ with the external magnetic field. Full line is the ideal evolution in Fig.1, shifted in $%
126\mu {\rm s}$ and normalized accordingly.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Previous results on certain sampling series have left open if divergence only occurs for certain subsequences or, in fact, in the limit. Here we prove that divergence occurs in the limit.
We consider three canonical reconstruction methods for functions in the Paley-Wiener space $\PWpio$. For each of these we prove an instance when the reconstruction diverges in the limit. This is a much stronger statement than previous results that provide only $\limsup$ divergence. We also address reconstruction for functions in the Hardy space $\Ho$ and show that for any subsequence of the natural numbers there exists a function in $\Ho$ for which reconstruction diverges in $\limsup$. For two of these sampling series we show that when divergence occurs, the sampling series has strong oscillations so that the maximum and the minimum tend to positive and negative infinity. Our results are of interest in functional analysis because they go beyond the type of result that can be obtained using the Banach-Steinhaus Theorem. We discuss practical implications of this work; in particular the work shows that methods using specially chosen subsequences of reconstructions cannot yield convergence for the Paley-Wiener Space $\mathcal{PW}^1_\pi$
address:
- 'Lehrstuhl für Theoretische Informationstechnik, Technische Universität München, Arcisstrasse 21, D080290, München, Germany'
- 'Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA 91125, U.S.A. '
author:
- Holger Boche
- Brendan Farrell
title: 'Strong Divergence of Reconstruction Procedures for the Paley-Wiener Space $\mathcal{PW}^1_\pi$ and the Hardy Space $\mathcal{H}^1$'
---
AMS Subject Classification: 94A20, 94A12
Keywords: Sampling series, Reconstruction, Paley-Wiener space, Hardy space
Introduction
============
Sampling theory originated with the study of reconstruction of a function in terms of its samples; the fundamental initial results of the theory state conditions on a function $f:\R\rightarrow\C$ for the expansion $$\label{eq:start}
f(t)=\sumbiinf f(k)\sinpitmk$$ to be justified and for equality to hold. Determining the appropriate function classes, the regions of convergence and proper ways to address the infinite series now constitute an entire area of research. The present paper addresses the pointwise approximation of sampling series for several standard reconstruction procedures for functions in the Paley-Wiener and Hardy spaces. We present four main results. Three of these (Theorems \[thm:one\], \[thm:two\] and \[thm:three\]) show stronger forms of divergence than were previously known for three different reconstruction procedures for functions in the Paley-Wiener space $\PWpio$; and the fourth (Theorem \[thm:four\]) is the first divergence result for a reconstruction procedure for the Hardy space $\Ho$. While the first three Theorems just mentioned concern strong divergence, the last holds for $\limsup$ divergence.
We now introduce the necessary definitions for a discussion of our results. $D$ denotes the unit disc $D:=\{z\in\C:\; |z|<1\}$, and $\partial D$ denotes its boundary. For $\sigma>0$ and $1\leq p<\infty$ the Paley-Wiener space $\PW^p_\sigma$ is the set of all functions $f$ that can be represented as $f(z)=\otwopi\int_{-\sigma}^{\sigma}g(\eiomega)e^{i\omega z}d\omega$, for all $z\in\C$ and some $g\in L^p(\partial D)$. The norm on $\PW^p_\sigma$ is $$\|f\|_{\PW^p_\sigma}=\left(\otwopi\int_{-\sigma}^{\sigma}|\fh(\omega)|^pd\omega\right)^{1/p} =
\left(\otwopi\int_{-\sigma}^{\sigma}|g(\eio)|^pd\omega\right)^{1/p},$$ where the second equality is due to the identity $\fh(\omega)=g(\eio)$ for $\fh$ being the Fourier transform of $f$.
For a function $f\in\PWpip$, $1\leq p<\infty$, one may consider the Shannon sampling series $$\label{eq:shannon}
\sumbiinf f(k)\sinpitmk.$$ For $1<p<\infty$, converges absolutely and uniformly on all $\R$ [@BSS88]. For $p=1$, one has the following theorem of Brown [@Bro67].
For all $f\in\PWpio$ and all $T>0$, $$\lim_{N\rightarrow\infty}\left( \max_{t\in[-T,T]}\left|f(t)-\sumbiN f(k)\sinpitmk\right|\right)=0.$$
The study of function representations using their samples is now its own field. Historical articles such as [@BDFHLSS] describe the origins of this field in the 1930’s and 1940’s and point out that a number of authors made groundbreaking contributions in that era, but, as dissemination was limited, many were not adequately recognized by the historical record until the last several decades. Yet, despite decades of research, there remain interesting and important open questions about the convergence of sampling series. Originally research was focused on $\PWpit$, but in recent years efforts have been made to investigate larger signal spaces. About twenty years ago, Butzer [@But93] suggested the study of convergence for reconstruction of elements from the function class $\PWpio$. In particular, he suggested studying the quantity $$\max_{t\in\R}\left|f(t)-\sum_{k=-N}^Nf(k)\sinpitmk\right|\label{eq:Butzer2}$$ as $N$ tends to infinity. In [@BM08] it is shown that there exists a function $\fo\in \PWpio$ such that $$\limsup_{N\rightarrow\infty}\left(\max_{t\in\R}\left|\fo(t)-\sum_{k=-N}^N\fo(k)\sinpitmk\right|\right)=+\infty.\label{eq:BM1}$$ In fact, in [@BM08] it is shown that a function leading to the divergence behavior in exists for any reconstruction process that relies primarily on integer sampling points. An important instance of the general result proved in [@BM08] is for the Valiron interpolation series $$\label{eq:Valiron}
f(t)=f(t_0)\frac{\sin \pi t}{\sin\pi t_0}+(t-t_0)\sumbiinf \frac{f(k)}{k-t_0} \sinpitmk,$$ where $t_0\in\R\backslash \Z$. Divergence is shown in [@BM08] for this expansion among others.
\[remark:Valiron\] A significant difference between the Valiron and the Shannon series is that the Valiron series is absolutely convergent for every fixed $t$. We return to the Valiron series in Corollary \[cor:Valiron\].
In [@BM08] and other works the divergence given in is only given in terms of the $\limsup$. Thus, two very natural questions remain:
- Does there exist a universal sequence $\{N_l\}_{l\in\N}$ such that for all $f\in \PWpio$, $$\label{eq:Q1}
\sup_{l}\left(\max_{t\in\R}\left|f(t)-\sum_{k=-N_l}^{N_l}f(k)\sinpitmk\right|\right)<\infty$$ holds?
- - If the universal sequence of Q1) does not exist, does there exist a specific sequence $\{N_l\}_{l\in\N}$ for each $f\in\PWpio$ such that holds.
Note that a negative answer to Q2) for $\PWpio$ implies a negative answer to Q1).
A negative answer to both of these questions is given by Theorem \[thm:one\] of the next section. In Section \[sec:four\] we address reconstruction of functions in the Hardy space $\Ho$. Theorem \[thm:four\] provides a divergence behavior that, due to a classical result on convergence of lacunary series, is unexpected and also provides a negative answer to Q1) for $\Ho$.
In Section \[sec:two\] we return to $\PWpio$ and address reconstruction for functions of sine type. In Section \[sec:three\], we address reconstruction of a function using the interpolating sequence that it generates. Lastly, in Section \[sec:diverge\] we show that both the maximum and minimum of the reconstruction procedures addressed in Sections \[sec:two\] and \[sec:three\] diverge strongly.
We say that $\{a_k\}_{k\in\N}$ diverges *strongly* if $\lim_{k\rightarrow\infty}a_k=+\infty$ or $\lim_{k\rightarrow\infty}a_k=-\infty$, which is stronger statement than $\limsup_{k\rightarrow\infty}a_k=+\infty$ or $\liminf_{k\rightarrow\infty} a_k=-\infty$. Theorems \[thm:one\], \[thm:two\] and \[thm:three\] provide strong divergence statements where previously only divergence statements were available. We emphasize that this is significant because it rules out the possibility that divergence occurs only as a result of a strongly divergent subsequence. This has obvious practical implications, and after the statement of Theorem \[thm:one\] we discuss the mathematical significance as well. Lastly, before turning to our results, we suggest to the reader the interesting section on “Analogue Signal Transmission” in Feynman’s book [@Fey98]. This chapter sets fundamental questions from sampling theory in the wider context of the physics of computation.
Strong Divergence for Reconstruction of Signals in $\PWpio$ {#sec:one}
===========================================================
\[thm:one\] There exists a function $\fo\in \PWpio$ such that $$\label{eq:diverge}
\lim_{N\rightarrow\infty}\left(\max_{t\in\R} \left| \sum_{k=-N}^N \fo(k)\sinpitmk \right|\right)=+\infty.$$
Before proving Theorem \[thm:one\], we first set it in a broader mathematical context, and then present a corollary and a conjecture. Theorem \[thm:one\] is interesting for functional analysis because of its relationship to the fundamental Banach-Steinhaus Theorem [@BS27]. The latter theorem says that a set of bounded linear operators acting on a Banach space is either uniformly bounded, or the supremum of the operators applied to elements of the space diverges for all elements in a residual set. Thus, if the operators being considered are countable and indexed, say $\{T_k\}_{k\in\N}$, then if $\sup_k\|T_k\|=+\infty$, $$\sup_k\|T_kx\|=+\infty$$ for all $x$ in a residual set. The strength of this statement is that it holds for such a large set of functions; its weakness, however, is that one only has $\sup_k\|T_kx\|=+\infty$, while in some settings it is natural to ask if $\lim_{k\rightarrow\infty}\|T_kx\|=+\infty$. This is the case when one is interested in bounded behavior of a subsequence. In particular, $\sup_{k}\|T_k z\|=+\infty$ does not exclude the possibility that there exists a subsequence $\{k_l\}_{l\in\N}$ such that $\sup_{l}\|T_{k_l}x\|<\infty$, or even $\sup_{l}\|T_{k_l}\|<\infty$. Thus, the investigation of a reconstruction method is incomplete if it only shows divergence of the $\sup$. The value of Theorem \[thm:one\], then, is that it demonstrates strong divergence.
Projections onto subsets of the Walsh functions provide an example where $$\sup_{k\in\N}\|P_k\|_{L^1([0,1])}=+\infty,$$ yet $\|P_{2^k}\|_{L^1([0,1])}=1$ for all $k$, where $P_l$ is the projection onto the first $l$ Walsh functions [@Fin49]. Thus Q1) holds even though the sequence of projections is divergent. Note also that $\|\cdot\|_{L^1([0,1])}$ is the norm corresponding to $\PWpio$.
It is possible that the methods developed here can be adapted to pursue further statements similar to the Banach-Steinhaus Theorem for other nonlinear operators, as is done in [@DNvW85]. While we have not indicated the rates of divergence in the statements of our theorems, the proofs are constructive and, therefore, yield quantitative divergence estimates. Thus, one could potentially couple these techniques with others developed, for example in [@DN85], to obtain quantitative versions of the Banach-Steinhaus Theorem for other operators.
There is also much to investigate about the relationship between a statement like Theorem \[thm:one\] and the Banach-Steinhaus Theorem. The latter theorem could not yield the strong limit proved in Theorem \[thm:one\], yet it would be interesting to know the nature of the set of functions $f$ for which divergence occurs, in particular if it is a residual set, as occurs in the Banach-Steinhaus Theorem. Similarly, the Banach-Steinhaus Theorem is not applicable to the nonlinear operator of Theorem \[thm:three\], yet here as well it would be interesting to understand the set for which the operator diverges.
A large class of standard reconstruction procedures was addressed in [@BM08] and $\limsup$ divergence results were presented. Theorem \[thm:one\] is a strengthening of Theorem 3 in [@BM08] from $\limsup$ to the limit, and the proof given here holds for the procedures addressed in [@BM08]. As a corollary, we state this explicitly for the Valiron reconstruction mentioned earlier. As pointed out in Remark \[remark:Valiron\], what is significant here is that despite absolute convergence for every fixed $t\in\R$, we still have global divergence of the $\|\cdot\|_{\infty}$ norm. While we state the corollary for the Valiron series, the same behavior holds for the wider class of [@BM08].
\[cor:Valiron\] For every fixed $t_0\in \R$, there exists a function $f_1\in\PWpio$ such that $$\lim_{N\rightarrow\infty}\left( \sup_{t\in\R}\left|f_1(t)-\left(f_1(t_0)\frac{\sin \pi t}{\sin\pi t_0}+(t-t_0)\sumbiN \frac{f_1(k)}{k-t_0} \sinpitmk\right)\right|\right)=+\infty.$$
We conjecture that the behavior seen in Theorem \[thm:one\] holds much more widely. To formulate the conjecture we need further notation, and we now define the terms needed for the rest of the paper. For a separable Hilbert space $H$, a set of vectors $\{\psi_{k}\}_{k\in\Z}$ is called a Riesz basis if $\{\psi_k\}_{k\in\Z}$ is complete in $H$ and there exist positive constants such that for all scalars $\{c_k\}_{k\in\Z}$ and $M,N\in \N$, $$A\sum_{k=-M}^N|c_k|^2\leq \left\|\sum_{k=-M}^Nc_k\psi_k\right\|^2\leq B\sum_{k=-M}^N|c_k|^2.$$ A sequence $\{\tk\}_{k\in\Z}$ is a complete interpolating sequence for $\PWpit$ if the interpolation problem $f(\tk)=c_k$, $k\in\Z$ has a unique solution $f\in \PWpit$ for every sequence $\{c_k\}_{k\in\Z}$ satisfying $\sumbiinf |c_k|^2<\infty$, see [@You01]. If $\{\tk\}_{k\in\Z}$ is a complete interpolating sequence, then one can use the functions $$\phi(z):=\lim_{R\rightarrow \infty}\prod_{|t_k|\leq R,\;\tk\neq 0}\left(1-\frac{z}{\tk}\right)$$ and $$\phi_k(z):=\phiz,\;\;\;k\in\Z$$ for reconstruction. This approach has been used in communications engineering for decades, [@Voe70; @VR73; @Log84].
Later in the paper we will consider a special case of this construction. For this we will need the following definitions. An entire function is of exponential type $\sigma$ if for all $\epsilon>0$ there exists a constant $C(\epsilon)>0$ such that for all $z\in\C$, $|f(z)|\leq C(\epsilon)e^{(\sigma+\epsilon)|z|}$.
\[def:sinetype\] An entire function is of sine type if
- the zeros of $f$ are separated and simple, and
- there exist positive constants $A,B$ and $H$ such that $Ae^{\pi|y|}\leq |f(x+iy)|\leq Be^{\pi|y|}$ whenever $x$ and $y$ are real and $|y|\geq H$.
With this notation set, we state our conjecture and then prove Theorem \[thm:one\].
\[conj:complete\] Let $\{\tk\}_{k\in\Z}$ be an arbitrary complete interpolating sequence. Then there exists a function $f_1\in\PWpio$ such that $$\lim_{N\rightarrow\infty}\left(\max_{t\in\R}\left| \sumbiN f_1(\tk)\phi_k(t)\right|\right)=+\infty.$$
In Theorem \[thm:two\] we prove an instance of this conjecture for a class of complete interpolating sequences that is of practical importance, yet we suspect that entirely new techniques will be necessary to prove the conjecture in its generality.
We begin with a sequence of functions $\{w_N\}_{N\in\N}$ contained in $\PWpio$. We set $$\label{eq:w}
\wN(k)=\left\{
\begin{array}{ll}
1&|k|\leq N,\;k\in\Z\\
1-\frac{|k|-N}{N}&N+1\leq |k|\leq 2N,\;k\in \Z\\
0&|k|\geq 2N,\; |k|\in \Z.
\end{array}\right.$$ and $$\wN(t)=\sum_{k=-2N+1}^{2N-1}w_N(k)\sinpitmk,$$ for which the bound $$\|\wN\|_{\PWpio}<3$$ was obtained in [@BM10]. We now consider the function $$g(t)=\sum_{l=1}^\infty \frac{1}{l^2}w_{N_l}(t),\;\;\;t\in\R,$$ where $N_l=2^{l^3}$, $l\in\N$. Here we have $$\|g\|_{\PWpio}\leq 3\sum_{l=1}^\infty \frac{1}{l^2}<+\infty.$$ Let $\gh$ denote the Fourier transform of $g$ and set $$\label{eq:five}
G(e^{i\omega})\eqD \sumbiinf g(k)e^{-ik\omega}.$$ We have $\gh(\omega)=G(e^{i\omega})$ for almost all $|\omega|\leq \pi$; that is, $G\in L^1(\partial D)$, where $L^1(\partial D)$ denotes the set of Lebesque-measurable functions on the unit circle $|z|=1$ satisfying $$\otwopi \inttwopi |G(e^{i\omega})|d\omega<\infty.\label{eq:G}$$ We set $F_1(e^{i\omega})=G(e^{i(\omega+\pi)})$ so that $$\label{eq:six}
\Fo(\eio)\eqD \sumbiinf g(k)(-1)^ke^{-i\omega k }$$ and $$\foh(\omega)=\Fo(\eio),\;\;\;\textnormal{ for }\;|\omega|\leq \pi,$$ so that $$\fo(t)=\otwopi \inttwopi \foh(\omega)e^{i\omega t}d\omega.$$ Then $\fo\in\PWpio$ and $$\fo(k)=(-1)^kg(k).$$ Note that equalities and hold in the sense of distribution theory.
Consider a fixed $N$. The functions $G$ and $\Fo$ are uniquely determined by their Fourier coefficients. We, therefore, have $$\begin{aligned}
\left|\sumbiN \fo(k)\sinpiNphmk\right|&=&\left|\sin\pi\left(\Nph\right)\sumbiN \fo(k)\frac{(-1)^k}{\pi(\Nph-k)}\right|\\
&=&\opi \sumbiN \frac{g(k)}{(\Nph-k)}.\end{aligned}$$ There exists a unique $\kh$ such that $N\in[N_{\kh},N_{\kh+1}]$. This gives us $$\begin{aligned}
\opi\sumbiN \frac{g(k)}{\Nph-k}&\geq& \frac{1}{(\kh+1)^2}\opi\sumbiN\frac{1}{\Nph-k}
=\frac{1}{(\kh+1)^2}\opi \sum_{k=0}^{2N}\frac{1}{k+\oh}\label{eq:kone}\\
&>& \frac{1}{(\kh+1)^2}\opi\int_0^{2N+1}\frac{dx}{x+\oh}
=\frac{1}{(\kh+1)^2}\opi\log(4N+3)\nonumber\\
&\geq& \frac{1}{(\kh+1)^2}\opi \log N_{\kh}=\frac{(\kh)^3}{(\kh+1)^2}\opi \log 2.\label{eq:ktwo}\end{aligned}$$ Thus, $$\left|\sumbiN \fo(k)\sinpiNphmk\right|\geq \opi\log 2\frac{(\kh)^3}{(\kh+1)^2},$$ from which follows $$\max_{t\in\R}\left|\sumbiN \fo(t)\sinpiNphmk\right|\geq \opi\log 2\frac{(\kh)^3}{(\kh+1)^2},$$ and, hence, $$\lim_{N\rightarrow\infty}\left( \max_{t\in\R}\left|\sumbiN \fo(t)\sinpiNphmk\right|\right)=+\infty.$$
Behavior for the Hardy space $\Ho$ {#sec:four}
==================================
The Hardy space $H_p$ consists of analytic functions on $D$ such that $$\|f\|_{H^p}:=\sup_{0\leq r<1}\left(\otwopi\int_{-\pi}^\pi |f(r\eiomega)|^pd\omega\right)^{1/p}<\infty,$$ which also defines the norm $\|\cdot\|_{H^p}$. The space $\Ho$ is then the space of functions $f$ with representation $f(z)=\otwopi\int_{-\pi}^{\pi}g(\omega)e^{i\omega z}d\omega$ for $z\in\C $ and some $g\in H^1$, see [@Hig96]. The norm on $\Ho$ is given by $\|f\|_{\Ho}=\|g\|_{H^1}$. $\Ho$ is a closed subspace of $\PWpio$ and so invites the same questions for its sampling series that are addressed for $\PWpio$. In particular, if a form of convergence of a certain sampling procedure does not hold for all $f\in\PWpio$, it is appropriate to ask whether the convergence holds for all $f\in\Ho$. A fundamental sampling procedure is the finite Fourier expansion of $\fh$: $$(\F_Nf)(\omega):=\sumbiN f(k)\eiok.$$
There is a rich history of results for this expansion. Kolmogorov famously showed [@Kol26] that there exists $\fo\in\PWpio$ so that for all $\omega\in[-\pi,\pi)$, $$\limsup_{N\rightarrow \infty}\left| \sumbiN \fo(k)\eiok\right|=+\infty.$$ It further holds, see Section VIII.3 of [@Zyg02], that for every subsequence $\{N_l\}_{l\in\N}$ there exists a function $f_2\in\PWpio$ (dependent on the subsequence) so that $$\label{eq:divsub}
\limsup_{l\rightarrow \infty}\left| \sum_{k=-N_l}^{N_l} f_2(k)\eiok\right|=+\infty$$ for almost all $\omega\in[-\pi,\pi)$. Yet, Theorem 5.11 in [@Zyg02] states that for all $f\in \Ho$ and for every subsequence $\{N_l\}_{l\in\N}$ satisfying $$\label{eq:lacunary}
\inf_{l}\frac{N_{l+1}}{N_l}=\lambda>1,$$ where such a sequence is called lacunary, $$\label{eq:Zygmund}
\lim_{l\rightarrow \infty} \sum_{k=0}^{N_l} f(k)\eiok=\fh(\omega)$$ for almost all $\omega\in [\pi,\pi)$. Thus, for every subsequence $\{N_l\}_{l\in\N}$ there exists a function $f\in\PWpio$ for which the divergence holds, yet if $\{N_l\}_{l\in\N}$ grows fast enough, one has the convergence for all functions in $\Ho$. The natural question is: does convergence hold globally for functions in $\Ho$, i.e. is the answer to Q1) for $\Ho$ positive? The following theorem shows that the answer is negative.
\[thm:four\] Let $\{N_l\}_{l\in\N}$ be an arbitrary sequence of natural numbers. There exists a function $f_1\in \Ho$ such that $$\label{eq:four}
\limsup_{l\rightarrow\infty}\left( \max_{t\in\R}\left|\sum_{k=0}^{N_l} f_1(k)\sinpitmk \right|\right)=+\infty.$$
\[conj:four\] Q2) holds for $\Ho$. That is, for each $f\in\Ho$ there exists a subsequence $\{N_l\}_{l\in\N}$ such that the $\limsup$ expression in remains bounded for that $f$. In practical terms this means adaptive procedures can be effective for $\Ho$.
By the Banach-Steinhaus Theorem, for a fixed sequence $\{N_l\}_{l\in\N}$, the set of functions in $\Ho$ for which the sampling procedure addressed in Theorem \[thm:four\] diverges is a residual set. Theorem \[thm:one\] proves the existence of a function $f_1$ in $\PWpio$ for which the reconstruction addressed there diverges for *any* subsequence. Conjecture \[conj:four\] claims that completely different behavior holds for $\Ho$: if the conjecture holds, then for every $f\in\Ho$ there exists a subsequence yielding boundedness.
We construct a sequence of functions by setting $q_N(k)=(-1)^kw_N(k)$ for $k\in\Z$, where $w_N$ is the function defined in . We further set $$q_N(t)=\sum_{k=-2N+1}^{2N-1} (-1)^kw_N(k)\sinpitmk.$$ Let $\{N_l\}_{l\in\N}$ be the arbitrary subsequence of the natural numbers for which we will show divergence of the reconstruction. For $l\in\N$ we define $\Nol$ to be the largest natural number satisfying $3\Nol\leq \Nl$. Now we set $$q^1_{N_l}(t)=q_{\Nol}(t-\Nl+\Nol),$$ so that $$\qh^1_{N_l}(\omega)=e^{i(\Nl-\Nol)\omega}\qh_{\Nol}(\omega),\;\;\omega\in[-\pi,\pi).$$ Just as in the proof of Theorem \[thm:one\], we have $$\|q^1_{N_l}\|_{\Ho}=\|q_{\Nol}\|_{\PWpio}<3.$$ We now select a subsequence $\{l_r\}_{r\in\N}$ such that
- $\frac{1}{r^2}\opi \log(\Nlr-\frac{3}{2})\geq r$
- - $N^1_{l_{r+1}}\geq (\Nlr)^2$
for all $r$ and define $$\fo(t)=\sumroi \frac{1}{r^2}q^1_{\Nlr}(t),$$ for which we have $$\|\fo\|_{\Ho}\leq \sumroi \frac{1}{r^2}\|q^1_{\Nlr}\|_{\Ho}<3\frac{\pi^2}{4}.$$ For $m\in\N$ we set $t_m=\Nlm+\oh$, for which $$\begin{aligned}
\sumNlm \fo(k)\sintmmk
&=& \sum_{r=1}^{m-1} \frac{1}{r^2} \sumNlm \qo_{\Nlr}(k)\sintmmk \label{eq:first}\\
&&+ \frac{1}{m^2}\sumNlm \qo_{N_{l_m}}(k)\sintmmk\label{eq:second} \\
&&+\sum_{r=m+1}^\infty \frac{1}{r^2}\sumNlm \qo_{\Nlr}(k)\sintmmk.\label{eq:third}\end{aligned}$$ For $r\leq m-1$ we have $$\sum_{k=0}^{N_{l_m}} q^1_{N_{l_r}}(k)\sintmmk =q^1_{N_{l_r}}(t_m).$$ Using this identity and the bound $$|q^1_{N_{l_r}}(t)|\leq \|q^1_{N_{l_r}}\|_{\PWpio}<3$$ for all $t\in \R$, for the term on the right in we have $$\label{eq:newA}
\left|\sum_{r=1}^{m-1}\sumNlm\frac{1}{r^2} \qo_{\Nlr}(k)\sintmmk \right|\leq \sum_{r=1}^{m-1}\frac{1}{r^2}| q_{N^1_{l_r}}(t_m-N_{l_r}+N^1_{l_r})|<3\frac{\pi^2}{4}.$$ For the term we have $$\begin{aligned}
\left|\frac{1}{m^2}\sumNlm \qo_{\Nlm}(k)\sintmmk \right|
&=& \left|\frac{1}{\pi m^2}\sumNlm (-1)^{k-\Nlm+\Nlm^1}w_{\Nlm^1}(k-\Nlm+\Nlm^1)\frac{(-1)^k\sin \pi t_m}{\Nlm+\oh-k} \right|\nonumber\\
&=&\frac{1}{\pi m^2}\sumNlm w_{\Nlm^1}(k-\Nlm+\Nlm^1)\frac{1}{\Nlm+\oh-k}\nonumber\\
&>& \frac{1}{\pi m^2}\sum_{k=\Nlm-2 N^1_{l_m}}^{\Nlm} \frac{1}{\Nlm+\oh-k}= \frac{1}{\pi m^2}\sum_{k=1}^{2 N^1_{l_m}} \frac{1}{k+\oh}\nonumber\\
&=&\frac{1}{\pi m^2}\log \left(3 N^1_{l_m}+\frac{3}{2}\right)>\frac{1}{\pi m^2}\log\left(\Nlm-\frac{3}{2}\right)\nonumber\\
&>&m, \label{eq:newB}\end{aligned}$$ where the last inequality holds because $3 N^{1}_{l_m}+3>\Nlm$. To address we note that for $r>m$ we have $N^1_{l_{r+1}}>(\Nlm)^2$, so that $$\begin{aligned}
\left|\sumNlm q^1_{\Nlr}(k)\sinNlm \right| &\leq& \sumNlm|q^1_{\Nlr}(k)|\frac{1}{\pi(\Nlm+\oh-k)}\nonumber\\
&\leq & \frac{1}{\pi \No_{l_r}}\sumNlm \frac{k}{\Nlm+\oh-k}= \frac{1}{\pi \No_{l_r}}\sumNlm \frac{\Nlm-k}{k+\oh}\label{eq:beachte}\\
&<& \opi \frac{\Nlm}{\No_{l_r}}\log \Nlm<\opi \frac{(\Nlm)^2}{\No_{l_r}}<\frac{1}{\pi}.\label{eq:newC}\end{aligned}$$ For the first inequality in it is important to note that if $r>m$ then $0\leq k\leq \Nlm-2\No_{l_m}\leq \Nlr-2\No_{l_r}$ yields $\frac{k}{\Nlr-2\No_{l_r}}\geq |q_{\Nlr}(k)|$ and $\Nlr-2\No_{l_r}\geq \No_{l_r}$, so $|q_{\Nlr}(k)|\leq \frac{k}{\No_{l_r}}$.
Combining , and we obtain $$\left|\sumNlm f_1(k)\sintmmk \right|\geq m-\frac{\pi}{4}-3\sum_{r=1}^{m-1}\frac{1}{r^2}
>m-\frac{\pi}{4}-\frac{3\pi^2}{4},$$ from which the result follows.
Theorem \[thm:four\] can be used to show that known results on the maximal operator are sharp.
To address the convergence in for a given subsequence $\{N_l\}_{l\in\N}$, we use the *maximal operator* $$(M_*f)(\omega):=\sup_{l\geq 1}\left|\sum_{k=0}^{N_l}f(k)\eiok\right|,$$ as developed in [@Ste61]. Note that the operator depends on the sequence $\{N_l\}_{l\in\N}$, and so the constants in what follows depend as well on the sequence. If $\{N_l\}_{l\in\N}$ is a sequence of natural numbers satisfying , then for every $\mu\in (0,1)$ there exists a constant $C_\mu$, so that for all $f\in\Ho$ $$\label{eq:mu}
\left(\frac{1}{2\pi}\int_{\pi}^\pi |(M_*f)(\omega)|^\mu d\omega\right)^{1/\mu}\leq C_\mu \|f\|_{\Ho}.$$ The convergence is a direct consequence of . Moreover, Theorem \[thm:four\] immediately implies that for every sequence satisfying there exists a function $f_1$ so that $$\frac{1}{2\pi}\int_{-\pi}^\pi |(M_*f_1)(\omega)| d\omega =+\infty.$$ Before justifying this statement we point out that it implies the necessity in of $\mu\in(0,1)$. The claim now follows from Theorem \[thm:four\] and the following calculation: $$\begin{aligned}
\max_{t\in\R}\left| \sum_{k=0}^{N_l}\fo(k)\sinpitmk \right|
&=&\max_{t\in\R} \left|\frac{1}{2\pi} \int_{-\pi}^\pi \sum_{k=0}^{N_l}\fo(k)e^{-i\omega k}e^{i\omega t}d\omega\right|\\
&\leq & \frac{1}{2\pi}\int_{-\pi}^\pi \left|\sum_{k=0}^{N_l}\fo(k)\eiok\right|d\omega\\
&\leq & \frac{1}{2\pi}\int_{-\pi}^\pi |(M_*\fo)(\omega)| d\omega.\end{aligned}$$
Reconstruction for Functions of Sine Type {#sec:two}
=========================================
Given a function $f\in\PWpio$ of sine type (Definition \[def:sinetype\]), one defines the invertible transformation $$F(z)=f(z)-A \sin(\pi z),\;\;\;z\in\C\label{eq:F}$$ for a constant $A>\|f\|_{\infty}$. Examples of the use of functions of this type in communications engineering include [@MC81; @Bar74; @Piw83]. We say that $F$ is the function determined by the sine wave crossings of $f$. But more generally, one can generate a function using the zeros of $F$ to reconstruct all functions in $\PWpit$ from their samples as follows. If $F$ is of sine type and $F$ has zeros $\{\tk\}_{k\in\Z}$, then the function $$\label{eq:phi}
\phi(z)=z\lim_{R\rightarrow\infty}\prod_{|\tk|\leq R,\tk\neq 0}\left(1-\frac{z}{\tk}\right)$$ converges uniformly on $|z|\leq R_1$ for all $R_1<\infty$, and $\phi$ is also an entire function of type $\pi$ [@Lev96]. One further defines, for $k\in\Z$, $$\phi_k(z)=\phiz.$$ The sequence $\{\phi_k\}_{k\in\Z}$ is known as the interpolating sequence, and, as just constructed, it is a Riesz basis for $\PWpit$, [@Lev96]. Thus, given this construction, for all $g\in\PWpit$ $$\lim_{N\rightarrow\infty}\left\| g-\sumbiN g(\tk)\phi_k\right\|_{\PWpit}=0.$$ Here we will address the pointwise convergence of such expansions for functions in $\PWpio$. The following theorem of Hryniv and Mykytyuk will be essential for our work with functions of sine type. Their beautiful theorem provides a correspondence between certain zero sequences and elements of $\PWpio$.
\[thm:HM\]
$\;\;$ Let $A>0$ be an arbitrary constant.
- Assume $w\in\PWpio$, set $\tk=k+w(k)$ for $k\in\Z$ and assume $\tk\neq t_l$ for $k\neq l$. Then there exists a function $f\in\PWpio$, $\|f\|_\infty<A$, such that $F$ has the zero sequence $\{\tk\}_{k\in\Z}$.
- If $f\in\PWpio$ and $\|f\|_{\infty}<A$, then there exists a function $w\in \PWpio$ such that $\{\tk\}_{k\in\Z}=\{k+w(k)\}_{k\in\Z}$ is the sequence of zeros of $F$, where $F$ is the function defined by .
With this result we prove the following theorem.
\[thm:two\] Let $f\in\PWpio$, $\|f\|_{\PWpio}<1$ be arbitrary. Let $\{t_k\}_{k\in \Z}$ be the zeros of the function $F$ determined by the sine wave crossings, and let $\{\phi_k\}_{k\in\Z}$ be the corresponding interpolating functions. There exists a function $\fo\in \PWpio$ such that $$\lim_{N\rightarrow \infty}\left(\max_{t\in \R}\left|\sumbiN \fo(t_k)\phik(t)\right|\right)=+\infty.\label{eq:two}$$
Let $g$ be the uniquely determined function in $\PWpio$ satisfying $$t_n=n+g(n)\;\;\;\textnormal{for all}\;\;n\in \N.$$ By the Riemann-Lebesque Lemma, $$\lim_{n\rightarrow\infty}g(n)=0.$$ We set $$C(0)=\max(1,\max_{k}|g(k)|)$$ and define a sequence of numbers by $$C(n)=\max_{|k|\geq n}|g(k)|$$ for $n\in\N$, $n>1$. The sequence $\{C(n)\}_{n=1}^\infty$ is monotonically decreasing and $$\lim_{n\rightarrow \infty}C(n)=0,$$ as well as $$C(n)\geq |g(k)|\;\;\;\textnormal{for all}\;\;|k|\geq n.$$ We will first construct a function $\go\in\PWpio$ and in turn the function $\fo$ that will yield our claim. Now let $\Nh_1$ denote the smallest positive integer satisfying $$4\cdot 2\pi C(0)C(\Nh_1)<1.$$ If $\Nh_k$ has been defined, we let $\Nh_{k+1}$ be the smallest positive integer such that $$4\cdot 2\pi C(0)C(\Nh_{k+1})<\frac{1}{2^{k+1}}$$ and $$\frac{1}{2^{k+1}}\log \Nh_{k+1}>k+1.$$
We now have that $$\go(t)=C(0)w_{N_1}(t)+\sum_{k=2}^\infty\frac{1}{2^{k-1}}w_{\Nh_k}(t)$$ is in $\PWpio$. Now $$\begin{aligned}
\|\go\|_{\PWpio}&<&3C(0)+3\sum_{k=2}^\infty\frac{1}{2^{k+1}}=3\left(C(0)+\sum_{k=2}^\infty\frac{1}{2^k}\right)\\
&=&3(C(0)+1)\leq 4C(0).\end{aligned}$$
We now set $$\Go(\eio)=\sumbiinf \go(k)\eiok,\;\;\;\Fo(\eio)=\Go(e^{i(\omega+\pi)})\;\;\textnormal{for}\;\omega\in[-\pi,\pi)$$ and $$\fho(\omega)=\Fo(\eio),\;\;\textnormal{for}\;\;|\omega|\leq \pi.$$ Lastly, we have $$\fo(t)=\otwopi \inttwopi \fho(\omega)e^{i\omega t}d\omega,\;\;\textnormal{for}\;t\in\R,$$ so that $$\fo(k)=(-1)^k\go(k).$$ Now, $$\begin{aligned}
|\fo(\tn-n)-\fo(n)|&\leq & \|f'_1\|_\infty|\tn-n|\leq \pi\|\fo\|_{\PWpio}|\tn-n|\\
&<& 4\pi C(0)|\tn-n|\leq 4\pi C(0)|g(n)|\\
&\leq & 4\pi C(0)C(n),\end{aligned}$$ which gives us $$|\fo(\tn)-(-1)^n\go(n)|\leq 4\pi C(0)C(n).$$ For every $n$ there is a unique $k_{n}$ such that $|n|\in[\Nh_{k_n},\Nh_{k_n+1}]$. Thus, recalling that $g$ is even, $$\go(n)\geq\frac{1}{2^{k_n+1}}\geq 4\cdot 2\pi C(0)C(n)\geq 4\cdot 2\pi C(0)|g(n)|,$$ so that $$\fo(\tn)=(-1)^n\dn$$ for $$\dn\geq \go(n)-4\pi C(0)C(n)\geq \go(n)-\frac{\go(n)}{2}=\frac{\go(n)}{2}.$$ Setting $\thN$ to be the midpoint of $(\tN,t_{N+1})$, we now have that for all $N>\Nh_1$
$$\begin{aligned}
\sumbiN \fo(t_k)\frac{\phi(\thN)}{\phi'(\tk)(\th_N-t_k)}
=\sum_{|k|\leq N_1}\fo(t_k)\frac{\phi(\thN)}{\phi'(\tk)(\th_N-t_k)}
+\sum_{N_1< |k|\leq N}\fo(t_k)\frac{\phi(\thN)}{\phi'(\tk)(\th_N-t_k)},\label{eq:twoterms}\end{aligned}$$
and it will suffice to address the second term in . We have $$\begin{aligned}
\left|\sum_{N_1< |k|\leq N}\fo(t_k)\frac{\phi(\thN)}{\phi'(\tk)(\th_N-t_k)}\right|
&=&|\phi(\thN)|\cdot \left| \sum_{N_1< |k|\leq N} (-1)^kd_k\frac{1}{(-1)^k\phi'(\tk)(\th_N-t_k)} \right|\\
&=& |\phi(\thN)| \sum_{N_1< |k|\leq N} \frac{d_k}{|\phi'(\tk)|(\th_N-t_k)} \end{aligned}$$ By Lemma 5 of [@BM12], there exists a constant $C_2>0$ such that $\inf_{N\in\Z}|\phi(N+\oh)|\geq C_2$. Since $\phi$ is of sine type, by page 164 of [@Lev96] we have that either $\phi'(t_k)=(-1)^kc_k$ or $\phi'(t_k)=(-1)^{k+1}c_k$ for a sequence of positive constants $\{c_{k}\}_{k\in \Z}$ satisfying $\sup_{k\in\Z}c_k\leq C_3<\infty$ and $0<C_4\leq \inf_{k\in\Z}c_k$. Thus, $$\begin{aligned}
\left|\sum_{N_1< |k|\leq N}\fo(t_k)\frac{\phi(\thN)}{\phi'(\tk)(\th_N-t_k)} \right|
&\geq & \frac{C_2}{C_3}\sum_{N_1< |k|\leq N} \frac{d_k}{(\th_N-t_k)} \\
&\geq & \frac{C_2}{2C_3}\sum_{N_1< |k|\leq N} \frac{\go(k)}{(\th_N-t_k)}.\end{aligned}$$
The separation property of the zeros of a sine type function, see p 163 of [@Lev96], we have that there exists a constant $\delta>0$ such that $$t_{k+1}-t_k>\delta \;\;\textnormal{ for all}\;\;k\in\Z.\label{eq:distdelta}$$ Using and writing $$\th_N-\tk=\th_N-t_N+t_N-t_{N+1}+\ldots+t_{k+1}-\tk,$$ we note that $\th_N-\tk\geq \delta(N+\oh-k)$ for all $-N\leq k\leq N$. We also not that there exists a unique $\kt$ such that $N\in [\Nh_{\kt},\Nh_{\kt+1})$. Using $\go(k)\geq \frac{1}{2^{\kt+1}}$ we have $$\begin{aligned}
\sum_{N_1< |k|\leq N} \frac{\go(k)}{\thN-\tk}
&\geq & \frac{1}{\delta 2^{\kt+1}}\sum_{N_1< |k|\leq N} \frac{1}{N+\oh-k}\nonumber\\
&>& \kt-\log (N_1). \nonumber\end{aligned}$$ Thus, $$\max_{t\in\R}\left|\sum_{N_1\leq|k|\leq N}\fo(t_k)\frac{\phi(t)}{\phi'(\tk)(t-t_k)} \right|\geq \kt-\log N_1$$ so that, recalling that $N_1$ is fixed, $$\lim_{N\rightarrow\infty}\max_{t\in\R}\left|\sumbiN\fo(t_k)\frac{\phi(t)}{\phi'(\tk)(t-t_k)} \right|=+\infty.$$
Reconstruction of the Generating Function {#sec:three}
==========================================
Theorem \[thm:two\] shows that when the interpolating sequence is generated by an arbitrary element of $\PWpio$, there exists another function in $\PWpio$ for which the reconstruction procedure diverges. Our final result addresses the case when one applies the reconstruction procedure to the function used to generate the interpolating sequence. Such a method is generally called reconstruction using sine wave crossings and has been used in communications engineering for several decades [@MC81; @Bar74; @Piw83]. Our result for this reconstruction method is the strengthening of Theorem 5 in [@BM12] from a $\limsup$ statement to a strong limit statement.
We note again that there is no theorem in the spirit of the Banach-Steinhaus Theorem that we could apply to the operator addressed in Theorem \[thm:three\]. The proof relies immediately on the specific properties of the operator.
\[thm:three\] There exists a function $g\in\PWpio$, $\|g\|_{\PWpio}<\oh$ satisfying $$\label{eq:thmthree}
\lim_{N\rightarrow \infty}\max_{t\in\R}\left|\sumbiN (\sin \pi \tk) \pk(t,g)\right|=+\infty,$$ where $\{\phi_k(\cdot,g)\}_{k\in\Z}$ denotes the interpolating sequence generated by $g$.
We have $$\sin \pi\tk=\sin \pi(n+g(k)),$$ so that, retaining $\th_N$ as the midpoint of the interval $(t_N,t_{N+1})$, $$\left|\sumbiN\sin\pi\tk \frac{\phi(\th_N,g)}{\phi'(\tk,g)(\th_N-\tk)}\right|
= |\phi(\th_N,g)|\cdot \left| \sumbiN \frac{\phi(\th_N,g)}{|\phi'(\tk,g)|(\th_N-\tk)}\right|.$$ We may begin with a $g\in\PWpio$ satisfying $\oh\geq g(k)\geq 0$ for all $k\in\Z$. We follow the construction used in the proof of Theorem \[thm:one\] and divide the function $g$ constructed there by a large enough constant to obtain $\|g\|_{\PWpio}\leq \oh$. We of course still have $$\lim_{N\rightarrow\infty}\sumbiN \frac{g(k)}{N+\oh-k}=+\infty.$$ Let $\phi$ be the generating function corresponding to the zeros $\tk=k+g(k)$, $k\in \Z$, and, since $\phi$ is of sine type, recall the bounds on $|\phi(t_N)|$ and $|\phi'(\tk)|$ described in the proof of Theorem \[thm:two\]. Note that $\phi$ depends on $g$, but if $g$ is fixed, then $\phi$ is as well.
With these facts we turn to $$\left|\sumbiN \sin(\pi\tk(g))\frac{\phi(\th_N)}{\phi'(\tk)(\th_N-\tk)} \right|.$$ We have $$\begin{aligned}
\left|\phi(\th_N)\sumbiN \sin(\pi \tk(g))\frac{1}{\phi'(\tk)(\th_N-\tk)} \right|
&\geq & C_1 \left|\sumbiN \frac{(-1)^{k}\sin\pi g(k)}{\phi'(\tk)(\th_N-\tk)} \right|\\
&\geq&\frac{C_1}{C_2}\left|\sumbiN \frac{\sin\pi g(k)}{(\th_N-\tk)} \right|,\end{aligned}$$ for appropriate constants $C_1$ and $C_2$. Note that here we have used that $0\leq \sin\pi g(k)\leq 1$. Using $\sin \pi g(k)\geq g(k)>0$, and the separation property again as in , we have $$\begin{aligned}
\sumbiN\frac{\sin \pi g(k)}{\th_N-\tk}&\geq& \sumbiN\frac{ \pi g(k)}{\th_N-\tk}\\
&\geq& \frac{\pi}{\delta}\sumbiN\frac{g(k)}{N+\oh-k}.\end{aligned}$$ We now use the exact same calculation as to to obtain a lower bound and finish the proof.
Behavior of Oscillations for Theorems \[thm:two\] and \[thm:three\] {#sec:diverge}
===================================================================
In this final section we show that the reconstruction sequences in Theorems \[thm:two\] and \[thm:three\] fluctuate very strongly: not only does the maximum absolute value diverge, but both the maximum and minimum diverge strongly to $+\infty$ and $-\infty$. Following the proof of this theorem, we close the section and the paper with a question.
Assume the hypotheses of Theorem \[thm:two\]. For the function $\fo$ that yields the strong divergence in that theorem, the following hold: $$\label{eq:divpos}
\lim_{N\rightarrow\infty}\left(\max_{t\in\R}\sumbiN \fo(\tk)\phi_k(t)\right)=+\infty$$ and $$\label{eq:divneg}
\lim_{N\rightarrow\infty}\left(\min_{t\in\R}\sumbiN \fo(\tk)\phi_k(t)\right)=-\infty.$$ For the function $g$ that yields divergence in Theorem \[thm:three\], the behavior analogous to and holds for the reconstruction method of Theorem \[thm:three\].
The proof follows immediately from the following technical lemma.
Assume that the generating function $\phi$ is of sine type and set $\phi_k(t)=\frac{\phi(t)}{\phi'(\tk)(\tk-t)}$ for $k\in\Z$. Suppose there exists a real-valued function $f\in\PWpio$ such that $$\label{eq:divabs}
\lim_{N\rightarrow\infty}\max_{t\in\R} \left|\sumbiN f(\tk)\phi_k(t)\right|=+\infty.$$ Then $$\label{eq:divposa}
\lim_{N\rightarrow\infty}\left(\max_{t\in\R}\sumbiN f(\tk)\phi_k(t)\right)=+\infty$$ and $$\label{eq:divnega}
\lim_{N\rightarrow\infty}\left(\min_{t\in\R}\sumbiN f(\tk)\phi_k(t)\right)=-\infty.$$
First note that $f$ is assumed to be real-valued. For each $N$ we must consider two cases.
Case 1.) For a fixed $N$ we assume $$\max_{t\in\R}\left|\sumbiN f(\tk)\phi_k(t)\right|=\sumbiN f(\tk)\phi_k(\tNs)\label{eq:caseone}$$ for a value $\tNs\in\R$. We assume that $\tNs\in(t_{k(N)},t_{k(N)+1})$, and we define $\tNh$ by $$\max_{t\in (t_{k(N)},t_{k(N)+1})}|\phi(t)|=|\phi(\tNh)|$$ and set $$\IN=[-N,N]\backslash \{t_{k(N)},t_{k(N)+1}\}.$$ Recall that $\phi(t)$ has the same sign for all $t$ between any two neighboring zeros. We now have $$\begin{aligned}
A&:=&\frac{1}{\phi(\tNs)}\sumIN f(\tk)\phik(\tNs)-\frac{1}{\phi(\tNs)}\sumIN f(\tk)\phik(\tNh)\\
&=&\sumIN \frac{f(\tk)}{\phi'(\tk)} \left(\frac{1}{\tNs-\tk}-\frac{1}{\tNh-\tk}\right)\\
&=&\sumIN \frac{f(\tk)}{\phi'(\tk)} \left(\frac{\tNh-\tNs}{(\tNs-\tk)(\tNh-\tk)}\right).\end{aligned}$$ As in the proof of Theorem \[thm:three\], we again use the separation properties of $\{\tk\}_{k\in\Z}$, to obtain $$0<\deltab\leq \inf_{k\in\Z}(t_{k+1}-\tk)\;\;\;\textnormal{and}\;\;\;\sup_{k\in\Z}(t_{k+1}-\tk)\leq \deltatop<\infty,$$ as well as the bound $\inf_k|\phi_k'(\tk)|\geq C_4>0$. With these bounds we obtain, $$\begin{aligned}
|A|&\leq&\sumIN \frac{|f(\tk)|}{|\phi'(\tk)|} \frac{|\tNh-\tNs|}{|\tNs-\tk|\cdot|\tNh-\tk|} \\
&\leq& \|f\|_{\PWpio}\frac{\deltatop}{C_4}\sumIN \frac{1}{|\tNs-\tk|\cdot|\tNh-\tk|} \\
&\leq&\|f\|_{\PWpio}\frac{2\deltatop}{C_4(\deltab)^2}\sum_{k=1}^\infty \frac{1}{k^2}\\
&=:&C_5\|f\|_{\PWpio}.\end{aligned}$$ Further, by defining $C_6<\infty$ so that $$\sup_{k\in\Z}\max_{t\in (t_{k},t_{k+1})}|\phi(t)| = C_6,$$ which is again possible by the properties of sine type functions [@Lev96], we have $$\begin{aligned}
\left|\frac{\phi(\tNh)}{\phi(\tNs)}\left(\sumIN f(\tk)\phi_k(\tNs)-\sumIN f(\tk)\phi_k(\tNh)\right)\right|
\leq|\phi(\tNh)|\cdot C_5\|f\|_{\PWpio}\leq C_7 \|f\|_{\PWpio} \end{aligned}$$ for $C_7=C_5\cdot C_6$. Since $\phi(\tNh)/\phi(\tNs)\geq 1$, we have $$\left|\sumIN f(\tk)\phik(\tNh) - \sumIN f(\tk)\phik(\tNs)\right|\leq C_7\|f\|_{\PWpio},$$ and, hence, $$\sumIN f(\tk)\phik(\tNh)\geq \sumIN f(\tk)\phik(\tNs)-C_7 \|f\|_{\PWpio}.$$ Finally, for a further constant $C_8$ we have $$\label{eq:up}
\sumbiN f(\tk)\phik(\tNh)\geq \sumbiN f(\tk)\phik(\tNs)-C_8 \|f\|_{\PWpio}.$$
In the interval $(t_{k(N)+1},t_{k(N)+2})$ the function $\phi(t)$ has the opposite sign as in the interval $(t_{k(N)},t_{k(N)+1})$. We combine this with the fact that there exist constants $C,c>0$ such that for all $k\in\Z$ [@You01], $$c \max_{t\in (t_{k-1},t_{k})}|\phi(t)|\leq \max_{t\in (t_{k},t_{k+1})}|\phi(t)|\leq C\max_{t\in (t_{k+1},t_{k+2})}|\phi(t)|.$$ With this fact, the calculation just given yields the existence of universal constants $C_9,C_{10}>0$, such that $$\label{eq:down}
\sumbiN f(\tk)\phi(\th_{N+1})\leq -C_9\sumIN f(\tk)\phi(\tNh)+C_{10}\|f\|_{\PWpio}.$$
Case 2. denotes the case when holds with a negative sign before the term on the right, and this can be treated in exactly the same manner as Case 1. Thus, by combining and and both cases, we obtain that there exist constants $C_{11},C_{12}>0$ depending only on $\phi$ such that $$\max_{t\in\R} \sumbiN f(\tk)\phi(t) \geq C_{11} \max_{t\in\R}\left| \sumbiN f(\tk)\phi(t)\right| -C_{12}\|f\|_{\PWpio}$$ and $$\min_{t\in\R} \sumbiN f(\tk)\phi(t) \leq -C_{11}\max_{t\in\R}\left| \sumbiN f(\tk)\phi(t)\right| +C_{12}\|f\|_{\PWpio}.$$ Since these inequalities hold for all $N$, the claim follows from the assumed divergence .
We now close by posing a question.
From Theorem \[thm:one\] we have that there exists a function $f_1\in\PWpio$ such that the Shannon series $$(S_N f)(t):=\sumbiN f(k)\frac{\sin \pi(t-k)}{\pi(t-k)}$$ diverges strongly. We also know, however, that $$\lim_{N\rightarrow\infty}\frac{\|S_Nf\|_{\infty}}{\log N}=0$$ for all $f\in\PWpio$. Thus, a natural question is the following: for what monotonically increasing positive functions $\psi:\N\rightarrow \R^+$ does there exist a function $f_1\in\PWpio$ such that $$\lim_{N\rightarrow\infty}\frac{\|S_Nf_1\|_{\infty}}{\psi(N)}=+\infty?$$ Characterizing these functions is an interesting further direction to explore.
ACKNOWLEDGMENT
The authors thank Przemysław Wojtaszczyk and Yuri Lyubarskii for valuable discussions of Conjecture \[conj:complete\] at the Strobl 2011 conference, and Ingrid Daubechies for valuable discussions of questions $Q1)$ and $Q2)$ at Strobl 2011 and at the “Applied Harmonic and Sparse Approximation” workshop at Oberwolfach in 2012. The first author thanks Rudolf Mathar for his insistence in several conversations on the importance of understanding the strong divergence behavior addressed here. The authors also thank the referees of the German Research Foundation (DFG) grant BO 1734/13-2 for highlighting these questions as well in their review.
H. Boche was supported by the German Research Foundation (DFG) through grant BO 1734/13-2. B. Farrell was partially supported by Joel A. Tropp under ONR awards N00014-08-1-0883 and N00014-11-1002 and a Sloan Research Fellowship.
[10]{}
S. Banach and H. Steinhaus. , 9:50–61, 1927.
I. Bar-David. , 24:36–44, 1974.
H. Boche and U. M[ö]{}nich. There exists no globally uniformly convergent reconstruction for the paley-wiener space $\mathcal{PW}^1_{\pi}$ of bandlimited functions sampled at [N]{}yquist rate. , 56(7):3170–3179, 2008.
H. Boche and U. M[ö]{}nich. Behavior of the quantization operator for bandlimited, nonoversampled signals. , 56(5):2433–2440, 2010.
H. Boche and U. M[ö]{}nich. Towards a general theory of reconstruction of bandlimited signals from sine wave crossings. , 92(3):737 – 751, 2012.
J. Brown. , 21:699, 1967.
P. Butzer. Personal communication with the first author, [RWTH]{} [A]{}achen, 1993.
P. Butzer, M. Dodson, P. Ferreira, J. Higgins, O. Lange, P. Seidler, and R. Stens. , 90(3-4):643–688, 2011.
P. Butzer, W. Splettst[ö]{}[ß]{}er, and R. Stens. , 90(1):1–70, 1988.
W. Dickmeis and R. Nessel. , 43:383–393, 1985.
W. Dickmeis, R. Nessel, and E. van Wickeren. , 52:1–20, 1985.
R. P. Feynman. . Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1998.
N. J. Fine. On the [W]{}alsh functions. , 65:372–414, 1949.
J. R. Higgins. , 1996.
R. O. Hryniv and Y. V. Mykytyuk. , 361(4):2207–2223, 2009.
A. Kolmogorov. , 178:303–305, 1926.
B. Levin. , 1996.
B. Logan. , 63:261–285, 1984.
E. Masry and S. Cambanis. Consistent estimation of continuous-time signals from nonlinear transformations of noisy samples. , 27(1):84–96, 1981.
K. Piwnicki. Modulation methods related to sine-wave crossings. , 31(4):503–508, 1983.
E. M. Stein. , 74:140–170, 1961.
H. Voelcker. .
H. Voelcker and A. A. G. Requicha. Clipping and signal determinism: Two algorithms requiring validation. , 21(6):738–744, 1973.
R. M. Young. , 2001.
A. Zygmund. . Cambride University Press, 2002.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- János Kollár
title: 'Esnault-Levine-Wittenberg indices'
---
These are notes of a lecture about the papers [@elw; @wit] and related results of [@mer-deg; @ros-deg; @MR2036596; @MR2587339; @MR3034413]. The papers do not define [*Esnault-Levine-Wittenberg indices,*]{} they focus on the two extreme cases ${\operatorname{elw}}_0(X)$ (traditionally called the [*index*]{}) and ${\operatorname{elw}}_{\dim X}(X)$. In retrospect, the method of [@MR2036596 Sec.5] is equivalent to the computation of ${\operatorname{elw}}_1$ for Del Pezzo surfaces.
The basic properties are at least implicitly in the above papers, with the possible exception of the birational invariance (\[elw.bir.lem\]) and the degree formula (\[rost.prop\]). However, the definition and systematic use of the ELW-indices streamlines several of the arguments.
It is interesting that the proof of the birational invariance does not rely on resolution. Instead, it uses what I call the [*Nishimura–Szabó lemmas*]{} (\[n-sz.lem\]–\[k-sz.lem\]).
H. Esnault, M. Levine and O. Wittenberg answered many of my questions and suggested several improvements. I tried to keep track and attribute specific comments whenever possible. Partial financial support was provided by the NSF under grant number DMS-0968337.
Definition and basic properties
===============================
\[elw.defn\] Let $X$ be a proper scheme defined over a field $k$. For $0\leq i\leq \dim X$ the [*Esnault-Levine-Wittenberg index*]{} ${\operatorname{elw}}_i(X) $ is defined as $${\operatorname{elw}}_i(X):=\bigl(\chi(X, F): \dim F\leq i\bigr)\subset {{\mathbb Z}},
\eqno{(\ref{elw.defn}.1)}$$ where $F$ runs through all coherent sheaves of dimension $\dim F:=\dim{\operatorname{Supp}}F\leq i$ over $X$. I will think of ${\operatorname{elw}}_i(X) $ as an ideal in ${{\mathbb Z}}$; it can be identified with its positive generator. It is convenient to set ${\operatorname{elw}}_{(-1)}(X)=(0) $ and $ {\operatorname{elw}}(X):={\operatorname{elw}}_{\dim X}(X)$. It is clear that $${\operatorname{elw}}_0(X)\subset {\operatorname{elw}}_1(X)\subset\cdots\subset {\operatorname{elw}}_{\dim X}(X).
\eqno{(\ref{elw.defn}.2)}$$
If $X$ has a $k$-point $p\in X(k)$ then $\chi\bigl(X, k(p)\bigr)=1$, hence ${\operatorname{elw}}_0(X)= {\operatorname{elw}}_1(X)=\cdots= {\operatorname{elw}}_{\dim X}(X)={{\mathbb Z}}$. Thus these notions are interesting only if $X$ does not have (or is not known to have) a $k$-point.
Usually $\operatorname{ind}(X):={\operatorname{elw}}_0(X) $ is called the [*index*]{} of $X$. Its generator is the smallest positive degree of a (not necessarily effective) 0-cycle on $X$.
We see in (\[deviss.lem\].1) that if $X$ is integral then $${\operatorname{elw}}(X)=\bigl({\operatorname{elw}}_{\dim X-1}(X),\chi(X, {{\mathcal O}}_X)\bigr).
\eqno{(\ref{elw.defn}.3)}$$ This implies that if $\ell$ is a prime such that ${\operatorname{ord}}_{\ell}{\operatorname{elw}}(X)<{\operatorname{ord}}_{\ell}{\operatorname{elw}}_{\dim X-1}(X)$ then $${\operatorname{ord}}_{\ell}{\operatorname{elw}}(X)={\operatorname{ord}}_{\ell}\chi(X, {{\mathcal O}}_X).
\eqno{(\ref{elw.defn}.4)}$$ This is why the results of [@elw] involving $\chi(X, {{\mathcal O}}_X) $ are equivalent to statements about ${\operatorname{elw}}(X) $.
The above definition makes sense for a proper scheme $X$ defined over a local Artin ring $A$. As a consequence of (\[deviss.lem\].1) we see that ${\operatorname{elw}}_i(X)={\operatorname{elw}}_i({\operatorname{red}}X)$ and ${\operatorname{red}}X$ can be viewed as a scheme over the residue field of $A$. Once the basic results are established over Artin rings, we concentrate on schemes over fields afterwards.
The following is useful in computations.
\[chi.of.sheaf.cor\] Let $X$ be a proper scheme over a local Artin ring $A$ and $F$ a coherent sheaf on $X$. Let $Z_i\subset {\operatorname{Supp}}F$ be the maximal dimensional irreducible components with generic points $\eta_i$. Then $$\chi(X, F)-{\textstyle{\sum}}_i {\operatorname{length}}_{\eta_i}F\cdot\chi\bigl(Z_i, {{\mathcal O}}_{Z_i}\bigr)
\in {\operatorname{elw}}_{\dim F-1}(X).
\eqno{(\ref{chi.of.sheaf.cor}.1)}$$
Proof. The $K$-group of coherent sheaves of dimension $\leq r$ is generated by the ${{\mathcal O}}_Z$ where $Z$ runs through all closed, integral subschemes of dimension $\leq r$. On this group $$F\mapsto \chi(X, F)-{\textstyle{\sum}}_i {\operatorname{length}}_{\eta_i}(F)\cdot\chi\bigl(Z_i, {{\mathcal O}}_{Z_i}\bigr)
\in {{\mathbb Z}}/{\operatorname{elw}}_{r-1}(X)$$ is linear. Thus it is enough to check that it vanishes on the generators. If $\dim Z=r$ then $${\textstyle{\sum}}_i {\operatorname{length}}_{\eta_i}({{\mathcal O}}_Z)\cdot\chi\bigl(Z_i, {{\mathcal O}}_{Z_i}\bigr)=\chi(X, {{\mathcal O}}_Z).$$ If $\dim Z<r$ then ${\textstyle{\sum}}_i {\operatorname{length}}_{\eta_i}(F)\cdot\chi\bigl(Z_i, {{\mathcal O}}_{Z_i}\bigr)=0$ and $\chi(X, F)\in {\operatorname{elw}}_{r-1}(X)$.
\[deviss.lem\] Let $X$ be a proper scheme defined over a local Artin ring $A$. The ELW-indices can also be computed the following ways.
1. ${\operatorname{elw}}_i(X) =\bigl(\chi(Z, {{\mathcal O}}_Z)\bigr)$ where $Z$ runs through all integral subvarieties of dimension $\leq i$.
2. ${\operatorname{elw}}_i(X) =\bigl(\chi(\bar Z, {{\mathcal O}}_{\bar Z})\bigr)$ where $Z$ runs through all integral subvarieties of dimension $\leq i$ and $\bar Z\to Z$ denotes the normalization.
3. ${\operatorname{elw}}_i(X) =\bigl(\chi( Z', {{\mathcal O}}_{Z'})\bigr)$ where $Z$ runs through all integral subvarieties of dimension $\leq i$ and $Z'\to Z$ is any proper birational morphism.
4. If $X$ is regular then ${\operatorname{elw}}(X) =\bigl(\chi(X, E)\bigr)$ where $E$ runs through all locally free sheaves on $X$.
Proof: The $K$-group of coherent sheaves of dimension $\leq i$ is generated by the ${{\mathcal O}}_Z$ where $Z$ runs through all integral subvarieties of dimension $\leq i$. This implies (1). If $X$ is regular then locally free sheaves also generate the $K$-group, giving (4).
Finally (2–3) follow from this, (\[elw.degd.lem\].3) and induction on $i$.
\[elw.degd.lem\] Let $f:X\to Y$ be a morphism of proper $k$-schemes and $F$ a coherent sheaf on $X$. Then
1. $\chi(X,F)\in {\operatorname{elw}}_{r}(Y)$ where $r=\dim f({\operatorname{Supp}}F)$.
2. If $Y$ is integral and $f$ is generically finite then $\chi(X,{{\mathcal O}}_X)-\deg (X/Y)\cdot \chi(Y,{{\mathcal O}}_Y)\in {\operatorname{elw}}_{\dim Y-1}(Y)$.
3. If $f$ is birational then $\chi(X,{{\mathcal O}}_X)- \chi(Y,{{\mathcal O}}_Y)\in {\operatorname{elw}}_{\dim Y-1}(Y)$.
4. If $X,Y$ are normal and $f$ is birational then $\chi(X,{{\mathcal O}}_X)- \chi(Y,{{\mathcal O}}_Y)\in {\operatorname{elw}}_{\dim Y-2}(Y)$.
Proof. By the Leray spectral sequence $\chi(X,F)=\sum (-1)^i \chi(Y, R^if_*F)$ and the latter is in ${\operatorname{elw}}_{r}(Y)$ where $r=\dim f({\operatorname{Supp}}F)$, showing (1).
If $X,Y$ are normal and $f$ is birational then $\dim{\operatorname{Supp}}R^if_*{{\mathcal O}}_X\leq \dim Y-2$ for $i>0$, thus $\sum_{i\geq 1} (-1)^i \chi(Y, R^if_*{{\mathcal O}}_X)\in {\operatorname{elw}}_{\dim Y-2}(Y)$, giving (4).
If $Y$ is integral and $f$ is generically finite then the generic rank of $f_*{{\mathcal O}}_X$ equals the degree of $f$, thus (2) follows from (\[chi.of.sheaf.cor\].1) and (3) is a special case.
Applying (\[elw.degd.lem\].1) to all subvarieties we get the following.
\[elw.morph.lem\] Let $f:X\to Y$ be a morphism of proper $k$-schemes. Then $ {\operatorname{elw}}_i(X)\subset {\operatorname{elw}}_i(Y)$ for every $i$.
Let $f:X\to Y$ be a morphism of proper $k$-schemes and $W\subset Y$ a closed subscheme such that $f$ is an isomorphism over $Y\setminus W$. Then $${\operatorname{elw}}_i(Y)={\operatorname{elw}}_i(X)+{\operatorname{elw}}_i(W).$$
Proof. It is clear that ${\operatorname{elw}}_i(Y)\supset {\operatorname{elw}}_i(X)+{\operatorname{elw}}_i(W)$. Conversely, let $Z\subset Y$ be an integral subvariety. If $Z\subset W$ then set $Z'=Z$ and note that $\chi(Z', {{\mathcal O}}_{Z'})\in {\operatorname{elw}}_i(W)$.
If $Z\not\subset W$ then let $Z'$ be the birational transform of $Z$ in $X$. Thus $\chi(Z', {{\mathcal O}}_{Z'})\in {\operatorname{elw}}_i(X)$. Using (\[elw.degd.lem\].3) we see that ${\operatorname{elw}}_i(Y)\subset {\operatorname{elw}}_i(X)+{\operatorname{elw}}_i(W)$.
A more interesting variant of (\[elw.morph.lem\]) is the following.
\[elw.map.lem\] Let $f:X{\dasharrow}Y$ be a rational map of proper $k$-schemes. Assume that $X$ is regular. Then $ {\operatorname{elw}}_i(X)\subset {\operatorname{elw}}_i(Y)$ for every $i$.
Proof. By induction on $i$. Let $Z\subset X$ be an integral subvariety of dimension $i$. By (\[n-sz.lem\]) we have a birational morphism $Z'\to Z$ and a morphism $Z'\to Y$. Thus $\chi(Z', {{\mathcal O}}_{Z'})\in {\operatorname{elw}}_i(Y)$ and $\chi(Z', {{\mathcal O}}_{Z'})-\chi(Z, {{\mathcal O}}_{Z})\in {\operatorname{elw}}_{i-1}(X)\subset {\operatorname{elw}}_{i-1}(Y)$ by induction. Thus $\chi(Z, {{\mathcal O}}_{Z})\in{\operatorname{elw}}_i(Y)$.
\[elw.bir.lem\] For every $i$, the ELW-index $X\mapsto {\operatorname{elw}}_i(X)$ is a birational invariant of proper, regular $k$-schemes.
An immediate consequence of (\[elw.degd.lem\].1) and (\[elw.map.lem\]) is the following quite useful result.
Let $X$ be a proper, regular $k$-scheme and $g:X{\dasharrow}Y$ a map. If $\chi(X, {{\mathcal O}}_X)\not\in {\operatorname{elw}}_r(Y)$ then $\dim g(X)>r$.
For $i=\dim X$, the following degree formula is in [@mer-deg; @MR2587339; @MR3034413]. (See (\[todd.gives.0.cyc.lem.n-1\]) for its precise relationship to the versions given in [@MR2587339; @MR3034413].)
\[rost.prop\] Let $f:X{\dasharrow}Y$ be a generically finite, rational map of proper $k$-schemes of the same dimension. Assume that $Y$ is integral and regular. Then $$\deg (X/Y)\cdot {\operatorname{elw}}_i(Y)\subset {\operatorname{elw}}_i(X)+{\operatorname{elw}}_{i-1}(Y).$$
Proof. Let $Z\subset Y$ be an integral subvariety of dimension $i$. By (\[k-sz.lem\]) we have a generically finite morphism $Z'\to Z$ of degree $d$ and a morphism $Z'\to X$. Thus $\chi(Z', {{\mathcal O}}_{Z'})\in {\operatorname{elw}}_i(X)$ by (\[elw.degd.lem\].1) and $\chi(Z', {{\mathcal O}}_{Z'})-\deg (X/Y)\cdot \chi(Z, {{\mathcal O}}_{Z})\in {\operatorname{elw}}_{i-1}(Y)
\subset {\operatorname{elw}}_{i-1}(Y)$ by (\[elw.degd.lem\].2). Thus $\deg (X/Y)\cdot \chi(Z, {{\mathcal O}}_{Z})\in{\operatorname{elw}}_i(X)+{\operatorname{elw}}_{i-1}(Y)$.
\[rost.cor\] Notation and assumptions as in (\[rost.prop\]). Fix a prime $\ell$ such that ${\operatorname{ord}}_{\ell}{\operatorname{elw}}_{i-1}(Y)>{\operatorname{ord}}_{\ell}{\operatorname{elw}}_{i}(Y)$ and $\ell{\hspace{-4pt}\not|\hspace{2pt}}\deg (X/Y)$.
Then ${\operatorname{ord}}_{\ell}{\operatorname{elw}}_{i}(X)={\operatorname{ord}}_{\ell}{\operatorname{elw}}_{i}(Y)$.
Cycle class map {#cycle-class-map .unnumbered}
---------------
[ ]{}
The ELW-indices determine how far the Euler characteristic is from being linear on the Chow groups.
For an algebraic cycle $Z=\sum_ia_iZ_i\subset X$ set $$\chi(Z):={\textstyle{\sum}}_ia_i\chi(Z_i, {{\mathcal O}}_{Z_i}){\quad\mbox{and}\quad}
\chi(\bar Z):={\textstyle{\sum}}_ia_i\chi(\bar Z_i, {{\mathcal O}}_{\bar Z_i}),$$ where $\bar Z_i\to Z_i$ denotes the normalization.
If $W\subset X$ is an integral subvariety of dimension $r$ then $\chi(\bar W, {{\mathcal O}}_{\bar W})-\chi(W, {{\mathcal O}}_{W})\in {\operatorname{elw}}_{r-1}(X)$ by (\[elw.degd.lem\].3). Thus $$\chi(\bar Z)-\chi(Z)\in {\operatorname{elw}}_{\dim Z-1}(X).$$
\[chi.on.alg.eq\] Let $B_r(X)$ denote the group of $r$ dimensional cycles in $X$ modulo algebraic equivalence. Then $Z\mapsto \chi(Z)$ and $Z\mapsto \chi(\bar Z)$ define the same well-defined linear map $$\chi: B_r(X)\to {\operatorname{elw}}_r(X)/{\operatorname{elw}}_{r-1}(X).$$
Proof. Two $r$-cycles $Z^1, Z^2$ are algebraically equivalent if there are
1. an irreducible, nonsingular curve $C$ with two points $p_1, p_2\in C(k)$,
2. a flat and proper morphism $g:W\to C$,
3. a morphism $\pi: W\to X$ and
4. an effective $r$-cycle $Z^c$ such that $\pi_*\bigl[g^{-1}(p_i)\bigr]=Z^i+Z^c$ for $i=1,2$.
Set $W^i:=g^{-1}(p_i)$ and let $W^i_j\subset W^i$ be the irreducible components with multiplicities $m^i_j$. By (\[chi.of.sheaf.cor\].1) $$\chi\bigl( W^i, {{\mathcal O}}_{W^i}\bigr)-{\textstyle{\sum}}_j m^i_j\cdot\chi\bigl( W^i_j, {{\mathcal O}}_{W^i_j}\bigr)
\in {\operatorname{elw}}_{r-1}(W)$$ for $i=1,2$. Furthermore, by (\[elw.degd.lem\].2) $\chi\bigl( \pi_*[W^i_j]\bigr)-\chi\bigl( W^i_j\bigr)
\in {\operatorname{elw}}_{r-1}(X)$. Therefore $$\chi\bigl( Z^i\bigr)+\chi\bigl( Z^c\bigr)-
{\textstyle{\sum}}_j m^i_j\cdot\chi\bigl( W^i_j, {{\mathcal O}}_{W^i_j}\bigr)
\in {\operatorname{elw}}_{r-1}(X).$$ Finally $\chi\bigl( W^1, {{\mathcal O}}_{W^1}\bigr)=\chi\bigl( W^2, {{\mathcal O}}_{W^2}\bigr)$ since $g$ is flat.
Let $Q^n\subset {{\mathbb P}}^{n+1}$ be the (empty) quadric $(x_0^2+\cdots+x_{n+1}^2=0)$ over ${{\mathbb R}}$. Note that the quadric $Q^2$ is isomorphic to the product $Q^1\times Q^1$. Let $C_2, C_4\subset Q^1\times Q^1$ be rational curves of bidegrees (1,1) (resp. (1,3)).
Both of these can be viewed as curves in $Q^3$ since $Q^2\subset Q^3$. The cycles $C_4$ and $2\cdot C_2$ have the same degree, hence they are algebraically equivalent over ${{\mathbb C}}$.
However, they are not algebraically equivalent over ${{\mathbb R}}$ since $\chi(C_4)=1$ but $2\chi(C_2)=2$.
This shows that $F\mapsto \chi(X,F)\in {{\mathbb Z}}/{\operatorname{elw}}_{\dim F-1}(X)$ can be viewed as the image of the push-forward map from the connective algebraic K-theory of $X$ to the base-field. This is given by taking $CK^q(X)$ to be the image of $K_0(M^q(X))$ in $K_0(M^{q-1}(X))$, where $M^q(X)$ is the category of coherent sheaves on X with support in codim at least q. This theory was first defined by [@MR2400737]. [@Dai-levine] shows that $CK^*$ is the universal theory with formal group law $u+v-buv \in {{\mathbb Z}}[b][[u,v]]$.
The Nishimura–Szabó lemmas on rational correspondences
======================================================
[ ]{}
[@nis] proved that if a regular $k$-scheme has a $k$-point then any proper $k$-scheme birational to it also has a $k$-point. Around 1992 Endre Szabó found a new short argument. (The proof is reproduced in [@MR1782331-app Prop.A.6] and [@ksc p.183].) Another application of the method was also used in [@MR1782331-app]; its generalization (\[k-sz.lem\]) is needed to prove the degree formula (\[rost.prop\]).
\[n-sz.lem\] Let $f:X{\dasharrow}Y$ be a rational map of proper $k$-schemes. Let $Z\subset X$ be a closed, integral subscheme that is not contained in ${\operatorname{Sing}}X$. Then there is a birational morphism $Z'\to Z$ such that there exists a morphism $f'_Z:Z'\to Y$.
We make no claim about $f'_Z$ beyond its existence. In particular, it can be a constant map. Thus (\[n-sz.lem\]) is interesting only if $Y$ is not known to have a $k$-point.
Proof. By induction on $\dim X$. If $\dim X\leq \dim Z+1$ then $f$ is defined at the generic point of $Z$, hence we can take $Z'$ to be the closure of the graph of $f|_Z$.
If $\dim X> \dim Z+1$, take the blow-up $B_ZX\to X$ and let $E_Z\subset B_ZX$ be the unique irreducible component of the exceptional divisor that dominates $Z$. Note that $E_Z\to Z$ is generically a projective space bundle, hence it has a rational section $Z_1\subset E_Z$ that maps birationally to $Z$. Note that $f$ gives a rational map $f_E:E_Z{\dasharrow}Y$. Induction gives a birational morphism $Z'_1\to Z_1\to Z$ and a morphism $Z'_1\to Y$.
The following is a variant of [@MR1782331-app].
\[k-sz.lem\] Let $f:X{\dasharrow}Y$ be a generically finite, rational map of proper $k$-schemes of the same dimension. Assume that $Y$ is integral. Let $Z\subset Y$ be a closed, integral subscheme that is not contained in ${\operatorname{Sing}}Y$. Then there are
1. a reduced, proper $k$-scheme $Z'$,
2. a generically finite morphism $Z'\to Z$ such that $\deg (Z'/Z)=\deg (X/Y)$ and
3. a morphism $Z'\to X$.
Proof. By induction on $\dim Y$.
Replacing $X$ by the normalization of the closure of the graph of $f$, we may assume that $X$ is normal and $f:X\to Y$ is a morphism.
If $\dim Y\leq \dim Z+1$ then $f$ is finite over the generic point of $Z$. Let $Z_i\subset X$ be the irreducible components of $f^{-1}(Z)$ that dominate $Z$ and $e_i$ the ramification index of $f$ along $Z_i$. Then $\deg (X/Y)=\sum_i e_i\deg(Z_i/Z)$, thus we can take $Z'$ to be the disjoint union of $e_i$ copies of $Z_i$ for every $i$.
If $\dim Y> \dim Z+1$, take the blow-up $B_ZY\to Y$ and let $E_Z\subset B_ZY$ be the unique irreducible component of the exceptional divisor that dominates $Z$. Note that $E_Z\to Z$ is generically a projective space bundle, hence it has a rational section $Z_1\subset E_Z$ that maps birationally to $Z$.
Replace $X$ by the by the normalization of the closure of the graph of $X{\dasharrow}B_ZY$. Let $F_i\subset X$ be the irreducible components of $f^{-1}(E_Z)$ that dominate $E_Z$ and $e_i$ the ramification index of $f$ along $F_i$. By induction, there are $Z'_{1i}\to Z_1$ and morphisms $ Z'_{1i}\to F_i\to X$. Thus we can take $Z'$ to be the disjoint union of $e_i$ copies of $Z'_{1i}$ for every $i$.
The computations of [@wit Lem.5.3] show that the above results also hold if the ambient variety has quotient singularities at the generic point of $Z$ and $k(Z)$ is perfect. (For $\dim Z>0$ this restricts us to characteristic 0.)
Here quotient singularity is understood in the strong sense: it should be Zariski locally a quotient of a regular scheme. The analogous assertion is not true for singularities that are quotients only étale locally. For instance, take $X=(x^2+y^2+z^2=0)\subset {{\mathbb A}}^3_{{{\mathbb R}}}$ and $Z=(0,0,0)$. After blowing up the origin, we get a surface with no real points. On the other hand, over ${{\mathbb C}}$ the singularity is isomorphic to ${{\mathbb C}}^2/(u,v)\sim(-u,-v)$.
Examples
========
Proposition \[chi.on.alg.eq\] makes it relatively easy to compute the ELW-indices when generators of the groups $B_r(X)$ are known.
Let $p$ be a prime and $X$ a nontrivial Severi–Brauer variety of dimension $p-1$. Then $${\operatorname{elw}}_{0}(X)=\cdots={\operatorname{elw}}_{p-2}(X)=p{\quad\mbox{and}\quad} {\operatorname{elw}}_{p-1}(X)=1.$$ To see this note that a Severi–Brauer variety of dimension $n-1$ has an (effective) 0-cycle of degree $n$ and it has a 0-cycle of degree 1 iff it is trivial. Thus ${\operatorname{elw}}_{0}(X)=p$. At the other end, $\chi(X,{{\mathcal O}}_X)=1$ shows that ${\operatorname{elw}}_{p-1}(X)=1 $. Thus the only question is when ${\operatorname{elw}}_i$ drops from $p$ to $1$. Assume that ${\operatorname{elw}}_{i-1}(X)=p $.
Let $K/k$ be a splitting field of degree $p$ and $L_K\subset X_K$ a linear subspace of dimension $i$. Let $L_1,\dots, L_p$ be its conjugates over $k$. Then $Z:=L_1+\cdots + L_p$ is defined over $k$ and it generates the Chow group $A_i(X)$ if $i<\dim X$. $Z$ could be singular, but its normalization $\bar Z$ has Euler characteristic $\chi(\bar Z, {{\mathcal O}}_{\bar Z})=p\chi(L_K, {{\mathcal O}}_{L_K})=p$.
By (\[chi.on.alg.eq\]), $Z\mapsto \chi(\bar Z, {{\mathcal O}}_{\bar Z})\in {{\mathbb Z}}/p$ is a well defined linear map on $A_i(X)$ that vanishes on the generator of $A_i(X)$. Thus ${\operatorname{elw}}_{i}(X)=p $.
(OW) notes that this is also a direct consequence of (\[todd.gives.0.cyc.lem.n-1\]).
Next we compute the ELW-indices for certain products of general curves.
\[prod.curves.prop\] Let $k$ be a field and $C_1,\dots, C_n$ smooth, irreducible, projective curves over $k$. Assume that
1. ${\operatorname{elw}}_0(C_i)\subset (m)$ for every $i$ for some $m\geq 1$ and
2. ${\operatorname{Pic}}\bigl(C_1\times\cdots\times C_n)
=\pi_1^*{\operatorname{Pic}}(C_1)+\cdots +\pi_n^*{\operatorname{Pic}}(C_n)$ where $\pi_i$ denotes the $i$-th coordinate projection.
Then ${\operatorname{elw}}_i\bigl(C_1\times\cdots\times C_n)\subset (m)$ for $i<n$.
Proof. Set $X_r:=C_1\times\cdots\times C_r$. The proof is by induction on $r$.
Let $F$ be a coherent sheaf on $X_n$. Consider the coordinate projection $\Pi_n: X_n\to X_{n-1}$. If $\dim F\leq n-2$ then $\dim \Pi_n({\operatorname{Supp}}F)\leq n-2$, hence, by (\[elw.degd.lem\].1) and induction, $\chi(X_n, F)\in {\operatorname{elw}}_{n-2}\bigl(X_{n-1}\bigr)\subset (m)$.
We are left with the case when $\dim F=n-1$. By (\[chi.of.sheaf.cor\].1) it is enough to show that $\chi(D, {{\mathcal O}}_D)\in (m)$ for every effective divisor $D\subset X_n$. Using the exact sequence $$0\to {{\mathcal O}}_{X_n}(-D)\to {{\mathcal O}}_{X_n}\to {{\mathcal O}}_D\to 0$$ we are reduced to proving that $$\chi(X_n, L)\equiv \chi(X_n, {{\mathcal O}}_{X_n})\mod m$$ for every line bundle $L$ on $X_n$. By assumption (2), there are line bundles $L_i$ on $C_i$ such that $L\cong \otimes_i\pi_i^*L_i$. Therefore $$\chi(X_n, L)=\textstyle{\prod}_i \chi(C_i, L_i)=
\textstyle{\prod}_i\bigl(\chi(C_i, {{\mathcal O}}_{C_i})+\deg L_i\bigr).$$ By assumption (1), $\deg L_i\in (m)$ for every $i$, thus $$\chi(X_n, L)\equiv \textstyle{\prod}_i\ \chi(C_i, {{\mathcal O}}_{C_i})=
\chi(X_n, {{\mathcal O}}_{X_n})\mod m.\qed$$
In applications the tricky part is to check the condition (\[prod.curves.prop\].2). Let us start over algebraically closed fields.
\[pic.of.prod.lem\] Let $Y_i$ be normal, irreducible, proper varieties over an algebraically closed field $k$. The following are equivalent.
1. ${\operatorname{Pic}}\bigl(Y_1\times\cdots\times Y_n)
=\pi_1^*{\operatorname{Pic}}(Y_1)+\cdots +\pi_n^*{\operatorname{Pic}}(Y_n)$ where $\pi_i$ denotes the $i$th coordinate projection.
2. For $i\neq j$, every morphism $Y_i\to {\operatorname{Pic}}(Y_j)$ is constant.
3. For $i\neq j$, every morphism ${\operatorname{Pic}}^0(Y_i)\to {\operatorname{Pic}}^0(Y_j)$ is constant.
The above conditions hold if the ${\operatorname{Pic}}^0(Y_i)$ are sufficiently general and independent, but they may be hard to check in concrete situations. Consider the case when $Y_i=C_i$ are smooth curves. For very general curves ${\operatorname{Pic}}^0(C_i)$ is a simple Abelian variety; see [@MR0453754; @MR529440; @MR1748293; @MR2040573] for explicit examples over ${{\mathbb Q}}$. Hence if, in addition, the $C_i$ all have different genera, then (\[pic.of.prod.lem\].3) holds. The same for sufficiently general and independent curves of the same genus $>0$.
Over arbitrary fields, an extra complication comes from the Brauer group.
\[brauer.map.say\] Let $Y$ be normal, irreducible, proper variety over a perfect field $k$ with algebraic closure $\bar k$. Let $L$ be a line bundle on $Y_{\bar k} $ that is isomorphic to its Galois conjugates. Equivalently, $[L]\in {\operatorname{Pic}}\bigl(Y_{\bar k}\bigr)(k)$. If the linear system $|L|$ is nonempty, then, as a subscheme of ${\operatorname{Hilb}}(Y)$; it is defined over $k$ and is isomorphic to a projective space over $\bar k$. Thus $|L|$ defines an element of the Brauer group $ \operatorname{Br}(k) $. This gives an exact sequence $${\operatorname{Pic}}(Y)\to {\operatorname{Pic}}\bigl(Y_{\bar k}\bigr)(k)
\stackrel{\operatorname{br}_Y}{\longrightarrow} \operatorname{Br}(k).$$ (See [@blr Chap.8] for a more conceptual construction.) It is clear that if $[L_i]\in {\operatorname{Pic}}\bigl((Y_i)_{\bar k}\bigr)(k)$ then $$\operatorname{br}_{\prod Y_i}\bigl(\otimes_i\pi_i^*L_i\bigr)=
{\textstyle{\sum}}_i \operatorname{br}_{Y_i}\bigl(L_i).$$
\[pic.of.prod.lem.2\] Let $Y_1,\dots, Y_n$ be normal, irreducible, proper varieties over a perfect field $k$. Assume that they satisfy the equivalent conditions (\[pic.of.prod.lem\].1–3) over $\bar k$. The following are equivalent.
1. ${\operatorname{Pic}}\bigl(Y_1\times\cdots\times Y_n)
=\pi_1^*{\operatorname{Pic}}(Y_1)+\cdots +\pi_n^*{\operatorname{Pic}}(Y_n)$.
2. The subgroups ${\operatorname{im}}\bigl[\operatorname{br}_{Y_i}: {\operatorname{Pic}}\bigl((Y_i)_{\bar k}\bigr)(k)
\to \operatorname{Br}(k)\bigr]$ are independent in $\operatorname{Br}(k)$. (Subgroups $A_i$ of an Abelian group $A$ are [*independent*]{} if $a_i\in A_i$ and $\sum a_i=0$ implies that $a_i=0$ for every $i$.)
Let $C_i\subset {{\mathbb P}}^2$ be plane conics over a perfect field $k$. Set $Y_n=C_1\times\cdots\times C_n$. The image of $\operatorname{br}_{C_i} $ equals the subgroup generated by $C_i$ in $\operatorname{Br}(k)$. Thus we see that the following are equivalent.
1. ${\operatorname{elw}}_{n-1}(Y_n)\subset (2)$ and
2. the classes $[C_i]\in \operatorname{Br}(k)$ are nonzero and independent.
If we are over ${{\mathbb Q}}$ then any finite collection of conics has a point in a quadratic extension. Thus we obtain the following.
Set $C_i:=(x^2+y^2=p_iz^2)$ where the $p_i$ are distinct primes congruent to $3$ modulo $4$ and $Y_n:=C_1\times\cdots\times C_n$. Then $${\operatorname{elw}}_{0}(Y_n)=\cdots ={\operatorname{elw}}_{n-1}(Y_n)=(2)
{\quad\mbox{and}\quad} {\operatorname{elw}}_{n}(Y_n)=(1).$$
We work over the field $k={{\mathbb C}}(t)$. A hyperelliptic curve over $k$ is given by an equation $$C:=\bigl(z^2=f(x,y)\bigr)\subset {{\mathbb P}}^2(1,1,m)$$ where $f(x,y)\in {{\mathbb C}}(t)[x,y]$ is homogeneous of degree $2m$. We will look at it as a cover of ${{\mathbb P}}^1\times {{\mathbb P}}^1$ where ${{\mathbb A}}^2$ (with coordinates $t,x/y$) is an affine chart. Thus we have a double cover $\pi:S\to {{\mathbb P}}^1\times {{\mathbb P}}^1$ with branch curve $B\subset {{\mathbb P}}^1\times {{\mathbb P}}^1$. For very general $S$, $$\pi^*{\operatorname{Pic}}\bigl ({{\mathbb P}}^1\times {{\mathbb P}}^1\bigr)\to {\operatorname{Pic}}(S)
{\quad\mbox{is an isomorphism;}\quad}$$ see [@MR699163; @rav-sri]. This implies that, for very general $f$, ${\operatorname{Pic}}(C)={{\mathbb Z}}\bigl[\pi^*{{\mathcal O}}_{{{\mathbb P}}^1}(1)\bigr]$. In particular, if $m$ is odd then ${\operatorname{elw}}_0(C)=2$ and ${\operatorname{elw}}_0(C)=1$.
For stronger examples over ${{\mathbb Q}}$, see [@MR2221085; @MR2578467].
Note that ${\operatorname{Pic}}(C)$ can be viewed as the generic fiber of the family of the Picard varieties of the fibers of $p_1\circ \pi:S\to {{\mathbb P}}^1$ where $p_1$ is the first coordinate projection of ${{\mathbb P}}^1\times {{\mathbb P}}^1$. The singular fibers correspond to the branch points of $p_1:B\to {{\mathbb P}}^1$.
Let us now choose our curves $C_i$ such that the branch points of the corresponding $p_1:B_i\to {{\mathbb P}}^1$ are all different. Then the different ${\operatorname{Pic}}^0(C_i)$ satisfy the condition (\[pic.of.prod.lem\].3). The Brauer group of ${{\mathbb C}}(t)$ is trivial. Thus if $Y_n:=C_1\times\cdots\times C_n$ is the product of such generic curves then $${\operatorname{elw}}_{0}(Y_n)=\cdots ={\operatorname{elw}}_{n-1}(Y_n)=(2)
{\quad\mbox{and}\quad} {\operatorname{elw}}_{n}(Y_n)=(1).$$
Let $C$ be a geometrically irreducible smooth real curve. The interesting case is when $C$ has no real points and even genus. Equivalently, when ${\operatorname{elw}}_0(C)=(2)$ and ${\operatorname{elw}}_1(C)=(1)$. Note that $\operatorname{Br}({{\mathbb R}})={{\mathbb Z}}/2$ and the Brauer map ${\operatorname{br}_C}: {\operatorname{Pic}}\bigl(C_{{{\mathbb C}}}\bigr)({{\mathbb R}})
\to \operatorname{Br}({{\mathbb R}})={{\mathbb Z}}/2$ is surjective, essentially by a result that goes back to [@witt-br] (see also [@MR0412193]). Thus we see that for any product $Y_n$ of such curves we have $${\operatorname{elw}}_{0}(Y_n)=(2) {\quad\mbox{and}\quad}{\operatorname{elw}}_{1}(Y_n)= \cdots =
{\operatorname{elw}}_{n}(Y_n)=(1).$$
The sequence of ELW-indices
===========================
There are several general questions about the ELW-indices that should be explored. Esnault asks about all possible sequences ${\operatorname{elw}}_0(X),\dots, {\operatorname{elw}}_{\dim X}(X)$.
\[mu.td.n\] Let ${\mu(\operatorname{Td}_{n})}$ denote the denominator appearing in the Todd class in dimension $n$. By [@hirz 1.7.3], it is $${\mu(\operatorname{Td}_{n})}=\textstyle{\prod}_p\ p^{[n/(p-1)]}{\quad\mbox{where $p\leq n+1$ is a prime.}\quad}
\eqno{(\ref{mu.td.n}.1)}$$ The sequence starts as ${\mu(\operatorname{Td}_{1})}=2$, ${\mu(\operatorname{Td}_{2})}=12$, ${\mu(\operatorname{Td}_{3})}=24$, ${\mu(\operatorname{Td}_{4})}=720$.
Note that ${\mu(\operatorname{Td}_{n})}$ is very close to $n!$. Indeed, we can rewrite the formula as $${\mu(\operatorname{Td}_{n})}=\textstyle{\prod}_p\ p^{[\frac{n}{p}+\frac{n}{p^2}+\cdots]}
{\quad\mbox{while}\quad}
n!=\textstyle{\prod}_p\ p^{[\frac{n}{p}]+[\frac{n}{p^2}]+\cdots}.$$ Each ${\mu(\operatorname{Td}_{n})}$ divides any later ${\mu(\operatorname{Td}_{m})}$. There is also the more delicate relation $$n!\cdot {\mu(\operatorname{Td}_{m})}\ \big| \ {\mu(\operatorname{Td}_{n+m-1})}.
\eqno{(\ref{mu.td.n}.2)}$$
Without much evidence, let me propose the following.
\[HE.conj\] Let $e_0,\dots, e_n$ be a sequence of natural numbers. Then there is a field $k$ and a $k$-scheme (or smooth $k$-variety) $X$ of dimension $n$ such that ${\operatorname{elw}}_r(X)=(e_r)$ for every $r$ iff the following hold.
1. $e_{r+1}|e_r$ for every $r$ and
2. $e_0| {\mu(\operatorname{Td}_{r})}\cdot e_r$ for every $r$.
Next we discuss some evidence supporting the above formulation of (\[HE.conj\]). The necessity of $e_{r+1}|e_r$ follows from (\[elw.defn\].2). The divisibility $e_0| {\mu(\operatorname{Td}_{r})}\cdot e_r$ is more subtle and I do not know a complete proof.
\[todd.gives.0.cyc.lem\] Let $X$ be a proper $k$-scheme over a field of characteristic 0. Then $${\mu(\operatorname{Td}_{r})}\cdot {\operatorname{elw}}_{r}(X) \subset {\operatorname{elw}}_0(X).
\eqno{(\ref{todd.gives.0.cyc.lem}.1)}$$
Proof. If $W$ is smooth, proper and of dimension $r$ then Riemann–Roch says that ${\mu(\operatorname{Td}_{r})}\cdot \chi(W, {{\mathcal O}}_W)$ is a polynomial in the Chern classes. In particular, $W$ has a 0-cycle of degree ${\mu(\operatorname{Td}_{r})}\cdot \chi(W, {{\mathcal O}}_W)$.
By (\[deviss.lem\].3), ${\operatorname{elw}}_{r}(X) $ is generated by the $\chi( Z', {{\mathcal O}}_{Z'})$ where $Z$ runs through all integral subvarieties of dimension $\leq r$ and $Z'\to Z$ is any resolution.
The only missing ingredient in positive characteristic is resolution of singularities. The method of [@elw] shows that one can use de Jong’s alterations [@deJ-alt; @MR1450427] (in the stronger form proved by Gabber, cf. [@2012arXiv1207.3648I Exp.IX]) to prove that (\[todd.gives.0.cyc.lem\].1) holds in ${{\mathbb Z}}[p^{-1}]$.
A more interesting part of (\[HE.conj\]) is the claim that there are no additional relations between the ${\operatorname{elw}}_i$. I do not even have a plausible argument, but the following example suggests that, for $n=\dim X$, there should not be any relations between ${\operatorname{elw}}_n(X)$ and ${\operatorname{elw}}_{n-1}(X)$ in general.
Let $X$ be a smooth projective variety of dimension $n$ such that ${\operatorname{Pic}}(X)={{\mathbb Z}}[H]$ and the intersection number $(C\cdot H)$ is divisible by $m\cdot {\mu(\operatorname{Td}_{n-1})}$ for some fixed number $m$ for every curve $C\subset X$. Such examples are K3 surfaces (where $m=(H^2)/2$ can take any value) or hypersurfaces of very high degree in ${{\mathbb P}}^4$ [@trento-1].
By Riemann–Roch, $\chi(X,L)-\chi(X,{{\mathcal O}}_X)$ is divisible by $m$ for any line bundle $L$, thus $m\ | \ \chi(D, {{\mathcal O}}_D)$ for every divisor $D\subset X$. Thus (\[elw.degd.lem\].2) implies that $${\operatorname{elw}}_{n-1}(X)\subset \bigl({\operatorname{elw}}_{n-2}(X), m\bigr).$$ Although I do not know if there are further relations between ${\operatorname{elw}}_n(X)$ and ${\operatorname{elw}}_{n-2}(X)$, the above computations suggest that once ${\operatorname{elw}}_n(X)$ and ${\operatorname{elw}}_{n-2}(X)$ are set, ${\operatorname{elw}}_{n-1}(X)$ could be any ideal satisfying ${\operatorname{elw}}_n(X)\supset{\operatorname{elw}}_{n-1}(X)\supset{\operatorname{elw}}_{n-2}(X)$.
An extreme case of (\[HE.conj\]) would be the following.
Find $n$-dimensional smooth, projective varieties such that ${\operatorname{elw}}(X)=1$ yet ${\operatorname{elw}}_i(X)=\bigl({\mu(\operatorname{Td}_{n})}\bigr)$ for $0\leq i<n$.
The higher dimensional examples of [@trento-1] are not convincing for the current purposes, but the following should be possible to prove.
Given $e_0,e_1, e_2$, there is K3 surface $S$ over any field (or even over ${{\mathbb Q}}$) such that ${\operatorname{elw}}_r(S)=(e_r)$ for $r=0,1,2$ iff
1. $e_2 \ | \ e_1\ | \ e_0 $
2. $e_2=\gcd(2, e_1)$
3. $\ e_0\ | \ 12 e_2$ and $\ e_0\ | \ 2 e_1$.
Another case when (\[todd.gives.0.cyc.lem\].1) is known connects our index formula (\[rost.prop\]) with the version in [@MR3034413]. The proof given in [@MR3034413] relies on [@MR1936372]; the claim also follows from properties of the numbers ${\mu(\operatorname{Td}_{r})}$.
\[todd.gives.0.cyc.lem.n-1\] Let $X$ be a proper, nonsingular variety of dimension $n$. Then $${\mu(\operatorname{Td}_{n-1})}\cdot {\operatorname{elw}}_{n-1}(X) \subset {\operatorname{elw}}_0(X).$$
Proof. Let $F$ be a coherent sheaf of dimension $\leq n-1$ on $X$. Since $X$ is nonsingular, there are two vector bundles $E_1, E_2$ such that $[F]=[E_1]-[E_2]$ in $K_0(X)$. Thus, by Riemann–Roch, $$\chi(X, F)=\int_X\operatorname{ch}(E_1)\cdot\operatorname{td}(T_X)-
\int_X\operatorname{ch}(E_2)\cdot\operatorname{td}(T_X).$$ Since ${\operatorname{rank}}E_1={\operatorname{rank}}E_2$, the terms ${\operatorname{rank}}E_1\cdot \operatorname{td}_n(T_X)$ and ${\operatorname{rank}}E_2\cdot \operatorname{td}_n(T_X)$ cancel each other.
All other terms have denominators dividing $(n-r-1)!\cdot {\mu(\operatorname{Td}_{r})}$ for some $r\leq n-1$. By (\[mu.td.n\].2) these all divide ${\mu(\operatorname{Td}_{n-1})}$.
ELW-indices for special fields
==============================
We consider various fields for which the sequences ${\operatorname{elw}}_0(X),\dots, {\operatorname{elw}}_{\dim X}(X)$ are special for all schemes, or at least for certain interesting classes of varieties.
Finite fields {#finite-fields .unnumbered}
-------------
[ ]{}
The following is proved in [@wit].
Let $X$ be a proper scheme over a finite field $k$. Then $${\operatorname{elw}}_0(X)= \cdots= {\operatorname{elw}}_{\dim X}(X)=
\bigl( h^0(\bar Z, {{\mathcal O}}_{\bar Z}): Z\subset X
\mbox{ and $Z$ is integral}\bigr).\qed$$
The proof is a combination of two lemmas.
\[H0.as.intrscetion.lem\] Let $X$ be a proper $k$-scheme. Then $${\operatorname{elw}}(X)\subset \bigl( h^0(\bar Z, {{\mathcal O}}_{\bar Z}): Z\subset X
\mbox{ and $Z$ is integral}\bigr),$$ where $Z$ runs through all closed, integral subschemes of $X$ and $\bar Z\to Z$ denotes the normalization.
Proof. Let $W$ be a proper, integral scheme and $F$ a coherent sheaf over $W$. Then $H^i(W, F)$ is a vector space over the field $H^0(W, {{\mathcal O}}_W) $, hence its dimension is divisible by $h^0(W, {{\mathcal O}}_W) $. Thus $\chi(W, F)$ is divisible by $h^0(W, {{\mathcal O}}_W) $. Applying this to $W=\bar Z$ we get that each $\chi (\bar Z, {{\mathcal O}}_{\bar Z})$ is contained in $\bigl( h^0(\bar Z, {{\mathcal O}}_{\bar Z}): Z\subset X\bigr) $. By (\[deviss.lem\].2) this implies our claim.
\[ff.0.cycle\] Let $W$ be a proper, normal, integral variety over a finite field $k$. Then ${\operatorname{elw}}_0(W)=h^0(W, {{\mathcal O}}_W)$.
Proof. Assume first that $ h^0(W, {{\mathcal O}}_W)=k$. Then $W$ is geometrically integral. If $\dim W=1$ then by Weil there are points in any large enough field extension; take two whose degrees are relatively prime.
If $\dim W>1$ then use Bertini to get a geometrically integral hyperplane section. More precisely, such hyperplane sections exists over any large enough field extension; take two whose degrees are relatively prime.
In general, set $K=H^0(W, {{\mathcal O}}_W)$. After base change to $K$, the irreducible components $W_i\subset W_K$ are geometrically integral. Thus there is a 0-cycle $Z_1\subset W_1$ of degree 1. The sum of its conjugates gives a 0-cycle of degree $\dim_kK$ on $W$.
Henselian fields with algebraically closed residue fields {#henselian-fields-with-algebraically-closed-residue-fields .unnumbered}
---------------------------------------------------------
[ ]{}
One of the main questions proposed and investigated in [@elw] is the following.
\[elw.conj\] Let $K$ be the quotient field of an excellent, Henselian DVR with algebraically closed residue field $k$. Let $X$ be a proper $K$-scheme. Then $${\operatorname{elw}}_0(X)= {\operatorname{elw}}_1(X)=\cdots= {\operatorname{elw}}_{\dim X}(X).\eqno{(\ref{elw.conj}.1)}$$
The conjecture is almost proved in [@elw].
Let $K$ be the quotient field of an excellent, Henselian DVR with algebraically closed residue field $k$. Let $X$ be a proper $K$-scheme. Then
1. If ${\operatorname{char}}k=0$ then ${\operatorname{elw}}_0(X)= \cdots= {\operatorname{elw}}_{\dim X}(X)$.
2. If ${\operatorname{char}}k=p>0$ then ${\operatorname{elw}}_0(X)= \cdots= {\operatorname{elw}}_{\dim X}(X)$ holds in ${{\mathbb Z}}[p^{-1}]$.
The key step is the following.
Let $R$ be a Henselian DVR with quotient field $K$ and algebraically closed residue field $k$. Let $X_R$ be a proper, regular $R$-scheme with generic fiber $X_K$. Then $\chi(X_K, {{\mathcal O}}_{X_K})\in {\operatorname{elw}}_0(X_K)$.
Proof. Write $X_0=\sum_{i\in I} m_i X^i_0$. If $r|m_i$ for every $i$ then set $Z_r:=\sum (m_i/r) X^i_0$. Note that $${{\mathcal O}}_X(-iZ_r)^r|_{Z_r}\cong {{\mathcal O}}_X(-irZ_r)|_{Z_r}\cong{{\mathcal O}}_{Z_r}.$$ Thus ${{\mathcal O}}_X(-iZ_r)|_{Z_r}$ is numerically trivial, hence $\chi\bigl(Z_r, {{\mathcal O}}_X(-iZ_r)|_{Z_r}\bigr)=\chi\bigl(Z_r, {{\mathcal O}}_{Z_r}\bigr) $. Therefore $$\chi(X_K, {{\mathcal O}}_{X_K})=\chi(X_0, {{\mathcal O}}_{X_0})=
{\textstyle{\sum}}_{i=1}^r\chi\bigl(Z_r, {{\mathcal O}}_X(-iZ_r)|_{Z_r}\bigr)=
r\chi\bigl(Z_r, {{\mathcal O}}_{Z_r}\bigr).$$ This implies that $\chi(X_K, {{\mathcal O}}_{X_K})\in (m_i:i\in I)$. We conclude by noting that through a general point of $X^i_0$ there is a multi-section of degree $m_i$.
Real closed fields {#real-closed-fields .unnumbered}
------------------
[ ]{}
For any scheme $X$ over ${{\mathbb R}}$, ${\operatorname{elw}}_0(X)=1$ iff $X({{\mathbb R}})\neq \emptyset$. Otherwise ${\operatorname{elw}}_0(X)=2$. Thus the only question is when the sequence ${\operatorname{elw}}_i$ drops form 2 to 1.
Let $\pi:S\to {{\mathbb P}}^2$ be a double cover ramified along a curve of degree $2d$. Let $H$ denote the pull-back of a line in ${{\mathbb P}}^2$. Then $\pi_*{{\mathcal O}}_S\cong {{\mathcal O}}_{{{\mathbb P}}^2}+{{\mathcal O}}_{{{\mathbb P}}^2}(-d)$, thus $$\chi(S, {{\mathcal O}}_S)=1+\tfrac{(d-1)(d-2)}{2}
{\quad\mbox{and}\quad} K_S\sim (d-3)H.$$ If $C\sim rH$ is a curve in $S$ then $$\chi(C, {{\mathcal O}}_C)=r(r+d-3)\tfrac{(H^2)}{2}=r(r+d-3).$$ Thus if $S({{\mathbb R}})=\emptyset$, $d\equiv 2 \mod 4$ and ${\operatorname{Pic}}(S)={{\mathbb Z}}[H]$ then $${\operatorname{elw}}_0(S)=2, \ {\operatorname{elw}}_1(S)=2,\ {\operatorname{elw}}_2(S)=1.$$ Such surfaces can be obtained as small perturbations of $$S_0:=\bigl(x^{12}+y^{12}+z^{12}+w^{2}=0\bigr)\subset {{\mathbb P}}^3(1,1,1,6).$$
Probably there are similar higher dimensional examples. It is, however, quite difficult to understand all subvarieties of codimension $\geq 2$ of a given variety. I do not know how to compute the ELW-indices for higher dimensional hypersurfaces.
\[RC.over.R.ques\] Let $X$ be a smooth rationally connected variety over ${{\mathbb R}}$. Is ${\operatorname{elw}}_1(X)=1$?
It is known that if $X$ is rationally connected and $X({{\mathbb R}})\neq \emptyset$ then it contains a rational curve [@MR1715330 1.7]. It is conjectured that $X$ contains a geometrically rational curve even if $X({{\mathbb R}})= \emptyset$; see [@ar-ko Rem.20]. (\[RC.over.R.ques\]) is a weaker variant of it. This is closely related to the conjecture that the function field of the empty real conic is $C_1$; see [@MR0053924 p.379].
\#1[0=]{} \#1[0=]{} \[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{}
[[Wit]{}13]{}
Carolina Araujo and J[á]{}nos Koll[á]{}r, *Rational curves on varieties*, Higher dimensional varieties and rational points ([B]{}udapest, 2001), Bolyai Soc. Math. Stud., vol. 12, Springer, Berlin, 2003, pp. 13–68.
Siegfried Bosch, Werner L[ü]{}tkebohmert, and Michel Raynaud, *Néron models*, Ergebnisse der Mathematik und ihrer Grenzgebiete (3) \[Results in Mathematics and Related Areas (3)\], vol. 21, Springer-Verlag, Berlin, 1990.
Alexandru Buium, *Sur le nombre de [P]{}icard des revêtements doubles des surfaces algébriques*, C. R. Acad. Sci. Paris Sér. I Math. **296** (1983), no. 8, 361–364.
Shuang Cai, *Algebraic connective [$K$]{}-theory and the niveau filtration*, J. Pure Appl. Algebra **212** (2008), no. 7, 1695–1715.
Jean-Louis Colliot-Th[é]{}l[è]{}ne and David A. Madore, *Surfaces de del [P]{}ezzo sans point rationnel sur un corps de dimension cohomologique un*, J. Inst. Math. Jussieu **3** (2004), no. 1, 1–16.
A. Johan de Jong, *Smoothness, semi-stability and alterations*, Inst. Hautes Études Sci. Publ. Math. (1996), no. 83, 51–93.
, *Families of curves and alterations*, Ann. Inst. Fourier (Grenoble) **47** (1997), no. 2, 599–621.
F. R. Demeyer and M. A. Knus, *The [B]{}rauer group of a real curve*, Proc. Amer. Math. Soc. **57** (1976), no. 2, 227–232.
S. [Dai]{} and M. [Levine]{}, *[Connective algebraic K-theory]{}*, ArXiv e-prints (2012).
H. [Esnault]{}, M. [Levine]{}, and O. [Wittenberg]{}, *[Index of varieties over Henselian fields and Euler characteristic of coherent sheaves]{}*, ArXiv e-prints (2012).
Olivier Haution, *Degree formula for the [E]{}uler characteristic*, Proc. Amer. Math. Soc. **141** (2013), no. 6, 1863–1869.
F. Hirzebruch, *Neue topologische [M]{}ethoden in der algebraischen [G]{}eometrie*, Ergebnisse der Mathematik und ihrer Grenzgebiete (N.F.), Heft 9, Springer-Verlag, Berlin, 1956.
L. [Illusie]{}, Y. [Laszlo]{}, and F. [Orgogozo]{}, *[Travaux de Gabber sur l’uniformisation locale et la cohomologie etale des schemas quasi-excellents. Seminaire a l’Ecole polytechnique 2006–2008]{}*, ArXiv e-prints (2012).
J[á]{}nos Koll[á]{}r, *Trento examples–1*, Classification of irregular varieties (Trento, 1990), Lecture Notes in Math., vol. 1515, Springer, Berlin, 1992, pp. 136–139.
, *Rationally connected varieties over local fields*, Ann. of Math. (2) **150** (1999), no. 1, 357–367.
János Kollár and Endre Szabó, *Fixed points of group actions and rational maps (appendex to: [R]{}eichstein, [Z.]{} and [Y]{}oussin, [B.]{}: [E]{}ssential dimensions of algebraic groups and a resolution theorem for [$G$]{}-varieties)*, Canad. J. Math. **52** (2000), no. 5, 1054–1056.
J[á]{}nos Koll[á]{}r, Karen E. Smith, and Alessio Corti, *Rational and nearly rational varieties*, Cambridge Studies in Advanced Mathematics, vol. 92, Cambridge University Press, Cambridge, 2004.
Serge Lang, *The theory of real places*, Ann. of Math. (2) **57** (1953), 378–391.
Alexander Merkurjev, *Degree formula*, http://www.math.uni-bielefeld.de/ rost/degree-formula.html, 2000.
, *Algebraic oriented cohomology theories*, Algebraic number theory and algebraic geometry, Contemp. Math., vol. 300, Amer. Math. Soc., Providence, RI, 2002, pp. 171–193.
Shigefumi Mori, *The endomorphism rings of some [A]{}belian varieties*, Japan. J. Math. (N.S.) **2** (1976), no. 1, 109–130.
, *The endomorphism rings of some abelian varieties. [II]{}*, Japan. J. Math. (N.S.) **3** (1977), no. 1, 105–109.
Hajime Nishimura, *Some remarks on rational points*, Mem. Coll. Sci. Univ. Kyoto. Ser. A. Math. **29** (1955), 189–192.
Markus Rost, *Notes on the degree formula*, http://www.math.uni-bielefeld.de/ rost/degree-formula.html, 2001.
G. V. Ravindra and V. Srinivas, *The [N]{}oether-[L]{}efschetz theorem for the divisor class group*, J. Algebra **322** (2009), no. 9, 3373–3391.
Ernst Witt, *Zerlegung reeller algebraischer [F]{}unktionen in [Q]{}uadrate, [S]{}chiefkörper über reellem [F]{}unktionenkörper*, J. Reine Angew. Math. **171** (1934), 4–11.
O. [Wittenberg]{}, *[Sur une conjecture de Kato et Kuzumaki concernant les hypersurfaces de Fano]{}*, ArXiv e-prints (2013).
K. Zainoulline, *Degree formula for connective [$K$]{}-theory*, Invent. Math. **179** (2010), no. 3, 507–522.
Yuri G. Zarhin, *Hyperelliptic [J]{}acobians without complex multiplication*, Math. Res. Lett. **7** (2000), no. 1, 123–132.
, *The endomorphism rings of [J]{}acobians of cyclic covers of the projective line*, Math. Proc. Cambridge Philos. Soc. **136** (2004), no. 2, 257–267.
, *Non-isogenous superelliptic [J]{}acobians*, Math. Z. **253** (2006), no. 3, 537–554.
, *Families of absolutely simple hyperelliptic [J]{}acobians*, Proc. Lond. Math. Soc. (3) **100** (2010), no. 1, 24–54.
Princeton University, Princeton NJ 08544-1000
[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We enumerate three specific permutation classes defined by two forbidden patterns of length four. The techniques involve inflations of geometric grid classes.'
author:
- |
Michael H. Albert\
Department of Computer Science\
University of Otago\
Dunedin, New Zealand\
M. D. Atkinson\
Department of Computer Science\
University of Otago\
Dunedin, New Zealand\
Vincent Vatter[^1]\
Department of Mathematics\
University of Florida\
Gainesville, Florida USA\
title: 'Inflations of Geometric Grid Classes: Three Case Studies'
---
Introduction {#infinite-simples-intro}
============
*Classes* of permutations are sets that are closed downwards under taking subpermutations. They are often presented as sets ${\mathcal{C}}$ that avoid a given set $B$ of permutations (i.e. the members of ${\mathcal{C}}$ have no subpermutation in the set $B$). We express this by the notation ${\mathcal{C}}={\operatorname{Av}}(B)$. We may take $B$ to be an antichain (a set of pairwise incomparable permutations), in which case we say that $B$ is the *basis* of ${\mathcal{C}}$.
Much of the inspiration for the early work on permutation classes was driven by the enumeration problem: given ${\mathcal{C}}={\operatorname{Av}}(B)$, how many permutations of each length does ${\mathcal{C}}$ contain? The answer to such a question could be a formula giving this number $|{\mathcal{C}}_n|$ in terms of the length, $n$, a generating function $\sum |{\mathcal{C}}_n|x^n$ or simply an asymptotic result about the behaviour of $|{\mathcal{C}}_n|$ as $n\rightarrow\infty$.
Recently, Albert, Atkinson, Bouvel, Ruškuc, and Vatter [@albert:geometric-grid-:] have developed the theory of geometric grid classes, and Albert, Ruškuc, and Vatter [@albert:inflations-of-g:] have continued this exploration by investigating the theory of inflations of such classes. Our aim in this paper is to:
- demonstrate the effectiveness of this approach, and
- illustrate how one might implement these techniques in a “real world” setting, bypassing what would otherwise be thorny theoretical issues.
It should be noted that this presentation is historically backward; the results of this paper preceded and inspired the two more theoretical papers cited above.
In this work, our examples are exclusively classes with two basis elements of length four, which we call *2 $\times$ 4 classes*. It must be admitted that the attention paid to 2 $\times$ 4 classes is not entirely in proportion to their intrinsic importance. Nevertheless, these classes represent a significant dataset which seems to contain some difficult enumerative problems. Thus they pose a good challenge for new approaches to the enumeration of restricted permutations.
There are $56$ essentially different (i.e. inequivalent under symmetries) 2 $\times$ 4 classes. Some of these classes nevertheless share the same enumeration (a phenomenon called *Wilf-equivalence*), so the 2 $\times$ 4 classes have only $38$ different enumerations [@bona:the-permutation:; @kremer:permutations-wi:; @kremer:postscript:-per:; @kremer:finite-transiti:; @le:wilf-classes-of:]. This paper brings the number of $2\times 4$ Wilf classes which have been enumerated to $24$ (see Wikipedia [@wikipedia:enumerations-of:], which contains a list of such enumerations).
A central part of our approach depends on analysing the simple permutations in a class. An *interval* in the permutation $\pi$ is a set of contiguous indices $I=\{a,a+1,\dots,b\}$ such that the set $\{\pi(i){\::\:}i\in I\}$ is also contiguous. Every permutation $\pi$ of length $n$ has *trivial intervals* of lengths $0$, $1$, and $n$, and other intervals are called *proper*. A permutation with no proper intervals is called *simple*. Another way to think about simple permutations arises repeatedly throughout our arguments. Any subset $p_1,\dots$ of entries of the permutation $\pi$ defines a minimal axes-parallel rectangle (or simply, *box*), whose left edge slices through the leftmost of these entries, top edge slices through the greatest of these entries, and so on. A simple permutation is one in which the box defined by any proper subset of two or more of its entries is *separated* by an entry outside the box, by which we mean that this entry lies either
- vertically amongst these entries but to the left (or right) of all of them (*vertical separation*), or
- horizontally amongst these entries but above (or below) all of them (*horizontal separation*).
Simple permutations are precisely those that do not arise from a non-trivial inflation, in the following sense. Given a permutation $\sigma$ of length $m$ and nonempty permutations $\alpha_1,\dots,\alpha_m$, the *inflation* of $\sigma$ by $\alpha_1,\dots,\alpha_m$, denoted $\sigma[\alpha_1,\dots,\alpha_m]$, is the permutation of length $|\alpha_1|+\cdots+|\alpha_m|$ obtained by replacing each entry $\sigma(i)$ by an interval that is order isomorphic to $\alpha_i$ in such a way that the intervals are order isomorphic to $\sigma$. For example, $$2413[1,132,321,12]=4\ 798\ 321\ 56.$$ We give two particular types of inflations special terminology and notation. The inflation $12[\alpha_1,\alpha_2]$ is called a *(direct) sum* and denoted by $\alpha_1\oplus\alpha_2$. A permutation is *sum decomposable* if it can be expressed as a nontrivial sum, and *sum indecomposable* otherwise. The inflation $21[\alpha_1,\alpha_2]$ is called a *skew sum*, similarly denoted $\alpha_1\ominus\alpha_2$, and accompanied by analogous terms *skew decomposable* and *skew indecomposable*. We extend the notion of direct and skew sum to classes, defining $${\mathcal{C}}\oplus{\mathcal{D}}=\{\pi\oplus\sigma{\::\:}\pi\in{\mathcal{C}}\mbox{ and }\sigma\in{\mathcal{D}}\},$$ with an analogous definition for ${\mathcal{C}}\ominus{\mathcal{D}}$.
The precise connection between simple permutations and inflations is furnished by the following result.
For every permutation $\pi$ there is a unique simple permutation $\sigma$ such that $\pi=\sigma[\alpha_1,\alpha_2,\ldots,\alpha_m]$. Furthermore, except when $\sigma=12$ or $\sigma=21$, the intervals of $\pi$ that correspond to $\alpha_1,\alpha_2,\ldots,\alpha_m$ are uniquely determined. In the case that $\sigma=12$ (respectively $\sigma=21$), the intervals are unique so long as we require the first of the two intervals to be sum (respectively skew) indecomposable.
One of the first general enumeration results is the following from [@albert:simple-permutat:]:
\[thm-fin-simples-alg\] If the class ${\mathcal{C}}$ contains only finitely many simple permutations, then ${\mathcal{C}}$ has an algebraic generating function.
This theorem has since been generalised in two different directions. Brignall, Huczynska, and Vatter [@brignall:simple-permutat:] introduced the notion of “query-complete sets of properties” to show that if a class satisfies the hypotheses of Theorem \[thm-fin-simples-alg\], then such subsets as the even permutations or the involutions in ${\mathcal{C}}$ have algebraic generating functions. More relevant to our investigation, [@albert:inflations-of-g:] significantly weakened the hypotheses of Theorem \[thm-fin-simples-alg\], showing that its conclusion holds even when ${\mathcal{C}}$ contains infinitely many simple permutations, so long as these simple permutations lie in a geometric grid class, a notion introduced in Section \[sec-geom-grid-review\]. Before this, we consider an example which gives the flavour of our approach without requiring much additional machinery.
Example \#1: Avoiding 4213 and 3142 {#sec-4213-3142}
===================================
Before describing our first example we need to introduce a family of simple permutations and quote a result. A [*parallel alternation*]{} is a permutation whose plot can be divided into two parts, by a single horizontal or vertical line, so that the points on either side of this line are both either increasing or decreasing and for every pair of points from the same part there is a point from the other part which [*separates*]{} them, i.e., there is a point from the other part which lies either horizontally or vertically between them. It is easy to see that a parallel alternation of length at least four is simple if and only if its length is even, it does not begin with its smallest entry, and it does not end with its greatest entry. Thus there are precisely four simple parallel alternations of each even length at least six, shown in Figure \[fig-par-alts\], and no simple parallel alternations of odd length.
[ccccccc]{}
(0,0)(80,80) (0,0)(80,80) (10,20)[0.04in]{} (20,40)[0.04in]{} (30,60)[0.04in]{} (40,80)[0.04in]{} (50,10)[0.04in]{} (60,30)[0.04in]{} (70,50)[0.04in]{} (80,70)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(80,80) (0,0)(80,80) (10,50)[0.04in]{} (20,10)[0.04in]{} (30,60)[0.04in]{} (40,20)[0.04in]{} (50,70)[0.04in]{} (60,30)[0.04in]{} (70,80)[0.04in]{} (80,40)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(80,80) (0,0)(80,80) (10,70)[0.04in]{} (20,50)[0.04in]{} (30,30)[0.04in]{} (40,10)[0.04in]{} (50,80)[0.04in]{} (60,60)[0.04in]{} (70,40)[0.04in]{} (80,20)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(80,80) (0,0)(80,80) (10,40)[0.04in]{} (20,80)[0.04in]{} (30,30)[0.04in]{} (40,70)[0.04in]{} (50,20)[0.04in]{} (60,60)[0.04in]{} (70,10)[0.04in]{} (80,50)[0.04in]{}
Schmerl and Trotter [@schmerl:critically-inde: Corollary 5.10] proved a result (in the more general context of irreflexive binary relational structures) which in our context states that every simple permutation of length $n$ which is not a parallel alternation contains simple subpermutations of every length $5\le m\le n$. Therefore, in order to establish that the permutation class ${\mathcal{C}}$ contains *only* parallel alternations, we just need to check that it does not contain any simple permutation of length $5$, i.e., that $${\mathcal{C}}\subseteq{\operatorname{Av}}(24153, 25314, 31524, 35142, 41352, 42513).$$ Clearly this holds for the class ${\operatorname{Av}}(4213,3142)$, because $4213$ is contained in $25314$ and $42513$ while $3142$ is contained in $24153$, $31524$, $35142$, and $41352$. Moreover, it is easily seen (because of the basis element $3142$) that ${\operatorname{Av}}(4213,3142)$ can contain only parallel alternations oriented as on the left of Figure \[fig-par-alts\], i.e., those of the form $$246\cdots(2m)135\cdots(2m-1).$$
With the simple permutations characterised, we now describe the allowed inflations. It is easy to see that $\pi\oplus\sigma\in{\operatorname{Av}}(4213,3142)$ for all $\pi,\sigma\in{\operatorname{Av}}(4213,3142)$, or in other words, that the class is [*sum closed*]{}. Thus, letting $f$ denote the generating function for nonempty permutations in ${\operatorname{Av}}(4213,3142)$ and $f_\oplus$ denote the generating function for sum decomposable permutations, we see that $f_\oplus=\left(f-f_\oplus\right)f$, from which it follows that $$f_\oplus=\frac{f^2}{1+f}.$$ For skew sums, we have that $\pi\ominus\sigma\in{\operatorname{Av}}(4213,3142)$ if and only if $\pi\in{\operatorname{Av}}(4213,3142)$ and $\sigma\in{\operatorname{Av}}(213)$. Letting $$c=\frac{1-2x-\sqrt{1-4x}}{2x}$$ denote the generating function for the Catalan numbers (with constant term zero), which is well-known as the generating function of nonempty permutations in ${\operatorname{Av}}(213)$, we have $f_\ominus=\left(f-f_\ominus\right)c$, so $$f_\ominus=\frac{cf}{1+c}.$$
Now we must count the inflations of the parallel alternations $246\cdots (2m)135\cdots (2m-1)$ for $m\ge 2$. This is relatively straightforward:
- the interval inflating $2m-1$ must avoid $213$,
- all other intervals inflating odd entries must be increasing, and
- even entries may be inflated by any element of ${\operatorname{Av}}(4213,3142)$.
This shows that the contribution of inflations of $246\cdots (2m)135\cdots (2m-1)$, for each $m\ge 2$, is $$f^m\left(\frac{x}{1-x}\right)^{m-1}c,$$ showing that $$f
=
x+f_\oplus+f_\ominus+\sum_{m=2}^\infty f^m\left(\frac{x}{1-x}\right)^{m-1}c
=
x+\frac{f^2}{1+f}+\frac{cf}{1+c}+\frac{xcf^2}{1-x-xf}.$$ From this we obtain:
\[thm-3142-4213\] The generating function $f$ for ${\operatorname{Av}}(4213,3142)$ satisfies $$\begin{array}{rclcc}
x^3f^6
&+&
(7x^3-7x^2+2x)f^5&&\\
&+&
(x^4+14x^3-21x^2+10x-1)f^4&&\\
&+&
(4x^4+8x^3-19x^2+11x-2)f^3&&\\
&+&
(6x^4-5x^3-2x^2+2x)f^2&&\\
&+&
(4x^4-7x^3+4x^2-x)f&&\\
&+&
x^4-2x^3+x^2
&=&0.
\end{array}$$
The first several terms of this sequence are $$1,2,6,22,89,379,1664,7460,33977,156727,730619,3436710,16291842,77758962,$$ sequence [A165541]{} in the [OEIS [@sloane:the-on-line-enc:]]{}. The discriminant of the polynomial defining the generating function has a smallest positive root $\rho \approx 0.1895$, which is therefore the radius of convergence of the generating function and as ${\operatorname{Av}}(4213,3142)$ is sum closed (and hence the sequence $f_n$ is supermultiplicative) we can conclude that $f_n^{1/n} \to 1/\rho \approx 5.2778$. More detailed on the asymptotic behaviour of $f_n$ could be determined by standard methods as found for instance in Flajolet and Sedgewick [@flajolet:analytic-combin: Section VII.7].
[ccccccc]{}
(0,0)(80,80) (0,0)(80,80) (10,20)[0.04in]{} (20,40)[0.04in]{} (30,60)[0.04in]{} (40,80)[0.04in]{} (50,10)[0.04in]{} (60,70)[0.04in]{} (70,50)[0.04in]{} (80,30)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(80,80) (0,0)(80,80) (10,30)[0.04in]{} (20,50)[0.04in]{} (30,70)[0.04in]{} (40,10)[0.04in]{} (50,80)[0.04in]{} (60,60)[0.04in]{} (70,40)[0.04in]{} (80,20)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(80,80) (0,0)(80,80) (10,40)[0.04in]{} (20,80)[0.04in]{} (30,10)[0.04in]{} (40,70)[0.04in]{} (50,20)[0.04in]{} (60,60)[0.04in]{} (70,30)[0.04in]{} (80,50)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(80,80) (0,0)(80,80) (10,50)[0.04in]{} (20,10)[0.04in]{} (30,80)[0.04in]{} (40,20)[0.04in]{} (50,70)[0.04in]{} (60,30)[0.04in]{} (70,60)[0.04in]{} (80,40)[0.04in]{}
This is not the only 2 $\times$ 4 class to which such elementary techniques apply. For example:
- ${\operatorname{Av}}(4213,1342)$ contains precisely two simple permutations of each length $n\ge 4$, both of which are wedge simple permutations oriented as the first two permutations shown in Figure \[fig-wedge-simples\]. This family of simple permutations is well enough behaved that we could enumerate the class, but this has already been done by Kremer and Shiu [@kremer:finite-transiti:] and can now be performed automatically using the Maple package [FinLabel]{} described in Vatter [@vatter:finitely-labele:].
- ${\operatorname{Av}}(4213,3124)$ contains precisely two simple permutations of each length $n\ge 4$, oriented as the rightmost two permutations shown in Figure \[fig-wedge-simples\]. This class was enumerated by Bóna [@bona:the-permutation:].
Grid Classes and Regular Languages {#sec-geom-grid-review}
==================================
Given a permutation $\pi$ of length $n$ and sets $X,Y\subseteq[n]$, we write $\pi(X\times Y)$ for the permutation that is order isomorphic to the subsequence of $\pi$ with indices from $X$ and values in $Y$. For example, $286435179([4,9]\times[5,9])$ consists of the subsequence of entries in indices $4$ through $9$ which have values between $5$ and $9$; in this case the subsequence is $579$, so $286435179([4,9]\times[5,9])=123$.
Suppose that $M$ is a $t\times u$ matrix[^2]with entries from $\{0,\pm 1\}$. A *gridded permutation* is a permutation $\pi$ equipped with *row* and *column divisions* denoted respectively by $1=c_1\le\cdots\le c_{t+1}=n+1$ and $1=r_1\le\cdots\le r_{u+1}=n+1$ (where $n$ is the length of $\pi$). This gridded permutation (or simply, gridding of $\pi$) is *compatible* with the matrix $M$ (in which case we sometimes call it an $M$-gridding of $\pi$) if $\pi([c_k,c_{k+1})\times[r_\ell,r_{\ell+1}))$ is increasing whenever $M_{k,\ell}=1$, decreasing whenever $M_{k,\ell}=-1$, and empty whenever $M_{k,\ell}=0$. The [*(monotone) grid class of $M$*]{}, written ${\operatorname{Grid}}(M)$, consists of all permutations which possess a gridding compatible with $M$. Figure \[fig-grid-geom-three-pane\] shows an example.
As illustrated by Murphy and Vatter [@murphy:profile-classes:], monotone grid classes can display chaotic and unstructured behaviour. However, it has recently been shown that these classes contain subclasses with especially amenable structure. To define these subclasses, consider the point set in $\mathbb{R}^2$ (called the *standard figure* of the ${0/\mathord{\pm} 1}$ matrix $M$) consisting of cells $C_{kl}$ whose contents are:
- the line segment from $(k-1,\ell-1)$ to $(k,\ell)$ if $M_{k,\ell}=1$ or
- the line segment from $(k-1,\ell)$ to $(k,\ell-1)$ if $M_{k,\ell}=-1$ or
- empty if $M_{kl}=0$.
The *geometric grid class* of $M$, denoted by ${\operatorname{Geom}}(M)$, is then the set of all permutations that can be drawn on this figure in the following manner. Choose $n$ points in the figure, no two on a common horizontal or vertical line. Then label the points from $1$ to $n$ from bottom to top and record these labels reading left to right. The centre pane of Figure \[fig-grid-geom-three-pane\] shows a permutation from a geometric grid class, while the right pane demonstrates that $2413$ is not in this geometric grid class.
It sometimes happens that ${\operatorname{Grid}}(M)={\operatorname{Geom}}(M)$; to characterise this phenomenon, we need to introduce a graph. The *row-column graph* of a $t\times u$ matrix $M$ is the bipartite graph on the vertices $x_1$, $\dots$, $x_t$, $y_1$, $\dots$, $y_u$ where $x_k$ is adjacent to $y_\ell$ if and only if $M_{k,\ell}\neq 0$. Albert, Atkinson, Bouvel, Ruškuc, and Vatter [[@albert:geometric-grid-: Theorem 3.2]]{} showed that ${\operatorname{Grid}}(M)={\operatorname{Geom}}(M)$ if and only if the row-column graph of $M$ is a forest (in this case we say that $M$ *is* a forest). As it happens, all gridding matrices encountered in this paper are forests.
Geometric grid classes are especially tractable because their elements can be encoded by words over a finite alphabet, and for the rest of this section we describe this encoding and its properties. We say that a ${0/\mathord{\pm} 1}$ matrix $M$ of size $t\times u$ is a *partial multiplication matrix* if there exist *column and row signs* $$f_1,\ldots,f_t,g_1,\ldots,g_u\in \{1,-1\}$$ such that every entry $M_{k,\ell}$ is equal to either $f_kg_\ell$ or $0$. It is not hard to prove that every geometric grid class is equal to ${\operatorname{Geom}}(M)$ for a partial multiplication matrix $M$, and this is especially trivial for forests.
[ccccc]{}
(0,0)(100,100) (35,0)(35,100) (0,45)(100,45) (0,0)(98,98) (10,20)[0.04in]{} (20,80)[0.04in]{} (30,60)[0.04in]{} (40,40)[0.04in]{} (50,30)[0.04in]{} (60,50)[0.04in]{} (70,10)[0.04in]{} (80,70)[0.04in]{} (90,90)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(100,100) (0,0)(100,100) (0,100)(100,0) (50,0)(50,100) (0,50)(100,50) (0,0)(100,0)(100,100)(0,100)(0,0) (10,10)[0.04in]{} (20,80)[0.04in]{} (30,70)[0.04in]{} (40,40)[0.04in]{} (60,60)[0.04in]{} (70,30)[0.04in]{} (80,20)[0.04in]{} (90,90)[0.04in]{}
&
------------------------------------------------------------------------
&
(0,0)(100,100) (20,20)(100,20) (20,20)(20,100) (25,75)(100,75) (70,70)(70,0) (0,0)(100,100) (0,100)(100,0) (50,0)(50,100) (0,50)(100,50) (0,0)(100,0)(100,100)(0,100)(0,0) (20,20)[0.04in]{} (25,75)[0.04in]{} (70,70)[0.04in]{} (60,10)[0.04in]{}
The column and row signs essentially specify an order in which the monotone entries in a cell of a gridded permutation should be read. Cells corresponding to $M_{k\ell}=f_kg_{\ell}$ are read from left to right (respectively right to left) if $f_k=1$ (respectively $f_k=-1$) and bottom to top (respectively top to bottom) if $g_{\ell}=1$ (respectively $g_{\ell}=-1$). These directions are sometimes marked on our diagrams. The [*base point*]{} of a cell is the corner from which its reading begins.
To describe the encoding of ${\operatorname{Geom}}(M)$ we introduce a *cell alphabet* $\Sigma$ associated to $M$ which consists of a unique letter $a_{kl}$ for each nonempty cell $C_{kl}$ of the standard figure of $M$. Then, to every word $w=w_1\cdots w_n\in\Sigma^\ast$ we associate a permutation ${\varphi}(w)$. First we choose arbitrary distances $$0<d_1<\cdots<d_n<1.$$ For each $1\le i\le n$, we choose a point $p_i$ corresponding to $w_i$ in the following manner. Let $w_i=a_{k\ell}$; the point $p_i$ is chosen from the line segment in cell $C_{k,\ell}$, at infinity-norm distance $d_i$ from the base point of this cell. Finally, ${\varphi}(w)$ denotes the permutation defined by the set $\{p_1,\dots,p_n\}$ of points. It can be seen that ${\varphi}$ does not depend on the particular choice of $d_1,\dots,d_n$, and thus ${\varphi}{\::\:}\Sigma^\ast\to {\operatorname{Geom}}(M)$ is a well-defined mapping.
The mapping ${\varphi}$ is many-to-one, and so for enumerative applications we must restrict its domain to a set $L\subseteq\Sigma^\ast$ on which ${\varphi}$ is injective. We seek to choose $L$ to be a *regular language*. The regular languages are those that can be obtained from the empty language and the singleton languages using the operations of union, concatenation, and Kleene star (where $K^*$ is the set of all concatenations of 0 or more words from $K$). Alternatively, regular languages can also be characterised as those accepted by deterministic finite state automata. From this latter viewpoint it follows (e.g., by the transfer matrix method) that regular languages have rational generating functions (either when enumerated by length, or with a separate variable $x_a$ for each $a \in \Sigma$). We refer readers to [@flajolet:analytic-combin: Section I.4 and Appendix A.7] for more information on regular languages.
The following theorem from [@albert:geometric-grid-:] demonstrates the connection between subclasses of geometric grid classes and regular languages. Essentially, it says that all such classes are extremely well behaved.
\[thm-geom-griddable-all\] Suppose that ${\mathcal{C}}\subseteq{\operatorname{Geom}}(M)$ is a permutation class and $M$ is a partial multiplication matrix with cell alphabet $\Sigma$. Then the following hold:
1. ${\mathcal{C}}$ is partially well-ordered.
2. ${\mathcal{C}}$ is finitely based.
3. There is a regular language $L\subseteq\Sigma^\ast$ such that the mapping ${\varphi}{\::\:}L\rightarrow{\mathcal{C}}$ is a bijection.
4. There is a regular language $L_S$, contained in the regular language $L$ from (iii), such that the mapping ${\varphi}$ is a bijection between $L_S$ and the simple permutations in ${\mathcal{C}}$.
Note that the proof of Theorem \[thm-geom-griddable-all\] is nonconstructive, so while we use the encoding ${\varphi}$ throughout this work, we construct the regular languages we use from first principles.
Albert, Atkinson, and Brignall [@albert:the-enumeration:2143:4231; @albert:the-enumeration:3] demonstrate four concrete examples of using these techniques to enumerate $2\times 4$ classes.
The examples considered in this paper are inflations of geometric grid classes. The theoretical issues of such classes were studied by Albert, Ruškuc, and Vatter [@albert:inflations-of-g:], who proved that every subclass of $\langle{\operatorname{Geom}}(M)\rangle$ has an algebraic generating function (essentially by showing that it is in bijection with a context-free language). From this perspective, Section \[sec-4213-3142\] considers the case of a class contained in $\langle{\operatorname{Geom}}(\begin{footnotesize}\begin{array}{rr}1&1\end{array}\end{footnotesize})\rangle$, while the next two sections consider classes whose simple permutations are contained in more complicated geometric grid classes.
Example \#2: Avoiding $4312$ and $3142$ {#sec-4312-3142}
=======================================
We begin our next example with a characterisation of its simple permutations.
\[prop-4312-3142-simples\] The simple permutations of ${\operatorname{Av}}(4312,3142)$ and ${\operatorname{Geom}}{\mbox{\begin{footnotesize}$\left(\begin{array}{rrr}0&1&1\\1&0&-1\end{array}\right)$\end{footnotesize}}}$ coincide.
First observe that $${\operatorname{Grid}}{\mbox{\begin{footnotesize}$\left(\begin{array}{rrr}0&1&1\\1&0&-1\end{array}\right)$\end{footnotesize}}}
=
{\operatorname{Geom}}{\mbox{\begin{footnotesize}$\left(\begin{array}{rrr}0&1&1\\1&0&-1\end{array}\right)$\end{footnotesize}}}
\subseteq
{\operatorname{Av}}(4312, 3142),$$ so it suffices to prove that the simple permutations in ${\operatorname{Av}}(4312, 3142)$ are contained this grid class. Specifically, we will show that in any simple permutation of ${\operatorname{Av}}(4312, 3142)$ the entries that follow the maximum make up a “wedge permutation” oriented as $<$ (which is equivalent to avoiding both 132 and 312), and those preceding the maximum form an increasing sequence.
So, let a simple permutation $\pi \in {\operatorname{Av}}(4312, 3142)$ be given. Because $\pi$ avoids 4312 there can be no 312 pattern after its maximum so, for the sake of contradiction, assume that there is a $132$ pattern. Specifically, choose such a pattern $acb$ where $a$ is as low as possible, and $c$ is as high as possible (for the chosen $a$). This yields the situation depicted on the left in Figure \[fig-ex2-simple-no-132-after-max\]. Now, in order that the cell bounded by $\{b,c\}$ not form an interval, there must be some entry $d$ in the cell immediately to its left. Taking the leftmost such entry yields the diagram on the right in Figure \[fig-ex2-simple-no-132-after-max\]. In this diagram we see that the entries of $\pi$ lying in the box bounded by $\{b,c,d\}$ (including those three entries) form a proper interval, contradicting the simplicity of $\pi$.
[ccc]{}
(0,0)(50,40) (0,20)(10,20)(10,30)(0,30) (10,0)(20,0)(20,10)(10,10) (10,20)(20,20)(20,30)(10,30) (10,30)(20,30)(20,40)(10,40) (20,0)(30,0)(30,10)(20,10) (20,30)(30,30)(30,40)(20,40) (30,0)(40,0)(40,10)(30,10) (30,10)(40,10)(40,20)(30,20) (30,30)(40,30)(40,40)(30,40) (40,20)(50,20)(50,30)(40,30) (10,40)[2.0]{} (20,10)[2.0]{} (20,10)[$a$]{} (30,30)[2.0]{} (30,30)[$c$]{} (40,20)[2.0]{} (40,20)[$b$]{} (0,0)(0,40) (0,0)(50,0) (10,0)(10,40) (0,10)(50,10) (20,0)(20,40) (0,20)(50,20) (30,0)(30,40) (0,30)(50,30) (40,0)(40,40) (0,40)(50,40) (50,0)(50,40)
&
------------------------------------------------------------------------
&
(0,0)(60,50) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (10,0)(20,0)(20,10)(10,10) (10,20)(20,20)(20,30)(10,30) (10,30)(20,30)(20,40)(10,40) (10,40)(20,40)(20,50)(10,50) (20,0)(30,0)(30,10)(20,10) (20,20)(30,20)(30,30)(20,30) (20,30)(30,30)(30,40)(20,40) (20,40)(30,40)(30,50)(20,50) (30,0)(40,0)(40,10)(30,10) (30,10)(40,10)(40,20)(30,20) (30,40)(40,40)(40,50)(30,50) (40,0)(50,0)(50,10)(40,10) (40,10)(50,10)(50,20)(40,20) (40,40)(50,40)(50,50)(40,50) (50,20)(60,20)(60,30)(50,30) (50,30)(60,30)(60,40)(50,40) (10,50)[2.0]{} (20,10)[2.0]{} (20,10)[$a$]{} (30,30)[2.0]{} (30,30)[$d$]{} (40,40)[2.0]{} (40,40)[$c$]{} (50,20)[2.0]{} (50,20)[$b$]{} (0,0)(0,50) (0,0)(60,0) (10,0)(10,50) (0,10)(60,10) (20,0)(20,50) (0,20)(60,20) (30,0)(30,50) (0,30)(60,30) (40,0)(40,50) (0,40)(60,40) (50,0)(50,50) (0,50)(60,50) (60,0)(60,50)
We can now argue in a similar fashion that the entries preceding the maximum entry of $\pi$ form an increasing sequence, i.e., that there cannot be a $21$ pattern before the maximum entry of $\pi$. Suppose to the contrary that there were one, and choose such a pattern $ba$ where $b$ is as high as possible, and $a$ is as low as possible (for the chosen $b$). Now the cell defined by $\{a,b\}$ must be split either to the left or to the right. The picture on the left of Figure \[fig-ex2-simple-no-21-after-max\] shows that $\{a,b\}$ cannot be split solely to the right, as then taking $c$ to be the rightmost such separator we see that $\{a,b,c\}$ would lie in a proper interval. Similarly, $\{a,b\}$ cannot be split solely to the left. Thus $\{a,b\}$ must be split on both the left and the right. Now, taking $c$ to be the rightmost separator and $d$ the leftmost separator, we have two cases, depicted in the centre and right of Figure \[fig-ex2-simple-no-21-after-max\]. In both cases it is clear that $\{a,b,c,d\}$ is contained in a proper interval, and this contradiction completes the proof.
[ccccc]{}
(0,0)(50,40) (0,10)(10,10)(10,20)(0,20) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (0,30)(10,40) (10,0)(20,0)(20,10)(10,10) (10,20)(20,20)(20,30)(10,30) (10,30)(20,30)(20,40)(10,40) (20,0)(30,0)(30,10)(20,10) (20,30)(30,30)(30,40)(20,40) (30,0)(40,0)(40,10)(30,10) (30,10)(40,10)(40,20)(30,20) (30,20)(40,20)(40,30)(30,30) (40,10)(50,10)(50,20)(40,20) (40,20)(50,20)(50,30)(40,30) (10,30)[2.0]{} (10,30)[$b$]{} (20,10)[2.0]{} (20,10)[$a$]{} (30,20)[2.0]{} (30,20)[$c$]{} (40,40)[2.0]{} (0,0)(0,40) (0,0)(50,0) (10,0)(10,40) (0,10)(50,10) (20,0)(20,40) (0,20)(50,20) (30,0)(30,40) (0,30)(50,30) (40,0)(40,40) (0,40)(50,40) (50,0)(50,40)
&
------------------------------------------------------------------------
&
(0,0)(60,50) (0,10)(10,10)(10,20)(0,20) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (0,40)(10,40)(10,50)(0,50) (10,0)(20,0)(20,10)(10,10) (10,10)(20,10)(20,20)(10,20) (10,20)(20,20)(20,30)(10,30) (10,40)(20,40)(20,50)(10,50) (10,40)(20,50) (20,0)(30,0)(30,10)(20,10) (20,20)(30,20)(30,30)(20,30) (20,30)(30,30)(30,40)(20,40) (20,40)(30,40)(30,50)(20,50) (30,0)(40,0)(40,10)(30,10) (30,30)(40,30)(40,40)(30,40) (30,40)(40,40)(40,50)(30,50) (40,0)(50,0)(50,10)(40,10) (40,10)(50,10)(50,20)(40,20) (40,20)(50,20)(50,30)(40,30) (40,30)(50,30)(50,40)(40,40) (50,10)(60,10)(60,20)(50,20) (50,20)(60,20)(60,30)(50,30) (50,30)(60,30)(60,40)(50,40) (10,30)[2.0]{} (10,30)[$d$]{} (20,40)[2.0]{} (20,40)[$b$]{} (30,10)[2.0]{} (30,10)[$a$]{} (40,20)[2.0]{} (40,20)[$c$]{} (50,50)[2.0]{} (0,0)(0,50) (0,0)(60,0) (10,0)(10,50) (0,10)(60,10) (20,0)(20,50) (0,20)(60,20) (30,0)(30,50) (0,30)(60,30) (40,0)(40,50) (0,40)(60,40) (50,0)(50,50) (0,50)(60,50) (60,0)(60,50)
&
------------------------------------------------------------------------
&
(0,0)(60,50) (0,10)(10,10)(10,20)(0,20) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (0,40)(10,40)(10,50)(0,50) (10,0)(20,0)(20,10)(10,10) (10,40)(20,40)(20,50)(10,50) (10,40)(20,50) (20,0)(30,0)(30,10)(20,10) (20,30)(30,30)(30,40)(20,40) (20,40)(30,40)(30,50)(20,50) (30,0)(40,0)(40,10)(30,10) (30,40)(40,40)(40,50)(30,50) (40,0)(50,0)(50,10)(40,10) (40,10)(50,10)(50,20)(40,20) (40,20)(50,20)(50,30)(40,30) (40,30)(50,30)(50,40)(40,40) (50,10)(60,10)(60,20)(50,20) (50,20)(60,20)(60,30)(50,30) (50,30)(60,30)(60,40)(50,40) (10,20)[2.0]{} (10,20)[$d$]{} (20,40)[2.0]{} (20,40)[$b$]{} (30,10)[2.0]{} (30,10)[$a$]{} (40,30)[2.0]{} (40,30)[$c$]{} (50,50)[2.0]{} (0,0)(0,50) (0,0)(60,0) (10,0)(10,50) (0,10)(60,10) (20,0)(20,50) (0,20)(60,20) (30,0)(30,50) (0,30)(60,30) (40,0)(40,50) (0,40)(60,40) (50,0)(50,50) (0,50)(60,50) (60,0)(60,50)
(-3,0)(120,85) (0,0)(80,80) (80,40)(120,0) (80,40)(120,80) (0,0)(0,80) (40,0)(40,80) (80,0)(80,80) (120,0)(120,80) (0,0)(120,0) (0,40)(120,40) (0,80)(120,80) (-3,2)(-3,38) (-3,42)(-3,78) (2,83)(38,83) (42,83)(78,83) (82,83)(118,83) (35,35)[2.0]{} (90,30)[2.0]{} (25,25)[2.0]{} (100,60)[2.0]{} (105,15)[2.0]{} (110,70)[2.0]{} (75,75)[2.0]{} (40,0)[${\textsf{a}}$]{} (80,40)[${\textsf{b}}$]{} (80,0)[${\textsf{c}}$]{} (120,40)[${\textsf{d}}$]{}
We now consider the encoding ${\varphi}$ over the cell alphabet $\Sigma=\{{\textsf{a}},{\textsf{b}},{\textsf{c}},{\textsf{d}}\}$ as indicated in Figure \[fig-2473516\], which also shows an example of ${\varphi}$. This mapping is not injective on $\Sigma^\ast$ for the following two reasons.
- The same gridded permutation may be the image of multiple words (in our example, this occurs because the pairs $\{{\textsf{a}},{\textsf{b}}\}$, $\{{\textsf{a}},{\textsf{d}}\}$, and $\{{\textsf{b}},{\textsf{c}}\}$ “commute”, i.e., they may be interchanged without affecting the gridded permutation obtained). A method to handle this issue in general (by appealing to the theory of “trace monoids”) is presented in [@albert:geometric-grid-: Section 7].
- A given permutation may have several different $M$-griddings. A (nonconstructive) method to handle this issue in general is presented in [@albert:geometric-grid-: Section 8].
In the class we are considering, ${\operatorname{Av}}(4312,3142)$, it is possible to deal with the issues concretely.
First we address (G1). For any particular gridded permutation, we prefer the lexicographically minimal word encoding it. For example, suppose that a word contained a factor of the form $\{{\textsf{b}},{\textsf{d}}\}^+{\textsf{a}}$ (here the $+$ superscript signifies that this portion of the word contains at least one letter). We could then replace this factor by a factor of the form ${\textsf{a}}\{{\textsf{b}},{\textsf{d}}\}^+$ and obtain a lexicographically lesser word which is mapped to the same permutation. Therefore we forbid factors of the form $\{{\textsf{b}},{\textsf{d}}\}^+{\textsf{a}}$. The other factor we need to forbid is ${\textsf{c}}{\textsf{a}}^\ast{\textsf{b}}$ (which could be replaced by a factor of the form ${\textsf{b}}{\textsf{c}}{\textsf{a}}^\ast$).
Now we address (G2), which requires us to choose a preferred (geometric) $M$-gridding for every permutation in ${\operatorname{Geom}}(M)$. Among all $M$-griddings of a permutation, we prefer the one that has the most entries in the first column, then the most entries in the second column, and then the most entries in the first row. Thus in terms of column divisions $1=c_1\le c_2\le c_3\le c_4=n+1$ and row divisions $1=r_1\le r_2\le r_3=n+1$, we seek to maximise $c_2$, then $c_3$, and then $r_2$. The words which correspond to such griddings can now be characterised as those which do not begin with ${\textsf{a}}^\ast{\textsf{d}}$, ${\textsf{b}}$, $\{{\textsf{a}},{\textsf{c}}\}^\ast{\textsf{b}}$, or ${\textsf{d}}$ and are not of the form ${\textsf{c}}\{{\textsf{a}},{\textsf{c}},{\textsf{d}}\}^\ast$.
With this language we may enumerate the grid class itself[^3], but we are interested instead in the simple permutations. The additional rules for the words encoding simple permutations of length at least four are:
[ccc]{}
(-6,0)(120,88) (80,20)(100,60) (0,0)(80,80) (80,40)(120,0) (80,40)(120,80) (0,0)(0,80) (40,0)(40,80) (80,0)(80,80) (120,0)(120,80) (0,0)(120,0) (0,40)(120,40) (0,80)(120,80) (-6,2)(-6,38) (-6,42)(-6,78) (2,86)(38,86) (42,86)(78,86) (82,86)(118,86) (40,0)[${\textsf{a}}$]{} (80,40)[${\textsf{b}}$]{} (80,0)[${\textsf{c}}$]{} (120,40)[${\textsf{d}}$]{}
&&
(-6,0)(120,88) (20,20)(100,60) (0,0)(80,80) (80,40)(120,0) (80,40)(120,80) (0,0)(0,80) (40,0)(40,80) (80,0)(80,80) (120,0)(120,80) (0,0)(120,0) (0,40)(120,40) (0,80)(120,80) (-6,2)(-6,38) (-6,42)(-6,78) (2,86)(38,86) (42,86)(78,86) (82,86)(118,86) (40,0)[${\textsf{a}}$]{} (80,40)[${\textsf{b}}$]{} (80,0)[${\textsf{c}}$]{} (120,40)[${\textsf{d}}$]{}
- To prevent intervals solely contained within an individual cell, we prohibit repetitions ${\textsf{a}}{\textsf{a}}$, ${\textsf{b}}{\textsf{b}}$, ${\textsf{c}}{\textsf{c}}$, or ${\textsf{d}}{\textsf{d}}$ as factors.
- To prevent intervals of the form shown in the first pane of Figure \[fig-4312-3142-grid-intervals\], we forbid words beginning with $\{{\textsf{c}},{\textsf{d}}\}^2$.
- To prevent intervals of the form shown in the second pane of Figure \[fig-4312-3142-grid-intervals\], we forbid words of the form $\{{\textsf{a}},{\textsf{b}},{\textsf{c}},{\textsf{d}}\}^\ast\{{\textsf{a}},{\textsf{c}},{\textsf{d}}\}^+$.
With these restrictions, we can then use the package [@delgado:automata-----a-:] for [@:gap----groups-a:] to count the simple permutations of this grid class[^4]. For future reference, we record that the multivariate generating function for these words of length at least four which begin with ${\textsf{a}}$ is $$s({x_{\textsf{a}}},{x_{\textsf{b}}},{x_{\textsf{c}}},{x_{\textsf{d}}})=\frac{{x_{\textsf{a}}}{x_{\textsf{b}}}{x_{\textsf{c}}}{x_{\textsf{d}}}}{1-{x_{\textsf{a}}}{x_{\textsf{c}}}-{x_{\textsf{b}}}{x_{\textsf{d}}}-{x_{\textsf{c}}}{x_{\textsf{d}}}-{x_{\textsf{a}}}{x_{\textsf{c}}}{x_{\textsf{d}}}- {x_{\textsf{b}}}{x_{\textsf{c}}}{x_{\textsf{d}}}},$$ while the words of length at least four which begin with ${\textsf{c}}$ have multivariate generating function ${x_{\textsf{c}}}s({x_{\textsf{a}}},{x_{\textsf{b}}},{x_{\textsf{c}}},{x_{\textsf{d}}})$. Note that our rules preclude words encoding simple permutations from beginning with ${\textsf{b}}$ or ${\textsf{d}}$.
Now we characterise the inflations. Because $3142$ is simple, it will not occur when inflating a $3142$-avoiding permutation by $3142$-avoiding intervals, so we need only avoid $4312$. Since the class is sum closed, we have that $f_\oplus=f^2/(1+f)$, as in Section \[sec-4213-3142\]. The skew decomposable permutations are a bit more complicated, but divide into a union: $$\left({\operatorname{Av}}(21)\ominus{\operatorname{Av}}(312)\right)\cup\left({\operatorname{Av}}_{\not\ominus}(4312,3142)\ominus{\operatorname{Av}}(12)\right),$$ where ${\operatorname{Av}}_{\not\ominus}(4312,3142)$ denotes the set of skew indecomposable permutations in this class. As the intersection of these two is simply ${\operatorname{Av}}(21)\ominus{\operatorname{Av}}(12)$, $f_\ominus=mc+(f-f_\ominus)m-m^2$, where $$m=\frac{x}{1-x}$$ denotes the generating function for the nonempty decreasing (or, increasing) permutations. Solving this shows $$f_\ominus
=
\frac{m(f+c-m)}{1+m}.$$ Inflations of simple permutation of length at least four are a bit more complicated, as there are several cases. In all such inflations, each entry which corresponds to a ${\textsf{b}}$ may only be inflated by an increasing permutation (but may be inflated by any such permutation), while each entry which corresponds to a ${\textsf{d}}$ may only be inflated by a $312$-avoiding permutation (but may be inflated by any such permutation). If the word begins with a ${\textsf{c}}$, then it follows that the entry corresponding to the first ${\textsf{c}}$ may be inflated by any permutation in ${\operatorname{Av}}(312)$, while each subsequent entry corresponding to a ${\textsf{c}}$ may only be inflated by a decreasing permutation. Otherwise, it follows from our rules that the word must begin with an ${\textsf{a}}$, and there are two cases. If the entry corresponding to the first ${\textsf{a}}$ is inflated by a permutation containing a descent, then each entry corresponding to a ${\textsf{c}}$ must be inflated with a decreasing permutation. Otherwise, if the entry corresponding to the first ${\textsf{a}}$ is inflated by an increasing permutation, then the entry corresponding to the first ${\textsf{c}}$ must be inflated by a permutation from ${\operatorname{Av}}(312)$, while each subsequent entry corresponding to a ${\textsf{c}}$ must be inflated by a decreasing permutation. From our multivariate generating function for these simple permutations, it follows that the contribution of their inflations is $$\left(\frac{f-m}{f}+\frac{c}{f}+\frac{c}{m}\right)s(f,m,m,c)
=
\frac{cm^2(c-m+f+cf)}{1-2cm-cm^2-mf-cmf}.$$ Combining this with the generating functions for $f_\oplus$ and $f_\ominus$ and solving for $f$ yields the generating function for the class (or, rather, its minimal polynomial).
\[thm-4312-3142\] The generating function $f$ for ${\operatorname{Av}}(4312,3142)$ satisfies $$\begin{array}{rclcc}
(x^3-2x^2+x)f^4
&+&
(4x^3-9x^2+6x-1)f^3&&\\
&+&
(6x^3-12x^2+7x-1)f^2&&\\
&+&
(4x^3-5x^2+x)f&&\\
&+&
x^3
&=&0.
\end{array}$$
The first several terms of this sequence are $$1,2,6,22,88,367,1568,6810,29943,132958,595227, 2683373,12170778,55499358,$$ sequence [A165538]{} in the [OEIS [@sloane:the-on-line-enc:]]{}. Though the form of the equation for $f$ is complicated, the close link with the Catalan numbers which can be seen in the previous development is enough to ensure that the radius of convergence is quite simple, exactly $1/5$, and in particular $f_n^{1/n} \to 5$.
Example \#3: Avoiding $4231$ and $3124$ {#sec-4231-3124}
=======================================
Before our final example we review a well-studied class. A permutation is [*layered*]{} if it is the direct sum of decreasing permutations (these decreasing permutations are called the [*layers*]{}). The class of [*layered permutations*]{} has the basis $\{312, 231\}$. To restrict the number of layers, we merely need to add an additional restriction, of the form $12\cdots k$.
\[prop-4231-3124-simples\] The simple permutations of ${\operatorname{Av}}(4231, 3124)$ and ${\operatorname{Geom}}{\mbox{\begin{footnotesize}$\left(\begin{array}{rrr}0&1&-1\\1&-1&0\end{array}\right)$\end{footnotesize}}}$ coincide.
First, it is straightforward to observe that $${\operatorname{Geom}}{\mbox{\begin{footnotesize}$\left(\begin{array}{rrr}0&1&-1\\1&-1&0\end{array}\right)$\end{footnotesize}}}
=
{\operatorname{Grid}}{\mbox{\begin{footnotesize}$\left(\begin{array}{rrr}0&1&-1\\1&-1&0\end{array}\right)$\end{footnotesize}}}
\subseteq{\operatorname{Av}}(4231, 3124),$$ so it suffices to prove that the simple permutations of ${\operatorname{Av}}(4231, 3124)$ are contained in this grid class.
Consider a simple permutation $\pi\in{\operatorname{Av}}(4231,3124)$ of length $n$. We analyse the entries of $\pi$ to the left of $n$ and to the right of $n$ separately, beginning with the entries on the right.
The entries to the right of $n$ form a layered permutation with at most two layers.
By the $4231$-avoidance of $\pi$ the entries to the right of $n$ avoid $231$, so to show that they are layered it suffices to show that they also avoid $312$. Suppose otherwise. Among all occurrences of $312$ choose one, $cab$, in which the ‘$1$’ and ‘$2$’ are as close together in position as possible (they will in fact be adjacent). Since $\pi$ is simple, $\{a,b\}$ cannot be an interval and thus (because they are adjacent) must be separated vertically. We claim that there is at least one such separator to the left of $n$. Let $x$ denote an arbitrary separator of $\{a,b\}$. We see that $x$ cannot lie horizontally between $n$ and $c$ by $4231$-avoidance. If $x$ were to lie horizontally between $c$ and $a$, then $\{x,a\}$ cannot be separated horizontally anywhere, nor vertically to the right of $n$, owing to the avoidance conditions, so must be separated vertically to the left of $n$. This separator therefore separates $\{a,b\}$ vertically to the left of $n$, as desired. The only other case is if $x$ lies to the right of $b$. In this case choose $x$ to be the bottommost such separator. Then $\{b,x\}$ must be separated. This can only occur to the left of $n$ (giving the separator we desire) or to the right of $x$. In this latter case it is easy to see that the region which consists of those points from $b$ to the right and which are also vertically between $x$ and $b$ (which contains $b$, $x$, and this new separator) can only be separated to the left of $n$, again giving the separator we desired. Therefore in all cases we may assume that there is an entry, $x$, to the left of $n$ which vertically separates $a$ and $b$. This situation is depicted on the left of Figure \[fig-4231-3124-claim-a\].
Now consider $\{n,c\}$. From Figure \[fig-4231-3124-claim-a\], we see that these entries can only possibly be separated vertically by an entry to the left of $n$ and to the right of $x$. Choose the leftmost such separator and label it $v$. We now have the situation depicted in the centre of Figure \[fig-4231-3124-claim-a\]. However, it is now clear that the entries $\{c,n,v\}$ lie in a proper interval, contradicting the simplicity of $\pi$. This contradiction shows that the entries to the right of $n$ must form a layered permutation.
[ccc]{}
(0,0)(60,50) (0,40)(10,40)(10,50)(0,50) (10,0)(20,0)(20,10)(10,10) (20,0)(30,0)(30,10)(20,10) (20,10)(30,10)(30,20)(20,20) (20,20)(30,20)(30,30)(20,30) (20,30)(30,30)(30,40)(20,40) (30,0)(40,0)(40,10)(30,10) (30,40)(40,40)(40,50)(30,50) (40,0)(50,0)(50,10)(40,10) (40,10)(50,10)(50,20)(40,20) (40,10)(50,20) (40,20)(50,20)(50,30)(40,30) (40,30)(50,30)(50,40)(40,40) (40,40)(50,40)(50,50)(40,50) (50,0)(60,0)(60,10)(50,10) (50,40)(60,40)(60,50)(50,50) (10,20)[2.0]{} (10,20)[$x$]{} (20,50)[2.0]{} (20,50)[$n$]{} (30,40)[2.0]{} (30,40)[$c$]{} (40,10)[2.0]{} (40,10)[$a$]{} (50,30)[2.0]{} (50,30)[$b$]{} (0,0)(0,50) (0,0)(60,0) (10,0)(10,50) (0,10)(60,10) (20,0)(20,50) (0,20)(60,20) (30,0)(30,50) (0,30)(60,30) (40,0)(40,50) (0,40)(60,40) (50,0)(50,50) (0,50)(60,50) (60,0)(60,50)
&&
(0,0)(70,60) (0,40)(10,40)(10,50)(0,50) (0,50)(10,50)(10,60)(0,60) (10,0)(20,0)(20,10)(10,10) (10,40)(20,40)(20,50)(10,50) (10,50)(20,50)(20,60)(10,60) (20,0)(30,0)(30,10)(20,10) (20,10)(30,10)(30,20)(20,20) (20,20)(30,20)(30,30)(20,30) (20,30)(30,30)(30,40)(20,40) (30,0)(40,0)(40,10)(30,10) (30,10)(40,10)(40,20)(30,20) (30,20)(40,20)(40,30)(30,30) (30,30)(40,30)(40,40)(30,40) (40,0)(50,0)(50,10)(40,10) (40,40)(50,40)(50,50)(40,50) (40,50)(50,50)(50,60)(40,60) (50,0)(60,0)(60,10)(50,10) (50,10)(60,10)(60,20)(50,20) (50,10)(60,20) (50,20)(60,20)(60,30)(50,30) (50,30)(60,30)(60,40)(50,40) (50,40)(60,40)(60,50)(50,50) (50,50)(60,50)(60,60)(50,60) (60,0)(70,0)(70,10)(60,10) (60,40)(70,40)(70,50)(60,50) (60,50)(70,50)(70,60)(60,60) (10,20)[2.0]{} (10,20)[$x$]{} (20,50)[2.0]{} (20,50)[$v$]{} (30,60)[2.0]{} (30,60)[$n$]{} (40,40)[2.0]{} (40,40)[$c$]{} (50,10)[2.0]{} (50,10)[$a$]{} (60,30)[2.0]{} (60,30)[$b$]{} (0,0)(0,60) (0,0)(70,0) (10,0)(10,60) (0,10)(70,10) (20,0)(20,60) (0,20)(70,20) (30,0)(30,60) (0,30)(70,30) (40,0)(40,60) (0,40)(70,40) (50,0)(50,60) (0,50)(70,50) (60,0)(60,60) (0,60)(70,60) (70,0)(70,60)
Having established that these entries are layered, it is easy to see that there are at most two layers. Otherwise the entries to the right of $n$ would contain a copy of $123$. The ‘$2$’ and ‘$3$’ in this copy of $123$ must be separated by an entry to the left of $n$ (because the entries to the right of $n$ form a layered permutation), but this would create a copy of $3124$. This contradiction completes the proof of Claim \[prop-4231-3124-simples\].a.
The entries to the left of $n$ and above $\pi(n)$ are increasing.
Let $a=\pi(n)$ and suppose to the contrary that the entries to the left of $n$ and above $a$ contain an inversion. Choose such an inversion $yx$ with $y$ as far left as possible and $x$ as close to $y$ as possible. This gives the situation depicted on the left of Figure \[fig-4231-3124-claim-b\]. As can be seen in this diagram, $\{x,y\}$ could only possibly be separated horizontally. Let $z$ denote a topmost such separator. This gives the situation depicted on the right of Figure \[fig-4231-3124-claim-b\]. However, as this diagram indicates, $y$ and $z$ now belong to a proper interval, contradicting the simplicity of $\pi$.
[ccc]{}
(0,0)(40,40) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (10,0)(20,0)(20,10)(10,10) (10,10)(20,10)(20,20)(10,20) (10,20)(20,20)(20,30)(10,30) (20,20)(30,20)(30,30)(20,30) (20,20)(30,30) (30,20)(40,20)(40,30)(30,30) (10,30)[2.0]{} (10,30)[$y$]{} (20,20)[2.0]{} (20,20)[$x$]{} (30,40)[2.0]{} (30,40)[$n$]{} (40,10)[2.0]{} (40,10)[$a$]{} (0,0)(0,40) (0,0)(40,0) (10,0)(10,40) (0,10)(40,10) (20,0)(20,40) (0,20)(40,20) (30,0)(30,40) (0,30)(40,30) (40,0)(40,40) (0,40)(40,40)
&&
(0,0)(50,50) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (0,40)(10,40)(10,50)(0,50) (10,0)(20,0)(20,10)(10,10) (10,10)(20,10)(20,20)(10,20) (10,20)(20,20)(20,30)(10,30) (10,40)(20,40)(20,50)(10,50) (20,0)(30,0)(30,10)(20,10) (20,10)(30,10)(30,20)(20,20) (20,20)(30,20)(30,30)(20,30) (20,40)(30,40)(30,50)(20,50) (30,20)(40,20)(40,30)(30,30) (30,20)(40,30) (30,30)(40,30)(40,40)(30,40) (40,20)(50,20)(50,30)(40,30) (40,30)(50,30)(50,40)(40,40) (10,30)[2.0]{} (10,30)[$y$]{} (20,40)[2.0]{} (20,40)[$z$]{} (30,20)[2.0]{} (30,20)[$x$]{} (40,50)[2.0]{} (40,50)[$n$]{} (50,10)[2.0]{} (50,10)[$a$]{} (0,0)(0,50) (0,0)(50,0) (10,0)(10,50) (0,10)(50,10) (20,0)(20,50) (0,20)(50,20) (30,0)(30,50) (0,30)(50,30) (40,0)(40,50) (0,40)(50,40) (50,0)(50,50) (0,50)(50,50)
Let $c$ denote the leftmost entry of $\pi$ greater than $\pi(n)$ (note that $c$ may equal $n$). The entries to the left of $c$ (which lie below $\pi(n)$ by Claim \[prop-4231-3124-simples\].b) are increasing.
The proof follows well-travelled lines. Suppose to the contrary that there is an inversion among these entries, and choose one such $ba$ where $b$ is as far left as possible, and $a$ is as small as possible. The cell bounded by $\{a,b\}$ can only be split above, and we may choose a split point $d$ which is as large as possible. Now the box bounded by $\{a,b,d\}$ defines a proper interval, a contradiction.
We are now in position to complete the proof of the proposition, after a brief recap of the structure we have established. Let $b=\pi(n)$, let $c$ denote the leftmost entry which is greater than $b$ (note that $c=n$ is a possibility), and finally let $a$ denote the bottommost entry which lies horizontally between $c$ and $n$ (such an entry need not exist, but this does not affect the argument). We then have the situation depicted on the left of Figure \[fig-4231-3124-finale\].
[ccccccc]{}
(0,0)(40,40) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (10,0)(20,0)(20,10)(10,10) (10,0)(20,10) (10,20)(20,20)(20,30)(10,30) (20,0)(30,0)(30,10)(20,10) (20,10)(30,10)(30,20)(20,20) (20,20)(30,20)(30,30)(20,30) (10,30)[2.0]{} (10,30)[$c$]{} (20,10)[2.0]{} (20,10)[$a$]{} (30,40)[2.0]{} (30,40)[$n$]{} (40,20)[2.0]{} (40,20)[$b$]{} (5,10)[$R_1$]{} (20,35)[$R_2$]{} (15,15)[$R_3$]{} (35,25)[$R_4$]{} (35,15)[$R_5$]{} (35,35)[$R_6$]{} (35,5)[$R_7$]{} (0,0)(0,40) (0,0)(40,0) (10,0)(10,40) (10,10)(40,10) (20,0)(20,30) (0,20)(40,20) (30,0)(30,40) (0,30)(40,30) (40,0)(40,40) (0,40)(40,40)
&&
(0,0)(50,50) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (0,40)(10,40)(10,50)(0,50) (10,0)(20,0)(20,10)(10,10) (10,20)(20,20)(20,30)(10,30) (10,30)(20,30)(20,40)(10,40) (20,0)(30,0)(30,10)(20,10) (20,10)(30,10)(30,20)(20,20) (20,20)(30,20)(30,30)(20,30) (20,30)(30,30)(30,40)(20,40) (30,10)(40,10)(40,20)(30,20) (30,20)(40,20)(40,30)(30,30) (30,30)(40,30)(40,40)(30,40) (40,0)(50,0)(50,10)(40,10) (40,30)(50,30)(50,40)(40,40) (40,40)(50,40)(50,50)(40,50) (10,40)[2.0]{} (10,40)[$c$]{} (20,10)[2.0]{} (20,10)[$a$]{} (30,50)[2.0]{} (30,50)[$n$]{} (40,30)[2.0]{} (40,30)[$x$]{} (50,20)[2.0]{} (50,20)[$b$]{} (0,0)(0,50) (0,0)(50,0) (10,0)(10,50) (10,10)(50,10) (20,0)(20,40) (0,20)(50,20) (30,0)(30,50) (0,30)(50,30) (40,0)(40,50) (0,40)(50,40) (50,0)(50,50) (0,50)(50,50)
&&
(0,0)(40,40) (0,20)(10,20)(10,30)(0,30) (0,30)(10,30)(10,40)(0,40) (10,0)(20,0)(20,10)(10,10) (10,0)(20,10) (10,20)(20,20)(20,30)(10,30) (20,0)(30,0)(30,10)(20,10) (20,10)(30,10)(30,20)(20,20) (20,20)(30,20)(30,30)(20,30) (30,10)(40,10)(40,20)(30,20) (30,20)(40,20)(40,30)(30,30) (10,30)[2.0]{} (20,10)[2.0]{} (30,40)[2.0]{} (40,20)[2.0]{} (0,0)(0,40) (0,0)(40,0) (10,0)(10,40) (10,10)(40,10) (20,0)(20,30) (0,20)(40,20) (30,0)(30,40) (0,30)(40,30) (40,0)(40,40) (0,40)(40,40) (0.5,1)(10,20) (10,20)(20,10) (10,30)(30,40) (30,10)(35,0) (35,40)(40,30)
The first three labeled regions are further restricted as follows:
- $R_1$ must be increasing by Claim \[prop-4231-3124-simples\].c.
- $R_2$ must be increasing by Claim \[prop-4231-3124-simples\].b.
- $R_3$ must be decreasing because $\pi$ avoids $3124$.
Next we claim that $R_4$ and $R_5$ are both empty. First suppose to the contrary that $R_4$ is nonempty, and take $x$ to be the topmost entry in this region. We then have the situation depicted in the centre of Figure \[fig-4231-3124-finale\], which shows that $\{b,x\}$ can only be separated by an entry in $R_5$. Let $y$ denote the bottommost such separator. It can then be seen that there is no way to separate $\{b,x,y\}$. Showing that $R_5$ is empty is very similar. Suppose to the contrary that this region is nonempty and let $x$ denote the bottommost entry in the region. It can be seen that there are two ways to separate $\{b,x\}$: vertically with an entry in $R_5$ or horizontally with an entry in $R_4$. In each case, though, these new entries cannot be separated from either $b$ or $x$.
Finally, regions $R_6$ and $R_7$ must both be decreasing because $\pi$ avoids $4231$ and $3124$ respectively. Then Claim \[prop-4231-3124-simples\].a shows that regions $R_5$ and $R_6$ must together form a layered permutation with at most two layers. The structure of $\pi$ is now displayed on the right of Figure \[fig-4231-3124-finale\], which shows that $\pi$ does indeed lie in the grid class desired, completing the proof.
Having restricted our attention to this (geometric) grid class, we now seek to place it in bijection with a regular language, following points (G1) and (G2) of Section \[sec-4312-3142\]. The first is the easiest to deal with since we need only forbid factors of the forms $\{{\textsf{c}},{\textsf{d}}\}^+{\textsf{a}}$ and ${\textsf{d}}^+{\textsf{b}}$.
(-3,0)(120,85) (0,0)(40,40) (40,40)(80,80) (40,40)(80,0) (80,80)(120,40) (0,0)(0,80) (40,0)(40,80) (80,0)(80,80) (120,0)(120,80) (0,0)(120,0) (0,40)(120,40) (0,80)(120,80) (-3,2)(-3,38) (-3,42)(-3,78) (2,83)(38,83) (42,83)(78,83) (82,83)(118,83) (40,0)[${\textsf{a}}$]{} (40,0)[${\textsf{b}}$]{} (80,40)[${\textsf{c}}$]{} (80,40)[${\textsf{d}}$]{}
To handle (G2), we use the same preference for $M$-griddings as in Section \[sec-4312-3142\]: among all $M$-griddings of a permutation, we prefer the one that has the most entries in the first column, then the most entries in the second column, and then the most entries in the first row. The words that correspond to these preferred griddings are those which do not begin with ${\textsf{b}}$ or ${\textsf{a}}^\ast{\textsf{c}}$, do not end with ${\textsf{d}}$, and are not of the forms ${\textsf{d}}\{{\textsf{a}},{\textsf{b}}\}^\ast$ or ${\textsf{a}}^\ast\{{\textsf{c}},{\textsf{d}}\}^+$.
This language allows us to enumerate the grid class itself[^5], and now we restrict to encoding the simple permutations.
[ccccc]{}
(-6,0)(120,88) (20,20)(120,80) (0,0)(40,40) (40,40)(80,80) (40,40)(80,0) (80,80)(120,40) (0,0)(0,80) (40,0)(40,80) (80,0)(80,80) (120,0)(120,80) (0,0)(120,0) (0,40)(120,40) (0,80)(120,80) (-6,2)(-6,38) (-6,42)(-6,78) (2,86)(38,86) (42,86)(78,86) (82,86)(118,86) (40,0)[${\textsf{a}}$]{} (40,0)[${\textsf{b}}$]{} (80,40)[${\textsf{c}}$]{} (80,40)[${\textsf{d}}$]{}
&&
(-6,0)(120,88) (60,60)(100,80) (0,0)(40,40) (40,40)(80,80) (40,40)(80,0) (80,80)(120,40) (0,0)(0,80) (40,0)(40,80) (80,0)(80,80) (120,0)(120,80) (0,0)(120,0) (0,40)(120,40) (0,80)(120,80) (-6,2)(-6,38) (-6,42)(-6,78) (2,86)(38,86) (42,86)(78,86) (82,86)(118,86) (40,0)[${\textsf{a}}$]{} (40,0)[${\textsf{b}}$]{} (80,40)[${\textsf{c}}$]{} (80,40)[${\textsf{d}}$]{}
&&
(-6,0)(120,88) (20,20)(60,60) (0,0)(40,40) (40,40)(80,80) (40,40)(80,0) (80,80)(120,40) (0,0)(0,80) (40,0)(40,80) (80,0)(80,80) (120,0)(120,80) (0,0)(120,0) (0,40)(120,40) (0,80)(120,80) (-6,2)(-6,38) (-6,42)(-6,78) (2,86)(38,86) (42,86)(78,86) (82,86)(118,86) (40,0)[${\textsf{a}}$]{} (40,0)[${\textsf{b}}$]{} (80,40)[${\textsf{c}}$]{} (80,40)[${\textsf{d}}$]{}
- To prevent intervals contained within individual cells, we prohibit repetitions ${\textsf{a}}{\textsf{a}}$, ${\textsf{b}}{\textsf{b}}$, ${\textsf{c}}{\textsf{c}}$, ${\textsf{d}}{\textsf{d}}$ as factors.
- To prevent intervals of the form shown in the first pane of Figure \[fig-4231-3124-grid-intervals\], we insist that the last ${\textsf{a}}$ is followed by a ${\textsf{b}}$, i.e., we prohibit words which end in ${\textsf{a}}\{{\textsf{c}},{\textsf{d}}\}^\ast$,
- To prevent intervals of the form shown in the second pane of Figure \[fig-4231-3124-grid-intervals\], we require that there is a ${\textsf{b}}$ after the last letter in $\{{\textsf{c}},{\textsf{d}}\}$; by the previous two rules, this means we need only prohibit words which end in ${\textsf{c}}{\textsf{d}}$ or ${\textsf{d}}{\textsf{c}}$.
- To prevent intervals of the form shown in the third pane of Figure \[fig-4231-3124-grid-intervals\], we prohibit words which begin with two or more letters from $\{{\textsf{a}},{\textsf{b}},{\textsf{c}}\}$.
Finally, we exclude the word ${\textsf{d}}{\textsf{c}}{\textsf{b}}$, which is mapped by ${\varphi}$ to the nonsimple permutation $312$.
We can then compute (again using the package [@delgado:automata-----a-:] for [@:gap----groups-a:]) that the multivariate generating function (counting occurrences of each letter) for the simple permutations in this geometric grid class is $$s({x_{\textsf{a}}},{x_{\textsf{b}}},{x_{\textsf{c}}},{x_{\textsf{d}}})=\frac{{x_{\textsf{b}}}{x_{\textsf{c}}}{x_{\textsf{d}}}({x_{\textsf{a}}}+{x_{\textsf{c}}}+{x_{\textsf{a}}}{x_{\textsf{b}}}+{x_{\textsf{a}}}{x_{\textsf{c}}}+{x_{\textsf{b}}}{x_{\textsf{c}}}+{x_{\textsf{c}}}{x_{\textsf{d}}}+{x_{\textsf{a}}}{x_{\textsf{b}}}{x_{\textsf{c}}}+{x_{\textsf{b}}}{x_{\textsf{c}}}{x_{\textsf{d}}})}
{1-{x_{\textsf{a}}}{x_{\textsf{b}}}-{x_{\textsf{b}}}{x_{\textsf{c}}}-{x_{\textsf{c}}}{x_{\textsf{d}}}-{x_{\textsf{a}}}{x_{\textsf{b}}}{x_{\textsf{c}}}-{x_{\textsf{b}}}{x_{\textsf{c}}}{x_{\textsf{d}}}}.$$ It remains only to determine the allowed inflations. First, the sum decomposable permutations all have unique representations in the form ${\operatorname{Av}}_{\not\oplus}(312)\oplus {\operatorname{Av}}(4231,3124)$, where ${\operatorname{Av}}_{\not\oplus}(312)$ denotes the sum indecomposable permutations in ${\operatorname{Av}}(312)$. As ${\operatorname{Av}}_{\not\oplus}(312)$ is well-known to be enumerated by the shifted Catalan numbers, we get $$f_\oplus=(xc+x)f.$$
The skew decomposable permutations are of the form ${\operatorname{Av}}(312)\ominus {\operatorname{Av}}(231,3124)$, and so (deviating slightly from our usual conventions) they all have a unique representation of the form ${\operatorname{Av}}(312)\ominus {\operatorname{Av}}_{\not\ominus}(231,3124)$. The skew indecomposable permutations in ${\operatorname{Av}}(231,3124)$ are counted by the Fibonacci numbers of odd index ($1$, $1$, $3$, $8$, $21$, $\dots$), giving that $$f_\ominus=\frac{x-2x^2+x^2}{1-3x+x^2}c.$$ We now come to inflations of simple permutations of length at least four. Each entry corresponding to an ${\textsf{a}}$ or ${\textsf{c}}$ may be inflated by an arbitrary member of ${\operatorname{Av}}(312)$. Each entry corresponding to a ${\textsf{b}}$ or non-initial ${\textsf{d}}$ may be inflated only by a decreasing interval. Finally, the first entry corresponding to a ${\textsf{d}}$ (and we see from the multivariate generating function $s$ that every word encoding simple permutations in this class contains at least one ${\textsf{d}}$) may be inflated by an arbitrary member of ${\operatorname{Av}}(231,3124)$, a class counted by the Fibonacci numbers of even index ($1$, $2$, $5$, $13$, $34$, $\dots$). Therefore, the generating function we are interested in is given by $$f
=
(xc+x)f
+
\frac{x-2x^2+x^2}{1-3x+x^2}c
+
\frac{s(c,m,c,m)}{m}\frac{x-x^2}{1-3x+x^2}.$$ As this equation is linear in $f$, it is trivial to obtain our final result.
The generating function for ${\operatorname{Av}}(4231, 3124)$ is $$\frac{1-8x+20x^2-20x^3+10x^4-2x^5 - (1-4x+2x^2)\sqrt{1-4x}}
{2(1-3x+x^2)(-1+5x-4x^2+x^3)}.
$$
The first several terms of this sequence are $$1, 2, 6, 22, 88, 363, 1508, 6255, 25842, 106327, 435965, 1782733, 7275351, 29648647,$$ sequence [A165535]{} in the [OEIS [@sloane:the-on-line-enc:]]{}. The radius of convergence is the smallest positive root of the cubic factor in the denominator, approximately 0.2451, and hence $f_n^{1/n} \to 4.0796\dots$
Conclusion {#sec-infinite-conclusion}
==========
It should be noted that the three examples presented in this paper are not the only $2\times 4$ classes which have been enumerated using these techniques. In [@albert:counting-1324-4:], the present authors used a precursor of this approach to enumerate ${\operatorname{Av}}(4231,1324)$; in the language of this paper, they proved that the simple permutations in this class are contained in $${\operatorname{Geom}}{\mbox{\begin{footnotesize}$\left(\begin{array}{rrr}1&0&-1\\0&\bullet&0\\-1&0&1\end{array}\right)$\end{footnotesize}}}.$$ (Here the $\bullet$ entry denotes a cell in the standard figure filled with a unique point; this notion is formally defined in [@albert:geometric-grid-: Section 10].)
Finally, we point out that there may be other $2\times 4$ classes to which these techniques apply. While there is a decision procedure to determine whether a given class lies in a monotone grid class (see Huczynska and Vatter [@huczynska:grid-classes-an:]), there is no known procedure to determine whether a given class lies in a *geometric* grid class (and indeed, there are indications that this question may be quite difficult). Needless to say, deciding whether the simple permutations of a given class lie in a geometric grid class is expected to be more difficult still.
[10]{}
Simple permutations and pattern restricted permutations. , 1-3 (2005), 1–15.
Geometric grid classes of permutations. *Trans. Amer. Math. Soc.*, to appear.
The enumeration of permutations avoiding $2143$ and $4231$. (2011), 87–98.
The enumeration of three pattern classes using monotone grid classes. , 3 (2012), P20 (34 pp.).
Counting $1324$, $4231$-avoiding permutations. , 1 (2009), Research Paper 136, 9.
Inflations of geometric grid classes of permutations. arXiv:1202.1833v1 \[math.CO\].
Restricted permutations. , 1-3 (1999), 27–38.
The permutation classes equinumerous to the smooth class. (1998), Research paper 31, 12 pp.
Simple permutations and algebraic generating functions. , 3 (2008), 423–441.
Automata — a [GAP]{} package, [V]{}ersion 1.10. `http://www.gap-system.org/Packages/automata.html`, 2007.
. Cambridge University Press, Cambridge, 2009.
. , 2012.
Grid classes and the [F]{}ibonacci dichotomy for restricted permutations. (2006), R54, 14 pp.
Permutations with forbidden subsequences and a generalized [S]{}chröder number. , 1-3 (2000), 121–130.
Postscript: “[P]{}ermutations with forbidden subsequences and a generalized [S]{}chröder number”. , 1-3 (2003), 333–334.
Finite transition matrices for permutations avoiding pairs of length four patterns. , 1-3 (2003), 171–183.
Wilf classes of pairs of permutations of length $4$. (2005), R25, 27 pp.
Profile classes and partial well-order for permutations. , 2 (2003), R17, 30 pp.
Critically indecomposable partially ordered sets, graphs, tournaments and other binary relational structures. , 1-3 (1993), 191–205.
The [O]{}n-line [E]{}ncyclopedia of [I]{}nteger [S]{}equences. Available online at [http://www.research.att.com/\~njas/sequences/]{}.
Finitely labeled generating trees and restricted permutations. , 5 (2006), 559–572.
. Enumerations of specific permutation classes — [W]{}ikipedia[,]{} [T]{}he [F]{}ree [E]{}ncyclopedia, 2012. .
[^1]: Vatter’s research was sponsored by the National Security Agency under Grant Number H98230-12-1-0207. The United States Government is authorized to reproduce and distribute reprints not-withstanding any copyright notation herein.
[^2]: Note that in order for the cells of the matrix $M$ to be compatible with plots of permutations, we use Cartesian coordinates for our matrices, indexing them first by column, from left to right starting with $1$, and then by row, from bottom to top.
[^3]: This grid class (which, because it can be viewed as a “juxtaposition” in the sense of Atkinson [@atkinson:restricted-perm:], can be shown to have basis $\{2143,3142,4132,4312\})$ has the generating function $$\frac{1-6x+11x^2-5x^3}{(1-x)(1-3x)(1-3x+x^2)}.
$$
[^4]: These simple permutations have the generating function $$\frac{x+x^2-4x^3-3x^3}{(1+x)(1-2x)},$$ showing that for $n\ge 3$ the simple permutations in this class are counting by the Jacobsthal numbers ([A001045]{} in the [OEIS [@sloane:the-on-line-enc:]]{}).
[^5]: The generating function for this grid class is $$\frac{1-5x+7x^2-x^3}{(1-x)(1-2x)(1-3x)},
$$ (sequence [A083323]{} in the [OEIS [@sloane:the-on-line-enc:]]{}). Our computations suggest that the basis of this class is $$\{4312, 4231, 4123, 3124, 32541, 21534, 21435\},
$$ but we have not verified this with a formal proof.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper we discuss some special generalizations of equationally Noetherian property which naturally arise in the universal algebraic geometry. We introduce weakly equationally Noetherian, ${{\mathrm{q}_\omega}}$-compact, ${{\mathrm{u}_\omega}}$-compact, and weakly ${{\mathrm{u}_\omega}}$-compact algebras and then examine properties of such algebras. Also we consider the connections between five classes: the class of equationally Noetherian algebras, the class of weakly equationally Noetherian algebras, the class of ${{\mathrm{u}_\omega}}$-compact algebras, the class of weakly ${{\mathrm{u}_\omega}}$-compact algebras, and the class of ${{\mathrm{q}_\omega}}$-compact algebras.
*Keywords:* Compactness Theorem, universal closure, quasivariety, algebraic structure, algebraic set, coordinate algebra, (weakly) ${{\mathrm{u}_\omega}}$-compact algebra, ${{\mathrm{q}_\omega}}$-compact algebra, (weakly) equationally Noetherian algebra, logically irreducible set.
*Mathematics Subject Classification:* 03C05+14A99+08B05
author:
- 'E.Daniyarova, A.Myasnikov, V.Remeslennikov'
title: 'Algebraic geometry over algebraic structures III: Equationally Noetherian property and compactness'
---
Introduction
============
This paper deals with the [*universal algebraic geometry*]{}. The universal algebraic geometry is a young branch of mathematics. The subject of universal algebraic geometry lies in the solutions of systems of equations over an arbitrary algebraic structure.
Investigations in universal algebraic geometry were started in works by B.I.Plotkin [@Plot1; @Plot2; @Plot3] and papers on algebraic geometry over groups by G.Baumslag, O.G.Kharlampovich, A.G.Myasnikov, and V.N.Remeslennikov [@KM1; @KM2; @KM3; @KM4]. After that there were a lot of papers on algebraic geometry over concrete groups, algebras, monoids and so on. Among them there are the famous works by O.G.Kharlampovich, A.G.Myasnikov [@KM1; @KM2; @KM3; @KM4] and Z.Sela [@Sela1; @Sela2; @Sela3] on algebraic geometry over free groups.
In recent years we have achieved more general and systematic point of view on the universal algebraic geometry as on a formalized theory. In this respect we have started a series of works on universal algebraic geometry. This paper is the third one of that series along with [@DMR1; @DMR2].
According to [@Gorbunov; @Malcev; @Marker], in [@DMR1] we give a framework of universal algebra and model theory as much as we need it in universal algebraic geometry. At the same time we discuss how notions and ideas from model theory work in universal algebraic geometry. In [@DMR2] we introduce the foundation of universal algebraic geometry, basic definitions and constructions of the algebraic geometry over an arbitrary algebraic structure $\B$.
This paper is supposed to be read after the previous ones [@DMR1; @DMR2], however for the sake of convenience we present in here some of the most essential notations and definitions (see Section \[sec:preliminaries\]).
We consider only first-order functional languages (signatures). Recall that algebraic structures in a functional language are called [*algebras*]{}. Typically we denote algebraic structures by capital calligraphic letters ($\A, \B, \Ce, \ldots$) and their universes by the corresponding capital Latin letters ($A, B, C,
\ldots$).
The main results of papers [@DMR1; @DMR2] are so-called the Unification Theorems (Theorem A and Theorem C) which give a description of coordinate algebras by means of several languages.
[**Theorem A.**]{}
*Let $\B$ be an equationally Noetherian algebra in a functional language ${{\mathtt{L}}}$. Then for a finitely generated algebra $\Ce$ of ${{\mathtt{L}}}$ the following conditions are equivalent:*
1. ${{\mathrm{Th}}}_{\forall} (\B) \subseteq {{\mathrm{Th}}}_{\forall} (\Ce)$, i.e., $\Ce \in {{\mathbf{Ucl}}}(\B)$;
2. ${{\mathrm{Th}}}_{\exists} (\B) \supseteq {{\mathrm{Th}}}_{\exists} (\Ce)$;
3. $\Ce$ embeds into an ultrapower of $\B$;
4. $\Ce$ is discriminated by $\B$;
5. $\Ce$ is a limit algebra over $\B$;
6. $\Ce$ is an algebra defined by a complete atomic type in the theory ${{\mathrm{Th}}}_{\forall} (\B)$ in ${{\mathtt{L}}}$;
7. $\Ce$ is the coordinate algebra of an irreducible algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$.
[**Theorem C.**]{}
*Let $\B$ be an equationally Noetherian algebra in a functional language ${{\mathtt{L}}}$. Then for a finitely generated algebra $\Ce$ of ${{\mathtt{L}}}$ the following conditions are equivalent:*
1. $\Ce \in {{\mathbf{Qvar}}}(\B)$, i.e., ${{\mathrm{Th}}}_{\rm qi} (\B) \subseteq {{\mathrm{Th}}}_{\rm qi} (\Ce)$;
2. $\Ce \in {{\mathbf{Pvar}}}(\B)$;
3. $\Ce$ embeds into a direct power of $\B$;
4. $\Ce$ is separated by $\B$;
5. $\Ce$ is a subdirect product of a finitely many limit algebras over $\B$;
6. $\Ce$ is an algebra defined by a complete atomic type in the theory ${{\mathrm{Th}}}_{\rm qi} (\B)$ in ${{\mathtt{L}}}$;
7. $\Ce$ is the coordinate algebra of an algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$.
Note that items 5) in both Theorem A and Theorem C give a description of coordinate algebras by means of limit algebras. The limit algebraic structures (groups, as the rule) become the object of intense interest in modern algebra [@CG; @GS; @Groves1; @Groves1-2; @Groves2-1; @Groves2; @G1].
Theorems A and C are formulated for so-called equationally Noetherian algebras (the definition see in Section \[sec:preliminaries\]). Equationally Noetherian algebras possess the best opportunity to study the algebraic geometry over them. If a given algebra $\B$ is equationally Noetherian then we have an advantage when investigating the algebraic geometry over $\B$. In this case we may use:
- Unification Theorems;
- the decomposition of any algebraic set over $\B$ into a finite union of irreducible algebraic sets (Theorem \[irr\] below);
- the possibility to study only finite system of equations;
- and some more results [@DMR1; @DMR2].
In the case when a given algebra $\B$ is not equationally Noetherian we lose some results for equationally Noetherian algebras, while some of them may remain in force. In this paper we introduce four generalizations of the equationally Noetherian property which naturally arise in universal algebraic geometry. These are
- weak equationally Noetherian property that retains (iii);
- ${{\mathrm{q}_\omega}}$-compactness that retains Unification Theorem C;
- ${{\mathrm{u}_\omega}}$-compactness that retains Unification Theorems A and C;
- and weak ${{\mathrm{u}_\omega}}$-compactness that retains (iv), namely, some weak form of Unification Theorem A.
We denote by ${{\mathbf{N}}}$ the class of all equationally Noetherian algebras in a given functional language ${{\mathtt{L}}}$. By ${{\mathbf{N}}}'$, ${{\mathbf{Q}}}$, ${{\mathbf{U}}}$, ${{\mathbf{U}}}'$, correspondingly, we denote the classes of algebras with properties above. The picture of connections between classes ${{\mathbf{Q}}}$, ${{\mathbf{U}}}$, ${{\mathbf{U}}}'$, ${{\mathbf{N}}}'$ and ${{\mathbf{N}}}$ is presented in Section \[sec:picture\].
There exist several equivalent approaches to ${{\mathrm{q}_\omega}}$- and ${{\mathrm{u}_\omega}}$-compact algebras. We introduce them in Section \[sec:compact\]. One of these approaches rises from some ideas of model theory. It relates to the Compactness Theorem and the notion of compact algebra.
Recall that a set of formulas $T$ in a language ${{\mathtt{L}}}$ is called [*satisfiable*]{} in a class ${{\mathbf{K}}}$ of algebraic structures in ${{\mathtt{L}}}$ (or $T$ is [*realized*]{} in ${{\mathbf{K}}}$) if one can assign some elements from a particular algebraic structure from ${{\mathbf{K}}}$ as values to the variables which occur in $T$ in such a way that all formulas from $T$ become true. The set $T$ is called [*finitely satisfiable*]{} in ${{\mathbf{K}}}$ if every finite subset of $T$ is realized in ${{\mathbf{K}}}$.
[**Compactness Theorem**]{} (K.F.Gödel, A.I.Malcev [@Gorbunov])[**.**]{} [*If a set of first-order formulas $T$ in a language ${{\mathtt{L}}}$ is finitely satisfiable in a class ${{\mathbf{K}}}$ of algebraic structures in ${{\mathtt{L}}}$, then $T$ is satisfiable in an ultraproduct of structures from ${{\mathbf{K}}}$.*]{}
Class ${{\mathbf{K}}}$ is called [*compact*]{} if every finitely satisfiable in ${{\mathbf{K}}}$ set of formulas $T$ is satisfiable in ${{\mathbf{K}}}$. This definition occurs in the book by V.A.Gorbunov [@Gorbunov]. It is natural to name an algebraic structure $\B$ compact if the class $\{\B\}$ is compact. However, according to W.Hodges [@Hodges], algebraic ${{\mathtt{L}}}$-structure is called compact if its universe is a Hausdorff topological space, in such a way that each function from ${{\mathtt{L}}}$ is interpreted by a continuous function. The same algebraic structures appear in [@Gorbunov] under the name of topologically compact structures.
Trying to avoid an ambiguity we call an ${{\mathtt{L}}}$-algebra $\B$ [*logically compact*]{} if every finitely satisfiable in $\B$ set of formulas $T$ in the language ${{\mathtt{L}}}$ is satisfiable in $\B$. When we modify this definition and consider only special types of sets of formulas $T$ we get definitions of special compactness, such as ${{\mathrm{q}_\omega}}$- and ${{\mathrm{u}_\omega}}$-compactness. Short review of the history of “${{\mathrm{q}_\omega}}$-compact” notion is represented in Subsection \[subsec:compact\].
First and foremost in this article we generalize the Unification Theorems to ${{\mathrm{u}_\omega}}$- and ${{\mathrm{q}_\omega}}$-compact algebras. In Section \[sec:unification\_theorems\_compact\] we give geometric definitions of ${{\mathrm{u}_\omega}}$- and ${{\mathrm{q}_\omega}}$-compactness. In Subsection \[subsec:A\] we prove that Theorem A is true for any ${{\mathrm{u}_\omega}}$-compact algebra $\B$ and every algebra $\B$ which satisfies Theorem A is ${{\mathrm{u}_\omega}}$-compact. The similar result that connects ${{\mathrm{q}_\omega}}$-compact algebras and Theorem C is presented in Subsection \[subsec:C\]. In Subsection \[subsec:weak\] for weakly ${{\mathrm{u}_\omega}}$-compact algebras we formulate and prove a weak analog of Theorem A.
Section \[sec:compact\] is devoted to ${{\mathrm{q}_\omega}}$- and ${{\mathrm{u}_\omega}}$-compact algebras. In Subsection \[subsec:compact\] we put definitions of ${{\mathrm{q}_\omega}}$- and ${{\mathrm{u}_\omega}}$-compact algebras in different equivalent forms and prove the equivalence of them in Subsection \[subsec:proof\]. For ${{\mathrm{u}_\omega}}$-compact algebra $\B$ Unification Theorems give a global view to all (irreducible) coordinate algebras over $\B$. However, it may happen that one has no ${{\mathrm{u}_\omega}}$-compact property but some “local ${{\mathrm{u}_\omega}}$-compact property” which gives result of Theorem A for a certain algebra $\Ce$ (not for all $\Ce$). This idea is developed in Subsection \[subsec:local\].
In Section \[sec:weak\] we discuss weak properties: weak equationally Noetherian property (Subsection \[subsec:weakEN\]) and weak ${{\mathrm{u}_\omega}}$-compactness (Subsection \[subsec:weakcompact\]). In Subsection \[subsec:logirr\] we introduce logically irreducible algebraic sets. Those sets naturally arise as generalization of irreducible ones. In particularly, we show that the notions of irreducible algebraic set and logically irreducible algebraic set over an algebra $\B$ coincide if and only if $\B$ is weakly ${{\mathrm{u}_\omega}}$-compact.
In the last Section \[sec:extension\] we continue discussion about connections between ${{\mathrm{u}_\omega}}$- and ${{\mathrm{q}_\omega}}$-compact algebras with the Compactness Theorem and corresponding technique from the model theory. By the way, we construct ${{\mathrm{u}_\omega}}$-compact elementary extension for an arbitrary algebra $\B$.
Preliminaries {#sec:preliminaries}
=============
In this section we remind basic notions and facts from universal algebraic geometry according to [@DMR1; @DMR2].
Let ${{\mathtt{L}}}$ be a first-order functional language, $X = \{x_1, x_2,
\ldots, x_n\}$ a finite set of variables, ${{\mathrm{T}}}_{{\mathtt{L}}}(X)$ the set of all terms of ${{\mathtt{L}}}$ with variables in $X$, $\T_{{\mathtt{L}}}(X)$ the absolutely free ${{\mathtt{L}}}$-algebra with basis $X$ and ${{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ the set of all atomic formulas of ${{\mathtt{L}}}$ with variables in $X$.
In universal algebraic geometry atomic formulas from ${{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ are named [*equations*]{} in ${{\mathtt{L}}}$ and subsets $S \subseteq
{{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ are named [*systems of equations*]{} in the language ${{\mathtt{L}}}$.
For a system of equations $S \subseteq {{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ and an algebra $\B$ in the language ${{\mathtt{L}}}$ we denote by ${{\mathrm{V}}}_\B(S)$ the set of all solutions of $S$ in $\B$: $${{\mathrm{V}}}_\B(S)=\{(b_1,\ldots,b_n)\in B^n \; \vert\quad \B\models
(t(b_1,\ldots,b_n)=s(b_1,\ldots,b_n))\quad \forall \;(t=s)\in S\}.$$ It is called the [*algebraic set*]{} over $\B$ defined by the system $S$. If $S$ contains of only one equation $(t=s)$ we write ${{\mathrm{V}}}_\B(t=s)$ instead of ${{\mathrm{V}}}_\B(\{(t=s)\})$.
Algebraic set is [*irreducible*]{} if it is not a finite union of proper algebraic subsets; otherwise it is [*reducible*]{}. The empty set is not considered to be irreducible. Hence, according to R.Hartshorne [@Hartshorne], all irreducible algebraic sets are non-empty in our paper.
Two systems $S_1, S_2 \subseteq {{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ are [*equivalent*]{} over $\B$ if ${{\mathrm{V}}}_\B(S_1)={{\mathrm{V}}}_\B(S_2)$. The [*radical*]{} ${{\mathrm{Rad}}}_\B(S)$ of a system of equations $S \subseteq {{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ is the maximal system which is equivalent to $S$ over $\B$. It is also called the radical of algebraic set $Y={{\mathrm{V}}}_\B(S)$ and denoted by ${{\mathrm{Rad}}}(Y)$. By $[S]$ we denote the congruent closure of $S$, i.e., the least congruent subset of ${{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ that contains $S$.
By $\Phi _{\mathrm{qf},{{\mathtt{L}}}} (X)$ we denote the set of all quantifier-free formulas in ${{\mathtt{L}}}$ with variables in $X$. We say that a formula $\phi \in \Phi _{\mathrm{qf},{{\mathtt{L}}}} (X)$ is a [*consequence*]{} of a system of equations $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ over an ${{\mathtt{L}}}$-algebra $\B$, if $\B\models \phi (b_1,\ldots,b_n)$ for all $(b_1,\ldots,b_n)\in {{\mathrm{V}}}_\B (S)$. For example, an atomic formula $(t=s)$, $t,s\in {{\mathrm{T}}}_{{\mathtt{L}}}(X)$, is a consequence of $S$ over $\B$ if and only if $(t=s)\in {{\mathrm{Rad}}}_\B(S)$.
For an arbitrary algebraic set $Y \subseteq B^n$ over $\B$ the radical ${{\mathrm{Rad}}}(Y)$ defines the congruence $\theta_{{{\mathrm{Rad}}}(Y)}$ on $\T_{{\mathtt{L}}}(X)$: $$t_1\sim_{\theta_{{{\mathrm{Rad}}}(Y)}}t_2 \quad \Longleftrightarrow \quad
(t_1=t_2)\in {{\mathrm{Rad}}}(Y), \quad t_1,t_2\in {{\mathrm{T}}}_{{\mathtt{L}}}(X).$$ The factor-algebra $\Gamma (Y)=\T_{{\mathtt{L}}}(X)/\theta_{{{\mathrm{Rad}}}(Y)}$ is called the [*coordinate algebra*]{} of the algebraic set $Y$.
Let $Y\subseteq B^n$ and $Z\subseteq B^m$ be algebraic sets over $\B$. One has $\Gamma (Y)\cong\Gamma (Z)$ if and only if algebraic sets $Y$ and $Z$ are isomorphic (we omit here the definition of isomorphism between algebraic sets). Isomorphic algebraic sets are irreducible and reducible simultaneously.
We say that an ${{\mathtt{L}}}$-algebra $\Ce$ is a [*coordinate algebra*]{} over $\B$ if $\Ce \cong \Gamma (Y)$ for some algebraic set $Y$ over $\B$, and $\Ce$ is an [*irreducible coordinate algebra*]{} over $\B$ if $\Ce \cong \Gamma (Y)$ for some irreducible algebraic set $Y$ over $\B$.
One of the principal goals of algebraic geometry over a given algebraic structure $\B$ is the problem of classification of algebraic sets over $\B$ up to isomorphism. This problem is equivalent to the problem of classification of coordinate algebras of algebraic sets over $\B$. Also it is important to classify coordinate algebras of irreducible algebraic sets over $\B$. Formulated in Introduction Unification Theorems A and C are very useful for solution of those problems.
In Theorems A and C we claim an algebra $\B$ is equationally Noetherian. Thus, let us remind that an ${{\mathtt{L}}}$-algebra $\B$ is called [*equationally Noetherian*]{}, if for every finite set $X$ and every system of equations $S \subseteq {{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$ there exists a finite subsystem $S_0\subseteq S$ such that ${{\mathrm{V}}}_\B(S_0)={{\mathrm{V}}}_\B(S)$. Properties of equationally Noetherian algebras are discussed in [@DMR1; @DMR2].
An ${{\mathtt{L}}}$-algebra $\C$ is [*separated*]{} by ${{\mathtt{L}}}$-algebra $\B$ if for any pair of non-equal elements $c_1, c_2 \in C$ there is a homomorphism $h \colon \C \to \B$ such that $h(c_1)\ne h(c_2)$. An algebra $\C$ is [*discriminated*]{} by $\B$ if for any finite set $W$ of elements from $C$ there is a homomorphism $h \colon\C \to
\B$ whose restriction onto $W$ is injective. We are interested in a familiar form of results, so it is useful to put by definition that the trivial algebra $\E$ is separated by an algebra $\B$ anyway, and $\E$ is discriminated by $\B$ if and only if $\B$ has a trivial subalgebra.
The definitions of limit algebras and algebras defined by complete atomic types need a large introduction, so we omit them (see [@DMR1]).
In this paper we use some operators which image a class ${{\mathbf{K}}}$ of ${{\mathtt{L}}}$-algebras into another one. For the sake on convenience we collect here the list of all these operators:\
${{\mathbf{S}}}({{\mathbf{K}}})$ — the class of subalgebras of algebras from ${{\mathbf{K}}}$;\
${{\mathbf{P}}}({{\mathbf{K}}})$ — the class of direct products of algebras from ${{\mathbf{K}}}$;\
${{\mathbf{P_{\!\! \omega}}}}({{\mathbf{K}}})$ — the class of finite direct products of algebras from ${{\mathbf{K}}}$;\
${{\mathbf{P_{\!\! s}}}}({{\mathbf{K}}})$ — the class of subdirect products of algebras from ${{\mathbf{K}}}$;\
${{\mathbf{P_{\! f}}}}({{\mathbf{K}}})$ — the class of filterproducts of algebras from ${{\mathbf{K}}}$;\
${{\mathbf{P_{\!\! u}}}}({{\mathbf{K}}})$ — the class of ultraproducts of algebras from ${{\mathbf{K}}}$;\
${{\underrightarrow{\mathbf{L}}}}({{\mathbf{K}}})$ — the class of direct limits of algebras from ${{\mathbf{K}}}$;\
${{\underrightarrow{\mathbf{L}}_{\! s}}}({{\mathbf{K}}})$ — the class of epimorphic direct limits of algebras from ${{\mathbf{K}}}$;\
${{\mathbf{L_{fg}}}}({{\mathbf{K}}})$ — the class of algebras in which all finitely generated subalgebras belong to ${{\mathbf{K}}}$;\
${{\mathbf{Pvar}}}({{\mathbf{K}}})$ — the least prevariety including ${{\mathbf{K}}}$;\
${{\mathbf{Qvar}}}({{\mathbf{K}}})$ — the least quasi-variety including ${{\mathbf{K}}}$, i.e., ${{\mathbf{Qvar}}}({{\mathbf{K}}})={{\mathrm{Mod}}}({{\mathrm{Th}}}_{\rm qi} ({{\mathbf{K}}}))$;\
${{\mathbf{Ucl}}}({{\mathbf{K}}})$ — the universal class of algebras generated by ${{\mathbf{K}}}$, i.e., ${{\mathbf{Ucl}}}({{\mathbf{K}}})={{\mathrm{Mod}}}({{\mathrm{Th}}}_{\forall} ({{\mathbf{K}}}))$;\
${{\mathbf{Res}}}({{\mathbf{K}}})$ — the class of algebras which are separated by ${{\mathbf{K}}}$;\
${{\mathbf{Dis}}}({{\mathbf{K}}})$ — the class of algebras which are discriminated by ${{\mathbf{K}}}$;\
${{\mathbf{K}}}{{\mathbf{\,_e}}}$ — the addition of the trivial algebra $\E$ to ${{\mathbf{K}}}$, i.e., ${{\mathbf{K}}}{{\mathbf{\,_e}}}={{\mathbf{K}}}\cup \{\E\}$;\
${{\mathbf{K}}}_\omega$ — the class of finitely generated algebras from ${{\mathbf{K}}}$.
Here we denote by ${{\mathrm{Th}}}_{\rm qi}({{\mathbf{K}}})$ (correspondingly, ${{\mathrm{Th}}}_{\forall}({{\mathbf{K}}})$, ${{\mathrm{Th}}}_{\exists}({{\mathbf{K}}})$) the set of all quasi-identities (correspondingly, universal sentences, existential sentences) which are true in all structures from ${{\mathbf{K}}}$.
For an arbitrary class ${{\mathbf{K}}}$ of ${{\mathtt{L}}}$-algebras one has: $$\begin{array}{rl}
{{\mathbf{Ucl}}}({{\mathbf{K}}})={{\mathbf{S}}}{{\mathbf{P_{\!\! u}}}}({{\mathbf{K}}}), &\; {{\mathbf{Dis}}}({{\mathbf{K}}}) \subseteq {{\mathbf{Ucl}}}({{\mathbf{K}}}),\\
{{\mathbf{Res}}}({{\mathbf{K}}})={{\mathbf{Pvar}}}({{\mathbf{K}}})={{\mathbf{S}}}{{\mathbf{P}}}({{\mathbf{K}}}), &\; {{\mathbf{Pvar}}}({{\mathbf{K}}})\subseteq{{\mathbf{Qvar}}}({{\mathbf{K}}}).
\end{array}$$
According to Gorbunov [@Gorbunov] and in contrast to [@DMR1], we assume that the direct product for the empty set of indexes coincides with the trivial ${{\mathtt{L}}}$-algebra $\E$. In particularly, when we say that an algebra $\Ce$ is a finite direct product of algebras from ${{\mathbf{K}}}$ (or a subdirect product of a finitely many algebras from ${{\mathbf{K}}}$) then $\Ce$ may be just the trivial algebra. However, while defining an filterproduct we assume that the set of indexes is non-empty.
Generalizations of the Unification Theorems {#sec:unification_theorems_compact}
===========================================
Unification Theorems A and C are formulated in Introduction above for an equationally Noetherian algebra $\B$. Those theorems have been proven in [@DMR1; @DMR2].
[**Question:**]{} [*Suppose that the algebra $\B$ is not equationally Noetherian. When Unification Theorems remain true for $\B$?*]{}
To answer this question we need to analyze the proofs of Theorems A and C. As it was mentioned in [@DMR2], for the reasoning of some implications in Theorems A and C the equationally Noetherian property is not required, namely, one has the following remark.
\[rem\] Let $\B$ be an algebra in a functional language ${{\mathtt{L}}}$ and $\Ce$ a finitely generated ${{\mathtt{L}}}$-algebra. Then
- $\Ce$ is the coordinate algebra of an irreducible algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$ IF AND ONLY IF $\Ce$ is discriminated by $\B$ (Theorem A: $7 \Longleftrightarrow 4$);
- IF $\Ce$ is discriminated by $\B$ THEN $\Ce$ is a limit algebra over $\B$ (Theorem A: $4 \Longrightarrow 5$);
- $\Ce$ is the coordinate algebra of an algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$ IF AND ONLY IF $\Ce \in {{\mathbf{Pvar}}}(\B)$ (Theorem C: $7 \Longleftrightarrow
2$);
- IF $\Ce$ is a subdirect product of a finitely many limit algebras over $\B$ THEN $\Ce \in {{\mathbf{Qvar}}}(\B)$ (Theorem C: $5 \Longrightarrow 1$); and so on.
The complete set of implications in Theorems A and C which always remain true is represented as follows: $$\mbox{Theorem~A:} \quad \{4 \Leftrightarrow 7\}\quad
\Longrightarrow \quad \{1 \Leftrightarrow 2\Leftrightarrow
3\Leftrightarrow 5\Leftrightarrow 6\};$$ $$\mbox{Theorem~C:} \quad \{5\} \quad \Longrightarrow \quad \{1
\Leftrightarrow 6\} \quad \Longleftarrow \quad \{2 \Leftrightarrow
3 \Leftrightarrow 4 \Leftrightarrow 7\}.$$
Further, when proving $1) \Longrightarrow 4)$ in both Theorems A and C, we use not equationally Noetherian property itself, but some weaker properties. What properties exactly? These are ${{\mathrm{u}_\omega}}$-compactness and ${{\mathrm{q}_\omega}}$-compactness.
We say ${{\mathtt{L}}}$-algebra $\B$ is [*${{\mathrm{q}_\omega}}$-compact*]{} if for any finite set $X$, any system of equations $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any equation $(t_0=s_0)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ such that $${{\mathrm{V}}}_\B(S) \; \subseteq \; {{\mathrm{V}}}_\B(t_0=s_0)$$ there exists a finite subsystem $S_0\subseteq S$ such that $${{\mathrm{V}}}_\B(S)\; \subseteq \; {{\mathrm{V}}}_\B(S_0) \; \subseteq \; {{\mathrm{V}}}_\B(t_0=s_0).$$ Here the finite subsystem $S_0$ may alter depending on equation $(t_0=s_0)$.
An ${{\mathtt{L}}}$-algebra $\B$ is termed [*${{\mathrm{u}_\omega}}$-compact*]{} if for any finite set $X$, any system of equations $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any equations $(t_1=s_1), \ldots, (t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ such that $${{\mathrm{V}}}_\B(S) \; \subseteq \; {{\mathrm{V}}}_\B(t_1=s_1)\:\cup \:\ldots\: \cup\:
{{\mathrm{V}}}_\B(t_m=s_m)$$ there exists a finite subsystem $S_0\subseteq S$ such that $${{\mathrm{V}}}_\B(S)\; \subseteq \; {{\mathrm{V}}}_\B(S_0) \; \subseteq \;
{{\mathrm{V}}}_\B(t_1=s_1)\:\cup \:\ldots\: \cup\: {{\mathrm{V}}}_\B(t_m=s_m).$$ Here the finite subsystem $S_0$ may alter depending on equations $(t_1=s_1),\ldots,(t_m=s_m)$.
It is clear that any equationally Noetherian algebra $\B$ is ${{\mathrm{u}_\omega}}$-compact, and any ${{\mathrm{u}_\omega}}$-compact algebra is ${{\mathrm{q}_\omega}}$-compact.
The definitions of ${{\mathrm{u}_\omega}}$-compactness and ${{\mathrm{q}_\omega}}$-compactness above are given in geometric form. We know some other approaches to these notions that will be discussed in Section \[sec:compact\]. In that section will be also represented the etymology of the notion of ${{\mathrm{u}_\omega}}$(${{\mathrm{q}_\omega}}$)-compactness.
The generalization of Unification Theorem A {#subsec:A}
-------------------------------------------
The significance of ${{\mathrm{u}_\omega}}$-compact algebras in universal algebraic geometry is shown in the following theorem.
\[uw\] Let $\B$ be ${{\mathrm{u}_\omega}}$-compact algebra in a functional language ${{\mathtt{L}}}$. Then for a finitely generated algebra $\Ce$ of ${{\mathtt{L}}}$ the following conditions are equivalent:
1. ${{\mathrm{Th}}}_{\forall} (\B) \subseteq {{\mathrm{Th}}}_{\forall} (\Ce)$, i.e., $\Ce \in {{\mathbf{Ucl}}}(\B)$;
2. ${{\mathrm{Th}}}_{\exists} (\B) \supseteq {{\mathrm{Th}}}_{\exists} (\Ce)$;
3. $\Ce$ embeds into an ultrapower of $\B$;
4. $\Ce$ is discriminated by $\B$;
5. $\Ce$ is a limit algebra over $\B$;
6. $\Ce$ is an algebra defined by a complete atomic type in the theory ${{\mathrm{Th}}}_{\forall} (\B)$ in ${{\mathtt{L}}}$;
7. $\Ce$ is the coordinate algebra of an irreducible algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$.
Moreover, if for an ${{\mathtt{L}}}$-algebra $\B$ and for all finitely generated ${{\mathtt{L}}}$-algebras $\Ce$ the conditions above are equivalent then $\B$ is ${{\mathrm{u}_\omega}}$-compact.
It follows from Remark \[rem\] that conditions 1)–7) are equivalent if and only if one has equivalence $1)\Longleftrightarrow 4)$. The latter means that a finitely generated algebra $\Ce$ is discriminated by $\B$ if and only if $\Ce \in {{\mathbf{Ucl}}}(\B)$, i.e., ${{\mathbf{Ucl}}}(\B)_\omega={{\mathbf{Dis}}}(\B)_\omega$. By Theorem \[theoremU\] below, one has the equality ${{\mathbf{Ucl}}}(\B)_\omega={{\mathbf{Dis}}}(\B)_\omega$ if and only if an algebra $\B$ is ${{\mathrm{u}_\omega}}$-compact.
The generalization of Unification Theorem C {#subsec:C}
-------------------------------------------
To prove an analog of Theorem C for ${{\mathrm{q}_\omega}}$-compact algebras we need the following results.
\[Corollary 5.7\] Let $\Ce$ be a limit algebra over an ${{\mathtt{L}}}$-algebra $\B$. Then there exists an ultrapower $\B^\ast$ of $\B$ such that $\Ce$ embeds into $\B^\ast$.
\[Lemma 3.9\] A finitely generated ${{\mathtt{L}}}$-algebra $\Ce$ is the coordinate algebra of an algebraic set over ${{\mathtt{L}}}$-algebra $\B$ if and only if $\Ce$ is a subdirect product of the coordinate algebras of irreducible algebraic sets over $\B$.
\[qw\] Let $\B$ be ${{\mathrm{q}_\omega}}$-compact algebra in a functional language ${{\mathtt{L}}}$. Then for a finitely generated algebra $\Ce$ of ${{\mathtt{L}}}$ the following conditions are equivalent:
1. $\Ce \in {{\mathbf{Qvar}}}(\B)$, i.e., ${{\mathrm{Th}}}_{\rm qi} (\B) \subseteq {{\mathrm{Th}}}_{\rm qi} (\Ce)$;
2. $\Ce \in {{\mathbf{Pvar}}}(\B)$;
3. $\Ce$ embeds into a direct power of $\B$;
4. $\Ce$ is separated by $\B$;
5. $\Ce$ is a subdirect product of limit algebras over $\B$;
6. $\Ce$ is an algebra defined by a complete atomic type in the theory ${{\mathrm{Th}}}_{\rm qi} (\B)$ in ${{\mathtt{L}}}$;
7. $\Ce$ is the coordinate algebra of an algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$.
Moreover, if for an ${{\mathtt{L}}}$-algebra $\B$ and for all finitely generated ${{\mathtt{L}}}$-algebras $\Ce$ the conditions above are equivalent then $\B$ is ${{\mathrm{q}_\omega}}$-compact.
By Remark \[rem\], it is sufficient to prove implications $1)
\Longrightarrow 2)$, $5') \Longrightarrow 1)$, and $7)
\Longrightarrow 5')$ for ${{\mathrm{q}_\omega}}$-compact algebra $\B$. By Theorem \[theoremQ\] below, we have the identity ${{\mathbf{Qvar}}}(\B)_\omega={{\mathbf{Pvar}}}(\B)_\omega$ that gives proof of $1)
\Longrightarrow 2)$. For implication $5') \Longrightarrow 1)$ we refer to Lemma \[Corollary 5.7\] and the fact that every quasi-variety is closed under ultraproducts, direct products and subalgebras.
For proving $7) \Longrightarrow 5')$ suppose that $\Ce$ is the coordinate algebra of an algebraic set over $\B$. By Lemma \[Lemma 3.9\], $\Ce$ is a subdirect product of coordinate algebras of irreducible algebraic sets over $\B$. By Remark \[rem\] (Theorem A: $7 \Longrightarrow 5$), coordinate algebras of irreducible algebraic sets over $\B$ are limit algebras over $\B$.
Suppose now that for some ${{\mathtt{L}}}$-algebra $\B$ we have equivalence $1) \Longleftrightarrow 2)$ for all finitely generated ${{\mathtt{L}}}$-algebras $\Ce$. It means that ${{\mathbf{Qvar}}}(\B)_\omega={{\mathbf{Pvar}}}(\B)_\omega$ and, by Theorem \[theoremQ\] below, the algebra $\B$ is ${{\mathrm{q}_\omega}}$-compact.
\[problem\] Unfortunately, we are not in a position to formulate Theorem C for ${{\mathrm{q}_\omega}}$-compact algebras in all its fullness, because item 5) essentially needs equationally Noetherian property. We have to weak 5), namely we should erase words “finitely many”.
To establish Remark \[problem\] we formulate the following problem.
[**Embedding Problem.**]{} [*Let $\B$ be ${{\mathrm{q}_\omega}}$-compact algebra in a functional language ${{\mathtt{L}}}$. The question: whether or not every coordinate algebra over $\B$ subdirectly embeds into a finite direct product of algebras from ${{\mathbf{Ucl}}}(\B)$? If the answer is “not”, then we ask whether or not the same holds for at least ${{\mathrm{u}_\omega}}$-compact algebras.*]{}
A.N.Shevlyakov in [@Shevl1] gives the negative answer to the Embedding Problem both for ${{\mathrm{q}_\omega}}$-compact and ${{\mathrm{u}_\omega}}$-compact algebras.
Let us put an addition to Remark \[rem\].
The following implications and equivalencies from Theorem \[qw\] hold for an arbitrary algebra $\B$: $$\xymatrix{ \{1 \Leftrightarrow 6\} && \{ 2 \Leftrightarrow 3 \Leftrightarrow 4 \Leftrightarrow 7\} \ar@{=>}[ll] \ar@/^/@{=>}[dl]\\
& \{5'\} \ar@/^/@{=>}[ul]}$$
Theorem \[qw\] gives a classification of coordinate algebras in terms of quasivarieties. Thereby, any characterizations of quasivariety ${{\mathbf{Qvar}}}({{\mathbf{K}}})$ of a class ${{\mathbf{K}}}$ of ${{\mathtt{L}}}$-algebras are helpful in universal algebraic geometry. In [@Gorbunov; @Malcev] one can find the identities: $$\begin{gathered}
{{\mathbf{Qvar}}}({{\mathbf{K}}})\:=\:{{\mathbf{S}}}{{\mathbf{P_{\! f}}}}({{\mathbf{K}}}){{\mathbf{\,_e}}}\:=\:{{\mathbf{S}}}{{\mathbf{P}}}{{\mathbf{P_{\!\! u}}}}({{\mathbf{K}}}) \:=\: {{\mathbf{S}}}{{\mathbf{P_{\!\! u}}}}{{\mathbf{P}}}({{\mathbf{K}}})\:=\:{{\mathbf{S}}}{{\mathbf{P_{\!\! u}}}}{{\mathbf{P_{\!\! \omega}}}}({{\mathbf{K}}})\:=\:\\\:=\:{{\mathbf{S}}}{{\underrightarrow{\mathbf{L}}_{\! s}}}{{\mathbf{P}}}({{\mathbf{K}}})\:=\:{{\underrightarrow{\mathbf{L}}_{\! s}}}{{\mathbf{S}}}{{\mathbf{P}}}({{\mathbf{K}}})\:=\:{{\underrightarrow{\mathbf{L}}_{\! s}}}{{\mathbf{P_{\!\! s}}}}({{\mathbf{K}}})\:=\:{{\underrightarrow{\mathbf{L}}}}{{\mathbf{S}}}{{\mathbf{P}}}({{\mathbf{K}}}).\end{gathered}$$
Weak generalization of Unification Theorem A {#subsec:weak}
--------------------------------------------
Let $\B$ be an algebra in a functional language ${{\mathtt{L}}}$. Let us consider the class ${{\mathbf{Ucl}}}(\B)_\omega$.
By Remark \[rem\], for any irreducible algebraic set $Y$ over $\B$ the coordinate algebra $\Gamma (Y)$ belongs to ${{\mathbf{Ucl}}}(\B)_\omega$. If $\B$ is ${{\mathrm{u}_\omega}}$-compact algebra then, by Theorem \[uw\], every algebra $\Ce$ from ${{\mathbf{Ucl}}}(\B)_\omega$ is the coordinate algebra of some irreducible algebraic set $Y$ over $\B$.
Let us apply a weak mode to ${{\mathrm{u}_\omega}}$-compactness and require that every coordinate algebra $\Ce$ from ${{\mathbf{Ucl}}}(\B)_\omega$ is irreducible. Suppose that some algebras from ${{\mathbf{Ucl}}}(\B)_\omega$ are not coordinate algebras for algebraic sets over $\B$ at all, however, if $\Gamma (Y) \in {{\mathbf{Ucl}}}(\B)$ then $Y$ is irreducible. Let us introduce a specific name for algebra $\B$ with this type of property.
We name an ${{\mathtt{L}}}$-algebra $\B$ [*weakly ${{\mathrm{u}_\omega}}$-compact*]{} if each non-empty algebraic set $Y$ over $\B$ which coordinate algebra $\Gamma (Y)$ belongs to ${{\mathbf{Ucl}}}(\B)$ is irreducible.
By Theorem \[uw\], every ${{\mathrm{u}_\omega}}$-compact algebra is weakly ${{\mathrm{u}_\omega}}$-compact. We will discuss weakly ${{\mathrm{u}_\omega}}$-compact algebras, their properties and equivalent approaches to them in Subsection \[subsec:weakcompact\].
For weakly ${{\mathrm{u}_\omega}}$-compact algebras we have just the following weak analog of Theorem A. It allows to describe irreducible coordinate algebras inside the class of all coordinate algebras.
\[wuw\] Let $\B$ be a weakly ${{\mathrm{u}_\omega}}$-compact algebra in a functional language ${{\mathtt{L}}}$ and $Y$ a non-empty algebraic set over $\B$. Then the following conditions are equivalent:
1. ${{\mathrm{Th}}}_{\forall} (\B) \subseteq {{\mathrm{Th}}}_{\forall} (\Gamma(Y))$, i.e., $\Gamma(Y) \in {{\mathbf{Ucl}}}(\B)$;
2. ${{\mathrm{Th}}}_{\exists} (\B) \supseteq {{\mathrm{Th}}}_{\exists} (\Gamma(Y))$;
3. $\Gamma(Y)$ embeds into an ultrapower of $\B$;
4. $\Gamma(Y)$ is discriminated by $\B$;
5. $\Gamma(Y)$ is a limit algebra over $\B$;
6. $\Gamma(Y)$ is an algebra defined by a complete atomic type in the theory ${{\mathrm{Th}}}_{\forall} (\B)$ in ${{\mathtt{L}}}$;
7. $Y$ is irreducible.
Moreover, if for an ${{\mathtt{L}}}$-algebra $\B$ and for every non-empty algebraic set $Y$ the conditions above are equivalent then $\B$ is weakly ${{\mathrm{u}_\omega}}$-compact.
It follows from Remark \[rem\] that conditions 1)–7) are equivalent if and only if one has implication $1) \Longrightarrow
7)$. By definition, implication $1) \Longrightarrow 7)$ take place if and only if $\B$ is weakly ${{\mathrm{u}_\omega}}$-compact.
${{\mathrm{q}_\omega}}$-compact and ${{\mathrm{u}_\omega}}$-compact algebras {#sec:compact}
============================================================================
In Section \[sec:unification\_theorems\_compact\] we gave the definitions of ${{\mathrm{q}_\omega}}$- and ${{\mathrm{u}_\omega}}$-compact algebras in geometric language. In Subsection \[subsec:compact\] we gather the numerous another approaches to these notions into two theorems. We will prove these theorems in Subsection \[subsec:proof\].
In Subsection \[subsec:local\] we introduce “local ${{\mathrm{q}_\omega}}$(${{\mathrm{u}_\omega}}$)-compact property” and show its use in universal algebraic geometry. Subsection \[subsec:Ecompact\] contains some accessory materials.
Criteria of ${{\mathrm{q}_\omega}}$- and ${{\mathrm{u}_\omega}}$-compactness {#subsec:compact}
----------------------------------------------------------------------------
At first we formulate the theorems and then give the necessary explanations.
\[theoremQ\] For an algebra $\B$ in a functional language ${{\mathtt{L}}}$ the following conditions are equivalent:
- $\B$ is ${{\mathrm{q}_\omega}}$-compact;
- for any finite set $X$, any system of equations $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any consequence $c=(t_0=s_0)\in {{\mathrm{Rad}}}_\B (S)$ there exists a finite subsystem $S_c\subseteq S$ such that $c\in {{\mathrm{Rad}}}_\B(S_c)$;
- for any finite set $X$, any subset $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any atomic formula $(t_0=s_0)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if an (infinite) formula $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; t_0 (\bar{x})=s_0(\bar{x}) \right)$$ holds in $\B$ then for some finite subsystem $S_c\subseteq S$ the quasi-identity $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S_c} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; t_0 (\bar{x})=s_0(\bar{x}) \right)$$ also holds in $\B$;
- for any finite set $X$, any subset $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any atomic formula $(t_0=s_0)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if the set of formulas $$T\; = \; S\cup \{\neg (t_0=s_0)\}$$ is finitely satisfiable in $\B$ then it is satisfiable in $\B$;
- every finitely generated algebra from ${{\mathbf{Qvar}}}(\B)$ is the coordinate algebra of an algebraic set over $\B$;
- ${{\mathbf{Qvar}}}(\B)_\omega={{\mathbf{Pvar}}}(\B)_\omega$;
- ${{\mathbf{Qvar}}}(\B)={{\mathbf{L_{fg}}}}{{\mathbf{Res}}}(\B)$;
- ${{\underrightarrow{\mathbf{L}}_{\! s}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)={{\mathbf{L_{fg}}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)$;
- ${{\underrightarrow{\mathbf{L}}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)={{\mathbf{L_{fg}}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)$;
- for any finite set $X$ and any system of equations $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ one has: $${{\mathrm{Rad}}}_\B(S)=\bigcup_{S_0 \subseteq S} \: {{\mathrm{Rad}}}_\B (S_0),$$ where $S_0$ runs all finite subsystems of $S$;
- for any finite set $X$ and any directed system $\{S_i, i\in I\}$ of radical ideals over $\B$ from ${{\mathrm{At}}}_{{\mathtt{L}}}(X)$ the union $S=\bigcup_{i\in I}{S_i}$ is a radical ideal over $\B$;
- for any finite set $X$ and any epimorphic direct system $\Lambda= (I,\Ce_i,h_{ij})$ of coordinate algebras over $\B$ with generating set $X$, and $h_{ij}(x)=x$, $x\in X$, the epimorphic direct limit $\underrightarrow{\lim} \: \Ce_i$ is a coordinate algebra over $\B$.
\[theoremU\] For an algebra $\B$ in a functional language ${{\mathtt{L}}}$ the following conditions are equivalent:
- $\B$ is ${{\mathrm{u}_\omega}}$-compact;
- for any finite set $X$, any system of equations $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any consequence $c$ of $S$ over $\B$ of the form $c=(t_1=s_1)\vee\ldots \vee (t_m=s_m)$, $t_i,s_i\in {{\mathrm{T}}}_{{\mathtt{L}}}(X)$, there exists a finite subsystem $S_c\subseteq S$ such that $c$ is a consequence of $S_c$ over $\B$;
- for any finite set $X$, any subset $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any atomic formulas $(t_1=s_1), \ldots, (t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if an (infinite) formula $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; \bigvee\limits_{i=1}^m t_i (\bar{x})=s_i(\bar{x}) \right)$$ holds in $\B$ then for some finite subsystem $S_c\subseteq S$ the universal sentence $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S_c} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; \bigvee\limits_{i=1}^m t_i (\bar{x})=s_i(\bar{x}) \right)$$ also holds in $\B$;
- for any finite set $X$, any subset $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any atomic formulas $(t_1=s_1), \ldots, (t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if the set of formulas $$T\; =\; S\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\}$$ is finitely satisfiable in $\B$ then it is satisfiable in $\B$;
- every finitely generated algebra from ${{\mathbf{Ucl}}}(\B)$ is the coordinate algebra of an irreducible algebraic set over $\B$;
- ${{\mathbf{Ucl}}}(\B)_\omega={{\mathbf{Dis}}}(\B)_\omega$;
- ${{\mathbf{Ucl}}}(\B)={{\mathbf{L_{fg}}}}{{\mathbf{Dis}}}(\B)$.
Item 2) in Theorem \[theoremQ\] (correspondingly, in Theorem \[theoremU\]) gives the definition of ${{\mathrm{q}_\omega}}$-compact (correspondingly, ${{\mathrm{u}_\omega}}$-compact) algebra in terms of radicals; item 3) — in terms of infinite formulas; item 5) — in terms of coordinate algebras.
Item 4) shows that the definition of ${{\mathrm{q}_\omega}}$(${{\mathrm{u}_\omega}}$)-compactness is a compact property relating to special types of sets of formulas $T$, as it is discussed in Introduction. The background of this notion is detailed in [@MR2] for groups. Here we will tell just a few words about it.
The answer for the following question has been attained by V.A.Gorbunov [@Gorbunov].
[**Malcev Problem.**]{} [*When the prevariety ${{\mathbf{Pvar}}}({{\mathbf{K}}})$ generated by class ${{\mathbf{K}}}$ is a quasivariety?*]{}
V.A.Gorbunov has introduced the notion of quasi-compact (${\rm
q}$-compact) class ${{\mathbf{K}}}$ and proved that ${{\mathbf{Pvar}}}({{\mathbf{K}}})={{\mathbf{Qvar}}}({{\mathbf{K}}})$ if and only if ${{\mathbf{K}}}$ is ${\rm q}$-compact. Let us compare that result with item 6) in Theorem \[theoremQ\].
The definition of ${\rm q}$-compact algebra $\B$ is much the same as the definition of ${{\mathrm{q}_\omega}}$-compact algebra given in item 4) of Theorem \[theoremQ\]. We just bound the set of variables $X$ for defining ${{\mathrm{q}_\omega}}$-compact algebras: $X$ must be finite. For ${\rm
q}$-compact algebras $X$ runs sets of all possible cardinalities.
While items 1)–7) in Theorems \[theoremQ\] and \[theoremU\] are symmetric, items 10)–12) in Theorem \[theoremQ\] are specific for ${{\mathrm{q}_\omega}}$-compact algebras; 8) and 9) in Theorem \[theoremQ\] are just corollaries of 7).
Items 10) and 11) in Theorems \[theoremQ\] are close. The family $\{{{\mathrm{Rad}}}(S_0)\}$, where $S_0$ runs all finite subsystems of a system $S$, gives an example of a directed system. Let us remind concerned definitions.
A partial ordering $(I, \leqslant)$ is [*directed*]{} if any two elements from $I$ have an upper bound. A family $\{\theta_i, i\in
I\}$ of congruencies on an ${{\mathtt{L}}}$-algebra $\M$ with $i\leqslant
j\Leftrightarrow\theta_i\subseteq\theta_j$ is called [*directed system of congruencies*]{}.
A system $S \subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ is [*radical ideal*]{} over $\B$ if $S={{\mathrm{Rad}}}_\B (S)$.
We say that a family $\{S_i, i\in I\}$ of radical ideals from ${{\mathrm{At}}}_{{\mathtt{L}}}(X)$ is a [*directed system*]{} if the family $\{\theta_{S_i}, i\in I\}$ is a directed system of congruencies on $\T_{{\mathtt{L}}}(X)$.
Let us prove just a little part of Theorems \[theoremQ\] and \[theoremU\].
\[lemmacompact\] Let $\B$ be an ${{\mathtt{L}}}$-algebra, $X$ a finite set, $|X|=n$, $S\subseteq{{\mathrm{At}}}_{{\mathtt{L}}}(X)$ a system of equations, and $(t_1=s_1),
\ldots, (t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ atomic formulas. Then the following conditions are equivalent:
- ${{\mathrm{V}}}_\B(S) \; \subseteq \; {{\mathrm{V}}}_\B(t_1=s_1)\:\cup \:\ldots\:
\cup\: {{\mathrm{V}}}_\B(t_m=s_m)$;
- $(t_1=s_1)\vee\ldots \vee
(t_m=s_m)$ is a consequence of $S$ over $\B$;
- the (infinite) formula $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; \bigvee\limits_{i=1}^m t_i (\bar{x})=s_i(\bar{x}) \right)$$ holds in $\B$;
- the set of formulas $$T\; =\; S\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\}$$ is not satisfiable in $\B$;
- there is no homomorphism $h\colon \langle \{c_1,\ldots,c_n \} \,|\, S\rangle \to
\B$ such that $$h(t_i(c_1,\ldots,c_n))\ne h(s_i(c_1,\ldots,c_n)) \quad \mbox{for
all} \quad i \in \{1,\ldots, m\}.$$
Straightforward.
One has equivalencies $1) \Longleftrightarrow 2)$, $1)
\Longleftrightarrow 3)$, $3) \Longleftrightarrow 4)$ in both Theorems \[theoremQ\] and \[theoremU\].
Equivalencies $1) \Longleftrightarrow 2)$, $1) \Longleftrightarrow
3)$ are easy. Note that the statement in item 3) has a form “$A$ implies $B$”. The equivalent statement is “$\neg B$ implies $\neg A$” which gives 4). So we have $3) \Longleftrightarrow 4)$.
From now on, we will use not only geometric definition of ${{\mathrm{q}_\omega}}$-compact (correspondingly, ${{\mathrm{u}_\omega}}$-compact) algebra, but also the definitions that items 2), 3), 4) in Theorem \[theoremQ\] (correspondingly, in Theorem \[theoremU\]) give us.
$\E$-compact algebras {#subsec:Ecompact}
---------------------
This subsection is a special excursus. We consider here the following problem.
[**Problem.**]{} [*When the conditions “$\E\in
{{\mathbf{Ucl}}}(\B)$” and “$\B$ has a trivial subalgebra” are equivalent?*]{}
It is important to note that for a large class of algebras the conditions “$\E\in {{\mathbf{Ucl}}}(\B)$” and “$\B$ has a trivial subalgebra” are equivalent, but not for all algebras.
We say an ${{\mathtt{L}}}$-algebra $\B$ is [*$\E$-compact*]{} if finite satisfiability in $\B$ of the set of all atomic formulas ${{\mathrm{At}}}_{{\mathtt{L}}}(\{x\})$ in one variable $x$ implies its satisfiability in $\B$.
\[lemma2\] An ${{\mathtt{L}}}$-algebra $\B$ is $\E$-compact if and only if the conditions “$\E\in {{\mathbf{Ucl}}}(\B)$” and “$\B$ has a trivial subalgebra” are equivalent.
It is sufficient to show that ${{\mathrm{At}}}_{{\mathtt{L}}}(\{x\})$ is satisfiable in $\B$ if and only if $\B$ has a trivial subalgebra, and ${{\mathrm{At}}}_{{\mathtt{L}}}(\{x\})$ is finitely satisfiable if and only if $\E \in
{{\mathbf{Ucl}}}(\B)$.
Suppose that ${{\mathrm{At}}}_{{\mathtt{L}}}(\{x\})$ is satisfiable in $\B$. Then there exists an element $b\in B$ with $\B \models (t(b)=s(b))$ for all $t,s\in{{\mathrm{T}}}_{{\mathtt{L}}}(\{x\})$. Therefore, subalgebra of $\B$ generated by the element $b$ is trivial. Conversely, if $\B$ has a trivial subalgebra $\E=\{e\}$ then the set of all atomic formulas ${{\mathrm{At}}}_{{\mathtt{L}}}(\{x\})$ is realized in $\B$ on the element $e$.
Assume now that ${{\mathrm{At}}}_{{\mathtt{L}}}(\{x\})$ is not finitely satisfiable in $\B$. Then there exists a finite set $S_0$ of atomic formulas such that the universal sentence $$\label{eq:e}
\forall \; x\quad \left(\bigvee_{(t=s)\in S_0} {\neg(\:t(x)=
s(x)\:)} \right)$$ holds in $\B$. However is false in $\E$, so $\E
\not\in {{\mathbf{Ucl}}}(\B)$. Conversely, if the set of all atomic formulas ${{\mathrm{At}}}_{{\mathtt{L}}}(\{x\})$ is finitely satisfiable in $\B$ then by Compactness Theorem it is realized in some ultrapower $\B^\ast$ of $\B$. Hence, $\E \in {{\mathbf{Ucl}}}(\B)$.
The condition “algebra $\B$ is $\E$-compact” means that $\B$ has a trivial subalgebra or $\E\not\in{{\mathbf{Ucl}}}(\B)$.
Let us note that in “good” signatures all algebras are $\E$-compact.
\[lemmaL\] Suppose a functional language ${{\mathtt{L}}}$ has at least one constant symbol. Then every algebra in ${{\mathtt{L}}}$ is $\E$-compact.
Let $\B$ be an ${{\mathtt{L}}}$-algebra. We need to show that condition $\E\in
{{\mathbf{Ucl}}}(\B)$ implies that $\B$ has a trivial subalgebra. Consider the set of formulas $$T=\{c = c'\} \cup \{F(c,\ldots,c)=c\},$$ where $c,c'$ run all constant symbols from ${{\mathtt{L}}}$ and $F$ runs all functional symbols from ${{\mathtt{L}}}$. If $\E\in {{\mathbf{Ucl}}}(\B)$, then $\B\models
T$. Therefore, there exists an element $b\in B$ such that $c^\B=b$ for all constant symbol $c$ from ${{\mathtt{L}}}$, and $F(b,\ldots,b)=b$ for all functional symbol $F$ from ${{\mathtt{L}}}$. Thereby, the element $b$ generates the trivial subalgebra in $\B$.
Suppose ${{\mathtt{L}}}$ is a finite functional language. Then every algebra in ${{\mathtt{L}}}$ is $\E$-compact.
After Lemma \[lemmaL\] we may assume that ${{\mathtt{L}}}$ has no constant symbols. Let $\B$ an ${{\mathtt{L}}}$-algebra. If $\E\in {{\mathbf{Ucl}}}(\B)$ then the existential sentence $$\exists \; x\quad \left(\bigwedge_{F\in \: {{\mathtt{L}}}} {F(x,\ldots,x)=x}
\right)$$ holds in $\B$. Thereby, $\B$ has a trivial subalgebra.
If ${{\mathtt{L}}}$ is an infinite functional language with no constant symbols, then it is easy to construct an ${{\mathtt{L}}}$-algebra $\B$ that is not $\E$-compact (see Example \[example\] below).
It follows from the definition that all equationally Noetherian algebras are $\E$-compact. Now we state that all ${{\mathrm{q}_\omega}}$- and ${{\mathrm{u}_\omega}}$-compact algebras are $\E$-compact. We need the following facts and definitions.
According to V.A.Gorbunov [@Gorbunov], an ${{\mathtt{L}}}$-algebra $\B$ is [*weakly atomic compact*]{}, if for any set $X$ and any subset $S \subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ finite satisfiability of $S$ in $\B$ implies realizability of $S$ in $\B$. We say that an ${{\mathtt{L}}}$-algebra $\B$ is [*weakly atomic $\omega$-compact*]{}, if for any finite set $X$ and any subset $S \subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ finite satisfiability of $S$ in $\B$ implies realizability of $S$ in $\B$. It is obvious that weak atomic $\omega$-compactness implies $\E$-compactness.
The following result has been proven by M.Kotov [@Kotov].
[**Lemma**]{} ([@Kotov])[**.**]{} [*Every ${{\mathrm{q}_\omega}}$-compact algebra in a functional language ${{\mathtt{L}}}$ is weakly atomic $\omega$-compact.*]{}
\[corkot\] Let $\B$ be ${{\mathrm{q}_\omega}}$-compact ${{\mathtt{L}}}$-algebra (in particularly, $\B$ may be ${{\mathrm{u}_\omega}}$-compact). Then the universal closure ${{\mathbf{Ucl}}}(\B)$ contains the trivial algebra $\E$ if and only if $\B$ has a trivial subalgebra.
Let us note that M.Kotov has proven more general result in his work. We formulate it on geometric language.
[**Lemma**]{} ([@Kotov])[**.**]{} [*Let $\B$ be an ${{\mathtt{L}}}$-algebra and $S$ a system of equations in ${{\mathtt{L}}}$. If $\B$ is ${{\mathrm{q}_\omega}}$-compact and ${{\mathrm{V}}}_\B(S)$ is a singleton set or the empty set, then there exists a finite subsystem $S_0\subseteq S$ which is equivalent to $S$ over $\B$. If $\B$ is ${{\mathrm{u}_\omega}}$-compact and ${{\mathrm{V}}}_\B(S)$ is a finite set or the empty set, then there exists a finite subsystem $S_0\subseteq S$ which is equivalent to $S$ over $\B$.*]{}
Local compact properties {#subsec:local}
------------------------
Let $X$ be a finite set. Fix a subset $S\subseteq{{\mathrm{At}}}_{{\mathtt{L}}}(X)$. We will give the definitions of local compact properties with respect to fixed $S$.
An ${{\mathtt{L}}}$-algebra $\B$ is called [*${{\mathrm{q}_S}}$-compact*]{} if for each atomic formula $(t_0=s_0)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if the set of formulas $$T\; = \; S\cup \{\neg (t_0=s_0)\}$$ is finitely satisfiable in $\B$ then it is satisfiable in $\B$.
An ${{\mathtt{L}}}$-algebra $\B$ is called [*${{\mathrm{u}_S}}$-compact*]{} if for any atomic formulas $(t_1=s_1), \ldots, (t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if the set of formulas $$\label{eq:T}
T\; =\; S\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\}$$ is finitely satisfiable in $\B$ then it is satisfiable in $\B$.
It is clear that algebra $\B$ is ${{\mathrm{q}_\omega}}$(${{\mathrm{u}_\omega}}$)-compact if and only if it is ${{\mathrm{q}_S}}$(${{\mathrm{u}_S}}$)-compact for every finite set $X$ and every $S\subseteq{{\mathrm{At}}}_{{\mathtt{L}}}(X)$.
The main results on local compact properties are the following.
\[localQ\] Let $\B$ be an algebra in a functional language ${{\mathtt{L}}}$, $X$ a finite set, $S\subseteq{{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and $\Ce=\langle X | \, S\rangle$. Then the following conditions are equivalent:
1. $\Ce$ is the coordinate algebra of an algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$;
2. $\Ce$ is separated by $\B$;
3. $\Ce\in {{\mathbf{Qvar}}}(\B)$ and $\B$ is ${{\mathrm{q}_S}}$-compact.
\[localU\] Let $\B$ be an algebra in a functional language ${{\mathtt{L}}}$, $X$ a finite set, $S\subseteq{{\mathrm{At}}}_{{\mathtt{L}}}(X)$, such that $[S]\ne {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and $\Ce=\langle X | \, S\rangle$. Then the following conditions are equivalent:
1. $\Ce$ is the coordinate algebra of an irreducible algebraic set over $\B$ defined by a system of equations in the language ${{\mathtt{L}}}$;
2. $\Ce$ is discriminated by $\B$;
3. $\Ce\in{{\mathbf{Ucl}}}(\B)$ and $\B$ is ${{\mathrm{u}_S}}$-compact.
Before giving a proof of these propositions we need some remarks. Firstly, equivalence $1) \Longleftrightarrow 2)$ in both Propositions \[localQ\] and \[localU\] have been proven in [@DMR2]. Secondly, let us answer the question: when the set of formulas is not finitely satisfiable in $\B$? It happens if and only if there exists a finite subset $S_0\subseteq
S$ such that the universal sentence $$\label{eq:unfor}
\forall \; y_1 \ldots \forall \; y_n \left( \bigwedge
\limits_{(t=s) \in S_0} t (\bar{y})=s (\bar{y}) \; \longrightarrow
\; \bigvee \limits_{i=1}^{m} t_i (\bar{x})=s_i
(\bar{y})\right),\quad\mbox{where} \; |X|=n,$$ holds in $\B$. For example, if $(t_i=s_i)\in [S]$ for some $i\in\{1,\ldots,m\}$, then there exists a finite subset $S_0\subseteq S$ such that $S_0\vdash (t_i=s_i)$, in particularly, universal formula holds in $\B$.
Thirdly, note that in Propositions \[localU\] we claim $[S]\ne
{{\mathrm{At}}}_{{\mathtt{L}}}(X)$, but in Propositions \[localQ\] such restriction is omitted. If $[S]={{\mathrm{At}}}_{{\mathtt{L}}}(X)$ then $\Ce=\langle X | \, S\rangle$ is the trivial algebra $\E$. Moreover, in this case every algebra $\B$ is ${{\mathrm{q}_S}}$- and ${{\mathrm{u}_S}}$-compact. Since the trivial algebra $\E$ is the coordinate algebra of an algebraic set over $\B$ anyway and $\E$ belongs to each quasi-variety [@DMR2], we have no difficulties with $\E$ in Propositions \[localQ\].
One can omit restriction $[S]\ne{{\mathrm{At}}}_{{\mathtt{L}}}(X)$ in Proposition \[localU\] if and only if $\B$ is $\E$-compact algebra. Indeed, the trivial algebra $\E$ is the coordinate algebra of an irreducible algebraic set over $\B$ if and only if $\B$ has a trivial subalgebra [@DMR2 Lemma 3.22]. By Lemma \[lemma2\], the conditions “$\E\in {{\mathbf{Ucl}}}(\B)$” and “$\B$ has a trivial subalgebra” are equivalent if and only if $\B$ is $\E$-compact.
Now we are going to prove Propositions \[localQ\] and \[localU\]. Arguments for them are the similar, so we will prove only Propositions \[localU\].
Let $\Ce\simeq \T_{{{\mathtt{L}}}}(X)/\theta_S$, $X=\{c_1,\ldots,c_n\}$, and $[S]\ne{{\mathrm{At}}}_{{\mathtt{L}}}(X)$. By definition $\Ce$ is discriminated by $\B$ if for any finite set of atomic formulas $(t_1=s_1), \ldots,
(t_m=s_m)\in{{\mathrm{At}}}_{{\mathtt{L}}}(X)\setminus [S]$ there exists a homomorphism $h\colon \Ce \to \B$, such that $h(t_i(c_1,\ldots,c_n))\ne
h(s_i(c_1,\ldots,c_n))$ for all $i\in\{1,\ldots,m\}$. The existence of such homomorphism $h\colon \Ce \to \B$ means that the set $T$ in is realized in $\B$. Note that if we take $(t_i=s_i)\in [S]$ for some $i\in\{1,\ldots,m\}$, then $T$ is not finitely satisfiable in $\B$. Anyway, we shown that if $\Ce$ is discriminated by $\B$ then $\B$ is ${{\mathrm{u}_S}}$-compact. The occurrence $\Ce\in{{\mathbf{Ucl}}}(\B)$ follows from the inclusion ${{\mathbf{Dis}}}(\B) \subseteq
{{\mathbf{Ucl}}}(\B)$.
Suppose now that $\Ce=\langle X | \, S\rangle$ is not discriminated by $\B$ and show that $\Ce\not\in{{\mathbf{Ucl}}}(\B)$ or $\B$ is not ${{\mathrm{u}_S}}$-compact. In this case for some atomic formulas $(t_1=s_1), \ldots, (t_m=s_m)\in{{\mathrm{At}}}_{{\mathtt{L}}}(X)\setminus [S]$ the set $T$ from is not realized in $\B$. If at the same time $T$ is finitely satisfiable in $\B$ then $\B$ is not ${{\mathrm{u}_S}}$-compact. Assume that $T$ is not finitely satisfiable in $\B$. Therefore, there exists a finite subset $S_0\subseteq S$ such that the universal formula holds in $\B$. On the other hand, the formula $$\bigwedge \limits_{(t=s) \in S_0} t (\bar{y})=s (\bar{y}) \;
\longrightarrow \; \bigvee \limits_{i=1}^{m} t_i (\bar{x})=s_i
(\bar{y})$$ is false in $\Ce$ under the interpretation $y_i\mapsto c_i$, $i=1,
\ldots,n$, hence $\Ce \not\in {{\mathbf{Ucl}}}(\B)$.
Proof of the criteria {#subsec:proof}
---------------------
In this subsection we prove Theorems \[theoremQ\] and \[theoremU\] that have been formulated in Subsection \[subsec:compact\]. Remain that equivalencies $1)
\Longleftrightarrow 2)$, $1) \Longleftrightarrow 3)$, $3)
\Longleftrightarrow 4)$ in both theorems have been proven in Subsection \[subsec:compact\].
At first we prove the following easy lemma that will be useful below.
\[lemmaT0\] Let $\B, \Ce$ be ${{\mathtt{L}}}$-algebras, $\Ce \in {{\mathbf{Ucl}}}(\B)$, and $T$ a set of quantifier-free formulas in ${{\mathtt{L}}}$. If $T$ is finitely satisfiable in $\Ce$ then it is finitely satisfiable in $\B$.
Suppose $T$ is finitely satisfiable in $\Ce$. Then for every finite subset $\{\phi_1,\ldots,\phi_m\}\subseteq T$ the existential sentence $$\label{eq:phi}
\exists \; x_1 \ldots \exists \; x_n \quad \left(
\phi_1(x_1,\ldots,x_n) \:\wedge\: \ldots \:\wedge\:
\phi_m(x_1,\ldots,x_n) \right)$$ holds in $\Ce$. Since $\Ce \in {{\mathbf{Ucl}}}(\B)$ then holds in $\B$ too. Thereby, $T$ is finitely satisfiable in $\B$.
We start with Theorem \[theoremU\]. Consider item 6). It states that ${{\mathbf{Ucl}}}(\B)_\omega={{\mathbf{Dis}}}(\B)_\omega$. As inclusion ${{\mathbf{Ucl}}}(\B)_\omega \supseteq {{\mathbf{Dis}}}(\B)_\omega$ holds for an arbitrary algebra $\B$, then item 6) is equivalent to inclusion ${{\mathbf{Ucl}}}(\B)_\omega \subseteq {{\mathbf{Dis}}}(\B)_\omega$. On the other hand, ${{\mathbf{Dis}}}(\B)_\omega$ is the class of all irreducible coordinate algebras over $\B$ [@DMR2 Corollary 3.39]. Hence, we have equivalence $5) \Longleftrightarrow 6)$.
Now let us show equivalence $4) \Longleftrightarrow 6)$. Suppose $\B$ is ${{\mathrm{u}_\omega}}$-compact and $\M$ is a finitely generated algebra from ${{\mathbf{Ucl}}}(\B)$. If $\M$ is a trivial algebra then, by Corollary \[corkot\], $\B$ has a trivial subalgebra, therefore, $\M$ is discriminated by $\B$.
For non-trivial algebra $\M$ let us find a presentation $\langle X
\mid S\rangle$, where $X$ is a finite set and $S \subseteq
{{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$, $[S] \ne {{\mathrm{At}}}_{{\mathtt{L}}}(X)$. As $\B$ is ${{\mathrm{u}_S}}$-compact we have $\M \in {{\mathbf{Dis}}}(\B)$, by Proposition \[localU\]. Thus we proved inclusion ${{\mathbf{Ucl}}}(\B)_\omega \subseteq {{\mathbf{Dis}}}(\B)_\omega$ and implication $4) \Longrightarrow 6)$.
We prove the converse implication $6) \Longrightarrow 4)$ by contradiction. Suppose that $\B$ is not ${{\mathrm{u}_\omega}}$-compact. Then there exists a finite set $X$, a subset $S \subseteq {{\mathrm{At}}}_{{{\mathtt{L}}}}(X)$, and atomic formulas $(t_1=s_1),\ldots,(t_m=s_m) \in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, such that the set of formulas $$T\; =\; S\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\},$$ is not realized in $\B$, but every its finite subset is realized in $\B$.
By Compactness Theorem $T$ is realized in some ultrapower $\B^I/D$ of $\B$. Let $c_1, \ldots, c_n$ be elements from $\B^I/D$, such that $\B^I/D \models T(c_1,\ldots,c_n)$, and $\Ce$ subalgebra of $\B^I/D$ generated by the set $\{c_1,\ldots,c_n\}$. Clearly, $\Ce$ is finitely generated algebra from ${{\mathbf{Ucl}}}(\B)$. Show that $\Ce$ is not discriminated by $\B$.
Let $\langle \{c_1,\ldots,c_n\} \mid R \rangle$ be a presentation of $\Ce$, i.e., $\Ce \simeq \T_{{\mathtt{L}}}(X)/\theta_R$, $R\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$. Since $\Ce \models T(c_1,\ldots,c_n)$, one has $S\subseteq
R$ and $(t_i=s_i)\not\in [R]$, $i=1,\ldots,m$. Put $$T'\; =\; R\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\}.$$ Since $T'$ is realized in $\Ce$ and $\Ce\in {{\mathbf{Ucl}}}(\B)$, then, by Lemma \[lemmaT0\], $T'$ is finitely satisfiable in $\B$. However, $T'$ is not satisfiable in $\B$. Thus $\B$ is not ${{\mathrm{u}_R}}$-compact. Hence, by Proposition \[localU\], $\Ce$ is not discriminated by $\B$. We proved $6) \Longrightarrow 4)$.
Equivalence $6) \Longleftrightarrow 7)$ is true in more general case. Let ${{\mathbf{K}}}$ and ${{\mathbf{K}}}'$ be two classes of ${{\mathtt{L}}}$-algebras (let us have in mind ${{\mathbf{K}}}={{\mathbf{Ucl}}}(\B)$ and ${{\mathbf{K}}}'={{\mathbf{Dis}}}(\B)$), ${{\mathbf{K}}}$ is universal axiomatizable and ${{\mathbf{K}}}'$ is closed under taking ${{\mathtt{L}}}$-subalgebras. Then ${{\mathbf{K}}}={{\mathbf{L_{fg}}}}{{\mathbf{K}}}'$ is equivalent to ${{\mathbf{K}}}_\omega={{\mathbf{K}}}'_\omega$. Indeed, ${{\mathbf{K}}}={{\mathbf{L_{fg}}}}{{\mathbf{K}}}'$ easy implies ${{\mathbf{K}}}_\omega={{\mathbf{K}}}'_\omega$. Inversely, if ${{\mathbf{K}}}_\omega={{\mathbf{K}}}'_\omega$ then ${{\mathbf{K}}}={{\mathbf{L_{fg}}}}{{\mathbf{K}}}_\omega={{\mathbf{L_{fg}}}}{{\mathbf{K}}}'_\omega={{\mathbf{L_{fg}}}}{{\mathbf{K}}}'$.
Now we begin to prove Theorem \[theoremQ\].
Equivalences $5) \Longleftrightarrow 6)$, $4) \Longleftrightarrow
6)$, $6) \Longleftrightarrow 7)$ may be proven by means of the similar reasoning as in Theorem \[theoremU\] (remind that ${{\mathbf{Pvar}}}(\B)={{\mathbf{Res}}}(\B)$).
Let us show equivalence $7) \Longleftrightarrow
8)\Longleftrightarrow 9)$. For an arbitrary algebra $\B$ we have ${{\mathbf{Qvar}}}(\B)={{\underrightarrow{\mathbf{L}}_{\! s}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)={{\underrightarrow{\mathbf{L}}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)$ [@Gorbunov Corollary 2.3.4] and ${{\mathbf{Res}}}(\B)={{\mathbf{S}}}{{\mathbf{P}}}(\B)$. So the identity ${{\mathbf{Qvar}}}(\B)={{\mathbf{L_{fg}}}}{{\mathbf{Res}}}(\B)$ is equivalent to ${{\underrightarrow{\mathbf{L}}_{\! s}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)={{\mathbf{L_{fg}}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)$ or ${{\underrightarrow{\mathbf{L}}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)={{\mathbf{L_{fg}}}}{{\mathbf{S}}}{{\mathbf{P}}}(\B)$.
Equivalence $2) \Longleftrightarrow 10)$ is easy. Equivalence $11)
\Longleftrightarrow 12)$ is due to V.A.Gorbunov [@Gorbunov Proposition 1.4.9]. So, it remains to prove implications $2) \Longrightarrow 11)$ and $11)
\Longrightarrow 10)$.
Let $\B$ be ${{\mathrm{q}_\omega}}$-compact algebra, $\{S_i, i\in I\}$ a directed system of radical ideals from ${{\mathrm{At}}}_{{\mathtt{L}}}(X)$ and $S=\bigcup_{i\in
I}{S_i}$. We show that $S={{\mathrm{Rad}}}(S)$, i.e., ${{\mathrm{Rad}}}(S) \subseteq
\bigcup_{i\in I}{S_i}$. Indeed, if $c$ is a consequence of $S$ then there exists a finite subsystem $S_0\subseteq S$ with $c\in
{{\mathrm{Rad}}}(S_0)$. Since $I$ is directed there exists an index $i\in I$ such that $S_0 \subseteq S_i$, therefore $c\in S_i$. Thus we have implication $2) \Longrightarrow 11)$.
To prove implication $11) \Longrightarrow 10)$ consider an arbitrary system $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$. The family $\{{{\mathrm{Rad}}}(S_0)\}$, where $S_0$ runs all finite subsystems of a system $S$, forms a directed system of radical ideals from ${{\mathrm{At}}}_{{\mathtt{L}}}(X)$. Hence $\bigcup_{S_0 \subseteq S}{{{\mathrm{Rad}}}(S_0)}$ is a radical ideal over $\B$. Also we have $$S \subseteq \bigcup_{S_0 \subseteq S} \: {{\mathrm{Rad}}}_\B (S_0) \subseteq
{{\mathrm{Rad}}}_\B(S),$$ therefore $\bigcup_{S_0 \subseteq S}{{{\mathrm{Rad}}}(S_0)} = {{\mathrm{Rad}}}_\B(S)$. So, implication $11) \Longrightarrow 10)$ has been proven.
Weakly equationally Noetherian and weakly ${{\mathrm{u}_\omega}}$-compact algebras {#sec:weak}
==================================================================================
A weak form of the equationally Noetherian property naturally arises in practice. We discuss algebras with this property in Subsection \[subsec:weakEN\].
In Subsection \[subsec:weak\] we have introduced weakly ${{\mathrm{u}_\omega}}$-compact algebras. Now in Subsection \[subsec:weakcompact\] we present some equivalent approaches to weakly ${{\mathrm{u}_\omega}}$-compact algebras.
In Subsection \[subsec:logirr\] we study logically irreducible algebraic sets. It is important ro note that logically irreducible algebraic sets inspired the notion of weakly ${{\mathrm{u}_\omega}}$-compact algebras.
Weak equationally Noetherian property {#subsec:weakEN}
-------------------------------------
An ${{\mathtt{L}}}$-algebra $\B$ is said to be [*weakly equationally Noetherian*]{}, if for any finite set $X$ every system $S\subseteq
{{\mathrm{At}}}_{{\mathtt{L}}}(X)$ is equivalent over $\B$ to some finite system $S_0
\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$. Here we do not assume that $S_0$ is a subsystem of $S$.
To make comparison equationally Noetherian and weakly equationally Noetherian properties it is required to reformulate corresponding definitions in the following form.
An ${{\mathtt{L}}}$-algebra $\B$ is termed [*weakly equationally Noetherian*]{}, if for any finite set $X$ and any system $S\subseteq
{{\mathrm{At}}}_{{\mathtt{L}}}(X)$ there exists finite system $S_0 \subseteq {{\mathrm{Rad}}}_\B(S)$ such that ${{\mathrm{V}}}_\B(S)={{\mathrm{V}}}_\B(S_0)$.
An ${{\mathtt{L}}}$-algebra $\B$ is termed [*equationally Noetherian*]{}, if for any finite set $X$ and any system $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ there exists finite system $S_0 \subseteq [S]$ such that ${{\mathrm{V}}}_\B(S)={{\mathrm{V}}}_\B(S_0)$.
Indeed, for every atomic formula $c=(t=s)\in [S]$ there exists a finite subsystem $S_c\subseteq S$ such that $S_c \vdash (t=s)$. Therefore, if ${{\mathrm{V}}}_\B(S)={{\mathrm{V}}}_\B(S_0)$ for a finite system $S_0
\subseteq [S]$ then one has $$\label{eq:V}
{{\mathrm{V}}}_\B(S)={{\mathrm{V}}}_\B(\bigcup_{c\in S_0}
{S_c}).$$
\[lemma4\] If an ${{\mathtt{L}}}$-algebra $\B$ is weakly equationally Noetherian and ${{\mathrm{q}_\omega}}$-compact then it is equationally Noetherian.
As $\B$ is weakly equationally Noetherian, for each system of equations $S$ there exists a finite system $S_0 \subseteq
{{\mathrm{Rad}}}_\B(S)$ with ${{\mathrm{V}}}_\B(S)={{\mathrm{V}}}_\B(S_0)$. As $\B$ is ${{\mathrm{q}_\omega}}$-compact, for each equation $c=(t_0=s_0)\in S_0$ there exists a finite subsystem $S_c\subseteq S$ with ${{\mathrm{V}}}_\B(S_c) \subseteq
{{\mathrm{V}}}_\B(t_0=s_0)$. Thereby, one has . It means that $\B$ is equationally Noetherian algebra.
If an ${{\mathtt{L}}}$-algebra $\B$ is weakly equationally Noetherian and $\Ce$ a subalgebra of some direct power of $\Ce$ then $\Ce$ is weakly equationally Noetherian too.
It follows from [@DMR2 Lemma 3.7].
It is clear that every weakly equationally Noetherian algebra is $\E$-compact.
\[lemma3\] If an ${{\mathtt{L}}}$-algebra $\B$ is weakly equationally Noetherian then $${{\mathbf{Ucl}}}(\B)\; \cap\; {{\mathbf{Res}}}(\B)_\omega = {{\mathbf{Dis}}}(\B)_\omega.$$
Since ${{\mathbf{Dis}}}(\B)\subseteq{{\mathbf{Res}}}(\B)$, ${{\mathbf{Dis}}}(\B)\subseteq{{\mathbf{Ucl}}}(\B)$, and ${{\mathbf{Res}}}(\B)={{\mathbf{Pvar}}}(\B)$ for any algebra $\B$ [@DMR1], we should check that ${{\mathbf{Ucl}}}(\B)\, \cap\, {{\mathbf{Pvar}}}(\B)_\omega \subseteq
{{\mathbf{Dis}}}(\B)_\omega$. Let us assume that $\Ce$ is a finitely generated algebra such that $\Ce\in{{\mathbf{Pvar}}}(\B)\setminus{{\mathbf{Dis}}}(\B)$ and prove $\Ce\not\in {{\mathbf{Ucl}}}(\B)$. If $\Ce$ is the trivial algebra $\E$ then, by definition, condition $\Ce\not\in{{\mathbf{Dis}}}(\B)$ implies that $\B$ has not a trivial subalgebra. Since $\B$ is weakly equationally Noetherian, then $\B$ is $\E$-compact, and, by Lemma \[lemma2\], $\E\not\in{{\mathbf{Ucl}}}(\B)$. Thereby, we may assume that $\Ce$ is non-trivial.
Let $\langle \{c_1,\ldots,c_n\} \mid S \rangle$ be a presentation of $\Ce$, i.e., $\Ce \simeq \T_{{\mathtt{L}}}(X)/\theta_S$, $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, $X=\{x_1,\ldots,x_n\}$. Since $\Ce\not\in{{\mathbf{Dis}}}(\B)$, there exits atomic formulas $(t_1=s_1),\ldots,(t_m=s_m)\in
{{\mathrm{At}}}_{{\mathtt{L}}}(X)\setminus [S]$ such that the (infinite) formula $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; \bigvee\limits_{i=1}^m t_i (\bar{x})=s_i(\bar{x}) \right)$$ holds in $\B$. As one can find a finite system $S_0\subseteq
{{\mathrm{At}}}_{{\mathtt{L}}}(X)$ with ${{\mathrm{V}}}_\B(S_0)={{\mathrm{V}}}_\B(S)$ then the universal sentence $$\label{eq:S0}
\forall \; x_1 \ldots \forall \; x_n \left(\bigwedge
\limits_{(t=s) \in S_0} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; \bigvee\limits_{i=1}^m t_i (\bar{x})=s_i(\bar{x}) \right)$$ holds in $\B$.
Since ${{\mathrm{V}}}_\B(S_0)={{\mathrm{V}}}_\B(S)$ we have ${{\mathrm{V}}}_\Ce(S_0)={{\mathrm{V}}}_\Ce(S)$ [@DMR2 Lemma 3.7]. Hence, $(c_1,\ldots,c_n)\in {{\mathrm{V}}}_\Ce(S_0)$ but $t_i(c_1,\ldots,c_n)\ne
s_i(c_1,\ldots,c_n)$ for all $i=1,\ldots,m$. Therefore, universal formula is not true in $\Ce$, and $\Ce\not\in{{\mathbf{Ucl}}}(\B)$.
Logically irreducible algebraic sets {#subsec:logirr}
------------------------------------
One of the approaches to ${{\mathrm{u}_\omega}}$-compact algebras deals with so-called logically irreducible algebraic sets.
We say that an algebraic set $Y$ over $\B$ is [*logically irreducible*]{} if its coordinate algebra $\Gamma (Y)$ belongs to ${{\mathbf{Ucl}}}(\B)$.
In Section \[sec:unification\_theorems\_compact\] we have discussed that every irreducible algebraic set over an arbitrary algebra $\B$ is logically irreducible. In Subsection \[subsec:weakcompact\] we will show that the notions of irreducible and logically irreducible algebraic sets coincide if and only if $\B$ is weakly ${{\mathrm{u}_\omega}}$-compact algebra.
Let $\B$ be an ${{\mathtt{L}}}$-algebra. For a finitely generated ${{\mathtt{L}}}$-algebra $\Ce$ the following conditions are equivalent:
- $\Ce$ is the coordinate algebra of a logically irreducible algebraic set over $\B$;
- $\Ce$ belongs to ${{\mathbf{Ucl}}}(\B)\, \cap \, {{\mathbf{Pvar}}}(\B)$.
Indeed, $\Ce$ is the coordinate algebra of an algebraic set over $\B$ if and only if $\Ce\in
{{\mathbf{Pvar}}}(\B)$ [@DMR2 Proposition 3.22].
\[lemma1\] The class of all coordinate algebras of logically irreducible algebraic sets over $\B$ coincides with ${{\mathbf{Ucl}}}(\B)\: \cap \:
{{\mathbf{Pvar}}}(\B)_\omega$.
For irreducible algebraic sets we have the following result.
\[lemma3.42\] Let $\B$ be an ${{\mathtt{L}}}$-algebra. Every non-empty algebraic set $Y$ over $\B$ is a union of maximal with respect to inclusion irreducible algebraic sets over $\B$.
Now we try to find a similar decomposition for algebraic sets into a union of maximal logically irreducible algebraic sets. It is clear that Lemma \[lemma3.42\] gives a decomposition. However, maximal with respect to inclusion irreducible algebraic set may be a proper subset of some logically irreducible algebraic set.
\[irr9\] Let $Y_1\subset Y_2 \subset \ldots$ be an ascending chain of logically irreducible algebraic sets in $B^n$ and $Y$ the least algebraic set containing all these sets. Then $Y$ is logically irreducible algebraic set.
Note that $Y={{\mathrm{V}}}_\B ({{\mathrm{Rad}}}(\bigcup\limits_{i}{Y_i}))$ and ${{\mathrm{Rad}}}(Y)=\bigcap\limits_{i}{{{\mathrm{Rad}}}(Y_i)}$. Hence, there exists embedding $h\colon \Gamma (Y) \to \prod\limits_{i}{\Gamma
(Y_i)}$ [@DMR1 Lemma 3.1]. Index $i$ runs the linearly ordered set $I$. For each $i\in I$ denote by $J_i$ the set $\{j\in
I, j\geqslant i\}$. The family of subsets $\{J_i, i\in I\}$ is centered, hence there exists an ultrafilter $D$ on $I$ containing $J_i$ for all $i\in I$. Let $f\colon \prod\limits_{i}{\Gamma
(Y_i)}\to \prod\limits_{i}{\Gamma (Y_i)}/D$ be a canonical homomorphism. Let us show that composition $f\circ h\colon \Gamma
(Y) \to \prod\limits_{i}{\Gamma (Y_i)}/D$ is embedding.
Indeed, we have $\Gamma (Y)=\T _{{\mathtt{L}}}(X)/\theta_{{{\mathrm{Rad}}}(Y)}$, where $X=\{x_1,\ldots,x_n\}$. If $t_1/\theta_{{{\mathrm{Rad}}}(Y)},
t_2/\theta_{{{\mathrm{Rad}}}(Y)}$ are distinct elements from $\Gamma (Y)$ then $(t_1=t_2)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)\setminus {{\mathrm{Rad}}}(Y)$. Since ${{\mathrm{Rad}}}(Y_1) \supset {{\mathrm{Rad}}}(Y_2) \supset \ldots$, then there exists an index $i_0\in I$ such that $(t_1=t_2)\not\in{{\mathrm{Rad}}}(Y_i)$ for all $i\in J_{i_0}$. It implies that $f\circ h (t_1/\theta_{{{\mathrm{Rad}}}(Y)})
\ne f\circ h (t_2/\theta_{{{\mathrm{Rad}}}(Y)})$. Thus $f\circ h$ is injective.
Since $\Gamma (Y_i) \in {{\mathbf{Ucl}}}(\B)$ for each $i\in I$ and $\Gamma
(Y) \in {{\mathbf{S}}}{{\mathbf{P_{\!\! u}}}}(\{\Gamma (Y_i), i\in I\})$, then $\Gamma (Y) \in
{{\mathbf{Ucl}}}(\B)$, i.e., $Y$ is logically irreducible algebraic set.
Let $\B$ be an ${{\mathtt{L}}}$-algebra. Every non-empty algebraic set $Y$ over $\B$ is a union of maximal with respect to inclusion logically irreducible algebraic sets over $\B$.
We will show that for each point $p\in Y$ there exists logically irreducible algebraic set $Z$ such that $p\in Z \subseteq Y$ and $Z$ is maximal with these properties. Denote by $\Omega$ the family of logically irreducible algebraic sets $Z$ with $p\in Z
\subseteq Y$ and show that $\Omega$ is not empty and has maximal elements.
Denote by $Z_p$ the closure in the Zariski topology of the set $\{p\}$. One has $p\in Z_p\subseteq Y$. Furthermore, $Z_p$ is irreducible algebraic set [@DMR2 Lemma 3.34]. Hence, $Z_p\in
\Omega$.
By Zorn Lemma it is sufficiently to show now that family $\Omega$ contains upper boundary for each ascending chain $Y_1\subset Y_2
\subset \ldots$ of element from $\Omega$. Let $Y_p$ be the least algebraic set that contains union $\bigcup\limits_{i}{Y_i}$. By Lemma \[irr9\], $Y_p$ is logically irreducible. As $Y_p\subseteq
Y$ one has $Y_p \in \Omega$.
Thereby, the union $\bigcup_{p\in Y} {Y_p}$ is desired.
Let us remind that for equationally Noetherian algebras we have the next result.
\[irr\] Let $\B$ be an equationally Noetherian algebra. Then any non-empty algebraic set $Y$ over $\B$ is a finite union of irreducible algebraic sets (irreducible components): $Y=Y_1 \cup \ldots \cup
Y_m$. Moreover, if $Y_i \not \subseteq Y_j$ for $i\ne j$ then this decomposition is unique up to a permutation of components.
It is natural to ask the following question.
[**Decomposition Problem.**]{} [*Let $\B$ be a “good” algebra (${{\mathrm{u}_\omega}}$-,${{\mathrm{q}_\omega}}$-compact, weakly equationally Noetherian, for instance). Is it true that every non-empty algebraic set over $\B$ is a finite union of logically irreducible algebraic sets?*]{}
In spite of the fact that ${{\mathrm{u}_\omega}}$-compact and weakly equationally Noetherian algebras are the closest algebras to equationally Noetherian ones we give for them the negative answer to the question above.
Indeed, a decomposition $Y=Y_1 \cup \ldots \cup Y_m$ of algebraic set $Y$ into a union of algebraic sets $Y_1,\ldots,Y_m$ implies the existence of a subdirect embedding $h\colon \Gamma(Y)\to
\Gamma(Y_1)\times\ldots\times\Gamma(Y_m)$ [@DMR2]. Suppose that the Decomposition Problem has the positive answer for ${{\mathrm{u}_\omega}}$-compact algebras. It involves that the Embedding Problem for ${{\mathrm{u}_\omega}}$-compact algebras has the positive answer too. However, A.N.Shevlyakov has proven the inverse result (see Subsection \[subsec:C\]). Moreover, he has proven also that the Decomposition Problem for weakly equationally Noetherian algebras has the negative answer [@Shevl1].
Weak ${{\mathrm{u}_\omega}}$-compactness {#subsec:weakcompact}
----------------------------------------
In the proposition below we gather the different approaches to weakly ${{\mathrm{u}_\omega}}$-compact algebras.
\[propwU\] For an algebra $\B$ in a functional language ${{\mathtt{L}}}$ the following conditions are equivalent:
- $\B$ is weakly ${{\mathrm{u}_\omega}}$-compact;
- every non-empty logically irreducible algebraic set over $\B$ is irreducible;
- every non-trivial coordinate algebra over $\B$ that belongs to ${{\mathbf{Ucl}}}(\B)$ is irreducible;
- ${{\mathbf{Ucl}}}(\B)\,\cap\,{{\mathbf{Res}}}(\B)_\omega=({{\mathbf{Dis}}}(\B){{\mathbf{\,_e}}})_\omega$.
Equivalence $1) \Longleftrightarrow 2)$ is evident by definition. Remind that the trivial algebra $\E$ is a coordinate algebra over $\B$ anyway, moreover, if $Y$ is an algebraic set over $\B$ such that $\E=\Gamma (Y)$ then $Y$ is irreducible or $Y=\emptyset$ [@DMR2 Lemma 3.22]. It implies that we have equivalence $2) \Longleftrightarrow 3)$.
Since ${{\mathbf{Dis}}}(\B)\subseteq{{\mathbf{Res}}}(\B)$, ${{\mathbf{Dis}}}(\B)\subseteq{{\mathbf{Ucl}}}(\B)$, ${{\mathbf{Res}}}(\B)={{\mathbf{Pvar}}}(\B)$, and $\E \in {{\mathbf{Res}}}(\B)$ for any algebra $\B$, then item 4) means that every non-trivial algebra $\Ce$ from ${{\mathbf{Ucl}}}(\B) \cap {{\mathbf{Pvar}}}(\B)_\omega$ belongs to ${{\mathbf{Dis}}}(\A)_\omega$.
As the class of all coordinate algebras of irreducible algebraic sets over $\B$ coincides with ${{\mathbf{Dis}}}(\B)_\omega$ [@DMR2 Corollary 3.37], and, by Corollary \[lemma1\], the class of all coordinate algebras of logically irreducible algebraic sets over $\B$ coincides with ${{\mathbf{Ucl}}}(\B)\: \cap \: {{\mathbf{Pvar}}}(\B)_\omega$, we have equivalence $3)
\Longleftrightarrow 4)$.
Every ${{\mathrm{u}_\omega}}$-compact (as well as ${{\mathrm{q}_\omega}}$-compact, weakly equationally Noetherian) algebra is $\E$-compact. However, there exist weakly ${{\mathrm{u}_\omega}}$-compact algebras that are not $\E$-compact (see Example \[example\] bellow). Suppose an algebra $\B$ is $\E$-compact. In this case one can omit “non-empty” in item 2), omit “non-trivial” in item 3), and write “${{\mathbf{Ucl}}}(\B)\,\cap\,{{\mathbf{Res}}}(\B)_\omega={{\mathbf{Dis}}}(\B)_\omega$” instead of “${{\mathbf{Ucl}}}(\B)\,\cap\,{{\mathbf{Res}}}(\B)_\omega=({{\mathbf{Dis}}}(\B){{\mathbf{\,_e}}})_\omega$” in item 4) in the formulation of Proposition \[propwU\]. In this case the empty set is not algebraic over $\B$, or if it is algebraic then its coordinate algebra $\E$ does not belong to ${{\mathbf{Ucl}}}(\B)$.
\[lemma5\] If an ${{\mathtt{L}}}$-algebra $\B$ is weakly ${{\mathrm{u}_\omega}}$-compact and ${{\mathrm{q}_\omega}}$-compact then it is ${{\mathrm{u}_\omega}}$-compact.
We need to show that ${{\mathbf{Ucl}}}(\B)_\omega \subseteq {{\mathbf{Dis}}}(\B)_\omega$. Assume that $\Ce$ is a finitely generated algebra and $\Ce\not\in{{\mathbf{Dis}}}(\B)$. Since ${{\mathbf{Dis}}}(\B)_\omega={{\mathbf{Ucl}}}(\B)\, \cap\,
{{\mathbf{Pvar}}}(\B)_\omega$, then $\Ce\not \in{{\mathbf{Ucl}}}(\B)$, and we have required, or $\Ce\not\in{{\mathbf{Pvar}}}(\B)$. By Theorem \[theoremQ\], $\Ce\not\in{{\mathbf{Pvar}}}(\B)$ implies that $\Ce\not\in{{\mathbf{Qvar}}}(\B)$, hence $\Ce\not \in{{\mathbf{Ucl}}}(\B)$.
The next question is naturally arises. Is there a geometric definition of weak ${{\mathrm{u}_\omega}}$-compactness?
We name an ${{\mathtt{L}}}$-algebra $\B$ [*geometrically weakly ${{\mathrm{u}_\omega}}$-compact*]{} if for any finite set $X$, any system of equations $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, and any equations $(t_1=s_1), \ldots,
(t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ such that $${{\mathrm{V}}}_\B(S) \; \subseteq \; {{\mathrm{V}}}_\B(t_1=s_1)\:\cup \:\ldots\: \cup\:
{{\mathrm{V}}}_\B(t_m=s_m)$$ and for each $i\in\{1,\ldots,m\}$ $${{\mathrm{V}}}_\B(S) \; \nsubseteq \; {{\mathrm{V}}}_\B(t_i=s_i)$$ there exists a finite subsystem $S_0\subseteq {{\mathrm{Rad}}}_\B(S)$ such that $${{\mathrm{V}}}_\B(S_0) \; \subseteq \; {{\mathrm{V}}}_\B(t_1=s_1)\:\cup \:\ldots\: \cup\:
{{\mathrm{V}}}_\B(t_m=s_m).$$
The definition above is evident generalization of both weak equationally Noetherian property and ${{\mathrm{u}_\omega}}$-compactness. It also has analogs in terms of radical, in terms of infinite formulas, and in terms of compactness.
\[theoremwU\] For an algebra $\B$ in a functional language ${{\mathtt{L}}}$ the following conditions are equivalent:
- $\B$ is geometrically weakly ${{\mathrm{u}_\omega}}$-compact;
- for any finite set $X$, any radical ideal $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ over $\B$, and any atomic formulas $(t_1=s_1), \ldots,
(t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)\setminus {{\mathrm{Rad}}}_\B (S)$ if $c=(t_1=s_1)\vee\ldots \vee (t_m=s_m)$ is a consequence of $S$ over $\B$ then there exists a finite subsystem $S_c\subseteq S$ such that $c$ is a consequence of $S_c$ over $\B$;
- for any finite set $X$, any radical ideal $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ over $\B$, and any atomic formulas $(t_1=s_1), \ldots,
(t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if an (infinite) formula $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; \bigvee\limits_{i=1}^m t_i (\bar{x})=s_i(\bar{x}) \right)$$ holds in $\B$, and for each $i \in\{1,\ldots,m\}$ an (infinite) formula $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; t_i(\bar{x})=s_i(\bar{x}) \right)$$ does not hold in $\B$, then for some finite subsystem $S_c\subseteq S$ the universal sentence $$\forall \; x_1 \ldots \forall \; x_n \left( \bigwedge
\limits_{(t=s) \in S_c} t (\bar{x})=s (\bar{x}) \; \longrightarrow
\; \bigvee\limits_{i=1}^m t_i (\bar{x})=s_i(\bar{x}) \right)$$ holds in $\B$;
- for any finite set $X$, any radical ideal $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ over $\B$, and any atomic formulas $(t_1=s_1), \ldots,
(t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)$ if the set of formulas $$T\; =\; S\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\}$$ is finitely satisfiable in $\B$ and for each $i \in\{1,\ldots,m\}$ the set of formulas $$T_i\; =\; S\cup \{\neg (t_i=s_i)\}$$ is realized in $\B$ then $T$ is satisfiable in $\B$.
Equivalences $1) \Longleftrightarrow 2)$, $1) \Longleftrightarrow
3)$, $3) \Longleftrightarrow 4)$ follows from Lemma \[lemmacompact\]. Note that the statement in item 3) has a form “$A \, \& \, \neg C$ implies $B$”. The equivalent statement is “$\neg B\, \& \, \neg C$ implies $\neg A$” which gives 4). So we have $3) \Longleftrightarrow 4)$.
Unfortunately, for weak ${{\mathrm{u}_\omega}}$-compactness we have no an analog of Theorem \[theoremU\] that holds for ${{\mathrm{u}_\omega}}$-compact algebras.
If an ${{\mathtt{L}}}$-algebra $\B$ is geometrically weakly ${{\mathrm{u}_\omega}}$-compact that it is weakly ${{\mathrm{u}_\omega}}$-compact. The converse statement does not hold.
Suppose that $\B$ is geometrically weakly ${{\mathrm{u}_\omega}}$-compact and $Y$ a non-empty algebraic set over $\B$ such that $\Gamma(Y)\in {{\mathbf{Ucl}}}(\B)$. We need to show that $\Gamma (Y) \in {{\mathbf{Dis}}}(\B)$. Let $S={{\mathrm{Rad}}}(Y)$, then $\Gamma (Y)$ has the presentation $\langle X
\mid S \rangle$. If $\Gamma (Y)$ is the trivial algebra, i.e., $S={{\mathrm{At}}}_{{\mathtt{L}}}(X)$, then $Y$ is irreducible [@DMR2 Lemma 3.22] and $\Gamma (Y) \in {{\mathbf{Dis}}}(\B)$.
Assume now that $\Gamma (Y)$ is non-trivial, i.e., $S\ne{{\mathrm{At}}}_{{\mathtt{L}}}(X)$. As the coordinate algebra $\Gamma (Y)$ is separated by $\B$, hence for each atomic formula $(t=s)\in
{{\mathrm{At}}}_{{\mathtt{L}}}(X)\setminus S$ the set of formulas $S\cup \{\neg (t=s)\}$ is realized in $\B$. Take atomic formulas $(t_1=s_1), \ldots,
(t_m=s_m)\in {{\mathrm{At}}}_{{\mathtt{L}}}(X)\setminus S$. As the set of formulas $T
=S\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\}$ is satisfiable in $\langle X \mid S \rangle$, and $\langle X \mid S
\rangle\in{{\mathbf{Ucl}}}(\B)$, then, by Lemma \[lemmaT0\], $T$ is finitely satisfiable in $\B$. It follows from item 4) of Lemma \[theoremwU\] that $T$ is satisfiable in $\B$. Thereby, algebra $\langle X \mid S \rangle$ is discriminated by $\B$.
Example \[example\] below shows that the converse statement does not hold.
The following example is similar to the example by M.V.Kotov [@Kotov].
\[example\] Let ${{\mathtt{L}}}=\{g_n, n\in \mathbb{N}\}$ be the infinite signature with unary functional symbols and $\A$ the ${{\mathtt{L}}}$-algebra with the universe $\mathbb{N}$ and $$g_n(x) = \left\{
\begin{array}{ll}
2n, & x = 2n + 1, \\
2n + 1, & x = 2n, \\
x, & \mbox{otherwise}.
\end{array}
\right.$$ It is clear that $\A$ has no trivial subalgebra. At the same time, the set of formulas $\{g_n(x)=x, n \in \mathbb{N}\}$ is finitely satisfiable in $\A$, therefore, by Compactness Theorem, it is satisfiable in some ultrapower $\A^\ast$ of $\A$. As $\A^\ast \in
{{\mathbf{Ucl}}}(\A)$, then $\E \in {{\mathbf{Ucl}}}(\A)$. Thereby, $\A$ is not $\E$-compact.
We state that $\A$ is weakly ${{\mathrm{u}_\omega}}$-compact. Indeed, take a non-trivial algebra $\Ce$ from ${{\mathbf{Ucl}}}(\A) \cap {{\mathbf{Pvar}}}(\A)$. Since ${{\mathbf{Pvar}}}(\A)={{\mathbf{S}}}{{\mathbf{P}}}(\A)$ then $\Ce$ is a subalgebra of a direct power of $\A$. For any $n,m\in \mathbb{N}$, $n\ne m$, the universal formula $$\forall \; x \quad (\, g_n(x)=x \; \vee \; g_m(x)=x \, )$$ holds in $\A$. Therefore, $\Ce$ has a finite universe $\{c_1,c'_1,
\ldots, c_d, c'_d\}$ with $c_i=g_{n_i}(c'_i)$ for all $i=\overline{1,d}$. The map $h\colon \Ce \to \A$, $h(c_i)=2n_i$, $h(c'_i)=2n_i+1$, $i=\overline{1,d}$, is a monomorphism. Thus, $\Ce \in {{\mathbf{Dis}}}(\A)$, and $\A$ is weakly ${{\mathrm{u}_\omega}}$-compact.
Let us check that $\A$ is not geometrically weakly ${{\mathrm{u}_\omega}}$-compact. Consider the systems of equations $S'(x)=\{g_n(x)=x,\: n\in
\mathbb{N}\setminus\{0\}\}$ and $S(x,y)=S'(x)\cup S'(y)$. We have ${{\mathrm{V}}}_\A(S)=\{(0,0), (0,1), (1,0), (1,1)\}$. Therefore, $$\begin{gathered}
{{\mathrm{V}}}_\A(S) \; \subseteq \; {{\mathrm{V}}}_\A(x=y) \: \cup \: {{\mathrm{V}}}_\A(x=g_0(y)),\\
{{\mathrm{V}}}_\A(S)\; \nsubseteq \; {{\mathrm{V}}}_\A(x=y), \quad {{\mathrm{V}}}_\A(S) \; \nsubseteq
\; {{\mathrm{V}}}_\A(x=g_0(y)).\end{gathered}$$ Furthermore, it is not hard to see that $${{\mathrm{Rad}}}_\A(S)= \left\{
\begin{array}{l}
x=g_{n_1}(g_{n_2}(\ldots g_{n_m}(x)\ldots)), \\
y=g_{n_1}(g_{n_2}(\ldots g_{n_m}(y)\ldots)), \; n_i\ne 0\\
x=g_{n_1}(g_{n_2}(\ldots g_{n_m}(y)\ldots)).
\end{array}
\right\}.$$ It is obvious that for any finite subsystem $S_0 \subseteq
{{\mathrm{Rad}}}_\A(S)$ we have $${{\mathrm{V}}}_\A(S_0) \; \nsubseteq \; {{\mathrm{V}}}_\A(x=y) \: \cup \: {{\mathrm{V}}}_\A(x=g_0(y)).$$
Connections between the classes of algebras ${{\mathbf{Q}}}$, ${{\mathbf{U}}}$, ${{\mathbf{U}}}'$, ${{\mathbf{N}}}'$, and ${{\mathbf{N}}}$ {#sec:picture}
==========================================================================================================================================
Let ${{\mathtt{L}}}$ be a functional language. We use the following denotations:\
${{\mathbf{N}}}\,\:$ — the class of all equationally Noetherian ${{\mathtt{L}}}$-algebras;\
${{\mathbf{N}}}\,'$ — the class of all weakly equationally Noetherian ${{\mathtt{L}}}$-algebras;\
${{\mathbf{Q}}}\,\:$ — the class of all ${{\mathrm{q}_\omega}}$-compact ${{\mathtt{L}}}$-algebras;\
${{\mathbf{U}}}\,\:$ — the class of all ${{\mathrm{u}_\omega}}$-compact ${{\mathtt{L}}}$-algebras;\
${{\mathbf{U}}}'$ — the class of all weakly ${{\mathrm{u}_\omega}}$-compact ${{\mathtt{L}}}$-algebras.
It is clear that $${{\mathbf{Q}}}\; \supseteq \; {{\mathbf{U}}}\; \supseteq \; {{\mathbf{N}}}\; \subseteq \; {{\mathbf{N}}}'.$$ Moreover, by Lemma \[lemma4\], $${{\mathbf{N}}}={{\mathbf{N}}}'\cap {{\mathbf{Q}}}= {{\mathbf{N}}}'\cap {{\mathbf{U}}}.$$
So, we have exactly the following picture for co-location of classes ${{\mathbf{N}}}$, ${{\mathbf{N}}}'$, ${{\mathbf{Q}}}$, ${{\mathbf{U}}}$:
(160,100) (50,0)(69,0)(85,15) (85,15)(100,31)(100,50) (100,50)(100,69)(85,85) (85,85)(69,100)(50,100) (50,100)(31,100)(15,85) (15,85)(0,69)(0,50) (0,50)(0,31)(15,15) (15,15)(31,0)(50,0)
(110,0)(129,0)(145,15) (145,15)(160,31)(160,50) (160,50)(160,69)(145,85) (145,85)(129,100)(110,100) (110,100)(91,100)(75,85) (75,85)(60,69)(60,50) (60,50)(60,31)(75,15) (75,15)(91,0)(110,0)
(40,50)(40,89)(80,89) (40,50)(40,11)(80,11)
(75,50)[${{\mathbf{N}}}$]{} (120,50)[${{\mathbf{N}}}'$]{} (20,50)[${{\mathbf{Q}}}$]{} (47,50)[${{\mathbf{U}}}$]{}
Let us find the place of the class ${{\mathbf{U}}}'$ in the picture above. By Theorem \[uw\], Lemma \[lemma3\] and Proposition \[propwU\], we have $${{\mathbf{U}}}\subseteq {{\mathbf{U}}}' \quad \mbox{and} \quad {{\mathbf{N}}}'\subseteq {{\mathbf{U}}}'.$$ It follows from Lemma \[lemma5\] that $${{\mathbf{Q}}}\, \cap\, {{\mathbf{U}}}'={{\mathbf{U}}}.$$
Hence, co-location of classes ${{\mathbf{N}}}$, ${{\mathbf{N}}}'$, ${{\mathbf{Q}}}$, ${{\mathbf{U}}}$, and ${{\mathbf{U}}}'$ are exactly the following:
(150,100) (50,0)(69,0)(85,15) (85,15)(100,31)(100,50) (100,50)(100,69)(85,85) (85,85)(69,100)(50,100) (50,100)(31,100)(15,85) (15,85)(0,69)(0,50) (0,50)(0,31)(15,15) (15,15)(31,0)(50,0)
(100,0)(119,0)(135,15) (135,15)(150,31)(150,50) (150,50)(150,69)(135,85) (135,85)(119,100)(100,100) (100,100)(81,100)(65,85) (65,85)(50,69)(50,50) (50,50)(50,31)(65,15) (65,15)(81,0)(100,0)
(100,5)(114,5)(125,15) (125,15)(135,26)(135,40) (135,40)(135,54)(125,65) (125,65)(114,75)(100,75) (100,75)(86,75)(75,65) (75,65)(65,54)(65,40) (65,40)(65,26)(75,15) (75,15)(86,5)(100,5)
(23,50)[${{\mathbf{Q}}}$]{} (53,50)[${{\mathbf{U}}}$]{}(83,50)[${{\mathbf{N}}}$]{} (113,50)[${{\mathbf{N}}}'$]{}(110,80)[${{\mathbf{U}}}'$]{}
In paper [@MR2] A.G.Myasnikov and V.N.Remeslennikov asked the questions for the class of groups:
[**Question 1:**]{}
[**Question 2:**]{}
Now we add new questions:
[**Question 3:**]{}
[**Question 4:**]{}
[**Question 5:**]{}
The answer to the first question has been given by B.I.Plotkin in [@Plot3]. He has constructed ${{\mathrm{q}_\omega}}$-compact group that is not equationally Noetherian. We will discuss that construction in this section below. Note that B.I.Plotkin uses notation [*logically Noetherian*]{} for ${{\mathrm{q}_\omega}}$-compact algebras and [*geometrically Noetherian*]{} for equationally Noetherian algebras.
The second and third questions have been solved by M.V.Kotov [@Kotov]. He has constructed examples that show ${{\mathbf{Q}}}\ne {{\mathbf{U}}}$ and ${{\mathbf{N}}}\ne{{\mathbf{N}}}'$. His examples are original algebraic structures in the language ${{\mathtt{L}}}=\{g_n, n \in
\mathbb{N}\}$ with countable set of unary functional symbols and with universe-sets $\mathbb{R}$ and $\mathbb{N}$.
At these results the fourth question remains open as well as the problem of differentiation of classes ${{\mathbf{Q}}}$, ${{\mathbf{U}}}$, ${{\mathbf{N}}}$, ${{\mathbf{N}}}'$ for classical varieties: groups, rings, monoids, semigroups. In [@Shevl1] A.N.Shevlyakov finds the neat examples in the variety of commutative idempotent semigroups in the language with countable set of constants. His examples distinguish classes ${{\mathbf{N}}}$, ${{\mathbf{N}}}'$, ${{\mathbf{Q}}}$, ${{\mathbf{U}}}$.
The algebra $\A$ from Example \[example\] gives an answer to the fifth question. It has been shown that $\A \in {{\mathbf{U}}}'$, but $\A$ is not $\E$-compact. Since all ${{\mathrm{u}_\omega}}$-compact and weakly equationally Noetherian algebras are $\E$-compact, then $\A \not \in
{{\mathbf{U}}}\,\cup\,{{\mathbf{N}}}'$. Another example for ${{\mathbf{U}}}'\ne{{\mathbf{U}}}\: \cup\: {{\mathbf{N}}}'$ has been constructed by A.N.Shevlyakov [@Shevl1] in the class of commutative idempotent semigroups in the language ${{\mathtt{L}}}$ with countable set of constants. It is important to note that all algebras in the language ${{\mathtt{L}}}$ are $\E$-compact, by Lemma \[lemmaL\].
Let us return to the construction given by B.I.Plotkin. He denotes by $H$ the discrete direct product of all finitely generated groups (in the language of groups ${{\mathtt{L}}}=\{\cdot, ^{-1},
e\}$). Since every finitely generated group $G$ imbeds into $H$, then $G$ is a coordinate group over $H$. By 10) in Theorem \[theoremQ\] below, $H$ is ${{\mathrm{q}_\omega}}$-compact. As there exists a finitely generated group $G$ that is not finitely presented, hence $H$ is not equationally Noetherian.
It is evident that this construction of $H$ may be repeated in other varieties of algebras, where exist finitely generated, not finitely presented algebras. Clearly, the algebraic geometry over objects like $H$ is quite elementary.
${{\mathrm{q}_\omega}}$-compact and ${{\mathrm{u}_\omega}}$-compact extensions {#sec:extension}
==============================================================================
In Introduction it is given the formulation of the Compactness Theorem and the notion of logical compactness. The Compactness Theorem has a great importance in model theory [@Hodges].
For an arbitrary algebra $\B$ it is possible with a use of the Compactness Theorem to construct an elementary extension $\B^\ast$ of $\B$ such that $\B^\ast$ is logically compact. This algorithm is close to the building of the algebraic closure to a given field $k$.
We use this idea to construct ${{\mathrm{u}_\omega}}$-compact elementary extension for an arbitrary algebra $\B$. At first, let us remind some more facts from model theory.
\[elem\_emb\] If $\B^I/D$ is an ultrapower of an algebra $\B$ then the diagonal map $d\colon \B
\to \B^I/D$, where $d(x)=\bar{x}/D$ and $\bar{x}(i)=x$ for all $i
\in I$, is an elementary embedding.
\[elem\_chain\] Suppose that $(I, <)$ is a linear order and $(\M_i, i\in I)$ is an elementary chain. Then $\M=\bigcup_{i\in I}{\M_i}$ is an elementary extension of each $\M_i$.
Denote by $\mathbb{T}$ the family of all sets of formulas $$T\; =\; S\cup \{\neg (t_1=s_1), \ldots , \neg (t_m=s_m)\},$$ where $S\subseteq {{\mathrm{At}}}_{{\mathtt{L}}}(X)$, $(t_1=s_1), \ldots, (t_m=s_m)\in
{{\mathrm{At}}}_{{\mathtt{L}}}(X)$, $\vert X \vert < \infty$. For a given ${{\mathtt{L}}}$-algebra $\B$ let $\mathbb{T}(\B)$ be such subfamily of $\mathbb{T}$ that $T\in \mathbb{T}(\B)$ if and only if $T$ is finitely satisfiable in $\B$ but not realized in $\B$. So, algebra $\B$ is ${{\mathrm{u}_\omega}}$-compact if and only if $\mathbb{T}(\B)=\emptyset$.
For ${{\mathtt{L}}}$-algebras $\B$ and $\Ce$ we write $\B \equiv_\forall \Ce$ if $\B$ and $\Ce$ are universally equivalent, i.e., ${{\mathbf{Ucl}}}(\B)={{\mathbf{Ucl}}}(\Ce)$.
\[lemmaT\] Let $\B$ and $\Ce$ be ${{\mathtt{L}}}$-algebras and $\B \leq \Ce$. If $\B
\equiv_\forall \Ce$ then $\mathbb{T}(\Ce)\subseteq
\mathbb{T}(\B)$.
Suppose $T\in \mathbb{T}$ and $T$ is finitely satisfiable in $\Ce$. Then, by Lemma \[lemmaT0\], $T$ is finitely satisfiable in $\B$. If $T \not\in \mathbb{T}(\B)$, then $T$ is realized in $\B$. As $\B \leq \Ce$, then $T$ is realized in $\Ce$ and $T
\not\in \mathbb{T}(\Ce)$.
\[el\_uw\] Let $\B$ be an ${{\mathtt{L}}}$-algebra. Then there exists an elementary extension $\B^\ast$ of $\B$, such that $\B^\ast$ is ${{\mathrm{u}_\omega}}$-compact (in particularly, $\B^\ast$ is ${{\mathrm{q}_\omega}}$-compact).
Consider a well-ordering $(I, <)$ on $\mathbb{T}(\B)$. Let us construct an elementary chain $(\B_i, i\in I)$. At first, take $\B_0=\B$. Then $\B_1$ is an ultrapower of $\B$ where $T_0$ is realized. By Compactness Theorem, such $\B_1$ exists and, by Theorem \[elem\_emb\], $\B_1$ is an elementary extension of $\B$. Further, $\B_2$ is an ultrapower of $\B_1$ where $T_1$ is realized, and so on. For an ordinal $\alpha=\beta+1$ we put $\B_\alpha$ as an ultrapower of $\B_\beta$ where $T_\beta$ is realized, and $\B_\alpha=\bigcup_{\beta < \alpha}{\B_\beta}$ for a limit ordinal $\alpha$. Desired algebra $\B^\ast$ is $\bigcup_{i\in I}{\B_i}$. Indeed, $\B^\ast$ is an elementary extension of $\B$, by Theorem \[elem\_emb\] and Proposition \[elem\_chain\].
Let us show that $\B^\ast$ is ${{\mathrm{u}_\omega}}$-compact. By Lemma \[lemmaT\], $\mathbb{T}(\B^\ast)\subseteq \mathbb{T}(\B)$. Every set of formulas $T$ from $\mathbb{T}(\B)$ is realized in $\B^\ast$. So $\mathbb{T}(\B^\ast)=\emptyset$.
For an arbitrary algebra $\B$ there exists ${{\mathrm{u}_\omega}}$-compact algebra $\B^\ast$ which is elementary equivalent to $\B$.
In Theorem \[el\_uw\] we constructed ${{\mathrm{u}_\omega}}$-compact extension $\B^\ast$ of $\B$ such that $\B^\ast$ is elementary equivalent to $\B$. One can modify the idea of Theorem \[el\_uw\] and find more constructive ${{\mathrm{u}_\omega}}$-compact extension $\B$ which is universally equivalent to $\B$.
Let $\B$ be an ${{\mathtt{L}}}$-algebra. Then there exists an extension $\Ce$ of $\B$ such that $\Ce$ is ${{\mathrm{u}_\omega}}$-compact and $\Ce\equiv_\forall
\B$. Moreover, one can get $\Ce$ by (transfinite) induction in series of extensions $$\B=\Ce_0 < \Ce_1 < \Ce_2 \ldots \: ,$$ where $\Ce_{\beta+1}$ is finitely generated extension of $\Ce_\beta$, and $\Ce_\alpha=\bigcup_{\beta < \alpha}{\Ce_\beta}$ is the union of the chain for a limit ordinal $\alpha$. Also $\Ce_\alpha\equiv_\forall \B$ for all $\alpha$.
Let us construct $\Ce$ by means of transfinite induction on $|\mathbb{T}(\B)|$. Take $\Ce_0=\B$. Consider an algebra $\B_1$ where $T_0$ is realized. Let $b_1,\ldots,b_n \in B_1$ be elements such that $\B_1 \models T(b_1,\ldots,b_n)$. Put $\Ce_1$ as the subalgebra of $\B_1$ generated by subalgebra $\B$ and elements $b_1,\ldots,b_n$. And so on.
If $\alpha=\beta+1$ then we take $\B_\alpha$ as an ultrapower of $\Ce_\beta$ where $T_\beta$ is realized, and $\Ce_\alpha$ is subalgebra of $\B_\alpha$ generated by $\Ce_\beta$ and finite set of element in $\B_\alpha$ which realize formulas from $T_\beta$. It is easy that $\Ce_\alpha \equiv_\forall \Ce_\beta$.
For a limit ordinal $\alpha$ we put $\Ce_\alpha=\bigcup_{\beta <
\alpha}{\Ce_\beta}$ as the union of the chain $(\Ce_\beta, \beta <
\alpha)$. In this case $\Ce_\alpha=\underrightarrow{\lim} \;
\Ce_\beta$ is also the direct limit of the direct system $(\Ce_\beta, \beta < \alpha)$, therefore $\Ce_\alpha\in
{{\mathbf{Ucl}}}(\{\C_\beta, \beta <
\alpha\})$ [@Gorbunov Theorem 1.2.9]. Since $\Ce_\beta <
\Ce_\alpha$ we have $\Ce_\beta \in {{\mathbf{Ucl}}}(\C_\alpha)$ for all $\beta < \alpha$. By induction, $\Ce_\beta \equiv_\forall
\Ce_\gamma$ for any $\beta,\gamma < \alpha$. Therefore, $\Ce_\alpha \equiv_\forall \Ce_\beta$ for every $\beta < \alpha$.
At the end of such process we get an extension $\Ce$ of $\B$ such that $\Ce\equiv_\forall \B$ and $\mathbb{T}(\Ce)=\emptyset$, i.e., $\Ce$ is ${{\mathrm{u}_\omega}}$-compact.
The following results are also useful in universal algebraic geometry.
\[lemmaQ\] Let $\B,\Ce$ be an ${{\mathtt{L}}}$-algebras. Suppose that $\B$ is ${{\mathrm{q}_\omega}}$-compact, $\Ce\in {{\mathbf{Qvar}}}(\B)$, and every finitely generated subalgebra $\B_0 < \B$ is separated by $\Ce$. Then $\Ce$ is ${{\mathrm{q}_\omega}}$-compact and and ${{\mathbf{Qvar}}}(\B)={{\mathbf{Qvar}}}(\Ce)$.
\[lemmaU\] Let $\B,\Ce$ be an ${{\mathtt{L}}}$-algebras. Suppose that $\B$ is ${{\mathrm{u}_\omega}}$-compact, $\Ce\in {{\mathbf{Ucl}}}(\B)$, and every finitely generated subalgebra $\B_0 < \B$ is discriminated by $\Ce$. Then $\Ce$ is ${{\mathrm{u}_\omega}}$-compact and ${{\mathbf{Ucl}}}(\B)={{\mathbf{Ucl}}}(\Ce)$.
We prove only statement about ${{\mathrm{u}_\omega}}$-compactness. Statement about ${{\mathrm{q}_\omega}}$-compactness may be proven in much the same way. By Theorem \[theoremU\], it is sufficient to show that ${{\mathbf{Ucl}}}(\Ce)_\omega\subseteq{{\mathbf{Dis}}}(\Ce)_\omega$ (inclusion ${{\mathbf{Ucl}}}(\Ce)_\omega\supseteq {{\mathbf{Dis}}}(\Ce)_\omega$ holds anyway). As $\Ce\in {{\mathbf{Ucl}}}(\B)$ then ${{\mathbf{Ucl}}}(\Ce) \subseteq {{\mathbf{Ucl}}}(\B)$ and ${{\mathbf{Ucl}}}(\Ce)_\omega \subseteq {{\mathbf{Ucl}}}(\B)_\omega$. Since $\B$ is ${{\mathrm{u}_\omega}}$-compact we have ${{\mathbf{Ucl}}}(\B)_\omega={{\mathbf{Dis}}}(\B)_\omega$. If every finitely generated subalgebra $\B_0 < \B$ is discriminated by $\Ce$ then ${{\mathbf{Dis}}}(\B)_\omega \subseteq {{\mathbf{Dis}}}(\Ce)_\omega$. Therefore, ${{\mathbf{Ucl}}}(\Ce)_\omega\subseteq{{\mathbf{Dis}}}(\Ce)_\omega$, as desired. Also we got ${{\mathbf{Ucl}}}(\B)_\omega={{\mathbf{Ucl}}}(\Ce)_\omega$ that implies ${{\mathbf{Ucl}}}(\B)={{\mathbf{Ucl}}}(\Ce)$.
For ${{\mathtt{L}}}$-algebra $\A$ we denote by ${{\mathtt{L}}}_\A = {{\mathtt{L}}}\cup\{c_a \mid a
\in A\}$ the language ${{\mathtt{L}}}$ extended by elements from $\A$ as new constant symbols [@DMR1 subsection 3.4]. An algebra $\M$ in ${{\mathtt{L}}}_\A$ is called [*$\A$-algebra*]{} if the map $h\colon \A \to
\M$, $h(a)=c^{\:\M}_a$, $a\in A$, is embedding.
Let $\A$ be an ${{\mathtt{L}}}$-algebra. Consider $\A$ as $\A$-algebra. If $\A$ is ${{\mathrm{q}_\omega}}$-compact (in the language ${{\mathtt{L}}}_\A$) then every $\A$-algebra $\Ce$ from ${{\mathbf{Qvar}}}_\A(\A)$ is ${{\mathrm{q}_\omega}}$-compact. If $\A$ is ${{\mathrm{u}_\omega}}$-compact (in the language ${{\mathtt{L}}}_\A$) then every $\A$-algebra $\Ce$ from ${{\mathbf{Ucl}}}_\A(\A)$ is ${{\mathrm{u}_\omega}}$-compact.
We use here Lemmas \[lemmaQ\] and \[lemmaU\]. Every finitely generated subalgebra $\A_0$ of $\A$ is an ${{\mathtt{L}}}_\A$-algebra, so $\A_0=\A$. Since $\Ce$ is $\A$-algebra then $\A$ is separated and discriminated by $\Ce$ in the obvious way. Thus, we have obtained the looked-for result.
[99]{}
G.Baumslag, A.Myasnikov, V.Remeslennikov, [*Algebraic geometry over groups I: Algebraic sets and ideal theory*]{}, J. Algebra, [**219**]{} (1999), pp.16–79,\
[http://ofim.okno.ru/$\sim$remesl/articles/algeom1.pdf](http://ofim.okno.ru/~remesl/articles/algeom1.pdf)
C.Champetier, V.Guirardel, [*Limit groups as limits of free groups: Compactifying the set of free groups*]{}, Israël J. Math., [**146**]{} (2005), pp.1–76, [arXiv:math/0401042](http://arxiv.org/abs/math/0401042)v2 \[math.GR\]
E.Daniyarova, A.Miasnikov, V.Remeslennikov, [*Unification theorems in algebraic geometry*]{}, Algebra and Discrete Mathematics, [**1**]{} (2008), pp.80–112, [arXiv:0808.2522](http://arxiv.org/abs/0808.2522)v1 \[math.AG\]
E.Daniyarova, A.Miasnikov, V.Remeslennikov, [*Algebraic geometry over algebraic structures II: Foundations*]{}, J. Algebra, submitted, [arXiv:1002.3562](http://arxiv.org/abs/1002.3562)v2 \[math.AG\]
A.Gaglione, D.Spellman, [*Some model theory of free groups and free algebras*]{}, Houston J. Math., [**19**]{} (1993), pp.327–356.
V.A.Gorbunov, [*Algebraic theory of quasivarieties*]{}, Nauchnaya Kniga, Novosibirsk, 1999; English transl., Plenum, 1998.
D.Groves, [*Limits of (certain) CAT(0) groups, I: Compactification*]{}, Algebraic and Geometric Topology, [**5**]{} (2005), pp.1325–1364, [arXiv:0404440](http://arxiv.org/abs/0404440)v3 \[math.GR\]
D. Groves, [*Limits of (certain) CAT(0) groups, II: The Hopf property and the shortening argument*]{}, Preprint, 2004, [arXiv:0408080](http://arxiv.org/abs/0408080)v1 \[math.GR\]
D.Groves, [*Limit groups for relatively hyperbolic groups, I: The basic tools*]{}, [arXiv:0412492](http://arxiv.org/abs/0412492)v2 \[math.GR\]
D.Groves, [*Limit groups for relatively hyperbolic groups, II: Makanin-Razborov diagrams*]{}, Geometry and Topology, [**9**]{} (2005), pp.2319–2358, [arXiv:0503045](http://arxiv.org/abs/0503045)v2 \[math.GR\]
V.Guirardel, [*Limit groups and group acting freely on $\mathbb{R}^n$-trees*]{}, Geometry and Topology, [**8**]{} (2004), pp.1427–1470, [arXiv:math/0306306](http://arxiv.org/abs/math/0306306)v5 \[math.GR\]
R.Hartshorne, [*Algebraic geometry*]{}, Graduate Texts in Math., [**52**]{}, Springer, 1977.
W.Hodges, [*Model theory*]{}, Cambridge University Press, 1993.
O.Kharlampovich, A.Myasnikov, [*Irreducible affine varieties over free group I: Irreducibility of quadratic equations and Nullstellensatz*]{}, J. Algebra, [**200 (2)**]{} (1998), pp.472–516.
O.Kharlampovich, A.Myasnikov, [*Irreducible affine varieties over free group II: Systems in trangular quasi-quadratic form and description of residually free groups*]{}, J. Algebra, [**200 (2)**]{} (1998), pp.517–570.
O.Kharlampovich, A.Myasnikov, [*Algebraic geometry over free groups: Lifting solutions into generic points*]{}, Contemp. Math., [**378**]{} (2005), pp.213–318, [arXiv:0407110](http://arxiv.org/abs/math/0407110)v4 \[math.GR\]
O.Kharlampovich, A.Myasnikov, [*Elementary theory of free nonabelian groups*]{}, J. Algebra, [**302 (2)**]{} (2006), pp.451–552.
M.V.Kotov, [*Equationally Noetherian property and close properties*]{}, to appear in Southeast Asian Bulletin of Mathematics.
A.I.Malcev, [*Algebraic structures*]{}, Nauka, Moscow, 1970.
D.Marker, [*Model theory: An introduction*]{}, Springer-Verlag New York, 2002.
A.Myasnikov, V.Remeslennikov, [*Algebraic geometry over groups II: Logical foundations*]{}, J. Algebra, [**234**]{} (2000), pp.225–276,\
[http://ofim.okno.ru/$\sim$remesl/articles/algeom2.pdf](http://ofim.okno.ru/~remesl/articles/algeom2.pdf)
B.Plotkin, [*Varieties of algebras and algebraic varieties. Categories of algebraic varieties*]{}, Siberian Advances in Math., [**7 (2)**]{} (1997), pp.64–97.
B.Plotkin, [*Varieties of algebras and algebraic varieties*]{}, Izrael J. Math., [**96 (2)**]{} (1996), pp.511–522.
B.Plotkin, [*Algebras with the same (algebraic) geometry*]{}, Proc. Steklov Inst. Math., [**242**]{} (2003), pp.165–196, [arXiv:math/0210194](http://arxiv.org/abs/math/0210194)v1 \[math.GM\]
Z.Sela, [*Diophantine geometry over groups I: Makanin-Razborov diagrams*]{}, Publications Mathematiques de l’IHES, [**93**]{} (2001), pp.31–105.
Z.Sela, [*Diophantine geometry over groups VI: The elementary theory of a free group*]{}, GAFA, [**16**]{} (2006), pp.707–730.
Z.Sela, [*Diophantine geometry over groups VII: The elementary theory of a hyperbolic group*]{}, Preprint.
A.N.Shevlyakov, [*Commutative idempotent semigroups at the service of the universal algebraic geometry*]{}, Southeast Asian Bulletin of Mathematics, submitted.
The information of the authors:
Evelina Yu. Daniyarova, Vladimir N. Remeslennikov
Omsk Department of Institute of Mathematics, Siberian Branch of the Russian Academy of Sciences
644099 Russia, Omsk, Pevtsova st. 13
Phone: +73812972251
e-mail: `[email protected], [email protected]`
Alexei G. Myasnikov
Department of Mathematics and Statistics, McGill University
Burnside Hall, Room 1005, 805 Shebrooke Street West, Montreal, Quebec, Canada, H3A 2K6
Phone: +15143985476
e-mail: `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop a theory of $L^p$ spaces based on outer measures generated through coverings by distinguished sets. The theory includes as special case the classical $L^p$ theory on Euclidean spaces as well as some previously considered generalizations. The theory is a framework to describe aspects of singular integral theory such as Carleson embedding theorems, paraproduct estimates and $T(1)$ theorems. It is particularly useful for generalizations of singular integral theory in time-frequency analysis, the latter originating in Carleson’s investigation of convergence of Fourier series. We formulate and prove a generalized Carleson embedding theorem and give a relatively short reduction of the most basic $L^p$ estimates for the bilinear Hilbert transform to this new Carleson embedding theorem.'
address:
- 'Yen Do, Department of Mathematics, Yale University, New Haven, CT 06511, USA'
- 'Christoph Thiele, Mathematisches Institut, Universität Bonn, Endenicher Alle 60, D-53115 Bonn, and Department of Mathematics, UCLA, Los Angeles, CA 90095, USA'
author:
- Yen Do
- Christoph Thiele
title: $L^p$ theory for outer measures and two themes of Lennart Carleson united
---
Introduction {#s.intro}
============
Two seminal papers of Lennart Carleson of the 1960’s each introduced a new tool into analysis that had profound influence. In his paper [@carleson1], [*Interpolation by bounded analytic functions and the corona problem*]{}, he introduced what later became known as Carleson measures. Carleson measures revolutionized singular integral theory, where they are for example related to the space $BMO$, and related areas in real and complex analysis. In his celebrated paper [@carleson], [*On convergence and growth of partial sums of Fourier series*]{}, Carleson introduced what we now call time-frequency analysis. Time-frequency analysis has remained until now an indispensable tool for its original application of controlling Fourier series pointwise as well as a number of other applications including $L^p$ estimates for the bilinear Hilbert transform. Our present paper shows that a natural $L^p$ theory for outer measures offers a unifying language for both Carleson measures and time-frequency analysis. The fundamental nature of our $L^p$ theory for outer measures might in hindsight be an explanation for the important role of Carleson measures.
This paper is divided into three parts. In the first part, Sections \[outermeasure\] and \[lptheory\], we carefully develop the basic $L^p$ theory for outer measure spaces. This part is in nature open ended and will hopefully lead to further investigations of outer measure spaces. We have focused only on those aspects of the theory that are directly relevant for the applications that we have in mind in the other parts of this paper.
Outer measures are subadditive set functions. In contrast to measures, outer measures do not necessarily satisfy additivity for disjoint finite or countable collections of sets. Some outer measures give rise to interesting measures by restriction to Caratheodory measurable sets, the most prominent example is classical Lebesgue theory. However, general outer measures need not give rise to interesting measures and one is led to studying outer measure spaces for their own sake. Lacking additivity for disjoint sets one can not expect a useful linear theory of integrals with respect to outer measure. A good replacement is a sub-linear or quasi sub-linear theory, which leads directly to norms or quasi norms rather than integrals. Naturally, $L^p$ norms are among the most basic norms to consider in the context of outer measures.
There is a rich literature on outer measures, for example on capacity theory. In contrast to previously developed theories based on the Choquet integral, we do not in general base our $L^p$ theory on the outer measure of super level sets $\{x:f(x)>\lambda\}$ for a function $f$. Instead, we use a more subtly defined quantity (Definition \[superleveldefinition\]) to replace the outer measure of a super level set. This new quantity, which we call [*super level measure*]{}, involves pre-defined averages over the generating sets of the outer measure. If the pre-defined averages are of $L^\infty$ type, the super level measure specializes to the outer measure of the super level set, but in general the two quantities are quite different. Once we have introduced the super level measure, the $L^p$ theory develops in standard fashion, and we develop it to the extend that we need for subsequent parts of the paper.
In the second part of this paper, Section \[t1section\], we describe how outer measures can be used in the context of Carleson measures. It is our first example of an outer measure space in which our refined definition of super level measure does not coincide with the classical case of the outer measure of super level set. The outer measure space in question is the upper half plane and the outer measure is generated by tents. The essentially bounded functions with respect to the outer measure in this upper half plane are Carleson measures. Moreover, the identification of a function on the boundary with the harmonic extension in the enterior of the upper half plane, that is the Carleson embedding map, turns out a basic example of a bounded map from a classical $L^p$ space to an outer $L^p$ space. We describe in Section \[t1section\] how classical estimates for paraproducts and $T(1)$ theorems can be proved by an outer Hölder inequality together with such embedding theorems. In this setting, the use of outer $L^p$ spaces is very much in the spirit of the use of tent spaces introduced in [@coifmanms]. It is an artifact in this particular situation that our notion of outer measure may be replaced with more classical concepts.
The full power of the new outer $L^p$ spaces becomes evident in its applications in time-frequency analysis, that we discuss in the third part of this paper. The underlying space for the outer measure becomes the Cartesian product of the upper half plane with a real line. In this setting there are no evident analogues of the tent spaces of [@coifmanms] that one could use in place of outer $L^p$ spaces. We formulate and prove a novel generalized Carleson embedding theorem, Theorem \[gen.carl.emb\], in Section \[gentents\]. It is a compressed and elegant way to state an essential part of time-frequency analysis. In Section \[bhtsection\] we then use the generalized Carleson embedding theorem to reprove bounds for the bilinear Hilbert transform.
The generalized Carleson embedding theorem can also be used as an ingredient to prove almost everywhere convergence of partial Fourier integrals of $L^p$ functions with $2<p<\infty$. One would need an additional Carleson embedding theorem, either analoguos to the interplay between energy and mass in [@lacey-thiele], or analoguos to some vector valued version of the Carleson embedding theorem as in [@demeter-thiele]. We also envision the generalized Carleson embedding theorem and variants thereof to be useful in further advances in time-frequency analysis. We were led to the theory of outer $L^p$ spaces while working on variation norm estimates as in [@oberlin-et-al] in the setting of biest type operators as in [@MTT-BiestFourier]. For brevity of the present paper, and because of the various possible routes towards Carleson’s theorem, we decided to restrict this exposition to a discussion of the bilinear Hilbert transform. This already captures many essential parts of Carleson’s time-frequency analysis.
Gaining a streamlined view on time-frequency analysis was the original motivation for the present paper, which is the outcome of a long evolution process. In traditional time-frequency analysis, one proves bounds of multilinear forms passing through model sums $$\Lambda=\sum_{P\in \P} c_P \prod_{j=1}^n a_j(P)\ \ ,$$ where the summation index runs through a discrete set, typically a collection of rectangles (tiles) in the phase plane. The coefficients $c_P$ are inherent to the multilinear form, while the sequences $a_j$ each depend on one of the input functions for the multilinear form in question. There is a multitude of examples in the literature for the tile sequences $a_j$, the most basic example being normalized wave packet coefficients $$\label{mostbasic}a_j(P)=\<f,\phi_P\>$$ for the $L^1$ normalized wave packets $$\phi_P(x) = 2^{-k}\phi(2^{-k}x-n)e^{2\pi i 2^{-k} x l}\ \ ,$$ where $k,n,l$ are integers and parameterize the space $\P$, and $\phi$ is a suitably chosen Schwartz class function. These coefficients are much in the spirit of the embedding maps considered in Sections \[gentents\] and \[bhtsection\] of the present paper. Use of such wavepackets in the study of the bilinear Hilbert transform appears in [@lacey-thiele1]. In the dyadic model as in [@thiele], one defines wave packets with respect to abstract Fourier analysis on the group $\Z/2\Z$. More generally one can have tile semi-norms $$\label{e.standard-sequence}
a_j(P)=\sup_{\phi\in \Phi}|\<f,\phi_P\>|$$ where one maximizes and possibly also averages over a suitably chosen set $\Phi$ of generating functions. This approach has been useful in [@thiele-uniform] and more explicitly in [@MTT-uniform]. To prove bounds on Carleson’s operator, [@lacey-thiele] uses modified wave packets $$\label{e.carleson-sequence}
a_j(P)=\<f,\phi_P1_{\{(x,N(x))\in P\}}\>$$ for the linearizing function $N$ of the linearized Carleson operator. In some instances such as in [@MTT-BiestFourier], the definition of $a_j(P)$ may involve itself a multi-linear operator whose analysis requires another level of time-frequency analysis. For variational estimates of the Carleson operator as in [@oberlin-et-al], one has variational wave packets $$\label{e.varcarleson-sequence}
a_j(P) = \<f,\phi_P \sum_{k} v_k1_{\{(x,N_k)\in P_2, (x,N_{k-1})\not\in P\}}\>$$ for a sequence of linearizing functions $N_0(x)<N_1(x)<\dots$ and a sequence of dualizing functions $v_1(x),v_2(x),\dots \ \ ,$ such that for some $r>2$ we have the uniform bound $$\sum_{k} |v_k(x)|^{r'} = O(1) \ \ .$$ A point of the present paper is that in many of these examples the bound on $\Lambda$ is a Hölder inequality with respect to an outer measure on the space $P$ : $$|\Lambda|\le C \sup_{P\in P}|c_P| \prod_{j=1}^n \|a_j\|_{L^p_j(\P,\dots)}$$ where the dots stand for specifications of the outer measure structures in each example. The rest of the proof of boundedness of $\Lambda$ then becomes modular in that one has to prove bounds for each $j$ separately on the outer $L^p$ norms of the sequences $a_j$, estimates which take for example the form $$\|a_j\|_{L^p_j(\P,\dots)}\le \|f_j\|_p\ ,$$ where $f_j$ may be the corresponding input function to the original multilinear form as for example in , and the $L^p$ norm is in the classical sense.
A novelty in the present paper is that we do not have to pass through a discrete model form, but rather work with an outer measure space on a continuum. This avoids both the cumbersome introduction of the discrete spaces as well as the usual technicalities in the discretization process.
The factorization of the multilinear form in time-frequency analysis into embedding theorems on the one hand and an outer Hölder’s inequality on the other hand is a clear modularization of the matter and promises to be useful in other applications of time-frequency analysis. Indeed, we were explicitly studying the modularization process because with Camil Muscalu we were considering a program outlined in [@do-muscalu-thiele] of estimating multilinear forms with nested levels of time-frequency analysis.
We are grateful to Mariusz Mirek for carefully reading an early version of this manuscript and pointing out many corrections. We are grateful to Pavel Zorin-Kranich for pointing out many corrections and an error on the last pages of a previous version posted on arxiv, that was overcome by proving a generalized Carleson embedding theorem with parameters $\alpha$ and $\beta$ and by removing any claim about explicit dependence on such parameters in the theorem on the bilinear Hilbert transform. Thanks to much stronger known uniform estimates for the bilinear Hilbert transform as in [@grafakos-li], tracking of the dependence on these parameters in our proof was not our key point. We also thank Yumeng Ou for pointing out an error in a previously posted version that was overcome by changing the exponent of $\beta$ in Corollary \[tentcorollary\], with some minor impact on the rest of the section. We thank an anonymous referee for a valuable list of suggestions to improve this exposition. We are grateful to Stefan Müller, Alexander Volberg, Igor Verbitzky and Nguyen Cong Phuc for discussions on capacity theory, which is a much studied example for outer measures. We finally are grateful for much feedback on outer measures during a season of conferences in which the ideas of this present paper were announced, and for the many suggestions on this exposition that we have received.
Outer measure spaces {#outermeasure}
====================
Outer measures
--------------
An outer measure or exterior measure on a set $X$ is a monotone and subadditive function on the collection of subsets of $X$ with values in the extended nonnegative real numbers, and with the value $0$ attained by the empty set.
Let $X$ be a set. An outer measure on $X$ is a function $\mu$ from the collection of all subsets of $X$ to $[0,\infty]$ that satisfies the following properties:
1. If $E\subset E'$ for two subsets of $X$, then $\mu(E)\le \mu(E')$.
2. $\mu(\emptyset)=0$.
3. If $E_1,E_2,\dots$ is a countable collection of sets in $X$, then $$\label{countsubadd}
\mu(\bigcup_{j=1}^\infty E_j)\le \sum_{j=1}^\infty \mu(E_j)\ .$$
In the examples we have in mind, the space $X$ is an infinite complete metric space and thus uncountable. The set of all subsets of $X$ has then even larger cardinality than the continuum, and can only be organized in abstract ways. The description of an outer measure then typically comes in two steps: First one specifies concretely a quantity that we may call pre-measure on a small collection of subsets, and then one passes abstractly from the pre-measure to the outer measure by means of covering an arbitrary subset by sets in the small collection. This covering process is the intuition behind the adjective [*outer*]{} in the term [*outer measure*]{}.
\[generate.outer\] Let $X$ be a set and $\E$ a collection of subsets of $X$. Let $\sigma$ be a function from $\E$ to $[0,\infty)$. Define for an arbitrary subset $E$ of $X$ $$\mu(E):=\inf_{\E' }\sum_{E'\in \E'} \sigma(E')\ ,$$ where the infimum is taken over all countable subcollections $\E'$ of $\E$ which cover the set $E$, that is whose union contains $E$. Here we understand that an empty sum is $0$. Then $\mu$ is an outer measure.
The concrete pre-measure requires the data $\E$, and $\sigma$. For simplicity we will often omit explicit mention of $\E$, since $\E$ is implicitly determined as the domain of $\sigma$. The proof of the proposition is basic and standard, we reproduce it here for emphasis.
We need to prove the three defining properties of outer measures.
The empty collection of subsets covers the empty set, which shows $\mu(\emptyset)= 0$ since the empty sum of nonnegative numbers is $0$.
If $F\subset F'$ for two subsets of $X$, then every cover of $F'$ is a cover of $F$ and hence $\mu(F)\le \mu(F')$. Let $F_1,F_2,\dots $ be a countable collection of subsets of $X$ and pick $\epsilon>0$. Find for each $i$ a countable subcollection $\E_i$ of $\E$ which covers $F_i$ and satisfies $$\sum_{E\in \E_i} \sigma(E)\le \mu(F_i)+ \epsilon 2^{-i}\ .$$ Then the union $\E'$ of the collections $\E_i$ covers the union of the sets $F_i$ and satisfies $$\sum_{E\in \E'} \sigma(E) \le (\sum_i \mu(F_i))+ \epsilon\ .$$
Since $\epsilon$ was arbitrary, we conclude that $\mu(\bigcup F_i)\le \sum_i \mu(F_i)$.
It is in general not true that for $E\in \E$ we have $\sigma(E)=\mu(E)$, however this identity can be established in many examples in practice. Clearly this identity holds precisely if for every set $E\in \E$ and every cover of $E$ by a countable subcollection $\E'$ of $\E$, we have $$\label{ecoverc}
\sigma(E)\le \sum_{E'\in \E'}\sigma(E')\ .$$ Then the most efficient cover of $E$ is by the trivial collection$\{E\}$, which establishes $\sigma=\left. \mu\right|_\E$.
We did not allow $\sigma$ to take value $\infty$. This is no restriction, since if we had $\sigma(E')=\infty$ for some $E'\in \E$, then using the set $E'$ in any cover of $E$ will make the sum $\sum_{E'\in \E'} \sigma(E')$ equal to $\infty$, a value that is as already the default even if no cover of $E$ exists at all.
If the collection $\E$ is countable, the contribution of sets $E\in \E$ with $\sigma(E)=0$ trivializes. Namely, we may consider the union $E_0$ of the countably many generating sets with pre-measure $0$. Then $E_0$ has outer measure zero, and we can construct an outer measure on $X\setminus E_0$ which reflects the structure of the outer measure on $X$ but does not contain any generating set with pre-measure $0$.
Examples for outer measures {#examplesection}
---------------------------
### Example 1: Lebesgue measure via dyadic cubes {#example-1-lebesgue-measure-via-dyadic-cubes .unnumbered}
Let $X$ be the Euclidean space $\R^m$ for some $m\ge 1$ and let $\E$ be the set of all dyadic cubes, that is all cubes of the form $$Q=[2^{k}n_1, 2^{k}(n_1+1))\times \dots \times [2^{k}n_m, 2^{k}(n_m+1))$$ with integers $k,n_1,\dots ,n_m$. For each dyadic cube $Q$ we set $$\sigma(Q)=2^{mk}\ .$$ Then $\sigma$ generates an outer measure which is the classical Lebesgue outer measure on $\R^m$. We have $\sigma(Q)=\mu(Q)$ for every dyadic cube. This latter fact requires a bit of work, in fact it is one of the more laborious items in the standard introduction of Lebesgue measure.
### Example 2: Lebesgue measure via balls {#example-2-lebesgue-measure-via-balls .unnumbered}
Let $X=\R^m$ as above and let $\E$ be the set of all open balls $B_r(x)$ with radius $r$ and center $x\in \R^m$. Let $\sigma(B_r(x))=r^m$ for each such ball. Then $\sigma$ generates a multiple of Lebesgue outer measure, and again we have $\sigma=\left. \mu\right|_\E$.
If one desires a countable generating set, one may restrict the collection of generating sets to the collection of balls which have rational radius and rational center. This choice will result in the same outer measure.
### Example 3: Outer measure generated by tents {#example-3-outer-measure-generated-by-tents .unnumbered}
Let $X=\R\times (0,\infty)$ be the open upper half plane and let $\E$ be the set of tents, that is open isosceles triangles of the form (see Figure 1 in Section \[t1section\]) $$T(x,s)=\{(y,t)\in \R\times (0,\infty): t<s, |x-y|<s-t\}$$ for some pair $(x,s)\in \R\times (0,\infty)$ which describes the tip of the tent. Note that the constraint $t<s$ is implied by the constraint $|x-y|<s-t$, but it is kept for emphasis. Define $\sigma(E)=s$ for any such tent, and note that $\sigma(E)$ is equal to $\frac 12\sigma_L(\pi(E))$ where $\pi(E)$ is the projection of $E$ onto the first coordinate and thus an open ball in $\R$, and $\sigma_L$ is the generator of Lebesgue outer measure on $\R$ described in Example $2$.
By projection onto the first coordinate it easily follows from Example $2$ that $\mu$ satisfies (\[ecoverc\]). Again one obtains the same outer measure restricting the collection of generating sets to the tents with rational tip.
### Example 4: Capacity {#example-4-capacity .unnumbered}
We restrict attention to a particular example of capacity, more examples can be found in the survey [@adams]. Let $X=\R^n$ with $n\ge 3$ and let $\E$ be the collection of open sets in $X$. Define the kernel $K(x):=|x|^{2-n}\ ,$ which is a multiple of the classical Newtonian kernel. Let $\sigma$ assign to each open set its capacity with respect to $K$, that is the least upper bound for the total mass $\|\nu\|$ of a positive Borel measure $\nu$ which has compact support in $E$ and satisfies $\|\nu*K\|_\infty\le 1$. Note that $\sigma(E)>0$ for every nonempty open set $E$, this can be seen by testing with a measure $\nu$ associated with a smooth nonnegative density supported in a small compact ball contained in $E$.
To see Property (\[ecoverc\]), assume $E$ is some open set covered by a countable collection $\E'$ of open sets. Let $ \nu$ be a measure supported on a compact set $F\subset E$ such that $\|\nu*K\|_\infty\le 1$. Then $$\|\nu\|\le \sum_{E'\in \E'}\|\nu 1_{E'}\|$$ $$\le
\sum_{E'\in \E'}\sup_{F\subset E'}
\|\nu 1_F\|$$ $$\le
\sum_{E'\in \E'} \sigma(E') \sup_{F\subset E'}
\|(\nu 1_{F})*K\|_{\infty}$$ $$\le \sum_{E'\in \E'} \sigma(E')
\|\nu*K\|_\infty\le \sum_{E'\in \E'} \sigma(E')\ .$$ Since $\E'$ was arbitrary, this proves $(\ref{ecoverc})$.
Remarks on measurable sets
--------------------------
Outer measures are used in classical textbooks such as [@wz] as a stepping stone towards the introduction of measures. In measure theory, one is interested in equality in under the additional assumption that the sets $E_i$ are pairwise disjoint. Such equality does not follow in general from the properties of outer measure. A sufficient additional criterion is that each of the sets $E_i$ is measurable, as in the following definition.
Let $\mu$ be an outer measure on a set $X$ generated by a pre-measure on a collection $\E$. An arbitrary subset $F$ of $X$ is called measurable if for every generating set $E\in \E$ we have $$\mu (F\cap E ) + \mu(F^c\cap E) = \mu(E) \ .$$
We note that if $F$ is measurable, then it also satisfies the Caratheodory criterion that for arbitrary subset $G$ of $X$ we have $$\mu (F\cap G ) + \mu(F^c\cap G) = \mu(G)\ .$$ We briefly sketch the argument. If $\mu(G)$ is infinite, then it is easy to see that one of the outer measures on the left hand side has to be infinite as well. If $\mu(G)$ is finite, pick $\epsilon>0$ and a cover $\E'$ of $G$ by generating sets such that $$\sum_{E\in \E'} \sigma(E)\le \mu(G)+\epsilon\ .$$ Then we have $$\mu(G)\le \mu(F\cap G)+\mu(F^c\cap G)\le \sum_{E\in \E'}\mu(F\cap E)+\sum_{E\in \E'}\mu(F^c\cap E)$$ $$\le \sum_{E\in \E'} \mu(F\cap E)+\mu(F^c\cap E)=\sum_{E\in \E'} \mu(E) \le \sum_{E\in \E'} \sigma(E)\le \mu(G)+\epsilon\ .$$ Since $\epsilon$ was arbitrary, it follows that the first inequality in this line of reasoning is indeed an equality.
In Example 1 above the measurable sets are called Lebesgue measurable. To see existence of many Lebesgue measurable sets, one observes that dyadic cubes are Lebesgue measurable. This follows from two observations: First one may estimate the outer measure of $F$ by coverings with cubes of side length at most that of the given cube $E$. Second, each such small cube is either contained in $E$ or disjoint from $E$ allowing to split the covering into two disjoint collections, of which one covers $F\cap E$ and the other covers $F\cap E^c$.
One can show in general that the collection of measurable sets is closed under countable union and countable intersection, thus from Lebesgue measurability of dyadic cubes one can conclude Lebesgue measurability of all Borel sets in $\R^m$.
In contrast, no set other than $\emptyset$ and $X$ is measurable in Example 3. For assume we are given a nontrivial subset $E$ of $X$, let $(x_0,s_0)$ be a point in the boundary of $E$ and consider a tent $T(x,s)$ which contains $(x_0,s_0)$ and satisfies $s<2s_0$. Then we find points $(y,t)\in E\cap T(x,s)$ and $(y',t')\in E^c\cap T(x,s)$ in the vicinity of $(x_0,s_0)$ such that $s<t+t'$. Then we have $$\mu(T(x,s))=\sigma(T(x,s))=s<t+t'\le \mu(T(x,s)\cap E)+\mu(T(x,s)\cap E^c)\ ,$$ where we used that if a set $F$ contains a point $(y,t)$, then $\mu(F)>t$ because any cover of $F$ needs to contain a tent with height at least $t$. The last display shows that the set $E$ is not measurable.
In Example 4, it is well known that no bounded open set $E$ is measurable.
Namely, let $E_1$ and $E_2$ be disjoint bounded open sets such that $\dist(E_1,E_2) > 0$ and set $E:=E_1\cup E_2$. Let $\nu$ be a positive Borel measure on a compact subset of $E$ with $\|\nu*K\|_\infty \le 1$. Since $E$ is bounded, for some finite constant $M$ that depends on the diameter of $E$ and $n$ it holds that $$\|\nu\| \le M \inf_{x\in E}(\nu*K)(x) \le M \|\nu*K\|_{\infty} \le M \ \ .$$
In particular, it follows that $\sigma(E),\sigma(E_1),\sigma(E_2) < \infty$ (they are positive from a previous discussion). Then, by inner regularity of Borel measures, we obtain $$\|\nu\|= \sum_{j=1}^2 \|\nu1_{E_j}\|\le \sum_{j=1}^2 \sigma(E_j)\|(\nu 1_{E_j})*K\|_{\infty} \ \ .$$ Using the fact that $\sigma$ satisfies the countably subadditive property , we obtain $\mu(E_j)=\sigma(E_j)$. Using harmonicity of $\nu1_{E_j}*K$ in the interior of $E_j^c$, it follows that $$\|\nu\|\le \sum_{j=1}^2 \mu(E_j)\|(\nu 1_{E_j})*K\|_{\infty}$$ $$\le \sum_{j=1}^2 \mu(E_j)\|(\nu 1_{E_j})*K\|_{L^\infty(\overline{E_j})}$$ $$\le \sum_{j=1}^2 \mu(E_j)(\|\nu*K\|_{\infty} -\inf_{x\in \overline{E_j}} ((\nu 1_{E_{3-j}})*K)(x))$$ $$\le \sum_{j=1}^2 \mu(E_j) -
\sum_{j=1}^2 \mu(E_j)\inf_{x\in \overline{E_j}} ((\nu 1_{E_{3-j}})*K)(x)\ .$$ Since $E$ is bounded, it follows that for some finite positive constant $M$ that depends on the diameter of $E$ and $n$ we have $$\inf_{x\in \overline{E_j}}((\nu1_{E_{3-j}})*K)(x) \ge \frac 1 M \|\nu1_{E_{3-j}}\|$$ It follows that $$\|\nu\| \le
\sum_{j=1}^2 \mu(E_j) - \sum_{j=1}^2 \mu(E_j)\frac 1{M} \|\nu1_{E_{3-j}}\|$$ $$\le \sum_{j=1}^2 \mu(E_j) - \frac 1 M \min(\mu(E_1),\mu(E_2)) \sum_{j=1}^2 \|\nu1_{E_{3-j}}\|$$ $$\le \sum_{j=1}^2 \mu(E_j) - \frac 1 M \min(\mu(E_1),\mu(E_2)) \|\nu\|$$ Since $\mu(E_1)>0$ and $\mu(E_2)>0$, it follows that for some constant $c>0$ that depends only on $E_1,E_2,n$ it holds that $$\|\nu\| \le \frac 1{1+c} \sum_{j=1}^2 \mu(E_j)$$ Taking supremum over all such $\nu$ it follows that $\mu(E)=\sigma(E)<\mu(E_1)+\mu(E_2)$, thus neither $E_1$ nor $E_2$ is measurable. While in Example 3 the lack of measurable sets is intuitively caused by the scarceness of the collection of generating sets, the collection $\E$ in this example is very rich and can hardly be blamed for the shortage of measurable sets.
Functions and sizes
-------------------
We propose an $L^p$ theory for functions on outer measure spaces. One possible way of introducing an $L^p$ norm of a nonnegative function $f$ and $1\le p<\infty$ is via the following definition: $$\label{choquet}
(\int_0^{\infty} p\lambda^{p-1} \mu(\{x\in X: f(x)>\lambda\})\, d\lambda)^{1/p}\ .$$ In many instances, this is the correct definition. However, we propose a different formula, which in many examples such as Lebesgue theory coincides with the above, but differs in full generality. The motivation for our definition is that it appears more useful in the applications that we have in mind.
Our different approach already finds a motivation in the efficiency of encoding of functions in classical Lebesgue theory. Classical coding describes functions as assignment of a value to every point in the space $X$. For an $L^p$ function this assignment has to be consistent with the measurability structure. The set of such assignments has a very large cardinality, which is only reduced after consideration of equivalence classes of $L^p$ functions. This detour over sets of large cardinality can be avoided by coding functions via their averages over dyadic cubes. There are only countably many such averages, and by the Lebesgue Differentiation theorem these averages contain the complete information of the equivalence class of the $L^p$ function.
Unlike in the above definition of $L^p$ norm, which regards the function $f$ as a pointwise assignment, we propose to build the $L^p$ theory on outer measure spaces via averages over generating sets. The theory then splits again into a concrete and abstract part, parallel to the construction of outer measures by generating sets. There will be a concrete procedure to assign to a function averages over generating sets, and further on there will be an abstract procedure to define the $L^p$ norms of functions from such averages. The concrete averaging procedure itself is based on some other measure theory (which by itself might be an outer measure theory, but in the current paper we will not delve into such higher level iteration of the theory). We will consider this other measure theory as concrete external input into the outer measure theory, while the genuine part of the outer measure theory is the abstract passage from the concrete averages to outer $L^p$ norms.
The class of functions that we will be able to take $L^p$ norms of will depend on the concrete averaging procedure we choose. To avoid too abstract a setup we shall assume that $X$ is a metric space, and that every set of the collection $\E$ is Borel. We shall assume the concrete averaging procedure will allow to average positive functions in the class $\B(X)$, the set of Borel measurable functions on $X$. If the set $X$ is countable, a case that exhibits many of the essential ideas of the theory, the space $\B(X)$ is the space of all functions on $X$.
As linearity is closely related with measurability, in the absence of measurability we will not require averages to be linear but merely sub linear or even quasi sub linear. We will call these averages “sizes”.
\[d.size\] Let $X$ be a metric space. Let $\sigma$ be a function on a collection $\E$ of Borel subsets of $X$ and let $\mu$ be the outer measure generated by $\sigma$. A size is a map $$S: \B(X)\to [0,\infty]^\E$$ satisfying for every $f,g\in \B(X)$ and every $E\in \E$ the following properties:
1. Monotonicity: if $|f|\le |g|$, then $S(f)(E)\le S(g)(E)$.
2. Scaling: $S(\lambda f)(E)= |\lambda| S(f)(E)$ for every $\lambda\in \C$.
3. Quasi-subadditivity: $$\label{quasi.sub}
S(f+g)(E)\le C[S(f)(E)+S(g)(E)]$$ for some constant $C$ depending only on $S$ but not on $f,g,E$.
Note that (1) above implies $S(f)(E)=S(|f|)(E)$ for all $f$ and $E$. Hence our theory is essentially one of nonnegative functions, and the size needs initially be only defined for nonnegative Borel functions and can then be extended via the above identity to all functions.
We discuss sizes for Examples 1 through 4, and give a number of forward looking remarks on particular aspects of the outer $L^p$ theory to be developed.
In Lebesgue theory in Example 1, we define for every Borel function $f\in \B(X)$ and every cube $Q$ $$S(f)(Q)=\mu(Q)^{-1}\int_Q |f(x)|\, dx\ .$$ The integral is in the Lebesgue sense. Note the coincidence that the measure theory used to define the size is the same as the measure theory associated with the outer measure $(X,\mu)$. This coincidence is a particular feature of Example 1 (and 2 below). The circularity of this setup does not invalidate our theory, certainly Lebesgue measure can be introduced without reference to the outer integration theory that we develop in this paper.
Note that $S(f)(Q)$ is finite for every locally integrable function on $\R^m$. For such function we may define the “martingale” $$M(f)(Q):=\mu(Q)^{-1} \int_Q f(x)\, dx\ .$$ A consistency condition applies for $M(f)$, namely, the value of $M(f)$ on a dyadic cube is equal to the average of the values on the dyadic subcubes of half the sidelength. By the dyadic Lebesgue Differentiation theorem, the martingale uniquely determines the value of the function $f$ at every Lebesgue point, and this uniquely determines the equivalence class of the measurable function $f$ in Lebesgue sense. As noted before, the martingale is a very efficient way of encoding the function $f$. The space $L^\infty(\R^m)$ can be described as all bounded maps from $\E$ to $\C$ which satisfy the consistency condition. This example is a strong indication that a useful general theory of outer measure may be built out of assigning values to elements $E\in \E$. Indeed, it would be possible in this example to built the theory entirely out of maps $M:\E\to \C$ satisfying the consistency condition, without reference to any Borel function $f$.
Turning to Example 2, we may similarly define $$S(f)(B)=\mu(B)^{-1}\int_B |f(x)|\, dx$$ for every ball $B$. Again, these averages determine $f$ by the Lebesgue Differentiation theorem. In this case there does not exist an easy algebraic consistency condition that identifies maps from $\E$ to $\C$ that arise from locally integrable functions $f$ as the average $$M(f)(B):=\mu(B)^{-1} \int_B f(x)\, dx\ .$$ This provides the evidence that it is impracticable to build a theory of functions on outer measure space entirely out of maps from $\E$ to $\C$ and without reference to a function $f$.
In Example 3 we make an assignment of a value to each tent by averaging a Borel measurable function on the tent: $$\label{primitivetentsize}
S(F)(T(x,s)):=s^{-1} \int_{T(x,s)} |F(y,t)|\, dy\, \frac{dt}{t}\ .$$
This averaging is based on weighted Lebesgue measure on $X$, which however is not the outer measure $(X,\mu)$ in this Example 3. In the literature, one often works with the class of Borel measures $\nu$ on $X$ rather than the class of Borel measurable functions, and defines $$S(\nu)(T(x,s)):=|s|^{-1} |\nu| (T(x,s))\ \ .$$ If the function $S(\nu)$ is bounded, the measure $\nu$ is called a Carleson measure in the literature, the concept of which dates back to the seminal paper [@carleson1]. The space of Carleson measures may be considered the space $L^\infty$ on the outer measure space, as will be discussed more thoroughly further below.
A specific Carleson measure of interest is the following. For some function $f\in L^\infty(\R)$ consider the function $F$ on $X$ defined by $$F(y,t)=\int f(z)t^{-1} \phi(t^{-1}(y-z))\, dz\ ,$$ where $\phi$ is some smooth and rapidly decaying function of integral zero. Then $|F(y,t)|^2 dy \frac{dt}t$ turns out to be a Carleson measure[^1]. The quadratic nature of this example suggests to define a size $$S(F)(T(x,s))=\left( s^{-1}\int _{T(x,s)}|F(y,t)|^2 \, dy\, \frac{dt}t\right)^{1/2}\ .$$ This example provides evidence why we do not try to base a theory of outer measure on linear averaging as could have been done in the example of martingales or the linear averaging over balls.
In Example 4, the most commonly (implicitly) used size is $$S(f)(E)=\sup_{x\in E}|f(x)|\ .$$ Rather than the $L^1$ or $L^2$ based averages from the previous examples, this is an $L^\infty$ based average. Such an $L^\infty$ average has the effect that the more generally defined outer $\L^p$ norms we will introduce specialize to the case of the integral (\[choquet\]), which is frequently referred to as the Choquet integral in the context of capacity theory. We conclude this very brief discussion of Example 4 with the remark that it may be interesting to compare the capacitary strong type inequalites [@adams], whose intensive study goes back to the work of Maz’ya, with the embedding theorems that we discuss further below.
A note on subadditivity
-----------------------
We have chosen to only demand quasi subadditivity in the definition of size. Many sizes will be subadditive, which means that the constant in (\[quasi.sub\]) can be chosen to be $1$. The general constant in (\[quasi.sub\]) allows for certain more general examples , for example $L^p$ type sizes with $p<1$. It also sets the stage for quasi-subadditivity throughout our discussion, which will simplify some of the arguments.
Note that $L^p$ type sizes occur naturally in factorizations. Generalizing the classical factorization $|f| =|f|^{\alpha} |f|^{1-\alpha}$ for a Borel measurable function $f$ and $0<\alpha<1$, one may consider modified sizes $S^{[\alpha]}$ defined, for every nonnegative function $f$, by $$\label{e.sizefactorization}
S^{[\alpha]}(f)(E) := \Big[S(f^{\frac 1{\alpha}})(E)\Big]^\alpha \ \ .$$ One then has the factorization $$S(f) = S^{[\alpha]}(f^{\alpha}) \ \times \ S^{[1-\alpha]}(f^{1-\alpha}) \ \ .$$ Even if $S$ is subadditive, the fractional size $S^{[\alpha]}$ with $0<\alpha<1$ might only be quasi-subadditive.
Essential supremum and super level measure
------------------------------------------
This section contains the most subtle points in the development of our $L^p$ theory on outer measure spaces, with definitions carefully adjusted to the precise setup and the applications we have in mind. To develop an $L^p$ theory we need a space $X$, which we assume to be a metric space. We need a pre-measure $\sigma$ on a collection $\E$ of Borel subsets, generating an outer measure $\mu$ on $X$. Finally, we need a size $S$. So as to not overburden the notation, we collect this data into a triple $(X,\sigma,S)$, because $\sigma$ determines the generating collection and the outer measure. We use the letters $\E$ and $\mu$ for these as standing convention. We call the triple $(X,\sigma,S)$ an outer measure space.
\[essential supremum\] Assume $(X,\sigma,S)$ is an outer measure space. Given a Borel subset $F$ of $X$, we define the outer essential supremum of $f\in \B(X)$ on $F$ to be $$\essup_F S(f):= \sup_{E\in \E} S(f1_F)(E)\ .$$
We emphasize that the values $S(f)(E)$ for fixed $f$ and all $E\in \E$ are in general not enough information to determine the essential supremum of $f$ on a Borel set $F$ other than $X$ or $\emptyset$. It is important to refer back to the function $f$ and truncate it according to the set $F$.
We also emphasize that, unlike in Examples 1 and 2, the outer essential supremum in general does not coincide with the essential supremum of $f$ on $F$ in the Borel sense. In Example 3 with size given by , we note that every Lebesgue integrable Borel function supported above a line $t=t_0>0$ in the space $X$ has finite outer essential supremum. Namely the size of such a function with respect to some tent vanishes if the tent is small and is bounded above by $t_0^{-2}$ times the Lebesgue integral of the function for arbitrary tent. On the other hand, if we define the size $S(f)(E)$ to be the supremum of $f$ on the set $E$, then the outer essential supremum defined above coincides with the classical supremum on the set $F$, under the mild assumption that $F$ can be covered by generating sets $E$.
The following properties of the outer essential supremum are inherited from the corresponding properties for the size. We have for every $f,g\in \B(X)$ and every Borel set $F\subset X$
1. Monotonicity: if $|f|\le |g|$, then $\essup_FS(f)\le \essup_FS(g)$.
2. Scaling: for $\lambda\in \C$ we have $\essup_FS(\lambda f)= |\lambda| \essup_FS(g)$ .
3. Quasi-subadditivity: for some constant $C<\infty$ independent of $f$, $g$, $F$, we have $$\essup_FS(f+g)
\le C (\essup_FS(f)+ \essup_FS(g))\ .$$
The use of the outer essential supremum is the main subtle point in the following definition.
\[superleveldefinition\] Let $(X,\sigma, S)$ be an outer measure space. Let $f\in \B(X)$ and $\lambda>0$. We define $$\label{largerlambdaset}
\mu(S(f)>\lambda)$$ to be the infimum of all values $\mu(F)$, where $F$ runs through all Borel subset of $X$ which satisfy $$\essup_{X\setminus F}S(f) \le \lambda\ .$$
We emphasize once more that in general $\mu(S(f)>\lambda)$ is not the outer measure of the Borel set where $|f|$ is larger than $\lambda$, even though it is precisely that in many special examples such as the case of Lebesgue outer measure or in cases where the outer essential supremum above coincides with the classical supremum.
We obtain the following properties of super level measure.
1. Monotonicity: if $|f|\le |g|$, then $$\mu(S(f)>\lambda)\le \mu(S(g)>\lambda)\ .$$
2. Scaling: for a complex number $\lambda'$ we have $$\mu(S(\lambda' f)>|\lambda'| \lambda) = \mu(S(f)>\lambda)\ .$$
3. Quasi-subadditivity: for some constant $C<\infty$ independent of $f$, $g$, $F$, $$\mu(S(f+g)>C\lambda)\le \mu(S(f)>\lambda)+\mu(S(g)>\lambda)\ .$$
Note that a constant $C=2$ would be necessary in general in the last inequality even if $S$ was sub-additive.
Outer $\L^p$ spaces {#lptheory}
===================
The definition of outer $L^p$ space and subsequent development of the theory of outer $L^p$ spaces follows classical lines of reasoning, once the crucial definitions of the outer essential supremum and the super level measure from the previous section have replaced their classical counterparts. The only minor deviation comes in the proof of the triangle inequality, since we do not have a satisfactory theory of duality in outer $L^p$ spaces. This manifests itself in a loss of a factor $2$ in the triangle inequality.
\[d.linfty\] Let $(X,\sigma, S)$ be an outer measure space. Let $f\in \B(X)$, then we define $$\|f\|_{L^\infty(X,\sigma,S)}
:=\essup_X S(f)=\sup_{E\in \E} S(f)(E)$$ and $L^\infty(X,\sigma,S)$ to be the space of elements $f\in \B(X)$ for which $\sup_{E\in \E} S(f)(E)$ is finite. For notational convenience we define $$L^{\infty,\infty}(X,\sigma,S):=L^\infty(X,\sigma,S)\ .$$
As the Example 3 of Carleson measures shows, $f\in L^\infty(X,\sigma,S)$ need not be an essentially bounded function on $X$ in the Borel sense.
\[d.levelset\] Let $0< p<\infty$. Let $(X,\sigma, S)$ be an outer measure space. We define for $f\in \B(X)$ : $$\|f\|_{L^p(X,\sigma,S)}
:=\left(\int_0^\infty p\lambda^{p-1} \mu(S(f)>\lambda) \, d\lambda\right)^{1/p}\ ,$$ $$\|f\|_{L^{p,\infty}(X,\sigma,S)}:=\left(\sup_{ \lambda >0} \lambda^{p} \mu(S(f)>\lambda)\right)^{1/p}\ .$$ Moreover we define $L^p(X,\sigma,S)$ and $L^{p,\infty}(X,\sigma,S)$ to be the spaces of elements in $\B(X)$ such that the respective quantities are finite.
Clearly $\mu(S(f)>\lambda)$ is monotone in $\lambda$, so that the integral in the definition of $\|f\|_{L^p(X,\sigma,S)}$ is well defined and a number in $[0,\infty]$. As in the classical case we trivially have $$\|f\|_{L^{p,\infty}(X,\sigma,S)}\le \|f\|_{L^p(X,\sigma,S)}\ .$$
The following properties hold, with elementary proofs that follow in most cases from the corresponding statements for super level measure.
Let $(X,\sigma,S)$ be an outer measure space and let $f,g$ be in $\B(X)$. Then we have for $0<p\le \infty$
1. Monotonicity: If $|f|\le |g|$, then $\|f\|_{L^p(X,\sigma,S)} \le \|g\|_{L^p(X,\sigma,S)}$.
2. Scaling: $\|\lambda f\|_{L^p(X,\sigma,S)}=|\lambda| \|f\|_{L^p(X,\sigma,S)}$ for any $\lambda\in \C$.
3. Quasi-subadditivity: there is a constant $C$ independent of $f,g$ such that $$\| f+g\|_{L^p(X,\sigma,S)}\le C(
\|f\|_{L^p(X,\sigma,S)}+\|g\|_{L^p(X,\sigma,S)})\ .$$
Moreover we have for $\lambda>0$ $$\|f\|_{L^p(X,\lambda \sigma,S)} = \lambda^{1/p}\|f\|_{L^p(X,\sigma,S)}\ .$$ Corresponding statements hold for the spaces $L^{p,\infty}(X,\sigma,S)$.
Note that the proof of quasi-subadditivity for $L^p$ with $p<\infty$ is based on quasi-subadditivity of super level measure, which yields a constant $C$ different from $1$ even if the size $S$ is subadditive. It might be interesting to study conditions under which one may have subadditivity for $L^p$.
We turn to the behaviour of outer $L^p$ spaces under mappings between outer measure spaces. Note that Borel measurable functions as well as classical $L^p$ functions are typically pulled back under a continuous map, while in contrast Borel measures are pushed forward under such maps. This is one of the motivations for us to use the class of Borel measurable functions to develop the theory of outer $L^p$ functions, even though much of the theory can be developed for Borel measures as well.
Let $X_1$ and $X_2$ be two metric spaces and let $\Phi:X_1\to X_2$ be a continuous map. For $j=1,2$ let $\E_j$ be a collection of Borel sets covering $X_j$ and let $\sigma_j:\E_j\to [0,\infty]$ be a function generating an outer measure $\mu_j$ on $X_j$. Let $S_1$ and $S_2$ be sizes turning $(X_1,\sigma_1,S_1)$ and $(X_2,\sigma_2,S_2)$ into outer measure spaces.
\[pullback\] Assume that for every $E_2\in \E_2$ we have $$\label{mupullback}
\mu_1(\Phi^{-1} E_2)\le A \mu_2(E_2)\ .$$ Further assume that for each $E_1\in \E_1$ there exists $E_2\in \E_2$ such that for every $f\in \B(X_2)$ we have $$\label{spullback}
S_1(f\circ\Phi)(E_1)\le B S_2(f)(E_2)\ .$$ Then we have for every $f\in \B(X_2)$ and $0<p \le \infty$ and some universal constant $C$: $$\|f\circ \Phi\|_{L^p(X_1,\sigma_1,S_1)}\le A^{1/p}B C \|f\|_{L^p(X_2,\sigma_2,S_2)}\ ,$$ $$\|f\circ \Phi\|_{L^{p,\infty}(X_1,\sigma_1,S_1)}\le A^{1/p}B C \|f\|_{L^{p,\infty}(X_2,\sigma_2,S_2)}\ .$$
First note that by scaling properties it is no restriction to prove the proposition with constants $A=B=1$ in and .
For every Borel set $F_2\subset X_2$ we have $$\mu_1(\Phi^{-1} F_2)\le \mu_2(F_2)\ .$$ Namely, given $F_2$ without loss of generality we may assume that $\mu_2(F_2)<\infty$. Let $\E'_2\subset \E_2$ be a cover of $F_2$ which attains, up to a factor $(1+\epsilon)$ with small $\epsilon>0$, the outer measure of $F_2$: $$\sum_{E_2\in \E'_2} \mu_2(E_2)\le \sum_{E_2\in \E'_2} \sigma_2(E_2)\le (1+\epsilon)\mu_2(F_2)\ .$$ Then we obtain $$\mu_1(\Phi^{-1}(F_2))\le \sum_{E_2\in \E'_2} \mu_1(\Phi^{-1}(E_2))
\le \sum_{E_2\in \E'_2}\mu_2(E_2)\le (1+\epsilon) \mu_2(F_2)\ .$$ This proves the claim, since $\epsilon>0$ can be chosen arbitrarily.
Assume $F\subset X_2$ is a Borel set such that $$\mu_2(F)\le
(1+\epsilon)\mu_2(S_2(f) >\lambda)$$ and for every $E_2\in \E_2$ we have $S_2(f1_{F^c})(E_2)\le \lambda$. Pick $E_1\in \E_1$, then there exists $E_2 \in \E_2$ such that we have $$S_1((f\circ \Phi)(1_{\Phi^{-1}(F^c)}))(E_1)=
S_1((f1_{F^c})\circ \Phi)(E_1)$$ $$\le S_2(f1_{F^c})(E_2)\le \lambda\ ,$$ and hence $$\mu_1(S_1(f\circ \Phi)\ge \lambda)\le
\mu_1((\Phi^{-1}(F^c))^c)$$ $$\le \mu_1(\Phi^{-1}F)\le \mu_2(F)
\le (1+\epsilon) \mu_2(S_2(f)>\lambda)\ .$$ This proves the desired inequalities for $p<\infty$. The case $p=\infty$ follows immediately from the assumption on sizes.
\[p.convex-interpolation\] Let $(X,\sigma,S)$ be an outer measure space and let $f\in \B(X)$. Assume $\alpha_1+\alpha_2=1, \ \ 0< \alpha_1,\alpha_2< 1$, and $$1/p = \alpha_1 / p_1 + {\alpha_2} / {p_2}$$ for $p_1,p_2\in (0,\infty]$ with $p_1\ne p_2$. Then $$\label{convexityc}
\|f\|_{L^{p}(X,\sigma,S)}
\le C_{p,p_1,p_2}\Big(\|f\|_{L^{p_1,\infty}(X,\sigma,S)}
\Big)^{\alpha_1} \Big(\|f\|_{L^{p_2,\infty}(X,\sigma,S)} \Big)^{\alpha_2} \ \ .$$
Assume without loss of generality $p_1<p_2$. We first consider the case $p_2<\infty$. If either of the norms on the right-hand-side of (\[convexityc\]) vanishes, then $\mu(S(f) > \lambda)$ vanishes for all $\lambda>0$ and then the left-hand-side of (\[convexityc\]) vanishes as well. By scaling we may then assume $$A:=\|f\|_{L^{p_1,\infty}(X,\sigma,S)}^{p_1}=\|f\|_{L^{p_2,\infty}(X,\sigma,S)}^{p_2}\ .$$ Optimizing the use of these two identites we have with $p_1<p<p_2$ $$\mu(S(f)>\lambda) \le A \min(\lambda^{-p_2}, \lambda^{-p_1}) \ \ ,$$ $$\|f\|_p \le (Ap(\int_0^{1} \lambda^{p-p_1-1} d\lambda + \int_{1}^\infty \lambda^{p-p_2-1} d\lambda))^{\frac 1 {p}}$$ $$\le C_{p,p_1,p_2} A^{1/p}=
C_{p,p_1,p_2}\Big(\|f\|_{L^{p_1,\infty}(X,\sigma,S,)}^{p_1}
\Big)^{\frac{\alpha_1}{p_1}} \Big(\|f\|_{L^{p_2,\infty}(X,\sigma,S)}^{p_2}
\Big)^{\frac{\alpha_2}{p_2}} \ \ .$$ This completes the proof in case $p_2<\infty$. If $p_2 = \infty$, we may assume by scaling that $\|f\|_{L^{\infty}(X,\sigma,S)}=1$. Then for $\lambda > 1$ we have $\mu(S(f)>\lambda) = 0 $. Consequently, $$\|f\|_{L^p(X,\sigma,S)}
\le (p \|f\|_{L^{p_1,\infty}(X,\sigma,S)}^{p_1}
\int_0^{1} \lambda^{p-p_1-1} d\lambda)^{\frac 1 {p}} \le C_{p,p_1,p_2}
\|f\|_{L^{p_1,\infty}(X,\sigma,S)}^{\alpha_1} \ \ .$$
\[p.hoelder-energy\]
Assume we have a metric space $X$, three collections $\E,\E_1,\E_2$ of Borel subsets, three functions $\sigma,\sigma_1,\sigma_2$ on these collections generating outer measures $\mu,\mu_1,\mu_2$ on $X$. Assume $\mu\le \mu_j$ for $j=1,2$. Assume $S,S_1,S_2$ are three respective sizes such that for any $E\in \E$ there exist $E_1\in \E_1$ and $E_2\in \E_2$ such that for all $f_1,f_2\in \B(X)$ we have $$\label{e.size-factorization}
S(f_1f_2)(E)\le S_1(f_1)(E_1) S_2(f_2)(E_2) \ \ .$$ Let $p, p_1,p_2\in (0,\infty]$ such that $1/p = 1/p_1 + 1/p_2$. Then $$\begin{aligned}
\label{e.hoelder-strong}
\|f_1f_2\|_{L^p(X,\sigma,S)} \le
2 \|f_1\|_{L^{p_1}(X,\sigma_1,S_1)}\|f_2\|_{L^{p_2}(X,\sigma_2,S_2)} \ \ .\end{aligned}$$
We assume $0< p_1,p_2<\infty$, the case $\max(p_1,p_2)= \infty$ can be argued similarly. Without loss of generality assume that the factors on the right hand side of are finite. For $j=1,2$ pick Borel sets $F_j\subset X$ such that for every $E_j\in \E_j$ we have $$S_j(f_j 1_{F_j^c})(E_j)\le \lambda^{p/p_j}$$ and $$\mu_j(F_j)\le \mu_j(S_j(f_j)>\lambda^{p/p_j})+\epsilon\ .$$ Define $F=F_1\cup F_2$. Let $E\in \E$ be arbitrary, then by there exists $E_1 \in \E_1$ and $E_2\in \E_2$ such that $$S(f_1f_2 1_{F^c})(E)\le
S_1(f_1 1_{F^c})(E_1) S_2(f_2 1_{F^c})(E_2)$$ $$\le S_1(f_1 1_{F_1^c})(E_1) S_2(f_2 1_{F_2^c})(E_2)\le \lambda^{p/p_1}\lambda^{p/p_2}=\lambda\ ,$$ the passage from the first to second line by monotonicity of the sizes.
It follows from subadditivity of $\mu$ and domination of $\mu$ by $\mu_1$ and $\mu_2$ that for all $\lambda>0$ $$\label{e.Clambda-break}
\mu(S(f_1f_2)>\lambda) \le \mu(F)\le \mu(F_1)+\mu(F_2)$$
$$\le \mu_1(F_1)+\mu_2(F_2)\le
2\epsilon+ \sum_{i=1}^2 \mu(S_i(f_i) > \lambda^{p/p_i}) \ \ .$$ To prove we may assume via scaling that $$\|f_1\|_{L^{p_1}(X,\sigma_1,S_1)} = \|f_2\|_{L^{p_2}(X,\sigma_2,S_2)} =1 \ \ .$$ Then follows from , using that $\epsilon>0$ is arbitrarily small, $$\int p\lambda^{p-1} \mu(S(f_1f_2)>\lambda) d\lambda \le
\int p\lambda^{p-1} \sum_{i=1}^2 \mu(S_i(f_i) > \lambda^{p/p_i}) d\lambda$$ $$= \sum_{i=1}^2 \int p_i \lambda^{p_i-1} \mu(S_i(f_i) > \lambda ) d\lambda
= 2 \ \ .$$
In the following proposition, let $L^p(Y,\nu)$ denote the classical space of complex valued functions on a measure space $(Y,\nu)$ such that $\|f\|_{L^p(Y,\nu)}:= (\int_Y |f(x)|^p d\nu)^{1/p}$ is finite.
The following proposition is an outer measure version of classical Marcinkiewicz interpolation, which in practice is used to obtain strong bounds in a range of exponents $p$ from weak bounds at the endpoints of the range.
\[p.marcinkiewicz\] Let $(X,\sigma,S)$ be an outer measure space. Assume $1\le p_1<p_2 \le \infty$. Let $T$ be an operator that maps $L^{p_1}(Y,\nu)$ and $L^{p_2}(Y,\nu)$ to the space of Borel functions on $X$, such that for any $f,g\in L^{p_1}(Y,\nu)+L^{p_2}(Y,\nu)$ and $\lambda\ge 0$ we have
1. Scaling: $|T(\lambda f)|=|\lambda T(f)|$.
2. Quasi subadditivity: $|T(f+g)|\le C(|T(f)|+|T(g)|)$.
3. Boundedness properties: $$\|T(f)\|_{L^{p_1,\infty}(X,\sigma,S)}\le A_1\|f\|_{L^{p_1}(Y,\nu)}\ ,$$ $$\|T(f)\|_{L^{p_2,\infty}(X,\sigma,S)}\le A_2\|f\|_{L^{p_2}(Y,\nu)}\ .$$
Then we also have $$\|T(f)\|_{L^{p}(X,\sigma,S)}\le A_1^{\theta_1}A_2^{\theta_2}C_{p_1,p_2,p}\|f\|_{L^{p}(Y,\nu)}\ ,$$ where $p_1<p<p_2$ and $\theta_1$, $\theta_2$ are such that $$\theta_1+\theta_2=1\ ,$$ $$\frac 1p = \frac {\theta_1} {p_1}+\frac {\theta_2}{p_2}\ .$$
We may normalize $\nu$ to become $\tilde{\nu}=\lambda^{-1}\nu$, with $\lambda$ chosen so that $$A_1 \lambda^{1/p_1}=A_2\lambda ^{1/p_2}:=A\ .$$ Then $$A_1^{\theta_1}A_2^{\theta_2}\lambda^{1/p}=A \lambda^{-\theta_1/p_1-\theta_2/p_2} \lambda^{1/p}=A\ .$$ Thus it suffices to prove the theorem with $A_1=A_2=A$. Further normalizing $T$ to become $\tilde{T}=A^{-1}T$, we observe that it suffices to prove the theorem with $A_1=A_2=1$.
If $f_1\in L^{p_1}(Y,\nu)$ and $f_2\in L^{p_2}(Y,\nu)$, then we have for every $E\in \E$ $$S(T(f_1+f_2))(E) \le C\Big( S(Tf_1)(E) + S(Tf_2)(E)) \Big)\ .$$ Then we also have for some possibly different constant $C$: $$\label{e.muTsublinear}
\mu(S(T(f_1+f_2))> C\lambda) \le \mu(S(Tf_1)>\lambda) + \mu(S(Tf_2)>\lambda) \ .$$ We first assume: $0<p_1<p_2<\infty$. Let $f\in L^p(Y,\nu)$. We decompose $f= f_{1,\lambda} + f_{2,\lambda}$ with $f_{1,\lambda}=f 1_{|f|>\lambda}$. It is clear that $f_{j,\lambda}\in L^{p_j}(Y,\nu)$. Using we obtain $$\mu(S(Tf)>C\lambda) \le C \sum_{j=1}^2 \lambda^{-p_j}\|f_{j,\lambda}\|_{p_j}^{p_j}$$ $$=C
\lambda^{-p_1} \int_Y |f|^{p_1} 1_{|f|> \lambda}\, d\nu(y)+
C \lambda^{-p_2} \int_Y |f|^{p_2} 1_{|f|\le \lambda}\, d\nu(y)\ ,$$ and therefore $$\|Tf\|_{L^p(X,\sigma,S)}=\left(p \int_0^\infty \lambda^{p-1} \mu(S(Tf)>\lambda)\, d\lambda\right)^{1/p}$$ $$\le
C\Big( \int_Y |f|^{p_1} (\int_0^{|f|} \lambda^{p-p_1-1} d\lambda) d\nu + \int_Y |f|^{p_2} (\int_{|f|}^\infty \lambda^{p-p_2-1} d\lambda )d\nu\Big)^{1/p} \ \ ,$$ $$\le C \|f\|_{L^p(Y,\nu)} \ \ .$$ It remains to consider the case $p_1<p_2=\infty$. We similarly decompose $f= f_{1,\lambda} + f_{2,\lambda}$ with $f_{1,\lambda}=f 1_{|f|>c\lambda}$ for suitable small $c$ to be determined momentarily. Then $$\| Tf_{2,\lambda}\|_{L^\infty(X,\sigma,S)} \le \|f_{2,\lambda}\|_{L^\infty(Y,\nu)} < c\lambda \ \ .$$ It follows from that with sufficiently small $c$ $$\mu(S(Tf)> \lambda) \le \mu(S(Tf_{1,\lambda})> \lambda/C) + \mu(S(Tf_{2,\lambda})> \lambda/C)
= \mu(S(Tf_{1,\lambda})> \lambda/C) \ \ .$$ Consequently, $$\|Tf\|_{L^p(X,\sigma,S)} \le C \Big(\int_0^\infty \lambda^{p-1} \mu(S(Tf)>\lambda)d\lambda\Big)^{1/p}$$ $$\le C \Big(\int_0^\infty \lambda^{p-1} \mu(S(Tf_{1,\lambda})>\lambda/C)d\lambda\Big)^{1/p}\ .$$
Then we proceed as before to obtain $$\|Tf\|_{L^p(X,\sigma,S)} \le
C \int_Y |f|^{p_1} (\int_{|f|}^\infty\lambda^{p-p_1-1} d\lambda )d\nu\Big)^{1/p} \le C \|f\|_{L^p(Y,\nu)}\ \ .$$
The following is a simple variant of a classical fact about measures: If a measure $\nu$ on a space is absolutely continuous with respect to another measure $\mu$, and if the Radon Nikodym derivative of $\nu$ with respect to $\mu$ is bounded, then the total mass of $\nu$ can be estimated by the total mass of $\mu$.
\[measuredomination\] Assume $(X,\sigma,S)$ is an outer measure space and assume that about every point in $X$ there is an open ball for which there exists $E\in \E$ which contains the ball. Let $\nu$ be a positive Borel measure on $X$. Assume that for every $f\in \B(X)$ and for every $E\in \E$ we have $${\int_E |f| \, d\nu} \le C S(f)(E) \sigma(E)\ .$$
Then, for every $f\in \B(X)$ with finite $\|f\|_{L^\infty(X,\sigma,S)}$ we have: $$|\int_X f\, d\nu |\le C \|f\|_{L^1(X,\sigma,S)}\ ,$$ where the implicit constant $C$ in particular is independent of $\|f\|_{L^\infty(X,\sigma,S)}$.
We may assume that $\mu(S(f)>\lambda)$ is finite for every $\lambda>0$, or else nothing is to prove. For each $k\in \Z$ consider a set $F_k$ such that $$\essup_{F_k^c}S(f) \le 2^k\ ,$$ $$\mu(F_k)\le 2\mu(S(f)>2^k)\ .$$ Cover $F_k$ by a countable subcollection $\E_k$ of $\E$ such that $$\sum_{E\in \E_k}{\sigma(E)}\le 2\mu(F_k)\ .$$
Let $F=\bigcup_k F_k$ and note that for every sufficiently small open ball $B$ about a point in $X$ we can find $E\in \E$ such that $B \subset E$, thus $$\int_{B\cap F^c} |f|1_{F^c}\, d\nu\le C S(f1_{F^c})(E) \sigma(E) =0\ .$$ Hence $$\int_X|f|\, d\nu=\int_F |f|\, d\nu\ .$$ Since we may assume $F_k=\emptyset$ for sufficiently large $k$ we have $$|\int_X f\, d\nu |\le \sum_k \int_{F_k\setminus \bigcup_{l>k} F_l} |f|\, d\nu
\le\sum_k \sum_{E\in \E_k} \int_{E\setminus \bigcup_{l>k} F_l} |f|\, d\nu$$ $$\le \sum_k \sum_{E\in \E_k} S(f1_{F_{k+1}^c}) \mu(E)\le
C\sum_k \sum_{E\in \E_k} 2^k \sigma(E)$$ $$\le C\sum_k 2^k \mu(S(f)>2^k)\le C\|f\|_{L^1(X,\sigma,S)}\ .$$ This completes the proof of the proposition.
Carleson embedding, paraproducts, and the $T(1)$ theorem {#t1section}
=========================================================
This section contains classical results rephrased in the language of outer measure spaces utilizing Example 3 of Section \[examplesection\]. Readers interested in reviewing the classical theory are referred to [@stein]. A novelty of our approach is the interpretation of Carleson embedding theorems as boundedness of certain maps from a classical $L^p$ to an outer $L^p$ space. As a consequence, outer Hölder’s inequality can be used to prove various multi-linear estimates such as paraproduct estimates or a core version of a $T(1)$ theorem.
Carleson embeddings
-------------------
We consider the upper half plane $X=\R\times (0,\infty)$, we let $\E$ be the collection of tents $$T(x,s)=\{(y,t)\in \R\times (0,\infty): t<s, |x-y|<s-t\}\ ,$$ and we set $\sigma(T(x,s))=s$ as in Example 3.
(220,100) (10,10)[(0,1)[70]{}]{} (5,10)[(1,0)[195]{}]{} (205,5)[$y$]{} (5,85)[$t$]{}
(40,10)[(1,1)[20]{}]{} (60,30)[(1,-1)[20]{}]{} (100,10)[(1,1)[40]{}]{} (140,50)[(1,-1)[40]{}]{} (50,40)[$(x,s)$]{} (130,60)[$(x',s')$]{} (60,30) (140,50)
Define for $1\le p<\infty$ the sizes $$S_p (F)(T(x,s)):=(s^{-1}\int_{T(x,s)} |F(y,t)|^p \, dy\, \frac{dt}t)^{1/p}\ \ ,$$ where we have used standard Lebesgue integration in $\R\times (0,\infty)$, and $$S_\infty (F)(T(x,s)):=\sup_{(y,t)\in T(x,s)} |F(y,t)| \ \ .$$
Let $\phi$ be a smooth function on the real line supported in $[-1,1]$ and define for a locally integrable function $f$ on the real line $$\label{definefphi}
F_\phi(f)(y,t):= \int f(x) t^{-1}\overline{\phi(t^{-1}(y-x))}\, dx
\ .$$ The mapping $f\to F_\phi(f)$ is an embedding of a space of functions on the real line into a space of functions in the upper half plane reminiscent of [*Carleson embeddings*]{}. Thus we call the following estimates [*Carleson embedding theorems*]{}, even though traditionally this notion is reserved for special instances and applications of such estimates. In particular, if $\nu$ is a Borel measure on the upper half plane satisfying the so-called Carleson measure condition $\nu(T(x,s))\le Ms$, then one could deduce from Theorem \[carlesonembedding\] a typical version of the classical Carleson embedding theorem, as follows. Below the first and last $L^p$ norm are classical Lebesgue norms while the second and third $L^p$ norms are outer $L^p$ norms over an outer measure generated by $\nu$ and $\sigma$ and the tent collection. $$\|F_\phi(f)\|_{L^p(X,\nu)} \le \|F_\phi\|_{L^p(X,\nu,S_\infty)} \le$$ $$\le M\|F_\phi\|_{L^p(X,\sigma,S_\infty)} \le$$ $$\le C_{p,\phi}M\|f\|_p$$
\[carlesonembedding\] Let $1<p\le \infty$. We have for $\phi$ as above $$\label{mfestimate}
\|F_\phi(f)\|_{L^{p}(X,\sigma,S_\infty)}\le C_{p,\phi}\|f\|_p\ .$$ If in addition $\int \phi=0$, then $$\label{sfestimate}
\|F_\phi(f)\|_{L^{p}(X,\sigma,S_2)}\le C_{p,\phi}\|f\|_p\ .$$
We first prove Estimate (\[mfestimate\]). The estimate will follow by Marcinkiewicz interpolation, Proposition \[p.marcinkiewicz\], between weak endpoint bounds at $p=\infty$ and $p=1$. Clearly we have for all $(y,t)\in X$: $$|F_\phi(f)(y,t)|\le \|f\|_\infty \|\phi\|_1\ .$$ Hence $$S_\infty(F_\phi)(T(x,s))\le \|f\|_\infty \|\phi\|_1$$ for every tent $T(x,s)$, and this implies the $L^\infty$ bound. To prove the weak type estimate at $L^1$, fix $f$ and $\lambda>0$. Consider the set $\Omega \subset \R$ where the Hardy Littlewood maximal function $Mf$ of $f$ is larger than $c_\phi \lambda$ for some constant $c_\phi$ that depends on $\phi$ and is specified later. The set $\Omega$ is open and thus the disjoint union of at most countably many open intervals $(x_i-s_i,x_i+s_i)$ for $i=1,2,\dots$. Let $E$ be the union of the tents $T(x_i,s_i)$. Then the geometry of tents implies that for $(x,s)\not\in E$ none of the intervals $(x_i-s_i,x_i+s_i)$ may contain the interval $(x-s,x+s)$ and hence there is a point $y\in (x-s,x+s)$ such that $Mf(y)\le c_\phi\lambda$. Then we see from a standard estimate of $\phi$ by a superposition of characteristic functions of intervals of length at least $2s$: $$F_\phi(f)(x,s)\le C_\phi Mf(y)\le \lambda\ ,$$ the latter by appropriate choice of $c_\phi$. Hence $$\essup_{E^c} S_\infty(F_\phi)\le \lambda\ .$$ On the other hand, by the Hardy Littlewood maximal theorem, $$\mu(E)\le \sum_i s_i\le |\{x: Mf(x)\ge c_\phi \lambda\}|\le C_\phi \|f\|_1 \lambda^{-1}\ .$$ This proves the weak type estimate at $L^1$ and completes the proof of Estimate (\[mfestimate\]).
We turn to Estimate (\[sfestimate\]), which is proven similarly by Marcinkiewicz interpolation between weak endpoint bounds at $\infty$ and $1$. Note first that if $\phi$ has integral zero, then the map $F_\phi$ goes under the name of “continuous wavelet transform” and is well known to be a multiple of an isometry in the following sense: $$\int_0^\infty \int_\R |F_\phi(g)(y,t)|^2 \, dy \,\frac{dt}t=C_\phi \|g\|_2^2$$ for every $g\in L^2(\R)$. This fact goes under the name of Calderón’s reproducing formula or Calderón’s resolution of the identity, see for example [@daubechies]. It can be proven by a calculation similar to our reduction of Theorem \[bilinearhttheo\] to Lemma \[bhttheorem\] below.
Consider a tent $T(x,s)$. For $(y,t)$ in the tent, we see from compact support of $\phi$ that $$F_\phi(f)(y,t)=F_\phi(f1_{[x-3s,x+3s]})(y,t)\ .$$ Applying Calderón’s reproducing formula with $g=f1_{[x-3s,x+3s]}$ gives $$\int \int_{T(x,s)} |F_\phi(f)(y,t)|^2 \, dy \,\frac{dt}t\le C_\phi \|f1_{[x-3s,x+3s]}\|_2^2\le C_\phi s \|f\|_\infty^2 \ .$$ Dividing by $s$ gives $$S_2(F_\phi(f))(T(x,s))\le C_\phi \|f\|_\infty\ ,$$ which proves the desired estimate for $p=\infty$.
To prove the weak type bound at $p=1$, fix $f\in L^1(\R)$ and $\lambda>0$ and consider again the set $\Omega=\{x: Mf(x)>c_\phi \lambda\}$, which is the disjoint union of open intervals $(x_i-s_i,x_i+s_i)$. Consider the Calderón-Zygmund decomposition of $f$ at level $c_\phi \lambda$: $$f=g+\sum_i b_i$$ which is uniquely determined by the demand that for each $i$ the function $b_i$ is supported on $[x_i-s_i,x_i+s_i]$, and has integral zero, while $g$ is constant on this interval. As a consequence, $g$ is bounded by $c_\phi \lambda$ and we have by the previous argument for any tent $$S_2(F_\phi(g))(T(x,s))\le \lambda/2\ .$$
Let $E$ be the union of tents $T(x_i,3s_i)$. Let $b=\sum_i b_i$. It remains to show that, with small choice of $c_\phi$, for every $(x,s)\in \R\times (0,\infty)$ it holds that $$S_2(F_\phi(b)1_{E^c})(T(x,s))\le \lambda/2 \ \ .$$ Let $B_i$ denote the compactly supported primitive of $b_i$. Then we have for $(y,t)\notin E$, using compact support of $\phi$, $$F_\phi(b)(y,t)= \int b(x) t^{-1}\overline{\phi(t^{-1}(y-x))}\, dy$$ $$=\int \sum_{i: s_i \le t} b_i(x) t^{-1}\overline{\phi(t^{-1}(y-x))}\, dx$$ $$= \int \sum_{i: s_i \le t} B_i(x) t^{-2}\overline{\phi'(t^{-1}(y-x))}\, dx\ .$$ Hence $$|F_\phi(b)(y,t)|\le \|\sum_{i: s_i \le t} t^{-1} B_i \|_\infty \|\phi'\|_1\ .$$ We claim that the $L^\infty$ norm on the right-hand-side is bounded by $4c_\phi\lambda$. Since the $B_i$ are disjointly supported, it suffices to see $\|t^{-1}B_i\|_\infty\le 4c_\phi \lambda$ for each $i$ with $s_i\le t$. However, this follows from $\|b_i\|_1\le 4c_\phi \lambda s_i$, which is a standard estimate for the Calderón Zygmund decomposition. Hence $$S_\infty(F_\phi(b)1_{E^c})(T(s,x))\le 4c_\phi \lambda \ .$$ To obtain a bound for $S_2$ in place of $S_\infty$, we use log convexity of $S_p$ and a bound on $S_1$. Let $T(x,s)$ be a tent and $b_i$ one summand of the bad function. Then we have from considerations of the support of $b_i$ and $\phi$: $$\int_{(y,t)\in T(x,s)\setminus E} |\int_\R b_i(z) t^{-1}\overline{\phi(t^{-1}(y-z))}\, dz|\, dy\, \frac{dt}t$$ $$\le \int_{t \ge s_i} \int _{|y-x_i| \le 2t} |\int_\R b_i(z) t^{-1}\overline{\phi(t^{-1}(y-z))}\, dz|\, dy\, \frac{dt}t\ .$$ Using partial integration we estimate this by $$\int_{t \ge s_i} \int _{|y-x_i| \le 2t} \|B_i\|_1 \|\phi'\|_\infty \, dy\, \frac{dt}{t^3}$$ $$\le C_\phi \int_{t \ge s_i} \|B_i\|_1 \, \frac{dt}{t^2}
\le C_\phi \|B_i\|_1 s_i^{-1}\le C_\phi \|b_i\|_1 \le \lambda s_i/6\ .$$
Adding over the disjointly supported $b_i$ inside $(x-3s,x+3s)$, which are all the summands of the bad function possibly contributing to $F_\phi(b)$ on $T(x,s)$, gives $$S_1(F_\phi(b)1_{E^c})(T(x,s)) \le \lambda /2\ .$$ By log convexity, we then obtain $$S_2(F_\phi(b)1_{E^c})(T(x,s)) \le \lambda/2 \ .$$ Together with the previously established bound for the good function we obtain by the triangle inequality $$S_2(F_\phi(f)1_{E^c})(T(x,s)) \le \lambda$$ and hence $$\essup_{E^c} S_2(F_\phi(f))\le \lambda\ .$$ On the other hand, we have by the Hardy Littlewood maximal theorem as before $$\mu(E)\le C_\phi \|f\|_1 \lambda^{-1}\ .$$ This completes the proof of the weak type $1$ endpoint bound for Estimate \[sfestimate\] and thus the proof of Theorem \[carlesonembedding\].
We will need to apply Theorem \[carlesonembedding\] in a slightly modified setting.
For two parameters $-1\le \alpha\le 1$ and $0< \beta\le 1$ define $$F_{\alpha,\beta,\phi}(f)(y,t):=F_\phi(f)(y+\alpha t, \beta t)\ .$$ To estimate the outer $L^p$ norm of $F_{\alpha,\beta,\phi}(f)$, first note that by a simple change of variables $$s^{-1}\int_{T(x,s)}|F_{\alpha,\beta,\phi}(f)(y,t)|^2\, dy \frac{dt}t
=s^{-1}\int_{T_{\alpha,\beta}(x,s)}|F_\phi(f)(y,t)|^2\, dy \frac{dt}t$$ where we have defined the modified tent $T_{\alpha,\beta}(x,s)$ to be the set of all points $(z,u)$ such that $(z-\alpha \beta^{-1} u,\beta^{-1} u)\in T(x,s)$. This modified tent is a tilted triangle, it has height $\beta s$ above the real line and width $2s$ near the real line. The tip of the tilted tent is the point $(x+\alpha s,\beta s)$, which is contained in a rectangle with base $[x-s,x+s]$ and height $s$ above the $x$-axis. We construct an outer measure space using the collection of modified tents by setting $$\sigma_{\alpha,\beta}(T_{\alpha,\beta}(x,s)):=s\ .$$ We then define for a Borel measurable function $G$ on $X$ $$S_{\alpha,\beta,2} (G)(T_{\alpha,\beta}(x,s)):=(s^{-1}\int_{T_{\alpha,\beta}(x,s)} |G(y,t)|^2 \, dy\, \frac{dt}t)^{1/2}\ .$$ We have by transport of structure $$\|F_{\alpha,\beta,\phi}(f)\|_{L^p(X,\sigma,S_2)}=
\|F_\phi(f)\|_{L^p(X,\sigma_{\alpha,\beta},S_{\alpha,\beta,2})}\ .$$ Given a standard tent $T(x,s)$, we may cover it by a modified tent $T_{\alpha,\beta}(x',s')$ of width $ 2s'=4\beta^{-1}s$. Hence $$\mu_{\alpha,\beta}(T(x,s))\le C \beta^{-1} \mu(T(x,s))\ .$$ Moreover, a modified tent $T_{\alpha,\beta}(x,s)$ is contained in a standard tent $T(x',s')$ of width $2s'=4s$. Hence $$S_{\alpha,\beta,2}(G)(T_{\alpha,\beta}(x,s))\le C S_2(G)(T(x',s'))\ .$$ Thus Proposition \[pullback\] applied to the identity map on $X$ gives $$\|F_\phi(f)\|_{L^p(X,\sigma_{\alpha,\beta}, S_{\alpha,\beta,2})}\le C\beta^{-1/p}
\|F_\phi(f)\|_{L^p(X,\sigma, S_2)}\ .$$ We have thus proven the following corollary.
\[tentcorollary\] Assume the setup as above. Let $1<p\le \infty$ and $-1\le \alpha\le 1$ and $0< \beta\le 1$ and assume $\int \phi=0$. Then $$\|F_{\alpha,\beta,\phi}(f)\|_{L^{p}(X,\sigma,S_2)}\le C_{p,\phi} \beta^{-1/p}\|f\|_p\ .$$
We shall need a slightly better dependence on the parameter $\beta$ in the last corollary. This is stated in the following lemma, where explicit values for $\epsilon$ are not difficult to obtain but unimportant for our purpose.
\[tiltedembedding\] Assume the setup as above. Let $1<p\le \infty$ and $-1\le \alpha\le 1$ and $0< \beta\le 1$ and assume $\int \phi=0$. Then there exists $\epsilon>0$ such that we have $$\|F_{\alpha,\beta,\phi}(f)\|_{L^{p}(X,\sigma,S_2)}\le C_{p,\phi} \beta^{-1/p+\epsilon}\|f\|_p\ .$$
Proof: This lemma follows by various applications of Marcinkiewicz interpolation using the bounds of Corollary \[tentcorollary\] and an improved weak type 2 bound: $$\|F_{\alpha,\beta,\phi}(f)\|_{L^{2,\infty}(X,\sigma,S_2)}=
\|F_{\phi}(f)\|_{L^{2,\infty}(X,\sigma_{\alpha,\beta},S_{\alpha,\beta,2})}
\le C_{\phi} \|f\|_2\ ,$$ where the right-hand-side does not depend on $\beta$. To see this bound, fix $f\in L^2(\R)$ and $\lambda>0$. Consider the collection $\I$ of all open intervals $(x-s,x+s)$ on the real line such that $$s^{-1}
\int_{T_{\alpha,\beta}(x,s)}|F_\phi(f)(y , t)|^2\, dy \frac{dt}t
> \lambda^2\ .$$ The union $\bigcup_{I\in \I}I$ is an open set which can be written as the disjoint union of countably many open intervals $(x_i-s_i,x_i+s_i)$. If we set $E=\bigcup_i T_{\alpha,\beta}(x_i,s_i)$, then it is clear that $$S_{\alpha,\beta,2}(F_{\phi}(f) 1_{E^c})(T_{\alpha,\beta}(x,s)) \le \lambda$$ for each $(x,s)$. Hence it suffices to show $$\label{tiltedtents}
\sum_is_i\le C\lambda^{-2}\|f\|_2^2\ .$$
We first show that if $\I_1\subset \I$ is a collection of disjoint intervals then $$\sum_{I_1\in \I_1}|I_1|\le C\lambda^{-2}\|f\|_2^2\ ,$$ It is clear that such $\I_1$ has to be countable. Enumerate the intervals in $\I_1$ as $(x_1'-s_1',x_1'+s_1')$, $(x_2'-s_2',x_2'+s_2')$ etc. Then we have by choice of the collection $\I$ $$\sum_i s_i'\le \sum_i
\lambda^{-2}
\int_{T_{\alpha,\beta}(x_i',s_i')}|F_\phi(f)(y, t)|^2\, dy
\frac{dt}t\ .$$ However, the tents $T_{\alpha,\beta}(x_i',s_i')$ are pairwise disjoint, and hence $$\sum_i s_i'\le
\lambda^{-2}
\int_{0}^\infty \int_\R |F_\phi(f)(y, t)|^2\, dy
\frac{dt}t\ .$$ By Calderón’s reproducing formula, the latter is bounded by $$C_\phi \lambda^{-2}\|f\|_2^2\ .$$
The above estimate shows in particular that $\sup_{I\in \I} |I|$ is finite, and we may select any $I_1\in \I$ such that $|I_1|>\frac 1 2 \sup_{I\in \I}|I|$. Let $\I'$ be the collection of intervals in $\I$ that does not intersect (or contain) $I_1$. Then select any $I_2 \in \I'$ such that its length is more than half of $\sup_{I\in \I'} |I|$. Iterate this argument we obtain a sequence $I_1,I_2,\dots$ of disjoint intervals in $\I$. We claim that $$\bigcup_{I\in \I} I \subset \bigcup_{j} 5I_j$$ here for any $m>0$ we define $mI_j$ to be the interval of length $m|I_j|$ with the same center as $I_j$. Certainly this claim will imply .
Suppose, towards a contradiction, that there exists $I\in \I$ such that $I\not\subset \bigcup_{j} 5I_j$. We first claim that $I$ intersects one of the intervals $I_1,I_2,\dots$. Indeed, since $|I_j|\to 0$ as $j\to\infty$, there exists $j\ge 1$ such that $|I| > 2|I_{j+1}|$ which means $I$ wasn’t available for selection after step $j$, i.e. $I$ has to intersect one of the intervals $I_1,\dots, I_j$. Now, let $k \ge 1$ be the smallest index such that $I\cap I_k \ne \emptyset$. It follows that $I$ is available for selection after step $k-1$, and hence $|I_k|\ge \frac 1 2 |I|$ and therefore $$I\subset 5I_k$$ which contradicts the above assumption.
This proves the lemma.
Paraproducts and the $T(1)$ theorem
-----------------------------------
A classical paraproduct is a bilinear operator, which after pairing with a third function becomes a trilinear form that is essentially of the type
$$\Lambda(f_1,f_2,f_3)= \int_{\R\times (0,\infty)}
\prod_{j=1}^3 F_{\phi_j}(f_j)(x,t)\, dx\frac{dt}t$$ with three compactly supported smooth functions $\phi_1,\phi_2,\phi_3$ of which two have vanishing integral while the third does not necessarily have vanishing integral. By symmetry we assume $\phi_1$ and $\phi_2$ to have vanishing integral. Paraproducts also appear in different forms in the literature, for example discretized versions of the above integral, or versions involving only two embedding maps $F_i$. In the latter case the third embedding can typically be inserted after using some manipulations on the integral expression.
Assuming $f_j$ are bounded, and thus $F_{\phi_j}(f_j)$ are bounded as well, we obtain by an application of Proposition \[measuredomination\] the estimate $$|\Lambda(f_1,f_2,f_3)| \le C\|\prod_{j=1}^3 F_{\phi_j}(f_j)\|_{L^1(X,\sigma,S_1)}\ .$$ By Hölder’s inequality, once the classical one for the sizes and once Proposition \[p.hoelder-energy\], we obtain $$|\Lambda(f_1,f_2,f_3)| \le C
\|F_{\phi_1}(f_1)\|_{L^{p_1}(X,\sigma,S_2)}\|F_{\phi_2}(f_2)\|_{L^{p_2}(X,\sigma,S_2)}
\|F_{\phi_3}(f_3)\|_{L^{p_3}(X,\sigma,S_\infty)}$$ for exponents $1< p_1,p_2,p_3\le \infty$. By applying the Carleson embedding theorems we obtain $$|\Lambda(f_1,f_2,f_3)| \le C
\|f_1\|_{p_1}\|f_2\|_{p_2}
\|f_3\|_{p_3}\ ,$$ which reproduces classical paraproduct estimates. Note that the last estimate does not depend on the $L^\infty$ bounds on $f_j$, and thus easily extends to unbounded functions. With well known and not too laborous changes in the above arguments one can also reproduce classical ${\rm BMO}$ bounds in place of $p_1=\infty$ or $p_2=\infty$.
We now state a simplified version of the classical $T(1)$ theorem originating in [@david-journe].
\[t1theorem\] Let $\phi$ be some nonzero smooth function supported in $[-1,1]$ with $\int \phi=0$ and define for $x\in \R$ and $s\in (0,\infty)$ $$\phi_{x,s}(y)=s^{-1}\phi(s^{-1}(y-x))\ .$$ Assume $T$ is a bounded linear operator in $L^2(\R)$ such that for all $x,y,s,t$ $$\label{t1assumption}
|\<T(\phi_{x,s}),\phi_{y,t}\>|\le \frac{\min(t,s)}{\max(t,s,|y-x|)^2}\ .$$ Then we have for the operator norm of $T$ the bound $$\|T\|_{L^2\to L^2}\le C$$ for some constant $C$ depending only on $\phi$ and in particular not on $T$. Moreover, for $1<p<\infty$, $$\|T f\|_{p}\le C_p \|f\|_p$$ for some constant $C_p$ depending only on $\phi$ and $p$.
To compare this with more classical formulations of the $T(1)$ theorem, the assumption (\[t1assumption\]) is typically deduced from Calderón-Zygmund kernel estimates if $|x-y|>t+s$ and thus the two test functions $\phi_{x,s}$ and $\phi_{y,t}$ are disjointly supported. It is deduced from one of the assumptions $T(1)=0$ and $T^*(1)=0$ and a weak boundedness assumption if $s$ or $t$ is within a factor of $2$ of the maximum of $s$, $t$, and $|y-x|$ and thus the two test functions are close. The assumptions $T(1)=0$ and $T^*(1)=0$ can be obtained from more general assumptions $T(1)\in {\rm BMO}$ and $T^*(1)
\in{\rm BMO}$ by subtracting paraproducts from $T$ first. A detailed exposition of the $T(1)$ theorem can be found in [@stein].
We note from Calderón’s reproducing formula $$f=C \int_0^\infty \int_\R F(x,s) \phi_{x,s} \, dx\, \frac{ds}{s}$$ with a weakly absolutely convergent integral in $L^2$ and $F=F_\phi(f)$ as defined in (\[definefphi\]). Thus we may write with the analoguous notation $G=F_\phi(g)$ $$\<T(f),g\>=$$ $$=\int_0^\infty \int_0^\infty \int_\R \int_\R
F(x,s) {\<T(\phi_{x,s}),\phi_{y,t}\>} \overline{G(y,t)}
\, dx\,dy \, \frac{ds}
{s}\, \frac{dt}{t}\ .$$
Here we implicitly used boundedness of $T$ and the Schwarz kernel theorem to move $T$ inside the integral representation of $f$. Note that we have again expressed the form $\<T(f),g\>$ in terms of the functions $F$ and $G$ on the outer space $X$, which leads towards the use of embedding theorems. However, we cannot apply Hölder’s inequality directly, but we first have to suitably express the double integral over the space $X$ as superposition of single integrals over $X$.
Set $$r:=\max(s,t,|y-x|)\ .$$ We split the domain of integration into the two regions $r> |x-y|$ and $r=|x-y|$ and estimate the two integrals separately. Splitting the first region further into two symmetric regions (overlapping in a set of measure zero), we may restrict attention to the region $s=r$. We estimate the integral over this region by $$|\int_0^\infty \int_\R
\int_{0}^s \int_{x-s}^{x+s}
F(x,s)
{\<T(\phi_{x,s}),\phi_{y,t}\>}
\overline{G(y,t)} \, dy\,\frac{dt}t
\, dx\, \frac{ds}s|$$ $$\le C
\int_0^\infty \int_\R
\int_{0}^s \int_{x-s}^{x+s}
|F(x,s)
\overline{G(y,t)}| \, dy\,{dt}
\, dx\, \frac{ds}{s^3}$$ $$= C
\int_0^1\int_{-1}^1
\int_0^\infty \int_\R
|F(x,s)
\overline{G(x+\alpha s,\beta s)} | \,
\, dx\, \frac{ds}{s}\,
\, d\alpha \, {d\beta }\ .$$ In the last line we have changed variables setting $y-x=\alpha s$ and $t=\beta s$. Setting ${G}_{\alpha,\beta}(x,s)=G(x+\alpha s,\beta s)$ we estimate the last display, using Propositions \[measuredomination\] and outer Hölder’s inequality with dual exponents $1<p,p'<\infty$, Proposition \[p.hoelder-energy\], $$\le C
\int_0^1\int_{-1}^1
\|F G_{\alpha,\beta}\|_{L^1(X,\sigma,S_1)}
\, d\alpha \, {d\beta }$$ $$\le C
\int_0^1\int_{-1}^1
\|F\|_{L^p(X,\sigma,S_2)}
\|G_{\alpha,\beta}\|_{L^{p'}(X,\sigma,S_2)}
\, d\alpha \, {d\beta }\ .$$ The norm of $F$ can be estimated by Theorem \[carlesonembedding\], while the norm of $G$ can be estimated by Lemma \[tiltedembedding\]. Hence we can estimate the last display by $$\le C
\int_0^1\int_{-1}^1
\beta^{-1/p'}\|f\|_{p}\|g\|_{p'}
\, d\alpha \, {d\beta }
\le
C\|f\|_{p}\|g\|_{p'}\ .$$
The region $r=|y-x|$ we also split into symmetric regions, ~~first~~ restricting to $r=y-x$ and $r=x-y$. By symmetry, it suffices to estimate the region $r=y-x$. We now split further into $t\le s$ and $s\le t$.
We obtain for the subregion $t\le s$ the estimate $$|\int_{0}^\infty\int_\R \int_{0}^r \int_0^s F(x,s)
\< T(\phi_{x,s}),\phi_{x+r,t}\>
\overline{G(x+r,t)} \frac{dt}t \frac {ds}s \, dx\, {dr}|$$ $$\le
\int_0^\infty \int_\R \int_{0}^r \int_0^s |F(x,s)
\overline{G(x+r,t)}| {dt} \frac{ds}s \, dx\, \frac{dr}{r^2}$$ $$=
\int_{0}^1 \int_0^\alpha
\int_0^\infty \int_\R
|F(x,\alpha r)
\overline{G(x+r,\beta r)} |
\, dx\, \frac{dr}r
{d\beta } \frac{d\alpha }\alpha$$ $$=
\int_{0}^1 \int_0^\alpha
\int_0^\infty \int_\R
|F_{0,\alpha}(x,r)
\overline{G_{1,\beta}(x,r)} |
\, dx\, \frac{dr}r
{d\beta } \frac{d\alpha }\alpha \ ,$$ where we have used the notation $F_{0,\alpha}$ and $G_{1,\beta}$ as above. We use Lemma \[tiltedembedding\] twice to estimate the last display by $$\le C
\int_{0}^1 \int_0^\alpha
\alpha^{\epsilon-1/p}\beta^{\epsilon-1/p'} \|f\|_p\|g\|_{p'}
{d\beta } \frac{d\alpha }\alpha \le C\|f\|_p\|g\|_{p'} \ .$$ The subregion $s\le t$ could be estimated similarly. This concludes the proof of the $L^2$ and $L^p$ estimate of Theorem \[t1theorem\].
We conclude this section by pointing at an alternative approach to Calderón Zygmund operators used in A. Lerner’s work [@lerner], who essentially controls a Calderón Zygmund operators by a superposition of “sparse” operators. These sparse operators lend themselves to an application of an outer Hölder inequality with spaces $ L^{\infty}(X,\sigma,S_1)\times L^{p}(X,\sigma,S_\infty)\times L^{p'}(X,\sigma,S_\infty)$ in lieu of the above $ L^{p}(X,\sigma,S_2)\times L^{p'}(X,\sigma,S_2)$ or implicit $ L^{\infty}(X,\sigma,S_\infty)\times L^{p}(X,\sigma,S_2)\times L^{p'}(X,\sigma,S_2)$.
Generalized Tents and Carleson Embedding {#gentents}
========================================
In this section we introduce a new outer measure space whose underlying set is the upper three space. The extra dimension relative to the classical tent spaces is a frequency parameter, which arises due to modulation symmetries in problems of time-frequency analysis. In contrast, the upper half plane merely represents dilation and translation symmetries. The generalized Carleson embedding theorem below is new, though its proof is an adaption of standard recipes in time-frequency analysis. The novelty lies in the concise formulation of an essential part of time-frequency analysis, and in the absence of any discretization in the formulation of Theorem \[gen.carl.emb\].
This section is the most technical one of the present paper, as it is devoted to a proof of Theorem \[gen.carl.emb\] and its discrete variant, Theorem \[gen.carl.emb.disc\]. We point out that the application of Theorem \[gen.carl.emb\] to the bilinear Hilbert transform discussed in the final section can be understood without detailed reading of the proof in the present section.
Let $X$ be the space $\R\times \R\times (0,\infty)$ with the usual metric as subspace of $\R^3$. Let $0<|\alpha|\le 1$ and $|\beta|\le 0.9$ be two real parameters and define for a point $(x,\xi,s)$ in $X$ the generalized tent $$\label{gentent}
T_{\alpha,\beta}(x,\xi,s):=\{(y,\eta,t)\in X: t\le s, |y-x|\le s-t, |\alpha(\eta-\xi)+\beta t^{-1}|\le t^{-1}\}\ .$$ For a first understanding the reader may focus on the example $\alpha=1$ and $\beta=0$. In this case the condition on the frequency variable $\eta$ becomes $-t^{-1} \le \eta-\xi \le t^{-1}$ which is symmetric around $\xi$, as can be seen below. The general case with other $(\alpha,\beta)$ leads to a condition $At^{-1}\le \eta - \xi \le Bt^{-1}$ for some $A<0<B$ depending on $\alpha,\beta$, and will correspond to an asymmetric variant of the Figure below.
(300,120) (20,40)[(1,0)[250]{}]{} (0,20)[(3,1)[120]{}]{} (60,40)[(0,1)[60]{}]{} (275,35)[$y$]{} (55,105)[$t$]{} (120,65)[$\eta$]{} (77,14)[(3,1)[126]{}]{} (157,14)[(3,1)[126]{}]{} (150,65)[(3,1)[60]{}]{} (140,35)[(1,1)[40]{}]{} (180,75)[(1,-1)[40]{}]{} (170,45)[(1,1)[40]{}]{} (210,85)[(1,-1)[40]{}]{} (110,25)[(1,1)[40]{}]{} (150,65)[(1,-1)[40]{}]{} (80,15)[(1,1)[20]{}]{} (160,15)[(-1,1)[20]{}]{} (100,35)[(1,0)[40]{}]{} (200,55)[(1,1)[20]{}]{} (280,55)[(-1,1)[20]{}]{} (220,75)[(1,0)[40]{}]{} (180,75) (165,85)[$(x,\xi, s)$]{} (100,35)(120,40)(150,65) (210,85)(215,78)(220,75) (140,35)(145,40)(150,65) (210,85)(235,78)(260,75)
The projection of the generalized tent onto the first two variables is a classical tent as in Example 3. We are only concerned with generalized tents in this section and will omit the adjective “generalized” when referring to $T_{\alpha,\beta}(x,\xi,s)$. The collection $\E$ of all tents generates an outer measure if we set $$\sigma(T_{\alpha,\beta}(x,\xi,s))=s\ .$$ By a similar argument as in Example 3, $\sigma$ satisfies , and hence the outer measure $\mu$ is an extension of the function $\sigma$ on $\E$.
To define a size on Borel functions on $X$, we use further auxiliary tents $$\label{definegent}
T^b(x,\xi,s):=\{(y,\eta,t)\in X: t\le s, |y-x|\le s-t, |\eta-\xi|\le b t^{-1}\}\ .$$ For $0<b< 1$ and a Borel measurable function $F$ on $X$ we define $$\label{definegens}
S^{b}(F)(T_{\alpha,\beta}(x,\xi,s)):=$$ $$(s^{-1}\int_{T_{\alpha,\beta}(x,\xi,s)\setminus T^b(x,\xi,s)}|F(y,\eta,t)|^2 \, dy\,
d\eta \,{dt})^{1/2}
+ \sup_{(y,\eta,t)\in T_{\alpha,\beta}(x,\xi,s)}|F(y,\eta,t)|\ .$$ One easily checks that this size satisfies the properties required in Definition \[d.size\]. The size $S^b$ increases as $b$ decreases.
The following is a version of a Carleson embedding theorem in the setting of generalized tents. We normalize the Fourier transform of a Schwartz function $\phi$ on the real line as $$\widehat{\phi}(\xi)=\int_\R e^{-i\xi x}\phi(x)\, dx\ .$$
\[gen.carl.emb\]
Let $0<|\alpha|\le 1$ and $|\beta|\le 0.9$. Let $0<b\le 2^{-8}$. Let $\phi$ be a Schwartz function with Fourier transform $\widehat{\phi}$ supported in $(-2^{-8}b,2^{-8}b)$, and let $2\le p\le \infty$. Define for $f\in L^p(\R)$ the function $F$ on $X$ by $$F(y,\eta,t):=\int_\R f(x) e^{i\eta (y-x)} t^{-1}\phi(t^{-1}(y-x)) \, dx\ .$$ There is some constant $C$ depending only on $\alpha$, $\beta$, $b$, $\phi$, and $p$, such that if $p > 2$, $$\|F\|_{\L^{p}(X,\sigma,S^b)}\le C \|f\|_p\ ,$$ and if $p=2$, $$\|F\|_{\L^{2,\infty}(X,\sigma,S^b)}\le C \|f\|_2\ .$$
By symmetry it is no restriction to assume $0<\alpha$ and we shall do so.
The dependence of the constant $C$ on the function $\phi$, conditioned on the fixed support condition on $\widehat{\phi}$, factors as dependence on the constant $$\sup_{x}\left[ |\phi(x)| (1+|x|)^{3} + |\phi'(x)| (1+|x|)^{2}\right]\ .$$ We do not claim that this explicit regularity of $\phi$ is sharp for the above theorem to hold.
From now on we fix the parameters $\alpha$ and $\beta$, and for simplicity of notation write $T$ for $T_{\alpha,\beta}$.
It is convenient to work with a discrete variant of Theorem \[gen.carl.emb\]. Fix the parameter $0<b\le 2^{-8}$. We introduce the discrete subset $X_\Delta$ of points $(x,\xi,s)\in X$ such that there exist integers $k,n,l\in \Z$ with $$x=2^{k-4} n,\ \xi=2^{-k-8} b l,\ s=2^k\ .$$ We denote by $\E_\Delta$ the collection of all tents $T(x,\xi,s)$ with $(x,\xi,s)\in X_\Delta$. This is a discrete subcollection of $\E$. However, each tent in $\E_\Delta$ by itself still forms a continuum in $X$.
We generate an outer measure $\mu_\Delta$ using $\E_\Delta$ as generating collection, setting as before $\sigma_\Delta(T(x,\xi,s))=s$ for each tent in $\E_\Delta$.
The following lemma will be used to relate this new measure to the previous one.
\[discretizationlemma\] If $(x',\xi',s')\in X$, then there exists a $(x,\xi,s)\in X_\Delta$ such that the tent $T(x,\xi,s)$ contains $(x',\xi',s')$ “centrally” in the sense $$2^{-3}s<s'\le 2^{-2}s \ ,$$ $$|x'-x|\le 2^{-4} s\ ,$$ $$|\xi'-\xi|\le 2^{-8}b s^{-1}\ .$$ Moreover, there exist two points $(x,\xi_-,s)\in X_\Delta$ and $(x,\xi_+,s)\in X_\Delta$ so that the corresponding tents contain $(x',\xi',s')$ centrally and satisfy $$T(x',\xi',s')\subset T(x,\xi_-,s)\cup T(x,\xi_+,s)\ ,$$ $$T(x',\xi',s')\cap T^b(x,\xi_-,s)\cap T^b(x,\xi_+,s) \subset T^b(x',\xi',s')
\ .$$
The interval $[2^2s',2^3s')$ contains a unique point of the form $2^k$ with $k\in \Z$. We set $s=2^k$. Then there is a point $x$ of the form $2^{k-4}n$ with some $n\in \Z$ such that $|x-x'|\le 2^{k-4}=2^{-4}s$. Likewise, there is a point $\xi$ of the form $2^{-k-8}bl$ with $l\in \Z$ such that $|\xi-\xi'|\le 2^{-k-8}b=2^{-8}bs^{-1}$.
Informally, this point $\xi$ may be chosen on either side of $\xi'$. Precisely, we may choose $\xi_-\le \xi'$ and $\xi_+\ge \xi'$ with $|\xi_--\xi'|\le 2^{-8}bs^{-1}$ and $|\xi_+-\xi'|\le 2^{-8}bs^{-1}$. If $(y,\eta, t)\in T(x',\xi',s')$, then we have $$t\le s'\le 2^{-2} s\ ,$$ $$|y-x|\le |y-x'|+|x'-x|\le s'-t+2^{-4}s\le s-t\ .$$ If in addition $\alpha(\eta- \xi')+\beta t^ {-1}\ge 0$, then (recall that $\alpha>0$) $$-t^{-1}\le \alpha(\xi'-\xi_+)\le \alpha(\eta-\xi_+)+\beta t^{-1} \le \alpha(\eta-\xi')+\beta t^{-1}\le t^{-1}\ ,$$ while if in addition $\alpha(\eta- \xi')+\beta t^ {-1}\le 0$, then $$- t^{-1}\le \alpha(\eta-\xi')+\beta t^{-1}\le
\alpha(\eta-\xi_-)+\beta t^{-1}\le \alpha(\xi'-\xi_-) \le t^{-1}\ .$$ Hence $(y,\eta,t)\in T(x,\xi_-,s)\cup T(x,\xi_+,s)$. Now let in addition $(y,\eta, t)$ be an element of $T^b(x,\xi_-,s)\cap T^b(x,\xi_+,s)$. If $\eta\ge \xi'$, then $$-bt^{-1} < 0 \le \eta-\xi' \le \eta-\xi_- \le bt^{-1}\ ,$$ while if $\eta\le \xi'$, then $$-bt^{-1}\le \eta-\xi_+\le \eta-\xi' \le 0 < bt^{-1} \ .$$ Hence $(y,\eta,t)\in T^b(x',\xi',s')$. This completes the proof of the lemma.
As a consequence of this lemma, if $T$ is a tent in $\E$, then we find two tents $T^+$, $T^-$ in $\E_\Delta$ such that $$T\subset T^+\cup T^-\ ,$$ $$\sigma_\Delta(T^+)+\sigma_\Delta(T^-)\le C\sigma(T)\ .$$ This implies for every subset $X'\subset X$ $$\mu(X')\le \mu_\Delta(X')\le C \mu(X')\ .$$ Hence the outer measures $\mu$ and $ \mu_\Delta$ are equivalent.
Moreover, we have for the same tents and every Borel function $F$ $$S^b(F)(T)\le C [S_\Delta^b (F)(T^+)+S_\Delta^b(F)(T^-)]\ ,$$ where we have defined $$S^b_\Delta(F)(T'):= S^b (F)(T')\ .$$ for any tent $T'$ in $\E_\Delta$.
This implies for every $1\le p\le \infty$ $$C^{-1}\L^{p}(X,\sigma,S^b)
\le \L^{p}(X,\sigma_\Delta,S^b_\Delta)\le
\L^{p}(X,\sigma,S^b)\ ,$$ $$C^{-1}\L^{p,\infty}(X,\sigma,S^b)
\le \L^{p,\infty}(X,\sigma_\Delta,S^b_\Delta)\le
\L^{p,\infty}(X,\sigma,S^b)\ .$$
Hence Theorem \[gen.carl.emb\] is equivalent to the following discrete version.
\[gen.carl.emb.disc\]
Let $0<\alpha\le 1$ and $-0.9\le \beta\le 0.9$. Let $0<b\le 2^{-8}$. Let $\phi$ be a Schwartz function with Fourier transform $\widehat{\phi}$ supported in the interval $(-2^{-8}b,2^{-8}b)$, and let $2\le p\le \infty$. Define for $f\in L^p(\R)$ the function $F$ on $X$ by $$F(y,\eta,t):=\int_\R f(x) e^{i\eta (y-x)} t^{-1}\phi(t^{-1}(y-x)) \, dx\ .$$ There is some constant $C$ depending only on $\alpha$, $\beta$, $b$, $\phi$, and $p$, such that if $p \neq 2$, $$\|F\|_{\L^{p}(X,\sigma_\Delta,S^b_\Delta)}\le C \|f\|_p\ ,$$ and if $p=2$, $$\|F\|_{\L^{2,\infty}(X,\sigma_\Delta,S^b_\Delta)}\le C \|f\|_2\ .$$
Since both theorems are equivalent, we will only prove the discrete version, Theorem \[gen.carl.emb.disc\]. Hence we will only work with the discrete quantities $\mu_\Delta$ and $S^b_\Delta$ and for simplicity of notation omit the subscribt $\Delta$. Since $b$ is fixed, we also denote $S:= S^b$.
The theorem follows by Marcinkiewicz interpolation, Proposition \[p.marcinkiewicz\], between the end point cases $p=2$ and $p=\infty$.
The endpoint $p=\infty$
-----------------------
We need to prove that for every $(x,\xi,s)\in X_\Delta$ and every $f\in L^\infty(\R)$ we have $$S(F)(T(x,\xi,s))\le C\|f\|_\infty\ \ .$$ The size $S$ is defined as a sum of an $L^2$ portion and an $L^\infty$ portion. It suffices to estimate both portions separately. Note that for all $y,\eta,t$ we trivially have $|F(y,\eta,t)|\le \|f\|_\infty \|\phi\|_1$ and this establishes the desired bound on the $L^\infty$ portion of $S$.
To estimate the $L^2$ portion of the size we first establish the estimate $$\label{l2tentbound}
\int_{T(x,\xi,s)\setminus T^b(x,\xi,s)}|F(y,\eta,t)|^2 \, dy\,d\eta\, dt\le
C\|{f}\|_2^2$$ for every function $f\in L^2(\R)$. Fix such a function $f$, we may assume by normalization that $\|f\|_2=1$. Replacing the domain of integration by a larger region we can estimate the left-hand-side of (\[l2tentbound\]) by $$\int_0^\infty \int_\R \int_{bt^{-1}\le |\eta-\xi|\le 2\alpha^{-1} t^{-1}} |F(y,\eta,t)|^2\, d\eta \, dy \, {dt}\ .$$ It suffices to estimate the integral over the region where $\eta> \xi$ , since by symmetry there is an analoguous estimate for the integral over region $\eta<\xi$. We replace the integration variable $\eta$ by $\gamma$ such that $\eta-\xi=\gamma t^{-1}$. Using Fubini we are reduced to estimating $$\int_{b}^{2\alpha^{-1}}
\int_0^\infty \int_\R |F(y,\xi+\gamma t^{-1},t)|^2 \, dy \, \frac {dt}t \, d\gamma\ .$$ We first estimate the inner double integral for fixed $\gamma$.
Define for each $y,\gamma,t$ the bump function $\phi_{y,\gamma,t}$ by $$\phi_{y,\gamma,t}(x)= e^{-i(\xi+\gamma t^{-1}) (y-x)} t^{-1}\overline{\phi(t^{-1}(y-x))} \ .$$ We are interested in the region $\gamma\ge b$, where the modulated function $\phi_{y,\gamma,t}e^{-i\xi.}$ has integral zero by support consideration of $\widehat{\phi}$. Hence its primitive is absolutely integrable with good bounds, which we will use later when applying partial integration.
We have $$( \int _0^\infty \int_\R
|\<f,\phi_{y,\gamma ,t}\> |^2 \, dy\, \frac{dt}t)^2$$ $$\le \| \int_0^\infty \int_\R \<f,\phi_{y,\gamma ,t}\>
\phi_{y,\gamma ,t} \, dy\, \frac{dt}t\|_2^2$$ $$\le \int _0^\infty \int_\R \int _0^\infty \int_\R
|\<f,\phi_{y,\gamma ,t}\>
\<\phi_{y,\gamma ,t}, \phi_{z,\gamma ,r} \>
\<\phi_{z,\gamma ,r},f\>|
\, dz\, \frac{dr}r
\, dy\, \frac{dt}t \ .$$ Estimating the smaller of the inner products with $f$ by the larger one and using symmetry we may estimate this by $$\label{smallbigf}
\le 2\int _0^\infty \int_\R |\<f,\phi_{y,\gamma ,t}\>|^2
[ \int _0^\infty \int_\R
|\<\phi_{y,\gamma ,t}, \phi_{z,\gamma ,r} \>|
\, dz\, \frac{dr}r
]\, dy\, \frac{dt}t \ .$$
We consider the inner double integral of (\[smallbigf\]). Considering first the region $t\le r$ and doing partial integration in the inner product $$\<\phi_{y,\gamma ,t}, \phi_{z,\gamma ,r} \>=
\<\phi_{y,\gamma ,t}e^{-i\xi .}, \phi_{z,\gamma ,r} e^{-i\xi.}\>\ ,$$ integrating the first and differentiation the second bump function, we estimate the integral over this region by $$C\int_\R \int_t^\infty
\int_\R
(1+|t^{-1}(y-x)|)^{-2}
r^{-2}(1+|r^{-1}(z-x)|)^{-2} dx \frac{dr}r\, dz$$ $$\le C\int_\R \int_t^\infty
(1+|t^{-1}(y-x)|)^{-2}
r^{-2} {dr}\, dx$$ $$\le C\int_\R
t^{-1}(1+|t^{-1}(y-x)|)^{-2}
\, dx\le C\ .$$ In the region $t\ge r$ we do partial integration in reverse, differentiating the first and integrating the second bump function, to obtain the estimate for the integral over this region by $$C\int_\R \int_0^t \int_\R
t^{-2}(1+|t^{-1}(y-x)|)^{-2}
(1+|r^{-1}(z-x)|)^{-2} dx \frac{dr}r\, dz$$ $$\le C\int_\R \int_0^t
t^{-2}(1+|t^{-1}(y-x)|)^{-2}
{dr}\, dx$$ $$\le C\int_\R
t^{-1}(1+|t^{-1}(y-x)|)^{-2}
\, dx\le C\ .$$ Inserting these two estimates into (\[smallbigf\]) gives $$( \int _0^\infty \int_\R
|\<f,\phi_{y,\gamma ,t}\> |^2 \, dy\, \frac{dt}t)^2
\le C
\int _0^\infty \int_\R
|\<f,\phi_{y,\gamma ,t}\> |^2 \, dy\, \frac{dt}t\ ,$$ which proves (\[l2tentbound\]).
We note that if we restrict the integral on the left hand side of (\[l2tentbound\]) to the region $\eta>\xi$, we may improve the bound on the right-hand-side to $$\label{fxibound}
C\|\widehat{f}1_{(\xi,\infty)}\|_2^2
\ .$$ This follows simply by support considerations on the Fourier transform side.
Now assume that $f\in L^\infty(\R)$ and write $f=f_1+f_2$ where $$f_1=f1_{[x-2s,x+2s]}\ .$$ By linearity we may split $F=F_1+F_2$ correspondingly. We have $\|f_1\|_2^2\le Cs\|f\|_\infty^2$, so by the above $L^2$ bound we have $$(s^{-1}\int_{T(x,\xi,s)\setminus T^b(x,\xi,s)}|F_1(y,\eta,t)|^2 \, dy\, d\eta \,{dt})^{1/2}
\le C\|f\|_\infty\ .$$ It remains to prove the analoguous estimate for $F_2$. But for $y\in [x-s,x+s]$ and $t<s$ we have $$F_2(y,\eta, t)\le \int_{[-s, s]^c} |f_2(y-z)| t^{-1} |\phi(t^{-1}z)|\, dz \le C (t/s)\|f\|_\infty\,$$ where we have crudely estimated the integral of the tail of $\phi$. But then $$(s^{-1}\int_{T(x,\xi,s)\setminus T^b(x,\xi,s)}|F_2(y,\eta,t)|^2 \, dy\, d\eta \,{dt})^{1/2}$$ $$\le C\|f\|_\infty (s^{-1}\int_{0}^s \int_{\xi-2\alpha^{-1}t^{-1}}^{\xi+2\alpha^{-1}t^{-1}}
\int_{x-s}^{x+s} (t/s)^{2} \, dy\, d\eta \,{dt})^{1/2}
\le C\|f\|_\infty\ .$$ This completes the proof of the endpoint $p=\infty$ of Theorem \[gen.carl.emb\].
The endpoint $p=2$
------------------
We need to find for each $\lambda>0$ a collection $\Q \subset X_\Delta$ such that $$\sum_{(x,\xi,s)\in \Q} s\le C\lambda^{-2} \|f\|_2^2$$ and for every $T'\in \E_\Delta$ we have $$\label{small2inftysize}
S(F 1_{X\setminus E})(T')\le \lambda\ ,$$ where $E=\bigcup_{(x,\xi,s)\in \Q} T(x,\xi,s)$.
We first reduce to the special case that the support of $\widehat{f}$ is compact. Choose an unbounded monotone increasing sequence $\xi_k$, $k=0,1,2,\dots$ with $\xi_0=0$ such that for $f_k$ defined by $\widehat{f}_k=\widehat{f}[1_{(-\xi_{k},-\xi_{k-1})}+1_{(\xi_{k-1},\xi_{k})}]$ we have $$\|f_k\|_2\le C2^{-10k}\|f\|_2\ .$$ Applying the special case to each of the functions $f_k$ with $\lambda_k=2^{-k}\lambda $ we obtain corresponding collections $\Q_k$. Then clearly $$\sum_{k=1}^\infty \sum_{(x,\xi,s)\in \Q_k} s\le
C \sum_k (2^{-k}\lambda)^{-2} 2^{-20k} \|f\|_2^2
\le C \lambda^{-2} \|f\|_2^2\ .$$ If $E$ denotes the union of all $T(x,\xi,s)$ with $(x,\xi,s)\in \bigcup_k \Q_k$, then by countable subadditivity of the size $S$ we have for every $T'\in \E_\Delta$ $$S(F1_{X\setminus E})(T')\le \sum_{k=1}^\infty S(F_k1_{X\setminus E})(T')
\le \sum_{k=1}^\infty 2^{-k}\lambda \le \lambda\ .$$ This completes the reduction to the case that $\widehat{f}$ has compact support, and we shall henceforth assume compact support of $\widehat{f}$.
By scaling of outer Lebesgue spaces we may assume $\|f\|_2=1$. Fix $\lambda>0$. We first set out to cover all points $(y,\eta,t)\in X$ with $|F(y,\eta,t)|> \lambda$ with tents. Note that there is an a priori upper bound on $t$ for any such point since by Cauchy-Schwarz we have directly from the definition of $F$: $$|F(y,\eta,t)|\le C t^{-1/2} \|f\|_2 \|\phi\|_2\le Ct^{-1/2}\ .$$ Assume there is a point $(y,\eta,t)$ with $|F(y,\eta,t)|> \lambda$, then by Lemma \[discretizationlemma\] we find a tent $T(x,\xi,s)$ centrally containing the point $(y,\eta,t)$. Because of the upper bound on $t$ and since $(x,\xi,s)\in X_\Delta$ and therefore $s=2^k$ for some integer $k$, we may choose $(y,\eta,t)$ and $(x,\xi,s)$ such that $s$ is maximal. Denote these points by $(y_1,\eta_1,t_1)$ and $(x_1,\xi_1,s_1)$ and the tent $T(x_1,\xi_1,s_1)$ by $T_1$.
We continue to select tents by iterating this procedure. Assume that we have already chosen points $(y_k,\eta_k,t_k)\in X$ and tents $T_k=T(x_k,\xi_k,s_k)$ for all $1\le k< n$. Assume there is a point $(y,\eta,t)$ with $|F(y,\eta,t)|> \lambda$ not contained in the union of the tents $T_k$ with $1\le k<n$. Then we choose such a point $(y_n,\eta_n,t_n)$ and a tent $T_n=T(x_n,\xi_n,s_n)$ centrally containing $(y_n,\eta_n,t_n)$ such that $s_n$ is maximal. We have $|F(y_n,\eta_n,t_n)|> \lambda$ and $$(y_n,\eta_n,t_n)\not\in \bigcup_{k=1}^{n-1} T_k\ .$$
We claim that $$\label{count.tiles}
\sum_{k=1}^n s_k \le C\lambda^{-2}\ .$$ To see the claim, let $K_m$ be the set of indices $k$ with $1\le k\le n$ such that $$2^m\lambda
\le |F(y_k,\eta_k,t_k)|\le 2^{m+1} \lambda \ .$$ Then we have $$\sum_{k=1}^n s_k
\le C \sum_{m=0}^\infty
2^{-2m} \lambda^{-2}
\sum_{k\in K_m} t_k |F(y_k,\eta_k,t_k)|^2\ .$$ The claim (\[count.tiles\]) will follow if we show for fixed $m\ge 0$ $$\label{seta}
\sum_{k\in K_m} t_k |F(y_k,\eta_k,t_k)|^2 \le C \ .$$ Define $$\label{wavepacket}
\phi_k(x):=\phi_{y_k,\eta_k,t_k}(x):=e^{-i\eta_k (y_k-x)}t_k^{-1/2}\overline{\phi(t_k^{-1}(y_k-x))}\ ,$$ so that $$t_k^{1/2}F(y_k,\eta_k,t_k)=\<f,\phi_k\>\ .$$ Let $A$ denote the left-hand-side of (\[seta\]). Assume we can show for every $k\in K_m$ $$\label{modifiedschurc}
\sum_{l\in K_m: s_l\le s_k} (t_l/t_k)^{1/2}|\<\phi_k,\phi_l\>|\le C \ .$$ Then we obtain $$A^2 \le \|\sum_{k\in K_m} \<f,\phi_k\>\phi_k\|_2^2$$ $$\le \sum_{k,l\in K_m} \<f,\phi_k\>\<\phi_k,\phi_l\> \<\phi_l,f\>$$ $$\le 2\sum_{k,l\in K_m: s_l\le s_k} |\<f,\phi_k\>\<\phi_k,\phi_l\> \<\phi_l,f\>|$$
$$\le C\sum_{k,l\in K_m: s_l\le s_k} (t_l/t_k)^{1/2}
|\<f,\phi_k\>|^2 |\<\phi_k,\phi_l\>| \le C A \ .$$ Here in the passage from the penultimate to ultimate line we have used that $k,l\in K_m$ and hence $t_k^{-1/2}|\<f,\phi_k\>|$ and $t_l^{-1/2}|\<f,\phi_l\>|$ are within a factor of $2$ of each other and in the last line we have used (\[modifiedschurc\]). Dividing by $A$ on both sides of the displayed inequality we have reduced the proof of the desired estimate (\[seta\]) to the proof of (\[modifiedschurc\]).
To prove (\[modifiedschurc\]), fix $k$. If $l\in K_m$ with $s_l\le s_k$ such that $\<\phi_k,\phi_l\>\neq 0$ then the supports of $\widehat{\phi}_k$ and $\widehat{\phi}_l$ overlap and hence there are numbers $- 2^{-8} \le \gamma,\delta \le 2^{-8}$ such that $$\eta_l+\delta b t_l^{-1}=\eta_k+\gamma b t_k^{-1}
\ .$$ Now suppose that there is another such $l'$ and we have analoguously $$\eta_{l'}+\delta' b t_{l'}^{-1}=\eta_k + \gamma' b t_k^{-1}
\ .$$ Assume without loss of generality that $T_{l}$ is selected prior to $T_{l'}$ and thus $s_k^{-1}\le s_{l}^{-1}\le s_{l'}^{-1}$. Then we have by central containment of $(y_l,\eta_l,t_l)$ in $T_l$ $$|\alpha (
\eta_{l'}- \xi_l)+\beta t_{l'}^{-1}|\le
|\alpha (
\eta_{l}-\xi_l)|+|\alpha(\eta_l-\eta_{l'})|+|\beta t_{l'}^{-1}|$$ $$\le b t_{l}^{-1}+|\eta_l-\eta_{l'}| + |\beta t_{l'}^{-1}|
\ .$$ Now using the information from the support of the bump functions we may estimate the latter by $$\le b t_l^{-1}
+\delta b t_l^{-1}+\delta' b t_{l'}^{-1}
+\gamma b t_k^{-1}+\gamma' b t_k^{-1}+|\beta t_{l'}^{-1}|
\le t_{l'}^{-1}
\ .$$ Here we have used again the central containment to estimate the inverse powers of $t_k$ and $t_l$ by that of $t_{l'}$.
This implies that $[x_l-2^{-8}s_l,x_l+2^{-8}s_l]$ and $[x_{l'}-2^{-8}s_{l'},x_{l'}+2^{-8}s_{l'}]$ are disjoint. For if they were not disjoint, then, since $s_{l'}\le s_l$, we would conclude $$|y_{l'}-x_l|\le |y_{l'}-x_{l'}|+|x_{l'}-x_{l}|$$ $$\le 2^{-4}s_{l'}+2^{-4}s_l < s_l-t_{l'}$$ and together with the previous estimate for $\xi_l-\eta_{l'}$ this implied that the point $(y_{l'},\eta_{l'},t_{l'})$ was in the tent $T_l$, contradicting the choice of this point.
The argument above in particular shows that $|x_k-x_l|\ge 2^{-8}s_k$. Let $\overline{x}$ be the midpoint of $x_l$ and $x_k$ and let $H_l$ and $H_{k}$ be the half lines emanating from the midpoint containing $x_l$ and $x_{k}$ respectively. We then have $$|\<\phi_k,\phi_l\>|\le \|\phi_k\|_{L^1(H_k)}\|\phi_l\|_{L^\infty(H_k)}+
\|\phi_k\|_{L^\infty(H_l)}\|\phi_l\|_{L^1(H_l)}\ .$$ Thanks to the rapid decay of the wave packets and $s_l\le s_k$ and the fact that $|x_k-x_l| \ge 2^{-8}s_k$ we can estimate the last display by $$C (s_ls_k)^{-1/2}\int (1+(\frac{x-x_k}{s_k})^2
)^{-2} 1_{[x_l-2^{-8}s_l,x_l+2^{-8}s_l]}(x)\, dx\ .$$
By disjointness of the intervals $[x_l-2^{-8}s_l,x_l+2^{-8}s_l]$ for different $l$ we obtain $$\sum_{l\in K_m,s_l\le s_k}
s_l^{1/2} |\<\phi_k,\phi_l\>|$$ $$\le C s_k^{-1/2}\int (1+(\frac{x-x_k}{s_k})^2)^{-1}\, dx\le C s_k^{1/2}\ .$$ This proves (\[modifiedschurc\]) since $t_l$ and $t_k$ are are comparable to $s_l$ and $s_k$, and hence completes the proof of (\[count.tiles\]).
If the iterative selection of tents $T_n$ stops because of lack of suitable points $(y,\eta,t)$ with large enough value $F(y,\eta,t)$, then clearly $F$ is bounded by $ \lambda$ outside the union $\bigcup_{k=1}^{n-1} T_k$. If the iterative selection does not stop, we claim that still $F$ is bounded above by $\lambda$ outside the union $\bigcup_{k=1}^\infty T_k$. Namely, assume the point $(y,\eta,t)$ is outside this union. Since by (\[count.tiles\]) we have $s_k\to 0$ as $k\to \infty$, we have $s_k<t$ for some $k$. By maximal choice of $s_k$ we have $F(y,\eta,t)\le \lambda$. This proves the desired bound on $F$. In the case of infinitely many selected tents $T_k$, it also follows by a limiting argument from (\[count.tiles\]) that $\sum_{k=1}^\infty s_k \le C\lambda^{-2}$ .
Summarizing, we have found a collection $\Q_0$ of tents such that $$\sum_{T\in \Q_0}^\infty \sigma(T) \le C \lambda^{-2}$$ and if we set $$E=\bigcup_{T\in \Q_0} T\ ,$$ then we have $$F(y,\eta,t)\le \lambda$$ for all points $(y,\eta,t)$ in the complement of $E$. In what follows, we shall no longer need the selected tents explicitly, and hence we shall free the symbols $T_k,x_k,\xi_k,s_k$ to have new meanings in the further selection process.
We need to select tents of large $L^2$ portion of the size. Given a number $\xi$, typically arising as second parameter of a tent $T(x,\xi,s)$, we split the space $X$ into upper half $$X_\xi^+=\{(y,\eta,s)\in X: \eta\ge \xi\}$$ and lower half $X_\xi^-=X\setminus X_\xi^+$. We first focus on $X_\xi^+$.
Call a point $(x,\xi,s)\in X_\Delta$ bad, if $$\label{lowerl2bound}
s^{-1} \int_{(T(x,\xi,s)\cap X_\xi^+)\setminus (T^b(x,\xi,s)\cup E)}|F(y,\eta,t)|^2\, dy\, d\eta \, {dt}\ge 2^{-8}\lambda^2\ .$$ By the estimate (\[l2tentbound\]) we obtain an a priori upper bound $2^{k_{\max}}$ for the third component $s$ of any bad point $(x,\xi,s)$. Given such an upper bound, the parameter $\xi$ becomes a multiple of $2^{-8-k_{\max}}b$ and is thus a discrete parameter. Since $\widehat{f}$ has compact support, we obtain from observation (\[fxibound\]) an upper bound for $\xi$ depending on the support of $\widehat{f}$. Hence there is a maximal possible value $\xi_{\max}$ for the second component of a bad point. We choose some bad point $(x_1,\xi_1,s_1)$ with $\xi_1=\xi_{\max}$ which maximizes $s_1$ under the constraint $\xi_1=\xi_{\max}$. Define the tents ${T}_1=T(x_1,\xi_1,s_1)$ and ${T}_1^b=T^b(x_1,\xi_1,s_1)$, and define $X_1^+=X_{\xi_1}^+$. Note that by maximizing $s_1$ for fixed $\xi_1$ and $x_1$ we guarantee that is sharp up to a factor of $2$ and hence the selected tent satisfies an upper bound $$s^{-1} \int_{(T(x,\xi,s)\cap X_\xi^+)\setminus (T^b(x,\xi,s)\cup E)}|F(y,\eta,t)|^2\, dy\, d\eta \, {dt}\le \lambda^2\ .$$
Now we iterate this selection: assume we have already chosen points $(x_k,\xi_k,s_k)\in X_\Delta$ for $1\le k<n$ and we have defined tents ${T}_k$, $T_k^b$ for $1\le k<n$. Define $E_n=E\cup \bigcup_{k=1}^{n-1} {T}_k$. We update the definition of a bad point $(x,\xi,s)$ to be a point in $X_\Delta$ with $$s^{-1}\int_{(T(x,\xi,s)\cap X_\xi^+)\setminus (T^b(x,\xi,s)\cup E_n)}
|F(y,\eta,t)|^2\, dy d\eta\, dt
\ge 2^{-8}\lambda^2\ .$$ Again, there is a (possibly new) maximal value $\xi_{\max}$ for the second component $\xi$ of a bad point. We pick one bad point $(x_n,\eta_n,s_n)$ with $\xi_n=\xi_{\max}$ which maximizes the value of $s_n$ among all bad points $(x,\xi,s)$ with $\xi=\xi_{\max}$. Then we define the tents ${T}_n=T(x_n,\xi_n,s_n)$ and ${T}_n^b=T^b(x_n,\xi_n,s_n)$ and define $X_n^+=X_{\xi_n}^+$. This completes the $n$-th selection step.
We introduce the notation $$T_n^*= (T_n\cap X_n^+)\setminus (T_n^b\cup E_n)\ .$$
We claim the analogue of (\[count.tiles\]), namely $$\label{count.tiles.plus}
\sum_{k=1}^n s_k \le C\lambda^{-2} \ .$$ To prove (\[count.tiles.plus\]), it suffices to show $$\sum_{k=1}^n \int_{T_k^*}|F(y,\eta,t)|^2 \, dyd\eta dt \le C\ .$$ With $\phi_{y,\eta,t}$ defined analoguously to (\[wavepacket\]) we may write for the left hand side of the last display $$A:=\sum_{k=1}^n \int_{T_k^*}|\<f,\phi_{y,\eta,t}\>|^2 \, dyd\eta \frac{dt}t \ .$$ Then we have by Cauchy-Schwarz $$\label{lacunarybessel}
A^2 \le \|\sum_{k=1}^n \int_{{T}_k^*} \<f,\phi_{y,\eta,t}\>\phi_{y,\eta,t}
dy\, d\eta \, \frac{dt}{t}
\|_2^2$$ $$=\sum_{k,l=1}^n \int_{{T}_k^*\times {T}_l^*}
\<f,\phi_{y,\eta,t}\>
\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>
\<\phi_{y',\eta',t'},f\>
dy\, d\eta \, \frac{dt}{t} dy'\, d\eta' \, \frac{dt'}{t'}$$ $$=\sum_{k,l=1}^n \int_{{T}_k^*\times {T}_l^*: B^{-1}t\le t'\le B t}
\dots
+2
\sum_{k,l=1}^n \int_{{T}_k^*\times {T}_l^*: B t'\le t}
\dots \ ,$$ where the large number $B=2^{8}\alpha^{-1}b^{-1}$ determines the cutoff in the last line between diagonal and off-diagonal part, the latter being estimated by twice the upper triangular part using symmetry. In the diagonal term we use symmetry to estimate the smaller of the inner products with $f$ by the larger one and obtain the upper bound $$2 \sum_{k,l=1}^n
\int_{{T}_k^*\times {T}_l^*: B^{-1}t \le t'\le B t}
|\<f,\phi_{y,\eta,t}\>|^2
|\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>|
dy\, d\eta \, \frac{dt}{t} dy'\, d\eta' \, \frac{dt'}{t'}$$ $$\le 2A \sup_{k,(y,\eta,t)\in T_k^*}\left(\sum_{l=1}^n
\int_{{T}_l^*: B^{-1}t\le t'\le B t}
|\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>| dy'\, d\eta' \, \frac{dt'}{t'} \right)$$ $$\le 2A \sup_{k,(y,\eta,t)\in T_k^*}\left(
\int_{B^{-1}t\le t'\le B t}\int_{\R^2}
|\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>| dy'\, d\eta' \, \frac{dt'}{t'} \right)\ .$$ Here we have used that the regions $T_l^*$ are pairwise disjoint. Integrating over the $t'$- interval of bounded $dt'/t'$-measure estimates the previous display by $$\le C A \sup_{k,(y,\eta,t)\in T_k^*} \left(\sup_{B^{-1}t\le t'\le tB}
\int_{\R^2}
|\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>| dy'\, d\eta'\right)\ .$$ For the $\eta'$ integration we use that $\widehat{\phi}_{y,\eta,t}$ is supported on an interval of length $t^{-1}$ and $t\sim t'$: $$\le C A \sup_{k,(y,\eta,t)\in T_k^*} \sup_{B^{-1}t\le t'\le tB}
\sup_{\eta'}\, \, t^{-1} \int_{\R}
\<|\phi_{y,\eta,t}|,|\phi_{y',\eta', t'}|\> dy' \ .$$ For the $y'$ integration we use that ${\phi}_{y,\eta,t}$ is an $L^2$ normalized wave packet adapted to an interval of length $t$. This estimates the last display by $CA$.
Turning to the off diagonal term in (\[lacunarybessel\]) we estimate it with Cauchy Schwarz and the upper bound on the selected tents by $$2
\sum_{k=1}^n
\left(
\int_{{T}_k^*}|\<f, \phi_{y,\eta,t}\>|^2\, dyd\eta \frac{dt}t
\right)^{1/2}
H_k^{1/2}
\le C
\sum_{k=1}^n \lambda s_k^{1/2} H_k^{1/2}\ ,$$ where $H_k$ is equal to $$\int_{{T}_k^*}
\left(\sum_{l=1}^n \int_{{T}_l^*: B t'\le t}
|\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>
\<\phi_{y',\eta',t'},f\>| dy'\, d\eta' \, \frac{dt'}{t'}
\right)^{2}
dy\, d\eta \, \frac{dt}{t}\ .$$ Using the pointwise bound on $F(y,\eta,t)$ outside $E$ we can estimate $H_k$ by $$\int_{{T}_k^*}
\left(\sum_{l=1}^n \int_{{T}_l^*: B t'\le t}
c\lambda t'^{1/2} |\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>| dy'\, d\eta' \, \frac{dt'}{t'}
\right)^{2}
dy\, d\eta \, \frac{dt}{t}\ .$$
Let $(y,\eta,t)\in T_k^*$ and $(y',\eta',t')\in T_l^*$ with $B t'\le t$. Assume that the inner product $\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>$ is not zero. Then we have $$\eta'+\gamma' (t')^{-1}=\eta+\gamma t^{-1}$$ for some $$-2^{-8}b \le \gamma,\gamma' \le 2^{-8}b \ .$$ and hence $$|\eta-\eta'|\le 2^{-4}b(t')^{-1}\ .$$ By definition of the reduced domains we have $$\alpha(\eta-\xi_k)+\beta t^{-1} \le t^{-1}\ ,$$ $$b(t')^{-1}\le \eta'-\xi_l\ .$$ This gives $$\xi_k-\xi_l= (\eta-\eta') - (\eta-\xi_k)+ (\eta'-\xi_l)$$ $$\ge
-2^{-4}b (t')^{-1}
- \alpha^{-1} (1-\beta) t^{-1} + b(t')^{-1}
\ .$$ Using $B t'\le t$ the last display strictly larger than $0$ and hence the tent $T_k$ has been chosen prior to $T_l$. Since $(y',\eta',t')$ is in the reduced tent $ T_l^*$, it is not in $ E_l$ and hence not in $T_k$. But $t'< t\le s_k$ and $$|\alpha(\eta'-\xi_k)+\beta(t')^{-1}|$$ $$\le |\alpha(\eta'-\eta)|+
|\alpha(\eta-\xi_k)+\beta(t)^{-1}|+|\beta (t')^{-1}-\beta(t^{-1})|$$ $$\le 2^{-4}b(t')^{-1}+t^{-1}+|\beta t^{-1}|+|\beta(t')^{-1}|\le (t')^{-1}$$ and hence we need to have $$|y'-x_k|\ge s_k-t'\ .$$ This implies $$\label{tentseparation}
|y'-x_k|\ge s_k-t\ .$$
Now pick a further point $(y'',\eta'',t'')\in T_{l'}^*$ with $Bt''\le t$ and nonzero inner product $\<\phi_{y,\eta,t},\phi_{y'',\eta'', t''}\>$. We have again $$|\eta-\eta''|\le 2^{-4}b(t'')^{-1}$$ and $$b(t'')^{-1}\le \eta''-\xi_{l'} \ .$$ Now we assume $Bt''\le t'$. Then we conclude $$|\eta'-\eta''|\le 2^{-2}b(t'')^{-1}$$ and $$\xi_l-\xi_{l'}=(\xi_l-\eta')+(\eta'-\eta'')+(\eta''-\xi_{l'})$$ $$\ge - 2\alpha^{-1}(t')^{-1} - 2^{-2}b(t'')^{-1}+b(t'')^{-1}>0\ .$$ Hence $T_l$ was chosen prior to $T_{l'}$ and in particular $(y'',\eta'',t'')$ is not in $T_l$. But we have $t''<t'\le s_l$ and $$|\alpha(\eta''-\xi_l)+\beta(t'')^{-1}|$$ $$\le |\alpha(\eta''-\eta')|+|\alpha(\eta'-\xi_l)+\beta(t')^{-1}|
+|\beta(t'')^{-1}-\beta(t')^{-1}|$$ $$\le 2^{-2} b (t'')^{-1}
+(t')^{-1}+ \beta(t'')^{-1}+\beta(t')^{-1} \le (t'')^{-1}\ .$$ Since $(y'',\eta'',t'')$ is not in $T_l$ we conclude $$|y''-x_l|>s_l-t''>s_l-t'\ge |y'-x_l|$$ and in particular $y''\neq y'$.
To summarize our finding, fix $(y,\eta,t)\in T_k^*$. Then for fixed $y'\in \R$, the minimal and maximal values of parameters $t'$ with $Bt'\le t$ such that there exists $l$ and $\eta'$ with $(y',\eta',t')\in T_l^*$ and $\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>\neq 0$ are at most a factor $B$ apart. It follows that for every $y'$ there exists an interval $I(y')=[T(y'), B T(y')]$ such that we need $t'\in I(y')$ for such $l,\eta'$ to exist.
Using also (\[tentseparation\]) and disjointness of the reduced domains $T_l^*$ , we may thus estimate $H_k$ by $$C \int_{{T}_k^*}
\left( \int_{|y'-x_x|>s_k-t}
\int_ {I(y')} \int_\R
\lambda t'^{1/2} |\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>| \, d\eta'
\, \frac{dt'}{t'}
\, dy' \right)^{2}
dy\, d\eta \, \frac{dt}{t}\ .$$ $$\le C\int_{{T}_k^*}
\left(
\int_{|y'-x_x|>s_k-t}
\sup_{t'\in I(y')}\int_\R
\lambda t'^{1/2} |\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>|
\, d\eta' \, dy' \right)^{2}
dy\, d\eta \, \frac{dt}{t}\ .$$
Further, by trivial reasoning with the Fourier support of the bump functions, if we fix $y'$ and $t'$ as in this integral, then there is an interval of length $2t'^{-1}$ which must contain $\eta'$ for the inner product $\<\phi_{y,\eta,t},\phi_{y',\eta', t'}\>$ to be nonzero. Using the estimate $$\<|\phi_{y,\eta,t}|,|\phi_{y',\eta', t'}|\> \le C (\frac{t'}t)^{1/2}
(1+\frac{|y'-y|}{t})^{-2}\ ,$$ we obtain for the previous display the upper bound $$C\int_{T_k^*}
\left( \int_{|y'-x_k|>s_k-t}\
\lambda (\frac{1}{t})^{1/2} (1+\frac{|y'-y|}{t})^{-2}
dy' \,
\right)^{2}
dy\, d\eta \, \frac{dt}{t}$$ $$\le C\int_{T_k^*}
\left(
\lambda {t}^{1/2} (1+\frac{s_k-|y-x_k|}{t})^{-1}
\right)^{2}
dy\, d\eta \, \frac{dt}{t}$$ $$\le C\int_0^{s_k} \int_{x_k-s_k}^{x_k+s_k}\int_{\xi_k-2\alpha^{-1}t^{-1}}^{\xi_k+2\alpha^{-1}t^{-1}}
\lambda^2 t(1+\frac{s_k-|y-x_k|}{t})^{-2}
d\eta \, dy\, \frac{dt}{t}$$ $$\le C\lambda^2s_k\ .$$ This completes our estimation of (\[lacunarybessel\]) and we have shown $$\frac 14 A^2\le CA + C\sum_{k=1}^n\lambda^2 s_k\le CA\ ,$$ where in the last inequality we have used the lower bound on the selected tents. Dividing by $A$ proves the desired estimate for $A$ and completes the proof of (\[count.tiles.plus\]) for the newly selected tents.
If the selection of tents stops lacking any further $(x,\xi,s)$ with $$s^{-1}\int_{(T(x,\xi,s)\cap X_\xi^+)\setminus (T^b_{x,\xi,s}\cup E_n)}
|F(y,\eta,t)|^2\, dy\, d\eta\, dt
\ge 2^{-8}\lambda^2\ ,$$ then clearly the converse inequality holds for all $(x,\xi,s)\in X_\Delta$. If the selection of tents does not stop, we collect $T_k$ for all $k\in \N$ and write $E_{(1)}=E\cup_{k=1}^\infty E_k$. Note that $\xi_k$ is a decreasing sequence, and as noted before the possible values of $\xi_k$ are in the discrete lattice $\Z b 2^{-8-k_{\max}}$. If $\xi_k\to -\infty$, then for every $(x,\xi,s)\in X_\Delta$ $$s^{-1}\int_{(T(x,\xi,s)\cap X_\xi^+)\setminus (T^b_{x,\xi,s}\cup E_{(1)})}
|F(y,\eta,t)|^2\, dy\, d\eta\, dt
\le 2^{-8}\lambda^2\ .$$ Namely, assume not, then $\xi> \xi_k$ for some $k$, and this would contradict the choice of $T_k$.
Now assume $\xi_k$ does not tend to $-\infty$, then the sequence stabilizes, that means eventually becomes constant, at some value $\xi_{(1)}$. We shall then choose further tents, and for emphasis we rename the previously selected tents into $ T_k =:T_{(1),k}=T(x_{(1),k},\xi_{(1),k},s_{(1),k})$.
Call a point $(x,\xi,s)\in X_\Delta$ bad if $$s^{-1}\int_{T(x,\xi,s)\cap X_\xi^+)\setminus (T^b(x,\xi,s)\cup E_{(1)})}
|F(y,\eta,t)|^2\, dy\, d\eta\, dt
\ge 2^{-8}\lambda^2\ ,$$ and let $\xi_{\max}$ be the maximal possible value of the second component of a bad point. Note that $\xi_{\max}$ is strictly less than $\xi_{(1)}$. For if not, then $\xi_{\max}=\xi_{(1)}$ by choice of the previously selected tents. Since $s_k\to 0$ by (\[count.tiles.plus\]) we have $s>s_k$ for some $k$ and some bad point $(x,\xi_{\max},s)$. This however contradicts the choice of the tent $T_{(1),k}$. We then choose a bad point $(x_{(2),1},\xi_{(2),1},s_{(2),1})$ such that $\xi_{(2),1}=\xi_{\max}$ and $s_{(2),1}$ is maximal among all such choices. We then iterate this selection process as before, obtaining tents $T_{(2),k}=T(x_{(2),k},\xi_{(2),k},s_{(2),k})$. Our proof of applies verbatim to yield $$(\sum_{k=1}^\infty s_{(1),k})+ (\sum_{k=1}^n s_{(2),k})
\le C\lambda^{-2} \ .$$
We now continue this double recursion in the obvious manner. If at some point the recursion stops, or yields for some fixed $m$ a sequence $\xi_{(m),k}$ tending to $-\infty$, then by the previous discussions we are left with no bad points. If the double iteration does not stop, we obtain a double sequence of tents $T_{(m),k}$ with $$\sum_{m=1}^\infty \sum_{k=1}^\infty s_{(m),k}\le C\lambda^{-2} \ .$$ Moreover, the sequence $\xi_{(m)}$ of stabilizing points decreases to $-\infty$, since they are strict monotone decreasing and in a discrete lattice. We can then observe that there are no bad points outside $\bigcup _{m=1}^\infty E_{(m)}$.
Summarizing, we have found a collection $\Q_+$ of tents such that $$\sum_{T\in \Q_+}^\infty \sigma(T) \le C \lambda^{-2}$$ and if we set $$E_+=E\cup \bigcup_{T\in \Q_+} T\ ,$$ then we have $$s^{-1}\int_{(T(x,\xi,s)\cap X_\xi^+)\setminus T^b(x,\xi,s)}
|F(y,\eta,t)1_{ E_+^c}(y,\eta,t)|^2\, dy\, d\eta\, dt
\le 2^{-8}\lambda^2\ ,$$ for all $(x,\xi,\delta)\in X_\Delta$.
We may repeat the above argument symmetrically to obtain a collection $\Q_-$ of tents such that $$\sum_{T\in \Q_-}^\infty \sigma(T) \le C \lambda^{-2}$$ and if we set $$E_-=E\cup \bigcup_{T\in \Q_-} T\ ,$$ then we have $$s^{-1}\int_{(T(x,\xi,s)\cap X_\xi^-)\setminus T^b(x,\xi,s)}
|F(y,\eta,t)1_{E_-^c}(y,\eta,t)|^2\, dy\, d\eta\, dt
\le 2^{-8}\lambda^2\ ,$$ for all $(x,\xi,\delta)\in X_\Delta$.
Setting finally $\Q=\Q_0\cup \Q_+\cup \Q_-$ we have clearly found the desired collection of tents. This completes the proof of the endpoint $p=2$ of Theorem \[gen.carl.emb\].
The bilinear Hilbert transform {#bhtsection}
==============================
The most immediate application of Theorem \[gen.carl.emb\] is to prove basic estimates for the bilinear Hilbert transform. Another possible application is towards Carleson’s theorem [@carleson] on almost everywhere convergence of Fourier series. However, the latter application requires more work, as Carleson’s operator lacks the symmetry that is exhibited by the bilinear Hilbert transform and therefore needs an additional embedding theorem. Hence we decided to restrict attention to the bilinear Hilbert transform, which suffices to illustrate some key points of time-frequency analysis originating in Carleson’s work on convergence of Fourier series.
Let $\beta=(\beta_1,\beta_2,\beta_3)$ be a vector in $\R^3$ with pairwise distinct entries. For three Schwartz functions $f_1,f_2,f_3$ on the real line we define $$\Lambda_\beta(f_1,f_2,f_3):=p.v. \int_\R [\int_\R [\prod_{j=1}^3 f_j(x-\beta_jt) ]\, dx ]\, \frac{dt} t\ .$$ Note that the inner integral produces a Schwartz function in the variable $t$, to which we apply the tempered distribution $p.v. 1/t$. By a change of variables, scaling $t$ and translating $x$, we may and do restrict attention to vectors $\beta$ which have unit length and are perpendicular to $(1,1,1)$. The resulting one parameter family of trilinear forms is dual to a family of bilinear operators called bilinear Hilbert transforms. To obtain explicit expressions for these bilinear operators, one applies another translation in the $x$ variable to make one of the components, say $\beta_i$ vanish. After interchanging the order of integrals one obtains an explicit pairing of a bilinear operator in $f_j$, $j\neq i$, with the function $f_i$.
Let $\alpha$ be a unit vector perpendicular to $(1,1,1)$ and $\beta$. The vector $\alpha$ is unique up to reflection at the origin, and has only non-zero components by the assumption that $\beta$ has pairwise distinct components. Note also that $|\beta_j|\le 0.9$ for each $j$. For if one component of $\beta_j$ in absolute value exceeds $0.9$, then since $\beta$ is perpendicular to $(1,1,1)$, at least one further component has to exceed $0.45$ in absolute value. But then the vector cannot be a unit vector.
The following a priori estimate for $\Lambda_\beta$ originates in [@lacey-thiele1].
\[bilinearhttheo\] For a unit vector $\beta$ perpendicular to $(1,1,1)$ with pairwise distinct entries, and for $2<p_1,p_2,p_3<\infty$ with $\sum_j \frac 1{p_j} =1$, there is a constant $C$ such that for all Schwartz functions $f_1,f_2,f_3$ we have $$|\Lambda_\beta (f_1,f_2,f_3)|\le C \prod_{j=1}^3 \|f_j\|_{p_j}\ .$$
We give a new proof of this theorem based on Theorem \[gen.carl.emb\] and an outer Hölder inequality. This proof is analoguous to the previously presented proof of boundedness of paraproducts. In our approach, much of the difficulty in proving bounds for the bilinear Hilbert transform has been moved into the proof of the generalized Carleson embedding theorem. What remains to be done is relatively easier and in particular conceptually quite simple. It is the strength of our approach that the main difficulty is packaged into a cleanly separated module; previous approaches do not suggest the formulation of as clean a statement as Theorem \[gen.carl.emb\]. In particular, our proof is the first one to succeed without the passage to a discrete model operator. This avoids a cumbersome setup of choices of the discretization.
One can prove a version of Theorem \[bilinearhttheo\] with a constant independent of $\beta$, see [@grafakos-li], but only at the expense of considerable additional work. One may also extend the range of exponents, see [@lacey-thiele2]. It would be interesting to discuss these results in the context of outer measure theory, but this is beyond the scope of the present paper.
Define for $j=1,2,3$ $$\label{bhtemb}
F_j(y,\eta,t):=\int_\R f_j(x) e^{i\eta (y-x)} t^{-1}\phi(t^{-1}(y-x)) \, dx\ ,$$ where $\phi$ is a real valued Schwartz function such that $\widehat{\phi}$ is nonnegative, non vanishing at the origin, and supported in $[-\epsilon,\epsilon]$ for suitably small $\epsilon$. It will suffice to choose $\epsilon=2^{-16}$.
The estimate of Theorem \[bilinearhttheo\] can be reformulated by means of the functions $F_j$.
\[bhttheorem\] Under the assumptions of Theorem \[bilinearhttheo\] there is a constant $C$ depending only on $\beta$, $p_1,p_2,p_3$ and $\phi$ as above such that $$\label{bhtbound}
\left|\int_0^\infty \int_\R\int_\R
\prod_{j=1}^3 F_j (y,\alpha_j\eta +\beta_jt^{-1}, t) \, d\eta \, dy\, dt\right|
\le C\prod_{j=1}^3 \|f_j\|_{p_j}\ .$$
We postpone the proof of Lemma \[bhttheorem\] and proceed to deduce Theorem \[bilinearhttheo\] from Lemma \[bhttheorem\].
Inserting the definition of $F_j$ and using that $\alpha$ and $ \beta$ are perpendicular to $(1,1,1)$ we obtain for the integral on the left-hand-side of (\[bhtbound\]): $$\int_0^\infty \int_\R\int_\R
t^{-3}\prod_{j=1}^3
[\int_\R f_j(x_j) e^{-i \beta_j t^{-1}x_j} e^{-i \alpha_j\eta x_j}
\phi(t^{-1}(y-x_j)) \, dx_j]
\, d\eta \, dy\, dt\ .$$ Recall that the integral of the Fourier transform of a Schwartz function $\varphi$ in $\R^3$ over the line through the origin spanned by $\alpha$ is proportional to the integral of the Schwartz function itself over the perpendicular hyperplane through the origin spanned by $(1,1,1)$ and $\beta$: $$\label{perpendicularintegrals}
\int_\R \widehat{\varphi}(\eta \alpha)\, d\eta=
c \int_\R \int_\R \varphi(u(1,1,1)+v \beta)\, du \, dv\ .$$ To apply this fact, we observe that the inner triple integral of the previous display over $x_1,x_2,x_3$ is the value of the Fourier transform of a certain Schwartz function in $\R^3$ at the point $\eta\alpha \in \R^3$, and the integral in $\eta$ is then the integral of this Fourier transformation over the line spanned by $\alpha$. Hence we obtain up to a nonzero constant factor for that display: $$\int_0^\infty \int_\R
\int_\R\int_\R[t^{-3} e^{- i t^{-1} v} \prod_{j=1}^3
f_j(u+\beta_jv) \phi(t^{-1}(y-u-\beta_j v))] \, du\, dv
\, dy\, dt\ .$$ Here we have used again in the argument of the exponential function that $\alpha$, $\beta$ and $(1,1,1)$ are pairwise orthogonal and that $\beta$ has unit length. Changing the order of integration so that the $y$ integration becomes innermost we obtain for the last display $$\int_{0}^\infty \int_\R \int_\R [\prod_{j=1}^3 f_j(u+\beta_j v)]
t^{-2}e^{-it^{-1} v}
\psi(t^{-1}v)\, du \, dv\, {dt}\ ,$$ where $$\psi (w):= \int_\R
\prod_{j=1}^3 \phi(z-\beta_j w)\, dz\ .$$ We claim that there are nonzero constants $a$ and $b$ such that for any Schwartz function $g$ on the real line we have $$\label{lincombclaim}
\int_{0}^\infty \int_\R g(v)t^{-2}e^{-it^{-1} v} \psi(t^{-1}v)\, dv \, dt=
a g(0)+b\, p.v. \int g(t)\frac{dt}t\ .$$ This claim turns the left hand side of (\[bhtbound\]) into a nontrivial linear combination of $$\int_\R [\prod_{j=1}^3 f_j(u)]\, du$$ and $$p.v. \int[\int f_1(u-\beta_1 t) f_2(u-\beta_2 t) f_3(u-\beta_3 t) \, du ] \frac {dt} t\ .$$ Since $L^p$ bounds for the former follow by Hölder’s inequality, we can deduce $L^p$ bounds for the latter from $L^p$ bounds as in (\[bhtbound\]). This will complete the reduction of Theorem \[bilinearhttheo\] to Lemma \[bhttheorem\], once we have verified the above claim.
To see the claim, it suffices to verify that the left-hand-side of (\[lincombclaim\]) can be written as a nonzero multiple of $$\int_{-\infty}^0 \widehat{g}(\zeta)\, d\zeta\ ,$$ since the characteristic function of the left half line is known to be a nontrivial linear combination of the Fourier transform of the Dirac delta distribution and the principal value integral against $dt/t$. Using Plancherel we identify the left-hand-side of (\[lincombclaim\]) as nonzero multiple of $$\int_{0}^\infty \int_\R \widehat{g}(\zeta) \widehat{\psi}( 1-t\zeta )\, d\zeta \, \frac{dt}t\ .$$
The claim will thus follow by Fubini if we can establish that $\widehat{\psi}$ is proportional to a function that is nonnegative, nonzero at $0$, and supported in $[-1/2, 1/2]$. We have $$\widehat{\psi}(\eta)= \int_\R\int_\R
\prod_{j=1}^3 \phi(z-\beta_j w) e^{i \beta_j \eta (z-\beta_j w)} \, dz\, dw .$$ This is an integral of a Schwartz function in $\R^3$ over the plane spanned by $(1,1,1)$ and $\beta$, which by the observation (\[perpendicularintegrals\]) again may be written as multiple of the integral of the Fourier transform of the Schwartz function over the line spanned by $\alpha$: $$\int_\R \prod_{j=1}^3 \widehat{\phi}(\alpha_j \xi-\beta_j \eta))\, d\xi\ \ .$$ Since $\alpha$ and $\beta$ are perpendicular unit vectors and the support of $\widehat{\phi}\otimes \widehat{\phi} \otimes \widehat{\phi}$ is in a neighborhood of $0$ with diameter less than $1/10$, this integral is non-zero only if $|\eta|$ is smaller than $1/2$. Moreover, $\widehat{\psi}$ is evidently nonnegative real and nonzero at $0$. This completes the proof of the claim and the reduction of Theorem \[bilinearhttheo\] to Lemma \[bhttheorem\].
We consider the space $X=\R\times \R\times (0,\infty)$ and the outer measure generated by the collection $\E$ of all tents $$T(x,\xi,s):=\{(y,\eta,t)\in X: t<s, |y-x|<s-t, |\eta-\xi|\le t^{-1}\}$$ parameterized by $(x,\xi,s)\in X$ and the premeasure $\sigma(T(x,\xi,s))=s$.
Define a size $S$ by setting $$S(G)(T(x,\xi,s))=s^{-1}\int_{T(x,\xi,s)} |G(y,\eta,t)|\, dy\, d\eta\, dt$$ for each $G\in \B(X)$.
By a straight forward application of Proposition \[measuredomination\] we may estimate the left-hand-side of (\[bhtbound\]) by $$C\|G_1G_2G_3\|_{L^1(X,\sigma,S)}\ ,$$ where we have defined $G_j$ for $j=1,2,3$ by $$G_j(y,\eta,t):=F_j(y,\alpha_j \eta+\beta_j t^{-1}, t)\ .$$
We intend to apply a threefold Hölder’s inequality, which requires us to define three appropriate sizes $S_j$. Set $$b=2^{-8}\min_{i\neq j}|\beta_i-\beta_j|\ .$$ Since no two components of $\beta$ are equal, we have $b>0$. Define for each $1\le j\le 3$ and $(x,\xi,s)\in X$ the region $$T^{(j)}(x,\xi,s)$$ $$:=\{(y,\eta,t)\in X: t \le s, |y-x| \le s-t, |\alpha_j^{-1}(\eta-\xi) - \alpha_j^{-1}\beta_jt^{-1}|\le b t^{-1}\}\ .$$
For fixed $(x,\xi,s)$ the three regions $T^{(j)}(x,\xi,s)$ are pairwise disjoint, by symmetry it suffices to establish this for $j=1,2$. Assume to get a contradiction that we have $\eta,t$ with $$|\alpha_1^{-1}(\eta-\xi) - \alpha_1^{-1}\beta_1 t^{-1}|,\ |\alpha_2^{-1}(\eta-\xi) - \alpha_2^{-1}\beta_2 t^{-1}|\le
b t^{-1}\ .$$ Multiplying by $|\alpha_1|,|\alpha_2|\le 1$ respectively and comparing yields $|\beta_1-\beta_2|\le 2b$. This however is a contradiction to the choice of $b$ and thus proves that the regions $T^{(j)}(x,\xi,s)$ are pairwise disjoint.
We now observe for each $T=T(x,\xi,s)$ with similar notation $T^{(j)}= T^{(j)}(x,\xi,s)$ $$s S(G)(T)
= \int_{T} |G(y,\eta,t)|\, dy\, d\eta\, dt$$ $$=
\int_{T\setminus(T^{(1)}\cup T^{(2)}\cup T^{(3)})}
|G(y,\eta,t)|\, dy\, d\eta\, dt
+\sum_{j=1}^3 \int_{T\cap T^{(j)}}|G(y,\eta,t)|\, dy\, d\eta\, dt$$ $$\le \prod_{j=1}^3\left(
\int_{T\setminus T^{(j)}}
|G_j(y,\eta, t)|^3\, dy\, d\eta\, dt \right)^{1/3}$$ $$+\sum_{j=1}^3
\sup_{( y,\eta,t )\in T^{(j)}}|G_j(y,\eta,t)|
\prod_{k\neq j}
\left(\int_{T\setminus T^{(k)}}|G_k(y,\eta,t)|^2\, dy\, d\eta\, dt \right)^{1/2}\ .$$ Define the size $$S_j(G)(T):=(
s^{-1}\int_{T\setminus T^{(j)} }|G(y,\eta,t)|^2 \, dy\,
d\eta \,{dt})^{1/2}
+ \sup_{(y,\eta,t)\in T}|G(y,\eta,t)|\ .$$ Then we conclude from the previous considerations that $$S(G)(T)
\le 4 \prod_{k=1}^3 S_k(G_k)(T)\ ,$$ where we have with log convexity estimated $L^3$ norms by $L^2$ and $L^\infty$ norms.
By the outer Hölder inequality, Proposition \[p.hoelder-energy\], we obtain for the left-hand-side of (\[bhtbound\]) the bound $$C \prod_{j=1}^3 \|G_j\|_{L^{p_j}(X,\sigma, S_j)}$$ with exponents $p_j$ as in Lemma \[bhttheorem\]. It remains to show for each $j$ that $$\|G_j\|_{L^{p_j}(X,\sigma,S_j)}\le C \|f_j\|_{p_j}\ .$$ This follows from the generalized Carleson embedding, Theorem \[gen.carl.emb\], after a re-parametrization of the space $X$ under the homeomorphism $$\Phi_j:X\to X,\ (y,\eta,t)\mapsto (y,\alpha_j \eta+\beta_j t^{-1}, t)\ .$$ Note that $\Phi_j$ maps $T_{\alpha_j,\beta_j}(x,\alpha_j^{-1} \xi,s)$ as defined in to $T(x,\xi,s)$ as above, and it maps $T^b(x,\alpha_j^{-1}\xi,s)$ to $T^{(j)}(x,\xi,s)$ as above and we have $F_j\circ\Phi_j=G_j$. This completes the proof of Lemma \[bhttheorem\].
[plain]{}
D. Adams, *Choquet integrals in potential theory*. Publ. Mat. 42 (1998), no. 1, 3–66.
L. Carleson, *Interpolations by bounded analytic functions and the corona problem*. Ann. of Math. (2) 76 (1962): 547-559. (1962)
L. Carleson, *On convergence and growth of partial sums of Fourier series*, Acta Math. 116 (1966) 135–157.
R. R. Coifman, Y. Meyer, E. M. Stein, *Some new function spaces and their applications to harmonic analysis*, J. Funct. Anal. 62 (1985), no. 2, 304–335.
G. David, J.-L. Journé, *A boundedness criterion for generalized Calder???n-Zygmund operators*, Ann. of Math. (2) 120 (1984), no. 2, 371–397.
I. Daubechies, *Ten lectures on wavelets* CBMS-NSF Regional Conference Series in Applied Mathematics 61, SIAM, Philadelphia, PA, (1992).
C. Demeter, C. Thiele, *On the two-dimensional bilinear Hilbert transform*, Amer. J. Math. 132 (2010), no. 1, 201–256.
Y. Do, C. Muscalu, C. Thiele, *Variational estimates for paraproducts*, Rev. Mat. Iberoamericana (2012), vol 28, no. 3, pp. 859–878.
L. Grafakos, X. Li, *Uniform bounds for the bilinear Hilbert transforms. I.*, Ann. of Math. (2) 159 (2004), no. 3, 889-933.
M. Lacey, C. Thiele, *$L^p$ estimates on the bilinear Hilbert transform*, Proc. Nat. Acad. Sci. U.S.A. 94. no. 1, (1997), 33–35
M. Lacey, C. Thiele, *On Calderón’s conjecture for the bilinear Hilbert transform*, Proc. Natl. Acad. Sci. USA 95 (1998), no. 9, 4828–4830 .
M. Lacey, C. Thiele, *A proof of boundedness of the Carleson operator*, Math. Res. Lett 7 (2000) 361–370 .
A. Lerner, *On an estimate of Calderón-Zygmund operators by dyadic positive operators*, J. Anal. Math. 121, (2013), 141–161 .
C. Muscalu, T. Tao, C. Thiele, *[$L^p$]{} estimates for the biest. [II]{}. [T]{}he [F]{}ourier case*, Math. Ann. 329 (2004), no. 3, 427–461.
C. Muscalu, T. Tao, C. Thiele, *Uniform estimates on multi-linear operators with modulation symmetry*,
J. Anal. Math. 88 (2002), 255–309.
R. Oberlin, A. Seeger, T. Tao, C. Thiele, J. Wright, *A variation norm Carleson theorem*, J. Eur. Math. Soc. (JEMS) 14 (2012), no. 2, 421–464.
E. Stein, *Harmonic Analysis: real-variable methods, orthogonality, and oscillatory integrals*, Princeton Mathematical Series 43, With the assistance of Timothy S. Murphy; Princeton University Press, Princeton, NJ, (1993).
C. Thiele, *The quartile operator and pointwise convergence of Walsh series.* Trans. Amer. Math. Soc. 352 (2000), no. 12, 5745–5766.
C. Thiele, *A uniform estimate*, Ann. of Math. (2) 156 (2002), no. 2, 519–563.
R. Wheeden, A. Zygmund, *Measure and integral* An introduction to real analysis. Pure and Applied Mathematics, Vol. 43. Marcel Dekker, Inc., New York-Basel, 1977.
[^1]: For details see the special case $p=\infty$ of
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, a fast multipole method (FMM) is proposed for 3-D Laplace equation in layered media. The potential due to charges embedded in layered media is decomposed into a free space component and four types of reaction field components, and the latter can be associated with the potential of a polarization source defined for each type. New multipole expansions (MEs) and local expansions (LEs), as well as the multipole to local (M2L) translation operators are derived for the reaction components, based on which the FMMs for reaction components are then proposed. The resulting FMM for charge interactions in layered media is a combination of using the classic FMM for the free space components and the new FMMs for the reaction field components. With the help of a recurrence formula for the run-time computation of the Sommerfeld-type integrals used in M2L translation operators, pre-computations of a large number of tables are avoided. The new FMMs for the reaction components are found to be much faster than the classic FMM for the free space components due to the separation of equivalent polarization charges and the associated target charges by a material interface. As a result, the FMM for potential in layered media costs almost the same as the classic FMM in the free space case. Numerical results validate the fast convergence of the MEs for the reaction components, and the $O(N)$ complexity of the FMM [with a given truncation number $p$]{} for charge interactions in 3-D layered media.'
address:
- 'LCSM(MOE), School of Mathematics and Statistics, Hunan Normal University, Changsha, Hunan, 410081, P. R. China.'
- 'Department of Mathematics, Southern Methodist University, Dallas, TX 75275, USA.'
author:
- Bo Wang
- Wen Zhong Zhang
- Wei Cai
title: 'Fast multipole method for 3-D Laplace equation in layered media'
---
Fast multipole method, layered media, Laplace equation, spherical harmonic expansion
Introduction
============
Solving the Laplace equation in layered media is connected to many important applications in science and engineering. For instance, finding the electric charge distribution over conductors embedded in a layered dielectric medium has important application in semi-conductor industry, especially in calculating the capacitance of interconnects (ICs) in very large-scale integrated (VLSI) circuits for microchip designs (cf. [@yu2014advanced; @seidl1988capcal-a; @ruehli1973efficient; @oh1994capacitance]). Due to complex geometric structure of the ICs, the charge potential solution to the Laplace equation is usually solved by an integral method with the Green’s function of the layered media (cf. [@oh1994capacitance; @zhao1998efficient]), which results in a huge dense linear algebraic system to be solved by an iterative method such as GMRES (cf. [@campbell1996gmres]), etc. Other applications of the Laplace equation can be found in medical imaging of brains (cf. [@xu2005]), elasticity of composite materials (cf. [@babu1994]), and electrical impedance tomography for geophysical applications (cf. [@borcea2002]).
Due to the full matrix resulted from the discretization of integral equations, it will incur an $O(N^{2})$ computational cost for computing the product of the matrix with a vector (a basic operation for the GMRES iterative solver). The fast multipole method (FMM) for the free space Green’s function (the Coulomb potential) has been used in the development of FastCap (cf. [@nabors1991fastcap]) to accelerate this product to $O(N)$. However, the original FMM of Greengard and Rokhlin (cf. [@greengard1987fast; @grengard1988rapid]) is only designed for the free space Green’s function. To treat the dielectric material interfaces in the IC design, unknowns representing the polarization charges from the dielectric inhomogeneities have to be introduced over the infinite material interfaces, thus creating unnecessary unknowns and contributing to larger linear systems. These extra unknowns over material interfaces can be avoided by using the Green’s function of the layered media in the formulation of the integral equations. To find fast algorithm to solve the discretized linear system, image charges are used to approximate the Green’s function of the layered media [@chow1991closed; @aksun1996robust; @alparslan2010closed], converting the reaction potential to the free space Coulomb potential from the charges and their images, thus, the free space FMM can be used [@jandhyala1995multipole; @gurel1996electromagnetic; @geng2001fast]. Apparently, this approach is limited to the ability of finding image charge approximation for the layered media Green’s function. Unfortunately, finding such an image approximation can be challenging if not impossible when many layers are present in the problem.
In this paper, we will first derive the multipole expansions (MEs) and local expansions (LEs) for the reaction components of the layered media Green’s function of the Laplace equation. Then, the original FMM for the interactions of charges in free space can be extended to those of charges embedded in layered media. The approach closely follows our recent work for the Helmholtz equation in layered media (cf. [@wang2019fast; @zhang2018exponential]), where the generating function of the Bessel function (2-D case) or a Funk-Hecke formula (3-D case) were used to connect Bessel functions and plane wave functions. The reason of using Fourier (2-D case) and spherical harmonic (3-D case) expansions of plane waves is that the Green’s function of layered media has a Sommerfeld-type integral representation involving the plane waves. Even though, the Laplace equation could be considered as a zero limit of the wave number $k$ in the Helmholtz equation, some special treatments of the $k\rightarrow0$ limit is required to derive a limit version of the extended Funk-Hecke formula, which is the key in the derivation of MEs, LEs and M2L for the reaction components of the Laplacian Green’s function in layered media. Similar to our previous work for the Helmholtz equation in layered media, the potential due to sources embedded in layered media is decomposed into free space and reaction components and equivalent polarization charges are introduced to re-express the reaction components. The FMM in layered media will then consist of classic FMM for the free space components and FMMs for reaction components, using equivalent polarization sources and the new MEs, LEs and M2L translations. Moreover, in order to avoid making pre-computed tables (cf. [@wang2019fast]), we introduce a recurrence formula for efficient computation of the Sommerfeld-type integrals used in M2L translation operators. As in the Helmholtz equation case, the FMMs for the reaction field components are much faster than that for the free space components due to the fact that the introduced equivalent polarization charges are always separated from the associated target charges by a material interface. As a result, the new FMM for charges in layered media costs almost the same as the classic FMM for the free space case.
The rest of the paper is organized as follows. In section 2, we will consider the limit case of the extended Funk-Hecke formula introduced in [@wang2019fast], which leads to an spherical harmonic expansion of the exponential kernel in the Sommerfeld-type integral representation of the Green’s function. By using this expansion, we present alternative derivation, via the Fourier spectral domain, for the ME, LE and M2L operators of the free space Green’s function. The same approach will be then used to derive MEs, LEs and M2L translation operators for the reaction components of the layered Green’s function. In Section 3, after a short discussion on the Green’s function in layered media consisting of free space and reaction components, we present the formulas for the potential induced by sources embedded in layered media. Then, the concept of equivalent polarization charge of a source charge is introduced for each type of the reaction components. The reaction components of the layered Green’s function and the potential are then re-expressed by using the equivalent polarization charges. Further, we derive the MEs, LEs and M2L translation operators for the reaction components based on the new expressions using equivalent polarization charges. Combining the original source charges and the equivalent polarization charges associated to each reaction component, the FMMs for reaction components can be implemented. Section 4 will give numerical results to show the spectral accuracy and $O(N)$ complexity of the proposed FMM for charge interactions in layered media. Finally, a conclusion is given in Section 5.
A new derivation for the ME, LE, and M2L operator of the Green’s function of 3-D Laplace equation in free space
===============================================================================================================
In this section, we first review the multipole and local expansions of the free space Green’s function of the Laplace equation and the corresponding shifting and translation operators. They are the key formulas in the classic FMM and can be derived by using the addition theorems for Legendre polynomials. Then, we present a new derivation for them by using the Sommerfeld-type integral representation of the Green’s function. The key expansion formula used in the new derivation is a limiting case of the extended Funk-Hecke formula introduced in [@wang2019fast]. This new technique shall be applied to derive MEs and LEs for the reaction components of the layered media Green’s function later on.
The multipole and local expansions of free space Green’s function
-----------------------------------------------------------------
Let us review some addition theorems (cf. [@grengard1988rapid; @epton1995multipole]), which have been used for the derivation of the ME, LE and corresponding shifting and translation operators of the free space Green’s function. In this paper, we adopt the definition $$\label{sphericalharmonics}
Y_n^m(\theta,\varphi)=(-1)^m\sqrt{\frac{2n+1}{4\pi}\frac{(n-m)!}{(n+m)!}}P_n^{m}(\cos\theta)e^{\ri m\varphi}:=\widehat P_n^{m}(\cos\theta)e^{\ri m\varphi}$$ for the spherical harmonics where $P_n^m(x)$ (resp. $\widehat P_n^m(x)$) is the associated (resp. normalized) Legendre function of degree $n$ and order $m$. Recall that $$P_n^m(x)=(-1)^m(1-x^2)^{\frac{m}{2}}\frac{d^m}{dx^m}P_n(x)$$ for integer order $0\leq m\leq n$ and $$P_n^{-m}=(-1)^m\frac{(n-m)!}{(n+m)!}P_n^m(x), \quad {\rm so}\quad \widehat P_n^{-m}(x)=(-1)^m\widehat P_n^m(x)$$ for $0<m\leq n$, where $P_n(x)$ is the Legendre polynomial of degree $n$. The so-defined spherical harmonics constitute a complete orthogonal basis of $L(\mathbb S^2)$ (where $\mathbb S^2$ is the unit spherical surface) and $$\langle Y_n^m, Y_{n'}^{m'}\rangle=\delta_{nn'}\delta_{mm'},\quad Y_n^{-m}(\theta,\varphi)=(-1)^m\overline{Y_n^m(\theta,\varphi)}.$$ It is worthy to point out that the spherical harmonics with different scaling constant defined as $$\widetilde Y_n^m(\theta,\varphi)=\sqrt{\frac{(n-|m|)!}{(n+|m|)!}}P_n^{|m|}(\cos\theta)e^{\ri m\varphi}=\ri^{m+|m|}\sqrt{\frac{4\pi}{2n+1}}Y_n^m(\theta,\varphi),\;\;$$ have been frequently adopted in published FMM papers (e.g., [@greengard1997new; @grengard1988rapid]). By using the spherical harmonics defined in , we will re-present the addition theorems derived in [@grengard1988rapid; @epton1995multipole]. For this purpose, we define constants $$c_n=\sqrt{\frac{2n+1}{4\pi}},\quad A_n^m=\frac{(-1)^nc_n}{\sqrt{(n-m)!(n+m)!}},\quad |m|\leq n.$$
\[addthmleg\] [**(Addition theorem for Legendre polynomials)**]{} Let $P$ and $Q$ be points with spherical coordinates $(r,\theta,\varphi)$ and $(\rho,\alpha,\beta)$, respectively, and let $\gamma$ be the angle subtended between them. Then $$P_n(\cos\gamma)=\frac{4\pi}{2n+1}\sum\limits_{m=-n}^n\overline{Y_n^{m}(\alpha,\beta)}Y_n^m(\theta,\varphi).$$
\[theorem:firstaddition\] Let $Q=(\rho,\alpha,\beta)$ be the center of expansion of an arbitrary spherical harmonic of negative degree. Let the point $P=(r,\theta,\varphi)$, with $r>\rho$, and $P-Q=(r', \theta', \varphi')$. Then $$\frac{Y_{n'}^{m'}(\theta', \varphi')}{r'^{n'+1}}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^n\frac{(-1)^{|m+m'|-|m'|}A_n^mA_{n'}^{m'}\rho^nY_n^{-m}(\alpha,\beta)}{c_n^2A_{n+n'}^{m+m'}}\frac{Y_{n+n'}^{m+m'}(\theta,\varphi)}{r^{n+n'+1}}.$$
\[theorem:secondaddition\] Let $Q=(\rho,\alpha,\beta)$ be the center of expansion of an arbitrary spherical harmonic of negative degree. Let the point $P=(r,\theta,\varphi)$, with $r<\rho$, and $P-Q=(r', \theta', \varphi')$. Then $$\frac{Y_{n'}^{m'}(\theta', \varphi')}{r'^{n'+1}}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^n\frac{(-1)^{n'+|m|}A_n^mA_{n'}^{m'}\cdot Y_{n+n'}^{m'-m}(\alpha,\beta)}{c_n^2A_{n+n'}^{m'-m}\rho^{n+n'+1}}r^nY_n^{m}(\theta,\varphi).$$
\[theorem:fourthaddition\] Let $Q=(\rho,\alpha,\beta)$ be the center of expansion of an arbitrary spherical harmonic of negative degree. Let the point $P=(r,\theta,\varphi)$ and $P-Q=(r', \theta', \varphi')$. Then $$r'^{n'}Y_{n'}^{m'}(\theta', \varphi')=\sum\limits_{n=0}^{n'}\sum\limits_{m=-n}^n\frac{(-1)^{n-|m|+|m'|-|m'-m|}c_{n'}^2 A_n^mA_{n'-n}^{m'-m}\cdot \rho^nY_n^{m}(\alpha,\beta) }{c_{n}^2c_{n'-n}^2A_{n'}^{m'}r^{n-n'}}Y_{n'-n}^{m'-m}(\theta,\varphi),$$ where $A_n^m=0$, $Y_n^m(\theta,\varphi)\equiv 0$ for $|m|>n$ is used.
Denote by $(r,\theta, \varphi)$ and $(r',\theta',\varphi')$ the spherical coordinates of given points ${\boldsymbol{r}}, {\boldsymbol{r}}'\in\mathbb R^3$. The law of cosines gives $$\label{cosineslaw}
|{\boldsymbol{r}}-{\boldsymbol{r}}'|^2=r^2+(r')^2-2r r'\cos\gamma,$$ where $$\cos\gamma=\cos\theta\cos\theta'+\sin\theta\sin\theta'\cos(\varphi-\varphi').$$ Then, the Green’s function of the Laplace equation in free space is given by $$G({\boldsymbol{r}}, {\boldsymbol{r}}')=\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\frac{1}{r\sqrt{1-2\mu\cos\gamma+\mu^2}}=\frac{1}{r'\sqrt{1-2\frac{\cos\gamma}{\mu}+\frac{1}{\mu^2}}},$$ where $\mu=r'/r$ and the scaling constant $1/4\pi$ has been omitted through out this paper. Furthermore, we have the following Taylor expansions $$\label{legendreexp}
\frac{1}{r\sqrt{1-2\mu\cos\gamma+\mu^2}}=\sum\limits_{n=0}^{\infty}P_n(\cos\gamma)\frac{\mu^n}{r}=\sum\limits_{n=0}^{\infty}P_n(\cos\gamma)\frac{r'^n}{r^{n+1}}, \;\; \mu=\frac{r'}{r}<1,$$ and $$\label{legendreexp2}
\frac{1}{r'\sqrt{1-2\frac{\cos\gamma}{\mu}+\frac{1}{\mu^2}}}=\sum\limits_{n=0}^{\infty}P_n(\cos\gamma)\frac{1}{r'\mu^n}=\sum\limits_{n=0}^{\infty}P_n(\cos\gamma)\frac{r^{n}}{r'^{n+1}},\;\; \mu=\frac{r'}{r}>1.$$ Straightforwardly, we have error estimates $$\label{meerror}
\left|\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}-\sum\limits_{n=0}^{p}\frac{P_n(\cos\gamma_j)(r')^n}{r^{n+1}}\right|\leq \frac{1}{r-r'}\Big(\frac{r'}{r}\Big)^{p+1}, \quad r>r',$$ and $$\label{leerror}
\left|\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}-\sum\limits_{n=0}^{\infty}P_n(\cos\gamma_j)\frac{r^{n}}{(r')^{n+1}}\right|\leq\frac{1}{r'-r}\Big(\frac{r}{r'}\Big)^{p+1}, \quad r>r',$$ by using the fact $|P_n(x)|\leq 1$ for all $x\in[-1, 1]$.
\[3dspherical\] 
Based on the discussion above, we are ready to present ME, LE and corresponding shifting and translation operators of the free space Green’s function. Let ${\boldsymbol{r}}_c^s$ and ${\boldsymbol{r}}_c^t$ be source and target centers close to source ${\boldsymbol{r}}'$ and target ${\boldsymbol{r}}$, i.e, $|{\boldsymbol{r}}'-{\boldsymbol{r}}_c^s|<|{\boldsymbol{r}}-{\boldsymbol{r}}_c^s|$ and $|{\boldsymbol{r}}'-{\boldsymbol{r}}_c^t|>|{\boldsymbol{r}}-{\boldsymbol{r}}_c^t|$. Following the derivation in - we have Taylor expansions $$\label{expansionbeforeme}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\frac{1}{|({\boldsymbol{r}}-{\boldsymbol{r}}_c^s)-({\boldsymbol{r}}'-{\boldsymbol{r}}_c^s)|}=\sum\limits_{n=0}^{\infty}\frac{P_n(\cos\gamma_s)}{ r_s}\Big(\frac{r'_s}{ r_s}\Big)^n,$$ and $$\label{localexpansionbeforeme}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\frac{1}{|({\boldsymbol{r}}-{\boldsymbol{r}}_c^t)-({\boldsymbol{r}}'-{\boldsymbol{r}}_c^t)|}=\sum\limits_{n=0}^{\infty}\frac{P_n(\cos\gamma_t)}{r_t'}\Big(\frac{r_t}{r_t'}\Big)^n,$$ where $(r_s, \theta_s,\varphi_s)$, $(r_t,\theta_t,\varphi_t)$ are the spherical coordinates of ${\boldsymbol{r}}-{\boldsymbol{r}}_c^s$ and ${\boldsymbol{r}}-{\boldsymbol{r}}_c^t$, $(r'_s, \theta'_s,\varphi'_s)$, $(r'_t,\theta'_t,\varphi'_t)$ are the spherical coordinates of ${\boldsymbol{r}}'-{\boldsymbol{r}}_c^s$ and ${\boldsymbol{r}}'-{\boldsymbol{r}}_c^t$( see Fig. \[3dspherical\]) and $$\begin{split}
&\cos\gamma_s=\cos\theta_s\cos\theta'_s+\sin\theta_s\sin\theta'_s\cos(\varphi_s-\varphi'_s),\\
&\cos\gamma_t=\cos\theta_t\cos\theta_t'+\sin\theta_t\sin\varphi_t'\cos(\varphi_t-\varphi_t').
\end{split}$$ Note that $P_n(\cos\gamma_s)$, $P_n(\cos\gamma_t)$ still mix the source and target information (${\boldsymbol{r}}$ and ${\boldsymbol{r}}'$). Applying Legendre addition theorem \[addthmleg\] to expansions and gives a ME $$\label{mefreespace}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^nM_{nm}r_s^{-n-1}Y_n^m(\theta_s,\varphi_s),$$ and a LE $$\label{lefreespace}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^nL_{nm}r^n_tY_n^m(\theta_t,\varphi_t),$$ where $$\label{freespaceexpcoef}
M_{nm}=c_n^{-2}r'^n_s\overline{Y_n^{m}(\theta'_s,\varphi'_s)},\quad L_{nm}=c_n^{-2}r'^{-n-1}_t\overline{Y_n^{m}(\theta'_t,\varphi'_t)}.$$
The FMM also need shifting and translation operators between expansions. Applying the addition Theorem \[theorem:secondaddition\] to expansion functions in ME provides a translation from ME to LE as follows $$\label{metole}
L_{nm}=\sum\limits_{n'=0}^{\infty}\sum\limits_{m'=-n'}^{n'}\frac{(-1)^{n'+|m|}A_{n'}^{m'}A_n^mY_{n+n'}^{m'-m}(\theta_{st}, \varphi_{st})}{c_{n'}^2A_{n+n'}^{m'-m}r_{st}^{n+n'+1}}M_{n'm'},$$ where $(r_{st}, \theta_{st}, \varphi_{st})$ is the spherical coordinate of ${\boldsymbol{r}}_c^s-{\boldsymbol{r}}_c^t$. Similarly, the following center shifting operators for ME and LE, $$\begin{aligned}
\displaystyle\tilde{M}_{nm}
=\sum\limits_{n'=0}^{n}\sum\limits_{m'=-n'}^{n'}\frac{(-1)^{|m|-|m-m'|}A_{n'}^{m'}A_{n-n'}^{m-m'}r_{ss}^{n'}Y_{n'}^{-m'}(\theta_{ss}, \varphi_{ss})}{c_{n'}^2A_{n}^{m}}M_{n-n',m-m'},\label{metome}\\
\displaystyle\tilde L_{nm}=\sum\limits_{n'=n}^{\infty}\sum\limits_{m'=-n'}^{n'} \frac{(-1)^{n'-n-|m'-m|+|m'|-|m|}c_{n'}^2A_{n'-n}^{m'-m}A_n^mr_{tt}^{n'-n}Y_{n'-n}^{m'-m}(\theta_{tt}, \varphi_{tt})}{c_{n'-n}^2c_n^2A_{n'}^{m'}}L_{n'm'},\label{letole}\end{aligned}$$ can be derived by using addition Theorem \[theorem:firstaddition\] and \[theorem:fourthaddition\]. Here, $(r_{ss}, \theta_{ss}, \varphi_{ss})$ and $(r_{tt},\theta_{tt}, \varphi_{tt})$ are the spherical coordinates of ${\boldsymbol{r}}_c^s-\tilde{{\boldsymbol{r}}}_c^s$ and ${\boldsymbol{r}}_c^t-\tilde{{\boldsymbol{r}}}_c^t$, $$\tilde{M}_{nm}=c_n^{-2}\tilde r'^n_s\overline{Y_{n}^{m}(\tilde\theta'_s,\tilde\varphi'_s)},\quad \tilde L_{nm}=c_n^{-2}\tilde r'^{-n-1}_t\overline{Y_{n}^{m}(\tilde\theta'_t,\tilde\varphi'_t)},$$ are the ME and LE coefficients with respect to new centers $\tilde{{\boldsymbol{r}}}_c^s$ and $\tilde{{\boldsymbol{r}}}_c^{t}$, respectively.
A very important fact in the expansions - is that the source and target coordinates are separated. It is one of the key features for the compression in the FMM (cf. [@greengard1987fast; @greengard1997new]). Besides using the addition theorems, this target/source separation can also be achieved in the Fourier spectral domain. We shall give a new derivation for and by using the integral representation of $1/|{\boldsymbol{r}}-{\boldsymbol{r}}'|$. More importantly, this methodology can be further applied to derive multipole and local expansions for the reaction components of the Green’s function in layered media to be discussed in section 3.
A new derivation of the multipole and local expansions
------------------------------------------------------
For the Green’s function $G({\boldsymbol{r}}, {\boldsymbol{r}}')$, we have the well known identity $$\label{freegreenintegral}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}
=\frac{1}{2\pi}\int_{0}^{\infty}\int_{0}^{2\pi}e^{\ri k_{\rho}((x-x')\cos\alpha+(y-y')\sin\alpha)-k_{\rho}|z-z'|}d\alpha dk_{\rho}.$$ By this identity, we straightforwardly have source/target separation in spectral domain as follows $$\label{positivecase}
\begin{split}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\frac{1}{2\pi}\int_{0}^{\infty}\int_{0}^{2\pi}e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot ({\boldsymbol{r}}-{\boldsymbol{r}}_c^s)}e^{-\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}'-{\boldsymbol{r}}_c^s)}d\alpha dk_{\rho},\\
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\frac{1}{2\pi}\int_{0}^{\infty}\int_{0}^{2\pi}e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot ({\boldsymbol{r}}-{\boldsymbol{r}}_c^t)}e^{-\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}'-{\boldsymbol{r}}_c^t)}d\alpha dk_{\rho},
\end{split}$$ for $z\geq z'$ where $${\boldsymbol{k}}_0=(\cos\alpha, \sin\alpha, \ri),
\label{k0}$$ and without loss of generality, here we only consider the case $z\geq z'$ as an example.
A FMM for the Helmholtz kernel in layered media has been proposed in [@wang2019fast] based on a similar source/target separation in the spectral domain. One of the key ingredients is the following extension of the well-known Funk-Hecke formula (cf. [@watson; @martin2006multiple]).
\[prop:Funk-Hecke\] Given ${\boldsymbol{r}}=(x, y, z)\in \mathbb R^3$, $k>0$, $\alpha\in[0, 2\pi)$ and denoted by $(r,\theta,\varphi)$ the spherical coordinates of ${\boldsymbol{r}}$, ${\boldsymbol{k}}=(\sqrt{k^2-k_z^2}\cos\alpha, \sqrt{k^2-k_z^2}\sin\alpha, k_z)$ is a vector of complex entries. Choosing branch for $\sqrt{k^2-k_z^2}$ in $e^{\ri {\boldsymbol{k}}\cdot{{\boldsymbol{r}}}}$ and $\widehat P_n^m(\frac{k_z}{k})$, then $$\label{extfunkhecke}
e^{\ri {\boldsymbol{k}}\cdot{{\boldsymbol{r}}}}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^n A_{n}^m({\boldsymbol{r}})\ri^n\widehat{P}_n^m\Big(\frac{k_z}{k}\Big)e^{-\ri m\alpha}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^n \overline{A_{n}^m({\boldsymbol{r}})}\ri^n\widehat{P}_n^m\Big(\frac{k_z}{k}\Big)e^{\ri m\alpha},$$ holds for all $k_z\in\mathbb C$, where $$A_{n}^m({\boldsymbol{r}})=4\pi j_n(kr)Y_n^m(\theta,\varphi).$$
This extension enlarges the range of the classic Funk-Hecke formula from $k_z\in (-k, k)$ to the whole complex plane by choosing the branch $$\label{branch}
\sqrt{k^2-k_z^2}=-\ri \sqrt{r_1r_2}e^{\ri\frac{\theta_1+\theta_2}{2}},$$ for the square root function $\sqrt{k^2-k_z^2}$. Here $(r_i,\theta_i), i=1, 2$ are the modules and principal values of the arguments of complex numbers $k_z+k$ and $k_z-k$, i.e., $$k_z+k=r_1e^{\ri\theta_1}, \quad -\pi<\theta_1\leq\pi,\quad k_z-k=r_2e^{\ri\theta_2},\quad -\pi<\theta_2\leq\pi.$$ It is worthy to point out that the normalized associated Legendre function $\widehat P_n^m(x)$ has also been extended to the whole complex plain by using the same branch.
[ Although we have $k_{\rho}{\boldsymbol{k}}_0=\lim\limits_{k\rightarrow 0^+}(\sqrt{k^2-k_z^2}\cos\alpha, \sqrt{k^2-k_z^2}\sin\alpha, k_z)$, with $k_z=\ri k_{\rho}$, taking limit directly in the expansion will induce singularity in the associated Legendre function. In the following, we will show how to cancel the singularity to obtain a limit version of , which gives an expansion for $e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot{\boldsymbol{r}}}$. ]{}For this purpose, we first need to recall the corresponding extended Legendre addition theorem (cf. [@wang2019fast]).
\[lemma2\] Let ${\boldsymbol{w}}=(\sqrt{1-w^2}\cos\alpha, \sqrt{1-w^2}\sin\alpha, w)$ be a vector with complex entries, $\theta, \varphi$ be the azimuthal angle and polar angles of a unit vector $\hat{{\boldsymbol{r}}}$. Define $$\beta(w)=w\cos\theta+\sqrt{1-w^2}\sin\theta\cos(\alpha-\varphi),$$ then $$\label{legendreadd}
P_n(\beta(w))=\frac{4\pi}{2n+1}\sum\limits_{m=-n}^n\widehat P_n^m(\cos\theta)\widehat P_n^m(w)e^{\ri m(\alpha-\varphi)},$$ for all $w\in\mathbb C$.
From this extended Legendre addition theorem, the following expansion can be obtained by choosing a specific ${\boldsymbol{\omega}}$ and then taking limit carefully.
\[lemma3\] Let ${\boldsymbol{k}}_0=(\cos\alpha, \sin\alpha, \ri)$ be a vector with complex entry, $\theta, \varphi$ be the azimuthal angle and polar angles of a unit vector $\hat{{\boldsymbol{r}}}$. Then $$\label{polynomialadd}
\frac{(\ri{\boldsymbol{k}}_0\cdot\hat{{\boldsymbol{r}}})^n}{n!}=\sum\limits_{m=-n}^nC_n^m\widehat P_n^m(\cos\theta)e^{\ri m(\alpha-\varphi)},$$ where $$\label{constantcnm}
C_n^m=\ri^{2n-m}\sqrt{\frac{4\pi}{(2n+1)(n+m)!(n-m)!}}.$$
For any $k\in \mathbb{R}^+$, define ${\boldsymbol{k}} = ( \sqrt{k^2+1} \cos\alpha, \sqrt{k^2+1} \sin\alpha, \ri)$. By lemma \[lemma2\], we have $$\begin{aligned}
\label{eq-ext-Legendre}
\begin{split}
k^nP_n\Big(\frac{{\boldsymbol{k}}\cdot\hat{{\boldsymbol{r}}}}{k}\Big)
&= \frac{4\pi}{2n+1}\sum\limits_{m=-n}^n\widehat P_n^m(\cos\theta)k^n\widehat P_n^m\Big(\frac{\ri}{k}\Big)e^{\ri m(\alpha-\varphi)}.
\end{split}
\end{aligned}$$ Consider the limit of the above identity as $k \to 0^+$. Note that $$\lim_{k \to 0^+} {\boldsymbol{k}}\cdot\hat{{\boldsymbol{r}}} = {\boldsymbol{k}}_0 \cdot \hat{{\boldsymbol{r}}} ,$$ together with the knowledge on the coefficient of the leading term in the Legendre polynomial $P_n(x)$ lead to $$\label{legendrelimit}
\lim_{k \to 0^+} k^n P_n\Big(\frac{{\boldsymbol{k}}\cdot\hat{{\boldsymbol{r}}}}{k}\Big)= \frac{(2n)!}{2^{n}(n!)^2}( {\boldsymbol{k}}_0 \cdot \hat{{\boldsymbol{r}}})^n.$$
Recall the Rodrigues’ formula of the associated Legendre function $$\widehat P_n^m(x)=\frac{c_{nm}}{2^nn!}(1-x^2)^{\frac{m}{2}}\frac{d^{n+m}}{dx^{n+m}}(x^2-1)^n,\quad c_{nm}=\sqrt{\frac{2n+1}{4\pi}\frac{(n-m)!}{(n+m)!}}$$ for $0 \le m \le n$, we have $$k^n \widehat P_n^m\Bigl(\frac{\ri }{k}\Bigr) = \frac{c_{nm}}{2^n n!}\frac{(2n)!}{(n-m)!}(k^2+1)^{\frac{m}{2}}\cdot k^{n-m}\widetilde{Q}_{n-m}\Bigl(\frac{\ri }{k}\Bigr)$$ where $\widetilde{Q}_n(z)$ is a *monic* polynomial of degree $n$. Hence, we get similarly $$\label{alegendrelimit}
\lim_{k \to 0^+} k^n \widehat P_n^m\Bigl(\frac{\ri}{k}\Bigr) = \frac{c_{nm}}{2^n n!}\frac{(2n)!\ri ^{n-m}}{(n-m)!}.$$ The identity $\widehat{P}_n^{-m}(x)=(-1)^m\widehat{P}_n^m(x)$ will give the limit for $-n\le m<0$ cases. Now, let $k\rightarrow 0^+$ in and use results and , we complete the proof.
\[prop:Funk-Hecke-limit\] Given ${\boldsymbol{r}}=(x, y, z)\in \mathbb R^3$, $\alpha\in[0, 2\pi)$ and denoted by $(r,\theta,\varphi)$ the spherical coordinates of ${\boldsymbol{r}}$, ${\boldsymbol{k}}_0=(\cos\alpha, \sin\alpha, \ri)$ is a vector of complex entries. Then $$\label{extfunkheckelim}
e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot{{\boldsymbol{r}}}}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^nC_n^m r^nY_n^m(\theta,\varphi)k_{\rho}^ne^{-\ri m\alpha}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^n C_n^mr^n\overline{Y_n^m(\theta,\varphi)}k_{\rho}^ne^{\ri m\alpha},$$ holds for all $r>0$, $k_{\rho}> 0$, where $C_n^m$ is the constant defined in .
By Taylor expansion, we have $$e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot{{\boldsymbol{r}}}}=\sum\limits_{n=0}^{\infty}\frac{(\ri {\boldsymbol{k}}_0\cdot\hat{{\boldsymbol{r}}})^n}{n!}k_{\rho}^nr^n.$$ Then, follows by applying lemma \[lemma3\] to each term in the above expansion.
By setting $k_z=\ri k_{\rho}$ and using the limit values given by and , one can also verify that the expansions for $e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot{{\boldsymbol{r}}}}$ in proposition \[prop:Funk-Hecke-limit\] are exactly the limiting cases of the expansions in proposition \[prop:Funk-Hecke\].
Applying spherical harmonic expansion to exponential functions $e^{-\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}-{\boldsymbol{r}}_c^s)}$ and $e^{\ri {\boldsymbol{k}}\cdot({\boldsymbol{r}}-{\boldsymbol{r}}_c^t)}$ in gives $$\label{meinspectraldomain}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n}M_{nm}\frac{(-1)^{n}c_n^2C_n^m}{2\pi}\int_0^{\infty}\int_0^{2\pi}k_{\rho}^{n}e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}-{\boldsymbol{r}}_c^s)}e^{\ri m\alpha}d\alpha dk_{\rho},$$ and $$\label{leinspectraldomain}
\frac{1}{|{\boldsymbol{r}}-{\boldsymbol{r}}'|}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n}\hat L_{nm}r_t^nY_n^m(\theta_t,\varphi_t),$$ for $z\geq z'$, where $M_{nm}$ is defined in and $$\hat L_{nm}=\frac{C_n^m}{2\pi}\int_0^{\infty}\int_0^{2\pi}k_{\rho}^{n}e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}_c^t-{\boldsymbol{r}}')}e^{-\ri m\alpha}d\alpha dk_{\rho}.$$ Recall the identity $$\label{wavefunspectralform}
r^{-n-1}Y_n^{-m}(\theta,\varphi)=\frac{(-1)^nc_n^2C_n^m}{2\pi}\int_0^{\infty}\int_0^{2\pi}k_{\rho}^{n}e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot{\boldsymbol{r}}}e^{-\ri m\alpha}d\alpha dk_{\rho},$$ for $z\geq 0$, we see that and are exactly the ME and LE in the case of $z\geq z'$.
To derive the translation from the ME to the LE , we perform further spliting $$e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}-{\boldsymbol{r}}_c^s)}=e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}-{\boldsymbol{r}}_c^t)}e^{\ri k_{\rho}{\boldsymbol{k}}_0\cdot({\boldsymbol{r}}_c^t-{\boldsymbol{r}}_c^s)},$$ in and apply expansion again to obtain the translation $$\begin{split}
L_{nm}=&C_n^m\sum\limits_{n'=0}^{\infty}\sum\limits_{m'=-n'}^{n'}M_{n'm'}\frac{(-1)^{n'}c_{n'}^2C_{n'}^{m'}}{2\pi}\int_0^{\infty}\int_0^{2\pi}k_{\rho}^{n+n'}e^{\ri k_{\rho}{\boldsymbol{k}}_0({\boldsymbol{r}}_c^t-{\boldsymbol{r}}_c^s)}e^{\ri(m'-m)\alpha}d\alpha dk_{\rho}.
\end{split}$$ By using the identity , we can also verify that the above integral form is equal to the entries of the M2L translation matrix defined in .
FMM for 3-D Laplace equation in layered media
=============================================
In this section, the potential of charges in layered media is formulated using layered Green’s function and then decomposed into a free space and four types of reaction components. Furthermore, the reaction components are re-expressed by using equivalent polarization charges. The new expressions are used to derive the MEs and LEs for the reaction components of the layered Green’s function in the same spirit as in the last section. Based on these new expansions and translations, FMM for 3-D Laplace kernel in layered media can be developed.
Potential due to sources embedded in multi-layer media
------------------------------------------------------
Consider a layered medium consisting of $L$-interfaces located at $z=d_{\ell
},\ell=0,1,\cdots,L-1$, see Fig. \[layerstructure\]. The piece wise constant material parameter is described by $\{\varepsilon_{\ell}\}_{\ell=0}^L$. Suppose we have a point source at $\boldsymbol{r}^{\prime}=(x^{\prime},y^{\prime},z^{\prime})$ in the $\ell^{\prime}$th layer ($d_{\ell^{\prime}}<z^{\prime}<d_{\ell^{\prime}-1}$), then, the layered media Green’s function $u_{\ell\ell'}({\boldsymbol{r}}, {\boldsymbol{r}}')$ for the Laplace equation satisfies $$\label{Laplaceeqlayer}
\boldsymbol{\Delta}u_{\ell\ell'}(\boldsymbol{r},\boldsymbol{r}^{\prime
})=-\delta(\boldsymbol{r},\boldsymbol{r}^{\prime}),$$ at field point $\boldsymbol{r}=(x,y,z)$ in the $\ell$th layer ($d_{\ell
}<z<d_{\ell}-1$) where $\delta(\boldsymbol{r},\boldsymbol{r}^{\prime})$ is the Dirac delta function. By using Fourier transforms along $x-$ and $y-$directions, the problem can be solved analytically for each layer in $z$ by imposing transmission conditions at the interface between $\ell$th and $(\ell-1)$th layer ($z=d_{\ell-1})$, *i.e.*, $$\label{transmissioncond}
u_{\ell-1,\ell'}(x,y,z)=u_{\ell\ell'}(x,y,z),\quad \varepsilon_{\ell-1}\frac{\partial u_{\ell-1,\ell'}(x,y,z)}{\partial z}=\varepsilon_{\ell}\frac{\partial \widehat u_{\ell\ell'}(k_{x},k_{y},z)}{\partial z},$$ as well as the decaying conditions in the top and bottom-most layers as $z\rightarrow\pm\infty$.
\[layerstructure\] 
Here, we give the expression for the analytic solution with detailed derivations included in the Appendix A. In general, the layered media Green’s function in the physical domain takes the form $$\label{layeredGreensfun}
u_{\ell\ell^{\prime}}(\boldsymbol{r},\boldsymbol{r}^{\prime})=\begin{cases}
\displaystyle u_{\ell\ell'}^{\text{r}}(\boldsymbol{r},\boldsymbol{r}^{\prime})+\frac{1}{4\pi
|{\boldsymbol{r}}-{\boldsymbol{r}}^{\prime}|},&\ell=\ell',\\
\displaystyle u_{\ell\ell'}^{\text{r}}(\boldsymbol{r},\boldsymbol{r}^{\prime}), & \text{otherwise},
\end{cases}$$ where $$\label{reactioncomponent}
u^{\text{r}}_{\ell\ell'}({\boldsymbol{r}}, {\boldsymbol{r}}')=\begin{cases}
\displaystyle u_{0\ell^{\prime}}^{11}(\boldsymbol{r},\boldsymbol{r}^{\prime})+u_{0\ell^{\prime}}^{12}(\boldsymbol{r},\boldsymbol{r}^{\prime}),\\
\displaystyle u_{\ell\ell^{\prime}}^{11}(\boldsymbol{r},\boldsymbol{r}^{\prime})+u_{\ell\ell^{\prime}}^{12}(\boldsymbol{r},\boldsymbol{r}^{\prime})+u_{\ell\ell^{\prime}}^{21
}(\boldsymbol{r},\boldsymbol{r}^{\prime})+u_{\ell\ell^{\prime}}^{22
}(\boldsymbol{r},\boldsymbol{r}^{\prime}), &
0<\ell<L,\\
\displaystyle u_{L\ell^{\prime}}^{21}(\boldsymbol{r},\boldsymbol{r}^{\prime})+u_{L\ell^{\prime}}^{22}(\boldsymbol{r},\boldsymbol{r}^{\prime}).
\end{cases}$$ The reaction component $u_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}')$ is given in an integral form $$\label{generalcomponents}
u_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}')=\frac{1}{8\pi^2 }\int_0^{\infty}\int_0^{2\pi}e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}')}\mathcal{Z}_{\ell\ell'}^{\mathfrak{ab}}(z, z')\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})d\alpha dk_{\rho},\quad \mathfrak{a, b}=1, 2,$$ where, $${\boldsymbol{k}}_{\alpha}=k_{\rho}(\cos\alpha,\sin\alpha),
\label{kalpha}$$ and $\{\mathcal{Z}_{\ell\ell'}^{\mathfrak{ab}}(z, z')\}_{\mathfrak{a, b}=1}^2$ are exponential functions defined as $$\label{zexponential}
\begin{split}
&\mathcal{Z}_{\ell\ell'}^{11}(z, z'):=e^{- k_{\rho} (z-d_{\ell}+z'-d_{\ell'})},\quad \mathcal{Z}_{\ell\ell'}^{12}(z, z'):=e^{-k_{\rho} (z-d_{\ell}+d_{\ell'-1}-z')},\\
&\mathcal{Z}_{\ell\ell'}^{21}(z, z'):=e^{-k_{\rho} (d_{\ell-1}-z+z'-d_{\ell'})},\quad \mathcal{Z}_{\ell\ell'}^{22}(z, z'):=e^{-k_{\rho} (d_{\ell-1}-z+d_{\ell'-1}-z')},
\end{split}$$ $\{\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})\}_{\mathfrak{a,b}=1}^2$ are reaction densities only dependent on the layer structure and the material parameter $k_{\ell}$ in each layer. The reaction densities can be calculated efficiently by using a recursive algorithm, see the Appendix A for more details. It is worthwhile to point out that the reaction components $u_{\ell\ell'}^{\mathfrak a2}$ or $u_{\ell\ell'}^{\mathfrak a1}$ will vanish if the source ${\boldsymbol{r}}'$ is in the top or bottom most layer.
Withe the expression of the Green’s function in layered media, we are ready to consider the potential due to sources in layered media. Let $\mathscr{P}_{\ell}=\{(Q_{\ell j},\boldsymbol{r}_{\ell j}),$ $j=1,2,\cdots
,N_{\ell}\}$, $\ell=0, 1, \cdots, L$ be $L$ groups of source charges distributed in a multi-layer medium with $L+1$ layers (see Fig. \[layerstructure\]). The group of charges in $\ell$-th layer is denoted by $\mathscr{P}_{\ell}$. Apparently, the potential at ${\boldsymbol{r}}_{\ell i}$ due to all other charges is given by the summation $$\label{potential1}
\hspace{-3pt}\Phi_{\ell}(\boldsymbol{r}_{\ell i})=\sum\limits_{\ell'=0}^{L}\sum\limits_{j=1}^{N_{\ell'}}Q_{\ell' j}u_{\ell\ell'}({\boldsymbol{r}}_{\ell i},{\boldsymbol{r}}_{\ell' j})
=\sum\limits_{j=1,j\neq i}^{N_{\ell}}\frac{Q_{\ell j}}{4\pi|{\boldsymbol{r}}_{\ell i}-{\boldsymbol{r}}_{\ell j}|}+\sum\limits_{\ell'=0}^{L}\sum\limits_{j=1}^{N_{\ell'}}Q_{\ell' j}u_{\ell\ell'}^{\text{r}}({\boldsymbol{r}}_{\ell i},{\boldsymbol{r}}_{\ell' j}),$$ where $u_{\ell\ell'}^{\text{r}}({\boldsymbol{r}},{\boldsymbol{r}}')$ are the reaction field components defined in -. As the reaction components of the Green’s function in multi-layer media have different expressions for sources and targets in different layers, it is necessary to perform calculation individually for interactions between any two groups of charges among the $L+1$ groups $\{\mathscr{P}_{\ell}\}_{\ell=0}^{L}$. Applying expressions and in , we obtain $$\label{totalinteraction}
\begin{split}
\Phi_{\ell}(\boldsymbol{r}_{\ell i})=&\Phi_{\ell}^{\text{free}}(\boldsymbol{r}_{\ell i})+\Phi_{\ell}^{\text{r}}(\boldsymbol{r}_{\ell i})\\
=&\Phi_{\ell}^{\text{free}}(\boldsymbol{r}_{\ell i})+\sum\limits_{\ell^{\prime}=0}^{L-1}[\Phi_{\ell\ell^{\prime}}^{11
}(\boldsymbol{r}_{\ell i})+\Phi_{\ell\ell^{\prime}}^{21
}(\boldsymbol{r}_{\ell i})]+\sum\limits_{\ell^{\prime}=1}^{L}[\Phi_{\ell\ell^{\prime}}^{12
}(\boldsymbol{r}_{\ell i})+\Phi_{\ell\ell^{\prime}}^{22
}(\boldsymbol{r}_{\ell i})],
\end{split}$$ where $$\label{freereactioncomponents}
\begin{split}
& \Phi_{\ell}^{\text{free}}(\boldsymbol{r}_{\ell i}):=\sum\limits_{j=1,j\neq
i}^{N_{\ell}}\frac{Q_{\ell j}}{4\pi|\boldsymbol{r}_{\ell
i}-\boldsymbol{r}_{\ell j}|},\quad \Phi_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i}):=\sum
\limits_{j=1}^{N_{\ell^{\prime}}}Q_{\ell^{\prime}j}u_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i},\boldsymbol{r}_{\ell^{\prime}j}).
\end{split}$$ Obviously, the free space component $\Phi_{\ell}^{\text{free}}(\boldsymbol{r}_{\ell i})$ can be computed using the traditional FMM. Thus, we will only focus on the computation of the reaction components $\{\Phi_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i})\}_{\mathfrak{a, b}=1}^2$.
Equivalent polarization sources for reaction components
-------------------------------------------------------
The expressions of the components given in show that the free space components only involve interactions between charges in the same layer. Interactions between charges in different layers are all included in the reaction components. Two groups of charges involved in the computation of a reaction component could be physically very far away from each other as there could be many layers between the source and target layers associated to the reaction component, see Fig. \[polarizedsource\] (left).
Our recent work on the Helmholtz equation [@zhang2018exponential; @wang2019fast], of which the Laplace equation can be considered as a special case where the wave number $k=0$, has shown that the exponential convergence of the ME and LE for the reaction components $u_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}')$ in fact depends on the distance between the target charge ${\boldsymbol{r}}$ and a polarization charge defined for the source charge ${\boldsymbol{r}}'$, which uses the distance between the source charge ${\boldsymbol{r}}'$ and the nearest material interface and always locates next to the nearest interface adjacent to the target charge. Fig. \[sourceimages\] illustrates the location of the polarization charge ${\boldsymbol{r}}'_{\mathfrak{ab}}$ for each of the four types of reaction fields $\tilde u_{\ell\ell'}^{\mathfrak{ab}},\mathfrak{a},\mathfrak{b}=1,2 $. Specifically, the equivalent polarization sources associated to reaction components $u^{\mathfrak{ab}}_{\ell\ell'}({\boldsymbol{r}}, {\boldsymbol{r}}')$, $\mathfrak{a}, \mathfrak{b} =1, 2$ are set to be at coordinates (see Fig. \[sourceimages\]) $$\label{eqpolarizedsource}
\begin{split}
&{\boldsymbol{r}}'_{11}:=(x', y', d_{\ell}-(z'-d_{\ell'})),\quad\quad{\boldsymbol{r}}'_{12}:=(x', y', d_{\ell}-(d_{\ell'-1}-z')),\\
&{\boldsymbol{r}}'_{21}:=(x', y', d_{\ell-1}+(z'-d_{\ell'})),\quad {\boldsymbol{r}}'_{22}:=(x', y', d_{\ell-1}+(d_{\ell'-1}-z')),
\end{split}$$ and the reaction potentials are $$\label{reactfieldpolarization}
\tilde u_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}'_{\mathfrak{ab}}):=\frac{1}{8\pi^2 }\int_0^{\infty}\int_0^{2\pi}e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}')}e^{-k_{\rho}|z-z^{\prime}_{\mathfrak{ab}}|}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})d\alpha dk_{\rho},\quad \mathfrak{a,b}=1, 2,$$ where $z'_{\mathfrak{ab}}$ denotes the $z$-coordinate of ${\boldsymbol{r}}'_{\mathfrak{ab}}$, i.e., $$z_{11}^{\prime}=d_{\ell}-(z'-d_{\ell'}),\; z_{12}^{\prime}=d_{\ell}-(d_{\ell'-1}-z'),\; z_{21}^{\prime}=d_{\ell-1}+(z'-d_{\ell'}),\;z_{22}^{\prime}=d_{\ell-1}+(d_{\ell'-1}-z').$$
![Equivalent polarized sources $\{{\boldsymbol{r}}_{\ell'j}^{11}\}$, $\{{\boldsymbol{r}}_{\ell'j}^{21}\}$ and boxes in source tree.[]{data-label="polarizedsource"}](polarizedcoord1121)
\
We can see that the reaction potentials represented by the equivalent polarization sources has similar form as the Sommerfeld-type integral representation of the free space Green’s function except for the extra density functions $\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})$. Moreover, recall the definition in we have $$z>z'_{1\mathfrak b}, \quad{\rm and}\quad z<z'_{2\mathfrak b},\quad \mathfrak b=1, 2.$$ Therefore, the absolute value in the integral form can be removed according to the index $\mathfrak a$. More precisely, define $$\label{mekernelimage}
{\mathcal E}^{+}({\boldsymbol{r}}, {\boldsymbol{r}}'):=e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}')}e^{k_{\rho}(z-z^{\prime})},\quad {\mathcal E}^{-}({\boldsymbol{r}}, {\boldsymbol{r}}^{\prime}):=e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}')}e^{-k_{\rho}(z-z^{\prime})},$$ then $$\begin{split}
\tilde u_{\ell\ell'}^{1\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b})=\frac{1}{8\pi^2 }\int_0^{\infty}\int_0^{2\pi}{\mathcal E}^{-}({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b})\sigma_{\ell\ell'}^{1\mathfrak b}(k_{\rho})d\alpha dk_{\rho},\\
\tilde u_{\ell\ell'}^{2\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b})=\frac{1}{8\pi^2 }\int_0^{\infty}\int_0^{2\pi}\mathcal E^{+}({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b})\sigma_{\ell\ell'}^{2\mathfrak b}(k_{\rho})d\alpha dk_{\rho}.
\end{split}$$ Recall the expressions , we verify that $$\label{expkernelexp}
{\mathcal E}^{-}({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b})=e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}')}\mathcal{Z}_{\ell\ell'}^{1\mathfrak{b}}(z, z'),\quad {\mathcal E}^{+}({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b})=e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}')}\mathcal{Z}_{\ell\ell'}^{2\mathfrak{b}}(z, z'), \quad \mathfrak b=1, 2.$$ Therefore, the reaction components is equal to the reaction potentials defined for associated equivalent polarization sources, i.e., $$\label{generalcomponentsimag}
\begin{split}
u_{\ell\ell'}^{1\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}')=\tilde u_{\ell\ell'}^{1\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b}),\quad
u_{\ell\ell'}^{2\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}')=\tilde u_{\ell\ell'}^{2\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b}), \quad \mathfrak b=1, 2.
\end{split}$$ A substitution into the expression of $\Phi_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i})$ in leads to $$\label{reactcompusingpolar}
\Phi_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i})=\sum
\limits_{j=1}^{N_{\ell^{\prime}}}Q_{\ell^{\prime}j}\tilde u_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i},\boldsymbol{r}^{\mathfrak ab}_{\ell^{\prime}j}),\quad \mathfrak{a,b}=1, 2,$$ where $$\label{equivpolarcoord}
\begin{split}
&\boldsymbol{r}^{11}_{\ell^{\prime}j}=(x_{\ell'j}, y_{\ell'j}, d_{\ell}-(z_{\ell j}-d_{\ell'})), \quad\;\;\;\, \boldsymbol{r}^{12}_{\ell^{\prime}j}=(x_{\ell'j}, y_{\ell'j}, d_{\ell}-(d_{\ell'-1}-z_{\ell j})),\\
&\boldsymbol{r}^{21}_{\ell^{\prime}j}=(x_{\ell'j}, y_{\ell'j}, d_{\ell-1}+(z_{\ell j}-d_{\ell'})), \quad \boldsymbol{r}^{22}_{\ell^{\prime}j}=(x_{\ell'j}, y_{\ell'j}, d_{\ell-1}+(d_{\ell'-1}-z_{\ell j})),
\end{split}$$ are coordinates of the associated equivalent polarization sources for the computation of reaction components $\Phi_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i})$, see Fig \[polarizedsource\] for an illustration of $\{{\boldsymbol{r}}_{\ell'j}^{11}\}_{j=1}^{N_{\ell'}}$ and $\{{\boldsymbol{r}}_{\ell'j}^{21}\}_{j=1}^{N_{\ell'}}$.
By using the expression , the computation of the reaction components can be performed between targets and associated equivalent polarization sources. The definition given by shows that the target particles $\{{\boldsymbol{r}}_{\ell i}\}_{i=1}^{N_{\ell}}$ and the corresponding equivalent polarization sources are always located on different sides of an interface $z=d_{\ell-1}$ or $z=d_{\ell}$, see Fig. \[polarizedsource\]. We still emphasize that the introduced equivalent polarization sources are separate with the target charges even in considering the reaction components for source and target charges in the same layer, see the numerical examples given in section 3.4. This property implies significant advantage of introducing equivalent polarization sources and using expression in the FMMs for the reaction components $\Phi_{\ell\ell^{\prime}}^{\mathfrak{ab}}(\boldsymbol{r}_{\ell i})$, $\mathfrak a,b=1,2$. The numerical results presented in Section 4 show that the FMMs for reaction components have high efficiency as a direct consequence of the separation of the targets and equivalent polarization sources by interface.
Fast multipole algorithm
------------------------
In the development of FMM for reaction components $\Phi_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}_{\ell i})$, we will adopt the expression with equivalent polarization sources. Therefore, multipole and local expansions and corresponding translation operators for $\tilde u_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}'_{\mathfrak{ab}})$ are derived first. Inspired by source/target separation in , similar separations $$\label{sourcetargetseparationsc}
\begin{split}
{\mathcal E}^{-}({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b})=\mathcal E^{-}({\boldsymbol{r}}, {\boldsymbol{r}}^{1\mathfrak b}_c)e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}_c^{1\mathfrak b}-{\boldsymbol{\rho}}'_{1\mathfrak b})-k_{\rho}(z^{1\mathfrak b}_c-z'_{1\mathfrak b})},\\
{\mathcal E}^{+}({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b})=\mathcal E^{+}({\boldsymbol{r}}, {\boldsymbol{r}}^{2\mathfrak b}_c)e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}_c^{2\mathfrak b}-{\boldsymbol{\rho}}'_{2\mathfrak b})+k_{\rho}(z^{2\mathfrak b}_c-z'_{2\mathfrak b})},
\end{split}$$ and $$\label{sourcetargetseparationtc}
\begin{split}
\mathcal E^-({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b})&=\mathcal E^{-}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{1\mathfrak b})e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}_c^t)-k_{\rho}(z-z_c^t)},\\
\mathcal E^+({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b})&=\mathcal E^{+}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{2\mathfrak b})e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}_c^t)+k_{\rho}(z-z_c^t)},
\end{split}$$ for $\mathfrak b=1, 2$ are introduced by inserting the source center ${\boldsymbol{r}}^{\mathfrak {ab}}_c=(x_c^{\mathfrak{ab}},y_c^{\mathfrak{ab}}, z_c^{\mathfrak{ab}} )$ and the target center ${\boldsymbol{r}}_c^t=(x_c^t, y_c^t, z_c^t)$, respectively. Here, we also use notations ${\boldsymbol{\rho}}_c^{\mathfrak{ab}}=(x_c^{\mathfrak{ab}}, y_c^{\mathfrak{ab}})$, ${\boldsymbol{\rho}}_c^t=(x_c^t, y_c^t)$ for coordinates projected in $xy$-plane. Now, applying proposition \[prop:Funk-Hecke-limit\] gives us the following spherical harmonic expansions: $$\label{planewavesphexpmultipole}
\begin{split}
&e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}_c^{2\mathfrak b}-{\boldsymbol{\rho}}'_{2\mathfrak b})+k_{\rho}(z^{2\mathfrak b}_c-z'_{2\mathfrak b})}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} C_n^m (r_c^{2\mathfrak b})^{n}\overline{Y_n^m(\theta^{2\mathfrak b}_c,\pi+\varphi^{2\mathfrak b}_c)}k_{\rho}^ne^{\ri m\alpha},\\
&e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}_c^{1\mathfrak b}-{\boldsymbol{\rho}}'_{1\mathfrak b})-k_{\rho}(z^{1\mathfrak b}_c-z'_{1\mathfrak b})}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} C_n^m (r_c^{1\mathfrak b})^n\overline{Y_n^m(\pi-\theta^{1\mathfrak b}_c,\pi+\varphi^{1\mathfrak b}_c)}k_{\rho}^ne^{\ri m\alpha},\\
\end{split}$$ and $$\label{planewavesphexplocal}
\begin{split}
&e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}_c^t)- k_{\rho}(z-z_c^t)}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} C_n^m r_t^nY_n^m(\theta_t,\varphi_t)k_{\rho}^ne^{-\ri m\alpha},\\
&e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}_c^t)+ k_{\rho}(z-z_c^t)}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} C_n^m r_t^nY_n^m(\pi-\theta_t,\varphi_t)k_{\rho}^ne^{-\ri m\alpha},
\end{split}$$ where $(r_c^{\mathfrak{ab}}, \theta_c^{\mathfrak{ab}}, \varphi_c^{\mathfrak{ab}})$ is the spherical coordinates of ${\boldsymbol{r}}'_{\mathfrak{ab}}-{\boldsymbol{r}}_c^{\mathfrak{ab}}$. By equalities $$Y_n^{m}(\pi-\theta,\varphi)=(-1)^{n+m}Y_n^{m}(\theta,\varphi),\quad Y_n^{m}(\theta,\pi+\varphi)=(-1)^{m}Y_n^{m}(\theta,\varphi),$$ the above spherical harmonic expansions - together with source/target separation and lead to $$\label{integrandme}
\begin{split}
\mathcal E^-({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b})&=\mathcal E^-({\boldsymbol{r}}, {\boldsymbol{r}}_c^{1\mathfrak b})\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} (-1)^nC_n^m ( r^{1\mathfrak b}_c)^n\overline{Y_n^m(\theta^{1\mathfrak b}_c,\varphi^{1\mathfrak b}_c)}k_{\rho}^ne^{\ri m\alpha},\\
\mathcal E^+({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b})&=\mathcal E^+({\boldsymbol{r}}, {\boldsymbol{r}}_c^{2\mathfrak b})\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} (-1)^{m}C_n^m ( r^{2\mathfrak b}_c)^n\overline{Y_n^m(\theta^{2\mathfrak b}_c,\varphi^{2\mathfrak b}_c)}k_{\rho}^ne^{\ri m\alpha},
\end{split}$$ and $$\label{integrandle}
\begin{split}
\mathcal E^-({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b})&=\mathcal E^-({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{1\mathfrak b})\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} C_n^m r_t^nY_n^m(\theta_t,\varphi_t)k_{\rho}^ne^{-\ri m\alpha},\\
\mathcal E^+({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b})&=\mathcal E^+({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{2\mathfrak b})\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} (-1)^{n+m}C_n^m r_t^nY_n^m(\theta_t,\varphi_t)k_{\rho}^ne^{-\ri m\alpha},
\end{split}$$ for $\mathfrak b=1, 2$. Then, a substitution of and into gives a ME $$\label{melayerupgoingimage}
\begin{split}
\tilde u_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}'_{\mathfrak{ab}})=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} M_{nm}^{\mathfrak{ab}}\widetilde{\mathcal F}_{nm}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{\mathfrak{ab}}), \quad M_{nm}^{\mathfrak{ab}}=c_n^{-2} (r_c^{\mathfrak{ab}})^n\overline{Y_n^{m}(\theta_c^{\mathfrak{ab}},\varphi_c^{\mathfrak{ab}})},
\end{split}$$ at equivalent polarization source centers ${\boldsymbol{r}}_c^{\mathfrak{ab}}$ and LE $$\label{lelayerimage}
\begin{split}
\tilde u_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}'_{\mathfrak{ab}})=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} L_{nm}^{\mathfrak{ab}}r_t^nY_n^m(\theta_t,\varphi_t)
\end{split}$$ at target center ${\boldsymbol{r}}_c^t$, respectively. Here, $\widetilde{\mathcal F}_{nm}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{\mathfrak{ab}})$ are given in forms of Sommerfeld-type integrals $$\label{mebasis}
\begin{split}
\widetilde{\mathcal F}_{nm}^{1\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{1\mathfrak b})=&\frac{(-1)^{n}c_n^2C_n^m}{8\pi^2}\int_0^{\infty}\int_0^{2\pi}{\mathcal E}^{-}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{1\mathfrak b})\sigma_{\ell\ell'}^{1\mathfrak b}(k_{\rho})k_{\rho}^{n}e^{\ri m\alpha}d\alpha dk_{\rho},\\
\widetilde{\mathcal F}_{nm}^{2\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{2\mathfrak b})=&\frac{(-1)^mc_n^2C_n^m}{8\pi^2}\int_0^{\infty}\int_0^{2\pi}{\mathcal E}^{+}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{2\mathfrak b})\sigma_{\ell\ell'}^{2\mathfrak b}(k_{\rho})k_{\rho}^{n}e^{\ri m\alpha}d\alpha dk_{\rho},
\end{split}$$ and the LE coefficients are given by $$\label{lecoeffimage}
\begin{split}
L_{nm}^{1\mathfrak b}=&\frac{C_n^m}{8\pi^2 }\int_0^{\infty}\int_0^{2\pi}{\mathcal E}^{-}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{1\mathfrak b})\sigma_{\ell\ell^{\prime}}^{1\mathfrak b}(k_{\rho})k_{\rho}^{n}e^{-\ri m\alpha}d\alpha dk_{\rho},\\
L_{nm}^{2\mathfrak b}=&\frac{(-1)^{n+m}C_n^m}{8\pi^2 }\int_0^{\infty}\int_0^{2\pi}{\mathcal E}^{+}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{2\mathfrak b})\sigma_{\ell\ell^{\prime}}^{2\mathfrak b}(k_{\rho})k_{\rho}^{n}e^{-\ri m\alpha}d\alpha dk_{\rho}.
\end{split}$$
Let us give some numerical examples to show the convergence behavior of the MEs in . Consider the MEs of $\tilde u_{11}^{11}({\boldsymbol{r}}, {\boldsymbol{r}}'_{11})$ and $\tilde u_{11}^{22}({\boldsymbol{r}}, {\boldsymbol{r}}'_{22})$ in a three-layer media with $\varepsilon_0=21.2$, $\varepsilon_1=47.5$, $\varepsilon_2=62.8$, $d_0=0, d_1=-1.2$. In all the following examples, we fix ${\boldsymbol{r}}'=(0.625, 0.5, -0.1)$ in the middle layer and use definition to determine ${\boldsymbol{r}}_{11}^{\prime}=(0.625, 0.5, -2.3)$, ${\boldsymbol{r}}_{22}^{\prime}=(0.625, 0.5, 0.1)$. The centers for MEs are set to be ${\boldsymbol{r}}_c^{11}=(0.6, 0.6, -2.4)$, ${\boldsymbol{r}}_c^{22}=(0.6, 0.6, 0.2)$ which implies $|{\boldsymbol{r}}_{11}'-{\boldsymbol{r}}_c^{11}|=|{\boldsymbol{r}}_{22}'-{\boldsymbol{r}}_c^{22}|\approx0.1436$. For both components, we shall test three targets given as follows $${\boldsymbol{r}}_1=(0.5, 0.625, -0.1),\quad {\boldsymbol{r}}_2= (0.5, 0.625, -0.6), \quad {\boldsymbol{r}}_3=(0.5, 0.625, -1.1).$$ The relative errors against truncation number $p$ are depicted in Fig. \[meconvergence\]. We also plot the convergence rates similar with that of the ME of free space Green’s function, i.e., $O\Big[\Big(\frac{|{\boldsymbol{r}}-{\boldsymbol{r}}_c^{\mathfrak{ab}}|}{|{\boldsymbol{r}}_{\mathfrak{ab}}'-{\boldsymbol{r}}_c^{\mathfrak{ab}}|}\Big)^{p+1}\Big]$ as reference convergence rates. The results clearly show that the MEs of the reaction components $u_{11}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}'_{\mathfrak{ab}})$ have spectral convergence rate $O\Big[\Big(\frac{|{\boldsymbol{r}}-{\boldsymbol{r}}_c^{\mathfrak{ab}}|}{|{\boldsymbol{r}}_{\mathfrak{ab}}'-{\boldsymbol{r}}_c^{\mathfrak{ab}}|}\Big)^{p+1}\Big]$ similar as that of free space Green’s function. Actually, their exponential convergence has been determined by the Euclidean distance between target and polarization source. Therefore, the MEs can be used to develop FMM for efficient computation of the reaction components as in the development of classic FMM for the free space Green’s function.
According to the definition of $\mathcal E^-({\boldsymbol{r}}, {\boldsymbol{r}}')$ and $\mathcal E^+({\boldsymbol{r}}, {\boldsymbol{r}}')$ in , the centers ${\boldsymbol{r}}_c^t$ and ${\boldsymbol{r}}_c^{\mathfrak{ab}}$ have to satisfy $$\label{imagecentercond}
z_c^{1\mathfrak b}<d_{\ell}, \quad z_c^{2\mathfrak b}>d_{\ell-1},\quad z_c^t>d_{\ell} \;\; {\rm for}\;\; \tilde u_{\ell\ell'}^{1\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}'_{1\mathfrak b}); \quad z_c^t<d_{\ell-1} \;\; {\rm for}\;\; \tilde u_{\ell\ell'}^{2\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}'_{2\mathfrak b}),$$ to ensure the exponential decay in $\mathcal E^-({\boldsymbol{r}}, {\boldsymbol{r}}^{1\mathfrak b}_c), \mathcal E^+({\boldsymbol{r}}, {\boldsymbol{r}}^{2\mathfrak b}_c)$ and $\mathcal E^{-}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{1\mathfrak b}), \mathcal E^{+}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}'_{2\mathfrak b})$ as $k_{\rho}\rightarrow\infty$ and hence the convergence of the corresponding Sommerfeld-type integrals in and . These restrictions can be met easily in practice, as we are considering targets in the $\ell$-th layer and the equivalent polarized coordinates are always located either above the interface $z=d_{\ell-1}$ or below the interface $z=d_{\ell}$. More details will discussed below in the presentation of the FMM algorithm.
We still need to consider the center shifting and translation operators for ME and LE . A desirable feature of the expansions of reaction components discussed above is that the formula for the ME coefficients and the formula for the LE have exactly the same form as the formulas of ME coefficients and LE for the free space Green’s function. Therefore, the center shifting for MEs and LEs of reaction components are exactly the same as free space case given in -.
Next, we derive the translation operator from the ME to the LE . Recall the definition of exponential functions in , ${\mathcal E}^{-}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{1\mathfrak b})$ and ${\mathcal E}^{+}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{2\mathfrak b})$ can have splitting $$\begin{split}
{\mathcal E}^{-}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{1\mathfrak{b}})&={\mathcal E}^{-}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}_c^{1\mathfrak b})e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}_c^t)}e^{-k_{\rho} (z-z_c^t)},\\
{\mathcal E}^{+}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{2\mathfrak b})&={\mathcal E}^{+}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}_c^{2\mathfrak b})e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}_c^t)}e^{k_{\rho} (z-z_c^t)}.
\end{split}$$ Applying spherical harmonic expansion again, we obtain $$e^{\ri{\boldsymbol{k}}_{\alpha}\cdot({\boldsymbol{\rho}}-{\boldsymbol{\rho}}_c^t)}e^{\pm k_{\rho} (z-z_c^t)}=\sum\limits_{n=0}^{\infty}\sum\limits_{m=-n}^{n} (\mp 1)^{n+m}C_n^m r_t^nY_n^m(\theta_t,\varphi_t)k_{\rho}^ne^{-\ri m\alpha}.$$ Substituting into , the ME is translated to LE via $$\label{metoleimage1}
L_{nm}^{1\mathfrak{b}}=\sum\limits_{n'=0}^{\infty}\sum\limits_{|m'|=0}^{n'}T_{nm,n'm'}^{1\mathfrak{b}}M_{n'm'}^{1\mathfrak{b}},\quad L_{nm}^{2\mathfrak{b}}=(-1)^{n+m}\sum\limits_{n'=0}^{\infty}\sum\limits_{|m'|=0}^{n'}T_{nm,n'm'}^{2\mathfrak{b}}M_{n'm'}^{2\mathfrak{b}},$$ and the M2L translation operators are given in integral forms as follows $$\label{metoleimage2}
\begin{split}
T_{nm,n'm'}^{1\mathfrak{b}}=&\frac{(-1)^{n'}D_{nm}^{n'm'}}{8\pi^2}\int_0^{\infty}\int_0^{2\pi}{\mathcal E}^{-}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}_c^{1\mathfrak{b}})\sigma_{\ell\ell'}^{1\mathfrak{b}}(k_{\rho})k_{\rho}^{n+n'}e^{\ri (m'-m)\alpha}d\alpha dk_{\rho},\\
T_{nm,n'm'}^{2\mathfrak{b}}=&\frac{(-1)^{m'}D_{nm}^{n'm'}}{8\pi^2}\int_0^{\infty}\int_0^{2\pi}{\mathcal E}^{+}({\boldsymbol{r}}_c^t, {\boldsymbol{r}}_c^{2\mathfrak{b}})\sigma_{\ell\ell'}^{2\mathfrak{b}}(k_{\rho})k_{\rho}^{n+n'}e^{\ri (m'-m)\alpha}d\alpha dk_{\rho},
\end{split}$$ where $$D_{nm}^{n'm'}=c_{n'}^2C_n^mC_{n'}^{m'}.$$ Again, the convergence of the Sommerfeld-type integrals in is ensured by the conditions in .
[ The framework of the traditional FMM together with ME , LE , M2L translation - and free space ME and LE center shifting and constitute the FMM for the computation of reaction components $\Phi_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}_{\ell i})$, $\mathfrak a, \mathfrak b=1, 2$. In the FMM for each reaction component, a large box is defined to include all equivalent polarization sources associated to the reaction component and corresponding target charges, and an adaptive tree structure will be built by a bisection procedure, see. Fig. \[polarizedsource\]. Note that the validity of the ME , LE and M2L translation used in the algorithm imposes restrictions on the centers, accordingly. This can be ensured by setting the largest box for the specific reaction component to be equally divided by the interface between equivalent polarized sources and corresponding targets, see. Fig. \[polarizedsource\]. Thus, the largest box for the FMM implementation will be different for different reaction components. With this setting, all source and target boxes of higher than zeroth level in the adaptive tree structure will have centers below or above the interfaces, accordingly. The fast multipole algorithm for the computation of a general reaction component $\Phi_{\ell\ell'}^{\mathfrak{ab}}({\boldsymbol{r}}_{\ell i})$ is summarized in Algorithm 1. Total interactions given by will be obtained by first calculating all components and then summing them up where the algorithm is presented in Algorithm 2.]{}
Efficient computation of Sommerfeld-type integrals
--------------------------------------------------
It is clear that the FMM demands efficient computation of the double integrals involved in the MEs, LEs and M2L translations. [In this section, we present an accurate and efficient way to compute these double integrals.]{} Firstly, the double integrals can be simplified by using the following identity $$J_n(z)=\frac{1}{2\pi \ri^n}\int_0^{2\pi}e^{\ri z\cos\theta+\ri n\theta}d\theta.$$ In particular, multipole expansion functions in can be simplified as $$\begin{split}
\widetilde{\mathcal F}_{nm}^{1\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{1\mathfrak b})=&\frac{(-1)^{n}c_n^2C_n^m\ri^me^{\ri m\phi_s^{1\mathfrak b}}}{4\pi}\int_0^{\infty}J_m(k_{\rho}\rho_s^{1\mathfrak b})e^{-k_{\rho}(z-z_c^{1\mathfrak b})}\sigma_{\ell\ell'}^{1\mathfrak b}(k_{\rho})k_{\rho}^{n}dk_{\rho},\\
\widetilde{\mathcal F}_{nm}^{2\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{2\mathfrak b})=&\frac{(-1)^mc_n^2C_n^m\ri^me^{\ri m\phi_s^{2\mathfrak b}}}{4\pi}\int_0^{\infty}J_m(k_{\rho}\rho_s^{2\mathfrak b})e^{-k_{\rho}(z_c^{2\mathfrak b}-z)}\sigma_{\ell\ell'}^{2\mathfrak b}(k_{\rho})k_{\rho}^{n} dk_{\rho},
\end{split}$$ and the expression for LE coefficients can be simplified as $$\begin{split}
L_{nm}^{1\mathfrak b}=&\frac{(-1)^mC_n^m\ri^{-m}e^{-\ri m\varphi_t^{1\mathfrak{b}}}}{4\pi }\int_0^{\infty}J_{m}(k_{\rho}\rho_t^{1\mathfrak{b}})e^{-k_{\rho}(z_c^t-z_{1\mathfrak b}')}\sigma_{\ell\ell^{\prime}}^{1\mathfrak b}(k_{\rho})k_{\rho}^{n} dk_{\rho},\\
L_{nm}^{2\mathfrak b}=&\frac{(-1)^{n}C_n^m\ri^{-m}e^{-\ri m\varphi_t^{2\mathfrak{b}}}}{4\pi }\int_0^{\infty}J_{m}(k_{\rho}\rho_t^{2\mathfrak{b}})e^{-k_{\rho}(z_{2\mathfrak b}'-z_c^t)}\sigma_{\ell\ell^{\prime}}^{2\mathfrak b}(k_{\rho})k_{\rho}^{n}dk_{\rho}
\end{split}$$ for $\mathfrak b=1, 2$, where $(\rho_s^{\mathfrak{ab}}, \varphi_s^{\mathfrak{ab}})$ and $(\rho_t^{\mathfrak{ab}}, \varphi_t^{\mathfrak{ab}})$ are polar coordinates of ${\boldsymbol{r}}-{\boldsymbol{r}}_c^{\mathfrak{ab}}$ and ${\boldsymbol{r}}_c^t-{\boldsymbol{r}}'_{\mathfrak{ab}}$ projected in the $xy$ plane.
\[algorithm1\]
Determine equivalent polarized coordinates for all source particles. form the free-space ME by merging children’s expansions using the free-space center shift translation operator . shift the LE of $j$’s parent to $j$ itself using the free-space shifting . collect interaction list contribution using the source box to target box translation operator in Eq. while $T_{nm,n'm'}^{\mathfrak{ab}}$ are computed using and recurrence formula . evaluate the LE at each particle location. compute Eq. of target particle $i$ in the neighboring boxes [using the mixed DE-SE quadrature for $I_{00}^{\mathfrak{ab}}(\rho, z)$.]{}
\[algorithm2\]
use [**Algorithm 1**]{} to compute $\Phi_{\ell\ell'}^{11}({\boldsymbol{r}}_{\ell i})$, $i=1, 2, \cdots, N_{\ell}$. use [**Algorithm 1**]{} to compute $\Phi_{\ell\ell'}^{12}({\boldsymbol{r}}_{\ell i})$, $i=1, 2, \cdots, N_{\ell}$. use [**Algorithm 1**]{} to compute $\Phi_{\ell\ell'}^{21}({\boldsymbol{r}}_{\ell i})$, $i=1, 2, \cdots, N_{\ell}$. use [**Algorithm 1**]{} to compute $\Phi_{\ell\ell'}^{22}({\boldsymbol{r}}_{\ell i})$, $i=1, 2, \cdots, N_{\ell}$.
Moreover, the M2L translation can be simplified as $$\label{me2lesimplified}
\begin{split}
T_{nm,n'm'}^{1\mathfrak{b}}=&\frac{(-1)^{n'}\widetilde D_{nm}^{n'm'}(\varphi_{ts}^{1\mathfrak b})}{4\pi}\int_0^{\infty}k_{\rho}^{n+n'} J_{m'-m}(k_{\rho}\rho_{ts}^{1\mathfrak b})e^{-k_{\rho}(z_c^t-z_c^{1\mathfrak b})}\sigma_{\ell\ell'}^{1\mathfrak{b}}(k_{\rho})dk_{\rho},\\
T_{nm,n'm'}^{2\mathfrak{b}}=&\frac{(-1)^{m'}\widetilde D_{nm}^{n'm'}(\varphi_{ts}^{2\mathfrak b})}{4\pi}\int_0^{\infty}k_{\rho}^{n+n'} J_{m'-m}(k_{\rho}\rho_{ts}^{2\mathfrak b})e^{-k_{\rho}(z_c^{2\mathfrak b}-z_c^t)}\sigma_{\ell\ell'}^{2\mathfrak{b}}(k_{\rho})dk_{\rho},
\end{split}$$ where $(\rho_{ts}^{\mathfrak{ab}}, {\varphi}_{ts}^{\mathfrak{ab}})$ is the polar coordinates of ${\boldsymbol{r}}_c^t-{\boldsymbol{r}}_c^{\mathfrak{ab}}$ projected in the $xy$ plane, $$\widetilde D_{nm}^{n'm'}(\varphi)=D_{nm}^{n'm'}\ri^{m'-m}e^{\ri(m'-m){\varphi}}.$$ Define integral $$\label{uniformintegral}
\begin{split}
I_{nm}^{\mathfrak{ab}}(\rho, z):=\int_0^{\infty}J_{m}(k_{\rho}\rho)\frac{k_{\rho}^{n}e^{-k_{\rho}z}}{\sqrt{(n+m)!(n-m)!}}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho}) dk_{\rho},
\end{split}$$ then $$\label{coefintable}
\begin{split}
&\widetilde{\mathcal F}_{nm}^{1\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{1\mathfrak b})=\frac{c_ne^{\ri m\varphi_s^{1\mathfrak b}}}{4\pi}I_{nm}^{1\mathfrak b}(\rho_s^{1\mathfrak b}, z-z_c^{1\mathfrak b}),\\
& \widetilde{\mathcal F}_{nm}^{2\mathfrak b}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{2\mathfrak b})=\frac{(-1)^{n+m}c_ne^{\ri m\varphi_s^{2\mathfrak b}}}{4\pi}I_{nm}^{2\mathfrak b}(\rho_s^{2\mathfrak b}, z_c^{2\mathfrak b}-z),\\
&L_{nm}^{1\mathfrak b}=\frac{(-1)^nc_n^{-1}e^{-\ri m\varphi_t^{1\mathfrak{b}}}}{4\pi }I_{nm}^{1\mathfrak b}(\rho_t^{1\mathfrak b}, z_c^t-z'_{1\mathfrak b}),\\
& L_{nm}^{2\mathfrak b}=\frac{(-1)^mc_n^{-1}e^{-\ri m\varphi_t^{2\mathfrak{b}}}}{4\pi }I_{nm}^{2\mathfrak b}(\rho_t^{2\mathfrak b}, z'_{2\mathfrak b}-z_c^t),\\
& T_{nm,n'm'}^{1\mathfrak b}=\frac{(-1)^{n+m}Q_{nm}^{n'm'}e^{\ri(m'-m)\varphi_{ts}^{1\mathfrak b}}}{4\pi}I_{n+n',m'-m}^{1\mathfrak b}(\rho_{ts}^{1\mathfrak b}, z_c^t- z_c^{1\mathfrak b}),\\
& T_{nm,n'm'}^{2\mathfrak b}=\frac{(-1)^{n+m+n'+m'}Q_{nm}^{n'm'}e^{\ri(m'-m)\varphi_{ts}^{2\mathfrak b}}}{4\pi}I_{n+n',m'-m}^{2\mathfrak b}(\rho_{ts}^{2\mathfrak b}, z_c^{2\mathfrak b}-z_c^t),
\end{split}$$ where $$Q_{nm}^{n'm'}:=\sqrt{\frac{(2n'+1)(n+n'+m'-m)!(n+n'-m'+m)!}{(2n+1)(n+m)!(n-m)!(n'+m')!(n'-m')!}}.$$ Therefore, we actually need efficient algorithm for the computation of the Sommerfeld-type integrals $I_{nm}^{\mathfrak{ab}}(\rho, z)$ defined in . It is clearly that they have oscillatory integrands. These integrals are convergent when the target and source particles are not exactly on the interfaces of the layered medium. High order quadrature rules could be used for direct numerical computation at runtime. However, this becomes prohibitively expensive due to a large number of integrals needed in the FMM. In fact, $(p+1)(2p+1)$ integrals will be required for each source box to target box translation. Moreover, the involved integrand decays more slowly as $n$ increases.
An important aspect in the implementation of FMM concerns scaling. Since $M_{nm}^{\mathfrak{ab}}\approx(|{\boldsymbol{r}}-{\boldsymbol{r}}_c^{\mathfrak{ab}}|)^n$, $L_{nm}^{\mathfrak{ab}}\approx(|{\boldsymbol{r}}^{\mathfrak{ab}}-{\boldsymbol{r}}_c^t|)^{-n}$, a naive use of the expansions and in the implementation of FMM is likely to encounter underflow and overflow issues. To avoid this, one must scale expansions, replacing $M_{nm}$ with $M_{nm}^{\mathfrak{ab}}/S^n$ and $L_{nm}^{\mathfrak{ab}}$ with $L_{nm}^{\mathfrak{ab}}\cdot S^n$ where $S$ is the scaling factor. To compensate for this scaling, we replace $\widetilde{\mathcal F}_{nm}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{\mathfrak{ab}})$ with $\widetilde{\mathcal F}_{nm}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{\mathfrak{ab}})\cdot S^n$, $T_{nm,n'm'}^{\mathfrak{ab}}$ with $T_{nm,n'm'}^{\mathfrak{ab}}\cdot S^{n+n'}$. Usually, the scaling factor $S$ is chosen to be the size of the box in which the computation occurs. Therefore, the following scaled Sommerfeld-type integrals $$\label{scaledsommerfeldint}
S^{n}I_{nm}^{\mathfrak{ab}}(\rho, z)=\int_0^{\infty}J_{m}(k_{\rho}\rho)\frac{(k_{\rho}S)^{n}e^{-k_{\rho}z}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})}{\sqrt{(n+m)!(n-m)!}}{\rm d}k_{\rho},\;\;n\geq 0,\;\;m=0, 1, \cdots, n,$$ are involved in the implementation of the FMM.
Recall the recurrence formula $$J_{m+1}(z)=\frac{2m}{z}J_{m}(z)-J_{m-1}(z),$$ and define $a_n=\sqrt{n(n+1)}$. We have $$\begin{split}
S^{n}I_{nm+1}^{\mathfrak{ab}}(\rho, z)=&\int_0^{\infty}J_{m+1}(k_{\rho}\rho)\frac{(k_{\rho}S)^{n}e^{-k_{\rho}z}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})}{\sqrt{(n+m+1)!(n-m-1)!}}{\rm d}k_{\rho}\\
=&\frac{2m S}{\rho }\int_0^{\infty}J_{m}(k_{\rho}\rho)\frac{(k_{\rho}S)^{n-1}e^{-k_{\rho}z}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})}{\sqrt{(n+m-1)!(n-m-1)!}}\sqrt{\frac{(n+m-1)!}{(n+m+1)!}}{\rm d}k_{\rho}\\
-&\int_0^{\infty}J_{m-1}(k_{\rho}\rho)\frac{(k_{\rho}S)^{n}e^{-k_{\rho}z}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})}{\sqrt{(n+m-1)!(n-m+1)!}}\sqrt{\frac{(n+m-1)!(n-m+1)!}{(n+m+1)!(n-m-1)!}}{\rm d}k_{\rho},
\end{split}$$ which gives the forward recurrence formula $$\label{recurrence}
S^{n}I_{nm+1}^{\mathfrak{ab}}(\rho, z)=\frac{2m}{a_{n+m}}\frac{S}{\rho}S^{n-1}I_{n-1m}^{\mathfrak{ab}}(\rho, z)-\frac{a_{n-m}}{a_{n+m}}S^{n}I_{nm-1}^{\mathfrak{ab}}(\rho, z),$$ for $m\geq 1, n\geq m+1$. This recurrence formula is stable if $$\frac{2m}{a_{n+m}}<\frac{\rho}{S}.$$
In the computation of $\widetilde{\mathcal F}_{nm}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{\mathfrak{ab}})\cdot S^n$ and $L_{nm}^{\mathfrak{ab}}\cdot S^n$, $\rho^{\mathfrak{ab}}_s$ and $\rho_t^{\mathfrak{ab}}$ could be arbitrary small. Therefore, the forward recurrence formula may not be able to applied to calculate them. Nevertheless, it is unnecessary to calculate $\widetilde{\mathcal F}_{nm}^{\mathfrak{ab}}({\boldsymbol{r}}, {\boldsymbol{r}}_c^{\mathfrak{ab}})\cdot S^n$ and $L_{nm}^{\mathfrak{ab}}\cdot S^n$ directly in the FMM. The coefficients $L_{nm}^{\mathfrak{ab}}\cdot S^n$ are calculated from ME coefficients via M2L translations and then the potentials are obtained via LEs . Therefore, we only need to consider the computation of the integrals involved in the M2L translation matrices $T_{nm,n'm'}^{\mathfrak{ab}}$. For any polarization source box in the interaction list of a given target box, one can find that $\rho_{ts}^{\mathfrak{ab}}$ is either $0$ or larger than the box size $S$. If $\rho_{ts}^{\mathfrak{ab}}=0$, we directly have $$I_{nm}^{\mathfrak{ab}}(\rho_{ts}^{\mathfrak {ab}},z)=0, \quad \forall m>0,\;\;\forall z>0.$$ In all other cases, we have $\rho_{ts}^{\mathfrak{ab}}\geq S$ and the forward recurrence formula can always be applied as we have $$\frac{2m}{\sqrt{(n+m+1)(n+m)}}<\frac{1}{\sqrt{3}}<\frac{\rho_{ts}^{\mathfrak{ab}}}{S},\quad n\geq m+1,\quad m\geq 1.$$
Given a truncation number $p$, we still need to use quadratures to calculate $4p+1$ initial values $ \{I_{n 0}^{\mathfrak{ab}}(\rho, z)\}_{n=0}^{2p}$ and $\{ I_{n 1}^{\mathfrak{ab}}(\rho, z)\}_{n=1}^{2p}$ for each M2L translation. Moreover, integrals $\{I_{00}^{\mathfrak{ab}}(\rho, z)\}_{\mathfrak{a,b}=1}^2$ are also required in the computation of the direct interactions between particles in neighboring boxes. These calculations require an efficient and robust numerical method. Note that $\{I_{00}^{\mathfrak{ab}}(\rho, z)\}_{\mathfrak{a,b}=1}^2$ are exactly the Sommerfeld integrals involved in the calculation of the layered Green’s function. A multitude of papers have been published until now, devoted to their efficient calculation (see [@michalski2016efficient] and the references there in).
Basically, we will adopt the mixed DE-SE quadrature (cf. [@michalski2016efficient; @takahasi1974double]) in this paper for efficient computations of the Sommerfeld-type integrals. Nevertheless, we still need to consider the case of large $n$ which has not been covered in the literature. We have found that the formulation is not adequate for two reasons: (i) the integrand may decay very slowly if $z$ is small; (ii) the integrand may have increasing oscillating magnitude as $n$ increases if $\rho>z$. As a matter of fact, the asymptotic formula and $$J_m(z)\sim \sqrt{\frac{2}{\pi z}}\cos\Big(z-\frac{m\pi}{2}-\frac{\pi}{4}\Big),\quad z\rightarrow\infty,$$ imply that the integrand in has an asymptotic form $$J_{m}(k_{\rho}\rho)\frac{(k_{\rho}S)^{n}e^{-k_{\rho}z}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})}{\sqrt{(n+m)!(n-m)!}}\sim \sqrt{\frac{2}{\pi}} C_{\ell\ell'}^{\mathfrak{ab}}\cos\Big(k_{\rho}\rho-\frac{m\pi}{2}-\frac{\pi}{4}\Big)\frac{(k_{\rho}\rho)^{n-\frac{1}{2}}S^ne^{-k_{\rho}(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})}}{\sqrt{(n+m)!(n-m)!}},$$ as $k_{\rho}\rightarrow\infty$. Given $\rho, z> 0$, define $$\label{gnmdefinition}
g_{nm}(k_{\rho}; \rho, z+\zeta_{\ell\ell'}^{\mathfrak{ab}})=\frac{(k_{\rho}\rho)^{n-\frac{1}{2}}S^ne^{-k_{\rho}(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})}}{\sqrt{(n+m)!(n-m)!}},$$ which has a maximum value $$\max\limits_{k_{\rho}\geq 0}g_{nm}(k_{\rho}; \rho, z+\zeta_{\ell\ell'}^{\mathfrak{ab}})=\frac{S^n}{\sqrt{(n+m)!(n-m)!}}\Big(\frac{2n-1}{2}\Big)^{n-\frac{1}{2}}\Big(\frac{\rho}{z+\zeta_{\ell\ell'}^{\mathfrak{ab}}}\Big)^{n-\frac{1}{2}}e^{\frac{1}{2}-n},$$ at $k_{\rho}=\frac{n}{z+\zeta_{\ell\ell'}^{\mathfrak{ab}}}-\frac{1}{2(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})}$ for $n\geq 1$. Applying Stirling formula $n!\sim \sqrt{2\pi n}n^n/e^n$ yields $$\label{miximumvalue}
\max\limits_{k_{\rho}\geq 0}g_{nm}(k_{\rho}; \rho, z+\zeta_{\ell\ell'}^{\mathfrak{ab}})\sim \sqrt{\frac{(2n-1)e}{2}}\frac{n!}{\sqrt{(n+m)!(n-m)!}}\Big(\frac{\rho}{z+\zeta_{\ell\ell'}^{\mathfrak{ab}}}\Big)^{n-\frac{1}{2}}S^n.$$ Considering the case $m=0$ and setting $S=\sqrt{\rho^2+z^2}$, we have $$\begin{split}
\max\limits_{k_{\rho}\geq 0}g_{n0}(k_{\rho}, \rho, z+\zeta_{\ell\ell'}^{\mathfrak{ab}})&\sim \sqrt{\frac{(2n-1)(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})e}{2\rho}}\Big(\frac{\rho S}{z+\zeta_{\ell\ell'}^{\mathfrak{ab}}}\Big)^{n}\\
&\geq\sqrt{\frac{(2n-1)(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})e}{2\rho}}\Big(\frac{\rho^2}{z+\zeta_{\ell\ell'}^{\mathfrak{ab}}}\Big)^{n}, \quad {\rm if}\;\;\rho>z+\zeta_{\ell\ell'}^{\mathfrak{ab}}.
\end{split}$$ From the above estimate, we can see that the formulation have very large cancellations in the integrand if $\rho/(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})$ and $n$ are large, see Fig. \[integrand\] (a) for an example. Therefore, simply applying a quadrature along the real axis will not be efficient.
To handle the case $\rho>(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})$, we change the contour to the imaginary axis as follows. We first reformulate the integral as $$\begin{split}
S^nI_{nm}^{\mathfrak{ab}}(\rho, z)=&\frac{1}{2}\int_{0}^{\infty}H_{m}^{(1)}(k_{\rho}\rho)\frac{(Sk_{\rho})^{n}e^{-k_{\rho}z}}{\sqrt{(n+m)!(n-m)!}}\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho}) dk_{\rho}\\
+&\frac{(-1)^{m+1}}{2}\int_{-\infty}^{0}H_{m}^{(1)}(k_{\rho}\rho)\frac{(-Sk_{\rho})^{n}e^{k_{\rho}z}}{\sqrt{(n+m)!(n-m)!}}\sigma_{\ell\ell'}^{\mathfrak{ab}}(-k_{\rho}) dk_{\rho},
\end{split}$$ by using identities $$J_m(z)=\frac{H_m^{(1)}(z)+H_m^{(2)}(z)}{2},\quad H_m^{(2)}(-x)=(-1)^{m+1}H_m^{(1)}(x).$$ As the density function $\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})$ is analytic in the right half complex plane, we can change the contour from the real axis to the one which wraps the positive imaginary axis to obtain $$\label{uniformintegralalongimax}
\begin{split}
S^nI_{nm}^{\mathfrak{ab}}(\rho, z)=&\frac{\ri}{2}\int_{0}^{\infty}H_{m}^{(1)}(\ri \eta\rho)\frac{(\ri\eta S)^{n}e^{-\ri\eta z}}{\sqrt{(n+m)!(n-m)!}}\sigma_{\ell\ell'}^{\mathfrak{ab}}(\ri\eta) d\eta\\
-&\frac{\ri}{2}\int_{0}^{\infty}H_{m}^{(1)}(\ri\eta\rho)\frac{(-1)^{m+1}(-\ri\eta S)^{n}e^{\ri\eta z}}{\sqrt{(n+m)!(n-m)!}}\sigma_{\ell\ell'}^{\mathfrak{ab}}(-\ri\eta) d\eta.
\end{split}$$ Then, a substitution of the identity (cf. [@Olver2010 Eq. (10.27.8)]) $$H_m^{(1)}(\ri z)=\frac{2\ri^{-m-1}}{\pi}K_m(z),\quad -\pi\leq\arg z\leq\frac{\pi}{2},$$ into gives $$\label{contourchange}
S^nI_{nm}^{\mathfrak{ab}}(\rho, z)=\frac{\ri^{n-m}}{\pi}\int_{0}^{\infty}K_m(\eta\rho)(S\eta)^n\frac{e^{-\ri\eta z}\sigma_{\ell\ell'}^{\mathfrak{ab}}(\ri\eta)+(-1)^{n+m}e^{\ri\eta z}\sigma_{\ell\ell'}^{\mathfrak{ab}}(-\ri\eta)}{\sqrt{(n+m)!(n-m)!}} d\eta.$$ According to the expressions given in -, all decaying terms in $\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})$ become bounded oscillating terms in $\sigma_{\ell\ell'}^{\mathfrak{ab}}(\pm \ri\eta)$. By the asymptotic formulation [@Olver2010 Eq. (10.25.3)]: $$K_m(z)\sim \sqrt{\frac{\pi}{2z}}e^{-z},\quad z\rightarrow\infty, \quad |\arg z|<\frac{3\pi}{2},$$ and the definition of $g_{nm}(k_{\rho}; \rho,z)$ in , the main part of the integrand has an asymptotic expression $$\frac{K_m(\eta\rho)(S\eta)^n}{\sqrt{(n+m)!(n-m)!}}\sim \sqrt{\frac{\pi}{2\eta\rho}}\frac{(S\eta)^ne^{-\eta\rho}}{\sqrt{(n+m)!(n-m)!}}= \sqrt{\frac{\pi}{2\rho}}g_{nm}(\eta; 1, \rho),\quad\eta\rightarrow\infty.$$ Recalling to get $$\max\limits_{\eta\geq 0}g_{nm}(\eta; 1, \rho)\sim\sqrt{\frac{(2n-1)e}{2}}\frac{n!}{\sqrt{(n+m)!(n-m)!}}\Big(\frac{1}{\rho}\Big)^{n-\frac{1}{2}}S^n,\quad \eta\rightarrow\infty.$$ As an example, we consider the case $m=0$ and set $S=\sqrt{\rho^2+z^2}$ again, i.e., $$\max\limits_{\eta\geq 0}g_{n0}(\eta; 1, \rho)\sim\sqrt{\frac{(2n-1)e}{2}}\Big(\frac{1}{\rho}\Big)^{n-\frac{1}{2}}S^n= \sqrt{\frac{(2n-1)e}{2}}\Big(1+\frac{z^2}{\rho^2}\Big)^{\frac{n}{2}}\sqrt{\rho}.$$ Apparently, the large cancellation in the case $\rho>z+\zeta_{\ell\ell'}^{\mathfrak{ab}}$ can be significantly suppressed by using the formulation . At the same time, the oscillating term $J_m(k_{\rho}\rho)$ is turned to be exponential decaying function $K_m(\eta\rho)$ and thus produce much fast decay when $\rho/(z+\zeta_{\ell\ell'}^{\mathfrak{ab}})$ is large. A comparison of the integrands along real and imaginary axises is plotted in Fig. \[integrand\].
To end this section, we will give some numerical results to show the accuracy and efficiency of the algorithm using mixed DE-SE quadrature together with formulations and for the computation of the Sommerfeld type integrals. We test the integral with densities $\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})\equiv 1$ as the asymptotic formula implies that $\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})$ tends to be either the constant $C_{\ell\ell'}^{\mathfrak{ab}}$ or $0$ rapidly as $k_{\rho}\rightarrow\infty$. Letting $S=r:=\sqrt{\rho^2+z^2}$, then the identity yields $$r^nI_{nm}^{\mathfrak{ab}}(\rho, z)=\sqrt{\frac{4\pi}{2n+1}}\frac{1}{r}\widehat{P}_{n}^m\Big(\frac{z}{r}\Big).$$ We fix $z=0.001$ and test $\rho=0.0005, 0.01, 0.1$ by using two different quadratures: (i) the composite Gaussian quadrature applied to the integral ; (ii) the mixed DE-SE quadrature applied to and for $\rho\leq z$ and $\rho>z$, respectively. For the composite
Gaussian quadrature, the asymptotic formula is used to determine the truncation points such that the magnitude of the integrand decays to smaller than $1.0e-15$. Then, a uniform mesh with mesh size equal to $2$ and $30$ Gauss points in each interval is used to achieve machine accuracy in regular case. Due to the small value of $z$, a very large truncation is needed if the formulation is used. The results are compared in Table. \[Table:gpofapproximationrp\]. We can see that the truncation is larger than 47834 in the case $\rho=0.0005$, $n=5$ and $m=0,1$. The truncation in all other tested cases is even larger. Thus, a large number of quadrature points have been used to achieve good accuracy if the composite Gauss quadrature is applied to . In contrast, the mixed DE-SE quadrature can obtain machine accuracy using no more than 100 points. Moreover, as the ratio $\rho/z$ increases, applying composite Gauss quadrature to can not give correct values due to the large cancellation in . Instead, the mixed DE-SE quadrature applied to can provide results with machine accuracy using a few quadrature points.
Apparently, the technique of using pre-computed tables together with polynomial interpolation can still be applied for efficient computation of the initial values $ \{I_{n 0}^{\mathfrak{ab}}(\rho, z)\}_{n=0}^{2p}$ and $\{ I_{n 1}^{\mathfrak{ab}}(\rho, z)\}_{n=1}^{2p}$ at run time. Then, $4p+1$ tables need to be pre-computed on the 2-D grid in a domain of interest. Efficient improvement by using pre-computed tables is validated by some numerical tests in next section.
Numerical results
=================
In this section, we present numerical results to demonstrate the performance of the proposed FMM. The algorithm is implemented based on an open-source adaptive FMM package DASHMM (cf. [@debuhr2016dashmm]) on a workstation with two Xeon E5-2699 v4 2.2 GHz processors (each has 22 cores) and 500GB RAM using the gcc compiler version 6.3.
We test problems in a three layers medium with interfaces placed at $z_{0}=0$, $z_{1}=-1.2$. Charges are set to be uniformly distributed in irregular domains which are obtained by shifting the domain determined by $r=0.5-a+\frac{a}{8}(35\cos^{4}\theta-30\cos^{2}\theta+3)$ with $a=0.1,0.15,0.05$ to new centers $(0,0,0.6)$, $(0,0,-0.6)$ and $(0,0,-1.8)$, respectively (see Fig. \[fmmconvergence\] (a) for the cross section of the domains). All particles are generated by keeping the uniform distributed charges in a larger cube within corresponding irregular domains. In the layered medium, the material parameters are set to be $\varepsilon_0=21.2$, $\varepsilon_1=47.5$, $\varepsilon_2=62.8$. Let $\widetilde{\Phi}_{\ell}(\boldsymbol{r}_{\ell i})$ be the approximated values of $\Phi_{\ell
}(\boldsymbol{r}_{\ell i})$ calculated by FMM. Define $\ell^{2}$ and maximum errors as $$Err_{2}^{\ell}:=\sqrt{\frac{\sum\limits_{i=1}^{N_{\ell}}|\Phi_{\ell
}(\boldsymbol{r}_{\ell i})-\widetilde{\Phi}_{\ell}(\boldsymbol{r}_{\ell
i})|^{2}}{\sum\limits_{i=1}^{N_{\ell}}|\Phi_{\ell}(\boldsymbol{r}_{\ell
i})|^{2}}},\qquad Err_{max}^{\ell}:=\max\limits_{1\leq i\leq{N_{\ell}}}\frac{|\Phi_{\ell}(\boldsymbol{r}_{\ell i})-\widetilde{\Phi}_{\ell
}(\boldsymbol{r}_{\ell i})|}{|\Phi_{\ell}(\boldsymbol{r}_{\ell i})|}.$$
For accuracy test, we put $N=912+640+1296$ charges in the irregular domains in three layers, see Fig. \[fmmconvergence\] (a). Convergence rates against $p$ are depicted in Fig. \[fmmconvergence\] (b). Clearly, the proposed FMM has spectral convergence with respect to truncation number $p$. The CPU time for the computation of all three free space components $\{\Phi^{free}_{\ell}(\boldsymbol{r}_{\ell i})\}_{\ell=0}^2$ and sixteen reaction components $\Phi^{\mathfrak{ab}}_{\ell\ell'}(\boldsymbol{r}_{\ell i})$ with fixed truncation number $p=5$ are compared in Fig. \[fmmconvergence\] (c) for up to 3 millions charges. It shows that all of them have an $O(N)$ complexity while the CPU time for the computation of reaction components has a much smaller linear scaling constant due to the fact that most of the equivalent polarization sources are well-separated with the targets. CPU time with multiple cores is given in Table \[Table:exthree\] and they show that the speedup of the parallel computing for reaction components is little bit lower than that for the free space components. Here, we only use parallel implementation within the computation of each component. Note that the computation of each component is independent of all other components. Therefore, it is straightforward to implement a version of the code which computes all components in parallel.
Conclusion
==========
In this paper, we have presented a fast multipole method for the efficient calculation of the interactions between charged particles embedded in 3-D layered media. The layered media Green’s function of the Laplace equation is decomposed into a free space and four types of reaction components. The associated equivalent polarization sources are introduced to re-express the reaction components. New MEs and LEs of $O(p^{2})$ terms for the far field of the reaction components and M2L translation operators are derived, accordingly. As a result, the traditional FMM framework can be applied to both the free space and reaction components once the polarization sources are used together with the original sources. The computational cost from the reaction component is only a fraction of that of the FMM for the free space component if a sufficient large number of charges are presented in the problem. Therefore, computing the interactions of many sources in layered media basically costs the same as that for the interactions in the free space.
For the future work, we will carry out error estimate of the FMM for the Laplace equation in 3-D layered media, which requires an error analysis for the new MEs and M2L operators for the reaction components. The application of the FMM in capacitance extraction of interconnects in VLSI will also be considered in a future work.
A stable recursive algorithm for computing reaction densities {#spherehar}
=============================================================
Denote the solution of the problem - in the $\ell$-th layer by $u_{\ell\ell'}({\boldsymbol{r}}, {\boldsymbol{r}}')$ and its partial Fourier transform along $x-$ and $y-$directions by $$\widehat{u}_{\ell\ell^{\prime}}(k_{x},k_{y},z)=\mathscr{F}[u_{\ell\ell'}({\boldsymbol{r}}, {\boldsymbol{r}}')](k_{x},k_{y},z):=\int_{-\infty
}^{\infty}\int_{-\infty}^{\infty}u_{\ell\ell^{\prime}}(\boldsymbol{r},\boldsymbol{r}^{\prime})e^{-\ri(k_{x}x+k_{y}y)}dxdy.$$ Then, $\widehat{u}_{\ell\ell^{\prime}}(k_{x},k_{y},z)$ satisfies second order ordinary differential equation $$\frac{d^2\widehat u_{\ell\ell'}(k_x,k_y, z)}{dz^2}-k_{\rho}^2\widehat u_{\ell\ell'}(k_x,k_y, z)=-e^{-\ri(k_xx'+k_yy')}\delta(z, z'), \quad z\neq d_{\ell}.\label{freqdomhelm}$$ Here, we consider the following general interface conditions $$\label{generalinterfacecond}
a_{\ell-1}\widehat{u}_{\ell-1,\ell^{\prime}}(k_{x},k_{y},z)=a_{\ell}\widehat{u}_{\ell
\ell^{\prime}}(k_{x},k_{y},z),\quad b_{\ell-1}\frac{d\widehat{u}_{\ell
-1,\ell^{\prime}}(k_{x},k_{y},z)}{dz}=b_{\ell}\frac{d\widehat{u}_{\ell
\ell^{\prime}}(k_{x},k_{y},z)}{dz},$$ in the frequency domain for $\ell=1, 2, \cdots, L$, where $\{a_{\ell}, b_{\ell}\}$ are given constants. Apparently, the classic transmission condition will lead to a special case of with $a_{\ell}=1$, $b_{\ell}=\varepsilon_{\ell}$. In the top and bottom-most layers, we also have decaying condition $$\widehat u_{0\ell'}(k_x,k_y, z)\rightarrow 0,\quad \widehat u_{L\ell'}(k_x,k_y, z)\rightarrow 0, \quad{\rm as}\;\; z\rightarrow\pm\infty.$$ This interface problem has a general solution $$\begin{cases}
\displaystyle\widehat{u}_{0\ell^{\prime}}(k_{x},k_{y},z)=\sigma^{1}_{0
\ell^{\prime}}e^{-k_{\rho}(z-d_0)},\\[10pt]\displaystyle\widehat{u}_{\ell^{\prime}\ell^{\prime}}(k_{x},k_{y},z)=\sigma^{1}_{\ell^{\prime}\ell^{\prime}}e^{-k_{\rho}(z-d_{\ell'})}+\sigma^{2}_{\ell^{\prime}\ell^{\prime}}e^{-k_{\rho}(d_{\ell'-1}-z)}+\delta_{\ell\ell'}\widehat{G}(k_x, k_y, z, z^{\prime}),\\[10pt]\displaystyle\widehat{u}_{L\ell^{\prime}}(k_{x},k_{y},z)=\sigma^{2}_{L\ell^{\prime}}e^{-k_{\rho}(d_{L-1}-z)},
\end{cases}
\label{gensolutionformulafreq}$$ where $\delta_{\ell\ell'}$ is the kronecker symbol, and $$\widehat{G}(k_x, k_y, z, z^{\prime})=\vartheta e^{-k_{\rho}|z-z^{\prime}|},\quad \vartheta=\frac{ e^{-\ri(k_{x}x^{\prime}+k_{y}y^{\prime})}}{2k_{\rho}},$$ is the Fourier transform of the free space Green’s function. We will use the decomposition $$\widehat{G}(k_x, k_y, z, z^{\prime})=\widehat{G}^{1}(k_x, k_y, z, z^{\prime})+\widehat{G}^{2}(k_x, k_y, z, z^{\prime}),$$ where the two components are defined as $$\label{freegreenfreq}
\widehat{G}^{1}(k_x, k_y, z, z^{\prime}):=H(z'-z)\vartheta e^{-k_{\rho}(z^{\prime}-z)},\quad\widehat{G}^{2}(k_x, k_y, z, z^{\prime}):=H(z-z')\vartheta e^{-k_{\rho}(z-z^{\prime})},$$ and $H(x)$ is the Heaviside function.
We first consider the $\ell$-th layer without source ($\ell\neq\ell'$), where the right hand side of becomes zero, the solution is given by $$\begin{split}
\widehat u_{\ell\ell'}&(k_x,k_y, z)=\sigma^{1}_{\ell\ell^{\prime}}(k_x, k_y)e^{-k_{\rho}(z-d_{\ell})}+\sigma^{2}_{\ell\ell^{\prime}}(k_x, k_y)e^{-k_{\rho}(d_{\ell-1}-z)}.
\end{split}$$ Applying the interface condition at $z=d_{\ell-1}$ gives $$\label{layertransmission}
\begin{split}
a_{\ell-1}\sigma^{1}_{\ell-1,\ell^{\prime}}+a_{\ell-1}e^{-k_{\rho}D_{\ell-1}}\sigma^{2}_{\ell-1,\ell^{\prime}}=a_{\ell}e^{-k_{\rho}D_{\ell}}\sigma^{1}_{\ell\ell^{\prime}}+a_{\ell}\sigma^{2}_{\ell\ell^{\prime}},\\
b_{\ell-1}\sigma^{1}_{\ell-1,\ell^{\prime}}-b_{\ell-1}e^{-k_{\rho}D_{\ell-1}}\sigma^{2}_{\ell-1,\ell^{\prime}}=b_{\ell}e^{-k_{\rho}D_{\ell}}\sigma^{1}_{\ell\ell^{\prime}}-b_{\ell}\sigma^{2}_{\ell\ell^{\prime}},
\end{split}$$ or in matrix form $$\widehat{\mathbb S}^{(\ell-1)}\begin{pmatrix}
\sigma^{1}_{\ell-1,\ell^{\prime}}\\
\sigma^{2}_{\ell-1,\ell^{\prime}}
\end{pmatrix}=\widetilde{\mathbb S}^{(\ell)}\begin{pmatrix}
\sigma^{1}_{\ell\ell^{\prime}}\\
\sigma^{2}_{\ell\ell^{\prime}}
\end{pmatrix},$$ where $$\widehat{\mathbb S}^{(\ell)}:=\begin{pmatrix}
a_{\ell} & a_{\ell}e_{\ell}\\
b_{\ell} & -b_{\ell}e_{\ell}
\end{pmatrix}, \quad \widetilde{\mathbb S}^{(\ell)}:=\begin{pmatrix}
a_{\ell}e_{\ell} & a_{\ell}\\
b_{\ell}e_{\ell} & -b_{\ell}
\end{pmatrix},\quad \ell=2,3,\cdots, L-1,$$ and $$\label{expdef}
e_{\ell}:=e^{- k_{\rho}D_{\ell}},\quad d_{-1}:=d_0,\quad d_{L+1}:=d_{L},\quad D_{\ell}=d_{\ell-1}-d_{\ell},\quad \ell=0, 1, \cdots, L.$$ Solving the above equations for $\{\sigma^{1}_{\ell-1,\ell'}, \sigma^{2}_{\ell-1,\ell'}\}$, we obtain $$\label{transmissionnosource}
\begin{pmatrix}
\sigma^{1}_{\ell-1,\ell^{\prime}}\\
\sigma^{2}_{\ell-1,\ell^{\prime}}
\end{pmatrix}=\mathbb T^{\ell-1,\ell}\begin{pmatrix}
\sigma^{1}_{\ell\ell^{\prime}}\\
\sigma^{2}_{\ell\ell^{\prime}}
\end{pmatrix}$$ for $\ell= 2, 3,\cdots, L-1$, where $$\label{transmissionmat}
\begin{split}
\mathbb T^{\ell-1,\ell}=&\begin{pmatrix}
a_{\ell-1} & a_{\ell-1}e_{\ell-1}\\
b_{\ell-1} & -b_{\ell-1}e_{\ell-1}
\end{pmatrix}^{-1}\begin{pmatrix}
a_{\ell}e_{\ell} & a_{\ell}\\
b_{\ell}e_{\ell} & -b_{\ell}
\end{pmatrix}
=\frac{1}{2e_{\ell-1}}\begin{pmatrix}
e_{\ell-1} & 0\\
0 & 1
\end{pmatrix}
\widehat{\mathbb T}^{\ell-1,\ell}\begin{pmatrix}
e_{\ell} & 0\\
0 & 1
\end{pmatrix},
\end{split}$$ and $$\widehat{\mathbb T}^{\ell-1,\ell}:=\begin{pmatrix}
\displaystyle\frac{a_{\ell}}{a_{\ell-1}}+\frac{b_{\ell}}{b_{\ell-1}} & \displaystyle\frac{a_{\ell}}{a_{\ell-1}}-\frac{b_{\ell}}{b_{\ell-1}}\\
\displaystyle\frac{a_{\ell}}{a_{\ell-1}}-\frac{b_{\ell}}{b_{\ell-1}} & \displaystyle\frac{a_{\ell}}{a_{\ell-1}}+\frac{b_{\ell}}{b_{\ell-1}}
\end{pmatrix}.$$ For the top and bottom most layers, we have $\sigma^{\downarrow}_{0\ell'}=0$ and $\sigma^{\uparrow}_{L\ell'}=0$, we can also verify that $$\begin{pmatrix}
\sigma^{1}_{0\ell'}\\
0
\end{pmatrix}=\mathbb T^{01}\begin{pmatrix}
\sigma^{1}_{1\ell'}\\
\sigma^{2}_{1\ell'}
\end{pmatrix},\quad \begin{pmatrix}
\sigma^{1}_{L-1,\ell'}\\
\sigma^{2}_{L-1,\ell'}
\end{pmatrix}=\mathbb T^{L-1,L}
\begin{pmatrix}
0\\
\sigma^{2}_{L\ell'}
\end{pmatrix}.$$
Next, we consider the solution in the layer with source ${\boldsymbol{r}}'$ inside, i.e., the solution in the $\ell'$-th layer. The general solution is given by $$\widehat u_{\ell'\ell'}(k_x, k_y, z)=\sigma^{1}_{\ell^{\prime}\ell^{\prime}}e^{\ri k_{\ell' z}(z-d_{\ell'})}+\sigma^{2}_{\ell^{\prime}\ell^{\prime}}e^{\ri k_{\ell' z}(d_{\ell'-1}-z)}+\widehat{G}(k_x, k_y, z, z^{\prime}).$$ At the interfaces $z=d_{\ell'-1}$ and $z=d_{\ell'}$, the interface conditions lead to equations $$\label{sourcelayerupper}
\begin{split}
&a_{\ell'-1}\big(\sigma^{1}_{\ell'-1,\ell^{\prime}}+e_{\ell'-1}\sigma^{2}_{\ell'-1,\ell^{\prime}}\big)=a_{\ell'}\big(e_{\ell'}\sigma^{1}_{\ell'\ell^{\prime}}+\sigma^{2}_{\ell'\ell^{\prime}}+\widehat{G}^{2}(k_x, k_y, d_{\ell'-1}, z')\big),\\
&b_{\ell'-1}\big(\sigma^{1}_{\ell'-1,\ell^{\prime}}-e_{\ell'-1}\sigma^{2}_{\ell'-1,\ell^{\prime}}\big)=b_{\ell'}\big(e_{\ell'}\sigma^{1}_{\ell'\ell^{\prime}}-\sigma^{2}_{\ell'\ell^{\prime}}\big)-\frac{ b_{\ell'}}{k_{\rho}}\partial_z\widehat{G}^{2}(k_x, k_y, d_{\ell'-1}, z'),
\end{split}$$ and $$\label{sourcelayerbottom}
\begin{split}
&a_{\ell'}\big(\sigma^{1}_{\ell'\ell^{\prime}}+e_{\ell'}\sigma^{2}_{\ell'\ell^{\prime}}\big)=a_{\ell'+1}\big(e_{\ell'+1}\sigma^{1}_{\ell'+1\ell^{\prime}}+\sigma^{2}_{\ell'+1,\ell^{\prime}}\big)-a_{\ell'}\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z'),\\
&b_{\ell'}\big(\sigma^{1}_{\ell'\ell^{\prime}}-e_{\ell'}\sigma^{2}_{\ell'\ell^{\prime}}\big)=b_{\ell'+1}\big(e_{\ell'+1}\sigma^{1}_{\ell'+1\ell^{\prime}}-\sigma^{2}_{\ell'+1,\ell^{\prime}}\big)+\frac{ b_{\ell'}}{k_{\rho}} \partial_z\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z').
\end{split}$$ Note that $$\partial_z\widehat{G}^{2}(k_x, k_y, d_{\ell'-1}, z')=- k_{\rho}\widehat{G}^{2}(k_x, k_y, d_{\ell'-1}, z'),\quad \partial_z\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z')= k_{\rho}\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z').$$ Then, equations - can be reformulated as $$\label{downarrowsource}
\begin{pmatrix}
\sigma^{1}_{\ell'-1,\ell^{\prime}}\\
\sigma^{2}_{\ell'-1,\ell^{\prime}}
\end{pmatrix}=\mathbb T^{\ell'-1,\ell'}\begin{pmatrix}
\sigma^{1}_{\ell'\ell^{\prime}}\\
\sigma^{2}_{\ell'\ell^{\prime}}
\end{pmatrix}+\breve{\mathbb S}^{(\ell'-1)}\begin{pmatrix}
a_{\ell'}\\
b_{\ell'}
\end{pmatrix}\widehat{G}^{2}(k_x, k_y,d_{\ell'-1}, z')$$ and $$\label{uparrowsource}
\begin{pmatrix}
\sigma^{1}_{\ell'\ell^{\prime}}\\
\sigma^{2}_{\ell'\ell^{\prime}}
\end{pmatrix}=\mathbb T^{\ell'\ell'+1}\begin{pmatrix}
\sigma^{1}_{\ell'+1,\ell^{\prime}}\\
\sigma^{2}_{\ell'+1,\ell^{\prime}}
\end{pmatrix}+\breve{\mathbb S}^{(\ell')}\begin{pmatrix}
-a_{\ell'}\\
b_{\ell'}
\end{pmatrix}\widehat{G}^{1}(k_x, k_y,d_{\ell'}, z'),$$ where $$\breve{\mathbb S}^{(\ell)}=\big(\widehat{\mathbb S}^{(\ell)}\big)^{-1}=\frac{1}{2}\begin{pmatrix}
1 & 0\\
0 & e_{\ell}^{-1}
\end{pmatrix}\begin{pmatrix}
\displaystyle\frac{1}{a_{\ell}} & \displaystyle\frac{1}{b_{\ell}}\\
\displaystyle\frac{1}{a_{\ell}} & \displaystyle-\frac{1}{b_{\ell}}
\end{pmatrix}:=\begin{pmatrix}
\breve S_{11}^{(\ell)} & \breve S_{12}^{(\ell)}\\
\breve S_{21}^{(\ell)} & \breve S_{22}^{(\ell)}
\end{pmatrix}.$$ Define $$\label{regularmat}
\widetilde{\mathbb T}^{\ell-1,\ell}=2e_{\ell-1}\mathbb T^{\ell-1,\ell},\quad C^{(\ell)}=\prod\limits_{j=0}^{\ell-1}\frac{1}{2e_j},\quad \mathbb A^{(\ell)}=\widetilde{\mathbb T}^{01}\widetilde{\mathbb T}^{12}\cdots \widetilde{\mathbb T}^{\ell-1,\ell}:=\begin{pmatrix}
\alpha_{11}^{(\ell)} & \alpha_{12}^{(\ell)}\\
\alpha_{21}^{(\ell)} & \alpha_{22}^{(\ell)}
\end{pmatrix},$$ for $\ell=1, 2,\cdots, L$. Then, recursions in , and result in the system $$\label{Lsystem}
\begin{split}
\begin{pmatrix}
\sigma^{1}_{0\ell'}\\
0
\end{pmatrix}=&C^{(L)}\mathbb A^{(L)}\begin{pmatrix}
0\\
\sigma^{2}_{L\ell'}
\end{pmatrix}+C^{(\ell'-1)}\mathbb A^{(\ell'-1)}\breve{\mathbb S}^{(\ell'-1)}\begin{pmatrix}
a_{\ell'}\\
b_{\ell'}
\end{pmatrix}\widehat{G}^{2}(k_x, k_y, d_{\ell'-1}, z')\\
&+C^{(\ell')}\mathbb A^{(\ell')}\breve{\mathbb S}^{(\ell')}\begin{pmatrix}
-a_{\ell'}\\
b_{\ell'}
\end{pmatrix}\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z').
\end{split}$$
It is not numerically stable to directly solve for $\sigma_{0\ell'}^{1}$ and $\sigma_{L\ell'}^{2}$ then apply recursions , and to obtain all other reaction densities due to the exponential functions involved in the formulations. According to the expression , the recursions , and are stable for the computation of the components $\sigma_{\ell\ell'}^{1}(k_{\rho})$. As for the computation of the components $\sigma_{\ell\ell'}^{2}(k_{\rho})$, we need to form linear systems similar as using recursions , and and then solve it.
We first solve the second equation in to get $$\sigma^{2}_{L\ell'}=\sigma_{L\ell'}^{21}\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z')+\sigma_{L\ell'}^{22}\widehat{G}^{2}(k_{\ell'z}, d_{\ell'-1}, z'),$$ where $$\label{bottommostdensity}
\begin{split}
&\sigma_{L\ell'}^{21}=-\frac{C^{(\ell'+1)}}{C^{(L)}\alpha_{22}^{(L)}}\begin{pmatrix}
\alpha^{(\ell')}_{21} & \alpha^{(\ell')}_{22}
\end{pmatrix}2e_{\ell'}\breve{\mathbb S}^{(\ell')}\begin{pmatrix}
-a_{\ell'}\\
b_{\ell'}
\end{pmatrix},\quad 0\leq\ell'<L,\\
&\sigma_{L\ell'}^{22}=-\frac{C^{(\ell')}}{C^{(L)}\alpha_{22}^{(L)}}\begin{pmatrix}
\alpha^{(\ell'-1)}_{21} & \alpha^{(\ell'-1)}_{22}
\end{pmatrix}2e_{\ell'-1}\breve{\mathbb S}^{(\ell'-1)}\begin{pmatrix}
a_{\ell'}\\
b_{\ell'}
\end{pmatrix},\quad 0<\ell'\leq L.
\end{split}$$ According to the recursion , and , all other reaction densities also have decompositions $$\label{desitydecomposition}
\begin{split}
\sigma_{\ell\ell'}^{1}=\sigma_{\ell\ell'}^{11}\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z')+\sigma_{\ell\ell'}^{12}\widehat{G}^{2}(k_x, k_y, d_{\ell'-1}, z'),\\ \sigma_{\ell\ell'}^{2}=\sigma_{\ell\ell'}^{21}\widehat{G}^{1}(k_x, k_y, d_{\ell'}, z')+\sigma_{\ell\ell'}^{22}\widehat{G}^{2}(k_x, k_y, d_{\ell'-1}, z').
\end{split}$$ For each $0\leq\ell<L$, we first calculate $\{\sigma_{\ell\ell'}^{11}, \sigma_{\ell\ell'}^{12}\}$ by using one of the recursions , and , then formulate a linear system for $\{\sigma_{0\ell'}^{1},\sigma_{\ell\ell'}^{2}\}$ as the linear system . Next, we solve the second equation in the linear system to obtain reaction densities $\{\sigma_{\ell\ell'}^{21},\sigma_{\ell\ell'}^{22}\}$. In summary, the formulations are given as follows: $$\begin{aligned}
&\sigma_{\ell\ell'}^{11}=\begin{cases}
\displaystyle T^{\ell'\ell'+1}_{11}\sigma_{\ell'+1,\ell'}^{11}+T^{\ell'\ell'+1}_{12}\sigma_{\ell'+1,\ell'}^{21}-\breve S_{11}^{(\ell')}a_{\ell'}+\breve S_{12}^{(\ell')}b_{\ell'},& \ell=\ell',\\
\displaystyle T^{\ell\ell+1}_{11}\sigma_{\ell+1,\ell'}^{11}+T^{\ell\ell+1}_{12}\sigma_{\ell+1,\ell'}^{21}, & {\rm else},
\end{cases}\\
&\sigma_{\ell\ell'}^{12}=\begin{cases}
\displaystyle T^{\ell'-1,\ell'}_{11}\sigma_{\ell'\ell'}^{12}+T^{\ell'-1,\ell'}_{12}\sigma_{\ell'\ell'}^{22}+\breve S_{11}^{(\ell'-1)}a_{\ell'}+\breve S_{12}^{(\ell'-1)}b_{\ell'},& \ell=\ell'-1,\\
\displaystyle T^{\ell\ell+1}_{11}\sigma_{\ell+1,\ell'}^{12}+T^{\ell\ell+1}_{12}\sigma_{\ell+1,\ell'}^{22}, & {\rm else},
\end{cases}\\
&\sigma_{\ell\ell'}^{21}=\begin{cases}
\displaystyle-\frac{1}{\alpha_{22}^{(\ell)}}\begin{pmatrix}
0 & 1
\end{pmatrix}\left[\frac{C^{(\ell'+1)}}{C^{(\ell)}}\mathbb A^{(\ell')}2e_{\ell'}\breve{\mathbb S}^{(\ell')}\begin{pmatrix}
-a_{\ell'}\\
b_{\ell'}
\end{pmatrix}+\mathbb A^{(\ell)}\begin{pmatrix}
\sigma_{\ell\ell'}^{11}\\
0
\end{pmatrix}\right],& \ell>\ell',\\[10pt]
\displaystyle-\frac{\alpha_{21}^{(\ell)}}{\alpha_{22}^{(\ell)}}\sigma_{\ell\ell'}^{11}, & {\rm else},
\end{cases}\\
&\sigma_{\ell\ell'}^{22}=\begin{cases}
\displaystyle-\frac{1}{\alpha_{22}^{(\ell)}}\begin{pmatrix}
0 & 1
\end{pmatrix}\left[\frac{C^{(\ell')}}{C^{(\ell)}}\mathbb A^{(\ell'-1)}2e_{\ell'-1}\breve{\mathbb S}^{(\ell'-1)}\begin{pmatrix}
a_{\ell'}\\
b_{\ell'}
\end{pmatrix}+\mathbb A^{(\ell)}\begin{pmatrix}
\sigma_{\ell\ell'}^{12}\\
0
\end{pmatrix}\right],& \ell\geq \ell',\\[10pt]
\displaystyle-\frac{\alpha_{21}^{(\ell)}}{\alpha_{22}^{(\ell)}}\sigma_{\ell\ell'}^{12}, & {\rm else}.
\end{cases}\label{downdowndensity}\end{aligned}$$ Substituting and into and taking inverse Fourier transform, we obtain expressions -.
From the definition and , we have $$\begin{split}
&T^{\ell\ell+1}_{11}=\frac{a_{\ell+1}b_{\ell}+a_{\ell}b_{\ell+1}}{2a_{\ell}b_{\ell}}e_{\ell+1},\quad T^{\ell\ell+1}_{12}= \frac{a_{\ell+1}b_{\ell}-a_{\ell}b_{\ell+1}}{2a_{\ell}b_{\ell}},\\
&2e_{\ell}\breve{\mathbb S}^{(\ell)}=\begin{pmatrix}
\displaystyle a_{\ell}^{-1}e_{\ell} & \displaystyle b_{\ell}^{-1}e_{\ell}\\
\displaystyle a_{\ell}^{-1} & \displaystyle-b_{\ell}^{-1}
\end{pmatrix},\quad
\frac{C^{(\ell_1)}}{C^{(\ell_2)}}=\begin{cases}
1 & \ell_1=\ell_2,\\
2^{\ell_2-\ell_1}e^{-k_{\rho}(d_{\ell_1-1}-d_{\ell_2-1})} & 0\leq\ell_1<\ell_2,
\end{cases}
\end{split}$$ and an asymptotic behavior $$\mathbb A^{(\ell)}\sim\begin{pmatrix}
\tilde\alpha_{11}^{(\ell)}e_0e_1\cdots e_{\ell} & \tilde\alpha_{12}^{(\ell)}e_0\\
\tilde\alpha_{21}^{(\ell)}e_{\ell} & \tilde\alpha_{22}^{(\ell)}
\end{pmatrix}, \quad k_{\rho}\rightarrow\infty,$$ where $\{\tilde\alpha_{11}^{(\ell)}, \tilde\alpha_{12}^{(\ell)}, \tilde\alpha_{21}^{(\ell)}, \tilde\alpha_{22}^{(\ell)}\}$ are constants independent of $k_{\rho}$. By using these formulations in -, we can show that all reaction densities $\{\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})\}_{\mathfrak{a,b}=1}^2$ have an asymptotic behavior $$\label{densityasymptotic}
\sigma_{\ell\ell'}^{\mathfrak{ab}}(k_{\rho})\sim C_{\ell\ell'}^{\mathfrak{ab}}e^{-k_{\rho}\zeta_{\ell\ell'}^{\mathfrak{ab}}},\quad k_{\rho}\rightarrow\infty,$$ where $C_{\ell\ell'}^{\mathfrak{ab}}$ and $\zeta_{\ell\ell'}^{\mathfrak{ab}}$ are constants independent of $k_{\rho}$. For example, we have $$\begin{split}
&\sigma_{L\ell'}^{21}(k_{\rho})\sim 2^{L-\ell'-1}\frac{\tilde\alpha_{22}^{(\ell')}}{\alpha_{22}^{(L)}}e^{-k_{\rho}(d_{\ell'}-d_{L-1})},\quad k_{\rho}\rightarrow\infty,\\
&\sigma_{L\ell'}^{22}(k_{\rho})\sim 2^{L-\ell'}\frac{\alpha_{22}^{(\ell')}}{\alpha_{22}^{(L)}}\Big(\frac{a_{\ell'}}{a_{\ell'-1}}+\frac{b_{\ell'}}{b_{\ell'-1}}\Big)e^{-k_{\rho}(d_{\ell'-1}-d_{L-1})},\quad k_{\rho}\rightarrow\infty.
\end{split}$$
If the number of layers is not large, we are able to write down explicit expressions of the reaction densities. Here, we give expressions for the case of a three layers media with $a_{\ell}=1$, $b_{\ell}=\varepsilon_{\ell}$ as an example.
- Source in the top layer: $$\label{densitythreelayer1}
\begin{split}
\sigma_{00}^{11}(k_{\rho})=&\frac{(\varepsilon_0-\varepsilon_1)(\varepsilon_1+\varepsilon_2)+(\varepsilon_0+\varepsilon_1)(\varepsilon_1-\varepsilon_2)e^{2d_1k_{\rho}}}{2\kappa(k_{\rho})},\\
\sigma_{10}^{21}(k_{\rho})=&\frac{\varepsilon_0(\varepsilon_1+\varepsilon_2)}{\kappa(k_{\rho})},\quad
\sigma_{10}^{11}(k_{\rho})=\frac{\varepsilon_0(\varepsilon_1-\varepsilon_2)e^{d_1k_{\rho}}}{\kappa(k_{\rho})},\quad
\sigma_{20}^{21}(k_{\rho})=\frac{2 \varepsilon_0\varepsilon_1 e^{d_1k_{\rho}}}{\kappa(k_{\rho})}.
\end{split}$$
- Source in the middle layer: $$\label{densitythreelayer2}
\begin{split}
\sigma_{01}^{12}(k_{\rho})=&\frac{\varepsilon_1(\varepsilon_1+\varepsilon_2) }{\kappa(k_{\rho})},\quad
\sigma_{01}^{11}(k_{\rho})=\frac{\varepsilon_1(\varepsilon_1-\varepsilon_2)e^{d_1k_{\rho}} }{\kappa(k_{\rho})},\\
\sigma_{11}^{11}(k_{\rho})=&\frac{(\varepsilon_1-\varepsilon_2)(\varepsilon_1+\varepsilon_0)}{2\kappa(k_{\rho})},\quad
\sigma_{11}^{21}(k_{\rho})=\frac{(\varepsilon_1-\varepsilon_2)(\varepsilon_1-\varepsilon_0)e^{d_1k_{\rho}}}{2\kappa(k_{\rho})},\\
\sigma_{11}^{12}(k_{\rho})=&\frac{(\varepsilon_1-\varepsilon_2)(\varepsilon_1-\varepsilon_0)e^{d_1k_{\rho}}}{2\kappa(k_{\rho})},\quad\sigma_{11}^{22}(k_{\rho})=\frac{(\varepsilon_1+\varepsilon_2)(\varepsilon_1-\varepsilon_0)}{2\kappa(k_{\rho})},\\
\sigma_{21}^{22}(k_{\rho})=&\frac{\varepsilon_1(\varepsilon_1-\varepsilon_0) e^{d_1k_{\rho}}}{\kappa(k_{\rho})},\quad
\sigma_{21}^{21}(k_{\rho})=\frac{\varepsilon_1(\varepsilon_0+\varepsilon_1) }{\kappa(k_{\rho})}.
\end{split}$$
- Source in the bottom layer: $$\label{densitythreelayer3}
\begin{split}
\sigma_{02}^{12}(k_{\rho})=&\frac{2 \varepsilon_1\varepsilon_2 e^{d_1k_{\rho}}}{\kappa(k_{\rho})},\quad
\sigma_{12}^{22}(k_{\rho})=\frac{\varepsilon_2(\varepsilon_1-\varepsilon_0)e^{d_1k_{\rho}}}{\kappa(k_{\rho})},\quad
\sigma_{12}^{12}(k_{\rho})=\frac{\varepsilon_2(\varepsilon_0+\varepsilon_1)}{\kappa(k_{\rho})},\\
\sigma_{22}^{22}(k_{\rho})=&\frac{(\varepsilon_1-\varepsilon_0)(\varepsilon_1+\varepsilon_2)+(\varepsilon_0+\varepsilon_1)(\varepsilon_2-\varepsilon_1)e^{2d_1k_{\rho}} }{2\kappa(k_{\rho})},
\end{split}$$
where $$\kappa(k_{\rho})=\frac{1}{2}\big[(\varepsilon_0+\varepsilon_1)(\varepsilon_1+\varepsilon_2)+(\varepsilon_0-\varepsilon_1)(\varepsilon_2-\varepsilon_1)e^{2d_1k_{\rho}}\big].$$ Apparently, these expressions also verify our conclusion on the asymptotic behavior of the reaction densities.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work was supported by US Army Research Office (Grant No. W911NF-17-1-0368) and US National Science Foundation (Grant No. DMS-1764187). The research of the first author is partially supported by NSFC (grant 11771137), the Construct Program of the Key Discipline in Hunan Province and a Scientific Research Fund of Hunan Provincial Education Department (No. 16B154).
References {#references .unnumbered}
==========
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
A. Alparslan, M. I. Aksun, K. A. Michalski, Closed-form [G]{}reen’s functions in planar layered media for all ranges and materials, IEEE Trans. Microw. Theory Tech. 58 (3) (2010) 602–613.
M. I. Aksun, A robust approach for the derivation of closed-form [G]{}reen’s functions, IEEE Trans. Microw. Theory Tech. 44 (5) (1996) 651–658.
I. Babu?ka, G. Caloz, J. E. Osborn, Special finite element methods for a class of second order elliptic problems with rough coefficients. SIAM J. Numer. Anal. 31 (4) (1994) 945-81.
L. Borcea, Electrical impedance tomography, Inverse Probl. 18 (6) (2002): R99–R136.
W. Cai, Computational Methods for Electromagnetic Phenomena: electrostatics in solvation, scattering, and electron transport, Cambridge University Press, New York, NY, 2013.
S. L. Campbell, I. C. Ipsen, C. T. Kelley, C. D. Meyer, [GMRES]{} and the minimal polynomial, [BIT]{} Numer. Math. 36 (4) (1996) 664–675.
M. H. Cho, J. F. Huang, D. X. Chen, W. Cai, A heterogeneous fmm for layered media Helmholtz equation I: Two layers in $R^2$, J. Comput. Phys. 369 (2018) 237–251.
Y. L. Chow, J. J. Yang, D. G. Fang, G. E. Howard, A closed-form spatial [G]{}reen’s function for the thick microstrip substrate, IEEE Trans. Microw. Theory Tech. 39 (3) (1991) 588–592.
J. DeBuhr, B. Zhang, A. Tsueda, V. Tilstra-Smith, T. Sterling, Dashmm: [D]{}ynamic adaptive system for hierarchical multipole methods, Commun. Comput. Phys. 20 (4) (2016) 1106–1126.
M. A. Epton, B. Dembart, Multipole translation theory for the three-dimensional [L]{}aplace and [H]{}elmholtz equations, SIAM J Sci. Comput. 16 (4) (1995) 865–897.
N. Geng, A. Sullivan, L. Carin, Fast multipole method for scattering from an arbitrary [PEC]{} target above or buried in a lossy half space, IEEE Trans. Antennas Propag. 49 (5) (2001) 740–748.
L. Greengard, V. Rokhlin, A fast algorithm for particle simulations, J. Comput. phys. 73 (2) (1987) 325–348.
L. Grengard, V. Rokhlin, The rapid evaluation of potential fields in three dimensions, in: Research Report YALEU/DCS/RR-515, Dept. of Comp. Sci., Yale University, New Haven, CT, Springer, 1987.
L. Greengard, V. Rokhlin, A new version of the fast multipole method for the Laplace equation in three dimensions, Acta Numer. 6 (1997) 229–269.
L. Gurel, M. I. Aksun, Electromagnetic scattering solution of conducting strips in layered media using the fast multipole method, IEEE Microw. Guided W. 6 (8) (1996) 277.
V. Jandhyala, E. Michielssen, R. Mittra, Multipole-accelerated capacitance computation for 3-[D]{} structures in a stratified dielectric medium using a closed-form [G]{}reen’s function, Int. J. Microw. Millimet. Wave Comput. Aided Eng. 5 (2) (1995) 68–78.
P. A. Martin, Multiple Scattering: interaction of time-harmonic waves with [N]{} obstacles, no. 107, Cambridge University Press, 2006.
K. Nabors, J. K. White, Fastcap: a multipole accelerated 3-D capacitance extraction program, IEEE Trans. Comput. Aided Des. 10 (11) (1991) 1447–1459.
K. S. Oh, D. Kuznetsov, J. E. Schuttaine, Capacitance computations in a multilayered dielectric medium using closed-form spatial [G]{}reen’s functions, IEEE Trans. Microw. Theory Tech. 42 (8) (1994) 1443–1453.
A. E. Ruehli, P. A. Brennan, Efficient capacitance calculations for three-dimensional multiconductor systems, IEEE Trans. Microw. Theory Tech. 21 (2) (1973) 76–82.
A. Seidl, H. Klose, M. Svoboda, J. Oberndorfer, W. Rosner, CAPCAL-a 3-D capacitance solver for support of CAD systems, IEEE Trans. Comput. Aided Des. 7 (5) (1988) 549–556.
B. Wang, D. Chen, B. Zhang, W. Z. Zhang, M. H. Cho, W. Cai, Taylor expansion based fast multipole method for 3-D Helmholtz equations in layered media, J. Comput. Phys. 401 (2019) 109008.
[ B. Wang, W. Z. Zhang, W. Cai, Fast multipole method for 3-D Helmholtz equation in layered media., SIAM J. Sci. Comput. 41 (6) (2019) A3954-A3981. ]{}
G. Watson, A Treatise of the Theory of Bessel Functions (second edition), Cambridge University Press, Cambridge, UK, 1966.
E. T. Whittaker amd G. N. Watson. A Course of Modern Analysis, Cambridge University Press, 4th edition, 1927.
M. Xu and R. R. Alfano, Fractal mechanisms of light scattering in biological tissue and cells, Opt. Lett. 30 (22) (2005) 3051–3053.
W. Yu, X. Wang, Advanced field-solver techniques for RC extraction of integrated circuits, Springer, 2014.
W. Z. Zhang, B. Wang, and W. Cai. Exponential convergence for multipole expansion and translation to local expansions for sources in layered media: 2-[D]{} acoustic wave. arXiv:1809.07716, to appear in SIAM Numer. Anal., May, 2020.
J. S. Zhao, W. M. Dai, S. Kadur, D. E. Long, Efficient three-dimensional extraction based on static and full-wave layered [G]{}reen’s functions, Proceedings of the 35th Annual Design Automation Conference, (1998) 224–229.
K. A. Michalski, J. R. Mosig, Efficient computation of sommerfeld integral tails–methods and algorithms, J. Electromagn. Waves Appl. 30 (3) (2016) 281–317.
H. Takahasi, M. Mori, Double exponential formulas for numerical integration, Publ. RIMS, Kyoto Univ. 9 (3) (1974) 721–741.
F. W. J. Olver, NIST handbook of mathematical functions, Cambridge University Press, 2010.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Rashba precession of spins moving along a one-subband quantum channel is calculated. The quantitative influence of unoccupied higher subbands depends on the shape of the transversal confinement and can be accounted for perturbatively. Coulomb interactions are included within the Tomonaga–Luttinger model with spin-orbit coupling incorporated. Increasing interaction strength at decreasing carrier density is found to [*enhance*]{} spin precession.'
address: 'Fakultät für Physik, Universität Freiburg, Hermann-Herder-Str. 3, 79104 Freiburg, Germany'
author:
- Wolfgang Häusler
title: Rashba spin splitting in different quantum channels
---
[2]{}
Complete understanding of the Datta-Das spin transistor [@dattadas] requires to know how different physical parameters influence the precession of spins which is caused by the spin-orbit coupling $$\label{hso}
H^{\rm so}=\alpha(\sigma_xp_z-\sigma_zp_x)\;.$$ $H^{\rm so}$ breaks spin rotation invariance proportional to the momentum of spins moving along the ‘active’ $x-z$–plane. For isotropic single particle energy dispersions ${\varepsilon}(|k|)$ in the plane, ignoring Coulomb interactions, this yields the well known spin split bands ${\varepsilon}_{\pm}(|k|)={\varepsilon}(|k|)\pm\alpha|k|$. The Rashba parameter $\alpha$ is proportional to the intrinsic or by means of gates externally applied electric field perpendicular to the layer [@rashba], here taken as the $y$–direction. Spin precession occurs then on the length scale $|k_+-k_-|^{-1}$ of the spin split momenta at the Fermi energy. It is special to the effective mass approximation ${\varepsilon}(|k|)=k^2/2m$, describing many semiconductors, that the Rashba length $2/|k_+-k_-|=(m\alpha)^{-1}$ does [*not*]{} depend on the Fermi energy nor the carrier density.
Coherent precession of many spins down to the spin selective drain of the transistor requires to diminish the directional spread of the electron momenta by introducing quasi one-dimensional (1D) ‘wave guides’ [@dattadas]. To leading order in $\alpha$ the wave guide simply projects the momenta in ${\varepsilon}_{\pm}(|k|)\to{\varepsilon}_{\pm}(k_x)$ onto the $x$–axes, leaving the basic features of the 1D case valid, cf. Fig.\[figurename\]. In particular will the linear kinetic energy dispersion in carbon nanotubes or in narrow gap semiconductors lead to enhanced precession with carrier density.
To higher order $H^{\rm so}$ will mix different transport channels in each wave guide. Up to O$(\alpha^5)$ this effect can be accounted for by a renormalizing $\alpha\to\alpha^*$ in ${\varepsilon}_{\pm}(k)$ and, within the effective mass approximation, by $m\to m^*$. In this latter case the Rashba length modifies according to $(m\alpha)^{-1}\to(m^*\alpha^*)^{-1}$. A quantitative estimate requires the intra-subband eigenfunctions $$\label{psins}
\psi_{kns}(x,z)={\rm e}^{{\rm i}kx}\phi_n(z)
(\cos(m\alpha z)|s\rangle+{\rm i}\,\sin(m\alpha z)|\!-\!\!s\rangle)\;.$$ which are plane waves of momentum $k$ along the wave guide and, without inter-subband scattering, slightly modified subband states $\:\phi_n\:$ (subband index $n$) in $z$–direction of the spin polarization $s=\pm$ on the axes. For a harmonic confinement (subband energy $\omega_0$), as relevant for example in gated samples [@tarucha], a perturbative estimate yields $\:\alpha^*=\alpha(1-\eta)\:$ and $\:m^*=m(1+8\eta^2)\:$ in the ground subband. Here, the dimensionless parameter $\eta=(mw\alpha/2)^2$ compares the wave guide width $w=2/\sqrt{m\omega_0}$ with the Rashba length. For a hard wall confinement on the other hand (again of width $w$), as possibly more relevant for wires fabricated by the cleaved edge technique [@yacoby; @rother], the renormalizations become $\:\alpha^*=\alpha(1-(1/6-1/\pi^2)\eta)\:$ and $\:m^*=m(1+3(4/3\pi)^6\eta^2)\:$, i.e. they are significantly reduced compared to the soft wall case since $1/6-1/\pi^2\approx
0.065$ and $3(4/3\pi)^6/8\approx 0.002$.
=0.8
Changing the gate voltage of the transistor when intending to vary the electrical field and therewith $\alpha$ changes at the same time the carrier density. Without interactions and when ${\varepsilon}(k)=k^2/2m^*$ this would be unimportant. Interactions on the other hand depend sensitively on the carrier density. Regarding interaction effects the Tomonaga-Luttinger (TL) model [@tl] provides the most precise low energy description of 1D metals. Though some of its characteristic power laws can affect spin properties [@egger-epl] we rather focus here on the question how interactions influence the length over which coherent spin rotation takes place.
As a second striking property quantum wires exhibit spin–charge separation which, interestingly and contrary to statements in the literature [@egger-epl; @moroz], is [*not*]{} spoiled unless spin-orbit coupling is not exceedingly strong $\eta\sim
1$ if the effective mass description applies. On the other hand, for non-quadratic dispersion relations, spin charge separation is in general destroyed, similar as is the case with Zeeman splitting [@frahm-aoki]. An example are carbon nanotubes where $v_\pm=v_{\rm F}\pm\alpha$ with $\alpha$ originating in this case from the curved surface [@cylinder] instead of the Rashba mechanism.
How to include the Rashba term in the TL-model ? In previous work [@egger-epl; @moroz] the Fermi velocities $v_+$ and $v_-$ have been set to different values, which for ${\varepsilon}(k)=k^2/2m^*$ does not describe $H^{\rm so}$ as can be seen also in Fig.\[figurename\]. Rather [*both*]{} velocities change slightly but obey $v_+=v_-$ leading immediately to charge–spin separation. Thus, in effective mass systems $H^{\rm so}$ acts solely in the topological spin sector of the corresponding TL low energy model (of length $L$), $$\label{bosontop}
\frac{\pi}{4L}\left(v_{\mbox{\tiny N}}N_{\sigma}^2+
v_{\mbox{\tiny J}}J_{\sigma}^2\right)-m^*\alpha^*v_{\rm F}
J_{\sigma}\;.$$ $N_{\sigma}$ and $J_{\sigma}$ denote the usual currents of velocities $v_{\mbox{\tiny N/J}}$ where the latter both differ from $v_{\rm F}$ to account for the Coulomb repulsion [@tl]. In strictly spin isotropic systems $v_{\mbox{\tiny
N}}=v_{\mbox{\tiny J}}$. Since we expect this isotropy being broken only weakly, both of these velocities should be similar in magnitude and also similar to the spin velocity $v_\sigma$. This latter quantity has been determined recently by extensive quantum Monte–Carlo simulations [@creffield]. With increasing interaction strength, equivalent to a decreasing carrier density, $v_\sigma/v_{\rm F}$ was found to decrease. With parameters for existing quantum wires [@yacoby] $v_\sigma/v_{\rm F}$ can drop below 0.5 .
Many quantities of interest can be calculated exactly using (\[bosontop\]). In particular it can be shown [@hausler] that spins polarized in $x$–direction along the wire precess over a length scale $$m^*\alpha^*\frac{v_{\sigma{\mbox{\tiny J}}}}{v_{\rm F}}\;.$$ With $v_{\mbox{\tiny J}}=v_\sigma$ we see that this length decreases with increasing interaction strength or decreasing particle density. A similar conclusion has been drawn for two-dimensional electrons after treating the interactions perturbatively [@raikh]. This trend is opposite to what is expected for linear single electron dispersions but agrees with experimental observations [@experiments].
[10]{} B. Datta and S. Das, Appl. Phys. Lett. (1990) 665. É.I. Rashba, Fiz. Tverd. Tela [**2**]{} (1960) 1224 \[Sov. Phys. Solid State [**2**]{} (1960) 1109\]. S. Tarucha, T. Honda, and T. Saku, Solid State Comm. [**94**]{} (1995) 413. A. Yacoby, [*et al.*]{}, Phys. Rev. Lett., (1996) 4612. M. Rother, [*et al.*]{}, Physica E [**6**]{} (2000) 551. F.D.M. Haldane, J. Phys. C [**14**]{} (1981) 2585. A. De Martino, R. Egger, Europhys. Lett. (2000) 3464. A.V. Moroz, K.V. Samokhin, C.H.W. Barnes, Phys. Rev. Lett. [**84**]{} (2000) 4164. H. Frahm and V.E. Korepin, Phys. Rev. B [**42**]{} (1990) 10553; T. Kimura, K. Kuroki, and H. Aoki, Phys. Rev. B [**53**]{} (1996) 9572. L.I. Magarill, D.A. Romanov, A.V. Chaplik, JETP [**86**]{} (1998) 771. C.E. Creffield, W. Häusler, and A.H. MacDonald, Europhys. Lett. [**53**]{} (2001) 221; W. Häusler, L. Kecke, this volume. W. Häusler, Phys. Rev. B [**63**]{} (2001) 121310(R). G.-H. Chen and M. E. Raikh, Phys. Rev. B [**60**]{} (1999) 4826. J. Nitta, [*et al.*]{}, Phys. Rev. Lett. [**78**]{} (1997) 1335; G. Engels, [*et al.*]{}, Phys. Rev. B [**55**]{} (1997) 1958(R); C.-M. Hu, [*et al.*]{}, Phys. Rev. B [**60**]{} (1999) 7736.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate disease spreading on eight empirical data sets of human contacts (mostly proximity networks recording who is close to whom, at what time). We compare three levels of representations of these data sets: temporal networks, static networks and a fully connected topology. We notice that the difference between the static and fully-connected networks—with respect to time to extinction and average outbreak size—is smaller than between the temporal and static topologies. This suggests that, for these data sets, temporal structures influence disease spreading more than static network structures. To explain the details in the differences between the representations, we use 32 network measures. This study concur that long-time temporal structures, like the turnover of nodes and links, are the most important for the spreading dynamics.'
author:
- Petter Holme
title: Temporal network structures controlling disease spreading
---
Introduction
============
The spread of infectious disease continues to be one of the major challenges to global health. Despite advances in genomics and access to large data sets, predicting outbreaks is still disturbingly difficult [@difficult]. Even knowing fairly much about an outbreak (at the time of writing the major concern regards the zika virus [@zika]) nobody can be very certain about its future. The basic mathematics of infectious disease outbreaks as emergent phenomena is well studied. No paper, to our knowledge, has strong alternatives to compartmental models—models dividing individuals into classes with respect to the disease and assigning transition rules between the classes. Straightforward implementations of compartmental models do, however, not explain the difficulties in predicting emergent outbreaks in real populations [@jansson]. There can be many reasons for this difficulty to predict the extinction time outbreaks. One obvious reason is that the data quality is still not good enough to make high-precision forecasting. There can however be other, more fundamental issues with how the compartmental models are integrated with models of contact patterns (describing how people meet in such a way that disease can spread). In this paper, we investigate different levels of representing contact structures: as temporal networks (including information both of the time of contact and the individuals involved), as static networks (including informations of pairs of people between which the disease can spread), and fully-connected networks (which is the traditional contact structure of theoretical epidemiology [@hethcote]).
Many studies have pointed out that to model disease spreading accurately, we need to understand both static networks structures [@keeling; @vespi_rev] and temporal-network structures [@masuda_holme_rev]. To make this point, a standard approach has been to first observe some structure in empirical data, then use models to prove this structure affects disease spreading, and finally conclude that this structure is important for epidemics. For example, Ref. [@liljeros] observed power-law distributions of degree (number of neighbors in the network) in sexual networks, Ref. [@vespi:threshold] showed that model networks with power-law degree distributions need not to have an epidemic threshold, thus concluding the degree distribution is an important structure. For another example, Ref. [@barabasi:burst] found power-laws in interevent time distributions, Ref. [@small_slow; @goh_vazquez] showed that outbreaks are slowed down by such fat-tailed distributions. Can we from this conclude that timing of contacts are important for disease spreading? Perhaps, but Ref. [@holme_liljeros] argued that other, longer time-scale temporal structures are even more important. However, also Ref. [@holme_liljeros] test two *a priori* chosen structures. There could of course be other structures present affecting the spreading processes even stronger. The idea of this paper is to scan the possible structures in a less restrictive way, so as to open for the discovery of new important temporal network structures. For the same reason—that it is hard to *a priori* reason about what the important temporal network structures are—we use empirical networks as our starting point rather than models generating the contact structure.
In this paper, we will run the Susceptible-Infectious-Recovered (SIR) compartmental disease spreading model (a canonical model for diseases that give immunity upon recovery) on eight human contact networks. We use temporal-network, static-network and fully-connected representations of these data sets. Then, to explain the deviations between the three representations, we explore 32 quantities measuring temporal-network structure.
Preliminaries
=============
In this section, we will clarify the methods and precise model definitions in common to the rest of the paper. We also mention some computational considerations. In general, we assume a temporal network $H$ as input. It can be described as a list of $C$ *contacts* $(i,j,t)$ where $i,j\in V$ are individuals and $t$ is the time of the contact (assuming a discretized time, as common for most data sets). The order of $i$ and $j$ does not matter. We set the smallest time to zero and label it $T$.
SIR simulation
--------------
We use the constant-duration version of the SIR model [@holme:ploscomp]. In this model, a contact between a susceptible and infectious individuals infects the susceptible (instantaneously) with a probability $\lambda$. Then the infectious recovers a time $\delta$ later, and stays recovered for the rest of the simulation. The infection seed is chosen randomly and taken to become infectious immediately prior to its first contact. When there are no infectious individuals left, the infection is extinct. The time from the infection is introduced until the last recovery is the *extinction time* $\tau$. The fraction of recovered individuals when the outbreak is extinct is the *outbreak size* $\Omega$.
For the static networks, we consider a disease spreading on a graph $G=(V,E)$ where $(i,j)\in E$ if there is a contact $(i,j,t)\in H$. We generate $C$ contacts between individuals connected by an link in $E$ at randomly chosen times between $0$ and $T$. Thus use as close as possible to the original data (assuming the maximum entropy principle—to maximize the randomness of the unknown structures). Analogously, for the fully-connected case, we also generate $C$ contacts at times in the interval $[0,T]$, but this time it can be between any pair of nodes.
Each data point of the $20\times 20$ parameter combinations is averaged over $200,000$ independent runs. We let the sequences of $\lambda$ and $\delta$ grow exponentially, as will be evident later, an exponential growth is needed to separate the data sets. For the same reason it is convenient to use the logarithm (we use the base-ten logarithm) of these values for discussion.
Data sets
---------
Data set $N$ $C$ $T$ $t$-res. $M$
---------------- -------- -------- ------- ---------- --------
*Prostitution* 16,730 50,632 6.00y 1d 39,044
*Conference* 113 20,818 2.50d 20s 2,196
*Hospital* 75 32,424 96.5h 20s 1,139
*Reality* 64 26,260 8.63h 5s 722
*School 1* 236 60,623 8.64h 20s 5,901
*School 2* 238 65,150 8.58h 20s 5,541
*Gallery 1* 200 5,943 7.80h 20s 714
*Gallery 2* 204 6,709 8.05h 20s 739
: \[tab:data\]Basic statistics of the data sets. $N$ is the number of individuals; $C$ is the number of contacts; $T$ is the total sampling time; $t$-res. is the time resolution of the data set and $M$ is the number of links in the projected static networks.
As mentioned above, this study is based on empirical data sets of human proximity. In other words, they are recording two persons in close proximity at a certain time. For obvious reasons, these are interesting for disease spreading. We list the basic statistics—sizes, sampling durations, etc.—of the data sets in Table \[tab:data\].
Our first data set (*Prostitution*) comes from rom self-reported sexual contacts between female prostitutes and male sex buyers [@prostitution]. This is a special form of proximity network in that a contact is sexual. Perhaps it should be classified as a separate type of network, but it is relevant for disease spreading. Several other data sets come from the Sociopatterns project (sociopatterns.org). These data sets by radio-frequency identification sensors that record a contact when two sensors are within 1–1.5 m. One of these datasets comes from a conference [@conference] (*Conference*), another from a school (*School*) [@school], a third from a hospital (*Hospital*) [@hospital] and a fourth from an art gallery (*Gallery*) [@gallery]. The *Gallery* data set comprises 69 days where we use the first two. *School* consists of two days and we use both.
A similar data set to the Sociopatterns data sets comes from the Reality mining study [@reality] (*Reality*). Here contacts within a cohort of university students were recorded by the Bluetooth channel of smartphones. The range of such connections is between 10 and 15 meters. We use the same subset of data set as in Ref. [@pfitzner].
Temporal network descriptors
----------------------------
To characterize the temporal-network structures of the data sets, we use 32 different quantities, which we call network descriptors. We choose these both to be relatively simple and straightforward to interpret and to cover as wide spectrum of structures as possible. Table \[tab:quantities\] presents an overview of the descriptors.
### Time evolution
We measure nine network descriptors characterizing the long-term behavior of the contact dynamics—briefly speaking, how the contact process differs from a stationary process. Some of these data sets (e.g. *Prostitution*, *Gallery 1* and *2*) are growing throughout the sampling period, and this has been argued to influence the spreading dynamics strongly [@holme_liljeros]. In such a system, the disease could burn out in the population even before some individuals have entered it.
The first of these measures focuses on the time when nodes and links first appear in the data. First, we measure the fraction of nodes (links) present at half the sampling time relative to the final number of nodes, $f_{TN}$ (links, $f_{TL}$). Some studies argues the order of events is a more natural measure of time than the actual time. Thus we also measure the corresponding quantities $f_{CN}$ and $f_{CL}$ where half the sampling time is replaced by the half the contacts.
The second class of network descriptors, focuses on the persistence nodes or links. Let $F_{TN}$ ($F_{TL}$) be the fraction of nodes (links) present in the first and last $5\%$ of the time. The corresponding quantities for the sequence of contacts are $F_{CN}$ and $F_{CL}$.
Yet a measure related to the time evolution is the largest gap $g$ on the contact sequence. (During a gap, the disease cannot spread, and for long enough gaps, the disease could die out.)
symbol description
--------------- -----------------------------------------------------------------------------------------------------------------
$f_{NC}$ Fraction of nodes present (had been involved in least one contact) when half of the contacts happened.
$f_{NT}$ Fraction of nodes present at half the sampling time.
$f_{LC}$ Fraction of links present when half of the contacts happened.
$f_{LT}$ Fraction of links present at half the sampling time.
$F_{NC}$ Fraction of nodes present at both the first and last $5\%$ of the contacts.
$F_{NT}$ Fraction of nodes present at both the first and last $5\%$ of the sampling time.
$F_{LC}$ Fraction of links present at both the first and last $5\%$ of the contacts.
$F_{LT}$ Fraction of links present at both the first and last $5\%$ of the sampling time.
$\mu_{Lt}$ Mean link interevent time.
$\sigma_{Lt}$ Standard deviation of interevent times of links.
$c_{Lt}$ Coefficient of variation of interevent times of links, also known as the average link burstiness [@burstiness].
$\gamma_{Lt}$ Skewness of interevent times of links.
$\mu_{Ld}$ Mean duration (time between first and last contact) of links.
$\sigma_{Ld}$ Standard deviation of the duration of links.
$c_{Ld}$ Coefficient variation of the duration of links.
$\gamma_{Ld}$ Skewness of the duration distribution of links.
$\mu_{Nt}$ Like $\mu_{Lt}$ but for nodes.
$\sigma_{Nt}$ Like $\sigma_{Lt}$ but for nodes.
$c_{Nt}$ Like $c_{Lt}$ but for nodes, i.e., the node burstiness.
$\gamma_{Nt}$ Like $\gamma_{Lt}$ but for nodes.
$\mu_{Nd}$ Like $\mu_{Ld}$ but for nodes.
$\sigma_{Nd}$ Like $\sigma_{Ld}$ but for nodes.
$c_{Nd}$ Like $c_{Ld}$ but for nodes.
$\gamma_{Nd}$ Like $\gamma_{Ld}$ but for nodes.
$g$ The longest gap between any two contacts in the data.
$\mu_k$ Average degree of the network of accumulated contacts.
$\sigma_k$ Standard deviation of the degree distribution of the network of accumulated contacts.
$c_k$ Coefficient of variation of the degree distribution of the network of accumulated contacts.
$\gamma_k$ Skewness of the degree distribution of the network of accumulated contacts.
$N$ Number of nodes.
$C$ Clustering coefficient of the network of accumulated contacts.
$r$ Degree assortativity of the network of accumulated contacts.
### Node and link activity
The node- and link-activity descriptors capture the bursty nature of human behavior. I.e., intense periods of activity separated by long periods of inactivity [@burstiness]. One can imagine many ways to measure burstiness. The common starting point is interevent times—the time gap between consecutive contacts of a node or link. We measure four descriptors characterizing this kind of time series—the mean $\mu$, standard deviation $\sigma$, coefficient of variation $c$ (i.e. the standard deviation divided by the mean), and the skewness $$\label{eq:skewness}
\gamma=\frac{(n^2-n)^{1/2}}{n-2}\frac{\mu_3}{\mu_2^{3/2}}$$ where $\mu_2$ and $\mu_3$ are the second and third moment of the distribution, respectively.
Some studies have pointed out that the duration (time from the first to the last observation) of nodes or links can be important for spreading phenomena [@holme_liljeros]. Therefore, we also study the distribution of node and link durations by the same four descriptors as the interevent times. In total, for this category, we define 16 network descriptors—$\mu$, $\sigma$, $c$ and $\gamma$, for both interevent-time and duration distributions and for both nodes and links.
{width="90.00000%"}
See Fig. \[fig:timeline\] for visualizations of the time structure of the data sets.
### Measures of static network structure
How contact structures affect dynamic processes, such as epidemic spreading, is more established for static network structures than for temporal structures. We will measure the static network structure for the networks of accumulated contacts, i.e. if there has been at least one contact between two nodes we consider them connected by an link.
Arguably, the most important static network structure is the degree distribution describing how frequent it is to observe a node of a particular degree. Essentially a broad, right-skewed degree distribution (such as frequently observed in real systems) speeds up spreading phenomena [@vespi_rev]. Usually researchers are interested in inferring the functional form of the degree distribution. For our purpose, we need to summarize the structures to numbers, no matter the functional forms. Therefore, we measure the same four quantities—$\mu$, $\sigma$, $c$ and $\gamma$—as for the interevent time and duration distributions.
In addition to the degree distribution, we also measure other static network descriptors. First, and simplest, the number of nodes, $N$. (But not the number of links since it is equal to $N\mu_{\rm deg}/2$. The next static network descriptor is the assortativity $r$. This is, in essence, the Pearson correlation of the degrees at either side of an link. One only has to symmetrize the arguments of the correlation coefficient (since the first and second arguments are different, but links are unordered with respect to the nodes—see Ref. [@newman:book]) for details. The assortativity captures the tendency for nodes of similar degree to connect to each other. A large assortativity means that high-degree nodes connect to other high-degree nodes, and low-degree nodes to other low-degree nodes. It has been shown to have an influence on disease dynamics—assortative networks having lower epidemic thresholds [@epi_ass]. Finally, we study the clustering coefficient—the number of triangles in the network normalized to the unit interval [@newman:book]. A high clustering coefficient is known to slow down disease spreading [@epi_clustering].
{width="90.00000%"}
See Fig. \[fig:structure\] for visualizations of the networks of accumulated contacts. Just like Fig. \[fig:timeline\], this figure does not tell us more than that there are rich structures in the network topology that can influence the outbreak dynamics.
Overlap statistics
------------------
We will look at groups of data sets with different behavior of the SIR model with respect to the three levels of representations of the contacts. A good candidate network descriptor should separate the two groups well. With more samples, we could use e.g. the mutual information or Kullback-Leibler divergence, but with only eight data points, we can use a simpler quantity relying on the extreme values of the quantity for the two groups. Let $A$ be one subset of the data sets and $B$ its complement. Let $v(G)$ be the value for a quantity as a function of the data set $G$. Furthermore, assume (without loss of generality) that $\max_{G\in A}v(G)\geq\max_{G\in B}v(G)$ Then, more specifically, we measure $$\label{eq:x}
x_v(A,B)=\frac{\min_{G\in A}v(G)-\max_{G\in B}v(G)}{\max_{G\in A}v(G)-\min_{G\in A\cup B}} .$$ In other words, if $\{v(G):G\in A\}$ and $\{v(G):G\in B\}$ do not overlap, then $x$ is the smallest difference between values in the two sets divided by the largest difference. If $x=1$, the separation is maximal. If $A$ and $B$ do overlap, $x$ will be negative, reaching a minimum $-1$ if the range of $\{v(G):G\in A\}$ and $\{v(G):G\in B\}$ are the same.
Results
=======
{width="90.00000%"}
{width="90.00000%"}
{width="90.00000%"}
Extinction time
---------------
One of our main quantities is the mean time to extinction $\tau$. Fig. \[fig:tte\_t\] shows the values for SIR simulated on temporal-network representations. $\tau$ is strictly increasing with the disease duration $\delta$ but has a maximum in the per-contact transmission probability $\lambda$. The maximum comes from two conflicting mechanisms [@holme_ext]. For small $\lambda$, decreasing $\lambda$ gives fewer chances for contagion and an increasing chance of the disease dying out. For large $\lambda$, the disease burns out fast in the population. The actual location of the peak varies much, from close to the maximum $\lambda=1$ for *Prostitution* data set to $\lg\lambda\approx -1.8$ for the *Hospital* data.
The effect of removing the temporal information by aggregating the contacts to a static network is seen in Fig. \[fig:tte\_diff\_n\]. This figure shows the deviation $\Delta\tau$ between $\tau$ of the static and temporal networks (so negative values means the outbreaks last longer in temporal networks). We see the *Prostitution*, *Gallery 1* and *2* are different than the others in that they do not have regions of negative $\Delta\tau$—the static networks always give longer outbreaks. If we proceed, removing the network structure by making the network fully mixed (i.e. fully connected), then not much more happens (Fig. \[fig:tte\_diff\_fm\]). $\Delta\tau$ becomes larger for some regions of, in particular, the *Gallery* data sets. The qualitative picture is, however, the same. Except for *Prostitution*, *Gallery 1* and *2*, extinction times are underestimated for the largest $\lg\lambda$ and $\lg\delta$ and overestimated for intermediate $\lg\lambda$ and $\lg\delta$. This seems to suggest that the extinction time is more dependent on temporal than topological structures—a hypothesis we hope that future studies can confirm. Below, we will explore what separates the behavior of the *Prostitution* and the two *Gallery* data sets apart from the rest.
{width="90.00000%"}
{width="90.00000%"}
{width="90.00000%"}
Outbreak size
-------------
The average expected outbreak size $\Omega$ is perhaps a yet more common quantity than $\tau$ to characterize outbreaks in computational studies of disease spreading. Fig. \[fig:omega\_t\] shows the values of $\Omega$ throughout the parameter space (for most data sets, these values were also presented in our Ref. [@holme]). $\Omega$ is monotonically increasing with $\lg\lambda$ and $\lg\delta$ which probably is inevitable on average (even though, for specific seeds $i$, a larger $\lg\delta$ can lead to that the disease burn out so fast around $i$ that it is already extinct when a contact leading away from $i$’s vicinity appears).
Figs. \[fig:omega\_diff\_n\] and \[fig:omega\_diff\_fm\] show, respectively, the deviation when the temporal and both temporal and topological information is removed. Unlike $\tau$, the static network structure creates a qualitative difference—but only for the *Prostitution* data. For this data set, the outbreak sizes are consistently underestimated for the fully connected networks, while for the static networks, $\lambda=\delta$ roughly separates two regions—for $\delta>\lambda$ the outbreak sizes are overestimated whereas for $\delta<\lambda$ they are underestimated. Below, we will look for a structural explanation behind this phenomenon.
![(Color online) The top-three temporal-network structures separating the *Prostitution*, *Gallery 1* and *Gallery 2* data sets from the rest of the data sets (a) and separating *Prostitution* from the rest (b). $x$ is the difference between the smallest value (of the network structural measure in question) of the set containing the largest value, and the largest value of the other set divided by the difference between the largest and smallest values in the union of the sets.[]{data-label="fig:explanatory"}](explanatory.pdf){width="\columnwidth"}
Structural explanations
-----------------------
In this section, we we try to find what network structures that affect the effects found above. First, that *Prostitution* and the *Gallery* data sets differ from the others in that they lack a region of parameter space where the representations without temporal structure overestimates the time to extinction. Second, that *Prostitution* has a different response to removing the network structure than all the other data sets.
First, we investigate which network descriptors that separates $A=\{\mbox{\textit{Prostitution}}, \mbox{\textit{Gallery 1}}, \mbox{\textit{Gallery 2}}\}$ from the rest ($A$ refers to Eq. \[eq:x\] and the discussion about it). The top three quantities $v$ with respect to $x_v$ along with their values, for the two groups of data sets are plotted in Fig. \[fig:explanatory\](a). These three quantities—the average life time of nodes $\mu_{Nt}$ and links $\mu_{Lt}$, and the fraction of nodes present at half of the contacts $f_{nC}$—are all temporal in nature, and all related to the turnover of individuals in the data, rather than higher frequency properties like the interevent time statistics. In more detail, we see that the data sets without regions of negative $\Delta\tau$ are characterized of a short average presence of the nodes and links in the data, and thus a high turnover of individuals. Representing such temporal networks as static networks destroys the long time-scale effects like that a node present early in the data cannot be infected by a node present only late in the data.
Our second investigation concerns how *Prostitution* differs from the other data sets (Fig. \[fig:explanatory\](b)). We find that the quantities with the largest $v$ values are: the number of nodes $N$, the average interevent time of nodes $\mu_{Nt}$ and the skewness of the degree distribution $\gamma_k$. These three quantities are very different from the ones to explain the other effect (in Fig. \[fig:explanatory\](a)). The number of nodes is probably not an explanation for this effect in itself, but it could help accentuating other effects. The long average interevent times of *Prostitution* come from a very skewed distribution of the number of contacts (a quantity we do not measure directly). The few-contact individuals can have long dormant periods, and thus increase the average interevent time. (Individuals with only one contact, of which there are around $35\%$, do not contribute to $\mu_{Nt}$.) The degree distribution is a very well-studied quantity, responsible for many peculiar features in static network epidemiology (such as the vanishing of epidemic thresholds or emergence of super spreaders [@vespi_rev]). If is therefore reassuring to see its skewness as one of the top explanatory descriptors. It, furthermore, makes sense that the difference between the static networks and the fully connected networks is best explained by static network quantities. However, except the *Prostitution* data, the static and fully connected-networks deviates from the temporal network in the same way, which means that the temporal structures are more influential with respect to disease spreading for these data sets, not only for $\tau$ but also for $\Omega$.
Discussion
==========
We have compared SIR simulations (the entire parameter space) on three levels of representations of empirical contact data—temporal networks, static networks and fully-connected networks. We used two quantities characterizing the evolution of the outbreak—the time to extinction and the average outbreak size. We see that going from a temporal network representation to static- or fully connected network representations can lead to both a severe under- and over-estimation of both the extinction time and the average outbreak size. In general, short disease durations and high transmission probabilities lead to an over-estimation when the temporal information is discarded. Going from a static-network representation to a fully-connected topology does not make much of a difference except for one data set (*Prostitution*) and one of the quantities (average outbreak size). Looking closer at the quantities determining the patterns of over- and underestimation of $\tau$ and $\Omega$ also gives at hand that quantities describing the time evolution of the network are the most influential structures (in agreement to Ref. [@holme]). Static network structure and shorter time scale temporal structure such as interevent times matters less. These observations are, of course, specific for the particular data sets we study. The results should be generalized with care. On the other hand, the contact data sets we use are as good as we can possibly obtain. There are no obvious structures in these data sets that disqualify them as representative of real data sets (except, perhaps, the limited sizes). At the very least, this should encourage more research into the role of time structures in disease spreading.
There are many possible extensions of this work. Even though we used a generous amount of 32 network descriptors, one can imagine many other—describing how static network quantities change over the sampling time, how the activity level of nodes and their network position are correlated, etc. Ultimately, one would like to use results from this type of study to construct generative models for outbreak scenarios, retaining the important structures, but not more. Indeed, some such models have already been proposed [@holme_saramaki; @holme_modern], but, to our knowledge, none that focuses on the longer time-scale features that we find important.
P.H. was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2013R1A1A2011947).
[10]{}
A.-L. Barabási. The origin of bursts and heavy tails in human dynamics. , 435:907, 2005.
T. Britton, M. Deijfen, A. N. Lager[å]{}s, and M. Lindholm. Epidemics on random graphs with tunable clustering. , 45:743–756, 2008.
J. M. Drake. Limits to forecasting precision for outbreaks of directly transmitted diseases. , 3(1), 11 2005.
N. Eagle and A. Pentland. Reality mining: Sensing complex social systems. , 10:255–268, 2006.
K.-I. Goh and A.-L. Barabási. Burstiness and memory in complex systems. , 81(4):48002, 2008.
H. W. Hethcote. The mathematics of infectious diseases. , 42(4):599–653, 2000.
P. Holme. Epidemiologically optimal static networks from temporal network data. , 9(7):1–10, 07 2013.
P. Holme. Extinction times of epidemic outbreaks in networks. , 8(12):1–7, 12 2013.
P. Holme. Information content of contact-pattern representations and predictability of epidemic outbreaks. , 5:14462, 2015.
P. Holme. Modern temporal network theory: a colloquium. , 88(9):1–30, 2015.
P. Holme and F. Liljeros. Birth and death of links control disease spreading in empirical contact networks. , 4:4999, 2014.
P. Holme and J. Saramäki. Temporal networks. , 519(3):97–125, 2012.
L. Isella, J. Stehlé, A. Barrat, C. Cattuto, J. F. Pinton, and W. van den Broeck. What’s in a crowd? analysis of face-to-face behavioral networks. , 271:166–180, 2011.
S. Janson, M. Luczak, and P. Windridge. Law of large numbers for the sir epidemic on a random graph with given degrees. , 45(4):726–763, 2014.
M. Karsai, M. Kivelä, R. K. Pan, K. Kaski, J. Kertész, A.-L. Barabási, and J. Saramäki. Small but slow world: How network topology and burstiness slow down spreading. , 83:025102, Feb 2011.
M. J. Keeling and K. T. Eames. Networks and epidemic models. , 2(4):295–307, 2005.
F. Liljeros, C. R. Edling, L. A. N. Amaral, H. E. Stanley, and Y. [Å]{}berg. The web of human sexual contacts. , 411:207–211, 2001.
N. Masuda and P. Holme. Predicting and controlling infectious disease epidemics using temporal networks. , 5:6, 2013.
B. Min, K.-I. Goh, and A. Vazquez. Spreading dynamics following bursty human activity patterns. , 83:036102, Mar 2011.
M. E. J. Newman. Mixing patterns in networks. , 67:026126, 2003.
M. E. J. Newman. . Oxford University Press, Oxford, 2010.
R. Pastor-Satorras, C. Castellano, P. [van Mieghem]{}, and A. Vespignani. Epidemic processes in complex networks. , 87:925–979, 2015.
R. Pastor-Satorras and A. Vespignani. Epidemic spreading in scale-free networks. , 86:3200–3203, 2001.
L. R. Petersen, D. J. Jamieson, A. M. Powers, and M. A. Honein. Zika virus. , 374(16):1552–1563, 2016.
R. Pfitzner, I. Scholtes, A. Garas, T. J. Tessone, and F. Schweitzer. Betweenness preference: Quantifying correlations in the topological dynamics of temporal networks. , 110:198701, 2013.
L. E. C. Rocha, F. Liljeros, and P. Holme. Information dynamics shape the sexual networks of internet-mediated prostitution. , 107:5706–5711, 2010.
J. Stehlé, N. Voirin, A. Barrat, C. Cattuto, L. Isella, J.-F. Pinton, M. Quaggiotto, W. van den Broeck, C. Régis, B. Lina, and P. Vanhems. High-resolution measurements of face-to-face contact patterns in a primary school. , 6:e23176, 2011.
W. van den Broeck, M. Quaggiotto, L. Isella, A. Barrat, and C. Cattuto. The making of sixty-nine days of close encounters at [The Science Gallery]{}. , 45:201–202, 2012.
P. Vanhems, A. Barrat, C. Cattuto, J.-F. Pinton, N. Khanafer, C. Régis, B.-A. Kim, B. Comte, and N. Voirin. Estimating potential infection transmission routes in hospital wards using wearable proximity sensors. , 8:e73970, 2013.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report a new search for dark matter in a data sample of an integrated luminosity of 7.7 fb$^{-1}$ of Tevatron $p\bar{p}$ collisions at $\sqrt{s} = 1.96$ TeV, collected by the CDF II detector. We search for production of a dark matter candidate, $D$, in association with a single top quark. We consider the hadronic decay mode of the top quark exclusively, yielding a final state of three jets with missing transverse energy. The data are consistent with the standard model; we thus set 95% confidence level upper limits on the cross section of the process $p\bar{p}\rightarrow t+D$ as a function of the mass of the dark-matter candidate. The limits are approximately 0.5 pb for a dark-matter particle with mass in the range of $0-150$ GeV/$c^{2}$.'
author:
- 'T. Aaltonen'
- 'B. Álvarez González$^z$'
- 'S. Amerio'
- 'D. Amidei'
- 'A. Anastassov$^x$'
- 'A. Annovi'
- 'J. Antos'
- 'F. Anzá'
- 'G. Apollinari'
- 'J.A. Appel'
- 'T. Arisawa'
- 'A. Artikov'
- 'J. Asaadi'
- 'W. Ashmanskas'
- 'B. Auerbach'
- 'A. Aurisano'
- 'F. Azfar'
- 'W. Badgett'
- 'T. Bae'
- 'A. Barbaro-Galtieri'
- 'V.E. Barnes'
- 'B.A. Barnett'
- 'P. Barria$^{hh}$'
- 'P. Bartos'
- 'M. Bauce$^{ff}$'
- 'F. Bedeschi'
- 'S. Behari'
- 'G. Bellettini$^{gg}$'
- 'J. Bellinger'
- 'D. Benjamin'
- 'A. Beretvas'
- 'A. Bhatti'
- 'D. Bisello$^{ff}$'
- 'I. Bizjak'
- 'K.R. Bland'
- 'B. Blumenfeld'
- 'A. Bocci'
- 'A. Bodek'
- 'D. Bortoletto'
- 'J. Boudreau'
- 'A. Boveia'
- 'L. Brigliadori$^{ee}$'
- 'C. Bromberg'
- 'E. Brucken'
- 'J. Budagov'
- 'H.S. Budd'
- 'K. Burkett'
- 'G. Busetto$^{ff}$'
- 'P. Bussey'
- 'A. Buzatu'
- 'A. Calamba'
- 'C. Calancha'
- 'S. Camarda'
- 'M. Campanelli'
- 'M. Campbell'
- 'F. Canelli$^{11}$'
- 'B. Carls'
- 'D. Carlsmith'
- 'R. Carosi'
- 'S. Carrillo$^m$'
- 'S. Carron'
- 'B. Casal$^k$'
- 'M. Casarsa'
- 'A. Castro$^{ee}$'
- 'P. Catastini'
- 'D. Cauz'
- 'V. Cavaliere'
- 'M. Cavalli-Sforza'
- 'A. Cerri$^f$'
- 'L. Cerrito$^s$'
- 'Y.C. Chen'
- 'M. Chertok'
- 'G. Chiarelli'
- 'G. Chlachidze'
- 'F. Chlebana'
- 'K. Cho'
- 'D. Chokheli'
- 'W.H. Chung'
- 'Y.S. Chung'
- 'M.A. Ciocci$^{hh}$'
- 'A. Clark'
- 'C. Clarke'
- 'G. Compostella$^{ff}$'
- 'M.E. Convery'
- 'J. Conway'
- 'M.Corbo'
- 'M. Cordelli'
- 'C.A. Cox'
- 'D.J. Cox'
- 'F. Crescioli$^{gg}$'
- 'J. Cuevas$^z$'
- 'R. Culbertson'
- 'D. Dagenhart'
- 'N. d’Ascenzo$^w$'
- 'M. Datta'
- 'P. de Barbaro'
- 'M. Dell’Orso$^{gg}$'
- 'L. Demortier'
- 'M. Deninno'
- 'F. Devoto'
- 'M. d’Errico$^{ff}$'
- 'A. Di Canto$^{gg}$'
- 'B. Di Ruzza'
- 'J.R. Dittmann'
- 'M. D’Onofrio'
- 'S. Donati$^{gg}$'
- 'P. Dong'
- 'M. Dorigo'
- 'T. Dorigo'
- 'K. Ebina'
- 'A. Elagin'
- 'A. Eppig'
- 'R. Erbacher'
- 'S. Errede'
- 'N. Ershaidat$^{dd}$'
- 'R. Eusebi'
- 'S. Farrington'
- 'M. Feindt'
- 'J.P. Fernandez'
- 'R. Field'
- 'G. Flanagan$^u$'
- 'R. Forrest'
- 'M.J. Frank'
- 'M. Franklin'
- 'J.C. Freeman'
- 'B. Fuks'
- 'Y. Funakoshi'
- 'I. Furic'
- 'M. Gallinaro'
- 'J.E. Garcia'
- 'A.F. Garfinkel'
- 'P. Garosi$^{hh}$'
- 'H. Gerberich'
- 'E. Gerchtein'
- 'S. Giagu'
- 'V. Giakoumopoulou'
- 'P. Giannetti'
- 'K. Gibson'
- 'C.M. Ginsburg'
- 'N. Giokaris'
- 'P. Giromini'
- 'G. Giurgiu'
- 'V. Glagolev'
- 'D. Glenzinski'
- 'M. Gold'
- 'D. Goldin'
- 'N. Goldschmidt'
- 'A. Golossanov'
- 'G. Gomez'
- 'G. Gomez-Ceballos'
- 'M. Goncharov'
- 'O. González'
- 'I. Gorelov'
- 'A.T. Goshaw'
- 'K. Goulianos'
- 'S. Grinstein'
- 'C. Grosso-Pilcher'
- 'R.C. Group$^{53}$'
- 'J. Guimaraes da Costa'
- 'S.R. Hahn'
- 'E. Halkiadakis'
- 'A. Hamaguchi'
- 'J.Y. Han'
- 'F. Happacher'
- 'K. Hara'
- 'D. Hare'
- 'M. Hare'
- 'R.F. Harr'
- 'K. Hatakeyama'
- 'C. Hays'
- 'M. Heck'
- 'J. Heinrich'
- 'M. Herndon'
- 'S. Hewamanage'
- 'A. Hocker'
- 'W. Hopkins$^g$'
- 'D. Horn'
- 'S. Hou'
- 'R.E. Hughes'
- 'M. Hurwitz'
- 'U. Husemann'
- 'N. Hussain'
- 'M. Hussein'
- 'J. Huston'
- 'G. Introzzi'
- 'M. Iori$^{jj}$'
- 'A. Ivanov$^p$'
- 'E. James'
- 'D. Jang'
- 'B. Jayatilaka'
- 'E.J. Jeon'
- 'S. Jindariani'
- 'M. Jones'
- 'K.K. Joo'
- 'S.Y. Jun'
- 'T.R. Junk'
- 'T. Kamon$^{25}$'
- 'P.E. Karchin'
- 'A. Kasmi'
- 'Y. Kato$^o$'
- 'W. Ketchum'
- 'J. Keung'
- 'V. Khotilovich'
- 'B. Kilminster'
- 'D.H. Kim'
- 'H.S. Kim'
- 'J.E. Kim'
- 'M.J. Kim'
- 'S.B. Kim'
- 'S.H. Kim'
- 'Y.K. Kim'
- 'Y.J. Kim'
- 'N. Kimura'
- 'M. Kirby'
- 'S. Klimenko'
- 'K. Knoepfel'
- 'K. Kondo[^1]'
- 'D.J. Kong'
- 'J. Konigsberg'
- 'A.V. Kotwal'
- 'M. Kreps'
- 'J. Kroll'
- 'D. Krop'
- 'M. Kruse'
- 'V. Krutelyov$^c$'
- 'T. Kuhr'
- 'M. Kurata'
- 'S. Kwang'
- 'A.T. Laasanen'
- 'S. Lami'
- 'S. Lammel'
- 'M. Lancaster'
- 'R.L. Lander'
- 'K. Lannon$^y$'
- 'A. Lath'
- 'G. Latino$^{hh}$'
- 'T. LeCompte'
- 'E. Lee'
- 'H.S. Lee$^q$'
- 'J.S. Lee'
- 'S.W. Lee$^{bb}$'
- 'S. Leo$^{gg}$'
- 'S. Leone'
- 'J.D. Lewis'
- 'A. Limosani$^t$'
- 'C.-J. Lin'
- 'M. Lindgren'
- 'E. Lipeles'
- 'A. Lister'
- 'D.O. Litvintsev'
- 'C. Liu'
- 'H. Liu'
- 'Q. Liu'
- 'T. Liu'
- 'S. Lockwitz'
- 'A. Loginov'
- 'D. Lucchesi$^{ff}$'
- 'J. Lueck'
- 'P. Lujan'
- 'P. Lukens'
- 'G. Lungu'
- 'J. Lys'
- 'R. Lysak$^e$'
- 'R. Madrak'
- 'K. Maeshima'
- 'P. Maestro$^{hh}$'
- 'S. Malik'
- 'G. Manca$^a$'
- 'A. Manousakis-Katsikakis'
- 'F. Margaroli'
- 'C. Marino'
- 'M. Martínez'
- 'P. Mastrandrea'
- 'K. Matera'
- 'M.E. Mattson'
- 'A. Mazzacane'
- 'P. Mazzanti'
- 'K.S. McFarland'
- 'P. McIntyre'
- 'R. McNulty$^j$'
- 'A. Mehta'
- 'P. Mehtala'
- 'C. Mesropian'
- 'T. Miao'
- 'D. Mietlicki'
- 'A. Mitra'
- 'H. Miyake'
- 'S. Moed'
- 'N. Moggi'
- 'M.N. Mondragon$^m$'
- 'C.S. Moon'
- 'R. Moore'
- 'M.J. Morello$^{ii}$'
- 'J. Morlock'
- 'P. Movilla Fernandez'
- 'A. Mukherjee'
- 'Th. Muller'
- 'P. Murat'
- 'M. Mussini$^{ee}$'
- 'J. Nachtman$^n$'
- 'Y. Nagai'
- 'J. Naganoma'
- 'I. Nakano'
- 'A. Napier'
- 'J. Nett'
- 'C. Neu'
- 'M.S. Neubauer'
- 'J. Nielsen$^d$'
- 'L. Nodulman'
- 'S.Y. Noh'
- 'O. Norniella'
- 'L. Oakes'
- 'S.H. Oh'
- 'Y.D. Oh'
- 'I. Oksuzian'
- 'T. Okusawa'
- 'R. Orava'
- 'L. Ortolan'
- 'S. Pagan Griso$^{ff}$'
- 'C. Pagliarone'
- 'E. Palencia$^f$'
- 'V. Papadimitriou'
- 'A.A. Paramonov'
- 'J. Patrick'
- 'G. Pauletta$^{kk}$'
- 'M. Paulini'
- 'C. Paus'
- 'D.E. Pellett'
- 'A. Penzo'
- 'T.J. Phillips'
- 'G. Piacentino'
- 'E. Pianori'
- 'J. Pilot'
- 'K. Pitts'
- 'C. Plager'
- 'L. Pondrom'
- 'S. Poprocki$^g$'
- 'K. Potamianos'
- 'F. Prokoshin$^{cc}$'
- 'A. Pranko'
- 'F. Ptohos$^h$'
- 'G. Punzi$^{gg}$'
- 'A. Rahaman'
- 'V. Ramakrishnan'
- 'N. Ranjan'
- 'I. Redondo'
- 'P. Renton'
- 'M. Rescigno'
- 'T. Riddick'
- 'F. Rimondi$^{ee}$'
- 'L. Ristori$^{42}$'
- 'A. Robson'
- 'T. Rodrigo'
- 'T. Rodriguez'
- 'E. Rogers'
- 'S. Rolli$^i$'
- 'R. Roser'
- 'F. Ruffini$^{hh}$'
- 'A. Ruiz'
- 'J. Russ'
- 'V. Rusu'
- 'A. Safonov'
- 'W.K. Sakumoto'
- 'Y. Sakurai'
- 'L. Santi$^{kk}$'
- 'K. Sato'
- 'V. Saveliev$^w$'
- 'A. Savoy-Navarro$^{aa}$'
- 'P. Schlabach'
- 'A. Schmidt'
- 'E.E. Schmidt'
- 'T. Schwarz'
- 'L. Scodellaro'
- 'A. Scribano$^{hh}$'
- 'F. Scuri'
- 'S. Seidel'
- 'Y. Seiya'
- 'A. Semenov'
- 'F. Sforza$^{hh}$'
- 'S.Z. Shalhout'
- 'T. Shears'
- 'P.F. Shepard'
- 'M. Shimojima$^v$'
- 'M. Shochet'
- 'I. Shreyber-Tecker'
- 'A. Simonenko'
- 'P. Sinervo'
- 'K. Sliwa'
- 'J.R. Smith'
- 'F.D. Snider'
- 'A. Soha'
- 'V. Sorin'
- 'H. Song'
- 'P. Squillacioti$^{hh}$'
- 'M. Stancari'
- 'R. St. Denis'
- 'B. Stelzer'
- 'O. Stelzer-Chilton'
- 'D. Stentz$^x$'
- 'J. Strologas'
- 'G.L. Strycker'
- 'Y. Sudo'
- 'A. Sukhanov'
- 'I. Suslov'
- 'K. Takemasa'
- 'Y. Takeuchi'
- 'J. Tang'
- 'M. Tecchio'
- 'P.K. Teng'
- 'J. Thom$^g$'
- 'J. Thome'
- 'G.A. Thompson'
- 'E. Thomson'
- 'D. Toback'
- 'S. Tokar'
- 'K. Tollefson'
- 'T. Tomura'
- 'D. Tonelli'
- 'S. Torre'
- 'D. Torretta'
- 'P. Totaro'
- 'M. Trovato$^{ii}$'
- 'F. Ukegawa'
- 'S. Uozumi'
- 'A. Varganov'
- 'F. Vázquez$^m$'
- 'G. Velev'
- 'C. Vellidis'
- 'M. Vidal'
- 'I. Vila'
- 'R. Vilar'
- 'J. Vizán'
- 'M. Vogel'
- 'G. Volpi'
- 'P. Wagner'
- 'R.L. Wagner'
- 'T. Wakisaka'
- 'R. Wallny'
- 'S.M. Wang'
- 'A. Warburton'
- 'D. Waters'
- 'W.C. Wester III'
- 'D. Whiteson$^b$'
- 'A.B. Wicklund'
- 'E. Wicklund'
- 'S. Wilbur'
- 'F. Wick'
- 'H.H. Williams'
- 'J.S. Wilson'
- 'P. Wilson'
- 'B.L. Winer'
- 'P. Wittich$^g$'
- 'S. Wolbers'
- 'H. Wolfe'
- 'T. Wright'
- 'X. Wu'
- 'Z. Wu'
- 'K. Yamamoto'
- 'D. Yamato'
- 'T. Yang'
- 'U.K. Yang$^r$'
- 'Y.C. Yang'
- 'W.-M. Yao'
- 'G.P. Yeh'
- 'K. Yi$^n$'
- 'J. Yoh'
- 'K. Yorita'
- 'T. Yoshida$^l$'
- 'G.B. Yu'
- 'I. Yu'
- 'S.S. Yu'
- 'J.C. Yun'
- 'A. Zanetti'
- 'Y. Zeng'
- 'C. Zhou'
- 'S. Zucchelli$^{ee}$'
title: Search for a dark matter candidate produced in association with a single top quark in collisions at TeV
---
Despite its successes, the standard model (SM) of particle physics leaves many important questions unanswered. For example the SM does not provide a candidate for dark matter (DM). Direct detection experiments such as DAMA[@ref:dama], CoGeNT[@ref:cogent1; @ref:cogent2] and CRESST[@ref:cresst] have reported signals suggestive of DM with mass in the few GeV/$c^2$ range, and with coupling to the SM sector of a strength enabling its detection at collider experiments. Many beyond-the-SM theories predict DM candidates to include such coupling between the DM and SM sectors.
In the framework of effective field theories, production of a DM particle ($D$) in association with a single top quark at hadron colliders has been recently studied[@ref:monotop; @ref:DM; @ref:models]. Here, we denote the final state containing one top quark and dark matter as [*monotop*]{}. Such studies are also inspired by the models of monojet produced in association with missing energy used to probe gravitons[@ref:monojet]. Monotop DM production is described by a set of Lagrangians incorporating all possible types of DM particles (scalar, vector, fermion, [*etc.*]{}) and their flavor-violating interactions with quarks[@ref:IVDM; @ref:IVDM2; @ref:IVDM3; @ref:IVDM4; @ref:IVDM5]. This effective theory can inclusively describe many beyond-the-SM models. For example, new physics processes with a monotop final state can also arise from the decay of a supersymmetric squark into a neutralino and a top quark, from the decay of a vector leptoquark into a massless neutrino and a top quark, or through flavor-changing neutral interactions with a new vector state escaping detection.
In the SM, top quarks are primarily produced in pairs at particle colliders. They can also be produced singly via weak interactions, resulting in a final state consisting of a single top quark with additional lighter-flavor quarks. SM single top-quark events in the missing energy plus jets channel have been studied within the standard-model hypothesis[@ref:sper3]. As the beyond-the-SM monotop theory predicts production of a single top quark in association with a DM particle, the published SM single top-quark results do not provide any conclusive information on the existence of monotop. In addition, searches for the associated production of top quarks with DM particles have only been performed in the context of events containing a pair of top quarks[@ref:sper1; @ref:sper2; @Aad:2011wc]. Therefore, a dedicated search for monotops produced in colliders is needed, as the observation of monotops would be a clear sign of new physics. In this Letter, we report the first direct search for monotop signatures at particle colliders, assuming the top quark to be produced through flavor-changing interactions of up and top quarks, in association with a DM candidate $D$. We assume that the $D$ particle has a mass in the range of $0-150$ GeV/$c^2$; we do not consider decays of the $D$ particle to up and top quarks in a higher mass range.
The top quark is short-lived and decays approximately 100% of the time into a $b$ quark and a $W$ boson, where $W \rightarrow l\nu, q\bar{q}'$. We consider the exclusive decay mode $t+D \rightarrow Wb +D$ in which $W\rightarrow q\bar{q}'$. This $W$ decay mode has the largest branching ratio and it allows for the full reconstruction of the top quark. In this channel, the missing transverse energy (${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$)[@ref:met] can be uniquely assigned to the DM particle’s passage through the detector.
Events are collected by CDF II[@ref:CDFII], a general purpose detector used to study Tevatron $p\bar{p}$ collisions with $\sqrt{s} = 1.96$ TeV. CDF II contains a tracking system consisting of a cylindrical open-cell drift chamber and silicon microstrip detectors immersed in a 1.4 T magnetic field parallel to the beam axis. Electromagnetic and hadronic calorimeters surrounding the tracking system measure particle energies. Drift chambers and muon scintillators located outside the calorimeters identify muons. We use a data sample corresponding to an integrated luminosity of $7.7\pm0.5$ fb$^{-1}$.
We consider only those events which triggered the data acquisition system due to the presence of two calorimeter clusters and significant ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$. We include data recorded between 2001 and 2010. Prior to 2007, the data acquisition system ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ threshold was 35 GeV[@ref:sper3]. After an upgrade to the system[@ref:L2upgrade] resulting in improved jet energy and ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ resolution, the requirement was lowered to ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}> 30$ GeV. Jets are reconstructed using the [jetclu]{} algorithm[@ref:jetclu] with a clustering radius of 0.4 in azimuth-pseudorapidity space ($\phi,\eta$)[@ref:pseudo]. Jet energies are corrected using standard techniques[@ref:jetcorr]. Jets originating from $b$ quarks are identified using a secondary-vertex-tagging algorithm[@ref:btag].
In order to retain only those events for which the trigger system is fully efficient, we select events with ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}>50$ GeV and three jets. Exactly one jet is identified as a $b$-jet. We require the jet transverse energy $E_T^{j_i}$, to be $E_T^{j_1}>35$ GeV, $E_T^{j_2}>25$ GeV, $E_T^{j_{3}}>15$ GeV, where the jets $j_i$ $(i=1,2,3)$ are ordered by decreasing energy. We require that either $j_i$ or $j_2$ have $|\eta| < 0.9$, and that all three jets have $|\eta| < 2.4$. We veto events with identified high-$p_T$ electrons or muons, removing monotop events inconsistent with a hadronically-decaying top quark.
We model the signal and background contributions to the selected sample using a variety of Monte Carlo (MC) simulation programs. In our simulation we assume a top-quark mass of $172.5$ GeV$/c^2$, consistent with the world’s best determination[@ref:masst; @ref:topmassreview]. We model monotop DM production in the flavor-violating process ($ug\rightarrow tD$) with [madgraph]{}[@ref:madgraph]. Additional showering and hadronization are described by [pythia]{}[@ref:pythia]. We have generated 11 signal samples assuming various DM mass in steps of 5 GeV/$c^2$ from 0 to 25 GeV/$c^{2}$, and then in steps of 25 GeV/$c^{2}$ from 25 to 150 GeV/$c^{2}$.
The event selection described above gives a data sample dominated by QCD multijet events, where the false ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ arises from the mismeasurement of jet energy. Simulation of this background is prohibitive due to the high production rate and large theoretical uncertainties. Instead, we use a method which relies on data and is based on a recently improved Tag Rate Matrix (TRM) method[@ref:tagrate]. The TRM method utilizes an estimate of the probability for QCD multijet events to have tagged jets. The probability is derived in a control region dominated by QCD multijet events. This probability is applied as a per-event weight to all events meeting our analysis selections excluding the $b$-jet requirement. From this sample of weighted events, we subtract the expected electroweak components (as modeled by applying the same TRM probability to simulated samples).Ê The resulting events form our model of the QCD multijet component of the analysis data sample.
We model other physics with samples generated by MC programs. Diboson and $t \bar{t}$ production are generated by [pythia]{} and normalized to the next-to-leading order (NLO) cross section predicted using the [mcfm]{} program[@ref:XsecDIB1; @ref:XsecDIB2] and the approximate next-to-next-to-leading order cross section[@ref:Xsectt1], respectively. The production of $W/Z$ plus light flavor and heavy flavor (HF) jets are simulated by [alpgen]{}[@ref:alpgen] with showering and hadronization performed by [pythia]{} and normalized to NLO cross sections. Single top, both $s-$ and $t$- channel production, are modeled using [madgraph]{} with [pythia]{} and normalized to NLO cross sections[@ref:XsecST1; @ref:XsecST2].
The light flavor jets misidentified as $b$-jets by the secondary-vertex-tagging algorithm are labeled as [*mistags*]{}. A data-driven method is used to estimate the mistag rate for the tagging algorithm[@ref:btag]. We apply the mistag rate to the MC events with light flavor jets to estimate the mistag contribution.
Figure\[ref:ewk-met\] shows the ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ distribution in a control region for events which pass our signal selection but have an identified high-$p_T$ electron or muon.
![The ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ distribution in a control region requiring an identified high-$p_T$ lepton; the gray area represents the statistical and systematic uncertainties on the background model. The lower panel displays the difference between the data and the expected backgrounds (Exp) divided by the latter.[]{data-label="ref:ewk-met"}](ewk_met_notitle.eps){width="40.00000%"}
After the selection described above, we are left with 6471 data events. We expect that approximately 70% of these events come from QCD multijet production. In order to further suppress the QCD contamination and the other SM backgrounds, we require the azimuthal distances between the ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ and $j_2$, $\Delta \phi({{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}, j_2) > 0.7$, as the ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ in QCD multijet background tends to align to the jet with less measured energy. We also require the invariant mass of the three jets to be consistent with the reconstructed top-quark mass, $110 < m_{jjj} < 200$ GeV/$c^2$, large ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ significance (${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}/\sqrt{\sum E_T} >3.5 \sqrt{\mathrm{GeV}}$, where $\sum E_T$ is the scalar sum of transverse energy deposited in the calorimeter) and $E_T^{j_3} > 25$ GeV. All selections have been chosen to optimize the significance $S/\sqrt{S+B}$, where $S$ and $B$ are the expected number of signal and backgrounds events, respectively. Table \[ref:EAC\] shows the number of events in the signal region for the data, the number of events for SM backgrounds, and the expected signal assuming different values of the DM particle’s mass. The events that fail these signal-region selections are used to form a control region that is used to validate the background models, as well as to determine the normalization of the QCD multijet background.
Processes Events
------------------------------ --------------------
$p\bar{p} \rightarrow t + D$
$m_D = 20$ GeV/$c^{2}$ 2116.9 $\pm$ 121.4
$m_D = 75$ GeV/$c^{2}$ 232.3 $\pm$ 22.9
$m_D = 100$ GeV/$c^{2}$ 129.8 $\pm$ 12.5
$m_D = 125$ GeV/$c^{2}$ 94.5 $\pm$ 9.3
$t \bar{t}$ 182.8 $\pm$ 20.2
Single top 24.3 $\pm$ 4.5
Diboson 15.7 $\pm$ 2.7
$W/Z$+HF 130.5 $\pm$ 33.8
Mistag 96.9 $\pm$ 39.4
QCD multijet 210.2 $\pm$ 54.5
Total background 660.2 $\pm$ 78.1
Data 592
: Number of expected signal and background events compared to data in the signal region. The expected signals, assuming different values for the mass of the DM particle, are also presented. The errors include statistical and systematic uncertainties.[]{data-label="ref:EAC"}
We consider several systematic uncertainties affecting the sensitivity of this search. The dominant systematic sources are the uncertainties on multijet normalization (25.5%), the mistag rate (16.6%) and the background cross sections ($6.5\% - 30\%$). We also consider uncertainties from the jet energy scale[@ref:jetcorr] ($2.8\% - 10.7\%$), the luminosity measurement[@ref:Lumi] (6%), parton density functions (2%), lepton veto (2%), $b$-tagging efficiency (5.2%), trigger efficiency ($0.4\% - 0.9\%$), and from the initial-state and final-state radiation (4%). We also assign systematic uncertainties, based on the variation in the shape of the distribution of kinematic quantities, under a $\pm 1\sigma$ variation of the jet energy scale and the uncertainty on the efficiency of the data acquisition system.
The ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ is chosen to discriminate the signals from the backgrounds. The ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ distribution due to a DM particle of mass of 125 GeV/$c^2$ and the SM backgrounds are shown in Fig.\[fig:met\]. The signal is expected to contribute significantly at high values of ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$. We find no significant excess of signal-like events in the data analyzed, and thus proceed to set 95% confidence level (C.L.) upper limits on the monotop DM production cross section. The limits are calculated with the ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ distribution as the shape discriminant using a Bayesian maximum likelihood method assuming a flat prior for the signal cross section[@ref:tom-CL]. We treat systematic uncertainties using a Bayesian marginal likelihood method. Figure\[fig:final\] shows the calculated upper limits on the monotop cross section as a function of the mass of the DM candidate compared to the theoretical predictions.
![The ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ distribution in the signal region. The data is compared to the sum of the SM contributions. The distribution of signal events with a DM mass of 125 GeV/$c^2$ is also shown.[]{data-label="fig:met"}](sigreg_met_notitle.eps)
![Exclusion curve of the monotop cross section as a function of the mass of DM particle[]{data-label="fig:final"}](upperLimit_notitle.eps)
In conclusion, we have performed the first search for the production of DM in association with a single top quark at hadron colliders. In an analysis of 7.7 fb$^{-1}$ of CDF II data we have found that the observed data is consistent with the expectation from SM backgrounds. We set 95% C.L. upper limits on the cross section of $p\bar{p} \rightarrow D+t$ as a function of the DM mass in the range of $0-150$ GeV/$c^{2}$. Future searches for new physics in monotop final states can probe resonant production of top quarks and DM candidates with exotic mediators. While these processes are predicted to have low production rates (making them difficult to probe with Tevatron data), they are expected to be within the reach of LHC experiments with sufficient data.
We thank the Fermilab staff and the technical staffs of the participating institutions for their vital contributions. We also thank Fabio Maltoni for useful discussions. This work was supported by the U.S. Department of Energy and National Science Foundation; the Italian Istituto Nazionale di Fisica Nucleare; the Ministry of Education, Culture, Sports, Science and Technology of Japan; the Natural Sciences and Engineering Research Council of Canada; the National Science Council of the Republic of China; the Swiss National Science Foundation; the A.P. Sloan Foundation; the Bundesministerium für Bildung und Forschung, Germany; the Korean World Class University Program, the National Research Foundation of Korea; the Science and Technology Facilities Council and the Royal Society, UK; the Russian Foundation for Basic Research; the Ministerio de Ciencia e Innovación, and Programa Consolider-Ingenio 2010, Spain; the Slovak R&D Agency; the Academy of Finland; and the Australian Research Council (ARC).
[99]{}
R. Bernabei [*et al.*]{} (DAMA Collaboration), Eur. Phys. J. C [**56**]{}, 333 (2008). C. E. Aalseth [*et al.*]{} (CoGeNT Collaboration), Phys. Rev. Lett. [**106**]{}, 131301 (2011). C. E. Aalseth [*et al.*]{} (CoGeNT Collaboration), Phys. Rev. Lett. [**107**]{}, 141301 (2011). G. Angloher [*et al.*]{}, arXiv:1109.0702. J. Kamenik and J. Zupan, Phys. Rev. D [**84**]{}, 111502 (2011). J. Andrea, B. Fuks, and F. Maltoni, Phys. Rev. D [**84**]{}, 074025 (2011). D. Alves [*et al.*]{}, arXiv: 1105.2838. Y. Bai, P. J. Fox, and R. Harnik, J. High Energy Phys. 12, 048 (2010). J. L. Feng, J. Kumar, D. Marfatia, and D. Sanford, Phys. Lett. B, [**703**]{}, 2 (2011). B. Batell, J. Pradler, and M. Spannowsky, arXiv:1105.1781. J. Kile and A. Soni, Phys. Rev. D [**84**]{} 035016, (2011). P. Agrawal, S. Blanchet, Z. Chacko, and C. Kilic, arXiv:1109.3516. S. Chen and Y. Zhang, arXiv:1106.4044. T. Aaltonen [*et al.*]{} (CDF Collaboration), Phys. Rev. D [**81**]{}, 072003 (2010). T. Aaltonen [*et al.*]{} (CDF Collaboration), Phys. Rev. Lett. [**106**]{}, 191801 (2011). T. Aaltonen [*et al.*]{} (CDF Collaboration), Phys. Rev. Lett. [**107**]{}, 191803 (2011). G. Aad [*et al.*]{} (ATLAS Collaboration), arXiv:1109.4725. Missing transverse energy, ${{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}_{T}}$ , is defined as the magnitude of the vector ${\vec{{\mbox{$\protect \raisebox{0.3ex}{$\not$}E$}}}_{T}}= - \sum_i E_T^i \vec{n}_i$ where $E_T^i$ is the magnitude of transverse energy contained in each calorimeter tower $i$, and $\vec{n}_i$ is the unit vector from the interaction vertex to the tower in the transverse $(x,y)$ plane. D. Acosta *et al.* (CDF collaboration), Phys. Rev. D **71**, 032001 (2005). A. Bhatti [*et al.*]{}, IEEE Transactions on Nuclear Science [**56**]{}, 3 (2009). F. Abe, [*et al.*]{} (CDF collaboration), Phys Rev. D [**45**]{}, 001448 (1992). CDF uses a cylindrical coordinate system with the z axis along the proton beam axis. Pseudorapidity is $\eta = − \mathrm{ln}(\mathrm{tan}( \theta /2))$, where $\theta$ is the polar angle relative to the proton beam direction, and $\phi$ is the azimuthal angle while $p_T=|p|\mathrm{sin} \theta$, $E_T=E \mathrm{sin} \theta$. A. Bhatti [*et al.*]{}, Nucl. Instrum. Methods A [**566**]{}, 375 (2006). D. Acosta [*et al.*]{} (CDF Collaboration), Phys. Rev. D [**71**]{}, 052003 (2005). Tevatron Electroweak Working Group for the CDF and D0 Collaborations, arXiv:1107.5255. A. B. Galtieri, F. Margaroli, and I. Volobuev, arXiv:1109.2163. J. Alwall, M. Herquet, F. Maltoni, O. Mattelaer, and T. Stelzer, J. High Energy Phys. 06, 128 (2011). T. Sjostrand, S. Mrenna, and P. Skands, J. High Energy Phys. 05, 026 (2006). T. Aaltonen [*et al.*]{} (CDF Collaboration), Phys. Rev. Lett. [**104**]{}, 141801 (2010). J. M. Campbell and R. K. Ellis, Phys. Rev. D [**60**]{}, 113006 (1999). J. M. Campbell and R. K. Ellis, Phys. Rev. D [**62**]{}, 114012 (2000). U. Langenfeld, S. Moch, and P. Uwer, Phys. Rev. D [**80**]{}, 054009 (2009). M. Mangano [*et al.*]{} J. High Energy Phys. 07, 001 (2003). B. W. Harris [*et al.*]{}, Phys. Rev. D [**66**]{}, 054024 (2002). Z. Sullivan, Phys. Rev. D [**70**]{}, 114012 (2004). D. Acosta [*et al.*]{}, Nucl. Instrum. Methods A [**494**]{}, 57 (2002). T. Aaltonen [*et al.*]{} (CDF Collaboration), Phys. Rev. D [**82**]{}, 112005 (2010).
[^1]: Deceased
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a complete classification of Dembowski-Ostrom polynomials from Dickson polynomials of arbitrary kind over finite fields of odd characteristic.'
address:
- 'Department of Mathematics, Indian Institute of Technology Jammu, Jammu 181221, India'
- 'Department of Mathematics, Indian Institute of Technology Jammu, Jammu 181221, India'
author:
- Sartaj Ul Hasan
- Mohit Pal
title: 'Dembowski-Ostrom polynomials from generalized Dickson polynomials'
---
Introduction {#S1}
============
Let $p$ be an odd prime. We shall denote, as usual, by $\mathbb F_q$, the finite fields with $q =p^e$ elements, where $e$ is a positive integer. For any non-negative integer $k$, and any element $a \in \mathbb F_q$, the $k$-th Dickson polynomial of first and second kind are given by $$D_k(X,a) =\sum_{i=0}^{\lfloor \frac{k}{2} \rfloor} \frac{k}{k-i} {k-i \choose i} (-a)^i X^{k-2i}$$ and $$E_k(X,a) =\sum_{i=0}^{\lfloor \frac{k}{2} \rfloor} {k-i \choose i} (-a)^i X^{k-2i},$$ respectively, where $a \in \mathbb{F}_q$ is a parameter. The Dickson polynomial of first kind was introduced by Dickson [@LED] in his Ph.D. thesis of 1897 and its variation, which is now called Dickson polynomial of second kind, was introduced by Schur [@IS] more than two decades later in 1923. It may be noted that these polynomials are closely related to the classical Chebyshev polynomials. We refer interested readers to the monograph [@Dickson-book] for more details about Dickson polynomials. The $k$-th reversed Dickson polynomial of the first kind was introduced in 2009 by Hou, Mullen, Sellers and Yucas [@HMSY] by interchanging the roles of $X$ and $a$ in the $k$-th Dickson polynomial of first kind. In 2013, Wang and Yucas [@WY] defined the $(m+1)$-th kind of Dickson polynomial, where $m$ is a non-negative integer as follows $$\label{domk}
D_{k,m}(X,a) =\sum_{i=0}^{\lfloor \frac{k}{2} \rfloor} \frac{k-mi}{k-i} {k-i \choose i} (-a)^i X^{k-2i},$$ where $a \in \mathbb{F}_q$ is a parameter and $D_{0,m}(X,a)=2-m$. It is clear from the definition that $D_{k,0}(X,a)= D_k(X,a)$ and $D_{k,1}(X,a)=E_k(X,a)$. The authors also gave a recurrence relation, $$\label{rr}
D_{k,m}(X,a) = mE_{k}(X,a)-(m-1)D_{k}(X,a)$$ to find the Dickson polynomial of $(m+1)$-th kind from the Dickson polynomial of first and second kind. The $k$-th reversed Dickson polynomial of $(m+1)$-th kind is also defined in the same way just by reversing the roles of the variable $X$ and the parameter $a$ in (\[domk\]), which would lead to the following expression $$\label{rdomk}
D_{k,m}(a,X) =\sum_{i=0}^{\lfloor \frac{k}{2} \rfloor} \frac{k-mi}{k-i} {k-i \choose i} (-X)^i a^{k-2i}.$$
Throughout this paper, we shall denote $D_{k,2}(X,a)$ by $F_k(X,a)$, $D_{k,3}(X,a)$ by $G_k(X,a)$ and $D_{k,4}(X,a)$ by $H_k(X,a)$, respectively. It is also clear from (\[rr\]) that for any finite field $\mathbb{F}_q$ of characteristic $p$, $D_{k,m+p}(X,a) = D_{k,m}(X,a)$ and thus, we shall restrict ourselves to the case $m<p$.
A [*Dembowski-Ostrom*]{} ($DO$) polynomial over a finite field $\mathbb F_q$ is a polynomial that admits the following shape $$\displaystyle \sum_{i,j} a_{ij}X^{p^i+p^j},$$ where $a_{ij} \in \mathbb{F}_q$. These polynomials were introduced by Peter Dembowski and Ted Ostrom [@DO]. Coulter and Matthews [@DOname] later termed these polynomials as Dembowski-Ostrom polynomials. A polynomial $f \in \mathbb{F}_q[X]$ is a [*planar function*]{} if for any $a\in \mathbb{F}_q^*$, the mapping $f_a(X)=f(a+X)-f(X)$ is a bijection from $\mathbb{F}_q$ to $\mathbb{F}_q$.
In 2010, Coulter and Matthews [@CM] completely classified $DO$ polynomials from Dickson polynomials of first and second kind over finite fields of odd characteristic. The authors also considered the problem that when the $DO$ polynomials obtained from Dickson polynomials of first and second kind are planar functions. Further results on planar $DO$ polynomials can be found in [@planar]. $DO$ polynomials from reversed Dickson polynomials over finite fields of characteristic 2 were studied by Zhang, Wu and Liu [@Zhang-Wu-Liu-2016]. Recently, Fernando, Hasan and Pal [@FHP] studied $DO$ polynomials from reversed Dickson polynomial of $(m+1)$-th kind over finite fields of odd characteristic. Motivated by these results, we consider the problem of classifying $DO$ polynomials from Dickson polynomials of $(m+1)$-th kind over finite fields of odd characteristic. Following the work of Coulter and Matthews [@CM], we take the composition of the Dickson polynomial with $X^d$, where $d$ is a positive integer and consider the polynomial $D_{k,m}(X^d,a)-D_{k,m}(0,a)$, which, in short, we shall denote by $\mathfrak D_{k,m}$. Here the difference is considered to exclude the constant term $D_{k,m}(0,a)$, which arises only when $k$ is even. We now give the structure of the paper. In Section \[S2\], we give a complete classification of $DO$ polynomials obtained from the polynomials $\mathfrak D_{k,2}$. In Section \[S3\] and \[S4\], we give a complete classification of $DO$ polynomials arising from the polynomials $\mathfrak D_{k,3}$ and $\mathfrak D_{k,4}$, respectively. The $DO$ polynomials stemming from $\mathfrak D_{k,m}$, when $m\geq 5$, have been considered in Section \[S5\]. In Section \[S7\], we give a list of all $DO$ polynomials obtained from $\mathfrak{D}_{k,m}$. It would be interesting to explore the planarity aspects of $DO$ polynomials obtained in Section \[S7\], which we plan to consider in future. However, it may be noted that irrespective of whether or not we obtain any new planar $DO$ polynomial from $\mathfrak{D}_{k,m}$, the results of this paper may still be of some independent interest to the general audience.
Dickson Polynomials of third Kind {#S2}
=================================
We know that $DO$ polynomials are closed under the left or write composition with $X^p$, therefore, it is sufficient to consider the cases when $(d,p)=1$. However, since $\displaystyle D_{kp,m}(X,a) \neq {D_{k,m}(X,a)}^p$ for $m\geq 1$, we shall not put the restriction that $(k,p)=1$. One may also note that the monomial $X^{rd}$ is $DO$ polynomial if and only if $rd=p^{\beta}(p^{\alpha }+1)$, where $\beta$ is the highest exponent of $p$ such that $p^{\beta} \mid r$. It is clear that whenever $(r,p)=1$, we must have $\beta=0$. These assumptions and conventions shall remain effective across the Sections.
We denote the polynomial $F_k(X^d,a)-F_k(0,a)$ by $\mathfrak F_k$, whose expression is given by $$\mathfrak F_{k} =\sum_{i=0}^{\lfloor \frac{k-1}{2} \rfloor} \frac{k-2i}{k-i} {k-i \choose i} (-a)^i X^{k-2i},$$ i.e., $$\label{fk}
\mathfrak F_k=X^{kd}-(k-2)aX^{(k-2)d}+\frac{(k-3)(k-4)}{2!}a^2X^{(k-4)d}- \cdots.$$ The following theorem gives the conditions on $k,d$ and $p$ for which the polynomial $\mathfrak F_{k}$ is a $DO$ polynomial.
Let $q$ be a power of odd prime $p$ and $a\in\mathbb{F}_q^*$. The polynomial $\mathfrak F_{k}$ is $DO$ polynomial over $\mathbb{F}_q$ if and only if one of the following holds.
1. $k=1$ and $d= p^n(p^{\alpha}+1)$ for non-negative integers $\alpha$ and $n$.
2. $k=2$ and $d= p^n(p^{\alpha}+1)/2$ for non-negative integers $\alpha$ and $n$.
3. $k=3$ and either
1. $p=3$ and $d= p^n(p^{\alpha}+1)$ for non-negative integers $\alpha$ and $n$; or
2. $p=5$ and $d= 2p^n$ for non-negative integer $n$.
4. $k=4$, $p=3$ and $d= p^n$ for non-negative integer $n$.
5. $k=5$ and either
1. $p=3$ and $d= 2p^n$ for non-negative integer $n$; or
2. $p=5$ and $d= 2p^n$ for non-negative integer $n$.
6. $k=6$, $p=3$ and $d= p^n$ for non-negative integer $n$.
7. $k=9$, $p=3$ and $d= 4p^n$ for non-negative integer $n$.
8. $k=12$, $p=3$ and $d= p^n$ for non-negative integer $n$.
The sufficiency of the theorem is easy to verify. So it is enough to prove the necessity.
We begin with the polynomial $\mathfrak F_{1}=X^d$ and it will be a $DO$ polynomial if $d$ is of the form $p^{\alpha}+1$. The polynomial $\mathfrak F_{2}=X^{2d}$ will be a $DO$ polynomial if $d$ is of the form $ (p^{\alpha}+1)/2$. The polynomial $\mathfrak F_{3}= X^{3d}-aX^{d}$ will be a $DO$ polynomial if $d = p^{\alpha}+1$ and $3d = p^s(p^{\beta}+1)$. Now $p^s \mid 3$ implies that either $p=3$ and $s=1$ or $p>3$ and $s=0$. In the event of $p=3$, $\mathfrak F_{3}$ will be $DO$ polynomial if $d$ is of the form $p^{\alpha}+1$. However, in case of $p>3$, we have $d = p^{\alpha}+1$ and $3d = p^{\beta}+1$ and these equations would lead to $3p^{\alpha}+2 = p^{\beta}$, which is true if and only if $\alpha = 0$, $\beta=1$ $p = 5$ and $d=2$. The polynomial $\mathfrak F_{4}=X^{4d}-2aX^{2d}$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$ and $4d = p^{\beta}+1$. Thus, we get $2p^{\alpha}+1 = p^{\beta}$, which is true if and only if $\alpha = 0$, $\beta=1$, $p = 3$ and $d=1$. For the polynomial $\mathfrak F_{5}=X^{5d}-3aX^{3d}+a^2X^{d}$, we are going to have two distinct cases, namely, $p=3$ and $p>3$. If $p=3$ then the coefficients of $X^{5d}$ and $X^d$ are nonzero. Thus, $\mathfrak F_{5}$ will be $DO$ polynomial if $d = 3^{\alpha}+1$ and $5d = 3^{\beta}+1$. Hence, $5\cdot 3^{\alpha}+4 = 3^{\beta}$, which forces $\alpha = 0$, $\beta=2$ and $d=2$. If $p>3$ then the coefficients of $X^{5d}$, $X^{3d}$ and $X^d$ are nonzero. Therefore, $\mathfrak F_{5}$ will be a $DO$ polynomial if $d = p^{\alpha}+1$, $3d = p^{\beta}+1$ and $5d = p^s(p^{\gamma}+1)$. On solving first two equations, we get $3\cdot5^{\alpha}+2 = 5^{\beta}$, which is true if and only if $\alpha = 0$, $\beta=1$, $p=5$ and $d=2$. Now putting these values into third equation, we have $10=5^s(5^{\gamma}+1)$, which gives $s=1$ and $\gamma=0$. Similarly, for the polynomial $\mathfrak F_{6}=X^{6d}-4aX^{4d}+3a^2X^{2d}$, we have two cases. If $p=3$ then the coefficients of $X^{6d}$ and $X^{4d}$ are nonzero. Thus, $\mathfrak F_{6}$ will be a $DO$ polynomial if $4d = 3^{\alpha}+1$ and $\displaystyle 6d = 3^s(3^{\beta}+1)$. Since $3^s \mid 6$, we have $s=1$ and hence, second equation reduces to $2d =3^{\beta}+1$. On solving these equations, we get $2\cdot3^{\beta}+1 = 3^{\alpha}$, which forces $\beta = 0$, $\alpha=1$ and $d=1$. If $p>3$ then the coefficients of $X^{2d}$, $X^{4d}$ and $X^{6d}$ are nonzero. In this case, $\mathfrak F_{6}$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$, $4d = p^{\beta}+1$ and $6d=p^{\gamma}+1$. On solving first two equations, we get $2\cdot p^{\alpha}+1 = p^{\beta}$, which forces $\alpha = 0$. Thus, we get $p^{\beta}=3$, which is a contradiction to our assumption that $p>3$.
For $k\geq 7$, consider the following cases.
**Case 1** $k \not\equiv 0,2~(\mbox{mod}~p)$.
In this case, the coefficients of $X^{kd}$ and $X^{(k-2)d}$ in $\mathfrak F_{k}$ are nonzero. Therefore, $\mathfrak F_{k}$ will be a $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-2)d = p^{\beta}+1$ and hence, $2d = p^{\alpha} - p^{\beta}$. Since $k \geq 7$, we have $\alpha, \beta \geq 1$, which implies that $p\mid d$. This gives us contradiction to the fact that $(p,d)=1$. Therefore, $\mathfrak F_{k}$ is not $DO$.
**Case 2** $k \equiv 2~(\mbox{mod}~p)$.
In this case, coefficient of $X^{kd}$ in $\mathfrak F_{k}$ is nonzero. Now if $p=3$ then $k\equiv 2 \pmod 3$ implies that $3\nmid (k-3)$ and $3\nmid (k-4)$ and thus, the coefficient of $X^{(k-4)d}$ is nonzero. If $p>3$ then $k\equiv 2 \pmod p$ implies that $k\not\equiv 3,4 \pmod p$ and again, the coefficient of $X^{(k-4)d}$ is nonzero. Thus, $\mathfrak F_{k}$ will be a $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. Solution to these equations yields $4d = p^{\alpha}-p^{\beta}$. Since $k \geq 7$, we have $\alpha,\beta \geq 1$, which implies that $p \mid d$. This gives a contradiction. Therefore, $\mathfrak F_{k}$ is not $DO$.
**Case 3** $k \equiv 0~(\mbox{mod}~p)$.
Here we take two subcases, namely, $p=3$ and $p>3$.
**Subcase 3.1** $p=3$.
In this case, $\mathfrak F_{9}=X^{9d}+2aX^{7d}+2a^3X^{3d}+a^4X^d$ will be a $DO$ polynomial if $9d = 3^s(3^{\alpha}+1)$, $7d = 3^{\beta}+1$, $3d=3^t(3^{\gamma}+1)$ and $d=3^{\delta}+1$. Since $3^s \mid 9$ and $3^t\mid 3$, we must have $s=2$ and $t=1$. Therefore, the first and third equations reduce to $d = 3^{\alpha}+1$ and $d=3^{\gamma}+1$. Hence, by solving first two equations, we get $3^{\beta} = 7\cdot3^{\alpha}+6$, which forces $\alpha = 1$, $ \beta = 3$ and $d = 4$. Similarly, the polynomial $\mathfrak F_{12}=X^{12d}+2aX^{10d}+a^3X^{6d}+2a^4X^{4d}$ will be a $DO$ polynomial if $12d = 3^s(3^{\alpha}+1)$, $10d = 3^{\beta}+1$, $6d~=~3^t(3^{\gamma}+1)$ and $4d = 3^{\delta}+1$. Now $3^s \mid 12$ and $3^t \mid 6$ implies that $s=1$ and $t=1$, respectively. Therefore, the first and third equations reduce to $4d = 3^{\alpha}+1$ and $2d = 3^{\gamma}+1$, respectively. Now solving the second and third equations, we get $3^{\beta} = 5\cdot 3^{\gamma}+4$, which forces $\gamma =0$ that, in turn, gives $d=1$, $\beta = 2$ and $\alpha = \delta = 1$. In the rest of this subcase, we shall prove that the polynomial $\mathfrak F_{k}$ is never $DO$ whenever $k\geq15$ and $k\equiv0 \pmod 3$. Suppose on the contrary that $\mathfrak F_{k}$ is $DO$. Since $k \equiv 0~(\mbox{mod}~3)$, therefore $k \not\equiv 1,2~(\mbox{mod}~3)$, hence the coefficients of $X^{kd}$ and $X^{(k-2)d}$ are nonzero. Since $\mathfrak F_{k}$ is $DO$ by our assumption, therefore, $kd=3^s(3^{\alpha}+1)$ and $(k-2)d=3^{\beta}+1$. Notice that since $k\geq 15$, we must have $\beta \geq 3$, which implies that $d\equiv1 \pmod 3$. From the first equation, since $3^s\mid k$, therefore, $k=u\cdot 3^s$ for some positive integers $u$ and $s$. Putting this value of $k$ in the first equation, we obtain $ud=3^{\alpha}+1$. We now consider two cases, namely, $\alpha =0$ and $\alpha \geq 1$. In the case $\alpha =0$, we have $ud=2$. Therefore, either $d=2$ and $u=1$ or $d=1$ and $u=2$. First case is not possible as $d\equiv1 \pmod 3$. In the case of $d=1$ and $u=2$, we have $k=2\cdot3^s$ and $k-2=3^{\beta}+1$. Solving these two equations, we have $2\cdot3^{s-1}-3^{\beta-1}=1$. Again we have two cases, namely, $s=1$ and $s\geq2$. If $s=1$, we have $3^{\beta-1}=1$ which is a contradiction as $\beta \geq 3$. In the case of $s\geq 2$, we have $2\cdot3^{s-1}-3^{\beta-1}=1$ and taking modulo $3$, we get $0\equiv1 \pmod 3$, a contradiction. In the case $\alpha\geq 1$, equation $ud=3^{\alpha}+1$ implies that $ud\equiv 1 \pmod 3$ and since $d\equiv 1 \pmod 3$, therefore $u\equiv 1 \pmod 3$. We now show that in this case, $k\not \equiv 6 \pmod 9$. Recall that $k=u \cdot 3^s$. If $u\equiv 1 \pmod 3$ then $u=3n+1$ for some non-negative integer $n$. If $n=0$, then $u=1$ and $k=3^s$. But for $k\geq15$, we must have $s\geq3$, which implies that $k\equiv0 \pmod 9$ and hence, $k \not \equiv 6 \pmod 9$. In the case $n\geq1$, we have $$k=u\cdot 3^s = (3n+1)3^s = n\cdot 3^{s+1}+3^s.$$ If $s\geq2$ then $k\equiv 0 \pmod 9$ and if $s=1$ then $k=9n+3$, which implies that $k\equiv 3 \pmod 9$. Therefore $k\not \equiv 6 \pmod 9$.
We now show that the fifth term in $\mathfrak F_k$ given by $$\frac{(k-5)(k-6)(k-7)(k-8)}{4!}a^4X^{(k-8)d}$$ will exist. Note that $k\equiv 0 \pmod 3$ implies that $3 \nmid (k-5)$, $3\nmid (k-7)$ and $3\nmid (k-8)$. Moreover, since we have already seen that $k\not\equiv6~(\mbox{mod}~9)$, therefore the highest power of $3$ that can divide the numerator of the coefficient of $X^{(k-8)d}$ is $1$ and the highest power of $3$ that divides $4!$ is $1$. Therefore, the coefficient of $X^{(k-8)d}$ is nonzero. Since $\mathfrak F_{k}$ is $DO$ polynomial, therefore $(k-2)d=3^{\beta}+1$ and $(k-8)d=3^{\gamma}+1$. Now since $k\geq15$, therefore $\beta, \gamma \geq 2$. On solving these two equations, we have $6d = 3^{\beta}-3^{\alpha}$ but then $3\mid d$ which will be a contradiction as $(d,3)=1$. Thus, our assumption that $\mathfrak F_{k}$ is $DO$ is wrong. This proves the desired result.
**Subcase 3.2** $p>3$.
In this case, $k\not\equiv2,3,4~(\mbox{mod}~p)$. Therefore, the coefficients of $X^{(k-2)d}$ and $X^{(k-4)d}$ are nonzero. Thus, $\mathfrak F_{k}$ will be a $DO$ polynomial if $(k-2)d=p^{\alpha}+1$ and $(k-4)d=p^{\beta}+1$. On solving these two equations, we have $2d=p^{\alpha}-p^{\beta}$. Since $k\geq7$, we must have $\alpha, \beta \geq1$, which implies that $p\mid d$. This leads to a contradiction. Therefore, $\mathfrak F_{k}$ is not $DO$ in this case.
Dickson Polynomials of fourth Kind {#S3}
===================================
We denote the polynomial $G_k(X^d,a)-G_k(0,a)$ by $\mathfrak G_k$. Then $$\mathfrak G_{k} =\sum_{i=0}^{\lfloor \frac{k-1}{2} \rfloor} \frac{k-3i}{k-i} {k-i \choose i} (-a)^i X^{k-2i}$$ i.e., $$\mathfrak G_k=X^{kd}-(k-3)aX^{(k-2)d}+\frac{(k-3)(k-6)}{2!}a^2X^{(k-4)d}- \frac{(k-4)(k-5)(k-9)}{3!}a^3X^{(k-6)d}+\cdots.$$ \[gk\]
Since, for $p=3$, $\mathfrak G_k = \mathfrak D_{k,0}$. Thus, for the case $p=3$, $\mathfrak G_k$ will be a $DO$ polynomial whenever $\mathfrak D_{k,0}$ is $DO$. Consequently, the case $p=3$, follows immediately from [@CM Theorem 2.1]. For $p\geq5$, we prove the following theorem.
Let $q$ be a power of odd prime $p\geq5$. The polynomial $\mathfrak G_k$ is $DO$ polynomial over $\mathbb{F}_q$ if and only if one of the following holds.
1. $k=1$ and $d= p^n(p^{\alpha}+1)$ for non-negative integers $\alpha$ and $n$.
2. $k=2$ and $d= p^n(p^{\alpha}+1)/2$ for non-negative integers $\alpha$ and $n$.
3. $k=3$ and $d= p^n(p^{\alpha}+1)/3$ for non-negative integers $\alpha$ and $n$.
4. $k=5$, $p=5$ and $d= 2p^n$ for non-negative integer $n$.
It is enough to prove only necessity of this theorem. Note that $\mathfrak G_1=X^d$, $\mathfrak G_2=X^{2d}$ and $\mathfrak G_3=X^{3d}$ will be $DO$ polynomial if $d$ is of the form $p^{\alpha}+1$, $(p^{\alpha}+1)/2$ and $(p^{\alpha}+1)/3$, respectively. The polynomial $\mathfrak G_4=X^{4d}-aX^{2d}$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$ and $4d = p^{\beta}+1$. Hence, $2p^{\alpha}+1 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=3$, a contradiction as $p>3$. Therefore, $\mathfrak G_4$ is not $DO$. The polynomial $\mathfrak G_5=X^{5d}-2aX^{3d}-a^2X^{d}$ will be $DO$ if $d = p^{\alpha}+1$ , $3d = p^{\beta}+1$ and $5d = p^s(p^{\gamma}+1)$. Solving first two equations, we have $3p^{\alpha}+2 = p^{\beta}$, which is true if and only if $\alpha = 0$, $\beta=1$, $p=5$ and $d=2$. Now putting these values into third equation, we have $10=5^s(5^{\gamma}+1)$, which gives $s=1$ and $\gamma=0$. The polynomial $\mathfrak G_6=X^{6d}-3aX^{4d}$ will be $DO$ if $4d = p^{\alpha}+1$ and $6d = p^{\beta}+1$. On solving these two equations, we have $3p^{\alpha}+1 = 2p^{\beta}$ which forces $\alpha = 0$ and $p^{\beta}=2$. This gives a contradiction. Therefore, $\mathfrak G_6$ is not $DO$. The polynomial $\mathfrak G_7=X^{7d}-4aX^{5d}+2a^2X^{3d}+2a^3X^{d}$ will be a $DO$ polynomial if $d = p^{\alpha}+1$, $3d = p^{\beta}+1$, $5d = p^s(p^{\gamma}+1)$ and $7d = p^t(p^{\delta}+1)$. On solving first two equations, we have $3p^{\alpha}+2 = p^{\beta}$, which is true if and only if $\alpha = 0$, $\beta=1$, $d=2$ and $p=5$. Now putting this into fourth equation, we get $t=0$ and $5^{\delta}=13$, which is a contradiction. Therefore, $\mathfrak G_7$ is not $DO$. For the polynomial $\mathfrak G_8=X^{8d}-5aX^{6d}+5a^2X^{4d}+2a^3X^{2d}$, we shall consider two cases, namely, $p=5$ and $p>5$. In the case $p=5$, coefficients of $X^{2d}$ and $X^{8d}$ are nonzero. Thus, $\mathfrak G_8$ will be a $DO$ polynomial if $2d = 5^{\alpha}+1$ and $8d = 5^{\beta}+1$. On solving the two equations, we have $4.5^{\alpha}+3 = 5^{\beta}$, which forces $\alpha = 0$ and $5^{\beta}=7$. This leads to a contradiction. Therefore, $\mathfrak G_8$ is not $DO$ polynomial in this case. For the case $p>5$, the coefficients of $X^{2d}$, $X^{4d}$, $X^{6d}$ and $X^{8d}$ in $\mathfrak G_8$ are nonzero. Thus, $\mathfrak G_8$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$, $4d = p^{\beta}+1$, $6d = p^{\gamma}+1$ and $8d = p^{\delta}+1$. On solving first two equations, we have $2p^{\alpha}+1 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=3$. Thus, we arrive at a contradiction. Therefore, $\mathfrak G_8$ is not $DO$.
For $k\geq 9$, we proceed by doing the following cases.
**Case 1** $k \not\equiv 0,2,3~(\mbox{mod}~p)$.
In this case, the coefficients of $X^{kd}$ and $X^{(k-2)d}$ are nonzero. Therefore, $\mathfrak G_k$ will be a $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-2)d = p^{\beta}+1$. On solving these two equations, we have $2d = p^{\alpha} - p^{\beta}$. Since $k \geq 9$, we must have $\alpha, \beta \geq 1$, which implies that $p \mid d$. This leads to a contradiction.
**Case 2** $k \equiv 0~(\mbox{mod}~p)$.
In this case, $k \not\equiv 2,3,4,6~(\mbox{mod}~p)$ and hence, the coefficients of $X^{(k-2)d}$ and $X^{(k-4)d}$ are nonzero. Thus, $\mathfrak G_k$ will be a $DO$ polynomial if $(k-2)d = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. Solving these two equations, we have $2d = p^{\alpha}-p^{\beta}$. Since $k \geq 9$, we have $\alpha, \beta \geq 1$, which implies that $p\mid d$. This gives a contradiction.
**Case 3** $k \equiv 2~(\mbox{mod}~p)$.
In this case, $k \not\equiv 0,3,4,6~(\mbox{mod}~p)$ and hence, the coefficients of $X^{kd}$ and $X^{(k-4)d}$ are nonzero. Therefore, $\mathfrak G_k$ will be $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. Solving these two equations, we have $4d = p^{\alpha}-p^{\beta}$. Since $k \geq 9$, we must have $\alpha, \beta \geq 1$, which implies that $p \mid d$. Thus, we arrive at a contradiction.
**Case 4** $k \equiv 3~(\mbox{mod}~p)$.
In this case $k \not\equiv 0,4,5,6,9~(\mbox{mod}~p)$ and hence the coefficients of $X^{kd}$ and $X^{(k-6)d}$ are nonzero, therefore $\mathfrak G_k$ will be $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-6)d = p^{\beta}+1$. Solving these two equations we have $6d = p^{\alpha}-p^{\beta}$. Since $k \geq 9$, therefore $\alpha, \beta \geq 1$ which implies that $p \mid d$, a contradiction.
Dickson polynomials of fifth Kind {#S4}
==================================
We denote the polynomial $H_k(X^d,a)-H_k(0,a)$ by $\mathfrak H_k$. Then $$\mathfrak H_{k} =\sum_{i=0}^{\lfloor \frac{k-1}{2} \rfloor} \frac{k-4i}{k-i} {k-i \choose i} (-a)^i X^{k-2i}$$ i.e., $$\mathfrak H_k=X^{kd}-(k-4)aX^{(k-2)d}+\frac{(k-3)(k-8)}{2!}a^2X^{(k-4)d}- \cdots.$$ \[hk\] Since, for $p=3$, $\mathfrak H_k = \mathfrak D_{k,1}$. Thus, for $p=3$, $\mathfrak H_k$ will be $DO$ polynomial whenever $\mathfrak D_{k,1}$ is $DO$. As a consequence, the case $p=3$ follows immediately from [@CM Theorem 3.1]. For $p\geq5$, we prove the following theorem.
Let $q$ be a power of odd prime $p\geq5$. The polynomial $\mathfrak H_k$ is $DO$ polynomial over $\mathbb{F}_q$ if and only if one of the following holds.
1. $k=1$ and $d= p^n(p^{\alpha}+1)$ for non-negative integers $\alpha$ and $n$.
2. $k=2$ and $d= p^n(p^{\alpha}+1)/2$ for non-negative integers $\alpha$ and $n$.
3. $k=4$ and $d= p^n(p^{\alpha}+1)/4$ for non-negative integers $\alpha$ and $n$.
4. $k=3,5$, $p=5$ and $d= 2p^n$ for non-negative integer $n$.
The sufficiency of the theorem is straightforward. It only remains to show the necessity. It may be noted that $\mathfrak H_1=X^{d}$, $\mathfrak H_2=X^{2d}$ and $\mathfrak H_4=X^{4d}$ will be a $DO$ polynomial if $d$ is of the form $p^{\alpha}+1$, $(p^{\alpha}+1)/2$ and $(p^{\alpha}+1)/4$, respectively. The polynomial $\mathfrak H_3=X^{3d}+aX^{d}$ will be a $DO$ polynomial if $d = p^{\alpha}+1$ and $3d = p^{\beta}+1$. Hence, $3p^{\alpha}+2 = p^{\beta}$, which is true if and only if $\alpha = 0$, $\beta=1$, $p=5$ and $d=2$. Again, the polynomial $\mathfrak H_5=X^{5d}-aX^{3d}-3a^2X^{d}$ will be a $DO$ polynomial if $d = p^{\alpha}+1$ , $3d = p^{\beta}+1$ and $5d = p^s(p^{\gamma}+1)$. Solving first two equations, we have $3p^{\alpha}+2 = p^{\beta}$, which is true if and only if $\alpha = 0$, $\beta=1$, $p=5$ and $d=2$. Now putting these values into third equation, we have $10=5^s(5^{\gamma}+1)$, which gives $s=1$ and $\gamma=0$. The polynomial $\mathfrak H_6=X^{6d}-2aX^{4d}-3a^2X^{2d}$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$, $4d = p^{\beta}+1$ and $6d = p^{\gamma}+1$. On solving first two equations, we get $2p^{\alpha}+1 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=3$ and this leads to a contradiction. Therefore, $\mathfrak H_6$ is not $DO$. For the polynomial $\mathfrak H_7=X^{7d}-3aX^{5d}-2a^2X^{3d}+5a^3X^{d}$, we consider the two cases, namely, $p=5$ and $p>5$. For the case $p=5$, $\mathfrak H_7$ will be DO if $3d = 5^{\alpha}+1$, $5d = 5^s(5^{\beta}+1)$ and $7d =5^{\gamma}+1$. Solving first and third equations, we have $7\cdot5^{\alpha}+4=3\cdot5^{\gamma}$, which forces $\alpha=0$ and $3\cdot5^{\gamma}=11$, a contradiction.
For the case, $p>5$, $\mathfrak H_7$ will be DO if $d = p^{\alpha}+1$, $3d = p^{\beta}+1$, $5d =(p^{\gamma}+1)$ and $7d = p^t(p^{\delta}+1)$. On solving first two equations, we have $3p^{\alpha}+2 = p^{\beta}$ which forces $\alpha = 0$, and $p^{\beta}=5$, which is a contradiction. Again the polynomial $\mathfrak H_8=X^{8d}-4aX^{6d}+8a^3X^{2d}$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$, $6d=p^{\beta}+1$ and $8d = p^{\gamma}+1$. On solving first two equation, we have $3p^{\alpha}+2 = p^{\beta}$ which is true if and only if $\alpha = 0$, $\beta=1$, $d=1$ and $p=5$. Now putting these values into third equation we have $5^{\gamma}=7$, which is a contradiction. For the polynomial $\mathfrak H_9=X^{9d}-5aX^{7d}+3a^2X^{5d}+10a^3X^{3d}-7a^4X^{d}$, we shall consider two cases, namely, $p=5$ and $p>5$. For the case $p=5$, the coefficients of $X^{d}$, $X^{5d}$ and $X^{9d}$ in $\mathfrak H_9$ are nonzero. Thus, $\mathfrak H_9$ will be a $DO$ polynomial if $d = 5^{\alpha}+1$, $5d = 5^s(5^{\beta}+1)$ and $9d = 5^{\gamma}+1$. Since $5^s \mid 5$, we have $s=1$. Thus, the second equation reduces to $d=5^{\beta}+1$. On solving first and third equation, we have $9\cdot 5^{\alpha}+8 = 5^{\beta}$, which forces $\alpha = 0$, $5^{\gamma}=17$, a contradiction. For the case $p>5$, the coefficients of $X^{3d}$, $X^{5d}$ and $X^{9d}$ are nonzero. Thus, $\mathfrak H_9$ will be a $DO$ polynomial if $3d = p^{\alpha}+1$, $5d = p^{\beta}+1$, and $9d = p^{\gamma}+1$. On solving first two equation, we have $5p^{\alpha}+2 = 3p^{\beta}$, which forces $\alpha = 0$ and $3p^{\beta}=7$. This gives a contradiction. Similarly, for the polynomial $\mathfrak H_{10}=X^{10d}-6aX^{8d}+7a^2X^{6d}+10a^3X^{4d}-15a^4X^{2d}$, we shall consider two cases, namely, $p=5$ and $p>5$. For the case $p=5$, the coefficients of $X^{6d}$, $X^{8d}$ and $X^{10d}$ are nonzero. Thus $\mathfrak H_{10}$ will be a $DO$ polynomial if $6d = 5^{\alpha}+1$, $8d=5^{\beta}+1$ and $10d=5^s(5^{\gamma}+1)$. On solving first two equations, we have $4.5^{\alpha}+1 = 3.5^{\beta}$, which forces $\alpha = 0$ and $3.5^{\beta}=5$ and this leads to a contradiction. If $p>5$ then the coefficients of $X^{2d}$ and $X^{4d}$ are nonzero. Therefore, $\mathfrak H_{10}$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$ and $4d=p^{\beta}+1$. On solving these two equations, we have $2p^{\alpha}+1 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=3$. This gives a contradiction and therefore, $\mathfrak H_{10}$ is not $DO$.
For $k\geq 11$, we proceed by considering various cases.
**Case 1** $k \not\equiv 0,2,4~(\mbox{mod}~p)$.
In this case, the coefficients of $X^{kd}$ and $X^{(k-2)d}$ in $\mathfrak H_k$ are nonzero. Therefore, $\mathfrak H_k$ will be a $DO$ polynomial if $kd = p^{\alpha}+1$ then $(k-2)d = p^{\beta}+1$. On solving these two equation we have $2d = p^{\alpha} - p^{\beta}$. Since $k \geq 11$, we have $\alpha, \beta \geq 1$, which implies that $p \mid d$. This leads to a contradiction. Therefore, $\mathfrak H_{k}$ is not $DO$.
**Case 2** $k \equiv 0~(\mbox{mod}~p)$.
In this case, $k \not\equiv 2,3,4,8~(\mbox{mod}~p)$ and hence, the coefficients of $X^{(k-2)d}$ and $X^{(k-4)d}$ in $\mathfrak H_{k}$ are nonzero. Therefore, $\mathfrak H_k$ will be $DO$ polynomial if $(k-2)d = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. Solving these two equations we have $2d = p^{\alpha}-p^{\beta}$. Since $k \geq 11$, we must have $\alpha, \beta \geq 1$, which implies that $p \mid d$. This gives a contradiction, therefore, $\mathfrak H_{k}$ is not $DO$.
**Case 3** $k \equiv 2~(\mbox{mod}~p)$.
In this case, $k \not\equiv 0,3,4,8~(\mbox{mod}~p)$ and hence, the coefficients of $X^{kd}$ and $X^{(k-4)d}$ in $\mathfrak H_k$ are nonzero. Therefore, $\mathfrak H_k$ will be $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. Solving these two equations we have $4d = p^{\alpha}-p^{\beta}$. Since $k \geq 11$, we have $\alpha, \beta \geq 1$ which implies that $p \mid d$. This leads to a contradiction, therefore, $\mathfrak H_{k}$ is not $DO$.
**Case 4** $k \equiv 4~(\mbox{mod}~p)$.
Consider the fifth term in $\mathfrak H_k$ given by $$\frac{(k-5)(k-6)(k-7)(k-16)}{4!}a^4X^{(k-8)d}.$$ Since, $k \not\equiv 0,5,6,7,8,16~(\mbox{mod}~p)$ and hence, the coefficients of $X^{kd}$ and $X^{(k-8)d}$ in $\mathfrak H_k$ are nonzero. Therefore, $\mathfrak H_k$ will be $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-8)d = p^{\beta}+1$. Solving these two equations, we have $8d = p^{\alpha}-p^{\beta}$. Since $k \geq 11$, we have $\alpha, \beta \geq 1$ which implies that $p \mid d$, a contradiction. Therefore, $\mathfrak H_{k}$ is not $DO$.
The case $m\geq 5$ {#S5}
===================
As alluded to in the Introduction, we denote the polynomial $D_{k,m}(X^d,a)-D_{k,m}(0,a)$ by $\mathfrak D_{k,m}$, whose expression is given by $$\mathfrak D_{k,m} =\sum_{i=0}^{\lfloor \frac{k-1}{2} \rfloor} \frac{k-mi}{k-i} {k-i \choose i} (-a)^i X^{k-2i}$$ i.e., $$\label{dkm}
\mathfrak D_{k,m}=X^{kd}-(k-m)aX^{(k-2)d}+\frac{(k-3)(k-2m)}{2!}a^2X^{(k-4)d}-\cdots.$$
In this section, we find conditions on $k$, $m$, $p$ and $d$ for which the polynomial $\mathfrak D_{k,m}$ is $DO$ polynomial. For the case $p=3$, we have either $m\equiv0 ~(\mbox{mod}~3)$ or $m\equiv1 ~(\mbox{mod}~3)$ or $m\equiv2~(\mbox{mod}~3)$. Therefore, in these cases, the polynomial $\mathfrak D_{k,m}$ will be a $DO$ polynomial whenever $\mathfrak D_{k,0}$, $\mathfrak D_{k,1}$ and $\mathfrak F_{k}$ are $DO$, respectively. Thus, for the case $p=3$, the classification follows from [@CM Theorem 2.1 and 3.1] and from the Section \[S2\] of this paper, respectively. Similarly, for $p=5$, we have either $m\equiv0~(\mbox{mod}~5)$ or $m\equiv1~(\mbox{mod}~5)$ or $m\equiv2~(\mbox{mod}~5)$ or $m\equiv3~(\mbox{mod}~5)$ or $m\equiv4~(\mbox{mod}~5)$. In these cases, the polynomial $\mathfrak D_{k,m}$ will be a $DO$ polynomial whenever $\mathfrak D_{k,0}$, $\mathfrak D_{k,1}$, $\mathfrak F_{k}$, $\mathfrak G_k$ and $\mathfrak H_k$ are $DO$, respectively. In the case $p>5$, the cases $m\equiv 0,1,2,3,4~(\mbox{mod}~p)$ are already classified in [@CM Theorem 2.1 and 3.1] and in Section \[S2\], \[S3\] and \[S4\] of this paper, respectively. For the case $p>5$ and $m\not \equiv 0,1,2,3,4 ~(\mbox{mod}~p)$, we prove the following theorem.
Let $q$ be a power of odd prime $p>5$ and $m\not\equiv0,1,2,3,4 \pmod p$. The polynomial $\mathfrak D_{k,m}$ is $DO$ polynomial over $\mathbb{F}_q$ if and only if one of the following holds.
1. $k=1$ and $d= p^n(p^{\alpha}+1)$ for non-negative integers $\alpha$ and $n$.
2. $k=2$ and $d= p^n(p^{\alpha}+1)/2$ for non-negative integers $\alpha$ and $n$.
We shall prove the necessity of this theorem as the sufficiency is straightforward. Notice that $\mathfrak D_{1,m}=X^d$ and $\mathfrak D_{2,m}=X^{2d}$ will be $DO$ polynomial if $d$ is of the form $p^{\alpha}+1$ and $(p^{\alpha}+1)/2$, respectively. For the polynomial $\mathfrak D_{3,m}=X^{3d}+(m-3)aX^d$, the coefficient of $X^{d}$ would be nonzero. Otherwise, $m\equiv 3 \pmod p$, which will lead to a contradiction to fact that $m\not\equiv 3 \pmod p$. Therefore, $\mathfrak D_{3,m}$ will be a $DO$ polynomial if $d = p^{\alpha}+1$ and $3d = p^{\beta}+1$. On solving these equations, we have $3p^{\alpha}+2 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=5$. This gives a contradiction and therefore, $\mathfrak D_{3,m}$ is not $DO$. For the polynomial $\mathfrak D_{4,m}=X^{4d}+(m-4)aX^{2d}$, the coefficient of $X^{2d}$ is nonzero. Therefore, $\mathfrak D_{4,m}$ will be $DO$ if $2d = p^{\alpha}+1$ and $4d = p^{\beta}+1$. On solving these equations, we have $2p^{\alpha}+1 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=3$. This leads to a contradiction. Therefore, $\mathfrak D_{4,m}$ is not $DO$. Note that in the polynomial $\mathfrak D_{5,m}=X^{5d}+(m-5)aX^{3d}+(5-2m)a^2X^{d}$, the coefficients of $X^d$ and $X^{3d}$ can not be simultaneously zero. Otherwise, $2m\equiv 5 \pmod p$ and $m\equiv 5 \pmod p$ would imply that $m\equiv0 \pmod p$, which is a contradiction as $m\not\equiv 0 \pmod p$. In the case, when $m\equiv 5 \pmod p$, $\mathfrak D_{5,m}$ will be a $DO$ polynomial if $d = p^{\alpha}+1$ and $5d = p^{\beta}+1$. On solving these equations, we have $5p^{\alpha}+4 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=9$. This gives a contradiction. Therefore, $\mathfrak D_{5,m}$ is not $DO$. Further, in the case, when $m \not\equiv 5~(\mbox{mod}~p)$, $\mathfrak D_{5,m}$ will be a $DO$ polynomial if $3d = p^{\alpha}+1$ and $5d = p^{\beta}+1$. On solving these equations, we obtain $5p^{\alpha}+2 = 3p^{\beta}$, which forces $\alpha = 0$ and $3p^{\beta}=7$. This leads to a contradiction and therefore, $\mathfrak D_{5,m}$ is not $DO$. For the polynomial $\mathfrak D_{6,m}= X^{6d}+(m-6)aX^{4d}+3(3-m)a^2X^{2d}$, the coefficient of $X^{2d}$ would be nonzero. Otherwise, $m\equiv3 \pmod p$, which is a contradiction. Therefore, $\mathfrak D_{6,m}$ will be a $DO$ polynomial if $2d = p^{\alpha}+1$ and $6d = p^{\beta}+1$. On solving these equations, we get $3p^{\alpha}+2 = p^{\beta}$, which forces $\alpha = 0$ and $p^{\beta}=5$ and we arrive at a contradiction. Therefore, $\mathfrak D_{6,m}$ is not $DO$.
Now for $k \geq 7$, we have following cases.
**Case 1** $k \not\equiv 0,2,m~(\mbox{mod}~p)$.
In this case, the coefficients of $X^{kd}$ and $X^{(k-2)d}$ are nonzero. Therefore, $\mathfrak D_{k,m}$ will be a $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-2)d = p^{\beta}+1$. On solving these equations, we obtain $2d =p^{\alpha}-p^{\beta}$. Note that we must have for $k\geq7$, $\alpha, \beta \geq 1$, which forces $p \mid d$. Thus, we arrive at a contradiction. Therefore, $\mathfrak D_{k,m}$ is not $DO$.
**Case 2** $k\equiv0~(\mbox{mod}~p)$.
In this case, $k\not\equiv 2,3,4~(\mbox{mod}~p)$. Also, we observe that $k \not\equiv m~(\mbox{mod}~p)$, otherwise, $m\equiv0~(\mbox{mod}~p)$, which is a contradiction. Similarly, $k \not\equiv 2m~(\mbox{mod}~p)$, otherwise $m\equiv0~(\mbox{mod}~p)$, which is again a contradiction. Therefore, in this case, the coefficients of $X^{(k-2)d}$ and $X^{(k-4)d}$ are nonzero. Thus, $\mathfrak D_{k,m}$ will be a $DO$ polynomial if $(k-2)d = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. On solving these equations, we have $2d =p^{\alpha}-p^{\beta}$. For $k\geq7$, we must have $\alpha, \beta \geq 1$, which forces $p \mid d$. This leads to a contradiction. Therefore, $\mathfrak D_{k,m}$ is not $DO$.
**Case 3** $k\equiv2~(\mbox{mod}~p)$.
In this case, $k\not\equiv 0,3~(\mbox{mod}~p)$. Also, we observe that $k \not\equiv 2m~(\mbox{mod}~p)$. Otherwise $m\equiv1~(\mbox{mod}~p)$, which would lead to a contradiction. In this case, the coefficients of $X^{kd}$ and $X^{(k-4)d}$ in $\mathfrak D_{k,m}$ are nonzero. Therefore, $\mathfrak D_{k,m}$ will be a $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. On solving these equations, we have $4d =p^{\alpha}-p^{\beta}$. Now for $k\geq7$, we must have $\alpha, \beta \geq 1$, which forces $p \mid d$. Thus, we arrive at a contradiction, Hence, $\mathfrak D_{k,m}$ is not $DO$.
**Case 4** $k \equiv m~(\mbox{mod}~p)$.
In this case, $k \not\equiv 0,3,4~(\mbox{mod}~p)$. Also, we observe that $k \not\equiv 2m~(\mbox{mod}~p)$. Otherwise $m\equiv0~(\mbox{mod}~p)$, which would lead to a contradiction. In this case, the coefficients of $X^{kd}$ and $X^{(k-4)d}$ in $\mathfrak D_{k,m}$ are nonzero. Therefore, $\mathfrak D_{k,m}$ will be a $DO$ polynomial if $kd = p^{\alpha}+1$ and $(k-4)d = p^{\beta}+1$. On solving these equations, we have $4d =p^{\alpha}-p^{\beta}$. For $k\geq7$, we have $\alpha, \beta \geq 1$, which forces $p \mid d$. This gives a contradiction. Therefore, $\mathfrak D_{k,m}$ is not $DO$.
$DO$ polynomials from Dickson polynomials of $(m+1)$-th kind {#S7}
=============================================================
The complete list of $DO$ polynomials obtained from the polynomials $\mathfrak D_{k,m}$ over a finite field $\mathbb F_q$ of odd characteristic $p$ is given as follows.
1. For $p=3$
1. When $m \equiv 0~(\mbox{mod}~3)$ [@CM Theorem 2.1]
1. $k=p^{\ell}$ then $X^{3^{n+\ell}(3^{\alpha}+1)}$ for non-negative integers $\alpha$, $n$ and $\ell$.
2. $k=2p^{\ell}$ then $X^{3^{n+\ell}(3^{\alpha}+1)}$ for non-negative integers $\alpha$, $n$ and $\ell$.
3. $k=4p^{\ell}$ then $X^{4\cdot 3^{n+\ell}}+2aX^{2\cdot 3^{n+\ell}}$ for non-negative integers $n$ and $\ell$.
4. $k=5p^{\ell}$ then $X^{10\cdot 3^{n+\ell}}+aX^{2 \cdot 3^{n+\ell+1}}+2a^{2}X^{2\cdot 3^{n+\ell}}$ for non-negative integers $n$ and $\ell$.
2. When $m \equiv 1~(\mbox{mod}~3)$ [@CM Theorem 3.1]
1. $k=1,2,4$ then $X^{3^n(3^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
2. $k=3$ then $X^{3^{n+1}(3^{\alpha}+1)}+aX^{3^n(3^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
3. $k=5$ then $X^{10\cdot 3^{n}}+2aX^{2\cdot 3^{n+1}}$ for non-negative integer $n$.
4. $k=6$ then $X^{2\cdot 3^{n+1}}+aX^{4\cdot 3^n}$ for non-negative integer $n$.
5. $k=7$ then $X^{28\cdot 3^{n}}+a^2X^{4\cdot 3^{n+1}}+2a^3X^{4\cdot 3^{n}}$ for non-negative integer $n$.
6. $k=9$ then $X^{4\cdot 3^{n+2}}+aX^{28\cdot 3^{n}}+a^3X^{4\cdot 3^{n+1}}+2a^4X^{4\cdot 3^{n}}$ for non-negative integer $n$.
7. $k=10$ then $X^{10\cdot 3^{n}}+a^2X^{2\cdot 3^{n+1}}+a^3X^{4\cdot 3^{n}}$ for non-negative integer $n$.
8. $k=12$ then $X^{4\cdot 3^{n+1}}+aX^{10\cdot 3^{n}}+a^4X^{4\cdot 3^{n}}$ for non-negative integer $n$.
3. When $m \equiv 2~(\mbox{mod}~3)$
1. $k=1,2$ then $X^{3^n(3^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
2. $k=3$ then $X^{3^{n+1}(3^{\alpha}+1)}+2aX^{3^n(3^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
3. $k=4$ then $X^{4\cdot 3^{n}}+aX^{2\cdot 3^{n}}$ for non-negative integer $n$.
4. $k=5$ then $X^{10\cdot 3^{n}}+a^2X^{2\cdot 3^n}$ for non-negative integer $n$.
5. $k=6$ then $X^{2\cdot 3^{n+1}}+2aX^{4\cdot 3^{n}}$ for non-negative integer $n$.
6. $k=9$ then $X^{4\cdot 3^{n+2}}+2aX^{28\cdot 3^{n}}+2a^3X^{4\cdot 3^{n+1}}+a^4X^{4\cdot 3^{n}}$ for non-negative integer $n$.
7. $k=12$ then $X^{4\cdot 3^{n+1}}+2aX^{10\cdot 3^{n}}+a^3X^{2\cdot 3^{n+1}}+2a^4X^{4\cdot 3^{n}}$ for non-negative integer $n$.
2. For $p=5$
1. When $m \equiv 0~(\mbox{mod}~5)$ [@CM Theorem 2.1]
1. $k=5^{\ell}$ then $\displaystyle X^{5^{n+\ell}(5^{\alpha}+1)}$ for non-negative integers $\alpha$, $n$ and $\ell$.
2. $k=2\cdot 5^{\ell}$ then $X^{5^{n+\ell}(5^{\alpha}+1)}$ for non-negative integers $\alpha$, $n$ and $\ell$.
3. $k=3\cdot 5^{\ell}$ then $X^{6\cdot 5^{n+\ell}}+2aX^{2\cdot 5^{n+\ell}}$ for non-negative integers $n$ and $\ell$.
2. When $m \equiv 1~(\mbox{mod}~5)$ [@CM Theorem 3.1]
1. $k=1,2$ then $X^{5^n(5^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
2. $k=3$ then $X^{6\cdot 5^{n}}+3aX^{2\cdot 5^n}$ for non-negative integers $\alpha$ and $n$.
3. $k=5$ then $X^{2\cdot 5^{n+1}}+aX^{6\cdot 5^{n}}+3a^2X^{2\cdot 5^{n}}$ for non-negative integer $n$.
4. $k=9$ then $X^{6\cdot 5^{n}}+a^2X^{2\cdot 5^{n}}$ for non-negative integer $n$.
3. When $m \equiv 2~(\mbox{mod}~5)$
1. $k=1,2$ then $X^{5^n(5^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
2. $k=3$ then $X^{6\cdot 5^{n}}+4aX^{2\cdot 5^{n}}$ for non-negative integer $n$.
3. $k=5$ then $X^{2\cdot 5^{n+1}}+2aX^{6\cdot 5^{n}}+a^2X^{2\cdot 5^{n}}$ for non-negative integer $n$.
4. When $m \equiv 3~(\mbox{mod}~5)$
1. $k=1,2,3$ then $X^{5^n(5^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
2. $k=5$ then $X^{2\cdot 5^{n+1}}+3aX^{6\cdot 5^{n}}+4a^2X^{2\cdot 5^{n}}$ for non-negative integer $n$.
5. When $m \equiv 4~(\mbox{mod}~5)$
1. $k=1,2,4$ then $X^{5^n(5^{\alpha}+1)}$ for non-negative integers $\alpha$ and $n$.
2. $k=3$ then $X^{6\cdot 5^{n}}+aX^{2\cdot 5^{n}}$ for non-negative integer $n$.
3. $k=5$ then $X^{2\cdot 5^{n+1}}+4aX^{6\cdot 5^{n}}+2a^2X^{2\cdot 5^{n}}$ for non-negative integer $n$.
3. For $p>5$, we obtain only one $DO$ polynomial of the form $\displaystyle X^{p^{n}(p^{\alpha}+1)}$ for some non-negative integers $\alpha$ and $n$.
Acknowledgement {#acknowledgement .unnumbered}
===============
The first named author would like to thank Professor Qiang Wang for suggesting this problem and for some helpful discussions.
[99]{}
R.S. Coulter, R.W. Matthews, [*Planar functions and planes of Lenz–Barlotti class II*]{}, Des. Codes Cryptogr. 10 (1997) 167–184.
R. S. Coulter and R. W. Mattews, [*Dembowski-Ostrom polynomials from Dickson polynomials*]{}, Finite Fields Appl. 16 : 369-379, 2010.
P. Dembowski and T. G. Ostrom, [*Planes of order $n$ with collineation groups of order $n^2$*]{}, Math. Z., 103 : 239 – 258, 1968.
L. E. Dickson, [*The analytic presentation of substitutions on a power of a prime number of letters with a discussion of the linear group*]{}, Ann. of Math. 11 (1897), 65–120.
N. Fernando, S. U. Hasan, M. Pal, [*Dembowski-Ostrom polynomials from reversed Dickson polynomials*]{}, arXiv.org: 2019.
X. Hou, G. L. Mullen, J. A. Sellers and J. L. Yucas, [*Reversed Dickson polynomials over finite fields*]{}. Finite Fields Appl. 15 : 748–733, 2009.
R. Lidl, G.L. Mullen, G. Turnwald, [*Dickson Polynomials*]{}, Pitman Monogr. Surveys Pure Appl. Math., vol. 65, Longman Scientific and Technical, Essex, England, 1993.
I. Schur, [*Über den Zusammenhang zwischen einemem Problem der Zahlentheorie und einem Satz üiber algebraische Funktionen*]{}, Sitzungsber. Akad. Wiss. Berlin (1923), 123–134.
Q. Wang and J. L. Yucas, [*Dickson polynomials over finite fields*]{}. Finite Fields Appl. 18 : 814–831, 2013.
X. Zhang, B. Wu, Z. Liu, [*Dembowski-Ostrom polynomials from reversed Dickson polynomials*]{}. J. Syst. Sci. Complex. [**29**]{} (2016), no. 1, 259 – 271.
G. Weng, X. Zeng, [*Further results on planar DO functions and commutative semifields*]{}, Des. Codes Cryptogr. 63 (2012) 413–423.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Many image processing problems are naturally expressed as energy minimization or shape optimization problems, in which the free variable is a shape, such as a curve in 2d or a surface in 3d. Examples are image segmentation, multiview stereo reconstruction, geometric interpolation from data point clouds. To obtain the solution of such a problem, one usually resorts to an iterative approach, a gradient descent algorithm, which updates a candidate shape gradually deforming it into the optimal shape. Computing the gradient descent updates requires the knowledge of the first variation of the shape energy, or rather the first shape derivative. In addition to the first shape derivative, one can also utilize the second shape derivative and develop a Newton-type method with faster convergence. Unfortunately, the knowledge of shape derivatives for shape energies in image processing is patchy. The second shape derivatives are known for only two of the energies in the image processing literature and many results for the first shape derivative are limiting, in the sense that they are either for curves on planes, or developed for a specific representation of the shape or for a very specific functional form in the shape energy. In this work, these limitations are overcome and the first and second shape derivatives are computed for large classes of shape energies that are representative of the energies found in image processing. Many of the formulas we obtain are new and some generalize previous existing results. These results are valid for general surfaces in any number of dimensions. This work is intended to serve as a cookbook for researchers who deal with shape energies for various applications in image processing and need to develop algorithms to compute the shapes minimizing these energies.'
author:
- Günay Doğan^1^
bibliography:
- 'shape\_calculus\_Dogan.bib'
title: |
Shape Calculus for Shape Energies\
in Image Processing
---
[^1]
Introduction {#S:intro}
============
Many image processing tasks are expressed as energy minimization problems in which the free variable is a shape, such as a curve in 2d or a surface in 3d, because the shape is a geometric representation for the object or the region of interest in the data. Examples of such tasks are image segmentation [@Aubert-etal-03; @Caselles-Kimmel-Sapiro-97; @Chan-Vese-01; @Chan-Vese-PWConst], surface regularization [@Clarenz-etal-04; @Schoenemann-Kahl-Cremers-09], geometric interpolation of data point clouds [@Zhao-etal-00] and multiview stereo reconstruction [@Faugeras-Keriven-98; @Jin-Yezzi-Soatto-03; @Kolev-Pock-Cremers-10]. In these problems, one defines an appropriate shape energy $J({\Gamma})$ that depends on the shape ${\Gamma}$, and the shape energy is designed such that its minimum corresponds to a solution of the image processing problem at hand. For example, as an image segmentation formulation (to locate distinct objects, regions or their boundaries in images), one can choose to use the Geodesic Active Contour model [@Caselles-Kimmel-Sapiro-97; @Caselles-Kimmel-Sapiro-97-3d], which is a weighted integral over candidate curves or surfaces ${\Gamma}$, as the shape energy, $$\label{E:gac-energy00}
J({\Gamma}) = {\int_{{\Gamma}}}g(x) dS,$$ where $g(x)$ is an image-based weight function, or one can choose the following variant of Mumford-Shah functional [@Chan-Vese-PWConst] $$\label{E:Chan-Vese-energy00}
J({\Gamma}) = \frac{1}{2} \sum_{i=1,2} \int_{{\Omega}_i} (I(x)-c_i)^2 dx
+ \nu {\int_{{\Gamma}}}dS, \qquad
c_i = \frac{1}{|{\Omega}_i|}\int_{{\Omega}_i} I(x) dx,$$ where $I(x)$ is the image function and ${\Omega}_1,{\Omega}_2$ are the inside and outside regions of the curve ${\Gamma}$. Then the curve minimizing the energy is a valid solution of segmentation problem.
The energy minimization formulation comes naturally for many image processing problems, because it gives a straight-forward way to penalize unwanted configurations of the shapes considered and to encourage the good configurations, especially when the problem has a data-fitting component or when the problem is hard to formulate in a direct manner. For example, the energy consists of a data term, which is minimized for curves that separate the image into regions of constant intensity. It also has a geometric penalty term, a length integral, which favors shorter curves over longer curves and acts as a regularizer in noisy images.
In order to find the minimizer ${\Gamma}^*$ of a given shape energy, one needs to implement an energy minimization or shape optimization algorithm. The shape optimization algorithms usually work iteratively; they start from an initial shape ${\Gamma}_0$ and deform the shape through several iterations with a velocity field ${{\vec V}}$ until a minimum of the energy is achieved. Thus, a crucial step to solve the minimization problem is the computation of the gradient descent velocities ${{\vec V}}$ at each iteration, namely deformation velocities that decrease the energy of the shape ${\Gamma}$. This requires understanding how a deformation of ${\Gamma}$ induced by a given velocity ${{\vec V}}$ changes the energy $J({\Gamma})$. An analytical tool that gives us this information is the shape derivative $dJ({\Gamma};{{\vec V}})$ [@Delfour-Zolesio-01; @Henrot-Pierre-05; @Simon-Murat-76; @Sokolowski]. It tells us the change in the energy $J({\Gamma})$ of the shape ${\Gamma}$ when ${\Gamma}$ is deformed by ${{\vec V}}$ (see Section \[S:shape-calculus\] for a rigorous definition of the shape derivative). If a given velocity ${{\vec V}}$ satisfies $dJ({\Gamma};{{\vec V}})<0$, then ${{\vec V}}$ is a gradient descent velocity. If the shape derivative is zero for all ${{\vec V}}$, then the shape ${\Gamma}$ is a stationary point, possibly a minimal shape. The shape derivative concept enables us to compute gradient descent velocities for a given shape ${\Gamma}$ and its energy $J({\Gamma})$ in a straight-forward manner. We review this briefly in Section \[S:grad-descent\]. The gradient descent velocity can be computed in other ways too, but we advocate shape differentiation in this paper as it is a very powerful technique and is widely applicable.
We can differentiate the energy more than once, and obtain the second shape derivative $d^2J({\Gamma};{{\vec V}},{{\vec W}})$, which is defined with respect to two perturbation velocities ${{\vec V}},{{\vec W}}$ [@Delfour-Zolesio-01; @Novruzi-Pierre-02; @Sokolowski]. The second shape derivative gives us second order shape sensitivity information. It can be used to perform stability analysis of a given stationary point [@Dambrine-Pierre-00] and it can tell us whether or not the stationary point is a minimum. Moreover it can be used to design fast Newton-type minimization schemes, which converge in fewer iterations. Only a few such schemes exist in image processing [@Bar-Sapiro-09; @Dogan-Morin-Nochetto-08; @Hintermueller-Ring-03; @Hintermueller-Ring-04] (more examples can be found in other areas of science and engineering [@Bui-Ghattas-12; @Eppler-Harbrecht-06; @Goto-Fujii-90; @Novruzi-Roche-00]). The existence of only a few schemes is due to the fact that the *explicit* formulas for the second shape derivatives of most energies in image processing are not known.
[**Contributions.**]{} In this paper, we use the shape differentiation methodology to derive the shape derivatives of large classes of shape energies. We aim our results to be as comprehensive as possible, so that this work will serve as a cookbook for the researchers who need to analyze shape energies and design algorithms that solve image processing problems using shape energies. In Section \[S:shape-energies\], we list the classes of shape energies that we consider and give examples of how they are used in the literature. The first shape derivatives or other derivatives of equivalent use have been derived for some of these energies. In some cases, these previous results are specific to a certain dimension, for example, curves in 2d or surfaces in 3d. In some cases, the results are specific to a certain geometric representation, such as a parametric surface. In most cases, the *explicit* formulas for the second shape derivative are not known. In fact, to our knowledge, the second shape derivatives have been computed for only two energies [@Hintermueller-Ring-03; @Hintermueller-Ring-04] in image processing (explicit formulas for some energies relevant to other application areas can be found in literature [@Bui-Ghattas-12; @Eppler-Harbrecht-06; @Goto-Fujii-90; @Novruzi-Roche-00])). In this work, we derive the first shape derivatives for all the classes of energies that we list in Section \[S:shape-energies\] and the second shape derivatives except for two of the energies (see Tables \[T:deriv-summary1\], \[T:deriv-summary2\] for a summary of the results, the new formulas derived that are previously unknown are denoted by $(\star)$). These new formulas are the main contribution of this paper. They are derived and laid out explicitly and are intended to serve researchers in image processing. These results are valid for hypersurfaces in any number dimensions and do not depend on the representation of the shape (parametric, level set or other). The only limitation of these results is that in their basic form, they are valid for closed surfaces or surfaces whose boundaries are on the image domain boundary. Other types of surfaces, such as open surfaces with boundaries inside the image domain or surfaces with junctions, require special consideration that we do not include in this paper.
The emphasis in this paper is on deriving the shape derivatives (assuming as much smoothness as needed of the shapes or the functions). Naturally, the existence of these shape derivative rely on certain differentiability requirements, and these may not be easy to satisfy in some practical situations. This will depend on the application and needs to be addressed on a case-by-case basis; therefore, such questions are not addressed in this paper. Moreover, existence of the shape derivatives does not imply the existence of the minimal shapes. This is a critical question that one must ask before using the shape derivatives to compute the minimum of the shape energies. For more information on existence and uniqueness of minimal shapes in shape optimization, we refer the reader to [@Bucur-Buttazzo-Henrot-98; @Bucur-Buttazzo-05].
**Outline.** We start with Section \[S:shape-energies\] explaining the major classes of shape energies used in image processing. Our goal is to compute the shape derivatives for these energies. In Section \[S:shape-calculus\], we introduce some basic differential geometry and some results and definitions from shape differential calculus. In Section \[S:shape-derivs-of-energies\], we use these results to compute the first and second shape derivatives of the energies introduced in Section \[S:shape-energies\]. These shape derivatives are the main contribution of the paper and are summarized in Tables \[T:deriv-summary1\], \[T:deriv-summary2\]. We conclude the paper with Section \[S:grad-descent\], where we briefly review how the shape derivatives can be used to compute gradient descent velocities for given shapes and energies.
Shape Energies in Image Processing {#S:shape-energies}
==================================
The search for geometric entities or geometric descriptions for objects based on given images is a main theme in image processing. Thus, researchers in this field are constantly devising new shape energies to address their problems, making it a very fertile field for applications of shape optimization. We find it useful to consider the numerous shape energies in image processing in four main classes: minimal surface energies, energies with integrals, higher-order energies, energies with PDEs (partial differential equations). Hybrids from these classes and exceptions are possible. These shape energy classes are explained below with examples from the literature. They are the starting point for the shape derivative calculations in Section \[S:shape-derivs-of-energies\].
**Minimal surface energies:** The first example of a shape energy in image processing was the Geodesic Active Contour Model proposed for image segmentation by Caselles, Kimmel and Sapiro in [@Caselles-Kimmel-Sapiro-97; @Caselles-Kimmel-Sapiro-97-3d] [^2]. The main idea of their work was to try to fit a curve or a surface to the edges of an object in the image. For this, they used an edge indicator function $g(x)$ such that $g\approx 0$ on edges and $g\approx 1$ elsewhere and tried to compute a surface minimizing the following energy $$\label{E:gac-energy0}
J({\Gamma}) = {\int_{{\Gamma}}}g(x) dS.$$ One can sometimes add an area or volume integral ${\int_{{\Omega}}}g(x) dx$ to to speed up the computations and to facilitate detection of concavities. Minimal surfaces computed from the energy make very satisfactory segmentations as they give continuous and smooth representations of the boundaries of objects or regions in the given images. Thus, the model is very popular and is widely implemented. The implementation is based on the first shape derivative or the first variation of the energy. Only recently the second shape derivative for was computed by Hintermüller and Ring [@Hintermueller-Ring-03] and was used to devise a second order minimization method resulting in faster convergence.
The energy is isotropic, i.e. it does not depend on the orientation or the normal of surface. In [@Kimmel-Bruckstein-03], Kimmel and Bruckstein proposed an anisotropic energy that fits the general form $$\label{E:anisotropic-energy0}
J({\Gamma}) = {\int_{{\Gamma}}}g(x,{n}) dS.$$ By setting $g=\langle \nabla I,{n}\rangle$, they aimed to better align solution curves with object boundaries in images and were able to attain improved segmentations. Before [@Kimmel-Bruckstein-03], the energy had been used by Faugeras and Keriven for multiview stereo reconstruction [@Faugeras-Keriven-98] (later by Jin et al. in [@Jin-Yezzi-Soatto-03] and by Kolev et al. [@Kolev-Pock-Cremers-10]). The first shape derivative of for general n-dimensional surfaces has been known in the literature for geometric flows [@Bellettini-Paolini-97; @Deckelnick-Dziuk-Elliott-05]. We derive the second shape derivative in this paper.
The key feature of the energy is the dependence on the geometry through the normal of the surface. The dependence may be through other geometric properties of the surface as well, such as the mean curvature $\kappa$ $$\label{E:curvature-energy0}
J({\Gamma}) = {\int_{{\Gamma}}}g(x,\kappa) dS.$$ The integral is usually used as part of a more involved energy, to impose higher regularity of the surface. Examples are the Willmore functional with $g=\frac{1}{2}\kappa^2$ [@Clarenz-etal-04; @WillmoreBook] or $g=\kappa^p$ in [@Schoenemann-Kahl-Cremers-09]. Sundaramoorthi et al. noted in [@Sundar-etal-09] that $g=\frac{1}{2}w(x)\kappa^2$ with an image-based weight $w(x)$ yielded better regularizations for image segmentation. Another more general functional form of $g(\kappa)$ was used in [@Droske-Bertozzi-10] to implement a corner-preserving regularization energy. The first shape derivative for and for the energy with the more general geometric weight $g=g(x,{n},\kappa)$ was derived in [@Dogan-Nochetto-12] by Doğan and Nochetto.\
**Energies with integrals:** If one views a surface as the boundary separating different regions in the image from each other (in order to identify distinct regions), a logical approach to designing the shape energy is to incorporate terms that compare the properties of the regions across the boundary and try to find surfaces maximizing the difference between the regions. The characteristics of each region can be quantified by computing the statistics of the image features in the region [@Cremers-Rousson-Deriche-07]. The statistics computations can often be expressed as various integrals over the regions. This results in shape energies with weight functions that depend on integrals over the regions. An example is the following energy proposed by Chan and Vese in [@Chan-Vese-PWConst]. It aims to find a partitioning of the image domain into a foreground region ${\Omega}_1$ and a background region ${\Omega}_2$ (inside and outside ${\Gamma}$ respectively), each with distinct averages $c_1, c_2$ of the image intensity $I(x)$ respectively. $$\label{E:ChanVese-energy}
J({\Gamma}) = \frac{1}{2} \sum_{i=1,2} \int_{{\Omega}_i} (I(x)-c_i)^2 dx
+ \nu {\int_{{\Gamma}}}dS, \qquad
c_i = \frac{1}{|{\Omega}_i|}\int_{{\Omega}_i} I(x) dx.$$
More general approaches to incorporating statistics into the shape optimization formulation are described in [@Cremers-Rousson-Deriche-07; @Paragios-Deriche-02a]. To develop a more general statistical formulation, one can consider a Bayesian interpretation of the estimation problem and try to maximize the a posteriori probability $p(\{{\Omega}_1,{\Omega}_2\}|I)$, namely the likelihood of having a certain partitioning $\{{\Omega}_1,{\Omega}_2\}$ given the image $I$ (multiple phases or regions $\{{\Omega}_i\}_{i=1}^m$ are possible, but not considered in this paper to simplify the presentation). We can write $p(\{{\Omega}_1,{\Omega}_2\}|I)$ as $$\label{E:bayes-rule}
p(\{{\Omega}_1,{\Omega}_2\}|I) \propto p(I|\{{\Omega}_1,{\Omega}_2\}) \
p(\{{\Omega}_1,{\Omega}_2\}),$$ and separate the a priori shape information $p(\{{\Omega}_1,{\Omega}_2\})$ from image-based cues encoded in $p(I|\{{\Omega}_1,{\Omega}_2\})$. A common example of the a priori shape term would be $p(\{{\Omega}_1,{\Omega}_2\}) \varpropto e^{-\nu|{\Gamma}|}$. Assuming no correlation between labelings of regions, one can simplify the conditional probability $$p(I|\{{\Omega}_1,{\Omega}_2\})
= p(I|{\Omega}_1) p(I|{\Omega}_2) = p_1(I) p_2(I) .$$ Maximizing the probability is equivalent to minimizing its negative logarithm. Thus we end up with the following energy $$J({\Gamma}) = -\int_{{\Omega}_1} \log p_1(I(x)) dx
-\int_{{\Omega}_2} \log p_2(I(x)) dx + \nu {\int_{{\Gamma}}}dS.$$ If the distributions are modeled as parametric ones, with parameters $\theta_i$ for $p_i$, then the energy can be rewritten as $$\label{E:stat-shape-energy}
J({\Gamma}) = -\int_{{\Omega}_1} \log p(I(x)|\theta_1) dx
-\int_{{\Omega}_2} \log p(I(x)|\theta_2) dx + \nu {\int_{{\Gamma}}}dS.$$ The parameters $\theta_i$ depend on the form of the probability density function and often involve integrals over the regions ${\Omega}_i$. For example, the Gaussian probability density function has the form $p_i(s) = \frac{1}{\sqrt{2\pi\sigma_i^2}} \exp\left(-\frac{(s-c_i)^2}{2\sigma_i^2}\right)$, where the parameters $c_i, \sigma_i$ are computed by the integrals $
c_i = \frac{1}{|{\Omega}_i|} \int_{{\Omega}_i} I(x) dx, \quad
\sigma_i^2 = \frac{1}{|{\Omega}_i|} \int_{{\Omega}_i} (I(x)-c_i)^2 dx.
$\
It is not hard to see that we can concoct more complicated statistical formulations where shape energies with integrals play a central role. Thus, shape energies with integral parameters have significant use in image processing. The prototype for energies with integrals is $$\label{E:domain-energy-w-integrals0}
J({\Omega}) = {\int_{{\Omega}}}g(x,I_w({\Omega})) dx, \qquad
I_w({\Omega}) = {\int_{{\Omega}}}w(x) dx,$$ or one whose weight function $g$ may depend on multiple integrals $$\label{E:domain-energy-w-multiple-integrals0}
J({\Omega}) = {\int_{{\Omega}}}g(x,I_{w_1}({\Omega}),\ldots,I_{w_m}({\Omega})) dx,
\qquad I_{w_i}({\Omega}) = {\int_{{\Omega}}}w_i(x) dx.$$ The first shape derivatives for the energies , were computed by Aubert et al. in [@Aubert-etal-03]. The second shape derivatives are computed in Section \[S:shape-derivs-of-energies\] in this paper, where we also deal with the case of nested integrals. Similar to domain energies with integrals, one can conceive of problems where it is necessary to deal with surface energies with integral parameters: $$\label{E:bound-energy-w-integral0}
J({\Gamma}) = {\int_{{\Gamma}}}g(x,I_w({\Gamma})) dS, \qquad
I_w({\Gamma}) = {\int_{{\Gamma}}}w(x) dS.$$ The first and second shape derivatives for the energy are not available in the literature and are computed in Section \[S:shape-derivs-of-energies\].\
**Higher order energies:** Some image processing problems require encoding nonlocal interactions between points of a surface or a domain. Shape energies involving such interactions may be written as higher order integrals over the surface or the domain. For example, an energy that encodes the interactions between any two points of a surface or a domain would have the form $$\label{E:h-o-energies0}
J({\Gamma}) = {\int_{{\Gamma}}}{\int_{{\Gamma}}}g(x,y) dS(x) dS(y), \qquad
J({\Omega}) = {\int_{{\Omega}}}{\int_{{\Omega}}}g(x,y) dx dy.$$ The weight function $g(x,y)$ describes the nature of the interaction between the points $x$ and $y$. If we want to account for nonlocal interactions of more points, say three, this can be formulated as a multiple integral with even higher order,\
like ${\int_{{\Omega}}}{\int_{{\Omega}}}{\int_{{\Omega}}}g(x,y,z)\ dx dy dz$.
Examples of higher order shape energies are not many in image processing, but they have been used successfully in applications such as road network extraction from images [@Rochery-Jermyn-Zerubia-06] and topology control of curves in image segmentation [@LeGuyader-Vese-08; @Sundar-Yezzi-07]. In [@Sundar-Yezzi-07], Sundaramoorthi and Yezzi used the following shape energy to prevent curves from changing topology by merging or splitting: $$\label{E:Sundar-Yezzi-energy0}
J({\Gamma}) = {\int_{{\Gamma}}}{\int_{{\Gamma}}}\frac{1}{|x-y|^\gamma} dS(x) dS(y), \quad \gamma > 0.$$ They added the energy as an additional term to their segmentation energy. Note that the value of the integral blows up as different parts of curve get close to each other, hence a topological change is prevented. In [@LeGuyader-Vese-08], Le Guyader and Vese accomplished the same goal by using a double integral over the domain ${\Omega}$, instead of the surface ${\Gamma}$. Rochery et al. proposed higher order active contours in [@Rochery-Jermyn-Zerubia-06], as a general formulation with multiple integrals over curves (but not general surfaces). They derived the first variation of the shape energy and illustrate its use with an application in road network detection. In this paper, the shape energies are considered for general surfaces ${\Gamma}$ and domains ${\Omega}$ in ${\mathbb R}^d$ and their first and second shape derivatives are derived for general weight functions $g(x,y)$ (note that the second shape derivatives are not known and first variations are reported for specific $g(x,y)$ in previous work). We also explain how shape derivatives for energies with order higher than two can be derived.\
**Energies with PDEs:** Large classes of images can be modeled as piecewise smooth functions with some discontinuities. For such images, the problems of image segmentation and image regularization can be formulated as finding the discontinuity set $K$ and approximating the image intensity function $I$ with a smooth function $u$ on the remaining parts $D-K$ of the image domain $D\subset {\mathbb R}^d$. Mumford and Shah proposed minimizing the following energy for this purpose [@Mumford-Shah-89] $$\label{E:original-MS-energy}
J(K,u) = \frac{1}{2}\int_D (I-u)^2
+ \frac{\mu}{2}\int_{D-K} |\nabla u|^2 + \nu |K|,
\quad \mu,\nu > 0$$ The set of discontinuities that is included in the formulation is very general and can include cracks and triple junctions. Therefore, a direct numerical realization of the minimization of is not practical. For this reason, Chan and Vese proposed an alternative energy in [@Chan-Vese-01] $$\label{E:MS-energy0}
J({\Gamma}) = \frac{1}{2} \sum_{i=1}^2 \int_{{\Omega}_i} \left( (u_i-I)^2
+ \mu |\nabla u_i|^2 \right) dx + \nu \int_{\Gamma}dS,$$ where $u_1, u_2$ are the smooth approximations to the image $I$ computed by $$\label{E:MS-pde0}
-\mu \Delta u_i + u_i = I \ \textrm{in} \ {\Omega}_i, \qquad
\frac{\partial u_i}{\partial{n}_i} = 0 \ \textrm{on} \ \partial {\Omega}_i,$$ and ${\Omega}_1, {\Omega}_2$ are the domains inside and outside the surface ${\Gamma}$ respectively. In [@Chan-Vese-01], the surface ${\Gamma}$ was represented with a level set function. Thus cracks and triple junctions were excluded (a method to represent cases with junctions was proposed in [@Vese-Chan-02] using multiple level set functions). Chan and Vese implemented a gradient descent method based on the first variation of . In [@Hintermueller-Ring-04], Hintermüller and Ring derived the first and second shape derivatives of . The second shape derivative was used in [@Dogan-Morin-Nochetto-08],[@Hintermueller-Ring-04] to develop fast Newton-type minimization methods for . A domain energy of the form ${\int_{{\Omega}}}g(x,u,\nabla u) dx$ with a Neumann PDE like was considered Goto and Fujii in [@Goto-Fujii-90], where they derived the first and second shape derivatives.
In [@Brox-Cremers-09], Brox and Cremers gave a statistical interpretation of the Mumford-Shah functional. They started from a Bayesian model for segmentation and introduced a local Gaussian probability density function for image intensity in each region ${\Omega}_i$: $p_i(I(x),x) = \frac{1}{\sqrt{2\pi}\sigma_i(x)}
\exp\left( -\frac{(I(x)-c_i(x))^2}{2\sigma_i(x)^2} \right)$ with spatially varying mean $c_i(x)$ and variance $\sigma_i(x)$. Then they obtained the following extended Mumford-Shah functional $$\label{E:Cremers-MS-energy}
\begin{aligned}
J_{BC}({\Gamma},\{c_i\},\{\sigma_i\})
= &\sum_i \int_{{\Omega}_i} \left( \frac{(I(x)-c_i(x))^2}{2\sigma_i(x)^2}
+ \frac{1}{2}\log(2\pi\sigma_i(x)^2) \right) dx \\
&+ \frac{\mu}{2}\sum_i\int_{{\Omega}_i} \left( |\nabla c_i(x)|^2
+ |\nabla\sigma_i(x)|^2 \right) dx + \nu{\int_{{\Gamma}}}dS,
\end{aligned}$$ by linking their model with the filtering theory of Nielsen et al. [@Nielsen-etal-97]. We generalize Brox and Cremer’s model and write $$\label{E:gen-MS-energy0}
J_0({\Gamma},\{u_{ki}\}) = \sum_i \int_{{\Omega}_i} \left( f(x,\{u_{ki}\})
+ \frac{\mu}{2} \sum_k |\nabla u_{ki}|^2 \right)dx
+ \nu {\int_{{\Gamma}}}dS,$$ where $u_{ki}$ is the smooth of approximation of a $k^{th}$ data channel or statistical descriptor over region ${\Omega}_i$, and $f(x,\{u_{ki}\})$ denotes a coupled data term, for example, $f(x,\{u_i,v_i\}) = \frac{(I(x)-u_i)^2}{2 v_i^2}$ in the case of or $f(x,\{u_{1i},u_{2i},u_{3i}\}) = \Sigma_{k=1}^{3}(I_k(x)-u_{ki}(x))^2$ for color image segmentation. We write the optimality condition of with respect to $\{u_{ki}\}$ $$\label{E:gen-MS-pde0}
-\Delta u_{ki} + f_{u_{ki}}(x,\{u_{li}\}) = 0 \ \mathrm{in} \ {\Omega}_i,
\qquad \frac{\partial u_{ki}}{\partial {n}} = 0
\ \mathrm{on} \ \partial{\Omega}_i,$$ and use the solution of to write the reduced shape energy $J({\Gamma}) = J_0({\Gamma},\{u_{ki}({\Gamma})\}) $ $$\label{E:gen-MS-energy1}
J({\Gamma}) = \sum_i \int_{{\Omega}_i} \left( f(x,\{u_{ki}({\Gamma})\})
+ \frac{\mu}{2} \sum_k |\nabla u_{ki}({\Gamma})|^2 \right) dx
+ \nu {\int_{{\Gamma}}}dS.$$ The energy is more general than and , and its first and second shape derivatives are not known; they are derived in Section \[S:energies-with-pdes\] in this paper.
In , the role of the elliptic PDE was in computing a piecewise smooth approximation $u$ to the image data $I$ on the domains ${\Omega}_i$. One can as well be interested in finding a smooth approximation to data defined on the surface ${\Gamma}$. This requires using an elliptic PDE defined on the surface ${\Gamma}$. Such a formulation was proposed in [@Jin-Yezzi-Soatto-03] by Jin, Yezzi and Soatto for stereoscopic reconstruction of 3d objects and their surface reflectance from 2d projections of the objects. The shape energy they used is essentially the following $$\label{E:Jin-energy0}
J({\Gamma}) = \frac{1}{2}{\int_{{\Gamma}}}(u(x)-d(x))^2 dS
+ \frac{\mu}{2} {\int_{{\Gamma}}}|{{\nabla_{{\Gamma}}}}u|^2 dS + \nu{\int_{{\Gamma}}}dS,$$ where $u$ is computed from the surface PDE, $-\mu{\Delta_{{\Gamma}}}u + u = d$ on ${\Gamma}$ (see §\[S:diff-geom\] for the definition of the surface gradient ${{\nabla_{{\Gamma}}}}$ and the surface Laplacian ${\Delta_{{\Gamma}}}$), and $d(x)$ is some data function based on the 2d images of the 3d scene. Jin, Yezzi and Soatto considered the parametric representation of a 2d surface in 3d in order to derive the first variation of the surface energy. Then they implemented a gradient descent algorithm using the level set method. The first shape derivative of for general surfaces in any number of dimensions, to our knowledge, is not available in literature. We consider the following more general energy $$\label{E:gen-Jin-energy0}
J({\Gamma}) = {\int_{{\Gamma}}}f(x,\{u_k({\Gamma})\}) dS
+ \frac{\mu}{2} \sum_k {\int_{{\Gamma}}}|{{\nabla_{{\Gamma}}}}u_k({\Gamma})|^2 dS + \nu {\int_{{\Gamma}}}dS,$$ where $\{u_k\}_{k=1}^m$ are computed from the optimality PDE: $-\mu {\Delta_{{\Gamma}}}u_k + f_{u_k}(x,\{u_l\}) = 0$. We compute the first shape derivative of in Section \[S:energies-with-pdes\]. This result includes the case of as well. Unlike [@Jin-Yezzi-Soatto-03], it is not restricted to parameterized surfaces and is valid in any number of dimensions.
[|l|]{}\
\
Geodesic active contour energy: $\displaystyle J({\Gamma}) = {\int_{{\Gamma}}}g(x) dS + \gamma{\int_{{\Omega}}}g(x) dx,$\
[@Caselles-Kimmel-Sapiro-97; @Caselles-Kimmel-Sapiro-97-3d; @Hintermueller-Ring-03]: $\displaystyle
dJ({\Omega};V)= {\int_{{\Gamma}}}\left((\kappa+\gamma)g(x)
+ \partial_{n}g(x)\right) V dS,
\qquad \quad (\partial_{n}g: \mathrm{normal \ derivative})
$\
------------------------------------------------------- ---------------------------------------------------------------------------------------------
[@Hintermueller-Ring-03]: $\quad d^2J({\Gamma};V,W) $ $\displaystyle = {\int_{{\Gamma}}}g {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W dS$
$\displaystyle \ + {\int_{{\Gamma}}}\left( \partial_{{n}{n}} g
+ (2\kappa + \gamma) \partial_{n}g
+ (\kappa^2 - \sum \kappa_i^2 + 2\gamma\kappa) g \right) V W dS.
$
------------------------------------------------------- ---------------------------------------------------------------------------------------------
: See the caption of Table \[T:deriv-summary2\] for an explanation of the labels $(\star)$,$[\#]$.[]{data-label="T:deriv-summary1"}
\
Normal-dependent surface energy: $\displaystyle J({\Gamma}) = {\int_{{\Gamma}}}g(x,{n}) dS,$\
[@Bellettini-Paolini-97; @Deckelnick-Dziuk-Elliott-05]: $\displaystyle
dJ({\Gamma};V)= {\int_{{\Gamma}}}\left(\kappa g + \partial_{n}g
+ {{\rm{div}_{{\Gamma}}}}(g_y)_{\Gamma}\right) V dS,
\qquad \quad (g_y: \mathrm{gradient \ w.r.t. \ 2nd \ arg.})
$\
$\displaystyle
\begin{aligned}
(\star):\
d^2J({\Gamma};V,W) =& {\int_{{\Gamma}}}{{\nabla_{{\Gamma}}}}V \cdot \left( (g-g_y \cdot {n})Id +
g_{yy}\right) \cdot {{\nabla_{{\Gamma}}}}W dS \\
&+ {\int_{{\Gamma}}}\left(\partial_{{n}{n}}g+2\kappa \partial_{n}g
+ (\kappa^2 - \sum \kappa_i^2) g \right)VW dS \\
&- {\int_{{\Gamma}}}(\kappa g_y + {n}^T g_{xy})
\cdot ( {{\nabla_{{\Gamma}}}}W \, V + {{\nabla_{{\Gamma}}}}V \, W ) dS .
\end{aligned}
$\
Curvature-dependent surface energy: $\displaystyle J({\Gamma}) = {\int_{{\Gamma}}}g(x,\kappa) dS,$\
[@Dogan-Nochetto-12]: $\displaystyle
dJ({\Gamma};V) = {\int_{{\Gamma}}}\left( -{\Delta_{{\Gamma}}}(g_z)
+ g \kappa - g_z \sum \kappa_i^2 + \partial_{n}g \right) V dS,
\qquad \ (g_z: \mathrm{deriv. \ w.r.t. \ 2nd \ arg.})
$\
$
\qquad\qquad\quad
{\Delta_{{\Gamma}}}(g_z) = {\Delta_{{\Gamma}}}g_z + g_{zz} {\Delta_{{\Gamma}}}\kappa + g_{zzz} |{{\nabla_{{\Gamma}}}}\kappa|^2
+ (g_{zxz} + g_{zzx})\cdot{{\nabla_{{\Gamma}}}}\kappa.
$\
(more results in §\[S:energies-with-integrals\])\
Energies with domain integrals: $\displaystyle J({\Omega}) = {\int_{{\Omega}}}g(x,I_w({\Omega}))dx, \qquad I_w({\Omega})={\int_{{\Omega}}}w(x)dx,$\
[@Aubert-etal-03]: $\displaystyle
dJ({\Omega};V) = {\int_{{\Gamma}}}\left( g(x,I_w({\Omega}))
+ I_{g_p}({\Omega}) w(x) \right) V dS,
\qquad \ (g_p: \mathrm{deriv. \ w.r.t. \ 2nd \ arg.})
$\
$\displaystyle
\begin{aligned}
(\star):\
d^2J({\Omega};V,W) =& {\int_{{\Gamma}}}\left( \partial_{n}g + I_{g_p} \partial_{n}w
+ \kappa(g + I_{g_p} w) \right) V W dS \\
&+ {\int_{{\Gamma}}}g_p V dS {\int_{{\Gamma}}}w W dS + {\int_{{\Gamma}}}w V dS {\int_{{\Gamma}}}g_p W dS
+ I_{g_{pp}} {\int_{{\Gamma}}}w V dS {\int_{{\Gamma}}}w W dS.
\end{aligned}
$\
Energies with surface integrals: $\displaystyle J({\Gamma}) {\int_{{\Gamma}}}g(x,I_w({\Gamma}))dS, \qquad I_w({\Gamma})={\int_{{\Gamma}}}w(x)dS,$\
$\displaystyle
(\star):\
dJ({\Gamma};V) = {\int_{{\Gamma}}}\left( (g + I_{g_p} w) \kappa
+ \partial_{n}g + I_{g_p} \partial_{n}w \right) V dS,
\qquad \ (g_p: \mathrm{deriv. \ w.r.t. \ 2nd \ arg.})
$\
$\displaystyle
\begin{aligned}
(\star):\
d^2J({\Gamma};V,W) =& {\int_{{\Gamma}}}(g + I_{g_p} w) {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W dS \\
&+ {\int_{{\Gamma}}}\left( \partial_{{n}{n}} g + I_{g_p} \partial_{{n}{n}} w +
2(\partial_{n}g + I_{g_p}\partial_{n}w)\kappa \right) V W dS \\
&+ {\int_{{\Gamma}}}\left( \partial_{n}g_p + g_p \kappa \right) V dS {\int_{{\Gamma}}}\left(
\partial_{n}w + w \kappa \right) W dS \\
&+ {\int_{{\Gamma}}}\left( \partial_{n}w + w \kappa \right) V dS {\int_{{\Gamma}}}\left(
\partial_{n}g_p + g_p \kappa \right) W dS \\
&+ I_{g_{pp}} {\int_{{\Gamma}}}\left( \partial_{n}w + w \kappa \right) V dS
{\int_{{\Gamma}}}\left( \partial_{n}w + w \kappa \right) W dS.
\end{aligned}
$\
[|l|]{}\
\
Higher order domain energies: $\displaystyle J({\Omega}) = {\int_{{\Omega}}}{\int_{{\Omega}}}g(x,y) dy dx,$\
($\star$,[@LeGuyader-Vese-08]): $\displaystyle
dJ({\Omega};V) = {\int_{{\Gamma}}}{\int_{{\Omega}}}\tilde{g}(x,y)dy V dS,
\qquad \left( \tilde{g}(x,y) = g(x,y)+g(y,x) \right)
$\
$\displaystyle
\begin{aligned}
(\star):\
d^2J({\Omega};V,W) = &{\int_{{\Gamma}}}{\int_{{\Gamma}}}\tilde{g}(x,y) W(y) dS(y) V(x) dS(x) \\
&+ {\int_{{\Gamma}}}\left( \kappa(x) {\int_{{\Omega}}}\tilde{g}(x,y)dy
+ {n}(x)\cdot {\int_{{\Omega}}}\tilde{g}_y(x,y)dy \right) V W dS.
\end{aligned}
$\
Higher order surface energies: $\displaystyle J({\Gamma}) = {\int_{{\Gamma}}}{\int_{{\Gamma}}}g(x,y) dS(y) dS(x),$\
($\star$,[@Rochery-Jermyn-Zerubia-06]): $\displaystyle
dJ({\Gamma};V)= {\int_{{\Gamma}}}\left(\kappa(x) {\int_{{\Gamma}}}\tilde{g}(x,y) dS(y)
+{n}(x)\cdot{\int_{{\Gamma}}}\tilde{g}_x(x,y) dS(y) \right) V(x) dS(x),
$\
$\displaystyle
\begin{aligned}
(\star):\
d^2J({\Gamma}; &V,W) = {\int_{{\Gamma}}}G(x,{\Gamma}){{\nabla_{{\Gamma}}}}V\cdot {{\nabla_{{\Gamma}}}}W dS(x)
\qquad \qquad \left(G(x,{\Gamma}) = {\int_{{\Gamma}}}\tilde{g}(x,y) dS(y) \right)
\\
&+ {\int_{{\Gamma}}}\left( {n}^T G_{xx}(x,{\Gamma})\,{n}+ 2\kappa\, G_{x}(x,{\Gamma})\cdot{n}+ (\kappa^2 - \Sigma\kappa_i^2) G(x,{\Gamma}) \right) V W dS(x)
\\
&+ {\int_{{\Gamma}}}\kappa {\int_{{\Gamma}}}\tilde{g}\kappa\, W dS(y) V dS(x)
+ {\int_{{\Gamma}}}{n}^T {\int_{{\Gamma}}}\tilde{g}_{xy}{n}\, W dS(y) V dS(x)
\\
& + {\int_{{\Gamma}}}\kappa {\int_{{\Gamma}}}\tilde{g}_y \cdot{n}\, W dS(y) V dS(x)
+ {\int_{{\Gamma}}}{n}\cdot{\int_{{\Gamma}}}\tilde{g}_x\kappa\, W dS(y) V dS(x),
\end{aligned}
$\
\
$\displaystyle
\begin{array}{ll}
\mathrm{Energies \ with \ domain \ PDEs:} &
\displaystyle{
J({\Omega}) = {\int_{{\Omega}}}\left(f(x,\{u_k\})
+ \frac{\mu}{2} \sum_k |\nabla u_k|^2 \right) dx,}
\\
&
-\mu\Delta u_l + f_{u_l} = 0 \ \mathrm{in} \ {\Omega},
\ \frac{\partial u_l}{\partial{n}} = 0 \ \mathrm{on} \ \partial{\Omega}, \ l=1,\ldots,m,
\end{array}
$\
($\star$,[@Brox-Cremers-09; @Goto-Fujii-90; @Hintermueller-Ring-04; @Vese-Chan-02]): $\displaystyle
dJ({\Omega};V)= {\int_{{\Gamma}}}\left(f(x,\{u_k\})
+ \frac{\mu}{2}\sum_k |{{\nabla_{{\Gamma}}}}u_k|^2 \right) V dS,
$\
-------------------------------------------------------------------------------- --------------------------------------------------------------------------------
($\star$,[@Goto-Fujii-90; @Hintermueller-Ring-04]): $\quad d^2J({\Omega};V,W)$ $\displaystyle
= {\int_{{\Gamma}}}\left(f\kappa + \frac{\partial f}{\partial{n}}
+ \mu \sum_k {{\nabla_{{\Gamma}}}}u_k^T \left( \frac{\kappa}{2}Id
- {{\nabla_{{\Gamma}}}}{n}\right) {{\nabla_{{\Gamma}}}}u_k \right) V W dS$
$\displaystyle \
+{\int_{{\Gamma}}}\sum_k \left( f_{u_k} u_{k,W}'
+ \mu {{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}u_{k,W}' \right) V dS.
$
-------------------------------------------------------------------------------- --------------------------------------------------------------------------------
: Next to each shape derivative formula is a label specifying whether or not the formula is new and citation to its source if it is not new. The label $(\star)$ indicates that it is a new result, not available in the literature in any form. The label $[\#]$ indicates that the result can be found in reference $[\#]$. The label $(\star,[\#])$ indicates that the result is new for general surfaces or domains in ${\mathbb R}^d$ or for general choices of the weight functions $g(x,{\Gamma}), g(x,{\Omega})$, but formulas for restricted situations can be found in reference $[\#]$. []{data-label="T:deriv-summary2"}
\
$\displaystyle
\begin{array}{ll}
\mathrm{Energies \ with \ surface \ PDEs:} &
\displaystyle{
J({\Gamma}) = \int_{\Gamma}\left( f(x,\{u_k\})
+ \frac{\mu}{2} \sum_k |{{\nabla_{{\Gamma}}}}u_k|^2 \right) dS,}
\\
&
-\mu{\Delta_{{\Gamma}}}u_l + f_{u_l} = 0 \ \mathrm{on} \ {\Gamma}, \ l=1,\ldots,m,
\end{array}
$\
($\star$,[@Jin-Yezzi-Soatto-03]): $\displaystyle
dJ({\Gamma};V) = {\int_{{\Gamma}}}\left(f \kappa
+ \frac{\partial f}{\partial{n}}
+ \mu \sum_k {{\nabla_{{\Gamma}}}}u_k^T \left(\frac{\kappa}{2}Id
- {{\nabla_{{\Gamma}}}}{n}\right){{\nabla_{{\Gamma}}}}u_k \right) V dS.
$\
Shape Differential Calculus {#S:shape-calculus}
===========================
In this section, we will reviews some basic differential geometry that we will refer to throughout the paper. We will prove some useful geometric formulas. Finally we will introduce the shape derivative concept and describe some related results that will enable us to differentiate the model energies from Section \[S:shape-energies\].
Review of Differential Geometry {#S:diff-geom}
-------------------------------
We assume that ${\Gamma}$ is a smooth orientable compact $(d-1)$ dimensional surface in ${\mathbb R}^d$ without boundary. Let us be given $h\in C^2({\Gamma})$ and a smooth extension $\tilde{h}$ of $h$, $\tilde{h}\in C^2(U)$ and $\tilde{h}\vert_{{\Gamma}}=h$ on ${\Gamma}$ where $U$ is a tubular neighborhood of ${\Gamma}$ in ${\mathbb R}^d$. The [*tangential gradient*]{} $\nabla_{\Gamma}h$ of $h$ is defined by: $$\nabla_{\Gamma}h=\big(\nabla \tilde{h}
- \partial_{n}\tilde{h}~{n}\big)\vert_{{\Gamma}},$$ where ${n}$ denotes the unit normal vector to ${\Gamma}$ and $\partial_{n}\tilde{h} = \nabla\tilde{h}\cdot{n}$ is the normal derivative. Similarly, given ${{\vec W}}\in [C^1({\Gamma})]^d$ and its smooth extension $\tilde{W}\in C^1(U)$, we define the *tangential divergence* of ${{\vec W}}$ by $${{\rm div}}_{{\Gamma}}{{\vec W}}= \big({{\rm div}}\tilde{W}-{n}\cdot D\tilde{W}\cdot {n}\big)
\vert_{{\Gamma}},$$ where $D\tilde{W}$ denotes the Jacobian matrix of $\tilde{W}$. We also define the tangential gradient ${{\nabla_{{\Gamma}}}}{{\vec W}}$ of ${{\vec W}}$, which is a matrix whose $i^{th}$ row is the tangential gradient ${{\nabla_{{\Gamma}}}}{{\vec W}}_i$ of the $i^{th}$ component of ${{\vec W}}$. Finally, the [*tangential Laplacian*]{} or [*Laplace-Beltrami operator*]{} $\Delta_{\Gamma}$ on ${\Gamma}$ is defined as follows: $$\label{E:Laplace-Beltrami}
\Delta_{\Gamma}h = {{\rm div}}_{\Gamma}(\nabla_{\Gamma}h)=
\big(\Delta\tilde{h} - {n}\cdot D^2\tilde{h}\cdot {n}-\kappa ~\partial_{n}\tilde{h}\big)\vert_{{\Gamma}}.$$ As seen above, in order to compute the full spatial derivatives of surface functions and geometric quantities, such as the normal ${n}$ and the mean curvature $\kappa$, defined only on the surface ${\Gamma}$, we need to extend them to a tubular neighborhood of ${\Gamma}$. This is accomplished using a signed distance function representation of the surface ${\Gamma}$: $$\label{E:signed-dist}
b(x,{\Gamma}) = \left \{ \begin{array}{ll}
\ \ \textrm{dist}(x,{\Gamma}) & \textrm{for }x\in {\mathbb R}^d-{\Omega}\\
\ \ 0 & \textrm{for } x \in {\Gamma}\\
-\textrm{dist}(x,{\Gamma}) & \textrm{for } x \in {\Omega}\end{array}\right.$$ where $\textrm{dist}(x,{\Gamma}) = \inf_{y\in{\Gamma}}|y-x|$ and ${\Omega}$ is the domain enclosed by ${\Gamma}$. Using we extend the normal ${n}$, the mean curvature $\kappa$ and the second fundamental form ${{\nabla_{{\Gamma}}}}{n}$ as follows [@Delfour-Zolesio-01 Chap. 8]: $$\label{E:geom-signed-dist}
{n}= \nabla b(x)|_{\Gamma}, \quad \kappa = \Delta b(x)|_{\Gamma}, \quad
{{\nabla_{{\Gamma}}}}{n}= D^2 b(x)|_{\Gamma}.$$ The extensions allow us to differentiate ${n}, \kappa, {{\nabla_{{\Gamma}}}}{n}$ in the normal direction in addition to the tangential direction, and the normal derivatives of the normal ${n}$ and the mean curvature $\kappa$ are given by $$\label{E:normal-deriv-curv}
\partial_{n}{n}= 0, \qquad
\partial_{n}\kappa = -\sum_{i=1}^{d-1} \kappa_i^2,$$ respectively, where $\kappa_i$ denote the principal curvatures of the surface. It is easy to show, using the extension , $$\partial_{n}{n}= \nabla(\nabla b) \nabla b |_{\Gamma}= D^2 b \nabla b |_{\Gamma}= 0,$$ because $0 = \nabla(1) = \nabla(\nabla b \cdot \nabla b) = 2D^2 b \nabla b$. The expression for $\partial_{n}\kappa$ can be computed similarly. The proof can be found in [@Dogan-Nochetto-12]. Note that Equation holds only for parallel surfaces defined by the signed distance function.
\[L:deriv-surf-f-u\] The following identities hold on ${\Gamma}$ for a function $u$ of class $C^2$ defined in (or extended properly to) a tubular neighborhood $U$ of the surface ${\Gamma}$, $$\begin{aligned}
\frac{\partial}{\partial{n}}({{\nabla_{{\Gamma}}}}u) &= D^2u{n}- {n}^T D^2u{n}{n}\label{E:normal-deriv-of-tangential-grad}
\\
{{\nabla_{{\Gamma}}}}\left(\frac{\partial u}{\partial{n}}\right)
&= {n}^T D^2u + {{\nabla_{{\Gamma}}}}u^T{{\nabla_{{\Gamma}}}}{n}- {n}^T D^2u{n}{n}.
\label{E:tangential-grad-of-normal-deriv}\end{aligned}$$ If the function $u$ is constant in the normal direction to ${\Gamma}$, i.e. $\frac{\partial u}{\partial{n}}=0$ on ${\Gamma}$, $$\label{E:D2u.n=...}
{n}^T D^2u = -{{\nabla_{{\Gamma}}}}u^T{{\nabla_{{\Gamma}}}}{n},$$ $$\label{E:mixed-normal-tangential-deriv}
{{\nabla_{{\Gamma}}}}\left(\frac{\partial u}{\partial {n}}\right) = 0, \qquad
\frac{\partial}{\partial {n}}\left({{\nabla_{{\Gamma}}}}u \right)
= -{{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}u.$$
We start by computing the normal derivative of the tangential gradient. For this, we resort to the signed distance function extension of the normal ${n}$. $$\begin{aligned}
\frac{\partial}{\partial{n}}({{\nabla_{{\Gamma}}}}u)
&= \partial_{x_j}(u_{x_i} - u_{x_k}{n}_k{n}_i) {n}_j,
\qquad i=0,\ldots,d-1 \\
&= \left( \partial_{x_j}(u_{x_i}
- u_{x_k} b_{x_k}b_{x_i}) b_{x_j} \right)|_{\Gamma}, \qquad i=0,\ldots,d-1 \\
&= \left( u_{x_ix_j} b_{x_j} - u_{x_kx_j} b_{x_k}b_{x_i} b_{x_j}
- u_{x_k} b_{x_kx_j}b_{x_i} b_{x_j} - u_{x_k} b_{x_k}b_{x_ix_j} b_{x_j} \right)|_{\Gamma}\\
&= \left( u_{x_ix_j} b_{x_j} - u_{x_kx_j} b_{x_k}b_{x_i} b_{x_j} \right)|_{\Gamma}\\
& = D^2u{n}- {n}^TD^2u{n}{n}.\end{aligned}$$ Note that $b_{x_ix_j} b_{x_j} = b_{x_jx_i} b_{x_j}
= \frac{1}{2}\partial_{x_i} (b_{x_j}b_{x_j}) = 0$ because $|\nabla b|^2 = b_{x_j}b_{x_j} = 1$.\
Now we compute the tangential gradient of the normal derivative, $$\begin{aligned}
{{\nabla_{{\Gamma}}}}\left( \frac{\partial u}{\partial{n}} \right)
&= {{\nabla_{{\Gamma}}}}\left( \nabla u\cdot{n}\right)
= \partial_{x_i}(u_{x_k}{n}_k)
- \partial_{x_j}(u_{x_k}{n}_k) {n}_j{n}_i, \qquad i=0,\ldots,d-1 \\
&= \left( \partial_{x_i}(u_{x_k}b_{x_k})
- \partial_{x_j}(u_{x_k}b_{x_k}) b_{x_j}b_{x_i} \right) \mathlarger{|}_{\Gamma}, \qquad i=0,\ldots,d-1 \\
&= ( u_{x_kx_i}b_{x_k} + u_{x_k}b_{x_kx_i}
- u_{x_kx_j}b_{x_k} b_{x_j}b_{x_i}
- u_{x_k} \underbrace{b_{x_kx_j} b_{x_j}}_{{}=0} b_{x_i} ) \mathlarger{|}_{\Gamma}\\
&= ( u_{x_kx_i}b_{x_k} + (u_{x_k}b_{x_kx_i}
- (u_{x_l} b_{x_l}) \underbrace{b_{x_k} b_{x_kx_i}}_{{}=0} )
- u_{x_kx_j}b_{x_k} b_{x_j}b_{x_i} )|_{\Gamma}\\
&= {n}^T D^2u + {{\nabla_{{\Gamma}}}}u^T{{\nabla_{{\Gamma}}}}{n}- {n}^T D^2u{n}{n}.\end{aligned}$$ We used the fact that $D^2b$ is symmetric and ${{\nabla_{{\Gamma}}}}{n}= D^2b|_{\Gamma}$.\
To prove , we start with the assumption $\frac{\partial u}{\partial{n}} = (\nabla u\cdot\nabla b)|_{\Gamma}=0$ and differentiate, $$\begin{aligned}
0 &= \partial_{x_j}(\nabla u\cdot\nabla b) = \partial_{x_j}(u_{x_i}b_{x_i})
= u_{x_ix_j}b_{x_i} + u_{x_i}b_{x_ix_j} \\
&= u_{x_ix_j}b_{x_i} + b_{x_ix_j}u_{x_i} - u_{x_k} b_{x_k}b_{x_i}b_{x_ix_j}
\qquad (b_{x_i}b_{x_ix_j} = 0) \\
&= u_{x_ix_j}b_{x_i} + b_{x_ix_j}(u_{x_i} - u_{x_k} b_{x_k}b_{x_i}) \\
\Rightarrow \
&\nabla b^T D^2u = -(\nabla u - \nabla u\cdot\nabla b \nabla b)^T D^2b,\end{aligned}$$ which on the surface ${\Gamma}$ is equivalent to $$\label{E:D2u.n}
{n}^T D^2u = -{{\nabla_{{\Gamma}}}}u^T {{\nabla_{{\Gamma}}}}{n}.$$ The identities follow trivially from substituted in , .
\[P:tangential-greens\] For a function $f \in C^1({\Gamma})$ and a vector $\vec{\omega} \in
C^1({\Gamma})^d$, we have the following tangential Green’s formula $$\label{E:greens-formula}
{\int_{{\Gamma}}}f {{\rm{div}_{{\Gamma}}}}\vec{\omega} + {{\nabla_{{\Gamma}}}}f \cdot \vec{\omega} dS
= {\int_{{\Gamma}}}\kappa f \vec{\omega} \cdot \nu dS.$$
Shape Differentiation {#S:shape-diff}
---------------------
We would like to understand how a quantity depending on a surface ${\Gamma}$ (or a domain ${\Omega}$) changes when ${\Gamma}$ is deformed by a given velocity field. For this, we consider a hold-all domain $\mathcal{D}$ (which may or may not be the image domain), containing the surface ${\Gamma}$, and a smooth vector field ${{\vec V}}$ defined on $\mathcal{D}$. The vector field ${{\vec V}}$ is used to define the continuous sequence of perturbed surfaces $\{{\Gamma}_t\}_{t\geq 0}$, with ${\Gamma}_0:={\Gamma}$. Each point $X\in{\Gamma}_0$ follows $$\label{ode:mapping}
\frac{dx}{dt}={{\vec V}}(x(t)),\quad \forall t\in [0,T],
\qquad
x(0)=X.$$ This defines the mapping $x(t,\cdot): X\in{\Gamma}\to x(t,X)\in {\mathbb R}^d$ and the perturbed sets\
${\Gamma}_t=\{x(t,X):~~X\in{\Gamma}_0\}$ (similarly perturbations of domains ${\Omega}_t=\{x(t,X):~~X\in{\Omega}_0\}$ for domains ${\Omega}(={\Omega}_0)$ contained by ${\Gamma}$).
Let $J({\Gamma})$ be a shape energy, namely a mapping that associates to surfaces ${\Gamma}$ a real number. The Eulerian derivative, or [*shape derivative*]{}, of the energy $J({\Gamma})$ at ${\Gamma}$ in the direction of the vector field ${{\vec V}}$, is defined as the limit $$dJ({\Gamma};{{\vec V}})=\lim_{t\to 0}
\frac{1}{t}\big(J({\Gamma}_t)-J({\Gamma})\big).$$ We define the shape derivatives $dJ({\Omega};{{\vec V}})$ for domain energies $J({\Omega})$ similarly. For more information on the concept of shape derivatives (including the definition and other properties), we refer to the book [@Delfour-Zolesio-01] by Delfour-Zolesio-01 and Zol[é]{}sio.
We now recall a series of results from shape differential calculus in ${\mathbb R}^d$.
\[L:deriv-domain\] Let $\phi\in W^{1,1}({\mathbb R}^d)$ and ${\Omega}\subset{\mathbb R}^d$ be an open and bounded domain with boundary ${\Gamma}=\partial{\Omega}$ of class $C^1$. Then the energy $ J({\Omega})=\int_{{\Omega}}\phi dx $ is shape differentiable. The shape derivative of $J({\Omega})$ is given by $$dJ({\Omega};{{\vec V}})={\int_{{\Gamma}}}\phi V dS,$$ where $V={{\vec V}}\cdot{n}$ is the normal component of the velocity.
\[L:deriv-bound\] Let $\psi\in W^{2,1}({\mathbb R}^d)$ and ${\Gamma}$ be of class $C^2$. Then the energy $ \ J({\Gamma})={\int_{{\Gamma}}}\psi dS \ $ is shape differentiable and the derivative $$dJ({\Gamma};{{\vec V}})={\int_{{\Gamma}}}\big(\nabla\psi\cdot{{\vec V}}+\psi {{\rm div}}_{{\Gamma}}{{\vec V}}\big)dS
= {\int_{{\Gamma}}}\big(\partial_{n}\psi + \psi \kappa\big) V dS,$$ depends on the normal component $V = {{\vec V}}\cdot{n}$ of the velocity ${{\vec V}}$.
Let us now consider more general energies $J({\Gamma})$. Specifically we are interested in computing shape derivatives for energies of the form $$\label{E:general-energies}
J({\Gamma})={\int_{{\Gamma}}}\varphi(x,{\Gamma})dS, \qquad
J({\Omega})={\int_{{\Omega}}}\phi(x,{\Omega}) dx,$$ in which the weight functions $\varphi, \phi$ depend not only on the spatial position $x$, but also on the shape ${\Gamma}, {\Omega}$. Examples are a weight function $\varphi(x,{\Gamma}) = \varphi(x,{n})$ that depends on the normal of the surface ${\Gamma}$ [@Faugeras-Keriven-98; @Kimmel-Bruckstein-03], or a weight function $\phi(x,{\Omega})=\phi(x,u({\Omega}))$ that depends on the solution $u$ of a PDE defined on ${\Omega}$ [@Chan-Vese-01; @Dogan-Morin-Nochetto-08; @Hintermueller-Ring-04]. To handle the computation of the shape derivatives of such energies we need to take care of the derivative of $\varphi,\phi$ with respect to the shape ${\Gamma},{\Omega}$. For this we recall the notions of [*material derivative*]{} and [*shape derivative*]{}.
\[D:material-deriv\] The [*material derivative*]{} $\dot{\varphi}({\Gamma};{{\vec V}})$ of $\varphi({\Gamma})$ at ${\Gamma}$ in direction ${{\vec V}}$ is defined as follows $$\dot{\varphi}({\Gamma};{{\vec V}})=
\lim_{t\to 0}\frac{1}{t}\big( \varphi(x(t,\cdot),{\Gamma}_t)
-\varphi(\cdot,{\Gamma}_0) \big),$$ where the mapping $ x(t,\cdot)$ is defined as in (\[ode:mapping\]). A similar definition holds for domain functions $\phi({\Omega})$.
\[D:shape-deriv\] The [*shape derivative*]{} $\phi({\Omega})$ at ${\Omega}$ in the direction ${{\vec V}}$ is defined to be $$\phi'({\Omega};{{\vec V}})=\dot{\phi}({\Omega};{{\vec V}})-\nabla \phi\cdot{{\vec V}}.$$ Accordingly, for surface functions $\varphi({\Gamma})$, the shape derivative is defined to be $$\varphi'({\Gamma};{{\vec V}})=\dot{\varphi}({\Gamma};{{\vec V}})-
\nabla_{\Gamma}\varphi\cdot{{\vec V}}\vert_{{\Gamma}}.$$
The shape derivative concept enables us to compute the change in the shape dependent quantities, such as the normal ${n}$ and the mean curvature $\kappa$, with respect to deformations of the shape by given velocity fields.
\[L:deriv-geom\] The shape derivatives of the normal ${n}$ and the mean curvature $\kappa$ of a surface ${\Gamma}$ of class $C^2$ with respect to velocity ${{\vec V}}\in C^2$ are given by $$\begin{aligned}
{n}' &= {n}'({\Gamma};{{\vec V}}) = -{{\nabla_{{\Gamma}}}}V \label{E:shape-deriv-normal}, \\
\kappa' &= \kappa'({\Gamma};{{\vec V}}) = -\Delta_{\Gamma}V, \label{E:shape-deriv-curv}\end{aligned}$$ where $V={{\vec V}}\cdot{n}$ is the normal component of the velocity. Moreover, the shape derivative of the tangential gradient of a function $u$ of class $C^1$ defined in (or extended properly to) a tubular neighborhood $U$ of the surface ${\Gamma}$ is $$\label{E:deriv-of-tangential-grad}
({{\nabla_{{\Gamma}}}}u)' = {{\nabla_{{\Gamma}}}}u' + {{\nabla_{{\Gamma}}}}u\cdot{{\nabla_{{\Gamma}}}}V {n}+ \frac{\partial u}{\partial{n}}{{\nabla_{{\Gamma}}}}V.$$
The derivations for ${n}'$ and $\kappa'$ can be found in [@Hintermueller-Ring-03 Sect. 3]. Once we have the result for ${n}'$, we can proceed with the following $$\begin{aligned}
({{\nabla_{{\Gamma}}}}u)' &= (\nabla u - \nabla u\cdot{n}{n})'
= \nabla u' - \nabla u'\cdot{n}{n}- \nabla u\cdot{n}'{n}- \nabla u\cdot{n}{n}' \\
&= {{\nabla_{{\Gamma}}}}u' + \nabla u\cdot{{\nabla_{{\Gamma}}}}V{n}+ \nabla u\cdot{n}{{\nabla_{{\Gamma}}}}V \\
&= {{\nabla_{{\Gamma}}}}u' + {{\nabla_{{\Gamma}}}}u\cdot{{\nabla_{{\Gamma}}}}V{n}+ \frac{\partial u}{\partial{n}}{{\nabla_{{\Gamma}}}}V.\end{aligned}$$
Now we state the shape derivatives of the general shape energies .
\[T:first-shape-deriv\] Let $\phi=\phi({\Omega})$ be given so that the material derivative $\dot{\phi}({\Omega};{{\vec V}})$ and the shape derivative $\phi'({\Omega};{{\vec V}})$ exist. Then, the shape energy $J({\Omega})$ in (\[E:general-energies\]) is shape differentiable and we have $$\label{E:deriv-domain}
dJ({\Omega};{{\vec V}})={\int_{{\Omega}}}\phi'({\Omega};{{\vec V}})dx + {\int_{{\Gamma}}}\phi V dS.$$ For surface functions $\varphi({\Gamma})$, the shape derivative of $J({\Gamma})$ in (\[E:general-energies\]) is given by $$dJ({\Gamma};{{\vec V}})={\int_{{\Gamma}}}\varphi'({\Gamma};{{\vec V}})dS+{\int_{{\Gamma}}}\kappa \varphi V dS,$$ whereas if $\varphi(\cdot,{\Gamma})=\psi(\cdot,{\Omega})\vert_{\Gamma}$, then we obtain $$\label{E:deriv-bound}
dJ({\Gamma};{{\vec V}})={\int_{{\Gamma}}}\psi'({\Omega};{{\vec V}})\vert_{\Gamma}dS
+{\int_{{\Gamma}}}\left(\frac{\partial\psi}{\partial{n}}+\kappa \psi \right) V dS.$$
We conclude this section with a Riesz representation theorem, the Hadamard-Zol[é]{}sio Theorem.
\[T:hadamard\] The shape derivative of a surface or domain energy always has a representation of the form $$dJ({\Gamma};{{\vec V}})=\langle G, V \rangle_{{\Gamma}},$$ where we denote by $\langle \cdot, \cdot \rangle_{{\Gamma}}$ a suitable duality pairing on ${\Gamma}$; that is, the shape derivative is concentrated on ${\Gamma}$.
Let us point out that an implication of this theorem is that the shape derivative $dJ({\Gamma};{{\vec V}})$ depends only on $V = {{\vec V}}\cdot {n}$, the normal component of the velocity. *For this reason, we will use $V$ in our derivations from now on and assume a normal extension when we need the velocity extended to a neighborhood of the surface. Hence, without loss of generality, we have the following assumptions on $V$ and ${{\vec V}}$* $$\label{E:vel-assumptions}
{{\vec V}}= V {n}, \qquad
\frac{\partial V}{\partial{n}} = 0 \quad \textrm{ on } {\Gamma}.$$
Deriving the first shape derivatives using only scalar velocities $V$, thus normal velocities , may at first appear as a restriction, since one is not always able to work with scalar velocities $V$ and may need to use arbitrary vector velocities ${{\vec V}}$. Because of Theorem \[T:hadamard\], this turns out not to be a problem; the first shape derivatives computed with scalar velocities $V$ will be the same as those computed with arbitrary ${{\vec V}}$ with $V={{\vec V}}\cdot{n}$. Moreover, in the case of second shape derivatives introduced in the next subsection, the formulas obtained with scalar velocities $V$ are the same around critical shapes as those obtained with arbitrary vector velocities ${{\vec V}}$ (given $V={{\vec V}}\cdot{n}$) [@Bucur-Zolesio-97; @Novruzi-Pierre-02].
The Second Shape Derivative {#S:second-deriv}
---------------------------
We continue to use the scalar velocity fields $V, W$ (corresponding to vector velocity fields ${{\vec V}}, {{\vec W}}$ by ) to perturb ${\Gamma}, {\Omega}$ and we define the second shape derivative as follows $$d^2J({\Gamma};V,W)= d\left(dJ({\Gamma};V)\right)({\Gamma};W), \quad
d^2J({\Omega};V,W)= d\left(dJ({\Omega};V)\right)({\Omega};W).$$ The second shape derivatives of functions $\phi({\Omega})$, $\varphi({\Gamma})$ can be defined similarly based on Definition \[D:shape-deriv\]. Now we can use this to compute the second shape derivative of the domain and the surface energies. We give these results below.
\[L:hess-domain\] Let $\phi\in W^{2,1}({\mathbb R}^d)$ and ${\Gamma}$ be of class $C^2$. Then the second shape derivative of the energy $$\label{E:func-domain-in-L1}
J({\Omega})=\int_{{\Omega}}\phi dx$$ at ${\Omega}$ with respect to scalar velocity fields $V$, $W$ is given by $$d^2J({\Omega};V,W) = {\int_{{\Gamma}}}\left( \partial_{n}\phi
+ \kappa \phi \right) V W dS.$$
\[L:hess-bound\] Let $\psi\in W^{3,1}({\mathbb R}^d)$ and ${\Gamma}$ be of class $C^2$. Then the second shape derivative of the energy $$\label{E:func-bound-in-L2}
J({\Gamma})={\int_{{\Gamma}}}\psi dS$$ at ${\Gamma}$ with respect to scalar velocity fields $V$, $W$ is given by $$d^2J({\Gamma};V,W) = {\int_{{\Gamma}}}\left( \psi {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W
+ \left( \partial_{{n}{n}} \psi + 2\kappa \partial_{n}\psi
+ (\kappa^2 - \sum \kappa_i^2)\psi \right) V W \right) dS.$$
We employ Lemmas \[L:hess-domain\] and \[L:hess-bound\] and Definition \[D:shape-deriv\] to state the second shape derivative for more general energies , in which the weight functions also depend on the shape. The assumptions are crucial to obtaining the following result.
\[T:second-shape-deriv\] Let $\phi=\phi(x,{\Omega})$ be given so that the first and the second shape derivatives $\phi'({\Omega};V)$, $\phi''({\Omega};V,W)$ exist. Then, the second shape derivative of the domain energies in (\[E:general-energies\]) is given by $$\label{E:hess-domain}
d^2J({\Omega};V,W)={\int_{{\Omega}}}\phi''dx +
{\int_{{\Gamma}}}\left(\phi'_W V+ \phi'_V W \right)dS
+ {\int_{{\Gamma}}}\left( \partial_{n}\phi + \kappa \phi \right) V W dS$$ where $\phi'_V = \phi'({\Omega};V)$, $\phi'' = \phi''({\Omega};V,W)$. For surface functions $\varphi({\Gamma})$ in (\[E:general-energies\]) with $\varphi(\cdot,{\Gamma})=\psi(\cdot,{\Omega})\vert_{{\Gamma}}$ we obtain $$\begin{aligned}
d^2J({\Gamma};V,W)&=&{\int_{{\Gamma}}}\psi'' dS
+ {\int_{{\Gamma}}}\left( \left(\partial_{n}\psi'_W + \kappa \psi'_W \right) V
+ \left(\partial_{n}\psi'_V + \kappa \psi'_V \right) W \right)dS
\label{E:hess-bound} \\
&+& {\int_{{\Gamma}}}\left( \psi {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W
+ \left( \partial_{{n}{n}} \psi + 2\kappa \partial_{n}\psi
+ (\kappa^2 - \sum \kappa_i^2)\psi \right) V W \right) dS. \nonumber\end{aligned}$$ where $\psi'_V = \psi'({\Gamma};V)\vert_{\Gamma}$, $\psi''_{V,W} =\psi''({\Gamma};V,W)\vert_{\Gamma}$.
For $J({\Omega}) = {\int_{{\Omega}}}\phi(x,{\Omega}) dx$, the first shape derivative at ${\Omega}$ in direction $V$ is $$\begin{array}{lcccc}
dJ({\Omega};V) &=& \underbrace{{\int_{{\Omega}}}\phi'({\Omega};V) dx}
& + &\underbrace{{\int_{{\Gamma}}}\phi V dS}. \\
& & J_1 & & J_2
\end{array}$$ To obtain the second shape derivative, we take the derivatives of $J_1$ and $J_2$. $$\begin{aligned}
dJ_1({\Omega};W) &= {\int_{{\Omega}}}\phi''({\Omega};V,W) dx
+ {\int_{{\Gamma}}}\phi'({\Omega};V)W dS, \\
dJ_2({\Omega};W) &= {\int_{{\Gamma}}}\phi'({\Omega};W)V dS
+ {\int_{{\Gamma}}}(\partial_{n}\phi V + \phi V \kappa) W dS.\end{aligned}$$ Plugging in $d^2J({\Omega};V,W)$ and reorganizing the terms yields the results. We obtain the second shape derivative for the surface energy similarly.
More details on the structure of the second shape derivative can be found in [@Bucur-Zolesio-97; @Delfour-Zolesio-01; @Novruzi-Pierre-02]. We should reemphasize that the use of assumptions has allowed us to derive the simpler forms in Theorem \[T:second-shape-deriv\], compared to these references, which investigate the second shape derivative with more general vector velocities or perturbations. For vector velocities with tangential components, the second shape derivative includes additional terms, but these terms disappear at critical shapes. Thus the formula obtained with scalar velocities is sufficient to characterize the second shape derivative at critical shapes, because in this situation, the second shape derivative depends only on the normal components of the vector velocity [@Bucur-Zolesio-97; @Novruzi-Pierre-02].
Shape Derivatives of the Model Energies {#S:shape-derivs-of-energies}
=======================================
We compute the first and second shape derivatives for each shape energy class introduced in Section \[S:shape-energies\].
Minimal Surface Energies {#S:surface-energies}
------------------------
We compute the first and the second shape derivatives of surface energies with weight functions that depend only on the local geometry at each point of the surface. We start with an isotropic energy, the Geodesic Active Contour Model, then consider anisotropic, i.e. normal-dependent energies; we describe the case of a curvature-dependent weight as well.
[**Isotropic surface energies.**]{} The archetype shape energy in image processing is the Geodesic Active Contour model [@Caselles-Kimmel-Sapiro-97; @Caselles-Kimmel-Sapiro-97-3d]: $$\label{E:gac-energy}
J({\Gamma}) := {\int_{{\Gamma}}}g(x)dS + \gamma {\int_{{\Omega}}}g(x) dx,$$ where ${\Gamma}$ is a surface in ${\mathbb R}^d$ and ${\Omega}$ is the domain enclosed by ${\Gamma}$. The first and second shape derivatives directly follow from Lemmas \[L:deriv-domain\], \[L:deriv-bound\], \[L:hess-domain\], \[L:hess-bound\] and the derivations can be found in [@Hintermueller-Ring-03].
The first shape derivative of the energy at ${\Gamma}$ with respect to velocity $V$ is given by $$dJ({\Gamma};V)= {\int_{{\Gamma}}}\Big((\kappa+\gamma)g(x)
+ \partial_{n}g(x)\Big) V dS.$$ The second shape derivative of with respect to velocities $V, W$ is given by $$d^2J({\Gamma};V,W) = {\int_{{\Gamma}}}\left(g {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W
+ \left( \partial_{{n}{n}} g
+ (2\kappa + \gamma) \partial_{n}g
+ (\kappa^2 - \sum \kappa_i^2 + 2\gamma\kappa) g \right) V W \right)dS.$$
[**Anisotropic surface energies.**]{} Now we compute the first and the second shape derivatives of the anisotropic surface energy $$\label{E:anisotropic-energy}
J({\Gamma}) := {\int_{{\Gamma}}}g(x,{n})dS.$$ The weight function $g(x,{n})$ depends on the orientation of the normal of the surface. We assume the derivatives $g_x, g_{xx}$ with respect to the first argument $x$ and the derivatives $g_y, g_{yy}$ with respect to the second argument ${n}$, also the mixed derivatives $g_{xy}, g_{yx}$ are well-defined. Applications of are in image segmentation [@Kimmel-Bruckstein-03] and in multiview stereo reconstruction [@Faugeras-Keriven-98; @Kolev-Pock-Cremers-10].
The first shape derivative of the anisotropic surface energy at ${\Gamma}$ with respect to velocity $V$ is given by $$\label{E:aniso-1st-deriv}
dJ({\Gamma};V)= {\displaystyle {\int_{{\Gamma}}}\left( \kappa g + \partial_{n}g \right)
V - g_y \cdot {{\nabla_{{\Gamma}}}}V dS }
= {\displaystyle {\int_{{\Gamma}}}\left(\kappa g + \partial_{n}g
+ {{\rm{div}_{{\Gamma}}}}(g_y)_{\Gamma}\right) V dS, }$$ where $g_y$ is the derivative of $g(x,{n})$ with respect to its second variable.
To derive the first derivative, we use Theorem \[T:first-shape-deriv\] with $\psi = g(x,{n})$. Then $$\psi'({\Gamma};V) = g_y \cdot {n}' = - g_y \cdot {{\nabla_{{\Gamma}}}}V$$ using . We also need to compute the normal derivative of $g(x,{n})$. We have $\partial_{n}\psi = \partial_{n}(g(x,{n}))
= \partial_{n}g + g_y^T\partial_{n}{n}=
\partial_{n}g$, because $\partial_{n}{n}= 0$ by Equation (assuming extension by signed distance function , ).\
We then substitute $\psi'$ and $\partial_{n}\psi$ in and obtain $$dJ({\Gamma};V)= {\int_{{\Gamma}}}\left( \kappa g + \partial_{n}g\right)
V - g_y \cdot {{\nabla_{{\Gamma}}}}V dS .$$ We can apply the tangential Green’s formula to the last term of the integral, and use the identity ${{\rm{div}_{{\Gamma}}}}(\vec{\omega})_{\Gamma}= {{\rm{div}_{{\Gamma}}}}(\vec{\omega})
- \kappa \vec{\omega} \cdot {n}$ [@Delfour-Zolesio-01 Chap. 8] to obtain .
\[P:hess-anisotropic\] The second shape derivative of the anisotropic surface energy at ${\Gamma}$ with respect to velocities $V,W$ is given by $$\begin{aligned}
d^2J({\Gamma};V,W) =& {\int_{{\Gamma}}}{{\nabla_{{\Gamma}}}}V \cdot \left( (g-g_y \cdot {n})Id +
g_{yy}\right) \cdot {{\nabla_{{\Gamma}}}}W dS \\
&+ {\int_{{\Gamma}}}\left(\partial_{{n}{n}}g+2\kappa \partial_{n}g
+ (\kappa^2 - \sum \kappa_i^2) g \right)VW dS \\
&- {\int_{{\Gamma}}}(\kappa g_y -g_y^T{{\nabla_{{\Gamma}}}}{n}+ {n}^T g_{xy}^T)
\cdot ( {{\nabla_{{\Gamma}}}}W \, V + {{\nabla_{{\Gamma}}}}V \, W ) dS .\end{aligned}$$ where $Id$ is the $d\times d$ identity matrix.
Use Theorem \[T:second-shape-deriv\] with $\psi = g(x,{n})$. Then $\partial_{n}\psi = \partial_{n}g, \
\partial_{{n}{n}} \psi = \partial_{{n}{n}} g,$ and $$\begin{aligned}
\psi'({\Gamma};V) &= - g_y \cdot {{\nabla_{{\Gamma}}}}V
\\
\psi''({\Gamma};V,W) &= - \left( g_y \right)' \cdot {{\nabla_{{\Gamma}}}}V
- g_y \cdot \left( {{\nabla_{{\Gamma}}}}V \right)'
\qquad (\mathrm{use} \eqref{E:deriv-of-tangential-grad}, \eqref{E:vel-assumptions})
\\
&= {{\nabla_{{\Gamma}}}}V \cdot g_{yy} \cdot {{\nabla_{{\Gamma}}}}W
- g_y \cdot {n}\, {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W
\\
&= {{\nabla_{{\Gamma}}}}V \cdot \left( g_{yy} - g_y
\cdot {n}\, Id \right) \cdot {{\nabla_{{\Gamma}}}}W
\\
\partial_{n}\psi'({\Gamma};V)
&= - \partial_{n}g_y \cdot {{\nabla_{{\Gamma}}}}V
- g_y \cdot \partial_{n}{{\nabla_{{\Gamma}}}}V
\\
&= -(g_{yx} {n}) \cdot {{\nabla_{{\Gamma}}}}V - g_y \cdot (-{{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}V)
\\
&= (g_y^T {{\nabla_{{\Gamma}}}}{n}- {n}^T g_{yx}^T) {{\nabla_{{\Gamma}}}}V,\end{aligned}$$ since $\partial_{n}{{\nabla_{{\Gamma}}}}V = -{{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}V$ by Lemma \[L:deriv-surf-f-u\] and assumptions (\[E:vel-assumptions\]). The function $g_{yx}$ is the Hessian obtained by differentiating with respect to the second variable ${n}$, and to the first variable $x$. Now we substitute the derivatives of $\psi$ in to obtain $$\begin{aligned}
d^2J({\Gamma};V,W)=&{\int_{{\Gamma}}}{{\nabla_{{\Gamma}}}}V \cdot \left( g_{yy} - g_y
\cdot {n}\, Id \right) \cdot {{\nabla_{{\Gamma}}}}W dS \\
&- {\int_{{\Gamma}}}\left( ({n}^T g_{yx}^T - g_y^T{{\nabla_{{\Gamma}}}}{n}) \cdot {{\nabla_{{\Gamma}}}}W
+ \kappa g_y \cdot {{\nabla_{{\Gamma}}}}W \right) V \\
&- {\int_{{\Gamma}}}\left( ({n}^T g_{yx}^T - g_y^T{{\nabla_{{\Gamma}}}}{n}) \cdot {{\nabla_{{\Gamma}}}}V
+ \kappa g_y \cdot {{\nabla_{{\Gamma}}}}V \right) W dS \\
&+ {\int_{{\Gamma}}}\left( g {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W
+ \left( \partial_{{n}{n}} g + 2\kappa \partial_{n}g
+ (\kappa^2 - \sum \kappa_i^2)g \right) V W \right) dS.\end{aligned}$$ Reorganizing the terms yields the result.
[**Curvature-dependent Energies.**]{} Next we compute the shape derivative of the curvature dependent surface energy $$\label{E:curvature-energy}
J({\Gamma}) := {\int_{{\Gamma}}}g(x,\kappa)dS.$$ The weight function $g(x,\kappa)$ depends on the mean curvature of the surface. We assume the derivative $g_x$ with respect to the first argument $x$ and the derivatives $g_z, g_{zz}, g_{zzz}$ with respect to the second argument $\kappa$, also the mixed derivatives $g_{zxz}, g_{zzx}$ are well-defined. Variants of the energy have been used to impose regularity of curves or surfaces in shape identification problems [@Clarenz-etal-04; @Droske-Bertozzi-10; @Sundar-etal-09].
\[P:deriv-gen-energy\] The first shape derivative of the curvature-dependent surface energy with $g = g(x,\kappa)$ at ${\Gamma}$ with respect to velocity ${{\vec V}}$ is given by $$\begin{aligned}
dJ({\Gamma};V) &= {\int_{{\Gamma}}}- g_z {\Delta_{{\Gamma}}}V
+ \left( g \kappa - g_z \sum \kappa_i^2 + \partial_{n}g \right) V dS, \\
&= {\int_{{\Gamma}}}\left( -{\Delta_{{\Gamma}}}(g_z)
+ g \kappa - g_z \sum \kappa_i^2 + \partial_{n}g \right) V dS,\end{aligned}$$ where ${\Delta_{{\Gamma}}}(g_z)$ denotes the total derivative $${\Delta_{{\Gamma}}}(g_z) = {\Delta_{{\Gamma}}}g_z + g_{zz} {\Delta_{{\Gamma}}}\kappa + g_{zzz} |{{\nabla_{{\Gamma}}}}\kappa|^2
+ (g_{zxz} + g_{zzx})\cdot{{\nabla_{{\Gamma}}}}\kappa$$ and $g_z$ is the derivative of $g(x,\kappa)$ with respect to its second variable; $g_{zz}, g_{zxz}, g_{zzx}, g_{zzz}$ are defined similarly.
The proof of Proposition \[P:deriv-gen-energy\] for the shape derivative of is given by Doğan and Nochetto in [@Dogan-Nochetto-12], where they derive the shape derivative also for the more general surface energy $$J({\Gamma}) = {\int_{{\Gamma}}}g(x,{n},\kappa) dS.$$
Shape Energies with Integrals {#S:energies-with-integrals}
-----------------------------
We compute the shape derivatives for the domain and surface energies with integrals. These energies are found in applications where we need to aggregate properties over regions or surfaces by integration, for example, to compute statistics of given data [@Aubert-etal-03; @Cremers-Rousson-Deriche-07; @Paragios-Deriche-02a], and the integrals are used as parameters of the weight function.
[**Energies with Domain Integrals.**]{} We start by computing the shape derivatives for the energy with a weight function that depends on a single domain integral: $$\label{E:d-energy-w-integral}
J({\Omega}) = {\int_{{\Omega}}}g(x,I_w({\Omega})) dx,
\qquad I_w({\Omega}) = {\int_{{\Omega}}}w(x) dx.$$ We then consider an energy with a weight that depends on multiple domain integrals: $$\label{E:energy-w-multiple-integrals}
J({\Omega}) = {\int_{{\Omega}}}g(x,I_{w_1},\ldots,I_{w_m}) dx, \qquad
I_{w_i} = I_{w_i}({\Omega}) = {\int_{{\Omega}}}w_i(x) dx, \ i=1,\ldots,m.$$ We also consider the case in which the dependence on domain integrals is recursive: $$\label{E:energy-w-nested-integrals}
J({\Omega}) = {\int_{{\Omega}}}g_0(x,I_{g_1}) dx,$$ where $$\begin{gathered}
I_{g_k} = I_{g_k}({\Omega}) = {\int_{{\Omega}}}g_k(x,I_{g_{k+1}}) dx,
\quad k=1,\ldots,m-1, \\
I_{g_m} = I_{g_m}({\Omega}) = {\int_{{\Omega}}}g_m(x) dx.\end{gathered}$$ The first shape derivatives of , and (only for two levels of recursion) were computed in [@Aubert-etal-03]. The second shape derivatives of , are computed in this section.
\[P:deriv-d-energy-w-integral\] The first shape derivative of the shape energy at ${\Omega}$ with respect to velocity $V$ is given by $$dJ({\Omega};V) = {\int_{{\Gamma}}}\left( g(x,I_w({\Omega}))
+ I_{g_p}({\Omega}) w(x) \right) V dS,$$ The second shape derivative of with respect to velocities $V, W$ is given by $$\begin{aligned}
d^2J({\Omega};V,W) =& {\int_{{\Gamma}}}\left( \partial_{n}g + I_{g_p} \partial_{n}w
+ \kappa(g + I_{g_p} w) \right) V W dS \\
&+ {\int_{{\Gamma}}}g_p V dS {\int_{{\Gamma}}}w W dS + {\int_{{\Gamma}}}w V dS {\int_{{\Gamma}}}g_p W dS \\
&+ I_{g_{pp}} {\int_{{\Gamma}}}w V dS {\int_{{\Gamma}}}w W dS,\end{aligned}$$ We use the short notation for the functions $g = g(x,I_w({\Omega}))$, $w = w(x)$, and $g_p, g_{pp}$ denote the derivatives of $g(x,I_w({\Omega}))$ with respect to its second variable. $I_{g_p}, I_{g_{pp}}$ are the integrals of $g_p, g_{pp}$ over the domain ${\Omega}$.
Let $\phi(x,{\Omega}) = g(x,I_w({\Omega}))$ in Theorem \[T:first-shape-deriv\], and calculate $\phi'({\Omega};V)$. $$\phi'({\Omega};V) = g_p(x,I_w) I_w' = g_p(x,I_w) {\int_{{\Gamma}}}w(y) V dS(y).$$ We have used the fact that Lemma \[L:deriv-domain\] applies to $I_w$, namely we have $I_w' = dI_w({\Omega};V)$. We substitute $\phi'({\Omega};V)$ in the general form $$dJ({\Omega};V) = {\int_{{\Omega}}}g_p(x,I_w) \left( {\int_{{\Gamma}}}w(y) V dS(y) \right) dx
+ {\int_{{\Gamma}}}g(x,I_w) V dS.$$ Since $I_{g_p} = {\int_{{\Omega}}}g_p(x,I_w) dx$, we can exchange the order of integration and obtain $$dJ({\Omega};V) = {\int_{{\Gamma}}}\left( g(x,I_w) + I_{g_p} w(x) \right) V dS.$$ To compute the second shape derivative, we use Theorem \[T:second-shape-deriv\]. We need $\phi''=\phi''({\Omega};V,W)$, which we compute by Lemma \[L:hess-domain\] $$\begin{aligned}
\phi'' &= (g_p)' {\int_{{\Gamma}}}w(y) V dS
+ g_p \left( {\int_{{\Gamma}}}w(y) V dS \right)' \\
&= g_{pp} {\int_{{\Gamma}}}w(z) V dS(z) {\int_{{\Gamma}}}w(y) W dS(y)
+ g_p {\int_{{\Omega}}}\left( \partial_{n}w(y)
+ \kappa w(y) \right) V W dS.\end{aligned}$$ Substitute in $$\begin{aligned}
d^2J({\Omega};V,W) =& {\int_{{\Omega}}}g_{pp} dx {\int_{{\Gamma}}}w V dS {\int_{{\Gamma}}}w W dS
+ {\int_{{\Gamma}}}g_p dx {\int_{{\Omega}}}\left( \partial_{n}w + \kappa w \right) V W dS \\
&+ {\int_{{\Gamma}}}g_p V dS {\int_{{\Gamma}}}w W dS + {\int_{{\Gamma}}}g_p W dS {\int_{{\Gamma}}}w V dS \\
&+ {\int_{{\Gamma}}}( \partial_{n}g + \kappa g ) V W dS.\end{aligned}$$ Reorganizing the various terms yields the result.
The first shape derivative of the energy at ${\Omega}$ with respect to velocity $V$ is given by $$dJ({\Omega};V) = \mathlarger{{\int_{{\Gamma}}}} \left(g(x,I_{w_1},\ldots,I_{w_m}) +
\sum^m_{i=1} I_{g_{p_i}} w_i(x) \right) V dS.$$ The second shape derivative of the energy at ${\Omega}$ with respect to velocity $V, W$ is $$\begin{aligned}
d^2J({\Omega};V,W) =& \mathlarger{{\int_{{\Gamma}}}}
\left( \partial_{n}g +
\sum^m_{i=1} I_{g_{p_i}} \partial_{n}w_i
+ \kappa(g + \sum^m_{i=1} I_{g_{p_i}} w_i) \right) V W dS \\
&+ \sum^m_{i=1} \left( {\int_{{\Gamma}}}g_{p_i} V dS {\int_{{\Gamma}}}w_i W dS
+ {\int_{{\Gamma}}}w_i V dS {\int_{{\Gamma}}}g_{p_i} W dS \right) \\
&+ \sum^m_{i,j=1} I_{g_{p_ip_j}} {\int_{{\Gamma}}}w_i V dS {\int_{{\Gamma}}}w_j W dS,\end{aligned}$$
The proof is essentially the same as that of Proposition \[P:deriv-d-energy-w-integral\], except that we need to keep track of indices and corresponding terms $w_i, I_{w_i}, g_{p_i}, I_{g_{p_i}},
g_{p_ip_j}, I_{g_{p_ip_j}}$.
The first shape derivative of the energy at ${\Omega}$ with respect to velocity $V$ is given by $$dJ({\Omega};V) = \mathlarger{{\int_{{\Gamma}}}} \left( g_0(x,I_{g_1})
+ \sum_{j=1}^{m-1} g_j(x,I_{g_{j+1}}) \prod_{i=0}^j I_{g_{i,p}} \right) V dS,$$ where $g_{i,p}$ denotes the derivative of $g_i(x,I_{g_{i+1}})$ with respect to its second argument.
Note that the shape derivative of $I_{g_m} = I_{g_m}({\Omega})$ is given by $$I_{g_m}' = {\int_{{\Gamma}}}g_m(x) V dS.$$ Now we compute the shape derivative of $I_{g_k} = I_{g_k}({\Omega})$. $$\begin{aligned}
I_{g_k}' = &{\int_{{\Omega}}}g_{k,p}(x,I_{g_{k+1}}) I_{g_{k+1}}' dx
+ {\int_{{\Gamma}}}g_k(x,I_{g_{k+1}}) V dS
= I_{g_{k,p}} I_{g_{k+1}}' + {\int_{{\Gamma}}}g_k V dS \\
= &I_{g_{k,p}} \left( I_{g_{k+1,p}} I_{g_{k+2}}' + {\int_{{\Gamma}}}g_{k+1} V dS \right)
+ {\int_{{\Gamma}}}g_k V dS \\
= &I_{g_{k,p}} I_{g_{k+1,p}} I_{g_{k+2}}' + I_{g_{k,p}} {\int_{{\Gamma}}}g_{k+1} V dS
+ {\int_{{\Gamma}}}g_k V dS \\
= &I_{g_{k,p}} I_{g_{k+1,p}} \ldots I_{g_m}'
+ I_{g_{k,p}} \ldots I_{g_{m-2,p}} {\int_{{\Gamma}}}g_{m-1} V dS + \ldots \\
&+ I_{g_{k,p}} {\int_{{\Gamma}}}g_{k+1} V dS + {\int_{{\Gamma}}}g_k V dS \\
= &I_{g_{k,p}} \ldots I_{g_{m-1,p}} {\int_{{\Gamma}}}g_m V dS
+ I_{g_{k,p}} \ldots I_{g_{m-2,p}} {\int_{{\Gamma}}}g_{m-1} V dS + \ldots \\
&+ I_{g_{k,p}} {\int_{{\Gamma}}}g_{k+1} V dS + {\int_{{\Gamma}}}g_k V dS.\end{aligned}$$ More concisely, $$I_{g_k}' = {\int_{{\Gamma}}}g_k V dS + \sum^m_{j=k+1} \prod^{j-1}_{i=k} I_{g_{i,p}} {\int_{{\Gamma}}}g_j V dS
= \mathlarger{{\int_{{\Gamma}}}} \left( g_k + \sum^m_{j=k+1} \prod^{j-1}_{i=k} I_{g_{i,p}} g_j \right) V dS.$$ Then the first shape derivative of energy is given by $dJ({\Gamma};V) = I_{g_0}'$.
[**Energies with Surface Integrals.**]{} Similar to the case with domain integrals, we compute the shape derivatives for energies with weight functions that depend on surface integrals $$\label{E:b-energy-w-integral}
J({\Gamma}) = {\int_{{\Gamma}}}g(x,I_w({\Gamma})) dS, \qquad I_w({\Gamma}) = {\int_{{\Gamma}}}w(x) dS.$$ The cases of multiple integral parameters and nested integrals in the weight function $g$ are straight-forward and are not included in this paper.
The first shape derivative of the energy at ${\Gamma}$ with respect to velocity $V$ is given by $$dJ({\Gamma};V) = {\int_{{\Gamma}}}\left( (g(x,I_w({\Gamma})) + I_{g_p} w(x)) \kappa
+ \partial_{n}g(x,I_w({\Gamma})) + I_{g_p} \partial_{n}w(x) \right) V dS.$$ The second shape derivative of at ${\Gamma}$ with respect to velocities $V, W$ is given by $$\begin{aligned}
d^2J({\Gamma};V,W) =& {\int_{{\Gamma}}}(g + I_{g_p} w) {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W dS \\
&+ {\int_{{\Gamma}}}\left( \partial_{{n}{n}} g + I_{g_p} \partial_{{n}{n}} w
+2(\partial_{n}g + I_{g_p}\partial_{n}w)\kappa
+(\kappa^2 - \Sigma\kappa_i^2)(g + I_{g_p}w) \right) V W dS \\
&+ {\int_{{\Gamma}}}\left( \partial_{n}g_p + g_p \kappa \right) V dS {\int_{{\Gamma}}}\left(
\partial_{n}w + w \kappa \right) W dS \\
&+ {\int_{{\Gamma}}}\left( \partial_{n}w + w \kappa \right) V dS {\int_{{\Gamma}}}\left(
\partial_{n}g_p + g_p \kappa \right) W dS \\
&+ I_{g_{pp}} {\int_{{\Gamma}}}\left( \partial_{n}w + w \kappa \right) V dS
{\int_{{\Gamma}}}\left( \partial_{n}w + w \kappa \right) W dS.\end{aligned}$$ We use the short notation for the functions $g = g(x,I_w({\Gamma}))$, $w = w(x)$, and $g_p, g_{pp}$ denote the derivatives of $g(x,I_w({\Gamma}))$ with respect to its second variable. $I_{g_p}, I_{g_{pp}}$ are the integrals of $g_p, g_{pp}$ over the surface ${\Gamma}$.
Let $\psi(x,{\Gamma}) = g(x,I_w({\Gamma}))$ in Theorem \[T:first-shape-deriv\]. By Lemma \[L:deriv-bound\], we have $$\psi' = \psi'({\Gamma};V)
= g_p(x,I_w) I_w' = g_p {\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS$$ Substitute in the general form $$dJ({\Gamma};V) = {\int_{{\Gamma}}}g_p dS {\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS
+ {\int_{{\Gamma}}}(\partial_{n}g + g \kappa) V dS.$$ We let $I_{g_p} = {\int_{{\Gamma}}}g_p(x,I_w) dS$, reorganize the terms, thus obtain the first shape derivative.\
Now we use Theorem \[T:second-shape-deriv\] and compute $\partial_{n}\psi = \partial_{n}g, \
\partial_{{n}{n}} \psi = \partial_{{n}{n}} g,$ and $$\begin{aligned}
\psi_V' = \psi'({\Gamma};V) = &g_p {\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS,
\
\partial_{n}\psi_V' = \partial_{n}g_p
{\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS, \\
\psi'' = \psi''({\Gamma};V,W)
= &(g_p)' {\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS +
g_p \left( {\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS \right)' \\
= &g_{pp}(x,I_w) {\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS
{\int_{{\Gamma}}}(\partial_{n}w + w\kappa)W dS \\
&+ g_p(x,I_w) {\int_{{\Gamma}}}w {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W dS \\
&+ g_p(x,I_w) {\int_{{\Gamma}}}\left( \partial_{{n}{n}} w + 2\kappa \partial_{n}w
+ (\kappa^2 - \sum \kappa_i^2)w \right) V W dS, \\\end{aligned}$$ using Lemma \[L:hess-bound\]. We substitute the derivatives of $\psi$ in and obtain $$\begin{aligned}
d^2J({\Gamma};V,W) =& {\int_{{\Gamma}}}g_{pp} dS {\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS
{\int_{{\Gamma}}}(\partial_{n}w + w\kappa)W dS \\
&+ {\int_{{\Gamma}}}g_p dS {\int_{{\Gamma}}}w {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W dS \\
&+ {\int_{{\Gamma}}}g_p dS {\int_{{\Gamma}}}\left( \partial_{{n}{n}} w + 2\kappa \partial_{n}w
+ (\kappa^2 - \sum \kappa_i^2)w \right) V W dS \\
&+ {\int_{{\Gamma}}}(\partial_{n}g_p + g_p\kappa) W dS
{\int_{{\Gamma}}}(\partial_{n}w + w\kappa)V dS \\
&+ {\int_{{\Gamma}}}(\partial_{n}g_p + g_p\kappa) V dS
{\int_{{\Gamma}}}(\partial_{n}w + w\kappa) W dS \\
&+ {\int_{{\Gamma}}}\left( g {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W
+\left(\partial_{{n}{n}} g + 2 \kappa \partial_{n}g +
(\kappa^2 - \sum \kappa_i^2) g \right) V W \right) dS.\end{aligned}$$ Reorganizing the various terms yields the result.
Higher-Order Energies {#S:higher-order-energies}
---------------------
In this section, we consider energies that have the form of a double integral with a weight function $g(x,y):{\mathbb R}^d\times{\mathbb R}^d\rightarrow{\mathbb R}$. These energies are used to model nonlocal interactions between two separate spatial locations $x, y$. Applications are found in road network detection [@Rochery-Jermyn-Zerubia-06] and topology control of curves for image segmentation [@LeGuyader-Vese-08; @Rocha-etal-09; @Sundar-Yezzi-05].
We introduce some notation that will simplify our derivations of the shape derivatives. We denote by $\tilde{g}(x,y)$ the symmetricization of the function $g(x,y)$: $$\label{E:sym-g}
\tilde{g}(x,y) = g(x,y) + g(y,x).$$ The derivatives of $\tilde{g}(x,y)$ are then given by $$\label{E:sym-g-derivs}
\begin{gathered}
\tilde{g}_x(x,y) = g_x(x,y) + g_y(y,x), \qquad
\tilde{g}_y(x,y) = g_y(x,y) + g_x(y,x), \\
\tilde{g}_{xy}(x,y) = g_{xy}(x,y) + g_{yx}(y,x),
\quad \ldots, \mathrm{\ and \ so \ on.}
\end{gathered}$$ We also introduce the domain integral of $\tilde{g}(x,y)$ and its derivative $$\label{E:sym-g-d-integral}
G(x,{\Omega}) = {\int_{{\Omega}}}\tilde{g}(x,y) dy, \qquad
G_x(x,{\Omega}) = {\int_{{\Omega}}}\tilde{g}_x(x,y) dy,$$ also the surface integral and its derivatives $$\label{E:sym-g-b-integral}
\begin{gathered}
G(x,{\Gamma}) = {\int_{{\Gamma}}}\tilde{g}(x,y) dS(y), \qquad
G_x(x,{\Gamma}) = {\int_{{\Gamma}}}\tilde{g}_x(x,y) dS(y), \\
G_{xx}(x,{\Gamma}) = {\int_{{\Gamma}}}\tilde{g}_{xx}(x,y) dS(y).
\end{gathered}$$ [**Higher-Order Domain Energies.**]{} We consider the following higher-order domain energy $$\label{E:domain-higher-order}
J({\Omega}) = {\int_{{\Omega}}}{\int_{{\Omega}}}g(x,y) \ dy dx,$$ and compute its first and the second shape derivatives.
\[P:domain-higher-order-deriv1\] The first shape derivative of the higher-order domain energy at ${\Omega}$ with respect to velocity $V$ is $$\label{E:domain-higher-order-deriv1}
dJ({\Omega};V) = {\int_{{\Gamma}}}G(x,{\Omega}) V dS = {\int_{{\Gamma}}}{\int_{{\Omega}}}\tilde{g}(x,y)dy V dS,$$ where $\tilde{g}(x,y)$ is defined by .
We define $\phi(x,{\Omega}) = {\int_{{\Omega}}}g(x,y) dy$, so that we work with a more concise form of the energy $$\label{E:domain-h-o-2}
J({\Omega}) = {\int_{{\Omega}}}\phi(x,{\Omega}) dx.$$ and use Theorem \[T:first-shape-deriv\]. We start by writing the shape derivative of $\phi$ $$\phi'({\Omega};V)(x) = {\int_{{\Gamma}}}g(x,y) V(y) dS(y),$$ We substitute $\phi'({\Omega};V)$ in the shape derivative of , $$\begin{aligned}
dJ({\Omega};V) &= {\int_{{\Omega}}}\phi'({\Omega};V) dx + {\int_{{\Gamma}}}\phi(x) V(x) dS(x) \\
&= {\int_{{\Omega}}}{\int_{{\Gamma}}}g(x,y) V(y) dS(y) dx
+ {\int_{{\Gamma}}}{\int_{{\Omega}}}g(x,y) dy V(x) dS(x) \\
&= {\int_{{\Omega}}}{\int_{{\Gamma}}}g(y,x) V(x) dS(x) dy
+ {\int_{{\Gamma}}}{\int_{{\Omega}}}g(x,y) dy V(x) dS(x) \\
&= {\int_{{\Omega}}}\left({\int_{{\Gamma}}}(g(x,y) + g(y,x)) dy\right) V(x) dS(x).\end{aligned}$$ and we let $\tilde{g}(x,y) = g(x,y)+g(y,x)$ to obtain the result .
Using the same strategy as in the proof of Proposition \[P:domain-higher-order-deriv1\], we can write the shape derivatives of even higher-order energies, for example, $$J({\Omega}) = {\int_{{\Omega}}}{\int_{{\Omega}}}{\int_{{\Omega}}}g(x,y,z) dz dy dx.$$ We set $\phi(x,{\Omega})={\int_{{\Omega}}}{\int_{{\Omega}}}g(x,y,z) dz dy$ and compute its shape derivative $$\phi'({\Omega};V)(x)
= {\int_{{\Gamma}}}\left( {\int_{{\Gamma}}}(g(x,y,z) + g(x,z,y) dz \right) V(y) dS(y),$$ using Proposition \[E:domain-higher-order-deriv1\]. Again we use the formula for the domain shape derivative in Theorem \[T:first-shape-deriv\] and substitute the current values of $\phi(x,{\Omega})$ and $\phi'(V;{\Omega})$: $$\begin{aligned}
dJ({\Omega};V) = &{\int_{{\Gamma}}}\left( {\int_{{\Omega}}}{\int_{{\Omega}}}(g(x,y,z)
+ g(x,z,y)) dz dx \right) V(y) dS(y) \\
&+ {\int_{{\Gamma}}}{\int_{{\Omega}}}{\int_{{\Omega}}}g(x,y,z) dz dy V(x) dS(x) \\
= &{\int_{{\Gamma}}}\left({\int_{{\Omega}}}{\int_{{\Omega}}}(g(x,y,z) + g(y,x,z)
+ g(y,z,x))dz dy \right) V(x) dS(x).\end{aligned}$$
\[P:domain-higher-order-deriv2\] The second shape derivative of the higher-order domain energy at ${\Omega}$ with respect to velocities $V,W$ is $$\label{E:domain-higher-order-deriv2}
\begin{aligned}
d^2J({\Omega};V,W) = &{\int_{{\Gamma}}}{\int_{{\Gamma}}}\tilde{g}(x,y) W(y) dS(y) V(x) dS(x) \\
&+ {\int_{{\Gamma}}}\left( \kappa(x) {\int_{{\Omega}}}\tilde{g}(x,y)dy
+ {n}(x)\cdot {\int_{{\Omega}}}\tilde{g}_x(x,y)dy \right) V W dS, \\
= &{\int_{{\Gamma}}}{\int_{{\Gamma}}}\tilde{g}(x,y) W(y) dS(y) V(x) dS(x) \\
&+ {\int_{{\Gamma}}}\left( \kappa(x) G(x,{\Omega})
+ {n}(x)\cdot G_x(x,{\Omega}) \right) V W dS,
\end{aligned}$$ where $\tilde{g}(x,y), \tilde{g}_x(x,y), G(x,{\Omega}),
G_x(x,{\Omega})$ are defined by , , .
The second shape derivative is computed using Theorem \[T:second-shape-deriv\]. We define $\phi(x,{\Omega})= {\int_{{\Omega}}}g(x,y) dy$, compute the derivatives $\phi'({\Omega};V), \phi''({\Omega};V,W), \partial_{n}\phi$ and substitute in formula . We have $\partial_{n}\phi(x,{\Omega}) = {n}(x)\cdot{\int_{{\Gamma}}}g_x(x,y)\, dy,$ also $$\begin{aligned}
\phi'(x) &= \phi'({\Omega};V)(x) = {\int_{{\Gamma}}}g(x,y) V(y) dS(y), \\
\phi''(x) &= \phi''({\Omega};V,W)(x)
= {\int_{{\Gamma}}}\left(\kappa g(x,y)V(y) + \partial_{n}\cdot (g(x,y)V(y))\right) W(y) dS(y) \\
&= {\int_{{\Gamma}}}\left(\kappa(y)g(x,y) + {n}(y)\cdot g_y(x,y)\right) V W dS(y).
\quad (\partial_{n}V = 0 \ \mathrm{by \ eqn }~\eqref{E:vel-assumptions})\end{aligned}$$ Then the second shape derivative is given by $$\begin{aligned}
d^2J({\Omega};V,W)=&{\int_{{\Omega}}}\phi''dx
+ {\int_{{\Gamma}}}\left( \kappa \phi + \partial_{n}\phi \right) V W dS
+ {\int_{{\Gamma}}}\left(\phi'_W V+ \phi'_V W \right)dS
\\
=&{\int_{{\Omega}}}{\int_{{\Gamma}}}\left(g(x,y)\kappa(y)
+ {n}(y)\cdot g_y(x,y)\right) V W dS(y) dx \\
&+ {\int_{{\Gamma}}}\left( \kappa(x){\int_{{\Omega}}}g(x,y)dy
+ {n}(x)\cdot{\int_{{\Omega}}}g_x(x,y)dy \right) V W dS(x) \\
&+ {\int_{{\Gamma}}}\left( {\int_{{\Gamma}}}g(x,y)W(y) dS(y) V(x)
+ {\int_{{\Gamma}}}g(x,y)V(y) dS(y) W(x)\right) dS(x).\end{aligned}$$ We change variables in the integrals, for example we let $${\int_{{\Omega}}}{\int_{{\Gamma}}}g(x,y) K(y) V(y) W(y) dS(y) dx
= {\int_{{\Gamma}}}\kappa(x) {\int_{{\Omega}}}g(y,x) dy V(x) W(x) dS(x),$$ and reorganize the terms in $d^2J({\Omega};V,W)$ and obtain $$\begin{aligned}
d^2J({\Gamma};V,W)=&{\int_{{\Gamma}}}\left(\kappa{\int_{{\Omega}}}(g(x,y)+ g(y,x))dy
+ {n}\cdot{\int_{{\Omega}}}(g_x(x,y)+ g_y(y,x))dy \right) V W dS(x) \\
&+ {\int_{{\Gamma}}}{\int_{{\Gamma}}}(g(x,y)+g(y,x))W(y) dS(y) V(x) dS(x) \\
= &{\int_{{\Gamma}}}\left(\kappa(x) {\int_{{\Omega}}}\tilde{g}(x,y) dy
+ {n}(x)\cdot{\int_{{\Omega}}}\tilde{g}_x(x,y) \right) V(x) W(x) dS(x).\end{aligned}$$ We substitute $G(x,{\Omega}), G_x(x,{\Omega})$ for the integrals of $\tilde{g}(x,y), \tilde{g}_x(x,y)$ respectively and obtain the second shape derivative.
[**Higher-Order Surface Energies.**]{} We consider the higher-order surface energy $$\label{E:bound-higher-order}
J({\Gamma}) = {\int_{{\Gamma}}}{\int_{{\Gamma}}}g(x,y) dS(y) dS(x),$$ and derive its first and second shape derivatives.
\[P:bound-higher-order-deriv\] The first shape derivative of the higher order surface energy $J({\Gamma})$ at ${\Gamma}$ with respect to velocity $V$ is $$\label{E:bound-higher-order-deriv}
\begin{aligned}
dJ({\Gamma};V)&= {\int_{{\Gamma}}}\left(\kappa(x) {\int_{{\Gamma}}}\tilde{g}(x,y) dS(y)
+ {n}(x)\cdot{\int_{{\Gamma}}}\tilde{g}_x(x,y) dS(y) \right) V(x) dS(x), \\
&= {\int_{{\Gamma}}}\left(\kappa(x) G(x,{\Gamma})
+{n}(x)\cdot G_x(x,{\Gamma}) \right) V(x) dS(x),
\end{aligned}$$ where $\tilde{g}(x,y), \tilde{g}_x(x,y)$ are defined by , and $G(x,{\Gamma}), G_x(x,{\Gamma})$ by .
We will use Theorem \[T:first-shape-deriv\] to calculate the first shape derivative of the energy . We define $\psi(x,{\Gamma})={\int_{{\Gamma}}}g(x,y) dS(y)$ (so that $J({\Gamma})={\int_{{\Gamma}}}\psi(x,{\Gamma}) dS(x)$) and calculate its derivatives of $\psi$ are $$\partial_{n}\psi = {n}(x)\cdot{\int_{{\Gamma}}}g_x(x,y) dS(y), \qquad
\psi'({\Gamma};V) = {\int_{{\Gamma}}}(g(x,y) \kappa(y)
+ g_y(x,y)\cdot{n}(y))V dS(y).$$ to be substituted in the formula for the first shape derivative: $$\begin{aligned}
dJ({\Gamma};V) =&{\int_{{\Gamma}}}\psi'({\Gamma};V) dS(x)
+ {\int_{{\Gamma}}}(\psi\kappa + \partial_{n}\psi_x) V dS(x) \\
=&{\int_{{\Gamma}}}{\int_{{\Gamma}}}\left( g(x,y) \kappa(y)
+ g_y(x,y)\cdot{n}(y) \right) V(y) dS(y) dS(x) \\
&+ {\int_{{\Gamma}}}\left( \kappa(x){\int_{{\Gamma}}}g(x,y) dS(y)
+ {n}(x) \cdot {\int_{{\Gamma}}}g_x(x,y) dS(y) \right) V(x) dS(x).\end{aligned}$$ We exchange the variables in the first integral, $$\begin{aligned}
dJ({\Gamma};V) =&{\int_{{\Gamma}}}{\int_{{\Gamma}}}\left( g(y,x) \kappa(x)
+ {n}(x) \cdot g_y(y,x) \right) V(x) dS(x) dS(y) \\
&+ {\int_{{\Gamma}}}\left( \kappa(x){\int_{{\Gamma}}}g(x,y) dS(y)
+ {n}(x) \cdot {\int_{{\Gamma}}}g_x(x,y) dS(y) \right) V(x) dS(x) \\
=&{\int_{{\Gamma}}}\left( \kappa(x){\int_{{\Gamma}}}(g(x,y)+g(y,x)) dS(y) \right.\\
&\quad \left.+ {n}(x) \cdot {\int_{{\Gamma}}}(g_x(x,y)
+ g_y(y,x)) dS(y) \right) V(x) dS(x) \\
=&{\int_{{\Gamma}}}\left( \kappa(x){\int_{{\Gamma}}}\tilde{g}(x,y) dS(y)
+ {n}(x) \cdot {\int_{{\Gamma}}}\tilde{g}_x(x,y) dS(y) \right) V(x) dS(x).\end{aligned}$$ We also replace the integrals of $\tilde{g}(x,y), \tilde{g}_x(x,y)$ by $G(x,{\Gamma}), G_x(x,{\Gamma})$ respectively.
\[P:bound-higher-order-hess\] The second shape derivative of the higher order surface energy $J({\Gamma})$ at ${\Gamma}$ with respect to velocities $V, W$ is $$\begin{aligned}
d^2J({\Gamma}; & V,W) = {\int_{{\Gamma}}}G(x,{\Gamma}){{\nabla_{{\Gamma}}}}V\cdot {{\nabla_{{\Gamma}}}}W dS(x)
\\
&+ {\int_{{\Gamma}}}\left( {n}^T G_{xx}(x,{\Gamma})\,{n}+ 2\kappa\, G_{x}(x,{\Gamma})\cdot{n}+ (\kappa^2 - \Sigma\kappa_i^2) G(x,{\Gamma}) \right) V W dS(x)
\\
&+ {\int_{{\Gamma}}}\kappa {\int_{{\Gamma}}}\tilde{g}\kappa\, W dS(y) V dS(x)
+ {\int_{{\Gamma}}}{n}^T {\int_{{\Gamma}}}\tilde{g}_{xy}{n}\, W dS(y) V dS(x)
\\
& + {\int_{{\Gamma}}}\kappa {\int_{{\Gamma}}}\tilde{g}_y \cdot{n}\, W dS(y) V dS(x)
+ {\int_{{\Gamma}}}{n}\cdot{\int_{{\Gamma}}}\tilde{g}_x\kappa\, W dS(y) V dS(x),\end{aligned}$$ where $\tilde{g}(x,y), \tilde{g}_x(x,y), \tilde{g}_y(x,y),
\tilde{g}_{xy}(x,y), G(x,{\Gamma}), G_{x}(x,{\Gamma}), G_{xx}(x,{\Gamma})$ are defined by the formulas ,, .
We define $\psi(x,{\Gamma})={\int_{{\Gamma}}}g(x,y) dS(y)$ and compute its normal derivatives $\partial_{n}\psi, \ \partial_{{n}{n}}\psi$ and its shape derivatives $\psi_V'=\psi'({\Gamma};V), \ \psi''=\psi''({\Gamma};V,W),
\ \partial_{n}\psi_V'$: $$\begin{gathered}
\partial_{n}\psi = {n}(x)\cdot{\int_{{\Gamma}}}g_x(x,y)dS(y), \quad
\partial_{{n}{n}}\psi = {n}(x)^T\left({\int_{{\Gamma}}}g_x(x,y)dS(y)\right){n}(x), \\
\psi_V' = {\int_{{\Gamma}}}(g\kappa(y) + g_y\cdot{n}(y)) V dS(y), \quad
\partial_{n}\psi_V'
= {n}\cdot{\int_{{\Gamma}}}(g_x\kappa(y) + g_{yx}^T{n}(y)) V dS(y), \\
\psi'' = {\int_{{\Gamma}}}g{{\nabla_{{\Gamma}}}}V\cdot {{\nabla_{{\Gamma}}}}W dS(y)
+ {\int_{{\Gamma}}}\left( {n}^T g_{yy} {n}+ 2\kappa g_y\cdot{n}+ (\kappa^2 - \Sigma\kappa_i^2) g \right) V W dS(y),\end{gathered}$$ which we subsitute in the general formula for the second shape derivative $$\begin{aligned}
d^2J({\Gamma};V,W)= &{\int_{{\Gamma}}}\psi'' dS
+ {\int_{{\Gamma}}}\left( \left(\partial_{n}\psi'_W + \kappa \psi'_W \right) V
+ \left(\partial_{n}\psi'_V + \kappa \psi'_V \right) W \right)dS \\
&+ {\int_{{\Gamma}}}\left( \psi {{\nabla_{{\Gamma}}}}V \cdot {{\nabla_{{\Gamma}}}}W
+ \left( \partial_{{n}{n}} \psi + 2\kappa \partial_{n}\psi
+ (\kappa^2 - \Sigma \kappa_i^2)\psi \right) V W \right) dS.\end{aligned}$$ and obtain $$\begin{aligned}
d^2J({\Gamma};V,W) = &{\int_{{\Gamma}}}{\int_{{\Gamma}}}g {{\nabla_{{\Gamma}}}}V(y) \cdot {{\nabla_{{\Gamma}}}}W(y) dS(y)dS(x)
\\
&+ {\int_{{\Gamma}}}{\int_{{\Gamma}}}\left( {n}(y)^T g_{yy}{n}(y) + 2\kappa(y) g_y\cdot{n}(y) \right. \\
& \qquad \ \left. {}+ (\kappa(y)^2 - \Sigma \kappa_i(y)^2) g \right) V(y) W(y) dS(y) dS(x)
\\
&+ {\int_{{\Gamma}}}\left( {n}(x)\cdot {\int_{{\Gamma}}}(g_x\kappa(y) + g_{yx}^T {n}(y))W(y)dS(y) \right. \\
& \qquad \ \left. {}+ \kappa(x) {\int_{{\Gamma}}}(g\kappa(y) + g_y\cdot{n}(y))W(y)dS(y) \right) V(x) dS(x)
\\
&+ {\int_{{\Gamma}}}\left( {n}(x)\cdot {\int_{{\Gamma}}}(g_x\kappa(y) + g_{yx}^T {n}(y))V(y)dS(y) \right. \\
& \qquad \ \left. {}+ \kappa(x) {\int_{{\Gamma}}}(g\kappa(y) + g_y\cdot{n}(y))V(y)dS(y) \right) W(x) dS(x)
\\
&+ {\int_{{\Gamma}}}{\int_{{\Gamma}}}g dS(y) {{\nabla_{{\Gamma}}}}V(x) \cdot {{\nabla_{{\Gamma}}}}W(x) dS(x)
\\
&+ {\int_{{\Gamma}}}\left( {n}(x)^T {\int_{{\Gamma}}}g_{xx} dS(y){n}(x) + 2\kappa(x) {n}(x)\cdot{\int_{{\Gamma}}}g_x dS(y) \right. \\
& \qquad \ \left. {}+ (\kappa(x)^2 - \Sigma \kappa_i(x)^2) {\int_{{\Gamma}}}g dS(y) \right) V(x) W(x) dS(x).\end{aligned}$$ We exchange variables $x \leftrightarrow y$ in the integrals that contain $V(y)$ and reorganize, $$\begin{aligned}
d^2J({\Gamma};V,W) = &{\int_{{\Gamma}}}\left({\int_{{\Gamma}}}(g(x,y)+g(y,x))dS(y)\right) {{\nabla_{{\Gamma}}}}V\cdot{{\nabla_{{\Gamma}}}}W dS(x)
\\
&+ {\int_{{\Gamma}}}\left\{ {n}(x)^T \left({\int_{{\Gamma}}}(g_{xx}(x,y)+g_{yy}(y,x)) dS(y)\right){n}(x) \right. \\
& \qquad \ {}+ 2\kappa(x){n}(x)\cdot \left({\int_{{\Gamma}}}(g_x(x,y)+g_y(y,x)) dS(y)\right) \\
& \qquad \ \left. {}+ (\kappa(x)^2 - \Sigma \kappa_i(x)^2) {\int_{{\Gamma}}}(g(x,y)+g(y,x)) dS(y) \right\} V W dS(x)
\\
&+ {\int_{{\Gamma}}}\kappa(x) \left({\int_{{\Gamma}}}(g(x,y)+g(y,x))\kappa(y) W(y)dS(y)\right) V(x) dS(x)
\\
&+ {\int_{{\Gamma}}}{n}(x)\cdot \left({\int_{{\Gamma}}}(g_{xy}(x,y)+g_{yx}(y,x)){n}(y) W(y)dS(y)\right) V(x) dS(x)
\\
&+ {\int_{{\Gamma}}}\kappa(x) \left({\int_{{\Gamma}}}(g_y(x,y)+g_x(y,x))\cdot{n}(y) W(y)dS(y)\right) V(x) dS(x)
\\
&+ {\int_{{\Gamma}}}{n}(x)\cdot \left({\int_{{\Gamma}}}(g_x(x,y)+g_y(y,x))\kappa(y) W(y)dS(y)\right) V(x) dS(x).\end{aligned}$$ We replace instances of $g(x,y)$ and its derivatives by $\tilde{g}(x,y)$ and its derivatives, as defined by . We also replace the integrals of $\tilde{g}, \tilde{g}_x, \tilde{g}_{xx}$ by $G(x,{\Gamma}), G_x(x,{\Gamma}), G_{xy}(x,{\Gamma})$ respectively.
Shape Energies with PDEs {#S:energies-with-pdes}
------------------------
In some problems, the shape energy may include a function $u$ that is obtained by solving a PDE on the surface ${\Gamma}$ or in the domain ${\Omega}$ enclosed by the surface ${\Gamma}$, namely, we consider energies of the form $$J({\Gamma}) = J_0({\Gamma},u({\Gamma})), \quad
\mathcal{A}_{\Gamma}(u) = f \ \mathrm{on} \ {\Gamma},
\quad \mathrm{or} \quad
J({\Omega}) = J_0({\Omega},u({\Omega})), \quad
\mathcal{A}_{\Omega}(u) = f \ \mathrm{in} \ {\Omega},$$ where $\mathcal{A}_{\Gamma},\mathcal{A}_{\Omega}$ denote some differential operators. The PDEs represented by $\mathcal{A}_{\Gamma},\mathcal{A}_{\Omega}$ might be in various forms and do not seem to be of interest in image processing except for a few specific cases relating to the Mumford-Shah functional [@Brox-Cremers-09; @Chan-Vese-01; @Dogan-Morin-Nochetto-08; @Hintermueller-Ring-04; @Jin-Yezzi-Soatto-03] (see Section \[S:shape-energies\]). We will consider these cases below. For other examples of shape energies with PDEs in areas outside image processing, we refer to the books [@Delfour-Zolesio-01; @Haslinger-Makinen-03; @Mohammadi-Pironneau-01; @Pironneau-84].
[**Energies with domain PDEs.**]{} We consider the generalization of the Mumford-Shah energy introduced in Section \[S:shape-energies\] and compute the first and second shape derivatives of the PDE-dependent part, namely the following domain energy, $$\label{E:gen-MS-energy}
J({\Omega}) = {\int_{{\Omega}}}f(x,\{u_k\}) dx +
\frac{\mu}{2} \sum_{k=1}^m {\int_{{\Omega}}}|\nabla u_k|^2 dx,$$ where the smooth approximation functions $\{u_k\}_{k=1}^m$ are computed from the PDEs $$\label{E:gen-MS-pde}
- \Delta u_l + f_{u_l}(x,\{u_k\}) = 0 \ \mathrm{in} \ {\Omega}, \quad
\frac{\partial u_l}{\partial {n}} = 0 \ \mathrm{on} \ {\Gamma}=\partial{\Omega},
\quad l=1,\ldots,m.$$ In , $f_{u_l}$ denotes the derivative of the coupled data function $f(x,\{u_k\})$ with respect to the argument $u_l$ and we assume $f$ is given such that unique solutions of and exist in $H^1({\Omega})$. The shape derivations for follow those of Hintermüller and Ring [@Hintermueller-Ring-04] for the Mumford-Shah energy [@Chan-Vese-01].
\[P:deriv-MS\] The first shape derivative of the energy at ${\Omega}$ with respect to velocity $V$ is $$dJ({\Omega};V)= {\int_{{\Gamma}}}\left(f(x,\{u_k\})
+ \frac{\mu}{2} \sum_k |{{\nabla_{{\Gamma}}}}u_k|^2 \right) V dS.$$ The second shape derivative of with respect to velocities $V, W$ is $$\label{E:gen-MS-hess}
\begin{aligned}
d^2J({\Omega};V,W) =
&{\int_{{\Gamma}}}\left(f\kappa + \frac{\partial f}{\partial{n}}
+ \mu\sum_k {{\nabla_{{\Gamma}}}}u_k^T \left( \frac{\kappa}{2}Id
- {{\nabla_{{\Gamma}}}}{n}\right) {{\nabla_{{\Gamma}}}}u_k \right) V W dS \\
& +\sum_k {\int_{{\Gamma}}}\left( f_{u_k} u_{k,W}'
+ \mu {{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}u_{k,W}' \right) V dS,
\end{aligned}$$ where $u_{k,W}' = u_k'({\Omega};W)$ is the shape derivative of $u_k$ at ${\Omega}$ with respect to $W$ computed from the PDEs $$\label{E:MS-u-deriv-pde}
-\mu \Delta u_k' + \sum_l f_{u_k u_l} u_{l,W}'
= 0 \ \mathrm{in} \ {\Omega}, \quad
\frac{\partial u_{k,W}'}{\partial {n}}
= {{\rm{div}_{{\Gamma}}}}(V{{\nabla_{{\Gamma}}}}u_k) - \frac{1}{\mu} f_{u_k} V \ \mathrm{on} \ {\Gamma}.$$
We use Theorem \[T:first-shape-deriv\] and let $\phi(x,{\Omega}) = f(x,\{u_k\}) + \frac{\mu}{2}\sum_k|\nabla u_k|^2$ so that $$\phi'({\Omega};V) = \sum_k f_{u_k}(x,\{u_l\}) u_k'
+ \mu \sum_k \nabla u_k \cdot \nabla u_k',$$ where $u_k' = u_k'({\Omega};V)$. From Theorem \[T:first-shape-deriv\] it follows that $$\label{E:part-MS-intermediate}
dJ({\Omega};V) = {\int_{{\Omega}}}\sum_k \left(f_{u_k} u_k'
+ \mu \nabla u_k \cdot \nabla u_k' \right) dx
+ {\int_{{\Gamma}}}\left( f + \frac{\mu}{2} \sum_k |{{\nabla_{{\Gamma}}}}u_k|^2 \right) V dS.$$ Note that $\nabla u_k = {{\nabla_{{\Gamma}}}}u_k$ on ${\Gamma}$ since $\frac{\partial u_k}{\partial{n}} = 0$ on ${\Gamma}$.\
Now we investigate $u_k'({\Omega};V)$. Consider the weak form of $$\label{E:MS-pde-weak-form}
\int_{{\Omega}} (\mu \nabla u_l \cdot \nabla \varphi + f_{u_l} \varphi) dx = 0,
\qquad \forall \varphi \in H^1({\Omega})$$ and take the shape derivative ($\varphi$ is shape-independent) $$\int_{{\Omega}} \left(\mu \nabla u_l' \cdot \nabla \varphi
+ \sum_k f_{u_l u_k}u_k' \varphi \right) dx
+ {\int_{{\Gamma}}}(\mu \nabla u_l \cdot\nabla \varphi + f_{u_l} \varphi) V dS = 0.$$ Again recall that $\frac{\partial u_l}{\partial{n}}=0$ and substitute $\nabla u|_{\Gamma}= {{\nabla_{{\Gamma}}}}u$, $$\label{E:MS-u-deriv-weak-form0}
{\int_{{\Omega}}}\left(\mu \nabla u_l' \cdot \nabla \varphi
+ \sum_k f_{u_l u_k}u_k' \varphi \right) dx =
-{\int_{{\Gamma}}}(\mu {{\nabla_{{\Gamma}}}}u_l {{\nabla_{{\Gamma}}}}\varphi + f_{u_l} \varphi) V dS.$$ This equation has a unique solution $u_l' \in
H^1({\Omega})$ and we use it as a test function in . In this way we see that the first integral in vanishes, which leaves us with the expression for the first shape derivative.\
We also write the strong form of the PDE for the shape derivative $u_l'$. For this, we first integrate the left hand side of by parts with tangential Green’s formula $${\int_{{\Omega}}}\left(\mu \nabla u_l' \cdot \nabla \varphi
+ \sum_k f_{u_l u_k}u_k' \varphi \right) dx =
{\int_{{\Gamma}}}\Big(\mu {{\rm{div}_{{\Gamma}}}}(V {{\nabla_{{\Gamma}}}}u_l) - f_{u_l} V \Big) \varphi dS.$$ Then we have the following PDE for $u_l'$ $$-\mu \Delta u_l' + \sum_k f_{u_l u_k} u_k'
= 0 \ \mathrm{in} \ {\Omega}, \quad
\frac{\partial u_l'}{\partial {n}}
= {{\rm{div}_{{\Gamma}}}}(V{{\nabla_{{\Gamma}}}}u_l) - \frac{1}{\mu} f_{u_l} V \ \mathrm{on} \ {\Gamma}.$$ To compute the second shape derivative, we now let $ \psi(x,{\Gamma}) = \big( f(x,\{u_k\}) +
\frac{\mu}{2}\sum_k |{{\nabla_{{\Gamma}}}}u_k|^2 \big) V$ and apply Theorem \[T:first-shape-deriv\]. We compute the derivatives $\frac{\partial \psi}{\partial{n}}$ and $\psi'({\Gamma};W)$. $$\frac{\partial \psi}{\partial{n}} =
\left( \frac{\partial f}{\partial{n}}
+ \sum_k f_{u_k} \frac{\partial u_k}{\partial{n}}
+ \frac{\mu}{2} \sum_k
\frac{\partial}{\partial{n}} |{{\nabla_{{\Gamma}}}}u_k|^2 \right) V
+ \Big( \ldots \Big) \frac{\partial V}{\partial{n}}.$$ Recall that $\frac{\partial u_k}{\partial{n}} = 0$ on ${\Gamma}$ and $\frac{\partial V}{\partial{n}} = 0$ by the assumptions . Also we have $\frac{\partial}{\partial{n}}({{\nabla_{{\Gamma}}}}u)
= -{{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}u$ by Lemma \[L:deriv-surf-f-u\]. Therefore $$\frac{\partial\psi}{\partial{n}} =
\left( \frac{\partial f}{\partial{n}}
- \mu\sum_k {{\nabla_{{\Gamma}}}}u_k^T {{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}u_k \right) V.$$ We compute the shape derivative $\psi' = \psi'({\Gamma};W)$, $$\psi' = \left( \sum_k f_{u_k} u_{k,W}'
+ \mu \sum_k{{\nabla_{{\Gamma}}}}u_k\cdot ({{\nabla_{{\Gamma}}}}u_k)_W'\right) V
+ \Big(\ldots \Big) V'.$$ By Lemma \[L:deriv-geom\] and $\frac{\partial u_k}{\partial{n}} = 0$, we have $({{\nabla_{{\Gamma}}}}u_k)' = {{\nabla_{{\Gamma}}}}u_{k,W}' + {{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}W{n}$ and one can trivially see $V' = ({{\vec V}}\cdot{n})' = {{\vec V}}\cdot(-{{\nabla_{{\Gamma}}}}W)
= - V{n}\cdot{{\nabla_{{\Gamma}}}}W = 0$. Therefore $$\begin{aligned}
\psi' &= \left( \sum_k f_{u_k} u_{k,W}'
+ \mu \sum_k \left({{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}u_{k,W}'
+ {{\nabla_{{\Gamma}}}}u_k\cdot{n}{{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}W\right) \right) V \\
& = \sum_k \left(f_{u_k} u_{k,W}' + \mu{{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}u_{k,W}'\right) V.
\qquad ({{\nabla_{{\Gamma}}}}u_k\cdot{n}= 0)\end{aligned}$$ The second shape deriative is obtained by plugging $\psi, \frac{\partial\psi}{\partial{n}}, \psi'$ in $$d^2J({\Omega};V,W) = {\int_{{\Gamma}}}\psi'({\Gamma};W) dS
+ {\int_{{\Gamma}}}\left( \psi\kappa + \frac{\partial\psi}{\partial{n}} \right) W dS.$$
[**Energies with Surface PDEs.**]{} In this section we derive the first shape derivative of the PDE-dependent surface energy $$\label{E:part-surf-MS-energy}
J({\Gamma}) = \frac{1}{2} \int_{\Gamma}\left( f(x,\{u_k\})
+ \frac{\mu}{2} \sum_k |{{\nabla_{{\Gamma}}}}u_k|^2 \right) dS,$$ where the smooth surface functions $\{u_k\}_{k=1}^m$ are computed from the PDE $$\label{E:surf-MS-pde}
-\mu {\Delta_{{\Gamma}}}u_l + f_{u_l}(x,\{u_k\}) = 0 \ \textrm{on} \ {\Gamma}, \quad
l=1,\ldots,m.$$ In , $f_{u_l}$ denotes the derivative of the coupled data function $f(x,\{u_k\})$ with respect to the argument $u_l$ and we assume $f$ is given such that unique solutions of the PDEs , exist in $H^1({\Gamma})$. The shape energy can be used for shape identification problems, in which smooth approximations $\{u_k\}$ of data channels or descriptors on the surface ${\Gamma}$ need to be estimated in addition to the surface ${\Gamma}$ itself (e.g. sterescopic segmentation [@Jin-Yezzi-Soatto-03]).
The first shape derivative of the energy at ${\Gamma}$ with respect to velocity $V$ is given by $$dJ({\Gamma};V) = {\int_{{\Gamma}}}\left(f \kappa
+ \frac{\partial f}{\partial{n}}
+ \mu \sum_k {{\nabla_{{\Gamma}}}}u_k^T \left(\frac{\kappa}{2}Id
- {{\nabla_{{\Gamma}}}}{n}\right){{\nabla_{{\Gamma}}}}u_k \right) V dS.$$
We take the first shape derivative using Theorem \[T:first-shape-deriv\]: $$\label{E:part-surf-MS-deriv-1}
\begin{aligned}
dJ({\Gamma};V) =& {\int_{{\Gamma}}}\left(\sum_k f_{u_k} u_k'
+ \mu\sum_k {{\nabla_{{\Gamma}}}}u_k\cdot({{\nabla_{{\Gamma}}}}u_k)'\right)dS \\
&+ {\int_{{\Gamma}}}\kappa\left(f(x,\{u_k\})
+ \frac{\mu}{2}\sum_k|{{\nabla_{{\Gamma}}}}u_k|^2\right) V dS \\
&+ {\int_{{\Gamma}}}\frac{\partial}{\partial n}\left( f(x,\{u_k\})
+ \frac{\mu}{2}\sum_k|{{\nabla_{{\Gamma}}}}u_k|^2\right) V dS.
\end{aligned}$$ Meaningful interpretation of the expression requires the functions $\{u_k\}$ to be defined off the surface ${\Gamma}$, because we need to be able to compute their full spatial gradient and the normal derivatives. But $\{u_k\}$ are computed with the PDE and are defined only on ${\Gamma}$. To be able to proceed with the derivations, we work with smooth extensions $\{\tilde{u}_k\}$ of $\{u_k\}$ in a tubular neighborhood $U$ of the surface ${\Gamma}$. We define the extension $\tilde{u}_k$ such that it is constant in the normal direction, i.e. $\frac{\partial \tilde{u}_k}{\partial{n}} = 0$. To keep notation simple, we will continue to refer to the extended function as $u_k$.
Using $\frac{\partial u_k}{\partial{n}} = 0$ and identity , we find $$\frac{1}{2}\frac{\partial}{\partial{n}} |{{\nabla_{{\Gamma}}}}u_k|^2
= -{{\nabla_{{\Gamma}}}}u_k^T{{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}u_k,
\qquad
\frac{\partial}{\partial n} \left( f(x,\{u_k\}) \right)
= \frac{\partial f}{\partial n}.$$ Then using equation and noting ${{\nabla_{{\Gamma}}}}u_k\cdot{n}=0$ and $\frac{\partial u_k}{\partial n} = 0$, we write $${{\nabla_{{\Gamma}}}}u_k\cdot({{\nabla_{{\Gamma}}}}u_k)' = {{\nabla_{{\Gamma}}}}u_k \cdot
\left({{\nabla_{{\Gamma}}}}u_k' + {{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}V {n}+ \frac{\partial u_k}{\partial n} {{\nabla_{{\Gamma}}}}V \right)
= {{\nabla_{{\Gamma}}}}u_k \cdot {{\nabla_{{\Gamma}}}}u_k'.$$ Now we can rewrite the shape derivative $$\label{E:part-surf-MS-deriv-2}
\begin{aligned}
dJ({\Gamma};V) =& {\int_{{\Gamma}}}\left(\sum_k f_{u_k} u_k'
+ \mu\sum_k{{\nabla_{{\Gamma}}}}u_k\cdot{{\nabla_{{\Gamma}}}}u_k'\right)dS \\
&+ {\int_{{\Gamma}}}\kappa\left(f + \frac{\mu}{2}\sum_k |{{\nabla_{{\Gamma}}}}u_k|^2\right) V dS
+ {\int_{{\Gamma}}}\left(\frac{\partial f}{\partial n}
- \mu{{\nabla_{{\Gamma}}}}u^T {{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}u\right) V dS,
\end{aligned}$$ in which the terms containing the shape derivatives $u'$ will vanish as we will see below. To show this, we start with the weak form of the surface PDE $$\label{E:surf-MS-pde-weak-form}
{\int_{{\Gamma}}}\left(\mu{{\nabla_{{\Gamma}}}}u_l\cdot{{\nabla_{{\Gamma}}}}\varphi + f_{u_l}\varphi \right) dS
= 0 , \qquad \forall \varphi \in H^1({\Gamma}),$$ also using normal extensions $\tilde{\varphi}$ of the test functions $\varphi$ with $\frac{\partial \tilde{\varphi}}{\partial n} = 0$ (which we continue to refer to as $\varphi$). We differentiate the two terms in . Start with the second term, $$\begin{aligned}
\left({\int_{{\Gamma}}}f_{u_l} \varphi dS\right)'
= {\int_{{\Gamma}}}\sum_k f_{u_l u_k} u_k' \varphi dS
+ {\int_{{\Gamma}}}\left( \kappa f_{u_l} +
\frac{\partial f_{u_l}}{\partial{n}} \right) \varphi V dS.\end{aligned}$$ Then the first term, $$\label{E:part-surf-MS-pde-deriv-1}
\begin{aligned}
\left({\int_{{\Gamma}}}\mu{{\nabla_{{\Gamma}}}}u_l\cdot{{\nabla_{{\Gamma}}}}\varphi dS\right)'
&= {\int_{{\Gamma}}}\left(\mu({{\nabla_{{\Gamma}}}}u_l)'\cdot{{\nabla_{{\Gamma}}}}\varphi
+ \mu{{\nabla_{{\Gamma}}}}u_l\cdot({{\nabla_{{\Gamma}}}}\varphi)' \right)dS \\
& + {\int_{{\Gamma}}}\kappa \mu{{\nabla_{{\Gamma}}}}u_l\cdot{{\nabla_{{\Gamma}}}}\varphi V dS
+ {\int_{{\Gamma}}}\frac{\partial}{\partial{n}}
(\mu{{\nabla_{{\Gamma}}}}u_l\cdot{{\nabla_{{\Gamma}}}}\varphi) V dS.
\end{aligned}$$ We use Lemmas \[L:deriv-surf-f-u\], \[L:deriv-geom\] to rewrite the following terms $$\begin{aligned}
({{\nabla_{{\Gamma}}}}u_l)'\cdot{{\nabla_{{\Gamma}}}}\varphi
&= {{\nabla_{{\Gamma}}}}u_l'\cdot{{\nabla_{{\Gamma}}}}\varphi
+ {{\nabla_{{\Gamma}}}}u_l\cdot{{\nabla_{{\Gamma}}}}V {{\nabla_{{\Gamma}}}}\varphi\cdot{n}+ \frac{\partial u_l}{\partial{n}}
{{\nabla_{{\Gamma}}}}\varphi\cdot{{\nabla_{{\Gamma}}}}V
= {{\nabla_{{\Gamma}}}}u_l'\cdot{{\nabla_{{\Gamma}}}}\varphi,
\\
{{\nabla_{{\Gamma}}}}u_l\cdot({{\nabla_{{\Gamma}}}}\varphi)'
&= {{\nabla_{{\Gamma}}}}u_l\cdot {{\nabla_{{\Gamma}}}}\varphi' = 0,
\qquad\qquad (\mathrm{note:} \ {{\nabla_{{\Gamma}}}}\varphi\cdot{n}=0, \
\frac{\partial u_l}{\partial{n}} = 0)
\\
\frac{\partial}{\partial{n}} ({{\nabla_{{\Gamma}}}}u_l\cdot{{\nabla_{{\Gamma}}}}\varphi)
&= -{{\nabla_{{\Gamma}}}}\varphi^T {{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}u_l
-{{\nabla_{{\Gamma}}}}u_l^T {{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}\varphi
= -2 {{\nabla_{{\Gamma}}}}u_l^T {{\nabla_{{\Gamma}}}}{n}{{\nabla_{{\Gamma}}}}\varphi,\end{aligned}$$ and substitute back in . Then the shape derivative of weak form is $$\label{E:surf-MS-u-deriv}
\begin{aligned}
{\int_{{\Gamma}}}\left(\mu{{\nabla_{{\Gamma}}}}u_l'\cdot{{\nabla_{{\Gamma}}}}\varphi \right. &
+ \left. \sum_k f_{u_l u_k} u_k' \varphi\right) dS \\
&= {\int_{{\Gamma}}}\left( \kappa f_{u_l}\varphi
+ \frac{\partial f_{u_l}}{\partial{n}}\varphi
+\mu {{\nabla_{{\Gamma}}}}u_l^T (\kappa - 2{{\nabla_{{\Gamma}}}}{n}) {{\nabla_{{\Gamma}}}}\varphi \right) V dS.
\end{aligned}$$ The function $f$ is given such that these coupled PDEs have a unique solution $\{u_l'\}$ in $H^1({\Omega})$. We plug in the solutions $\{u_l'\}$ as test functions in and find that $${\int_{{\Gamma}}}\left(\mu{{\nabla_{{\Gamma}}}}u_l\cdot{{\nabla_{{\Gamma}}}}u_l' + f_{u_l} u_l' \right) dS = 0.$$ This removes the first term in the shape derivative .
Gradient Descent Flows {#S:grad-descent}
======================
The main motivation for deriving the shape derivatives $dJ({\Gamma};V), d^2J({\Gamma};V,W)$ of a given shape energy $J({\Gamma})$ is to design algorithms for minimization of the energy $J({\Gamma})$ and for computing the optimal shape ${\Gamma}^*$. In this section, we briefly review how to develop gradient descent flows, namely energy-decreasing evolutions of the shapes, using shape derivatives for this purpose. We refer to [@Almgren-Taylor-Wang-93; @Almgren-Taylor-95; @Ambrosio-95; @Ambrosio-Gigli-Savare-05; @Dogan-etal-07; @Taylor-Cahn-94] for more information on this topic.
We note, in Theorem \[T:hadamard\], that the shape derivative has the following form $$dJ({\Gamma};V) = {\int_{{\Gamma}}}G({\Gamma}) V dS,$$ where $G({\Gamma})$ (or alternatively $G({\Omega})$) is the shape gradient depending on the shape energy. It is easy to see that, formally by setting $V=-G({\Gamma})$, we obtain a gradient descent velocity $$dJ({\Gamma};V) = -{\int_{{\Gamma}}}G^2 dS \leqslant 0.$$ The velocity $V=-G({\Gamma})$ is the most commonly used gradient descent velocity for shape optimization problems in image processing. Given a method to compute the gradient descent velocity $V$, we can now perform the minimization by starting from an initial surface ${\Gamma}_0$ and updating it iteratively, recomputing the velocity for the new shape ${\Gamma}_{k+1}$ at each step: $$\label{E:surface-update}
{\vec{X}}_{k+1} = {\vec{X}}_k + \tau_k {{\vec V}}_k, \qquad
\forall {\vec{X}}\in {\Gamma}_k,$$ where $\tau_k>0$ is a step size parameter that can be fixed or chosen by a line search algorithm. The vector velocity ${{\vec V}}$ can be computed from the normal velocity $V$; for a surface ${\Gamma}$ with normal ${n}$, a natural choice is ${{\vec V}}=V{n}$ as the tangential component of the velocity ${{\vec V}}$ does not change the shape of the surface. An alternative to the explicit update is to embed the surface in a Eulerian representation, such as a level set function $\varphi$, extend the velocity $V$ off the surface, and compute the level set evolution solving the following PDE: $\frac{\partial\varphi}{\partial t} = V |\nabla\varphi|$ [@Osher-Fedkiw-03; @Osher-Sethian-88; @Sethian-99].
Other gradient descent velocities than $V=-G({\Gamma})$ are possible [@Ambrosio-Gigli-Savare-05; @Charpiat-etal-07; @Dogan-etal-07; @Sundar-Yezzi-Mennucci-07; @Sundar-etal-09; @Taylor-Cahn-94]. We can introduce a scalar product $b(\cdot,\cdot)$ associated with a Hilbert space $B({\Gamma})$ on the surface ${\Gamma}$ and use it to compute a different gradient flow by solving the following equation $$\label{E:general-vel}
b(V,W) = -{\int_{{\Gamma}}}G({\Gamma}) W dS, \qquad \forall W \in B({\Gamma}).$$ It is easy to verify that the solution $V$ of equation is a gradient descent velocity; we substitute it in the shape derivative and see that $dJ({\Gamma};V) = -b(V,V) \leqslant 0$ (as the scalar product $b(\cdot,\cdot)$ is positive definite).
The velocity $V=-G({\Gamma})$ mentioned above is actually the $L^2$ gradient descent velocity obtained by setting the scalar product equal to the $L^2$ scalar product, $b(V,W) = {\int_{{\Gamma}}}V W dS$, in . We can take advantage of other scalar products $b(\cdot,\cdot)$ in order to obtain velocities that improve the descent process in various ways [@Charpiat-etal-07; @Dogan-etal-07; @Sundar-Yezzi-Mennucci-07; @Sundar-etal-09]. For example, an $H^1$ scalar product, $$b(V,W) = {\int_{{\Gamma}}}\alpha(x) {{\nabla_{{\Gamma}}}}V\cdot{{\nabla_{{\Gamma}}}}W + \beta(x)V W dS,
\quad (\alpha(x),\beta(x)>0),$$ results in smoother velocities that are advantageous in applications of segmentation and tracking [@Sundar-etal-06; @Sundar-etal-09].
Another option is to use the second shape derivative as the basis of the scalar product, for example, set $$\label{E:Newton-vel}
b(V,W) = d^2J({\Gamma};V,W),$$ This choice results in a Newton’s method for shape optimization. It can yield quadratic convergence in the neighborhood of the solution. However, direct use of the scalar product may not always be possible, because the second shape derivative $d^2J({\Gamma};{{\vec V}},{{\vec W}})$ may not always satisfy the properties of a scalar product, for example, it may not be positive definite. In this case, one can still design a scalar product $b(\cdot,\cdot)$ based on the second shape derivative and retain partially the favorable convergence properties. This was pursued successfully in [@Dogan-Morin-Nochetto-08; @Hintermueller-Ring-03; @Hintermueller-Ring-04] and used to achieve a significant reduction in the number iterations needed for convergence to the optimal shape.
[^1]: ^1^ Theiss Research, La Jolla, CA 92037, USA, and Applied and Computational Mathematics Division, Information Technology Laboratory, National Institute of Standards and Technology, Gaithersburg, MD 20899, USA ([[email protected]]{}). Partially supported by NSF Grant DMS-0505454 and NIST Grant 70NANB10H046.
[^2]: Although the Snakes model of Kass, Witkin, Terzopoulos [@Kass-etal-88] can also be viewed as a first example, it is not truly a shape energy, because the value of the energy depends on the parametrization of the curve and it can be different for different parametrizations even though the shape is the same.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'One of the most fundamental problems in optomechanical cooling is how small the thermal phonon number of a mechanical oscillator can be achieved under the radiation pressure of a proper cavity field. Different from previous theoretical predictions, which were based on an optomechanical system’s time-independent steady states, we treat such cooling as a dynamical process of driving the mechanical oscillator from its initial thermal state, due to its thermal equilibrium with the environment, to a stabilized quantum state of higher purity. We find that the stabilized thermal phonon number left in the end actually depends on how fast the cooling process could be. The cooling speed is decided by an effective optomechanical coupling intensity, which constitutes an essential parameter for cooling, in addition to the sideband resolution parameter that has been considered in other theoretical studies. The limiting thermal phonon number that any cooling process cannot surpass exhibits a discontinuous jump across a certain value of the parameter.'
author:
- Bing He
- Liu Yang
- Qing Lin
- Min Xiao
title: Radiation Pressure Cooling as a Quantum Dynamical Process
---
Preparing the approximate pure quantum states of a sizable mechanical oscillator is a feasible way toward macroscopic quantumness. Practically starting from its thermal equilibrium with the environment, such process is implemented by coupling the oscillator to a cavity field generated by a red-detuned external drive, to reduce the associated thermal phonon number to a low level, similar to cooling the oscillator to a lower temperature. An important feature we will illustrate is that the cooling result depends on how fast the optomechanical system (OMS) evolves to the finally stable quantum state.
So far numerous experiments have realized the cooling to a few and even less than one mechanical quanta [@ex1; @ex2; @ex3; @ex4; @ex5; @ex6; @ex7; @ex8; @ex9; @ex10; @ex11; @ex12; @ex13]. Following the earlier study of quantum fluctuations under radiation pressure [@00; @01], the theoretical description of such optomechanical cooling (see, e.g. [@c1; @c2; @c0; @ck; @c3; @c4; @c5; @c6]) was based on a linearization procedure as that described in [@rev]; that is to decompose the cavity field mode $\hat{a}$ into the sum of the classical mean value $\alpha$ and its quantum fluctuation $\delta \hat{a}$. The linearized Hamiltonian gives the cooling action as a beamsplitter (BS) type coupling between the mechanical mode $\hat{b}$ and the fluctuation $\delta \hat{a}$ with their coupling intensity $g$ magnified by $\alpha$, which was generally treated as a constant of steady-state value. In an actual cooling process, however, the cavity mean field $\langle \hat{a}(t)\rangle=\alpha(t)$ is built up from zero (when the mechanical oscillator is in thermal equilibrium with its environment) and takes time to evolve to stable value. Then the effective coupling strength $g|\alpha|$ used in the previous studies should be more appropriately taken as a variable, since $\alpha(t)$ keeps changing during a cooling process. Due to the impossibility of finding the time-dependent $\alpha(t)$ analytically, it is difficult to study the cooling as a dynamical process if adopting the above-mentioned linearization.
In the present work we put forward a quantum dynamical theory for optomechanical cooling. Using this completely quantum approach that linearizes a weakly coupled OMS’s dynamics without resorting to its classical mean values such as $\alpha(t)$, one can numerically predict the involved physical quantities that evolve with time. The residual occupation of the thermal excitation of mechanical oscillator, as found in our approach, is connected with the whole cooling process. Compared with the previous theoretical predictions, this dynamical approach provides richer information about how good a quantum OMS can be fully cooled down to.
Fig. 1(a) illustrates an exemplary optomechanical cooling setup. In a frame with the system modes rotating at the cavity frequency $\omega_c$ and the mechanical frequency $\omega_m$, respectively, the process is governed by the following Hamiltonians [@supp]: (1) $H_e(t)=iE(\hat{a}^{\dagger}e^{i\Delta t}-\hat{a}e^{-i\Delta t})$ for an external drive with the intensity $E$ and the detuning $\Delta=\omega_c-\omega_l$ of its frequency $\omega_l$; (2) $H_{om}(t)=-g\hat{a}^\dagger\hat{a}(\hat{b}e^{-i\omega_m t}+\hat{b}^\dagger e^{i\omega_m t})$ for the optomechanical coupling; (3) the stochastic Hamiltonian $H_{sr}(t)=i\sqrt{2\kappa }\{\hat{a}^{\dagger }\hat{\xi}
_{c}(t)-\hat{a}\hat{\xi}^{\dagger }
_{c}(t)\}+i\sqrt{2\gamma_m }\{\hat{b}^{\dagger }\hat{\xi}
_{m}(t)-\hat{b}\hat{\xi}^{\dagger }
_{m}(t)\}$ accounting for the cavity (mechanical) damping at the rate $\kappa$ ($\gamma_m$), with the correlations of the quantum noise operators satisfying [@rev] $$\begin{aligned}
&&\langle \hat{\xi}^\dagger_{c}(t)\hat{\xi}_{c}(t')\rangle=0, ~~\langle \hat{\xi}^\dagger_{m}(t)\hat{\xi}_{m}(t')\rangle=n_{th}\delta(t-t'),
\label{noise}\end{aligned}$$ where a zero thermal occupation for the cavity reservoir is assumed and $n_{th}$ is the thermal occupation of the mechanical reservoir. Then we take an interaction picture with respect to $H_e(t)$. The transformed Hamiltonian $H^{in}(t)=U_0^\dagger(t)(H_{om}+H_{sr})U_0(t)$, where $U_0(t)={\cal T}e^{-i\int_0^t d\tau H_e(\tau)}$, gives the dynamical equations $$\begin{aligned}
\dot{\hat{a}}&=&-\kappa\hat{a}+g E f(t)e^{-i\omega_mt} \hat{b}+g Ef(t)e^{i\omega_mt} \hat{b}^\dagger\nonumber\\
&+&i\kappa f(t) E+\sqrt{2\kappa}\hat{\xi}_c(t),\\
\dot{\hat{b}}&=&-\gamma_m\hat{b}-g Ef^{\ast}(t)e^{i\omega_mt}\hat{a}+g E f(t)e^{i\omega_mt} \hat{a}^\dagger\nonumber\\
&+&ig|f(t)E|^2e^{i\omega_m t}+\sqrt{2\gamma_m}\hat{\xi}_m(t),\end{aligned}$$ with $f(t)=(e^{i\Delta t}-1)/\Delta$. A cubic term in $H^{in}(t)$ is neglected in deriving these linear dynamical equations, since we are dealing with a weakly coupled OMS [@supp].
We first look at the mechanical oscillator’s thermal equilibrium with the environment. In this initially prepared state with no external drive ($E=0$), the mechanical mode takes the exact form $\hat{b}(t)=e^{-\gamma_m t}\hat{b}+\sqrt{2\gamma_m}\int_0^t d\tau e^{-\gamma_m(t-\tau)}\hat{\xi}_{m}(\tau)$ from Eq. (3). It is the second noise drive term that maintains the invariant phonon number $\langle \hat{b}^\dagger\hat{b}\rangle=n_{th}$ under the thermal equilibrium, while the contribution from the first term lowers with time. The noise actions are thus essential to a quantum OMS.
In a general situation it is clearer to use the equation $$\begin{aligned}
\frac{d}{dt}\vec{\hat{c}}(t)=\hat{M}(t)\vec{\hat{c}}(t)+\vec{\lambda}(t)+\vec{\hat{\eta}}(t)
\label{d-equation}\end{aligned}$$ about the complete set $\vec{\hat{c}}(t)=(\hat{a}(t),\hat{a}^\dagger(t),\hat{b}(t),\hat{b}^\dagger(t))^T$ of the system modes, where $\vec{\lambda}=(\lambda_c,\lambda_c^\ast, \lambda_m, \lambda_m^\ast)^T$ with $\lambda_c(t)=i\kappa f(t)E$ and $\lambda_m(t)=ig|f(t)E|^2e^{i\omega_mt}$, $\vec{\hat{\eta}}=(\hat{\eta}_c, \hat{\eta}_c^\dagger, \hat{\eta}_m,\hat{\eta}_m^\dagger)^T$ with $\hat{\eta}_c(t)=\sqrt{2\kappa}\hat{\xi}_c(t)$ and $\hat{\eta}_m(t)=\sqrt{2\gamma_m}\hat{\xi}_m(t)$, and the detailed matrix $\hat{M}(t)$ can be found in [@supp]. The general solution to Eq. (\[d-equation\]) reads $$\begin{aligned}
\vec{\hat{c}}(t)&=&{\cal T}e^{\int_0^t d\tau \hat{M}(\tau)}\vec{\hat{c}}\nonumber\\
&+&\int_0^t d\tau~ {\cal T}e^{\int_\tau^t dt' \hat{M}(t')}\big(\vec{\lambda}(\tau)+\vec{\hat{\eta}}(\tau)\big),
\label{solution}\end{aligned}$$ with $\vec{\hat{c}}=(\hat{a},\hat{a}^\dagger,\hat{b},\hat{b}^\dagger)^T$. The time-ordered exponentials appear because we have $[\hat{M}(t),\hat{M}(t')]\neq 0$ for $t\neq t'$. In terms of the notation $d_{ij}(t,\tau)$ for the matrix elements $[{\cal T}e^{\int_\tau^t dt' \hat{M}(t')}]_{ij}$, the evolving thermal phonon number $n_m(t)=\langle \hat{b}^\dagger\hat{b}(t)\rangle-|\langle \hat{b}(t)\rangle|^2$ (the subtraction of the coherent drive $\vec{\lambda}(t)$’s contribution is to be discussed below) consists of two parts, $$\begin{aligned}
n^{(s)}_m(t)&=&|d_{41}(t,0)|^2+|d_{44}(t,0)|^2n_{th}\nonumber\\
&+&|d_{43}(t,0)|^2(n_{th}+1)
\label{s}\end{aligned}$$ from taking the expectation value of the first term on the right side of Eq. (\[solution\]) with respect to the initial cavity vacuum state and mechanical thermal state, and $$\begin{aligned}
n^{(n)}_m(t)&=&2\kappa\int_0^{t} d\tau |d_{41}(t,\tau)|^2+2\gamma_m \int_0^{t} d\tau |d_{44}(t,\tau)|^2n_{th}\nonumber\\
&+& 2\gamma_m \int_0^{t} d\tau |d_{43}(t,\tau)|^2(n_{th}+1)
\label{n}\end{aligned}$$ from averaging the following noise drive term over the reservoir states by means of Eq. (\[noise\]). A meaningful scenario beyond thermal equilibrium is cooling—the noise contribution in Eq. (\[n\]) finally stabilizes to a thermal phonon number $n_{m,f}$ less than the initial occupation $n_{th}$, while the contribution $n_m^{(s)}(t)$ from the evolved system operators gradually tends to zero with time.
Going back to the specific terms in Eqs. (2)-(3), one finds that an increased magnitude of the third term on their right sides (from a squeezing type coupling) can enhance the phonon numbers in both Eq. (\[s\]) and Eq. (\[n\]). The cooling action, on the other hand, manifests as the second term of a BS coupling in these equations. To let the BS action dominate, one could set the detuning $\Delta$ to be the mechanical frequency $\omega_m$, so that the factor $f(t)e^{-i\omega_mt}$ in Eq. (2) will become $\Gamma_b(t)/\omega_m=(1-e^{-i\omega_m t})/\omega_m$ containing a non-oscillating term. Meanwhile, the factor $f(t)e^{i\omega_mt}$ in the squeezing coupling term will be $\Gamma_s(t)/\omega_m=(e^{i2\omega_m t}-e^{i\omega_m t})/\omega_m$. A cooling can be performed with a sufficiently large $\omega_m/\kappa$, because in Eq. (5) it suppresses the integrals of the matrix elements carrying the oscillating factor $\Gamma_s(t)$.
Our dynamical approach directly gives the picture of how the changes of the system parameters will turn the system from heating to cooling; see the time evolutions of the thermal phonon numbers in Figs. 2(a) and 2(b). In the transitional regimes where the cooling (BS) and heating (squeezing) effect compete with each other, the phonon numbers exhibit oscillations with time, as those in the insets of the figures. The corresponding cavity photon numbers in Figs. 2(c) and 2(d) evolve synchronously with the thermal phonon numbers, and the phenomenon is discussed further in [@supp].
New understanding of optomechanical cooling can be obtained from this dynamical picture, as it brings about another important parameter for cooling in addition to the widely concerned sideband resolution $\omega_m/\kappa$. A relevant phenomenon shown in Fig. 3(a) is that a higher sideband resolution beyond a certain value will be actually worse for cooling if the drive intensity $E$ is fixed. Our predicted thermal phonon number $n_{m,f}$ comes from the noise contributions as the integrals in Eq. (\[n\]), and is therefore determined by the whole process from $t=0$ to the approximate end time $t=t_s$ when $n_{m}^{(n)}(t)$ begins to be stable and $n_{m}^{(s)}(t_s)\rightarrow 0$ (the associated dynamical behaviors at $t_s$ asymptotically approach those at $t=\infty$). A slower process will accumulate a higher noise contribution $n_{m,f}$, which can be the time average of $n_{m}^{(n)}(t)$ ($t>t_s$) for the finally oscillating ones in Fig. 2(b). At the BS coupling resonance $\Delta=\omega_m$, the coefficient of the BS (squeezing) coupling term in Eqs. (2)-(3) takes the form $J \Gamma_{b(s)}(\kappa t)$. The parameter defined as $J=(g/\omega_m)\cdot (E/\kappa)$ decides how fast the cooling of a specific OMS will be; see the illustration in Fig. 2(a). Were there no quantum noises, an uncoupled system with $J=0$ would take the longest time scale $1/\gamma_m$ ($\gamma_m\ll \kappa$) to reach its stability; cf. the solution $\hat{b}(t)=e^{-\gamma_m t}\hat{b}$ to Eq. (3) when $E, \hat{\xi}_m(t)=0$. With the drive intensity $E$ fixed as in Fig. 3(a), a lower sideband resolution $\omega_m/\kappa$ corresponding to a higher $J$ makes the cooling faster. It is also shown in Fig. 2(b) that, once the parameter $J$ is fixed, the OMSs with different sideband resolution $\omega_m/\kappa$ will evolve to the stabilized phase almost together.
Apart from speeding up the cooling of an OMS, increasing $J$ will make the coexisting squeezing (heating) effect stronger. These two tendencies strike a balance somewhere in the parameter space, so that the best cooling under a fixed drive intensity $E$ takes place at an optimum $\omega_m/\kappa$ \[see Fig. 3(b)\]. The latter tendency will dominate when the parameter $\omega_m/\kappa$ is continuously lowered, which will increase $n_{m,f}$ significantly on the left of the optimum $\omega_m/\kappa$. On the other hand, the suppression of the factor $\Gamma_s(t)$ by faster oscillations due to larger $\omega_m/\kappa$ will diminish the squeezing effect, leading to the tendency of the stabilized $n_m$ in Fig. 2(b). Among the three different terms in Eq. (7), the first one independent of $n_{th}$ monotonically decreases with increased $\omega_m/\kappa$, as illustrated in Figs. 3(c) and 3(d). The experimental investigation of the effect related to this contribution was recently reported in [@ex12]. Such pure cavity noise contribution decided by the squeezing coupling intensity $J\Gamma_s(\kappa t)$ increases significantly with the drive intensity $E$, which distinguishes our results fundamentally from the previous predictions; see the comparisons in Fig. 3(d) and Fig. S-4(b) in [@supp].
Another important issue is how to reach the best cooling from the initial thermal equilibrium of an OMS. To answer the question, we first examine the limit of $\omega_m/\kappa\rightarrow \infty$. In this limit the stabilized thermal phonon number given by $\Delta=\omega_m$ can be found analytically from Eqs. (2)-(3) as $$\begin{aligned}
n_{m,f}
&=& \Gamma_m n_{th}\frac{1}{|\eta_{+}-\eta_{-}|^2}\nonumber\\
&\times & \big\{4\mbox{Re} \{\frac{\eta_{+}^\ast \eta_{-}}{(\lambda_+^\ast+\lambda_-)}\}-\frac{|\eta_{+}|^2}{\mbox{Re}\lambda_{-}}-\frac{|\eta_{-}|^2}{\mbox{Re}\lambda_{+}}\big\},
\label{l}\end{aligned}$$ where $\lambda_{\pm}=\frac{1}{2}(-1-\Gamma_m\pm \sqrt{(1-\Gamma_m)^2-4J^2})$ and $\eta_{\pm}=-1+\Gamma_m\pm \sqrt{(1-\Gamma_m)^2-4J^2}$ with $\Gamma_m=\gamma_m/\kappa$. It is solely from the contribution of the second term in Eq. (\[n\]), since the squeezing action with its intensity $J\Gamma_s(\kappa t)$ has been completely averaged out by an infinite $\omega_m/\kappa$. Analogous to a first-order phase transition, this limiting value exhibits a discontinuous jump across the point $J=1/2(1-\Gamma_m)$ (in Fig. 4 it is around $J=0.5$), which separates the weak coupling regime from the strong coupling one that has been experimentally observed [@ex8; @ex10; @strong]. In the strong coupling regime, the limiting value becomes the constant $n_{m,f}=\Gamma_m n_{th}$. Inside the weak coupling regime the stabilized phonon numbers for different $\omega_m/\kappa$ can be close to the limiting value and drop quickly with increased $J$ (see the inset in Fig. 4). With the parameters of the experimental setup in [@ex9] (equivalent to $J\approx 0.16$), for example, an ideal cooling process would not reach a phonon number lower than the limit of $n_{m,f}\approx 0.26$.
The parametric conditions for achieving good cooling manifest more clearly with Fig. 4, in which the limiting value in Eq. (\[l\]) gives the boundary any cooling process cannot surpass. For a setup with the built-in sideband resolution $\omega_m/\kappa$, increasing the drive power can make its cooling better and faster, moving the achieved thermal phonon number along a curve of fixed $\omega_m/\kappa$ to larger $J$, though the enhanced power may change the phonon number less obviously for a system with $\omega_m/\kappa\gg 1$ (in the strong coupling regime its phonon curve asymptotically approaches the limit line). The higher the parameter $\omega_m/\kappa$ is, the stronger the drive can be applied to achieve a better cooling; this explains the phonon number tendency on the right of the optimum points in Fig. 3(b). However, the cooling cannot be improved further if $J$ arrives at a turning point (the ending dots of the phonon curves) where the coexisting squeezing effect begins to be significant. After crossing the point, the thermal phonon number will increase from the minimum value there, and its evolution will continue to become oscillating and then growing with time as in Fig. 2(a). Such a reference point indicates that the cooling processes in Fig. 2(b) are not the optimum ones—the parameter $J=1.0$ used there has been on the right side of the turning point for $\omega_m/\kappa=8$.
The evolving Gaussian states of the weakly coupled OMSs, which follow the linear dynamical equations, can be depicted with their Wigner functions. A cooling process starts from a mechanical oscillator’s thermal state with its Wigner function being $W(q_m, p_m)=[1/\pi(1+2n_{th})]\exp\{-[1/(1+2n_{th})](q_m^2+p_m^2)\}$. After turning on a drive, the OMS will evolve under optomechanical coupling to two-mode Gaussian states, whose Wigner functions can be found by numerically calculating the correlation matrix of the two system modes. As illustrated in Fig. 1(b), a perfect “ground-state cooling" to $n_{m,f}=0$ evolves the mechanical mode’s Wigner function to $W(q_m, p_m)=\frac{1}{\pi}e^{-(q_m-q_m^0)^2-(p_m-p_m^0)^2}$ of a coherent state, which differs from that of a vacuum state only by the displacements $q_m^0(t)$, $p_m^0(t)$ determined by the coherent drive term $\vec{\lambda}(t)$ in Eq. (4). Without impairing the purity [@purity] of the target coherent state, the contribution from the coherent drives should be excluded from the thermal phonon number $n_{m}(t)$ to be reduced from the occupation $n_{th}$ of an initial thermal state.
In summary, we have developed a quantum dynamical approach to optomechanical cooling. The motivation for the development is to reflect the fact that such cooling is a process for a mechanical oscillator to evolve from its initial thermal state to another state with higher purity, which takes time. Cooling starts after turning on a red-detuned drive that realizes a BS type coupling between the fast damping cavity mode and the slowly decaying mechanical mode, so that the oscillator’s thermal excitation being converted to cavity photons could be totally eliminated after a time controlled by the parameter $J$. Only with this scenario giving the phonon number in Eq. (\[s\]), the mechanical oscillator would be cooled down to a vacuum state modified by its inevitable motion under radiation pressure to a coherent state. The simultaneous quantum noise actions, however, cause the evolution to deviate from going to such a pure quantum state and add to the thermal phonon number in Eq. (\[n\]). The thermal occupation left in the end is decided by how soon the coupled cavity and mechanical modes evolve together to a dynamical stability, after which the system cannot be cooled down further. Like the fully quantum mechanical treatment of OMS in [@law] and recent studies of the quantum dynamical features of other physical systems (see, e.g. [@dd1; @dd2]), the properties in a cooling process illustrated here are for a genuine quantum OMS rather than the quantum fluctuations around the trajectories of a classical one; this approach can also be used for blue-detuned drives [@s-laser]. This quantum dynamical picture of cooling applies to an OMS truly approaching its macroscopic quantum states.
We acknowledge the funding supports from NSFC (Grant No. 11574093 and Grant No. 61435007). L. Y. is sponsored by the Fundamental Research Funds for the Central Universities (No. HEUCFJ170402). This research is also supported by the Arkansas High Performance Computing Center and the Arkansas Economic Development Commission.
B. H. and L. Y. contributed equally to this work.
[Supplementary Material for “Radiation Pressure Cooling as a Quantum Dynamical Process"]{} {#supplementary-material-for-radiation-pressure-cooling-as-a-quantum-dynamical-process .unnumbered}
==========================================================================================
I. DYNAMICAL EQUATIONS OF OPTOMECHANICAL SYSTEMS IN A ROTATING FRAME
--------------------------------------------------------------------
A quantum optomechanical system (OMS) is modeled by the coupled cavity mode $\hat{a}$ with the frequency $\omega_c$ and mechanical mode $\hat{b}$ with the frequency $\omega_m$, having the self-oscillation Hamiltonians $\omega_c\hat{a}^\dagger\hat{a}$ and $\omega_m\hat{b}^\dagger\hat{b}$ ($\hbar=1$) for the respective mode. Initially, when there is no external drive field, the mechanical oscillator is in a thermal equilibrium with the environment at a temperature corresponding to the occupation $n_{th}$. An external drive with the intensity $E$ and frequency $\omega_l$, as seen from the Hamiltonian $iE(\hat{a}^{\dagger}e^{-i\omega_l t}-\hat{a}e^{i\omega_l t})$, will build up the cavity field. Under the radiation pressure of the cavity field, the size of the cavity will be changed from $L$ to $L+\hat{x}_m$, where $\hat{x}_m=\sqrt{\frac{1}{2m\omega_m}}(\hat{b}+\hat{b}^\dagger)$ is the displacement of the mechanical oscillator with the effective mass $m$ (the notation of $\hbar=1$ is adopted), and is treated as a q-number for a quantum OMS. The consequentially modified cavity oscillation term $n\pi c/(L+\hat{x}_m)\hat{a}^\dagger\hat{a}\approx \omega_c\hat{a}^\dagger\hat{a}-(\omega_c/L)\hat{x}_m\hat{a}^\dagger\hat{a}$ gives the optomechanical coupling Hamiltonian as the second term written as $-g\hat{a}^\dagger\hat{a}(\hat{b}+\hat{b}^\dagger)$, where $g=\sqrt{\frac{1}{2m\omega_m}}\frac{\omega_c}{L}$.
In addition to the above-mentioned elements, the OMS is an open quantum system coupled to the environment known as the cavity and mechanical reservoirs, which are modeled as ensembles of oscillators with the Hamiltonian $$\begin{aligned}
H_r=\int_0^\infty d\omega_1\omega_1\hat{\xi}^\dagger_c\hat{\xi}_c(\omega_1)+\int_0^\infty d\omega_2\omega_2\hat{\xi}^\dagger_m\hat{\xi}_m(\omega_2)\end{aligned}$$ where $[\hat{\xi}_{c,m}(\omega),\hat{\xi}^\dagger_{c,m}(\omega')]=\delta(\omega-\omega')$. The Hamiltonian for the coupling between the system and the reservoirs takes the general form $$\begin{aligned}
H_{s-r}=i\int d\omega_1 \kappa(\omega_1)(\hat{a}-\hat{a}^\dagger)\{\hat{\xi}^\dagger_c(\omega_1)+\hat{\xi}_c(\omega_1)\}+i\int d\omega_2 \gamma_m(\omega_2)(\hat{b}-\hat{b}^\dagger)\{\hat{\xi}^\dagger_m(\omega_2)+\hat{\xi}_m(\omega_2)\}.\end{aligned}$$ In the rotating frame with respect to the cavity frequency $\omega_c$ and the mechanical frequency $\omega_m$, as well as to the frequencies of the reservoir modes, the system-reservoir coupling Hamiltonian will become $$\begin{aligned}
H^{(i)}_{s-r}&=&i\int d\omega_1 \kappa(\omega_1)(\hat{a}e^{-i\omega_c t}-\hat{a}^\dagger e^{i\omega_c t})\{\hat{\xi}^\dagger_c(\omega_1)e^{i\omega_1 t}+\hat{\xi}_c(\omega_1)e^{-i\omega_1 t}\}\nonumber\\
&+&i\int d\omega_2 \gamma_m(\omega_2)(\hat{b}e^{-i\omega_m t}-\hat{b}^\dagger e^{i\omega_m t})\{\hat{\xi}^\dagger_m(\omega_2)e^{i\omega_2 t}+\hat{\xi}_m
(\omega_2)e^{-i\omega_2 t}\},
\label{couple}\end{aligned}$$ as the result of the unitary transformation with $U_I(t)=\exp\{-i\int_0^t d\tau (H_r+\omega_c\hat{a}^\dagger\hat{a}+\omega_m\hat{b}^\dagger\hat{b})\}$. Similar to the practices in [@supp1], we make a rotating-wave approximation to neglect the rapidly oscillating terms in Eq. (\[couple\]), together with an approximation of smooth system-reservoir couplings to reduce the associated coupling intensities to constants, i.e, $\kappa(\omega)\rightarrow \sqrt{\kappa/\pi}$ and $\gamma_m(\omega)\rightarrow \sqrt{\gamma_m/ \pi}$. Then the system-reservoir coupling will be simplified to $$\begin{aligned}
H_{sr}(t)=i\sqrt{2\kappa }\{\hat{a}^{\dagger }\hat{\xi}_{c}(t)-\hat{a}\hat{\xi}^{\dagger}_{c}(t)\}+i\sqrt{2\gamma_m }\{\hat{b}^{\dagger }
\hat{\xi}_{m}(t)-\hat{b}\hat{\xi}^{\dagger }_{m}(t)\},
\label{couple1}\end{aligned}$$ where $\hat{\xi}_{c,m}(t)=\frac{1}{\sqrt{2\pi}}\int d\omega \hat{\xi}_{c,m}(\omega)e^{-i(\omega-\omega_{c,m})t}$. For the relatively slow interaction processes described by Eq. (\[couple1\]), the quantum noise represented by $\hat{\xi}_{c,m}(t)$ can be approximated as a white one satisfying $[\hat{\xi}_{c,m}(t),\hat{\xi}^\dagger_{c,m}(t')]=\delta(t-t')$ [@supp1].
Meanwhile, under the above-mentioned rotation by $U_I(t)$, the external drive Hamiltonian and the optomechanical coupling Hamiltonian will become $H_e(t)=iE(\hat{a}^{\dagger}e^{i\Delta t}-\hat{a}e^{-i\Delta t})$, where $\Delta=\omega_c-\omega_l$, and $H_{om}(t)=-g\hat{a}^\dagger\hat{a}(\hat{b}e^{-i\omega_m t}+\hat{b}^\dagger e^{i\omega_m t})$, respectively. The action $U(t)={\cal T}\exp\{-i\int_0^t d\tau H(\tau)\}$ of the total Hamiltonian $H(t)=H_e(t)+H_{om}(t)+H_{sr}(t)$, which is called the formal solution to Quantum Stochastic Schrödinger equations in [@supp1], will lead to the following nonlinear dynamical equations of the system modes: $$\begin{aligned}
\dot{\hat{a}}&=&-\kappa\hat{a}+ig(\hat{b}e^{-i\omega_m t}+\hat{b}^\dagger e^{i\omega_m t})\hat{a}+E e^{i\Delta t}+\sqrt{2\kappa}\hat{\xi}_c(t),\nonumber\\
\dot{\hat{b}}&=&-\gamma_m\hat{b}+ige^{i\omega_m t}\hat{a}^\dagger\hat{a}+\sqrt{2\gamma_m}\hat{\xi}_m(t).
\label{e1}\end{aligned}$$ Among the literature, the dynamics of an OMS is also described in another rotating frame only with respect to the drive frequency $\omega_l$ (see, e.g. [@rev]). Then its dynamical equations read $$\begin{aligned}
\dot{\hat{a}}&=&-\kappa\hat{a}-i\Delta \hat{a}+ig(\hat{b}+\hat{b}^\dagger)\hat{a}+E +\sqrt{2\kappa}\hat{\xi}_c(t),\nonumber\\
\dot{\hat{b}}&=&-\gamma_m\hat{b}-i\omega_m\hat{b}+ig\hat{a}^\dagger\hat{a}+\sqrt{2\gamma_m}\hat{\xi}_m(t),
\label{e2}\end{aligned}$$ which directly uses the system-reservoir coupling Hamiltonian in Eq. (\[couple1\]).
II. LINEARIZED DYNAMICS OF WEAKLY COUPLED QUANTUM OPTOMECHANICAL SYSTEMS
------------------------------------------------------------------------
For the total Hamiltonian $H(t)=H_e(t)+H_{om}(t)+H_{sr}(t)$ in our used rotating frame, we take another interaction picture with respect to the part of $H_e(t)$, to have the cavity mode $\hat{a}$ displaced by the unitary operation $U_0(t)={\cal T}\exp\{-i\int_0^t d\tau H_e(\tau)\}$ to $\hat{a}-i\frac{e^{i\Delta t}-1}{\Delta}E$. Then the remaining part of the Hamiltonian $H(t)$ will be transformed to $$\begin{aligned}
H^{in}(t)&=& U_0^\dagger(t)\big(H_{om}(t)+H_{sr}(t)\big)U_0(t)\nonumber\\
&=&\underbrace{ig E\frac{e^{i\Delta t}-1}{\Delta}e^{-i\omega_mt} \hat{a}^\dagger \hat{b}-ig E \frac{e^{-i\Delta t}-1}{\Delta}e^{i\omega_mt} \hat{a} \hat{b}^\dagger
+ig E\frac{e^{i\Delta t}-1}{\Delta}e^{i\omega_mt} \hat{a}^\dagger \hat{b}^\dagger-ig E \frac{e^{-i\Delta t}-1}{\Delta}e^{-i\omega_mt} \hat{a} \hat{b}}_{H_{couple}(t)}\nonumber\\
&-&g\underbrace{\big|\frac{e^{i\Delta t}-1}{\Delta}E\big|^2(e^{-i\omega_mt}\hat{b}+e^{i\omega_m t}\hat{b}^\dagger)}_{H_{displace}(t)}
+i\sqrt{2\kappa}\{(\hat{a}^\dagger +i\frac{e^{-i\Delta t}-1}{\Delta}E)\hat{\xi}_p(t)
-(\hat{a}-i\frac{e^{i\Delta t}-1}{\Delta}E)\hat{\xi}_p^\dagger(t)\}\nonumber\\
&+&i\sqrt{2\gamma_m }\{\hat{b}^{\dagger}\hat{\xi}_{m}(t)-\hat{b}\hat{\xi}_{m}^\dagger (t)\}
-\underbrace{g\hat{a}^\dagger\hat{a}(e^{-i\omega_mt}\hat{b}+e^{i\omega_m t}\hat{b}^\dagger)}_{H_{nl}(t)}.
\label{Hi}\end{aligned}$$ Among this transformed Hamiltonian $H^{in}(t)$, the effect of the cubic term $H_{nl}(t)$ is insignificant compared with the others. Because it is non-commutative with the rest of the Hamiltonian, its action will modify the mode operators in the rest part of $H^{in}(t)$ by the unitary action [@s-laser] $$\begin{aligned}
U_{nl}(t,\tau)&=&{\cal T}e^{i\int_\tau^t dt' g\hat{a}^\dagger\hat{a}(e^{-i\omega_m t'}\hat{b}+e^{i\omega_m t'}
\hat{b}^\dagger)}\nonumber\\
&=& e^{i\int_\tau^t dt' g\hat{a}^\dagger\hat{a}\hat{b}e^{-i\omega_m t'}}e^{i\int_\tau^t dt' g\hat{a}^\dagger\hat{a}\hat{b}^\dagger e^{i\omega_m t'}}
e^{i\int_\tau^t dt'\frac{g^2}{\omega_m}(1-e^{i\omega_m(t'-\tau)})(\hat{a}^\dagger\hat{a})^2},
\label{NL}\end{aligned}$$ as it can be seen from the expectation value of any system operator $\hat{O}$: $$\begin{aligned}
\langle \hat{O}(t)\rangle&=&\mbox{Tr}_s\{\hat{O}\rho(t)\}=\mbox{Tr}_s\{\hat{O}\mbox{Tr}_r(U(t)\rho(0)\rho_r U^\dagger(t))\}\nonumber\\
&=&\mbox{Tr}_{s,r}\{{\cal T}e^{i\int_0^t d\tau U_{nl}(t,\tau)\{H^{in}(\tau)+H_{nl}(\tau)\}U^\dagger_{nl}(t,\tau)}U^\dagger_0(t)\hat{O}U_0(t){\cal T}e^{-i\int_0^t d\tau U_{nl}(t,\tau)\{H^{in}(\tau)+H_{nl}(\tau)\}U^\dagger_{nl}(t,\tau)}\nonumber\\
&\times & (U_{nl}(t,0)\rho(0)\rho_r U^\dagger_{nl}(t,0))\}\nonumber\\
&\approx &\mbox{Tr}_{s,r}\{{\cal T}e^{i\int_0^t d\tau \{H^{in}(\tau)+H_{nl}(\tau)\}}U^\dagger_0(t)\hat{O}U_0(t){\cal T}e^{-i\int_0^t d\tau \{H^{in}(\tau)+H_{nl}(\tau)\}}\rho(0)\rho_r\},
\label{expect}\end{aligned}$$ where $U(t)={\cal T}\exp\{-i\int_0^t d\tau H(\tau)\}$ is the action including that of the stochastic part $H_{sr}(t)$ [@supp1], and $\rho_r$ represents the total quantum state of the reservoirs. Because a concerned dynamical evolution starts from a vacuum state $|0\rangle_c$ of the cavity and a thermal state $\rho_m(0)=\sum_{n=0}^{\infty} \frac{n_{th}^n}{(1+n_{th})^{n+1}}\left\vert n\right\rangle_m\left\langle n\right\vert$ of the mechanical oscillator ($\rho(0)$ is their product and $n_{th}$ is the initial thermal occupation), the action $U_{nl}(t,0)\rho(0)U^\dagger_{nl}(t,0)$ keeps this initial state invariant on the third line of Eq. (\[expect\]), as a consequence of the relation $U_{nl}(t,0)|0\rangle_c=|0\rangle_c$. Note that, in Eq. (\[expect\]), the operators $\hat{c}=\hat{a},\hat{b}$ in $H^{in}(\tau)+H_{nl}(\tau)$ are transformed by the action in Eq. (\[NL\]) to $U_{nl}(t,\tau)\hat{c}U^\dagger_{nl}(t,\tau)$, unlike the general form $U^{\dagger}(t)\hat{o}U(t)$ for a unitary operation on an operator $\hat{o}$. According to Eq. (\[NL\]), the modified operators differ from the original ones $\hat{c}$ by the corrections in the orders of $g/\omega_m\ll 1$, so these corrections can be well neglected for a weakly coupled OMS. This step noted with an approximation sign on the last line of Eq. (\[expect\]) is the only theoretical approximation made in our approach, in addition to those commonly used ones to obtain the system-reservoir coupling Hamiltonian in Eq. (\[couple1\]). The neglecting of the cubic term $H_{nl}(t)$ has a physically different meaning from that of omitting the higher-order fluctuation terms in the previously used linearization around the steady states of the averaged Eq. (\[e2\]), because it is not necessary for our completely quantum treatment to require that the cavity mean field should overwhelm the corresponding cavity field fluctuation.
By using the associated Ito’s rules for the stochastic part [@supp1], one will obtain the following dynamical equation $$\begin{aligned}
\frac{d}{dt}\left(
\begin{array}{c}
\hat{a}\\
\hat{a}^\dagger\\
\hat{b}\\
\hat{b}^\dagger
\end{array}
\right)
&=&
\underbrace{\left(
\begin{array}{cccc}
-\kappa & 0 & gE f(t)e^{-i\omega_m t} & gE f(t)e^{i\omega_m t}\\
0 &-\kappa & gE f^\ast(t)e^{-i\omega_m t}& gE f^\ast(t)e^{i\omega_m t}\\
-gE f^\ast (t)e^{i\omega_m t}&gE f(t)e^{i\omega_m t}&-\gamma_m & 0\\
gE f^\ast(t)e^{-i\omega_m t} &-gE f(t)e^{-i\omega_m t} &0 &-\gamma_m\\
\end{array}
\right)}_{\hat{M}(t)}\left(
\begin{array}{c}
\hat{a}\\
\hat{a}^\dagger\\
\hat{b}\\
\hat{b}^\dagger
\end{array}
\right) \nonumber\\
&+&
\underbrace{\left( \begin{array}{c}
i\kappa f(t)E \\
-i\kappa f^\ast(t)E\\
ig|f(t)E|^2e^{i\omega_mt}\\
-ig|f(t)E|^2e^{-i\omega_mt}\\
\end{array}
\right)}_{\vec{\lambda}(t)}+\underbrace{\left( \begin{array}{c}
\sqrt{2\kappa}\hat{\xi}_c(t) \\
\sqrt{2\kappa}\hat{\xi}^\dagger_c(t)\\
\sqrt{2\gamma_m}\hat{\xi}_m(t)\\
\sqrt{2\gamma_m}\hat{\xi}^\dagger_m(t)\\
\end{array}
\right)}_{\vec{\hat{\eta}}(t)}
\label{dynamics}\end{aligned}$$ with the effective Hamiltonian $H^{in}(t)+H_{nl}(t)$. This extended form of Eqs. (2)-(3) in the main text contains the factor $f(t)=\frac{e^{i\Delta t}-1}{\Delta}$ in the dynamical matrix $\hat{M}(t)$ and the coherent drive term $\vec{\hat{\lambda}}(t)$. The numerical solution to this differential equation can be found by calculating the time-ordered exponentials (in Eq. (5) of the main text) as the products of the matrices $(I+\hat{M}(t_i)\delta t)$ at the different $t_i$ [@s-laser], where $\delta t$ is a small iteration step. The corresponding quantum master equation for the reduced system state $\rho(t)$ reads $$\begin{aligned}
\dot{\rho}(t)&=&-i[H_{couple}(t)-H_{displace}(t),\rho(t)]-\kappa\big(\hat{o}_a^\dagger\hat{o}_a\rho(t)+\rho(t)\hat{o}_a^\dagger\hat{o}_a-2\hat{o}_a\rho(t)\hat{o}_a^\dagger\big)\nonumber\\
&-&\gamma_m(n_{th}+1)\big(\hat{b}^\dagger\hat{b}\rho(t)+\rho(t)\hat{b}^\dagger\hat{b}-2\hat{b}\rho(t)\hat{b}^\dagger\big)-\gamma_m n_{th}\big(\hat{b}\hat{b}^\dagger\rho(t)+\rho(t)\hat{b}\hat{b}^\dagger-2\hat{b}^\dagger\rho(t)\hat{b}\big)\end{aligned}$$ with the notations in Eq. (\[Hi\]), where $\hat{o}_a=\hat{a}-i\frac{e^{i\Delta t}-1}{\Delta}E$.
As a comparison, in the previous approach of linearization with the decomposition $\hat{a}=\alpha+\delta \hat{a}$ or the displacement $\hat{a}\rightarrow \hat{a}+\alpha$ with a steady-state value $\alpha$ of the cavity mode $\hat{a}$, the linearized Hamiltonian excluding the coherent and noise drive terms takes the form [@supp2] $$\begin{aligned}
H_{linear}=G(\delta \hat{a} e^{-i\Delta t}+\delta \hat{a}^\dagger e^{i\Delta t})(\hat{b}e^{-i\omega_m t}+\hat{b}^\dagger e^{i\omega_m t}),
\label{lnh}\end{aligned}$$ with $|G|=g|\alpha|$ being a constant. A variation form of this linearized Hamiltonian has been used to study real-time cooling processes [@c6]. In reality, however, the constant optomechanical coupling intensity $G$, which equals to $g$ multiplied by the steady value $\alpha$ appearing after a certain period of time, can not exist before that time, for instance, shortly after turning on the cooling laser beam at $t=0$. The cavity field will certainly take time to evolve to stable value. So it is more appropriate to use an evolving field $\alpha(t)$, which develops from zero to stable value throughout a cooling process, in that linearization approach. Our quantum dynamical approach totally dispenses with such time-dependent classical mean values, which should be found by solving the nonlinear classical dynamical equations \[those obtained by taking the averages of Eq. (\[e1\]) or Eq. (\[e2\])\]. The linearization of the system dynamics is realized by a unitary transformation $\hat{a}\rightarrow \hat{a}-i\frac{e^{i\Delta t}-1}{\Delta}E$ of the cavity mode, which is completely inequivalent to the previous linearization by the shift $\hat{a}\rightarrow \hat{a}+\alpha$ with a time-independent classical steady-state value $\alpha$.
III. FURTHER DISCUSSION ON THE EFFECTS OF THE COHERENT DRIVES
-------------------------------------------------------------
A majority of physical effects in the concerned cooling processes come from the coherent drive term $\vec{\lambda}(t)$ and the noise drive term $\vec{\hat{\eta}}(t)$ in Eq. (\[dynamics\]). Here we discuss more of their properties that are complementary to those described in the main text. We first look at the former, which has been less explored by the previous studies. In terms of the five dimensionless parameters $G_m=g/\kappa, {\cal E}=E/\kappa, s_m=\omega_m/\kappa, Q=\omega_m/\gamma_m, \delta=\Delta/\omega_m$, as well as the dimensionless time $\tau$ going from $0$ to $\kappa t$, the evolving cavity photon number $$\langle \hat{a}^\dagger\hat{a}(t)\rangle=\langle U^\dagger(t)\hat{a}^\dagger\hat{a} U(t)\rangle=\langle {\cal T}e^{i\int_0^t d\tau \{H^{in}(\tau)+H_{nl}(\tau)\}}U^\dagger_0(t)\hat{a}^\dagger\hat{a}U_0(t){\cal T}e^{-i\int_0^t d\tau \{H^{in}(\tau)+H_{nl}(\tau)\}}\rangle$$ is given as the sum of the following three parts: (1) the system operator contribution $$\begin{aligned}
\langle \hat{a}^\dagger\hat{a}(t)\rangle_s=d_{21}(\kappa t,0)d_{12}(\kappa t,0)+d_{23}(\kappa t,0)d_{14}(\kappa t,0)(n_{th}+1)+d_{24}(\kappa t,0)d_{13}(\kappa t,0)n_{th};
\label{c1}\end{aligned}$$ (2) the noise drive contribution $$\begin{aligned}
\langle \hat{a}^\dagger\hat{a}(t)\rangle_n&=&2\int_0^{\kappa t} d\tau d_{21}(\kappa t,\tau)d_{12}(\kappa t,\tau)+\frac{2s_m}{Q}\int_0^{\kappa t} d\tau d_{23}(\kappa t,\tau)d_{14}(\kappa t,\tau)(n_{th}+1)\nonumber\\
&+&\frac{2s_m}{Q}\int_0^{\kappa t} d\tau d_{24}(\kappa t,\tau)d_{13}(\kappa t,\tau)n_{th};
\label{c2}\end{aligned}$$ (3) the contribution $$\begin{aligned}
\langle \hat{a}^\dagger\hat{a}(t)\rangle_c &=&\big |{\cal E}\int_0^{\kappa t} d\tau e^{i\delta s_m \tau}+i \frac{{\cal E}}{s_m \delta}\int_0^{\kappa t} d\tau d_{11}(\kappa t,\tau)F(\tau)-i\frac{{\cal E}}{s_m \delta}\int_0^{\kappa t} d\tau d_{12}(\kappa t,\tau)F^\ast(\tau)\nonumber\\
&+&iG_m|\frac{{\cal E}}{s_m\delta}|^2 \int_0^{\kappa t} d\tau d_{13}(\kappa t,\tau)|F(\tau)|^2e^{is_m\tau}-iG_m|\frac{{\cal E}}{s_m\delta}|^2 \int_0^{\kappa t} d\tau d_{14}(\kappa t,\tau)|F(\tau)|^2 e^{-is_m\tau}\big|^2
\label{c3}\end{aligned}$$ purely from the coherent drive term, with $F(\tau)=e^{i\delta s_m \tau}-1$. In Eq. (\[c3\]) we have also included the contribution from the pure drive action $U_0(t)={\cal T}\exp\{-i\int_0^t d\tau H_e(\tau)\}$. The notation $d_{ij}(t,\tau)=[{\cal T}e^{\int_\tau^t dt' \hat{M}(t')}]_{ij}$ represents a matrix element for $i,j=1,2,3,4$. For a sufficiently high drive intensity $E$, the third contribution from the coherent drive term dominates over the two others.
We here study the cavity photon number evolutions in the weak coupling regime of small effective coupling intensity $J=(g/\omega_m)(E/\kappa)$, as a complement of the discussion on the relatively large $J$ in Figs. 2(a) and 2(c) of the main text. In Fig. S-1(a) about $J=0$, the numerical calculation gives a perfect agreement with the known analytical result of a cavity with fixed size—the photon number tends to a steady value after a period in the order of $1/\kappa$. By intuition, a small $J$ would only lead to perturbative effect, but this is true only at the beginning period of the dynamical evolutions—the plots in Fig. S-1(b) first show a tendency of getting close to the corresponding steady value in Fig. S-1(a), but keep the growing oscillations due to the existence of optomechanical interaction. The finally stabilized photon number evolutions are stably oscillating ones as shown in the upper inset of Fig. S-1(b), in which the photon numbers tend to such a stable dynamical evolution phase together with the corresponding thermal phonon numbers in these cooling processes; compare the corresponding plots in the lower and upper insets of Fig. S-1(b). The photon number curves in the upper inset of Fig. S-1(b) look stuck together because the oscillation period is much shorter than the time scale to reach the finally stable oscillations. The corresponding dimensionless displacements $q_m^0(t)= \frac{1}{\sqrt{2}}\langle \hat{b}(t)+\hat{b}^\dagger(t)\rangle$ of the mechanical oscillator for the three different situations in Fig. S-1(b) are plotted in Fig. S-2; they respectively become stable oscillations together with the cavity photon numbers and thermal phonon numbers. Such stable oscillation looks similar to those of classical OMSs due to the saturated nonlinear effect in the blue-detuned regime (see Sec. VIII in [@rev]), but has a different physical origin from that type of classical motion since we are currently concerned with the red-detuned regime. Another interesting phenomenon in the processes is the considerably increased average cavity photon numbers from that of an uncoupled system, when the systems become stabilized. We illustrate its mechanism in Fig. S-1(c), which shows that the increased photon numbers in cavity is due to the constructive interference of two factors, the displacement term irrelevant to the optomechanical coupling \[the first term, which is from the unitary action $U_0(t)$, inside the absolute sign in Eq. (\[c3\])\] and the cavity mode amplitude induced by the optomechanical coupling (the summation of the rest terms, which come from the coherent drive term $\vec{\lambda}(t)$ in Eq. (\[dynamics\]), in the same equation). The latter can be regarded as a relatively small effect for a weakly coupled OMS. Such constructive interference, which is predicted with a completely quantum mechanical treatment for an OMS entering its macroscopic quantum states, indicates that a small coupling $J$ can still affect the cavity photon number considerably. Enhanced cavity photon numbers in oscillation is a signature of macroscopic quantum states for the OMS.
There are two intrinsic time scales for reaching the dynamical stability in the cooling of an OMS, $1/\kappa$ for the cavity mode and $1/\gamma_m$ for the mechanical mode. When these two modes are uncoupled ($J=0$), the cavity mode tends to a steady state within a time in the order of $1/\kappa$ as in Fig. S-1(a), while the mechanical mode under a thermal equilibrium with the environment satisfies the equation $$\begin{aligned}
\dot{\hat{b}}&=&-\gamma_m\hat{b}+\sqrt{2\gamma_m}\hat{\xi}_m(t),\end{aligned}$$ which is from Eq. (\[e1\]). According to the above equation, the mechanical mode takes the time of $1/\gamma_m\gg 1/\kappa$ to damp to zero, a stable value, while the second noise drive term keeps the thermal phonon number $n_{th}$ invariant in this situation. Once these two system modes are coupled ($J\neq 0$), the time scale $t_s$ for the OMS to approach the stable evolution phase will be decided by their effective coupling intensity $J$ and be reduced from $1/\gamma_m$ due to the coupling to the cavity mode with a faster damping rate. For example, in the limit of $\omega_m/\kappa \rightarrow \infty$, the eigenvalues of the dynamical matrix $\hat{M}$ in Eq. (\[dynamics\]) degenerate to the analytical form $\lambda_{\pm}=\frac{1}{2}(-\kappa-\gamma_m\pm \sqrt{(\kappa-\gamma_m)^2-4J^2\kappa^2})$. The dynamical stability controlled by the two eigenvalues will be realized after a long time in the order of $1/\lambda_+\sim 1/\gamma_m$, if the coupling $J$ is very small. This explains the time scale to reach the stable evolution phases in Fig. S-1(b). By the time (in the order of $1/\kappa$) when the uncoupled cavity field in Fig. S-1(a) becomes steady, the coupled system is far from being in dynamical stability, so the cavity mode will keep the unstable oscillation of growing to higher amplitude until the oscillation under the cavity field’s damping completely stabilizes after much longer time. In contrast, the tendency of first dropping to a lower oscillation amplitude (close to the steady state of $J=0$) and then growing to a higher amplitude is less obvious in Fig. 2 of the main text, because at the beginning period the effects of the larger values of $J$ are no longer perturbative. Moreover, as the intensity of $J$ becomes still larger, the cavity photon numbers (if they finally stabilize) may exhibit more complicated periodic patterns due to the more obvious manifestation of the higher harmonic components $n\omega_m$ ($n>1$) from the dynamical matrix $\hat{M}(t)$ in Eq. (\[dynamics\]).
A stably oscillating average mechanical mode $\langle \hat{b}(t)\rangle$ in the end, instead of a stopped one, is a major difference from the previous picture. The classical dynamical equation for the cavity mean value, which is due to a finally stabilized mechanical oscillation $\langle \hat{b}(t)\rangle=\beta e^{-i\omega_m t}$ as in Fig. S-2, can be found from Eq. (\[e2\]) as $$\begin{aligned}
\dot{\alpha}&=&-\kappa\alpha-\big(i\Delta-ig(\beta e^{-i\omega_m t}+\beta e^{i\omega_m t})\big)\alpha+E.
\label{cl-solution}\end{aligned}$$ The solution to this equation is the series $\alpha(t)=e^{i\phi(t)}\sum_{n=-\infty}^\infty\alpha_n e^{in\omega_m t}$ with $$\begin{aligned}
\alpha_n=E\frac{J_n(-\frac{g\beta}{\omega_m})}{i(n\omega_m+\Delta)+\kappa},
\label{series}\end{aligned}$$ where $J_n(x)$ is the Bessel function of the first kind and $\phi(t)$ is a global phase [@supp2; @supp0]. The result $|\alpha_0 +\alpha_{-1}e^{-i\omega_m t}+\alpha_{1}e^{i\omega_m t}|^2$, which gives higher cavity photon numbers than those of uncoupled systems ($J=0$) due to the interference with the extra sidebands, can qualitatively explain the stabilized photon numbers. For example, because the finally stabilized factors $g\beta$ are close to a same value for the three different oscillations in Fig. S-2, the corresponding photon numbers that are to be stabilized will become close to one another as shown by the upper inset in Fig. S-1(b).
Static expectation values of the cavity and mechanical modes, i.e. $\frac{d}{dt}\langle\hat{a}(t)\rangle=\frac{d}{dt}\langle\hat{b}(t)\rangle=0$, for the final state of a cooling process were assumed in the previous studies; see, e.g. the vanishing amplitude $\beta \rightarrow 0$ of the mechanical oscillation as described in Sec. VIII of Ref. [@rev]. According to the classical mean field $\alpha(t)$ determined with the above Eqs. (\[cl-solution\]) and (\[series\]), such vanishing mechanical oscillation amplitude will lead to a cavity photon number indifferent to that of an uncoupled cavity ($J=0$). If the mechanical oscillator is assumed to stop moving at a nonzero displacement $\langle \hat{q}^0_m\rangle=\bar {x}_m$, the cavity field amplitude \[only from the contribution of the term containing $J_0(-\frac{g\beta}{\omega_m})$\] will become $\alpha_0=E/\{i(\Delta-g\bar{x}_m)+\kappa\}$, an exact steady-state solution to be reached within a time $t\sim 1/\kappa$ \[similar to the evolution in Fig. S-1(a)\]. The static picture thus implies the separate dynamical evolutions of the cavity and mechanical modes, which are very different from the illustrations in Fig. S-1(b).
The cooling of a mechanical oscillator starts from its thermal state with the Wigner function $W(q_m, p_m)=\frac{1}{\pi(1+2n_{th})}\exp\{-\frac{1}{(1+2n_{th})}(q_m^2+p_m^2)\}$. A perfect cooling is to the ground state $|0\rangle_m$ of the mechanical oscillator with its Wigner function $W(q_m, p_m)=\frac{1}{\pi}\exp\{-(q_m^2+p_m^2)\}$, having the associated thermal phonon number reduced from $n_{th}$ to $0$. Due to the inevitable motion of the mechanical mode under the radiation pressure for cooling, in such perfect cooling its Wigner function will actually become $W(q_m, p_m)=\frac{1}{\pi}\exp\{-(q_m-q_m^0(t))^2+(p_m-p_m^0(t))^2)\}$, which is that of a coherent state with the displacements $q_m^0(t)$, $p_m^0(t)$ in the phase space. An arbitrary time-dependent function $q_m^0(t)$ or $p_m^0(t)$ does not affect the cooling result, since the purity of this target coherent state is irrelevant to the average position and average momentum of the oscillator. Actually, cooling a mechanical oscillator is to realize its pure macroscopic quantum states, no need to let the cavity and mechanical mode be frozen in the phase space. More discussion on the picture can be found in the next section.
IV. FURTHER DISCUSSION ON THE EFFECTS OF THE NOISE DRIVES
---------------------------------------------------------
In the calculation of the thermal phonon numbers, we have excluded the coherent drive term $\vec{\lambda}(t)$’s contribution. The displacement Hamiltonian $H_{displace}(t)$ in Eq. (\[Hi\]) certainly contributes to the total occupation $\langle \hat{b}^\dagger\hat{b}\rangle$ of the mechanical mode, which will not drop to a lower level throughout a cooling process as indicated by the overall mechanical motion like those in Fig. S-2. Even if the mechanical oscillator were stopped, the resulting nonzero displacement due to the radiation pressure would still significantly contribute to $\langle \hat{b}^\dagger\hat{b}\rangle$ different from our concerned thermal phonon number. In the previous linearization of an OMS’s dynamical equations by shifting cavity mode by its steady-state value $\alpha$, such effect exhibits as a displacement term $ig|\alpha|^2$ from Eq. (\[e2\]). After excluding the coherent drive term’s contribution, the result of a cooling process is determined by the coexisting action of the noise drives, as the finally stabilized thermal phonon number in Eq. (7) of the main text.
The noise contributions from different sources can be identified as the three different terms in Eq. (7) of the main text. They are, respectively, $n_{m,1}^{(n)}$ from the cavity noise drive via squeezing (SQ) type coupling, $n_{m,2}^{(n)}$ from the mechanical reservoir noise drive via BS type coupling, and $n_{m,3}^{(n)}$ from the mechanical noise drive via SQ type coupling. The first contribution, which exists even at the zero temperature due to the remnant squeezing effect, can be lowered with increased sideband resolution $\omega_m/\kappa$ as seen from Figs. S-3(c) and S-3(d). The cooling limit when the parameter $\omega_m/\kappa$ tends to infinity comes from the second contribution proportional to equilibrium thermal occupation $n_{th}$. Due to the small ratio $\gamma_m/\kappa$ and a BS resonant coupling used for cooling, the third contribution can be totally neglected.
Such identification of different quantum noise origins for stabilized thermal phonon number is not so explicit with the other approaches. The earliest approach to quantum optomechanical cooling borrows the picture for cooling of trapped ions and for cavity assisted laser cooling of atomic and molecular motion. Similar to that picture, the mechanical oscillator is modeled as a quantum noise spectrometer responding to the cavity field [@supp3], which is phenomenologically treated as a reservoir, to derive an effective master equation $$\begin{aligned}
\dot{\rho}_m(t)&=&-i[\omega_m\hat{b}^\dagger\hat{b},\rho_m(t)]-(\gamma_m (n_{th}+1)+A_{-})\big(\hat{b}^\dagger\hat{b}\rho_m(t)+\rho_m(t)\hat{b}^\dagger\hat{b}-2\hat{b}\rho_m(t)
\hat{b}^\dagger\big)\nonumber\\
&-&(\gamma_m n_{th}+A_+)\big(\hat{b}\hat{b}^\dagger\rho_m(t)+\rho_m(t)\hat{b}\hat{b}^\dagger-2\hat{b}^\dagger\rho_m(t)\hat{b}\big)
\label{reduced}\end{aligned}$$ for the mechanical oscillator’s quantum state $\rho_m$. The transition rate $A_{\pm}=g^2|\alpha|^2\frac{2\kappa}{\kappa^2+(\Delta\pm \omega_m)^2}$ in the equation is decided by a time-independent amplitude $\alpha$ of the cavity mean field [@c1; @c2], which is approximated with a zeroth-order term in the solution given by Eq. (\[series\]) or a steady-state field in a cavity of fixed size as in Fig. S-1(a). The corresponding equation of motion for the averaged mechanical mode $\langle \hat{b}(t)\rangle=\mbox{Tr}\{\hat{b}\rho_m(t)\}$ reads $$\begin{aligned}
\frac{d}{dt}\langle \hat{b}\rangle=-i\omega_m \langle \hat{b}\rangle-\{\gamma_m+(A_{-}-A_+)\}\langle \hat{b}\rangle,
\label{b-mode}\end{aligned}$$ implying that the mechanical oscillator will gradually become motionless after a time $t\sim 1/(\gamma_m+\Gamma_{opt})$ where $\Gamma_{opt}=A_{-}-A_+$. In the weak coupling regime with $\gamma_m+\Gamma_{opt} \ll \kappa$, this will lag far behind the time $t\sim 1/\kappa$ for the cavity field to reach the assumed steady-state value $\alpha=\langle \hat{a}\rangle=E/(i\Delta+\kappa)$. The Heisenberg-Langevin equations (such as those for the coupled cavity field fluctuation and the mechanical mode in [@rev]) derived previously with the steady-state value $\alpha$ also reflect this notion of asynchronous evolution, which is an approximation for the actual quantum optomechanical cooling processes. For instance, shortly after the moment of turning on the driving laser, the mechanical oscillator in a thermal state \[the initial state $\rho_m(0)$ for Eq. (\[reduced\])\] cannot immediately have the transition rate $A_{\pm}$ proportional to the steady-state value $|\alpha|^2$, because the cavity field has not evolved to that assumed steady state yet. Even the evolution in Fig. S-1(a) will take time to become steady, though the time scale is much shorter than that for the process depicted by Eq. (\[b-mode\]). The quantum state $\rho_m$ of the mechanical mode will be definitely changed under the coupling with a developing cavity field before it arrives at any possible stable evolution or steady state, so that the details of the transient period become relevant to the cooling result.
In terms of the motion of two coupled harmonic oscillators (as described by the averaged Eqs. (\[e1\]) and (\[e2\]) for a weakly coupled system $g/\omega_m\ll 1$) and the dynamics of the associated system-mode fluctuations, the pictures for the different treatments of quantum optomechanical cooling are summarized as below. Here the described motions are from the same static initial condition for the oscillators, i.e. the expectation values $\langle \hat{a}\rangle=\langle \hat{b}\rangle=0$ determined from the cavity vacuum state and mechanical thermal state at $t=0$.
—(A) the effective approach based on steady states:
$\bullet$ Oscillator A representing the cavity field first stops at $t\sim 1/\kappa$, and then oscillator B representing the mechanical vibration will stop after $t\sim 1/(\gamma_m+\Gamma_{opt})$ to end a cooling process.
$\bullet$ An alternative picture is to consider both of the fluctuations $\delta\hat{a}=\hat{a}-\alpha$ and $\delta\hat{b}=\hat{b}-\beta$, where the time-independent steady cavity field amplitude $\alpha$ and steady mechanical displacement $\beta$ are found from the averaged form of Eq. (\[e2\]). These fluctuations have been used to interpret optomechanical cooling [@ck] and optomechanical entanglement [@supp4]. According to the linearized dynamical equations to govern the evolutions of the fluctuations driven by the noises, which are derived with the steady-state values $\alpha$ and $\beta$, the relevant physical quantities (such as $\langle \delta \hat{x}_m^2\rangle$ and $\langle \delta \hat{p}_m^2\rangle$) will inevitably evolve to time-independent values in the end.
—(B) our dynamical approach:
$\bullet$ The coupled oscillators A and B start moving due to an external force. Then they will gradually turn into stable oscillations together, unlike a process to trap the mechanical oscillator’s motion to the ground state in an external simple harmonic potential so that $\langle\hat{x}_m \rangle=\langle \hat{p}_m\rangle=0$. All finally realized physical quantities in a cooling process are determined in the transient period toward the dynamical stability of oscillators A and B.
$\bullet$ In terms of the fluctuation $\delta\hat{b}(t)=\hat{b}(t)-\langle \hat{b}(t)\rangle$ around the time-dependent expectation value $\langle \hat{b}(t)\rangle$ that can give the motions in Fig. S-2, our concerned thermal occupation $\langle\hat{b}^\dagger\hat{b}\rangle-|\langle\hat{b}\rangle|^2=\langle (\hat{b}^\dagger-\langle \hat{b}\rangle^\ast)(\hat{b}-\langle \hat{b}\rangle)\rangle$ can also be expressed as $\langle \delta \hat{b}^\dagger\delta \hat{b}\rangle$. The contributions from the different noise drive terms to the fluctuation $\delta \hat{b}(t)$ will lead to the stable oscillations as those in Fig. S-3 (when viewed with small scales), so there will not be exactly time-independent steady values. In a cooling process, the quantity $\langle \delta \hat{b}^\dagger\delta \hat{b}(t)\rangle$ will be lowered, while the purity of the macroscopic quantum state $\rho_m(t)$ of the mechanical mode will become higher. Note that the thermal phonon number $\langle \delta \hat{b}^\dagger\delta \hat{b}(t)\rangle$ also includes the contribution from the system-operator part as Eq. (6) in the main text, which will asymptotically tend to zero by the time $t\sim t_s$ to reach the dynamical stability.
Two qualitative differences from our approach will still exist even when the system dynamics is linearized by shifting $\hat{a}(t)\rightarrow \hat{a}(t)+\alpha(t)$ with a time-dependent mean field $\alpha(t)$. (1) The mean-field solution $\alpha(t)$ of the nonlinear dynamical equations can become undetermined, for example, in the regimes around bistability. (2) The magnitude of $\alpha(t)$ should be much higher than that of the cavity field fluctuation so that a linearized Hamiltonian like that in Eq. (\[lnh\]) \[with the steady-state value $\alpha$ in the equation replaced by an evolving one $\alpha(t)$\] can be obtained. The difference in the linearized Hamiltonians indicates that, even if such simulation with an evolving $\alpha(t)$ could be performed, the simulated evolution of thermal occupation or thermal phonon number will not be similar to the prediction by our quantum dynamical approach.
The results of cooling according to the distinct pictures can have big difference. From the steady-state solution $\dot{\rho}_m(t)=0$ of Eq. (\[reduced\]), the thermal occupation at a final steady state will be found as $$\begin{aligned}
n_{m,f}=\frac{\Gamma_{opt}n_m^0+2\gamma_m n_{th}}{\Gamma_{opt}+2\gamma_m},
\label{weak-1}\end{aligned}$$ where $$\begin{aligned}
\Gamma_{opt}=\frac{2g^2|\alpha|^2}{\kappa}\frac{1}{1+(\frac{\kappa}{2\omega_m})^2}
\label{rate}\end{aligned}$$ and $n^0_m=(\kappa/2\omega_m)^2$, $|\alpha|^2=(E/\kappa)^2/(1+(\Delta/\kappa)^2)$ [@c2]. The coefficients in these formulas have slight variations from those in the previous literature, due to our notation for the damping terms ($-\kappa \hat{a}$, $-\gamma_m \hat{b}$ in the dynamical equations) as compared with the corresponding ones $-\frac{\kappa}{2}\hat{a}$ and $-\frac{\gamma_m}{2} \hat{b}$ in [@c1; @c2], and our definition of detuning $\Delta$ also differs by an opposite sign. Similar steady-state occupation can also be found for other systems (see, e.g. [@supp5]). In Eq. (\[weak-1\]) the first part from the cavity noise’s back-action is almost independent of a cooling rate $\Gamma_{opt}\gg \gamma_m$ and is thus beyond the control by the drive power. A contrasting feature of our results is that the thermal phonon number can be totally controlled by a parameter $J$ proportional to the drive intensity $E$, since each contribution to the finally stabilized thermal occupation shown in Fig. S-3 is decided by how soon its evolution becomes stable.
The explicit comparison between a result from Eq. (\[weak-1\]) and the prediction by our dynamical approach is shown in Fig. S-4(a). Given a sufficiently small ratio $\gamma_m/\kappa$ or a sufficiently high quality factor of the mechanical oscillator as in the figure, our predicted thermal phonon number to be stabilized changes slightly with increased sideband resolution $\omega_m/\kappa$, in contrast to an obvious optimum sideband resolution $\omega_m/\kappa$ found with Eq. (\[weak-1\]). Another obvious difference is in an assumed situation of initial zero temperature, where the residual phonon number is known as pure quantum limit or quantum back-action limit. This contribution as the first term in Eq. (7) of the main text increases significantly with the parameter $J$. A previous development from the above-mentioned treatment, on the other hand, predicts the value in the strong coupling regime as [@c3] $$\begin{aligned}
n_{m,f}=\frac{\kappa^2}{4\omega_m^2}+\frac{g^2|\alpha|^2}{2\omega_m^2},
\label{p}\end{aligned}$$ where a zero thermal occupation for the cavity reservoir is assumed, and the different coefficients from those in [@c3] are due to the same reason for Eqs. (\[weak-1\]) and (\[rate\]). Our result, the solid curve in Fig. S-4(b), indicates a much more significant effect of the noise from the cavity reservoir.
The last issue is how to know the cooling result. The correlation function $\langle \hat{i}(t+\tau)\hat{i}(t)\rangle-\langle \hat{i}(t+\tau)\rangle\langle \hat{i}(t)\rangle$ [@supp6], where $\hat{i}(t)=\hat{a}_{out}(t)+\hat{a}^\dagger_{out}(t)\rangle$ and $\hat{a}_{out}(t)=E(t)/\sqrt{\kappa}+\sqrt{\kappa}\hat{a}(t)$, after the time $t \sim t_s$ for reaching the dynamically stable phase is in one-to-one correspondence with the finally achieved thermal occupation or thermal phonon number. It primarily consists of the following correlation of the cavity mode: $$\begin{aligned}
{\cal C}(\tau)& = & \langle \hat{a}^\dagger(t+\tau)\hat{a}(t)\rangle- \langle \hat{a}^\dagger(t+\tau)\rangle \langle \hat{a}(t)\rangle.\end{aligned}$$ The Fourier transform of this function, ${\cal C}(\omega)=\int d\tau {\cal C}(\tau)e^{i\omega \tau}$, constitutes the main component in a measured noise spectrum. Such correlation function is only determined by the noise drive term in Eq. (\[dynamics\]). If the contribution from the coherent drive term in the equation is also included, the Fourier transform of $\langle \hat{a}^\dagger(t+\tau)\hat{a}(t)\rangle$ will have an extra delta-function term from the finally stable cavity field oscillation as the background (see, e.g. [@c1]). The noise spectra for the cooling processes taking place close to the limiting boundary (the dashed curve in Fig. 4 of the main text) are illustrated in Fig. S-5. The single-peak one in Fig. S-5(a) reflects the thermometry results in the past experiments.
[99]{} J. D. Teufel, J. W. Harlow, C. A. Regal, and K. W. Lehnert, “Dynamical backaction of microwave fields on a nanomechanical oscillator", Phys. Rev. Lett. 101, 197203 (2008). S. Gröblacher, J. B. Hertzberg, M. R. Vanner, S. Gigan, K. C. Schwab, and M. Aspelmeyer, “Demonstration of an ultracold micro-optomechanical oscillator in a cryogenic cavity", Nature Phys. 5, 485 (2009). A. Schliesser, O. Arcizet, R. Riviére, G. Anetsberger, and T. J. Kippenberg, “Resolved sideband cooling and position measurement of a micromechanical oscillator close to the Heisenberg uncertainty limit", Nature Phys. 5, 509 (2009). Y.-S. Park and H. Wang, “Resolved-sideband and cryogenic cooling of an optomechanical resonator", Nature Phys. 5, 489 (2009). A. D. O’Connell, M. Hofheinz, M. Ansmann, R. C. Bialczak, M. Lenander, E. Lucero, M. Neeley, D. Sank, H. Wang, M. Weides, J. Wenner, J. M. Martinis, and A. N. Cleland, “Quantum ground state and single-phonon control of a mechanical resonator", Nature 464, 697 (2010). T. Rocheleau, T. Ndukum, C. Macklin, J. B. Hertzberg, A. A. Clerk, and K. C. Schwab, “Preparation and detection of a mechanical resonator near the ground state of motion", Nature 463, 72 (2010). R. Riviére, S. Delèglise, S. Weis, E. Gavartin, O. Arcizet, A. Schliesser, and T. J. Kippenberg, “Optomechanical sideband cooling of a micromechanical oscillator close to the quantum ground state," Phys. Rev. A 83, 063835 (2011). J. D. Teufel, T. Donner, D. Li, J. H. Harlow, M. S. Allman, K. Cicak, A. J. Sirois, J. D. Whittaker, K. W. Lehnert, and R. W. Simmonds, “Sideband Cooling Micromechanical Motion to the Quantum Ground State", Nature 475, 359 (2011). J. Chan, T. P. Mayer Alegre, A. H. Safavi-Naeini, J. T. Hill, A. Krause, S. Gröblacher, M. Aspelmeyer, and O. Painter, “Laser cooling of a nanomechanical oscillator into its quantum ground state", Nature 478, 89 (2011). E. Verhagen, S. Delèglise, S. Weis, A. Schliesser, and T. J. Kippenberg, “Quantum-coherent coupling of a mechanical oscillator to an optical cavity mode", Nature 482, 63 (2012). A. H. Safavi-Naeini, J. Chan, J. T. Hill, T. P. Mayer Alegre, A. Krause, and O. Painter, “Observation of Quantum Motion of a Nanomechanical Resonator", Phys. Rev. Lett. 108, 033602 (2012). R. W. Peterson, T. P. Purdy, N. S. Kampel, R. W. Andrews, P.-L. Yu, K. W. Lehnert, and C. A. Regal, “Laser Cooling of a Micromechanical Membrane to the Quantum Backaction Limit", Phys. Rev. Lett. 116, 063601 (2016). J. B. Clark, F. Lecocq, R. W. Simmonds, J. Aumentado, and J. D. Teufel, “Sideband cooling beyond the quantum backaction limit with squeezed light", Nature 541, 191 (2017). C. Fabre, M. Pinard, S. Bourzeix, A. Heidmann, E. Giacobino, and S. Reynaud, “Quantum-noise reduction using a cavity with a movable mirror", 49, 1337 (1994). S. Mancini and P. Tombesi, “Quantum noise reduction by radiation pressure", Phys. Rev. A 49, 4055 (1994). I. Wilson-Rae, N. Nooshi, W. Zwerger, and T. J. Kippenberg, “Theory of Ground State Cooling of a Mechanical Oscillator Using Dynamical Backaction", Phys. Rev. Lett. 99, 093901 (2007). F. Marquardt, J. P. Chen, A. A. Clerk, and S. M. Girvin, “Quantum Theory of Cavity-Assisted Sideband Cooling of Mechanical Motion", Phys. Rev. Lett. 99, 093902 (2007). C. Genes, D. Vitali, P. Tombesi, S. Gigan, and M. Aspelmeyer, “Ground-state cooling of a micromechanical oscillator: Comparing cold damping and cavity-assisted cooling schemes", Phys. Rev. A 77, 033804 (2008). A. Dantan, C. Genes, D. Vitali, and M. Pinard, “Self-cooling of a movable mirror to the ground state using radiation pressure", Phys. Rev. A 77, 011804(R) (2008). J. M. Dobrindt, I. Wilson-Rae, and T. J. Kippenberg, “Parametric Normal-Mode Splitting in Cavity Optomechanics", Phys. Rev. Lett. 101, 263602 (2008). I. Wilson-Rae, N. Nooshi, J. Dobrindt, T. J. Kippenberg, and W. Zwerger, “Cavity-Assisted Back Action Cooling of Mechanical Resonators", New J. Phys. 10, 095007 (2008). P. Rabl, C. Genes, K. Hammerer, and M. Aspelmeyer, “Phase-noise induced limitations on cooling and coherent evolution in optomechanical systems", 80, 063819 (2009). Y.-C. Liu, Y.-F. Xiao, X. Luan, and C. W. Wong, “Dynamic Dissipative Cooling of a Mechanical Resonator in Strong Coupling Optomechanics", Phys. Rev. Lett. 110, 153606 (2013). M. Aspelmeyer, T. J. Kippenberg, and F. Marquardt, “Cavity optomechanics", Rev. Mod. Phys. 86, 1391 (2014). See Supplemental Material at \[URL will be inserted by publisher\] for the detailed discussion, which includes Refs. \[26\]-\[32\]. C. W. Gardiner and P. Zoller, [*Quantum Noise*]{}, Springer-Verlag, Berlin, Heidelberg (2000). G. J. Milburn and M. J. Woolley, “An introduction to quantum optomechanics", Acta Physica Slovaca 61, 483 (2011). F. Marquardt, J. G. E. Harris, and S. M. Girvin, “Dynamical Multistability Induced by Radiation Pressure in High-Finesse Micromechanical Optical Cavities", Phys. Rev. Lett. 96, 103901 (2006). A. A. Clerk, M. H. Devoret, S. M. Girvin, F. Marquardt, and R. J. Schoelkopf, “Introduction to Quantum Noise, Measurement and Amplification", Rev. Mod. Phys. 82, 1155 (2010). D. Vitali, S. Gigan, A. Ferreira, H. R. Böhm, P. Tombesi, A. Guerreiro, V. Vedral, A. Zeilinger, and M. Aspelmeyer, “Optomechanical Entanglement between a Movable Mirror and a Cavity Field", Phys. Rev. Lett. 98, 030405 (2007). S. Zippilli, G. Morigi, and A. Bachtold, “Cooling Carbon Nanotubes to the Phononic Ground State with a Constant Electron Current", Phys. Rev. Lett. 102, 096804 (2009). H. A. Haus, [*Electromagnetic Noise and Quantum Optical Measurements*]{}, Springer-Verlag, Berlin, Heidelberg (2000). S. Gröblacher, K. Hammerer, M. R. Vanner, and M. Aspelmeyer, “Observation of strong coupling between a micromechanical resonator and an optical cavity field", Nature 460, 724 (2009). M. G. A. Paris, F. Illuminati, A. Serafini, and S. De Siena, “Purity of Gaussian states: Measurement schemes and time evolution in noisy channels", Phys. Rev. A 68, 012314 (2003). C. K. Law, “Interaction between a moving mirror and radiation pressure: A Hamiltonian formulation", 51, 2537 (1995). N. Goldman, G Juzeliunas, P. Öhberg, and I. B. Spielman, “Light-induced gauge fields for ultracold atoms", Rep. Prog. Phys. 77, 126401 (2014). L. Dong, C. Zhu, and H. Pu, “Photon-Induced Spin-Orbit Coupling in Ultracold Atoms inside Optical Cavity", Atoms 3, 182 (2015). B. He, L. Yang, and M. Xiao, “Dynamical phonon laser in coupled active-passive microresonators", 94, 031802(R) (2016).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we present TailorNet, a neural model which predicts clothing deformation in 3D as a function of three factors: pose, shape and style (garment geometry), while retaining wrinkle detail. This goes beyond prior models, which are either specific to one style and shape, or generalize to different shapes producing smooth results, despite being style specific. Our hypothesis is that (even non-linear) combinations of examples smooth out high frequency components such as fine-wrinkles, which makes learning the three factors jointly hard. At the heart of our technique is a decomposition of deformation into a high frequency and a low frequency component. While the low-frequency component is predicted from pose, shape and style parameters with an MLP, the high-frequency component is predicted with a mixture of shape-style specific pose models. The weights of the mixture are computed with a narrow bandwidth kernel to guarantee that only predictions with similar high-frequency patterns are combined. The style variation is obtained by computing, in a canonical pose, a subspace of deformation, which satisfies physical constraints such as inter-penetration, and draping on the body. TailorNet delivers 3D garments which retain the wrinkles from the physics based simulations (PBS) it is learned from, while running more than $1000$ times faster. In contrast to classical PBS, TailorNet is easy to use and fully differentiable, which is crucial for computer vision and learning algorithms. Several experiments demonstrate TailorNet produces more realistic results than prior work, and even generates temporally coherent deformations on sequences of the AMASS [@mahmood2019amass] dataset, despite being trained on static poses from a different dataset. To stimulate further research in this direction, we will make a dataset consisting of 55800 frames, as well as our model publicly available at <https://virtualhumans.mpi-inf.mpg.de/tailornet/>.'
author:
- |
Chaitanya PatelZhouyingcheng LiaoGerard Pons-Moll\
Max Planck Institute for Informatics, Saarland Informatics Campus, Germany\
[{cpatel, zliao, gpons}@mpi-inf.mpg.de]{}
title: 'TailorNet: Predicting Clothing in 3D as a Function of Human Pose, Shape and Garment Style '
---
oldmaketitlemaketitle
Introduction {#sec:introduction}
============
Related Work {#sec:related}
============
Garment Model Aligned with SMPL {#sec:garmentmodel}
===============================
Method {#sec:method}
======
Dataset {#sec:dataset}
=======
Experiments {#sec:experiments}
===========
Discussion and Conclusion {#sec:conclusion}
=========================
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we answer affirmatively the problem proposed by A. Selvitella in his paper “Nondegenracy of the ground state for quasilinear Schrödinger Equations” (see Calc. Var. Partial Differ. Equ., [**53**]{} (2015), pp 349-364): every ground state of equation $$\begin{aligned}
-\Delta u-u\Delta |u|^2+\omega u-|u|^{p-1}u=0&&\text{in }\mathbb{R}^N\end{aligned}$$ is nondegenerate for $1<p<3$, where $\omega>0$ is a given constant and $N\ge1$. We also derive further properties on the linear operator associated to ground states of above equation.
address: 'University of Jyväskylä, Department of Mathematics and Statistics, P.O. Box 35, FI-40014 University of Jyväskylä, Finland.'
author:
- 'Chang-Lin Xiang'
title: Remarks on Nondegeneracy of Ground States for Quasilinear Schrödinger Equations
---
[ ]{}
Introduction and main result
============================
Consider the quasilinear Schrödinger equation $$\begin{aligned}
i\pa_{t}U=-\De U-U\De|U|^{2}-|U|^{p-1}U & & \text{in }\R^{N}\times\R_{+},\label{eq: Quasili-Schrodinger eq.}\end{aligned}$$ where $U:\R^{N}\times\R_{+}\to\C$ is the wave function, $i$ is the imaginary unit, $N\ge1$ and $p>1$. Equation (\[eq: Quasili-Schrodinger eq.\]) arises in various domains of physics, such as superfluid film equation in plasma physics. More physical background of equation (\[eq: Quasili-Schrodinger eq.\]) can be found in e.g. Colin et al. [@Colin-Jeanjean-; @Squassina-2010] and the references therein. Equation (\[eq: Quasili-Schrodinger eq.\]) has been studied extensively in the literature, see e.g. [@Colin-2002; @Colin-Jeanjean-; @Squassina-2010; @Kenig-Ponce-Vega-2004; @Lange-Poppenberg-Teismann-1999; @Poppenberg-2001; @Poppenberg-Schmitt-; @Wang-2002] and the references therein. A special class of solutions to equation (\[eq: Quasili-Schrodinger eq.\]) that represent particles at rest is the so called standing waves, that is, solutions of the form $$U(x,t)=e^{i\om t}u(x),$$ where $\om>0$ is a given constant which stands for the time frequency, and $u:\R^{N}\to\C$ is a complex valued function that is independent of time $t\in\R_{+}$. It is elementary to verify that if $U(x,t)=e^{i\om t}u(x)$ is a standing wave, then $u$ solves the stationary equation $$\begin{aligned}
-\Delta u-u\Delta|u|^{2}+\omega u-|u|^{p-1}u=0 & & \text{in }\R^{N}.\label{eq: Object equ.}\end{aligned}$$
Equation (\[eq: Object equ.\]) is known [@Colin-Jeanjean-; @Squassina-2010] as the Euler-Lagrange equation of the energy functional $\E_{\om}:\X_{\C}\to\R$ defined as $${\cal E}_{\om}(u)=\frac{1}{2}\int_{\R^{N}}|\na u|^{2}{{\rm d}}x+\int_{\R^{N}}|u|^{2}|\na|u||^{2}{{\rm d}}x+\frac{\om}{2}\int_{\R^{N}}|u|^{2}{{\rm d}}x-\frac{1}{p+1}\int_{\R^{N}}|u|^{p+1}{{\rm d}}x,$$ where $\X_{\C}$ is the function space given by $$\X_{\C}=\left\{ u\in H^{1}(\R^{N}):\int_{\R^{N}}|u|^{2}|\na|u||^{2}{{\rm d}}x<\wq\right\} .$$ Here we assume throughout the paper that $1<p<p_{\max}$ holds, where the critical exponent $p_{\max}$ is defined as $$p_{\max}\equiv\begin{cases}
\frac{3N+2}{N-2} & \text{if }N\ge3\\
\wq & \text{if }N=1,2.
\end{cases}\label{eq: p-max}$$ Technically speaking, the condition that $p$ be strictly less than $p_{\max}$ ensures that the power nonlinearity in equation (\[eq: Object equ.\]) is $\X_{\C}$-subcritical. Indeed, by a simple application of Sobolev embedding theorems, we infer that $\X_{\C}$ is continuously embedded into $L^{p+1}(\R^{N})$ for $1<p<\wq$ if $N=1,2$ and $1<p\le(3N+2)/(N-2)$ if $N\ge3$.
In view of the variational structures of equation (\[eq: Object equ.\]), critical point theory has been devoted to find solutions for equation (\[eq: Object equ.\]). Here, as in Colin et al. [@Colin-Jeanjean-; @Squassina-2010], a function $u\in\X_{\C}$ is said to be a solution to equation (\[eq: Object equ.\]), if for any function $\phi\in C_{0}^{\wq}(\R^{N})$, the space of smooth functions in $\R^{N}$ with compact support, there holds $$\text{{\rm Re}}\int_{\R^{N}}\Big(\na u\cdot\na\bar{\phi}+\na|u|^{2}\cdot\na(u\bar{\phi})+\om u\bar{\phi}-|u|^{p-1}u\bar{\phi}\Big){{\rm d}}x=0$$ (here $\text{{\rm Re}}z$ is the real part of $z\in\C$). The existence of solutions to equation (\[eq: Object equ.\]) is now well known, see e.g. [@Colin-Jeanjean-2004; @Colin-Jeanjean-; @Squassina-2010; @Liu-Wang-Wang-2003; @Liu-Wang-Wang-2004; @Liu-; @Wang-2003] and the references therein.
In this paper, we consider ground state to equation (\[eq: Object equ.\]). Following the convention of Colin et al. [@Colin-Jeanjean-; @Squassina-2010] (see also Selvitella [@Selvitella-2011; @Selvitella-2015]), we say that a solution $u\in\X_{\C}$ to equation (\[eq: Object equ.\]) is a *ground state*, if $u$ satisfies $$\E_{\om}(u)=\inf\left\{ \E_{\om}(v):v\in\X_{\C}\text{ is a nontrivial solution to equation (\ref{eq: Object equ.})}\right\} .$$ We remark that the notion of ground state here is different from that defined in [@Chang; @et; @al-2007; @Frank-Lenzmann-2013; @Frank-Lenzmann-Silvestre-2013; @Kwong1989; @Lenzmann-2009]. We are concerned about the nondegeneracy (see below) of ground states. Before proceeding further, let us summarize the existence result of ground states to equation (\[eq: Object equ.\]) together with a list of basic properties for later use.
\[thm: properties of GS\] Assume that $1<p<p_{\max}$ with $p_{\max}$ defined as in (\[eq: p-max\]). Then for any given constant $\om>0$, there exists a ground state to equation (\[eq: Object equ.\]). Moreover, for any ground state $u\in\X_{\C}$ to equation (\[eq: Object equ.\]), there exist a constant $\theta\in\R$, a decreasing positive function $v:[0,\wq)\to(0,\wq)$ and a point $x_{0}\in\R^{N}$ such that $u$ is of the form $$\begin{aligned}
u(x)=e^{i\theta}v(|x-x_{0}|) & & \text{for }x\in\R^{N}.\end{aligned}$$ Furthermore, the following properties hold.
\(1) (Smoothness) $u\in C^{\wq}(\R^{N})$.
\(2) (Decay) For any multi-index $\al\in\N^{N}$ with $|\al|\ge0$, there exist positive constants $C_{\al}>0$ and $\de_{\al}>0$ such that $$\begin{aligned}
|\pa^{\al}u(x)|\le C_{\al}\exp(-\de_{\al}|x|) & & \text{for all }x\in\R^{N}.\end{aligned}$$
\(3) (Uniqueness) In the case $N=1$, the ground states to equation (\[eq: Object equ.\]) is unique up to phase and translation. In particular, there exists a unique positive even ground state for equation (\[eq: Object equ.\]).
For a complete proof of Theorem \[thm: properties of GS\], we refer to Colin et al. [@Colin-Jeanjean-; @Squassina-2010] and Selvitella [@Selvitella-2011].
In this paper, our aim is to study nondegeneracy of ground states for equation (\[eq: Object equ.\]). The motivation comes from the fact that the nondegeneracy of ground states for equation (\[eq: Object equ.\]) plays an important role when studying the existence of concentrating solutions in the semiclassical regime. We refer the readers to Selvitella [@Selvitella-2015] for more applications of nondegeneracy results. We also follow the convention of Selvitella [@Selvitella-2015] (see also Ambrosetti and Malchiodi [@Ambrosetti-Malchiodi-book]) and define nondegeneracy of ground states for equation (\[eq: Object equ.\]) as follows.
\[def: nondegeneracy\]Let $u\in\X_{\C}$ be a ground state of equation (\[eq: Object equ.\]). We say that $u$ is nondegenerate if the following properties hold:
\(1) (ND) ${\text{\rm Ker}}\E_{\om}^{\prime\prime}(u)=\text{{\rm span}}\left\{ iu,\pa_{x_{1}}u,\cdots,\pa_{x_{N}}u\right\} $;
\(2) (Fr) $\E_{\om}^{\prime\prime}(u)$ is an index 0 Fredholm map.
The first result on nondegeneracy of ground states for equation (\[eq: Object equ.\]) was obtained by Selvitella [@Selvitella-2011] in a perturbative setting, where uniqueness of ground states for equation (\[eq: Object equ.\]) was also considered. In his quite recent paper [@Selvitella-2015], Selvitella proved, under the assumption $$p\ge3,$$ that every ground state of equation (\[eq: Object equ.\]) is nondegenerate in the sense of Definition \[def: nondegeneracy\] above, see Theorem 1.2 of [@Selvitella-2015]. Selvitella also commented (see Remark 1.3 of [@Selvitella-2015]) that his nondegeneracy result could also be true for the case $1<p<3$. However, his approach can not handle this case. In this paper, we give an affirmative answer to his question. We obtain the following result.
\[thm: main result\] For $1<p<3$, every ground state of equation (\[eq: Object equ.\]) is nondegenerate in the sense of Definition \[def: nondegeneracy\] above.
We remark that our argument is applicable to the whole range $1<p<p_{\max}$.
As already remarked by Selvitella (see Remark 1.3 of [@Selvitella-2015]), except Proposition 3.10 of [@Selvitella-2015] that requires him to assume $p\ge3$, all the rest of his arguments can be applied to the whole range $1<p<p_{\max}$ to prove Theorem \[thm: main result\]. So in this paper, we will follow the line of Selvitella [@Selvitella-2015] to prove Theorem \[thm: main result\]. However, since his approach can not handle the whole range $1<p<p_{\max}$, we will apply a different idea from that of Selvitella [@Selvitella-2015]. Precisely, let $u$ be a positive radial ground state of equation (\[eq: Object equ.\]). Define the linear operator $\L_{+}$ associated to $u$ by $$\L_{+}\eta=-\De\eta-2u\De(u\eta)+\om\eta-(\De u^{2}+pu^{p-1})\eta.$$ We note that $\L_{+}$ is a self-adjoint operator acting on $L^{2}(\R^{N})$ with form domain $\X_{\C}$ and operator domain $H^{2}(\R^{N})$. It turns out that the key to prove Theorem \[thm: main result\] is to show that $\L_{+}$ satisfies $${\text{\rm Ker}}\L_{+}=\text{{\rm span}}\left\{ \pa_{x_{1}}u,\cdots,\pa_{x_{N}}u\right\} .\label{eq: Kernel 1.1}$$ In the approach of Selvitella [@Selvitella-2015] to (\[eq: Kernel 1.1\]), ordinary differential equation analysis plays a central role, in which the assumption $p\ge3$ is required. To prove (\[eq: Kernel 1.1\]) for $p$ in the whole range $1<p<p_{\max}$, we will use a spectrum analysis to the operator $\L_{+}$. In this way, we obtain deeper results on the operator $\L_{+}$ than (\[eq: Kernel 1.1\]).
Our idea comes from the spectrum analysis of Chang et al. [@Chang; @et; @al-2007], even through we can not use their refined arguments to derive (\[eq: Kernel 1.1\]) directly. Roughly speaking, Chang et al. [@Chang; @et; @al-2007] considered the following problem. Let $Q$ the unique positive radial solution in $H^{1}(\R^{N})$ to the equation $$\begin{aligned}
-\De Q+\om Q-Q^{q}=0 & & \text{in }\R^{N},\label{eq: NLS}\end{aligned}$$ where $1<q<(N+2)/(N-2)$. Here we assume $N\ge3$ for simplicity. Chang et al. [@Chang; @et; @al-2007] studied the spectrum of the linear operator ${\cal A}_{+}$ around $Q$ given by $${\cal A}_{+}\eta=-\De\eta+\om\eta-qQ^{q-1}\eta,\label{eq: operator of Chang}$$ acting on $L^{2}(\R^{N})$ with form domain $H^{1}(\R^{N})$ and operator domain $H^{2}(\R^{N})$. For the importance of the spectrum of ${\cal A}_{+}$, we refer the readers to Chang et al. [@Chang; @et; @al-2007]. We remark that Chang et al. [@Chang; @et; @al-2007] studied far more than ${\cal A}_{+}$ in their work. We will give a brief comparison between the two self-adjoint operators ${\cal A}_{+}$ and $\L_{+}$ below. We also refer the readers to [@Frank-Lenzmann-2013; @Frank-Lenzmann-Silvestre-2013; @Lenzmann-2009] for spectrum analysis for linearized operators around ground states of nonlocal problems.
It seems that the spectrum $\si(\L_{+})$ of $\L_{+}$ has not been studied in the literature. According to the analysis in next section, we obtain the following properties for the spectrum of $\L_{+}$:
\(1) the continuous spectrum of $\L_{+}$ is contained in $[\om,\wq)$ (see Lemma \[lem: essential spectrum\]);
\(2) $\inf\si(\L_{+})<0$ is the first eigenvalue of $\L_{+}$ and is simple (see Lemma \[lem: first eigenvalue is simple and negative\]);
\(3) as a consequence of (1) and (2), $0$ belongs to the discrete spectrum of $\L_{+}$ and is not the first eigenvalue of $\L_{+}$.
We will give the proof of Theorem \[thm: main result\] together with above properties of $\si(\L_{+})$ in the next section. Before we close this section, let us address some differences between the two self-adjoint operators ${\cal A}_{+}$ defined as in (\[eq: operator of Chang\]) and $\L_{+}$.
First we point out that the second property (2) is not obvious. In fact, even the fact $\inf\si(\L_{+})<0$ is not obvious. For the operator ${\cal A}_{+}$, a simple observation gives that $$\langle{\cal A}_{+}Q,Q\rangle=-(q-1)\int_{\R^{N}}Q^{q+1}{{\rm d}}x<0,$$ which implies $\inf\si(A_{+})<0$. Furthermore, it is standard (see e.g. Lieb and Loss [@Lieb-Loss-book]) to show that $\inf\si({\cal A}_{+})$ is the first eigenvalue of ${\cal A}_{+}$ and is simple. However, in our case, a direct calculation gives us $$\langle\L_{+}u,u\rangle=8\int_{\R^{N}}u^{2}|\na u|^{2}{{\rm d}}x-(p-1)\int_{\R^{N}}u^{p+1}{{\rm d}}x.$$ We will confirm that $\langle\L_{+}u,u\rangle<0$ in the case $N=2$, based on a Pohozaev type identity, see the end of next section. While in higher dimensions $N\ge3$, it is still not clear from above expression whether $\langle\L_{+}u,u\rangle$ is negative or not. Hence we can not infer that $\inf\si(\L_{+})<0$ holds by such a simple observation as above. Similarly, due to the presence of the quasilinear term $-2u\De(u\eta)$ in $\L_{+}$, it is not obvious as well that a nonnegative eigenfunction of $\L_{+}$ is in fact positive everywhere.
Second, we point out that we do not known whether $0$ is the second eigenvalue of $\L_{+}$ or not. Then we can not give exact estimates on the numbers of nodal domains of radial functions $v$ with $v\in{\text{\rm Ker}}\L_{+}$. Thus we can not use the arguments of Chang et al. [@Chang; @et; @al-2007] directly (see the proof of Lemma 2.1 of Chang et al. [@Chang; @et; @al-2007]). As to the operator ${\cal A}_{+}$, it is known (see e.g. Chang et al. [@Chang; @et; @al-2007]) that $0$ is the second eigenvalue of ${\cal A}_{+}$. This is due to fact that, by uniqueness, $Q$ is also a minimizer (up to rescaling) of the ’Weinstein’ functional $$\begin{aligned}
W(f)=\frac{\|\na f\|_{2}^{(q-1)N/2}\|f\|_{2}^{(N+2-(N-2)q)/2}}{\|f\|_{q+1}^{q+1}}, & & f\in H^{1}(\R^{N}),f\not\equiv0.\end{aligned}$$ In our case, except the case $N=1$ (see Colin et al. [@Colin-Jeanjean-; @Squassina-2010]), the uniqueness of positive solutions to equation (\[eq: Object equ.\]) is unknown in general when $N\ge2$. In fact, even the uniqueness of ground states to equation (\[eq: Object equ.\]) is unknown in general when $N\ge2$. Some partial results on uniqueness were obtained in the literature. Since this is out of the scope of this paper, we refer the interested readers to Selvitella [@Selvitella-2011; @Selvitella-2015] and the references therein.
Our nations are standard. We write $\R_{+}=(0,\wq)$ and denote $\N=\{0,1,2,\cdots\}$ the set of nonnegative integers. For any $1\le s\le\infty$, $L^{s}(\R^{N})$ is the Banach space of complex valued Lebesgue measurable functions $u$ such that the norm $$\|u\|_{s}=\begin{cases}
\left(\int_{\R^{N}}|u|^{s}{{\rm d}}x\right)^{\frac{1}{s}} & \text{if }1\le s<\infty\\
{\text{\rm \,esssup\,}}_{\R^{N}}|u| & \text{if }s=\infty
\end{cases}$$ is finite. A function $u$ belongs to the Sobolev space $H^{k}(\R^{N})$ ($k\in\N$) if $u\in L^{2}(\R^{N})$ and its weak partial derivatives up to order $k$ also belong to $L^{2}(\R^{N})$. We equip $H^{k}(\R^{N})$ with the norm $$\|u\|_{H^{k}}=\sum_{\al\in\N^{N},|\al|\le k}\|\pa^{\al}u\|_{2}.$$ For the properties of the Sobolev functions, we refer to the monograph [@Ziemer]. By abuse of notation, we write $f(x)=f(r)$ with $r=|x|$ whenever $f$ is a radially symmetric function in $\R^{N}$.
Proof of main result
====================
In this section we prove Theorem \[thm: main result\]. Since we deal with the same problem as that of Selvitella [@Selvitella-2015], we will follow the line of Selvitella [@Selvitella-2015]. Similar lines can also be found in e.g. [@Chang; @et; @al-2007; @Frank-Lenzmann-2013; @Frank-Lenzmann-Silvestre-2013; @Lenzmann-2009] and the monograph [@Ambrosetti-Malchiodi-book].
Proof of Theorem \[thm: main result\]
-------------------------------------
Let $u\in\X_{\C}$ be an arbitrary ground state for equation (\[eq: Object equ.\]). By Definition \[def: nondegeneracy\], to prove Theorem \[thm: main result\] we have to show that $\E_{\om}^{\prime\prime}(u)$ satisfies property (ND) and property (Fr). The property (Fr) can be proved by the same argument as that of Selvitella [@Selvitella-2015], since which is applicable to the whole range $1<p<p_{\max}$. So we omit the details.
We focus on the proof of the property (ND), that is, we prove in the following that $${\text{\rm Ker}}\E_{\om}^{\prime\prime}(u)=\text{{\rm span}}\left\{ iu,\pa_{x_{1}}u,\cdots,\pa_{x_{N}}u\right\} .\label{eq: ND}$$
By Theorem \[thm: properties of GS\], every ground state of equation (\[eq: Object equ.\]) can be regarded as a positive, radial and symmetric-decreasing ground state. Hence we assume in the sequel that $u=u(|x|)>0$ is a positive, radial and symmetric-decreasing ground state for equation (\[eq: Object equ.\]). We also assume $N\ge2$ in the sequel. In the case $N=1$ the proof of (\[eq: ND\]) is similar and even simpler. Then the linearized operator $\E_{\om}^{\prime\prime}(u)$ is giving by $$\E_{\om}^{\prime\prime}(u)\xi=-\De\xi-2u\De\left(u\text{Re}\xi\right)+\om\xi-\left(\De u^{2}\right)\xi-(p-1)u^{p-1}\text{Re}\xi-u^{p-1}\xi$$ acting on $L^{2}(\R^{N})$ with form domain $\X_{\C}$ and operator domain $H^{2}(\R^{N})$.
Note that $\E_{\om}^{\prime\prime}(u)$ is not even $\C$-linear. To overcome this difficulty, it is preferable to introduce the linear operator $\L_{+}$ given by $$\L_{+}\eta=-\De\eta-2u\De(u\eta)+\om\eta-(\De u^{2}+pu^{p-1})\eta,\label{operator: real part of linearization}$$ acting on $L^{2}(\R^{N})$ with form domain $\X_{\C}$ and operator domain $H^{2}(\R^{N})$, and the linear operator $\L_{-}$ given by $$\L_{-}\zeta=-\De\zeta+\om\zeta-(\De u^{2}+u^{p-1})\zeta$$ acting on $L^{2}(\R^{N})$ with form domain $H^{1}(\R^{N})$ and operator domain $H^{2}(\R^{N})$. Then for any $\xi\in H^{2}(\R^{N})$ we obtain $$\E_{\om}^{\prime\prime}(u)\xi=\L_{+}\text{Re}\xi+i\L_{-}\text{Im}\xi$$ (here $\text{Im}z$ is the imaginary part of $z\in\C$). Therefore, to prove (\[eq: ND\]), it is sufficient to prove the following result.
\[prop: two kernels\] Let $\L_{+}$ and $\L_{-}$ be defined as above. We have that $${\text{\rm Ker}}\L_{+}=\text{{\rm span}}\left\{ \pa_{x_{1}}u,\cdots,\pa_{x_{N}}u\right\} \label{eq: kernel 1}$$ and $${\text{\rm Ker}}\L_{-}=\text{{\rm span}}\left\{ u\right\} .\label{eq: kernel 2}$$
First we prove (\[eq: kernel 2\]). The proof is standard. In fact, we can use the argument of Selvitella [@Selvitella-2015] since which is applicable to $p$ in the whole range of $1<p<p_{\max}$. We give a proof here for the reader’s convenience.
First we use spherical harmonics to decompose functions $v\in H^{j}(\R^{N})$ for $j\in\N$. Denote by $-\De_{\S^{N-1}}$ the Laplacian-Beltrami operator on the unit $N-1$ dimensional sphere $\S^{N-1}$ in $\R^{N}$. Write $$\begin{aligned}
M_{k}=\frac{(N+k-1)!}{(N-1)!k!}\quad\forall\, k\ge0, & \text{and } & M_{k}=0\quad\forall\, k<0.\end{aligned}$$ Denote by $Y_{k,l}$, $k=0,1,\ldots$ and $1\le l\le M_{k}-M_{k-2}$, the spherical harmonics such that $$-\De_{\S^{N-1}}Y_{k,l}=\la_{k}Y_{k,l}$$ for all $k=0,1,\ldots$ and $1\le l\le M_{k}-M_{k-2}$, where $$\begin{aligned}
\la_{k}=k(N+k-2) & & \forall\, k\ge0\end{aligned}$$ are eigenvalues of $-\De_{\S^{N-1}}$ with multiplicities $M_{k}-M_{k-2}$. In particular, we deduce that $\la_{0}=0$ is of multiplicity 1 with $Y_{0,1}=1$, and $\la_{1}=N-1$ is of multiplicity $N$ with $Y_{1,l}=x_{l}/|x|$ for $1\le l\le N$.
Then for any function $v\in H^{j}(\R^{N})$, we have $$v(x)=v(r\Om)=\sum_{k=0}^{\wq}v_{k}(r)Y_{k}(\Om)$$ with $r=|x|$ and $\Om=x/|x|$, where $$\begin{aligned}
v_{k}(r)=\int_{\S^{N-1}}v(r\Om)Y_{k}(\Om){{\rm d}}\Om & & \forall\, k\ge0.\label{eq: k-th component of v}\end{aligned}$$ Note that $v_{k}\in H^{j}(\R_{+},r^{N-1}{{\rm d}}r)$ holds for all $k\ge0$ since $v\in H^{j}(\R^{N})$.
Next, apply above decomposition to any function $v\in H^{1}(\R^{N})$. We conclude that $\L_{-}v=0$ if and only if $$\L_{-,k}v_{k}\equiv-v_{k}^{\prime\prime}-\frac{N-1}{r}v_{k}^{\prime}+\frac{\la_{k}}{r^{2}}v_{k}+\om v_{k}-(\De u^{2}+u^{p-1})v_{k}=0$$ for all $k\ge0$, where $v_{k}$ is defined as in (\[eq: k-th component of v\]). Note that $\L_{-,k}$ is a self-adjoint operator acting on $L^{2}(\R_{+},r^{N-1}{{\rm d}}r)$ for all $k\in\N$.
First we consider $k=0$. In this case we have $\la_{0}=0$. By a direct computation, we obtain that $\L_{-,0}u=0$. Since $u(r)>0$ for all $r>0$, we conclude in a standard way that $u$ is the first eigenfunction and $0$ is the first simple eigenvalue of $\L_{-,0}$. Thus we have $${\text{\rm Ker}}\L_{-,0}=\text{span}\{u\}.\label{eq: kernel of L---0}$$
Next consider $k\ge1$. We claim that for all $k\ge1$, there holds $${\text{\rm Ker}}\L_{-,k}=\{0\}.\label{eq: kernel of L---k}$$ Indeed, since $\la_{k}>0$, we deduce that $\L_{-,k}>\L_{-,0}$ holds in the sense of quadratic form, which implies that $\L_{-,k}w=0$ if and only if $w\equiv0$. This proves the claim.
Finally, we infer from (\[eq: kernel of L—0\]) and (\[eq: kernel of L—k\]) that (\[eq: kernel 2\]) holds. This finishes the proof of (\[eq: kernel 2\]).
It remains to prove (\[eq: kernel 1\]). We still use spherical harmonic as above. Then $\L_{+}v=0$ for $v\in\X_{\C}(\R^{N})$ if and only if for all $k=0,1,\ldots$, we have $$\begin{aligned}\L_{+,k}v_{k} & \equiv-(1+2u^{2})\left(v_{k}^{\prime\prime}+\frac{N-1}{r}v_{k}^{\prime}-\frac{\la_{k}}{r^{2}}v_{k}\right)-4uu^{\prime}v_{k}^{\prime}+\om v_{k}\\
& \qquad-(2u\De u+\De u^{2}+pu^{p-1})v_{k}=0.
\end{aligned}
\label{operator: L-+--k}$$ For a detailed calculation of $\L_{+,k}$, we refer to Selvitella [@Selvitella-2015]. Note the fact that $$\begin{aligned}
\pa_{x_{l}}u=u^{\prime}(|x|)\frac{x_{l}}{|x|}=u^{\prime}(r)Y_{1,l} & & \text{for }1\le l\le N.\end{aligned}$$ Thus to prove (\[eq: kernel 1\]), it is sufficient to prove that $$\begin{aligned}
\L_{+,0}v_{0}=0 & & \text{if and only if }v_{0}\equiv0,\label{eq: kernel 1.1}\end{aligned}$$ and that $$\begin{aligned}
\L_{+,1}v_{1}=0 & & \text{if and only if }v_{1}\in\text{span}\left\{ u^{\prime}\right\} ,\label{eq: kernel 1.2}\end{aligned}$$ and that $$\begin{aligned}
\L_{+,k}v_{k}=0 & & \text{if and only if }v_{k}\equiv0\label{eq: kernel 1.3}\end{aligned}$$ for all $k\ge2$.
(\[eq: kernel 1.2\]) and (\[eq: kernel 1.3\]) can be proved in the same way as that of (\[eq: kernel of L—0\]) and (\[eq: kernel of L—k\]). Consider $k=1$. In this case we have $\la_{1}=N-1$. We deduce from $\L_{+}\pa_{x_{1}}u=0$ that $\L_{+,1}u^{\prime}=0$. Since $u^{\prime}(r)<0$ for all $r>0$, we conclude in a standard way that $u^{\prime}$ is the first eigenfunction and $0$ is the first simple eigenvalue of $\L_{+,1}$. This proves (\[eq: kernel 1.2\]). To conclude (\[eq: kernel 1.3\]), it is enough to notice that $\L_{+,k}>\L_{+,1}$ for any $k>1$. This proves (\[eq: kernel 1.3\]).
We leave the proof of (\[eq: kernel 1.1\]) in the next subsection. The proof of Proposition \[prop: two kernels\] is complete. Thus the proof of Theorem \[thm: main result\] is complete.
We remark that (\[eq: kernel 2\]) can be proved in a more compact way. Indeed, note that $\L_{-}u=0$ since $u$ solves equation (\[eq: Object equ.\]). Thus $u$ is an eigenfunction of $\L_{-}$ with eigenvalue 0. Moreover, recall that $u$ is a positive eigenfunction. We can conclude in a standard way that $0$ is the first eigenvalue of $\L_{-}$ and is simple. Hence ${\text{\rm Ker}}\L_{-}=\text{span}\{u\}$. See similar discussions in Chang et al. [@Chang; @et; @al-2007].
Proof of (\[eq: kernel 1.1\])
------------------------------
Let us first briefly review the proof of (\[eq: kernel 1.1\]) of Selvitella [@Selvitella-2015]. Suppose that $v_{0}$ belongs to $L^{2}(\R_{+},r^{N-1}{{\rm d}}r)$, $v_{0}\not\equiv0$ and satisfies $\L_{+,0}v_{0}=0$. His proof (see Lemma 4.4 of Selvitella [@Selvitella-2015]) contains two ingredients. First he proved that $v_{0}(r)$ changes sign at least once for $r>0$, and then by the disconjugacy interval argument of Kwong [@Kwong1989] he deduced that $v_{0}(r)$ is unbounded for $r>0$ sufficiently large, which contradicts to $v_{0}\in L^{2}(\R_{+},r^{N-1}{{\rm d}}r)$. In this way Selvitella [@Selvitella-2015] proved (\[eq: kernel 1.1\]). To prove that $v_{0}$ changes sign at least once on $\R_{+}$, Selvitella [@Selvitella-2015] used an ordinary differential equation analysis, in which the assumption $p\ge3$ is needed (see Section 3 of Selvitella [@Selvitella-2015]). While the disconjugacy interval argument applies to the whole range $1<p<p_{\max}$.
Taking into account above review, we infer that (\[eq: kernel 1.1\]) can be deduced from the following result together with the disconjugacy interval argument as that of Kwong [@Kwong1989] and Selvitella [@Selvitella-2015].
\[prop: changing sign proposition\] Let $\L_{+,0}$ be defined as in (\[operator: L-+–k\]) with $k=0$. Suppose that $v$ belongs to $L^{2}(\R_{+},r^{N-1}{{\rm d}}r)$, $v\not\equiv0$ and satisfies $\L_{+,0}v=0$. Then $v(r)$ changes sign at least once for $r>0$.
Proposition \[prop: changing sign proposition\] can be viewed as a substitute of Proposition 3.10 of Selvitella [@Selvitella-2015]. We use a spectrum analysis to prove Proposition \[prop: changing sign proposition\].
First we note that $\L_{+,0}$ is the restriction of $\L_{+}$ on the sector $L_{\rad}^{2}(\R^{N})$, the subspace of radial functions in $L^{2}(\R^{N})$. Indeed, for any $v\in L_{\rad}^{2}(\R^{N})$, we have $$\begin{aligned}\L_{+}v & =-\De v-2u\De(uv)+\om v-(\De u^{2}+pu^{p-1})v\\
& =-(1+2u^{2})\left(v^{\prime\prime}+\frac{N-1}{r}v^{\prime}\right)-4uu^{\prime}v^{\prime}+\om v-(2u\De u+\De u^{2}+pu^{p-1})v\\
& =\L_{+,0}v
\end{aligned}$$ since $\la_{0}=0$. Thus we immediately find the following result which is equivalent to Proposition \[prop: changing sign proposition\].
\[prop: equiv result\] Suppose that $v\in{\text{\rm Ker}}\L_{+}\cap L_{\rad}^{2}(\R^{N})$ is a nontrivial function. Then $v(x)=v(r)$ with $r=|x|$ changes sign at least once for $r>0$.
The idea to prove Proposition \[prop: equiv result\] is as follows. Note that $0$ belongs to the spectrum $\si(\L_{+})$ of $\L_{+}$, since it is straightforward to verify that $$\text{span}\left\{ \pa_{x_{1}}u,\cdots,\pa_{x_{N}}u\right\} \subset{\text{\rm Ker}}\L_{+}.$$ We will show that $0$ belongs to the discrete spectrum $\si_{\text{disc}}(\L_{+})$ of $\L_{+}$, that is, $0$ is an isolated eigenvalue of $\L_{+}$ and the corresponding eigenfunction space is finite dimensional. We also show that $0$ is not the first eigenvalue of $\L_{+}$. Then we have $\int_{\R^{N}}ve_{1}{{\rm d}}x=0$, where $e_{1}$ is the first eigenfunction of $\L_{+}$. This fact will imply that $v=v(r)$ changes sign for $r>0$, once we prove that $e_{1}$ does not change sign in $\R^{N}$.
It is easy to verify that $\L_{+}$ is a self-adjoint operator acting on $L^{2}(\R^{N})$ with form domain $\X_{\C}$ and domain $H^{2}(\R^{N})$. Hence we have $\si(\L_{+})\subset\R$. Furthermore, by Weyl’s theorem (see Theorem 7.2 of Hislop and Sigal [@Hislop-Sigal-1996]) we have $\si(\L_{+})=\si_{\text{disc}}(\L_{+})\cup\si_{\text{cont}}(\L_{+})$, and $\si_{\text{disc}}(\L_{+})\cap\si_{\text{cont}}(\L_{+})=\emptyset$, where $\si_{\text{cont}}(\L_{+})$ denotes the continuous spectrum of $\L_{+}$. Let us now start the proof of Proposition \[prop: equiv result\] with an estimate on $\si_{\text{cont}}(\L_{+})$. Recall that a constant $\la$ belongs to $\si_{\text{cont}}(\L_{+})$ if and only if there exists a sequence $\phi_{n}\in H^{2}(\R^{N})$, $n=1,2,\ldots$, such that $$\begin{aligned}
& & \|\L_{+}\phi_{n}-\la\phi_{n}\|_{2}\to0\qquad\text{as }n\to\wq,\text{ and }\label{eq: asym. eigenf.}\\
& & \|\phi_{n}\|_{2}=1\qquad\text{ for all }n\in\N,\text{ and }\label{eq: normalizations}\\
& & \phi_{n}{\rightharpoonup}0\qquad\text{ weakly in }L^{2}(\R^{N})\text{ as }n\to\wq.\label{eq: weak convergence}\end{aligned}$$
\[lem: essential spectrum\] We have $\si_{\text{{\rm cont}}}(\L_{+})\subset[\om,\wq).$
Since $\L_{+}$ is self-adjoint, we have $\si(\L_{+})\subset\R$. So it is sufficient to prove that if $\la<\om$, then $\la\not\in\si_{\text{cont}}(\L_{+})$. We argue by contradiction. Suppose, on the contrary, that $\la<\om$ is a real number and $\la\in\si_{\text{cont}}(\L_{+})$. Then there exists a sequence $\{\phi_{n}\}_{n=1}^{\wq}\subset H^{2}(\R^{N})$ such that (\[eq: asym. eigenf.\])-(\[eq: weak convergence\]) hold. We claim that, up to a subsequence, $$\begin{aligned}
\phi_{n}\to0 & & \text{strongly in }L^{2}(\R^{N}).\label{esti: estimate 1}\end{aligned}$$ Then we reach to a contradiction to (\[eq: normalizations\]) and Lemma \[lem: essential spectrum\] is proved.
We prove (\[esti: estimate 1\]) as follows. Note that $\De u^{2}+pu^{p-1}$ is bounded in $\R^{N}$ by Theorem \[thm: properties of GS\]. Thus we obtain that $$\sup_{n}\int_{\R^{N}}(\om-\la+|\De u^{2}+pu^{p-1}|)|\phi_{n}|^{2}{{\rm d}}x<\wq.$$ On the other hand, we have $$\begin{aligned}o(1) & =\langle(\L_{+}-\la)\phi_{n},\phi_{n}\rangle\\
& =\int_{\R^{N}}\left(|\na\phi_{n}|^{2}+|\na(u\phi_{n})|^{2}+(\om-\la-\De u^{2}-pu^{p-1})|\phi_{n}|^{2}\right){{\rm d}}x.
\end{aligned}
\label{eq: first order esti}$$ The first equality of above follows from (\[eq: asym. eigenf.\]) and (\[eq: normalizations\]). Therefore we derive directly from (\[eq: first order esti\]) that $|\na\phi_{n}|\in L^{2}(\R^{N})$ is bounded uniformly for all $n\in\N$. Hence $\phi_{n}\in H^{1}(\R^{N})$ is bounded uniformly for all $n$ in view of (\[eq: normalizations\]). In particular, we deduce, after possibly passing to a subsequence, that $$\begin{aligned}
\phi_{n}\to0 & & \text{strongly in }L_{\loc}^{2}(\R^{N}).\label{eq: local strong conver.}\end{aligned}$$ Next we recall that the function $\De u^{2}+pu^{p-1}$ decays exponentially to zero at infinity by Theorem \[thm: properties of GS\]. Combining this fact together with (\[eq: local strong conver.\]) gives us that $$\int_{\R^{N}}|\De u^{2}+pu^{p-1}||\phi_{n}|^{2}{{\rm d}}x\to0\label{eq: vanishing term 1}$$ as $n\to\wq$. Combining (\[eq: vanishing term 1\]) with (\[eq: first order esti\]) and recalling that $\om>\la$, we obtain that $$\lim_{n\to\wq}\int_{\R^{N}}|\phi_{n}|^{2}{{\rm d}}x=0,$$ which contradicts to the assumption (\[eq: normalizations\]). The proof of Lemma \[lem: essential spectrum\] is complete.
A direct consequence of Lemma \[lem: essential spectrum\] is that $0\in\si_{\text{disc}}(\L_{+})$. Lemma \[lem: essential spectrum\] also allows us to derive a variational characterization for eigenvalues of $\L_{+}$ that are below the infimum of $\si_{\text{cont}}(\L_{+})$. Indeed, suppose that we have eigenvalues $$\inf\si(\L_{+})\equiv\mu_{1}\le\mu_{2}\le\cdots\le\mu_{n}<\inf\si_{\text{{\rm cont}}}(\L_{+}).$$ The fact $\mu_{1}>-\wq$ follows easily from the elementary estimate $$\inf_{\xi\in\X_{\C},\|\xi\|_{2}=1}\langle\L_{+}\xi,\xi\rangle>-\wq.$$ Then we have $$\mu_{1}=\inf\left\{ \langle\L_{+}\xi,\xi\rangle:\xi\in\X_{\C},\|\xi\|_{2}=1\right\} .$$ Denoting by $W_{k}$, $2\le k\le n$, the linear space spanned by the first $n-1$ eigenfunctions corresponding to $\mu_{1}$, $\ldots$, $\mu_{n-1}$, we have by induction $$\mu_{n}=\inf\left\{ \langle\L_{+}\xi,\xi\rangle:\xi\in\X_{\C},\|\xi\|_{2}=1,\text{ and }\xi\bot W_{n}\right\} .$$ Furthermore, for any function $\xi\in\X_{\C}$ with $\|\xi\|_{2}=1$, $\xi\bot W_{n}$ and $\langle\L_{+}\xi,\xi\rangle=\mu_{n}$, $\xi$ is a linear combination of eigenfunctions corresponding to $\mu_{n}$.
Next we prove that $\L_{+}$ satisfies Perron-Frobenius property. That is, if $\inf\si(\L_{+})$ is an eigenvalue, then it is simple and the corresponding eigenfunction can be chosen strictly positive. In the case of equation (\[eq: NLS\]), it is easy to verify that the operator ${\cal A}_{+}$ satisfies the Perron-Frobenius property as follows. Let $\xi_{1}\in L^{2}(\R^{N})$ is an eigenfunction of ${\cal A}_{+}$ with eigenvalue $a_{1}=\inf\si({\cal A}_{+})<0$. Then $\xi_{1}$ solves equation $$-\De\xi+(\om-a_{1})\xi_{1}=qQ^{q-1}\xi_{1}.$$ Since above equation is linear, we can assume that $\xi_{1,+}=\max\{\xi_{1},0\}\not\equiv0$. Then we obtain, by multiplying above equation by $\xi_{1,+}$, that $\langle{\cal A}_{+}\xi_{1,+},\xi_{1,+}\rangle=0$, which implies that $\xi_{1,+}$ is also an eigenfunction of ${\cal A}_{+}$ with eigenvalue $a_{1}$. Therefore, $\xi_{1,+}$ satisfies the equation of $\xi_{1}$ as well. Finally, note that $\om-a_{1}>0$. Thus we have $$\xi_{1,+}=\frac{1}{-\De+(\om-a_{1})}qQ^{q-1}\xi_{1,+}.$$ Since $Q>0$ for all $x\in\R^{N}$ and since the integral kernel of the operator $\frac{1}{-\De+(\om-a_{1})}$ is positive away from the origin, we conclude from above formula that $\xi_{1,+}(x)>0$ for all $x\in\R^{N}$. Thus $\xi_{1}=\xi_{1,+}$ is strictly positive in $\R^{N}$. This shows that ${\cal A}_{+}$ satisfies the Perron-Frobenius property. However, in our case, we do not know whether the first eigenvalue $\mu_{1}$ of $\L_{+}$ is negative or not. Due to the presence of the quasilinear term $u\De(u\cdot)$, we can not use above simple argument to assert that $\L_{+}$ satisfies Perron-Frobenius property. Nevertheless, we can still deduce the following result.
\[lem: first eigenvalue is simple and negative\] The first eigenvalue $\mu_{1}$ of $\L_{+}$ is negative and simple.
We have to show that $\mu_{1}<0$ holds and that eigenfunctions corresponding to $\mu_{1}$ is of constant sign. We argue by contradiction. Suppose that $\mu_{1}\ge0$ holds. Then the fact $0\in\si_{\text{disc}}(\L_{+})$ implies that $\mu_{1}=0$. Note that ${\text{\rm Ker}}\L_{+}\neq\emptyset$ is the eigenfunction space corresponding to $0$. For any $\phi\in{\text{\rm Ker}}\L_{+}$, we have that $$-\De\phi-2u\De(u\phi)+\om\phi-(\De u^{2}+pu^{p-1})\phi=0.$$ Since $u$ is a real valued function, we can assume, with no loss of generality, that $\phi$ is a real valued function as well. Furthermore, we can assume that the positive part $\phi_{+}=\max(\phi,0)$ is not identically zero. Then multiply above equation by $\phi_{+}$. We obtain by integrating by parts that $$\langle\L_{+}\phi_{+},\phi_{+}\rangle=0.$$ That is, $\langle\L_{+}\phi_{+},\phi_{+}\rangle$ achieves the first eigenvalue 0. Thus $\phi_{+}$ is a combination of eigenfunctions of $0$, which implies that $\phi_{+}$ satisfies equation $$-\De\phi_{+}-2u\De(u\phi_{+})+\om\phi_{+}-(\De u^{2}+pu^{p-1})\phi_{+}=0.\label{eq: equ. of first eignef.}$$ We claim that equation (\[eq: equ. of first eignef.\]) implies that $$\begin{aligned}
\phi_{+}(x)>0 & & \text{for all }x\in\R^{N}.\label{eq: positivity claim}\end{aligned}$$ Rewrite equation (\[eq: equ. of first eignef.\]) in the form $$\begin{aligned}
-\De\phi_{+}-\sum_{i=1}^{N}b_{i}(x)\cdot\pa_{x_{i}}\phi_{+}+c(x)\phi_{+}=0 & & \text{in }\R^{N}.\label{eq: equ. of first eignef.2}\end{aligned}$$ By Theorem \[thm: properties of GS\], both functions $$\begin{aligned}
b_{i}(x)\equiv-\frac{4u}{1+2u^{2}}\pa_{x_{i}}u,\quad(1\le i\le N) & \text{and} & c(x)\equiv\frac{\om-2u\De u-\De u^{2}-pu^{p-1}}{1+2u^{2}}\end{aligned}$$ are bounded smooth functions. Thus elliptic regularity theory gives us that $\phi_{+}\in C^{\wq}(\R^{N})$ holds. Now, by a famous generalized comparison principle for second order elliptic equations due to Serrin (see Theorem 2.10 of Han and Lin [@Han-Lin-Book Chapter 2]), we deduce from equation (\[eq: equ. of first eignef.2\]) that (\[eq: positivity claim\]) holds. This proves the claim.
Recall that $\pa_{x_{1}}u\in{\text{\rm Ker}}\L_{+}$. Take $\phi=\pa_{x_{1}}u=u^{\prime}(|x|)x_{1}/|x|$. Since $u^{\prime}(|x|)<0$ for $|x|>0$, we have that $\phi_{+}(x)\equiv0$ for any $x\in\R^{N}$ with $x_{1}\ge0$. We obtain a contradiction to (\[eq: positivity claim\]). Hence we conclude that $\mu_{1}<0$.
Finally, by similar arguments as above, we infer that any eigenfunction corresponding to $\mu_{1}$ is either positive or negative in $\R^{N}$. This proves that $\mu_{1}$ is simple. The proof of Lemma \[lem: first eigenvalue is simple and negative\] is complete.
Now we are able to prove Propositions \[prop: changing sign proposition\] and \[prop: equiv result\].
It is enough to prove Proposition \[prop: equiv result\] due to the equivalence. For any function $v\in{\text{\rm Ker}}\L_{+}\cap L_{\rad}^{2}(\R^{N})$, $v\not\equiv0$, we obtain from above that $$\int_{\R^{N}}v\bar{e}_{1}{{\rm d}}x=0$$ holds for any eigenfunction $e_{1}$ of $\L_{+}$ with eigenvalue $\mu_{1}$. Since $e_{1}$ can be chosen strictly positive in $\R^{N}$, we infer that $v(x)=v(r)$ with $r=|x|$ must change sign for $r>0$. This proves Proposition \[prop: equiv result\]. So follows Proposition \[prop: changing sign proposition\].
We end this section by showing that $\inf\si(\L_{+})<0$ holds for $N=2$ via direct computations. Precisely, we show that $$\langle\L_{+}u,u\rangle<0.\label{esti: N equals to 2}$$ (\[esti: N equals to 2\]) follows from a Pohozaev type identity. Since $u$ solves equation (\[eq: Object equ.\]), an elementary calculation gives the following Pohozaev type identity $$\om\int_{\R^{2}}|u|^{2}{{\rm d}}x=\frac{2}{p+1}\int_{\R^{2}}|u|^{p+1}{{\rm d}}x.$$ Here we used the fact $N=2$. On the other hand, multiplying equation (\[eq: Object equ.\]) by $u$ and integrating by parts yields $$\int_{\R^{2}}|\na u|^{2}{{\rm d}}x+4\int_{\R^{2}}|u|^{2}|\na u|^{2}{{\rm d}}x+\om\int_{\R^{2}}|u|^{2}{{\rm d}}x=\int_{\R^{2}}|u|^{p+1}{{\rm d}}x.$$ Recall that $$\langle\L_{+}u,u\rangle=8\int_{\R^{2}}|u|^{2}|\na u|^{2}{{\rm d}}x-(p-1)\int_{\R^{2}}|u|^{p+1}{{\rm d}}x.$$ Combining above three identities, we deduce that $$\langle\L_{+}u,u\rangle=-2\int_{\R^{2}}|\na u|^{2}{{\rm d}}x-\frac{(p-1)^{2}}{p+1}\int_{\R^{2}}|u|^{p+1}{{\rm d}}x<0.$$ This proves (\[esti: N equals to 2\]). Thus we conclude that $\inf\si(\L_{+})<0$ holds for $N=2$.
*Acknowledgment.* The author is financially supported by the Academy of Finland, project 259224.
[10]{} <span style="font-variant:small-caps;">A. Ambrosetti and A. Malchiodi,</span> *Perturbation Methods and Semilinear Elliptic Problems on $\mathbf{R}^{n}$.* Progress in Mathematics, 240. Birkhäuser Verlag, Basel, 2006.
<span style="font-variant:small-caps;">S.-M. Chang, S. Gustafson, K. Nakanishi and T.-P. Tsai,</span> *Spectra of linearized operators for NLS solitary waves.* SIAM J. Math. Anal. **39** (2007/08), no. 4, 1070-1111.
<span style="font-variant:small-caps;">M. Colin</span>, *On the local well-posedness of quasilinear Schrödinger equations in arbitrary space dimension.* Comm. Partial Differ. Equ. **27** (2002), 325-354 .
<span style="font-variant:small-caps;">M. Colin and L. Jeanjean</span>, *Solutions for a quasilinear Schrödinger equation: a dual approach.* Nonlinear Anal. **56** (2004), 213-226.
<span style="font-variant:small-caps;">M. Colin, L. Jeanjean and M. Squassina</span>, *Stability and instability results for standing waves of quasi-linear Schrödinger equations.* Nonlinearity **23** (2010), 1353-1385.
<span style="font-variant:small-caps;">R. Frank and E. Lenzmann,</span> *Uniqueness of non-linear ground states for fractional Laplacians in $\R$.* Acta Math. **210** (2013), no. 2, 261-318.
<span style="font-variant:small-caps;">R. Frank, E. Lenzmann and L. Silvestre,</span> *Uniqueness of radial solutions for the fractional Laplacian.* Preprint at arXiv:1302.2652v1.
<span style="font-variant:small-caps;">B. Gidas, W.M. Ni and L. Nirenberg</span>, *Symmetry of positive solutions of nonlinear elliptic equations in $\mathbf{R}^{n}$.* Mathematical analysis and applications, Part A, pp. 369402, Adv. in Math. Suppl. Stud., 7a, Academic Press, New York-London, 1981.
<span style="font-variant:small-caps;">Q. Han and F.-H. Lin,</span> *Elliptic partial differential equations.* Courant Lecture Notes in Mathematics, 1. New York University, Courant Institute of Mathematical Sciences, New York; American Mathematical Society, Providence, RI, 1997.
<span style="font-variant:small-caps;">P.D. Hislop and I.M. Sigal,</span> *Introduction to spectral theory.* *With applications to Schrödinger operators.* Applied Mathematical Sciences, 113. Springer-Verlag, New York, 1996.
<span style="font-variant:small-caps;">C.E. Kenig, G. Ponce and L. Vega</span>, *The Cauchy problem for quasi-linear Schrödinger equations.* Invent. Math. **158** (2004), 343-388.
<span style="font-variant:small-caps;">M.K. Kwong,</span> *Uniqueness of positive solutions of $\De u-u+u^{p}=0$ in $\mathbf{R}^{n}$.* Arch. Rational Mech. Anal. **105** (1989), no. 3, 243-266.
<span style="font-variant:small-caps;">H. Lange, M. Poppenberg and H. Teismann,</span> *Nash-Moser methods for the solution of quasi-linear Schrödinger equations.* Comm. Partial Differ. Equ. **24** (1999), 1399-1418.
<span style="font-variant:small-caps;">E. Lenzmann,</span> *Uniqueness of ground states for pseudorelativistic Hartree equations.* Anal. PDE **2** (2009), no. 1, 1-27.
<span style="font-variant:small-caps;">E.H. Lieb and M. Loss,</span> *Analysis.* 2nd edn. Graduate Studies in Mathematics, Vol. 14 (Amer. Math. Soc., 2001).
<span style="font-variant:small-caps;">J.-Q. Liu, Y.-Q. Wang and Z.-Q. Wang,</span> *Soliton solutions for quasi-linear Schrödinger equations II.* J. Differ. Equ. **187** (2003), 473-493.
<span style="font-variant:small-caps;">J.-Q. Liu, Y.-Q. Wang and Z.-Q. Wang,</span> *Solutions for quasi-linear Schrödinger equations via the Nehari method.* Comm. Partial Differ. Equ. **29** (2004), 879-901.
<span style="font-variant:small-caps;">J.-Q. Liu and Z.-Q. Wang,</span> *Solitons solutions for quasi-linear Schrödinger equations.* Proc. Am. Math. Soc. **131** (2003), 441-448.
<span style="font-variant:small-caps;">M. Poppenberg,</span> *On the local well posedness of quasi-linear Schrödinger equations in arbitrary space dimension.* J. Differ. Equ. **172** (2001), 83-115.
<span style="font-variant:small-caps;">M. Poppenberg, K. Schmitt and Z.-Q. Wang,</span> *On the existence of soliton solutions to quasilinear Schrödinger equations.* Calc. Var. Partial Differ. Equ. **14** (2002), 329-344.
<span style="font-variant:small-caps;">A. Selvitella</span>, *Uniqueness and nondegeneracy of the ground state for a quasilinear Schrödinger equation with a small parameter.* Nonlinear Anal. **74**(5) (2011), 1731-1737.
<span style="font-variant:small-caps;">A. Selvitella</span>, *Nondegeneracy of the ground state for quasilinear Schrödinger equations.* Calc. Var. Partial Differ. Equ. **53** (2015), 349-364.
<span style="font-variant:small-caps;">W.P. Ziemer</span>, *Weakly differentiable functions.* Graduate Texts in Mathematics, 120. Springer-Verlag, New York, 1989.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The production of ions via laser ablation for the loading of radiofrequency (RF) ion traps is investigated using a nitrogen laser with a maximum pulse energy of 0.17 mJ and a peak intensity of about 250 MW/cm$^2$. A time-of-flight mass spectrometer is used to measure the ion yield and the distribution of the charge states. Singly charged ions of elements that are presently considered for the use in optical clocks or quantum logic applications could be produced from metallic samples at a rate of the order of magnitude $10^5$ ions per pulse. A linear Paul trap was loaded with Th$^+$ ions produced by laser ablation. An overall ion production and trapping efficiency of $10^{-7}$ to $10^{-6}$ was attained. For ions injected individually, a dependence of the capture probability on the phase of the RF field has been predicted. In the experiment this was not observed, presumably because of collective effects within the ablation plume.'
author:
- 'K. Zimmermann, M. V. Okhapkin, O. A. Herrera-Sancho, and E. Peik'
title: Laser ablation loading of a radiofrequency ion trap
---
Introduction {#sec:intro}
============
Experiments with ions in radiofrequency Paul traps or Penning traps are usually prepared by producing ions with low kinetic energy inside the trapping volume. For non-volatile species, originally a combination of a small atomic beam oven with an electron gun was used to ionize atoms by electron impact [@fischer; @neuhauser]. This technique has two disadvantages: the atomic beam may unintentionally deposit an inhomogeneous layer of the element to be trapped on the trap electrodes, thus creating an inhomogeneous surface potential. The electron beam may leave stray charges on the surface of insulators situated in the vicinity of the trap. Both effects make the precise and stable localization of the ion in the trap more tedious. These adverse effects can be strongly mitigated by using resonant laser photoionization of the neutral atom [@drewsen1; @blatt], allowing to eliminate the electron beam and to reduce the atomic flux because of the higher ionization cross section. This approach adds to the complexity of the setup, however, because one or even two lasers are required to excite resonance transitions of the neutral atom.
Here we present a systematic study of ion production via photoablation using a nitrogen laser. This type of laser was selected because it provides the most compact source of nanosecond pulses with a peak power of tens of kW and because its wavelength of 337 nm will generally be absorbed more strongly in metals than the infrared radiation available from other efficient pulsed lasers. The motivation for this study was to establish a trap loading method that requires less source material and produces less stray charges than electron impact ionization. The advantage of laser ablation results from the fact that a small source volume is heated rapidly to obtain the required vapor pressure over a very short time interval. In comparison with resonant photoionization, laser ablation loading is technically less complex. Ideally, laser ablation and photoionization may be combined, as has been investigated for Ca$^+$ [@hendricks; @sheridan], because this provides isotope-selective ion production with a minimum amount of stray charge. In our experiments, ions of most elements that are presently considered for the use in optical clocks or quantum logic applications were produced effectively from metallic samples. This method also allows one to produce ions of elements that are not easily evaporated like e.g. tungsten or thorium and it also produces doubly charged ions. We are especially interested in the production of Th$^+$ for an investigation of a nuclear optical clock based on the low energy transition to an isomeric state in $^{229}$Th [@porsev]. Previous studies have already reported laser ablation from metals for the loading of ion traps [@knight; @kwong1; @hashimoto; @leib; @kwapien; @hendricks; @chapman], but typically higher laser pulse energies at infrared wavelengths were used and experiments were carried out only for a small selection of elements. Laser desorption of organic molecules has been studied in conjunction with ion trap mass spectrometry (see for example Refs. [@hemberger; @robb]).
Photoablation is categorized by two different regimes which are distinguished by comparing the laser pulse duration with the characteristic time of electron-phonon interaction. When the laser pulse duration is shorter than this characteristic time, which is typically several picoseconds in metals, bonds may be broken by the strong electric field and ablation can be considered as a direct solid-to-vapor transition. In contrast, in the interaction of metals with nanosecond laser pulses as investigated in this paper, the pulse duration is longer than the thermalization time constants of electrons and lattice and ablation proceeds as a thermal process. If the laser intensity exceeds a threshold of the order of $10^8$ W/cm$^2$ (corresponding to an energy density in the range of $0.1$ J/cm$^2$), the target surface is locally molten and the material partly vaporized [@willmott; @chichkov]. Above threshold, the ablation layer thickness quickly reaches the range of 10 nm per laser pulse. The resulting plume contains atoms, molecules, and clusters as well as electrons and ions. Electrons are heated by inverse bremsstrahlung in the interaction with the laser light, leading to further ionization and heating of the ions. Finally, ions leave the plume away from the target surface at velocities in the range of $10^4$ m/s (see e.g. Refs. [@thestrup; @laska]). It has been shown that the distribution of ions over the charge states depends on the laser intensity and shifts to higher charges for higher intensity [@laska].
In order to investigate the suitability of this process for ion production of relevant elements with a low-power nitrogen laser, we use a time-of-flight (TOF) mass spectrometer to measure the ion yield and the distribution over the charge states. These results are described in sections 2 and 3. In a second experiment, described in sections 4 and 5, a linear Paul trap was used to trap Th$^+$ ions produced by laser ablation and the loading efficiency and dependence on the RF phase was investigated.
Laser source and time-of-flight spectrometer {#sec:lasersource}
============================================
A commercially available nitrogen laser (model SRS NL-100) is used at a wavelength of 337 nm with a maximum pulse energy of 170 $\mu\textrm{J}$, a pulse duration of 4 ns, and a maximum repetition rate of 20 Hz. To obtain intensities higher than the ablation threshold, the transverse multimode laser beam was focused to a spot of dimensions ${100\, \cdot 150\, \mu \textrm{m}}^2$ (to within $\pm10$% uncertainty) using a lens system as indicated in figure \[fig:setup\]. Thus the obtained peak intensity is about $280\, {\textrm{MW}}/{\textrm{cm}^2}$ and the energy density is $1.1\, {\textrm{J}}/{\textrm{cm}^2}$.
In order to characterize the distribution of charged ablation products we use a TOF mass spectrometer [@wiley]. This technique is well suited here because the ions are produced during a very short time interval. If, subsequently, the ensemble of ions is accelerated in a static electric field to the same kinetic energy, different masses $m$ and charges $q$ will have a distribution of velocities which is proportional to $\sqrt{q/m}$. The flight path of the ions is divided into four regions as outlined in figure \[fig:setup\]. At first, ions are produced in the ablation region. The ion cloud is then moved by a small electric field to the acceleration region where the ions gain most of their kinetic energy. They are injected into the field-free drift region where the mass-to-charge resolution is obtained and are finally detected according to their arrival time.
The ablation targets were mounted in a cubic vacuum chamber segment that was evacuated to $10^{-6}\, \textrm{Pa}$. The sample is in electrical contact with a repeller plate held at a potential of +2000 V, defining the ablation region with a spacing of 10 mm to the first grid at a voltage of +1900 V. A second, grounded grid, mounted 10 mm behind the first grid, creates the acceleration region. Attached to the cubic chamber, a 1.25 m long tube provides the drift region between the second grid and a third grid, also at ground potential, at the end of the tube. The positive ions are then accelerated towards the cathode of a channeltron electron multiplier that was held at a voltage of $-2000$ V. The signal current of the channeltron was converted to a voltage using a transimpedance amplifier and was measured with a digital storage oscilloscope. A small fraction of the ablation laser beam was detected by a photodiode in order to start the oscilloscope scan.
![Setup of the time-of-flight mass spectrometer and typical experimental result. All distances are given in mm. 1 - ablation region, 2 - acceleration region, 3 - drift region, 4 - detection region with channeltron. Dashed lines indicate grids. The inset shows a TOF spectrum obtained in the ablation of thorium metal.[]{data-label="fig:setup"}](figure1){width="0.95\columnwidth"}
A typical result for the ablation of thorium is shown in figure \[fig:setup\]. It shows a signal of Th$^+$ with the maximum of the peak in good agreement with the calculated flight time. A Th$^{2+}$ signal is clearly seen as well, but no evidence of Th$^{3+}$ at the expected time of $\approx19\,\mu$s is found. The width in the arrival times for these ions is much larger than the laser pulse duration and is a consequence of field inhomogeneity in our TOF setup. In addition, an unresolved contribution from ThO$^+$ ions contributes to a shoulder trailing the Th$^+$ signal. Much narrower signals of light molecular ions from the residual gas or desorbed from the surface are observed at $\approx 10\,\mu$s and show similar distributions in ablation measurements for all tested materials. The peaks can be assigned to molecular ions of water, nitrogen, and oxygen.
Ion yield and distribution of charge states {#sec:ionyield}
===========================================
A number of elements that are of interest for current ion trap experiments were tested and the ablation yields for the observed singly and doubly positively charged ions are noted in Table \[tab:relative\_ablation\]. The number of ions was calculated from the pulse area of the channeltron signal and was normalized to Th$^+$, corresponding to a detected ion number of about $10^5$ per laser pulse. From all tested metals, including the refractory metals molybdenum, tantalum, tungsten, and also from a silicon crystal, singly charged atomic ions could be produced effectively. Doubly charged ions were observed for Mo and Th only. For samples of Ca, Sr and Yb, oxide molecules could be detected as well. Since the yield of the oxide ions was not stable for the investigated metals as seen in figure \[fig:strontium\_and\_ytterbium\_cleaning\] and discussed below, the detected amount of oxides is not included in the table. In general, the total number of detected ions scales roughly inversely proportional with the heat of evaporation of the corresponding metal.
[lccccc]{} element & symbol & 1+ & 2+\
aluminum & Al & 0.5 &\
silicon & Si & 0.8 &\
calcium & Ca & 4.7 &\
copper & Cu & 1.1 &\
strontium & Sr & 1.4 &\
molybdenum & Mo & 0.6 & 0.1\
indium & In & 1.8 &\
ytterbium & Yb & 0.8 &\
tantalum & Ta & 0.7 &\
tungsten & W & 0.4 &\
gold & Au & 0.6 &\
thorium & Th & 1.0 & 0.3\
The TOF spectra detected from Sr and Yb samples showed a significant abundance of oxide ions. The employed metal samples were heavily oxidized due to exposure to air for prolonged periods. Usually, these metals must be handled in a protective gas atmosphere to prevent the formation of oxide layers. Right after focussing the laser on the untreated surface of the Sr sample, the amount of detected SrO$^+$ ions is considerably higher than the amount of Sr$^+$ ions. Doubly charged SrO$^{2+}$ ions were detected as well. Repeated laser pulses on the same spot decrease the amount of detected SrO$^+$ and SrO$^{2+}$ ions significantly until much less molecular ions than atomic Sr$^+$ ions are detected after 1000 shots. Figure \[fig:strontium\_and\_ytterbium\_cleaning\] shows the number of oxide and atomic ions in the produced ablation plasma as a function of the number of laser pulses on the same spot for two samples of Sr and Yb. For both elements, laser ablation turns out to provide an effective method to remove the oxide layer. The approach of cleaning the surface with multiple laser ablation pulses prior to an actual experiment is an effective method to load ion traps with highly reactive elements without prior handling in a protective gas atmosphere.
![Evolution of the yield of molecular and atomic ions from laser ablation of oxidized samples of Sr (left) and Yb (right) as a function of the number of laser pulses applied to the same surface area. The lines superimposed on the data points are a guide to the eye.[]{data-label="fig:strontium_and_ytterbium_cleaning"}](figure2){width="0.95\columnwidth"}
Loading of a linear Paul Trap {#sec:loading}
=============================
In order to test the suitability of the method for loading an ion trap, the laser and optics shown in Fig. 1 were used together with a linear RF Paul trap and the ablation beam was focussed on a target located between the trap electrodes. A segmented linear Paul trap was constructed to store a high number of ions. It consists of cylindrical rods with a diameter of 8 mm, mounted at a diagonal distance of 7 mm between the surfaces of the rods as it is shown in Fig. 3. The trap electrodes are made of copper-beryllium and consist of three parts, where the middle section for storage of the ions is 20 mm with an overall length of the three segments of 160 mm. The experiments were conducted with Th$^+$ ions.
The operation principle of a Paul trap relies on an alternating electric quadrupole field. The equations of motion for an ion with mass $m$ and charge $e$ in a segmented linear Paul trap can be described by a set of Mathieu equations [@paul]. The stability of the radial confinement is determined by the dimensionless parameters: $$\begin{aligned}
\label{eq:stab_param}
a_x=\frac{4e}{m\Omega^2}\frac{U_{DC}}{r_0^2}-\frac{4e}{m\Omega^2}\frac{\kappa U_{EC}}{z_0^2}\, , && q_x&=\frac{2e}{m\Omega^2}\frac{U_{RF}}{r_0^2}.\end{aligned}$$ Here $U_{RF}$ is the amplitude of the RF voltage (at angular frequency $\Omega$), $r_0$ and $z_0$ are the radial and axial distances from the trap center to the end of the middle electrode. $U_{DC}$ is a static voltage that may be applied to two electrodes of the middle section in order to lift the degeneracy of the radial secular frequencies. $U_{EC}$ is the static voltage that is applied to all electrodes of the two outer sections for the axial confinement within the middle section with the geometrical factor $\kappa\approx 0.3$. In the experiments reported here a frequency $\Omega \approx {2 \pi}\cdot 2\ \textrm{MHz}$ and $U_{RF}$ between 400 V and 900 V were used, resulting in a $q$-parameter in the range 0.2–0.4.
The ablation target is placed in the central position between the two lower trap electrodes (see Fig. 3). The ablation laser beam passes through the gap between the upper electrodes and is focussed on the target. The grounded ablation target produces additional gradients of the trapping potential, with the effect of repelling the ions along the axis. For ions in the middle section at an axial distance of about 80 mm from the target, this perturbation has no detectable influence. The yield of captured ions depends on the distance of the ablation target from the axis of the trap. It could be increased by about a factor of two by moving the target 1 mm higher, i.e. closer to the trap axis, but with the disadvantage of significantly higher shot-to-shot fluctuations of the ion number. With the target closer to the trap axis, the ions are produced in a region of smaller trap potential, but the trap is also perturbed more strongly by the target. To maintain a high loading efficiency the position of the focus of the ablation laser beam on the target is shifted periodically because the ablation induces changes in the surface profile.
![Schematic of trap electrodes and Th sample. The Th target is installed in the central position between the two lower trap electrodes of the loading section. The ablation laser beam passes through the gap between the upper electrodes. Units are in mm.[]{data-label="fig:electrodes and sample"}](figure3){width="0.50\columnwidth"}
The trap is installed in a low permeability stainless steel vacuum chamber with a diameter of 300 mm. Two CaF$_2$ windows in the top and the bottom of the chamber are facing the middle section for the observation of fluorescence light emitted by the ions. A fused silica window in the top provides access for the ablation laser and additional windows around the circumference of the chamber pass laser beams for excitation of the trapped ions. To maintain a pressure in the range of $10^{-8}$ Pa, a turbomolecular pump is supported by a NEG (Non-Evaporable Getter) pump that removes water, oxygen and hydrogen but does not pump noble gases that will be used for collisional cooling of the ions. A buffer gas system is attached to provide a supply of helium of high purity.
Different methods were applied to estimate the number of trapped ions. A nondestructive and quantitative method uses electronic detection of the mirror current induced in one of the trap electrodes when the ion cloud is resonantly excited to a collective oscillation at the secular frequency [@fischer]. A lower limit of $10^5$ stored ions loaded after 5–10 ablation pulses was deduced from this signal.
A simple quantitative method to measure the ion number is the counting of ions with a channeltron after release from the trap. After a loading and trapping period, the amplitude of the RF voltage is reduced until the ions are extracted towards the channeltron. The opening of the channeltron was facing towards the middle section of the trap perpendicular to the trap axis from a distance of 5 cm. These experiments were performed at buffer gas pressures in the range of $10^{-4}$ Pa, limited for the operation of the channeltron without discharges. Typically, on the order of $10^4$ ions were detected after one ablation pulse, consistent with the result from electronic detection described above. The absolute efficiency of this detection method is uncertain because ions leave the trap not only in the direction towards the channeltron. The main advantages of the method turned out to be the experimental simplicity and its insensitivity with respect to the ion species to be detected. This method was used for the experiments on the RF phase dependence of the loading efficiency that are described in section 5.
Future experiments with this setup will focus on laser excitation of the stored Th$^+$ ions and optical fluorescence detection provides a convenient measure of the relative ion density in the center of the trap. The ions were excited with a tunable extended-cavity diode laser on the strongest resonance line at 401.9 nm which connects the (6*d*$^{2}$7*s*)*J*=3/2 ground state with the (6*d*7*s*7*p*)$ \linebreak $*J*=5/2 state at 24874 cm$^{-1}$. Helium buffer gas at a pressure of $0.2$ Pa was used for collisional cooling and quenching of metastable states. The fluorescence signal was detected at the same wavelength as the excitation with a photomultiplier and was recorded as a function of the number of ablation laser pulses. Figure \[fig:loading\_characteristics\] shows the results for two different DC axial confinement voltages $U_{EC}$. The experiment shows that laser ablation loading in this trap geometry is cumulative over several ablation shots before the ion number saturates. The saturation ion number depended on the trap parameters and but also on the ablation conditions like the laser pulse energy.
![Fluorescence signal of stored Th$^+$ ions for multiple ablation pulses with different DC axial confinement voltages, showing cumulative loading of the trap.[]{data-label="fig:loading_characteristics"}](figure4){width="0.95\columnwidth"}
Investigation of the RF phase dependence of the loading efficiency {#sec:phase}
==================================================================
Theoretically, the trajectory of an ion in a RF trap is described by a solution of the Mathieu equation. The stability of the solution (i.e. whether the distance of the ion from the trap center stays finite for all times) depends only on the values of the parameters $a$ and $q$ (see Eq. (1)). For a real ion trap with a given extension of the trapping volume that is limited by the electrode surfaces, the stability of a trajectory depends on the initial position and velocity of the ion and on the initial phase of the RF field as well. This phase dependence of ion loading has already been studied theoretically in early works, for the rotationally symmetric Paul trap with ions created at rest inside the trap [@fischer], and for ions injected through holes in the electrodes or through the gap between electrodes [@schuessler1; @schuessler2]. In both cases the phase-dependence is found to be pronounced especially for high values of $q$. For an initially homogeneously filled trap volume, Ref. [@fischer] predicts a modulation of the capture efficiency with the RF phase between 100% and 50% for $q=0.28$. Very few experimental studies of this effect have been reported [@hemberger; @robb; @vargas], because the most commonly used method of ion creation via electron impact requires a time that is much longer than the period of the RF field.
In the experiment reported here, laser ablation loading of a linear Paul trap produces ions on the nanosecond timescale and is therefore capable of resolving a phase dependent loading efficiency. Since the pulse timing of the ablation laser showed a considerable jitter of about $3\,\mu$s, it was not possible to trigger it on a definite value of the RF phase. Instead, the phase of the RF field was measured for every ablation shot and the observed ion number on the channeltron was stored and averaged over 45$^\circ$ wide bins. The phase convention $U_{RF}\propto \cos \Omega t$ was used here, i.e. maxima of the RF field strength appear at phases $0$ and $\pm 180$ degree. The $q$ value was chosen in the range that provides the most efficient trapping conditions. The result is shown for trap parameters $q=0.4$ and $a=0$ as the lower-lying curve (triangles) in figure \[fig:breakdown\]. Measurements for $q=0.2$ and $q=0.3$ lead to similar results. With a statistical uncertainty that allowed to resolve a relative change of the ion number of 4%, no significant phase dependence of the loading rate was observed for these parameters.
![Trapped ion number as a function of the timing of the ablation pulse with respect to the RF phase. The triangles show data for events with constant RF amplitude, the black circles include events when the RF amplitude was momentarily reduced (see inset) because of a discharge breakdown that was ignited by photoelectrons from the ablation pulse. The connecting curves are drawn as a guide to the eye.[]{data-label="fig:breakdown"}](figure5){width="0.95\columnwidth"}
In some events close to the field maximum at phase $0$, the ablation pulses were accompanied by a temporary breakdown of the RF field. In these events, the loading efficiency was about 20% higher than in the case without breakdowns. The black circles in figure \[fig:breakdown\] show data that includes those events. This observation can be understood as follows: Free electrons produced by the ablation laser pulse lead to the ignition of a discharge between the trap electrodes and the grounded ablation target. In these events the RF voltage decreased to less than 50% of its steady state value and recovered within about 5–10 cycles of the driving frequency or about 3–5 $\mu$s. A typical time dependence of $U_{RF}$ during a breakdown is shown in the inset of figure \[fig:breakdown\] together with the signal from a photodiode that detects the ablation pulse. Similar effects have been described in [@kwong1; @hashimoto]. The ignition of the discharge did not appear in a completely deterministic way and evidently showed a dependence on the RF phase. It is initiated with highest probability when the trapping voltage is close to its maximum and in one polarity only, i.e. once during the RF cycle. In order to test whether the increased loading efficiency during the RF breakdown was associated with the time dependence of the RF amplitude (as proposed in [@hemberger]), a fast voltage controlled attenuator was used to produce a similar envelope of the RF amplitude, correlated with the laser ablation pulse but without the ignition of a discharge. In this test, no enhancement of the ion loading rate was observed. We conclude that the increase of the trapped ion number that was associated with the discharge was due to further electron impact ionization of neutral Th within the ablation plume.
![Results of a numerical simulation of the trap loading as a function of RF phase for the experimental parameters of Fig. 5 (lower data points). The upper curve is obtained for an ion distribution with ten times reduced initial kinetic energy.[]{data-label="fig:numeric_distribution"}](figure6){width="0.95\columnwidth"}
In order to obtain a more complete understanding of these observations, a numerical simulation of the loading process was undertaken. The trap was modeled as infinitely long and ion trajectories were calculated in two dimensions in the plane orthogonal to the trap axis. The initial velocity distribution of the ions in the expanding ablation plasma was derived from an experiment [@thestrup] performed at laser parameters that were similar to the ones used here, indicating a mean kinetic energy of 100 eV. This is in agreement with observations based on the TOF experiments described in Sections 2 and 3. Simulations were performed for this velocity distribution and for one with the kinetic energy reduced by a factor of 10, to a value below the trap depth. The initial position distribution was assumed to be a point, corresponding to the tip of the ablation target. The simulation did not include any damping of the motion. Collisional damping from helium buffer gas in the investigated pressure range of $10^{-4}$ Pa results in a damping time in the range of seconds [@schwarz], i.e. in the range of $10^6$ RF cycles.
The lower-lying data points in Figure 6 show the calculated capture probability for parameters identical to Fig. 5, obtained from the integration of individual ion trajectories with randomly chosen initial velocity for each value of the phase. The low average capture probability obtained here is in disagreement with the experimental finding of comparable ion numbers in the TOF and trap experiments. For the slower ion distribution (upper curve in Fig. 6), the maximum capture probability approaches unity and a phase-dependent modulation of about 20% is obtained in the simulation. A similar modulation can be seen for the more realistic velocity distribution, though at a lower statistical significance, because of the small overall capture probability.
In comparison with the results from the experiments, it can be seen that the simulation underestimates the capture probability and predicts a phase dependence that is not observed experimentally. Since the initial density of neutral and charged particles in the ablation plume is high [@willmott], velocity changing collisions and Debye shielding may be expected to play an important role in the loading of the trap. Shielding will reduce the influence of the trapping field on some ions as long as the density is high. A collision between atoms of identical mass can lead to strong momentum transfer. Both effects may therefore produce a sub-ensemble of ions close to trap center at a kinetic energy that is lower than expected from the model with noninteracting particles. Inclusion of these effects would make a simulation of the process much more involved. A RF phase dependence of the loading efficiency that also depended on the ion density has been observed in an experiment with organic molecules [@robb].
Conclusion
==========
In conclusion, we have presented a study of the production of metal ions by laser ablation with a compact, low-power nitrogen laser. Significant numbers of singly charged ions could be produced for all the elements studied here. These were selected because they are of interest for ion trap experiments in the domains of optical clocks and quantum information processing. The method also proved to be effective for refractory metals and for thorium. Since ion production occurs on a nanosecond time scale it was possible to investigate the theoretically predicted dependence of the loading efficiency on the phase of the RF field. Contrary to this prediction that was derived for individual ions, no significant phase dependence was observed in the experiment, presumably because of collective processes within the dense ablation plume.
Based on a rough estimate of the vaporized volume of $10^{-11}$ cm$^3$ of metal, i.e. about $10^{11}$ atoms, and a detected ion number below $10^5$ per ablation pulse, an overall ion production and trapping efficiency of order of magnitude $10^{-7}$ to $10^{-6}$ can be deduced for the experiments reported here. This may be compared to an efficiency of about $10^{-8}$ to $10^{-7}$ that has been reported for the loading of Th$^+$ ions from a heated wire source [@kaelber].
As mentioned in the introduction, the production of free electrons can be a critical factor for the operation of the ion trap because of stray charges deposited on insulators. With a work function of 3.5 eV for Th it can be estimated that both, photoelectron emission and thermal electron emission contribute and that an order of magnitude of $10^{10}$ electrons will be produced per pulse. This number is significantly lower than those typically used in loading from an atomic beam with an electron gun and it could be reduced further by using an ablation laser with a photon energy below the work function.
The efficiency in our laser ablation experiments seems to be primarily limited by the low fraction of charged ablation products and by the high kinetic energy of the ion cloud that expands from the ablation plume. A further optimization of the loading efficiency should therefore include a study of the laser heating of the plasma. A significant gain may also be expected by combining laser ablation with photoionization close to the center of the trap that will produce additional ions from ablated neutral atoms.
We thank Chr. Tamm for helpful discussions and D. Griebsch and Th. Leder for their expert technical support. This work was partially supported by DFG within the cluster of excellence QUEST. OAHS acknowledges support from ITCR, MICIT and DAAD.
[99]{}
E. Fischer, Z. Phys. [**156**]{}, 1 (1959)
W. Neuhauser, M. Hohenstatt, P. Toschek, H. Dehmelt, Phys. Rev. Lett. [**41**]{}, 233 (1978)
N. Kjaergaard, L. Hornekaer, A.M. Thommesen, Z. Videsen, M. Drewsen Appl. Phys. B [**71**]{}, 207 (2000)
S. Gulde, D. Rotter, P. Barton, F. Schmidt-Kaler, R. Blatt, W. Hogervorst, Appl. Phys. B [**73**]{}, 861 (2001)
R. J. Hendricks, D. M. Grant, P. F. Herskind, A. Dantan, M. Drewsen, Appl. Phys. B [**88**]{}, 507 (2007)
K. Sheridan, W. Lange, M. Keller, Appl. Phys. B [**104**]{}, 755 (2011)
S. G. Porsev, V. V. Flambaum, E. Peik, Chr. Tamm, Phys. Rev. Lett. [**105**]{}, 182501 (2010)
R. D. Knight, Appl. Phys. Lett. [**38**]{}, 221 (1981)
V. H. S. Kwong, Phys. Rev. A [**39**]{}, 4451 (1989)
Y. Hashimoto, L. Matsuoka, H. Osaki, Y. Fukushima, S. Hasegawa, Jap. Journ. Appl. Phys. [**45**]{}, 7108 (2006)
D. R. Leibrandt, R. J. Clark, J. Labaziewicz, P. Antohi, W. Bakrt, K. R. Brown, I. L. Chuang, Phys. Rev. A [**76**]{}, 055403 (2007)
T. Kwapien, U. Eichmann, W. Sandner, Phys. Rev. A [**75**]{}, 063418 (2007)
C. J. Campbell, A.V. Steele, L. R. Churchill, M.V. DePalatis, D. E. Naylor, D. N. Matsukevich, A. Kuzmich, M. S. Chapman, Phys. Rev. Lett. [**102**]{}, 233004 (2009)
G. C. Eiden, A. W. Garrett, M. E. Cisper, N. S. Nogar, P. H. Hemberger, Int. J. Mass Spectrom. [**136**]{}, 119 (1994)
D. B. Robb, M. W. Blades, Int. J. Mass Spectrom. [**190/191**]{}, 69 (1999)
P. R. Willmott, J. R. Huber, Rev. Mod. Phys. [**72**]{}, 315 (2000)
B. N. Chichkov, C. Momma, S. Nolte, F. von Alvensleben, A. Tünnermann, Appl. Phys. A [**63**]{}, 109 (1996)
B. Thestrup, B. Toftmann, J. Schou, B. Doggett, J. G. Lunney, Appl. Surf. Sci. [**197**]{}, 175 (2002)
L. Laska et al., Plasma Phys. Control. Fusion [**45**]{}, 585 (2003)
W. C. Wiley, I. H. McLaren, Rev. Sci. Instr. [**26**]{}, 1150 (1955)
W. Paul, Rev. Mod. Phys. [**62**]{}, 531 (1990)
Chung-Sing O, H. A. Schuessler, J. Appl. Phys. [**52**]{}, 1157 (1981)
Chung-Sing O, H. A. Schuessler, Appl. Phys. B [**27**]{}, 129 (1982)
R. R. Vargas, R. A. Yost, in: Practical Aspects of ion trap mass spectrometry, Vol. II, Eds.: R. E. March, J.F.J. Todd, CRC Press 1995
S. Schwarz, Lect. Notes Phys. [**749**]{}, 97 (2008)
W. Kälber, G. Meisel, J. Rink, R. C. Thompson, J. Mod. Opt. [**39**]{}, 335 (1992)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Mona Berciu, Ilya Elfimov and George A. Sawatzky'
title: 'Electronic polarons and bipolarons in Fe-based superconductors: a pairing mechanism'
---
[**Superconductivity is a fascinating example of how “more is different” [@PWA]. It is due to electrons binding into bosonic Cooper pairs, which exhibit coherent behavior across a macroscopic sample. Finding the mechanism responsible for this binding is one of the more difficult tasks of condensed matter physics. For conventional superconductors the solution was given by the BCS theory [@BCS] as being due to exchange of phonons. For the cuprate high-T$_c$ superconductors a widely-accepted explanation is still missing despite intense effort. The recently discovered Fe-based high-T$_c$ superconductors pose now a new challenge [@Fe1; @Fe2; @Fe3; @Fe4; @Fe5]. We present here a quantum mechanical theory for pnictides describing the influence of the large electronic polarizability of the heavy anions. We demonstrate that its inclusion results in electronic polarons as the low-energy quasi-particles and also unveils a pairing mechanism for these electronic polarons.** ]{}
At first sight, one may expect the pairing mechanism of pnictides to be related to that of the cuprates, given the somewhat similar layered structures. However, there is an essential difference between the CuO$_2$ layer where the important physics takes place in a cuprate, and its counterpart, the FeAs layer of a pnictide: while the former is two-dimensional (2D), with all Cu and O atoms in the same layer, the latter is not. Instead, the layer hosting the Fe atoms (which are arranged on a simple square lattice) is sandwiched between two layers which share equally the As atoms, as sketched in Fig. \[fig1\](a). Each Fe has 4 nearest neighbor (nn) As atoms at a distance $R=2.4\AA$, arranged in a somewhat distorted tetrahedron, two in the upper and two in the lower layer.
The different geometry has important consequences. In the cuprates, the states near the Fermi energy consist of strongly hybridized Cu $3d$ and O $2p$ orbitals. In contrast, hardly any hybridization appears between the Fe $3d$ and As $4p$ orbitals, with the former giving essentially all the contribution to the low-energy states of the pnictides [@G-old]. This lack of hybridization is due not only to the lattice structure but also to the substantial spread in space of $4p$ orbitals, as compared to $2p$ orbitals.
As a result, one might assume that the simplest Hamiltonian describing the low-energy physics of pnictides is a Hubbard Hamiltonian for the Fe $3d$ electrons, namely: $${ \label{e1} }
{\cal H}_{\rm Fe}= -\sum_{i,j,
\sigma}^{} \left(t_{ij}
c_{i,\sigma}^\dagger c_{j,\sigma} + h.c.\right)+U_H \sum_{i}^{}
\hat{n}_{i\uparrow} \hat{n}_{i\downarrow}$$ where $c^\dagger_{i,\sigma}$ creates an electron on the Fe site $i$ with spin $\sigma$ and $\hat{n}_{i\sigma} =
c^\dagger_{i\sigma}c_{i\sigma}$. For simplicity, here we only count “doping” electrons, i.e. extra charges on top of the $3d^6$ configuration of the Fe in the undoped compound. Of course, a multiple-band Hamiltonian can also be used for the proper description of the various $3d$ orbitals, however the essential physics we want to discuss is already captured within this simpler starting point. The hopping integral is certainly finite for nn Fe sites, with $t_{i, i+x}=t$. We will also consider the effects of including 2$^{nd}$ nn hopping, with $t_{i, i+x+y}=t'$ \[the indexing of various sites is indicated in Fig. \[fig1\](b)\]. The hopping to As sites is neglected, given the small hybridization.
![(a) 3D sketch of the idealized FeAs layer. The lattice constant $a=2.8\AA$ and angle $\theta$ (see text) are indicated; (b) Top view of the same Fe-As layer. Several sites are indexed. The Fe and As atoms are at different “depths”.[]{data-label="fig1"}](FIG1.pdf){width="0.90\columnwidth"}
Of course, Hamiltonian (\[e1\]) is the most common starting point for cuprates (where it does include the O contribution, as $c^\dagger_{i,\sigma}$ are operators for Zhang-Rice singlets [@ZR]). In contrast, using (\[e1\]) or a similar starting point either in the strong-coupling [@S1; @S2; @S3; @S4; @S5; @S6] or weak-coupling limit [@W1; @W2; @W3; @W4] as the low-energy Hamiltonian for pnictides implies that the As anions play no role in their physics.
However, the As atoms are not irrelevant. As pointed out recently in Ref. , these are big, highly-polarizable ions which are strongly influenced by the extra charges in their vicinity. One expects each such charge to be surrounded by polarized As ions, giving rise to [*electronic polarons*]{}. This results in a strong screening of the on-site Coulomb repulsion, suggesting that these materials are not in the large $U$ limit of a Mott-Hubbard insulator. It also results in a strong nn attraction, which may be the key component in the pairing mechanism. The arguments of Ref. are based on semi-classical estimates. Here we propose a quantum model which allows us to investigate this phenomenology away from the linear regime, and to also study dynamic properties of these polarons and bound bipolarons.
The polarization of the As$^{3-}$ ions by the electric fields of the external charges is due to matrix elements coupling their filled $4p$ to (primarily) their empty $5s$ orbitals (more details are in the supplementary material). This is equivalent to a hole being virtually excited from the $5s$ into the $4p$ orbitals. We therefore describe each As by 4 operators: $s^\dagger_\sigma, p_{x,\sigma}^\dagger,
p_{y,\sigma}^\dagger, p_{z,\sigma}^\dagger$, which create a hole of spin $\sigma$ in the respective orbital. The ground-state of an As ion is $s_\uparrow^\dagger s_\downarrow^\dagger|0\rangle$ and the $p$-orbitals lie at an energy $\Omega$ above it, so that the unperturbed As ions are described by: $${ \label{e2} }
{\cal H}_{\rm As} = \Omega \sum_{i,\lambda,\sigma}^{}
p^\dagger_{i,\lambda,\sigma} p_{i,\lambda,\sigma},$$ where $i$ is the location of the As ion \[see Fig. \[fig1\](b)\] $\lambda=x,y,z$ and $\sigma$ is the spin of the hole. Hopping between As ions is very small, given the large nn distance between them of about $4\AA$, and we ignore it. Indeed, LDA calculations find narrow $4p$ bands [@G-old], justifying this atomistic description of the As ions.
We make two approximations regarding the interactions: (i) only As atoms nn to an Fe hosting a charge are polarized by its electric field, and (ii) we ignore dipole-dipole interactions between As clouds. (Estimates of the corresponding corrections are given in the supplementary material). Our interaction Hamiltonian is, then: $$\begin{aligned}
\nonumber {\cal H}_{\rm int} &&= g \sum_{i, \sigma}^{} \hat{n}_i
\left[s_{i,\sigma}^\dagger\left(-\sin \theta p_{i,2,\sigma} + \cos
\theta p_{i,3,\sigma}\right)\right.\\
\nonumber &&+
s_{i-y,\sigma}^\dagger\left(-\sin \theta p_{i-y,1,\sigma} - \cos
\theta p_{i-y,3,\sigma}\right) \\
\nonumber && + s_{i-x-y,\sigma}^\dagger\left(\sin \theta
p_{i-x-y,2,\sigma} + \cos \theta p_{i-x-y,3,\sigma}\right)\\
{ \label{e3} } &&\left.+ s_{i-x,\sigma}^\dagger\left(\sin \theta
p_{i-x,1,\sigma} - \cos \theta p_{i-x,3,\sigma}\right)+ h.c.\right]\end{aligned}$$ where the sum is over all unit cells $i$ in the lattice. This interaction describes the $5s\leftrightarrow 4p$ transitions resulting in the polarization of the As along the appropriate Fe-As direction (see Fig. \[fig1\]) if doping charges, counted by $\hat{n}_i = \hat{n}_{i\uparrow} + \hat{n}_{i\downarrow}$, are on any Fe$_i$ site.
Our model Hamiltonian ${\cal H} = {\cal H}_{\rm Fe} + {\cal H}_{\rm
As} + {\cal H}_{\rm int}$ is characterized by 5 energy scales. The hopping integral $t=0.25$eV and the $5s-4p$ energy difference $\Omega=6$eV are estimated from LDA results (for details on determining all the parameters, see the supplementary material). We use either $t'=0$ or $t'=-t/2$, the latter being the appropriate value for hopping between $d_{xy}, d_{xz}, d_{yz}$ orbitals which contribute most near the Fermi level. The interaction energy $g=2.5$eV is extracted from the As$^{3-}$ polarizability $\alpha_p=10\AA^3$, and the Hubbard onsite repulsion $U_H$ is left as a free parameter. We present results for a wide range of these parameters, proving that our results are only weakly sensitive to their precise values.
![Sketch of a single electronic polaron and its nn hopping. The electric field created by the extra electron residing on the Fe excites holes of nn As into the $4p$ orbital pointing towards the Fe (more distant As atoms stay in the $5s$ ground state). The hopping integral is renormalized by the overlap of the polarization clouds in the initial and final states. []{data-label="fig2"}](FIG2.pdf){width="0.90\columnwidth"}
We begin with a study of the low-energy spectrum of the electronic polaron \[sketched in Fig. \[fig2\]\] created when there is one doping electron in the system. We use first order perturbation theory in the hopping, which is the smallest energy (a quantitative justification and full details of the calculations are provided in the supplementary material). The polaron eigenenergies are: $${ \label{e4} }
E_{P}(\vec{k}) = 4(\Omega -\sqrt{\Omega^2+4g^2}) + \epsilon_{\rm eff}(\vec{k}).$$ The first term is the interaction energy between the electron and the induced As dipole moments. For $g\ll \Omega$ it equals $-4
[\alpha_p E^2/2]\approx $-8.2eV, $E$ being the electric field at the As sites. This linear expression was used in Ref. . Since in fact $g/\Omega \approx 0.4$, non-linear effects become important and reduce it slightly to $\approx-7.1$eV. $\epsilon_{\rm eff}(\vec{k})=- 2 t_{\rm eff} \left[ \cos (k_x
a) + \cos
(k_ya)\right] -4t'_{\rm eff}\cos (k_x a) \cos (k_y a)$ is a dispersion identical to that of a free particle, but with renormalized hopping integrals due to the overlap of the As clouds as the polaron moves between different sites (see Fig. \[fig2\]). Expressions for $t_{\rm eff}$ and $t'_{\rm eff}$ are given in Eqs. (23), (24) in the supplementary material. We plot their values in Fig. \[fig3\], for a wide range of $\Omega$ and $\alpha_p$ values.
![(a) $t_{\rm eff}/t$ and (b) $t'_{\rm eff}/t'$ vs. $\Omega$, for a polarizability $\alpha_p=7, 10$ and $12\AA^3$. The dots show the values used here.[]{data-label="fig3"}](FIG3.pdf){width="0.80\columnwidth"}
For a fixed $\alpha_p$, a smaller $\Omega$ implies a larger $g/\Omega$ ratio \[Eq. (14) in supplementary material\], [*i.e.*]{} a stronger effective coupling. This explains the smaller $t_{\rm eff}$, $t'_{\rm eff}$ as $\Omega$ decreases to be typical polaronic physics: stronger interactions lead to heavier polarons [@polarons]. However, note that these ratios are not exponential in the coupling, like for normal lattice polarons, but rather weakly dependent on the various parameters. This is due to non-linear effects, namely the existence of a maximum allowed value of the induced dipole moment. Both $t_{\rm eff}/t$ and $t'_{\rm eff}/t'$ are around 50% for a wide range of values centered on our parameters, suggesting a light polaron with an effective mass roughly twice as big as the free band carrier.
Such renormalization of the band-structure compared to that predicted by LDA should be detected by ARPES. Interestingly, a recent measurement [@ARPES] showed reasonable fits to LDA upon rescaling the bandwidth by a factor of 2.2, similar to our typical values (the material studied has P instead of As, and these smaller anions have a smaller polarizability $\alpha_p \approx 7\AA^3$). To first order in perturbation theory, we also predict a quasiparticle weight which is independent of $\vec{k}$, with $Z=0.38$ for our parameters. The remaining spectral weight should be observed at much high energies typical of the excited As polarization clouds, of the order ${1\over 2} \left[ \Omega+
\sqrt{\Omega^2 + 4g^2}\right] = 5eV$ and higher (more details are in the supplementary material).
We now study a system with 2 doping electrons, also within first order perturbation in the hopping. We focus on singlet solutions (eigenstates are either singlets or triplets of the two electrons). Also, we first analyze results for $t'=0$, then investigate the role of a finite $t'$.
If there were no interactions between the two polarons, then for a total momentum $\vec{k}$ one would get a continuum of eigenstates with energies $E_P(\vec{k}-\vec{q}) + E_P(\vec{q})$ for all $\vec{q}$ in the Brillouin zone, with the ground-state at $\vec{k}=\vec{q}=0$. However, there are short range interactions between the polarons. If the charges are 2$^{nd}$ nn or closer to each other, some As ions are simultaneously nn to both charges and acquire a dipole moment different from that of As ions which interact with a single charge. This leads to different energies $U_0$, $U_1$ and $U_2$ for configurations with the two charges being, respectively, on-site, nn and 2$^{nd}$ nn to each other. We sketch below the derivation of $U_0$ and $U_1$. The effective hopping between these configurations is also different from $t_{\rm eff}$, because of the overlap with these differently polarized clouds. These interaction energies and some of the effective hoppings are summarized in Fig. \[fig4\] (expressions and plots for all these quantities are provided in the supplementary material).
![Sketches of on-site, $1^{st}$, $2^{nd}$ and $3^{rd}$ nn bipolarons. The first three configurations have interaction energies $U_0, U_1$ and $ U_2$, respectively. Several of the special effective hopping integrals are also indicated.[]{data-label="fig4"}](FIG4.pdf){width="\columnwidth"}
An on-site bipolaron also has 4 polarization clouds, however effectively $g\rightarrow 2g$ in Eq. (\[e3\]), as the on-site charge is doubled, so its electrostatic energy can be immediately obtained from Eq. (\[e4\]) to be $E_{BP,0} = U_H + 4(\Omega
-\sqrt{\Omega^2+16g^2})$. The energy of two static distant polarons is $E_{BP,\infty}=8(\Omega -\sqrt{\Omega^2+16g^2})$ , therefore $U_0 =
E_{BP,0} - E_{BP,\infty} = U_{\rm H}- 4\left[\sqrt{\Omega^2 +16g^2} +
\Omega - 2 \sqrt{\Omega^2 +4g^2}\right]< U_H$ always. This is obvious in the linear regime: because the electric field is twice as large, the energy of the bipolaron is 4 times that of a polaron, [*i.e.*]{} twice as large as of two free polarons. This leads to a significant screening of $U_H$, of several eV, as shown in Fig. \[fig5\](a) (these values are smaller than in Ref. because of non-linear effects). Such a strong on-site attraction is well-known to arise for polarons in general [@Alex].
A nn bipolaron has 6 polarization clouds (see Fig. \[fig4\]), the two central ones being larger than the regular polaron clouds. In the linear regime, the energy of each central cloud is $-\alpha_p (\vec{E}_1+\vec{E}_2)^2/2$, so the energy difference with respect to two regular clouds is $-\alpha_p
\vec{E}_1\cdot\vec{E}_2<0$ if the Fe-As-Fe angle is less than 90$^o$, as is here. This means an attractive nn interaction, as indeed seen for typical parameters in Fig. \[fig5\](b). However, because of non-linear effects, this interaction becomes repulsive at strong coupling, as shown in Fig. \[fig5\](d). Interestingly, this figure reveals that these materials are close to optimal, with $U_1$ near its minimum, for typical values of the parameters (see also Fig. 3 in the supplementary material).
This nn attraction [*is not*]{} typical polaron physics. On a 2D lattice like in cuprates, the interaction between electronic polarons is [*repulsive*]{} because the Cu-O-Cu angle is 180$^o$ (the same holds true for polarons coupled to a breathing-mode phonon of the O atoms). For a Holstein bipolaron, there is a weak nn attraction, but it is due to spin exchange and is ${\cal
O}(t^2)$, not ${\cal O}(t^0)$ like here [@Alex]. Its non-2D geometry is the essential ingredient in bringing about this strong nn attraction, for the Fe-based superconductors. This mechanism is absent in cuprates.
![(a) Renormalization of on-site interaction, $U_0-U_H$; (b) nn energy $U_1$ and (c) $2^{nd}$ nn energy $U_2$ vs. $\Omega$ for various polarizabilities. (d) $U_1$ vs. $g/\Omega$ when $\Omega=4,6,8$eV. The dots show our typical values.[]{data-label="fig5"}](FIG5.pdf){width="\columnwidth"}
![(a) Eigenstates below the two-polaron continuum, $\Delta
E_{BP}= E_{BP}(0)+8t_{\rm eff}$ vs. $U_H$. The symmetry of the three bound eigenstates is labeled in the inset. (b) Probability for on-site, $1^{st}$, $2^{nd}$ and $3^{rd}$ nn separation in the ground state, vs. $U_H$. The dashed lines show the same quantities for the $d$-state. We use $t'=0, \alpha_p=10\AA^3$, $\Omega=6eV$ (similar results are found for all $\alpha_p=7-12\AA^3, \Omega=4-8$ eV). []{data-label="fig6"}](FIG6.pdf){width="0.8\columnwidth"}
First order perturbation in hopping mixes these static configurations. Following the calculation described in the supplementary material, we obtain the eigenergies $E_{BP}(\vec{k})$ for a total momentum $\vec{k}$. In Fig. \[fig6\](a) we plot $\vec{k}=0$ eigenenergies for all bound bipolaron states, [*i.e.*]{} which have energies below the two-polaron continuum starting at $-8t_{\rm eff}$. We find 3 such states. The GS first shows linear dependence on $U_H$ and then flattens out. Its wavefunction has $s$-type symmetry (unchanged sign upon $90^o$ rotation). There is a second $s$-state at small $U_H$, which then joins the continuum. The 3$^{rd}$ bound bipolaron state has an energy independent of $U_H$, and is $d$-type (wavefunction changes sign upon $90^o$ rotation).
The nature of these bound states is revealed in Fig. \[fig6\](b). For low $U_H$ values, the GS primarily consists of an on-site bipolaron, with hardly any contribution from nn or more distant configuration. This explains why its energy here scales with $U_H$ (more precisely with $U_0$). When $U_0\approx 0$ there is a fast crossover to a state dominated by the nn bipolaron configurations (a combined total of 90% probability). This is expected, since $U_1 <0$ irrespective of $U_H$, favoring such a pair when $U_0$ becomes repulsive. The onsite contribution is now exponentially small. This explains the weak $U_H$ dependence here, as coming from virtual hopping to the on-site configuration. The dashed lines show the contributions for the $d$-type state. As expected, it is dominated by nn bipolarons. It has a zero on-site probability, consistent with its symmetry and explaining the lack of dependence on $U_H$. The $2^{nd}$ nn contribution is also zero ($U_2 >0$ as well). There are small $3^{rd}$ and more distant bipolaron contributions.
![Dispersion of the two bound bipolaron states along high-symmetry axes in the Brillouin zone, for (a) $t'=0$ and (b) $t'=-t/2$. The two-polaron continuum is also shown. Parameters are $U_H=10$ eV, $\alpha_p=10\AA^3, \Omega=6$ eV (similar results are found for all $\alpha_p=7-12\AA^3, \Omega=4-8$ eV). The symmetry of the ground state changes from $s$ to $d$ if $t'\ne 0$. []{data-label="fig7"}](FIG7.pdf){width="0.9\columnwidth"}
The unscreened Hubbard repulsion is very large, $U_H\sim 10-20$ eV. For these values there is hardly any dependence on $U_H$, so its precise value is of little importance. We use $U_H=10$ eV from now, and ask how mobile are these bound, predominantly nn, bipolaron pairs. Their dispersion in the Brillouin zone in shown in Fig. \[fig7\](a), where we also show the two-polaron continuum (eigenstates above the continuum are not shown). The $s$-pair has a bandwidth $E_{GS}(\pi,\pi)-E_{GS}(0,0) \approx 0.3$ eV, implying a bipolaron mass about 7 times that of a free carrier mass (the bandwidth of a free electron is $8t=2$eV). This is not a huge enhancement, since it means that the bipolaron is about 3.5 times heavier than a single polaron. In Fig. \[fig8\](b) we plot the bipolaron mass for various $\alpha_p$ and $\Omega$ values, showing only limited variation over a wide parameter range. The higher energy $d$-pair is heavier, with a much narrower bandwidth.
If we include 2$^{nd}$ nn hopping $t'=-t/2$, two effects are apparent. First, the two-polaron continuum is pushed to higher energies, effectively increasing the binding energies of both bound bipolaron states. (The binding energies for the GS bipolaron are shown in Fig. \[fig8\](a) for various parameters). Second, the $d$-state becomes the ground-state. This is not surprising, since the 2$^{nd}$ nn hopping links directly the two nn bipolaron configurations which give the bulk contribution to these eigenstates. Since $t'<0$, this mixing raises the energy of an $s$ state and lowers that of a $d$-state. Thus, if the effective $t'$ between the two nn bipolaron configurations is large enough, the $d$ state has to become the ground-state.
![Ground-state bipolaron (a) binding energy, and (b) effective mass in units of the free carrier mass vs. $\Omega$, for various polarizabilities. The full lines correspond to $t'=0$, dashed lines to $t'=-t/2$. Here $U_H=10$eV. []{data-label="fig8"}](FIG8.pdf){width="0.8\columnwidth"}
Let us now comment briefly on the triplet eigenstates. Since there is no onsite triplet bipolaron configuration, their energies are essentially identical to the energies of the singlet eigenstates in the limit $U_H\rightarrow \infty$. This implies that at this level of approximation and for large enough $U_H$, singlet and triplet bound bipolarons are almost degenerate. However, it is well known that second-order perturbation theory produces an exchange interaction which strongly favors the singlet eigenstates (this is the interaction responsible for binding the S1 Holstein bipolaron, see for example Ref. ). It follows that in reality, the ground-state must be a singlet.
The binding energies shown in Fig. \[fig8\](a) are substantial, even without inclusion of this singlet exchange energy. As discussed in the supplementary material, relaxation of our approximations (that only As ions nn to a charge are polarized and that dipole-dipole interactions are ignored) further enhances $U_1$, and therefore these binding energies, to several eV. Yet more enhancement is expected if we include even higher orbitals than $5s$ when describing the As ion polarization. All this suggests the appearance of pre-formed pairs well above room temperature. On the other hand, in our model Hamiltonian we have ignored a nn repulsion energy which comes from the bare Coulomb interaction itself (in reality, this nn Coulomb repulsion will be reduced by other screening mechanisms, such as the bond polarizabilities involving the As $4p$ and Fe $3s$ and $4p$ states). This nn repulsion will decrease $U_1$ substantially, and may even change its sign, making nn bipolaron pairs unstable. It is difficult to obtain accurate quantitative estimates of all these terms, to find out whether bound bipolaron states exist and what are their binding energies.
It is important to point out that the presence of this As-mediated nn attractive interaction $U_1$ may be essential even if preformed bipolaron pairs do not exist. This would put these materials in a BCS-like framework, with the phonon glue replaced by a virtual excitonic glue. Because $U_1$ is so large one does not really need much of a retardation effect to overscreen the bare Coulomb repulsion. In favor of this scenario is experimental data indicating higher T$_c$ in samples with shorter Fe-As distances and smaller Fe-As-Fe angles, which is precisely what increases the value of $U_1$ (more speculation on this is presented in the supplementary material).
However we do not want to rule out the presence of preformed pairs of the kind discussed above. There is ample evidence that singlet pairs could exist to quite high temperatures in these systems. For example, the magnetic susceptibility [@Dresden] shows the same strong increase with temperature above the spin density wave (SDW) or superconducting transition temperature. This is very difficult to explain if we start from a free single-particle-like picture, but easy to understand if we assume the presence of singlet preformed pairs well above T$_c$, with a binding energy of the order of 100meV or more. The magnetic susceptibility of such a system would increase with increasing $T$ with an activated kind of behavior. NMR Knight shift data also displays this kind of behavior [@Dresden1].
We therefore suggest to take the scenario of preformed singlet bipolarons seriously. We note that a superconducting state would then behave more like a Bose Einstein condensate. At first glance one might think that its T$_c$ would have to be very low, however we note that these bipolarons are very light, with a mass which is about 3-4 times the single polaron mass, and thus very high condensation temperatures could be expected. The other rather interesting aspect to consider is that perhaps this scenario might also explain the low amplitude SDW observed at low dopings, as being a different kind of condensate of singlet bipolarons due to rather strong exchange interactions. The low amplitude would be a result of the pair-wise singlet formation tendency competing with the pair-pair exchange interactions which would favor the SDW. We propose to study these issues next.
Acknowledgments: We thank J. Zaanen, J. van den Brink, C. Varma, D. Bonn and A. Damascelli for many stimulating discussions and insightful opinions. This work was supported by NSERC, by CIfAR Nanoelectronics and Quantum Materials, and by the Killam Trust and the Alfred P. Sloan Foundation (M.B.).
Competing interests statement: The authors declare no competing financial interests.
[99]{}
Anderson, P. W. More is different. [*Science*]{} [ **177**]{}, 393-396 (1972). Bardeen, J., Cooper, L. N., &Schrieffer, J. R. Microscopic Theory of Superconductivity. [*Phys. Rev.* ]{}[ **106**]{}, 162 - 164 (1957).
Kamihara Y. [*et al.*]{} Iron-Based Layered Superconductor: LaOFeP, [*J. Am. Chem. Soc.*]{} [**128**]{}, 10012 -10013 (2006).
Kamihara Y. [*et al.*]{} Iron-Based Layered Superconductor La\[O$_{1-x}$F$_x$\]FeAs (x = 0.05-0.12) with T$_c$ = 26 K. [*J. Am. Chem. Soc.*]{} [**130**]{}, 3296 -3297 (2008).
Takahashi, H. [*et al.*]{} Superconductivity at 43 K in an iron-based layered compound LaO$_{1-x}$F$_x$FeAs. [*Nature*]{} [**453**]{}, 376-378 (2008).
Chen, X. H. [*et al.*]{} Superconductivity at 43 K in SmFeAsO$_{1-x}$F$_x$. [*Nature*]{} [**453**]{}, 761-762 (2008).
Chen, G. F. [*et al.*]{} Superconductivity at 41 K and Its Competition with Spin-Density-Wave Instability in Layered CeO$_{1-x}$F$_x$FeAs. [*Phys. Rev. Lett.*]{} [**100**]{}, 247002 (2008).
Sawatzky, G. A. [*et al.*]{} Heavy anion solvation of polarity fluctuations in Pnictides. Preprint at $\langle$http://arxiv.org/abs/0808.1390$\rangle$ (2008).
Zhang, F. C. & Rice, T. M. Effective Hamiltonian for the superconducting Cu oxides. [*Phys. Rev. B*]{} [**37**]{}, 3759 - 3761 (1988).
Cao, C. [*et al.*]{} Proximity of antiferromagnetism and superconductivity in LaO$_{1-x}$F$_x$FeAs: effective Hamiltonian from ab initio studies. [*Phys. Rev. B*]{} [**77**]{}, 220506 (2008).
Haule, K. [*et al.*]{} Correlated electronic structure of LaO$_{1-x}$F$_x$FeAs. [*Phys. Rev. Lett.*]{} [**100**]{}, 226402 (2008).
Yin, Z. P. et al. Electron-hole symmetry and magnetic coupling in antiferromagnetic LaOFeAs. [*Phys. Rev. Lett.*]{} [ **101**]{}, 047001 (2008).
Si, Q. &Abrahams, E. Strong Correlations and Magnetic Frustration in the High T$_c$ Iron Pnictides. [*Phys. Rev. Lett.*]{} [**101**]{}, 076401 (2008).
Xu, C. [*et al.*]{} Ising and spin orders in iron-based superconductors. [*Phys. Rev. B* ]{} [**78**]{}, 020501 (2008).
Chen, W. [*et al.*]{} Strong Coupling Theory for Superconducting Iron Pnictides. Preprint at $\langle$http://arxiv.org/abs/0808.3234$\rangle$ (2008).
Maier, T. A. & Scalapino, D. J. Theory of neutron scattering as a probe of the superconducting gap in the iron pnictides. [*Phys. Rev. B*]{} [**78**]{}, 020514(R) (2008).
Raghu, S. [*et al.*]{} Minimal two-band model of the superconducting iron oxypnictides. [*Phys. Rev. B*]{} [**77**]{}, 220503(R) (2008).
Lee, P. A. & Wen, X.-G. Spin-triplet p-wave pairing in a 3-orbital model for FeAs superconductors. Preprint at $\langle$http://arxiv.org/abs/0804.1739$\rangle$ (2008).
Singh, D. J. & Du, M.-H. Density functional study of LaFeAsO$_{1-x}$F$_x$: a low carrier density superconductor near itinerant magnetism. [*Phys. Rev. Lett.*]{}[**100**]{}, 237003 (2008).
Goodvin, G. L., Berciu, M. & Sawatzky, G. A. The Green’s function of the Holstein polaron, [*Phys. Rev. B*]{} [**74**]{}, 245104 (2006).
Lu, D. H. [*et al.*]{} Electronic structure of the iron-based superconductor LaOFeP. [*Nature*]{} [**455**]{}, 81-84 (2008).
Macridin, A., Sawatzky, G. A. & Jarrell, M. Two-dimensional Hubbard-Holstein bipolaron. [ *Phys. Rev. B*]{}[**69**]{}, 245111 (2004).
Klingeler, R. [*et al.*]{} Evidence for antiferromagnetic correlations in superconducting LaFeAsO$_{1-x}$F$_x$. Preprint at $\langle$http://arxiv.org/abs/0808:0708$\rangle$ (2008).
Grafe, H.-J. [*et al.*]{} $^{75}$As NMR studies of superconducting LaO$_{0.9}$F$_{0.1}$FeAs. [*Phys. Rev. Lett.* ]{}[ **101**]{}, 047003 (2008).
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'DIPTBTP.bib'
---
Adjunct Professor of Finance, Stuart School of Business, Illinois Institute of Technology and Principal, Market Pattern Research, Inc.
Alumni Professor of Financial Modeling and Stochastic Optimization (Emeritus), University of British Columbia, Vancouver, BC, and Distinguished Visiting Research Associate, Systemic Risk Centre, London School of Economics, UK
Introduction
============
[@moffitt:ziemba:2017a] show that expected returns of $10\%$-$25\%$ can be achieved under certain conditions from betting all the tickets in a lottery that pays its entire jackpot in equal shares to winning ticket holders. For many large government lotteries, this strategy of “buying the pot” is not feasible because the logistical problems are insurmountable. In the California Powerball Lottery, for example, the number of ticket combinations is over $175,000,000$ and the rules do not allow betting large numbers of combinations on single paper tickets.
The Canadian 6/49 Lotto, however, has a large but manageable number of ticket combinations ($13,983,816$) and allows paper tickets that have multiple combinations. The 6/49 is played in other countries, including the UK. Here we focus on the Canadian version.
The purpose of this paper is threefold: (1) to modify the *pure jackpot model* in [@moffitt:ziemba:2017a] to accomodate the irregular payout features of the Canadian 6/49 Lotto, (2) to derive conditions under which the expected return from buying the pot is positive, and (3) to discuss the implications of our findings for lottery design.
Previous Work and Instances of Buying the Pot {#S:PreviousWorkAndInstancesOfBuyingThePot}
=============================================
Each lottery has the following rules — players buy tickets and the winning ticket is selected using an equiprobable drawing. Those who hold the winning ticket share equally in a jackpot that consists of a carryover pot from the previous lottery plus an after tax portion the monies wagered. If there is no winner, the jackpot pool carries over to the next drawing. There can be multiple carryovers.
[@moffitt:ziemba:2017a] use the following assumptions and notation to analyze the *pure jackpot model*:
- Each lottery has $t$ tickets costing $\$1$ apiece.
- A single winning ticket $w$, $1 \le w \le t$ is drawn from $i=1,\ldots,t$ using probabilities $p_i = 1/t$.
- The syndicate buys one of each ticket for a total of $t$ tickets, and $c$ individuals (the “crowd”) independently buy one ticket apiece using probabilities $q_i$, $1 \le i \le t$.
- A cash jackpot $v = a + (t + c)(1 - x)$ is awarded in equal shares to all holders of the winning ticket $w$, where $a \ge 0$ is the current carryover from the previous lottery draws, $c$ is the number of tickets bet by the crowd, and $x$ is the the (fractional) take.
[@moffitt:ziemba:2017a] show the following for the pure jackpot model:
1. [*Recursion:*]{} When $t$ and $c$ are large, $q_i = 1/t$ for each $i$, and $X$ is the random number of winning tickets held by the crowd, the expected value $E\left[ \frac{n}{n+X} \right]$, $n$ an integer $\ge 1$, is to close approximation equal to $$\label{E:EVrecursion}
E\left[ \frac{n}{n+X} \right] = \begin{cases}
\frac{1}{\lambda(c)} \left( 1 - e^{-\lambda(c)} \right) & n=1 \\
\frac{n}{\lambda(c)} \left\{ 1 - E\left[ \frac{n-1}{n-1+X} \right] \right\} & n > 1,
\end{cases}$$ where $\lambda(c) = c/t$. \[Enum:Recursion\]
2. [*Condition under which Buying the Pot has Positive Expected Return:*]{} The expected gain for a syndicate that bets one of each ticket is positive $$(a + (t + c)(1 - x)) \, E\left[ \frac{1}{1+X} \right] \, - \, t \, > 0 \label{E:FairSplitCondition}$$ provided that $a/(t + c) - x \ge 0$. Since $a/(t + c) - x$ is the after tax value of a ticket assuming the pot $a$ is fairly split, this condition implies that a syndicate earns more than a fair split of the jackpot. In a lottery with no take, the returns to the syndicate typically range between 10% and 25%. \[Enum:BTPEdge\]
3. [*Optimal Strategies:*]{}
1. The best returning strategy for the crowd consists of using $q_i = 1/t$ for each $i$.
2. Let $E_q[X/(1+X)]$ be the expectation for a crowd that bets with probability vector $q = (q_1, \ldots, q_t)'$, and let $1_t/t$ be the probability t-vector that has $1/t$ for each entry. Then if $q \ne 1_t/t$ $$E_q[X/(1+X)] < E_{1_t/t}[X/(1+X)]. \label{E:OptimalityOfCrowdProportionalBetting}$$
\[Enum:OptimalStrategy\]
Several studies of lottery strategy and design have appeared in the economic literature. [@chernoff1980analysis; @Chernoff1981] studied the Massachusetts Numbers Game, proposing that playing unpopular numbers might be a winning strategy. However, the results from a test were disappointing because of learning (unpopular numbers became less unpopular) and gambler’s ruin (betting funds were exhausted). [@ziemba1986dr] carry this further and study various Canadian lotto games, their unpopular numbers and the uniformity of betting. [@10.2307/2290349] has additional discussion of this latter point and [@10.2307/3314913], [@10.2307/2290073] and [@Ziemba2008183] further analyze unpopular numbers. [@citeulike:1337256] review the behavioral evidence in efficient markets for a persistence of betting at unfavorable odds. [@RePEc:inm:ormnsc:v:38:y:1992:i:11:p:1562-1585] investigate the use of Kelly optimal wagering on unpopular tickets and find that this strategy has positive expectation, but the waiting time to achieve reliable gains with high probability is millions of years! [@10.2307/1942724] discuss behavioral bases of betting and along with [@Walker2008459], discuss design considerations for lotteries. None of these studies consider the strategy that in a short time achieves reliable gains — buying the pot.
There are anecdotal accounts of successful buyings of the pot. One putative attempt involved a syndicate that tried but failed to buy all tickets. But they were lucky, having had time to bet only about $70\%$ of all tickets according to one source and $85\%$ according to another ([@NYTimes:US:BuyThePot1992]). The syndicate ostensibly bet about \$5 million and won about \$27 million.
There are examples of when it was optimal to buy the pot or betting was advantageous. In June 1984 four western Canadian provinces jointly ran the *Lotto West 6/8/56*, in which players choose six numbers from a field of 56, but eight winning numbers and a bonus number are drawn. The jackpot is shared among all tickets that select six of the eight drawn, second price among all that had five of six, and other payouts to those having five of six plus the bonus, four of six or three of six. These rules make the jackpot about twelve times easier to hit than the 6/49 Lotto: 1 in $1,159,587$ (See [@ziemba1986dr]).
In 1987, the provinces went their own ways, at which time the BC Lotto Corporation had about \$10 million in unclaimed prize money. Rather than donate it, they created a version of Lotto 6/8/56 to give it back on March 27, 1987. As before eight numbers were drawn from 56, but players could now choose 1, 2, 3, 4, 5 or 6 numbers on a ticket. A schedule of payouts was published for 1/1, 2/2, /3/3, 4/4, 5/5 and for 3/6, 4/6, 5/6 and 6/6. With these payouts, the expected payback on a \$1 ticket was $\$0.385$. To promote the game, the Corporation offered six tickets for the price of one, for an expected return of \$0.385 times 6, or \$2.31, a $131\%$ edge ([@drz:ColumnsOnRacing]). Ziemba and colleagues at the University of British Colombia knew that individual tickets had a positive expected return, and in a makeshift effort, they bought about 13,000 of the combinations. They made a nice return, but spent hours buying and then locating the winning tickets. Some $\$3.5$ million was paid out of the unclaimed prize fund.
A game where it was optimal to buy the pot was the 5/40 Lotto played in British Colombia and Rhode Island. Ninety-one percent of the net pool went to 5/5 with a minimum shared pool of $\$150,000$ and maximum of $\$300,000$. There were small prizes for 1+, 2+, 3, 4 and 4+ where “+” means getting the sixth bonus number correct. There were $658,008$ combinations. But the jackpot that had built up slowly fell because the public viewed it as unwinnable, so it became a prime target for buying the pot; see [@drz:ColumnsOnRacing].
Rules of the 6/49 Lotto {#S:RulesOfThe649Lotto}
=======================
A *ticket* in the 6/49 Lotto is a unique choice of $6$ different numbers from integers $1$ to $49$. Thus the total number of tickets is the number of combinations of $49$ things taken $6$ at a time: $$t = \binom{49}{6} = \frac{49!}{43! 6!} = 13, 983, 816. \label{Nbr-6/49-Tickets}$$
The Canadian 6/49 Lotto holds drawings twice a week and lumps together the monies wagered for purposes of awarding prizes, whose allocation is described below. On the drawing day, 6 numbers (the “*winning numbers*”) are selected equiprobably and without replacement from 1, 2, …49. Following that, a $7^{th}$ “*bonus number*” is selected.
We introduce notation to describe types of prize-wining tickets. A $x$/6- ticket is one that contains exactly $x$ of the six winning numbers but does not contain the bonus number and a $x$/6+ ticket is one that contains exactly $x$ of the 6 numbers plus the bonus number. A x/6 ticket contains x of the 6 numbers, irrespective of the status of the bonus number; it is therefore a union of types x/6- and x/6+. A 5/6-, for example, contains exactly 5 of the 6 winning numbers with the other not being the bonus number, and a 5/6+ ticket contains exactly 5 of the 6 winning numbers plus the bonus number. For example, if the six numbers drawn were 46, 13, 4, 21, 38, 25 and the bonus number was 43 then ticket 1-4-20-21-32-43 would be a 2/6+ ticket because it contains 4 and 21 from the six plus the bonus number. Similarly, ticket 4-13-21-25-43-46 would be a 5/6+ ticket.
Rules for the Original Lottery: 1982-2004
-----------------------------------------
Table \[Ta:PrizeAllocIn649:1982-2004\] has the initial 6/49 payout scheme (1982-2004) for 3/6, 4/6, 5/6-, 5/6+ and the *Jackpot* 6/6. The cost of a single ticket was \$1, with the lottery sponsors taking 55% of each daily *betting pool* and committing the remaining 45% (the “*prize pool*”) for player payouts. The 45% *prize pool* was allocated as follows: all 3/6 tickets were paid \$10, and the remainder was paid to holders of 4/6, 5/6-, 5/6+ and 6/6 using percentage allocation rules in Table \[Ta:PrizeAllocIn649:1982-2004\]. That game is analyzed thoroughly in [@ziemba1986dr]. See also [@10.2307/2290073]. For other analyses of such games, see [@citeulike:1337256] and [@Haigh2008481].
[lrlll]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
Prize & Combinations & Probability & Allocation Rule & Type\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
$6/6$ & $1$ & $p_1 \sim 7.1e$-$08$ & 50% of the Pools Fund & Share\
$5/6+$ & $6$ & $p_2 \sim 4.3e$-$07$ & 15% of the Pools Fund & Share\
$5/6-$ & $252$ & $p_3 \sim 1.8e$-$05$ & 12% of the Pools Fund & Share\
$4/6$ & $13,545$ & $p_4 = 0.000969$ & 23% of the Pools Fund & Share\
$3/6$ & $246,820$ & $p_5 = 0.017650$ & $\$10$ per ticket & Fixed\
No Win & $13,723,192$ & $p_8 = 0.981362$ & Non-winner = \$0 & Returns 0\
------------------------------------------------------------------------
\
Figure \[G:PoolShare\] shows the leveraging effect of the fixed \$10 3/6 prize when there are average numbers, popular numbers, and unpopular numbers. The impact of popular vs. unpopular numbers selected in the drawing is significant, producing a 17% versus a 36% jackpot share. The large prizes 5/6-, 5/6+ and 6/6 for unpopular numbers in the drawing are typically seven times larger than for popular ones. See examples in [@ziemba1986dr].
![[]{data-label="G:PoolShare"}](lottery649-popularity.pdf)
Rules for the Current Lottery: 9/18/2013 -
-------------------------------------------
In the 6/49 Lotto, new rules were introduced in June, 2004 and again on September 18, 2013. We discuss only the latter rules. These included (1) a single ticket cost of \$3, (2) three fixed prizes, the same 3/6 paying \$10, a 2/6+ paying \$5 and a 2/6- that earns a free play at the next drawing, and (3) altered payout percentages for 4/6, 5/6-, 5/6+ and 6/6 (Table \[Ta:PrizeAllocIn6492013-\]), (4) an increase in the take from 55% to 60%, and (5) a greater allocation to 6/6 winners. The intention of these changes was to increase sales by growing jackpots faster, and creating of many small consolation prices (2/6-, 2/6+ and 3/6). This is a typical convex prize structure where most of the daily payout goes to the smallest (to make them feel that the lottery is winnable) and to the largest (to show that a huge gain can be made). Ziemba has used this in lottery consulting over the years. [@RePEc:eee:jfinec:v:13:y:1984:i:2:p:253-282] call this a “silver lining” for non-winners.
We call the number of tickets bet at a drawing (twice a week in the 6/49), the *ticket pool*, contributors to which are the crowd in amount $c$ and the syndicate in amount $t$. Thus the total number of tickets bet is $c + t$. The *betting pool* $d_{\!_{BP}}$ is the total number of dollars contributed by the bettors. The *betting pool* is divided among the lottery sponsors and the bettors as follows:
[*Sponsors.*]{} Sponsors (the state, the lottery organization) receive $0.60*d_{\!_{BP}}$, with the remaining $0.40*d_{\!_{BP}}$, the *prize pool*, awarded as prizes or added to the carryover pool as indicated below. The “lottery take” $0.60*d_{\!_{BP}}$ is used to cover expenses of running the lottery and to provide funds for community and government services and for donations. The lottery itself, however, is run by a non-governmental company.
[*Prize Distribution.*]{} The *prize pool* has eight classes ($i = 1, 2, \ldots, 8$) of payouts grouped into four types: (A) fixed dollar (2/6+ and 3/6), (B) free play in the next lottery (2/6-), (C) payouts that split among 4/6, 5/6-, 5/6+ and 6/6 tickets the remaining *prize pool* after deductions for type (A) and (B) payouts, and (D) non-winner tickets that receive no payout.
Table \[Ta:PrizeAllocIn6492013-\] details these payouts by showing in the first column the type of ticket, in the second column a notation for the number of each class determined after the random, equiprobable drawing of 6 numbers and a bonus, the third column showing the notation for the class, the fourth column the number of tickets matching a randomly drawn ticket, the fifth column having the probability that a randomly chosen ticket is in the class, the sixth column having the allocation rule and the last, whether the ticket payout is fixed, shared as part of a pool, or returns nothing.
The 2/6+ and 3/6 tickets receive \$5 and \$10, respectively, and 2/6- tickets receive a free play in the next lottery, but a charge of \$1.41 is applied to the *prize pool*. See Example \[Exmp:ExampleOfPrizePayouts\] for details. The payouts are shown in the first four lines of the table for 4/6, 5/6-, 5/6+ and 6/6 tickets. These type (C) tickets share the remainder of the $0.40*d_{\!_{BP}}$ after deductions for tickets of types (A) and (B). The amount $0.40*d_{\!_{BP}} - (\text{payouts to 2/6+, 3/6 and charges for 2/6-})$ is called the *Pools Fund*. Type (C) tickets share in a pool whose percentage of the total bets varies greatly, depending on the winning numbers of 2/6+, 3/6 and free plays. The lottery also guarantees a \$5,000,000 pool to holders of 6/6 tickets.
[lclrlll]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
& \# Crowd & & \# Combinations & & Allocatio & Share\
Type & Tickets & Class & for any ticket & Probability & Rule & Status\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
(C) & $N_1$ & $6/6$ & $1$ & $p_1 \sim 7.1e$-$08$ & 79.5% of the Pools Fund & Share\
(C) & $N_2$ & $5/6+$ & $6$ & $p_2 \sim 4.3e$-$07$ & 6% of the Pools Fund & Share\
(C) & $N_3$ & $5/6-$ & $252$ & $p_3 \sim 1.8e$-$05$ & 5% of the Pools Fund & Share\
(C) & $N_4$ & $4/6$ & $13,545$ & $p_4 = 0.000969$ & 9.5% of the Pools Fund & Share\
(A) & $N_5$ & $3/6$ & $246,820$ & $p_5 = 0.017650$ & $\$10$ per ticket & Fixed\
(A) & $N_6$ & $2/6+$ & $172,200$ & $p_6 = 0.012314$ & $\$5$ per ticket & Fixed\
(B) & $N_7$ & $2/6-$ & $1,678,950$ & $p_7 = 0.120064$ & free play ($\$1.41$ deduction) & Fixed\
(D) & $N_8$ & No Win & $11,872,042$ & $p_8 = 0.848984$ & Non-winner = \$0 & Returns 0\
------------------------------------------------------------------------
\
Any unclaimed monies in the *Pools Fund* are added to the current jackpot and carried over to the next drawing. From Table \[Ta:PrizeAllocIn6492013-\], it is clear that the majority contribution to the carryover is the $79.5\%$ that occurs when there is no 6/6 winner. But 5/6+ and 5/6- tickets also have low probabilities of occurring and when there are no 5/6+ or 5/6- winners, those shares of 6% and 5%, respectively, are added to the carrover pool for the next lottery.
The probabilities of these tickets occurring in an equiprobable lottery are denoted by $p_1$, $p_2$, etc. This notation is useful in the analytical expressions developed below.
\[Exmp:ExampleOfPrizePayouts\]
The carryover is \$30,000,000 and the crowd bets $10,000,000$ tickets, of which $1,000,000$ are assumed to be free plays, yielding a net cash contribution of $\$27,000,000$. Assuming the crowd chooses quick picks with probabilities of $1/t$, numbers for each ticket will have a binomial distribution. Random selections under the binomial are displayed in column 3 of Table \[Ta:ExamplePrizeAllocIn6492013-\], which gives the probability of this class of ticket winning. The first column is the winning ticket type, the second, the number of combinations, the fourth the total payouts to the crowd, the fifth the number of tickets held by the syndicate and the sixth, the total payouts to the syndicate. Using these numbers, we calculate the *prize pool*, the fixed payouts to crowd and syndicate, and the pools fund as follows:
- Prize Pool: $\$27,580,579 = 0.40 * (13,983,816 * 3 \, + \, 0.90*30,000,000)$.
- Crowd Fixed: $\$4,077,490 = 176,933*\$10 + 123,569*\$5 + 1,198,805*\$1.41$.
- Syndicate Fixed: $\$5,696,520 = 246,820*\$10 + 172,200*\$5 + 1,678,950*\$1.41$.
- Pools Fund: $\$17,806,569 = \$27,580,579 - \$4,077,490 - \$5,696,520$.
[lrrrrr]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
& & \# Crowd & Crowd & \# Syndicate & Syndicate\
Type & Combinations & Tickets & Payout & Tickets & Payout\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
$6/6$ & $1$ & 0 & \$0 & 1 & \$44,156,222\
$5/6+$ & $6$ & 6 & \$534,135 & 6 & \$534,135\
$5/6-$ & $252$ & 185 & \$375,960 & 252 & \$514,534\
$4/6$ & $13,545$ & 9,773 & \$708,909 & 13,545 & \$982,518\
$3/6$ & $246,820$ & 176,933 & \$1,769,330 & 246,820 & \$2,468,200\
$2/6+$ & $172,200$ & 123,569 & \$617,845 & 172,200 & \$861,000\
$2/6-$ & $1,678,950$ & 1,198,805 & \$0 & 1,678,950 & \$0\
------------------------------------------------------------------------
\
Summing all payouts in the syndicate payout column gives \$49,516,609, for a gain of $$\$7,565,161 = \$49,516,609 - \$3*13,983,816.$$ plus 1,678,200 free plays in the next lottery. The cash payout from non-6/6 tickets is \$5,360,387, despite a crowd and syndicate bet of \$68,951,454. Clearly, the jackpot must be large in order for buying the pot to be justifiable.
Expected Return from Buying the Pot {#S:ExpectedReturnFromBuyingThePot}
===================================
Notation and Terminology
------------------------
Table \[Ta:FixedParametersForThe649\] gives the fixed parameters of the lottery, namely those that do not involve betting strategies of the syndicate or crowd.
[ll]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
Notation & Description\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
$t$ & Number of tickets in the lottery = 13,983,816.\
$a$ & Carryover pool in dollars, $a \ge 0$.\
$p_i$ & Probability that a ticket is of class $i$ assuming that\
& the winning ticket is drawn equiprobably (see Table \[Ta:PrizeAllocIn6492013-\]).\
$f$ & Fraction of tickets that are “free plays.”\
$c$ & Number of tickets bet by the crowd.\
------------------------------------------------------------------------
\
Table \[Ta:RandomVariablesThatInvolveBettorStrategies\] has the notation for the random variables that account for stochasticity and strategy in playing the lottery.
[ll]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
Notation & Description\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
$N_i$ & Random variable for the number of tickets of class $i$ bet by\
& the crowd (see Table \[Ta:PrizeAllocIn6492013-\]).\
$N$ & Vector $N = (N_1, N_2, \ldots, N_8)'$.\
$d_{\!_{AB}}$ & Dollars awarded or deducted for tickets of types (A) and (B).\
$d_{\!_{BP}}$ & Dollars in the *betting pool*.\
$d_{\!_{PP}}$ & Dollars in the *prize pool*.\
$d_{\!_{PF}}$ & Dollars in the *Pools Fund*.\
------------------------------------------------------------------------
\
Using the notation in Tables \[Ta:FixedParametersForThe649\] and \[Ta:RandomVariablesThatInvolveBettorStrategies\], the number of dollars in each fund is $$\begin{aligned}
d_{\!_{AB}} &= 10 (N_5 + t p_5) \, + \, 5 (N_6 + t p_6) \, + \, 1.41 (N_7 + t p_7) \label{dAB} \\
d_{\!_{BP}} &= 3 ( t + c (1 - f)) \label{dBP} \\
d_{\!_{PP}} &= d_{\!_{BP}} 0.4 \label{dPP} \\
d_{\!_{PF}} &= d_{\!_{PP}} - d_{\!_{AB}} \label{dPF}\end{aligned}$$
Since $f$ is non-random, the second entry of the above table is the only non-stochastic entry.
Equiprobable Betting by the Crowd {#S:EquiprobableBettingByTheCrowd}
---------------------------------
We calculate first the expected return to a syndicate that buys the pot when the crowd chooses tickets independently and equiprobably. As we discuss in Section \[S:Non-equiprobableBettingByTheCrowd\], this is the crowd’s optimal strategy, although they do not employ it in practice — and the cost of this “mistake” is considerable.
### Syndicate’s Expected Value for Equiprobable Crowd Betting
In Appendix \[S:TheSyndicate’sExpectedValueWhenTheCrowdBetsEquiprobably\], we develop a formula for the syndicate’s expected gain $G(c)$ from the wagering of $\$41,951,448 = \$3 * 13,983,816$ on $13,983,816$ tickets: $$\begin{aligned}
E[ \, G(c) \, ] & = (a + 0.795*\mu(c)) \, \lambda(c)^{-1} (1 - exp(-\lambda(c))) \label{E:SyndicateExpectedReturn} \\
& + \left( 0.06 \nu(c) \, + \, 0.145 \frac{1}{1 + c/t} \right) \mu(c) \nonumber \\
& + \$3,329,200 \, - \, \$3 t, \nonumber\end{aligned}$$ where $$\begin{aligned}
\lambda(c) &= c/t, \\
\mu(c) &= 0.40 \cdot 3 \left( ( t + c \cdot (1 - f)) \right) - \, (t + c) \cdot (p_5 \$10 \, + \, p_6 \$5 \, + \, p_7 \$1.41), \\
\nu(c) &= E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right], \qquad \text{for } X \sim Bin(c,6/t).\end{aligned}$$ The term $\nu(c)$ is calculated using the recursive formula in Appendix \[S:TheSyndicate’sExpectedValueWhenTheCrowdBetsEquiprobably\] and appears as the last column of Table \[Ta:BettingThresholds\].
### Parameters that Lead to Positive Expected Return {#SSS:ParametersThatLeadToPositiveExpectedReturn}
Consider the implications of the 6/49 rules and of Formula . Because the lottery sponsors take such a high percentage of the betting pool (60%), a large jackpot is needed for a syndicate to have a positive expected return. When a syndicate bets one of each ticket, previous analysis showed the syndicate’s numbers of winning tickets are known exactly, irrespective of the winning numbers from the drawing. There will always be exactly 1 winning ticket, exactly 6 5/6+ tickets, exactly 252 5/6- tickets, and so on. The RHS of first line of formula dominates the others when a jackpot $a$ is large.
Table \[Ta:BettingThresholds\] shows the results of applying formula for 10 levels of total crowd betting (*c*) to solve for the sizes of carryover pools (*a*) that produce expected returns of 0%, 10% and 20% for the syndicate. Since the cost of buying the pot is $\$3 * 13,983,816 =\$41,951,448$, a return of 10% is $\$4,195,145$. When the crowd bets \$30 million, for example, any carryover larger than \$36.92 million is a potential play for the syndicate, and carryovers of \$42.80 and \$48.67 million have expected returns of 10% and 20%, respectively. The last three columns of Table \[Ta:BettingThresholds\] provide insight into the payout structure. The sixth column shows the expected amount in the *Pools Fund* and the next column is its percentage in the *prize pool*. Thus when the crowd bets \$40 million, the expected *Pools Fund* is \$19.97 million, which is 49.68% of the prize pool. Thus, the charges for fixed payout tickets amount to \$50.32% of the *prize pool*. The final column is the expected value for the 5/6+ factor: $$\text{EV56+} = E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right]. \label{E:ExpValOf5/6+Factor}$$ The expected value declines when the crowd bets more, as one expects since $X_{\text{5/6+}}$ is generally larger.
[rccccccc]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
& Crowd & & Expected & (Pools Fund) &\
& \$ Bet & Breakeven & +10% & +20% & Pools Fund & (Prize Pool) & Eqn.\
& (millions) & (millions) & (millions) & (millions) & (millions) & (%) & (%)\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
3.3 & 9 & 30.33 & 35.05 & 39.76 & 13.28 & 58.39 & 82.71\
6.7 & 18 & 33.46 & 38.74 & 44.01 & 15.51 & 54.02 & 70.15\
10.0 & 27 & 36.92 & 42.80 & 48.67 & 17.74 & 51.32 & 60.71\
13.3 & 36 & 40.71 & 47.22 & 53.73 & 19.97 & 49.68 & 53.40\
16.7 & 45 & 44.81 & 51.99 & 59.17 & 22.20 & 48.73 & 47.60\
20.0 & 54 & 49.21 & 57.10 & 64.99 & 24.44 & 48.27 & 42.90\
23.3 & 63 & 53.90 & 62.52 & 71.15 & 26.67 & 48.14 & 39.02\
26.7 & 72 & 58.84 & 68.24 & 77.63 & 28.90 & 48.25 & 35.77\
30.0 & 81 & 64.03 & 74.23 & 84.42 & 31.13 & 48.53 & 33.01\
33.3 & 90 & 69.45 & 80.46 & 91.48 & 33.37 & 48.92 & 30.64\
------------------------------------------------------------------------
\
Recall from Example \[Exmp:ExampleOfPrizePayouts\] that the crowd bet a net \$27,000,000 on 10 million tickets and the carryover was \$30,000,000 — yet the syndicate won over \$6 million. According to Table \[Ta:BettingThresholds\], the syndicate should not bet under these conditions, since a minimum carryover of \$36.92 million is necessary. There is no problem here, since the numbers in the table are expected values and it is quite possible for a syndicate to win despite making an unfavorable bet. The syndicate in that example just got lucky.
Non-equiprobable Betting by the Crowd {#S:Non-equiprobableBettingByTheCrowd}
-------------------------------------
The calculations in Section \[S:EquiprobableBettingByTheCrowd\] assumed that the crowd bets independently using $q = \frac{1}{t} 1_t$, where $1_t$ is a t-vector of all ones. What happens when the crowd bets using $q \ne \frac{1}{t} 1_t$?
In part \[Enum:OptimalStrategy\], we stated a result from [@moffitt:ziemba:2017a], that for *pure jackpot* lotteries (ones having a single prize, a non-stochastic jackpot[^1] $v$) the expected payoff is $$E_q \left[ v \frac{1}{1 + N_1} \right] \, = v E_q \left[ \frac{1}{1 + N_1} \right] \, > \, v E_{1_t/t} \left[ \frac{1}{1 + N_1} \right] = \, E_{1_t/t} \left[ v \frac{1}{1 + N_1} \right], \label{E:CrowOptimalqi}$$ where $q \ne 1/t 1_t$, $N_1$ is the random number of 6/6 tickets held by the crowd. However, formula does not apply in the present case because $v$ is stochastic, depending on the size of the *Pools Fund*.
Consider a non-stochastic configuration of single ticket bets $n_j = (n_{j1}, n_{j2}, \ldots, n_{jt})'$ for individuals $j = 1, \ldots, c$, each having zeroes except for a single $1$ in some position. Define $z_k = \sum_{j=1}^{j=c} n_{jk}$ and t-vector $z = (z_1, \ldots, z_t)'$. Clearly, $\sum_{k=1}^{k=t} z_k = c$. To compute the expected values of ticket types 1, 2, …7 with respect to an equiprobable drawing, observe that as $i$ ranges over all ticket drawings $i = 1, \ldots t$, for **any** $n_j$, the number of 6/6 is 1, the number of 5/6+ is 6, the number of 5/6 is 252, and so on as indicated in Table \[Ta:PrizeAllocIn6492013-\]. Since the drawing is equiprobable, dividing each of these by $t$ gives the probability that **any** non-stochastic ticket will be of the indicated type under an equiprobable drawing. Define indicator functions on single ticket t-vectors $n$ as: $$I_{x/6}(n) = \begin{cases}
1 & \text{if ticket } n \text{ is a x/6 ticket}, \\
0 & \text{otherwise.}
\end{cases}$$ Applying this to fixed payout types 3/6, 2/6+ and 2/6, we obtain for $d_{\!_{AB}}$ in formula $$\begin{aligned}
E_e[d_{\!_{AB}}] &= E_e \left[ \sum_{j=1}^{j=c} \left\{ \$10 I_{3/6}(n_j) \, + \, 5 I_{2/6+}(n_j) \, + \, 1.41 I_{2/6} \right\} \right] + \$5,696,520 \nonumber \\
&= \left( \$10 p_{3/6} \, + \$5 \, p_{2/6+} \, + \, \$1.41 p_{2/6} \right) c \, + \, \$5,696,520, \nonumber \\
&= \$0.4073651 \cdot c * \, + \, \$5,696,520. \label{E:ExpectedFixedTicketPayout}\end{aligned}$$ where the notation $E_e$ emphasizes that the expectation is taken over equiprobable drawings and $\$5,696,520$ is the fixed payout/deduction for the syndicate.[^2] The (stochastic) jackpot is $v = a \, + \, 0.795 d_{\!_{PF}}$ and the random 6/6 payout to the syndicate is $$\begin{aligned}
v \frac{1}{1 + N_1} &= \left( a \, + \, 0.795 d_{\!_{PF}} \right) \frac{1}{1 + N_1} \nonumber \\
&= \left( a \, + \, 0.795 (0.4 (3( t \, + \, c(1 - f))) - d_{\!_{AB}}) \right) \frac{1}{1 + N_1} \nonumber \\
&= \left( a \, + \, 0.954( t \, + \, c(1 - f)) \right) \frac{1}{1 + N_1} - \frac{d_{\!_{AB}}}{1 + N_1} \nonumber \\
&= \left( a \, + \, 0.954( t \, + \, c(1 - f)) - \$5,696,520 \right) \frac{1}{1 + N_1} - \label{E:SyndicateTermOfJackpotPayout} \\
& \qquad \frac{\$10 N_5 \, + \, 5 N_6 \, + \, 1.41 N_7}{1 + N_1} \label{E:CrowdTermOfJackpotPayout}\end{aligned}$$ In , the factor multiplying $1/(1 + N_1)$ is fixed. Its expectation using is $$\begin{aligned}
& E_q \left[ \left( a \, + \, 0.954( t \, + \, c(1 - f) - \$5,696,520) \right) \frac{1}{1 + N_1} \right] \nonumber \\
& \qquad > \left( a \, + \, 0.954( t \, + \, c(1 - f) - \$5,696,520) \right) \frac{1}{\lambda} ( 1 - \exp(-\lambda) ), \label{E:SyndicateCostBound}\end{aligned}$$ where $\lambda = c/t$. Thus for this term at least, the syndicate gets more than a fair split of the jackpot since $$\frac{1}{\lambda} ( 1 - \exp(-\lambda) ) > \frac{t}{t + c}.$$ The second term depends on $N_1$, $N_5$, $N_6$ and $N_7$, which respectively, are the numbers of 6/6, 3/6, 2/6+ and 2/6 tickets held by the crowd, and these are dependent on the crowd betting probabilities $q = (q_1, \ldots, q_t)'$. But we do not have the data to model the joint distribution of $(N_1, N_5, N_6, N_7)$ which is needed to evaluate .
However, we have circumstantial evidence that $N_5$, $N_6$ and $N_7$ are positively correlated with $N_1$. Therefore we make a crude assumption that the joint crowd payouts for 3/6, 2/6+ and 2/6 tickets are increased linearly with the winning ticket, that is, the payout for ticket $i$ is proportional to $q_i$: $$\frac{\$10 N_5 \, + \, \$5 N_6 \, + \, \$1.41 N_7}{1 + N_1} \cdot q_i / (1/t).$$ Thus if the winning ticket $i$ is bet with twice the frequency of an equiprobable bet, so that $t q_i = 2$, then the fixed payouts/deductions will be twice that expected in the equiprobable case (see the discussion leading to equation ).
Using $H = \$10 p_5 \, + \, \$5 p_6 \, + \, \$1.41 p_7$, we calculate $$\begin{aligned}
E_{1_t/t} \left[ \min \left( c H t q_i, d_{PF} \right) \frac{1}{1 + N_1} \right]
&\le E_{1_t/t} \left[ c H t q_i \frac{1}{1 + N_1} \right] \nonumber \\
&= H t E_{1_t/t} \left[ c q_i \frac{1}{c q_i} (1 - e^{-cq_i}) \right] \nonumber \\
&= H t \sum_{i=1}^{i=t} \frac{1}{t} (1 - e^{-cq_i}) \nonumber \\
&\le H t (1 - e^{-c/t}) \label{E:CrowdCost} \\
&= \frac{c H}{\lambda} (1 - e^{-\lambda}), \label{E:CrowdCostBound}\end{aligned}$$ where $\lambda = c/t$ and the step follows from Jensen’s inequality since $1 - e^{-cq}$ is a concave function of $q$. Jensen’s inequality can be stated as follows. A function $f: [a,b] \rightarrow \mathbb{R}$ that satisfies $f(ta + (1-t)b) \le t f(a) + (1-t) f(b)$ for all $t \in (0,1)$ is called *convex*, and if the inequality is strict, *strictly convex*. For a random variable $X$ and *convex* function $f$, Jensen’s inequality asserts that $f(E[X]) \le E[f(X)]$. Further, if $X$ is not degenerate and $f$ is strictly convex, then $f(E[X]) < E[f(X)]$. A function $f$ is (*strictly*) *concave* if $-f$ is (*strictly*) *convex*i, so Jensen’s inequality is reversed for *concave* functions.
Putting together with yields $$\begin{aligned}
E \left[v \frac{1}{1 + N_1} \right] &\ge \left( a \, + \, 0.954( t \, + \, c(1 - f)) - \$5,696,520 \, - \, c H \right) \nonumber \\
&\qquad \cdot \frac{1}{\lambda} (1 - e^{-\lambda}) \label{E:SyndicateExpectedReturn}\end{aligned}$$ where $\lambda = c/t$ and $H = \$10 p_5 \, + \, \$5 p_6 \, + \, \$1.41 p_7$.
This calculation shows that the syndicate obtains a better result than when the crowd bets proportionally, as in the corresponding result for *pure jackpot* lotteries.
Design Considerations for Lotteries {#SS:DesignConsiderationsForLotteries}
===================================
Lottery design includes the goal to maximize the the sponsors’ earnings. Assuming fairly constant fixed costs of running the lottery, sponsors should strive to make the lottery popular, thereby increasing profitability. The most recent changes to payouts were made with that goal in mind — these changes increased the “convexity” of payouts, meaning many little prizes and greater jackpot growth. Ziemba recommended these designs in his work in the 1980’s and [@Walker2008459] later also recommended them. Convex designs encourage players because more “get something back,” while at the same time growing large jackpots quickly.
This design is supported by research in behavioral finance. Lopes’ SP/A (Security-Potential/Aspiration) model ([@Lopes1987255]), an improved version of the classic Friedman/Savage (1948) utility curves, argues that many unsophisticated gamblers prefer strategies of buying safe prospects with a few longshots (the “Cautiously Hopeful” pattern of SP/A). Regarding large *jackpots*, Daniel Kahneman has written
> “For emotionally significant events, the size of the probability simply doesn’t matter. What matters is the possibility of winning. People are excited by the image in their mind. The excitement grows with the size of the prize, but it doesn’t diminish with the size of the probability.” Source: [@NYTimes:YourMoney:KahnemanQuote:Online].
There is another aspect of lottery design, namely, discouraging syndicates from buying the pot. There are two ways to accomplish this: (1) creating a large number of tickets making it logistically difficult to buy the pot, and (2) using convex designs, which reduces the likelihood that pot buying situations will occur. Method (1) is not feasible except for large lotteries like the California Powerball lottery. The reason is that if the number of tickets sold are too small relative to the total number of tickets, the jackpot may build slowly and seldom be won. On the other hand, method (2) can be effective regardless of the size of the lottery. To illustrate, consider a *pure jackpot* lottery with the same carryover, take and crowd betting as in Table \[Ta:BettingThresholds\]. The results are shown in Table \[Ta:BettingThresholdsForPureJockpotAnd649\]. The first column has the number of tickets, which after a 10% deduction for free plays, equals the crowd contribution to the *betting pool* shown in the second column. Then assuming a take of 60%, breakeven thresholds of 0%, 10% and 20% for the pure jackpot lottery are shown in columns 3-5 and for the 6/49 in columns 6-8. The results show that buying the pot thresholds are lower in the pure lottery, but not as much as might be expected.
But one can see the reason by a simple argument. When the sponsors takes 60%, only 40 cents is returned as prizes for each dollar wagered. Therefore, a syndicate needs to recover 60% of the covering bet, or $0.6*\$3*13,983,816 = \$25,170,869$, regardless of the lottery’s rules. As we have shown, the syndicate earns its fair share of consolation prizes, but the free plays it earns are not worth too much since after the lottery is hit the next lottery when those tickets will be used will have a small purse. Assuming the the crowd bets $\$1,000,000$ on the next lottery the expected value of these $1,678,950$ tickets will under optimal wagering be worth about $\$150,000$.
[rccccccc]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
& Crowd & &\
& \$ Bet & Breakeven & +10% & +20% & Breakeven & +10% & +20%\
& (millions) & (millions) & (millions) & (millions) & (millions) & (millions) & (millions)\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
3.3 & 9 & 26.77 & 31.48 & 36.20 & 30.33 & 35.05 & 39.76\
6.7 & 18 & 28.76 & 34.04 & 39.31 & 33.46 & 38.74 & 44.01\
10.0 & 27 & 31.14 & 37.02 & 42.89 & 36.92 & 42.80 & 48.67\
13.3 & 36 & 33.90 & 40.41 & 46.92 & 40.71 & 47.22 & 53.73\
16.7 & 45 & 37.02 & 44.20 & 51.38 & 44.81 & 51.99 & 59.17\
20.0 & 54 & 40.49 & 48.38 & 56.26 & 49.21 & 57.10 & 64.99\
23.3 & 63 & 44.28 & 52.91 & 61.53 & 53.90 & 62.52 & 71.15\
26.7 & 72 & 48.37 & 57.77 & 67.17 & 58.84 & 68.24 & 77.63\
30.0 & 81 & 52.75 & 62.94 & 73.13 & 64.03 & 74.23 & 84.42\
33.3 & 90 & 57.38 & 68.39 & 79.41 & 69.45 & 80.46 & 91.48\
------------------------------------------------------------------------
\
We conclude the discussion by examining the impacts of design choices in the 6/49 Lotto. The 6/49 Lotto’s convex design according to Table \[Ta:BettingThresholdsForPureJockpotAnd649\] raised the bar for buying-the-pot strategies, making carryover thresholds roughly 12% to 20% higher. We now compare the impacts of the 6/49’s design features toward increasing the threshold for buying the pot. We identify four factors: (1) the take, (2) the payouts for small prizes, (3) the payouts for large, non 6/6 prizes, and (4) free plays. Then we compare by
1. Changing the take only, using alternatives 55%, 60% (current) and 65%.
2. Eliminating fixed payouts 3/6 and 2/6+ only.
3. Eliminating 4/6, 5/6 and 5/6+ payouts only.
4. Eliminating free plays only.
Table \[Ta:BettingThresholdsForDesignFactors\] shows breakeven carryover thresholds for these design factors. The factor is indicated in the first column and the other 5 columns are carryover thresholds when the crowd bets the indicated millions of dollars, 20, 40, etc. In the second column (corresponding to a crowd bet of \$20 million), the numbers in parenthesis are differences of threshold carryovers from the current 6/49 values (second row, second column). Since the relative impacts of these factors are the same for the five crowd betting amounts, their impacts on the buying the pot strategy can be assessed using this column. The greatest factor impact is due to free plays; removing them drops the threshold by $\$3.39$ million ($\sim 10\%$). The largest inhibitor is clearly the take — increasing it by $0.05\%$ from to $65\%$ has a large impact on breakeven carryovers.
[lcccccccccc]{}
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
&\
& 20 & 40 & 60 & 80 & 100\
& million & million & million & million & million\
------------------------------------------------------------------------
\
------------------------------------------------------------------------
\
TAKE=0.55 & 30.56 (-2.90) & 36.98 & 44.64 & 53.41 & 63.15\
CURRENT 6/49 & 33.46 ( 0.00) & 40.71 & 49.21 & 58.84 & 69.45\
TAKE=0.65 & 36.37 ( 2.91) & 44.44 & 53.79 & 64.27 & 75.74\
NO 2/6+, 3/6 & 32.88 (-0.58) & 39.65 & 47.76 & 57.08 & 67.44\
NO 4/6, 5/6 & 32.99 (-0.47) & 39.87 & 48.07 & 57.48 & 67.91\
NO FREE PLAY & 30.07 (-3.39) & 36.28 & 43.73 & 52.29 & 61.81\
------------------------------------------------------------------------
\
Based on these statistics, we make recommendations for state lotteries using ratings of the form $(+=-\pm\mp, +=-\pm\mp)$. The first sign is for popularity, the second for inhibiting buyers of the pot. For example, $(+,-)$ indicates that a factor increases the lottery’s popularity, but encourages buying the pot.
1. $(\mp,+)$ If possible, add combinations to the lottery by increasing the numbers.
2. $(+,+)$ Initiate a free play feature.
3. $(\mp,+)$ Increase the take.
4. $(+,=)$ Offer many small prizes.
5. $(+,=)$ Increase the allocation of the *Pools Fund* to 6/6 winners.
6. $(=,+)$ Decrease the awards to hard-to-win non 6/6 tickets.
Increasing the allocation to 6/6 allows quicker build-up of jackpots, which encourages greater crowd participation. However, we did not address the question of build-up speed of the jackpot, nor the acceleration of betting on larger jackpots. These need to be studied in order to design prizes and allocations to optimize betting flows.
Conclusions
===========
In this paper, we have shown conditions under which buying the pot in the 6/49 Lotto has positive expected return when the crowd bets equiprobably. We also indicated that equiprobable betting is optimal for the crowd, that is, expected return is lower when it does not bet equiprobably. We illustrated the advantages of lotteries with convex designs by calculating 6/49 carryover thresholds and comparable *pure jackpot* carryover thresholds. We then rated various design features for their likelihood of increasing a lottery’s popular, and decreasing the likelihood of buyers of the pot.
The Syndicate’s Expected Value when the Crowd bets Equiprobably {#S:TheSyndicate'sExpectedValueWhenTheCrowdBetsEquiprobably}
===============================================================
Assuming that the lottery’s tickets are equiprobable, $(N_1, \ldots N_8)'$ has a multinomial distribution $$(N_1, \ldots N_8)' \sim Multin(c + t, p),$$ where $p = (p_1, p_2, \ldots, p_8)'$. The distribution of deductions $d_{\!_{AB}}$ from the *prize pool* is given by $$d_{\!_{AB}} = \beta_{\!_{AB}}' N,$$ where $\beta_{\!_{AB}} = (0,0,0,0,10,5,1.41,0)'$ and $N \sim Multin(c + t, p)$.
Substituting and into equation gives the *Pools Fund* as $$d_{\!_{PF}} = 0.40 \cdot 3 \cdot ( t + c \cdot (1 - f)) \, - \, d_{\!_{AB}}.$$ and in RHS of this expression, only $d_{\!_{AB}}$ is random.
Using results from [@moffitt:ziemba:2017a], the expected value $G$ of the syndicate’s net gain, given $d_{\!_{PF}}$, as $$\begin{aligned}
E[ G(c) \, | \, d_{\!_{PF}}] &= 0.795 \cdot (a + d_{\!_{PF}}) E \left[ \frac{1}{1 + X_{\text{6/6}}} \right] \label{E:6/6Expectation} \\
& + 0.06 \cdot d_{\!_{PF}} E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right] \label{E:5/6+Expectation} \\
& + 0.05 \cdot d_{\!_{PF}} E \left[ \frac{252}{252 + X_{\text{5/6-}}} \right] \label{E:5/6-Expectation} \\
& + 0.095 \cdot d_{\!_{PF}} E \left[ \frac{13545}{13545 + X_{\text{4/6}}} \right] \label{E:4/6Expectation} \\
& + \$2,468,200 \, + \, \$86,100 \, - \, \$3 t \label{E:2-3/6Expectation}\end{aligned}$$ where $$\begin{aligned}
(X_{\text{6/6}}, X_{\text{5/6+}}, X_{\text{5/6-}}, X_{\text{4/6}})' \sim Multin(c,(1, 6, 252, 13545)'/t).\end{aligned}$$ Since buying one of each ticket gives the same exact payout regardless of the winning ticket (numbers of tickets shown in Table \[Ta:PrizeAllocIn6492013-\]), we know that a covering strategy pays \$2,468,200 and \$86,100, respectively, for 3/6 and and 2/6+ tickets. This explains the term .
Using the formulas from we obtain for $\lambda(c) = c/t$ $$E \left[ \frac{1}{1 + X_{\text{6/6}}} \right] = \lambda(c)^{-1} (1 - exp(-\lambda(c)))$$ and values $\nu(c) = E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right]$ using recursion. These calculations take care of terms and .
Using the Law of Large Numbers, the expectation in the term can be approximated by $$\frac{252}{252 \, + \, 252 c/t} = \frac{1}{1 + c/t}. \label{E:FairSplitExpectation3}$$ and in term by $$\frac{13545}{13545 \, + \, 13545 c/t} = \frac{1}{1 + c/t}. \label{E:FairSplitExpectation4}$$ Basically, these two approximations amount to fair split of the corresponding share of the *Funds Pool*. Thus $$\begin{aligned}
E[ G(c) \, | \, d_{\!_{PF}}] & = (a + 0.795 \cdot d_{\!_{PF}}) \lambda(c)^{-1} (1 - exp(-\lambda(c))) \label{E:6/6ExpectationC} \\
& + 0.06 \cdot d_{\!_{PF}} E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right] \label{E:5/6+ExpectationC} \\
& + 0.145 \cdot d_{\!_{PF}} \frac{1}{1 + c/t} \label{E:5/6-ExpectationC} \\
& + 2,468,200 \, + \, 86,100 \, - \, \$3 t \label{E:2-3/6ExpectationC} \end{aligned}$$ where $\lambda(c) = c/t$. To complete the calculation, we need to eliminate the dependence of $E[ G(c) \, | \, d_{\!_{PF}}]$ on $d_{\!_{PF}}$ by determining its distribution and performing an integration. But this is straightforward: the first three terms , and are linear in $d_{\!_{PF}}$ so that the expectation $\mu(c) = E[d_{\!_{PF}}]$ should be substituted for $d_{\!_{PF}}$. The expectation $E[d_{\!_{PF}}]$can be calculated by substituting into and into taking expectations: $$\mu(c) = E[d_{\!_{PF}}] = 0.40 \cdot 3 \cdot ( t + c \cdot (1 - f)) \, - \, E[d_{\!_{AB}}]$$ We calculate $E[d_{\!_{AB}}]$ as follows. For any ticket $i$, the number of tickets that are 3/6, 2/6+ and 2/6- are respectively $248,820$, $172,200$, and $1,678,950$, respectively. Therefore, the probability that ticket $i$ is a 3/6, 2/6+ or 2/6- ticket, *given that a winning ticket is drawn equiprobably*, is $p_5 = 248,820/t$, $p_6 = 172,200/t$ and $p_7 = 1,678,950/t$, respectively. Now consider any choice of $c$ tickets. By linearity of expectations, the expected number of 3/6 tickets is $c * p_5$, of 2/6+ tickets is $c * p_6$ and of 2/6- tickets, $c * p_7$. Therefore, $$\nu(c) = E[d_{\!_{AB}}] = (t + c) (p_5*\$10 + p_6*\$5 + p_7*\$1.41).$$
Summarizing, the expected gain $G(c)$ to a syndicate that covers the pool is $$\begin{aligned}
E[ \, G(c) \, ] & = (a + 0.795 \cdot \mu(c)) \lambda(c)^{-1} (1 - exp(-\lambda(c))) \label{E:SyndicateExpectedReturnAppendix} \\
& + \, \left( 0.06 \nu(c) \, + \, 0.145 \frac{1}{1 + c/t} \right) \mu(c) \nonumber \\
& + \, \$2,553,300 \, - \, \$3 t, \nonumber\end{aligned}$$ where $$\begin{aligned}
\lambda(c) &= c/t, \\
\mu(c) &= 0.40 \cdot 3 \left( ( t + c \cdot (1 - f)) \right) - \, (t + c) \cdot (p_5 \$10 \, + \, p_6 \$5 \, + \, p_7 \$1.41), \\
\nu(c) &= E \left[ \frac{6}{6 + X_{\text{5/6+}}} \right], \qquad \text{for } X \sim Bin(c,6/t).\end{aligned}$$
[^1]: We are assuming that the crowd’s number of tickets, $c$, is known.
[^2]: $\$5,696,520 = \$10*246,820 + \$5*172,200 + \$1.41*1,678,950$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we study the convergence properties of the one-level parallel Schwarz method applied to the one-dimensional and two-dimensional Helmholtz and Maxwell’s equations. One-level methods are not scalable in general. However, it has recently been proven that when impedance transmission conditions are used in the case of the algorithm applied to the equations with absorption, under certain assumptions, scalability can be achieved and no coarse space is required. We show here that this result is also true for the iterative version of the method at the continuous level for strip-wise decompositions into subdomains that can typically be encountered when solving wave-guide problems. The convergence proof relies on the particular block Toeplitz structure of the global iteration matrix. Although non-Hermitian, we prove that its limiting spectrum has a near identical form to that of a Hermitian matrix of the same structure. We illustrate our results with numerical experiments.'
author:
- 'N. Bootland [^1]'
- 'V. Dolean [^2]'
- 'A. Kyriakis'
- 'J. Pestana'
bibliography:
- 'paper\_refs.bib'
title: '[[Analysis of parallel Schwarz algorithms for time-harmonic problems using block Toeplitz matrices]{}]{}[^3]'
---
domain decomposition methods, Helmholtz equations, Maxwell equations, Schwarz algorithms, one-level methods, block Toeplitz matrices
65N55, 65N35, 65F10, 15A18, 15B05
Introduction
============
Time-harmonic wave propagation problems, such as those arising in electromagnetic and seismic applications, are notoriously difficult to solve for several reasons. At the continuous level, the underlying boundary value problems lead to non self-adjoint operators (when impedance boundary conditions are used). The discretisation of these operators by a Galerkin method requires an increasing number of discretisation points as the wave number gets larger in order to avoid the pollution effect, meaning a shift in the numerical wave velocity with respect to the continuous one [@Babuska:1997:IPE]. This leads to increasingly large linear systems with non-Hermitian matrices that are difficult to solve by classical iterative methods [@Ernst:12:NAM].
In the past two decades, different classes of efficient solvers and preconditioners have been devised; see the review paper [@Gander:2018:SIREV] and references therein. One important class is based on domain decomposition methods [@Dolean:15:DDM], which are a good compromise between direct and iterative methods. Some of these domain decomposition methods rely on improving the transmission conditions, that pass data between subdomains, to give optimised transmission conditions; see the seminal work on Helmholtz equations [@Gander:2007:OSM] and its extensions to Maxwell’s equations [@Dolean:2015:ETC; @Dolean:09:OSM; @Dolean:08:DDM; @ElBouajaji:12:OSM] as well as to elastic waves [@Brunet:2019:NDD; @Mattesi:2019:ABC]. For large-scale problems, in order to achieve robustness with respect to the number of subdomains (scalability) and the wave number, two-level domain decomposition solvers have been developed in recent years: they are based on the idea of using the absorptive counterpart of the equations as a preconditioner, which in turn is solved by a domain decomposition method. These methods were successfully applied to Helmholtz and Maxwell’s equations, which arise naturally in different applications [@Bonazzoli:2019:ADD; @Dolean:2020:IFD; @Graham:2017:RRD].
However, an alternative idea emerged in the last few years by observing that, when using Robin or impedance transmission conditions, under certain assumptions involving the physical and numerical parameters of the problem (i.e. absorption, size of the subdomains, etc.) one-level Schwarz algorithms can scale (behave independently of the number of subdomains) without the addition of a second level [@Gong:2020:DDP; @Graham:2018:DDI]. We would like to explore this idea at the continuous level (independent of the discretisation) for a strip-wise decomposition into subdomains as it arises naturally in the solution of wave-guide problems. The main contributions of the paper are the following:
- We provide analysis of the limiting spectrum, as the number of subdomains grows, for a one-level Schwarz method applied to a strip-wise decomposition. While our analysis is limited to this simple yet realistic configuration, it is valid at the continuous level both for one-dimensional and two-dimensional Helmholtz and Maxwell’s equations.
- We build on the formalism introduced in [@Chaouqui:2018:OSC] for the iteration matrices at the interface but we are able to characterise the entire spectrum of these iteration matrices by using their block Toeplitz structure.
- Despite the fact that the block Toeplitz structure is non-Hermitian, and thus results from the standard literature do not apply in a straightforward manner, we prove that the limiting spectrum of the iteration matrices as their size grows (corresponding to an increasing number of subdomains) tends to the limit predicted by the eigenvalues of the symbol of the block Toeplitz matrix, except perhaps for two additional simple eigenvalues.
- We show that the limiting spectrum is descriptive of what is observed in practice numerically, even for a relatively small number of subdomains.
The structure of the paper is as follows: In we present our results on the limiting spectrum of a non-Hermitian block Toeplitz matrix whose characteristic polynomial verifies a three-term recurrence. In we apply these results to the analysis of the iterative Schwarz algorithm in the one-dimensional and two-dimensional cases. We illustrate the theory with numerical results in . Finally, draws together our conclusions.
A non-Hermitian block Toeplitz structure {#sec:toeplitz}
========================================
Consider a non-Hermitian block Toeplitz matrix $\mathcal{T} \in \mathbb{C}^{2m\times2m}$ of the form
\[NonHermitianBlockToeplitzStructure\] $$\begin{aligned}
\label{NonHermitianBlockToeplitzStructure-Matrix}
\mathcal{T}
= \left[\begin{array}{ccccc}
A_{0} & A_{1} & & & \\
A_{-1} & A_{0} & A_{1} & & \\
& \ddots & \ddots & \ddots & \\
& & A_{-1} & A_0 & A_1 \\
& & & A_{-1} & A_0
\end{array}\right],\end{aligned}$$ where $$\begin{aligned}
\label{NonHermitianBlockToeplitzStructure-Blocks}
A_0 = \left[\begin{array}{cc} 0 & b \\ b & 0 \end{array}\right], \ A_1 = \left[\begin{array}{cc} a & 0 \\ 0 & 0 \end{array}\right], \ A_{-1} = \left[\begin{array}{cc} 0 & 0 \\ 0 & a \end{array}\right],\end{aligned}$$
for some non-zero complex coefficients $a$ and $b$. We will see in the sections that follow that such non-Hermitian block Toeplitz structures arise naturally for iterative Schwarz algorithms applied to wave propagation problems. We are interested in a characterisation of the complete spectrum of when its dimension becomes large, equating to the many subdomain case for the Schwarz method.
The so-called Szegő formula enables the asymptotic spectrum, i.e. the spectrum as $m \to \infty$, of a wide class of Hermitian block Toeplitz matrices to be characterised by the eigenvalues of an associated matrix-valued function called the (block) symbol [@Tillinota]. For non-Hermitian matrices, analogous results do not exist in general [@Tillinota], but do hold when the union of the essential ranges of the eigenvalues of the block symbol has empty interior and does not disconnect the complex plane [@DNS12]. Unfortunately, $\mathcal{T}$ in has symbol $F(z) = A_{-1}z+ A_0 + A_1z^{-1}$ and, for relevant values of $a$ and $b$, the union of essential ranges is a closed curve. Additional characterisations of the asymptotic spectrum of (block) banded Toeplitz matrices are available [@Hirs67; @ScSp60; @Wido74], but these do not provide explicit formulae for the eigenvalues, as we shall in . Other formulae for the eigenvalues [@SaMo13] and determinant [@Time87] of block tridiagonal Toeplitz matrices are known, however, they are applicable only when $A_1$ (or $A_{-1}$) is nonsingular. Accordingly, in this section we derive the limiting spectrum of $\mathcal{T}$.
In order to establish a result on the spectrum, we first show that the characteristic polynomials of for increasing $m$ obey a three-term recurrence relation.
\[lemma:ThreeTermRecurrence\] Let $p_{m}(z)$ denote the characteristic polynomial of the block Toeplitz matrix $\mathcal{T} \in \mathbb{C}^{2m\times2m}$ defined in . Then $p_{m}(z)$ satisfies the three-term recurrence relation $$\begin{aligned}
\label{ThreeRecurrenceRelation-General}
p_{m}(z) + B(z) p_{m-1}(z) + A(z) p_{m-2}(z) & = 0, & \text{for} \ m \ge 2,\end{aligned}$$ with $A(z) = a^2 z^2$ and $B(z) = - z^2 + b^2 - a^2$ and where $p_{0}(z) = 1$ and $p_{1}(z) = z^{2} - b^{2}$. Furthermore, this recurrence relation is encoded in the generating function $$\begin{aligned}
\label{GeneratingFunctionDefinition}
\sum_{m=0}^{\infty} p_{m}(z) t^{m} = \frac{N(t,z)}{D(t,z)},\end{aligned}$$ where
\[GeneratingFunctionDAndN\] $$\begin{aligned}
\label{GeneratingFunctionD}
D(t,z) & = 1 + B(z)t + A(z)t^2,\\
\label{GeneratingFunctionN}
N(t,z) & = p_{0}(z) + (p_{1}(z) + B(z)p_{0}(z))t.\end{aligned}$$
Thus, in our case, $D(t,z) = 1 - (z^2-b^2+a^2)t + a^2 z^2 t^2$ while $N(t,z) = 1 - a^2 t$.
We first prove the recurrence relation. Let $D_{m}$ be the $2m\times2m$ matrix whose determinant is the characteristic polynomial of $\mathcal{T}$ in the variable $z$. Note that the first two characteristic polynomials are
\[FirstTwoCharacteristicPolynomials\] $$\begin{aligned}
p_{1}(z) &= \det(D_{1}) = \left|\begin{array}{cc} -z & b \\ b & -z \end{array}\right| = z^{2} - b^{2}, \\
p_{2}(z) &= \det(D_{2}) = \left|\begin{array}{cccc} -z & b & a & 0 \\ b & -z & 0 & 0 \\ 0 & 0 & -z & b \\ 0 & a & b & -z \end{array}\right| = (z^{2} - b^{2})^{2} - a^{2} b^{2}.\end{aligned}$$
To derive a recurrence relation, let us also define the intermediary determinants $$\begin{aligned}
r_{m}(z) \vcentcolon= \left|\begin{array}{c;{2pt/2pt}cccc} b & a & 0 & 0 & \cdots \\ \hdashline[2pt/2pt] 0 & \multicolumn{4}{c}{\multirow{4}{*}{$D_{m-1}$}} \\ a & & & & \\ 0 & & & & \\ \vdots & & & & \end{array}\right| = \left|\begin{array}{c;{2pt/2pt}cc;{2pt/2pt}cc} b & a & 0 & 0 & \cdots \\ \hdashline[2pt/2pt] 0 & -z & b & a & 0 \\ a & b & -z & 0 & 0 \\ \hdashline[2pt/2pt] 0 & 0 & 0 & \multicolumn{2}{c}{\multirow{2}{*}{$D_{m-2}$}} \\ \vdots & 0 & a & & \end{array}\right| = b \, p_{m-1}(z) + a^{2} \, r_{m-1}(z),\end{aligned}$$ so that $$\begin{aligned}
p_{m}(z) = z^{2} \, p_{m-1}(z) - b \, r_{m}(z) = (z^{2} - b^{2}) \, p_{m-1}(z) - a^{2} b \, r_{m-1}(z).\end{aligned}$$ We can then rearrange this relation to give an expression for $r_{m-1}(z)$ in terms of $p_{m}(z)$ and $p_{m-1}(z)$. Substituting this into the recurrence for $r_{m}(z)$ above, along with the equivalent expression for $r_{m}(z)$, yields the desired recurrence relation $$\begin{aligned}
\label{ThreeTermRecurrence}
p_{m+1}(z) & = (z^{2} - b^{2} + a^{2}) \, p_{m}(z) - a^{2}z^{2} \, p_{m-1}(z),\end{aligned}$$ where $A(z) \vcentcolon= a^2 z^2$ and $B(z) \vcentcolon= - z^2 + b^2 - a^2$. Finally, note that setting $p_{0} = 1$ is consistent with this recurrence relation and initial characteristic polynomials .
To show the equivalence of the generating function, we multiply by $t^{m}$ and sum over $m\ge2$ before adding relevant terms to isolate $\sum_{m=0}^{\infty} p_{m}(z)t^{m}$ as follows $$\begin{aligned}
& \sum_{m=2}^{\infty} \left[p_{m}(z) + B(z) p_{m-1}(z) + A(z) p_{m-2}(z)\right]t^{m} = 0 \\
\iff & \sum_{m=0}^{\infty} \left[1 + B(z) t + A(z) t^{2}\right]p_{m}(z)t^{m} = p_{0}(z) + \left(p_{1}(z)+B(z)p_{0}(z)\right)t \\
\iff & \sum_{m=0}^{\infty} p_{m}(z)t^{m} = \frac{p_{0}(z) + \left(p_{1}(z)+B(z)p_{0}(z)\right)t}{1 + B(z) t + A(z) t^{2}}.\end{aligned}$$ Substituting in the appropriate values gives $D(t,z) = 1 - (z^2-b^2+a^2)t + a^2 z^2 t^2$ and $N(t,z) = 1 - a^2 t$ in our case, as required.
We now introduce a useful tool that will help us to characterise the spectrum of : the $q$-analogue of the discriminant known as the $q$-discriminant [@Tran:2014:CBD]. The $q$-discriminant of a polynomial $P_{n}(t)$ of degree $n$ with leading coefficient $p$ is defined as $$\begin{aligned}
\label{q-discriminant}
\mathrm{Disc}_{t}(P_{n};q) = p^{2n-2} q^{n(n-1)/2} \prod_{1 \le i < j \le n} (q^{-1/2}t_{i} - q^{1/2}t_{j})(q^{1/2}t_{i} - q^{-1/2}t_{j}),\end{aligned}$$ where $t_{i}$, $1 \le i \le n$, are the roots of $P_{n}(t)$. A key point is that the $q$-discriminant is zero if and only if a quotient of roots $t_{i}/t_{j}$ equals $q$. Note that as $q\rightarrow1$ the $q$-discriminant becomes the standard discriminant of a polynomial.
In particular, we will consider the $q$-discriminant of the denominator $D(t,z)$ as a quadratic in $t$. Direct calculation using the quadratic formula yields $$\begin{aligned}
\label{q-discriminant-of-D}
\mathrm{Disc}_{t}(D(t,z);q) = q \left(B(z)^{2} - (q + q^{-1} + 2)A(z)\right),\end{aligned}$$ for any $q\neq0$. If $q$ is a quotient of the two roots in $t$ of $D(t,z)$ then is zero and so $q$ must satisfy $$\begin{aligned}
\label{DefiningCurve}
\frac{B(z)^{2}}{A(z)} = q + q^{-1} + 2,\end{aligned}$$ where, in general, $q$ will depend on $z$. We now state our main result on the limiting spectrum of $\mathcal{T}$ as its dimension becomes large in which we adapt some ideas from [@Tran:2014:CBD] for finding roots of polynomials verifying a three-term recurrence but now with a different generating function.
\[theorem:LimitingSpectrum\] The limiting spectrum, as $m\rightarrow\infty$, of the block Toeplitz matrix $\mathcal{T} \in \mathbb{C}^{2m\times2m}$, defined in , lies on the curve defined by $$\begin{aligned}
\label{LimitingCurve}
\lambda_{\pm}(\theta) &= a \cos(\theta) \pm \sqrt{b^2 - a^2 \sin^2(\theta)}, & & \theta \in [-\pi,\pi],\end{aligned}$$ except perhaps for simple eigenvalues $$\begin{aligned}
\label{LimitingPoints}
\lambda = \pm \sqrt{\tfrac{1}{2}b^{2} - a^{2}},\end{aligned}$$ which can only occur if $|a^{2}|>|\frac{1}{2}b^{2} - a^{2}|$.
Suppose that $z_{m}$ is a root of the characteristic polynomial $p_{m}(z)$ for $m\ge2$. If $z_{m}=0$ then we must have that $a^{2} = b^{2}$. To see this, assume for a contradiction that $a^{2} \neq b^{2}$, then $B(0)\neq0$ while $A(0)=0$ and $p_{m}(0)=0$ and thus the recurrence relation gives that $p_{m-1}(0)=0$. Following this recursion down to $m=2$ gives that $p_{1}(0)=0$, which is false as $b\neq0$. Further, if $p_{m}(0) = 0$ then also $p_{m+1}(0) = 0$ by since $A(0)=0$ and so a sequence of zero roots occurs as $m$ increases giving $0$ in the limiting spectrum. This case is covered by choosing $\theta = \frac{\pi}{2}$ in and noting that $a^{2} = b^{2}$ must hold. As such, for the remainder of the proof we assume that $z_{m}\neq0$.
Now consider the denominator $D(t,z_{m})$. Since $A(z_{m})\neq0$ by the assumption that $z_{m}\neq0$, the denominator as a quadratic in $t$ has two roots $t_{1}$ and $t_{2}$. Note that neither of these two roots can be zero since $t_{1}t_{2}A(z_{m}) = 1$, by the definition of $t_{1}$ and $t_{2}$ as roots of $D(t,z_{m})$. If $t_{1}=t_{2}$ then the (regular) discriminant of $D(t,z_{m})$ is zero, giving $B(z_{m})^2-4A(z_{m})=0$. Solving for $z_{m}$ given our expressions for $A(z)$ and $B(z)$ yields solutions $z_{m} = \pm(a \pm b)$ for all choices of signs. These cases are also covered by when $\theta=0$ or $\theta=\pi$.
As such, we now assume that $t_{1} \neq t_{2}$ and so $D(t,z_{m})=A(z_{m})(t-t_{1})(t-t_{2})$. Considering the generating function we observe that $$\begin{aligned}
\nonumber
\frac{N(t,z_{m})}{D(t,z_{m})} & = \frac{1-a^{2}t}{A(z_{m})(t-t_{1})(t-t_{2})} = \frac{1-a^{2}t}{A(z_{m})(t_{1}-t_{2})}\left(\frac{1}{t-t_{1}} - \frac{1}{t-t_{2}}\right) \\
\nonumber
& = \frac{1-a^{2}t}{A(z_{m})(t_{1}-t_{2})} \sum_{m=0}^{\infty} \frac{t_{1}^{m+1}-t_{2}^{m+1}}{t_{1}^{m+1}t_{2}^{m+1}} t^{m} \\
\label{ExpansionOfGeneratingFunction}
& = \frac{1}{A(z_{m})(t_{1}-t_{2})} \sum_{m=1}^{\infty} \left[ \frac{t_{1}^{m+1}-t_{2}^{m+1}}{t_{1}^{m+1}t_{2}^{m+1}} - a^{2} \frac{t_{1}^{m}-t_{2}^{m}}{t_{1}^{m}t_{2}^{m}} \right] t^{m} + 1.\end{aligned}$$ Thus, if $z_{m}$ is a root of $p_{m}(z)$ then the coefficient of $t^{m}$ in must be zero. Now suppose $t_{1}=qt_{2}$ for some $q\neq0$ (as neither $t_{1}$ nor $t_{2}$ is zero), then this condition on the coefficient of $t^{m}$ translates into $$\begin{aligned}
\frac{q^{m+1}-1}{q^{m+1}t_{2}^{m+1}} - a^{2} \frac{q^{m}-1}{q^{m}t_{2}^{m}} = 0 \implies q^{m+1} - 1 = a^{2}t_{2}q(q^{m}-1).\end{aligned}$$ We will later use the $q$-discriminant condition for $D(t,z_{m})$ to characterise the root $z_m$ as $m \to \infty$, noting that holds precisely when $q$ is as defined above. For now, since $t_{1}t_{2}A(z_{m}) = 1$, we deduce that $t_{2} = \pm (A(z_{m})q)^{-1/2}$ and thus $q$ must solve $$\begin{aligned}
\left(q^{m+1} - 1\right)^{2} = \frac{a^4}{A(z_{m})} \left(q^{m} - 1\right)^{2}q.\end{aligned}$$ Let us define the coefficient, depending on $z_{m}$, $$\begin{aligned}
\label{q-PolynomialCoeffient}
c_{m} = \frac{a^4}{A(z_{m})} = \frac{a^2}{z_{m}^2}.\end{aligned}$$ Then $q$ must be a root of the $2m+2$ degree polynomial $$\begin{aligned}
\label{q-Polynomial}
f_{m}(q) = q^{2m+2} - c_{m}q^{2m+1} + 2(c_{m}-1)q^{m+1} - c_{m}q + 1.\end{aligned}$$ In order to characterise the roots of we will make use of the following corollary of Rouché’s theorem: for a polynomial $f$ of degree $d$ with coefficients $\left\lbrace\alpha_{j}\right\rbrace_{j=0}^{d}$, if $R>0$ is such that for an integer $0 \le k \le d$ we have $$\begin{aligned}
\label{RoucheCondition}
|\alpha_{k}|R^{k} > |\alpha_{0}| + \ldots + |\alpha_{k-1}|R^{k-1} + |\alpha_{k+1}|R^{k+1} + \ldots + |\alpha_{d}|R^{d},\end{aligned}$$ then there are exactly $k$ roots of $f$, counted with multiplicity, having absolute value less than $R$. In particular, we will use this result for the polynomial $f_{m}(q)$ with $k=0$ or $k=1$.
We first point out some facts about . Note that $q=0$ is not a root of $f_{m}$ and, moreover, by symmetry of the coefficients, if $q_{m}$ is a root of $f_{m}$ then $$\begin{aligned}
f_m(q_{m}^{-1}) = q_{m}^{-(2m+2)} f_{m}(q_{m})=0\end{aligned}$$ and so $q_{m}^{-1}$ is also a root. In particular, note that all such $q_{m}^{-1}$ satisfy the same polynomial and thus the multiplicities of the roots $q_{m}$ and $q_{m}^{-1}$ are identical. This means that we only need to study roots with $|q_{m}|\le1$, with roots outside the unit disc being precisely the reciprocal values of those inside the unit disc.
We will use to determine how many roots of $f_{m}(q)$ in do not approach the unit circle as $m \to \infty$. This information, along with , will allow us to determine conditions for $z_m$. A significant challenge is that the coefficient $c_m$ depends on $m$, and so we must consider the behaviour of the sequence $(c_{m})$. Specifically, let $\left(\hat{c}_{m}\right)$ be a subsequence of $c_{m}$ values and suppose it falls into one of the four cases:
1. $\left(\hat{c}_{m}\right)$ is such that $|\hat{c}_{m}| \rightarrow \infty$, \[Case1\]
2. $\left(\hat{c}_{m}\right)$ is such that there exists $c$ and $C$ fixed with $1 < c \le |\hat{c}_{m}| \le C$, \[Case2\]
3. $\left(\hat{c}_{m}\right)$ is such that $|\hat{c}_{m}| \rightarrow 1$ from above, \[Case3\]
4. $\left(\hat{c}_{m}\right)$ is such that $|\hat{c}_{m}| \le 1$. \[Case4\]
Note that any sequence $(c_{m})$ may feature a combination of different subsequence cases but must contain a subsequence of at least one type. Further, all other subsequences must be a combination of the four cases above. We will show that all cases lead to the conclusion that all but possibly two roots of $f_{m}(q)$ tend to the unit circle. The two separate roots only persist in cases \[Case1\] and \[Case2\] and, moreover, case \[Case1\] only occurs in a very specific situation which results in zero being in the limiting spectrum of $\mathcal{T}$.
#### Case 1
To start the analysis we suppose we are in case \[Case1\] where $|\hat{c}_{m}| \rightarrow \infty$ and consider the required inequality to apply the corollary of Rouché’s theorem. Let $R = 2^{-2/m}$ then we have that $$\begin{aligned}
& 1 + 2|c_{m}-1|R^{m+1} + |c_{m}|R^{2m+1} + R^{2m+2} \\
& \le 1 + 2(|c_{m}|+1)2^{-2-2/m} + |c_{m}|2^{-4-2/m} + 2^{-4-4/m} \\
& = |c_{m}|2^{-2/m} \left( 2^{-1} + 2^{-4} + |c_{m}|^{-1}(2^{2/m} + 2^{-1} + 2^{-4-2/m}) \right) \\
& < |c_{m}|R,\end{aligned}$$ where the final inequality is true for all $m\ge2$ when $|c_{m}|>\frac{81}{14}$. Thus, for large enough $m$, there is only a single root $s_{m}$ such that $|s_{m}| < 2^{-2/m}$ and it is simple. As $m\rightarrow\infty$ this shows that all but two roots, $s_{m}$ and $s_{m}^{-1}$, tend to one in modulus.
We now investigate the behaviour of $s_{m}$ and $s_{m}^{-1}$. Let $R = 2|c_{m}|^{-1}$ and suppose, w.l.o.g., that $|c_{m}|\ge4$ and $m\ge4$. Then we have that $$\begin{aligned}
& 1 + 2|c_{m}-1|R^{m+1} + |c_{m}|R^{2m+1} + R^{2m+2} \\
& \le 1 + 2 (|c_{m}|+1) \frac{2^{m+1}}{|c_{m}|^{m+1}} + \frac{2^{2m+1}}{|c_{m}|^{2m}} + \frac{2^{2m+2}}{|c_{m}|^{2m+2}} \\
& \le 1 + 2\left(2^{-(m-1)} + 2^{-(m+1)}\right) + 2^{-(2m-1)} + 2^{-(2m+2)} \\
& < 1 + 2^{3-m} \\
& < |c_{m}|R,\end{aligned}$$ and so we deduce that $s_{m}$ is such that $|s_{m}| < 2|c_{m}|^{-1}$ for large enough $m$, showing that $s_{m}\rightarrow0$ as $m\rightarrow\infty$. This bound on $|s_{m}|$ gives that $$\begin{aligned}
s_{m}^{2m+2} - c_{m}s_{m}^{2m+1} + 2(c_{m}-1)s_{m}^{m+1} \rightarrow 0.\end{aligned}$$ Hence, since $f_{m}(s_{m})=0$ for all $m$, we must have that $1-c_{m}s_{m} \rightarrow 0$ and thus $s_{m} - c_{m}^{-1} \rightarrow 0$ as $m\rightarrow\infty$. We would now like to interpret what this shows for the corresponding root $z_{m}$ using the $q$-discriminant condition but we cannot simply insert $q=0$ and taking the limit $q=s_{m}\rightarrow0$ shows that $z_{m}\rightarrow0$ which can already be deduced from the fact that $|c_{m}|\rightarrow\infty$. Instead, we use the definition of $c_{m} = {a^{2}}/{z_{m}^{2}}$ and denote $\delta_{m} =c_{m}s_{m}-1$ where $\delta_{m}\rightarrow0$ as $m\rightarrow\infty$. Then, with $q=s_{m}=c_{m}^{-1}(1+\delta_{m})$, becomes $$\begin{aligned}
\nonumber
& \frac{B(z_{m})^{2}}{A(z_{m})} = c_{m}^{-1}(1+\delta_{m}) + c_{m}(1+\delta_{m})^{-1} + 2 \\
\nonumber
\iff & \frac{(- z_{m}^2 + b^2 - a^2)^{2}}{a^{2}z_{m}^{2}} = \frac{z_{m}^{2}}{a^{2}}(1+\delta_{m}) + \frac{a^{2}}{z_{m}^{2}}(1+\delta_{m})^{-1} + 2 \\
\label{SingularlyPerturbedEquationForz0}
\iff & b^{4} - 2 a^2 b^{2} - 2 b^{2} z_{m}^2 = \delta_{m} (z_{m}^{4}-a^4) + \mathcal{O}(\delta_{m}^{2}),\end{aligned}$$ where we have used the binomial expansion $(1+\delta_{m})^{-1}=1-\delta_m + \mathcal{O}(\delta_{m}^{2})$, which is valid for large $m$ since $\delta_m \to 0$. Now note that is a singular perturbation and as $\delta_{m}\rightarrow0$ all roots $z_m$ go to infinity except for those which satisfy the left-hand side being zero, namely we have the limiting roots $$\begin{aligned}
\label{Limiting-z-different-case}
z = \pm \sqrt{\tfrac{1}{2}b^{2} - a^{2}}.\end{aligned}$$ Now since $z_{m}\rightarrow0$ by the assumption that $|c_{m}|\rightarrow\infty$, this case of unbounded $c_{m}$ can only occur when the limiting roots defined in are zero, that is when $a^{2} = \frac{1}{2}b^{2}$.
#### Case 2
We now turn to the analysis of case \[Case2\] where $1 < c \le |\hat{c}_{m}| \le C$ and let $\varepsilon>0$ be small. Suppose, w.l.o.g., that $c \ge \frac{1+\varepsilon}{1-\varepsilon}$, or else take a smaller $\varepsilon$. Now consider $R = 1-\varepsilon$ and let $M_{C}$ be such that $$\begin{aligned}
2(C+1) R^{m+1} + C R^{2m+1} + R^{2m+2} < \varepsilon\end{aligned}$$ for all $m \ge M_{C}$. Then $$\begin{aligned}
1 + 2|c_{m}-1|R^{m+1} + |c_{m}|R^{2m+1} + R^{2m+2} & \le 1 + 2(C+1) R^{m+1} + C R^{2m+1} + R^{2m+2} \\
& < 1 + \varepsilon \\
& \le |c_{m}|R\end{aligned}$$ for $m \ge M_{C}$. Thus there is only a single root $s_{m}$ with modulus less than $R$ and it is simple. Since $\varepsilon>0$ was arbitrary, taking the limit $\varepsilon\rightarrow0$ so that $R\rightarrow1$ shows that all remaining roots of $f_{m}$ tend to one in modulus. Now let $R = (1+\varepsilon)c^{-1}$ with $\varepsilon$ small enough so that $R<1$, then the same argument above applies similarly. As such, for large enough $m$, taking $\varepsilon = \frac{c-1}{c+1}$ shows that $|s_{m}|<\frac{2}{c+1}<1$. Since $|s_{m}|$ is bounded away from one, an identical argument to that made in case \[Case1\] shows that $s_{m} - c_{m}^{-1} \rightarrow 0$ as $m\rightarrow\infty$ and so the corresponding $z_{m} = \sqrt{a^{2}/c_{m}}$ must tend to the limiting roots in , but now these roots cannot be zero as $|c_{m}|$ is bounded. Furthermore, since we are considering case \[Case2\] we require $|c_{m}|>1$, and so $|a^{2}|>|z_{m}^{2}|$, to hold for arbitrarily large $m$. For this to hold in the limit we require $|a^{2}|>|\frac{1}{2}b^{2} - a^{2}|$.
#### Case 3
Consider now case \[Case3\], where $|\hat{c}_{m}| \rightarrow 1$ from above, and let $\varepsilon>0$ be small. Let $R = 1-\varepsilon$, and define $\bar{C} = \max_{m} |c_{m}|$, $\bar{M}$ such that $$\begin{aligned}
2(\bar{C}+1) R^{m+1} + \bar{C} R^{2m+1} + R^{2m+2} < \frac{\varepsilon}{2}\end{aligned}$$ for all $m\ge\bar{M}$, and $\hat{M}$ such that $$\begin{aligned}
|c_{m}| \le \frac{1-\frac{\varepsilon}{2}}{1-\varepsilon}\end{aligned}$$ for all $m\ge\hat{M}$. Then, $$\begin{aligned}
& |c_{m}|R + 2|c_{m}-1|R^{m+1} + |c_{m}|R^{2m+1} + R^{2m+2} \\
& \le \frac{1-\frac{\varepsilon}{2}}{1-\varepsilon} R + 2(\bar{C}+1) R^{m+1} + \bar{C} R^{2m+1} + R^{2m+2} \\
& < 1,\end{aligned}$$ for all $m \ge \max \lbrace \bar{M}, \hat{M} \rbrace$. Thus, for large enough $m$, there are no roots of $f_{m}$ with modulus less than $R = 1-\varepsilon$. Since $\varepsilon>0$ was arbitrary, taking the limit $\varepsilon\rightarrow0$ so that $R\rightarrow1$ shows that all roots of $f_{m}$ tend to one in modulus.
#### Case 4
Finally, consider case \[Case4\] where $|\hat{c}_{m}| \le 1$. Let $\varepsilon>0$ be small, and define $R = 1-\varepsilon$. Let $M_{1}$ be such that $$\begin{aligned}
4 R^{m+1} + R^{2m+1} + R^{2m+2} < \varepsilon\end{aligned}$$ for all $m \ge M_{1}$. Then we have that $$\begin{aligned}
|c_{m}|R + 2|c_{m}-1|R^{m+1} + |c_{m}|R^{2m+1} + R^{2m+2} & \le R + 4 R^{m+1} + R^{2m+1}+ R^{2m+2} \\
& < 1,\end{aligned}$$ for $m \ge M_{1}$. Thus, as in case \[Case3\], we deduce that all roots of $f_{m}$ tend to one in modulus as $m\rightarrow\infty$.
We have already seen that in cases \[Case1\] and \[Case2\] we have simple roots $z_{m}$ which tend to the limiting values in when $|a^{2}|>|\frac{1}{2}b^{2} - a^{2}|$. In all cases, and thus all sequences of possible $c_{m}$ values, all remaining roots $z_{m}$ correspond to $q$ values which tend to the unit circle. Thus $z_{m}$ must tend to the limiting curve defined by where $q = e^{i\phi}$ for some $\phi \in [-\pi,\pi]$. This limiting curve in the complex plane is given parametrically as $$\begin{aligned}
& \frac{B(z)^{2}}{A(z)} = e^{i\phi} + e^{-i\phi} + 2, & & \phi \in [-\pi,\pi] \\
\iff & \frac{(- z^2 + b^2 - a^2)^{2}}{a^{2}z^{2}} = 4 \cos^{2}\left(\frac{\phi}{2}\right), & & \phi \in [-\pi,\pi] \\
\iff & z^2 - b^2 + a^2 = \pm 2 az \cos\left(\frac{\phi}{2}\right), & & \phi \in [-\pi,\pi] \\
\iff & z^2 - 2 a \cos(\theta) z - b^2 + a^2 = 0, & & \theta \in [-\pi,\pi] \\
\iff & z = a \cos(\theta) \pm \sqrt{b^2 - a^{2} \sin^{2}(\theta)}, & & \theta \in [-\pi,\pi].\end{aligned}$$ Thus, as roots $z_{m}$ of $p_{m}(z)$ are eigenvalues $\lambda$ of $\mathcal{T}\in\mathbb{C}^{2m\times2m}$, we deduce that the limiting spectra of $\mathcal{T}$ lies on the curve defined by as $m\rightarrow\infty$, except perhaps for simple eigenvalues which can only occur if $|a^{2}|>|\frac{1}{2}b^{2} - a^{2}|$.
We note that, while the so-called Szegő formula does not apply in our non-Hermitian case, we have just proven that all eigenvalues of $\mathcal{T}$, except perhaps two, lie on the equivalent curve defined by eigenvalues of the (block) symbol of $\mathcal{T}$, which is precisely that defined in .
The one-dimensional problem {#sec:1dcase}
===========================
We now turn our attention to analysing the one-level method. In this section we study the parallel Schwarz iterative method for the one-dimensional Maxwell’s equations with Robin boundary conditions defined on the domain : $$\begin{aligned}
\label{eq:1d0}
\left\{\begin{array}{r@{}ll}
{\cal L}u \vcentcolon= -\partial_{xx}u + (ik\tilde \sigma - k^2)u &{}= 0, & x \in (a_1,b_N),\\
{\cal B}_lu \vcentcolon= -\partial_x u + \alpha u &{}= g_1, & x = a_1,\\
{\cal B}_ru \vcentcolon= \partial_x u + \alpha u &{}= g_2, & x = b_N,
\end{array}\right.\end{aligned}$$ where $u$ represents the complex amplitude of the electric field, $k$ is the wave number, and $\tilde\sigma = \sigma Z$ with $\sigma$ being the conductivity of the medium and $Z$ its impedance. Here $\alpha$ is the impedance parameter which is chosen such that the local problems are well-posed and is classically set to $ik$, in which case the problem corresponds to a [*“one-dimensional wave-guide”*]{} and the incoming wave or excitation can be represented by $g_1$, for example, with $g_2$ being set to $0$. Note that, when $\alpha = ik$, the problem is well-posed even if $\tilde\sigma=0$ but in the following we will assume that $\tilde\sigma>0$. In order to simplify notation we will omit the tilde symbol for $\sigma$. We remark that can also be seen as an absorptive Helmholtz equation where the absorption term $ik\sigma$ comes from the physics of the problem.
Let us also consider two sets of points $\{a_j\}_{j=1,..,N+1}$ and $\{b_j\}_{j=0,..,N}$ defining the overlapping decomposition $\Omega = \cup_{j=1}^N \Omega_j$ such that $\Omega_{j} = (a_j,b_j)$, as illustrated in (and considered in [@Chaouqui:2018:OSC]), where $$\begin{aligned}
\label{eq:ab}
b_j-a_j &= L +2\delta, & b_{j-1}-a_j &= 2\delta, & a_{j+1}-a_j &= b_{j+1}-b_j = L, & \delta &> 0. \end{aligned}$$ Note that the length of each subdomain is fixed and equal to $L+2\delta$ while the overlap is always $2\delta$. We consider solving by a Schwarz iterative algorithm and denote by $u_j^n$ the approximation to the solution in subdomain $j$ at iteration $n$, starting from an initial guess $u_j^0$. We compute $u_j^n$ from the previous values $u_j^{n-1}$ by solving the following local boundary value problem
\[eq:sch1dall\] $$\begin{aligned}
\label{eq:sch1d}
\left\{\begin{array}{r@{}ll}
{\cal L} u_j^n &{}= 0, & x \in \Omega_j,\\
{\cal B}_l u_j^n &{}= {\cal B}_l u_{j-1}^{n-1}, & x = a_j,\\
{\cal B}_r u_j^n &{}= {\cal B}_r u_{j+1}^{n-1}, & x = b_j,
\end{array}\right.\end{aligned}$$ in the case $2\le j \le N$ while for the first ($j=1$) and last ($j=N$) subdomain we have $$\begin{aligned}
\label{eq:sch1d1N}
\begin{array}{cc}
\left\{\begin{array}{r@{}ll}
{\cal L}u_1^n &{}= 0, & x \in \Omega_1,\\
{\cal B}_l u_1^n &{}= g_1, & x = a_1,\\
{\cal B}_r u_1^n &{}= {\cal B}_r u_{2}^{n-1}, & x = b_1,
\end{array}\right.
&
\left\{\begin{array}{r@{}ll}
{\cal L}u_N^n &{}= 0, & x \in \Omega_N,\\
{\cal B}_l u_N^n &{}= {\cal B}_l u_{N-1}^{n-1}, & x = a_N,\\
{\cal B}_r u_N^n &{}= g_2, & x = b_N.
\end{array}\right.
\end{array}\end{aligned}$$
In the following we wish to analyse the convergence of the iterative method defined in , which is also called the [*parallel Schwarz method*]{}, for a growing number of subdomains $N$ and for the absorptive problem, i.e. $\sigma>0$.[^4]
(-0.5,0) – (11,0) node\[right\] [$x$]{}; in [0,0.5,2,2.5,4,4.5,6,6.5,8,8.5,10,10.5]{} [ (,-0.1) – (,0.1); ]{} (0,0.5) node\[below\] [$a_{1}$]{}; (2,0.5) node\[below\] [$a_{2}$]{}; (4,0.5) node\[below\] [$a_{j}$]{}; (6,0.5) node\[below\] [$a_{j+1}$]{}; (8,0.5) node\[below\] [$a_{N}$]{}; (10,0.5) node\[below\] [$a_{N+1}$]{}; (0.5,-0.05) node\[below\] [$b_{0}$]{}; (2.5,-0.05) node\[below\] [$b_{1}$]{}; (4.5,-0.05) node\[below\] [$b_{j-1}$]{}; (6.5,-0.05) node\[below\] [$b_{j}$]{}; (8.5,-0.05) node\[below\] [$b_{N-1}$]{}; (10.5,-0.05) node\[below\] [$b_{N}$]{}; (3.25,0.5) node\[below\] [$\cdots$]{}; (3.25,-0.05) node\[below\] [$\cdots$]{}; (7.25,0.5) node\[below\] [$\cdots$]{}; (7.25,-0.05) node\[below\] [$\cdots$]{}; (0,-0.75) – (2.5,-0.75); (4,-0.75) – (6.5,-0.75); (8,-0.75) – (10.5,-0.75); (1.25,-0.75) node\[below\] [$\Omega_{1}$]{}; (5.25,-0.75) node\[below\] [$\Omega_{j}$]{}; (9.25,-0.75) node\[below\] [$\Omega_{N}$]{}; (4,0.75) – (4.5,0.75); (4.5,0.75) – (6,0.75); (6,0.75) – (6.5,0.75); (4.25,0.8) node\[above\] [$2\delta$]{}; (5.25,0.8) node\[above\] [$L-2\delta$]{}; (6.25,0.8) node\[above\] [$2\delta$]{};
In order to do this we define the local errors in each subdomain $j$ at iteration $n$ as $e^n_j = u|_{\Omega_j} - u^n_j$. They verify the boundary value problems for the interior subdomains and the homogeneous analogues of for the first and last subdomains (i.e. but with boundary conditions $g_{1}=0$ and $g_{2}=0$). The convergence study will be done in two steps: first we prove that the Schwarz iteration matrix is a block Toeplitz matrix and then that its spectral radius is less than one.
\[lemma:basic1d\] If $e^n_j = u|_{\Omega_j} - u^n_j$ is the local error in each subdomain $j$ at iteration $n$ and $$\begin{aligned}
{{\cal R}^n} &\vcentcolon= [{\cal R}^{n}_+(b_1), \, {\cal R}^{n}_-(a_2), \, {\cal R}^{n}_+(b_2), \ldots, {\cal R}^{n}_-(a_{N-1}), \,{\cal R}^{n}_+(b_{N-1}), \, {\cal R}^{n}_-(a_N)]^T,\end{aligned}$$ where $$\begin{aligned}
\label{eq:rpm}
{\cal R}^{n}_-(a_j) &\vcentcolon= {\cal B}_l e_{j-1}^{n}(a_j), & {\cal R}^{n}_+(b_j) &\vcentcolon={\cal B}_r e_{j+1}^{n}(b_j),\end{aligned}$$ is the Robin interface data, then $$\begin{aligned}
{{\cal R}^n} = {\cal T}_{1d} {{\cal R}^{n-1}},\end{aligned}$$ where ${\cal T}_{1d} $ is a block Toeplitz matrix of the form with $a$ and $b$ given by
\[eq:ab0\] $$\begin{aligned}
a &= \frac{(\zeta + \alpha )^2 e^{2\zeta \delta} - (\zeta - \alpha)^2 e^{-2\zeta \delta}}{(\zeta + \alpha)^2 e^{\zeta (2\delta+L)} - (\zeta - \alpha)^2 e^{-\zeta (2\delta+L)}}, \\
b &= -\frac{(\zeta^2-\alpha^2)(e^{\zeta L} - e^{-\zeta L} )}{ (\zeta + \alpha)^2 e^{\zeta (2\delta+L)} - (\zeta - \alpha)^2 e^{-\zeta (2\delta+L)}},\end{aligned}$$
where $\zeta = \sqrt{ik\sigma-k^2}$.
We first see that the solution to ${\cal L}e_j^n=0$ is given by $$\begin{aligned}
\label{eq:ajbj}
e_j^n(x) &= \alpha_j^n e^{-\zeta x} + \beta_j^n e^{\zeta x}, & \zeta &= \sqrt{ik\sigma-k^2}. \end{aligned}$$ Note that we choose the principle branch of the square root here so that $\zeta$ always has positive real and imaginary parts. Now the interface iterations at $x=a_j$ and $x=b_j$ from can be written in terms of the error as $$\begin{aligned}
\label{eq:blbr}
\left[\begin{array}{c} {\cal B}_l e_j^n(a_j) \\ {\cal B}_r e_j^n(b_j) \end{array}\right] = \left[\begin{array}{c} {\cal B}_l e_{j-1}^{n-1}(a_j) \\ {\cal B}_r e_{j+1}^{n-1}(b_j)\end{array}\right]. \end{aligned}$$ By introducing into the left-hand side of and by using the notation from we obtain $$\begin{aligned}
\left[\begin{array}{cc}
(\zeta + \alpha) e^{-\zeta a_j} & - (\zeta - \alpha) e^{\zeta a_j} \\
-(\zeta - \alpha ) e^{-\zeta b_j} & (\zeta + \alpha) e^{\zeta b_j}
\end{array}
\right] \left[\begin{array}{c} \alpha_j^n \\ \beta_j^n \end{array}\right] = \left[\begin{array}{c} {\cal R}^{n-1}_-(a_j)\\ {\cal R}^{n-1}_+(b_j)\end{array}\right],\end{aligned}$$ which we can solve for the unknowns $\alpha_j^n$ and $\beta_j^n$ to give $$\begin{aligned}
\label{eq:ajbj0}
\left[\begin{array}{c} \alpha_j^n \\ \beta_j^n \end{array}\right] & = \frac{1}{D_j} \left[\begin{array}{cc}
(\zeta+ \alpha) e^{\zeta b_j} & (\zeta - \alpha) e^{\zeta a_j} \\
(\zeta - \alpha) e^{-\zeta b_j} & (\zeta + \alpha) e^{-\zeta a_j}
\end{array}
\right] \left[\begin{array}{c} {\cal R}^{n-1}_-(a_j) \\{\cal R}^{n-1}_+(b_j) \end{array}\right],\end{aligned}$$ where $D_j = (\zeta + \alpha)^2 e^{\zeta (b_j-a_j)} - (\zeta - \alpha)^2 e^{\zeta (a_j-b_j)}$. Note that, since $b_j-a_j = L+2\delta$, then $D_{j}$ is actually independent of $j$ and thus we simply denote it by $D$. The algorithm is based on Robin transmission conditions, hence the quantities of interest which are transmitted at the interfaces between subdomains are the Robin data . Therefore, we need to compute the current interface values ${\cal R}^{n}_-(a_j)$ and ${\cal R}^{n}_+(b_j)$ by replacing the coefficients from into and then applying the formulae in , giving
\[eq:rjajbj\] $$\begin{aligned}
\begin{split}
{\cal R}^{n}_-(a_j) & = {\cal B}_l e_{j-1}^{n}(a_j) = (\zeta +\alpha) \alpha^n_{j-1}e^{-\zeta a_j} - (\zeta - \alpha) \beta^n_{j-1}e^{\zeta a_j} \\
& = \frac{1}{D} \Bigl[ ((\zeta + \alpha)^2 e^{\zeta (b_{j-1}-a_j)} - (\zeta - \alpha)^2 e^{\zeta (a_j-b_{j-1})}){\cal R}^{n-1}_-(a_{j-1}) \Bigr.\\
& \mathrel{\phantom{=}} \mathrel+ \Big. (\zeta^2-\alpha^2)( e^{\zeta (a_{j-1}-a_{j})}-e^{\zeta (a_{j}-a_{j-1})} ) {\cal R}^{n-1}_+(b_{j-1}) \Bigr],
\end{split} \\
\begin{split}
{\cal R}^{n}_+(b_j) & = {\cal B}_r e_{j+1}^{n}(b_j) = -(\zeta -\alpha) \alpha^n_{j+1}e^{-\zeta b_j} + (\zeta +\alpha) \beta^n_{j+1}e^{\zeta b_j} \\
& = \frac{1}{D} \Bigl[ (\zeta^2-\alpha^2)(e^{\zeta (b_{j}-b_{j+1})}-e^{\zeta (b_{j+1}-b_{j})} )
{\cal R}^{n-1}_-(a_{j+1}) \Bigr. \\
& \mathrel{\phantom{=}} \mathrel+ \Bigl. ((\zeta + \alpha)^2 e^{\zeta (b_j-a_{j+1})} - (\zeta - \alpha)^2 e^{\zeta (a_{j+1}-b_j)}) {\cal R}^{n-1}_+(b_{j+1}) \Bigr].
\end{split}\end{aligned}$$
The iteration of interface values can be summarised as follows:
\[eq:rniterall\] $$\begin{aligned}
\label{eq:rniter}
\begin{split}
\left[\begin{array}{c} {\cal R}^{n}_-(a_j) \\ {\cal R}^{n}_+(b_j) \end{array}\right] & = T_1 \left[\begin{array}{c} {\cal R}^{n-1}_-(a_{j-1}) \\ {\cal R}^{n-1}_+(b_{j-1}) \end{array}\right] + T_2 \left[\begin{array}{c} {\cal R}^{n-1}_-(a_{j+1}) \\ {\cal R}^{n-1}_+(b_{j+1}) \end{array}\right], \\ T_1 & = \left[\begin{array}{cc} a & b \\ 0 & 0 \end{array} \right], \ T_2 = \left[\begin{array}{cc} 0 & 0 \\ b & a \end{array} \right],
\end{split}\end{aligned}$$ where $a$ and $b$ are given by . Note that since the homogeneous counterparts of the boundary conditions from translate into ${\cal R}^{n}_-(a_1)=0$ and ${\cal R}^{n}_+(b_N)=0$ for all $n$, we can remove these terms. As such, the iterates for $j\in \{1,2,N-1,N\}$ are prescribed slightly differently as $$\begin{aligned}
\label{eq:rniterends}
\begin{split}
\left[\begin{array}{c} 0 \\ {\cal R}^{n}_+(b_1) \end{array}\right] &= T_2 \left[\begin{array}{c} {\cal R}^{n-1}_-(a_{2}) \\ {\cal R}^{n-1}_+(b_{2}) \end{array}\right], \\
\left[\begin{array}{c} {\cal R}^{n}_-(a_2) \\ {\cal R}^{n}_+(b_2) \end{array}\right] &= T_1 \left[\begin{array}{c} 0 \\{\cal R}^{n-1}_+(b_1) \end{array}\right] + T_2 \left[\begin{array}{c} {\cal R}^{n-1}_-(a_3) \\ {\cal R}^{n-1}_+(b_3) \end{array}\right], \\
\left[\begin{array}{c} {\cal R}^{n}_-(a_{N-1}) \\ {\cal R}^{n}_+(b_{N-1}) \end{array}\right] &= T_1 \left[\begin{array}{c} {\cal R}^{n-1}_-(a_{N-2}) \\ {\cal R}^{n-1}_+(b_{N-2}) \end{array}\right] + T_2 \left[\begin{array}{c} {\cal R}^{n-1}_-(a_N) \\ 0\end{array}\right], \\
\left[\begin{array}{c} {\cal R}^{n}_-(a_N) \\ 0 \end{array}\right] &= T_1 \left[\begin{array}{c} {\cal R}^{n-1}_-(a_{N-1}) \\ {\cal R}^{n-1}_+(b_{N-1}) \end{array}\right].
\end{split}\end{aligned}$$
With the notation from , global iteration over interface data belonging to all subdomains becomes ${{\cal R}^n} = {\cal T}_{1d} {{\cal R}^{n-1}}$ where $$\begin{aligned}
\label{eq:itert1}
{{\cal T}_{1d}} = \left[\begin{array}{ccccccc}
0 & \hat T_2 & & & & & \\
\tilde T_1 & 0_{2\times 2} & T_2 & & & & \\
& \ddots & \ddots & \ddots & & & \\
& & T_1 &0_{2\times 2}& T_2 & &\\
& & & \ddots & \ddots & \ddots & \\
& & & & T_1 & 0_{2\times 2}& \tilde T_2 \\
& & & & & \hat T_1 & 0
\end{array}\right]\end{aligned}$$ with $\tilde T_1 = \left[\begin{array}{cc} b & 0 \end{array} \right]^T$, $\tilde T_2 = \left[\begin{array}{cc} 0 & b \end{array} \right]^T$, $\hat T_1 = \left[\begin{array}{cc} a & b \end{array} \right]$, $\hat T_2 = \left[\begin{array}{cc} b & a \end{array} \right]$. We conclude from this that the Schwarz algorithm is given by a stationary iteration with iteration matrix ${\cal T}_{1d}$ given by and, therefore, convergence is determined by the spectral radius $\rho({\cal T}_{1d})$. We also notice that ${\cal T}_{1d}$ is a block Toeplitz matrix of the form where $a$ and $b$ are given by .
Before proving convergence of the Schwarz algorithm, we first utilise to provide a useful intermediary result, which will also aid our analysis in the two-dimensional case.
\[lemma:LimitingSpectralRadius1D\] The following relation holds: $$\begin{aligned}
\max_{\theta \in [-\pi,\pi]}\left|a \cos(\theta) \pm \sqrt{b^2 - a^2 \sin^2(\theta)}\right| = \max\{ |a + b|, | a -b|\},\end{aligned}$$ and thus the convergence factor $R_{1d} \vcentcolon= \lim_{N\rightarrow\infty} \rho({\cal T}_{1d})$ of the Schwarz algorithm as the number of subdomains tends to infinity verifies $$\begin{aligned}
\label{eq:r1d}
R_{1d} \le
\left\{\begin{array}{ll}
\max \left\lbrace |a + b|, | a - b| \right\rbrace & \text{if } \left|a^2-\frac{1}{2}b^2\right|^{1/2} \ge |a|, \\
\max \left\lbrace |a + b|, | a - b|, |a| \right\rbrace & \text{if } \left|a^2-\frac{1}{2}b^2\right|^{1/2} < |a|.
\end{array}\right.\end{aligned}$$ Further, consider the change of variables $$\begin{aligned}
\label{eq:params}
z &= 2\delta\zeta, & l &= \frac{L}{2\delta}, & \gamma &= 2\delta \alpha, & v &= \frac{z-\gamma}{z+\gamma},\end{aligned}$$ and let $z \vcentcolon= x+iy$. Then the condition $g_{\pm}(z;\delta,l) > 0$, where $$\begin{aligned}
\label{eq:g}
g_{\pm}(z;\delta,l) &= (e^{2lx} - 1)(e^{2x} - |v|^2) \pm 4 \sin(ly) (\Im v\cos{y} - \Re v \sin{y}) e^{x(l+1)},\end{aligned}$$ will ensure that $ \max\{ |a + b|, | a - b| \} <1$. Similarly, the condition $g(z;\delta,l) > 0$, where $$\begin{aligned}
\label{eq:ga}
\begin{split}
g(z;\delta,l) &= (e^{2lx} - 1)(e^{2x(l+2)} - |v|^4) + 4 \sin(ly) \\
& \mathrel{\phantom{=}} \mathrel\cdot \left[((\Re v)^2-(\Im v)^2) \sin(y(l+2)) - 2 \Re v \Im v \cos(y(l+2))\right]e^{2x(l+1)},
\end{split}\end{aligned}$$ will ensure that $|a|<1$.
Since $\mathcal{T}_{1d}$ is of the form $\mathcal{T}$ in , provides its limiting spectrum and thus allows us to bound $R_{1d}$ by the largest eigenvalue in magnitude. We first bound $\lambda_{\pm}(\theta) = a \cos(\theta) \pm \sqrt{b^2 - a^2 \sin^2(\theta)}$. It is straightforward to see that these values are the eigenvalues of the matrix $$\begin{aligned}
T = \left(\begin{array}{cc}
a \cos(\theta) & b - a\sin(\theta) \\
b + a\sin(\theta) & a \cos(\theta)
\end{array}
\right).\end{aligned}$$ A simple computation shows that the matrix $$\begin{aligned}
T^*T = \left(\begin{array}{cc}
|a|^2 +|b|^2 +(a\bar b+\bar ab)\sin(\theta) & (a\bar b+\bar ab)\cos(\theta)\\
(a\bar b+\bar ab)\cos(\theta) & |a|^2 +|b|^2 - (a\bar b+\bar ab)\sin(\theta)
\end{array}
\right)\end{aligned}$$ has the eigenvalues $\mu_{\pm} = |a\pm b|^2$. We can now conclude that $$\begin{aligned}
|\lambda_{\pm}(\theta)| \le \|T\| = \sqrt{\|T^*T\|} = \sqrt{\max\{\mu_+,\mu_-\} } = \max\{|a+b|,|a-b|\}.\end{aligned}$$ Additionally, states that eigenvalues $\lambda = \pm(\tfrac{1}{2}b^{2} - a^{2})^{1/2}$ may belong to the limiting spectrum but only if they have magnitude strictly less than $|a|$. Together, these two cases yield .
Let us consider now the complex valued functions $F_{\pm} \colon \mathbb{C} \rightarrow \mathbb{C}$ $$\begin{aligned}
F_{\pm}(z) = \frac{(z + \gamma)^2 e^{z} - (z- \gamma )^2 e^{-z}}{(z + \gamma)^2 e^{(l+1)z} - (z - \gamma)^2 e^{-(l+1)z}} \pm \frac{(z^2-\gamma^2)(e^{lz} -e^{-lz} ) }{(z + \gamma)^2 e^{(l+1)z} - (z-\gamma)^2 e^{-(l+1)z}}.\end{aligned}$$ It is easy to see that ${a \mp b} = F_{\pm} (z)$ when $z$, $l$ and $\gamma$ are as defined in . Similarly, we define the function $G \colon \mathbb{C} \to \mathbb{C}$ to be the first term in $F_{\pm}(z)$ so that $a = G(z)$. Let us simplify in the first instance the expression of $|F_{\pm}(z)|$ without using any assumption on $z \vcentcolon= x+iy$. For this we consider the transformation $v$ along with its polar form $$\begin{aligned}
\label{eq:v}
v &\vcentcolon= \frac{z-\gamma}{z+\gamma}, & v &= w(\cos(\varphi)+i\sin (\varphi)), & w=|v|.\end{aligned}$$ After some lengthy but elementary calculations we find that
\[eq:fpmandgpm\] $$\begin{aligned}
\label{eq:fpm}
|F_{\pm}(z)|^2 &= 1 - \frac{(e^{x(l + 1)}-w)^2+2w(1\mp\cos\left({(l + 1)y} - \varphi\right))e^{x(l + 1)} }{(e^{2x(l + 1)}-w^2)^2+4w^2\sin^2((l + 1)y - \varphi)e^{2x(l + 1)} } g_{\pm}(z;\delta,l)\\
\begin{split}
\label{eq:fpmgpm}
g_{\pm}(z;\delta,l) &= (e^{2lx} - 1)(e^{2x} - w^2) \pm 4 w\sin(ly)\sin(\varphi - y)e^{x(l + 1)}.
\end{split}\end{aligned}$$
We observe that the fraction in is positive, since the individual terms involved are, and thus $\max\{|a-b|,|a+b|\} < 1 \Leftrightarrow |F_{\pm}(z)|^2 < 1 \Leftrightarrow g_{\pm}(z;\delta,l) > 0$. We can now rewrite $g_{\pm}(z;\delta,l)$ in using and convert $v$ to Cartesian form to obtain the required expression in . A near identical argument can be used to derive conditions for $|G(z)|^2 < 1$ and results in the criterion that $g(z;\delta,l) > 0$, where $g(z;\delta,l)$ is defined by . Thus the required conclusions follow.
We are now ready to state our main convergence result for the one-dimensional problem in the case when $\alpha = ik$, namely that of classical impedance conditions.
\[theorem:1d\] If $\alpha = ik$ (the case of classical impedance conditions), then for all $k>0$, $\sigma>0$, $\delta>0$ and $L>0$ we have that $R_{1d} < 1$. Therefore the convergence will ultimately be independent of the number of subdomains (we say that the Schwarz method will scale).
By we see that it is enough to study the sign of $g_{\pm}(z;\delta,l)$ and of $g(z;\delta,l)$. We can see that if $\alpha = ik$ and $\kappa = 2\delta k$ then for $z\vcentcolon= x+iy$ becomes $$\begin{aligned}
\Re v &= \frac{-\kappa^2 + x^2 + y^2}{(\kappa+y)^2 + x^2}, & \Im v &= \frac{-2\kappa x}{(\kappa+y)^2 + x^2}, & |v|^2 &= \frac{(\kappa-y)^2 + x^2}{(\kappa+y)^2 + x^2} < 1,\end{aligned}$$ the final inequality holding since $\kappa > 0$ and $y > 0$. We emphasise that $x$ and $y$ are the real and imaginary parts of $z=2\delta\zeta$ and so are positive by the nature of $\zeta$ in . Now we can further simplify using these expressions for $v$ to obtain
\[eq:gpm\] $$\begin{aligned}
\label{eq:gpmg}
g_{\pm}(z;\delta,l) &= \frac{4e^{x(l+1)}}{(\kappa+y)^2 + x^2} \tilde{g}_{\pm}(z;\delta,l) \\
\begin{split}
\label{eq:gpmgt}
\tilde{g}_{\pm}(z;\delta,l) &= [(\kappa^2+x^2+y^2) \sinh(x) + 2\kappa y \cosh(x)]\sinh(lx) \\
& \mathrel{\phantom{=}} \mathrel\pm [(\kappa^2-x^2-y^2) \sin(y) - 2\kappa x \cos(y)]\sin(ly).
\end{split}\end{aligned}$$
Proving positivity of $g_{\pm}(z;\delta,l)$ is then equivalent to positivity of $\tilde g_{\pm}(z;\delta,l)$. To proceed we relate $x$ and $y$ by considering the real part of $z^2 = (x+iy)^2 = 2i\kappa\delta\sigma - \kappa^2$ which yields $y^2 = \kappa^2 + x^2$. Let us now eliminate $y$ using this identity to obtain $$\begin{aligned}
\begin{split}
\tilde{g}_{\pm}(z;\delta,l) &= 2 \left[(\kappa^2+x^2) \sinh(x) + \kappa \sqrt{\kappa^2+x^2} \cosh(x)\right]\sinh(lx) \\
& \mathrel{\phantom{=}} \mathrel\mp 2 \left[x^2 \sin(\sqrt{\kappa^2+x^2}) + \kappa x \cos(\sqrt{\kappa^2+x^2})\right]\sin(l\sqrt{\kappa^2+x^2}).
\end{split}\end{aligned}$$ To show that this is positive we want to lower bound the hyperbolic term in the first line (which is positive) while making the trigonometric term in the second line as large as possible in magnitude and negative. To do this we make use of some elementary bounds which hold for $t > 0$: $$\begin{aligned}
\label{eq:HypTrigBounds}
|\sin(t)| &< t < \sinh(t), & |\cos(t)| &\le 1 < \cosh(t).\end{aligned}$$ We can now derive the positivity bound on $\tilde{g}_{\pm}(z;\delta,l)$, noting that $x>0$, as follows $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> 2 \left[(\kappa^2+x^2)x + \kappa \sqrt{\kappa^2+x^2}\right] lx - 2 \left[x^2 \sqrt{\kappa^2+x^2} + \kappa x \right] l\sqrt{\kappa^2+x^2} = 0.\end{aligned}$$ Turning to $g(z;\delta,l)$, we can follow a similar process, simplifying to find that
\[eq:g1d\] $$\begin{aligned}
\label{eq:g1dg}
g(z;\delta,l) &= \frac{4e^{2x(l+1)}}{((\kappa+y)^2 + x^2)^2} \tilde{g}(z;\delta,l) \\
\begin{split}
\label{eq:g1dgt}
\tilde{g}(z;\delta,l) &= \bigl[ ((\kappa^2+x^2+y^2)^2+4\kappa^2y^2) \sinh(x(l+2)) \\
& \qquad\qquad\;\ \mathrel+ 4\kappa y(\kappa^2+x^2+y^2) \cosh(x(l+2)) \bigr] \sinh(lx) \\
& \mathrel{\phantom{=}} \mathrel+ \bigl[ ((-\kappa^2+x^2+y^2)^2-4\kappa^2x^2) \sin(y(l+2)) \\
& \qquad\qquad\;\ \mathrel+ 4\kappa x(-\kappa^2+x^2+y^2) \cos(y(l+2)) \bigr] \sin(ly).
\end{split}\end{aligned}$$
Using the identity $y^2 = \kappa^2 + x^2$ along with the elementary bounds we obtain $$\begin{aligned}
\begin{split}
\tilde{g}(z;\delta,l) &= 4 \left[ y^2(y^2+\kappa^2) \sinh(x(l+2)) + 2\kappa y^3 \cosh(x(l+2)) \right] \sinh(lx) \\
& \mathrel{\phantom{=}} \mathrel+ 4 \left[ x^2(x^2-\kappa^2) \sin(y(l+2)) + 2 \kappa x^3 \cos(y(l+2)) \right] \sin(ly)
\end{split} \\
&> 4 \left[ y^2(y^2+\kappa^2) x(l+2) + 2\kappa y^3 \right] lx - 4 \left[ x^2(x^2+\kappa^2) y(l+2) + 2 \kappa x^3 \right] ly \\
&= 4l(l+2)x^2y^2\kappa^2 + 8lxy\kappa^3 \\
&> 0.\end{aligned}$$ Thus, we conclude that for any choice of parameters the required sufficient criteria from on $g_{\pm}(z;\delta,l)$ and $g(z;\delta,l)$ hold and hence $R_{1d} < 1$. Therefore the algorithm will always converge in a number of iterations ultimately independent of the number of subdomains. Nonetheless, note that as any problem parameter shrinks to zero the bounds become tight and so $R_{1d}$ can be made arbitrarily close to one.
In order to verify this result, we compute numerically the spectrum of the iteration matrix and compare it with the theoretical limit for different values of $\sigma$. We choose here $k = 30$, $L=1$ and $\delta = L/10$. From we notice that the spectrum of the iteration matrix tends to the theoretical limit when the number of subdomains becomes large and the algorithm remains convergent. Additionally, when $\sigma$ grows the behaviour of the algorithm improves, which is consistent with the fact that when the absorption in the equations is important (solutions are less oscillatory) or the overlap is large (more information is exchanged) the systems are easier to solve. We also remark an empirical observation that the convergence factor monotonically increases towards the limit given in , thus indicating that the algorithm will always converge for any $N$.
The two-dimensional problem {#sec:2dcase}
===========================
Consider the domain $\Omega = (a_1,b_N)\times (0,\hat L)$ on which we wish to solve the two-dimensional problem and a decomposition into $N$ overlapping subdomains defined by $\Omega_j = (a_j,b_j)\times (0,\hat L)$, where $a_j$ and $b_j$ are as given in . We will analyse the case of the Helmholtz and then Maxwell’s equations.
The Helmholtz equation {#sec:2dhelmholtz}
----------------------
The definition of the parallel Schwarz method for the iterates $u_j^n$ in the case of the two-dimensional Helmholtz problem is $$\begin{aligned}
\label{eq:2dh}
\left\{ \begin{array}{r@{}ll}
(ik\sigma-k^2) u_j^{n} - (\partial_{xx} + \partial_{yy}) u_j^{n} &{}= f, & (x,y) \in (a_j,b_j) \times (0,\hat{L}), \\
{\cal B}_l u_j^{n}(a_j,y) &{}= {\cal B}_l u_{j-1}^{n-1}(a_j,y), & y \in (0,\hat L), \\
{\cal B}_r u_j^{n}(b_j,y) &{}= {\cal B}_r u_{j+1}^{n-1}(b_j,y), & y \in (0,\hat L), \\
u_j^n(x,y) &{}= 0, & x \in (a_j,b_j), \ y \in \lbrace0,\hat L\rbrace,
\end{array} \right. \end{aligned}$$ where the boundary operators ${\cal B}_l$ and ${\cal B}_r$ are as defined in . We consider here the case of impedance conditions, i.e. $\alpha = ik$. Note that this configuration corresponds to a [*“two-dimensional wave-guide”*]{} problem. By linearity, it follows that the local errors $e_j^{n} = u|_{\Omega_j} - u_j^n$ satisfy the homogeneous analogue of . To proceed, we make use of the Fourier sine expansion of $e_j^{n}$, as the solution verifies Dirichlet boundary conditions on the top and bottom of each rectangular subdomain: $$\begin{aligned}
\label{eq:Fourier}
e_j^{n}(x,y) &= \sum_{m=1}^{\infty} v_j^{n}(x,\tilde{k}) \sin(\tilde{k}y), & \tilde{k} &= \frac{m\pi}{\hat{L}}, \ m \in \mathbb{N}.\end{aligned}$$ Inserting this expression into the homogeneous counterpart of we find that, for each Fourier number $\tilde k$, $v_j^{n}(x, \tilde k)$ verifies the one-dimensional problem $$\begin{aligned}
\label{eq:2dv}
\left\{ \begin{array}{r@{}ll}
(ik\sigma+\tilde k^2-k^2)v_j^{n} - \partial_{xx}v_j^{n} &{}= 0, & x \in (a_j,b_j), \\
{\cal B}_l v_j^{n}(x,\tilde k) &{}= {\cal B}_l v_{j-1}^{n-1}(x,\tilde k), & x = a_j,\\
{\cal B}_r v_j^{n}(x,\tilde k) &{}= {\cal B}_r v_{j+1}^{n-1}(x,\tilde k), & x = b_j,
\end{array} \right. \end{aligned}$$ which is of exactly the same type as where $ik\sigma-k^2$ is replaced by $ik\sigma+\tilde k^2-k^2$. Therefore, the result from applies here if we replace $\alpha$ with $ik$ and $\zeta$ with $$\begin{aligned}
\label{eq:lamk}
\zeta(\tilde k) = \sqrt{ik\sigma+\tilde k^2-k^2}.\end{aligned}$$ Let us denote the resulting iteration matrix, which propagates information for each Fourier number $\tilde k$ independently, by ${\cal T}_{1d}^{\mathrm{H}}(\tilde k)$ and let $R_{1d}^{\mathrm{H}}(\tilde k) \vcentcolon= \lim_{N\rightarrow\infty} \rho({\cal T}_{1d}^{\mathrm{H}}(\tilde k))$ with $R_{2d}^{\mathrm{H}} = \sup_{\tilde k} R_{1d}^{\mathrm{H}}(\tilde k)$. We can now state our main convergence result for the two-dimensional Helmholtz problem.
\[theorem:2dhelmholtz\] If $\alpha = ik$ (the case of classical impedance conditions), then for all $k>0$, $\sigma>0$, $\delta>0$ and $L>0$ we have that $R_{1d}^{\mathrm{H}}(\tilde k) < 1$ for all evanescent modes $\tilde k > k$. Furthermore, under the assumption that between them $\sigma$, $\delta$ and $L$ are sufficiently large we have that $R_{2d}^{\mathrm{H}} < 1$. In particular, this is true when $\sigma \ge k$ for all $\delta>0$ and $L>0$. Therefore the convergence will ultimately be independent of the number of subdomains (we say that the Schwarz method will scale).
By we see that it is enough to study the sign of $g_{\pm}(z;\delta,l)$ and $g(z;\delta,l)$. To assist, we use the scaled notation $\kappa = 2\delta k$, $\tilde \kappa = 2\delta \tilde k$ and $s = 2\delta\sigma$ akin to . Now $g_{\pm}(z;\delta,l)$ can be formally simplified identically to , however, in this case with $\zeta$ as in the real part of $z^2$ gives the identity $\tilde \kappa^2 - \kappa^2 = x^2 - y^2$. Utilising this identity along with the bounds yields $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> \left[(\kappa^2+x^2+y^2)x + 2\kappa y\right] lx - \left|(\kappa^2-x^2-y^2)y -2\kappa x\right| ly \\
&\ge l (\kappa^2+x^2+y^2) (\tilde \kappa^2-\kappa^2).\end{aligned}$$ Hence we always have $\tilde{g}_{\pm}(z;\delta,l) > 0$ for the evanescent modes $\tilde k > k$ (equivalent to $\tilde \kappa > \kappa$). Similarly, $g(z;\delta,l)$ can be simplified identically to and we find that $$\begin{aligned}
\begin{split}
\tilde{g}(z;\delta,l) &> l(l+2) \left( x^2((\kappa^2+x^2+y^2)^2+4\kappa^2y^2) - y^2|(-\kappa^2+x^2+y^2)^2-4\kappa^2x^2| \right) \\
& \mathrel{\phantom{=}} \mathrel+ 4l\kappa xy \left( \kappa^2+x^2+y^2 - |-\kappa^2+x^2+y^2| \right)
\end{split} \\
&\ge l(l+2) (\kappa^2+x^2+y^2)^2 (\tilde \kappa^2-\kappa^2),\end{aligned}$$ and so we always have $\tilde{g}(z;\delta,l) > 0$ for the evanescent modes $\tilde k > k$ too. Together this shows that $R_{1d}^{\mathrm{H}}(\tilde k) < 1$ for all evanescent modes. Note that, for the remaining modes $\tilde k \le k$, it is possible that $R_{1d}^{\mathrm{H}}(\tilde k) \ge 1$ for some choices of problem parameters.
We now refine the above bounds. In order to do so we make use of the identities $4x^2y^2 = \kappa^2s^2$ and $x^2+y^2 = \sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2}$ which arise since (by considering both real and imaginary parts of $z^2 = (x+iy)^2 = i\kappa s+\tilde \kappa^2-\kappa^2$) we have that $$\begin{aligned}
\label{eq:2x^2and2y^2}
2x^2 & = \sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2} + \tilde \kappa^2 - \kappa^2, & 2y^2 & = \sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2} - \tilde \kappa^2 + \kappa^2.\end{aligned}$$ Now, if we make use of the substitution $\kappa^2 + x^2 = \tilde \kappa^2 + y^2$ for the terms involving hyperbolic functions and the substitution $\kappa^2-y^2=\tilde \kappa^2-x^2$ for the terms involving trigonometric functions, we obtain the following: $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> \left[(\tilde \kappa^2+2y^2)x + 2\kappa y\right] lx - \left|(\tilde \kappa^2-2x^2)y -2\kappa x\right| ly \\
&\ge l \left( x^2(\tilde \kappa^2+2y^2) - y^2|\tilde \kappa^2-2x^2| \right) \\
&= \left\lbrace \begin{array}{l}
l \tilde \kappa^2 (x^2+y^2) \\
l \left( 4x^2y^2 + \tilde \kappa^2(\tilde \kappa^2-\kappa^2) \right)
\end{array} \right. &
\begin{array}{l}
\text{if } \tilde \kappa^2 \le 2x^2, \\
\text{if } \tilde \kappa^2 > 2x^2,
\end{array} \\
&= \left\lbrace \begin{array}{l}
l \tilde \kappa^2 \sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2} \\
l \left( \tilde \kappa^4 + \kappa^2(s^2-\tilde \kappa^2) \right)
\end{array} \right. &
\begin{array}{l}
\text{if } \tilde \kappa^2 \le 2x^2, \\
\text{if } \tilde \kappa^2 > 2x^2,
\end{array}\end{aligned}$$ and $$\begin{aligned}
\begin{split}
\tilde{g}(z;\delta,l) &> l(l+2) \left( x^2(\tilde \kappa^2+2y^2)^2 - y^2(\tilde \kappa^2-2x^2)^2 \right) \\
& \mathrel{\phantom{=}} \mathrel+ 4l\kappa xy \left( \tilde \kappa^2+2y^2 - |\tilde \kappa^2-2x^2| \right)
\end{split} \\
\begin{split}
&= l(l+2) \left( \tilde \kappa^4 (x^2-y^2) + 4x^2y^2(2\tilde \kappa^2+y^2-x^2)\right) \\
& \mathrel{\phantom{=}} \mathrel+ 4l\kappa xy \left( \tilde \kappa^2+2y^2 - |\tilde \kappa^2-2x^2| \right)
\end{split} \\
&= \left\lbrace \begin{array}{l}
l(l+2) \left( \tilde \kappa^4 (\tilde \kappa^2-\kappa^2) + 4x^2y^2(\tilde \kappa^2+\kappa^2)\right) \\
\quad \mathrel+ 8l\kappa^3xy \\
l(l+2) \left( \tilde \kappa^4 (\tilde \kappa^2-\kappa^2) + 4x^2y^2(\tilde \kappa^2+\kappa^2)\right) \\
\quad \mathrel+ 8l\kappa xy(x^2+y^2)
\end{array} \right. &
\begin{array}{l}
\text{if } \tilde \kappa^2 \le 2x^2, \\ \\
\text{if } \tilde \kappa^2 > 2x^2,
\end{array} \\
&= \left\lbrace \begin{array}{l}
l(l+2) \left( \tilde \kappa^6 + \kappa^4s^2 + \tilde \kappa^2\kappa^2(s^2-\tilde \kappa^2)\right) \\
\quad \mathrel+ 4l\kappa^4s \\
l(l+2) \left( \tilde \kappa^6 + \kappa^4s^2 + \tilde \kappa^2\kappa^2(s^2-\tilde \kappa^2)\right) \\
\quad \mathrel+ 4l\kappa^2s\sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2}
\end{array} \right. &
\begin{array}{l}
\text{if } \tilde \kappa^2 \le 2x^2, \\ \\
\text{if } \tilde \kappa^2 > 2x^2.
\end{array}\end{aligned}$$ From the penultimate expression in each case we see that for evanescent modes $\tilde k > k$ (i.e. $\tilde \kappa > \kappa$) we always have $\tilde{g}_{\pm}(z;\delta,l) > 0$ and $\tilde{g}(z;\delta,l) > 0$. Furthermore, from the final expressions we see that all modes $\tilde k \le \sigma$ (i.e. $\tilde \kappa \le s$) also give the desired positivity. Thus we deduce that when $\sigma \ge k$ we have positivity for all modes $\tilde k$ and hence $R_{2d}^{\mathrm{H}} < 1$. We also remark that modes $\tilde k \le k$ which are relatively close to $k$ are identified as those giving the worst bounds, suggesting these are the most problematic modes for the algorithm.
If $\sigma < k$ we may still have positivity of $\tilde{g}_{\pm}(z;\delta,l)$ and $\tilde{g}(z;\delta,l)$ for all modes so long as $x$ or $lx$ is large enough so that the hyperbolic term, which is always positive, is larger than the magnitude of the trigonometric term in both and . Using and converting back to the original variables we have that $$\begin{aligned}
\label{eq:x}
x &= 2\delta \sqrt{\tfrac{1}{2}\left(\sqrt{(k^2-\tilde k^2)^2+\sigma^2k^2} + \tilde k^2 - k^2\right)},\end{aligned}$$ while $lx$ has an identical expression except with $2\delta$ replaced by $L$. Thus we see that, between the parameters $\sigma$, $\delta$ and $L$, so long as they are sufficiently large we will have $\tilde{g}_{\pm}(z;\delta,l) > 0$ and $\tilde{g}(z;\delta,l) > 0$ for all modes $\tilde k$ and thus $R_{2d}^{\mathrm{H}} < 1$ as desired.
In order to verify this result, we compare numerically the spectral radius of the iteration matrix with the theoretical limit for different values of $\sigma$. We choose here $k = 30$, $L=1$, $\hat{L}=1$ and $\delta = L/10$. From we see that, as predicted, the Schwarz algorithm is not convergent for all Fourier modes when $\sigma$ is small, but becomes convergent for $\sigma$ sufficiently large. We also see that the algorithm converges well for the evanescent modes ($\tilde{k} > k$), as expected from our theory.
The transverse electric Maxwell’s equations {#sec:2dmaxwell}
-------------------------------------------
We now apply the same ideas to the transverse electric Maxwell’s equations with damping in the frequency domain. For an electric field $\mathbf{E}=(E_x,E_y)$, these equations are expressed as $$\begin{aligned}
\label{eq:2d0}
\begin{split}
& {\cal L}\mathbf{E} \vcentcolon= -k^2\mathbf{E}+ \nabla \times (\nabla \times \mathbf{E}) +ik \sigma \mathbf{E} = \mathbf{0} \\
& \Leftrightarrow \left\{\begin{array}{r@{}l}
- k^2 E_x - \partial_{yy} E_x + \partial_{xy} E_y + ik\sigma E_x &{}= 0, \\
- k^2 E_y - \partial_{xx} E_y + \partial_{xy} E_x + ik\sigma E_y &{}= 0,
\end{array}\right.
\end{split}\end{aligned}$$ for $(x,y) \in \Omega$. The boundary conditions on the top and bottom boundaries ($y=0$ and $y=\hat L$) are perfect electric conductor (PEC) conditions, the equivalent of Dirichlet conditions for Maxwell’s equations: $$\begin{aligned}
\label{eq:2dPEC}
\mathbf{E} \times \mathbf{n} = \mathbf{0} \Leftrightarrow E_x &= 0, & y &= \{0, \hat L\}.\end{aligned}$$ On the left and right boundaries ($x=a_1$ and $x=b_N$) we use impedance boundary conditions[^5]: $$\begin{aligned}
\label{eq:2dimp}
\begin{split}
& (\nabla \times \mathbf{E} \times \mathbf{n}) \times \mathbf{n} + ik \mathbf{E} \times \mathbf{n} = \mathbf{g} \\
& \Leftrightarrow \left\{\begin{array}{r@{}ll}
{\cal B}_l\mathbf{E} \vcentcolon= (-\partial_x + ik) E_y + \partial_y E_x &{}= g_1, & x = a_1,\\
{\cal B}_r\mathbf{E} \vcentcolon= (\partial_x + ik) E_y - \partial_y E_x &{}= - g_2, & x = b_N.
\end{array}\right.
\end{split}\end{aligned}$$ The same conditions will be used at the interfaces between subdomains, akin to the classical algorithm defined in [@Despres:1992:DDM]. The Maxwell problem – constitutes a [*“two-dimensional wave-guide”*]{} model.
Let us denote by $\mathbf{E}_j^n$ the approximation to the solution in subdomain $j$ at iteration $n$. Starting from an initial guess $\mathbf{E}_j^0$, we compute $\mathbf{E}_j^n$ from the previous values $\mathbf{E}_j^{n-1}$ by solving the following local boundary value problems $$\begin{aligned}
\label{eq:sch2d}
\left\{\begin{array}{r@{}ll}
{\cal L} \mathbf{E}_j^n &{}= \mathbf{0}, & x \in \Omega_j,\\
{\cal B}_l\mathbf{E}_j^n &{}= {\cal B}_l \mathbf{E}_{j-1}^{n-1}, & x = a_j,\\
{\cal B}_r\mathbf{E}_j^n &{}= {\cal B}_r \mathbf{E}_{j+1}^{n-1}, & x = b_j,\\
E_{x,j}^n &{}= 0, & y \in \{0,\hat L\},
\end{array}\right.\end{aligned}$$ for the interior subdomains ($1 < j < N$), while for the first ($j=1$) and last ($j=N$) subdomain we impose ${\cal B}_l \mathbf{E}_1^n {}= g_1$ when $x = a_1$ and ${\cal B}_r \mathbf{E}_N^n {}= -g_2$ when $x = b_N$. To study the convergence of the Schwarz algorithm we define the local error in each subdomain $j$ at iteration $n$ as $\mathbf{e}^n_j = \mathbf{E}|_{\Omega_j} - \mathbf{E}^n_j$. Note that these errors verify boundary value problems which are the homogeneous counterparts of .
Due to the PEC boundary conditions on the top and bottom boundaries of each rectangular subdomain we can use the following Fourier series ansatzes to compute the local solutions of ${\cal L}\mathbf{e}^n_j = 0$: $$\begin{aligned}
\label{eq:Fourier2}
e_{x,j}^n &= \sum_{m=1}^{\infty} v^n_j(x,\tilde k) \sin(\tilde k y), & e_{y,j}^n &= \sum_{m=1}^{\infty} w^n_j(x,\tilde k) \cos(\tilde k y), & \tilde k &= \frac{m\pi}{\hat L}, \ m \in \mathbb{N}.\end{aligned}$$ By plugging the expressions for $e_{x,j}^n$ and $e_{y,j}^n$ into ${\cal L}\mathbf{e}^n_j = \mathbf{0}$, a simple computation shows that, for each Fourier number $\tilde k$, we have the general solutions $$\begin{aligned}
\label{eq:vjwj}
v^n_j(x,\tilde k) &= -\alpha^n_j \frac{\tilde k}{{\zeta}} e^{-\zeta x} + \beta^n_j \frac{\tilde k}{{\zeta}} e^{\zeta x}, & w^n_j(x,\tilde k) &= \alpha^n_j e^{-\zeta x} + \beta^n_j e^{\zeta x},\end{aligned}$$ where $\zeta(\tilde k) = \sqrt{ik\sigma+\tilde k^2-k^2}$. From these formulae we can see easily that $$\begin{aligned}
\label{eq:vwrel}
\partial_x v^n_j &= \tilde k w^n_j, & \partial_x w^n_j &= \frac{\zeta^2}{\tilde k}v^n_j. \end{aligned}$$ In order to benefit again from the analysis in the one-dimensional case, we first prove the following result.
For each Fourier number $\tilde k$, we have that both $v^n_j(x,\tilde k)$ and $w^n_j(x,\tilde k)$ are solutions of the following one-dimensional problem: $$\begin{aligned}
\label{eq:2dmax}
\left\{ \begin{array}{r@{}ll}
(ik\sigma+\tilde k^2 -k^2) u_j^{n} - \partial_{xx}u_j^{n} &{}= 0, & x \in (a_j,b_j), \\
{\cal B}_{l,\sigma} u_j^{n}(x,\tilde k) &{}= {\cal B}_{l,\sigma} u_{j-1}^{n-1}(x,\tilde k), & x = a_j, \\
{\cal B}_{r,\sigma} u_j^{n}(x,\tilde k) &{}= {\cal B}_{r,\sigma} u_{j+1}^{n-1}(x,\tilde k), & x = b_j,
\end{array} \right.\end{aligned}$$ where ${\cal B}_{l,\sigma} = -\partial_x + ik + \sigma$ and ${\cal B}_{r,\sigma} = \partial_x + ik + \sigma$.
Let us notice first that, because of , we have $$\begin{aligned}
\partial_x e_{x,j}^n+ \partial_y e_{y,j}^n = \sum_{m=1}^{\infty} \left(\partial_x v_{j}^n - \tilde k w_{j}^n\right)\sin(\tilde k y) = 0.\end{aligned}$$ If we use this in the error equation ${\cal L} \mathbf{e}^n_j = \mathbf{0}$ we obtain that both $v^n_j(x,\tilde k)$ and $w^n_j(x,\tilde k)$ satisfy, for each $\tilde k$, the one-dimensional equation $(ik\sigma+\tilde k^2-k^2) u_j^{n} - \partial_{xx} u_j^{n} = 0$. Let us analyse now the boundary conditions. With the help of , we consider the right boundary and note that the left one can be treated similarly: $$\begin{aligned}
{\cal B}_r \mathbf{e}^n_j &= (\partial_x + ik) e_{y,j}^n -\partial_y e_{x,j}^n = \sum_{m=1}^{\infty} ( (\partial_x +ik)w_{j}^n -\tilde k v_{j}^n) \cos(\tilde k y) \\
&= \sum_{m=1}^{\infty} \left ( \frac{ik}{\tilde k} \partial_x v_{j}^n + \left(\frac{\zeta^2}{\tilde k} -\tilde k\right ) v_{j}^n\right) \cos(\tilde k y) = \sum_{m=1}^{\infty} \frac{ik}{\tilde k} {\cal B}_{r,\sigma} v_{j}^n \cos(\tilde k y).\end{aligned}$$ Thus, imposing transfer of boundary data with ${\cal B}_r \mathbf{e}^n_j$ is equivalent to that with ${\cal B}_{r,\sigma} v^n_j$, for each Fourier number $\tilde k$.
It is now clear that the analysis of the two-dimensional case can again be derived from the one-dimensional case. That is, the result from applies here if we replace $\alpha$ with $ik+\sigma$ and with $\zeta$ being defined by . Let us denote the resulting iteration matrix, for each $\tilde k$, by ${\cal T}_{1d}^{\mathrm{M}}(\tilde k)$ and let $R_{1d}^{\mathrm{M}}(\tilde k) \vcentcolon= \lim_{N\rightarrow\infty} \rho({\cal T}_{1d}^{\mathrm{M}}(\tilde k))$ with $R_{2d}^{\mathrm{M}} = \sup_{\tilde k} R_{1d}^{\mathrm{M}}(\tilde k)$. We can now state our main convergence result for the two-dimensional Maxwell problem.
\[theorem:2dmaxwell\] For all $k>0$, $\sigma>0$, $\delta>0$ and $L>0$ we have that $R_{1d}^{\mathrm{M}}(\tilde k) < 1$ for all evanescent modes $\tilde k > k$. Furthermore, under the assumption that between them $\sigma$, $\delta$ and $L$ are sufficiently large we have that $R_{2d}^{\mathrm{M}} < 1$. In particular, this is true when $\sigma \ge k$ for all $\delta>0$ and $L>0$. Therefore the convergence will ultimately be independent of the number of subdomains (we say that the Schwarz method will scale).
By we see that it is enough to study the sign of $g_{\pm}(z;\delta,l)$ and $g(z;\delta,l)$. To assist, we use the scaled notation $\kappa = 2\delta k$, $\tilde \kappa = 2\delta \tilde k$ and $s = 2\delta\sigma$ akin to . We can see that if $\alpha = ik +\sigma$ then for $z\vcentcolon= x+iy$ becomes $$\begin{aligned}
\Re v &= \frac{-\kappa^2 - s^2 + x^2 + y^2}{(\kappa+y)^2 + (s+x)^2}, &
\Im v &= \frac{2sy - 2\kappa x}{(\kappa+y)^2 + (s+x)^2}, &
|v|^2 &= \frac{(\kappa-y)^2 + (s-x)^2}{(\kappa+y)^2 + (s+x)^2},\end{aligned}$$ where $|v|^2 < 1$. We can now simplify $g_{\pm}(z;\delta,l)$ in using these formulae to give
\[eq:gpm2d\] $$\begin{aligned}
\label{eq:gpm2dg}
g_{\pm}(z;\delta,l) &= \frac{4e^{x(l+1)}}{(\kappa+y)^2 + (s+x)^2} \tilde{g}_{\pm}(z;\delta,l) \\
\begin{split}
\label{eq:gpm2dgt}
\tilde{g}_{\pm}(z;\delta,l) &= [(\kappa^2+s^2+x^2+y^2) \sinh(x) + 2(\kappa y+sx) \cosh(x)]\sinh(lx) \\
& \mathrel{\phantom{=}} \mathrel\pm [(\kappa^2+s^2-x^2-y^2) \sin(y) + 2(sy-\kappa x) \cos(y)]\sin(ly).
\end{split}\end{aligned}$$
Proceeding as before, using $\tilde \kappa^2 - \kappa^2 = x^2 - y^2$ and the bounds , we derive that $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> l (\kappa^2+s^2+2s+x^2+y^2) (\tilde \kappa^2-\kappa^2)\end{aligned}$$ which is positive for all evanescent modes $\tilde k > k$. Similarly, simplifying $g(z;\delta,l)$ in we find that
\[eq:g2d\] $$\begin{aligned}
\label{eq:g2dg}
g(z;\delta,l) &= \frac{4e^{2x(l+1)}}{((\kappa+y)^2 + (s+x)^2)^2} \tilde{g}(z;\delta,l) \\
\begin{split}
\label{eq:g2dgt}
\tilde{g}(z;\delta,l) &= \bigl[ ((\kappa^2+s^2+x^2+y^2)^2+4(\kappa y+sx)^2) \sinh(x(l+2)) \\
& \qquad\qquad \mathrel+ 4(\kappa y+sx)(\kappa^2+s^2+x^2+y^2) \cosh(x(l+2)) \bigr] \sinh(lx) \\
& \mathrel{\phantom{=}} \mathrel+ \bigl[ ((-\kappa^2-s^2+x^2+y^2)^2-4(\kappa x-sy)^2) \sin(y(l+2)) \\
& \qquad\qquad \mathrel+ 4(\kappa x-sy)(-\kappa^2-s^2+x^2+y^2) \cos(y(l+2)) \bigr] \sin(ly),
\end{split}\end{aligned}$$
from which we can obtain the bound $$\begin{aligned}
\begin{split}
\tilde{g}(z;\delta,l) &> l(l+2) \left( (\kappa^2+s^2+x^2+y^2)^2 + 4s^2(x^2+y^2) + 8\kappa sxy \right) (\tilde \kappa^2-\kappa^2) \\
& \mathrel{\phantom{=}} \mathrel+ 4ls \left( \kappa^2+s^2+x^2+y^2 \right) (\tilde \kappa^2-\kappa^2).
\end{split}\end{aligned}$$ Again, this is positive for all evanescent modes and thus we deduce that $R_{1d}^{\mathrm{M}}(\tilde k) < 1$ for all $\tilde k > k$.
We now refine these bounds, as in the proof of and using the same identities and substitutions. For $g_{\pm}(z;\delta,l)$ we first obtain $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> l \left( x^2(s^2+\tilde \kappa^2+2y^2) - y^2\left|s^2+\tilde \kappa^2-2x^2\right| + 2x(\kappa y+sx) - 2y\left|\kappa x-sy\right| \right),\end{aligned}$$ and split into four cases based on the sign of each term we take the absolute value of. Consider first the case $s^2 + \tilde \kappa^2 \le 2x^2$ and $\kappa x \le sy$, then $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> l \left( (s^2+\tilde \kappa^2)(x^2+y^2) + 4\kappa xy + 2s(x^2-y^2) \right) \\
&= l \left( (s^2+\tilde \kappa^2)\sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2} + 2\tilde \kappa^2s \right).\end{aligned}$$ Now consider the case $s^2 + \tilde \kappa^2 > 2x^2$ and $\kappa x > sy$ where we find that $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> l \left( 4x^2y^2 + (s^2+\tilde \kappa^2)(x^2-y^2) + 2s(x^2+y^2) \right) \\
&= l \left( \tilde \kappa^2(s^2+\tilde \kappa^2-\kappa^2) + 2s\sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2} \right).\end{aligned}$$ The remaining cases follow as combinations of the previous two cases and we deduce, in the case $s^2 + \tilde \kappa^2 \le 2x^2$ and $\kappa x > sy$, that $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> l (s^2+\tilde \kappa^2+2s)\sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2},\end{aligned}$$ while the case $s^2 + \tilde \kappa^2 > 2x^2$ and $\kappa x \le sy$ gives $$\begin{aligned}
\tilde{g}_{\pm}(z;\delta,l) &> l \tilde \kappa^2(s^2+\tilde \kappa^2-\kappa^2+2s).\end{aligned}$$ Turning to $\tilde{g}(z;\delta,l)$, we first derive that $$\begin{aligned}
\begin{split}
\tilde{g}(z;\delta,l) &> l(l+2) \bigl[ x^2((s^2+\tilde \kappa^2+2y^2)^2+4(\kappa y+sx)^2) \\
& \mathrel{\phantom{=}} \mathrel- y^2((s^2+\tilde \kappa^2-2x^2)^2+4(\kappa x-sy)^2) \bigr] \\
& \mathrel{\phantom{=}} \mathrel+ 4l \left( x(\kappa y+sx)(s^2+\tilde \kappa^2+2y^2) - y\left| (\kappa x-sy)(s^2+\tilde \kappa^2-2x^2) \right| \right),
\end{split}\end{aligned}$$ from which we see that we need to analyse just two sets of combined cases. First consider when both $s^2 + \tilde \kappa^2 \le 2x^2$ and $\kappa x \le sy$ or both $s^2 + \tilde \kappa^2 > 2x^2$ and $\kappa x > sy$, yielding $$\begin{aligned}
\begin{split}
\tilde{g}(z;\delta,l) &> l(l+2) \bigl[ (s^2+\tilde \kappa^2)^2(x^2-y^2) + 4x^2y^2(2s^2+2\tilde \kappa^2+y^2-x^2) \\
& \mathrel{\phantom{=}} \mathrel+ 4s(x^2+y^2)(2\kappa xy + s(x^2-y^2)) \bigr] + 4l (x^2+y^2)(2\kappa xy + s(s^2+\tilde \kappa^2))
\end{split} \\
\begin{split}
&= l(l+2) \Bigl[ \kappa^2s^2(s^2+\kappa^2) + \tilde \kappa^2(s^2+\tilde \kappa^2)(s^2+\tilde \kappa^2-\kappa^2) \\
& \mathrel{\phantom{=}} \mathrel+ 4\tilde \kappa^2s^2\sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2} \Bigr] + 4ls (s^2+\tilde \kappa^2+\kappa^2)\sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2}.
\end{split}\end{aligned}$$ On the other hand, in the second set of cases when both $s^2 + \tilde \kappa^2 \le 2x^2$ and $\kappa x > sy$ or both $s^2 + \tilde \kappa^2 > 2x^2$ and $\kappa x \le sy$ we have $$\begin{aligned}
\begin{split}
\tilde{g}(z;\delta,l) &> l(l+2) \bigl[ (s^2+\tilde \kappa^2)^2(x^2-y^2) + 4x^2y^2(2s^2+2\tilde \kappa^2+y^2-x^2) \\
& \mathrel{\phantom{=}} \mathrel+ 4s(x^2+y^2)(2\kappa xy + s(x^2-y^2)) \bigr] \\
& \mathrel{\phantom{=}} \mathrel+ 4l \left( 2\kappa xy(s^2+\tilde \kappa^2+y^2-x^2) + s(4x^2y^2+(s^2+\tilde \kappa^2)(x^2-y^2)) \right)
\end{split} \\
\begin{split}
&= l(l+2) \Bigl[ \kappa^2s^2(s^2+\kappa^2) + \tilde \kappa^2(s^2+\tilde \kappa^2)(s^2+\tilde \kappa^2-\kappa^2) \\
& \mathrel{\phantom{=}} \mathrel+ 4\tilde \kappa^2s^2\sqrt{(\tilde \kappa^2-\kappa^2)^2 + \kappa^2 s^2} \Bigr] + 4ls \left( \kappa^2 (s^2+\kappa^2) + \tilde \kappa^2(s^2+\tilde \kappa^2-\kappa^2) \right).
\end{split}\end{aligned}$$ Summarising, we see that all cases give $\tilde{g}_{\pm}(z;\delta,l) > 0$ and $\tilde{g}(z;\delta,l) > 0$ for all modes $\tilde k$ satisfying $\tilde k^2 \ge k^2 - \sigma^2$ (i.e. $\tilde \kappa^2 \ge \kappa^2 - s^2$). From this we can deduce that when $\sigma \ge k$ we have positivity for all modes $\tilde k$ and hence $R_{2d}^{\mathrm{M}} < 1$. Note that $\sigma \ge k$ is far from a necessary requirement and it is clear that there is some slack in these bounds. We also remark from this analysis that modes $\tilde k \le \sqrt{k^2-\sigma^2}$ which are relatively close to $\sqrt{k^2-\sigma^2}$ yield the poorest bounds, suggesting they are the most problematic for the algorithm. Indeed, we may have $R_{1d}^{\mathrm{M}}(\tilde k) \ge 1$ when $\tilde k \le \sqrt{k^2-\sigma^2}$ for some choices of problem parameters. However, as in we can force positivity of $\tilde{g}_{\pm}(z;\delta,l)$ and $\tilde{g}(z;\delta,l)$ for all modes so long as $x$ or $lx$ is large enough. Since $x$ and $lx$ take the same expressions as in we can similarly deduce that, so long as the parameters $\sigma$, $\delta$ and $L$ between them are sufficiently large, we will have $\tilde{g}_{\pm}(z;\delta,l) > 0$ and $\tilde{g}(z;\delta,l) > 0$ for all modes $\tilde k$ and thus the required conclusion that $R_{2d}^{\mathrm{M}} < 1$.
Numerical simulations on the discretised equation {#sec:num}
=================================================
Although extensive numerical results are beyond the scope of this paper, in the following we will show some simulations which confirm our theory within the more practical setting of using an iterative Krylov method to accelerate convergence, with the Schwarz method being used as a preconditioner. We focus here on the two-dimensional Helmholtz equation, as described in , where a (horizontal) plane wave is incoming from the left boundary and homogeneous Dirichlet boundary conditions are imposed on the top and bottom boundaries, giving a wave-guide problem. A second test case we consider is the propagation of such a wave in free space (i.e. when impedance boundary conditions are imposed on the whole boundary). While not covered by our theory, we will nonetheless observe similar conclusions, illustrating that the results apply more widely than within the restrictions of our theoretical assumptions. In our simulations, each subdomain is a unit square split uniformly with a fixed number of grid points in each direction. New subdomains are added on the right so that, with $N$ subdomains, the whole domain is $\Omega = (0,N)\times (0,1)$.
To discretise we use a uniform square grid in each direction and triangulate with alternating diagonals to form P1 elements. As we increase $k$ we increase the number of grid points proportional to $k^{3/2}$ in order to ameliorate the pollution effect [@Babuska:1997:IPE]. We use an overlap of size $2h$, with $h$ being the mesh size. All computations are performed using FreeFem (<http://freefem.org/>), in particular using the `ffddm` framework. We solve the discretised problem using GMRES where the Schwarz method with Robin conditions is used as a preconditioner. In particular, we use right-preconditioned GMRES and terminate when a relative residual tolerance of $10^{-6}$ is reached. The construction of the domain decomposition preconditioner is described in detail in [@Bonazzoli:2019:ADD; @Dolean:2020:IFD]. The one-level ORAS preconditioner is $\mathbf{M}^{-1} = \sum_{i=1}^{N} \mathbf{R}_i^T \mathbf{D}_i \mathbf{A}_i^{-1} \mathbf{R}_i$ where $\left\{\mathbf{R}_i\right\}_{1 \le i \le N}$ are the Boolean restriction matrices from the global to the local finite element spaces and $\left\{\mathbf{D}_i\right\}_{1 \le i \le N}$ are local diagonal matrices representing the partition of unity. The key ingredient of the ORAS method is that the local matrices $\left\{\mathbf{A}_i\right\}_{1 \le i \le N}$ incorporate more efficient Robin transmission conditions.
Note that, unlike in [@Graham:2018:DDI] where the emphasis is placed on the independence of the one-level method to the wave number, we focus here on the scalability aspect, i.e. the independence of the one-level method with respect to the number of subdomains $N$ as soon as the absorption parameter $k\sigma$ is positive. We will observe that, beyond a sufficiently large value of $N$, the iteration count does not increase further, though in general this value will depend on the parameters of the problem, namely the wave number and absorption as well as the overlap and subdomain size. As a side effect, when the absorption is sufficiently large, i.e. of order $k$, wave number independence is also achieved.
In we detail the GMRES iteration count for an increasing number of subdomains $N$ and different values of $k$ for the wave-guide problem and the wave propagation in free space problem. We set the conductivity parameter as $\sigma = 1$ (giving an absorption parameter $k$). We see that, after an initial increase, the iteration counts become independent of the number of subdomains and also independent of the wave number, which is consistent with the results obtained in [@Graham:2018:DDI] where the absorption parameter for optimal convergence is of order $k$. Another possible explanation of this is that when the absorption parameter increases, the waves are damped and their amplitude will decrease with the distance to the boundary on which the excitation is imposed. Hence, when additional subdomains are added, the solution will not vary much in these subdomains.
--------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
$k$/$N$ 8 16 24 32 40 48 56 64 8 16 24 32 40 48 56 64
20 19 22 25 30 30 30 30 30 19 21 25 25 25 25 25 25
40 18 21 24 29 29 29 29 29 17 19 24 25 25 25 25 25
60 19 21 24 29 29 29 29 29 16 19 24 25 25 25 25 25
80 19 21 24 28 28 28 28 28 16 18 24 25 25 25 25 25
100 19 21 24 28 28 28 28 28 16 18 24 25 25 25 25 24
--------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
: Preconditioned GMRES iteration counts for varying wave number $k$ and number of subdomains $N$ when $\sigma = 1$.[]{data-label="Table:Itcount1"}
Conclusions {#sec:conc}
===========
In this work we have analysed a purely iterative version of the Schwarz domain decomposition algorithm, in the limiting case of many subdomains, at the continuous level for the one-dimensional and two-dimensional Helmholtz and Maxwell’s equations with absorption. The key mathematical tool which facilitated this study is the limiting spectrum of a sequence of block Toeplitz matrices having a particular structure, for which we proved a new result in the non-Hermitian case. The algorithm is convergent in the one-dimensional case as soon as we have absorption and, for sufficiently many subdomains $N$, its convergence factor becomes independent of the number of subdomains, meaning the algorithm is also scalable. In practice, this is achieved for relatively small $N$. In the two-dimensional case these conclusions remain true for the evanescent modes of the error (i.e. $\tilde k > k$) or when, between them, $\sigma$, $\delta$ and $L$ are sufficiently large. In particular, we proved that the stationary iteration will always converge when $\sigma \ge k$, giving an absorption parameter $k^2$. The concept of the limiting spectrum proved to be a very elegant mathematical tool and can be used, for example, in constructing more sophisticated transmission conditions, to analyse the algorithm at the discrete level, or to design improved preconditioners.
[^1]: Department of Mathematics and Statistics, University of Strathclyde, Glasgow, UK (, , ).
[^2]: Department of Mathematics and Statistics, University of Strathclyde, Glasgow, UK and Laboratoire J.A. Dieudonné, CNRS, University Côte d’Azur, Nice, France ().
[^3]: The first two authors gratefully acknowledge support from the EPSRC grant EP/S004017/1. The fourth author gratefully acknowledges support from the EPSRC grant EP/R009821/1.
[^4]: When $\sigma=0$, impedance transmission conditions are also transparent conditions, with the resulting iteration matrix being nilpotent. Therefore, the algorithm will converge in a number of iterations equal to the number of subdomains in this case.
[^5]: Note that in rewriting the impedance conditions we can use the three-dimensional definition of the operators, i.e. $\mathbf{E}=(E_x,E_y,0)$ and $\mathbf{n}=(1,0,0)$ for the right boundary and $\mathbf{n}=(-1,0,0)$ for the left boundary.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The map-matching is an essential preprocessing step for most of the trajectory-based applications. Although it has been an active topic for more than two decades and, driven by the emerging applications, is still under development. There is a lack of categorisation of existing solutions recently and analysis for future research directions. In this paper, we review the current status of the map-matching problem and survey the existing algorithms. We propose a new categorisation of the solutions according to their map-matching models and working scenarios. In addition, we experimentally compare three representative methods from different categories to reveal how matching model affects the performance. Besides, the experiments are conducted on multiple real datasets with different settings to demonstrate the influence of other factors in map-matching problem, like the trajectory quality, data compression and matching latency.'
author:
- Pingfu Chao
- Yehong Xu
- Wen Hua
- Xiaofang Zhou
bibliography:
- 'reference.bib'
title: 'A Survey on Map-Matching Algorithms'
---
Introduction {#sec:intro}
============
Nowadays, the ubiquity of positioning devices enables the tracking of user/vehicle trajectories. However, due to the intrinsic inaccuracy of the positioning systems, a series of preprocessing steps are required to correct the trajectory errors. As one of the major preprocessing techniques, the map-matching algorithm finds the object’s travel route by aligning its positioning data to the underlying road network. It is the prerequisite of various location-based applications, such as navigation, vehicle tracking, map update and traffic surveillance.
The map-matching problem has been studied for more than two decades. Despite hundreds of papers are proposed, to the best of our knowledge, only several works were conducted[@quddus2007current; @hashemi2014critical; @wei2013mapc; @kubicka2018comparative] surveying them. More importantly, even the most recent surveys[@kubicka2018comparative] fail to categorise the existing methods comprehensively. They either classify them based on applications[@kubicka2018comparative] that are not very distinctive to each other, or follow the previous categorisation[@quddus2007current] that is obsolete. Besides, various new techniques are introduced to the map-matching problem recently, including new models (weight-based[@sharath2019dynamic], multiple hypothesis theory[@taguchi2018online]), new tuning techniques (machine learning[@osogami2013map], information fusion[@li2013high; @hu2017if]), new data types (DGPS, inertial sensor, semantic road network) and new research topics (lane-level, parallel). Hence, it is about time to conduct a new survey to summarise existing solutions and provide guidance to future research.
Note that the existing map-matching problem covers various scenarios, ranging from indoor to outdoor and from pedestrian, vehicle to multimodal. However, to ensure a unified setting for survey and comparison, in this paper, we target the vehicle trajectory map-matching in an outdoor environment due to its popularity. We categorise the existing work from technical perspective. In addition, we discuss the main properties of the methods and future research directions according to the experiment results conducted on multiple matching algorithms. Overall, our contributions are listed as follows:
- We review the map-matching solutions proposed since the last comprehensive survey[@quddus2007current] and propose a new categorisation of the algorithms based on their methodology. Our proposed categorisation can better distinguish the existing methods from the technical perspective, which is beneficial for future study.
- We enumerate several map-matching challenges that are caused by low-quality trajectory data. The challenges are exemplified and explained concretely, which leads to future research directions.
- To further demonstrate the challenges, we implement three representative map-matching algorithms and conduct extensive experiments on datasets with different sampling rate, map density and compression level. Our claims about the relationship between data quality and map-matching quality are fully supported by the experiments.
The rest of the paper is organised as follows: In Section \[sec:preliminaries\], we first formally define the map-matching problem and enumerate the existing surveys and their limitations. Then, we propose our new categorisation in Section \[sec:survey\]. We further discuss the current challenges which are demonstrated through experiments in Section \[sec:experiment\] and we draw conclusions in Section \[sec:conclusion\].
Preliminaries {#sec:preliminaries}
=============
Problem Definition
------------------
We first define the map-matching problem and relevant datasets, including trajectory (input), road network (input) and route (output):
(Trajectory)\[def:trajectory\] A **trajectory** $Tr$ is a sequence of chronologically ordered spatial points $Tr:p_{1} \rightarrow p_{2} \rightarrow ... \rightarrow p_{n}$ sampled from a continuously moving object. Each point $p_{i}$ consists of a 2-dimensional coordinate $<x_{i},y_{i}>$, a timestamp $t_{i}$, a speed $spd_{i}$ (optional) and a heading $\theta_{i}$ (optional). i.e.: $p_{i} = <x_{i},y_{i},t_{i},spd_{i},\theta_{i}>$.
(Road Network) A **road network** (also known as map) is a directed graph $G = (V,E)$, in which a vertex $v = (x,y) \in V$ represents an intersection or a road end, and an edge $e = (s,e,l)$ is a directed road starting from vertices $s$ to $e$ with a polyline $l$ represented by a sequence of spatial points.
(Route) A route $R$ represents a sequence of connected edges, i.e. $R: e_{1} \rightarrow e_{2} \rightarrow ... \rightarrow e_{n}$, where $e_{i} \in G.E (1 \leq i \leq n)$ and $e_{k}.e = e_{k+1}.s$.
(Map-Matching)\[def:map-matching\] Given a road network $G(V,E)$ and a trajectory $Tr$, the map-matching find a route $\mathcal{MR}_{G}(Tr)$ that represents the sequence of roads travelled by the trajectory.
For simplicity, we omit the subscript $G$ and use $\mathcal{MR}(Tr)$ instead to represent the matching result as different trajectories are usually map-matched on the same map. In general, the map-matching route is expected to be continuous as it represents the vehicle’s travel history. However, it is quite often that $\mathcal{MR}(Tr)$ contains disconnected edges due to incorrect map-matching, which will be discussed in Section \[sec:experiment\].
Related Work
------------
Intuitively, since the vehicle usually runs on the roads, a fully accurate trajectory sampled from a vehicle should always lie on the map. Therefore, apart from some unexpected map errors, which happens less frequently and is addressed by map update process[@chao2019trajectories], the difficulty of map-matching problem solely depends on the quality of the input trajectories. As studied in many papers, the quality issues in trajectories are pervasive, which mainly caused by inaccurate measurement and low sampling rate. In terms of the *measurement error*, due to the unstable connection between GPS device and satellites, the location of GPS samples usually deviate from its actual position by a random distance. Meanwhile, the *sampling error* is mainly caused by lowering the sampling frequency.
To deal with the quality issues, the map-matching problem has been studied for more than two decades. In terms of the working scenarios and applications, the current map-matching solutions can be classified into online mode and offline mode. In online map-matching, the vehicle positions are sampled continuously and are processed in a streaming fashion, which means each time the map-matching is only performed on the current sample with a limited number of preceding or succeeding samples[@goh2012online; @yin2018feature] as reference. The process is usually simple and fast for interactive performance. In contrary, the offline map-matching is performed after the entire trajectory is obtained, so it aims for optimal matching route with less constraint on processing time.
From the methodology perspective, Quddus et al.[@quddus2007current] first conducted a comprehensive review of the map-matching algorithms proposed before 2007. The paper classified the methods into four categories, namely *geometric*, *topology*, *probabilistic* and *advanced*. The *geometric* methods only focus on the distance between trajectory elements and the road network, while the *topology* methods take into consideration the connectivity and shape similarity. The *probabilistic* methods try to model the uncertainty of trajectory, including the measurement error and the unknown travel path between two samples, and they aim to find a path that has the highest probability to generate the given trajectory. The *advanced* category contains methods that are based on some advanced models, like Kalman Filter, particle filter and fuzzy logic. This categorisation shows the evolution of map-matching research, which starts from simple, fast but inaccurate geometric-based methods to more complicated but accurate probability/advanced solutions. It is by far the most comprehensive survey of this field. However, after more than ten years’ development, most of the methods mentioned in the paper has been outperformed by their new successors and the previous categorisation also requires a revisit. Several surveys proposed afterwards reviewed the methods in certain perspectives. Hashemi et al. [@hashemi2014critical] targeted at the online map-matching scenario. Kubička et al. discussed the map-matching problem based on the applications[@kubicka2018comparative], namely *navigation*, *tracking* and *mapping*. Other categorisations also appear recently (*incremental max-weight*, *global max-weight* and *global geometry*[@wei2013mapc]) which shows that there is still no consensus on how to classify the algorithms technically. However, all of the existing categorisations inherit the same idea from Quddus’ survey[@quddus2007current] with minor variations, which fail to categorise the recent methods for multiple reasons, explained in Section \[sec:survey\].
Survey of Map-Matching Algorithm {#sec:survey}
================================
According to our study, previous categorizations fail to classify the current solutions due to three main reasons: (1) Categories for some primary methods, such as *geometric* category[@quddus2007current], are no longer the focus due to their weak performance. (2) Application-based classification[@hashemi2014critical; @kubicka2018comparative] cannot fully distinguish the methods. Many of the map-matching algorithms, like the Hidden Markov Model (HMM) and Multiple Hypothesis Technique (MHT), apply to both online and offline scenarios for different applications. (3) Classifying algorithms by embedded mathematical tools are not feasible since many recent algorithms employ multiple mathematical tools. Furthermore, the same tool implemented in different algorithms may be used for different purposes, for example, an extended Kalman filter can be used to either estimate biases in GPS or fuse measurements from different sources[@li2013high].
Therefore, we establish a new classification that classifies the map-matching algorithms by their core matching model, which is employed to coordinate their techniques to finally achieve map-matching. In a map-matching algorithm, the map-matching model is the overall framework or matching principle for the map-matching process. A model usually consists of a set of computation components, like the calculation of distance, transition and user behaviour modelling, and a workflow connecting them. Those components are fixed while their definition and implementation vary among different methods. Existing map-matching models can be categorised into four classes: *similarity model*, *state-transition model*, *candidate-evolving model* and *scoring model*.
Similarity Model
----------------
The similarity model refers to a general approach that returns the vertices/edges that is *closest* to the trajectory geometrically and/or topologically. Intuitively, since a vehicle’s movement always follows the topology of the underlying road network and the vehicle can never leap from one segment to another, the trajectory should also similar to those of the true path on the map. Therefore, the main focus in this category is how to define the *closeness*.
### Distance-based
Most of the earliest point-to-curve and curve-to-curve matching algorithms[@quddus2007current] follow this idea. Specifically, the point-to-curve solution projects each trajectory point to the geometric-closest edge, whereas the curve-to-curve matching algorithms project each trajectory segment to the closest edge where the *closeness* is defined by various similarity metrics. Fréchet distance is the most commonly-used distance function[@wei2013mapf] since it considers the monotonicity and continuity of the curves. However, it is sensitive to trajectory measurement errors since its value can be dominated by the outliers. As an alternative, Longest Common Subsequence (LCSS)[@zhu2017trajectory] divide a trajectory into multiple segments and find the shortest path on the map for each pair of start and end points of a trajectory segment. The shortest paths are then concatenated and form the final path while their corresponding LCSS scores are summed as the final score. Then, the path whose LCSS score is higher than a predefined threshold is regarded as the final matching result.
### Pattern-based
The pattern-based algorithms utilise the historical map-matched data to answer new map-matching queries by finding similar travel patterns[@zheng2012reducing]. The assumption is that people tend to travel on the same paths given a pair of origin and destination points. Therefore, by referring to the historical trajectories that are similar to the query trajectory, its candidate paths can be obtained without worrying about the sparseness of trajectory samples. Specifically, a historical trajectory or a trajectory obtained by concatenating multiple historical trajectories will be referred to if each point of this trajectory is in the safe region around the query trajectory. The algorithm finally uses a scoring function to decide the optimal route. However, due to the sparsity and disparity of historical data, the query trajectory may not be fully covered by historical trajectories especially in some rarely travelled regions, which leads to a direct matching process.
State-Transition Model
----------------------
The state-transition models build a weighted topological graph which contains all possible routes the vehicle might travel. In this graph, the vertices represent the possible *states* the vehicle may be located at a particular moment, while the edges represent the *transitions* between states at different timestamps. Different from the road network, the weight of a graph element represents the possibility of a state or a transition, and the best matching results comes from the optimal path in the graph globally. There are three major ways of building the graph and solving the optimal path problem, namely Hidden Markov model (HMM), Conditional Random Field (CRF) and the Weighted Graph Technique (WGT).
### Hidden Markov model
HMM is one of the most widely used map-matching models as it simulates the road network topology meanwhile considers the reasonability of a path. HMM focuses on the case when states in the Markov chain are unobservable (hidden) but can be estimated according to the given observations associated with them. This model fits in the map-matching process naturally. Each trajectory sample is regarded as the observation, while the vehicle actual location on the road, which is unknown, is the hidden states. In fact, due to the trajectory measurement error, all the roads near the observation can potentially be the actual vehicle location (state), each of which with a probability (emission probability). As the trajectory travels continuously, the transition between two consecutive timestamps is concluded by the travel possibility (transition probability) between their candidate states. Therefore, the objective is to find an optimal path which connects one candidate in every timestamp. The final path is obtained by the Viterbi algorithm which utilises the idea of dynamic programming. The major difference between various HMM-based algorithms is their definition of emission probability and transition probability. Unlike the emission probability, which is defined identically in most papers, the definition of the transition probability varies since the travel preference can be affected by plenty of factors. Some works[@newson2009hidden] prefers a candidate pair whose distance is similar to the distance between the observation pair, while others consider velocity changes[@goh2012online], turn restriction[@osogami2013map], closeness to the shortest path, the heading mismatch and travel penalty on U-turns, tunnels and bridges. Besides, HMM is also applied to online scenario[@goh2012online]. However, to build a reasonable Markov chain, online HMM-based algorithms usually suffer from latency problems, which means a point is matched after a certain delay.
### Conditional random field
CRF is utilized in many areas as an alternative to HMM to avoid the selection bias problem[@hunter2014path]. As both CRF and HMM are statistic models, the major difference is that CRF models interactions among observations while HMM models only model the relation between an observation with the state at the same stage and its closest predecessor. Hunter et al.[@hunter2014path] proposed a CRF-based map-matching algorithm that can be applied to both online and offline situations with high accuracy. Its overall approach is similar to HMM-based algorithms but with different transition probability which considers the maximum speed limit and the driving patterns of drivers. However, the problem shared by both HMM and CRF is the lack of a recovery strategy for the match deviation. Since once a path is confirmed, it will be contained by all future candidate paths, which hurts the online scenarios especially.
### Weighted graph technique
WGT refers to a model that infers the matching path from a weighted candidate graph, where the nodes are candidate road points of location measurements and edges are only formed between two nodes corresponding to two consecutive samples. In most WGT-based algorithms, candidate points are the closest points on road segments in a radius of measurements[@lou2009map; @hu2017if], which is similar to HMM. The process of the WGT can be summarized as three steps: (1) Initializing the candidate graph. (2) Weighting edges in the graph using a scoring function. (3) Inferring a path based on the weighted graph.
Algorithms fall in this category mainly differs from each other in weighting functions. Lou et al.[@lou2009map] firstly propose the WGT. It weights an edge simply based on a spatial cost and a temporal cost, where the spatial cost is modelled on the distance between candidate $c_{i}$ to its observed position $p_{i}$ and the shortest length between $c_{i}$ and $c_{i+1}$ whereas the temporal cost is modelled on the velocity reasonability. Based on Lou’s design, the following work further considers mutual influences between neighbouring nodes, road connectivity, travel time reasonability[@hu2017if] and other road features (traffic lights, left turns, etc.).
Candidate-Evolving Model
------------------------
Candidate-evolving model refers to a model which holds a set of candidates (also known as particles or hypotheses) during map-matching. The candidate set is initiated based on the first trajectory sample and keeps evolving by adding new candidates propagated from old ones close to the latest measurements while pruning irrelevant ones. Interpreting a candidate as a vote, by maintaining the candidate set, the algorithms are able to find a segment with the most votes, thereby, determining the matching path. Compare to the state-transition model, the candidate-evolving model is more robust to the off-track matching issue since the current matching is influenced not only by a previously defined solution, but also by other candidates. The particle filter (PF) and the Multiple Hypothesis Technique (MHT) are two representative solutions.
### Particle filter
PF is a state estimation technique that combines Monte Carlo sampling methods with Bayesian Inference. This technique has been utilized to support map-matching by the way of sensor fusion and measurement correction[@wang2016improved], while it is also applicable to directly address map-matching problem[@bonnifait2009multi]. The general idea of the PF model is to recursively estimate the Probability Density Function (PDF) of the road network section around the observation as time advances. Here, the PDF is approximated by $N$ discrete particles, each particle maintains a weight representing how consistent it is to the location observation. The process of a PF can be summarized as follows: Initially, $N$ particles are sampled with the same weight representing different locations in the local road network. The weight of each particle keeps getting updated as new observations are received. Then the PDF for the road network section around the new observations is calculated and the area with the highest probability is determined as the matched region. A resampling stage starts afterwards, where a new set of particles are derived based on the current set. The particles with higher weights are more likely to propagate according to moving status to feed particles for the next cycle, while those with low weights are likely to die out.
### Multiple hypothesis technique
Similar to PF, the MHT also tries to maintain a list of candidate road matches for the initial trajectory point and the list is expected to be as large as possible to ensure correct result coverage. However, different from the PF which iterate through all possibilities over time, the MHT is a much simpler model that inherits the idea of maintaining hypotheses but manages to reduce computation during the process. An MHT evaluates each candidate road edge (or point) based on a scoring function instead of trying to approximate the complicated PDF for the neighbour map area. Thereby, the computation cost of the MHT is dramatically reduced. According to the intuition, the MHT can be easily adopted in online scenario[@taguchi2018online]. Moreover, since it possesses all the possibility of previous hypotheses, Taguchi et al.[@taguchi2018online] propose a prediction model which extends the hypotheses to further predict the future route, which can achieve better online map-matching accuracy without introducing latency.
Scoring Model
-------------
### Naïve weighting
A group of algorithms[@quddus2015shortest; @sharath2019dynamic] apply the weight without using a particular model. Instead, they simply assign a group of candidates to each trajectory segment (or location observation) and find a road edge from each group that maximizes the predefined scoring function. The found segment in every timestamp is either returned if applied to the online scenario or waited to be joint with other matched segments if applied in the offline scenario. Most recent work in this category[@sharath2019dynamic] achieves a lane-level map-matching performance. The algorithm first identifies lanes in each road by utilising the road width information in the map and partition them into grids accordingly. The algorithm then finds candidate lane grids around the observed location and scores these grids at each timestamp. The grid results in the maximum score are then returned. The scoring function is a linear combination of four features, i.e. the proximity between the grid and trajectory sample, the estimated location of the vehicle at the next time stage, the reachability from the grid and the intention of a turn. These features are modelled individually, their scores can be obtained from the corresponding models in every timestamp. In addition, feature scores are weighted differently in the scoring function whose coefficients are computed by a training process before map-matching starts.
Challenges and Evaluations {#sec:experiment}
==========================
Despite various of map-matching models are proposed to deal with trajectory quality issues, the current solutions still fail to achieve decent matching quality in all scenarios. Therefore, in this section, we will discuss several major challenges caused by data quality issues that are affecting the map-matching results. We will demonstrate them both visually and experimentally to exemplify their significance.
Experimental Settings
---------------------
As listed in Table \[table:dataset summary\], we use four datasets for our experiments. The *Global*[@kubivcka2015dataset] dataset is a public dataset for map-matching evaluation. It contains 100 GPS trajectories sampled from 100 different areas all over the world, each of which is provided with a dedicate underlying map. Besides, we extract three sub-areas, namely *Beijing-U*, *Beijing-R* and *Beijing-M*, from a commercial dataset which contains taxi trajectories in Beijing. The reason for choosing these four datasets is their diversity in terms of trajectory quality and map density. The *Global* dataset has the best trajectory accuracy and its maps are also very sparse. The *Beijing-U* and *Beijing-R* represent two maps extracted from urban and rural areas, respectively. They have roughly the same size but different map density ($27.3 vs 13.9$), so they can be used to evaluate the influence of map density to map-matching results. *Beijing-M* is a larger map area with more trajectories for large-scale performance test.
\[table:dataset summary\]
Name
----------- -------------- --------------- ------------- ------------------ --------- ------------------ ---------------
*Trajectory* *Trajectory* *Sampling* *\# of vertices* *\# of* *Map* *Map Density*
*Count* *Point Count* *rate(sec)* *+ mini nodes* *edges* *Size($km^{2}$)* ($km/km^{2}$)
Global 100 1 N/A N/A N/A N/A
Beijing-U 7,905 247,544 11.0 7,672 4,484 9.9 27.3
Beijing-R 3,106 119,612 8.6 3,927 1,326 9.9 13.9
Beijing-M 73,072 3,285,934 10.3 41,353 22,580 57.0 24.2
Our experiments are performed on a single server with two Intel(R) Xeon(R) CPU E5-2630 with 10 cores/20 threads at 2.2GHz each, 378GB memory and Ubuntu 16.04. Both the route matching result $\mathcal{MR}(Tr)$ and the corresponding ground-truth are regarded as sets of road edges and are evaluated by F-measure, which is commonly used in map-matching evaluation[@wei2013mapc; @sharath2019dynamic]. The candidate map-matching algorithms used in the experiments include the most popular offline HMM map-matching[@newson2009hidden], the most-recent offline WGT algorithm[@yang2018fast] and an online Scoring method[@quddus2015shortest].
Data Quality Challenges
-----------------------
According to our observations from the experiments, the current data quality issues affect the map-matching in three major ways: the unnecessary detours, the matching breaks and the matching uncertainty.
### Unnecessary detour
As an example shown in Fig. \[fig:unecessary detour\], the matching result sometimes may contain unnecessary detours, which happens more frequently when the trajectory sampling rate is very high. In most scenarios, the detour is caused by two consecutive trajectory samples being too close to each other so that the succeeding point happens to be matched to the upper stream of its preceding point. Therefore, the shortest path between these two points has to go through a long detour. To avoid such issue, the measurement error should be considered when finding the shortest path, which means a certain degree of backtrace should be allowable. Alternatively, instead of simply project trajectory samples to the candidate roads to find candidate points, the actual matching point should follow a distribution, according to the trajectory measurement error, along the candidate road.
In general, as depicted in Fig. \[fig:accuracy sampling\], the detour problem strongly affects the matching quality when the sampling rate is high. The result shows that it is not always the case that a higher sampling rate leads to higher matching quality especially when the measurement error becomes the major problem. Therefore, a better way of modelling the measurement error is still required.
### Matching break
The matching break is a common problem in map-matching, which is mainly caused by trajectory outliers. This happens more frequently in the state-transition matching model when the correct state falls out of the candidate range of the outlier. In this case, the states of two consecutive observations may be unreachable, leading to disconnected matching route, as shown in the green circled area in Fig. \[fig:matching break\]. Currently, most of the solutions[@newson2009hidden] try to overcome this problem by identifying and removing the outliers to remedy the broken route. In Fig. \[fig:accuracy compression\], we apply online scoring method on *Beijing-M* with random down-sample and trajectory compression (Douglas-Peucker algorithm), respectively. The result shows that simple trajectory compression fails to prune outliers as they are usually preserved as outstanding point, which means more preprocessing step is required to remove such outliers. However, considering the detour problem in high sampling rate data, the trajectory compression achieves better performance compared with simply down-sample the trajectory as it better preserve the shape of the trajectory, which is still beneficial.
### Matching uncertainty
Although the main goal of map-matching algorithms is to reduce the uncertainty of trajectory, the matching uncertainty varies in different scenarios. One of the main factor, which is not mentioned by any of previous work, is the map density. Intuitively, the map-matching of trajectory is much harder when the map area is full of roads compared with an emptier region. As shown in Fig. \[fig:accuracy density\], the map density can significantly affect the matching quality as the *Beijing-U* has much worse performance than *Beijing-R* given both of them have a similar trajectory quality. On the other hand, the trajectory quality also plays an important role since the performance on *Global* is better than on *Beijing-U* with similar map density. Therefore, achieving decent performance on dense map area is still a challenging task for future map-matching research.
Conclusion {#sec:conclusion}
==========
In this paper, we conduct a comprehensive survey of the map-matching problem. We reveal the inability of all previous surveys in classifying new map-matching solutions. On top of that, we propose a new categorisation of existing methods from the technical perspective, which consists of similarity model, state-transition model, candidate-evolving model and scoring model. In addition, we list three major challenges (unnecessary detour, matching break and matching uncertainty) that the current map-matching algorithms are facing. To exemplify and demonstrate their influence on the current map-matching algorithms, we conduct extensive experiments over multiple datasets and map-matching algorithms. Overall, this paper concludes the current state of the map-matching problem and provides guidance to future research directions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this study, we investigate the sensitivity of the masses and decay constants of the light $f_{2}(1270)$ and $a_{2}(1320)$ tensor mesons to the temperature using QCD sum rule approach. In our calculations, we take into account the additional operators appearing in operator product expansion at finite temperature. It is obtained that at deconfinement temperature the decay constants and masses decrease with amount of $6\%$ and $96\%$ compared to their vacuum values, respectively. Our results on the masses at zero temperature are consistent with the vacuum sum rules predictions as well as the experimental data.'
address: |
$^{\dag}$Department of Physics, Doğuş University, Ac[i]{}badem-Kad[i]{}köy, 34722 Istanbul, Turkey\
$^{*}$Department of Physics, Kocaeli University, 41380 Izmit, Turkey
author:
- 'K. Azizi$^{\dag1}$, A. Türkan$^{*2}$, H. Sundu$^{*3}$, E. Veli Veliev$^{*4}$, E. Yaz[i]{}c[i]{}$^{*5}$'
title: Thermal behaviors of light unflavored tensor mesons in the framework of QCD sum rule
---
Introduction
============
In recent years, many researchers are focused on heavy-ion collision experiments in order to deeply understand the hadronic dynamics [@Apel; @Longacre; @Doser; @Kubota; @Ablikim; @Aaij]. It is believed that a transition occurs from hadronic matter to Quark Gluon Plasma (QGP) phase around critical temperature $T_{c}=175
MeV$. Therefore, thermal QCD calculations on the properties of hadrons can also be useful in understanding the phase diagram and other properties of strong interactions. In spite of considerable theoretical studies on properties of (pseudo)scalar and (axial)vector mesons at finite temperature (for instance see [@Mallik; @Veliev; @Veliev3; @Loewe; @arzu]), there are few theoretical studies on the thermal properties of the tensor mesons ( as an example see [@arzu2]).
In this article, we use the thermal QCD sum rule approach to investigate the sensitivity of the masses and decay constants of the light $f_{2}(1270)$ and $a_{2}(1320)$ tensor mesons to temperature (for the original work see [@arzu1]). In the literature, there are also few studies on the vacuum properties of tensor mesons using different models [@Aliev; @Aliev2; @Aliev3; @Sundu; @Ebert]. The QCD sum rule approach was firstly proposed by Shifman, Vainshtein and Zakharov [@Shifman] in vacuum as one of the most applicable tools to hadron physics. This method then was extended to finite temperature by Bochkarev and Shaposhnikov [@Bochkarev]. In this extension, the Wilson expansion and the quark-hadron duality approximation are valid, but the quark and gluon condensates are replaced by their thermal expectation expressions. In thermal QCD sum rule, due to the breaking of the Lorentz invariance some extra operators, which are expressed in terms of 4-vector velocity of the medium and the energy-momentum tensor [@Hatsuda; @Shuryak] are appeared in the Wilson expansion. Taking into account these new operators at finite temperature we obtain the sum rule for the light tensor mesons under consideration.
Thermal QCD sum rule for light mesons
=====================================
In this section our aim is to obtain sum rules for the masses and the decay constants of the $f_{2}$ and $a_{2}$ tensor mesons. For this reason, we start with the following thermal correlation function:
$$\begin{aligned}
\label{correl.func.101}
\Pi _{\mu\nu,\alpha\beta}(q,T)=i\int
d^{4}xe^{iq\cdot(x-y)}{\langle} {\cal T}[J _{\mu\nu}(x) \bar
J_{\alpha\beta}(y)]{\rangle}\mid_{y=0},\end{aligned}$$
where $\cal T$ indicates the time ordering operator and $J_{\mu\nu}$ is the interpolating current of the tensor mesons. This current for the mesons under consideration is given by $$\begin{aligned}
\label{tensorcurrent2}
J _{\mu\nu}^{f_2}(x)&=&\frac{i}{2\sqrt{2}}\Big[\bar u(x)
\gamma_{\mu} \olra{\cal D}_{\nu}(x) u(x)+\bar u(x) \gamma_{\nu}
\olra{\cal D}_{\mu}(x) u(x)
\nonumber\\
&+& \bar d(x)\gamma_{\mu} \olra{\cal D}_{\nu}(x) d(x)+ \bar d(x)
\gamma_{\nu} \olra{\cal D}_{\mu}(x) d(x)\Big],\end{aligned}$$ and $$\begin{aligned}
\label{tensorcurrent2}
J _{\mu\nu}^{a_2}(x)&=&\frac{i}{2\sqrt{2}}\Big[\bar u(x)
\gamma_{\mu} \olra{\cal D}_{\nu}(x) u(x)+\bar u(x) \gamma_{\nu}
\olra{\cal D}_{\mu}(x) u(x)
\nonumber\\
&-& \bar d(x)\gamma_{\mu} \olra{\cal D}_{\nu}(x) d(x)- \bar d(x)
\gamma_{\nu} \olra{\cal D}_{\mu}(x) d(x)\Big],\end{aligned}$$ where $ \olra{\cal D}_{\mu}(x)$ is the four-derivative with respect to the space-time acting on the left and right, simultaneously. We will set $y=0$ after applying derivatives with respect to $y$.
To obtain the thermal QCD sum rule for the tensor mesons we need to calculate the correlation function both in QCD and hadronic representations. To get the QCD side we use the operator product expansion (OPE) to separate the perturbative and non-perturbative contributions. The perturbative part in spectral representation is written as $$\begin{aligned}
\label{QCD Side}
\Pi^{pert}(q,T) =\int ds\frac{\rho(s)}{s-q^2},\end{aligned}$$ where $\rho(s)$ is the spectral density and it is given by the imaginary part of the correlation function, $\rho(s)=\frac{1}{\pi}Im[\Pi^{pert}(s,T)]$. We use the perturbative parts of the light quark propagator to get the spectral density (see [@arzu1] for more details). It is obtained as $$\begin{aligned}
\label{spectraldenstyf2a2}
\rho_{f_2
(a_2)}(s)=\frac{(m_{u}^{2}+m_{d}^{2})s}{32\pi^2}+\frac{3s^{2}}{160\pi^{2}}.\end{aligned}$$ From a similar manner, using the non-perturbative parts of the quark propagator we obtain the non-perturbative contributions as: $$\begin{aligned}
\label{nonpertf2a2}
\Pi^{non-pert}_{f_2(a_2)}=\frac{m_d
m_0^{2}}{144q^{2}}\langle\bar{d}d\rangle+\frac{m_u
m_0^{2}}{144q^{2}}\langle\bar{u}u\rangle -\frac{2\langle
u\Theta^{f}u\rangle(q\cdot u)^{2}}{9q^{2}},\end{aligned}$$ where $\Theta^{f}_{\mu\nu}$ is the fermionic part of the energy momentum tensor and $u_{\mu}$ is the four-velocity of the heat bath. According to the general idea of the QCD sum rules, after calculation of also the hadronic side of the correlator we match both the hadronic and OPE representations of the correlation function. As a result, we obtain the following sum rule: $$\begin{aligned}
\label{rhomatching}
f_{f_2(a_2)}^2(T)m_{f_2(a_2)}^6(T)e^{-m_{f_2(a_2)}(T)/M^{2}}=\int_{(m_u+m_d)^2}^{s_0(T)}
ds \rho_{f_2(a_2)}(s) e^{-s/M^{2}}
+\hat{B}\Pi_{f_2(a_2)}^{non-pert},\end{aligned}$$ where $\hat B$ denotes the Borel transformation with respect to $q^2$, $M^2$ is the Borel mass parameter and $s_0(T)$ is the temperature-dependent continuum threshold. It is given as $$\label{eqn16}
s_{0}(T)=s_{0}\frac{\langle\bar{q}q\rangle}{\langle0|\bar{q}q|0\rangle}\Big{(}1-\frac{(m_{q}+m_{d})^{2}}{s_{0}}\Big{)}+(m_{q}+m_{d})^{2}, \\$$ where $s_{0}$ on the right hand side is the hadronic threshold at zero temperature.
Numerical Results
=================
To obtain the values of the masses and decay constants we need to determine the working regions of two auxiliary parameters: the Borel mass parameter and the hadronic threshold at zero temperature. We choose the intervals $s_{0}=(2.2-2.5) GeV^2$ and $s_{0}=(2.4-2.7) GeV^2$ for the continuum thresholds in $f_2$ and $a_2$ channels, respectively. Also the working region of the Borel mass is taken as $1.4 GeV^2 \leq M^2\leq 3.0 GeV^2$. In these intervals, the dependence of the results on the auxiliary parameters are relatively weak.
Using the working regions of the auxiliary parameters and other input values, we obtain that the masses and decay constants are well described by the following fit functions in terms of temperature: $$\begin{aligned}
\label{tetamumu}
m_{f_{2}(a_{2})}(T)==Ae^{\alpha T}+B,\end{aligned}$$ and $$\begin{aligned}
\label{tetamumu}
f_{f_{2}(a_{2})}(T)=Ce^{\beta T}+D,\end{aligned}$$ where the temperature $T$ is in units of GeV. The parameters A, B, C, D, $\alpha$ and $\beta$ are given in Table 1.
$$\begin{array}{|c|c|c|c|c|c|c|}
\hline \hline \hline
&\mbox{A (GeV)} & \mbox{$\alpha (GeV^{-1}$)}& \mbox{B (GeV)} &\mbox{C} &\mbox{$\beta (GeV^{-1}$)} &\mbox{D}\\
\hline
\mbox{$f_2(1270)$} & -1.055\times10^{-5} & 72.674 &1.265 & -4.280\times10^{-6} & 42.808 & 0.043\\
\hline
\mbox{ $a_{2}(1320)$} & -1.255\times10^{-5} & 71.839 & 1.322 & -4.061\times10^{-6} & 42.662 & 0.042\\
\hline \hline \hline
\end{array}$$
We obtain that the values of the masses and decay constants are stable until temperature $0.1$ $GeV$, but after this point they start to decrease with altering the temperature (see [@arzu1] for more details). Also at deconfinement temperature, the decay constants and masses decrease with amount of $6\%$ and $96\%$ compared to their vacuum values, respectively.
The obtained results of masses and decay constants at zero temperature are $m_{f_{2}}=(1.28\pm0.08)~ GeV$, $f_{f_{2}}=0.041\pm0.002$, $m_{a_{2}}=(1.33\pm0.10)~ GeV$ and $f_{a_{2}}=0.042\pm0.002$, which are compatible with the vacuum predictions [@Aliev; @Aliev2; @Ebert] as well as the existing experimental data [@Beringer]. Our predictions on the temperature behaviors of decay constants and masses can be verified in the future experiments.
This work has been supported in part by the Scientific and Technological Research Council of Turkey (TUBITAK) under the research projects 110T284 and 114F018.
References {#references .unnumbered}
==========
[9]{} Apel W D et al. (Serpukhov-CERN Collaboration) 1975 *Phys. Lett.* B **57** 398 Longacre R S et al. 1986 *Phys. Lett.* B **177** 223 Doser M et al. (ASTERIX Collaboration) 1988 *Phys. Lett.* B **215** 792 Kubota Y et al. (CLEO Collaboration) 1994 *Phys. Rev. Lett.* **72** 1972 Ablikim M et al. (BES Collaboration) 2005 *Phys. Lett.* B **607** 243 Aaij R et al. (LHCb Collaboration) 2011 *Phys. Lett.* B **698** 14 Mallik S and Sarkar S 2002 *Eur. Phys. J.* C **25** 445 Veliev E V 2008 *J. Phys.* G **35** 035004; Veliev E V and Aliev T M 2008 *J. Phys.* G **35** 125002 Veliev E V, Azizi K, Sundu H and Aksit N 2012 *J. Phys.* G **39** 015002; Veliev E V and Kaya G 2009 *Eur. Phys. J.* C **63** 87 Dominguez C A, Loewe M, Rojas J C and Zhang Y 2010 *Phys. Rev.* D **81** 014007; Dominguez C A, Loewe M and Rojas J C 2007 *JHEP* **08** 040 Veliev E V, Azizi K, Sundu H, Kaya G and Türkan A 2011 *Eur. Phys. J.* A **47** 110 Azizi K, Sundu H, Türkan A and Veliev E V 2014 *J. Phys.* G **41**, 035003 Azizi K, Türkan A, Veliev E V, Sundu H, 2014 arXiv:1410.5237 \[hep-ph\] Aliev T M and Shifman M A 1982 *Phys. Lett.* B **112** 401 Aliev T M, Azizi K and Bashiry V 2010 *J. Phys.* G **37** 025001 Aliev T M, Azizi K, Savc[i]{} M *Phys.Lett.* B **690** 164 Sundu H and Azizi K 2012 *Eur. Phys. J.* A **48** 81 Ebert D, Faustov R N and Galkin V O 2009 *Phys. Rev.* D **79** 114029
Shifman M A, Vainshtein A I and Zakharov V I 1979 *Nucl. Phys.* B **147** 385; 1979 *Nucl. Phys.* B **147** 448 Bochkarev A I and Shaposhnikov M E 1986 *Nucl. Phys.* B **268** 220 Hatsuda T, Koike Y and Lee S H 1993 *Nucl. Phys.* B **394** 221 Shuryak E V 1993 *Rev. Mod. Phys.* **65** 1 J. Beringer et al. (Particle Data Group), Phys. Rev. *D86*, 010001 (2012).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'These notes summarise a talk surveying the combinatorial or Hamiltonian quantisation of three dimensional gravity in the Chern-Simons formulation, with an emphasis on the role of quantum groups and on the way the various physical constants ($c,G,\Lambda,\hbar)$ enter as deformation parameters. The classical situation is summarised, where solutions can be characterised in terms of model spacetimes (which depend on $c$ and $\Lambda$) together with global identifications via elements of the corresponding isometry groups. The quantum theory may be viewed as a deformation of this picture, with quantum groups replacing the local isometry groups, and non-commutative spacetimes replacing the classical model spacetimes. This point of view is explained, and open issues are sketched.'
---
EMPG-11-11\
28 pt
[**Quantum gravity and non-commutative spacetimes in three dimensions: a unified approach**]{}
18 pt
[ Bernd J. Schroers[^1]\
Department of Mathematics and Maxwell Institute for Mathematical Sciences\
Heriot-Watt University\
Edinburgh EH14 4AS, United Kingdom ]{}
Talk given at ‘Geometry and Physics in Cracow’, September 2010
Introduction and motivation
===========================
Historical remarks
------------------
Giving a talk on three dimensional (3d) gravity at a meeting in Cracow is like carrying coal to Newcastle: the beginnings of the subject are usually traced back to the paper [@Staruszkiewicz] by Andrzej Staruszkiewicz, alumnus and later professor at the Jagellonian University in Cracow. Staruszkiewicz’s paper, published in 1963, is about classical 3d gravity and its special features. The subject of 3d [*quantum*]{} gravity started only five years later with the realisation by Ponzano and Regge [@PonzanoRegge] that angular momentum theory plays an important role in this context.
Gravity in 3d is now a large subject in its own right, which I can not possibly review here. However, in this introductory part of the talk I will at least attempt to identify a few of the main themes and relate them to the approach followed here. Influential papers by Deser, ’t Hooft and Jackiw written in the 1980s [@DJtH; @DeserJackiw1; @DeserJackiw2; @tHooftscattering] on classical and quantum scattering of particles demonstrated the possibility of carrying out non-perturbative calculations of quantum scattering processes in 3d gravity. As we shall see, they also contain indications of the relevance of the braid group in describing such processes. These indications are elaborated in the later literature, see for example [@Carlipscattering; @BM; @BMS], and turn out to be closely related to the quantum group approach pursued in this talk.
The Chern-Simons formulation of 3d gravity, observed in [@AT] and elaborated in [@Witten], establishes a connection between 3d gravity and a host of areas in mathematical physics, including topological field theory, knot theory, the theory of Poisson-Lie groups and of quantum groups. Since this talk is based on the Chern-Simons approach, we will see many of these connections.
The early paper by Ponzano and Regge, mentioned above, provides the foundation of the spin foam approach to 3d quantum gravity. This is perhaps the approach to 3d quantum gravity that contains the most directly useful lessons for 4d quantum gravity. I will not discuss this approach in this talk, and shall not attempt to summarise the large literature on it. However, it is worth pointing out that there are close links with Chern-Simons theory (spin foam state sums may be viewed as discretisation of the path integral) and to quantum groups, see [@TuraevViro] for an early paper and [@Foxton; @BarrettNaishGuzman] for examples of recent papers with many references.
The possibility that non-commutative geometry is needed to describe spacetime at the quantum level has long been a theme in quantum gravity research [@DFR], see [@BianchiRovelli] for a recent discussion with some references. It is therefore interesting to ask if one can use the relatively tractable 3d situation to establish the role of non-commutative geometry in quantum gravity in a mathematically convincing way. Early discussions of non-commutative spacetime coordinates appear in the paper [@tHooftdiscrete]. Spacetime non-commutativity in 3d quantum gravity is studied, in different approaches, in [@MatWell; @BatistaMajid; @FreidelLivine; @JoungMouradNoui]. Putting these approaches into one coherent picture is one of the objectives of this talk.
Finally, I should mention two further important themes of 3d gravity research which I will not be able to touch on in this talk. One is the study of BTZ black holes, an introduction to which can be found in the book [@Carlipbook]. The other is the relation to 3d hyperbolic geometry, where the papers and books [@Mess; @Messcomment; @BenedettiBonsante; @Thurston] may provide good starting points.
Topological degrees of freedom and interactions in 3d gravity {#intro}
-------------------------------------------------------------
The Einstein field equations (without cosmological constant and in units where the speed of light is 1) $$R_{ab}-\frac 1 2 R g_{ab} = -8\pi GT_{ab}$$
determine the Ricci tensor of a spacetime in terms of the energy momentum tensor. In spacetime dimensions greater than three, the Ricci tensor does not fix the Riemann tensor and it is possible to have metrically non-trivial (i.e. curved) spacetimes satisfying the vacuum ($T_{ab}=0$) field equations. In three spacetime dimensions, this is not possible. The Ricci tensor determines the Riemann tensor and, as a result, the only vacuum solutions of the Einstein equations with vanishing cosmological constant are flat [@Carlipbook]. This result simplifies Einstein’s theory of gravity in 3d dramatically, but does not render it trivial. There are non-trivial solutions of the Einstein equations in the presence of matter, and, if the topology of the three-dimensional manifold representing the universe is non-trivial, there may be vacuum solutions which, though flat, have non-trivial holonomy. These observations are often summarised in the slogan that in 3d gravity there are no gravitational waves but that the theory has topological degrees of freedom.
The simplest solution of the Einstein equations illustrating the previous paragraph is the spacetime surrounding a point-particle. The energy-momentum tensor is a Dirac delta-function with support on the world line of the particle.The metric solving the field equations is flat away from the world line and is singular on the world line. More precisely it is a direct product of a cone (space) and ${\mathbb{R}}$ (representing time) [@Carlipbook]. The line element, in terms of polar coordinates $(r,\phi)$, with $r>0$, and a time coordinate $t$ is simply \[pointsolution\] ds\^2 = c\^2dt\^2 -dr\^2 - r\^2 d\^2. However, the range of $\phi$ is $[0,2\pi-\mu)$, where the parameter $\mu$ is related to the particle’s mass $m$ and to Newton’s constant $G$ via $$\mu = 8\pi G m.$$ In three dimensions , the physical dimension of $G$ is that of an inverse mass so that $\mu$ is a dimensionless, angular parameter. The effect of a particle on the geometry of spacetimes is, then, to cut out a wedge of size $\mu$ from the spacetime surrounding the particle’s world line.
It is instructive to consider the effect of the geometry on light test particles. Such particles travel on geodesics, which are simply straight lines on the cone after it has been cut open. It is easy to check that geodesics passing the particle of mass $m$ on one side are deflecting relative to particles who pass it on the other side by the angle $\mu$ (in the coordinate system $(t,r,\phi)$). This relative deflection is illustrated in Fig. \[conedeflection\] and is independent of the distance of closest approach between the heavy particle of mass $m$ and the test particles (impact parameter). The interaction is topological in the sense that it only depends on whether the test particle passes on the left or the right of the heavy particle, and not on the relative distance. This kind of interaction is familiar from the Aharonov-Bohm interaction between electrons and a magnetic flux, and this analogy can be made precise: both interactions can be related to the braiding of the world lines of the interacting particles [@BMS].
![Geodesics in the space surrounding a conical singularity with deficit angle $\mu$[]{data-label="conedeflection"}](conegeodesics){width="5truecm"}
Physical constants entering 3d quantum gravity
----------------------------------------------
The four physical constant entering 3d quantum gravity are the speed of light $c$, Newton’s constant $G$, Planck’s constant $\hbar$ and the cosmological constant $\Lambda$. From these, we can form two length constants (remembering that the dimension of $G$ is an inverse mass), namely \[constants\] \_P=, \_C = . In this talk we will deal with both Lorentzian and Euclidean gravity, and we parametrise Euclidean and Lorentzian metrics in a unified fashion by allowing $c^2<0$ in the Euclidean situation. As a result, both the length parameters in may be imaginary, depending on the sign of $c^2$ and $\Lambda$. From the ratio of the two length parameters we can form a dimensionless quantity. We define the deformation parameter \[qpara\] q= e\^[-]{}, which may take values on the real line or the unit circle in the complex plane.
It is useful to clarify the role played by the various constants in 3d gravity in general terms at this stage. The observation of the previous section that, in the absence of matter, solutions of the Einstein equations are locally flat generalises in the presence of a cosmological constant to the statement that vacuum solutions are locally isometric to model space times, which depend on the parameters $c$ and $\Lambda$. For Lorentzian gravity with vanishing cosmological constant, for example, the model spacetime is Minkowski space while for Euclidean gravity with positive cosmological constant it is the four-sphere with the round metric. The isometry groups of the model spacetimes inherit a dependence on $c$ and $\Lambda$. In the examples above they are, respectively, the Poincaré group in 3d and the 4d rotation group $SO(4)$. Newton’s constant $G$ enters when one studies the dynamics of spacetime and plays the role of a parameter in the Poisson structure and that of a coupling constant to matter. Finally, $\hbar$ enters in the quantisation and the dimensionless parameter $q$ in , combining all four constant, controls the quantum theory when all the constants $1/c,G, \Lambda,\hbar$ are non-zero.
Motivation and outline of the talk
----------------------------------
The goal of this talk is give a unified account of aspects of classical and quantum gravity in 3d, in which the physical parameters of the previous section enter as deformation parameters. Our account of classical gravity is based on the formulation of 3d gravity as a Chern-Simons gauge theory, where the local isometry groups play the role of the gauge groups. As well shall see, the parameters $c$ and $\Lambda$ enter in this description via the structure constants of the Lie algebra of the gauge group, while the parameter $G$ enters via the inner product (or trace) on the Lie algebra which is used in the Chern-Simons action. We sketch the description of the phase space of 3d gravity as the moduli space of flat connections, and review the description of its Poisson structure in a formulation, due to Fock and Rosly [@FockRosly], which makes essential use of classical $r$-matrices.
The description of the Poisson structure in terms of $r$-matrices is tailor-made for the quantisation via the combinatorial or Hamiltonian scheme pioneered in [@AGSI], [@AGSII] and [@AS]. In this scheme, the quantisation is controlled by quantum groups which are deformations of the local isometry groups of the model spacetimes, with deformation parameters $G$ and $\hbar$ in addition to $c$ and $\Lambda$. These quantum groups naturally act on non-commutative spaces, which one may interpret as deformations of the classical model spacetimes. This framework thus provides a concrete mathematical setting for exploring the proposal that, in quantum gravity, spacetime should be mathematically modelled in terms of non-commutative geometry. We end our talk with an evaluation of the successes and limitations of this approach to 3d quantum gravity.
Model spacetimes and isometry groups
====================================
The following treatment of the model spacetimes follows closely that in [@PapageorgiouSchroers1]. We use Roman letters $a,b,c \ldots$ for 3d spacetimes indices, with range for $\{0,1,2\}$ (in both the Euclidean and Lorentzian case). The model spacetimes arising in 3d gravity can be described in a simple an unified fashion in terms of the metric \[4dmetric\] g\_=(-c\^2,1,1,1 ) in an auxiliary ${\mathbb{R}}^4$. Here we use Greek indices for the range $\{0,1,2,3\}$. The model spacetimes can be realised as embedded hypersurfaces via \[hypersur\] H\_[c,]{}= {(t,x,y, w)\^4| -c\^2t\^2 + x\^2 + y\^2 + w\^2 = } . This two-parameter family includes the three-sphere $S^3$ ($c^2<0$, $ \Lambda >0$), doubles covers of hyperbolic space $H^3$ ($c^2 <0$, $\Lambda <0$), de Sitter space dS$^3$ ($c^2>0$, $ \Lambda >0$) and anti-de Sitter space AdS$^3$ ($c^2>0$, $\Lambda <0$). Double covers of Euclidean space $E^3$ and Minkowski $M^3$ space arise in the limit $\Lambda\rightarrow 0$, which one should take [*after*]{} multiplying the defining equation in by $\Lambda$. In Fig. \[modelspaces\] we show the embedded model spacetimes (with one spatial dimension suppressed).
![(Double covers of) Model spacetimes for 3d gravity, shown as 2d models embedded in a 3d auxiliary space with coordinates $(t,x,w)$ according to (the second spatial coordinate $y$ is suppressed). Euclidean and Minkowski space at the top, spherical and de Sitter space in the middle, hyperbolic and Anti-de Sitter space at the bottom[]{data-label="modelspaces"}](minkspace "fig:"){width="5.5truecm"}\
![(Double covers of) Model spacetimes for 3d gravity, shown as 2d models embedded in a 3d auxiliary space with coordinates $(t,x,w)$ according to (the second spatial coordinate $y$ is suppressed). Euclidean and Minkowski space at the top, spherical and de Sitter space in the middle, hyperbolic and Anti-de Sitter space at the bottom[]{data-label="modelspaces"}](sphere "fig:"){width="5.5truecm"} ![(Double covers of) Model spacetimes for 3d gravity, shown as 2d models embedded in a 3d auxiliary space with coordinates $(t,x,w)$ according to (the second spatial coordinate $y$ is suppressed). Euclidean and Minkowski space at the top, spherical and de Sitter space in the middle, hyperbolic and Anti-de Sitter space at the bottom[]{data-label="modelspaces"}](deSitter "fig:"){width="5.5truecm"}\
![(Double covers of) Model spacetimes for 3d gravity, shown as 2d models embedded in a 3d auxiliary space with coordinates $(t,x,w)$ according to (the second spatial coordinate $y$ is suppressed). Euclidean and Minkowski space at the top, spherical and de Sitter space in the middle, hyperbolic and Anti-de Sitter space at the bottom[]{data-label="modelspaces"}](hyperbolicspace "fig:"){width="5truecm"} ![(Double covers of) Model spacetimes for 3d gravity, shown as 2d models embedded in a 3d auxiliary space with coordinates $(t,x,w)$ according to (the second spatial coordinate $y$ is suppressed). Euclidean and Minkowski space at the top, spherical and de Sitter space in the middle, hyperbolic and Anti-de Sitter space at the bottom[]{data-label="modelspaces"}](antideSitter "fig:"){width="5.5truecm"}
In order to be able to take the limit $\Lambda \rightarrow 0$ for the associated isometry groups it is best to work with the inverse metric \[invmet\] g\^=(-,1,1,). The Lie algebra generators of the isometry groups of can conveniently be defined in terms of the Clifford algebra associated to [@PapageorgiouSchroers1]. Thus we define generators $\gamma^\mu$ via {\^,\^}=-2g\^, so that the six Lie algebra generators are given by M\^=1 4 \[\^,\^\]. They have the commutation relations \[masterlie\] \[ M\^, M\^ \] = g\^ M\^ + g\^ M\^- g\^ M\^-g\^ M\^. The advantage of the Clifford algebra approach is that one can immediately write down two naturally defined invariant bilinear forms. One, denoted $\langle\cdot , \cdot \rangle$ is defined by carrying out the Clifford multiplication and projecting onto the invariant, central element $\gamma^5=\gamma^0\gamma^1\gamma^2\gamma^3$. Multplying by $-4$ for later convenience, the resulting inner product is non-zero whenever the indices on the basis vectors are complementary, for example $$\langle M^{12},M^{03}\rangle=-1, \quad \langle M^{12},M^{01}\rangle=0 \quad \text{etc.}$$ Another bilinear form $(\cdot, \cdot)$ is obtained by carrying out the Clifford multiplication and projecting onto the identity. Again rescaling by $-4$ for convenience we have a non-zero answer whenever the indices on the basis vectors match: $$(M^{12},M^{12})=1,\quad (M^{01},M^{01})=-\frac{1}{c^2},
\quad (M^{13},M^{13})=\Lambda \quad \text{etc.}$$ As we shall see shortly, this is the Killing form on the Lie algebra
We now express the above generators in more conventional 3d notation. For this purpose we define the three-dimensional totally antisymmetric tensor with downstairs indices via $\epsilon_{012}=1$. Then we define the rotation generator $\tJ_0$, the boost generators $\tJ_1,\tJ_2$ and translation generators $\tP_a$ via \[liegens\] \_a=1 2 \_[abc]{}M\^[bc]{},\_a=g\_[ab]{}M\^[b3]{}, where we used the spacetime part of the 4d metric $g_{\mu\nu}$ to lower indices, and refer to [@PapageorgiouSchroers1] for a discussion of physical dimensions and interpretation of these generators (which are denoted by the same letters, but without tilde there). The Lie algebra brackets are now \[t21alg\] \[\_a,\_b\]=\_[abc]{}J\^c, =\_[abc]{} \^[ c]{},=-c\^2 \_[abc]{}\^c, with indices raised via the inverse metric $g^{ab}$. The combination $-c^2\Lambda$ which occurs in the Lie brackets plays an important role in what follows, and we introduce \[algebraicconstant\] =-c\^2. The bilinear form $(\cdot, \cdot)$, already advertised as the Killing form, is \[nongravin\] (\_a,\_b)=\_[ab]{},(\_a,\_b)=\_[ab]{}, where \_[ab]{}=- g\_[ab]{} = (1, - ,-). The metric $\kappa_{ab}$ is the most natural one on the Lie algebra $so(3)$ respectively $so(2,1)$ spanned by $\tJ_0,\tJ_1$ and $\tJ_2$. Note that it differs from the spacetime metric $g_{ab}$, but that it has the right physical dimensions and that imaginary $c$ gives the usual Euclidean metric, as required.
It is one of the coincidences of 3d that spacetime and the Lie algebra of rotations and/or boosts are both three-dimensional. Both are equipped with Euclidean respectively Lorentzian metrics, but our derivation shows that, in a physically natural normalisation and construction, the spacetime and Lie algebra metrics come out differently. This is potentially confusing in calculations where indices are raised and contracted with these metrics, and most papers on 3d gravity use conventions where the two kinds of metrics coincide. We can achieve this by switching from the physical Lie algebra basis used thus far to a geometrical basis according to $$\begin{aligned}
\tJ_0 &\rightarrow J_{0}=-\frac{|c|^2}{c^2} \tJ_0,\quad \tJ_1 \rightarrow J_1 = |c| \tJ_1, \quad \tJ_2 \rightarrow J_2 = |c| \tJ_2, \nonumber \\
\tP_0 &\rightarrow P_{0}=-\frac{|c|^2}{c^2} \tP_0,\quad \tP_1 \rightarrow P_1 = |c| \tP_1, \quad \tP_2 \rightarrow P_2 = |c| \tP_2.\end{aligned}$$ In this geometrical basis, all the generators $J_a$ are dimensionless, and all the translation generators $P_a$ have the dimension of inverse time. One checks that the Killing metric now takes the form \[3dmetric\] (J\_a,J\_b)= \_[ab]{}:=(1,-, -), which is diag$(1,1,1)$ in the Euclidean and diag$(1,-1,-1)$ in the Lorentzian case. Moreover, the Lie brackets take the [*same*]{} form as in , \[21alg\] \[J\_a,J\_b\]=\_[abc]{}J\^c, =\_[abc]{} P\^[ c]{},= \_[abc]{}J\^c, but all indices are now raised with the Lie algebra metric $\eta_{ab}$. This is convenient and we shall work in this basis for the remainder of this talk. We denote the Lie algebra with these brackets by $\glambda$. The conventions regarding the metric then agree with [@MeusburgerSchroers6], but the convention regarding the naming of $\lambda$ agrees with [@Witten] and differs from [@MeusburgerSchroers6], where $\Lambda$ was used for what we call $\lambda$ now. Conventions regarding the naming of the cosmological constant and the combination differ in the literature, and the reader will need to take good care when comparing results from different sources.
The other bilinear form introduced in the Clifford language gives the following non-zero pairings \[3dgravin\] J\_a, P\_b=c\^2\_[ab]{}. This pairing is non-degenerate for any value of $\lambda$ and is crucial for the Chern-Simons formulation of 3d gravity, as we shall see.
In Table \[isogroups\] we list Lie groups whose Lie algebras are . We have used the isomorphisms $SU(2)/{\mathbb{Z}}_2 = SO(3)$ and $SL(2,{\mathbb{R}})/{\mathbb{Z}}_2= SO(2,1)_0$, the identity component of $SO(2,1)$. The isometry groups are determined by their Lie algebras only up to coverings, and our choice in Table \[isogroups\] is one of convenience. In the following, we write $\Glambda$ for this family of Lie groups.
--------------- ------------------------------- ------------------------------------------------
Cos. constant Euclidean ($c^2<0$) Lorentzian ($c^2>0$)
$\Lambda =0$ $SU(2)\ltimes {\mathbb{R}}^3$ $SL(2,{\mathbb{R}}) \ltimes {\mathbb{R}}^3$
$\Lambda> 0$ $ SU(2)\times SU(2)$ $SL(2,{\mathbb{C}})$
$ \Lambda< 0$ $ SL(2,{\mathbb{C}}) $ $SL(2,{\mathbb{R}})\times SL(2,{\mathbb{R}}) $
--------------- ------------------------------- ------------------------------------------------
: Local isometry groups in 3d gravity[]{data-label="isogroups"}
The Chern-Simons formulation of 3d gravity
==========================================
In Cartan’s approach to Riemannian geometry [@Sharpe] the fundamental geometrical object is a connection which combines an orthonormal frame field (or vielbein) $e_a$ and the spin connection $\omega_{ab}$ on the orthonormal frame bundle into the so-called Cartan connection. Concretely, in the case of 3d geometry, we combine the dreibein with the translation generators $P_a$ of and the local connection one-forms $\omega^a = \frac 1 2 \epsilon^{abc}\omega_{bc}$ with the rotation and/or Lorentz generators $J^a$ into the local one-form \[Cartan\] A=e\_aP\^a + \_aJ\^a, taking values in the Lie algebra $\glambda$. The curvature \[Cartancurv\] F\_A=dA+1 2\[AA\] =R+C+T of the Cartan connection combines the Riemann curvature of the spin connection $\omega= \omega_a J^a$, $$R= d\omega + \frac 1 2 [\omega\wedge\omega],$$ a cosmological term $$C=\frac \lambda 2\epsilon^{abc} e_a\wedge e_b J_c,$$ and the torsion $$T=(de^c +\epsilon^{abc}\omega_a\wedge e_b) P_c.$$
In the Cartan approach to general relativity (in any dimension), the Einstein-Hilbert action is expressed in terms of the vielbein and the connection, which are treated as independent variables. The action is called the Palatini action when interpreted in this way. In this approach, the condition of vanishing torsion (in the absence of spin sources) follows as a variational equation rather than as an a priori condition. It turns out that, in three dimensions, the Einstein-Hilbert (or Palatini) action is simply the Chern-Simons action for the Cartan connection , with the bilinear form used as an inner product [@AT; @Witten]. However, beyond the equality of the actions, the relationship between the Chern-Simons formulation and the Einstein formulation of 3d gravity is subtle: non-invertible dreibeins $e_a$ may occur in the Chern-Simons formulation but are ruled out in the Einstein approach, based on metrics. This changes the nature of gauge orbits in the two cases, so that the physical phase spaces are, in general, different. This was pointed out in a 1+1 dimensional context in [@SchS] and was demonstrated in an explicit example involving four particles in 3d gravity in [@Matschull2]. Our approach to 3d gravity in the remainder of this talk is based on the Chern-Simons formulation.
We discuss the Chern-Simons action in terms of the general bilinear form \[genform\] (,)\_ = , + (,) on the Lie algebra $\glambda$. This form is non-degenerate iff [@MeusburgerSchroers7] \[nondegcond\] \^2 - \^20, and the associated action $$\begin{aligned}
\label{genact}
I_{\alpha\beta}(A)&=\int_M (A\wedge dA)_{\alpha\beta}+ \frac 1 3 (A\wedge[A,A])_{\alpha\beta}
\nonumber \\
&=\alpha \int_M\left(2 e^a\wedge R_a + \frac \lambda 3 \epsilon_{abc}
e^a\wedge e^b\wedge e^c \right) \nonumber \\
&+\beta \int_M\left(\omega^a\wedge d\omega_a +\frac{1}{3}\epsilon_{abc}\omega^a
\wedge \omega^b\wedge\omega^c + \lambda e^a\wedge T_a\right),\end{aligned}$$ contains the gravitational action (the terms proportional to $\alpha$), the Chern-Simons action for the spin connection and additional terms involving torsion. This general action was first considered by Mielke and Baekler [@MielkeBaekler] and recently revisited in [@BonzomLivine], where the analogy between the terms proportional to $\beta$ and the Immirzi term in 4d was stressed. The variational equations which follow from the general action are simply the flatness condition for the Cartan connection, i.e. the vanishing of , provided the form is non-degenerate. This appears to imply that the family of actions leads to equivalent physics provided the condition holds. However, as argued in [@MeusburgerSchroers7], the induced canonical structure of the phase space does depend on the ratio of $\alpha$ and $\beta$. Since we are only interested in the Chern-Simons formulation of 3d gravity here, we set \[alphachoice\] =, = 0 from now onwards.
The gauge formulation of 3d gravity can easily and naturally be extended to include minimal coupling between the gauge field and point particles. This was first discussed in in detail in [@SousaGerbert] and is reviewed in our notation in [@MeusburgerSchroers7], where the dependence of the coupling on the parameters $\alpha$ and $\beta$ is also discussed. We are not able to discuss the coupling to particles, the Poisson structure and the division by gauge equivalence in the space available here. Instead, we summarise the results in the next section, and motivate them in general, geometric terms.
Classical $r$-matrices and Poisson brackets on the space of holonomies
======================================================================
Having established that, in the Chern-Simons formulation, classical solutions of the field equations are flat $\Glambda$-connections, we can characterise the phase space of 3d gravity on a manifold $M^3$ in the Chern-Simons formulation as the space of flat $ \Glambda$-connections on $M^3$, modulo gauge transformations. In order to make this precise and concrete, we consider 3d universes of topology $M^3={\mathbb{R}}\times S$, where $S$ is a two-dimensional manifold representing space. Then one can show [@Witten] that the phase space is the moduli space of flat $G_\lambda$-connections on $S$ (i.e. the space of flat $G_\lambda$-connections moduli gauge transformations), equipped with the Atiyah-Bott symplectic structure [@AtiyahBott; @Atiyah] , which is defined in terms of the bilinear form used in the Chern-Simons action. With the choice this bilinear form is \[gravinprod\] , . Therefore, in the Chern-Simons formulation, and assuming the factorisation $M^3={\mathbb{R}}\times S$, the task of constructing a theory of quantum gravity amounts to quantising the moduli space of flat $\Glambda$-connections on $S$, with a symplectic structure induced by .
Despite the elegance and generality of this result, a precise mathematical description of this moduli space and a rigorous quantisation remains a difficult task. In the case where $S$ is a compact surface of genus $g\geq 2$, the moduli space can characterised in terms of the moduli space $\mathcal{A}_S$ of flat $SU(2)$ connections in the Euclidean case and in terms of Teichmüller Space $\mathcal{T}_S$ (a component of the moduli space of flat $SL(2,{\mathbb{R}})$ connections) in the Lorentzian case. In Table \[phasesummary\] we reproduce a summary of the results given in [@KrasnovSchlenker], where further references can be found. The results in the Lorentzian case are due to [@Mess; @Messcomment].
---------------- ------------------------------------------------------------- ----------------------------------------------------------
Cos. constant Euclidean ($c^2<0$) Lorentzian ($c^2>0$)
$ \Lambda=0$ $T^*\mathcal{A}_S$ $T^*\mathcal{T}_S$
$\Lambda > 0$ $\mathcal{A}_S\times \mathcal{A}_S$ $\mathcal{T}_S\times \mathcal{T}_S\sim T^*\mathcal{T}_S$
$ \Lambda < 0$ $\mathcal{T}_S\times \mathcal{T}_S\subset T^*\mathcal{T}_S$ $\mathcal{T}_S\times \mathcal{T}_S\sim T^*\mathcal{T}_S$
---------------- ------------------------------------------------------------- ----------------------------------------------------------
: Phase space of 3d gravity for universes of the form ${\mathbb{R}}\times S$, with $S$ compact and genus $\geq 2$ (quoted from [@KrasnovSchlenker]).[]{data-label="phasesummary"}
For each of the symplectic manifolds in the table, one may in principle attempt a quantisation and subsequent interpretation in terms of 3d quantum gravity. In this talk I summarise a description of the moduli space and its Poisson structure which is closely based the parametrisation in terms of $\Glambda$-valued holonomies, and which uses a concrete and unified description of the Poisson structure, which is tailor-made for quantisation. The idea for this description is due to Fock and Rosly [@FockRosly]. It is the foundation of the combinatorial or Hamiltonian quantisation programme for Chern-Simons theory, described in [@AGSI; @AGSII; @AS].
Fock and Rosly’s description of the phase space starts with the observation that flat connections on a manifold are characterised by their holonomies along non-contractible paths. The moduli space of flat connections on a surface $S$ can thus be parametrised by the set of holonomies along closed paths which generate the fundamental group of $S$, modulo gauge transformations at the common starting and end point of those paths. So far we have assumed that $S$ is a compact manifold without boundary, but in the Fock and Rosly description it is easy to include punctures decorated with co-adjoint orbits of $\Glambda$. This is desirable in the context of 3d gravity, since a co-adjoint orbit of $\Glambda$ physically correspond to the phase space of a point particle, and the ‘decoration’ of a puncture with a co-adjoint orbit is precisely the effect of minimal coupling between the Cartan connection and the point particle’s degrees of freedom. Moreover, this minimal coupling correctly reproduces the gravitational coupling between a point particle and the gravitational field, with momentum acting as a source of curvature and spin acting as a source for torsion. For details we refer the reader to the papers [@SousaGerbert; @MeusburgerSchroers7] and for a relatively brief but pedagogical account to the talk [@sissatalk].
The effect of the minimal coupling to co-adjoint orbits on the holonomies can be summarised as follows. Using the inner product , co-adjoint orbits can be written as adjoint orbits. For particles with mass $m$ and spin $s$, these orbits are of the form $$O_{ms}=\{g(-\mu J_0 - \sigma P_0)g^{-1} | g \in \Glambda\},$$ where $$\mu =8\pi G{m}, \quad \sigma =8\pi G s.$$ Decorating a puncture on $S$ with such an orbit forces the holonomy around the puncture to lie in the conjugacy class $${\cal C}_{\mu \sigma}=\{g(\exp(-\mu J_0 - \sigma P_0))g^{-1} | g \in \Glambda\}.$$ For a genus $g$ surface $S$ with $n$ punctures and orbit labels $\mu_i,\sigma_i$, $i=1\ldots n$, a set of generators of the fundamental groups is shown in Fig. \[fundamentalgroup\]. The moduli space of flat $\Glambda$-connections can be written in terms of the extended phase space \[extended\] = \^[2g]{}\_[\_n \_n]{}…[C]{}\_[\_1 \_1]{}, by imposing the condition that a suitable composition of the generating loops is contractible (and hence has trivial holonomy), and by dividing by conjugation at the base point: $$\begin{aligned}
{\cal P} & = \{(A_g,B_g,\ldots, A_1,B_1, M_n,\ldots M_1)\in \tilde {\cal P}| \nonumber \\
& \quad [A_g,B_g^{-1}]\ldots[ A_1,B_1^{-1}]M_n\ldots M_1=1\}/\mbox{conjugation}.
\label{truephase}\end{aligned}$$
{width="6truecm"} \[fundamentalgroup\]
The trick introduced by Fock and Rosly is to define a (symplectic) Poisson structure on the extended phase space $\tilde {\cal P}$ in such a way that the $\Glambda$-conjugation action on $\tilde {\cal P} $ is symplectic and that the symplectic quotient by it gives $\cal P$ with the Atiyah-Bott symplectic structure. The Poisson structure on $\tilde {\cal P}$ is defined in terms of a classical $r$-matrix, i.e. an element $r\in \glambda\otimes \glambda$ which satisfies the classical Yang-Baxter equation (CYBE) \[cybe\] \[r\_[12]{},r\_[13]{}\]+\[r\_[12]{},r\_[23]{}\]+\[r\_[13]{},r\_[23]{}\]=0, where we have used standard notation, explained, for example in textbooks like [@CP] or [@Majidbook1]. The information about the inner product used in the definition of the Atiyah-Bott symplectic structure (or, equivalently, in the Chern-Simons action) is encoded in $r$ via the following compatibility requirement:
[**Definition:**]{} An $r$-matrix is compatible with a Chern-Simons action if it satisfies the CYBE and if its symmetric part is equal to the Casimir associated to the $\Ad$-invariant, non-degenerate symmetric bilinear form used in the Chern-Simons action.
In our case, the relevant Casimir operator for the ‘gravitational’ bilinear form is \[casimir\] K = 16G (J\_aP\^a+P\_aJ\^a). A family of compatible $r$-matrices is given by [@Schroers; @MeusburgerSchroers6] \[rmatsols\] r= 32 G (P\_aJ\^a + \_[abc]{}n\^aJ\^bJ\^c) , n\_an\^a=-, where we use the metric to lower and contract indices.
Two comments are in order here. The first concerns the dependence of the solution on the real vector $\bn=(n^0,n^1,n^2)$ which has to satisfy the given constraint but is otherwise arbitrary. Thus, for $\lambda < 0$, the vector $\bn$ is any vector of length $\sqrt{-\lambda}$ in the Euclidean (hyperbolic) case, but is necessarily time-like in the Lorentzian (de Sitter) case. For $\lambda =0$, $\bn$ vanishes in the Euclidean case but may be any light-like vector in the Lorentzian case. For $\lambda >0$, $\bn$ is space-like in the Lorentzian (anti de-Sitter) case, while there is no real solution in the Euclidean case. However, the Euclidean case with $\lambda >0$ (and hence $\Lambda >0$) is the only case where the model space ($S^3$) and the local isometry group $SU(2)\times SU(2)$ are both compact, and the Chern-Simons theory is simply two copies of $SU(2)$ Chern-Simons theory, which is extensively studied in the literature, see [@EMSS] for an early paper. I will not say much about this case in the following, although it seems interesting and worthwhile to relate the many results about $SU(2)$ Chern-Simons theory to the framework discussed here, and to interpret them in terms of 3d gravity. Presumably this would involve using a complex vector $\bn$ and imposing a suitable reality condition after quantisation.
The second comment concerns the non-uniqueness of the solutions . These solutions all amount to equipping the Lie algebras $\glambda$ with the structure of a classical double, see [@CP; @K-S] for general background and [@MeusburgerSchroers6] for an explanation in the context of 3d gravity. However, other $r$-matrices are known, which are also compatible with the bilinear form but which do not belong to the family , see [@MeusburgerSchroers7] for examples and the forthcoming paper [@OseiSchroers2] for a systematic discussion. This gives rise to an ambiguity in the implementation of the Fock-Rosly prescription and the subsequent quantisation, but presumably leads to the same quantum theory. This issue has not been conclusively settled, and is also discussed in [@OseiSchroers2]. One advantage of working with the $r$-matrices associated to classical doubles is that one may quantise by going to the associated quantum double. This is what we will review in the next section.
The Fock-Rosly Poisson structure on $\tilde {\cal P}$ is determined in terms of a compatible $r$-matrix. The formulae for the brackets are explicit but lengthy, and we refer the reader to [@FockRosly] or [@AS] for details. Some understanding of it can be gained from the observation, made in [@AMII], that the Poisson brackets can be ‘decoupled’ after a suitable coordinate change, and that, as a symplectic manifold, $\tilde {\cal P}$ is isomorphic to a direct sum of $g$ copies of the Heisenberg double of the Poisson-Lie group $\Glambda$ (with the Sklyanin Poisson-Lie structure defined by $r$) and the manifolds $ {\cal C}_{\mu_i \sigma_i}$, $i=1,\ldots n$ viewed as symplectic leaves of the dual Poisson-Lie group $\Glambda^*$: \[extendeddecoupled\] () …() \_[\_n \_n]{}…[C]{}\_[\_1 \_1]{}. The general definitions of the Sklyanin, Heisenberg double and dual Poisson structures can be found in the paper [@AMII] and also in the textbook [@CP] or the lecture notes [@K-S]. We will give some further background in the next Section, but here we note that all of these structures for the family of groups $\Glambda$ with the $r$-matrices are explicitly given in in [@MeusburgerSchroers6]. For example, in the case of vanishing cosmological constant (and $\bn$ vanishing), one finds [@MeusburgerSchroers1; @MeusburgerSchroers2] $$\text{Hei}(SL({\mathbb{R}})\ltimes {\mathbb{R}}^3)\simeq T^*(SL(2,{\mathbb{R}})\times SL(2,{\mathbb{R}})).$$ In the Fock-Rosly description of the phase space one still needs to impose a constraint in $\tilde {\cal P}$, and take a quotient. We will not pursue this here since we are mainly interested in the quantum theory. Our approach to quantisation is to quantise $\tilde {\cal P}$ first, and then to take the quotient at the quantum level.
Quantum groups and 3d quantum gravity
=====================================
The combinatorial quantisation programme and associated quantum groups
----------------------------------------------------------------------
The task of constructing a quantum theory of 3d gravity in the Chern-Simons approach followed here is that of quantising the Poisson algebra of functions on the physical phase space , and of finding a unitary, irreducible representation (UIR) of the quantised algebra. By ‘quantisation’ of a Poisson manifold $M$ we mean, generally speaking, a deformation $F_h(M)$ of the algebra of functions on that manifold with a multiplication depending on a parameter $h$ in such a way that the commutator of two elements in $F_h(M)$ to first order in that parameter equals the Poisson bracket of the classical limit of those elements [@CP]. Details, for example the precise class of functions (${\cal C}^\infty$ or some algebraic subset), depend on the Poisson manifold in question.
In the combinatorial approach, one simplifies this task by first quantising the extended phase space , and then imposing the reduction to at the quantum level by a suitable condition on the Hilbert space carrying the UIR of the quantisation of . An important advantage of the combinatorial approach is that one really only needs to carry out the quantisation of the building blocks entering the decomposition of the extended phase space , and that these, in turn, can all be constructed from [*one*]{} quantum group $H$ and its representations.
The quantum group $H$ in question is the quantisation of the so-called dual Poisson-Lie group $\Glambda^*$ of $\Glambda$ (with the Sklyanin Poisson-Lie defined by the $r$-matrix ). This is explained in general terms in [@AGSI; @AGSII] and in the particular case of semi-direct products like the Euclidean or Poincaré groups in [@MeusburgerSchroers2]. It can be motivated as follows.
The dual Poisson-Lie group $\Glambda^*$ is a non-linear analogue of the Kirillov-Kostant-Souriau (KKS) Poisson structure on the dual $\glambda^*$ of the Lie algebra $\glambda$ [@AMI; @K-S]. Since the quantisation of the KKS structure on $\glambda^*$ is the universal enveloping algebra $U(\glambda)$, it is not surprising that the quantisation of the Poisson algebra of $\Glambda^*$ is a deformation of $U(\glambda)$. Thus we see already at this general level that the quantum groups $H$ are Hopf algebras obtained by deforming the local isometry groups $\Glambda$ (or more precisely, of their group algebras). We therefore refer to them as quantum isometry groups in the following. There is further similarity between the canonical Poisson structure on $\glambda^*$ and $\Glambda^*$: the symplectic leaves of the former are co-adjoint orbits while the symplectic leaves of the latter are conjugacy classes in $\Glambda$ [@CP; @K-S]. Given the non-degenerate bilinear pairing on $\glambda$, co-adjoint orbits may be thought of as adjoint orbits in $\glambda$, and conjugacy classes in $\Glambda$ may be thought of as non-linear deformations of these.
The irreducible representations of a Lie algebra can be obtained by quantising the KKS Poisson algebra and imposing the conditions which define the co-adjoint orbits in terms of suitable Casimir operators. This analogy, and the general comments of the previous paragraph, go some way to motivating the result that the quantisation of the conjugacy classes ${\cal C}_{\mu_i\sigma_i}$ in the decomposition gives UIRs $ V_{\mu_i\sigma_i}$ of the quantum group $H$ (with possible quantisation conditions on the labels $\mu_i,\sigma_i$). The quantisation of the classical Heisenberg double of $G_\lambda$ is the Heisenberg double of the Hopf algebra $H$ [@Majidbook1]. Its unique irreducible representation, in the cases where they have been studied, is a quantum group analogue of the regular representation of a group, and we therefore denote it by $\text{Reg}(H)$. We thus arrive at the following Hilbert space for the quantisation of the extended phase space \[exthilbert\] = (H)\^gV\_[\_n\_n]{}…V\_[\_1 \_1]{}. This space is, by construction, a (reducible) representation of the quantum group $H$. The Hilbert space for quantisation of the physical phase space is the invariant part under this $H$-action [@AGSI; @AGSII; @AS]: \[truehilbert\] [H]{} = \_H ().
In order to carry out the combinatorial quantisation programme in practice one needs to construct the quantum group $H$ and to find the representations appearing in . The construction of the quantum group $H$ is facilitated by the fact that the $r$-matrices equip $\glambda$ with the structure of a classical double of either $sl(2,{\mathbb{R}})$ (in the Lorentzian case) or $su(2)$ (in the Euclidean case) with suitable bialgebra structures, given in [@MeusburgerSchroers6]. Following the principle that the quantisation of the double is quantum double of the quantisation [@Drinfeld], the family of quantum groups $H$ can thus easily be found. We list them in Table \[qisogroups\], which should be seen as a quantised and ‘gravitised’ version of Table \[isogroups\] of the classical isometry groups. We will not give definitions or lists of generators and relations for any of these quantum groups here, but refer to the standard textbooks [@CP; @Majidbook1]. However, to gain some physical understanding it is worth noting that half the generators should be interpreted as rotation/boost generators and the other half as momentum generators. Thus, for example in the Lorentzian case of vanishing cosmological constant \[double\] D(U(su(1,1)))=U(su(1,1))(SU(1,1)), as an algebra, where ${\mathbb{C}}(SU(1,1))$ are complex-valued, smooth functions on $SU(1,1)$. The generators $J_a$ of $U(su(1,1)))$ are simply the rotation generator $J_0$ and the boost generators $J_1,J_2$ already encountered in , while elements of ${\mathbb{C}}(SU(1,1))$ should be thought of as functions or coordinates on the non-linear momentum space $SU(1,1)$, see [@sissatalk] for details and references, and also below for further remarks. Finally, the parameter $q$ appearing in the table is the one introduced at the beginning of this talk . It combines all four physical parameters entering quantum gravity with a cosmological constant.
16 pt
---------------- -------------------------------------- -------------------------------------------
Cos. constant Euclidean ($c^2<0)$ Lorentzian ($c^2>0$)
$\Lambda = 0$ $D(U(su(2)))$ $D(U(su(1,1)))$
$\Lambda > 0$ $ D(U_q(su(2)))$, $q$ root of unity $D(U_q(su(1,1)))$ $q\in {\mathbb{R}}$
$\Lambda < 0$ $D(U_q(su(2)))$, $q\in {\mathbb{R}}$ $D(U_q(sl(2,{\mathbb{R}})))$, $q\in U(1)$
---------------- -------------------------------------- -------------------------------------------
: Quantum isometry groups in 3d quantum gravity, $ q=e^{-\frac{\hbar G\sqrt{\Lambda}}{c}}$[]{data-label="qisogroups"}
The combinatorial quantisation programme has been carried out to various degrees of completeness in the different cases. For the Euclidean case with vanishing cosmological constant, the importance of the quantum double $D(U(su(2))$ was first pointed out in [@BM], and the proof that it plays the role of the quantum isometry group $H$ in the combinatorial approach to Euclidean quantum gravity without cosmological constant was given in [@Schroers]. The Lorentzian case was considered in [@BMS] and the general situation of Chern-Simons theory with certain semidirect product gauge groups was considered in [@MeusburgerSchroers2]. The situation where the classical gauge group is $SL(2,{\mathbb{C}})$ (i.e. Euclidean with $\Lambda <0$ or Lorentzian with $\Lambda >0$) was studied in [@BNR], with the relevant quantum group already constructed in [@PodlesWoronowicz]. The Euclidean case with $\Lambda >0$ is essentially the Turaev-Viro model. Finally, the very interesting Anti-de Sitter case (Lorentzian and $\Lambda >0$) has, unfortunately, not received much attention in the framework sketched here.
Non-commutative momentum addition, braiding and non-commutative spacetimes
--------------------------------------------------------------------------
Having constructed the quantum groups which control the construction of 3d quantum gravity according to the combinatorial scheme it is natural to ask what one can learn from them about the physics of 3d quantum gravity.
Formally, the role of the quantum isometry groups listed in Table \[qisogroups\] is strictly auxiliary. The physical Hilbert space is, by definition, invariant under the action of those quantum groups. Physical observables which act on this Hilbert space (see [@Meusburger6] for a discussion of classical examples) are not obviously related to the quantum isometry groups. As already mentioned (and discussed further in the Conclusion), the $r$-matrix used in the Fock-Rosly scheme, and hence the associated quantum group, is not uniquely determined. Both of these observations suggest that the quantum groups in Table \[qisogroups\] have only an indirect physical significance.
On the other hand, the quantum isometry groups, their representations and even their quantum $R$-matrices can be directly related to physical properties of particles in 3d quantum gravity. We will illustrate this for the case of vanishing cosmological constant. In that case, the quantum doubles appearing in Table \[qisogroups\] are quantum doubles of the Lie groups $SU(2)$ in the Euclidean case and $SU(1,1)$ (which is isomorphic to $SL(2,{\mathbb{R}})$) in the Lorentzian case. These quantum doubles are semi-direct products as algebras as shown in , and have a representation theory which is very similar to those of the Euclidean and Poincaré groups [@BM; @KM; @KBM]. The only difference is that the ‘mass shell’ in momentum space which characterises UIRs of the Euclidean and Poincaré group become conjugacy classes in the non-linear momentum spaces ($SU(2)$ in the Euclidean case and $SU(1,1)$ in the Lorentzian case). Physically, this means that momenta are no longer vectors but group elements of $SU(2)$ or $SU(1,1)$ and that momentum ‘addition’ is implemented by group multiplication in $SU(2)$ or $SU(1,1)$ instead of vector addition. These non-linear and non-commutative properties of momentum addition for gravitating particle reflect the use of holonomies for characterising particle properties, as y used in early papers on 3d gravity [@DJtH; @Carlipscattering]. We can even see it in the simplest non-trivial example of 3d spacetime, namely the cone shown in Fig. \[conedeflection\]. The spacetime is fully characterised by the deficit angle $\mu$, which is the mass of the particle in units of the Planck mass $1/8\pi G$. However, the angular nature of this parameter fits very well into the picture of $SU(1,1)$-valued momenta: we simply think of $\mu$ as a rotation, i.e. a particular element of $SU(1,1)$.
A closely related property of gravitating particles is their scattering, as analysed in some of the early papers on 3d quantum gravity [@DeserJackiw2; @tHooftscattering]. It turns out that the $S$-matrix for the scattering of two massive and spinning particles can also be interpreted in terms of quantum groups and the sort of topological interactions discussed in Sect. \[intro\]. As shown in [@BMS], the $S$-matrix is naturally related to the $R$-matrix of the quantum double $D(U(su(1,1)))$.
Finally, the curved and non-abelian nature of the momentum manifold suggests that naturally defined positions coordinates (which should generate translations on momentum space) should be non-commutative. One can argue this more formally by demanding that momentum and position algebras should be dual as Hopf algebras, leading to the family of Hopf algebras shown in Table \[qmompos\]. A particular, and much studied example is the ‘spin spacetime’ with generators $X_0,X_1,X_2$ and commutation relations =\_P \_[abc]{}X\^c, where $\ell_P= 8\pi \hbar G$ is the Planck length in 3d gravity, and both the Euclidean and Lorentzian interpretation apply. This non-commutativity of positions was already considered in [@tHooftdiscrete] and [@MatWell], and appears naturally in the quantum group theoretical framework considered here. It can also be derived in other approaches, namely in a path integral for particles where gravitational field degrees of freedom have been integrated out [@FreidelLivine] or in a coset construction [@JoungMouradNoui], which is analogous to the way the classical spacetimes can be obtained as homogeneous spaces of the classical isometry groups $\Glambda$. Finally, the role of the quantum double $D(SU(2))$ as a quantum isomtetry group of the 3d (Euclidean) was noted in [@BatistaMajid], where the latter was studied from the point of view of non-commutative differential geometry.
It is interesting that physical arguments, path integrals, coset constructions and general quantum group theoretical considerations all lead to the same non-commutative spacetimes. One way of exploring the physical significance of this non-commutativity is to study representations of the quantum doubles in Table \[qisogroups\] in position space. The requires Fourier-transforming the usual formulation of the representations in momentum space, in analogy to the way the UIRs of the Poincaré group can be Fourier transformed into the solution space of the familiar wave equations of relativistic physics (Klein-Gordon, Dirac, Maxwell etc). This was carried out for $D(SU(2))$ in [@MajidSchroers] and is considered for the Lorentzian case in [@SchroersWilhelm].
16 pt
---------------- -------------------------------------------------------------- --------------------------------------------------------------------------------
Cos. const. Euclidean ($c^2<0)$ Lorentzian ($c^2>0$)
$\Lambda = 0$ ${\mathbb{C}}(SU(2)) \;/\; U(su(2))$ ${\mathbb{C}}(SU(1,1)))\;/\; U(su(1,1))$
$\Lambda > 0$ $ {\mathbb{C}}_q(SU(2))\;/\; U_q(su(2))$, $q$ root of unity ${\mathbb{C}}_q(SU(1,1))\;/\; U_q(su(1,1))$ $q\in {\mathbb{R}}$
$\Lambda < 0$ ${\mathbb{C}}_q(SU(2))\;/\; U_q(su(2))$, $q\in {\mathbb{R}}$ ${\mathbb{C}}_q(SL(2,{\mathbb{R}}))\;/\; U_q(sl(2,{\mathbb{R}}))$, $q\in U(1)$
---------------- -------------------------------------------------------------- --------------------------------------------------------------------------------
: Momentum/position algebras in 3d quantum gravity, $ q=e^{-\frac{\hbar G\sqrt{\Lambda}}{c}}$[]{data-label="qmompos"}
Outlook and conclusion
======================
We have seen that the combinatorial quantisation of the Chern-Simons formulation of 3d gravity gives a unified picture of the various regimes of 3d gravity, with the physical parameters $c,\Lambda, G $ and $\hbar$ entering as deformation parameters in distinctive ways. Quantum groups naturally replace the classical isometry groups in this approach to 3d quantum gravity, and non-commutative spacetimes replace the classical model spacetimes. In general, the relation between the quantum isometry groups and the physical Hilbert space of 3d quantum gravity is a formal one, but we have seen that aspects of the quantum isometry groups like the non-commutative momentum addition and the braiding via the quantum $R$-matrix have a direct physical interpretation. It is worth noting that it is possible to take a Galilean limit $c\rightarrow \infty$ in the framework discussed here [@PapageorgiouSchroers1; @PapageorgiouSchroers2], and that the non-commutative quantum space is the Moyal plane in that case, with a time-dependent non-commutativity of the spatial coordinates.
In order to clarify the physical interpretation of quantum isometry groups and the associated non-commutative spacetimes it may be useful to consider universes with a boundary instead of the spatially compact universes considered in this talk. The treatment of boundaries in the classical theory is discussed in [@Matschull2; @MeusburgerSchroers4; @MeusburgerSchroers5] but a general treatment of the quantisation has not been given. Another approach would be to work directly on the physical phase space as in [@Meusburger6; @MeusburgerSchonfeld], and to attempt the quantisation there.
Other quantum groups than quantum doubles have been discussed in relation to 3d quantum gravity, notably bicrossproducts or $\kappa$-Poincaré algebras which were originally introduce in 4d [@LNRT; @LNR; @MajidRuegg]. As shown in [@MeusburgerSchroers7], the $\kappa$-Poincaré algebra with the usual time-like deformation parameter is [*not*]{} compatible with 3d gravity in the combinatorial framework. On the other hand, $\kappa$-Poincaré algebras with space-like deformation parameters are possible. This and other quantisation ambiguities of 3d quantum gravity are discussed in the forthcoming paper [@OseiSchroers2].
[99]{}
A. Staruszkiewicz, Gravitation theory in three-dimensional space, Acta Physica Polonica 6 (1963) 735-740.
G. Ponzano and T. Regge, Semiclassical limit of Racah coefficients, in Spectroscopic and group theoretical methods in physics, ed. F. Bloch, S. G. Cohen, A. De-Shalit, S. Sambursky and I. Talmi, North Holland, Amsterdam, 1968.
S. Deser, R. Jackiw and G. ’t Hooft, Three-dimensional Einstein gravity: dynamics of flat space, Ann. Phys. 152 (1984) 220-235.
S. Deser and R. Jackiw, Three dimensional cosmological gravity: dynamics of constant curvature, Ann. Phys. 153 (1984) 405-416.
S. Deser and R. Jackiw, Classical and quantum scattering on a cone, Commun. Math. Phys 118 (1988) 495-509.
G. ’t Hooft, Non-perturbative two-particle scattering amplitudes in 2+1 dimensional gravity, Commun. Math. Phys 117 (1988) 685-700.
S. Carlip, Exact quantum scattering in 2+1 dimensional gravity, Nucl. Phys. B324 (1989) 106-122.
F. A. Bais and N. M. Muller, Topological field theory and the quantum double of $SU(2)$, Nucl. Phys., B530 (1998), 349–400.
F. A. Bais, N. M. Muller and B. J. Schroers, Quantum group symmetry and particle scattering in (2+1)-dimensional quantum gravity, Nucl. Phys. B 640 (2002) 3–45.
A. Achucarro and P. Townsend, [A Chern–Simons action for three-dimensional anti-de Sitter supergravity theories]{}, [ Phys. Lett.]{} B [ 180]{} (1986) 85–100.
E. Witten, [2+1 dimensional gravity as an exactly soluble system]{}, [ Nucl. Phys.]{} B [311]{} (1988) 46–78.
V. G. Turaev and O. Y. Viro, State sum invariants of 3-manifolds and quantum 6j symbols, Topology 31 (1992) 865–902.
T. Foxton, Spin networks, Turaev-Viro theory and the loop representation, Class. Quantum Grav. 12 (1995) 951–964.
J. W. Barrett and I. N. Guzman, The Ponzano-Regge model, Class. Quant. Grav. 26 (2009) 155014.
S. Doplicher, K. Fredenhagen, J. E. Roberts, The Quantum structure of space-time at the Planck scale and quantum fields, Commun. Math. Phys. 172 (1995) 187–220.
E. Bianchi and C. Rovelli, A note on the geometrical interpretation of quantum groups and non-commutative spaces in gravity, arXiv:1105.1898 \[gr-qc\].
G. ’t Hooft, Quantisation of point particles in 2+1 dimensional gravity and space-time discreteness, Class. Quant. Grav. 13 (1996) 1023-1039. H. J. Matschull and M. Welling, Quantum mechanics of a point particle in (2+1)-dimensional gravity, Class. Quant. Grav. 15 (1998) 2981–3030.
E. Batista and S. Majid, Noncommutative geometry of angular momentum space $U(su_2)$, J. Math. Phys. 44 (2003) 107–137.
L. Freidel and E. R. Livine, Effective 3d quantum gravity and non-commutative quantum field theory. Phys. Rev. Lett. 96 (2006) 221301. E. Joung, J. Mourad, and K. Noui, Three dimensional quantum geometry and deformed symmetry J. Math. Phys. 50 (2009) 052503.
S. Carlip, [ Quantum gravity in 2+1 dimensions]{}, Cambridge University Press, Cambridge, 1998.
G. Mess, Lorentz spacetimes of constant curvature, preprint IHES/M/90/28, 1990.
L. Andersson, T. Barbot, R. Benedetti, F. Bonsante, W. M. Goldman, F. Labourie, K. P. Scannell and J. -M. Schlenker, Notes on a paper of Mess, Geometriae Dedicata 126 (2007), 47–70, see also arXiv:0706.0640.
R. Benedetti and F. Bonsante, Canonical Wick rotations in 3-dimensional gravity, Memoirs of the American Mathematical Society, Volume 198 (2009) Number 926, see also e-Print: math/0508485 \[math-dg\] and e-Print: math/0412470 \[math-dg\].
W. Thurston, Three-dimensional geometry and topology, Princeton Univesity Press, Princeton 1997.
V. V. Fock and A. A. Rosly, [Poisson structures on moduli of flat connections on Riemann surfaces and $r$-matrices]{}, [ ITEP preprint]{} (1992) [ 72-92]{} (see also [math.QA/9802054]{}).
A. Y. Alekseev, H. Grosse and V. Schomerus, [ Combinatorial quantization of the Hamiltonian Chern-Simons Theory,]{} Commun. Math. Phys. [ 172]{} (1995) 317–358.
A. Yu. Alekseev, H. Grosse and V. Schomerus, [ Combinatorial quantization of the Hamiltonian Chern-Simons Theory II,]{} Commun. Math. Phys. [ 174]{} (1995) 561–604.
A. Yu. Alekseev and V. Schomerus, [ Representation theory of Chern-Simons observables,]{} Duke Math. Journal [ 85]{} (1996) 447–510.
G. Papageorgiou and B. J. Schroers, A Chern-Simons approach to Galilean quantum gravity in 2+1 dimensions, JHEP11 (2009) 009.
C. Meusburger and B. J. Schroers, Quaternionic and Poisson-Lie structures in 3d gravity: the cosmological constant as deformation parameter, J. Math. Phys. 49 (2008) 083510.
R. W. Sharpe, Differential Geometry, Springer Verlag, New York, 1997.
P. Schaller and T. Strobl, Diffeomorphisms versus non abelian gauge transformations: an example of (1+1)-dimensional gravity, Phys. Lett. B337 (1994) 266–270. H. -J. Matschull, [On the relation between (2+1) Einstein gravity and Chern-Simons Theory]{}, [ Class. Quant. Grav.]{} [ 16]{} (1999) 2599–2609.
E. W. Mielke and P. Baekler, Topological Gauge Model Of Gravity With Torsion, Phys. Lett. A 156 (1991), 399.
V. Bonzom and E. R. Livine, A Immirzi-like parameter for 3d quantum gravity, Class. Quant .Grav. 25 (2008) 195024.
C. Meusburger and B. J. Schroers, Generalised Chern-Simons actions for 3d gravity and kappa-Poincare symmetry, Nucl. Phys. B 806 (2009) 462–488.
P. de Sousa Gerbert, On spin and (quantum) gravity in 2+1 dimensions, Nucl. Phys. B346 (1990) 440–472.
M. Atiyah, R. Bott, Yang-Mills equations over Riemann surfaces, [ Phil. Trans. R. Soc. London]{} [ A 308]{} (1982) 523.
M. F. Atiyah, The geometry and physics of knots, Cambridge University Press, Cambridge, [1990]{}.
K. Krasnov and J. -M. Schlenker, Minimal surfaces and particles in 3-manifolds, Geom. Dedicata 126 (2007) 187–254.
B. J. Schroers, Lessons from (2+1)-dimensional quantum gravity, Proceedings PoS (QG-Ph) 035 for workshop ”From Quantum to Emergent Gravity: Theory and Phenomenology“, Trieste 2007; see also arXiv:0710.5844 \[gr-qc\].
V. Chari and A. Pressley, A guide to quantum groups, Cambridge University Press, Cambridge, 1994.
S. Majid, Foundations of quantum group theory, Cambridge University Press, Cambridge 1995.
B. J. Schroers, [Combinatorial quantisation of Euclidean gravity in three dimensions]{}, in: N. P. Landsman, M. Pflaum, M. Schlichenmaier (Eds.), [ Quantization of singular symplectic quotients]{}, Birkhäuser, in: Progress in Mathematics, Vol. 198, 2001, 307–328; [ math.qa/0006228]{}.
S. Elizur, G. Moore, A. Schwimmer, and N. Seiberg, Remarks on the Canonical Quantization of the Chern-Simons-Witten Theory, Nucl. Phys. B326 (1989), 108.
P. K. Osei and B. J. Schroers, Classical $r$-matrices for the generalised Chern-Simons action of 3d gravity, in preparation.
A. Yu. Alekseev and A. Z. Malkin, [ Symplectic structure of the moduli space of flat connections on a Riemann surface]{}, Commun. Math. Phys. [ 169]{} (1995) 99–119.
Y. Kosmann-Schwarzbach, Lie Bialgebras, Poisson Lie groups, and Dressing Transformations, Lect. Notes Phys. 638 (2004) 107–173.
C. Meusburger and B. J. Schroers, Poisson structure and symmetry in the Chern-Simons formulation of (2+1)-dimensional gravity, [ Class. Quant. Grav.]{} [20]{} (2003) 2193–2233.
C. Meusburger and B. J. Schroers, The quantisation of Poisson structures arising in Chern-Simons theory with gauge group $G\ltimes \mathfrak{g}^*$, Adv. Theor. Math. Phys. 7 (2004) 1003–1043.
A. Y. Alekseev, A. Z. Malkin, [ Symplectic structures associated to Lie-Poisson groups]{}, Commun. Math. Phys.[ 162]{} (1994) 147–173.
V. G. Drinfeld, Quantum Groups, Proceedings of the ICM (1987), 798–820.
E. Buffenoir, K. Noui and P. Roche, Hamiltonian Quantization of Chern-Simons theory with $SL(2,{\mathbb{C}})$ Group, [Class. Quant. Grav.]{} [19]{} (2002) 4953–5016.
P. Podles and S. L. Woronowicz, Quantum deformation of Lorentz group, Commun. Math. Phys. 130 (1990) 381–431.
C. Meusburger, Cosmological measurements, time and observables in (2+1)-dimensional gravity, Class. Quant. Grav. 26 (2009) 055006.
T. Koornwinder, N. Muller, The quantum double of a (locally) compact group, Journal of Lie Theory 7 (1997) 101–120.
T. Koornwinder, N. Muller, F. A. Bais, Tensor product representations of the quantum double of a compact group, [ Commun. Math. Phys]{} [198]{} (1998) 157–186.
S. Majid and B. J. Schroers, q-deformation and semi-dualisation in 3d quantum gravity J. Phys. A 42 (2009) 425402.
B. J. Schroers and M. Wilhelm, Non-commutative Fourier transform and wave equations in 2+1 dimensional quantum gravity, in preparation.
G. Papageorgiou and B. J. Schroers, Galilean quantum gravity with cosmological constant and the extended q-Heisenberg algebra, JHEP11 (2010) 020.
C. Meusburger and B. J. Schroers, Phase space structure of Chern-Simons theory with a non-standard puncture, Nucl. Phys. B 738 (2006) 425–456.
C. Meusburger and B. J. Schroers, Boundary conditions and symplectic structure in the Chern-Simons formulation of 2+1 dimensional gravity, Class. Quantum Grav. 22 (2005) 3689–3724.
C. Meusburger and T. Schonfeld, Gauge fixing in (2+1)-gravity: Dirac bracket and spacetime geometry. Class. Quant. Grav. 28 (2011) 125008.
J. Lukierski, A. Nowicki, H. Ruegg and V. Tolstoi, $q$-deformation of Poincaré algebra, Phys. Lett. B 264 (1991) 331–338.
J. Lukierski, A. Nowicki and H. Ruegg, New quantum Poincaré algebra and $\kappa$-deformed field theory, Phys. Lett. B293 (1992) 344–352.
S. Majid and H. Ruegg, Bicrossproduct structure of the $\kappa$-Poincaré group and non-commutative geometry, Phys. Lett. B. 334 (1994) 348.
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this report I surveyed physics potential of the å option of a Linear Collider with the following questions in mind: [*What new discovery can be expected at a åcollider in addition to what will be learned at its ‘parent’ Linear Collider?* ]{} By taking account of the hard energy spectrum and polarization of colliding photons, produced by Compton back-scattering of laser light off incoming $e^-$ beams, we find that a åcollider is most powerful when new physics appears in the neutral spin-zero channel at an invariant mass below about 80% of the c.m. energy of the colliding system. If a light Higgs boson exists, its properties can be studied in detail, and if its heavier partners or a heavy Higgs boson exists in the above mass range, they may be discovered at a å collider. CP property of the scalar sector can be explored in detail by making use of linear polarization of the colliding photons, decay angular correlations of final state particles, and the pattern of interference with the Standard model amplitudes. A few comments are given for SUSY particle studies at a å collider, where a pair of charged spinless particles is produced in the $s$-wave near the threshold. Squark-onium may be discovered. An $e^\pm \gamma$ collision mode may measure the Higgs-$Z$-$\gamma$ coupling accurately and probe flavor oscillations in the slepton sector. As a general remark, all the Standard Model background simulation tools should be prepared in the helicity amplitude level, so that simulation can be performed for an arbitrary set of Stokes parameters of the incoming photon beams.'
address: 'KEK Theory Group, Tsukuba, Ibaraki 305-0801, Japan'
author:
- Kaoru Hagiwara
title: |
Higgs, SUSY and the Standard Model\
at $\gamma\gamma$ Colliders
---
KEK-TH-731\
10 November 2000\
Photon Linear Collider; Polarization; CP; Higgs; SUSY
Why do we need a Photon Linear Collider?
========================================
The photon linear collider (PLC) makes use of the hard energy spectrum of the photons produced by Compton backscattering of a high power laser light off the linac $e^-$ beam[@gkst83; @gkst84; @telnov90]. Therefore, a PLC should be considered as an option of a future Linear Collider. This observation naturally leads us to the following questions:
- [*What new discovery can we expect at a Photon Linear Collider in addition to what we will learn at its ‘parent’ Linear Collider?* ]{}
- [*Does the PLC option make a Linear Collider project more attractive?* ]{}
In this report, I try to find the answer to the above two questions.
I start my study by comparing the $\gamma\gamma$ channel with the $e^+e^-$ annihilation channel, which are both source of various new particles. All charged particles are pair produced in both channels, and neutral particles can be produced as $s$-channel resonances. There is, however, an important difference in the spin of the accessible $s$-channel resonances. The $e^+e^-$ annihilation channel cannot couple to a spin 0 resonance because of the electronic chirality conservation, whose breaking is suppressed by the tiny electron mass. The lowest spin of a particle that can be produced in the $e^+e^-$ annihilation channel should hence be 1. On the other hand, the $\gamma\gamma$ channel can couple to a spin 0 resonance, while it cannot couple to a spin 1 resonance due to spin statistics of the $J_z=0$ two-photon system[@landauyang]. It is this stunning difference between the two channels that makes a PLC complementary to $e^+e^-$ colliders. We can probe the scalar sector in the $s$-channel of the colliding two photons at a PLC, whereas it can be probed only in association with another particle production at $e^+e^-$ colliders. This possibility gives a PLC the unique potential of becoming the best observatory of the scalar sector, or the Higgs sector. We should take this opportunity very seriously, because the scalar sector is the least known sector of the Standard Model (SM), and because its detailed understanding is probably the most important key in our search for physics beyond the SM.
There are many excellent reports[@watanabe97; @watanabe99; @jikia99; @melles99] on the role of a PLC as the laboratory of the Higgs sector, and I will give only a few general remarks in section 2. In addition to the precision measurements of the two-photon decay partial width and branching fractions of the Higgs bosons, I would like to emphasize the importance of probing new interactions, including CP violating ones, in the scalar sector. PLC is particularly well tailored for studying the CP property of resonances and interactions, because the two $J_z=0$ two-photon initial states can form a CP-even and a CP-odd state, and they can be prepared by using linear polarizations of the laser beams. The power and limitation of PLC with linearly polarized laser beams are discussed. Once the Higgs property is determined well in $\gamma\gamma$ collisions, measurement of the Higgs-$\gamma$-$Z$ coupling may be done in the $e\gamma$ collision mode[@chois95].
The role of a PLC in our study of supersymmetry is discussed in section 3. Because the available highest c.m. energy in the $\gamma\gamma$ mode is about 80% of the original $e^+e^-$ collision c.m. energy, = < < 0.83, \[def:tau\] one generally expects that the $\epem$ mode is the discovery channel for those SUSY particles which may escape detection at the LHC. Here $x=4E_e w_o /m_e^2$ is the normalized laser frequency which is bounded from above, $x<2+2\sqrt{2}$, in order not to loose the photon beam by soft $\epem$ pair production. A PLC should therefore provide us with measurements which cannot be matched by the other experiments. Formation of squark-onia, precision measurements of SUSY particle properties, and measurements of CP violation in the chargino sector are examined. In the $\ea$ mode, single production of s-electron in association with a neutralino may turn out to be the best laboratory of s-lepton flavor physics. The mode can be the precision SUSY factory if s-electron pair can be produced.
In section 4, I give a few remarks on the role of a PLC in the study of the properties of the SM particles and their interactions. Such studies could be the main theme of high energy physics if neither the LHC nor $\epem$ collider fail to identify the physics beyond the SM. I also emphasize the importance of preparing all the SM background simulation programs in the helicity amplitude level so that they can be simulated for arbitrary polarization of the colliding photons.
Throughout this report, I use a very simple approximation to a PLC where the Compton scattering formula is used to generate colliding photons in the exactly backward direction. This description gives a good approximation only for the hard part of the photon beams[@ginzburg00; @cain]. In Fig. \[fig:aalumi\], I show the normalized effective luminosity function \[daalumi\] in this approximation for four values of the laser frequency parameter $x$, $x=2+2\sqrt{2}$, 1.6, 0.8 and 0.4 from the right to the left, and for three cases of the electron and laser polarizations, (a) to (c). Only the region where $\sqrt{\tau}/\sqrt{\tau}_\max > 0.75$ are shown, where the approximation may hold. In (a) and (b), circularly polarized laser beams ($P_c=\pm 1$) are used to make collisions of definite helicity photons, and the luminosity distribution is given for the collision of two right-handed photons $(++)$, denoted by solid lines, that of right- and left-handed photons $(+-)+(-+)$, denoted by long-dashed lines, and that of two left-handed photons $(--)$, denoted by short-dashed lines. In (a), both photons are obtained by setting $P_e P_c=-1$, whereas in (b) they are obtained by setting $P_e P_c=-P_e=-0.8$, as a more realistic case. Although the laser light may be 100% circularly polarized ($|P_c|=1$), the colliding beams have finite polarization. It is worth noting here that nearly optimal monochromaticity of the colliding photon polarizations is obtained with the electron beam polarization of $|P_e|=0.8$ which may well be realized. The curves in (c) are obtained for $P_e=0$ and $P_t=1$ when the two laser photons have parallel linear polarization. Here $P_t$ stands for the degree of linear polarization. Solid lines show collisions when two photons have parallel linear polarizations, and dashed lines are for perpendicular linear polarizations. The former two-photon state is CP-even while the latter state is CP-odd. We can clearly see that capability of distinguishing the two cases is small for high laser frequencies ($x$) or when $\sqrt{\tau}\simgt 0.6$.
Before closing this section, I should note that the actual luminosity function may depend strongly on the property of the incoming electron beams and on the electron-to-photon conversion efficiency. We may express &=& \^2 L\_[ee]{}\^ ( ), \[aalumi\]\
&=& L\_[ee]{}\^ ( ), \[ealumi\] where $\sqrt{\tau}=\sqrt{s_\aa }/\sqrt{s_{ee}}$ in Eq. (\[aalumi\]) while $\sqrt{\tau}=\sqrt{s_\ea }/\sqrt{s_{ee}}$ in Eq. (\[ealumi\]). $L_{ee}^\geom$ is the geometric luminosity of the colliding beams in the absence of beam-beam effects, and $\kappa$ denotes the conversion factor of order 0.5. Because the geometric luminosity can be significantly larger than actual collision luminosity in the or modes, there is a possibility that the luminosity integrated over the high $\sqrt{\tau}/\sqrt{\tau}_\max $ region shown in Fig. \[fig:aalumi\] can be comparable or even larger than the corresponding luminosity of collisions. When we compare physics capability of a PLC with its parent or LC, I assume that the integrated $\aa$ luminosity in the high $\sqrt{\tau}/\sqrt{\tau}_\max $ region is about the same as the luminosity of collisions.
It should be remarked here that the luminosity distributions as shown in Fig. 1 do not include contributions from the beamstrahlung. At a PLC, since the electron beams are tuned to maximize the åluminosity, more beamstrahlung may be produced than in the corresponding collisions. It is therefore important to estimate the effects of beamstrahlung in all quantitative studies, especially in the relatively low $\sqrt{\tau}$ region where we expect to have high degree of linear polarization transfer.
Neutral Higgs bosons
====================
A PLC associated with a 500 GeV LC will have the strongest case when a light Higgs boson exists below 200 GeV. Despite failures of discovering the Higgs boson so far, this is the most favored scenario of particle physics at the moment, because it is favored by the electroweak precision measurements[@lepewwg00], and because it is predicted by all supersymmetric theories with grand unification of the three gauge couplings[@susymh]. Such a Higgs boson may still be found at LEP200 or at Tevatron if its mass is very near to the present lower mass bound ($\sim 110\gev$) and if it has nearly maximum coupling to the $W$ and $Z$ bosons. It should certainly be discovered at LHC through gluon-gluon or $WW/ZZ$ fusion, unless its couplings to gluons and $WW/ZZ$ are both very small, or if it has little decay branching fractions into all the observable channels, such as å, $\tau^+\tau^-$ and $WW^*/ZZ^*$. It should be definitely discovered at an LC as long as it has a significant coupling to the $W$ and $Z$ bosons, the condition which is required for a good fit to the electroweak data[^1] and for the perturbative unification of the gauge couplings in SUSY models.
A 500 GeV LC will measure its couplings to $Z$ and $W$ bosons very accurately, and measure its decay branching fractions[@higgsatlc], Br$(bb)$, Br$(\tau^+\tau^-)$ and Br$(WW^*)$ with a good accuracy, as well as Br$(cc)$ and Br$(gg)$. The PLC will in addition measure the partial width $\Gamma(H\to\aa)$ at a few % level[@jikia99; @melles99]. Because the $H\aa$ coupling receives contributions from all the massive charged particles that couple to the Higgs boson, its accurate measurement will give us decisive information on the scalar sector when combined with accurate measurements in collisions. For instance, by using the couplings that are measured in collisions, one may estimate the Higgs coupling to the top-quarks and other new states by using the $\Gamma(H\to\aa)$ data.
I note in passing that the Higgs-$Z\gamma$ coupling can be most accurately measured in the $e\gamma$ collision mode of a PLC[@chois95]. It is clear that accurate knowledge of both the Higgs-$\aa$ and the Higgs-$Z\gamma$ couplings will be powerful in probing the quantum numbers of the charged particles whose masses originate from the electroweak symmetry breakdown.
In addition, I would like to emphasize the importance of the CP measurements in the scalar sector. Because the gauge interactions do not allow CP violation,[^2] non-gauge interactions should be responsible for the observed CP violation, and ultimately, for the origin of the matter dominated universe. The non-gauge interactions among scalar bosons and between scalars and fermions are among the most likely sources of CP violation, and precise measurements of the CP properties of the Higgs bosons and their interactions may open a completely new road in our investigation. The PLC will be an excellent laboratory for CP violation in the scalar sector, because it allows us to prepare the $J_z=0$ initial states with definite CP parity. By denoting the two colliding photon helicities as $\lambda_1$ and $\lambda_2$, the CP transformation changes their signs **[CP]{} \_1, \_2 > = -\_1, -\_2 > \[cpaa\] and hence the two states with definite CP parity are obtained as **[CP]{} ( ++ > – > ) = ( ++ > – > ). \[cpaa2\] The CP-even state has two linearly polarized photons with the polarization planes parallel, while the CP-odd state has perpendicular linear polarization directions. If CP is a good symmetry of the scalar sector, the CP-even Higgs boson can couple only to the CP-even initial state, whereas the CP-odd Higgs boson can couple only to the CP-odd initial state.****
In Fig. 1(c), I show the åluminosity when the two laser beams are both linearly polarized ($P_t=1$) along the same direction in the perfect backward scattering configuration. The Compton scattering with unpolarized beams then produces collisions of high energy linearly polarized photons which are partially CP-even (parallel, or $\parallel$) and partially CP-odd (perpendicular, or $\perp$). When the initial laser polarization planes are made to the perpendicular orientation, the luminosity functions of the CP-even and CP-odd configurations are reversed. We note that in this simple Compton scattering scheme, the difference between CP-even and CP-odd luminosity functions is significant only for relatively low laser frequencies ($x\simlt 1.6$) or at relatively low åinvariant mass, $\sqrt{\tau}\simlt 0.6$. The linear polarization of the PLC will hence be useful for studying the CP property of the neutral scalar boson whose mass is less than about 50% of the initial collision energy, $\sqrt{s}_{ee}$. Because of the necessity of relatively low $\sqrt{\tau}$ values to achieve high degree of linear polarization transfer, backgrounds from beamstrahlung photons should be estimated in quantitative studies.
It is worth noting here that it is an easy task for LC to distinguish between a CP-even and CP-odd neutral Higgs bosons. Such discrimination can e.g. be made by a simple angular correlation study in the process $\epem \to ZH$ followed by the decays $Z\to \ffbar$[@hs93; @kksz94; @hikk00]. What is difficult in the mode is to detect small CP violation effects in the study of dominantly CP-even Higgs bosons. The observable effects are expected to be rather small in collisions because the small CP-odd component can contribute to the process only in the one-loop order whereas the dominant CP-even component contributes at the tree-level. At a PLC, both components are expected to contribute in the one-loop order, and hence we can generally expect bigger CP asymmetries[@gunion94]. The use of the linearly polarized laser light allows us to make the precision CP measurement a counting experiment when the Higgs boson mass is less than about 50% of the collision energy[@choilee99]. Although we may need to rely more on the final state decay angular correlation studies for higher mass bosons, the advantage of larger CP asymmetry expected at a PLC will persist for all the neutral spin-less states that couple to the åchannel.
A PLC will be powerful in studying/discovering the neutral Higgs bosons (or its partners) which have suppressed couplings to the $Z$ and $W$ bosons[@aksw99]. Such states are expected in multiple Higgs doublet models including the SUSY-SM, and in fact their existence at or below the TeV scale makes the unification of the three gauge couplings possible in the minimal SUSY-SM[@mssmunif]. The precision electroweak experiments constrain the mass of the Higgs boson which has significant couplings to the $W$ and $Z$ bosons to be less than about 200 GeV, or else there should be subtle cancellation among new physics contributions. The degree of subtleness of this cancellation increases as the mass of the Higgs boson increases. Therefore it is most natural for us to expect that a light Higgs boson of mass in the range $100\sim 200~\gev$ has nearly the maximal couplings to the $W$ and $Z$ bosons, and hence its heavier partners do not have significant couplings to the weak bosons. Such states are difficult to produce singly at collisions, and they can be discovered at a PLC if their masses lie in the range $0.5\sqrt{s_{ee}}\simlt m_{\rm Higgs} \simlt 0.8\sqrt{s_{ee}}$.
As an example of how heavier Higgs bosons may be found at a PLC, I show in Fig. 2(a) the cross section of the process å\[aatt\] as a function of the invariant mass of the final state, $m(\ttbar)$. The cross section is calculated for a 500 GeV LC with the å luminosity function of Fig. 1(a) at the highest laser frequency ($x=4.83$). It should be noted that because of the high threshold ($2m_t/\sqrt{s_{ee}}\approx 0.7$), only the peak region of the åluminosity functions contributes where the purity of the collisions of right-handed photons ($++$) is high. The thick short dashed line shows the prediction of QED. The thick solid line (long-dashed line) shows the prediction where a CP-odd (CP-even) scalar boson of mass 400 GeV is produced as an $s$-channel resonance. For definiteness, we use the MSSM (minimal SUSY-SM) prediction for the total and partial widths of a 400 GeV CP-odd Higgs boson, A[^3]. As a comparison, predictions for the CP-even Higgs boson case are obtained simply by reversing the CP-parity of the state while keeping all the other properties. It is clearly seen that the interference pattern with the QED amplitude is very sensitive to the CP-parity of the resonance.
In fact the whole difference appears in the helicity amplitudes $\gamma_+ \gamma_+ \to t_L \bar{t}_L$ where both $t$ and $\bar{t}$ quarks are left-handed. The thick lines show the distributions when all the helicities are summed up, whereas the thin lines show those when the $t_L \bar{t}_L$ events (denoted as ‘LL’) are selected. Note, however, that even though the signal is clearer when the LL events are selected, only those events when one of the $W$’s decay leptonically can be used to distinguish the LL events from the dominant RR events. Still, about 40% of all the events may be used for such helicity analysis. It may be worth noting here that even the 6-jet events can be used to distinguish between the LL+RR modes and the LR+RL modes. In our example, because of suppressed $J_z=\pm 2$ åluminosity distribution in Fig. 1(a), very small fraction of all the pairs have the polarization LR+RL.
A more serious problem at a PLC is that the invariant mass of the colliding åsystem, $\sqrt{s_\aa }=m(\ttbar )$, can only be measured through the final top-quark momentum measurement. Accurate knowledge of the top-quark mass and their decay properties which will be obtained in the collision experiments will be used to refine such measurements. Fig. 2(b) shows the $m(\ttbar )$ distributions when a Gaussian smearing with the error m() = 3 \[mtterror\] is applied. Comparisons between the distributions of Fig. 2(a) and Fig. 2(b) show that the sharp peaks and dips in the original curves are smeared out and accurate knowledge of the measurement error, $\Delta m(\ttbar )$, will be needed to determine the mass and the widths of the Higgs resonances. The quoted error of 3 GeV may be too optimistic especially for semi-leptonic modes where at least one hard neutrino is missing. More work is needed to make the error as small as possible, for both 6-jet and semi-leptonic modes.
From Fig. 1(c), we find that linearly polarized laser beams produce little CP discriminating power at large $x$, or at large $\sqrt{\tau}$. When a scalar resonance is found in the mass range of $0.6 \simlt m_\res /\sqrt{s}_{ee} \simlt 0.83$, we should refer to the final state analysis to determine its CP property. As an example of CP-sensitive observables, I show in Fig. 2(c) the $m(\ttbar )$ dependences of $\left< \sin\Delta\phi \right>$, where $\Delta\phi$ stands for the difference of the azimuthal angles of the $t$-decay and $\bar{t}$-decay planes along the momentum axis. The thin lines are predictions before smearing and the thick lines after smearing with the error of Eq. (\[mtterror\]). The asymmetry has opposite signs between the CP-odd and the CP-even resonances, whose predictions are shown by the solid and the long-dashed curves, respectively. The asymmetry is found to be rather small because of cancellation between the contributions from longitudinally and transversally polarized $W$’s. One can certainly find final-state observables which have significantly higher sensitivity to the CP-properties of the resonance, by taking account of the $W$-decay distributions.
In summary, if the resonance mass is low enough, say $m_\res /\sqrt{s}_{ee} \simlt 0.6$, we can use the linear polarization of laser beams to determine its CP property. If the resonance has significant decay branching fraction into spinful heavy particles, such as or $W^+W^-/ZZ$, it is relatively easy to determine its CP property by making use of the decay angular correlations. Only when the mass is in the range $0.6 \simlt m_\res /\sqrt{s}_{ee} \simlt 0.83$ and when it rarely decays into heavy spinful particles, we will have difficulty in determining its CP property. The possibility of giving high degree of linear polarization to the colliding high-energy photons, that has been discussed at this workshop[@serbo00] may turn out to be useful in such cases. I also feel that more serious work may be needed to determine if the $\tau^+\tau^-$ decay mode of a heavy resonance can be used to study its spin and CP-parity at a PLC. In general, it is important to make CP measurements at all possible channels, so that we can probe CP-violation in the mixing, in the production, and in various decay channels[@achl00].
SUSY particles and charged Higgs bosons
=======================================
At a PLC, we can produce a pair of squarks, sleptons, charged Higgs bosons and charginos above the threshold. Because the åcollision at a PLC can reach the c.m. energy of at most about 80% of the corresponding collision energy, it is unlikely that these particles are discovered at a PLC. The question is what advantage does a PLC have over its parent LC when studying their properties.
In case of squarks, sleptons and charged Higgs bosons, we note that the pair can be produced in the $s$-wave at a PLC, whereas the pair can only be produced in the $p$-wave near threshold at collisions[^4]. This can lead to a higher production rate at a PLC depending on the factor of $\kappa^2 L_{ee}^\geom$ in Eq. (\[aalumi\]). It is possible that the production rate larger than that at the parent LC can be achieved at a PLC if the factor of $\kappa^2 L^\geom_{ee}$ can be made significantly larger than the actual luminosity. In such cases, precision measurements of the charged scalar boson properties may be performed at a PLC. In addition, in case of squarks, we may find the $s$-wave $J=0$ squark-onia at a PLC. There might be a case when the squark-pair production is sensitive to the Higgs-boson exchange between stop quarks[@hkmn90].
In case of chargino-pair production near the threshold, they are produced in the $s$-wave both in and åcollisions. The only difference is that the pair is in the spin-triplet state in case of collisions while it is in the spin-singlet state at a PLC. I do not know if this difference leads to a significant difference in the study of their properties. Because the $s$-wave spin-singlet state is CP-odd, a PLC may be useful in probing the CP property of the chargino-photon couplings. This however requires a PLC with relatively low laser frequencies (hence a higher energy), and the sensitivity should be compared with that of the final state analysis in the mode[@tsukamoto95; @choizerwas].
There is one point which might be worth noting. Both the sfermion-pair and chargino-pair production cross sections are uniquely determined by QED in the leading order of åcollisions. This uniqueness of the tree-level amplitudes may help us identify radiative effects. Certainly a combination of precision measurements of production cross sections at both and åcollisions will give us useful additional information on the interactions of the SUSY particles and the charged Higgs bosons.
Finally, the collision mode of a PLC may become a unique laboratory of slepton flavor physics[@feng96] if $\se_L$ or $\se_R$ ($\snu_{eL}$) can be produced in association with a neutralino (chargino). Flavor oscillation in the slepton sector can be most clearly studied in this channel where a SUSY particle state with a definite flavor and chilarity quantum numbers can be produced at the time of collisions.
Before concluding this section, let me comment on the possibility of precision SUSY tests in the process[@feng97] $e^-_\alpha e^-_\beta \to \se^-_\alpha \se^-_\beta$ for the three distinct channels, $\alpha\beta = LL$, $RR$ and $LR$. The process is most suited for precision measurements of the s-electron masses and their couplings to the gauginos, which may give us precious information on the heavy SUSY particle masses[@hikasa95; @nojirim96; @feng97]. Dedicated study of the collider option is worth serious attention, which requires studies independent of the beams which are optimized for the PLC option.
The Standard Model
==================
When we consider the SM processes as a probe of new physics, I think that a PLC can have an advantage over its parent LC when we study in detail the $J=0$ channel, by using the monochromaticity of high $\sqrt{\tau}$ region with high laser frequency, or when we study the CP property by using the linear polarization with relatively low laser frequency. Measurements of $W$ and top-quark EDM[@ch95] are examples of the latter. As for the $J=0$ channel, $ZZ$[@jikiazz], $W^+W^-$[@jikiaww] and $\ttbar$[@jikiatt] final states are all important because they should couple to the electroweak symmetry breaking sector to obtain their masses. That the electroweak precision measurements favor the SM Higgs boson of mass below about 200 GeV implies that if the Higgs boson (whose coupling to $W$ and $Z$ bosons are significant) is much heavier than 200 GeV or absent, there should be new physics that couple to $W$ and $Z$ bosons. Because it should affect the $W$ and $Z$ properties significantly, we should be able to identify its effect at LHC or at a lepton collider. The combined study of the $J=1$ channel at LC and the $J=0$ channel at a PLC may be most fruitful in the search of such new interactions.
I would like to note also that detailed studies of purely neutral gauge boson scattering processes, $\aa \to \aa$, $Z\gamma$ and $ZZ$, may give us useful information on new physics that affect these channels either in the tree-level or through radiative effects. The complete helicity amplitudes for all these mode in the SM are known[@gounaris99] and they should be useful in determining the properties of new physics that affect these processes.
The SM processes are also important for monitoring of the luminosity and the polarization of colliding photons, and also as backgrounds in new particle studies. Because we will need both circular and linear polarization of laser light, it is important for us to prepare all the SM simulation tools in the helicity amplitude level. All the SM processes should be generated for an arbitrary set of the Stokes parameters of the incoming photons. For instance, since massless leptons and quarks cannot be produced at large scattering angles from $J_z=0$ two photons in the lowest order of QED, $J_z=0$ luminosity functions may be monitored by using higher order processes, such as $l^+l^-\gamma$, $l^\pm \gamma (l^\mp)$ and $l^+l^-{l'}^+{l'}^-$, or by using $W^+W^-$[@watanabe92]. Hadron jet shape from $\aa \to \qqbar (g)$ processes may also be sensitive to the ratio of $J_z=0$ and $J_z=\pm2$ collisions because we expect $J_z=0$ photons to produce fatter jets. The $J_z=0$ luminosity function may be measured more efficiently by reversing the laser and electron polarizations simultaneously but in one side only, which leaves all the distributions the same while replacing the $J_z=0$ and $|J_z|=2$ distributions[@telnov93]. Linear polarization may be monitored by azimuthal angle distributions of high $p_T$ $l^+l^-$ events. All these studies should be done in the presence of realistic beamstrahlung backgrounds in order to estimate the monitoring errors.
Conclusions
===========
If a light Higgs boson of mass below 200 GeV is found, a PLC should be built in association with the first stage of an LC at about 500 GeV collision energy. Precision measurements of its åwidth and its CP property and the search for its high mass partners are the main targets of the PLC. If a Higgs boson is not found or found at a significantly higher mass, I think that a PLC will be most powerful when combined with the highest-energy LC. Studies of the processes $\aa \to ZZ, W^+W^-, \ttbar$ in the $J=0$ channel at highest åcollision energies with high laser frequencies, and precise measurements of $W$ and top-quark properties at low laser frequencies may be most fruitful in discovering new physics in such cases.
Acknowledgments {#acknowledgments .unnumbered}
===============
I wish to thank Eri Asakawa for valuable discussions and for preparing of the figures in this report. I learned a lot from Eri Asakawa, Seongyoul Choi and Jae-Sik Lee through collaborative research works. Discussions with Ilya Ginzburg, Thoru Takahashi, Valery Telnov, Isamu Watanabe and Kaoru Yokoya have been essential for me to appreciate the high physics potential of a PLC.
[999]{} I.Ginzburg, G.Kotkin, V.Serbo, V.Telnov, [*Nucl. Instr. Meth.*]{} [**205**]{} (1983) 47.
I.Ginzburg, G.Kotkin, V.Serbo, V.Telnov, [*Nucl. Instr. Meth.*]{} [**219**]{} (1984) 5.
V.Telnov, [*Nucl. Instr. Meth.*]{} [**A294**]{} (1990) 72.
L.F.Landau, [*Dok. Akad. Nauk USSR*]{} [**60**]{} (1948) 207; C.N.Yang, [*Phys. Rev.*]{} [**77**]{} (1950) 242.
I.Watanabe et al., “$\gamma\gamma$ collider as an option of JLC”, KEK Report 97-17 (March 1998).
I.Watanabe, talk at LCWS99, Sitges, Spain, 28 April–5 May, 1999.
G.Jikia S.Söldner-Rembold, hep-ph/9910366, [*Nucl. Phys. B(Proc. Suppl.)*]{} [**82**]{} (2000) 373.
M.Melles, hep-ph/9906467, [*Nucl. Phys. B(Proc. Suppl.)*]{} [**82**]{} (2000) 379.
S.Y.Choi, [*Z. Phys.*]{} [**C68**]{} (1995) 163; E.Gabrielli, V.A.Ilyin, B.Mele, hep-ph/9902362, [*Phys. Rev.*]{} [**D60**]{} (1999) 113005; hep-ph/9912321; I.F.Ginzburg, hep-ph/9907549, [*Nucl. Phys. B(Proc. Suppl.)*]{} [**82**]{} (2000) 367; I.F.Ginzburg, M.Krawczyk, P.Osland, hep-ph/9909455, talk at LCWS99, Sitges, Spain, 28 April–5 May, 1999; I.F.Ginzburg, I.P.Ivanov, hep-ph/0004069.
I.F.Ginzburg, G.L.Kotkin, hep-ph/9905462, [*Eur. Phys. J.*]{} [**C13**]{} (2000) 295.
K.Yokoya, CAIN, http://www-acc-theory.kek.jp/members/cain/default.html
A.Grutu, talk at ICHEP2000, Osaka, Japan, 27 July–2 August, 2000.
J.R.Espinosa, M.Quirós, [*Phys. Rev. Lett.*]{} [**81**]{} (1998) 516.
See e.g. S.Yamashita, talk at LCWS99, Sitges, Spain, 28 April–5 May, 1999.
K.Hagiwara, M.L.Stong, [*Z. Phys. C.*]{} [**62**]{} (1994) 99.
M.Kramer, J.Kühn, M.L.Stong, P.M.Zerwas, [*Z. Phys. C.*]{} [**64**]{} (1994) 21;
K.Hagiwara, S.Ishihara, J.Kamoshita, B.A.Kniehl, hep-ph/0002043, [*Eur. Phys. J.*]{} [**C14**]{} (2000) 457.
J.F.Gunion, J.G.Kelly, hep-ph/9404343, [*Phys. Lett.*]{} [**B333**]{} (1994) 110; B.Grzadkowski, J.F.Gunion, hep-ph/9501339, [*Phys. Lett.*]{} [**B350**]{} (1995) 218; G.J.Gounaris, G.P.Tsirigoti, hep-ph/9703446, [*Phys. Rev.*]{} [**D56**]{} (1997) 3030.
S.Y.Choi, J.S.Lee, hep-ph/9912330, [*Phys. Rev.*]{} [**D62**]{} (2000) 036005.
E.Asakawa, J.Kamoshita, A.Sugamoto, I.Watanabe, hep-ph/9912373, [*Eur. Phys. J.*]{} [**C14**]{} (2000) 335.
C.Giunti, C.W.Kim, U.W.Lee, [*Mod. Phys. Lett.*]{} [**A6**]{} (1991) 1745; U.Amaldi, W.de Boer, H.Fürstenau, [*Phys. Lett.*]{} [**B260**]{} (1991) 447; P.Langacker, M.Luo, [*Phys. Rev.*]{} [**D44**]{} (1991) 817; J.Ellis, S.Kelley, D.V.Nanopoulos, [*Phys. Lett.*]{} [**B260**]{} (1991) 131.
V.G.Serbo, talk at $\aa$2000, DESY, Germany, 14–17 Jun 2000.
E.Asakawa, S.Y.Choi, K.Hagiwara, J.S.Lee, hep-ph/0005313, to be published in [*Phys. Rev.*]{} [**D**]{}.
K.Hagiwara, K.Kato, A.D.Martin, C.K.Ng, [*Nucl. Phys.*]{} [**B344**]{} (1990) 1.
T.Tsukamoto, K.Fujii, H.Murayama, M.Yamaguchi, Y.Okada, [*Phys. Rev.*]{} [**D51**]{} (1995) 3153; J.L.Feng, M.E.Peskin, H.Murayama, X.Tata, [*Phys. Rev.*]{} [**D52**]{} (1995) 1418.
S.Y.Choi, A.Djouadi, H.Dreiner, J.Kalinowski, P.M.Zerwas, hep-ph/9806279, [*Eur. Phys. J.*]{} [**C7**]{} (1999) 123; S.Y.Choi, A.Djouadi, H.S.Song, P.M.Zerwas, hep-ph/9812236, [*Eur. Phys. J.*]{} [**C8**]{} (1999) 669; S.Y.Choi, M.Guchait, J.Kalinowski, P.M.Zerwas, hep-ph/0001175, [*Phys. Lett.*]{} [**B479**]{} (2000) 235; S.Y.Choi, A.Djouadi, M.Guchait, J.Kalinowski, H.S.Song, P.M.Zerwas, hep-ph/0002033, [*Eur. Phys. J.*]{} [**C14**]{} (2000) 535.
N.Arkani-Hamed, H.C.Cheng, J.L.Feng, L.J.Hall, hep-ph/9603431, [*Phys. Rev. Lett.*]{} [**77**]{} (1996) 1937; J.J.Cao, T.Han, X.Zhang, G.R.Lu, hep-ph/9808466, [*Phys. Rev.*]{} [**D59**]{} (1999) 095001.
H.-C.Cheng, J.L.Feng, N.Polonsky, hep-ph/9706438, [Phys. Rev.]{} [**D56**]{} (1997) 6875; hep-ph/9706476, [Phys. Rev.]{} [**D57**]{} (1998) 152; J.L.Feng, hep-ph/0002055, [*Int. J. Mod. Phys.*]{} [**A15**]{} (2000) 2355.
K.Hikasa, Y.Nakamura, hep-ph/9501382, [*Z. Phys.*]{} [**C70**]{} (1996) 139; Erratum [**C71**]{} (1996) 356.
M.M.Nojiri, K.Fujii, T.Tsukamoto, hep-ph/9606370, [*Phys. Rev.*]{} [**D54**]{} (1996) 6756; M.M.Nojiri, D.M.Pierce, Y.Yamada, hep-ph/9707244, [*Phys. Rev.*]{} [**D57**]{} (1998) 1539; S.Kiyoura, M.M.Nojiri, D.M.Pierce, Y.Yamada, hep-ph/9803210, [*Phys. Rev.*]{} [**D58**]{} (1998) 075002.
S.Y.Choi, K.Hagiwara, hep-ph/9506430, [*Phys. Lett.*]{} [**B359**]{} (1995) 369; S.Y.Choi, K.Hagiwara, M.S.Baek, hep-ph/9605334, [*Phys. Rev.*]{} [**D54**]{} (1996) 6703.
G.V.Jikia, [*Phys. Lett.*]{} [**B298**]{} (1993) 224; [*Nucl. Phys.*]{} [**B405**]{} (1993) 24.
G.Jikia, hep-ph/9612380, [*Nucl. Phys.*]{} [**B494**]{} (1997) 19.
B.Kamal, Z.Merebashvili, A.P.Contogouris, hep-ph/9503489, [*Phys. Rev.*]{} [**D51**]{} (1995) 4808; G.Jikia, A.Tkabladze, hep-ph/9601384, [*Phys. Rev.*]{} [**D54**]{} (1996) 2030; G.Jikia, A.Tkabladze, hep-ph/0004068.
G.Jikia, A.Tkabladze, hep-ph/9312228, [*Phys. Lett.*]{} [**B323**]{} (1994) 453; hep-ph/9312274, [*Phys. Lett.*]{} [**B332**]{} (1994) 441; G.J.Gounaris, P.I.Porfyriadis, F.M.Renard, hep-ph/9812378, [*Phys. Lett.*]{} [**B452**]{} (1999) 76; hep-ph/9902230, [*Eur. Phys. J.*]{} [**C9**]{} (1999) 673; G.J.Gounaris, J.Layssac, P.I.Porfyriadis, F.M.Renard, hep-ph/9904450, [*Eur. Phys. J.*]{} [**C10**]{} (1999) 499; hep-ph/9909243, [*Eur. Phys. J.*]{} [**C13**]{} (2000) 79.
Y.Yasui, I.Watanabe, J.Kodaira, I.Endo, hep-ph/9212312, [*Nucl. Instrum. Meth.*]{} [**A335**]{} (1993) 385.
V.Telnov, Proc. of LCWS93, Waikola, Hawaii, 26–30 April 1993 (World Scientific, 1993) p.551.
[^1]: Otherwise, there should be some sort of cancellation between the heavy Higgs boson and new physics contributions to the precisely measured electroweak parameters.
[^2]: Disregarding the CP-odd vacuum angle of QCD, whose effect is known to be negligibly small.
[^3]: $\Gamma_A=1.75~\gev$, Br$(A\to\ttbar)=0.95$, Br$(A\to\aa)=1.5\times 10^{-5}$ for $\tan\beta=3$, $m_\susy=1~\tev$, $M_2=500~\gev$ and $\mu=-500~\gev$ as chosen in ref. [@aksw99].
[^4]: The exception to this rule is the production of $\se_L^\pm \se_R^\mp$ pairs where the electronic chilarity of the initial channel is transfered to the final state. With proper choice of initial $e^\pm$ beam polarizations, these pairs can be produced at $s$-wave near the threshold.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'David Rolnick[^1]'
- 'Kevin Aydin[^2]'
- 'Shahab Kamali[^3]'
- 'Vahab Mirrokni[^4]'
- 'Amir Najmi[^5]'
bibliography:
- 'references.bib'
title: 'GeoCUTS: Geographic Clustering Using Travel Statistics'
---
Introduction {#sec:intro}
============
Large-scale online services routinely conduct live experimentation to improve their products. As described in [@diane-overlapping], browser cookies are the standard unit of analysis for experiments run by many web services. Typically, cookies are randomly selected into disjoint treatment and control groups and then subjected to different treatments (e.g. a different color background). The statistical and practical significance of metric differences between the two groups of cookies are important factors in the decision whether or not to launch the experimental treatment to all users.
While this cookie-based approach has been used extensively in the industry, it has some important limitations when measuring long-term effects on users. As detailed in [@henning], the fundamental problem is that
> “\[cookies\]...are a poor proxy for users. A cookie is simply an anonymous id attached to a browser and a device. Users can clear their cookies whenever they want, and they frequently use multiple devices and multiple browsers.”
The authors further observe that
> “Using a signed-in id may seem to mitigate this issue, but users can have multiple sign-ins and many queries are made while signed-out.”
Thus the effect measured by a cookie experiment is diluted by the fact that a user may be in the treatment group on one device or browser and in the control group in another.
A different approach to treating *whole users* in an experiment is to experiment on randomized geographical regions. This is a form of *cluster sampling*. Assuming a user remains within one such region, the user will receive consistent experimental treatment regardless of which device or browser she uses. The use of geo regions is common in advertising, with Nielsen’s DMA (Designated Market Area) regions being the best known. These have been used to run randomized *geo experiments* for measuring the effectiveness of internet advertising [@vaver].
DMAs have the advantage of being well-established as a means of subdividing a user population. However, they were created by Nielsen to correspond to television audiences in the US. Their partitioning is not necessarily optimal for internet usage. Furthermore, DMAs are restricted to the US without a direct international equivalent, and their granularity is fixed, with approximately 200 across the US.
We should also mention that the US Office of Management and Budget defines 381 Metropolitan Statistical Areas (MSAs) for the US. Each MSA is an economically integrated set of adjacent counties with at least one urban core area of 50,000. Since MSAs suffer from the limitations of DMAs noted above, we do not discuss them further in this paper.
The increase in mobile device usage poses an additional problem for the use of DMAs (and MSAs) as the unit of experimentation. This is because users can move into and out of a region. By doing so, they cause *leakage* between the treatment and control arms of the experiment. The growth of mobile devices has exacerbated the problem, though the same user can also use a single non-mobile device (e.g. laptop) in different geo regions or use two different fixed devices (e.g. one at work and one at home).
Ideally, geographic partitions should be based not on television viewership or population density but on the distribution and movement of users of the product in question. It would be useful to define these geo regions to strike a balance between the following design criteria:
1. Regions should be large enough to minimize leakage during the experiment (the probability of a user issuing queries from more than one region).
2. Regions should be numerous.
3. Regions should be relatively balanced in size.
\#1 above helps us ensure users in a geo region will receive consistent treatment during experimentation, while \#2 and \#3 allow us to make assumptions of exchangeability necessary to draw inferences from an experiment. Even though we do not require perfection, each of these requirements is necessary if geo regions are to be a better basis for experimentation than cookies.
While treating whole users is the primary motivation to develop a geo-based experimental framework, there is an important secondary motivation as well. When experiments are run with large, noticeable changes (e.g. a new feature or significant redesign of an existing feature) it is often the case that users influence each other by word of mouth. In such a situation, part of the treatment administered to a randomly assigned group of users may leak out to affect the control group according to the patterns of social networks. To fully account for these effects would require social network information and is beyond the scope of this paper; however, such networks tend to be biased towards geographically local connections [@backstrom2010find]. Thus we expect that our geo partitions, while designed to minimize splitting single users, are also somewhat well-suited to mitigate secondary effects from interaction between treatment and control groups.
In this paper, we present an algorithm, Geographic Clustering Using Travel Statistics (GeoCUTS) designed to optimize the three above criteria. We perform a comprehensive evaluation to study the impact of different design choices on GeoCUTS and also compare its performance against alternatives. We consider various massive-scale datasets to evaluate our algorithm under different circumstances, and consider both novel and existing metrics to compare GeoCUTS with alternative algorithms with respect to the balance and leakage between clusters. The results of our experiments suggest that the performance of GeoCUTS is equal to or surpasses that of hand-designed regions. Moreover, our automatic approach is much simpler and easier, provides extra flexibility in controlling the number of clusters, and extends to countries besides the US.
The contributions of this paper are as follows:
- Modeling user movement patterns as a graph using geographic information.
- Designing the GeoCUTS algorithm to partition this graph into regions.
- Describing a statistical framework and introducing metrics to evaluate the quality of the clusters.
- Showing through a comprehensive evaluation that GeoCUTS matches or outperforms the output of expert annotation while allowing flexibility essential to randomized geo experiments.
Prior results
=============
Many authors have considered the problem of causal inference on networks, where treatment is imposed at the level of nodes and interference is assumed to occur between nodes that are sufficiently close. For instance, in a social network, an individual’s response may be influenced both by their own treatment and by that of their friends or friends of friends. Ugander et al. [@ugander2013graph] and Gui et al. [@gui2015network] design low-variance estimators for these cases, relying upon clusters within the network. Eckles et al. [@eckles2014design] demonstrate the effectiveness of *network bucket testing*, randomly assigning treatments to different clusters, which gives a natural low-variance estimator. This approach is analyzed further by Backstrom and Kleinberg [@backstrom2011network] and by Katzir et al. [@katzir2012framework], who consider making clusters using weighted random walks on the network.
In this paper, we build upon existing literature for finding clusters efficiently within a large network. Here we focus on a variant known as [*balanced partitioning*]{}. In balanced partitioning, our goal is to find a set of clusters of almost equal size and to minimize the total weight of edges that cross clusters (i.e., minimize the cut). This is a challenging problem that is computationally hard even for medium-sized graphs [@AR06] as it captures the problem of graph bisection [@gareyjohnson].
Balanced graph partitioning is not only an NP-hard problem, no constant approximation algorithm is known for it and solving this problem even for graphs with hundreds of nodes is not easy. In fact, the logarithmic approximation algorithms for this problem are based on solving linear programming and semi-definite programming relaxations for these problems. These LP and SDP relaxations are hard to solve for graphs with thousands of nodes, and we aim to solve this problem for much larger graphs. As a result, we need to settle for heuristic algorithms that can be implemented in a distributed manner.
While the topic of large-scale balanced graph partitioning has attracted significant attention in the literature [@DGRW12; @DGRW11; @tsourakakis2014fennel; @ugander2013balanced; @stanton2012streaming], many prior authors have studied large-scale but non-distributed solutions to this problem. The need for distributed algorithms has been observed by several practical and theoretical research papers [@aydin2015distributed; @tsourakakis2014fennel; @ugander2013balanced]. Here we summarize some related work in this area.
Zhu and Ghahramani [@zhu2002learning] introduced the approach of label propagation. According to this method, clusters are initialized with known labels, after which these labels are propagated outward through the other nodes according to the transition probabilities of a Markov chain. The approach of label propagation was generalized to balanced clusters in the work of Ugander and Backstrom [@ugander2013balanced]. These authors formulated the swapping of nodes between clusters as a linear program which only depends on the number of clusters therefore allowing its implementation on very large graphs. More recently, Aydin et al. [@aydin2015distributed] achieve a scalable clustering in large networks by embedding the nodes along a line and use this embedding in future optimization steps. This approach has been proved to be effective for highly connected networks and expanders such as social networks [@aydin2015distributed].
Our work in this paper concentrates upon clustering with a geographically defined network, in contrast to the focus upon social networks. Recent work suggests that for graphs that have a geographic nature, applying the idea of “natural cuts” [@DGRW12; @DGRW11] is effective in solving the balanced partitioning problem. Inspired by this heuristic, we develop a distributed implementation, which we detail further in Section \[graphclustering\].
Algorithm {#sec:algorithm}
=========
We now present the GeoCUTS algorithm. The input is a set of locations from which a single user has issued queries. The algorithm proceeds in two phases. In Phase 1, we build a graph from the given data by discretizing locations and assigning edges between nodes for which there is frequent transit (see Algorithm \[alg:graph-building\]). This algorithm is designed to be run massively in parallel. Each step in the computation can be distributed among individual vertices simply with knowledge of their neighborhood. In Phase 2, we find a clustering of this graph by applying a geographic clustering algorithm that combines recently developed techniques.
Phase 1: Graph building
-----------------------
.1 in
**Discrete locations.** The first step in building our graph is location discretization. While given user locations are assumed to be on a continuum, in real-world settings accuracy is finite. Therefore, we round each location to the nearest gridpoint in a lattice, where the width of the lattice may be specified; a coarser lattice yields a faster but less precise algorithm. We have chosen to discretize to a grid, rather than for instance to the nearest city, as described in Ugander and Backstrom [@ugander2013balanced]. The motivation is that a grid has a natural geometrical structure that we will utilize in our algorithm, and in addition our method is applicable to location data from users in rural areas or along transit corridors. .075 in **Node weights.** We define a graph for which the nodes are the gridpoints above; in the following discussion we shall identify each node with the corresponding range of user locations. The weight of a node is a measure of the number of user visits to that location; specifically, if a user visits node $A$ a total of $a$ times, then the user’s influence upon $A$ is $\sqrt{a}$: $$\text{weight}(A) = \sum_{\text{user}\, u} \sqrt{\text{\# visits $u$ to $A$}}.$$We use $\sqrt{a}$ as a slight normalization. Alternatively, it would be possible to increment the node’s influence by $a$ itself; however, some individual users may have much more location data available than others, and this would bias the graph strongly towards these users to the exclusion of others. At the opposite extreme, one might assign a fixed influence to each user and distribute this proportionally among the weights of the nodes visited; however, this would bias the graph towards users for whom little location data is known, increasing the susceptibility to noise. .075 in **Edge weights.** The edges in this graph correspond to the intensity of transit between nodes. Specifically, if a user visits node $A$ a total of $a$ times, and node $B$ a total of $b$ times, then the user’s influence upon edge $AB$ shall be $\sqrt{ab}$: $$\text{weight}(AB) = \sum_{\text{user}\, u} \sqrt{(\text{\# visits $u$ to $A$})\cdot (\text{\# visits $u$ to $B$})}.$$We use the geometric mean to increment edge weights since it is minimized when either endpoint is visited seldom, and is maximized when the endpoints are visited equally. .075 in **Sparsity.** We retain only certain edges within the graph, specifically those for which the geographical distance between the two nodes is less than a given parameter. The reason for this trimming of edges is twofold: Firstly, it greatly reduces the size of the graph, allowing the number of edges to be linear, instead of at worst quadratic, in the number of nodes. Secondly, edges between nodes at great geographical distance can prove actively counterproductive to the algorithm. This is because the majority of user transit is geographically local, and therefore the desired output of the algorithm consists of clusters that are largely geographically connected. While there may be a relatively large amount of transit between, say, San Francisco and New York City, it would be a suboptimal to assign these cities to the same cluster. Therefore, any edge weight assigned to such long-distance connections would merely add noise to the objective function of our algorithm and make it more challenging to find the optimal clustering. .075 in **Normalization.** As a final and critical step, we renormalize the weights in the graph. We consider the cases of renormalizing the weight $x$ to $\sqrt{x}$ and to $\log(x)$. As shall be discussed further in Section \[sec:results\], we observed best results after log-normalizing both node and edge weights, though the normalization of nodes was most critical. Our motivation is that we are more concerned with finding a natural clustering than with the exact balance of the pieces. Our clustering algorithm is designed to balance within hard limits, which would generally yield clusters constrained by size rather than graphical structure. For example, New York City, which contains many users, might be divided into many small clusters, irrespective of user traffic between them. By normalizing node weights, we ensure a condition of “soft balance” in which natural clusters are allowed to vary in size with a preference for balance.
Phase 2: Graph clustering {#graphclustering}
-------------------------
As a crucial first step for several graph mining algorithms, graph clustering is an active research area and numerous algorithms have been developed for this problem. Since our objective is to minimize [*leakage*]{} (that is, the interaction between clusters) while maintaining clusters with roughly similar size, we have chosen an algorithm that solves the [*balanced partitioning problem*]{}. While this problem is computationally hard even for small instances, we seek to solve this problem in a distributed manner for large-scale graphs. Next, we formally define the balanced partitioning problem, and then present our distributed algorithms for this problem.
Consider a number $k$, and a graph $G(V, E)$ of $n$ vertices with edge lengths (or edge weights), and node weights, and a real number $\alpha \geq 0$. We denote the edge lengths by $\ell : E \rightarrow R$, and the node weights by $w : V \rightarrow R$. Let $w(S)$ for any subset $S$ of nodes be the total weight of nodes in $S$, i.e., $w(S) = \sum_{i\in S} w(i)$. A partition of nodes of $G$ into $k$ parts $\{V_i: i\in [k]\}$ is said to be $\alpha$-balanced if and only if $(1-\alpha)
w(V) / k \leq w(V_i) \leq (1+\alpha)
w(V) / k$. In particular, a zero-balanced (or fully balanced) partition is one where all partitions have the same weight. The weight of the cut for this partitioning is the total sum of all edges whose endpoints fall in different parts: $$\displaystyle\sum_{\substack{i \\ j<i}}\, \sum_{\substack{u\in V_i,v\in V_j:\\ (u,v)
\in E(G)}} \ell(u,v).$$
Our goal in the balanced partitioning problem is to find an $\alpha$-balanced partitioning whose cut size is (approximately) minimized. In most cases, we are not given a specific $\alpha$ as input, and we aim to find a partitioning that is as balanced as possible, i.e., with minimum $\alpha$. .075 in
[**Partitioning large scale geographic graphs.**]{} As our main contribution in this section, we present a distributed implementation of balanced partitioning by combining the idea of natural cuts [@DGRW12; @DGRW11] and linearly embedding the nodes of a geo graph [@aydin2015distributed; @article:MJFS01:analysis-hilbert; @book:global-opt-parallel]. In particular we use Hilbert curve embedding of geo graphs due to its nice space-filling properties [@article:MJFS01:analysis-hilbert]. Following the structure of algorithms based on natural cuts [@DGRW12; @DGRW11], our distributed algorithm iterates on a contraction stage and generates the output by applying a post-processing contraction stage, in which we [*merge*]{} parts of the contracted graph to compute the final partitioning.
.075 in [**Contraction stage.**]{} Inspired by prior work [@DGRW11; @DGRW12], we propose and implement a distributed version of the contraction stage of balanced partitioning based on natural cuts. This stage consists of a sequence of four main steps: (i) identifying a seed set, (ii) finding a natural cut around each seed node, (iii) computing connected components of the graph after removing edges of the natural cuts, and (iv) finally contracting nodes in each connected component to one node and computing an updated smaller graph with new node weights and edge weights. The main feature of our distributed implementation is that all of the natural cuts are computed in parallel, and the graph contraction based on these cuts also happens in a distributed manner. More specifically, the contraction stage of the algorithm is as follows:
1. [**Identify a seed set.**]{} Here, the goal is to identify a set of $k$ seed nodes $S$ from which we compute natural cuts. We follow the following strategy for computing this set $S$: embed nodes of the graph into a line using the Hilbert curve, and divide the line into $k$ pieces, each with almost the same number of nodes. Then, output a node close to the center of each piece of the Hilbert curve embedding. This method ensures that the set of seed nodes are spread uniformly across different parts, and thus natural cuts cover different parts of the graph.
2. [**Parallel computation of natural cuts around seed nodes.**]{} After selecting the seed set, we compute a “natural cut” around each seed node in parallel. Consider a seed node $v$, and let $Q~= \frac{w(V)}{k}$, i.e., $Q$ is the maximum weight of a cluster in an $\alpha$-balanced partitioning. The idea is first to compute a core $C(v)$ around node $v$ by performing a BFS around node $v$ until we cover $Q/10$ nodes. We contract $C(v)$ to a node $s$. Then we continue the BFS until the total size of the neighborhood reaches $Q$ and form a graph $G'(v)$ around node $v$. We take the rest of the graph $G\backslash G'(v)$ and contract it to one node, denoted by $t$. Finally, we compute a minimum $(s,t)$-cut in this graph $G'(v)$. We call this $(s,t)$-cut a [*natural cut*]{} around seed node $v$. A desirable property of this cut is that it has less than $Q$ total weight on its nodes, and can be used as a building block for computing parts of a balanced partitioning. Note that we compute all these natural cuts in parallel in a distributed manner by applying a MapReduce framework, uploading the graph in a distributed hash-table service, and accessing the neighborhood of nodes via a read-only service [@aydin2015distributed; @KiverisLMRV14].
3. [**Distributed connected components.**]{} As the next part of the contraction stage, we remove all edges of the graph that appear in at least one of the natural cuts computed around the seed nodes, and then compute connected components in the remaining graph (after removing those edges). We apply a distributed implementation of connected components that employs a distributed hash-table service, and that has been shown to be effective and scalable in practice [@KiverisLMRV14; @RastogiMCS13].
4. [**Contracting nodes of each connected component.**]{} After computing connected components, we can easily construct a contracted graph as follows: we put a node $u_i$ for each connected component $T_i$ with node weight $w(T_i)$, i.e., the sum of the weight of nodes in $T_i$. We also set the weight of the edge between two nodes $u_i$ and $u_j$ to $w(u_i, u_j)$.
[**Merging stage.**]{} If the size of the contracted graph is large, we iterate on the contraction stage until the size of the graph is small enough to fit in memory.[^6] When the contracted graph fits in memory, we can produce an output by applying any in-memory heuristic for balanced partitioning of graphs with node weights. The algorithm that we employ at this stage is similar to the greedy assembly algorithm proposed in [@DGRW12; @DGRW11]. We refer for details of this algorithm to [@DGRW12; @DGRW11].
Statistical evaluation {#sec:stats}
======================
In order to evaluate the quality of the clusters, we define a set of metrics to measure how well a given partitioning supports the purpose of experimentation. The goal in designing any experimental methodology is to maximize statistical power. While the optimal partitioning may depend on the treatment assigned, we require our clusters to support a class of experiments, the effects of which are unknown in advance. Our metrics are thus designed to be treatment-independent and based solely on the structure of the partitions.
Cluster quality metric for leakage
----------------------------------
One of our chief considerations in partitioning is leakage. The reason is that the fidelity of a geo region as a unit of experimentation depends upon it. If we experiment on a leaky region, users therein will receive a diluted dose of treatment. Let $a_{ij}$ represent the number of queries performed by User $i$ in Region $j$. If we treat only Cluster $k$, the fraction of treatment received by User $i$ is given by $$\frac{a_{ik}}{\sum_j a_{ij}}.$$ We do not know precisely how much a user will respond to partial treatment; no doubt it depends on the type of treatment. A moderate assumption is that the user will respond in a manner prorated to the treatment she receives. That is, if Cluster $k$ receives one unit of treatment (assume for now it is the only cluster being treated), then User $i$ will respond with $\frac{a_{ik}}{\sum_j a_{ij}}$ units of response.
Of course, our goal is not to have to measure the response of individual users. Instead, we wish to treat and measure the response of whole clusters. The response measured in Cluster $k$ is simply the weighted average of all user responses in that cluster, the weight being the fraction of mass or activity the user comprises in the cluster. This leads to an overall response for Cluster $k$ being $$\sum_i \left( \frac{a_{ik}}{\sum_\ell a_{\ell k}} \right)
\left( \frac{a_{ik}}{\sum_j a_{ij}} \right) \\
= \sum_i \frac{a_{ik}^2}{(a_{\cdot k})(a_{i \cdot})} = Q_k,$$ where $a_{i \cdot}$ is the total number of queries issued by User $i$ across all clusters and $a_{\cdot k}$ is the number of queries in Cluster $k$ issued by all users. This is our definition of $Q_k$, the quality or *Q-metric* of Cluster $k$. It is the prorated response we would measure in Cluster $k$ were we to apply to it one unit of treatment (and to no other cluster).
In reality, an experiment is likely to treat multiple clusters. Let $p$ represent the fraction of clusters treated by the experiment. A user may therefore receive treatment outside of Cluster $k$. Furthermore this reduction in treatment leakage out of Cluster $k$ also contributes to treatment leakage into control clusters. The result is that the average measured difference between response of a treatment cluster $T$ and a control cluster $C$ is given by the convex combination $$(1-p)Q_T + pQ_C.$$ Thus the average experimental effect we measure using clusters as our randomized experimental unit is proportional to the average “quality” $\bar{Q}$ of the clusters. In this sense, $\bar{Q}$ measures the average attenuation of any experimental effect we induce on a given cluster. And this is proportional to the statistical power of our experiments.
The case for why the average Q-metric is meaningful in running experiments assumes a simple prorated response by users. Such a linear dose-response curve is not always justified. For many treatments, we expect no response at all until the dose goes above a threshold, or perhaps a sigmoid dose-response curve. Without knowledge of the specific treatment we cannot improve our model of how user response is attenuated due to leakage. However, we may choose the conservative route of only using clusters with very high Q-metric as experimental units. This approach trades some statistical power for robustness, and it is this approach that we will take in analyzing many of our experiments.
Cluster balance metric
----------------------
While it is certainly possible to experiment with units of different size, there are practical reasons to wish for greater balance. One is that when assigning each cluster to a randomized treatment group with probability $p$, we hope to have somewhere near fraction $p$ of the population in treatment. This is unlikely to occur exactly if the clusters are of unequal size. Let $X_i$ represent the binary variable for assignment of Cluster $i$ to treatment where $\Pr(X_i=1)=p$. Let $S$ be the size of the chosen treatment group. The expectation of $S$ is $\mathbb{E}\left(\sum_i w_i X_i\right) = p$. However if we assign cluster to treatment independently $$\mathrm{var\ } S = \sum_i w_i^2 \mathrm{var\ } X_i = \sum_i w_i^2 p(1-p) = p(1-p)|w|^2$$ where $|w|^2$ is the $\ell^2$ norm of the cluster weights. Thus the average degree of imbalance is proportional to $|w|^2$. We can effect greater balance if we choose exactly $k$ clusters for treatment out of a total of $n$ clusters where $p=k/n$. In this case, $$\mathrm{var\ } S = \frac{k}{n} \left(\frac{n-k}{n-1} \right)
\left(|w|^2 - \frac{1}{n} \right)
\approx p(1-p) \left( |w|^2 - \frac{1}{n} \right),$$ the approximation being for $k \ll n$. Again, the $\ell^2$ norm of cluster weights relates to imbalance in selecting the treatment group. This imbalance is minimized when every cluster has the same size. We refer to the quantity $\left( |w|^2 - \frac{1}{n} \right)$ as the *B-metric* of the clustering. It is equal to 0 if the clustering is perfectly balanced and is otherwise positive, with greater size indicating greater imbalance.
Effective number of clusters
----------------------------
As mentioned in the introduction, the goal of our clustering is to maximize experimental power. However this is not possible to do without some knowledge of the experimental effect being administered. For instance, if the clusters grouped together users whose responses to the experiment were uncorrelated, the size and number of clusters would not affect estimation variance, and hence nor experimental power. To see this, suppose we measure $\hat{\theta}$ as the average effect across a set of clusters. If $w_i$ is the fraction of all users in Cluster $i$ then under uncorrelation $\mathrm{var\ } \hat{\theta_i} \propto \frac{1}{w_i}$. Thus,
$$\begin{aligned}
\hat{\theta} &= \sum_i w_i\hat{\theta_i}\\
\mathrm{var\ } \hat{\theta} &= \sum_i w_i^2 \mathrm{var\ } \hat{\theta_i} = \sum_i w_i^2 \frac{\sigma^2}{w_i} = \sigma^2,\end{aligned}$$
where $\sigma^2$ is a constant. Of course we cannot assume that the users in any cluster will have uncorrelated responses so this is just one extreme. On the opposite extreme, if every user within a cluster behaves identically then $\mathrm{var\ } \hat{\theta_i}$ is constant, say $\sigma^2$. In this case, $$\mathrm{var\ } \hat{\theta} = \sum_i w_i^2 \mathrm{var\ } \hat{\theta_i} = \sum_i w_i^2 \sigma^2 = \sigma^2 |w|^2.$$ We require knowledge of how correlation within the cluster varies with cluster size to determine the effect of clustering on statistical power, and this is not available without knowledge of the experimental treatment. We could take $|w|^2$ as proportional to estimation variance in the worst case. This could be exceedingly conservative, but the argument emphasizes the utility of measuring $|w|^2$ through the B-metric.
Empirical Results {#sec:results}
=================
In this section we evaluate our algorithm and compare it against alternative algorithms and baselines.
The dataset {#subsec:data}
-----------
![The graph built by the GeoCUTS algorithm, with nodes shown on corresponding locations of the US. White represents large edge weights (high traffic areas), while black represents low edge weights. Larger edge weights often do not match larger vertex weights, showing the difference between GeoCUTS and an algorithm simply measuring population density. Gaps in the colored regions represent locations for which no data is available. For example, in uninhabited regions such as deserts, data points come disproportionately from narrow strips corresponding to major roads.[]{data-label="fig:us-graph"}](us-graph.png)
A good partitioning algorithm minimizes the leakage introduced by the movement of users while keeping clusters roughly balanced. Hence, identifying movement trends is crucial in our algorithm. We wish to use clusters to run randomized experiments on user queries, hence a natural choice for constructing movement patterns is to use the approximate location of queries.
Our dataset consists of an anonymized set of randomly selected cookies. Each cookie is represented with a hashed value. The available data is an approximate location specified as a bounding box that covers the location at which a query was issued. The size of the bounding boxes is not uniform but all of them are large enough to contain locations of queries issued by a large number of distinct cookies. In some cases the boxes are so large that they cover an entire country or a continent. We remove such extreme cases from our dataset. Figure \[fig:us-boxes\] shows the distribution of queries from the United States in terms of the size of their bounding boxes. This distribution varies markedly between countries and regions and tends to be skewed towards larger boxes in rural areas, where geo-location is less accurate and where the smaller number of queries also makes it necessary for bounding boxes to be larger in order to ensure anonymization.
![The GeoCUTS algorithm applied to user queries from the United States. The algorithm automatically identifies metropolitan areas, correctly predicting, for example, that the Bay Area includes San Francisco, Berkeley, and Palo Alto, but not Sacramento.[]{data-label="fig:us-clusters"}](us_lognorm_025_highactive_clusters.jpg)
To build a graph, we form a grid on the geographical area we wish to partition. Each graph node corresponds to a grid cell. We establish edges between two nodes based on the number of movements between their corresponding cells, that is, when the same cookie issues queries in both cells. Hence, for each query we only need to identify the cell it is issued from; for this purpose, we assume that the unknown query location is issued at the center of the known bounding box. To ensure that inaccuracies in estimating positions do not negatively impact our algorithm, the grid cells should be large relative to the typical sizes of the bounding boxes. We will take this assumption into account when we discuss the granularity of the grid cells.
\[uspartitions\] ![The distribution of sizes of bounding boxes for queries within the United States. Most bounding boxes are small, with an area of no more than 10 square miles, while about 4% have an area of 100 square miles or greater.[]{data-label="fig:us-boxes"}](us-boundingbox-distribution.png "fig:")
We build the graph based on 28 days of cookie movements between grid cells with respect to user queries. The approximate locations of a very large number of queries are considered to model movement trends. Due to the scale of our datasets, while we have a very rough position for each individual cookie’s queries, the aggregate movement patterns are quite accurate.
In the rest of this section, we evaluate the performance of the GeoCUTS algorithm on different datasets, for each of which a separate graph was built and clustered. Figures \[fig:us-clusters\] and \[fig:fr-clusters\] show the clusters generated for data from the United States and France, respectively, and Figure \[fig:us-graph\] shows the GeoCUTS graph before clustering.
![The GeoCUTS algorithm applied to user queries from France. It correctly identifies metropolitan areas such as Paris, Bordeaux, and Lyon, and regions such as Alsace and Normandy.[]{data-label="fig:fr-clusters"}](fr_highlyactive_clusters.jpg)
Mobility {#subsec:mobility}
--------
Cookies that do not move are not interesting for our problem. In an extreme scenario where all cookies are stationary and issue queries from one location only, any arbitrary clustering algorithm performs perfectly well in terms of leakage. Hence for our first dataset we consider only *highly mobile* cookies: cookies that issue queries in at least two different cells of the grid.
We also note that some cookies are churned rather shortly after they are created. In practice, multiple cookies represent the same user over the period of our analysis. Hence, cookies with low query frequency under-represent the true movement. Query bounding boxes are in fact samples from the actual movement path, and a small number of samples is not enough to reconstruct the path. Hence for our second dataset we consider [*highly active*]{} cookies only: cookies with a query in more than 10 out of the 28 days.
Some highly active users may still have a limited movement and issue queries from the same geographical area. On the other hand, while some highly mobile users may issue fewer queries, they tend to move over a wider range. Therefore we expect leakage to be higher for highly mobile users, which is validated in our experiments.
For each of the United States and France, we collect two datasets - one for highly mobile and one for highly active cookies - and form a graph for each dataset. Both graphs have the same number of nodes (e.g. about 11,000 for the US for grid size $0.25$ degrees). Unless otherwise specified, node and edge weights are log-normalized - in a subsequent section we will compare various normalization methods.
Comparison against other clusterings {#subsec:q-metric-cutsize}
------------------------------------
[|c||c|cc|cc|cc|]{} (a) & & & &\
& & Avg & Query-weighted avg& Avg & Query-weighted ag& Avg & Query-weighted avg\
US&Highly Active & $86.9\%$ & $92.1\%$ & $88.1\%$ & $91.7\%$ & $84.0\%$ & $90.5\%$\
&Highly Mobile& $79.0\%$ & $85.4\%$ & $80.1\%$ & $85.0\%$ & $76.0\%$ & $80.7\%$\
France&Highly Active& $84.2\%$ & $88.8\%$ &-&-&$82.8\%$ &$86.3\%$\
&Highly Mobile & $74.4\%$ & $78.6\%$ & - & - & $75.0\%$ & $77.2\%$\
[|c||c|ccc|ccc|ccc|]{} (b) & & & &\
& & $\geq 0.75$ & $\geq 0.8$ & $\geq 0.85$ & $\geq 0.75$ & $\geq 0.8$ & $\geq 0.85$ & $\geq 0.75$ & $\geq 0.8$ & $\geq 0.85$\
US&Highly Active & $99.8\%$ & $99.0\%$ & $96.6\%$ & $100.0\%$ & $99.6\%$ & $97.9\%$ &$99.8\%$ & $98.8\%$ & $94.3\%$\
&Highly Mobile& $96.1\%$ & $86.4\%$ & $51.7\%$ & $95.3\%$ & $81.4\%$ & $48.5\%$ & $93.6\%$ & $60.0\%$ & $10.5\%$\
France&Highly Active & $99.8\%$ & $97.1\%$ & $79.7\%$ &-&-&- &$99.7\%$ & $88.8\%$ & $64.0\%$\
&Highly Mobile&$78.0\%$ & $42.3\%$ & $11.3\%$ & -&-&-& $74.4\%$ & $24.3\%$ & $5.0\%$\
[|c||c|ccc|ccc|]{} & & &\
& & $\geq 0.75$ & $\geq 0.8$ & $\geq 0.85$ & $\geq 0.75$ & $\geq 0.8$ & $\geq 0.85$\
$\sim 25$ clusters &Highly Active & $100\%$ & $100\%$ & $94.3\%$&$99.6\%$ & $99.1\%$ & $69.1\%$\
&Highly Mobile& $93.6\%$ & $55.5\%$ & $9.5\%$ & $83.0\%$ & $30.1\%$ & $0\%$\
$\sim 50$ clusters &Highly Active & $99.8\%$ & $97.1\%$ & $79.7\%$ &$99.7\%$ & $88.8\%$ & $64.0\%$\
&Highly Mobile& $78.0\%$ & $42.3\%$ & $11.3\%$ & $74.4\%$ & $24.3\%$ & $5.0\%$\
We previously defined Q-metric to measure leakage. GeoCUTS partitions the map in a way that reduces cut size. Clusters with lower values of cut size tend to have higher values of Q-metric. In this section we compare GeoCUTS with other alternatives in terms of Q-metric and cut size.
The average and query-weighted average of the Q-metric for each clustering algorithm are listed in Table \[tab:qmetric\](a). According to every metric, GeoCUTS beats the baseline clustering corresponding to a coarse grid of regions. Where applicable, GeoCUTS and DMAs perform similarly well. It is important to note that in every evaluation, we compared only clusterings with similar numbers of clusters. Thus, in constructing the grid baseline, we picked the coarseness of the grid so that the number of regions in the grid approximated the number of clusters formed by GeoCUTS. We also used $\sim 200$ clusters for GeoCUTS in the US, in order to provide an effective comparison with DMAs.
The fraction of clusters and queries for different lower bounds of Q-metric are shown in Table \[tab:qmetric\](b). For example $80\%$ of queries in the highly active set are issued from minimum-cut clusters with a Q-metric of at least $0.8$. For the grid, $62\%$ of queries are issued from clusters with a Q-metric of at least $0.8$. As already noted, highly mobile graphs are more challenging to partition compared to highly active graphs. The data indicates that the gap between our algorithm and baseline is larger for highly mobile graphs.
In Table \[tab:qmetric\], Q-metric values for DMAs are shown only for the US as they are undefined for other countries such as France. For both highly active and highly mobile graphs, GeoCUTS performs comparably to DMAs. Unlike DMAs, GeoCUTS offers flexibility in the number of clusters and functioning outside the US.
[GeoCUTS]{} [DMA]{} [Grid]{}
---------------------- ------------- --------- ----------
US Highly Active 0.015 0.015 0.015
US Highly Mobile 0.018 0.017 0.013
France Highly Active 0.111 - 0.115
: B-metrics compared to other clusterings. \[tab:bmetric\]
While the Q-metric quantifies the leakage, we still need to compare the clustering algorithms in terms of balance. An algorithm that produces highly unbalanced clusters may outperform other alternatives if only the Q-metric is considered. For example, if we partitioned the US into 200 clusters where 199 of them were in Alaska and one cluster covered the rest of the country, we would obtain an almost perfect Q-metric as relatively few users would cross between clusters. Obviously, however, such a clustering would not be useful for our applications. We compare B-metrics in Table \[tab:bmetric\]. The results indicate that GeoCUTS performs equally well in terms of balance as the alternatives for highly active graphs, and performs slightly better for the highly mobile graph. In summary, while we perform better in terms of leakage, we do not compromise balance.
We next compare the cut-sizes generated by GeoCUTS to DMA and grid clusters as baselines. Table \[tab:cut-size\] shows the cut size comparison on the US log-normalized graph with grid size $0.25$. It is clear that the GeoCUTS algorithm produces much better cut sizes compared to DMA and grid partitioning.
GeoCUTS DMA Grid LE Hilbert
--------------- ---------- ------ ------- ------ ---------
Highly Active **3.5%** 6.7% 14.8% 4.2% 7.3%
Highly Mobile **3.8%** 7.2% 14.0% 4.4% 7.4%
: Cut size comparison against different clustering algorithms. “Grid” denotes the grid partition, “LE” denotes the Linear Embedding algorithm [@aydin2015distributed].[]{data-label="tab:cut-size"}
Finally, we compare GeoCUTS results with Linear Embedding [@aydin2015distributed]. Since GeoCUTS itself uses Hilbert Curve for seed-selection we also compared it with partitions generated along a Hilbert Curve [@article:MJFS01:analysis-hilbert]. Table \[tab:cut-size\] compares the performance of alternative clustering algorithms; again GeoCUTS algorithm outperforms others on the same graph.
Normalization {#subsec:normalization}
-------------
We compared the performance of the GeoCUTS algorithm for different types of normalization during the graph-building phase. Specifically, we built graphs over US queries using logarithmic normalization of both vertices and edges, square root normalization, and also no normalization step at all. As expected, a stronger normalization was associated with better Q-metrics but worse B-metrics, demonstrating that normalization may be seen as mediating the tradeoff between diminished leakage and increased balance. The results of our comparison are shown in Table \[tab:qandbmetrics\](a).
Cell width {#subsec:cell-width}
----------
Finally, we compare the performance of GeoCUTS across varying coarsenesses of location discretization. The majority of our experiments are conducted with locations rounded to the nearest 0.25 degree (latitude and longitude). Here we compare log-normalized graphs in which locations are discretized to 0.1, 0.25, and 0.5 degrees. It is worth noting that all of these sizes are considerably larger than the side length of a typical bounding box for location data. As expected, the coarsest discretization of 0.5 performs the worst in Q-metric and best in B-metric, as coarser discretization enforces balance but reduces the ability to decrease leakage. The results of our comparison are shown in Table \[tab:qandbmetrics\](b).
(a) Log Square root None
------------------------- ------------ ------------- ------------
Q-metric, Highly Active **0.9205** 0.8809 0.8395
Q-metric, Highly Mobile **0.8539** 0.8069 0.7645
B-metric, Highly Active 0.0165 0.0047 **0.0006**
B-metric, Highly Mobile 0.0182 0.0053 **0.0011**
: Comparison of weighted average Q-metrics and B-metrics across (a) varying normalizations, (b) varying coarsenesses of location discretization.[]{data-label="tab:qandbmetrics"}
(b) 0.1 0.25 0.5
------------------------- -------- ------------ ------------
Q-metric, Highly Active 0.9157 **0.9205** 0.8580
Q-metric, Highly Mobile 0.8471 **0.8539** 0.7813
B-metric, Highly Active 0.0157 0.0165 **0.0115**
B-metric, Highly Mobile 0.0175 0.0182 **0.0132**
: Comparison of weighted average Q-metrics and B-metrics across (a) varying normalizations, (b) varying coarsenesses of location discretization.[]{data-label="tab:qandbmetrics"}
Conclusion {#sec:conclusion}
==========
We have presented an algorithm, GeoCUTS, designed for optimizing the assignment of regions to user queries. These regions can be used in running A/B bucket experiments to measure user response under treatment. Clustering users based on geographic region offers two major advantages: 1) assigning identical treatments to different browser cookies of the same user, and 2) decreasing the probability of interference if users with different treatments interact. Unlike existing systems, GeoCUTS can be run in any region of the world and for any number of clusters. Furthermore, the algorithm is based on user traffic as measured by queries.
We have presented a system of metrics for statistically evaluating the performance of clustering algorithms for A/B bucket testing. According to these and prior metrics, GeoCUTS performs significantly above baseline, and comparably to the state-of-the-art hand-designed clustering given by DMAs. We have also analyzed the parameter settings used in the GeoCUTS algorithm, showing how normalization of node and edge weights allows us to interpolate between the extremes of highly balanced clusters and minimal leakage.
[^1]: Massachusetts Institute of Technology, Cambridge, MA USA, `[email protected]`
[^2]: Google, Inc., New York, NY USA, `[email protected]`
[^3]: Google, Inc., Mountain View, CA USA, `[email protected]`
[^4]: Google, Inc., New York, NY USA, `[email protected]`
[^5]: Google, Inc., Mountain View, CA USA, `[email protected]`
[^6]: For our data sets, we never needed to iterate on this stage, since after the first stage, the graph fits in memory.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The formation and dynamics of cavities in liquids leads to focusing of kinetic energy and emission of longitudinal stress waves during the cavity collapse. Here we report that cavitation in elastic solids may additionally emit shear waves that could affect soft tissues in human bodies/brains. During collapse of the cavity close to an air-solid boundary, the cavity moves away from the boundary and forms a directed jet flow, which confines shear stresses in a volume between the bubble and the free boundary. Elastographic and high-speed imaging resolve this process and reveal the origin of a shear wave in this region. Additionally, the gelatin surface deforms and a conical crack evolves. We speculate that tissue fracture observed in medical therapy may be linked to the non-spherical cavitation bubble collapse.'
author:
- 'J. Rapet'
- 'Y. Tagawa'
- 'C.D. Ohl'
bibliography:
- 'Bib\_shear.bib'
title: 'Shear-wave generation from cavitation in soft solids'
---
In many medical applications energy is deposited within the tissue resulting in the formation of a cavity which expands and collapses. This phenomena is termed cavitation. The understanding of its dynamics helps to improve precision and mitigate side effects [@Vogel2005Nanosurgery; @brennen2015cavitation]. Examples of cavitation based therapy are histotripsy where pulsed finite amplitude ultrasound waves are focused into tissues and cornea surgery where laser pulses locally vaporize corneal tissue and produce clean intrastromal cuts [@juhasz1999corneal; @lubatschowski2000application]. The former leads to regions of intense cavitation where tissue is rapidly and locally turned into a paste[@roberts2006pulsed; @kim2011non]. The importance of bubble-tissue interaction stimulated research on individual bubbles oscillating in a tissue mimicking elastic solid. There the emission of stress and tension waves during bubble generation and collapse were documented [@brujan2006stress].
Tissue mimicking materials such as gelatin, polyacrylamide gels (PAA) or other hydrogels provide an elastic restoring force. Thus, they can transport not only longitudinal waves (with a wave velocity of $\approx 1500\,$m/s) but also transversal waves propagating at a considerably smaller velocity of $\approx 1-50$ m/s. As the shear wave velocity depends on the mechanical properties of the tissue, the tissue type may be characterized in diagnostic ultrasound from the shear front propagation [@gennisson2013ultrasound]. Beside the established methods based on acoustic focusing[@song2012comb], shear wave generation is an active research field with a number of new actuation mechanisms being reported. For example using an electric current in combination with a magnetic field displaces the elastic solid by the Lorentz force[@grasland2014imaging] or by locally displacing the hydrogel with bubbles from electrolysis[@montalescot2016electrolysis]. Recently, the generation of shear waves from local heating was demonstrated: when a laser beam heats up the surface of a soft elastic absorber[@grasland2016generation] two regimes are observed. In the thermoelastic regime the waves arise from the thermal expansion and in the ablative regime from impulse transfer due to vaporization of the surface.
Shear waves of larger amplitude, often induced by impacts[@cooper1989biophysics; @taylor2014investigation], cause the negative effects on soft tissues, e.g. in blast-induced traumatic brain injuries[@taylor2014investigation; @taber2006blast]. In the present work, we report on mechanism of shear wave generation from the non-spherical collapse of a single cavitation bubble in a soft solid. To induce the non-spherical collapse, the bubble is created in the solid near an air-solid boundary.
High-speed photography is used to record the bubble dynamics and the shear wave propagation. The latter is compared with a simple model to support the hypothesis of the mechanism of shear wave generation.
While pressure waves typically result in a change of index of refraction and can be visualized with shadowgraphy or Schlieren imaging, shear waves are more difficult to picture. Here, we utilize birefringence in the solid. Gelatin as our tissue phantom becomes doubly refractive under stress and is then suitable for photoelastic photography [@tomlinson2015photoelastic]. Those two refracted rays possess directions of polarization coinciding with the local principal stress directions. A circular polariscope is particularly suitable for measuring the stress distribution in the medium. It is obtained from two linear polarizers and two $\lambda/4$ waveplates as sketched in Figure \[fig:Setup\](a). The intensity of the light exiting the circular polariscope is a function of the retardation $\delta$ (Eq. \[eq:retardation\]) and can be expressed using trigonometry or Jones calculus [@ramesh2000digital] for 2-dimensional stress distributions $$\label{eq:circ-pola}
I = k^2\,\sin^2\bigg(\frac{\delta}{2}\bigg)\quad .$$ Monochromatic illumination assures that the relative retardation $\delta$ is only a function of the difference in amplitude of the two principal stresses and can be expressed as $$\label{eq:retardation}
\delta = \frac{2\pi h}{\lambda}C(\sigma_1-\sigma_2)\quad ,$$ where $h$ is the thickness of the sample, $\lambda$ the wavelength of the incoming light, $C$ the stress-optic coefficient, $\sigma_1$ and $\sigma_2$ the principal stresses. Thus the light is extinguished only where the sample is unstressed or where the principal stress difference $(\sigma_1 - \sigma_2)$ is causing a phase difference of $\delta = 2m\pi$ $ (m = 0,1,2,...)$. The second condition marks the presence of an isochromatic, i.e an area of constant principal stress difference.
The experimental setup shown in Figure \[fig:Setup\](a) allows to record the cavitation bubble dynamics and the shear wave emission through photoelastic imaging with a high speed camera. Single laser induced cavitation bubbles are created by focusing a laser pulse (Litron Lasers, Nano series, Q-switched Nd:YAG, 6 ns, wavelength $1064\,$nm ) into the gelatin with a microscope objective (Olympus $10\times$ Plan Achromat, N.A. $= 0.25$). At the focal point of the lens the bubble is created through an optical breakdown. The dynamics of the bubble (expansion, collapse and rebound) and the shear waves are recorded with a high speed camera (Shimadzu HPV-X2) equipped with a macro lens (Canon MP-E 65mm f/2.8 1-5$\times$ Macro). Single wavelength illumination is provided by a continuous green laser (Shaan’’xi Richeng Ltd, DPSS Green Dot Laser Module, wavelength $532\,$nm).
The soft solids samples are prepared from powdered gelatin (Gelatin 250 bloom, Yasin Gelatin CO.,LTD). It is mixed with deionized water at a mass ratio of 4% (gelatin to water) and dissolves in a flask on a hot plate with a magnetic stirrer. The hot mixture is poured into an optical glass cuvette (Aireka Scientific Co. Ltd, $3.5\,$ml, QO10204-4) to minimize birefringence distortion from the container and insure a flat gelatin-glass interface for clear observation and accurate laser focusing. The samples cool down to room temperature and are stored in a fridge. Before use, we assure that the samples have reached room temperature again.
Figure \[fig:Pola-side\] presents a typical result from the experimental setup Figure \[fig:Setup\](a). It consists of selected frames from a high-speed recording of a bubble expanding and collapsing in gelatin near a free (air) boundary. The time $t = 0$ denotes the time of bubble nucleation. Then the bubble expands and compresses the gelatin nearby resulting in a bright area around the bubble ($t = 5\,\mu$s). Upon reaching the maximum size, the bubble pushes and deforms the free surface ($t=35\,\mu$s) before it shrinks again. Thereby its upper wall bulges in and forms a funnel while the position of the interface at the bottom of the bubble remains approximately static ($t= 65\,\mu$s). Between the bubble and the free surface, two vertical dark lines develop on top of the brighter stressed area. These two lines mark the presence of isochromatics. The dark lines extend vertically following the bubble collapse with a velocity of $V_v = 25\,$m/s$\,\pm\,1$ m/s, which roughly also is the velocity of the upper bubble interface. Their horizontal position changes, too. They move outward with a velocity of $V_s =2\,$m/s$\,\pm\,0.2\,$m/s.
After reaching the minimum size the bubble rebounds ($t = 85\,\mu$s) and its volume oscillations cease shortly after the second collapse at $t = 119\,\mu$s (not shown), while it remains moving downwards.
At $t = 155\,\mu$s, a conical-shaped crack appears starting from the air-gelatin boundary and follows the path of the downwards translating bubble ($t = 225\,\mu$s). We speculate that the shrinking and moving bubble is the source for shear stress generation. As the speed of this motion is considerably higher than the shear wave velocity, shear stress is confined in the region bounded by the air interface and the bubble. Later only this confined stress propagates from this region as a shear wave into the surrounding medium.
This hypothesis is tested with a numerical simulation of the stress generation and propagation using finite element solver for an elastic solid (Solid Mechanics Module, COMSOL Multiphysics). For the sake of simplicity, the stress field resulting from a body force moving downward with a constant velocity is modeled. Figure \[fig:Simu\](a) depicts the 2D axisymmetric simulation domain, properties of the gelatin are taken from literature ($\rho = 1000$ kg/m$^3$, $\epsilon = 0.45$, $E = 10$ kPa, $C_p = 1500$ m/s, $C_s = 1.8$ m/s)[[@czerner2015determination; @gennisson2013ultrasound]]{}. The boundary conditions are stress free on the top, an axis of symmetry, and to the right and bottom of the domain impedance matched boundaries. At $t = 0$, a rectangular source of force (body load, $F = 1\,$N) moves downward from the free surface along the axis of symmetry with a velocity of $V_Z = 25$ m/s for a duration $\Delta t = 32\,\mu$s. To compare the simulations with the photoelastic images we split the axisymmetric model into slices perpendicular to the line of sight. In each slice of finite thicknesses, the polariscope equation is solved assuming a constant stress along the depth.
Figure \[fig:Simu\](b) presents selected frames from the numerical results. Black areas correspond to the gas/vapour domains (bubble and free surface drawn from the experimental results and added on the numerical results), dark gray (low light intensity) shows the unstressed gelatin and areas of stress are depicted with gray scales, with higher brightness indicating areas of higher stress. Overall the simulation results show good agreement considering the coarse simplifications. In particular we observe a bright region which extend vertically following the rectangular source of force with a velocity of $V_z$ and slowly expands radially. The main difference between simulation and experiment is that the isochromatics are not reproduced. We explain this that the pre-existing stress fields from the earlier bubble dynamics are not considered in the simulations, i.e. there the body force moves within an unstressed solid. The simulation supports the explanation that shear waves can be induced with a source of stress moving with the bubble upper interface.
{width="100.00000%"}
The motion of the bubble collapsing away from the free surface (Figure \[fig:Pola-side\]) is also reported for liquids, i.e. in water where a jet pierces through the bubble [@robinson2001interaction]. It suggests that the emergence of a jet penetrating through the bubble in gelatin could cause the generation of the shear waves. To obtain a better photographic evidence we repeated the experiment in absence of the polariscope optics, see Figure \[fig:Setup\](b) for the experimental setup. Now two high-speed cameras (Photron, Fastcam, Mini AX200) equipped with camera lenses (Canon MP-E 65 mm f/2.8 1-5$\times$ Macro) record simultaneously the bubble and the surface from the side and top.
Figure \[fig:2-cam\] presents selected frames from a bubble with very similar size and position as shown in Figure \[fig:Pola-side\]. Again the bubble expands and pushes the free surface resulting in an upward displacement visible on the surface with a gelatin jet pointing upward ((a):$t = 15\,\mu$s) and expanding laterally as the bubble grows ((a):$t = 44\,\mu$s, (b):$t = 37\,\mu$s). The bubble then collapses and moves away from the free surface. During collapse, a jet is penetrating through the center of the bubble with a velocity of $V_s = 19\,$m/s$\,\pm\,1\,$m/s ((b):$t = 68\,\mu$s). Then the jet impacts on the lower bubble interface while the bubble continues to translate downward ((b):$t = 86\,\mu$s and $t= 168\,\mu$s). These observations indicate that the jet formed by the bubble moving away from the air-gelatin boundary is causing the shear wave. At a later time the restoring force of the gelatin pulls back the residues of the bubble towards its position at creation, see (b):$t = 222\,\mu$s.
The crack originating from the air-gelatin boundary indicates that the upwards pointing tip of the jet on the gelatin-air interface ((a):$t = 15 \,\mu$s) retracts first towards the original level and then penetrates into the gelatin, visible as a conical crack in (a):$t = 178\,\mu$s. The opening angle of the crack remains while the crack tip propagates into the gelatin trailing the cavitation bubble (Figure \[fig:Pola-side\](b):$t = 222\,\mu$s). The oscillations of the free surface result in surface waves (Figure \[fig:2-cam\] sketch and (a)) traveling radially with a velocity of $V_s = 4.5\,$m/s$\,\pm\,0.5\,$m/s.
In conclusion, we have demonstrated that a single cavitation bubble in soft media can generate shear waves. The jetting of the shrinking and collapsing bubble is causing shear stresses to be confined between the air-gelatin interface and the bubble. Shear waves are then emitted from this region. We also observe the formation of a crack following the downward motion of the bubble. We speculate that this could be a potential source of tissue damage when cavitation bubbles collapse non-spherically. We speculate that confinement of shear stress near to free interfaces may lead to cracks and tissue damage.
*Acknowledgements* Y.T. acknowledges financial support from JSPS KAKENHI Grants No. R2801 and 17H01246. This project has received funding from the European Union Horizon 2020 Research and Innovation programme (No 813766).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
An experiment is described that confirms the security of a well-studied class of cryptographic protocols (Dolev-Yao intruder model) can be verified by two-way nondeterministic pushdown automata (2NPDA). A nondeterministic pushdown program checks whether the intersection of a regular language (the protocol to verify) and a given Dyck language containing all canceling words is empty. If it is not, an intruder can reveal secret messages sent between trusted users. The verification is guaranteed to terminate in cubic time at most on a 2NPDA-simulator. The interpretive approach used in this experiment simplifies the verification, by separating the nondeterministic pushdown logic and program control, and makes it more predictable. We describe the interpretive approach and the known transformational solutions, and show they share interesting features. Also noteworthy is how abstract results from automata theory can solve practical problems by programming language means.
#### Keywords
protocol verification, ping-pong protocols, cryptographic protocols, nondeterministic programming, two-way pushdown automata, memoizing interpreters
author:
- Robert Glück
title: |
An Experiment in Ping-Pong Protocol Verification\
by Nondeterministic Pushdown Automata
---
Introduction {#sec:intro}
============
Soon after the introduction of public-key encryption [@RSA:78], it was found that an adversary can obtain a secret message sent on a network between trusted users, not by breaking the cryptographic algorithm, but by breaking the communication protocol through complex interactions with the users. A key finding by Dolev and Yao [@DolevYao:81; @DolevYao:83] was that the security problem of cryptographic ping-pong protocols can be mapped onto a decidable grammar problem. They gave an algorithm for constructing, for any given ping-pong protocol, a nondeterministic finite-state automaton (regular language) representing all possible interactions between the trusted users and the adversary, and a special-purpose algorithm for deciding the security question by computing the collapsing-state relation by a closure algorithm. Their original algorithm decided the security question in time $O(n^8)$, where $n$ is the size of the automaton [@DolevYao:81; @DolevYao:83]. This was later improved to $O(n^3)$ [@DolevEvenKarp:82]. The security of protocols is very important because public-key crypto systems are widely used for electronic communication and underpin various Internet standards.
Recently, Nepeivoda [@Nepeivoda:16] showed that the security of ping-pong protocols can also be verified by program transformation. Instead of regular expressions, the protocol in question is mapped onto a prefix grammar encoded as a first-order functional program in such a way that it can be used to decide the security question by a program transformer, specifically a supercompiler. This method builds upon work solving other verification problems by general-purpose program transformation (, [@AhmedLisitsaNemytykh:13; @LisitsaNemytykh:07]). The main steps of these two methods are shown on the left- and right-most branches in Fig. \[fig:VERIFYapproaches\].
This paper takes another programming language approach — an interpreter for a nondeterministic language is used instead of a program transformer. We show how to write a two-way nondeterministic pushdown (2NPDA) program that searches for insecure communications in a finite-state automaton constructed by the Dolev-Yao algorithm, and interpret the program by an existing simulator for nondeterministic pushdown programs, which decides the security question in time $O(n^3)$. This approach leads to a surprisingly straightforward solution that is asymptotically as efficient as any of the other verification methods above. The verification is simplified by separating the problem into an easy-to-write pushdown logic that specifies the solution in terms of finding a path in a directed graph and a control component that calculates the actual solution; such separation is known in logic programming as “algorithm = logic + control” [@Kowalski:79]. Herein, a nondeterministic pushdown program and a simulator that embeds memoization as a control component are used. The simulator guarantees termination and polynomial-time performance on a random-access machine with a uniform cost model [@AHU:68; @Glueck:16:pdasim]. The simulators that we use simulate one machine by another; in programming language terms these are interpreters. This experiment also shows how, relying on proven results from automata theory, a class of cryptographic protocols can be verified by means of nondeterministic programming. The method is shown by the bold shapes in Fig. \[fig:VERIFYapproaches\]. The deterministic (det) and nondeterministic (ndet) source and implementation languages of the simulator (SIM) and the supercompiler (SCP) are discussed in a later section. The interpretation and transformation approaches are two sides of the same coin, and we show they share interesting features.
(4,7) rectangle node\[align=center\][ping-pong [\
]{}protocol]{} (6,8); (5,7) –node\[left\][*mapping* ]{} (3,6); (5,7) –node\[right\][ *mapping*]{} (7,6);
(2,5) rectangle node\[align=center\][finite-state [\
]{}automaton]{} (4,6); (3,5) –node\[left\][*edge set *]{} (1,4); (3,5) –node\[right\][ *edges on tape*]{} (5,4); (6,5) rectangle node\[align=center\][prefix [\
]{}grammar]{} (8,6); (7,5) –node\[right,align=center\] [ *program text*]{} (9,4);
(-0.05,3) rectangle node\[align=center\][state-relation [\
]{}closure algo.]{} (2.05,4); (1,3) –node\[right\][*answer*]{}(1,2.05);
(4,3) rectangle node\[align=center\][pathfinder]{} node\[below right\][ ndet]{} (6,4); (4,2) rectangle node\[\][SIM]{} node\[above right\][ ndet]{} node\[below right\][ det]{} (6,3); (5,2) –node\[right\][*answer*]{}(5,1.07);
(5.7,2.8) node\[below right\][ interpreter]{}; (9.67,3.75) node\[below right\][ transformer]{}; (8,3) rectangle node\[\][SCP]{} node\[above right\][ det]{} node\[below right\][ det]{} (10,4); (9,3) –node\[right\][*answer*]{}(9,2);
(1,1.5) circle \[radius=0.55\] node\[align=center\][y/n]{}; (5,0.5) circle \[radius=0.57\] node\[align=center\]
------------------------------------------------------------------------
accept\
reject
; (8.5,1) rectangle node\[align=center\][resid. [\
]{}pgm]{} (9.5,2);
\[fig:VERIFYapproaches\]
The approach taken here is similar to other resource-bounded computation models [@Jones:97:complexity] where certain properties are guaranteed for all programs regardless of how they are written (, all programs written in a reversible language are easily invertible [@AxGl11FoSSaCS]). Even though the nondeterministic pushdown computation model used here is subuniversal (not Turing-complete), it is not particularly weak. The multihead 2NPDA characterize the polynomial-time (“tractable”) algorithms. Pushdown programming is a technique for solving problems that may deserve more attention, perhaps supported by program transformation.
In Sect. \[sec:pingpong\], we briefly review ping-pong protocols and the security problem. In Sect. \[sec:ndetpgmoverview\], we introduce the nondeterministic pushdown language and, in Sect. \[sec:ndetpgm\], we present the protocol verifier. In Sect. \[sec:vptsim\] and \[sec:related\], we discuss the methods and related work, respectively.
Review of Ping-Pong Protocols {#sec:pingpong}
=============================
Ping-pong protocols are a class of two-party cryptographic protocols. Their purpose is to transmit secret text between two users in a network. The initiator of a communication applies an initial sequence of operators to the text of a message and sends the message to the intended recipient. In each step of their communication, a participant applies an operator sequence to the text most recently received and returns the result. This ping-pong action continues several times as specified by the protocol. Operators that can be applied to a text include name stamps and cryptographic operators (see [@DolevEvenKarp:82] for more information).
Public-key encryption is a cryptographic system that allows an *encryption key* to be revealed to the public without revealing the corresponding *decryption key* (, RSA [@RSA:78]). Consequently, a text can be enciphered by anyone using the encryption key publicly revealed by the intended recipient of the text, but only the intended recipient can decipher the text because only this recipient has the corresponding decryption key. There is no need to secretly exchange keys between the participants.
#### Protocol Operators.
A network is assumed to have three legitimate users ($X, Y, Z$) with equal rights. Each user can initiate a communication with another user by sending an initial message. A message sent in the network consists of three fields: the sender’s name, the receiver’s name and the text. The text is the part of a message to which a user can apply operators as specified by the communication protocol. All users have the same set of operators ($\Sigma$) that they can apply to the text of a message and each user also has a private operator ($D_X, D_Y, D_Z$) for decrypting a text with their private key.
\[def:opsets\] The operator sets of users $X, Y, Z$ are $$\begin{aligned}
\Sigma_X &=& \Sigma~\cup~\{ D_X \}~~~~~~~\textrm{decrypt by private key of \mymath{X}}{\,},\\
\Sigma_Y &=& \Sigma~\cup~\{ D_Y \}~~~~~~~\textrm{decrypt by private key of \mymath{Y}}{\,},\\
\Sigma_Z &=& \Sigma~\cup~\{ D_Z \}~~~~~~~\textrm{decrypt by private key of \mymath{Z}}{\,},\end{aligned}$$ where $\Sigma$ is the common operator set available to every user: $$\begin{aligned}
\Sigma
&=& \{E_X, E_Y, E_Z,~~~~~~~~\textrm{encrypt by public key of \mymath{X}, \mymath{Y}, \mymath{Z}} \\
&& ~\;P_X, P_Y, P_Z,~~~~~~~~~~\textrm{prepend name of \mymath{X}, \mymath{Y}, \mymath{Z}} \nonumber\\
&& ~\;M_X, M_Y, M_Z,~~~~~~\textrm{match and delete prepended name of \mymath{X}, \mymath{Y}, \mymath{Z}} \nonumber\\
&& ~\;M \}~~~~~~~~~~~~~~~~~~~~~\textrm{delete any prepended name}{\,}.\nonumber\end{aligned}$$
Operators $E_X, E_Y, E_Z$ encrypt a text with the public key of a user, operators $P_X, P_Y, P_Z$ prepend a user name to a text, operators $M_X, M_Y, M_Z$ delete a prepended user name if the name matches, and $M$ deletes any prepended user name from a text. The cryptographic operators are defined for any text. If the keys mismatch, they return just gibberish ( decryption of an encoded text with the wrong key: $D_Y E_X$). The cryptographic operators of public-key encryption are inverse to each other ( $D_X E_X = E_X D_X = \epsilon$ where $\epsilon$ denotes the empty sequence of operators). A protocol aborts when prepended names mismatch ( expecting another name stamp: $M_Y P_X$). The order of applying the operators is from right to left.
\[def:opids\] Cryptographic operators of public-key encryption [@RSA:78] have the following identities for any user $U$, as have the operators for matching and deleting prepended user names [@DolevEvenKarp:82]. $$\begin{aligned}
\label{eq:identities}
& D_U E_U ~~=~~ E_U D_U ~~=~~M_U P_U ~~=~~ M P_U ~~=~~ \epsilon{\,}. & $$
Consider a Dolev-Yao ping-pong protocol [@DolevEvenKarp:82] as an example that is defined between two users $A,B \in \{ X,Y,Z \}$ where $A$ is the initiator and $B$ the recipient. The protocol consists of two operator words, $\alpha_1 \in \Sigma_A^*$ and $\alpha_2 \in \Sigma_B^*$. In each step one of the two participants applies an operator word to the text and sends the result to the other participant. The following two words define the protocol’s ping-pong action. $$\begin{aligned}
\mbox{Protocol 1:}\\
\alpha_1(A,B) &=& E_B \\
\alpha_2(A,B) &=& E_A D_B\end{aligned}$$ Initially, $A$ sends a message to $B$ after encrypting the secret text by $B$’s public key, that is $A$ applies $\alpha_1(A,B) = E_B$ to the text. Recipient $B$ sends a message back to $A$ after decrypting the received text by $B$’s private key $D_B$ and encrypting the result by $A$’s public key $E_A$, that is $B$ applies $\alpha_2(A,B) = E_A D_B$ to the text it received. The second step completes this ping-pong protocol. The complete operator word applied to the text is $\alpha_2(A,B) \circ \alpha_1(A,B) = E_A D_B E_B$. Under this cryptographic protocol, recipient $B$ can read $A$’s secret text and initiator $A$ can compare the original text with the text echoed by $B$. The communication between $A$ and $B$ is performed without sending any plain text in the network. This cryptographic protocol appears to be a secure way of exchanging messages but, as we shall see, it is not.
=\[font=\]
\(9) \[\] [0]{}; (1) \[right of=9\] [1]{}; (2) \[below left of=1\] ; (3) \[below right of=1\] ;
\(9) edge node [$E_Y$]{} (1) (1) edge node [\[0ex\]\[0ex\][$D_Y$]{}]{} (3) edge node [$D_X$]{} (2) edge \[loop above\] node [$\Sigma_Z$]{} (1); (2) – node\[left\][$E_Y$]{} (-0.02,-1.57) – (2.6,-1.57) – (2.6,0) – (1); (2) – node\[right\][$E_Z$]{} (0.6,-1.57) – (0.9,-1.57); (3) – node\[left\] [$E_X$]{} (1.38,-1.57) – (1.7,-1.57); (3) – node\[right\][$E_Z$]{} (2.0,-1.57) – (2.2,-1.57);
\[fig:FSAprotocolone\]
#### Dolev-Yao Intruder Model.
Assume that two of the users on the network are well-behaved ($X, Y$), which means they only apply the operations specified by the protocol. The third user is a saboteur ($Z$) who can apply any operator sequence $\Sigma_Z^*$ to a message text. The saboteur waits patiently for a chance to crack the communication between the well-behaved users by listening to the network, intercepting and altering any message. This scenario is sufficient for checking the security of two-party ping-pong protocols in the Dolev-Yao intruder model. A single saboteur is sufficient in this model because a single saboteur can do whatever a group of saboteurs can do [@DolevEvenKarp:82]. The model assumes only a few limitations on the behavior of the saboteur. The saboteur can impersonate any sender ($X, Y, Z$), apply any operator sequence $\Sigma_Z^*$ to a text, and send an altered message to any user. It is assumed that private keys cannot be stolen from the users. Consider as an example how Protocol 1 can be cracked by saboteur $Z$. Let $X$ initiate the communication with $Y$ by sending a text encrypted by $\alpha_1(X,Y)$ in the network. Assume that $Z$ intercepts $X$’s initial message, changes the sender to $Z$, and sends the unchanged text to $Y$, who believes $Z$ initiated a communication by sending the initial $\alpha_1(Z,Y)$. As specified by the protocol, the well-behaved user $Y$ responds to $Z$ by applying $\alpha_2(Z,Y)$ to the received text, that is after decrypting the text by $E_X$ and encrypting it by $E_Z$ for the perceived sender $Z$. The saboteur $Z$ can now simply decrypt the message by $D_Z$. The secret has been revealed, not by cracking the public-key encryption algorithm, but the protocol! The entire operator sequence applied to the original text is reduced to $\epsilon$ by the operator identities: $$\begin{aligned}
&
D_Z\circ \alpha_2(Z,Y) \circ \alpha_1(X,Y) ~~=~~
\fbox{$D_Z E_Z$} \fbox{$D_X E_X$} ~~=~~
\epsilon {\,}.
&\end{aligned}$$
#### The Security Question.
The Dolev-Yao intruder model formalizes the interaction of the well-behaved users ($X,Y$) and the saboteur ($Z$) using a nondeterministic finite-state automaton (FSA). The FSA for Protocol 1 in Fig. \[fig:FSAprotocolone\] generates all operator sequences that can be applied to a text sent by initiator $X$. Assuming that the initiator is $X$ is sufficient for checking the security of the protocol. The initial state is $\snode$, the accepting state is $\tnode$, and the edges are labeled with operators. $X$ starts by sending a message to $Y$ that the saboteur tries to obtain by cracking the protocol. After the initial operator sequence $\alpha_1(X,Y)= E_Y$ is applied to the text, $Z$ can intercept the message and apply to it an arbitrary operator sequence $\Sigma_Z^*$, or $X$ and $Y$ can apply $\alpha_2(X,Y)$, $\alpha_2(X,Z)$, $\alpha_2(Y,X)$ or $\alpha_2(Y,Z)$ to the text in response to a message received from $X,Y,Z$. The FSA can be constructed for any ping-pong protocol by an algorithm [@DolevEvenKarp:82]. [^1] The security question translates into the following grammar problem: A protocol is secure if the intersection of $\mathcal{L}(\mathrm{FSA})$, the regular language defined by its FSA, and $\mathcal{L}(\Gnterm)$, the context-free language of all reducible operator words, is empty. Thus, the security question of ping-pong protocols is a decidable grammar problem, namely the emptiness of the intersection of a regular language and a context-language: $$\begin{aligned}
\mathcal{L}(\mathrm{FSA}) \cap \mathcal{L}(\Gnterm) &\stackrel{?}{=}& \emptyset{\,}.\end{aligned}$$ The context-free grammar $\Gnterm$ generating all reducible words is the same for all ping-pong protocols: $$\begin{aligned}
\Gnterm &::=& D_U~\Gnterm~E_U~\mid~E_U~\Gnterm~D_U~\mid~M_U~\Gnterm~P_U~\mid~M~\Gnterm~P_U~\mid~\Gnterm~\Gnterm~\mid~\epsilon~~~~~\mbox{for all $U \in \{X,Y,Z\}${\,}.}\end{aligned}$$ The saboteur in the Dolev-Yao intruder model can apply any operator sequence $\Sigma_Z^*$ to a message text, so $\Gnterm$ must generate all reducible words. A word in $\Gnterm$ is balanced with respect to “opening” and “closing” pairs of operators. Any word in $\mathcal{L}(\Gnterm)$ can be *reduced* to $\epsilon$ by repeatedly applying the identity rules in Def. \[def:opids\], by substituting $\epsilon$ successively for every occurrence of pairs to which the identity rules apply. It can easily be shown that these reductions can be performed in any order without changing the result. The *reduction strategy* we are going to use later is to repeatedly apply the identity rules to the rightmost, innermost pair to which they apply. [^2] There is no difference between handling a mismatch of keys and of prepended names (the operator pair does not reduce).
A ping-pong protocol is *secure* iff there is no accepting path in its FSA representation whose word is reducible to $\epsilon$ by the operator identities; otherwise, the protocol is insecure [@DolevEvenKarp:82].
=\[font=\] (8) [0]{}; (9) \[right of=8\] ; (1) \[right of=9\] [1]{}; (3) \[below right of=1\] ; (6) \[below of=3\] ; (7) \[right of=6\] ; (2) \[below left of=1\] ; (5) \[below of=2\] ; (4) \[left of=5\] ; (8) edge node [$P_X$]{} (9) (9) edge node [$E_Y$]{} (1) (1) edge \[\] node [\[0ex\]\[0ex\][$D_Y$]{}]{} (3) edge \[\] node [$D_X$]{} (2) edge \[loop above\] node [$\Sigma_Z$]{} (1) (2) edge \[above left\] node [$M_Y$]{} (4) edge \[right\] node [$M_Z$]{} (5) (3) edge \[left\] node [$M_X$]{} (6) edge node [$M_Z$]{} (7); (4) \[rounded corners\] – node\[left\][$E_Y$]{} (0.3,-2.57) – (4.5,-2.57) – (4.5,0) – (1); (5) \[rounded corners\] – node\[right\][$E_Z$]{} (1.3,-2.57) – (1.5,-2.57); (6) \[rounded corners\] – node\[left\] [$E_X$]{} (2.7,-2.57) – (2.9,-2.57); (7) \[rounded corners\] – node\[right\][$E_Z$]{} (3.7,-2.57) – (3.9,-2.57);
=\[font=\] (0) [0]{}; (8) \[right of=0\] ; (9) \[right of=8\] ; (1) \[right of=9\] [1]{}; (3) \[below right of=1\] ; (6) \[below of=3\] ; (10) \[below of=6\] ; (7) \[right of=6\] ; (13) \[below of=7\] ; (2) \[below left of=1\] ; (5) \[below of=2\] ; (12) \[below of=5\] ; (4) \[left of=5\] ; (11) \[below of=4\] ; (0) edge node [$E_Y$]{} (8) (8) edge node [$P_X$]{} (9) (9) edge node [$E_Y$]{} (1) (1) edge \[\] node [\[0ex\]\[0ex\][$D_Y$]{}]{} (3) edge \[\] node [$D_X$]{} (2) edge \[loop above\] node [$\Sigma_Z$]{} (1) (3) edge \[left\] node [$M_X$]{} (6) edge node [$M_Z$]{} (7) (6) edge \[left\] node [$D_Y$]{} (10) (7) edge node [$D_Y$]{} (13) (2) edge \[above left\] node [$M_Y$]{} (4) edge \[right\] node [$M_Z$]{} (5) (5) edge node [$D_X$]{} (12) (4) edge \[left\] node [$D_X$]{} (11); (11) \[rounded corners\] – node\[left\][$E_Y$]{} (1.3,-3.57) – (5.5,-3.57) – (5.5,0) – (1); (12) \[rounded corners\] – node\[right\][$E_Z$]{} (2.3,-3.57) – (2.5,-3.57); (10) \[rounded corners\] – node\[left\][$E_X$]{} (3.7,-3.57) – (3.9,-3.57); (13) \[rounded corners\] – node\[right\][$E_Z$]{} (4.7,-3.57) – (4.9,-3.57);
\[fig:FSAprotocoltwothree\]
Protocol 1 can be made secure by prepending the name of initiator $A$ to the text before encrypting it with $B$’s public key, that is $A$ applies $\alpha_1(A,B) = E_B P_A$ to the text. Recipient $B$ now encrypts using $A$’s public key only if the text has $A$’s name prepended, as checked by match $M_A$, that is $B$ applies $\alpha_2(A,B) = E_A M_A D_B$ to the text it received. The FSA of Protocol 2 in Fig. \[fig:FSAprotocoltwothree\] has no accepting path whose word is reducible [@DolevEvenKarp:82].
$$\begin{aligned}
\mbox{Protocol 2:} \\
\alpha_1(A,B) &=& E_B P_A \\
\alpha_2(A,B) &=& E_A M_A D_B\end{aligned}$$
$$\begin{aligned}
\mbox{Protocol 3:} \\
\alpha_1(A,B) &=& E_B P_A E_B \\
\alpha_2(A,B) &=& E_A D_B M_A D_B\end{aligned}$$
One may think that extra encryption makes a protocol even more secure, but this is not necessarily the case. Suppose we want to improve Protocol 2 by encrypting the text once more. Let initiator $A$, before prepending the name to the text by $P_A$, encrypt it with $B$’s public key, that is $\alpha_1(A,B) = E_B P_A E_B$. Receiver $B$ now decrypts the text before encrypting it with $A$’s public key, that is $\alpha_2(A,B) = E_A D_B M_A D_B$. This innocent looking “improvement” makes the protocol insecure! This shows that informal arguments are error prone. A formal way to decide the security of ping-pong protocols is needed. The operators which a saboteur can inject into the communication are not immediately obvious in Fig. \[fig:FSAprotocoltwothree\] in the FSA of Protocol 3 [@DolevEvenKarp:82]. For example, it can be cracked by a patient saboteur who twice injects operators and fakes senders and receivers: $$\begin{aligned}
D_Z \circ \alpha_2(Z,Y) \circ E_Y P_Z M_X D_Z \circ \alpha_2(Z,Y) \circ E_Y P_Z \circ \alpha_1(X,Y) &=& \nonumber\\
\fbox{$D_Z E_Z$} \fbox{$D_Y \fbox{$M_Z \fbox{$D_Y E_Y$} P_Z$} \fbox{$M_X \fbox{$D_Z E_Z$} \fbox{$D_Y \fbox{$M_Z \fbox{$D_Y E_Y$} P_Z$} E_Y$} P_X$} E_Y$} &=&
\epsilon{\,}.\end{aligned}$$
Nondeterministic Programming and a Pushdown Language {#sec:ndetpgmoverview}
====================================================
*Two-way nondeterministic pushdown automata* (2NPDA) will be our programming model. Pushdown automata are simple versatile devices comprising three components: a read-only input tape, a potentially infinite stack, and a finite-state control. They can move the head on the tape in both directions, push and pop symbols to and from the stack, and test the symbol on top of the stack and the symbol read on the tape. The alphabets of tape symbols, stack symbols and states are finite. *Multihead pushdown automata* can read and move multiple heads independently on the tape. Even though these devices are subuniversal, they are not particularly weak. The *multihead 2NPDA* are equivalent to the polynomial-time algorithms [@WagnerWechsung:86]. Any $k$-head 2NPDA can be simulated in at most $O(n^{3k})$ steps on a random-access machine with a uniform cost model where $n$ is the length of the tape [@AHU:68]. We shall see that a single head ($k=1$) is sufficient to check the security of ping-pong protocols, which means this takes at most cubic time.
Instead of using a traditional multi-valued transition function for defining the operation of a 2NPDA, we introduce the *multihead nondeterministic pushdown language*, shown in Fig. \[fig:pdafcl-syntax\], with commands to move the head (`left`, `right`), push and pop symbols (`push`, `pop`), and halt in an accepting or rejecting final state (`accept`, `reject`). Predicates can compare two symbols (`=`), test the emptiness of the stack (`bottom`) and the ends of the tape (`leftend`, `rightend`). Global variables contain the symbol currently on top of the stack (`top`) and read by the tape head (`hd`). The variables are updated when the stack top changes or the head moves. Similarly, additional tape heads (`hd2`, `hd3`, $\ldots$) can be moved (`left2`, `left3`, $\ldots$) and tested (`leftend2`, `leftend3`, $\ldots$). The deterministic control-flow operators are as usual (`if`, `goto`). A program consists of labeled command sequences. Execution begins at the first command of a program. The left and right ends of the tape are marked with the tape symbols `>` and `<`, respectively. The language resembles flowchart languages except for its additional `choice` command [@Jones:97:complexity] which nondeterministically executes either command sequence $\cmdsv_1$ or $\cmdsv_2$, that is, the next state after the choice is not uniquely determined by the current state:
`choice `$\cmdsv_1$` or `$\cmdsv_2$` end`
[lcl]{} &::= & ()\^+\
&::= &\
&::= &\
&& …\
&& …\
&&\
& ::= & …\
[lcl]{} &::= & \
&& …\
&& …\
&&\
&&\
&&\
&&\
\[fig:pdafcl-syntax\]
Initially, the stack is empty and the tape heads scan the left end of the tape containing the input word. An input word is *accepted* by a nondeterministic pushdown program if there exists at least one computation sequence for the program that terminates in an `accept` command. The nondeterministic operation allows the simultaneous construction of every computation sequence for a given input. The formal language accepted by a program is the set of all input words it accepts. This is the usual definition for nondeterministic pushdown automata. The semantics of the pushdown language will not be formally defined here due to lack of space and the semantics being straightforward.
A program that contains no `left` command is *one way*; one that contains no `choice` command is *deterministic*. One-head programs accept important classes of formal languages. For example, the one-way nondeterministic pushdown (1NPDA) programs accept the context-free languages.
A textbook interpretation of a pushdown program may not terminate (, push forever on the stack) or take exponential time before terminating. However, memoizing simulation methods ensure termination and polynomial-time performance for all pushdown programs, because the number of possible surface configurations is polynomially bounded. Thus, every pushdown program has a definite answer (accept, reject). We refer the reader to [@AHU:68; @Glueck:16:pdasim] for a presentation of the simulation methods, and to [@Cook:72; @Jones:77] for the deterministic case. Accordingly, we shall be programming in a resource-bounded and decidable nondeterministic programming language, following the approach marked bold in Fig. \[fig:VERIFYapproaches\].
Protocol Security Checked by Nondeterministic Pushdown Programs {#sec:ndetpgm}
===============================================================
Combinatorial search problems may often be simply written using nondeterministic programs. Before we show how to verify the security of ping-pong protocols, we show how to find a path between two nodes in a directed graph by a pushdown program. We then extend the pathfinding program into the desired protocol verifier, discuss the nondeterministic programs and report on simulation results for the Dolev-Yao protocols. Verification by program transformation is discussed in the next section.
Pathfinding in a Directed Graph by a Pushdown Program
-----------------------------------------------------
`init: push ’0’; right (* push start node, pos 1st edge )`\
`loop: if top = hd (* both nodes match? )`\
` then choice (* make a guess: traverse or skip )`\
` pop; right; (* traverse edge )`\
` if hd = ’1’ then accept end; (* path 0 -> 1 found )`\
` push hd; right (* push next node, pos next edge )`\
` or`\
` 2-right end; (* skip edge )`\
` else 2-right end; (* mismatch: skip edge )`\
` if rightend then move-to-leftend end; (* return to 1st edge )`\
` goto loop`
\[fig:PDApathfind\]
Given a *directed graph* $G = (V, E)$ with nodes $u,v \in V$ and directed edges $E = \{(u_1,v_1), \ldots, (u_n,v_n)\}$, where an edge $(u_i,v_i)$ leads from node $u_i$ to node $v_i$, the task of the pushdown program is to check whether there exists a path from a source node $s \in V$ to a target node $t \in V$ in $G$.
Let the names of the source and target nodes be fixed as $s=\snode$ and $t=\tnode$. For simplicity, we assume that the node names in $V$ are included in the tape- and stack-symbol alphabets of the automaton. $G$ can be represented by a tape of length $O(n)$ that lists all edges in $E$, where `>` and `<` mark the two tape ends: $$\texttt{>}~u_1~v_1~\ldots~u_n~v_n~\texttt{<}$$ The nondeterministic program shown in Fig. \[fig:PDApathfind\] guesses a path in $G$ from $\snode$ to $\tnode$, if it exists. Initially, the stack is empty and the head is positioned at the left end (`>`) of the tape, so $\snode$ is pushed on the empty stack by `push ’0’` and the head is positioned at $u_1$ of the first edge $(u_1~v_1)$ on the tape by `right` in the first line of the program. The invariant of the main loop that follows, is that the current node of the path that the program is exploring is kept on top of the stack. This node is updated by the main loop when an edge is traversed to a new node. No other nodes are pushed on the stack, so a stack of height 1 suffices for following a path in $G$. Furthermore, a single head suffices for scanning the edges on the tape.
The main loop moves the head `hd` over the sequence of edges on the tape, until an edge originating in the current node on top of the stack is found (`top` = `hd`). The nondeterministic choice at this point is to either traverse the edge originating in `top` or to continue the search for another edge originating in `top`. If the edge $(\texttt{top},v)$ is traversed and $v=\tnode$, then a path from $\snode$ to $\tnode$ exists and the computation halts with `accept`; otherwise, the search continues with $v$ as the new current node. When the right end (`<`) of the tape is reached during the search, that is predicate `rightend` is true, the head is repositioned at the first edge by `move-to-leftend`, and the search continues with the first edge. The main loop cycles over the edges on the tape traversing or skipping edges nondeterministically. An edge takes two positions on the tape, so two right moves skip an edge (shorthand notation `2-right`). Command `move-to-leftend` is assumedly built-in. It can be implemented by commands
`repeat: left; if leftend then right else goto repeat end`
The nondeterministic logic of finding a path in a directed graph is straightforward: Follow all paths starting from $\snode$ and accept if $\tnode$ is reached. The nondeterministic program describes how to follow all paths in a pushdown computation model without concern for efficiency and termination (, cycles in graph). The control is separate (in the simulator). Both of the simulation methods for nondeterministic pushdown automata [@AHU:68; @Glueck:16:pdasim] perform a universal search in the space of nondeterministic computations (find all computation sequences), even though an existential search (halt after first accept) decides the problem. Both use a control component consisting of memoization for avoiding redundant computations. Together, the pushdown program and the simulator constitute the algorithm for finding a path in a directed graph.
Protocol Security Checking by Pathfinding
-----------------------------------------
`init: push ’0’; right (* push start node, pos 1st edge )`\
`loop: if top = hd (* both nodes match? )`\
` then choice (* make a guess: traverse or skip )`\
` pop; right; (* traverse edge, check identities )`\
` if (hd = ’DX’ top = ’EX’) (hd = ’EX’ top = ’DX’) `\
` (hd = ’DY’ top = ’EY’) (hd = ’EY’ top = ’DY’) `\
` (hd = ’DZ’ top = ’EZ’) (hd = ’EZ’ top = ’DZ’) `\
` (hd = ’MX’ top = ’PX’) (hd = ’M’ top = ’PX’) `\
` (hd = ’MY’ top = ’PY’) (hd = ’M’ top = ’PY’) `\
` (hd = ’MZ’ top = ’PZ’) (hd = ’M’ top = ’PZ’)`\
` then pop (* reduce operator pair to )`\
` else push hd end; (* trace unreducible operator )`\
` right; (* move to next node )`\
` if hd = ’1’ bottom then accept end; (* insecure path )`\
` push hd; right (* push next node, pos next edge )`\
` or`\
` 3-right end; (* skip edge )`\
` else 3-right end; (* mismatch: skip edge )`\
` if rightend then move-to-leftend end; (* return to 1st edge )`\
` goto loop`
\[fig:PDAverify\]
To verify ping-pong protocols, we extend the nondeterministic pathfinding program (Fig. \[fig:PDApathfind\]) to operator-labeled graphs that represent the security problem of a ping-pong protocol (, Figs. \[fig:FSAprotocolone\] and \[fig:FSAprotocoltwothree\]). Given an *operator-labeled directed graph* $G = (V, E)$ with nodes $u,v \in V$ and directed edges $E = \{(u_1,o_1,v_1),$ $\dots,$ $(u_n,o_n,v_n)\}$ labeled with the operators $o_i \in \Sigma_{\mathit{XYZ}} = \Sigma_X \cup \Sigma_Y \cup \Sigma_Z$ of Def. \[def:opsets\], the task of the pushdown program is to decide whether there exists a path from a source node $s \in V$ to a target node $t \in V$ in $G$ along which the operator word is reducible to $\epsilon$ by the operator identities of Def. \[def:opids\].
As above, let the names of the source and target nodes be fixed as $s=\snode$ and $t=\tnode$. Assume the node names in $V$ and the operator names in $\Sigma_{\mathit{XYZ}}$ are included in the tape- and stack-symbol alphabets of the automaton. $G$ can be represented by a tape of length $O(n)$ that lists all edges in $E$: $$\texttt{>}~u_1~o_1~v_1~\ldots~u_n~o_n~v_n~\texttt{<}$$ An edge is represented by three symbols on the tape. As with pathfinding, the main loop of the pushdown program cycles over the edges on the tape and guesses a path from $\snode$ to $\tnode$, if it exists. In addition, the program collects the operators labeling the edges along the path and tries to reduce the operator word to $\epsilon$ by applying the operator identities. The program is shown in Fig. \[fig:PDAverify\]. The main parts are as follows:
- Main loop with nondeterministic choice to guess a path.
- Applying the operator identities.
The current node of the path being explored is kept on top of the stack while searching for matching edges. The operators that could not be reduced so far along the current path are on the stack below the current node. The task of the main loop is to find edges originating in the current node by cycling over the edges on the tape. An edge takes three positions on the tape, so three right moves skip an edge (shorthand notation 3-right). When an edge $(\texttt{top},o,v)$ originating in the current node `top` is nondeterministically selected for traversal by `choice`, the current node is popped and an attempt is made to reduce the operator $o$ labeling the selected edge and the last unreduced operator $o_\textit{top}$, now on top of the stack, by trying all operator identities of Def. \[def:opids\]. If an operator identity applies, $o~o_\textit{top} = \epsilon$, then $o_\textit{top}$ is popped from the stack in the then-branch of the second conditional `if` (, `hd = ’MX’ `$\wedge$` top = ’PX’` in the test). Otherwise, the new operator $o$ is pushed on the stack in the else-branch because the operator pair is not reducible. This is the rightmost, innermost reduction strategy discussed in Sect. \[sec:pingpong\]. Next, $v$ is pushed and the search continues with $v$ as the new current node, unless $v=\tnode$ and the stack is empty. If a path exists from $\snode$ to $\tnode$ along which all operators can be reduced to $\epsilon$, which means the stack is empty (predicate `bottom` is true) at the target node, the program halts with `accept`. This tells us that the protocol is *insecure*. The protocol is *secure* if no path from $\snode$ to $\tnode$ is labeled with a reducible word, that is, the input is rejected. The program defines how to search for an insecure path using a nondeterministic choice.
#### Tape Representation.
The representation of the protocol graphs in Figs. \[fig:FSAprotocolone\] and \[fig:FSAprotocoltwothree\] on the tape of the pushdown program is shown in Fig. \[fig:FSAtape\]. The node names (`0`, `1`, ...) and operator names (`EX`, `EY`, ...) are included in the tape symbols, so each takes one position on the tape. The length of the tape is $O(|E|)$ given a graph $G=(V,E)$. The representation of the 18 edges of Protocol 1 takes 56 symbols, the 23 edges of Protocol 2 take 71 symbols, and the 28 edges of Protocol 3 take 86 symbols including the two endmarkers. Names are introduced for intermediate nodes because each edge can only be labeled by a single operator. For example, the initial word of user $X$ in Protocol 3 between source node `0` and target node `1` takes three labeled edges with two intermediate nodes, which we call `8` and `9`.
(If node names are encoded as numbers in a fixed set of tape symbols (, $\texttt{0}, \texttt{1}$), the tape length would be $O(|E|\cdot\log\:|V|)$, and the program would need to be adapted to deal with the encoding and separators between the edges on the tape, , edge $(3, E_X, 5) = \texttt{1 1 EX 1 0 1 \$}$.)
\
`> 0 EY 1 `\
` 1 EX 1 1 PX 1 1 MX 1 1 EY 1 1 PY 1 1 MY 1 `\
` 1 EZ 1 1 PZ 1 1 MZ 1 1 M 1 1 DZ 1`\
` 1 DX 2 2 EY 1 2 EZ 1 1 DY 3 3 EX 1 3 EZ 1 < `\
\
\
`> 0 PX 8 8 EY 1 `\
` 1 EX 1 1 PX 1 1 MX 1 1 EY 1 1 PY 1 1 MY 1 `\
` 1 EZ 1 1 PZ 1 1 MZ 1 1 M 1 1 DZ 1`\
` 1 DX 2 2 MY 5 5 EY 1 2 MZ 6 6 EZ 1 `\
` 1 DY 3 3 MX 4 4 EX 1 3 MZ 7 7 EZ 1 <`\
\
\
`> 0 EY 8 8 PX 9 9 EY 1 `\
` 1 EX 1 1 PX 1 1 MX 1 1 EY 1 1 PY 1 1 MY 1 `\
` 1 EZ 1 1 PZ 1 1 MZ 1 1 M 1 1 DZ 1`\
` 1 DX 2 2 MY 4 4 DX 11 11 EY 1 2 MZ 5 5 DX 12 12 EZ 1 `\
` 1 DY 3 3 MX 6 6 DY 10 10 EX 1 3 MZ 7 7 DY 13 13 EZ 1 <`
\[fig:FSAtape\]
#### Experiments and Implementation.
The pushdown program in Fig. \[fig:PDAverify\] defines a 1-head, 2-way and nondeterministic pushdown automaton (1-head 2NPDA). This is easy to see because the program has a single head variable (`hd`), moves in both directions (`left`, `right`), and has a nondeterministic choice (`choice`). Because a $k$-head 2NPDA can be simulated in at most $O(n^{3k})$ steps where $n$ is the length of the tape, checking the security of ping-pong protocols by a $1$-head 2NPDA takes at most $O(n^{3})$ steps. The pushdown simulator [@Glueck:16:pdasim] takes a set of transition rules as the definition of a pushdown automaton. It takes 690 transition rules to define the program shown in Fig. \[fig:PDAverify\] as a 1-head 2NPDA with 8 control states, 30 tape symbols, and 30 stack symbols. The number of transition steps and surface configurations (state $\times$ stack-top symbol $\times$ tape symbol) the simulator takes to verify the three Dolev-Yao protocols are listed in Table \[tab:simruntape\]. The simulator looks for a universal solution, exploring all computation sequences leading to an accept (insecure protocol) and not halting at the first accept being found. Thus, the outcome and the order of the edges on the tape have no significant influence on the performance of the simulation.
*verification* *edges* *tape* *configs* *steps* *answer*
---------------- --------- -------- ----------- --------- -------------------
Protocol 1 18 56 584 8100 accept (insecure)
Protocol 2 23 71 740 6031 reject (secure)
Protocol 3 28 86 1184 11412 accept (insecure)
: Simulation of the pushdown program as 2NPDA with the Dolev-Yao example protocols.
\[tab:simruntape\]
#### Discussion.
The nondeterministic pushdown verifier is surprisingly simple (in the author’s opinion), especially when considering that Dolev and Yao’s first algorithm took $O(n^8)$ steps [@DolevYao:81; @DolevYao:83], while the verifier in Fig. \[fig:PDAverify\] is guaranteed to take at most $O(n^3)$ steps thanks to fundamental results of automata theory. The verifier checks whether the intersection of a regular language represented on the tape as FSA (the protocol to verify) and a fixed Dyck-like language containing all canceling (insecure) words is empty. Clearly, the verifier is not limited to protocol graphs constructed by the algorithm [@DolevEvenKarp:82]. Any operator-labeled FSA can be placed on the tape and the operator identities tested in the program can be adapted easily to other identities. Thus, any security question that can be captured by an operator-labeled FSA (any regular language) intersected with a Dyck-like language induced by a fixed set of operator identities can be decided by the verifier after adaption to the specific operator identities.
The experiment also supports the proposition that programming languages can make abstract theoretical results more accessible and applicable [@Jones:97:complexity; @Reus:16] (other examples are reversible programming languages [@AxGl11FoSSaCS]). The simulation approach is also amenable to optimization by program specialization, which typically reduces the interpretive overhead by an order of magnitude. However, a memoizing interpreter for a nondeterministic language may pose additional challenges to non-trivial specialization. A downside, from a programming perspective, is that the pushdown language has no convenient data structures, just a linear tape with symbols. A search for matching edges needs to cycle over all edges on the tape and to return the head from the right end to the left end.
Protocol Verification by Transformation and Pushdown Simulation {#sec:vptsim}
===============================================================
Verification by general-purpose program transformers is another approach that has been proposed and used successfully to check the security of cryptographic protocols by supercompilation [@AhmedLisitsaNemytykh:13; @Nepeivoda:16].
Typically, the security problem of a protocol (, a ping-pong protocol) is encoded as a functional program with an additional trace parameter that constrains all nondeterministic choices such that they become deterministic. Given a trace, the functional program maps an initial state into a final state which can then be tested for validity (, whether the state is insecure). By specializing the program a static (known) initial state and a dynamic (unknown) trace, a supercompiler will explore the control flow of all possible traces for the given initial state.[^3] If the program is specialized into a residual program from which it is immediately seen that no trace can steer the residual program into an invalid final state then the original protocol is considered secure. A difficulty with this approach is the preparation of a functional program that specializes well, which may require considerable knowledge about supercompilation. There are obstacles, namely finding the right encoding of the security problem in a universal source language and taming the power of the supercompiler with folding, generalization and other sophisticated optimizations, especially if success depends on the way the program is written. The nondeterminism inherent in the security problem is first mapped into a deterministic program and then reintroduced by specializing the deterministic program a dynamic trace, which can make the transformation hard to predict. A nondeterministic choice in the original problem does not necessarily correspond to a nondeterministic (dynamic) choice taken by the supercompiler when exploring all possible control flows in the program. The control over these choices is therefore indirect. On the other hand, case studies [@AhmedLisitsaNemytykh:13; @LisitsaNemytykh:07] have shown that a supercompiler [@Turchin:86] can solve a variety of security problems, including cryptographic ping-pong protocols [@Nepeivoda:16]. In our case study, the protocol verifier is a pushdown program that explores all paths in a protocol graph by nondeterministic means. The security problem is not encoded as a functional program, but given as input data (Fig. \[fig:FSAtape\]) to the verifier (Fig. \[fig:PDAverify\]). The verifier is the same for all protocols. A branching in the graph on the tape is directly modeled by a nondeterministic choice in the pushdown program. The verifier has exactly one nondeterministic choice point; all other choices in the program are deterministic (`if`). The nondeterministic `choice` is an integral part of the language semantics and gives direct control of the problem-specific nondeterminism inherent in the security problem. The verifier is written in a language for which decidability is guaranteed, which is a major advantage of this approach. To the best of the author’s knowledge, this is the first study in which the pushdown programming model has been applied to ping-pong protocols, but the model’s practicality for verifying a larger class of protocols has not yet been demonstrated. At least in principle, the multihead pushdown programs can decide any computationally “tractable” verification problem. Both approaches explore all possible paths of a security problem regardless of its representation. In one approach, this is achieved by a supercompiler specializing a program representing the problem and building an internal process graph, whereas in the other approach it is achieved by a simulator of pushdown programs using memoization. The nondeterministic choice is an integral part of the pushdown language semantics, while in the case of supercompilation the nondeterminism is induced into the source program by the supercompiler’s non-standard “transformation semantics” [@AbrGlu:00:nsint], not the standard semantics of the source language. Neither approach shows why a protocol is insecure or how to fix it, but a step in this direction has been taken by building attack models [@Nepeivoda:16].
*method* SIM SCP
-------------------------- ------------------------ ----------------
*approach* interpretation transformation
*sound answer* yes yes
*complete answer* yes case by case
*time complexity class* $=$ polynomial exponential?
*space complexity class* $\subseteq$ polynomial unknown
subuniversal universal
*source language* nondeterministic deterministic
tail-recursive recursive
: Pushdown simulation and supercompilation characteristics.
\[tab:SIMSCPprop\]
Table \[tab:SIMSCPprop\] summarizes the two approaches. The entries in the table are for a general-purpose supercompiler [@Turchin:86], bearing in mind there are different supercompiler variants (, [@LisitsaNemytykh:07; @Nepeivoda:16; @sorm99:intro]). The simulator (SIM) interprets a pushdown program for a given protocol graph (sequence of edges) on the input tape and the supercompiler (SCP) transforms the program representation of the security problem into a residual program. Provided that SIM is correct, the answer (accept/reject) is sound and complete (multihead 2NPDA are decidable). Provided that SCP is correct, the generated residual program is a sound answer, but depending on the particular protocol encoding, the residual program may not always answer the security question (, due to overgeneralization it contains valid and invalid final states even though the protocol is secure, or due to infinite specialization no program is generated at all). Thus, completeness of the answer is marked as ‘case by case’ in the table. (An exception is the verification of ping-pong protocols which was shown to be decidable for all multi-party ping-pong protocols [@Nepeivoda:16].) Exponential time complexity was conjectured for ping-pong protocol verification by supercompilation [@Nepeivoda:16], while the space complexity class is unknown. The time and space complexity classes of multihead 2NPDA are guaranteed [@WagnerWechsung:86]. The source language of SIM is a subuniversal (not Turing-complete), nondeterministic and tail-recursive pushdown language (Sect. \[sec:ndetpgmoverview\]), while the source language of SCP is a universal (Turing-complete), deterministic and recursive first-order functional language. The deterministic and nondeterministic property of their source languages is also indicated in Fig. \[fig:VERIFYapproaches\].
Related Work {#sec:related}
============
It has been know for several decades that nondeterministic programs are well suited for combinatorial search problems, and in many cases even easier to write than deterministic ones [@Floyd:67]. An obstacle is the effective control of the often exponential-time complexity of straightforward runs and nonterminating computation paths. For two-way nondeterministic pushdown automata, a polynomial-time and terminating bottom-up simulation algorithm could be given [@AHU:68]. Another simulation algorithm [@Glueck:16:pdasim] follows top-down, all reachable computation paths, as does the one for two-way deterministic pushdown automata [@Jones:77]. In these algorithms, exponential time is converted into polynomial time by sharing computations. A linear-time simulation by instrumented two-way deterministic pushdown programs is given in [@Mogensen:94]. Certain methods of model checking [@ABE:16:handbook] also make use of pushdown systems. A classic case where investigations in pure theory of pushdown automata led to a practically significant algorithm is pattern matching [@KnuMorPra:77]. The case study presented here appears to be one of the first to show how a nondeterministic pushdown language can be used as a decidable programming model for protocol verification.
A series of case studies examined the verification of protocols by program transformation (, [@LisitsaNemytykh:07]), and in particular cryptographic protocols by supercompilation (, [@AhmedLisitsaNemytykh:13; @Nepeivoda:16]). A few orthogonal supercompilation principles [@Turchin:86] were shown to solve a number of seemingly different verification problems. Related approaches for specific verification problems have been investigated (, [@GlueckLeuschel:99:PSI; @Klimov:12]) and comparable results can be conjectured (, [@FKG:01:NGC]). Interpreters have been used to improve the transformation of programs (, [@Glueck:94:JFP; @GlueckJoergensen:94:SAS]), which is another way to factorize supercompilation-based verification. The literature on transformation-based verification is larger than the one cited here, including approaches based on unfold/fold rules (, [@FioravantiPettorossiProietti:02]). The present study examined a related, but different programming language solution, namely nondeterministic programs in a decidable computation model with guaranteed resource bounds. Related principles underlying both approaches, verification by transformation and interpretation, were discussed above (Sect. \[sec:vptsim\]). Extensions of logic languages with tabulation can ensure termination and optimal known complexity for queries to a large class of practical programs [@SwiftWarren:12].
Conclusions and Further Work {#sec:conclusion}
============================
This study broadens previous studies on the verification of security by program transformation in that another programming language approach, namely program interpretation, is used. We confirmed that the security of a well-studied class of cryptographic protocols can be verified by a 1-head 2NPDA. The interpretive approach used in this experiment considerably simplified the verification, by separating nondeterministic pushdown logic from control concerns, which shows again the power of a declarative style of programming.
Program transformation and interpretation are two sides of the same coin, and we identified principles that both verification approaches share (, explore all possible paths of a security problem). Also noteworthy is how abstract results from automata theory can be applied to practical problems when combined with a programming language approach, and that this can yield more natural and simple solutions. This study is one of the few examples in the literature where pushdown automata have been used to answer questions other than those of formal language theory. This situation is perhaps surprising because the multihead 2NPDA programming model is equivalent to the class of polynomial-time algorithms and decidable within guaranteed time and space bounds determined by the number of heads. Though we showed how a class of cryptographic protocols can be verified by nondeterministic programming, further work is needed before a more complete picture emerges as to the practicality of the interpretive pushdown approach. The pushdown computation model is subuniversal, and thus cannot be expected to be capable of solving all verification problems in reach of a general-purpose program transformer, such as a supercompiler using sophisticated generalization techniques and capable of generating complex recursive programs as answers. On the other hand, the multihead nondeterministic pushdown model is theoretically powerful enough to decide all polynomial-time verification problems, but whether this is as straightforward as in the case of ping-pong protocols, only further investigations will show. Verification of multi-party extended protocols [@Nepeivoda:16] in further studies is warranted.
From a programming perspective, the “machine-code” transition rules of classic presentations of pushdown automata are too low-level. A step towards a more user-friendly abstraction was undertaken in this paper by employing an imperative flowchart language with deterministic and nondeterministic control-flow operators. Still, the language inherits the linear input tape from automata theory. More practical data structures and languages abstractions could be considered, arrays and index calculations [@Mogensen:94], tree-structured data or cons-free functional programming languages [@Jones:97:complexity].
#### Acknowledgements.
Thanks to Antonina Nepeivoda for a concise explanation of the Dolev-Yao model and to the anonymous reviewers for their constructive feedback.
[10]{} \[2\] \[1\][`#1`]{} \[2\][`#2`]{} \[2\][[\#2](#1)]{} \[1\][doi:]{} \[2\][\#2]{}
& (): **. In & , editors: [ ** ]{}, , , pp. , .
, & (): **. In & , editors: [ ** ]{}, [**]{} , pp. , .
, & (): **. (), pp. , .
, & (): **. In , , & , editors: [ ** ]{}, . .
& (): ** In , editor: [**]{}, , , pp. , .
(): **. In , & ,editors:[ ** ]{}, ,pp..
, & (): **. (), pp. , .
& (): **. In: [**]{}, , pp. , .
& (): **. (), pp. , .
, & (): **. In , editor: [**]{}, , , pp. , .
(): **. (), pp. , .
, & (): **. (), pp. , .
(): **. (), pp. , .
(): **. In & , editors: [ ** ]{}, , , pp. , .
& (): **. In , editor: [**]{}, , , pp. , .
& (): **. In , & , editors: [**]{}, , , pp. , .
(): **. (), pp. , .
(): **. , , .
(): **. In , & , editors: [ ** ]{}, , , pp. , .
, & (): **. (), pp. , .
(): **. (), pp. , .
& (): **. (), pp. , .
(): **. (), pp. , .
(): **. (), pp. , .
(): **. , .
, & (): **. (), pp. , .
& (): **. In , & , editors: [ ** ]{}, , , pp. , .
& (): **. (), pp. , .
(): **. (), pp. , .
& (): **. , .
[^1]: We use the reversed and simplified version of the FSA constructed by the algorithm.
[^2]: $\mathcal{L}(\Gnterm)$ is an ambiguous language: parenthesizing is not necessarily unique, `` or ``.
[^3]: The trace program may have additional parameters depending on the particular class of protocols and the desired answer.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article we study the Hofer geometry of a compact Lie group $K$ which acts by Hamiltonian diffeomorphisms on a symplectic manifold $M$. Generalized Hofer norms on the Lie algebra of $K$ are introduced and analyzed with tools from group invariant convex geometry, functional and matrix analysis. Several global results on the existence of geodesics and their characterization in finite dimensional Lie groups $K$ endowed with bi-invariant Finsler metrics are proved. We relate the conditions for being a geodesic in the group $K$ and in the group of Hamiltonian diffeomorphisms. These results are applied to obtain necessary and sufficient conditions on the moment polytope of the momentum map, for the commutativity of the Hamiltonians of geodesics. Particular cases are studied, where a generalized non-crossing of eigenvalues property of the Hamiltonians hold.'
address:
- 'Departamento de Matemática, FCEyN-UBA, and Instituto Argentino de Matemática, CONICET. Buenos Aires, Argentina.'
- 'Instituto Argentino de Matemática, CONICET. Buenos Aires, Argentina.'
author:
- Gabriel Larotonda
- Martín Miglioli
title: 'Hofer’s metric in compact Lie groups'
---
=1
[^1]
Introduction
============
The Finsler length structure on the group $\operatorname{Ham}(M,\omega)$ of Hamiltonian diffeomorphisms of a symplectic manifold $M$ with symplectic form $\omega$, was introduced by Hofer in the paper [@hofer]. The Lie algebra of this group is the set of Hamiltonian vector fields, which can be identified (by means of the symplectic gradient) with the set of Hamiltonian functions in $M$ -modulo constant functions-. The norm of a vector field is then the quotient $L^\infty(M)$-norm of the generating Hamiltonian function (modulo constant functions). A natural problem of current interest in the literature is the study of geodesics in this Finsler manifold. There has been a significant amount of progress, and fairly deep work on the properties of this metric, mostly from the point of view of symplectic topology, see [@bp; @lm95; @lm; @pol01; @ps] and also the textbook [@mds] and the references therein.
In this article we consider (almost) effective Hamiltonian actions of compact semi-simple Lie groups $K$ on a symplectic manifold $(M,\omega)$. This action defines an inclusion (modulo the discrete kernel of the action) $K\hookrightarrow \operatorname{Ham}(M,\omega)$, which allows us to introduce a Hofer metric on $K$ by means of $$\|x\|_{\mu(M)}=\max_{y\in\mu(M)}\langle y,x\rangle-\min_{y\in\mu(M)}\langle y,x\rangle.$$ Here $x\in \k=T_1K$ and $\mu$ is the momentum map of the Hamiltonian action. For this norm, the intersection of its unit ball with a maximal abelian subalgebra $\t\subseteq \mathfrak k$ is the polar dual of a certain polytope $P$, which is derived from the moment polytope of the action $\mu$. The rectifiable distance in $K$ is defined as the infima of the lenghts of paths joining given endpoints, and we define a *geodesic* in $K$ as a distance minimising path.
We first show that one-parameter groups in $K$ are always geodesics for this distance, provided the speed is in the domain of injectivity of the exponential map of $K$. These paths are mapped to paths in $\operatorname{Ham}(M,\omega)$ with autonomous Hamiltonian. Since the norm is not strictly convex, it is natural to expect other geodesics in $K$, and this is the case for all these norms; we show that these geodesics correspond to quasi-autonomous Hamiltonianans. The extreme points of the polytope $P\subseteq \k$ encode the privileged directions that the speeds of geodesics can take. Relevant geometrical properties of the geodesics can be expressed in terms of these extreme points. Of particular relevance are the polytopes with only regular extreme points in $\k$, which are fully characterized both in terms of the Lie algebra (by polar duality) and in terms of the geometry of geodesics in $K$. We obtain similar results for the pull-back of the one-sided Hofer norm, which is usually only positively homogeneous: $$\|x\|_{\mu(M)}^+=\max_{y\in\mu(M)}\langle y,x\rangle.$$ In fact, the general results on geodesics in this article are stated for any Finsler length structure given by (left or right) translation of an $\operatorname{Ad}$-invariant Finsler norm in $\k$, therefore we include non-symmetric distances in our discussion. It is worthile mentioning here that the notion of majorization of real vectors $\overrightarrow{v}\prec \overrightarrow{w}$ (identified with the eigenvalues of the operators $\operatorname{ad}v, \operatorname{ad}w$, where $v,w\in \k$) plays a significant role in the proofs concerning minimality of geodesics, and it is related to the condition $v\in \operatorname{conv}(\O_w)$, where the later set is the convex closure of the coadjoint orbit of $w$ in $\k$.
There are several relevant applications related to this setting of actions of compact Lie groups: for instance, as shown in [@entov], the geometry of the canonical Hamiltonian action $\operatorname{SU}(n)\to \operatorname{Ham}(Gr_{r,n},\omega)$ can be used as a tool to study the eigenvalue inequalities in the quantum version of Horn’s problem [@belkale]. Here $Gr_{r,n}$ is the Grassmannian of $r$-dimensional planes in $\C^n$, and $\omega$ is the canonical Kirilov-Kostant-Souriau symplectic form. We plan to extend some of the results in this article to the case of infinite dimensional groups using ideas connected to the results in [@blz] and [@lar19]. Some of the techniques developed in this article might also be relevant to the study of Finsler length structures derived from mechanics with non-smooth energy.
The article is organized as follows: in Section \[sectionnorms\] we define generalized Hofer norms. We study the faces and norming functionals of the unit balls of these norms, based on two different theories. We first analyse the structure of these balls with functional analytic techniques via an embedding in certain function spaces, an approach that will be useful in the study of the stability under geodesy at the end of Section \[sectionquasiaut\]. We also study these norms using results from convex geometry.
In Section \[hofergroup\] we recall basic results on Hamiltonian actions and Hofer’s metric on groups of Hamiltonian diffeomorphisms, and we pull-back these metrics to compact groups using the homomorphism $K\to \operatorname{Ham}(M,\omega)$. These are the motivations and main examples for the norms and length structures on groups studied in this article. Nevertheless, this part of symplectic geometry is not necessary for the understanding of several results in the article which are solely based on convexity and Finsler length structures.
In Section \[convexgeocart\] a characterization of the intersection of the unit balls of the Hofer norms with maximal abelian algebras is given based on group invariant convex analysis and symplectic convexity theorems.
In Section \[l\] we first recall several results obtained in [@lar19] for groups endowed with Finsler length structures obtained from $\operatorname{Ad}$-invariant norms which are valid for the groups studied in this article. Then we prove several global results on geodesics in the case of finite dimensional groups endowed with continuous Finsler metrics. We show that geodesics in groups $K$ with Hofer’s metric are quasi-autonomous, which provides a link between the conditions for length minimization in $K$ and the corresponding conditions in $\operatorname{Ham}(M,\omega)$.
Finally, in Section \[geodesicommute\] we study actions of groups with commuting Hamiltonians. We start with the important special case of actions on regular coadjoint orbits and related groups: the Hamiltonians of length minimizing curves have the interesting feature of “non-crossing of eigenvalues”. We characterize the $\operatorname{Ad}$-invariant norms such that in groups with Finsler structures defined from these norms, all geodesics have commuting speeds. Based on this result we characterize the compact groups of Hamiltonian diffeomorphisms such that length minimizing curves have commuting Hamiltonians. We then show how conditions on Kirwan’s polytope can characterize this property. The paper ends with a study of how these properties behave when we consider the direct product of Hamiltonaian actions. It is proved then that it suffices to have geodesics with commuting Hamiltonians for one of the actions, to obtain the same property for the geometry induced in $K$ by the direct product of actions.
The generalized Hofer norm and its convex geometry {#sectionnorms}
==================================================
In this section we define the generalized Hofer norms and we study them with two approaches. In the first we embed the normed space in a quotient of a space of continuous functions and use functional analytic techniques. In the second we use the polar duality from convex geometry.
\[convexb\] A subset $E\subseteq V$ of a vector space $V$ is called *full* if it affinely generates the space. A set $B\subseteq V$ is a *convex body* if $B$ is a compact convex set with non empty interior. If additionally, $B$ is centrally symmetric ($v\in B\Rightarrow -v\in B$), then it is called a *symmetric convex body*. Equivalently, a symmetric convex body is a convex balanced absorbing set in $V$.
\[iota\] Let $(V,\langle \cdot,\cdot\rangle)$ be a finite dimensional inner product space and let $E\subseteq V$ be a compact full subset. Consider the norm $\|\cdot\|_E$ on $V$ given by the embedding $$\iota:V\hookrightarrow C(E)/\R{\mathbf 1}, \quad x\mapsto [\varphi_x]:=\varphi_x+\R{\mathbf 1},$$ where $\varphi_x(y)=\langle x,y\rangle$ for $y\in E$. Then $$\label{generhofernorm}
\|x\|_E=\max_{y\in E}\varphi_x(y)-\min_{y\in E}\varphi_x(y)=2\|[\varphi_x]\|_{\infty},$$ where $\|[\varphi_x]\|_{\infty}=\inf\{\|\varphi_x-\lambda {\mathbf 1}\|_{\infty} :\lambda\in\mathbb R\}$ is the quotient norm. We call $\|\cdot\|_E$ a *generalized Hofer norm*.
This is a norm since $E$ is full. If there is a group acting isometrically on $V$ and leaving the set $E$ invariant then the action is also isometric for the norm $\|\cdot\|_E$.
\[otrasnorm\] Another (semi)-norm which is invariant for isometric actions on $V$ that we will consider is the *generalized second Hofer norm* given by the supremum norm $$\label{generhofernorm2}
\|x\|'_E= \max\{\max_{y\in E}\varphi_x(y),-\min_{y\in E}\varphi_x(y)\}=\|\varphi_x\|_{\infty}.$$ It is also of interest to consider a third invariant norm which is only a Finsler norm (i.e. only positively homogeneous), given by $$\|x\|^+_E=\max_{y\in E}\varphi_x(y).$$ We will call this the *one-sided Hofer norm*. This defines a Finsler norm for a bounded set $E\subseteq V$ such that its convex hull contains $0$ in its interior, or equivalently, such that the cone generated by $E$ is all of $V$. It also possible to consider $$\|x\|^-_E=-\min_{y\in E}\varphi_x(y).$$ but note that this one can be obtained from the previous by replacing $E$ with $-E$. Clearly $\|x\|'_E=\max \{\|x\|_E^+,\|x\|_E^-\}$ is the second Hofer norm and $\|x\|_E=\|x\|_E^++\|x\|_E^-$ is the first Hofer norm.
Maximal faces and norming functionals
-------------------------------------
We begin studying the faces of the sphere and norming functionals of Hofer’s norm by relating this norm on the space $V$ to the norm on the much larger space $C(E)/\R{\mathbf 1}$ (where $E$ is a compact full set) by means of the embedding $\iota:V\hookrightarrow C(E)/\R{\mathbf 1}$ given by $x\mapsto \varphi_x+\R{\mathbf 1}$, where $\varphi_x=\langle x,\cdot\rangle$.
Let $V$ be a normed space and denote $\|\varphi\|=\sup\{\varphi(v):\|v\|=1\}$ for $\varphi\in V^*$. The dual space $V^*$ with this norm is a Banach space. We say that $\varphi\in V^*$ is a *norming functional* of $v\in V$ if $\varphi(v)=\|v\|$ and $\|\varphi\|=1$. A functional $\varphi$ is *extremal* if $\varphi$ is an extreme point of the unit ball $B_{V^*}$. If $\|\cdot\|$ is only positively homogeneous (to remark it we say that it is a *Finsler norm*) the same definitions apply, and the norm given to $V^*$ is only positively homogeneous. In any case we refer to it as the *dual norm*.
$\S$ Note that the difference bewtween the ball of a norm and that of a Finsler norm is that the last one might not be balanced (i.e. symmetric). In both cases, it is an absorbing, open convex set containing $0\in V$.
A *face* $F$ of the unit ball $B_V$ of a normed space $V$ is the intersection of the unit ball $B_V$ with the hyperplane determined by a unit norm functional $\varphi\in V^*, \|\varphi\|=1$, i.e. $$F_{\varphi}=B_V\cap\{v\in V:\varphi(v)=1\}.$$ We say that the face is *maximal* if $\varphi$ is extremal. Every face is contained in a maximal face: if $\varphi$ is a unit norm functional then $\|\varphi\|=1$ and since $B_{V^*}$ is compact and convex there exists by the Krein-Milman theorem extremal functionals $\{\varphi_i\}_{i=1,\dots,n}\subseteq B_{V^*}$ such that $\varphi$ is a convex combination of the $\varphi_i$: $$\varphi=\sum_i \lambda_i \varphi_i, \quad \lambda_i\ge 0,\quad \sum_i \lambda_i=1.$$ It is then easy to check that if $\varphi(v)=\|v\|$ then $\varphi_i(v)=\|v\|$ for all $i$. Therefore if $v\in F_{\varphi}$, $v\in F_{\varphi_i}$ for all $i$ and in fact $F_{\varphi}$ is the intersection of all the maximal faces that contain it.
The *cone generated* by a face $F_{\varphi}$ is $\R_+ F_{\varphi}$. Note that this cone consists exactly of those $v\in V$ such that $\varphi(v)=\|v\|$.
The following elementary characterization will be useful:
\[mismacara\] In a vector space $V$, $\|v_1+\dots +v_n\|=\|v_1\|+\dots+\|v_n\|$ holds if and only if $v_1,\dots,v_n$ belong to the cone generated by a face.
If the $v_i$ are in the cone of $\varphi$, then $$\|v_1\|+\dots+\|v_n\|\ge \|v_1+\dots+v_n\|\ge \varphi(v_1+\dots +v_n)=\|v_1\|+\dots + \|v_n\|.$$ On the other hand, if $\|v_1+\dots +v_n\|=\|v_1\|+\dots+\|v_n\|$ holds, by means of Hahn-Banach’s theorem pick a unit norm functional that norms the sum of the $v_i$, that is $\varphi(\sum_i v_i)=\|\sum_i v_i\|$. Then $$\|\sum_i v_i\|=\varphi(\sum_i v_i)=\sum_i \varphi(v_i)\le \sum_i \|v_i\|=\|\sum_i v_i\|$$ since $\varphi(v_i)\le \|v_i\|$ for all $i$; this is only possible if equality holds for each $i$. Therefore $\varphi$ is a norming functional for all the $v_i$.
### Norming functionals as Borel measures
Let $X$ be a compact Hausdorff topological space and let $C(X)$ be the continuous real valued functions on $X$. Endow $C(X)/\R{\mathbf 1}$ with (twice) the quotient $L^\infty$ norm (the factor $2$ is there to be consistent with (\[generhofernorm\])). By Riesz-Markov’s theorem, its dual space can be identified with the regular finite Borel signed measures in $X$ such that $\mu(X)=0$ (that is because the identification $\mu\mapsto \varphi_\mu$ is given by integration $\varphi_\mu(f)=\int_X fd\mu$, and we require that $\varphi_\mu({\mathbf 1})=0$). The norm of $\varphi_\mu$ is given by the total variation of $\mu$, therefore unit norm functionals are characterized by having total variation equal to *two*. From here it is immediate the following characterization:
\[normingcont\] For $f\in C(X)/\R{\mathbf 1}$ different from zero, its norming functionals are given by $\varphi=\mu^+-\mu^-$, where $\mu^+$ and $\mu^-$ are probability measures in $X$, supported in $\operatorname{argmax}(f)$ and $\operatorname{argmin}(f)$ respectively. Since the extreme points of the probability measures are the Dirac measures, the maximal faces are given by norming functionals $\varphi=\delta^+-\delta^-$, with delta measures $\delta^+,\delta^-$ supported in $x^+,x^-\in X$ respectively.
\[argcont\] A set $\{f_i\}_{i\in I}\subseteq C(X)/\R{\mathbf 1}$ is a subset of a cone generated by a face if and only if $\cap_{i\in I}\operatorname{argmin}(f_i)\neq\emptyset$ and $\cap_{i\in I}\operatorname{argmax}(f_i)\neq\emptyset$.
If both intersections are non empty, pick $x^-,x^+$ respectively in each of them and consider $\varphi$ in the dual given by $\varphi(f)=f(x^+)-f(x^-)$. Then $\varphi$ is a unit norm functional and $$\sum_i \|f_i\|\ge \|\sum_i f_i\|\ge \varphi(\sum_i f_i)=\sum_i f_i(x^+)-f_i(x^-)=\sum_i\|f_i\|,$$ therefore by Lemma \[mismacara\] the $f_i$ are in the cone generated by the face given by $\varphi$. Reciprocally, if there are, say, $f=f_k$ and $g=f_l$ such that the maximal argument of $f$ does not intersect the maximal argument of $g$, then $\max (f+g)<\max f +\max g$, therefore $$\|f+g\|=\max(f+g)-\min(f+g)<\max f+ \max g-\min f - \min g=\|f\|+\|g\|$$ and the conclusion follows by Lemma \[mismacara\].
For $f,g\in C(X)/\R{\mathbf 1}$ we have $\|f+g\|=\|f\|+\|g\|$ if and only if $\operatorname{argmin}(f)\cap\operatorname{argmin}(g)\neq\emptyset$ and $\operatorname{argmax}(f)\cap\operatorname{argmax}(g)\neq\emptyset$.
Putting together the previous characterizations (and recalling that the norm we are considering is twice the quotient norm), it is clear that
The maximal faces of the ball of $C(X)/\R{\mathbf 1}$ are given by the sets $$F_{x^-,x^+}=\{[f]:\|[f]\|_{\infty}=2,x^-\in\operatorname{argmin}(f),x^+\in\operatorname{argmax}(f)\}$$ for a choice of points $x^-,x^+\in X$.
Since a face of the ball of $V$ is contained in a face of $C(E)/\R{\mathbf 1}$ by means of the map $\iota$ of Definition \[iota\], we have the following result (see Figure \[fig: norming2\]):
\[arghofer\] Let $V$ be a vector space with the norm defined by the map $\iota$ and the compact full set $E\subseteq V$. A set $S\subseteq V$ is a subset of a cone generated by a face if and only if $$\bigcap_{x\in S}\operatorname{argmin}_E(\varphi_{x})\neq\emptyset \quad \textrm{ and }\quad \bigcap_{x\in S}\operatorname{argmax}_E(\varphi_{x})\neq\emptyset.$$
\[conosmaxmin\] Given a compact $E\subseteq V$ and $x^-,x^+\in E$ we define the cone $$C_{x^-,x^+}(E):=\{x\in V:x^-\in\operatorname{argmin}_E(\varphi_{x})\mbox{ and }x^+\in\operatorname{argmax}_E(\varphi_{x})\}.$$
\[maxfacemaxmin\] Each cone generated by a maximal face is equal to $C_{x^-,x^+}(E)$ for some $x^-,x^+\in E$.
Let $\R_+F_{max}$ be the cone generated by a maximal face. By Corollary \[arghofer\] it is contained in $C_{x^-,x^+}(E)$ for some $x^-,x^+\in E$. Since $C_{x^-,x^+}(E)$ satisfies the condition of Corollary \[arghofer\] it is contained in the cone $\R_+F$ generated by a face $F$. By maximality of $F_{max}$ we get $F_{max}=F$ and the conclusion follows.
The cones of Definition \[conosmaxmin\] have good properties with respect to sum of sets.
\[interconosmaxmin\] Let $E_1,\dots,E_n$ be compact sets in $V$ and let $x^-_i,x^+_i\in E_i$ for $i=1,\dots,n$. If we define $E=E_1+\dots + E_n$, $x^-=x^-_1+\dots +x^-_n$ and $x^+=x^+_1+\dots +x^+_n$ then $$C_{x^-,x^+}(E)=\bigcap_{i=1,\dots,n}C_{x^-_i,x^+_i}(E_i).$$
It is easy to verify that that for $x\in V$ the functional $\varphi_x$ has a maximum at $x^-_i,x^+_i\in E_i$ in $E_i$ for $i=1,\dots,n$ if and only if it has a maximum at $x^-_1+\dots +x^-_n$ in $E_1+\dots + E_n$. The same holds for the minimizers and the proof follows.
\[hofer3argmax\] Similar results can be obtained for the one-sided Hofer norm $\|\cdot\|_E^+$ (Remark \[otrasnorm\]), for a compact set $E$ such that its connvex hull contains $0$ in its interior. A set $S\subseteq V$ is a subset of a cone generated by a face if and only if $$\bigcap_{x\in S}\operatorname{argmax}_E(\varphi_{x})\neq\emptyset.$$ Given a compact $E\subseteq V$ and $x^+\in E$ we define the cone $$C_{x^+}(E):=\{x\in V:x^+\in\operatorname{argmax}_E(\varphi_{x})\}.$$ Each cone generated by a maximal face is equal to $C_{x^+}(E)$ for some $x^+\in E$. These cones have also good properties with respect to sum of sets. Let $E_1,\dots,E_n$ be compact sets in $V$ and let $x^+_i\in E_i$ for $i=1,\dots,n$. If we define $E=E_1+\dots + E_n$ and $x^+=x^+_1+\dots +x^+_n$ then $$C_{x^+}(E)=\bigcap_{i=1,\dots,n}C_{x^+_i}(E_i).$$
We now characterize norming functionals for the Hofer norm, see Figure \[fig: norming\] and Figure \[fig: norming2\]. The convex hull of a set $X$ is denoted by $\operatorname{conv}(X)$.
\[normingfunc\] The norming functionals of $x \in (V,\|\cdot\|_E)$ are $\varphi_{y^+ -y^-}$, with $$y^+ \in \operatorname{conv}(\operatorname{argmax}_E \varphi_x)\quad\textrm{ and }\quad y^-\in\operatorname{conv}(\operatorname{argmin}_E \varphi_x).$$
Suppose first that $y^+ \in \operatorname{conv}(\operatorname{argmax}_E \varphi_x)$ and $y^-\in\operatorname{conv}(\operatorname{argmin}_E \varphi_x)$, then $\varphi_x(y^+)=\max_E(\varphi_x)$ and likewise with $y^-$. It is immediate from the definitions that $\|\varphi_{y^+-y^-}\|\le 1$, and on the other hand $$\varphi_{y^+-y^-}(x)=\langle y^+-y^-,x\rangle=\max_E(\varphi_x)-\min_E(\varphi_x)=\|x\|_E,$$ thus $\varphi_{y^+-y^-}$ has unit norm and it is norming for $x$. Suppose now $\varphi$ is a norming functional of $x$ in $V\simeq \iota(V)\subseteq C(E)/\R{\mathbf 1}$. We can extend it by the Hahn-Banach theorem to all $C(E)/\R{\mathbf 1}$, so that it is given by integration with $\mu^+-\mu^-$ for probability measures $\mu^+,\mu^-$ supported in $\operatorname{argmax}(\varphi_x)$ and $\operatorname{argmin}(\varphi_x)$ respectively. Hence $$\begin{aligned}
\varphi(z)&=\int_E\langle w,z\rangle\, d(\mu_1-\mu_2)(w)\\
&=\langle \int_E w\,d\mu_1(w) -\int_Ew\,d\mu_2(w),z\rangle\\
&=\langle \operatorname{cent}(\mu_1) -\operatorname{cent}(\mu_2),z\rangle,\end{aligned}$$ where $\operatorname{cent}(\mu)$ denotes the center of mass of the probability measure $\mu$. The result follows if we take $y^+=\operatorname{cent}(\mu_1)$ and $y^-=\operatorname{cent}(\mu_2)$.
In Theorem \[normingpolar\] we will give another proof of the previous result, based on polar duality.
Convex sets, polar duality and Minkowski norms
----------------------------------------------
We present some basic results on convex geometry and polar duality which will be used in this section to characterize the Hofer norms. We refer to Chapter I of [@bron] for basic results on convex sets and Chapter II of the same book for basic results on convex polytopes. Another reference is [@bar]. Let $V$ be a finite dimensional inner product space. For a non-zero vector $x\in V$ and a scalar $a\in\R$ we define the hyperplane $$H_{x,a}=\{y\in H:\langle x,y\rangle=a\}.$$ For $a=1$ we set $H_x:=H_{x,1}$. We define the negative halfspace as $$H_{x,a}^-=\{y\in H:\langle x,y\rangle\leq a\}.$$ The *polar duality* is given by the following bijection between non-zero points in $V$ and hyperplanes in $V$ not containing zero: $x\mapsto H_x$.
\[defih\] The *support function* of a bounded subset $E\subseteq V$ is the function $$h_E:V\to\R,\quad h_E(u)=\sup_{x\in E}\langle x,u\rangle.$$ Note that $h_E=h_{\operatorname{conv}(E)}$, and that if $\operatorname{conv}(E)$ contains $0$ in its interior, then $h_E$ is a Finsler norm, our one-sided Hofer norm (Remark \[otrasnorm\]).
If $0\neq x\in V$, the hyperplane given by $$H(E,x):=\{v\in V:\langle v,x\rangle=h_E(x)\}$$ is the *supporting hyperplane* of $E$ for $x$ (See Figure \[fig: norming\]). For $x\in E$, the set $$F_x(E):=E\cap H(E,x)=\operatorname{argmax}_E(\varphi_x)$$ is called the *face* of $E$ defined by $x$, or also the *support set* of $E$ for $x$. For $v\in F_x(E)$ we say that $H(E,x)$ supports $E$ at $v$.
In the literature the faces defined above are usually called exposed faces.
\[minkowskigauge\] The *Minkowski gauge* or *gauge* of a bounded convex set $B\subseteq V$ which contains the origin in its interior is the function $$g_B:V\to\R,\quad g_B(x)=\inf\{t>0:x\in tB\}.$$
The set $B$ is a symmetric convex body (Definition \[convexb\]) if and only if the gauge function $g_B$ is a norm on $V$ whose unit ball is $B$. Otherwise it is a Finsler norm, i.e. only positively homogeneous.
\[normingsupport\] If $x\in V$ is such that $g_B(x)=1$ then $H_y=\varphi_y^{-1}(1)=\{z\in V:\langle z,y\rangle=1\}$ is a supporting hyperplane of $B$ at $x$ if and only if $\varphi_y$ is a norming functional of $x$.
The *polar* of a nonempty bounded set $E\subseteq V$ is $$E^\circ =\{x\in V:\langle x,y\rangle\leq 1 \mbox{ for all }y\in E\}.$$
Note that if $E$ is invariant by an isometric action so is its polar $E^\circ$. A *polytope* is the convex hull of a finite set of points.
These are the result of applying the polar operation to some standard sets
- If $B_p$ is the unit ball of the $\ell^p$ space for $1\leq p\leq \infty$ then $B_p^\circ=B_q$ where $q$ is conjugate to $p$.
- The polar of an ellipsoid $E\subseteq \mathbb R^n$ with axes of length $a_1,\dots,a_n$ is the ellipsoid with axes of length $1/a_1,\dots, 1/a_n$.
- The polar of a polytope $P\subseteq V$ containing $0$ in its interior and which has $n$ faces and $k$ vertices is a polytope with $k$ faces and $n$ vertices.
\[polar\] These are some standard properties that will be used later; let $E,F\subseteq V$ be compact convex sets containing $0\in V$ in the interior, then
1. $E^{\circ\circ}=E$, this is the bipolar property.
2. $(\lambda E)^\circ=\lambda^{-1}E^\circ$ for $\lambda >0$.
3. If $E\subseteq F$, then $F^\circ\subseteq E^{\circ}$.
4. $(E\cup F)^\circ=E^\circ\cap F^\circ$.
5. $(E\cap F)^\circ=\operatorname{conv}(E^\circ\cup F^\circ)$.
6. For a polytope $E=\operatorname{conv}\{x_1,\dots,x_n\}$ we have $$E^\circ=\{y\in V:\langle x_i,y\rangle\leq 1\mbox{ for }i=1,\dots,n\}.$$
7. For $E=\{y\in V:\langle x_i,y\rangle\leq 1\mbox{ for }i=1,\dots,n\}$ its polar is the polytope $$E^\circ=\operatorname{conv}\{0,x_1,\dots,x_n\}.$$
The next result can be found in Theorem 6.4 and Corollary 6.5 of [@bron], and Theorem 14.5 of [@rock].
\[supportdual\] Let $E\subseteq V$ be a compact convex set containing $0$ in its interior. Then
1. $h_E=g_{E^\circ}$ and $g_E=h_{E^\circ}$.
2. The supporting hyperplanes of $E$ are the hyperplanes determined by the points of the boundary $\operatorname{bd}E^\circ$ of $E^\circ$, i.e. $H_x$ with $x\in\operatorname{bd}E^\circ$.
The following are equivalent, see Figure \[fig: duality\]
- the hyperplane $H_y$ supports $E$ at $x\in E$.
- the hyperplane $H_x$ supports $E^\circ$ at $y\in E^\circ$.
- $x\in\operatorname{bd}E$, $y\in \operatorname{bd}E^\circ$ and $\langle x,y\rangle=1$.
The following, which relates the polar operation to orthogonal projections and sections with subspaces will also be used. Its proof is elementary therefore omitted.
\[projpolar\] If $B$ is a convex body in the inner product space $V$, $W$ is a subspace of $V$, and $p_W$ is the orthogonal projection onto $W$, then the following holds:
- $(B\cap W)^\circ=p_W(B^\circ).$
- $B^\circ\cap W=p_W(B)^\circ.$
Item ii) actually holds for any subset $B$ of $V$ and item i) follows from polar duality.
Convex structure of the unit ball of Hofer norms
------------------------------------------------
In this section, we characterize the generalized Hofer norms in terms of the supporting and gauge functions $h_E,g_E$.
\[hofernorms\] Let $E\subseteq V$ be a compact full set. Then we have $$\begin{aligned}
\|y\|_E & =h_{E-E}(y)=g_{(E-E)^\circ}(y).\\
\|y\|_E'&=h_{E\cup-E}(y)=g_{(E\cup-E)^\circ}(y).\end{aligned}$$ If in addition $0$ is in the interior of $\operatorname{conv}(E)$ $$\begin{aligned}
\|y\|_E^+ & =h_{E}(y)=g_{E^\circ}(y).\end{aligned}$$
For the first Hofer norm observe that $$\begin{aligned}
\|y\|_E&=\max_{x\in E}\langle x,y\rangle-\min_{x\in E}\langle x,y\rangle= \max_{x\in E}\langle x,y\rangle-(-\max_{x\in E}\langle x,-y\rangle)\\
&=h_E(y)-(-h_E(-y))=h_E(y)+h_E(-y)&\mbox{ by Definition \ref{defih}}\\
&=h_E(y)+h_{-E}(y)=h_{E-E}(y)&\\
&=g_{(E-E)^\circ}(y)&\mbox{ by Theorem \ref{supportdual}.}\end{aligned}$$
For the second Hofer norm $$\begin{aligned}
\|y\|'_E&=\max\{\max_{x\in E}\langle x,y\rangle,-\min_{x\in E}\langle x,y\rangle\} \\
&=\max\{\max_{x\in E}\langle x,y\rangle,\max_{x\in E}-\langle x,y\rangle\}\\
&=\max\{\max_{x\in E}\langle x,y\rangle,\max_{x\in -E}\langle x,y\rangle\}=\max_{x\in E\cup-E}\langle x,y\rangle=h_{E\cup-E}(y)\\
&=g_{(E\cup-E)^\circ}(y)&\mbox{ by Theorem \ref{supportdual}}.\end{aligned}$$ For the one-sided Hofer norm $$\begin{aligned}
\|y\|_E^+&=\max_{x\in E}\langle x,y\rangle=h_{E}(y)=g_{E^\circ}(y).\end{aligned}$$
\[ballh\] Note that the unit ball of Hofer’s norm is exactly $$(E-E)^\circ=(\operatorname{conv}(E-E))^\circ=(\operatorname{conv}(E)-\operatorname{conv}(E))^\circ.$$ Therefore all $\operatorname{Ad}$-invariant norms are Hofer norms if we take $E=\frac{1}{2}B^\circ$ where $B$ is the unit ball of the norm.
With these tools, we give a second proof of the characterization of the norming functionals of vectors in the unit sphere (Theorem \[normingfunc\]), or equivalently, the supporting hyperplanes of the unit ball for points in its sphere.
\[normingpolar\] The norming functionals of $x \in V$ are given by $\varphi_{y^+ -y^-}$, with $$y^+ \in \operatorname{conv}(\operatorname{argmax}_E \varphi_x)\quad \textrm{ and }\quad y^-\in\operatorname{conv}(\operatorname{argmin}_E \varphi_x).$$
Rescaling, we can assume that $1=\|x\|_E=g_{\operatorname{conv}(E-E)^\circ}(x)$, so that that $x\in\operatorname{bd}(E-E)^\circ$. By Remark \[normingsupport\] a functional $\varphi_y$ is a norming functional of $x$ if and only if the hyperplane $H_y$ supports $(E-E)^\circ$ at $x$. Theorem \[supportdual\] implies that $H_y$ supports $(E-E)^\circ$ at $x$ if and only if $H_x$ supports $(E-E)^{\circ\circ}=\operatorname{conv}(E-E)=\operatorname{conv}(E)-\operatorname{conv}(E)$ at $y\in \operatorname{conv}(E-E)$. Moreover, $H_x$ supports $\operatorname{conv}(E-E)$ at $y\in \operatorname{conv}(E-E)$ if and only if $y\in\operatorname{argmax}_{\operatorname{conv}(E)-\operatorname{conv}(E)}(\varphi_x)$, and this happens if and only if $$y\in\operatorname{argmax}_{\operatorname{conv}(E)}(\varphi_x)-\operatorname{argmin}_{\operatorname{conv}(E)}(\varphi_x).$$ That is, $y=y^+-y^-$ with $$y^+\in \operatorname{argmax}_{\operatorname{conv}(E)}(\varphi_x)=\operatorname{conv}(\operatorname{argmax}_E \varphi_x)$$ and likewise with $y^-$. This finishes the proof.
For the one-sided Hofer norm the norming functionals of $x \in V$ are given by $\varphi_{y^+}$, with $$y^+ \in \operatorname{conv}(\operatorname{argmax}_E \varphi_x).$$
Hofer’s metric on compact Lie groups {#hofergroup}
====================================
In this section, we present actions of compact semi-simple Lie groups $K$ on compact connected manifolds $M$ with symplectic form $\omega$, as a nice setting for the convex geometry that was discussed in the previous sections. It should serve as motivation and also as a source of examples. We refer to [@ps] for general background on the geometry of Hamiltonian actions.
Hamiltonian diffeomorphisms
---------------------------
Let $(M,\omega)$ be a connected closed symplectic manifold and let $H:[0,1]\times M \to \R$ be a smooth function. We denote $H_t(m):=H(t,m)$. This function $H$ induces a time dependent Hamiltonian vector field $X_{H_t}$ by Hamilton’s equations $$\begin{aligned}
dH_t=\omega(\cdot,X_{H_t})=-\iota_{X_{H_t}}\omega,\label{hamiltoneq}\end{aligned}$$ and hence an isotopy $\phi^H_t:M\to M$, $t\in [0,1]$ by the prescription that $$\phi^H_0=\phi\mbox{ and } \frac{d}{dt}\phi^H_t(m)=X_{H_t}(\phi^H_t(m)).$$ For a symplectic map $\varphi$. The Hamiltonian diffeomorphism group $\operatorname{Ham}(M,\omega)$ is by definition the set of diffeomorphisms $\phi:M\to M$ which can be written as $\phi=\phi^H_1$ for some $H$ and $\phi^H_0=\operatorname{id}$ as above.
The set $\operatorname{Ham}(M,\omega)$ is an infinite dimensional group under composition, all elements of which are symplectomorphisms of $(M,\omega)$. Its Lie algebra are the Hamiltonian vector fields in $\X(M)$, which can be identified with the smooth functions in $M$ modulo constant functions, via (\[hamiltoneq\]), hence $$T_{\operatorname{id}}\operatorname{Ham}(M,\omega)\simeq \X_{\operatorname{Ham}}(M)\simeq C^{\infty}(M)/\R{\mathbf 1}.$$ Since $\psi^*X_H=X_{H\circ \psi}$ the adjoint action in this group is given by $\operatorname{Ad}_\psi [H]=[H\circ \psi]$, where $[H]$ will denote the class of $H$ modulo constant functions. If the Hamiltonian is time independent, i.e. $H_t=H$ for $t\in[0,1]$, it is called *autonomous*. Note that the flow of such $H$ is ruled by the equation $$\frac{d}{dt}\phi_t(m)=X_H(\phi_t(m))=D(R_{\phi_t})_1(X_H)$$ when we interpret the differential of the right translation $R_g$ in the group of diffeomorphisms, as composition from the right. Therefore $\phi_t$ is the flow of the right invariant field $X_g=D(R_g)_1(X_H)$ and with the initial condition $\phi_0=\operatorname{id}=1_{\operatorname{Ham}(M,\omega)}$ it is clear that $\phi_1=\exp(X_H)$, where $\exp$ is the exponential map of the group of Hamiltonian diffeomorphism. However, this exponential map is not well-suited as a chart for the group, since it is not a local diffeomorphism in any reasonable neighbourhood of the $0$ vector field, see [@ps].
### Hofer’s norm
The $\operatorname{Ad}$-invariant $L^\infty$ norm $$\label{normahofer}
\|[H]\|= \max_MH -\min_M H$$ on the Lie algebra $T_{\operatorname{id}}\operatorname{Ham}(M,\omega)\simeq C^{\infty}(M)/\R{\mathbf 1}$ of the group $\operatorname{Ham}(M,\omega)$ is *Hofer’s norm*. It induces a Finsler length structure on curves $(\phi_t)_{t\in [0,1]}$ in $\operatorname{Ham}(M,\omega)$ by means of $$\begin{aligned}
\operatorname{length}(\phi_t^H)& =\int_0^1\|\frac{d}{dt}\phi_t^H\|dt =\int_0^1\|H_t\circ\phi^H_t\|dt\\
& =\int_0^1 \left(\max_M H_t-\min_M H_t \right) dt,\end{aligned}$$ and hence a bi-invariant distance $$\operatorname{dist}(\phi_0,\phi_1) =\inf\left\{\operatorname{length}(\phi^t_H) :\phi^0_H=\phi_0,\quad\phi^1_H=\phi_1\right\}.$$ As was shown for $\mathbb R^{2n}$ in [@hofer] and for general symplectic manifolds in [@lm], $\operatorname{dist}$ is a non degenerate, bi-invariant metric on $\operatorname{Ham}(M,\omega)$.
Hofer proved that the path of any autonomous Hamiltonian on $\C^n$ is length minimizing (among homotopic paths with fixed endpoints) as long as the corresponding Hamilton’s equation has no non-constant time-one periodic orbit. This result was generalized in [@lm] to general symplectic manifolds.
The *symplectic energy* on $\operatorname{Ham}(M,\omega)$ is $\operatorname{E}(\psi)=\operatorname{dist}(\operatorname{id},\psi)$. It is invariant under conjugation by symplectomorphisms. This function is also referred to as the Hofer energy.
\[quasiauto\] A path $(\phi_t)_{t\in [0,1]}\subseteq \operatorname{Ham}(M,\omega)$ is a geodesic of the Hofer metric if each $t\in [0,1]$ has a neighbourhood $I$ such that $\phi|_I$ is minimal, i.e. shorter than any other path joining its endpoints.
A Hamiltonian $H_t$ is called *quasi-autonomous* if there exists two points $x^-,x^+ \in M$ such that $$H_t(x^-) = \min_M H_t,\quad H_t(x^+) = \max_M H_t$$ for all $t\in [0,1]$. The Hamiltonian is called *slow* if it has no contractible periodic orbits of period less than or equal to one.
It is worth mentioning here the following results concerning quasi-autonomous Hamiltonians and the Hofer metric (but we won’t be making use of them). See [@mds Section 12.3] and the references therein for the proofs:
\[geodquasia\] Let $(\phi_t)_{t\in [0,1]}\subseteq \operatorname{Ham}(M,\omega)$ be a regular path ($C^1$ and with non-vanishing derivative). If $\phi$ is short for the Hofer metric, then the corresponding Hamiltonian $H_t$ is quasi-autonomous. If the Hamiltonian is quasi-autonomous, then $\phi$ is locally short (locally here refers to the time-variable). There exists a $C^2$ neighbourhood $V$ of the identity of $Ham(M,\omega)$ such that every point of $V$ can be joined to the identity by a minimizing geodesic.
A very similar norm used in the literature is defined without taking the quotient of the Hamiltonian functions by the constant functions. It is defined via the normalization $\int_MH\omega^n=0$, where $\omega^n$ is the Liouville measure on $M$. The norm is the $\operatorname{Ad}$-invariant $L^\infty$ norm $$\|H\|'= \max\{\max H,-\min H\}= \|H\|_{\infty}$$ which we are going to relate to the second Hofer norm in Remark \[hofernormalizada\].
Hamiltonian actions {#hamactions}
-------------------
Let $K$ be compact semi-simple Lie group with Lie algebra $\k=T_1K$ and dual space $\k^*$, which we identify with $\k$ through the duality pairing given by the (opposite of) the $\operatorname{Ad}$-invariant Killing form $\langle \cdot,\cdot\rangle$. This is an inner product in $\k$ because $\k$ is compact; since it is also $\operatorname{Ad}$-invariant, we can identify the coadjoint action with the adjoint action. The Lie algebra $\k$ acts by skew-symmetric transformations, that is $\operatorname{ad}x$ is skew-adjoint for this inner product, for any $x\in\k$.
Assume that the action of $K$ on $M$ is a *symplectic action*, that is, there exists a smooth map $\Phi:K\times M\to M$ that we denote $\Phi(g,m)=g\cdot m$ such that for each fixed $g\in K$, the automorphism $\Phi_g=\Phi(g,\cdot)$ is a symplectomorphism of $M$. We also assume that the action is *almost effective* ($g\mapsto \Phi_g$ has discrete kernel). For fixed $m\in M$ we denote by $\pi_m:K\to M$ the map $\pi_m(g)=\Phi(g,m)$.
For $x\in \k$, let $x_M(m)=\frac{d}{dt}\mid_{t=0}\exp(tx)\cdot m\in T_mM$ denote the infinitesimal action on $M$, the assumption that the action is almost effective implies that $\k \ni x\mapsto x_M\in \mathfrak X(M)$ is injective; the assumption that the action is symplectic implies that the field $x_M$ is symplectic, i.e. $\mathcal L_{x_M}\omega=0$.
A *moment map* for a symplectic $K$ action on $(M,\omega)$ is a map $\mu:M\to\k^*$ defined by $$\mu^x(m)=\langle\mu(m),x\rangle, \quad \mu^x:M\to \mathbb R$$ such that $\mu$ intertwines the $K$-action on $M$ and the coadjoint action on $\k^*$, i.e. $\mu(g\cdot m)= \operatorname{Ad}_g\mu(m)$ for all $g\in K,m\in M$, and such that $\mu$ satisfies Hamilton’s equation $$\begin{aligned}
d\mu^x=-\iota(x_M)\omega=-\omega (x_M,\cdot)\mbox{ for all }x\in\k.\label{hamiltoneq2}\end{aligned}$$ A symplectic $K$ action is called *Hamiltonian* if it admits a moment map.
A symplectic $K$ action on a symplectic manifold $(M,\omega)$ is the same thing as a homomorphism $K\to \operatorname{Symp}(M,\omega)$ to the group of symplectomorphisms of $(M,\omega)$ such that the map $K\times M\to M$, $(g,m)\to g\cdot m$ is smooth. The action has a moment map if and only if the image of the identity component of $K$ is contained in $\operatorname{Ham}(M,\omega)$, i.e. there is a map $\mu:M\to\k^*$ which satisfies Hamilton’s equations (\[hamiltoneq2\]). Averaging with the Haar probability measure $dk$ on $K$ yields an equivariant moment map $${\widetilde}{\mu}(x)=\int_{k\in K}\operatorname{Ad}_k\mu(k^{-1}\cdot x)\, dk.$$
Note that if $x\in\k$ then $x_M$ is a Hamiltonian vector field and $H=\mu^x$ is a Hamiltonian; from (\[hamiltoneq2\]) it is apparent that $X_{\mu^x}=x_M$. Also note that the Poisson bracket of the Hamiltonian functions $\mu^x$ and $\mu^y$ is given by $\{\mu^x,\mu^y\}=\mu^{[x,y]}$ for $x,y\in\k$.
The action defines a homomorphism with discrete kernel $$K\to \operatorname{Ham}(M,\omega).$$ At the level of Lie algebras this morphism is injective $$\begin{aligned}
T_1(K)=\k&\hookrightarrow C^{\infty}(M)/\R{\mathbf 1}\simeq T_{\operatorname{id}} \operatorname{Ham}(M,\omega)\label{liealgaction}\\
x&\mapsto [\mu^x]=\mu^x +\R{\mathbf 1}\nonumber\mapsto x_M.\end{aligned}$$
Given $u\in K$ and a path $x:[0,1]\to\k$, $t\mapsto x_t$, we calculate the isotopy $\phi_t\in \operatorname{Ham}(M,\omega)$ given by the initial condition $\phi_0=\Phi_u$ and the time dependent Hamiltonian $H_t=\mu^{x_t}$ (as noted, the Hamiltonian vector fields are $(x_t)_M$).
\[actiondiffeo\] If $\gamma:[0,1]\to K$ is the smooth solution to $\dot{\gamma}_t\gamma_t^{-1}=x_t$ and $\gamma_0=u$, then the isotopy is given by $\phi_t=\Phi_{\gamma_t}$.
We have to check that $\dot{\phi}_t\phi^{-1}_t=(x_t)_M$: for each $m\in M$, $$\begin{aligned}
\dot{\phi}_t(m) & = (\gamma_t\cdot m)^{\,\cdot}=D(\pi_m)_{\gamma_t}(\dot{\gamma}_t)=D(\pi_m)_{\gamma_t}(x_t \gamma_t)= D(\pi_m)_{\gamma_t}(\frac{d}{dr}\Bigr|_{r=0}\exp(rx_t)\gamma_t)\\
&=\frac{d}{dr}\Bigr|_{r=0} \pi_m(e^{rx_t}\gamma_t)=\frac{d}{dr}\Bigr|_{r=0} (e^{rx_t}\gamma_t)\cdot m=\frac{d}{dr}\Bigr|_{r=0} e^{rx_t}\cdot(\gamma_t\cdot m)\\
& =(x_t)_M(\gamma_t\cdot m)= (x_t)_M(\phi_t(m))=((x_t)_M\circ \phi_t)(m),\end{aligned}$$ and therefore $\dot{\phi}_t=(x_t)_M\circ\phi_t$ as claimed.
\[chamber\] The image of the moment map $\mu(M)$ is a union of (co)adjoint orbits in $\k$ which we denote by $\O_{\lambda}=\operatorname{Ad}_K(\lambda)$ for $\lambda\in\k$. Let $T$ be a maximal torus of $K$ and $\t$ its Lie algebra, a Cartan subalgebra. The image of the moment map can therefore be parametrized by a subset $A=\mu(M)\cap\t_+$ of a closed positive Weyl chamber $\t_+$, corresponding to a choice positive simple roots. Hence we have $$\mu(M)=\bigsqcup_{\lambda\in A}\operatorname{Ad}_K(\lambda)=\bigsqcup_{\lambda\in A}\O_\lambda.$$
### Hofer’s metric given by Hamitonian actions {#subsectionhoferham}
We use the inclusion of Lie algebras (\[liealgaction\]) to pull-back the Finsler length structure given by Hofer’s norm. The $L^\infty$ norm on $C^{\infty}(M)/\R{\mathbf 1}$ restricted to the image $T_1(K)=\k$ is therefore by definition of the Hofer norm (\[normahofer\]) given by $$\begin{aligned}
\|\mu^x\| &=\max_{m\in M}\mu^x(m)-\min_{m\in M}\mu^x(m)=\max_{m\in M}\langle\mu(m),x\rangle-\min_{m\in M}\langle\mu(m),x\rangle\label{hofer1}\\
&=\max_{y\in\mu(M)}\langle y,x\rangle-\min_{y\in\mu(M)}\langle y,x\rangle= \|x\|_{\mu(M)},\nonumber\end{aligned}$$ where the last equality is by definition of the generalized Hofer norm (Definition \[generhofernorm\]) for the $\operatorname{Ad}$-invariant set $E=\mu(M)\subseteq \k$.
\[fullsiiae\] Since $M$ is compact the image of the moment map is bounded. Also, because the image of the moment map is full in $\k$ if and only if the action is almost effective: The image of the moment map is not full if and only if its image is contained in a hyperplane, which in turn happens if and only if there exists $x\in\k$ such that $\varphi_x$ is constant on $\mu(M)$. This is equivalent to the map $\mu^x:M\to\R$ being a constant Hamiltonian which generates the zero vector field. Finally, this is also equivalent to the one-parameter group $\exp(tx)$ acting trivially, i.e. the action not being almost effective.
Therefore, the Hofer norm just described is the norm $\|\cdot\|_{\mu(M)}$ for a moment map $\mu:M\to\k^*\simeq\k$, given by equation (\[hofer1\]) above. For a general $\operatorname{Ad}$-invariant set $E$ in $\k$ we obtain an $\operatorname{Ad}$-invariant norm $\|\cdot\|_{E}$ in $\k$. This norm is $\operatorname{Ad}$-invariant and induces a bi-invariant Finsler metric on $K$ by left (or right) translation given by $$\begin{aligned}
\operatorname{length}(\gamma)&=\int_0^1\|\dot{\gamma}_t\|_{\gamma_t}dt=\int_0^1\|\dot{\gamma}_t\gamma_t^{-1}\|_{\mu(M)} dt=\int_0^1\|\gamma_t^{-1}\dot{\gamma}_t\|_{\mu(M)}dt,\end{aligned}$$ $$\begin{aligned}
\label{longitud}
\operatorname{dist}(u,v) &:=\inf\left\{\operatorname{length}(\gamma_t) :\gamma_0=u,\quad\gamma_1=v\right\}.\end{aligned}$$
\[hofernormalizada\] In the case of actions by semi-simple Lie groups, all Hamiltonians are normalized in the following sense: $$\int_M \mu^x \omega^n=0.$$ This is because $$\int_M \langle\mu,x\rangle \omega^n=\int_{\mu(M)}\langle y,x\rangle d\nu(y),$$ where $\nu=\mu_*(\omega^n)$ is the pushforward of the normalized Liouville measure $\omega^n$ by the moment map. The measure $\nu$ is $\operatorname{Ad}$ invariant hence its center of mass is also invariant. The center of mass is zero since the group is semi-simple, therefore $\int_{\mu(M)}\langle y,x\rangle d\nu(y)=0$. The resulting $L^\infty$ norm is $$\label{hofer2}
\|\mu^x\|_{\infty}=\max\{\sup_{m\in M}\mu^x(m),-\inf_{m\in M}\mu^x(m)\}=\|x\|_{\mu(M)},$$ by the definition of the second generalized Hofer norm in \[generhofernorm2\] for the $\operatorname{Ad}$-invariant set $E=\mu(M)\subseteq\k$. Therefore, this second Hofer norm just described is the norm defined in equation (\[generhofernorm2\]) as $\|\cdot\|'_{\mu(M)}$ for a momentum map $\mu:M\to\k^*\simeq\k$.
\[hoferlp\] We can define also an $L^p$-norm for $1\le p<\infty$ via the embedding $$\iota:V\hookrightarrow L^p(E,\nu)/\R{\mathbf 1}, \quad x\mapsto [\varphi_x]=\varphi_x+\R{\mathbf 1},$$ where $\nu$ is an $\operatorname{Ad}$-invariant measure. Consider the case of an $L^p$-norm defined by the inclusion $$\iota:\k\hookrightarrow L^p(M,\nu)/\R{\mathbf 1}.$$ Since $L^p(M,\nu)$ is uniformly convex when $1<p<\infty$, the quotient $L^p(M,\nu)/\R{\mathbf 1}$ is uniformly convex. Therefore the $L^p$-norm is uniformly convex and this implies by [@lar19 Theorem 4.15] that $K$ with the induced metric is uniquely geodesic. We won’t be pursuing the geometry of these norms in this paper.
\[regular\] Recall that for a compact semisimple group $K$, each element $\lambda\in K$ is semi-simple. A *regular* element $\lambda$ is such that its commutant $\t={\mathfrak z}(\lambda)$ is a maximal abelian subalgebra; equivalently in this setting, $\t$ is a Cartan subalgebra of $\k$.
In the case of $K=\operatorname{SU}(n)$, an element of $\k=\su(n)$ is regular if and only if all its eigenvalues are distinct.
\[coadj\] An important special case is that of a compact connected semi-simple Lie group $K$ acting via $\operatorname{Ad}$ in a Hamiltonian way on a (co)adjoint orbit $\O_\lambda$ containing $\lambda\in\k^*\simeq \k$.
For the action to be almost effective, it is necessary and sufficient that $\lambda$ is non-zero in each simple summand of $\k$ (because in that case each simple factor of $K$ acts nontrivially on the orbit, and this is equivalent to the fact that the (co)adjoint orbit is full in $\k$ (see [@bgh Lemma 17] and Remark \[fullsiiae\]). In particular this occurs if $\lambda$ is regular.
The adjoint orbit is endowed with the Kostant-Kirillov-Souriau symplectic form $\omega$. The kernel of this homomorphism is the center $Z(K)$ of $K$ which is a discrete subgroup. We have therefore an inclusion $$K/Z(K)\hookrightarrow \operatorname{Ham}(\O_\lambda,\omega).$$ At the level of Lie algebras this inclusion is $x\mapsto [\mu^x]$ where $\mu^x:\O_\lambda\to\R$ is the Hamiltonian map given by the $\mu^x(y)=\langle x, y\rangle$. The Hamiltonian is the component of the moment map $\iota:\O_\lambda\hookrightarrow\k^*\simeq\k$ along $x$, that is $\mu^x=\varphi_x\circ\iota$. Moreover, since the action of $K$ is given by $g\mapsto g xg^{-1}$, it is apparent that if $x\in \k$ then the induced Hamiltonian vector field is given by $$x_{\O_{\lambda}}(m)=\frac{d}{dt}\Bigr|_{t=0}\exp(tx)\cdot m = \frac{d}{dt}\Bigr|_{t=0}\operatorname{Ad}_{e^{tx}}(m)=(\operatorname{ad}x)(m)=[x,m].$$ The symplectic form is $$\omega_x( y_{\O_{\lambda}}(x),z_{\O_{\lambda}}(x) )=\langle x,[y,z]\rangle$$ where the pairing is given by the opposite Killing form of $\k$ as before. The induced Hofer norm on $T_1(K)=\k$ is therefore $$\begin{aligned}
\label{normcadjoint}
\|x\|_{\O_\lambda}&=\max_{y\in\O_\lambda}\mu^x(y)-\min_{y\in\O_\lambda}\mu^x(y)=\max_{k\in K}\mu^x(\operatorname{Ad}_k\lambda)-\min_{k\in K}\mu^x(\operatorname{Ad}_k\lambda)\\
&=\max_{k\in K}\langle x,\operatorname{Ad}_k\lambda\rangle-\min_{k\in K}\langle x,\operatorname{Ad}_k\lambda\rangle.\nonumber\end{aligned}$$ Since the action is almost effective $\|\cdot\|_{\O_\lambda}$ is a norm. Since it is also $\operatorname{Ad}$-invariant it induces a bi-invariant Finsler metric on $K$.
\[sun\] In the case of the special unitary group $\operatorname{SU}(n)$ we have the inclusion $$\operatorname{SU}(n)/\Z_n\simeq\operatorname{PU}_n\hookrightarrow \operatorname{Ham}(\O_\lambda,\omega)$$ where $\O_\lambda$ is a (co)adjoint orbit of $\operatorname{SU}(n)$ passing through the given diagonal matrix $\lambda = i\operatorname{diag}(\lambda_1,\ldots,\lambda_n)\in\su(n)$ with $\lambda_1\leq \lambda_2 \leq \ldots \le \lambda_n$. At the Lie algebra level this inclusion is $$T_1(\operatorname{SU}(n))=\su(n)\hookrightarrow C^{\infty}(\O_\lambda)/\R{\mathbf 1}\simeq T_{\operatorname{id}} \operatorname{Ham}(\O_\lambda,\omega)$$ and $\mu^x:\O_\lambda\to\R$ is the linear Hamiltonian given by the opposite Killing form $$\mu^x(y)=-2n\operatorname{tr}(xy^*)=2n\operatorname{tr}(xy).$$ The Hofer norm (\[normcadjoint\]) is given by $$\|v\|_{\O_\lambda}=2n\max\{\operatorname{tr}(\lambda uvu^{-1}):u\in \operatorname{SU}(n)\}-2n\min\{\operatorname{tr}(\lambda wvw^{-1}):w\in \operatorname{SU}(n)\},$$ which is exactly ($2n$-times) the $\lambda$-numerical diameter of $v$, see the next remark.
The group $\operatorname{SU}(2)/{\{\operatorname{id},-\operatorname{id}\}}\simeq\operatorname{SO}(3)$ acts on a nontrivial (co)adjoint orbit $\O_\lambda\simeq S^2$ which is the two dimensional sphere endowed with the area form $\omega$. The action is given by rotations, see Example 1.4.H in [@pol01].
Another bi-invariant norm related to the Hofer norm for skew-symmetric matrices $v\in\su(n)$, is the $c$*-numerical radius* $$\omega(v)_c=\max\{|\operatorname{tr}(cuvu^{-1})|:u\in U_n\},$$ given by a fixed matrix $c$ such that $\operatorname{tr}(c)\ne 0$ and $c$ is not a scalar multiple of the identity (see [@li2]). This is our second Hofer norm applied to non semi-simple groups. The usual numerical radius of $v$ can be recovered by picking the one-dimensional projection $c=e_1\otimes e_1$.
\[semiads\] For fixed $0\ne \lambda\in \k$, and $v\in \k$, consider the function $$\|v\|_{\O_\lambda}^+=\max_{k\in K}\langle v , \operatorname{Ad}_k\lambda\rangle.$$ Note that by the $\operatorname{Ad}$-invariance of the norm, this can be also computed as $$\|v\|_{\O_\lambda}^+=\max_{k\in K}\langle \operatorname{Ad}_k v ,\lambda\rangle=\max\{\langle y , \lambda\rangle: y\in \operatorname{conv}(\O_v)\}.$$ When $\lambda$ is non-zero on each simple-summand of $\k$ (in particular if $\lambda$ is regular), then $0$ is in the interior of its closed convex hull (see Remark \[onesided\] below). Therefore this is a Finlser norm, in fact it is our one-sided Hofer norm of Remark \[otrasnorm\]. In the context of the group of Hamiltonian symplectomorphisms, these norms were studied in several papers; in particular see [@md] where the name *one-sided* was used. This norm is positively homogeneous, but if there is a Weyl group element $w$ such that $w.\lambda=-\lambda$ then it is homogeneous. Note by replacing $\lambda$ with minus $\lambda$, we can also obtain $\|v\|_{\O_\lambda}^-=-\min_{k\in K}\langle \operatorname{Ad}_k v ,\lambda\rangle$.
Convex geometry of Hofer’s norm on Cartan algebras {#convexgeocart}
==================================================
In this section we study generalized Hofer norms on the Lie algebra of a compact semisimple Lie group. We use several convexity theorems and the convex analysis of $\operatorname{Ad}$-invariant convex functions. This convex analysis expresses properties of $\operatorname{Ad}$-invariant functions in terms of properties of its restrictions to Cartan algebras and positive Weyl chambers, which are fundamental domains for the adjoint action.
We first recall Kirwan’s nonabelian convexity theorem (see [@guisja]), which will be useful for our purposes of characterizing the Hofer norm restricted to a Cartan subalgebra $\t\subseteq \k$. As before $\t_+\subseteq \t$ is a choice of closed positive Weyl chamber.
\[kirwan\] If $K\curvearrowright M$ is a Hamiltonian action with $K$ a compact connected Lie group and $(M,\omega)$ compact and connected, then $\mu(M)\cap\t_+$ is a convex polytope, i.e. the convex hull of a finite set of points $x_1,\dots,x_n$ in $\t_+$, that is $$\mu(M)\cap\t_+=\operatorname{conv}\{x_1,\dots, x_n\}.$$
This theorem is a generalization of the Atiyah-Guillemin-Sternberg theorem, which states that for a Hamiltonian action of a torus on a compact connected manifold the image of the moment map is the convex hull of the image of the fixed point set of the action. The Atiyah-Guillemin-Sternberg theorem is a generalization of Kostant’s convexity theorem which will be useful here:
\[kostant\] If $K$ is a compact Lie group and $T$ is a maximal torus in $K$ and $p:\k\to\t$ is the projection of the Lie algebra of $K$ onto the Lie algebra of $T$, then $$p(\O_\lambda )=\operatorname{conv}(\W . \lambda ),$$ where $\W .\lambda$ is the Weyl group orbit of $\lambda\in\k$.
The projection $p$ is taken along the orthogonal direction given by (minus) the Killing form of $\k$.
The Hofer polytope
------------------
With these tools we characterize the intersection with Cartan algebras of unit balls (of Hofer norms given by compact, full and $\operatorname{Ad}$-invariant sets $E\subseteq \k$).
\[projconv\] If $E$ is $\operatorname{Ad}$-invariant, and $p:\k\to \t$ is the orthogonal projection, then $$p(\operatorname{conv}(E))=\operatorname{conv}\{w.x:x\in E\cap\t_+, \,w\in\W\},$$ and if $E\cap\t_+=\operatorname{conv}\{x_1,\dots,x_n\}$, then $$p(\operatorname{conv}(E))=\operatorname{conv}\{w.x_i:i=1,\dots,n,\,w\in\W\}.$$
This follows from the identities $$\begin{aligned}
p(\operatorname{conv}(E))&=p(\operatorname{conv}(\cup_{\lambda\in E\cap\t_+}\O_\lambda))&\\
&=\operatorname{conv}(p(\cup_{\lambda\in E\cap\t_+}\O_\lambda))&\mbox{ since }p\mbox{ is affine}\\
&=\operatorname{conv}(\cup_{\lambda\in E\cap \t_+}p(\O_\lambda))\\
&=\operatorname{conv}(\cup_{\lambda\in E\cap \t_+}\{w.\lambda:w\in\W\})&\mbox{ by Theorem \ref{kostant}}\\
&=\operatorname{conv}\{w.\lambda:\lambda\in E\cap \t_+,\,w\in\W\}.\end{aligned}$$ Now observe that if $E\cap\t_+=\operatorname{conv}\{x_1,\dots,x_n\}$ $$\begin{aligned}
p(\operatorname{conv}(E))&=\operatorname{conv}(\cup_{\lambda\in E\cap\t_+}\{w.\lambda:w\in\W\})& \\
&=\operatorname{conv}(\cup_{w\in \W}w.(E\cap \t_+))=\operatorname{conv}(\cup_{w\in \W}\operatorname{conv}(w.(E\cap\t_+)))\\
&=\operatorname{conv}(\cup_{w\in \W}\operatorname{conv}\{w.x_i:i=1,\dots,n\})\\
&=\operatorname{conv}\{w.x_i:i=1,\dots,n,\,w\in\W\}\end{aligned}$$ where the third and last equalities follow from basic properties of the convex hull operation.
\[proyecartan\] Let $B\subseteq \k$ be an $\operatorname{Ad}$-invariant convex body, and $p:\k\to \t$ the orthogonal projection to a Cartan sub-algebra, then Kostant’s Theorem \[kostant\] implies that $B\cap \t=p(B)$. This follows from $p(b)\in p(\O_b)=\operatorname{conv}(\O_b\cap \t)\subseteq\operatorname{conv}(B\cap \t)=B\cap\t$ for $b\in B$.
\[hoferpoly\] Set $A=E\cap\t_+$, then $$\begin{aligned}
(\operatorname{conv}(E)-\operatorname{conv}(E))\cap\t&=\operatorname{conv}\{w_1.x_1-w_2.x_2:x_1,x_2\in A,\,w_1,w_2\in\W\}.\\
(\operatorname{conv}(E\cup -E))\cap\t&=\operatorname{conv}\{w.x:x\in A\cup -A,w\in\W\}.\\
\operatorname{conv}(E)\cap\t&=\operatorname{conv}\{w.x:x\in A,w\in\W\}.\end{aligned}$$
If furthermore $A=\operatorname{conv}\{x_1,\dots,x_n\}$ is a polytope, then $$\begin{aligned}
(\operatorname{conv}(E)-\operatorname{conv}(E))\cap\t&=\operatorname{conv}\{w.x-w'.x':x,x'\in \{x_1,\dots,x_n\},\,w,w'\in\W\}.\\
(\operatorname{conv}(E)\cup\operatorname{conv}(E))\cap\t&=\operatorname{conv}\{w.x:x\in \{x_1,\dots,x_n,-x_1,\dots,-x_n\},w\in\W\}.\\
\operatorname{conv}(E)\cap\t&=\operatorname{conv}\{w.x:x\in \{x_1,\dots,x_n\},w\in\W\}.\end{aligned}$$
For the first equality note that by the previous remark, $$\begin{aligned}
(\operatorname{conv}(E)-\operatorname{conv}(E))\cap\t & = p(\operatorname{conv}(E)-\operatorname{conv}(E)) \\
& = p(\operatorname{conv}(E))-p(\operatorname{conv}(E))\\
& = \operatorname{conv}\{w_1.x_1-w_2.x_2:x_1,x_2\in A,\,w_1,w_2\in\W\}&\end{aligned}$$ where the last equality is due to Lemma \[projconv\]. For the fourth assertion note that $$\begin{aligned}
(\operatorname{conv}(E)-\operatorname{conv}(E))\cap\t & = p(\operatorname{conv}(E)-\operatorname{conv}(E)) = p(\operatorname{conv}(E))-p(\operatorname{conv}(E))\\
& = \operatorname{conv}\{w.x-w'.x':x,x'\in \{x_1,\dots,x_n\},\quad w,w'\in\W\}\end{aligned}$$ holds, where we used the second assertion of Lemma \[projconv\] in the last equality.
For the second equality note that by Remark \[proyecartan\], $$\begin{aligned}
(\operatorname{conv}(E\cup -E)\cap\t & = p(\operatorname{conv}(E\cup -E)) \\
& = p(\operatorname{conv}(\operatorname{conv}(E)\cup \operatorname{conv}(-E))) \\
& = \operatorname{conv}(p(\operatorname{conv}(E)\cup \operatorname{conv}(-E))) \\
& = \operatorname{conv}(p(\operatorname{conv}(E))\cup p(\operatorname{conv}(-E))) \\
& = \operatorname{conv}(p(\operatorname{conv}(E))\cup -p(\operatorname{conv}(E))) \\
& = \operatorname{conv}(\{w.x:x\in A,\,w\in\W\}\cup -\{w.x:x\in A,\,w\in\W\}) \\
& = \operatorname{conv}(\{w.x:x\in A\cup -A,\,w\in\W\}) \end{aligned}$$ where the penultimate equality is due to Lemma \[projconv\]. For the fifth assertion note that $$\begin{aligned}
(\operatorname{conv}(E\cup -E)\cap\t & = p(\operatorname{conv}(E\cup -E)) \\
& = \operatorname{conv}\{\pm w.x:x\in \operatorname{conv}\{x_1,\dots,x_n\},\,w\in\W\}\\
& = \operatorname{conv}\{w.x:x\in \{x_1,\dots,x_n,-x_1,\dots,-x_n\},\,w\in\W\}\end{aligned}$$ holds, where we used the previous identity and properties of the convex hull operation.
The proof of the third and fourth equalities is simpler and we omit it.
In Figure \[fig: ext\] we illustrate the first couple of equalities of Proposition \[hoferpoly\] in the case $E=\O_\lambda$ where $\lambda\in\su(3)$ is singular and non zero.
\[hoferpolytopes\] For $E\cap\t_+=\operatorname{conv}\{x_1,\dots,x_n\}$, we call the polytope $$P=\operatorname{conv}\{w.x-w'.x':x,x'\in \{x_1,\dots,x_n\},\quad w,w'\in\W\}$$ of Proposition \[hoferpoly\] the *Hofer norm polytope*. We call $$P'=\operatorname{conv}\{w.x:x\in \{x_1,\dots,x_n,-x_1,\dots,-x_n\},\,w\in\W\}$$ the *second Hofer norm polytope*. Finally, we call $$P^+=\operatorname{conv}\{w.x:x\in \{x_1,\dots,x_n\},\,w\in\W\}$$ the *one-sided Hofer norm polytope*.
l\[polnorm\] If $\operatorname{ext}(P)=\{y_1,\dots,y_m\}$ are the extreme points of the Hofer norm polytope $P$ then this set is Weyl group invariant. If $$B=(\operatorname{conv}(E)-\operatorname{conv}(E))^\circ$$ is the unit ball of the Hofer norm (Remark \[ballh\]), from Remarks \[projpolar\] and \[proyecartan\], we obtain $$\label{convep}
B\cap \t=(\operatorname{conv}(E)-\operatorname{conv}(E))^\circ\cap\t=P^\circ.$$ The same holds for the other Hofer norms with $B'=(\operatorname{conv}(E)\cap-\operatorname{conv}(E))^\circ $ and $B^+=\operatorname{conv}(E)^\circ$ and the polytopes $P'$ and $P^+$ respectively (for the second Hofer norm and the one-sided Hofer norm, see Remark \[otrasnorm\]). The polytopes $P$ and $P'$ are centrally symmetric, and in general $P^+$ is not.
From (\[convep\]) and Proposition \[hofernorms\] we obtain the following
Let $\|\cdot\|_{\mu(M)}$ be the Hofer norm derived from the $\operatorname{Ad}$-invariant set $E=\mu(M)$, and let $P$ be the corresponding Hofer norm polytope. Then Hofer norm restricted to the Cartan subalgebra $\t$ is given by the Minkowski gauge $g_{P^\circ}$, that is $$\|x\|_{\mu(M)}=g_{P^\circ}(x)=\inf\{t>0: x\in tP^{\circ}\} \qquad \forall x\in \t.$$
From Remark \[polar\] it follows that $$P^\circ=\{x\in V:\langle y_i,x\rangle\leq 1\mbox{ for }i=1,\dots,m\}.$$
### Direct sums of manifolds and polytopes
If $M_1,\dots,M_n$ are symplectic manifolds equipped with Hamiltonian actions of $K$, with moment maps $\mu_1,\dots,\mu_n$, then the induced action on $M_1\times\dots\times M_n$ is also Hamiltonian with moment map $\mu:M_1\times\dots\times M_n\to\k$ given by $$\mu(m_1,\dots,m_n)=\mu_1(m_1)+\dots+\mu_n(m_n).$$ Therefore the image of $\mu$ is given by $$\begin{aligned}
\label{sumaimagmoment}
\mu(M_1\times\dots\times M_n)=\mu_1(M_1)+\dots+\mu_n(M_n).\end{aligned}$$ To find the Hofer norm polytope of this action we need the following
\[sumadeadinv\] If $E_1,\dots,\operatorname{E}_n$ are $\operatorname{Ad}$-invariant and compact subsets of $\k$, and $p:\k\to \t$ is the orthogonal projection, then $$\begin{aligned}
(\operatorname{conv}(E_1&+\dots+\operatorname{E}_n)-\operatorname{conv}(E_1+\dots+\operatorname{E}_n))\cap\t\\
&=(\operatorname{conv}(E_1)-\operatorname{conv}(E_1))\cap\t+\dots +(\operatorname{conv}(E_n)-\operatorname{conv}(E_n))\cap\t.\end{aligned}$$
Again, by Remark \[proyecartan\] $$\begin{aligned}
(\operatorname{conv}(E_1&+\dots+\operatorname{E}_n)-\operatorname{conv}(E_1+\dots+\operatorname{E}_n))\cap\t \\
& = p((\operatorname{conv}(E_1+\dots+\operatorname{E}_n)-\operatorname{conv}(E_1+\dots+\operatorname{E}_n))) \\
& = p((\operatorname{conv}(E_1)-\operatorname{conv}(E_1))+\dots +(\operatorname{conv}(E_n)-\operatorname{conv}(E_n)))\\
& = p(\operatorname{conv}(E_1)-\operatorname{conv}(E_1))+\dots + p(\operatorname{conv}(E_n)-\operatorname{conv}(E_n))\\
& =(\operatorname{conv}(E_1)-\operatorname{conv}(E_1))\cap\t+\dots +(\operatorname{conv}(E_n)-\operatorname{conv}(E_n))\cap\t.\end{aligned}$$
The next proposition is now a straightforward consequence of Lemma \[sumadeadinv\].
\[sumahoferpoly\] If $M_1,\dots,M_n$ are symplectic manifolds equipped with Hamiltonian almost effective actions of a compact semi-simple group $K$ such that the Hofer norm polytopes are $P_1,\dots,P_n$. Then the induced action on $M_1\times\dots\times M_n$ has Hofer norm polytope $P_1+\dots+P_n$.
Proposition \[sumahoferpoly\] is also valid in the case of the one-sided Hofer Finsler norm (Remark \[semiads\]), with polytopes $P^+_1,\dots,P^+_n$ and $P^+_1+\dots+P^+_n$.
### Faces of balls of $\operatorname{Ad}$-invariant norms
In this section we will use Lewis’ results on convex analysis of $\operatorname{Ad}$-invariant functions and their restriction to Cartan subalgebras [@lewis] to describe the faces of the ball of $\operatorname{Ad}$-invariant norms.
Since Lewis’ results are stated in terms of subgradients of gauge function we provide next the correspondence between the faces of the balls and the gradients of the gauge functions: *the subdifferential of a gauge $g_B$ at a unit norm $x$ corresponds via $y\mapsto \varphi_y$ to the supporting functionals of the polar of the ball at $x$*.
Recall that the subdifferential of a convex function $f:V\to\R$ at $x_0\in V$ is defined as $$\partial f(x_0)=\{y\in V : f(x)-f(x_0)\geq \langle x-x_0,y\rangle\qquad \forall\, x\in V\}.$$ Each such $y$ which satisfies the condition stated above is called a *subgradient*. It defines a supporting hyperplane to the graph of $f$ at $(x_0,f(x_0))$.
\[gradface\] Let $B\subseteq V$ be a convex body containing $0\in V$ and let $g_B$ be the gauge function associated to $B$. For a non-zero $x\in V$ $$\partial g_B(x)=F_x(B^\circ)\quad \textrm{ and }\quad \partial g_{B^\circ}(x)=F_x(B)$$ i.e. $\partial g_B(x)$ is the exposed face of $B^\circ$ defined by $x$.
If $x\neq 0$, we first claim that $$\partial g_B(x)=\{y\in V:\langle y,x\rangle=g_B(x),g_{B^\circ}(y)= 1\}.$$ This is Corollary 23.5.3 in [@rock], we give a direct proof for the convenience of the reader. Note that since $g_B=h_{B^{\circ}}$ (Proposition \[hofernorms\]), $\langle y,x\rangle=g_B(x)$ and $g_{B^{\circ}}(y)= 1$ imply $\langle x,y\rangle=\sup_{w\in B^{\circ}}\langle x, w\rangle=g_B(x)$ and $\langle y,z\rangle\le g_B(z) \quad\forall z\in V$ respectively. Thus for such $y\in V$ and any $z\in V$, we have $\langle z,y\rangle- \langle x,y\rangle \le g_B(z)-g_B(x)$ which shows that $y\in \partial g_B(x)$. Reciprocally, if $y\in \partial g_B(x)$, then replacing with $z=0$ and $z=2x$ in $$\sup_{w\in B^{\circ}}\langle w,z\rangle - \sup_{w\in B^{\circ}}\langle w,x\rangle= g_B(z)-g_B(x)\ge \langle z,y\rangle -\langle x,y\rangle,$$ we obtain $g_B(x)=\langle x,y\rangle$; in particular $g_{B^{\circ}}(y)\ge 1$. But then from $g_B(z)\ge \langle z,y\rangle$ for all $z$ it also follow that $g_{B^{\circ}}(y)\le 1$, thus proving the claim.
Therefore, noting that $x/{g_B(x)}\in\operatorname{bd}B$ $$\begin{aligned}
\partial g_B(x)&=\{y\in V:\langle y,x\rangle=g_B(x),g_{B^\circ}(y)= 1\}\\
&=\{y\in V:\langle y,x\rangle=g_B(x),y\in\operatorname{bd}B^\circ\}\\
&=\{y\in V:\langle y,x/{g_B(x)}\rangle=1,y\in\operatorname{bd}B^\circ\}\\
&=\{y\in V:{H_{x/{g_B(x)}}}\mbox{ supports } B^\circ\mbox{ at }y\}&\mbox{ by Theorem \ref{supportdual}} \\
&=F_{x/{g_B(x)}}(B^\circ)=F_{x}(B^\circ).&\mbox{ by Definition \ref{defih}}\end{aligned}$$ The other identity is immediate from the polar duality.
The next theorem relates the subdifferential of the gauges to the subdifferentials of the gauges restricted to Cartan subalgebras $\t$. It’s proof can be found in [@lewis Theorem 3.5]. We define the map $\delta:\k\to\t^+$ by $\O_x\cap\t^+=\{\delta(x)\}$, that is, $\delta(x)$ is the unique element in the positive Weyl chamber $\t^+$ that is $\operatorname{Ad}$-conjugated to $x$.
\[lewis\] Let $K$ be a semi-simple compact group and let $g_\k:\k\to\R_+$ be an $\operatorname{Ad}$-invariant gauge function. Then $y\in\partial g_\k(x)$ if and only if $\delta(y)\in\partial g_\t(\delta(x))$ and there is $u\in K$ such that $\operatorname{Ad}_u(x),\operatorname{Ad}_u(y)\in\t_+$.
If $\g=\k\oplus\p$ is the Cartan decomposition of a semi-simple Lie algebra $\g$, then Lewis’ theorem was stated for an $\operatorname{Ad}$ invariant gauge on $\p$. We can take the complexified Lie algebra $\g=\k\oplus i\k$ and apply the theorem to $\p=i\k$. In the case $K=\operatorname{SU}(n)$, the matrices $x$ and $y$ such that there is $u\in K$ with $\operatorname{Ad}_u(x),\operatorname{Ad}_u(y)\in\t_+$ are said to be *simultaneously ordered diagonalizable*.
We now restate Lewis’ theorem in a form convenient to its application in Section \[geodesicommute\]; there is a related formulation due to Lewis in [@lewis2] for the orthogonal group of matrices.
\[lewisfaces\] Let $K$ be a semi-simple compact group, let $B$ be an $\operatorname{Ad}$-invariant convex body in $\k$ containing $0$, and let $\t$ be a Cartan algebra in $\k$ and $\t_+$ a positive Weyl chamber. Then for non-zero $x\in\t_+$ $$F_x(B)=\operatorname{Ad}_{Z(x)}(F_x(B\cap\t)\cap\t_+),$$ and for general $x'\in\k$, if $v\in K$ is such that $\operatorname{Ad}_v(x')\in\t_+$, then $$F_{x'}(B)=\operatorname{Ad}_{v^{-1}}\operatorname{Ad}_{Z(\operatorname{Ad}_v(x'))}(F_{\operatorname{Ad}_v(x')}(B\cap\t)\cap\t_+).$$
We restate Theorem \[lewis\] in the case that $x\in\t_+$. Let $g=g_B$ be the gauge function, see Definition \[minkowskigauge\]. The theorem states that $y\in\partial g(x)$ if and only if $\gamma(y)\in\partial g_\t(x)$ and for some $u\in K$ we have $\operatorname{Ad}_u(x),\operatorname{Ad}_u(y)\in\t_+$. Observe that $\operatorname{Ad}_u(x)\in\t_+$ for $u\in K$ if and only if $u\in Z(x)$, so the condition $\operatorname{Ad}_u(y)\in \t_+$ reduces to $y\in\operatorname{Ad}_{Z(x)}\t_+$. The conditions $\gamma(y)\in\partial g_\t(x)$ and $y\in\operatorname{Ad}_{Z(x)}\t_+$ are equivalent to $y\in\operatorname{Ad}_{Z(x)}(\partial g_\t(x)\cap\t_+)$. Hence $$\partial g(x)=\operatorname{Ad}_{Z(x)}(\partial g_\t(x)\cap\t_+).$$ For a general $x'\in\k$ take $v\in K$ such that $\operatorname{Ad}_v(x')\in\t_+$. Then, since the adjoint action is isometric for the norm $g$ we get $$\partial g(x')=\operatorname{Ad}_{v^{-1}}\operatorname{Ad}_{Z(\operatorname{Ad}_v(x'))}(\partial g_\t(\operatorname{Ad}_v(x'))\cap\t_+).$$ In the case that $g=g_B$ and $g_{\t}=g_{B\cap\t}$ we have $$F_x(B^\circ)=\partial_B(x)=\operatorname{Ad}_{Z(x)}(\partial g_{B\cap\t}(x)\cap\t_+)=\operatorname{Ad}_{Z(x)}(F_x(B^\circ\cap\t)\cap\t_+),$$ where we used $F_x(B^\circ)=\partial_B(x)$ and $F_x(B^\circ\cap\t)=\partial g_{B\cap\t}$ (this follows from Proposition \[gradface\]). If we take polars, by Remark \[projpolar\] and the bipolar property, we obtain the required statement.
If $B=\operatorname{conv}(E-E)^\circ$, then $B\cap\t=\operatorname{conv}(E-E)^\circ\cap\t=(\operatorname{conv}(E-E)\cap\t)^\circ=P^\circ$, so we get the following:
If $B=\operatorname{conv}(E-E)^\circ$ is the unit ball of Hofer’s norm and $P=\operatorname{conv}(E-E)\cap\t$ is Hofer’s norm polytope then for non-zero $x\in\t_+$ $$F_x(B)=\operatorname{Ad}_{Z(x)}(F_x(P^\circ)\cap\t_+).$$
The same characterization holds for the one-sided Hofer norm and its polytope $P^+$.
The geometry of groups with bi-invariant Finsler metrics {#l}
========================================================
In this section we focus our study of geodesics of Lie groups with Finsler metrics obtained by (left or right) translation of $\operatorname{Ad}$-invariant Finsler norms. We want to emphasize that in this section, we only require that norms are positively homogeneous, i.e. $\|\lambda x\|=\lambda \|x\|$ for $\lambda \ge 0$, thus including our one-sided Hofer norms. Therefore the distance obtained is not necessarily symmetric.
We need to begin this section with some remarks concerning compact semi-simple Lie algebras and their root decomposition.
\[milnor\]A connected finite dimensional Lie group $K$ admits a bi-invariant *Riemannian* metric if and only if it is isomorphic to the cartesian product of a compact group and an additive vector group (see [@milnor Lemma 7.5]). Therefore in that case $\k\simeq \k_0\oplus \mathfrak a$ where $a$ is an abelian sub-algebra and $\k_0$ is semi-simple. To simplify the discussion, we will only consider here compact groups $K$ with semisimple Lie algebra $\k$. Thus $K$ is always unimodular, in particular $\operatorname{tr}\operatorname{ad}v=0$ for any $v\in \k$. The $\operatorname{Ad}$-invariant inner product is given by (minus) the Killing form of $\k$, as before, and for this $\operatorname{Ad}$-invariant inner product, the operator $\operatorname{ad}v$ is skew-adjoint for any $v\in \k$ ([@milnor Lemmas 6.3 & 7.2]).
What follows, in the form of a Remark, is in fact a recollection of useful facts that we will use about the real root decomposition of a real semi-simple Lie algebra. It will also help to fix and clarify the notation.
\[rootdecom\] Let $\Delta$ be the set of (real) roots of $\k$ with respect to a fixed Cartan subalgebra $\t $, and denote $\Delta_+$ the positive roots. There is an orthonormal set (the real root vectors) $$\{u_\alpha,v_\alpha:\alpha \in\Delta_+\}\subseteq \k$$ with respect to this inner product, such that for each $h\in\t$ $$\begin{aligned}
\label{weights}
[h,u_\alpha] & =-\alpha(h)v_\alpha \qquad [h,v_\alpha]=\alpha(h)u_\alpha \qquad [u_\alpha,v_\alpha]= h_\alpha,\end{aligned}$$ where $h_\alpha\in \t$ is the unique element such that $\langle h_\alpha,\cdot\rangle=-\alpha( \cdot )$ (see for instance [@knapp Chapter 6]). Set $$Z_\alpha=\R u_\alpha\oplus\R v_\alpha.$$ Then $\k=\t\oplus\bigoplus_{\alpha\in\Delta_+}Z_\alpha$, and moreover for each $h\in \t$, we have $$\label{diago}
\operatorname{ad}h=i\sum_{\alpha\in\Delta_+} \alpha(h)(u_{\alpha}\otimes v_{\alpha}-v_{\alpha}\otimes u_{\alpha})=i \sum_{\alpha\in\Delta_+} \alpha(h) T_{\alpha}$$ were we write $T_{\alpha}=(u_{\alpha}\otimes v_{\alpha}-v_{\alpha}\otimes u_{\alpha})$ for short. Note that $T_{\alpha}T_{\beta}=0$ when $\alpha\ne \beta$.
\[perm\] The Weyl group of $K$ acts transitively on the roots (see the remark in [@serre p. 47]). Thus, if $\sigma$ is a permutation of the roots, there exists $k_{\sigma}\in K$ such that $\operatorname{Ad}_{k_{\sigma}}u_{\alpha}=u_{\sigma \alpha}$ and likewise with $v_{\alpha}$. Let $w\in \t$, and assumme that $T_{\sigma}$ is the linear transform in $\k$ obtained by permuting the eigenvalues of $\operatorname{ad}w$, then $$\begin{aligned}
T_{\sigma} & = i\sum_{\alpha\in\Delta_+} \sigma\alpha(w)(u_{\alpha}\otimes v_{\alpha}-v_{\alpha}\otimes u_{\alpha})= i\sum_{\alpha\in\Delta_+} \alpha(w)(u_{\sigma \alpha}\otimes v_{\sigma \alpha}-v_{\sigma\alpha}\otimes u_{\sigma\alpha})\\
& = \operatorname{Ad}_{k_{\sigma}} i\sum_{\alpha\in\Delta_+} \alpha(w)(u_{\alpha}\otimes v_{\alpha}-v_{\alpha}\otimes u_{\alpha})\operatorname{Ad}_{k\sigma}^{-1}\\
&= \operatorname{Ad}_{k_{\sigma}}\operatorname{ad}w \operatorname{Ad}_{k_{\sigma}}^{-1}=\operatorname{ad}(\operatorname{Ad}_{k_{{\sigma}}}w).\end{aligned}$$ Then $T_{\sigma}=\operatorname{ad}(\operatorname{Ad}_{k_{{\sigma}}}w)$. It is clear that $T_{\sigma}$ commutes with $\operatorname{ad}w$, and since $\k$ is semi-simple, $\operatorname{Ad}_{k_{\sigma}}w$ commutes with $w$ and in particular $w_{\sigma}=\operatorname{Ad}_{k_{{\sigma}}}w\in \t$.
Geodesics of groups with bi-invariant Finsler metrics
-----------------------------------------------------
First we recall here some useful facts regarding norming functionals in $\k$, their proofs are quite elementary and can be found in [@lar19 Section 4].
\[gausslem\] Let $\k$ be the Lie algebra of $K$ equipped with an $\operatorname{Ad}$-invariant norm (or Finsler norm). As mentioned in the introduction, it is relevant to remark here that the norm does not need to be homogeneous, it is only necessary that it is positively homogeneous. For $v\in\k$ let $\varphi$ be a norming functional of $v$ of $-v$. Then
1. We have $\varphi([w,v])=0$ for all $w\in\k$. Equivalently, $\varphi\circ \operatorname{ad}v\equiv 0$ on $\k$.
2. $\varphi(e^{\lambda \operatorname{ad}v}w)=\varphi(w)$ for all $w\in\k$ and $\lambda\in\R$.
3. $\varphi(e^{-v}D\exp_v w)=\varphi(w)$ for all $w\in\k$.
When the norm $\| \cdot \|$ is a norm derived from an inner product, the third item above is in fact Gauss’ Lemma of Riemannian geometry (it is well-known that the Riemannian exponential map, for a bi-invariant Riemannian metric in a Lie group, coincides with the group exponential). We next show how to describe the boundary of a ball in $K$. We first recall a definition.
Let $B$ be a convex body in a vector space $V$ and let $v\in\operatorname{bd}B$. The *solid tangent cone* at $v$ is $$TC_v:=\bigcap\{H_v^-:H_v \mbox{ is a supporting hyperplane of }B\mbox{ at }v\}.$$
If $v\in (V,\|\cdot\|)$ and $\|v\|=r>0$, then taking $B=B_r(0)$ it is apparent that $$TC_v:=\bigcap\{\varphi^{-1}(-\infty,r]: \varphi \mbox{ is a norming functional of }v\}.$$
Therefore, by Gauss’ Lemma above, for each $0\ne v\in \k$, if $w\in TC_v$ then $ e^{-v}D\exp_v w\in TC_v$. Moreover, if $v$ is such that the differential of the exponential map is invertible (for instance, if $v$ is smaller that the injectivity radius of $K$, see Definition \[injrad\] below), then it is clear that $$D\exp_v(TC_v)=e^vTC_v.$$ Geometrically, the differential of the exponential map at $v$ acts on the tangent cone at $v$ as the left translation.
\[normicon\] Any functional $\varphi$ can be described as $\varphi(v)=\langle v,a\rangle$ for some $a\in \k$, via the inner product in $\k$ (Remark \[milnor\]). That is $\varphi=-\operatorname{tr}(\operatorname{ad}a\;\cdot\;)$. If $\varphi(z)=\|z\|$ for some $z\in \k$, then $\varphi([v,z])=0$ for all $v\in\k$ (Remark \[gausslem\]). Thus for any $v\in \k$, $$0=\varphi([v,z])=\langle [v,z],a\rangle =\langle v,[z,a]\rangle$$ and then $[z,a]=0$. Therefore $\operatorname{ad}a$ and $\operatorname{ad}z$ commute, and they are simultaneously diagonalizable. Let $\{e_i\}_{i=1,\dots,N}$ be an orthonormal basis of $\k^{\mathbb C}$ that diagonalizes both simultaneously, let $q_{i,j}=e_i\otimes e_j$ and $q_i=q_{i,i}$ the corresponding rank-one orthogonal projections. If $\operatorname{ad}z=i\sum_{j\in J} z_j q_j$ is the spectral decomposition of the skew-adjoint operator $\operatorname{ad}z$, we can write $$A=\operatorname{ad}a=i\sum_{j\in J} a_j q_j +iB$$ for certain $a_j\in \mathbb R$ and $B^*=B=\sum_{k,l\notin J}B_{kl}q_{kl}$, with $Bq_j=0$ for all $j\in J$.
\[normicon2\] If $\varphi$ norms $z\in \k$ as in the previous remark, and $\|\varphi\|=1$, drop the term $B$ and all the $a_j=0$, and consider $\psi=-\operatorname{tr}({\widetilde}{A}\quad \cdot)$, where ${\widetilde}{A}=i\sum_j a_jq_j$. Then $\psi$ is still norming for $z$ and has unit norm. If we drop some of the non-zero $a_j$, its norm shrinks, and moreover $a_jz_j>0$ for each non-zero $z_j$.
Note first that $\psi$ is still norming for $z$: $$\psi(z)=\sum_j a_j z_j\operatorname{tr}(q_j)=\operatorname{tr}(A\circ\operatorname{ad}z)=\varphi(z)=\|z\|.$$ Now for any $x\in \k$, write $\operatorname{ad}x$ as a block matrix in terms of the $q_j$ and its orthogonal complement, $$A=\operatorname{ad}a=\left(
\begin{array}{cc} ia_j & 0 \\
0 & B
\end{array} \right)\qquad
{\widetilde}{A}=\left(
\begin{array}{cc} ia_j & 0 \\
0 & 0
\end{array} \right)\qquad
\operatorname{ad}x=\left(
\begin{array}{cc} x_{ij} & \ast \\
\ast& x_0
\end{array} \right)$$ where $x_0q_j=0$ for all $j$. Then $$\begin{aligned}
1 & =\|\varphi\|=\sup\limits_{\|x\|=1} \varphi(x)=\sup\limits_{\|x\|=1} -\operatorname{tr}(\operatorname{ad}a\circ\operatorname{ad}x)= \sup\limits_{\|x\|=1} \sum_j a_j x_{jj}+ \operatorname{tr}(Bx_0)\\
& \ge \sup\limits_{\|x\|=1, \, x_0=0} \sum_j a_j x_{jj}+ \operatorname{tr}(Bx_0)= \sup\limits_{\|x\|=1, \, x_0=0} \sum_j a_j x_{jj}\\
& = \sup\limits_{\|x\|=1} \sum_j a_j x_{jj} = \sup\limits_{\|x\|=1} -\operatorname{tr}({\widetilde}{A}\circ\operatorname{ad}x) = \sup\limits_{\|x\|=1}\psi(x)=\|\psi\|.\end{aligned}$$ This proves that $\|\psi\|\le 1$, but since $\psi(z)=\|z\|$, it must be $\|\psi\|=1$. Now consider ${\widetilde}{\psi}$ which is exactly as $\psi$ but omitting $a_d$ with $d\in J$. Then if $x\in \k$ and $\operatorname{ad}x=\sum_{k,l}x_{kl}q_{kl}$ with the $q_{kl}$ as in the previous Remark, we have $$\begin{aligned}
\|{\widetilde}{\psi}\| & =\sup\limits_{\|x\|=1} {\widetilde}{\psi}(x)=\sup\limits_{\|x\|=1} \sum_{j\in J,j\ne d}a_jx_{jj}=\sup\limits_{\|x\|=1,\, x_{dd}=0} \sum_{j\in J}a_jx_{jj}\\
& \le \sup\limits_{\|x\|=1} \sum_{j\in J}a_jx_{jj}= \sup\limits_{\|x\|=1} \psi(x)=\|\psi\|=1.\end{aligned}$$ Finally, we claimed that since $\psi(z)=\|z\|$ it must be $a_jz_j>0$ for each non-zero $z_j$. Otherwise, suppose that $a_dz_d<0$ then $$\|z\|=\psi(z)=\sum_{j\in J} a_jz_j <\sum_{j\ne d} a_jz_j=\sum_{j\ne d} {\widetilde}{a}_jz_j={\widetilde}{\psi}(z)\le \|{\widetilde}{\psi}\|\|z\|\le \|z\|,$$ a contradiction.
We now charachterize the linear order given by Finsler $\operatorname{Ad}$-invariant norms; results in the same vein can be found in [@atiyah Section 12], [@bhatia Proposition 6] and [@tam Proposition 2.8]. For $z,w\in \k$ we denote $\overrightarrow{w}=(w_1,w_2,\dots,w_N)$ the string of real numbers such that $i w_j$ are the eigenvalues of $\operatorname{ad}w$ in the complexification $\k^{\mathbb C}$ of $\k$, and likewise with $\overrightarrow{z}$.
\[majo\] Let $z,w\in\k$, let $N=\dim(\k)$. The following are equivalent:
1. $z\in \operatorname{conv}\O_w$, more precisely there exist (at most) $N+1$ points $k_i\in K$ and $N+1$ real numbers $\lambda_i \ge 0$ with $\sum_i \lambda_i= 1$ such that $$z=\sum_{i=1}^{N+1} \lambda_i\, \operatorname{Ad}_{k_i}w.$$
2. $\overrightarrow{z}\prec\overrightarrow{w}$ (strong majorization).
3. $\|z\|\le \|w\|$ for all $\operatorname{Ad}$-invariant Finsler norms in $\k$.
4. $\max_{k\in K} \langle z, \operatorname{Ad}_k x\rangle\le \max_{k\in K} \langle w, \operatorname{Ad}_k x\rangle$ for all $x\in \k$.
If moreover equality holds for some Finsler norm, then $z$ and all the $\operatorname{Ad}_{k_i}w$ lie in the same face of the ball for that norm (and in fact lie in the intersection of all the faces that $z$ lies in). If that norm is strictly convex then $z=\operatorname{Ad}_k w$ for some $k\in K$.
The fact that any element in the orbit can be written with a prescribed number of combinations ($N+1$) is a consequence of Caratheodory’s theorem. Let us establish first the equivalence $(1)\Leftrightarrow (2)$. Assume $(1)$, passing to the adjoint representation we have $$\operatorname{ad}z= \sum_{i=1}^N \lambda_i \operatorname{Ad}_{k_i} \operatorname{ad}w \operatorname{Ad}_{k_i}^{-1}.$$ Each $\operatorname{Ad}_{k_i}$ is a unitary operator acting in $\k^{\mathbb C}$ therefore $\operatorname{ad}z$ belongs to the convex hull of the coadjoint orbit of $\operatorname{ad}w$, and this in turn (and by Schur-Horn’s theorem) implies that strong majorization $\overrightarrow{z}\prec \overrightarrow{w}$ holds. Now assume $(2)$ holds, let $\t$ be a Cartan subalgebra containing $w$, let $\Delta^+$ be the positive simple roots and let $k\in K$ such that $\operatorname{Ad}_k z\in \t$. The spectrum of $$\operatorname{ad}(\operatorname{Ad}_k z)=\operatorname{Ad}_k \operatorname{ad}z \operatorname{Ad}_k^{-1}$$ is also the string $\overrightarrow{z}$, and the assumption implies $\overrightarrow{z}$ is in the convex hull of the permutations of the eigenvalues of $\operatorname{ad}w$ (see [@bhatia2 Theorem II.1.10]). Equivalently (and again invoking Caratheodory’s theorem), there are $N+1$ such elements with $$\operatorname{ad}(\operatorname{Ad}_k z)=\sum_{i=1}^{N+1} \lambda_i T_{\sigma_i}$$ for a certain string of non-negative numbers $(\lambda_i)_{i=1,\dots , N+1}$ such that $\sum_i\lambda_i=1$. Now each $T_{\sigma_i}$ is obtained permuting the eigenvalues of $\operatorname{ad}w$ or equivalently, permuting the roots $\alpha\in\Delta_+$. By Remark \[perm\], any such permutation is obtained by an inner automorphism, therefore $$\operatorname{ad}(\operatorname{Ad}_k z)=\sum_{i=1}^{N+1}\lambda_i \operatorname{ad}(\operatorname{Ad}_{h_i}w)=\operatorname{ad}(\sum_{i=1}^{N+1}\lambda_i \operatorname{Ad}_{h_i}w),$$ and by the semi-simplicity of $\k$, we obtain $z=\sum_{i=1}^{N+1}\lambda_i \operatorname{Ad}_{k_i}w \in \operatorname{conv}\O_w$, where $k_i=k^{-1}h_i$.
Clearly $(1)\Rightarrow (3)$, and $(3)\Rightarrow (4)$ when $x\in\k$ is a regular element (so we obtain a non-degenerate Finsler norm, see Remarks \[semiads\] and \[onesided\]). Since regular elements are dense in $\k$, it is straightforward to see that $(4)$ must then hold for any $x\in \k$, if $(3)$ holds. Now assume that $(1)$ does not hold, then by Hahn-Banach’s theorem there exists a linear functional $\varphi$ separating $z$ and the convex capsule of the orbit, i.e. $$\varphi (y) \le r<\varphi(z) \qquad \forall\, y\in \operatorname{conv}(\O_w).$$ Let $x\in \k$ such that $\varphi=\langle x,\,\cdot\rangle$, then by Remark \[semiads\] $$\max_{k\in K}\langle w, \operatorname{Ad}_k x\rangle=\max_{k\in K}\langle\operatorname{Ad}_k w,x\rangle =\max\{\langle y , x\rangle: y\in \operatorname{conv}(\O_w)\}\le r<\langle z,x\rangle.$$ This shows that $(4)$ cannot hold, finishing the proof of the equivalences. Now assumme that any of the conditions hold, and we have equality of norms $\|z\|=\|w\|$ holds for some norm, then $$\|w\|=\|z\|=\|\sum \lambda_i \operatorname{Ad}_{k_i}w\|\le \sum \lambda_i \|\operatorname{Ad}_{k_i}w\|=\sum \lambda_i \|w\|\le \|w\|.$$ We note that there is a common norming functional $\varphi$ for all the $\lambda_i \operatorname{Ad}_{k_i}w$ by Lemma \[mismacara\]; since $\lambda_i\ge 0$ this amounts for the $\operatorname{Ad}_{k_i}w$ in the same face of a sphere of the norm. Then also $$\varphi(z)=\varphi (\sum_i \lambda_i \operatorname{Ad}_{k_i}w)=\sum \lambda_i \varphi(\operatorname{Ad}_{k_i}w)=\sum \lambda_i \|w\|=\|w\|=\|z\|.$$ If $\varphi$ is any functional norming $z$, then $$\|z\|=\varphi(z)=\sum \lambda_i \varphi(\operatorname{Ad}_{k_i}w)\le \sum \lambda_i \|\operatorname{Ad}_{k_i}w\|=\|w\|=\|z\|$$ shows that it must be $\varphi(\operatorname{Ad}_{k_i}w)=\|\operatorname{Ad}_{k_i}w\|$ for all $i$, thus these vectors are in fact in the intersection of all the faces where $z$ lives. If the norm is strictly convex all the $\operatorname{Ad}_{k_i}w$ are aligned, but being normed by the same functional they must be equal therefore $z=\operatorname{Ad}_k w$ as claimed.
It suffices to check condition $(2)$ above only for *strictly convex norms* to obtain the equivalences. This is because any $\operatorname{Ad}$-invariant Finsler norm $\|\cdot\|$ can be approximated explicitly with a strictly convex ($\operatorname{Ad}$-invariant, Finsler) norm by means of $$\|x\|_{\varepsilon}=\|x\|+\varepsilon \|x\|_2.$$ Here $\|x\|_2^2=-\operatorname{tr}(\operatorname{ad}x\circ \operatorname{ad}x)$ is the norm derived from the Killing form of $\k$. Likewise, it suffices to check $(4)$ for regular $x\in \k$.
\[onesided\] Since $K$ is compact it is unimodular and then $0=\sum_j w_j=i\operatorname{tr}(\operatorname{ad}w)$ (Remark \[milnor\]). This easily implies that all the partial sums $\sum_{k=1}^m w_k$, with the $w_k$ rearranged in decreasing order, must be non-negative. Thus the vector $\overrightarrow{w}$ strongly majorizes the zero vector in $\mathbb R^N$, i.e. $\overrightarrow{0}\prec \overrightarrow{w}$, and by the previous proposition, $0\in \operatorname{conv}(\O_w)$ for any $w\in \k$. In particular the one-sided Hofer norms (Remark \[semiads\]) are in fact-nonnegative, regardless their degeneracy or non-degeneracy. Assumming that the orbit $\O_w$ is full, then $0$ must be an interior point of $\operatorname{conv}(\O_w)$, and then we obtain a true Finlser norm. This can be seen using the argument in [@bgh Lemma 6]: if $0$ is in the boundary of $\operatorname{conv}(\O_w)$, by Hahn-Banach’s separation theorem, there exists $0\ne x\in\k$ such that $\varphi_x(0)=0$ and $\varphi_x(\operatorname{conv}(\O_w))\geq 0$. But then it must be $\varphi_x(\O_w)= 0$ because otherwise $$\int_{k\in K}\varphi_x(\operatorname{Ad}_k(w))dk=\varphi_x(\int_{k\in K}\operatorname{Ad}_k(w)dk) >0$$ contradicting that $\int_{k\in K}\operatorname{Ad}_k(w)dk$ is a fixed point of the adjoint action, therefore it is $0$ because $\k$ is semi-simple. Since $\varphi_x(\O_w)= 0$, the orbit is not full.
### Domain of injectivity of the exponential
Since $K$ is a finite-dimensional Lie group, the exponential map of $K$ is a local diffeomorphism for some open ball of the norm $\|\cdot \|$. More precisely, let $\mathbf D\subseteq \k$ be a maximal open convex $\operatorname{Ad}$-invariant set, such that $V=\exp(\mathbf D)$ is open in $\k$ and $\exp:{\mathbf D}\to V$ is a diffeomorphism. It will be convenient to denote $\|\cdot\|_{\infty}$ to the Minkowski gauge of the set $\mathbf D$; it is an $\operatorname{Ad}$-invariant Finsler norm in $\k$ and we will refer to it as the *uniform norm*. Then we also define
\[injrad\] Let $K$ be a Lie group, $\| \cdot \|$ an $\operatorname{Ad}$-invariant Finsler norm on $\k$, and for $R>0$ let $B_R=\{v\in\k:\|v\|<R\}$, $V_R=\exp(B_R)$. If $\exp:B_R\to V_R$ is a diffeomorphism between open sets and $R$ is maximal, we call $R$ the *radius of injectivity* for the given norm.
Note that if $B_R\subseteq \mathbf D$ then $\exp:B_R\to V_R$ is a diffeomorphism between open sets, thus one looks for balls of the given norm that fit inside the domain of injectivity of the exponential map.
The condition $B_R\subseteq \mathbf D$ is equivalent to $B_R\cap\t\subseteq \mathbf D\cap\t$. In the case of Hofer norms $B_R\cap\t=R\{y_1,\dots,y_m\}^\circ$, where $\{y_1,\dots,y_m\}$ are the extreme points of the Hofer norm polytope. Note also that $\mathbf D\cap\t$ can be taken as the interior of $\cup_{w\in\W}w.C$, where $C$ is a Weyl alcove.
\[espectro\] For the group $\operatorname{SU}(n)$ we take $\mathbf D=\{z\in {\mathfrak su}_n: \|z\|_{\infty}<\pi\}$ where $\|\cdot\|_{\infty}$ now is exactly the usual spectral norm. Equivalently, $\mathbf D\cap\t$ is $$\{\operatorname{diag}(x_1,\dots,x_n)\in\t:|x_i|<\pi,\mbox{ for }i=1,\dots,n\}.$$ For the group $\operatorname{SU}(n)/\Z_n$ we take $$\mathbf D\cap\t=\{\operatorname{diag}(x_1,\dots,x_n)\in\t:|x_i|<\pi/n,\mbox{ for }i=1,\dots,n\}.$$
We now show that the convex body $\mathbf D$ (which depends only on $K$) is optimal in terms of lengths of segments (one-parameter subgroups), for any bi-invariant distance:
\[expono\] Let $z,w\in \k$ such that $e^z=e^w$. Assumme that $z\in \mathbf D/2$, Then $w$ commutes with $z$ and $\|z\|\le \|w\|$ for any $\operatorname{Ad}$-invariant Finsler norm $\|\cdot\|$ in $\k$. If equality of norms holds for a strictly convex norm, then $w=z$.
Assumme first that $w$ is regular, let $\mathfrak z(w)=\t$ denote the Cartan subalgebra. Note that $$\exp(e^{t \operatorname{ad}w}z)=\exp(\operatorname{Ad}_{e^{tw}}z)=e^{tw}e^ze^{-tw}=e^{tw}e^w e^{-tw}=e^w.$$ Therefore, differentiating at $t=0$ we obtain $D\exp_z([w,z])=0$, and since $z\in \mathbf D$, we conclude that $[w,z]=0$, and $z\in \t$. Since $e^{\operatorname{ad}w}=e^{\operatorname{ad}z}$, then $\exp(\operatorname{ad}(w-z))=\exp(\operatorname{ad}w-\operatorname{ad}z)=1$, implying that $\sigma(\operatorname{ad}(w-z))\subseteq 2\pi i \mathbb Z$. This implies, using equation (\[diago\]), that we can write $$i\sum_{\alpha\in \Delta_+} w_{\alpha} T_{\alpha} =\operatorname{ad}w=i\sum_{\alpha\in \Delta_+} (z_{\alpha}+ 2\pi n_{\alpha})T_{\alpha}$$ wiht $w_{\alpha},z_{\alpha}\in\mathbb R$ and $n_{\alpha}\in \mathbb Z$. It will be convenient to number the roots, so we let $J=card(\Delta_+)$ and we have $w_j=z_j+2\pi n_j$ for all $j\in J$, where some of the $z_j$ might be zero. Note that since $z\in \mathbf{D}/2$, then $2z\in \mathbf{D}$ therefore the exponential map is injective and a diffeomorphism in $t2z$ for $t\in [-1,1]$, and in particular it must be $\sigma(\operatorname{ad}2z)\subseteq (-2\pi i,2\pi i)$ (by inspection of the formula of the differential of the exponential map, see Remark \[DAD\] below). Thus we have $|2z_j|<2\pi$ for all $j\in J$ or equivalently $-\pi<z_j<\pi$. We can asumme that the $z_j$ are given in order $z_1\ge z_2\ge \dots \ge z_J$ (recall also $\sum z_j=0)$. Let us reorder the $w_j$ in decreasing order also, so there is a permutation $\sigma$ of $\{1,\dots, J\}$ such that if $v_j=w_{\sigma j}$ then $v_1\ge v_2\ge \dots \ge v_J$ (and we also have $\sum v_j=0$). From here it is also clear that $\sum n_j=0$. Let’s spare for a moment those $j$ such that $n_j=0$, and for the others, note that if $n_j>0$ and $n_r<0$ then $$z_j+2\pi n_j> -\pi+2\pi=\pi >-\pi =-2\pi +\pi > 2\pi n_r+z_r.$$ This shows that the $v_j=z_{\sigma j}+2\pi n_{\sigma j}$, with positive $n_{\sigma j}$, are always bigger than those with negative $n_{\sigma j}$. We split the indices in two sets: let $j_0$ be such that if $j\in {1,\dots, j_0}$ then $n_{\sigma j}>0$ and otherwise $n_{\sigma j}<0$ when $j_0+1\le j\le J$. We compute the sum of the $z_j$ up to any such $1\le j\le j_0$, we have $$\sum_{k=1}^j z_k \le \sum_{k=1}^j \pi \le j\pi.$$ On the other hand, it is clear that the sum of the first $j$ bigger $v_k$, for $k\le j$, must be of those $v_k$ with $n_{\sigma k}>0$, therefore $$\sum_{k=1}^j v_j=\sum_{k=1}^j z_{\sigma k}+ 2\pi n_{\sigma k}\ge -j\pi+2\pi j=j\pi.$$ Thus if $j\le j_0$, $$\label{mayorizando}
z_1+z_2+\dots + z_j\le v_1+ v_2+\dots +v_j.$$ Now assumme that $j\ge j_0+1$, and note that $$\sum_{k=1}^j z_k=-\sum_{k=j+1}^N z_k< \pi( N-(j+1)).$$ Likewise $$\sum_{k=1}^j v_k=-\sum_{k=j+1}^N v_k =-\sum_{k=j+1}^N (z_{\sigma k}+ 2\pi n_{\sigma k})$$ and note that now all the $n_{\sigma k}\le -1$, since $k\ge j+1\ge j_0+1>j_0$. Therefore $$\sum_{k=1}^j v_k \ge -\pi(N-(j+1))+2\pi(N-(j+1))=\pi(N-(j+1)),$$ and equation (\[mayorizando\]) is also valid for $j\ge j_0$. Let $\overrightarrow{v}=(v_1,v_2,\dots,v_N)$ and likewise $\overrightarrow{z}=(z_1,z_2,\dots,z_N)$. Then equation (\[mayorizando\]) together with $\sum z_j=\sum v_j=0$ tells us that $\overrightarrow{z}\prec \overrightarrow{v}$, that is $\overrightarrow{v}$ majorizes $\overrightarrow{z}$. Then by [@bhatia2 Theorem II.1.10], $\overrightarrow{z}$ is in the convex hull of all vectors obtained by permutating the coordinates of $\overrightarrow{v}$. Clearly, we can add those $w_j$ such that $w_j=z_j$ ($n_j=0$) and this still holds true. Since the $v_j$ are just a permutation of the $w_j$, then $\overrightarrow{z}$ is in fact in the convex hull of all vectors obtained permutating the coordinates of $\overrightarrow{w}=(w_1,w_2,\dots,w_J)$. By Remark \[perm\], we have $$\operatorname{ad}z=\sum_{{\sigma}} \lambda_{{\sigma}} \operatorname{ad}(\operatorname{Ad}_{k_{\sigma}}w)=\operatorname{ad}(\sum_{\sigma} \lambda_{\sigma} \operatorname{Ad}_{k_{\sigma}}w),$$ and since $\k$ is semi-simple, it must be $z=\sum \lambda_{\sigma} \operatorname{Ad}_{k_{\sigma}}w$. Then $$\|z\|\le \sum \lambda_{\sigma} \|\operatorname{Ad}_{k_{\sigma}}w\|=\sum \lambda_{\sigma} \|w\|=\|w\|$$ proving the claim for regular $w$. If $w$ is not regular, fix the bi-invariant distance in $K$ given by the uniform norm (the Minkowski Finsler norm of the convex set $\mathbf D$). For each $\varepsilon>0$ pick $w_{\varepsilon}\in\k$ such that $\|w_{\varepsilon}\|_{\infty}<\varepsilon$ and $w+w_{\varepsilon}$ is regular (regular elements are dense). Observe that $$\operatorname{dist}(e^w,e^{w+w_{\varepsilon}})\le \|w_{\varepsilon}\|_{\infty}<\varepsilon$$ by Theorem \[charactgeod\]. Therefore there exists $y_{\varepsilon}\in \k$ such that $e^we^{y_{\varepsilon}}=e^{w+w_{\varepsilon}}$. Again, note that $$\|y_{\varepsilon}\|_{\infty}=\operatorname{dist}(1,e^{y_{\varepsilon}})=\operatorname{dist}(1,e^{-w}e^{w+w_{\varepsilon}})=\operatorname{dist}(e^w,e^{w+w_{\varepsilon}})\le \|w\|_{\infty}<\varepsilon.$$ Now consider the map $f:v\mapsto e^{z+v}$. Since $f(0)=e^z$ and $Df_0=D\exp_z$, the hypotesis $z\in \mathbf{D}/2$ guarantees that $f$ is a local diffeomorphism from a $0$-neighbourhood to a neighbourhood of $e^z$. Therefore there exists a unique $z_{\varepsilon}\in \k$ in that neighbourhood, such that $e^{z+z_{\varepsilon}}=e^ze^{y_{\varepsilon}}$. Note also that when $y_{\varepsilon}\to 0$, then also $z_{\varepsilon}\to 0$. In particular, for small $\varepsilon>0$, $z+z_{\varepsilon}\in \mathbf{D}/2$ just like $z$. Then from $$e^{z+z_{\varepsilon}}=e^ze^{y_{\varepsilon}}=e^we^{y_{\varepsilon}}=e^{w+w_{\varepsilon}}.$$ and the previous proof, we can conclude that $\|z+z_{\varepsilon}\|\le \|w+w_{\varepsilon}\|$ for any $\operatorname{Ad}$-invariant Finsler norm in $\k$. Letting $\varepsilon\to 0$ gives us the desired inequality $\|z\|\le \|w\|$.
Now assumme that there is an equality of norms for a strictly convex norm, then by Proposition \[majo\], $z=\operatorname{Ad}_k w$ for some $k\in K$, and in particular $w\in \mathbf D/2$ also. But $e^w=e^z$ and the injectivity of the exponential map implies $z=w$.
If $z,w$ are as in the previous theorem, then by Proposition \[majo\], we have $$z=\sum_{i=1}^{N+1} \lambda_i\, \operatorname{Ad}_{k_i} w,$$ for some $k_i\in K$, $\lambda_i\ge 0$ with $\sum_i\lambda_i=1$. We also mention here that the proof of the previous theorem shows that when $w$ is regular, the $k_i$ are in the Weyl group of $K$.
For linear Lie groups such as $K=\operatorname{SU}(n)$, the passage to the adjoint representation is unnecessary. Thus from $e^z=e^w$ with $z\in \mathbf D$ we can conclude that $[w,z]=0$ and therefore $$w=z+ 2\pi i \sum_j n_j q_j.$$ With the same proof as the previous theorem, we now obtain the same result for $z\in \mathbf D$, i.e. $\|z\|_{\infty}<\pi$ (and not just $z\in \mathbf D/2$):
\[lingr\] Let $K\subseteq M_n(\mathbb C)$ be compact semi-simple linear Lie group. Let $z,w\in \k$ such that $e^z=e^w$, and assumme that $z\in \mathbf D$. Then $\|z\|\le \|w\|$ for any $\operatorname{Ad}$-invariant Finsler norm $\|\cdot\|$ in $\k$, and if equality of norms holds for a strictly convex norm, then $w=z$.
### Local Hopf-Rinow theorem and the characterization of geodesics {#1}
Let us take a look at geodesics of a Lie group with a bi-invariant Finsler metric. We recall here the fundamental results about geodesics, for proofs see [@lar19 Section 4]. For a given Finsler norm, let $R>0$ be an injectivity radius. The results are stated in terms of the left logarithmic derivatives $\gamma^{-1}_t\dot{\gamma}_t$ and hold also if stated in terms of the right logarithmic derivative since the norm is $\operatorname{Ad}$-invariant.
We call a curve $\gamma:[0,1]\to K$ *short* or we say that $\gamma$ is a *geodesic* if it minimizes the length functional $$L(\gamma)=\operatorname{length}(\gamma)=\int_a^b\|\dot{\gamma}_t\|_{\gamma_t}dt=\int_a^b\|\dot{\gamma}_t\gamma_t^{-1}\|dt=\int_a^b\|\gamma_t^{-1}\dot{\gamma}_t\|dt$$ among all curves in $K$ with the same endpoints.
\[charactgeod\] Let $u_0,u_1=u_0e^z\in K$ with $\|z\|< R$.
1. If $\delta(t)=u_0e^{tz}$, $t\in [0,1]$, then $\delta$ is shorter than any other piecewise $C^1$ path $\gamma$ in $K$ joining $u_0,u_1$ and $\operatorname{dist}(u_0,u_1)=\|z\|$.
2. If $v,w\in\k$ then $$\operatorname{dist}(e^v,e^w) \leq \|w-v\|$$ and if $w,v$ commute and $\|w-v\| \leq R$, then equality holds (this is known as the exponential metric decreasing property).
3. Let $\Gamma:[a,b]\to \k$ be a piecewise $C^1$ short path joining $0,z$, let $\gamma=e^{\Gamma}$. Then $\|\gamma^{-1}_t\dot{\gamma}_t\|=\|\dot{\Gamma}_t\|$ for all $t$, and $\gamma$ is short in $K$ with the same length than $\Gamma$. Moreover if $\varphi$ is norming functional for $z$, then $$\varphi(\gamma^{-1}_t\dot{\gamma}_t)=\|\gamma^{-1}_t\dot{\gamma}_t\|=\|\dot{\Gamma}_t\|=\varphi(\dot{\Gamma}_t)\quad \forall t\in [a,b],$$ thus $\gamma^{-1}\dot{\gamma}$ (normalized) sits inside a face of the unit sphere of the norm.
4. $\gamma:[a,b]\to K$ is a piecewise $C^1$ short path joining $1,e^z$ in $K$ if and only if $\gamma=e^{\Gamma}$ for a piecewise $C^1$ path $\Gamma:[a,b]\to \k$ joining $0,z$ (with $\|\Gamma_t\|\le R$) and $$\varphi(\dot{\Gamma}_t)=\|\gamma_t^{-1}\dot{\gamma}_t\|=\varphi(\gamma_t^{-1}\dot{\gamma}_t)$$ for some unit norm functional $\varphi$ and all $t\in [a,b]$ (and then this holds for any norming functional of $z$).
5. If $z/\|z\|$ is an extremal point of the unit sphere of $\k$, then the only short piecewise $C^1$ path joining $1,e^z$ in $K$ is (a reparametrization of) the segment $\delta(t)=e^{tz}$.
These results were established in [@lar19] with some generality; for finite dimensional groups we can improve the existence of short paths invoking the metric version of Hopf-Rinow’s theorem. As usual, here $K$ denotes a connected compact Lie group with semi-simple Lie algebra $\k$. The last item of this theorem extends significanly (to this family of Lie groups) the results obtained in [@alv] for the group $U(n)$.
\[existen\] Let $\operatorname{dist}$ be a bi-invariant metric in $K$ (i.e. from an $\operatorname{Ad}$-invariant Finsler norm $\|\cdot\|$ in $\k$). Then
1. For each $u_1,u_2\in K$ there exists a short polygonal path $\delta$ joining them, i.e. a concatenation of segments $t\mapsto u_ie^{tz_i}$ such that $\|z_i\|<R$ and $$L(\delta)=\sum \|z_i\|=\operatorname{dist}(u_1,u_2).$$
2. If the norm is strictly convex, there exists $w\in \k$ such that $\|w\|=\operatorname{dist}(u_1,u_2)$ and the segment $t\mapsto u_1e^{tw}$ is a short path joining them. Any short path is a reparametrization of a segment, and if $\operatorname{dist}(u_1,u_2)<R$, there is exactly one short segment joining them.
3. If $z\in \mathbf D/2$ ($z\in \mathbf D$ for linear Lie groups), then $\operatorname{dist}(1,e^z)=\|z\|$. If the norm is strictly convex, $t\mapsto e^{tz}$ is the unique short path joining them.
Let $\operatorname{dist}_g$ denote the bi-invariant distance induced by the $\operatorname{Ad}$-invariant metric given by the Killing form in $\k$. It is well-known that such $\operatorname{dist}_g$ metric has one-parameter groups as Riemannian geodesics, therefore it is geodesically complete. By Hopf-Rinow’s theorem $(K,\operatorname{dist}_g)$ is metrically complete, and therefore $(K,\operatorname{dist})$ is metrically complete since both metrics are uniformly equivalent (since both are bi-invariant and the tangent norms are uniformly equivalent, $\k$ being finite dimensional). Since any of these metrics induce the original topology of $K$, and since $K$ is a finite dimensional manifold, it is locally compact and we can apply Cohn-Vossen’s theorem [@burago Theorem 2.5.28] to the metric space $(K,\operatorname{dist})$. This theorem tells us that any approximating sequence of paths in $K$ has a limit point $\gamma:[0,1]\to K$ such that $\gamma$ is rectifiable and $L(\gamma)=\operatorname{dist}(u_1,u_2)$. Partition $\gamma$ into finite pieces, in points denoted $\gamma_i=\gamma_{t_i}$, such that $\operatorname{dist}( \gamma_i,\gamma_{i+1})<R$. Write $\gamma_{i+1}=\gamma_ie^{z_i}$ using Theorem \[charactgeod\](1), hence if $\delta$ is the concatenation of these paths, $$L(\delta)=\sum \|z_i\|= \sum \operatorname{dist}(\gamma_i,\gamma_{i+1})=\operatorname{dist}(u_1,u_2)$$ which shows that $\delta$ is minimizing. If the norm is strictly convex, it can be shown that the $z_i$ commute, hence we can replace the concatenation of segments with a segment; the proof of this and the local uniqueness can be found in [@lar19 Theorem 4.15]. Now assumme that $z$ is in (half of) the domain of injectivity of the exponential map. Assumme first that the norm is strictly convex. By the previous item of this theorem, there exists $w\in \k$ such that $t\mapsto e^{tw}$ joins $1,e^z$ and such that $\|w\|=\operatorname{dist}(1,e^z)\le \|z\|$. But Theorem \[expono\] also tells us that $\|z\|\le \|w\|$, therefore $\|z\|=\|w\|=\operatorname{dist}(1,e^z)$. Now let $\|\cdot\|$ be any $\operatorname{Ad}$-invariant norm, let $\varepsilon >0$ and let $\|v\|_g=\sqrt{\langle v,v\rangle}$ be an $\operatorname{Ad}$-invariant Riemannian metric in $\k$. Consider $$|v|_{\varepsilon}=\|v\|+ \varepsilon \|v\|_g,$$ and note that $|\cdot|_{\varepsilon}$ is $\operatorname{Ad}$-invariant and strictly convex. Therefore by what we just proved, if $z\in \mathbf D/2$ then $$\|z\|\le |z|_{\varepsilon}=\operatorname{dist}_{\varepsilon}(1,e^z)\le L_{\varepsilon}(\gamma)=L(\gamma)+\varepsilon L_g(\gamma)$$ for any piecewise smooth path $\gamma$ joining $1,e^z$ in $K$. Letting $\varepsilon\to 0^+$ first, and taking the infimum over the paths $\gamma$, shows that $\|z\|\le \operatorname{dist}(1,e^z)$ as claimed. If the norm is strictly convex, any other short path is also a segment $t\mapsto e^{tw}$ by the second item of this theorem. Thus $e^w=e^z$ and $\|w\|=\|z\|$, and by Theorem \[expono\], we conclude that $w=z$. For linear Lie groups we can replace half of $\mathbf D$ with the full set $\mathbf D$ by Corollary \[lingr\].
From the last assertion of the theorem we can give a nice characterization of the product of exponentials. This is connected with the noncommutative Horn inequalities as studied by Belkale et al, see [@belkale; @entov] and the references therein.
Let $x,y,z\in \k$ with $z\in\overline{\mathbf D}/2$ such that $e^xe^y=e^z$. Then
1. $\|z\|\le \|x+y\|$ for any $\operatorname{Ad}$-invariant Finsler norm in $\k$.
2. Let $N=\dim(\k)$, then there exist (at most) $N+1$ points $k_i\in K$ and $N+1$ real numbers $\lambda_i \ge 0$ with $\sum_i \lambda_i= 1$ such that $$z=\sum_{i=1}^{N+1} \lambda_i\, \operatorname{Ad}_{k_i}(x+y).$$
3. If equality holds for some Finsler norm, then $z$ and all the $\operatorname{Ad}_{k_i}(x+y)$ lie in the same face of the ball for that norm (in fact, in the intersection of all the faces where $z$ sits).
4. If equality holds for a strictly convex norm then $x,y$ commute, $z=\operatorname{Ad}_k(x+y)$ for some $k\in K$ (thus $x+y\in \overline{\mathbf D}/2$) and $k$ commutes with $e^z$. If moreover $z\in \mathbf D/2$ then $x,y$ commute and $z=x+y$.
Let $\beta(t)=e^{tx}e^{ty}$, which joins $1,e^z$ in $K$. Note that $\beta_t^{-1}\dot{\beta}_t=e^{-t\operatorname{ad}y}(x+y)$ therefore $L(\beta)=\|x+y\|$ for any $\operatorname{Ad}$-invariant Finsler norm in $\k$. By the previous theorem, the third assertion also holds for the closure of $\mathbf D/2$ therefore we must have $$\|z\|\le L(\beta)=\|x+y\|$$ for any such Finsler norm. Assertions $2$ and $3$ follow from Proposition \[majo\]. If the norm is strictly convex then $z=\operatorname{Ad}_k(x+y)$ by Proposition \[majo\], but also $x,y$ commute by [@lar19 Theorem 4.17]. This implies $e^z=e^{\operatorname{Ad}_k(x+y)}=ke^{x+y}k^{-1}=ke^xe^yk^{-1}=ke^zk^{-1}$ thus $k$ commutes with $e^z$. When $z\in \mathbf D/2$, the condition $e^{\operatorname{Ad}_k z}=e^z$ is only possible if $\operatorname{Ad}_k z=z=\operatorname{Ad}_k(x+y)$ therefore $z=x+y$.
For linear Lie groups such as $K=\operatorname{SU}(n)$, and by Corollary \[lingr\], the same results stated in the previous corollary hold for $z\in \overline{\mathbf D}$, i.e. $\|z\|_{\infty}\le \pi$ (and not just $z\in \mathbf \overline{\mathbf D}/2$).
Characterization of all short paths
-----------------------------------
Before we proceed with the characterization of other short paths (for the case of non-strictly convex norms), we recall two results concerning the exponential map and the adjoint representation of the group $K$ and its differentials:
\[remark\] If $g_t$ is a smooth path in $K$, $$\label{DAD}
\frac{d}{dt}\operatorname{Ad}_{g_t}v=\operatorname{Ad}_{g_t}[g_t^{-1}\dot{g}_t,v]\qquad \forall v\in \k.$$ This follows by noting that to compute the derivative we can take $g_t=g_0e^{tg_0^{-1}\dot{g}_0}$, and differentiate at $t=0$, thus $$\operatorname{Ad}_{g_t}v=\operatorname{Ad}_{g_0} \operatorname{Ad}_{e^{tg^{-1}_0\dot{g}_0}}v=\operatorname{Ad}_{g_0}e^{t\operatorname{ad}g^{-1}_0\dot{g}_0}v=\operatorname{Ad}_{g_0}v+ t\operatorname{Ad}_{g_0}\operatorname{ad}(g^{-1}_0\dot{g}_0)v+o(t^2).$$ On the other hand if $v,w\in \k$ then it is well-known that $$\label{dexp}
e^{-v}D\exp_v w=\int_0^1 e^{-\lambda \operatorname{ad}v}w\,d\lambda=F(\operatorname{ad}v)w$$ where $F(\lambda)=\frac{1-e^{-\lambda}}{\lambda}$ is extended by $F(0)=1$ to be an holomorphic function in $\mathbb C$. In particular, note that if $v\in \mathbf D$ (the domain of injectivity of $\exp$), then $F(\operatorname{ad}v)$ must be nonsingular and in particular $\pm 2\pi i\notin \sigma(\operatorname{ad}v)$; otherwise we would have $0\in \sigma(F(\operatorname{ad}v))$. Moreover, it must be $\sigma(\operatorname{ad}v)\subseteq i(-2\pi ,2\pi )$, otherwise replacing $v$ with $tv$ for some $t\in (0,1)$ we would obtain a contradiction (recall that $\mathbf D$ is convex).
With these tools at hand we now characterize short paths $\gamma$ without the restriction of having length less that the radius of inyectivity of the group. In the next theorem $K$ is a connected compact semi-simple Lie group with the metric induced by an $\operatorname{Ad}$-invariant Finsler norm in $\k$.
\[cuasi\] Let $\gamma:[a,b]\to K$ be a piecewise $C^1$ path in $K$. If $\gamma$ is short for the bi-invariant metric, then for (almost) all $t$ $$\label{cuasiaut}
\varphi(\gamma_t^{-1}\dot{\gamma}_t)=\|\gamma_t^{-1}\dot{\gamma}_t\|$$ for some unit norm functional $\varphi$. Reciprocally, if the equality holds for some $\varphi$ and (almost) all $t\in [t_0,t_1]$, and $L(\gamma)_{t_0}^{t_1}\le R$, then $\gamma$ is short in $[t_0,t_1]\subseteq [a,b]$.
By the invariance of the metric, it suffices in all cases to consider paths starting at $u=1$. If $\gamma$ is short, assume first that its length is smaller than $R$, then for any $t\in [a,b]$ we have $\operatorname{dist}(\gamma(a),\gamma(t))= L(\gamma)_a^t<R$. Therefore we can lift $\gamma_t= e^{\Gamma_t}$ for a rectifiable path $\Gamma:[a,b]\to\k$, which does not leave the ball $B_R$. Then by Theorem \[charactgeod\], for any norming functional of $z$ and any $t$ we have $$\varphi(\Gamma_t)=\int_{a}^t \varphi(\dot{\Gamma})=\int_{a}^t \|\gamma^{-1}\dot{\gamma}\|=L(\gamma)_{a}^t=\operatorname{dist}(\gamma_{a},\gamma_{a}e^{\Gamma_t})=\|\Gamma_t\|.$$ Therefore again by the previous theorem and by Gauss’ Lemma \[gausslem\], $$\begin{aligned}
\|z\| & =\varphi(z)=\int_{a}^{b} \varphi(\dot{\Gamma}_s)ds=\int_{a}^b \varphi(e^{-\Gamma_s}D\exp_{\Gamma_s}\dot{\Gamma}_s) ds=\int_a^b \varphi(\gamma^{-1}_s\dot{\gamma}_s)ds \\
& \le \int_a^b\|\gamma^{-1}_s\dot{\gamma}_s\|ds =L(\gamma)_a^b=\operatorname{dist}(1,e^z)=\|z\|,\end{aligned}$$ and this is only possible if $\varphi(\gamma_t^{-1}\dot{\gamma}_t)=\|\gamma_t^{-1}\dot{\gamma}_t\|$ for (almost) all $t$. If $\gamma$ is short but its longer than $R$, we can still partition $\gamma$ in pieces of length smaller than $R$ and its still short on each piece, in any of these intervals $[a_i,a_{i+1}]$ $(i=1,\dots, k)$. There we have $\operatorname{dist}(\gamma(a_i),\gamma(t))\le L(\gamma)_{a_i}^t<R$. Write $\gamma_{a_{i+1}}=\gamma_{a_i}e^{z_i}$, thus we can lift $\gamma_t=\gamma_{a_i} e^{\Gamma^i_t}$ for a rectifiable $\Gamma^i:[a_i,a_{i+1}]\to\k$, which does not leave the ball $B_R$. Using the translation invariance of the metric (from $\gamma_{a_i}$ to $1$), and repeating the argument above, we have for each $i$ $$\label{eachi}
\varphi_i(\gamma_t^{-1}\dot{\gamma}_t)=\|\gamma_t^{-1}\dot{\gamma}_t\|\qquad \textrm{ if }\, \varphi_i \; \textrm{ norms }z_i \,\textrm{ and }\; \|\varphi_i\|=1,$$ for (almost) all $t\in [a_i,a_{i+1}]$. Now let $\beta(s)=e^{sz_1}e^{sz_2}\dots e^{sz_k}$, then $\beta(0)=1=\gamma(0)$, $\beta(1)=\gamma(1)$ and $$\operatorname{dist}(1,\gamma_1)\le L(\beta)\le \sum_i \|z_i\|=\sum_i \operatorname{dist}(\gamma_i,\gamma_{i+1})=\sum_i L(\gamma|_{[a_i,b_i]}) = L(\gamma)_a^b=\operatorname{dist}(1,\gamma_1),$$ therefore $\beta$ is short also. Since $\beta$ is smooth, there exists $c\in (0,1)$ such that $$\sum_i\|z_i\|=\operatorname{dist}(1,\gamma_1)=L(\beta)=\int_0^1 \|\beta_s^{-1}\dot{\beta}_s\|ds= \|\beta_c^{-1}\dot{\beta}_c\|.$$ Now let $w_0=z_k$, $w_1=e^{-c\operatorname{ad}z_k}z_{k-1}$, $w_2=e^{-c\operatorname{ad}z_k}e^{-c\operatorname{ad}z_{k-1}}z_{k-2}$ and in general $$w_j=e^{-c\operatorname{ad}z_k}e^{-c\operatorname{ad}z_{k-1}}\dots e^{-c\operatorname{ad}z_{k-j+1}} z_{k-j}$$ for $j=0,\dots, k-1$. Note that $\|w_j\|=\|z_{k-j}\|$ for each $j$. A straitghtforward computation shows that $$\beta_c^{-1}\dot{\beta}_c=\sum_{j=0}^{k-1} w_j,$$ therefore $$\sum_j \|w_j\|=\sum_i \|z_i\|=L(\beta)=\|\beta_c^{-1}\dot{\beta}_c\|=\|\sum_j w_j\|,$$ and by Lemmma \[mismacara\], there exists a unit norm functional such that $\varphi(w_j)=\|w_j\|$ for all $j=0,\dots, k-1$. In particular $\varphi(z_k)=\varphi(w_0)=\|w_0\|=\|z_k\|$, thus also by Gauss’ Lemma (Remark \[gausslem\]) we have $$\varphi(z_{k-1})=\varphi(e^{-c \operatorname{ad}z_k}z_{k-1})=\varphi(w_1)=\|w_1\|=\|z_{k-1}\|.$$ Proceedings backwards in this fashion, we can conclude that $\varphi(z_i)=\|z_i\|$ for all $i=1,\dots, k$. Thus by (\[eachi\]) we have that (\[cuasiaut\]) holds for this $\varphi$, for (almost) all $t\in [a,b]$.
Now assume that (\[cuasiaut\]) holds for some $\varphi$, described as $\varphi(v)=\langle v,a\rangle=-\operatorname{tr}(\operatorname{ad}v\circ \operatorname{ad}a)$ for some $a\in \k$ (Remark \[normicon\]), and let $t_0\le t\le t_1$ then $\operatorname{dist}(\gamma_{t_0},\gamma_t)\le L(\gamma)_{t_0}^t\le L(\gamma)_{t_0}^{t_1}=R$. Using the invariance of the metric, we can assumme that $\gamma_{t_0}=1$. Then there exists a rectifiable lift $\Gamma$ of $\gamma$ such that $\Gamma\subseteq B_R$, $\Gamma_{t_0}=0$, $e^{\Gamma_{t_1}}=\gamma_{t_1}$. Note that $\|\Gamma_{t_1}\|=R=\operatorname{dist}(1,\gamma_{t_1})$. By Remark \[normicon\], since $\varphi$ norms $\gamma_t^{-1}\dot{\gamma}_t$, we have that $\gamma_t^{-1}\dot{\gamma}_t$ and $a$ commute for all $t$. By formula (\[DAD\]) of Remark \[remark\] $$\frac{d}{dt} \operatorname{Ad}_{\gamma_t}a=\operatorname{Ad}_{\gamma_t}[\gamma_t^{-1}\dot{\gamma}_t,a]=0,$$ which shows that $\operatorname{Ad}_{\gamma_t}a=a$ for all $t$ (since $\gamma_{t_0}=1$). Then by the formula (\[dexp\]) for the differential of the exponential map of $K$, $$\begin{aligned}
e^{-\Gamma_t}D\exp_{\Gamma_t}[\Gamma_t,a] & = (\int_0^1 e^{-\lambda \operatorname{ad}\Gamma_t}\operatorname{ad}\Gamma_t \,d\lambda\, )(a)=-e^{-\lambda \operatorname{ad}\Gamma_t}\Bigr|_0^1(a)\\
&=-e^{-\operatorname{ad}\Gamma_t}a+a=-\operatorname{Ad}_{\gamma_t^{-1}}a+a=0.\end{aligned}$$ Since $\Gamma_t$ is inside the injectivity radius of the exponential map, it follows that $[\Gamma_t,a]=0$ for all $t\in [t_0,t_1]$. Then for fixed $t$, the operators $\operatorname{ad}\Gamma_t$ and $\operatorname{ad}a$ commute, and since they are both skew-adjoint operators acting on $\k$, they can be simultaneously diagonalized in an orthonormal basis, say $\{e_1,\dots,e_k\}$ of $\k$. Let $p_i=e_i\otimes e_i$ denote the rank-one orthogonal projection with range $e_i$, then $\operatorname{ad}a=\sum_{i=1}^k a_i p_i$ and $\operatorname{ad}\Gamma_t=\sum_i \lambda_i p_i$. Note that $\operatorname{Ad}_{\gamma_t}=e^{\operatorname{ad}\Gamma_t}=\sum_i e^{\lambda_i}p_i$. Now we apply the formula (\[dexp\]) for the differential of the exponential map to the path $t\mapsto e^{\operatorname{ad}\Gamma_t}$, and noting that $\operatorname{ad}\dot{\Gamma}_t=(\operatorname{ad}\Gamma_t)^{\, \cdot}$ we obtain $$e^{-\operatorname{ad}\Gamma_t} D\exp_{\operatorname{ad}\Gamma_t}\operatorname{ad}\dot{\Gamma}_t =\int_0^1 e^{-s\operatorname{ad}(\operatorname{ad}\Gamma_t)}\operatorname{ad}\dot{\Gamma}_t ds = \int_0^1 e^{-s\operatorname{ad}\Gamma_t}\operatorname{ad}\dot{\Gamma}_t e^{s\operatorname{ad}\Gamma_t}ds.$$ Due to (\[DAD\]) we also have $$\label{aii}
\operatorname{ad}\gamma_t^{-1}\dot{\gamma}_t=\operatorname{Ad}_{\gamma_t^{-1}}\frac{d}{dt}\operatorname{Ad}_{\gamma_t}=e^{-\operatorname{ad}\Gamma_t}D\exp_{\operatorname{ad}\Gamma_t}(\operatorname{ad}\dot{\Gamma_t}).$$ Then since $\operatorname{ad}a$ and $e^{\operatorname{ad}\Gamma_t}$ are diagonal in the orthonormal basis $\{e_i\}_{i=1,\dots,k}$ (for this particular $t$), we have $$\begin{aligned}
p_i ( e^{-s\operatorname{ad}\Gamma_t}\operatorname{ad}\dot{\Gamma}_t e^{s\operatorname{ad}\Gamma_t}\circ \operatorname{ad}a) p_i & = e^{-s\lambda_i} p_i \operatorname{ad}\dot{\Gamma}_t \,e^{s\lambda_i}\,a_i\, p_i= p_i \operatorname{ad}\dot{\Gamma}_t\, a_i\, p_i \\
& = a_i(\operatorname{ad}\dot\Gamma_t)_{ii} \,p_i=p_i(\operatorname{ad}a\circ \operatorname{ad}\dot{\Gamma}_t)p_i.\end{aligned}$$ Integrating $s$ in $[0,1]$ it follows that $$(\operatorname{ad}a\circ e^{-\operatorname{ad}\Gamma_t} D\exp_{\operatorname{ad}\Gamma_t}\operatorname{ad}\dot{\Gamma}_t)_{ii}=(\operatorname{ad}a\circ \operatorname{ad}\dot{\Gamma}_t)_{ii}$$ and using (\[aii\]) shows that $$(\operatorname{ad}a\circ (\operatorname{ad}\gamma_t^{-1}\dot{\gamma}_t))_{ii}= (\operatorname{ad}a\circ \operatorname{ad}\dot{\Gamma}_t)_{ii}$$ for all $i$. Thus $$\begin{aligned}
\varphi(\gamma_t^{-1}\dot{\gamma}_t)&=\langle a,\gamma_t^{-1}\dot{\gamma_t}\rangle=-\operatorname{tr}(\operatorname{ad}\gamma_t^{-1}\dot{\gamma}_t\circ\operatorname{ad}a)=-\sum_i (\operatorname{ad}\gamma_t^{-1}\dot{\gamma}_t\circ\operatorname{ad}a)_{ii}\\
&= -\sum_i (\operatorname{ad}a\circ\operatorname{ad}\dot{\Gamma}_t)_{ii}=-\operatorname{tr}(\operatorname{ad}a\circ\operatorname{ad}\dot{\Gamma}_t) = \langle\dot{\Gamma_t},a\rangle=\varphi(\dot{\Gamma}_t).\end{aligned}$$ Since this holds for any $t\in [t_0,t_1]$, we have $\varphi(\dot{\Gamma}_t)=\|\gamma_t^{-1}\dot{\gamma}_t \|$ there, and then $$\operatorname{dist}(\gamma_{t_0},\gamma_{t_1})=\|\Gamma_{t_1}\|\ge \varphi(\Gamma_{t_1})=\varphi(\int_{t_0}^{t_1} \dot{\Gamma}_t \, dt)=\int_{t_0}^{t_1} \varphi(\dot{\Gamma}_t)\, dt= \int_{t_0}^{t_1} \|\gamma_t^{-1}\dot{\gamma}_t \|dt=L_{t_0}^{t_1}(\gamma)$$ which proves that $\gamma$ is short in that interval.
It is clear from the proof of the previous theorem, that
- If $\gamma=e^{\Gamma}$ is short, then there exists $a\in \k$ such that $\operatorname{ad}\Gamma_t$ (and then also $\operatorname{ad}\dot{\Gamma}_t$, $\operatorname{Ad}_{\gamma_t}$, $\operatorname{ad}\gamma_t^{-1}\dot{\gamma}_t$) commute with $\operatorname{ad}a$ for all $t$. From the faithfulness of the adjoint representation, this is equivalent to $\Gamma_t,\dot{\Gamma}_t,\gamma_t,\gamma_t^{-1}\dot{\gamma}_t$ commuting with $a$ for each $t$. For linear Lie groups it then follows that $a$ and $\Gamma_t$ can be simultaneously diagonalized for each $t$, and this basis also diagonalizes $\gamma_t=e^{\Gamma_t}$. However, a word of caution: the orthonormal basis depends on $t$.
- Once $\gamma$ is short in a certain interval $[a,b]$, and assuming that we first write $\gamma_b=\gamma_a e^{z_1}e^{z_2}\dots e^{z_k}$, with $\gamma(t_{i+1})=\gamma(t_i)e^{z_i}$, and the $z_i$ giving the distance among the endpoints and $\|z_i\|\le R$, then the equality (\[cuasiaut\]) if full-filled *for any* norming functional of $z_1+\dots+z_n$. If $\gamma_b=\gamma_ae^z$ with $z\in \mathbf D$, when can we ensure that $z$ and the $z_i$ are in the same face of the sphere? Equivalently, $z$ and $\gamma^{-1}\dot{\gamma}$ are in the same face of the sphere? If $\|z\|<R$ this follows from Theorem \[charactgeod\], see also the next remark.
- If $\gamma_0=1,\gamma_1=e^z$ with $z\in \mathbf D$, and $\gamma$ is short for a given norm, does it follow that $\gamma=e^{\Gamma}$ with $\Gamma\subset \mathbf D$? The argument used in repeated occasions is that if $\|z\|<R$ and $\gamma$ is short for that norm, then there is a lift $\Gamma\subset B_R$; but this radius depends on the norm.
By Theorem \[charactgeod\], we know that if $\Gamma$ is short in $\k$ for a given norm, then its exponential $\gamma=e^{\Gamma}$ is short in $K$ for the bi-invariant metric of that norm. Is there any concrete example of Lie group $K$ with $\operatorname{Ad}$-invariant Finsler norm such that: there exists a path $\gamma=e^{\Gamma}:[a,b]\to K$, with $\gamma$ short in $K$ joining $1,e^z$ but $\Gamma$ (which joins $0,z$ in $\k$) not short in $\k$? For strictly convex norms this is not possible since the only short paths are segments; as we will see in the next section it is also impossible if all the faces of the unit sphere are abelian (Corollary \[corabelian\]). Another relevant example, studied by Antezana, Ghighlioni and Stojanoff in [@ags] is the full unitary group (or in our setting, $K=\operatorname{SU}_n$) with the spectral norm: examining [@ags Theorem 2.1] it is apparent that lifted short paths are also short there.
Norms with abelian faces
------------------------
In Theorem \[existen\] we showed that when the faces of the unit ball are singletons, the short paths are one-parameter groups. We will show here that the situation is somewhat similar if we allow the faces of the unit ball to be inside *abelian* subalgebras of $\k$. In Section \[geodesicommute\] we will give a full characterization of those norms with this important structural property.
\[abelianfaces1\] Let $B$ the unit ball of the bi-invariant norm in $\k$. Then all the faces of $B$ are abelian if and only if for all piecewise $C^1$ short curves $\gamma\subseteq K$, the logarithmic derivatives $x_t=\gamma_t^{-1}\dot{\gamma}_t$ of $\gamma$ commute for all $t$.
Assume first that each face of the unit ball is abelian. By Theorem \[cuasi\], if $\gamma$ is short, it has logarithmic derivative (after normalizing) inside a face of the ball, and these commute for all $t$.
Now assume that for some $w\in\k$ the face $F_w(B)$ is non-abelian. Let $x,y\in F_w(B)$ with $[x,y]\ne 0$. Let $R>0$ be the radius of injectivity and for $0<b<R$ let $\gamma:[0,b]\to K$ be the path which solves $$\gamma_t^{-1}\dot{\gamma}_t=tx+(b-t)y$$ for $t\in[0,b]$. Then by Theorem \[cuasi\] $\gamma$ is a geodesics, furthermore the logarithmic derivatives $\gamma_0^{-1}\dot{\gamma}_0=x$ and $\gamma_b^{-1}\dot{\gamma}_b=y$ do not commute.
\[corabelian\] Let $B$ be the unit ball of the bi-invariant norm in $\k$ and assume that the faces of $B$ are all abelian. Let $\gamma:[a,b]\to K$ be a short piecewise $C^1$ path. Then
- There exists $z\in \k$ such that $\delta_t=\gamma_a e^{tz}$ is also short with the same endpoints.
- If $L(\gamma)\le R$ (thus $\gamma_b=\gamma_ae^z$ for some $\|z\|\le R$), then $\gamma=\gamma_a e^{\Gamma}$ where $\Gamma:[a,b]\to \k$ and $[\Gamma_t,\Gamma_s]=0$ for all $s,t\in [a,b]$, thus $\gamma_t^{-1}\dot{\gamma}_t=\dot{\Gamma}_t$. In particular the logarithmic derivatives of $\gamma$ commute and $\Gamma$ is short in $\k$, with the same length than $\gamma$.
As always we can assume that $\gamma_a=1$, partition $\gamma$ in small pieces such that $\gamma(t_{i+1})=\gamma(t_i)e^{z_i}$ with $\|z_i\|\le R$ as in the proof of Theorem \[cuasi\]. Let $\beta(s)=e^{sz_1}e^{sz_2}$, then $\beta$ joins $1,e^{z_1}e^{z_2}=\gamma(t_2)$ and $L(\beta)=\|z_1+z_2\|$, thus $$\operatorname{dist}(1,\gamma_{t_2})\le L(\beta)\le\|z_1\|+\|z_2\|=\operatorname{dist}(1,\gamma_{t_1})+\operatorname{dist}(\gamma_{t_1},\gamma_{t_2})=\operatorname{dist}(1,\gamma_{t_2}),$$ which shows that $\beta$ is short among its endpoints. For small $s$, let $\beta_s=e^{B_s}$ with $B_s=s(z_1+z_2)+\frac{s^2}{2}[z_1,z_2]+o(s^3)$, then by Theorem \[charactgeod\](1), $\|B_s\|=\operatorname{dist}(1,\beta_s)=L_0^s(\beta)=s\|z_1+z_2\|$. On the other hand, by Theorem \[charactgeod\](4), we also have $\varphi(\dot{B}_s)=\|\beta_s^{-1}\dot{\beta}_s\|=\|z_1+z_2\|$, then integrating we have $\varphi(B_s)=s\|z_1+z_2\|=\|B_s\|$, and this shows that $B_s$ (normalized) is inside a face of the ball, which implies that the $B_s$ commute for all (small) $s$ (therefore they also commute with $\dot{B}_s$). Then using the formula (\[dexp\]) we have $$e^{-s\operatorname{ad}z_2}(z_1+z_2)=\beta_s^{-1}\dot{\beta}_s=e^{-B_s}D\exp_{B_s}\dot{B}_s=\dot{B_s}=z_1+z_2+s[z_1,z_2]+o(s^2).$$ Differentiating at $s=0$ shows that $[z_1,z_2]=0$. Thus $\gamma(t_2)=e^{z_1}e^{z_2}=e^{z_1+z_2}$ and $\gamma(t_3)=e^{z_1}e^{z_2}e^{z_3}=e^{z_1+z_2}e^{z_3}$. We repeat the argument now using $\beta_s=e^{s(z_1+z_2)}e^{sz_3}$, this shows that $[z_3,z_1+z_2]=0$. Thus $\gamma_b=e^{z_1}e^{z_2}e^{z_3}\dots e^{z_k}=e^{\sum z_i}$, and if we let $z=\sum z_i$, then $\|z\|=\sum \|z_i\|=\operatorname{dist}(\gamma_a,\gamma_b)$ and the first claim follows.
For the second assertion let $\Gamma$ be the smooth lift of $\gamma$, $\Gamma_a=0$, then by Theorem \[charactgeod\](4) we have $\varphi(\dot{\Gamma}_t)=\|\gamma_t^{-1}\dot{\gamma}_t\|$ for some unit norm $\varphi$, and again integrating $\varphi(\Gamma_t)=L_0^t(\gamma)=\operatorname{dist}(1,e^{\Gamma_t})=\|\Gamma_t\|$ shows that $\Gamma_t$ (normalized) is inside a face of the sphere, which by hypothesis is abelian. But then $\Gamma_t$ and $\dot{\Gamma}_t$ also commute for all $t$ and it is apparent that $$L(\Gamma)=\int_a^b \|\dot{\Gamma}\|dt=\int_a^b\|e^{-\Gamma}D\exp_{\Gamma}\dot{\Gamma}\|dt=\int_a^b\|\gamma^{-1}\dot{\gamma}\|dt=L(\gamma)$$ thus if $\Gamma_b=z$ then $\|z\|\le L(\Gamma)=L(\gamma)=\operatorname{dist}(1,e^z)=\|z\|$ and $\Gamma$ is short.
Geodesic are quasi-autonomous {#sectionquasiaut}
-----------------------------
Throughout, the action of the semi-simple compact group $K$ on the symplectic manifold $M$ complies the hypothesis of Section \[hamactions\]. Recall (Definition \[quasiauto\]) that a Hamiltonian $H_t$ is called *quasi-autonomous* if there exists $x^-,x^+ \in M$ such that $H_t(x^-) = \min_M H_t$, $H_t(x^+) = \max_M H_t$ for all $t\in[a,b]$. As before, we use $R$ to indicate the injectivity radius of the exponential map of $K$, for the given norm. In the next theorem, the geometry of the group $K$ is the one given by the generalized Hofer norm (\[hofer1\]) obtained by the almost effective action.
Let $K\curvearrowright (M,\omega)$ be a Hamiltonian almost effective action. Let $\gamma:[a,b]\to K$ be piecewise $C^1$, and denote its right logarithmic derivative by $x_t=\dot{\gamma}_t\gamma^{-1}_t$. Then if $\gamma$ is short, $(\mu^{x_t})_{t\in [0,1]}$ is a quasi-autonomous Hamiltonian, and if $\mu$ is quasi-autonomous, $\gamma$ is locally short (in each interval of length $\le R$).
By Theorem \[cuasi\], $\gamma$ is locally short if and only if there is a common norming functional for $x_t=\dot{\gamma}_t\gamma^{-1}_t$, for all $t$. By Corollary \[arghofer\] the set of logarithmic derivatives $\{x_t\}_{t\in [0,1]}\subseteq \k$ is contained in a cone generated by a face if and only if there exist $x^-$ and $x^+$ such that $x^-\in\cap_{t\in [0,1]}\operatorname{argmin}(\varphi_{x_t})$ and $x^+\in\cap_{t\in [0,1]}\operatorname{argmax}(\varphi_{x_t})$, where $\varphi_{x_t}:\mu(M)\to\R$. We can choose $m^-,m^+\in M$ such that $x^-=\mu(m^-)$ and $x^+=\mu(m^+)$. The result follows since $\mu^{x_t}=\varphi_{x_t}\circ\mu$ for all $t\in[0,1]$.
A similar characterization of geodesics (in fact, of their logarithmic derivative $\gamma_t^{-1}\dot{\gamma}_t=x_t$) can be obtained for the one-sided norm induced by the action of $K$ in $M$, if one replaces the condition of quasi-autonomous for the Hamiltonian $H_t=\mu^{x_t}$, with the condition that there exists a point $x^+\in M$ such that $H_t(x^+)=\max_M H_t$ for all $t\in [a,b]$.
Spheres with abelian faces {#geodesicommute}
==========================
In this section we characterize those $\operatorname{Ad}$-invariant norms which have unit balls with abelian faces in terms of conditions on its intersection with a Cartan subalgebra. When this intersection is a polytope the condition reduces to the regularity of the extreme points of its polar dual. Based on this result and Kirwan’s Theorem \[kirwan\] we characterize the compact semi-simple groups of Hamiltonian diffeomorphims such that geodesics have commuting Hamiltonians. We start with the important special case of Hofer norms derived from coadjoint actions on regular coadjoint orbits. Groups with length structures derived from these norms have the property that the logarithmic speed of its geodesics lie in a Weyl chamber.
Regular coadjoint actions and non-crossing of eigenvalues {#subsectionspeedweyl}
---------------------------------------------------------
In this section we follow the setting of Example \[coadj\] on (co)adjoint orbits in $\O\subseteq \k\simeq \k^*$.
The (co)adjoint orbit $\O=\O_{\lambda}\subseteq \k$ of the action is *regular* if $\lambda\in\k$ is a regular element (Remark \[regular\]).
In order to characterize the faces of the unit ball in $\k$ via Corollary \[arghofer\] we need to study, for a nonzero $x$ in $\k$, the sets $\operatorname{argmax}_{\O}(\varphi_x)$ and $\operatorname{argmin}_{\O}(\varphi_x)$ of $\varphi_x=\mu^x:\O\to\R$. Let $m_t=\operatorname{Ad}_{e^{tv}}m$ be a path through $m\in\O$ with $\dot{m}_0=[v,m]\in T_m\O$, differentiating $\varphi_x(m_t)=\langle x,m_t\rangle$ at $t=0$ we obtain $$D(\varphi_x)_m(\dot{m}_0)=\langle x, [v,m]\rangle =-\langle x,[m,v]\rangle =\langle [m,x],v\rangle,$$ since $\operatorname{ad}m$ is skew-adjoint for the Killing form. Since $v\in \k$ is arbitrary, this implies that $m\in\O$ is a critical point of $\varphi_x$ if and only if $m\in \z(x)$, where $\z(x)$ is the centralizer of $x$, i.e. $$\label{critpoint}
Crit(\varphi_x)=\O\cap\z(x).$$
The proof of following proposition can be found in [@bgh Lemma 23].
\[maximumorbit\] Fix a maximal torus $T\subseteq K$, a nonzero vector $x\in \t=Lie(T)$ and a point $m\in \O\cap\t$; thus $m\in Crit(\varphi_x)$. Then $m$ is a maximum point of $\varphi_x$ if and only if there is a Weyl chamber in $\t$ whose closure contains both $x$ and $m$.
With these tools at hand, we next characterize the faces of the unit ball of $\|\cdot\|_{\O}$.
\[speedweylchamber\] Let $\O$ be a regular (co)adjoint orbit and let $\|\cdot\|_{\O}$ be the associated $\operatorname{Ad}$-invariant Hofer norm. A set of vectors has a common norming functional if and only if it is contained in a Weyl chamber (given by a choice of torus and positive simple roots). Hence, the maximal cones generated by faces of the unit ball are Weyl chambers.
By Corollary \[arghofer\] a set of elements $S\subseteq\k$ have the same norming functional if and only if the set $\{\varphi_u:u\in S\}$ has a common maximizer $x^+\in \O$ and a common minimizer $x^-\in\O$.
We denote by $\t_+$ the Weyl chamber that contains $x^+$, it is unique since $x^+$ is regular. If $u\in S$, then the functional $\varphi_u$ has a maximum at $x^+$, hence by equation (\[critpoint\]) $u$ commutes with $x$ and therefore $u\in\t$. By Proposition \[maximumorbit\] we conclude that $u\in\t_+$ if and only if the functional $\varphi_u$ has a maximum at $x^+$. Let us denote by $x'\in\k$ the element defined by $\{x'\}=\O\cap -\t_+$ , that is, the element of $\O$ in the opposite Weyl chamber. A functional $\varphi_u$ has a minimum at $x'$ if and only if $\varphi_{-u}$ has a maximum at $x'$. This holds if and only if $-u$ and $x'$ belong to the same Weyl chamber, which is equivalent to $u\in\t_+$. Hence, $\varphi_u$ has a maximum at $x^+$ and $\varphi_u$ has a minimum at $x'$, are both equivalent to $u\in\t_+$. If we take $x^-=x'$ as the common minimizer we get a maximal face which is $\t_+$.
If $\t_+$ is a Weyl chamber we denote by $x^+$ and $x^-$ the elements defined by $\{x^+\}=\O\cap \t_+$ and $\{x^-\}=\O\cap -\t_+$. Then we can reverse the argument in the previous paragraph to conclude that the functionals $\varphi_u$ with $u\in \t_+$ have $x^+\in \O$ as maximizer and $x^-\in\O$ as minimizer.
From this characterization of the faces of the unit ball in $\k$ we can determine the geodesics in $K$. Let $\gamma:[a,b]\to K$ be a curve in a group $K$ endowed with the length structure obtained from Hofer’s norm $\|\cdot\|_{\O}$ for a regular (co)adjoint orbit $\O$. As a combination of the previous proposition and Theorem \[cuasi\], we obtain
\[speedchamber\] If $\gamma$ is short then all its logarithmic derivatives are contained in the same Weyl chamber $\t_+$ (given by a choice of torus and positive simple roots). If its derivatives are contained in a Weyl chamber, then $\gamma$ is locally short (in each interval of $\operatorname{length}\le R$, where $R$ is the injectivity radius of the norm).
Consider the case of $\operatorname{SU}(n)$ acting on a regular (co)adjoint orbit $\O_{\lambda}$ containing $\lambda=i\operatorname{diag}(\lambda_1,\dots,\lambda_n)$, with $\lambda_1<\dots<\lambda_n$. Recall that here we obtain in $\k$ the $\lambda$-numerical radius as Hofer norm (Example \[sun\]). The local condition for the logarithmic derivatives to be the speeds of geodesics is that there exist a fixed orthonormal basis such that the speeds are simultaneously diagonal in this basis for all $t$ (Corollary \[corabelian\]) and if $x_i(t)$ are the eigenvalues of $x_t=\dot{\gamma}_t\gamma^{-1}_t$, then $$x_1(t)\leq\dots\leq x_n(t) \qquad \forall t\in [a,b].$$ This is because in $\operatorname{SU}(n)$ the condition of being in the same Weyl chamber is given by the non-crossing of the eigenvalues.
The Hofer norm associated to *singular* (co)adjoint orbits can be studied using the following result on maximizers and minimizers of linear functionals restricted to the orbits, see [@bgh Lemma 22]. Let $Z(x)$ be the centraliser of $x$ in $K$ and let $F_x(\O)$ be the face of $\O$ defined by $x$. Then
- $\operatorname{argmax}(\varphi_x)$ is a $Z(x)$-orbit,
- $\operatorname{ext}(F_x(\operatorname{conv}(\O))=\operatorname{argmax}(\varphi_x)$, so $\operatorname{ext}(F_x(\operatorname{conv}(\O))$ is a $Z(x)$-orbit.
- $F_x(\operatorname{conv}(\O))\subseteq \z(x)$.
Invariant norms with abelian faces {#ssabelianface}
----------------------------------
Let $K$ be a compact connected semi-simple Lie group and let $\langle\cdot,\cdot\rangle$ be as usual the opposite Killing form of $\k^\C$ restricted to $\k$. Let $\Delta$ be the set of (real) roots of $\k$ with respect to a fixed Cartan subalgebra $\t $, as in Remark \[rootdecom\]. For $x\in\t$ we have $$\begin{aligned}
\label{centralizer}
\z(x) & =\t\oplus\bigoplus_{\alpha\in\Delta_+,\alpha(x)=0}Z_\alpha,\end{aligned}$$ here $\z(x)$ is the Lie algebra of $Z(x)$ as usual (the centralizer of $x\in \k$). We define the *smallest Weyl chamber wall containing* $x$ as the linear space $$W_x=\bigcap_{\{\alpha\in\Delta_+ : \alpha(x)=0\}}\ker(\alpha).$$
\[abelianfaces\] Let $K$ be a semi-simple compact connected group and let $B$ be an $\operatorname{Ad}$-invariant convex body in $\k$ containing $0$. Let $\t$ be a Cartan algebra in $\k$ and $\t_+$ a positive Weyl chamber. Then all faces of $B$ are abelian if and only if for all $x\in\t^+$ we have $$F_x(B\cap\t)\cap\t_+\subseteq W_x,$$ and in this case $F_x(B)=F_x(B\cap\t)\cap\t_+.$
We assume for simplicity that $x\in\t^+$. Theorem \[lewisfaces\] states that $$F_x(B)=\operatorname{Ad}_{Z(x)}(F_x(B\cap\t)\cap\t_+).$$
Next we show that this set is abelian when the inclusion stated in the theorem holds. If $x$ is in the interior of the Weyl chamber then $Z(x)$ is trivial and $F_x(B)=F_x(B\cap\t)\cap\t_+$, hence the face is abelian. If $x$ is not regular, and hence $Z(x)$ is not trivial, then by (\[weights\]) and (\[centralizer\]) the centralizer $Z(x)$ acts trivially on $W_x$. Therefore, the inclusion $F_x(B\cap\t)\cap\t_+\subseteq W_x$ implies that $F_x(B)=\operatorname{Ad}_{Z(x)}(F_x(B\cap\t)\cap\t_+)=F_x(B\cap\t)\cap\t_+$, so the face is abelian. Note also that the last assertion of the theorem follows from this argument.
If the inclusion in the statement of the theorem is not satisfied there exists $y\in F_x(B\cap\t)\cap\t_+$ such that $y\notin W_x$; it follows that there exists $\alpha$ such that $\alpha(y)\neq 0$ and $\alpha(x)=0$. Consider the orbit $\operatorname{Ad}_{Z(x)}(y)\subseteq \k$ as a submanifold with its differentiable structure, and observe that $$[\z(x),y]\subseteq T_y(\operatorname{Ad}_{Z(x)}(y)).$$ By (\[centralizer\]), we have $v_\alpha,u_\alpha\in\z(x)$; the equations $[y,u_\alpha]=-\alpha(y)v_\alpha$ and $[y,v_\alpha]=\alpha(y)u_\alpha$ hold by (\[weights\]), therefore we conclude that $$\{u_\alpha,v_\alpha\}\subseteq T_y(\operatorname{Ad}_{Z(x)}(y)).$$ Since $[u_\alpha,v_\alpha]=h_\alpha\neq 0$ the tangent $T_y(\operatorname{Ad}_{Z(x)}(y))$ is not an abelian space, so that $\operatorname{Ad}_{Z(x)}(y)$ cannot be an abelian set. The inclusion $\operatorname{Ad}_{Z(x)}(y)\subseteq \operatorname{Ad}_{Z(x)}(F_x(B\cap\t)\cap\t_+)=F_x(B)$ implies that $F_x(B)$ is not an abelian set.
\[maximalabelianfaces\] From the previous theorem we know that if the faces of $B$ are abelian then $$F_x(B)=F_x(B\cap\t)\cap\t_+$$ for a Weyl chamber $\t_+$ given by a choice of torus and positive simple roots such that $x\in\t_+$. Therefore the balls of the norms studied in Section \[subsectionspeedweyl\] have maximal commuting cones generated by faces.
Polytopes with regular extreme points {#ssabelianfacepolytope}
-------------------------------------
The aim of this section is to prove the following theorem. Note that an element $x\in \k$ is regular if and only if $\alpha(x)\neq 0$ for all $\alpha\in\Delta_+$.
\[abelian\] Let $K$ be a semi-simple compact connected group, let $B$ be an $\operatorname{Ad}$-invariant convex body in $\k$ containing $0$, such that $B\cap\t=P^\circ$ is a polytope. Then all faces of $B$ are abelian if and only if all extreme points of $P$ are regular.
For the proof of this theorem we need a couple of preliminary lemmas on polytopes invariant under finite reflection groups. These lemmas will be later applied to the Hofer norm polytopes defined in Definition \[hoferpolytopes\], which are a polytopes invariant under the Weyl group.
\[frl\] Let $\W$ be a finite reflection group acting isometrically on an inner product vector space $(V,\langle \cdot,\cdot\rangle)$ and generated by reflections $\{r_u\}_{u\in\Phi}$. Here $r_u$ is the refection which fixes the mirror hyperplane $M_u=\{u\}^\perp$ and $\Phi$ is a finite subset of the unit sphere that is called the positive root system. For a finite reflection group the space can be subdivided into chambers bounded by mirror hyperplanes. We can choose any (closed) chamber $C$ and call it the fundamental chamber. Given a positive root system $\Phi$, a fundamental chamber is given by $$C=\{x\in V:\langle x,u\rangle\geq 0 \mbox{ for all }u\in\Phi\}.$$ Let $P$ be a polytope with extreme points $\operatorname{ext}(P)=\{y_1,\dots,y_m\}$ and assume that $P$ is invariant under $\W$.
\[lemregext\] Let $\Pi\subseteq\Phi$ and assume that $\cup_{u\in\Pi}M_u\cap\{y_1,\dots,y_m\}=\emptyset$. If $H_x$ supports $P$ at $y$ and $y\in\cap_{u\in\Pi}M_u$ it follows that $x\in\cap_{u\in\Pi}M_u$.
Assume $P$ is invariant under a reflection $r_u$ and the mirror hyperplane does not contain any extreme point of $P$, i.e. $M_u\cap\{y_1,\dots,y_m\}=\emptyset$, see Figure \[fig: refl\]. Then, if $H_x$ is a hyperplane that supports $P$ at an $y\in M_u$ we claim that $x\in M_u$. To prove this we write $y$ as a convex combination of the $y_1,\dots,y_m$, i.e. $y=\lambda_1y_1+\dots\lambda_ny_n$, with $n\le m$, $\lambda_1,\dots,\lambda_n>0$, and $\sum_{i=1}^n\lambda_i=1$. If we write $$y=\frac{1}{2}y+ \frac{1}{2}r_u(y)=\frac{1}{2}(\lambda_1y_1+\dots\lambda_ny_n)+\frac{1}{2}r_u(\lambda_1y_1+\dots\lambda_ny_n),$$ we see that we can assume that $y=\lambda_1y_1+\dots\lambda_ny_n$, with $\lambda_1,\dots,\lambda_n>0$, $\sum_{i=1}^n\lambda_i=1$ and $r_u(y_j)\in\{y_1,\dots,y_n\}$ for $i=1,\dots,n$. Since $H_x$ supports $P$ at $y\in M_u$ we get $\varphi_x(y)=1$ and $\varphi_x(y_1),\dots\varphi_x(y_m)\leq 1$. From $y=\lambda_1y_1+\dots+\lambda_ny_n$, with $\lambda_1,\dots,\lambda_n>0$, $\sum_{i=1}^n\lambda_i=1$, it follows that $\varphi_x(y_1),\dots, \varphi_x(y_n)= 1$. Since $y_1\notin M_u$ we know that $r_u(y_1)\in\{y_2,\dots,y_n\}$ so that $y_1-r_u(y_1)$ is non zero. Since $y_1-r_u(y_1)$ is orthogonal to the mirror $M_u$ and $\varphi_x(y_1-r_u(y_1))=1-1=0$ we conclude that $x$ is orthogonal to $y_1-r_u(y_1)$ and is therefore contained in $M_u$. The Lemma follows by applying this result to each of the mirrors $\{M_u\}_{u\in\Pi}$.
\[lemnonregext\] Assume that there is an extreme point $y_1$ of $P$ such that $y_1\in C$ and $y_1$ is contained in a mirror. Then there is a supporting hyperplane $H_{x'}$ of $P$ at $y_1$ such that $x'$ is in the interior of the fundamental chamber.
By assumption $\Pi:=\{u\in\Phi:y_1\in M_u\}\neq\emptyset$. The set of reflections $\{r_u:u\in\Pi\}$ generates the stabilizer $\operatorname{Stab}(y_1)$ of $y_1$ (see e.g. [@bor Theorem 12.6]). Since $P$ is a polytope there exists $x$ which determines a supporting hyperplane of $P$ at $y_1$ such that $\varphi_x(y_1)=1$ and $\varphi_x(y_j)<1$ for $j=2,\dots,m$. If we denote by $x'$ the average $$x'=\frac{1}{\vert \operatorname{Stab}(y_1) \vert}\sum_{w\in \operatorname{Stab}(y_1)} w.x,$$ then $\varphi_{x'}(y_1)=1$, $\varphi_{x'}(y_j)<1$ for $j=2,\dots,m$, and $x'\in\cap_{u\in\Pi}M_u$. To show that $x'\in C$ we need to verify that $\langle x',u\rangle\geq 0 $ for all $u\in\Phi$. If $u$ is not in $\Pi$ then $y_1$ is not in $M_u$, therefore $\varphi_{x'}(y_1)=1$ and $\varphi_{x'}(r_u(y_1))<1$ which implies that $\varphi_{x'}(y_1-r_u(y_1))>0$. On the other hand, since $y_1$ is in the positive chamber $C$ it follows that $y_1-r_u(y_1)=d.u$ for $d>0$. Hence $\varphi_{x'}(d.u)>0$ which is equivalent to $\langle x',u\rangle>0$.
The $x'$ can be perturbed so that $H_{x'}$ is a supporting hyperplane of $P$ at $y_1$ and $x'$ is in the interior of the fundamental chamber.
*(of Theorem \[abelian\])*. We apply the previous lemmas to the Weyl group action on $\t$. It is easy to check that the condition of Theorem \[abelianfaces\] can be stated using the more general notation of the previous lemmas as follows: $$F_y(P^\circ)\cap C\subseteq W_y:=\bigcap_{u\in\Phi:y\in M_u}M_u$$ for all nonzero $y$ in $C=\t_+$. Here, for example, $M_{h_{\alpha}}=\ker(\alpha)$, with $h_\alpha$ as in (\[weights\]). We can multiply $y$ by a positive scalar and assume that it is in $\operatorname{bd}P$. Then by Theorem \[supportdual\] $$\begin{aligned}
F_y(P^\circ)&=\{x\in V:H_y\mbox{ supports } P^\circ\mbox{ at }x\}\\
&=\{x\in V:H_x\mbox{ supports } P\mbox{ at }y\}.\end{aligned}$$ By Lemma \[lemregext\], if the extreme points of $P$ are all regular and $y\in W_y=\bigcap_{u\in\Phi:y\in M_u}M_u$, then $x\in W_y$ holds when $H_x$ supports $P$ at $y$, i.e. $x$ is in $F_y(P^\circ)$.
If there is a non-regular extreme point of $P$ then by invariance of $P$ under the reflection group $\W$ we can choose a non-regular extreme point $y_1$ of $P$ in $C$. Lemma \[lemnonregext\] implies that there is a supporting hyperplane $H_{x'}$ of $P$ at $y_1$ such that $x'$ is in the interior of $C$. Then $x'\in F_y(P^\circ)\cap C$ and $x'$ does not belong to $W_y$, so that the condition of Theorem \[abelianfaces\] is not satisfied and there is a face of $B$ which is not abelian.
Let $K$ be a semi-simple compact connected group, let $B$ be an $\operatorname{Ad}$-invariant convex body in $\k$ containing $0$, such that $B\cap\t=P^\circ$ is a polytope. Let $g_B$ be the associated $\operatorname{Ad}$-invariant norm and endow $K$ with the corresponding Finsler length structure.
\[geodpolytope\] The extreme points of $P$ are regular if and only if all short curves $\gamma$ in $K$ have commuting logarithmic derivatives.
By Theorem \[abelian\] the extreme points of $P$ are regular if and only if $B$ has abelian faces, and by Proposition \[abelianfaces1\] this holds if and only if all short curves have commuting logarithmic derivatives.
We now specialize Theorem \[geodpolytope\] to compact semi-simple groups of Hamiltonian diffeomorphisms.
\[polytopecommuting\] Let $K\curvearrowright M$ be an almost effective Hamiltonian action with moment map $\mu:M\to\k^*\simeq\k$ and endow $K$ with the pullback metric of Section \[subsectionhoferham\]. Let $\mu(M)\cap\t^+=\operatorname{conv}\{x_1,\dots,x_n\}$ be Kirwan’s polytope given by Theorem \[kirwan\], and let $$P=\operatorname{conv}\{w.x-w'.x':x,x'\in \{x_1,\dots,x_n\},\quad w,w'\in\W\}$$ be the Hofer norm polytope derived from it. Then all short curves in $K$ have commuting Hamiltonians if and only if all the extreme points of $P$ are regular.
By Proposition \[actiondiffeo\] a curve $\gamma$ has Hamiltonians $\mu^{x_t}$, where $x_t=\dot{\gamma}_t\gamma_t^{-1}$ is the right logarithmic derivative. Recalling that $\{\mu^{x_s},\mu^{x_t}\}=\mu^{[x_s,x_t]}$ the theorem follows.
Note that the same result Holds if we endow $K\to\operatorname{Ham}(M,\omega)$ with the second Hofer norm by taking the second Hofer norm polytope $$P'=\operatorname{conv}(\{w.x:x\in \{x_1,\dots,x_n,-x_1,\dots,-x_n\},\quad w\in\W\})$$ of Definition \[hoferpolytopes\], and also if we consider the one-sided Finlser Hofer norm with its polytope $$P^+=\operatorname{conv}\{w.x:x\in \{x_1,\dots,x_n\},\,w\in\W\}.$$
An example of a group $K$ with non-commuting Hamiltonians is $\operatorname{SU}(4)$ acting on the singular (co)adjoint orbit containing $x=i\operatorname{diag}(3,-1,-1,-1)$. The Hofer norm polytope is the convex hull of the permutations of the matrix $i\operatorname{diag}(4,0,0,-4)$. Its extreme points are the permutations of the matrix given by $i\operatorname{diag}(4,0,0,-4)$, which are all singular. We can give a short informal explanation for this which is similar to the proof of Proposition \[speedweylchamber\]. If the maximum of $\varphi_x$ is at $x^+=i\operatorname{diag}(3,-1,-1,-1)$ and the minimum is at $x^-=i\operatorname{diag}(-1,-1,-1,3)$ then $x^+$ block diagonalizes $x$ and $x^-$ block diagonalizes $x$, so that $$x=i(\lambda_1P_{\C e_1}\oplus A\oplus \lambda_4P_{\C e_4}),$$ with $A$ a selfadjoint operator on $\C e_2\oplus\C e_3$ such that its spectrum $spec(A)$ satisfies $\lambda_1\geq spec(A)\geq \lambda_4$.
Properties determined by Kirwan’s polytope and product actions
--------------------------------------------------------------
In Section \[ssabelianfacepolytope\] we characterized Hofer norms with abelian faces using the regularity of the norm polytope (Theorem \[abelian\]), and before that in Section \[subsectionspeedweyl\] we studied the case of faces generating maximal abelian cones. In this section we show how conditions on Kirwan’s polytope can characterize these properties, and then to finish the paper we study how these properties behave if we take products of Hamiltonian actions.
We are here in the context of finite reflection groups of Section \[ssabelianfacepolytope\], in particular see Definition \[frl\]. We recall from [@eaton Section 4] or [@mozz Lemma 2.9] the following generalization of the rearrangement inequality $$\sum_{i=1}^nx_{\pi(i)}y_i\leq\sum_{i=1}^nx_iy_i$$ for $x_1\leq\dots\leq x_n$, $y_1\leq,\dots\leq y_n$ and certain permutations $\pi$.
\[rearrengementineq\] Suppose $x,y\in V$, then $$\sup_{w\in\W}\langle x,w.y\rangle=\langle x,y\rangle$$ if and only if $x$ and $y$ belong to the same Weyl chamber. In this case $$\langle x,w.y\rangle=\langle x,y\rangle$$ if and only if there exists $w'\in\operatorname{Stab}(x)$ such that $w.y=w'.y$. That is, the set of maximizers of $\varphi_x$ in $\W.y$ is exactly $\operatorname{Stab}(x).y$.
Let $E\subseteq V$ be a convex set with $x\in E$. The normal cone to $E$ at $x$ is defined as $$N(x,E):=\{y\in V:\langle y,z-x\rangle\leq 0 \mbox{ for all }z\in E \}.$$ It is a closed convex cone.
We will also need the following simple fact about a polytope and its normal cones.
\[normalcones\] Let $P\subseteq V$ be a convex polytope. If $A\subseteq\operatorname{ext}(P)$ and $$\bigcup_{y\in A}N(y,P)=V$$ then $A=\operatorname{ext}(P)$.
Note that if $P$ is a convex polytope that contains $0$ in its interior, then for $y\in\operatorname{ext}(P)$ we have $\R_+F_y(P^\circ)=N(y,P)$.
### Conditions on Kirwan’s polytope
If a Hamiltonian action has the same Hofer norm polytope as the action on a *regular* (co)adjoint orbit, then it has the same geodesics; therefore they are characterized by Theorem \[speedchamber\]. We next characterize the Hofer norm polytopes which are derived from regular coadjoint orbits.
Given a Weyl chamber $\t_+$ let $-\t_+=w^*.\t_+$ be the opposite Weyl chamber (there is a unique such $w^*\in\W$). We say that $y\in\t_+$ is *symmetric* if $w^*.y=-y$.
\[extremoshofercadjunta\] A $\W$-invariant symmetric polytope $P$ is the Hofer norm polytope of $\|\cdot\|_{\O}$ for a coadjoint orbit $\O$ if and only if $\operatorname{ext}(P)=\W.y$ for a symmetric $y\in\t_+$.
A $\W$-invariant symmetric polytope $P$ is the Hofer norm polytope of $\|\cdot\|_{\O}$ with $\O$ a regular coadjoint orbit if and only if $\operatorname{ext}(P)=\W.y$ for a symmetric regular $y\in\t_+$.
For $x\in\t_+$ let $\O_x$ be a coadjoint orbit. The Hofer norm polytope of $\|\cdot\|_{\O_x}$ is given by $$P=\operatorname{conv}\{w.x-w'.x:w,w'\in\W\}.$$ If we take $y=x-w^*x\in\t_+$ we claim that $P$ has an extreme point at $y\in\t^+$ with normal cone at this point which equals $\cup_{w\in\operatorname{Stab}(y)}w.\t_+$. To see this note that by Lemma \[rearrengementineq\] for $z\in\t_+$ and $w\in\W$ $$\varphi_z(x)\geq\varphi_z(w.x),$$ and $$\varphi_z(-w^*.x)=\varphi_{-z}(w^*.x)\geq\varphi_{-z}(ww^*.x)=\varphi_{z}(-ww^*.x).$$ Therefore $$\varphi_z(x-w^*.x)\geq\varphi_z(w.x-w'.x),$$ for $w,w'\in\W$ and we conclude that the normal cone to $P$ at $y$ includes $\t_+$.
Since this argument is Weyl group invariant, for $w\in\W$ the normal cone to $P$ at $w.y$ includes $w.\t_+$. In the case of singular $y$ there can be repetitions: if $\operatorname{Stab}(y)$ is not trivial then the normal cone to $P$ at $y$ includes $\cup_{w\in\operatorname{Stab}(y)}w.\t_+$. We have $\cup_{w\in\W}w.\t_+=\t$, hence by Lemma \[normalcones\] $\W.y$ are all the extreme points of $P$ and the normal cone to $P$ at $w'.y$ is $w'.\cup_{w\in\operatorname{Stab}(y)}w.\t_+$.
If $y$ is symmetric then we can take $\O_{{(1/2)}y}$ and the Hofer norm polytope of $\|\cdot\|_{\O_{{(1/2)}y}}$ satisfies $\operatorname{ext}(P)=\W.y$.
The proof of the second assertion is similar and we omit it.
\[kirwant+\] Let $E\subseteq \k$ be an $\operatorname{Ad}$-invariant set such that $E\cap\t_+=\operatorname{conv}\{x_1,\dots,x_n\}$. The Hofer norm polytope derived from $E$ has extreme points $\W.y$ for regular $y$ if there is an $x_i$, say $x_1$, such that
- The point $x_1-w^*.x_1\in\t_+$ is regular.
- For $x\in\t_+$ we have $\varphi_x(x_1)\geq\varphi_x(x_j)$ for $j\in\{2,\dots,n\}$, that is, $\t_+$ is contained in the normal cone of $\operatorname{conv}\{x_1,\dots,x_n\}\subseteq\t$ at $x_1$.
The Hofer norm polytope is given by $$P=\operatorname{conv}\{w.x-w'.x':x,x'\in \{x_1,\dots,x_n\},\quad w,w'\in\W\}.$$ We are going to prove that $\operatorname{ext}(P)=\W.(x_1-w^*.x_1)$. Since $x_1-w^*.x_1$ is regular $\W.(x_1-w^*.x_1)$ has $|\W|$ points. We claim that $P$ has an extreme point at $x_1-w^*.x_1\in\t^+$ with a normal cone at this point which equals $\t_+$. To see this note that by the second assumption in the statement of the theorem and by Lemma \[rearrengementineq\] for $x\in\t_+$, $j\in\{1,\dots,n\}$ and $w\in\W$ $$\varphi_x(x_1)\geq\varphi_x(x_j)\geq\varphi_x(w.x_j),$$ and $$\varphi_x(-w^*.x_1)=\varphi_{-x}(w^*.x_1)\geq\varphi_{-x}(w^*.x_j)\geq\varphi_{-x}(ww^*.x_j)=\varphi_{x}(-ww^*.x_j).$$ Therefore $$\varphi_x(x_1-w^*.x_1)\geq\varphi_x(w.x-w'.x'),$$ for $x,x'\in \{x_1,\dots,x_n\}$ and $w,w'\in\W$ and we conclude that the normal cone to $P$ at $x_1-w^*.x_1$ includes $\t_+$. Since this argument is Weyl group invariant, for $w\in\W$ the normal cone to $P$ at $w.(x_1-w^*.x_1)$ includes $w.\t_+$. We have $\cup_{w\in\W}w.\t_+=\t$, hence by Lemma \[normalcones\] the orbit $\W.(x_1-w^*.x_1)$ are all the extreme points of $P$ and the normal cone to $P$ at $w.(x_1-w^*.x_1)$ is $w.\t_+$.
An example of this is the action of $\operatorname{SU}(3)$ on the singular (co)adjoint orbit containing $\lambda=i\operatorname{diag}(2,-1,-1)$. The Hofer norm polytope is the convex hull of $0$ and the permutations of $i\operatorname{diag}(3,0,-3)$. Its extreme points are the permutations of $i\operatorname{diag}(3,0,-3)$. This is the same Hofer norm polytope as the one derived from the (co)adjoint action on the regular (co)adjoint orbit containing $\frac{i}{2}\operatorname{diag}(-3,0,3)$. Several other examples can be computed from the cases of $\operatorname{SU}(3)$ acting on products of $\mathbb{P}^2$ studied in [@ms]. In fact, it is easy to see from all the figures in [@ms] that all the Kirwan polytopes listed there lead by Theorem \[kirwant+\] to Hofer norm polytopes with extreme points $\W.y$ for regular $y$.
\[cororeg\] Let $\O_x$ for $x\in\t_+$ be a coadjoint orbit. The Hofer norm polytope derived from $\O$ has extreme points $\W.y$ for regular $y$ if and only if $x-w^*.x\in\t_+$ is regular.
An example of the previous corollary is the case of (co)adjoint orbits of $\operatorname{SU}(n)$. Let $y$ be an extreme point of $P$. The set of $x$ such that $\varphi_x$ has a unique maximum at $y$ is an open cone so we can assume that all eigenvalues of $x$ are different, and without loss of generality we assume that they are ordered increasingly: $x_1<\dots <x_n$. Hence if the eigenvalues of the (co)adjoint orbit are $\lambda_1\le \dots\le \lambda_n$ the extreme point of the Hofer norm polytope which maximizes $\varphi_x$ is $$i\operatorname{diag}(\lambda_1,\dots,\lambda_n)-i\operatorname{diag}(\lambda_n,\dots,\lambda_1)=i\operatorname{diag}(\lambda_1-\lambda_n,\lambda_2-\lambda_{n-1},\dots,\lambda_n-\lambda_1).$$ This is the same maximum as the one that would be obtained from looking at the Hofer norm polytope derived from the (co)adjoint orbit with eigenvalues $x_1=\frac{i}{2}(\lambda_1-\lambda_n)<\dots<x_n=\frac{i}{2}(\lambda_n-\lambda_1)$. If these eigenvalues are all distinct then the Hofer norm polytope is equal to the Hofer norm polytope of a regular (co)adjoint orbit.
For the one-sided Hofer norm polytope the condition for being derived from a regular coadjoint orbit, is that there exists a regular $x_i\in P^+$, say $x_1$ such that for $x\in\t_+$ we have $\varphi_x(x_1)\geq\varphi_x(x_j)$ for $j\in\{2,\dots,n\}$, where $$P^+=\operatorname{conv}\{w.x:x\in \{x_1,\dots,x_n\},\,w\in\W\}.$$
The extreme points of the Hofer norm polytope are Weyl group invariant, so we can partition them into orbits $$\operatorname{ext}(P)=\W.y_1\sqcup\dots\sqcup\W.y_m$$ for $y_1,\dots,y_m\in\t_+$. If $y_1,\dots,y_m$ are regular then by Theorem \[abelian\] the unit ball $B$ has abelian faces. We next give a sharper characterization of these faces.
If $\operatorname{ext}(P)=\W.y_1\sqcup\dots\sqcup\W.y_m$ for regular $y_1,\dots,y_m\in\t_+$, then for $i\in\{1,\dots,m\}$ $$\begin{aligned}
F_{y_i}(B)&=\{x\in\t_+:\varphi_{x}(y_i)=1\mbox{ and }\varphi_{x}(y_i)\leq 1\mbox{ for }i\neq j\}.\end{aligned}$$
From Theorem \[abelianfaces\] we know that if the faces of $B$ are abelian then $$F_x(B)=F_x(B\cap\t)\cap\t_+$$ for a Weyl chamber $\t_+$ given by a choice of torus and positive simple roots such that $x\in\t_+$. We take for simplicity $y_i=y_1$, note that $$\begin{aligned}
F_{y_1}(B)&=F_{y_1}(B\cap\t)\cap\t_+=F_{y_1}(P^\circ)\cap\t_+\\
&=\{x\in\operatorname{bd}P^\circ:H_{y_1}\mbox{ supports }P^\circ\mbox{ at }x\}\\
&=\{x\in\operatorname{bd}P^\circ:H_{x}\mbox{ supports }P\mbox{ at }y_1\}\\
&=\{x\in\t_+:\varphi_{x}(y_1)=1\mbox{ and }\varphi_{x}(y_j)\leq 1\mbox{ for }j=2,\dots,n\}.\end{aligned}$$ Note that $H_{x}$ supports $P$ at $y_1$ is equivalent to $\varphi_{x}(y_1)=1$ and $\varphi_{x}(w.y_j)\leq 1$ for $w\in\W$ and $j=1,\dots,n$. This in turn is equivalent by Lemma \[rearrengementineq\] to $\varphi_{x}(y_1)=1$ and $\varphi_{x}(y_j)\leq 1$ for $w\in\W$ and $j=2,\dots,n$ since $x\in\t_+$. This establishes the last equality. The third and fourth equalities follow from previous results on polar duality.
We now have an alternative proof of Proposition \[speedweylchamber\]: the extreme points of the Hofer norm polytope are $\W.y$ for a regular $y\in\t_+$, hence $F_{y}(B)=\{x\in\t_+:\varphi_{x}(y)=1\}$. The cone generated by this face is $\t_+$.
Next we obtain conditions in terms of Kirwan’s polytope which imply that $y_1,\dots,y_m\in\t_+$ are regular. This follows from the previous discussion therefore we omit the proof.
Let $E\subseteq \k$ be an $\operatorname{Ad}$-invariant set such that $E\cap\t_+=\operatorname{conv}\{x_1,\dots,x_n\}$. The Hofer norm polytope derived from $E$ has regular extreme points if the extreme points $y$ of $$A=\operatorname{conv}\{x_i-w^*.x_j:i,j=1,\dots,n\}$$ such that the normal cone $N(y,A)$ intersects the interior of $\t_+$ are all regular.
### Products of actions
We now study products of Hamiltonian actions where the image of the moment map is given by (\[sumaimagmoment\]). A property of regularity on one of the factors implies the same property in the product, in the following two cases:
Consider the case of the Hofer norm $\|\cdot\|_{\O_1+\dots+\O_n}$ for coadjoint orbits $\O_1,\dots,\O_n$. This norm arises by (\[sumaimagmoment\]) from the canonical symplectic action of $K$ on $\O_1\times\dots\times\O_n$.
\[sumaextremounico\] Let $P_1,\dots,P_n$ be $\W$-invariant convex polytopes in $\t$ such that $\operatorname{ext}(P_i)=\W.y_i$ for $y_i\in\t_+$ and $i=1,\dots,n$. Then $\operatorname{ext}(P_1+\dots+P_n)=\W.(y_1+\dots+y_n)$. Hence, if there is a $y_i$ that is regular, then the extreme points of $P_1+\dots+P_n$ are the Weyl group orbit of a regular element.
Let $x$ be a point in the interior of $\t_+$. Then, by Lemma \[rearrengementineq\] $\varphi_x(y_i)>\varphi_x(w.y_i)$ for $i=1,\dots,n$ and $w\in\W\setminus\operatorname{Stab}(y_i)$. Hence $\varphi_x$ has a unique maximum in $P_i$ at $y_i$. This implies that $\varphi_x$ has a unique maximum in $P_1+\dots+P_n$ at $y_1+\dots+y_n$, i.e. $y_1+\dots+y_n$ is an extreme point of $P_1+\dots+P_n$ and $x\in N(P_1+\dots+P_n,y_1+\dots+y_n)$. Since the normal cones are closed $\t_+\subseteq N(P_1+\dots+P_n,y_1+\dots+y_n)$ and since this argument is Weyl group invariant $$w.\t_+\subseteq N(P_1+\dots+P_n,w.(y_1+\dots+y_n))$$ for $w\in\W$. By Lemma \[normalcones\] all the extreme points are $\W.(y_1+\dots+y_n)$. If there is a $j\in\{1,\dots,n\}$ such that $y_j$ is regular, then this point is in the interior of the cone $\t_+$ so that $y_1+\dots+y_n$ is also in the interior, i.e. is regular.
From Proposition \[extremoshofercadjunta\] we get the following
\[speedweylchamber2\] Let $\O_1,\dots,\O_n$ be (co)adjoint orbits and let $\|\cdot\|_{\O_1+\dots+\O_n}$ be the $\operatorname{Ad}$-invariant Hofer norm defined by $\O_1+\dots+\O_n$. If at least one coadjoint orbit is regular, then the Hofer norm polytope derived from $\|\cdot\|_{\O_1+\dots+\O_n}$ has extreme points equal to $\W.y$ for a symmetric regular $y\in\t_+$, i.e. it is the Hofer norm polytope derived from a regular coadjoint orbit.
By Corollary \[arghofer\] a cone $C\subseteq\k$ generated by a face has the same norming functional, so the set $\{\varphi_x:x\in C\}$ has a common maximizer $x^+\in \O_1+\dots+\O_n$ and a common minimizer $x^-\in\O_1+\dots+\O_n$. Let us write $x^-=x^-_1+\dots +x^-_n$ and $x^+=x^+_1+\dots +x^+_n$ with $x^-_i,x^+_i\in \O_i$ for $i=1,\dots,n$. Then by Proposition \[interconosmaxmin\] $$C_{x^-,x^+}(\O_1+\dots+\O_n)=\bigcap_{i=1,\dots,n}C_{x^-_i,x^+_i}(\O_i).$$ If $\O_1$ is regular then by Theorem \[speedweylchamber\] $C_{x^-_1,x^+_1}(\O_1)$ is contained in a Weyl chamber (given by a choice of torus and positive simple roots).
Conversely, if $\t_+$ is a Weyl chamber (given by a choice of torus and positive simple roots) we choose $x^+_i$ to be the intersection of $\O_i$ with $\t_+$ and $x^-_i$ to be the intersection of $\O_i$ with $-\t_+$. By Proposition \[maximumorbit\] the Weyl chamber $\t_+$ is contained in $C_{x^-_i,x^+_i}(\O_i)$ for $i=2,\dots,n$ and by Proposition \[speedweylchamber\] $C_{x^-_1,x^+_1}(\O_1)=\t_+$. Hence if we write $x^-=x^-_1+\dots +x^-_n$ and $x^+=x^+_1+\dots +x^+_n$ we get by Proposition \[interconosmaxmin\] $$C_{x^-,x^+}(\O_1+\dots+\O_n)=\bigcap_{i=1,\dots,n}C_{x^-_i,x^+_i}(\O_i)=\t_+,$$ which is a set with the same norming functionals. We conclude again that the Weyl chambers (given by a choice of torus and positive simple roots) are the cones generated by maximal faces.
We now turn to $\W$-invariant polytopes such that its extreme points are more than one $\W$-orbit.
\[sumaextremoregular\] Let $P_1,\dots,P_n$ be $\W$-invariant convex polytopes in $\t$. If the extreme points of one of the polytopes are all regular then the extreme points of $P_1+\dots+P_n$ are all regular.
Let $y$ be an extreme point of $P_1+\dots+P_n$. There exists an $x\in \t$ such that $\varphi_x$ attains its unique maximum in $P_1+\dots+P_n$ at $y$. Since the normal cone to the polytope $P_1+\dots+P_n$ at $y$ is open we can chose a regular $x$, therefore $\operatorname{Stab}(x)$ is trivial. This regular $x$ is in a unique Weyl chamber which we denote by $\t_+$. We have $y=y_1+\dots +y_n$, where $y_1,\dots,y_n$ are points where $\varphi_x$ attains its unique maximums in $P_1,\dots,P_n$ respectively. Since for $i=1,\dots,n$ we have $\W.y_i\subseteq P_i$ and $\varphi_x$ attains a maximum at $y_i$ Lemma \[rearrengementineq\] implies that $y_1,\dots,y_n$ are in $\t_+$. Since one of the $y_i$ is in the interior of $\t_+$ so is their sum $y=y_1+\dots +y_n$, hence $y$ is regular.
To fininsh this paper, from Proposition \[sumaextremoregular\] and Theorem \[polytopecommuting\] we obtain the following structural property of geodesics (with different metrics) in the group $K$:
\[sumahoferpolyregular\] Let Let $K\curvearrowright M_1,\dots,K\curvearrowright M_n$ be Hamiltonian almost effective actions of a compact semi-simple group $K$ and let $K\curvearrowright M_1\times\dots\times M_n$ be the product action. Endow $K$ with the pullback metrics of Section \[subsectionhoferham\]. Assume that each short curve in $K$ with the metric derived from the action on one factor has commuting Hamiltonians: then all short curves in $K$ with the metric derived from the product action have commuting Hamiltonians.
Another proof of the previous theorem can be given as follows. If the extreme point of one of the Hofer norm polytopes, say $P_1$, are regular, then all the maximal faces of the sphere of the norm $\|\cdot\|_{\mu_1(M)}$ are abelian. The cones of the form $C_{x^-_1,x^+_1}(\mu_1(M))$ for $x^-_1,x^+_1\in \mu_1(M)$ are contained in cones generated by faces and are therefore abelian. The cone generated by faces of the sphere of the norm $\|\cdot\|_{\mu(M)}$ are contained in $C_{x^-,x^+}(\mu(M))$ for $x^-,x^+\in \mu(M)$. Also $$C_{x^-,x^+}(\mu(M))=\bigcap_{i=1,\dots,n}C_{x^-_i,x^+_i}(\mu_i(M))$$ for $x^-_i,x^+_i\in \mu_i(M)$ and $i=1,\dots,n$ by Proposition \[interconosmaxmin\]. Since the first set of the intersection is abelian the result follows.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research was supported by Instituto Argentino de Matemática (CONICET) and Universidad de Buenos Aires.
[XXXxXxX]{}
J. Antezana, G. Larotonda, A. Varela, *Optimal paths for symmetric actions in the unitary group*. Comm. Math. Phys. 328 (2014), no. 2, 481–497.
J. Antezana, E. Ghighlioni, D. Stojanoff, *Minimal curves in $U(n)$ and $Gl(n)^+$ with respect to the spectral and the trace norms*, preprint (2019) arXiv:1907.03368.
M. F. Atiyah, R. Bott, *The Yang-Mills equations over Riemann surfaces*. Philos. Trans. Roy. Soc. London Ser. A 308 (1983), no. 1505, 523–615.
A. Barvinok, *A course in convexity*. Graduate Studies in Mathematics, 54. American Mathematical Society, Providence, RI, 2002.
P. Belkale. *Quantum generalization of the Horn conjecture*. J. Amer. Math. Soc. 21 (2008), no. 2, 365–408.
R. Bhatia, J. A. R. Holbrook. *A softer, stronger Lidskiĭ theorem*. Proc. Indian Acad. Sci. Math. Sci. 99 (1989), no. 1, 75–83.
R. Bhatia, *Matrix analysis*. Graduate Texts in Mathematics, 169. Springer-Verlag, New York, 1997.
M. Bialy, L. Polterovich. *Geodesics of Hofer’s metric on the group of Hamiltonian diffeomorphisms*. Duke Math. J. 76 (1994), no. 1, 273–292.
L. Biliotti, A. Ghigi, P. Heinzner. *Coadjoint orbitopes.* Osaka J. Math. 51 (2014), no. 4, 935–968.
A. Borovik, A. Borovik. *Mirrors and reflections*. The geometry of finite reflection groups. Universitext. Springer, New York, 2010.
J. M. Borwein, A. S. Lewis, Q. J. Zhu, *Convex spectral functions of compact operators. II. Lower semicontinuity and rearrangement invariance. Optimization and related topics* (Ballarat/Melbourne, 1999), 179–196, Appl. Optim., 47, Kluwer Acad. Publ., Dordrecht, 2001.
A. Br[ø]{}ndsted. *An introduction to convex polytopes*. Graduate Texts in Mathematics, 90. Springer-Verlag, New York-Berlin, 1983.
D. Burago, Y. Burago, S. Ivanov. *A course in metric geometry*. Graduate Studies in Mathematics, 33. American Mathematical Society, Providence, RI, 2001.
M. L. Eaton, M. D. Perlman, *Reflection groups, generalized Schur functions, and the geometry of majorization.* Ann. Probability 5 (1977), no. 6, 829–860.
Entov, Michael. *K-area, Hofer metric and geometry of conjugacy classes in Lie groups*. Invent. Math. 146 (2001), no. 1, 93–141.
V. Guillemin, R. Sjamaar. *Convexity properties of Hamiltonian group actions*. CRM Monograph Series, 26. American Mathematical Society, Providence, RI, 2005.
Hofer, H. On the topological properties of symplectic maps. Proc. Roy. Soc. Edinburgh Sect. A 115 (1990), no. 1-2, 25–38.
R. R. Holmes, T.-Y. Tam, *Distance to the convex hull of an orbit under the action of a compact Lie group*. J. Austral. Math. Soc. Ser. A 66 (1999), no. 3, 331–357.
A. W. Knapp. *Lie groups beyond an introduction*. Second edition. Progress in Mathematics, 140. Birkhäuser Boston, Inc., Boston, MA, 2002.
F. Lalonde, D. McDuff. *Hofer’s $L^\infty$-geometry: energy and stability of Hamiltonian flows. I, II.* Invent. Math. 122 (1995), no. 1, 1–33, 35–69.
F. Lalonde, D. McDuff. *The geometry of symplectic energy*. Ann. of Math. (2) 141 (1995), no. 2, 349–371.
G. Larotonda. *The metric geometry of infinite dimensional Lie groups and their homogeneous spaces*, preprint (2019) arXiv:1805.02631.
A. S. Lewis. *Eigenvalue-constrained faces*. Linear Algebra Appl. 269 (1998), 159–181.
A. S. Lewis. *Convex analysis on Cartan subspaces*. Nonlinear Anal. 42 (2000), no. 5, Ser. A: Theory Methods, 813–820.
C.-K. Li, N.-M. Tsing. *Norms that are invariant under unitary similarities and the $C$-numerical radii*. Linear and Multilinear Algebra 24 (1989), no. 3, 209–222.
D. McDuff. *Geometric variants of the Hofer norm*. J. Symplectic Geom. 1 (2002), no. 2, 197–252.
N. McCarthy, D. Ogilvie, N. Zobin, V. Zobin. *Convex geometry of Coxeter-invariant polyhedra.* Trends in Banach spaces and operator theory (Memphis, TN, 2001), 153–179, Contemp. Math., 321, Amer. Math. Soc., Providence, RI, 2003.
D. McDuff, D. Salamon. *Introduction to symplectic topology*. Third edition. Oxford Graduate Texts in Mathematics. Oxford University Press, Oxford, 2017.
J. Milnor. *Curvatures of left invariant metrics on Lie groups*. Advances in Math. 21 (1976), no. 3, 293–329.
J. Montaldi, A. Shaddad *Non-Abelian momentum polytopes for products of $CP^ 2$* preprint (2019) arXiv:1809.09020.
L. Polterovich. *The geometry of the group of symplectic diffeomorphisms.* Lectures in Mathematics ETH Zurich. Birkhauser Verlag, Basel, 2001.
L. Polterovich, E. Shelukhin. *Autonomous Hamiltonian flows, Hofer’s geometry and persistence modules*. Selecta Math. (N.S.) 22 (2016), no. 1, 227–296.
R. T. Rockafellar, *Convex analysis*. Princeton Mathematical Series, No. 28 Princeton University Press, Princeton, N.J. 1970
J.-P. Serre, *Complex semisimple Lie algebras*. Translated from the French by G. A. Jones. Reprint of the 1987 edition. Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2001.
[^1]: The authors were supported by grants PIP 2010-0757 (CONICET), 2010-2478 (ANPCyT)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Circular variables arise in a multitude of data-modelling contexts ranging from robotics to the social sciences, but they have been largely overlooked by the machine learning community. This paper partially redresses this imbalance by extending some standard probabilistic modelling tools to the circular domain. First we introduce a new multivariate distribution over circular variables, called the multivariate Generalised von Mises (mGvM) distribution. This distribution can be constructed by restricting and renormalising a general multivariate Gaussian distribution to the unit hyper-torus. Previously proposed multivariate circular distributions are shown to be special cases of this construction. Second, we introduce a new probabilistic model for circular regression inspired by Gaussian Processes, and a method for probabilistic Principal Component Analysis with circular hidden variables. These models can leverage standard modelling tools (e.g. kernel functions and automatic relevance determination). Third, we show that the posterior distribution in these models is a mGvM distribution which enables development of an efficient variational free-energy scheme for performing approximate inference and approximate maximum-likelihood learning.'
author:
- |
Alexandre K. W. Navarro\
Department of Engineering\
University of Cambridge\
Cambridge, UK\
`[email protected]`\
Jes Frellsen\
Department of Engineering\
University of Cambridge\
Cambridge, UK\
`[email protected]`\
Richard E. Turner\
Department of Engineering\
University of Cambridge\
Cambridge, UK\
`[email protected]`\
bibliography:
- 'references.bib'
title: 'The Multivariate Generalised von Mises Distribution: Inference and Applications'
---
Introduction
============
Many data modelling problems in science and engineering involve circular variables. For example, the spatial configuration of a molecule [@boomsma08; @frellsen09], robot, or the human body [@chirikjian2000] can be naturally described using a set of angles. Phase variables arise in image and audio modelling scenarios [@wadhwa2013], while directional fields are also present in fluid dynamics [@jona-lasinio2012], and neuroscience [@benyishai1995]. Phase-locking to periodic signals occurs in a multitude of fields ranging from biology [@gao2010] to the social sciences [@brunsdon_using_2006].
It is possible, at least in principle, to model circular variables using distributional assumptions that are appropriate for variables that live in a standard Euclidean space. For example, a naïve application might represent a circular variable in terms of its angle $\arv \in [0,2\pi)$ and use a standard distribution over this variable (presumably restricted to the valid domain). Such an approach would, however, ignore the topology of the space e.g. that $\arv =0$ and $\arv=2 \pi$ are equivalent. Alternatively, the circular variable can be represented as a unit vector in $\reals^{2}$, $\vt{x} = [\cos(\arv),\sin(\arv)]^{\top}$, and a standard bivariate distribution used instead. This partially alleviates the aforementioned topological problem, but standard distributions place probability mass off the unit circle which adversely affects learning, prediction and analysis.
In order to predict and analyse circular data it is therefore key that machine learning practitioners have at their disposal a suite of bespoke modelling, inference and learning methods that are specifically designed for circular data [@lebanon2005]. The fields of circular and directional statistics have provided a toolbox of this sort [@mardia_directional_2000]. However, the focus has been on fairly simple and small models that are applied to small datasets enabling MCMC to be tractably deployed for approximate inference.
The goal of this paper is to extend the existing toolbox provided by statistics, by leveraging modelling and approximate inference methods from the probabilistic machine learning field. Specifically, the paper makes three technical contributions. First, in \[sec:mgvm\] it introduces a central multivariate distribution for circular data—called the multivariate Generalised von Mises distribution—that has elegant theoretical properties and which can be combined in a plug-and-play manner with existing probabilistic models. Second, in \[sec:applications\] it shows that this distribution arises in two novel models that are circular versions of Gaussian Process regression and probabilistic Principal Component Analysis with circular hidden variables. Third, it develops efficient approximate inference and learning techniques based on variational free-energy methods as demonstrated on four datasets in \[sec:results\].
Circular distributions primer\[sec:review\]
===========================================
In order to explain the context and rationale behind the contributions made in this paper, it is necessary to know a little background on circular distributions. Since *multidimensional* circular distributions are not generally well-known in the machine learning community, we present a brief review of the main concepts related to these distributions in this section. The expert reader can jump to \[sec:mgvm\] where the multivariate Generalised von Mises distribution is introduced.
A univariate circular distribution is a probability distribution defined over the unit circle. Such distributions can be constructed by wrapping, marginalising or conditioning standard distributions defined in Euclidean spaces and are classified as *wrapped*, *projected* or *intrinsic* according to the geometric interpretation of their construction.
More precisely, the *wrapped* approach consists of taking a univariate distribution $\p(\xx)$ defined on the real line, parametrising any point $\xx \in \reals$ as $\xx = \arv + 2 \pi k$ with $k \in \integers$ and summing over all $k$ so that $\p(\xx)$ is wrapped around the unit circle. The most commonly used wrapped distribution is the Wrapped Gaussian distribution [@ferrari_wrapping_2009; @jona-lasinio2012].
An alternative approach takes a standard bivariate distribution $\p(\xx,\yy)$ that places probability mass over $\reals^2$, transforms it to polar coordinates $[\xx, \yy]^{\top} \rightarrow [r\cos\arv, r\sin\arv]^{\top}$ and marginalises out the radial component $\int_{0}^{\infty}\p(r\cos\arv, r\sin\arv)r \del r$. This approach can be interpreted as projecting all the probability mass that lies along a ray from the origin onto the point where it crosses the unit circle. The most commonly used projected distribution is the Projected Gaussian [@wang_directional_2013].
Instead of marginalising the radial component, circular distributions can be constructed by conditioning it to unity, $p(\xx, \yy | \xx^2+\yy^2=1)$. This can be interpreted as restricting the original bivariate density to the unit circle and renormalising. A distribution constructed in this way is called “intrinsic” (to the unit circle). The construction has several elegant properties. First, the resulting distribution inherits desirable characteristics of the base distribution, such as membership of the exponential family. Second, the form of the resulting density often affords more analytical tractability than those produced by wrapping or projection. The most important intrinsic distribution is the von Mises (vM), $p(\arv|\mu,\kappa) \propto \exp(\kappa \cos(\arv-\mu))$, which is obtained by conditioning an isotropic bivariate Gaussian to the unit circle. The vM has two parameters, the mean $\mu \in [0,2 \pi)$ and the concentration $\kappa \in \reals^{+}$. If the covariance matrix of the bivariate Gaussian is a general real positive definite matrix, we obtain the Generalised von Mises (GvM) distribution [@gatto_generalized_2007][^1] $$\begin{aligned}
p(\arv) \propto \exp(\kappa_1 \cos(\arv-\mu_1) + \kappa_2 \cos( 2 (\arv-\mu_2)))\,.
\label{eq:3}\end{aligned}$$ The GvM has four parameters, two mean-like parameters $\mu_i \in [0,2 \pi)$ and two concentration-like parameters $\kappa_i \in \reals^{+}$ and is an exponential family distribution. The GvM is generally asymmetric. It has two modes when $4\cstd_{2} \geq \cstd_{1}$, otherwise it has one mode except when it is a uniform distribution $\cstd_{2} = \cstd_{1} = 0$.
The GvM is arguably more tractable than the distributions obtained by wrapping or projection as its unnormalised density takes a simple form. In comparison, the unnormalised density of the wrapped normal involves an infinite sum and that of the projected normal is complex and requires special functions. However, the normalising constant of the GvM (and its higher moments) are still complicated, containing infinite sums of modified Bessel functions [@gatto_computational_2008].
In this paper the focus will be on the extensions to vectors of dependent circular variables that lie on a (hyper-) torus (although similar methods can be applied to multivariate hyper-spherical models). An example of a multivariate distribution on the hyper-torus is the multivariate von Mises (mvM) by @mardia_multivariate_2008 $$\begin{aligned}
\mvm(\varv) \propto \exp\Big\{\vcstd^{\top} \cos(\varv) + \sin(\varv)^{\top}\mat{G}\sin(\varv)\Big\}\,.\end{aligned}$$ The terms $\cos(\varv)$ and $\sin(\varv)$ denote element-wise application of sine and cosine functions to the vector $\varv$, $\vcstd$ is a element-wise positive $D$-dimensional real vector, $\vcmean$ is a $D$-dimensional vector whose entries take values on $[0, 2\pi)$, and $\mat{G}$ is a matrix whose diagonal entries are all zeros.
The mvM distribution draws its name from the its property that the one dimensional conditionals, $p(\arv_d|\varv_{\neq d})$, are von Mises distributed. As shown in the Supplementary Material, this distribution can be obtained by applying the *intrinsic* construction to a $2D$-dimensional Gaussian, mapping $\vt{x}\to (r \cos\varv^{\top}, r\sin\varv^{\top})^{\top}$ and assuming its precision matrix has the form $$\begin{aligned}
\mat{W} = \cov^{-1} =
\begin{bmatrix}
\Lambda & \mat{A} \\
\mat{A}^{\top} & \Lambda
\end{bmatrix}
\label{eq:sparsityMvM}\end{aligned}$$ where $\Lambda$ is a diagonal $D$ by $D$ matrix and $\mat{A}$ is an antisymmetric matrix. Other important facts about the mvM are that it bears no simple closed analytic form for its normalising constant, it has $D+(D-1)D/2$ degrees of freedom in its parameters and it is not closed under marginalisation.
We will now consider multi-dimensional extensions of the GvM distribution.
The multivariate Generalised von Mises\[sec:mgvm\]
==================================================
In this section, we present the multivariate Generalised von Mises (mGvM) distribution as an *intrinsic* circular distribution on the hyper-torus and relate it to existing distributions in the literature. Following the construction of *intrinsic* distributions, the multivariate Generalised von Mises arises by constraining a $2D$-dimensional multivariate Gaussian with arbitrary mean and covariance matrix to the $D$-dimensional torus. This procedure yields the distribution $$\begin{gathered}
\mgvm(\varv; \vcmean, \vcstd, \matW) \propto
\exp
\Big\{
\vcstd^{\top}\cos(\varv - \vcmean)
\\
- \frac{1}{2}
\begin{bmatrix}
\cos(\varv)\\
\sin(\varv)
\end{bmatrix}^{\top}
\begin{bmatrix}
\matW^{cc} & \matW^{cs} \\
(\matW^{cs})^{\top} & \matW^{ss}
\end{bmatrix}
\begin{bmatrix}
\cos(\varv) \\
\sin(\varv)
\end{bmatrix}
\Big\}
\label{eq:mgvm_1}\end{gathered}$$ where $\matW^{cc}, \matW^{cs}, \matW^{ss}$ are the blocks of the underlying Gaussian precision matrix $\matW=\cov^{-1}$, $\vcmean$ is a $D$-dimensional angle vector and $\vcstd$ is a $D$-dimensional concentration vector. \[eq:mgvm\_1\] is over-parametrised with $2D + 3(D-1)D/2$ parameters, $D$ more than the degrees of freedom of the most succinct form of the mGvM given in Supplemental Material.
The mGvM distribution generalises the multivariate von Mises by @mardia_multivariate_2008; it collapses to the mvM when $\matW$ has the form of \[eq:sparsityMvM\]. Whereas the one-dimensional conditionals of the mvM are von Mises and therefore unimodal and symmetric, those of the mGvM are generalised von Mises and therefore can be bimodal and asymmetric. The mGvM also captures a richer set of dependencies between the variables than the mvM, notice that the mvM is not the most general form of mGvM that has vM conditionals. The tractability of the one-dimensional conditionals of the mGvM can be leveraged for approximate inference using variational mean-field approximations and Gibbs sampling (see \[sec:inference\]). The mGvM is a member of the exponential family and a maximum entropy distribution subject to multidimensional first and second order circular moments constraints. We will now show that the mGvM can be used to build rich probabilistic models for circular data.
Some applications of the mGvM\[sec:applications\]
=================================================
In this section, we outline two novel and important probabilistic models in which inference produces a posterior distribution that is a mGvM. The first model is a circular analogue of Gaussian Process regression and the second is a version of Principal Component Analysis for circular latent variables.
Regression of circular data\[sec:regression\]
---------------------------------------------
Consider a regression problem in which a set of noisy output circular variables $\{\psi_n \}_{n=1}^N$ have been collected at a number of input locations $\{\vt{s}_n \}_{n=1}^N$. The treatment will apply to inputs that can be multi-dimensional and lie in any space (e.g. they could be circular themselves). The goal is to predict circular variables $\{ \psi^{*}_m \}_{m=1}^M$ at unseen input points $\{ \vt{s}^{*}_m \}_{m=1}^M$. Here we leverage the connection between the mGvM distribution and the multivariate Gaussian in order to produce a powerful class of probabilistic models for this purpose based upon Gaussian Processes. In what follows the outputs and inputs will be represented as vectors and matrices respectively, that is $\vt{\psi}$, $\mathcal{S}$, $\vt{\psi}^*$ and $\mathcal{S}^*$.
In standard Gaussian Process regression [@rasmussen_gaussian_2006] a multivariate Gaussian prior is placed over the underlying unknown function values at the input points $p(\vt{f}|\mathcal{S}) = \gp(\vt{f}; 0, \matK(\vt{s},\vt{s}^{\prime}))$, and a Gaussian noise model is assumed to produce the observations at each input location, $p(y_n|f_n,\vt{s}_n) = \mathcal{N}(y_n;f_n,\sigma_y^2)$. The prior over the function values is specified using the Gaussian Process’s covariance function $K(\vt{s},\vt{s}^{\prime})$ that encapsulates prior assumptions about the properties of the underlying function, such as smoothness, periodicity, stationarity etc. Prediction then involves forming the posterior predictive distribution, $p(\vt{f}^* |\vt{y}, \mathcal{S}, \mathcal{S}^{*})$, which also takes a Gaussian form due to conjugacy.
Here an analogous approach is taken. The circular underlying function values and observations are denoted $\varv$ and $\vt{\psi}$. The prior over the underlying function is given by a mGvM in overparametrised form $p(\varv|\mathcal{S}) = \mgvm(\varv; 0, 0, \matK(\vt{s}, \vt{s}^{\prime})^{-1})$ and the observations are assumed to be von Mises noise corrupted versions of this function $p(\psi_n| \arv_n, \vt{s}_n) = \vm(\psi_{n}; \arv_{n}, \cstd)$. In order to construct a sensible prior over circular function values we use a construction that is inspired by a multi-output GP to produce bivariate variables at each input location. We then leverage the intrinsic construction of the mGvM to constrain each regressed point to the unit circle to allow the mGvM to inherit the properties from the GP covariance function it was built from. This is central to creating a flexible and powerful mGvM regression framework, as GP covariance functions that can handle exotic input variables such as circular variables, strings or graphs [@gartner2003graph; @duvenaud2011additive].
Inference proceeds subtly differently to that in a GP due to an important difference between multivariate Gaussian and multivariate Generalised von Mises distributions. That is, the former are consistent under marginalisation whilst the latter are not: if a subset of mGvM variables are marginalised out, the remaining variables are not distributed according to a mGvM. Technically, this means that for analytic tractability of inference we have to handle the joint posterior predictive distribution $p(\varv, \varv^* |\vt{\psi}, \mathcal{S}, \mathcal{S}^{*})$, which is a mGvM due to conjugacy, rather than $p(\varv^* |\vt{\psi}, \mathcal{S}, \mathcal{S}^{*})$, which is not. Whilst this is somewhat less elegant than GP regression as it requires the prediction locations to be known up front, in many applications this is not a great restriction. This model type is termed transductive [@quinonero2005].
Latent angles: dimensionality reduction and representation learning\[sec:latentAngles\]
---------------------------------------------------------------------------------------
Next consider the task of learning the motion of an articulated rigid body from noisy measurements on a Euclidean space. Articulated rigid bodies can represent a large class of physical problems including mechanical systems, human motion and molecular interactions. The dynamics of rigid bodies can also be fully described by rotations around a fixed point plus a translation and, therefore, can be succinctly represented using angles see [@chirikjian2000]. For simplicity, we will restrict our treatment to a rigid body with $D$ articulations on a 2-dimensional Euclidean space and rotations only, as the discussion trivially generalises to higher dimensional spaces and translations can be incorporated through an extra linear term. Extensions for 3-dimensional models follow directly from the 2-dimensional case, which can be seen as a first step towards these more complex models.
The Euclidean components of any point on an articulated rigid body can be described using the angles between each articulation and their distances. More precisely, for an upright, counter-clockwise coordinate system, the horizontal and vertical components of a point in the $d$-th articulator can be written as $x_{d} = \sum_{j=1}^{d} l_{j} \sin(\vt{\varphi}_{j})$ and $y_{d} = -\sum_{j=1}^{d} l_{j} \cos(\vt{\varphi}_{j})$, where $l_{j}$ is the length of a link $j$ to the next link or the marker. Without loss of generality, we can model only the variation around the mean angle for each joint, i.e. $\varphi_d=\arv_d-\cmean_d$ which results in the general model for noisy measurements $$\begin{bmatrix}
\vyy \\
\vxx
\end{bmatrix}
=
\begin{bmatrix}
-\mat{L} \\
\mat{L}
\end{bmatrix}
\begin{bmatrix}
\cos(\vt{\varphi}) \\
\sin(\vt{\varphi})
\end{bmatrix}
+ \verror
=
\begin{bmatrix}
\matA \\
\matB
\end{bmatrix}
\begin{bmatrix}
\cos(\varv) \\
\sin(\varv)
\end{bmatrix}
+ \verror
\label{eq:rigid_body}$$ where $\mat{L}$ is the matrix that encodes the distances between joints, $\matA$ and $\matB$ are the distance matrix rotated by the vector $\vcmean$ and $\verror\sim\gaussian(0,\stdev^2I)$. The prior over the joint angles can be modelled by a multivariate Generalised von Mises. Here we take inspiration from Principal Component Analysis, and use independent von Mises distributions $$\begin{aligned}
p(\varv_{1,\ldots,N}) &= \textstyle\prod\limits_{n = 1}^{N}\prod\limits_{d = 1}^{D}
\vm\left(\arv_{d,n};0,\cstd_{d}\right).
\label{eq:vonMisesPriors}\end{aligned}$$ Due to conjugacy, the posterior distribution over the latent angles is a mGvM distribution. This can be informally verified by noting that the priors on the latent angles $\varv$ are exponentials of linear functions of sines and cosines, while the likelihood is the exponential of a quadratic function in sine and cosines. This leads to the posterior being an exponential quadratic function of sines and cosines and, hence, mGvM.
The model can be extended to treat the parameters in Bayesian way by including sparse priors over the coefficient matrices $\matA$ and $\matB$ and the observation noise. A standard choice for this task is to define Automatic Relevance Detection priors [@mackay_bayesian_1994] over the columns of these matrices defined as $\gaussian(\matA_{m,d};0,\stdev^{2}_{\matA,d})$ and $\gaussian(\matB_{m,d};0,\stdev^{2}_{\matB,d})$ in order to perform automatic structure learning. Additional Inverse Gamma priors over $\stdev^{2}_{\matA,d}$, $\stdev^{2}_{\matB,d}$ and $\stdev^{2}$ are also employed.
The dimensionality of the latent angle space can be lower than the dimensionality of the observed space, in which case learning and inference perform dimensionality reduction that maps real-valued data to a lower-dimensional torus. Besides motion capture, toroidal manifolds can also prove useful when modelling other relevant applications, such as electroencephalogram (EEG) and audio signals [@turner_probabilistic_2011]. Further connections between dimensionality reduction with the mGvM and Probabilistic Principal Component Analysis (PPCA) proposed by @tipping_probabilistic_1999 (including limiting behaviour and geometrical relations between these models) are explored in the Supplementary Material. As a consequence of these similarities, we denote this model as circular Principal Component Analysis (cPCA).
Approximate inference for the mGvM\[sec:inference\]
===================================================
The multivariate Generalised von Mises does not admit an analytic expression for its normalizing constant, therefore we need to resort to approximate inference techniques. This section presents two approaches that exploit the tractable univariate conditionals of the mGvM: Gibbs sampling and mean-field variational inference.
Gibbs sampling
--------------
A Gibbs sampling procedure for sampling the mGvM of Equation can be derived leveraging the GvM form of the one-dimensional mGvM conditionals. In particular, the Gibbs sampler updates for the $d$-th conditional of the mGvM will have the form $$p(\arv_{d}|\varv_{\neq d}) = \gvm(\arv_{d}; \tilde{\cstd}_{1,d}, \cstd_{2,d},\tilde{\cmean}_{1,d}, \cmean_{2,d})$$ where $\tilde{\cstd}_{1,d}$ and $\tilde{\cmean}_{1,d}$ are functions of $\vcstd$, $\vcmean$ and $\varv_{\neq d}$ given in the Supplementary Material.
The Gibbs sampler can be used to support approximate maximum-likelihood learning by using it to compute the expectations required by the EM algorithm [@wei_tanner_1990]. However, it is well-known that Gibbs sampling becomes less effective as the joint distribution becomes more correlated and the dimensionality grows. This is particularly significant when using the distribution in high-dimensional cases with rich correlational structure, such as those considered later in the paper.
Mean-field Variational inference
--------------------------------
As a consequence of the problems encountered when using Gibbs sampling, the variational inference framework emerges as an attractive, scalable approach to handling inference in when the posterior distribution is a mGvM.
The variational inference framework [@jordan1999] aims to approximate an intractable posterior $p(\varv|\vt{\psi}, \theta)$ with a distribution $q(\varv|\rho)$ by minimising the Kullback-Leiber divergence from the distribution $q$ to $p$. If the approximating distribution is chosen to be fully factored, i.e. $\q(\varv) = \prod_{d=1}^{d}\q_{d}(\arv_{d})$, the optimal functional form for $\q_{d}(\arv_{d})$ can be obtained analytically using calculus of variations. The functional form of each mean-field factor is inherited from the one-dimensional conditionals and consequently is a Generalised von Mises of the form $$\q_{d}(\arv_{d}) = \gvm(\arv_{d}; \bar{\cstd}_{1,d}, \cstd_{2,d},\bar{\cmean}_{1,d}, \cmean_{2,d})$$ where the formulas for the parameters $\bar{\cstd}_{1,d}$ and $\bar{\cmean}_{1,d}$ are similar in nature to the Gibbs sampling update and given in the Supplementary Material.
Furthermore, since the moments of the Generalised von Mises can be computed through series approximations [@gatto_computational_2008], the errors from series truncation are negligible if a sufficiently large number of terms is considered. It is possible to obtain gradients of the variational free energy and optimise it with standard optimisation methods such as Conjugate Gradient or Quasi-Newton methods instead of resorting to coordinate-ascent under the variational Expectation-Maximization algorithm which often is slow to converge.
Despite these improvements, we found empirically that accurate calculations of the moments of a Generalised von Mises distribution can become costly when the magnitude of the concentration parameters exceeds $\approx$100 and the posterior concentrates. This numerical instability occurs when the infinite expansion for computing the moments contains a large number of significant terms that have alternating signs leading to accumulation of numerical errors. It is possible to use other approximate integrations schemes if these cases arise during inference. An alternative way to alleviate this problem is to consider a sub-optimal form of factorised approximating distribution. An obvious choice is to use von Mises factors as this results in tractable updates and requires simpler moment calculations. A von Mises field can also be motivated as a first order approximation to a GvM field by requiring that the log approximating distribution is linear in sine and cosine terms, as shown in the Supplementary Material.
In addition to inference, we can use the same variational framework for learning in cases where the mGvM we wish to approximate is a posterior of tractable likelihoods and priors, as in the cPCA model. To achieve this, we form the variational free-energy lower bound on the log-marginal likelihood as $$\log p(\vt{\psi}|\theta) \geq \free(q,\theta) = \expect{\log p(\vt{\psi},\varv| \theta)}_{q(\varv|\rho)} + \entropy(q),$$ where $\entropy(q)$ is the entropy of the approximating distribution and $q(\varv|\rho)$, $p(\varv, \vt{\psi}| \theta)$ is the model log-joint distribution, $\free(q,\theta,\rho)$ is the variational free-energy, $\theta$ are the model parameters and $\rho$ represents the parameters of the approximating distribution. The same bound cannot be used directly for doubly-intractable mGvM models, such as the circular regression model, and it constitutes an area for further work.
Experimental results\[sec:results\]
===================================
To demonstrate approximate inference on the applications outlined in \[sec:applications\] we present experiments on synthetic and real datasets. A comprehensive description and the data sets used in the all experiments conducted are available at <http://tinyurl.com/mgvm-release>. Further experimental details are also provided in the Supplementary Material.
Comparison to other circular distributions
------------------------------------------
For illustrative purposes we qualitatively compared multivariate Wrapped Gaussian and mvM approximations to a base mGvM and mGvM approximations to these two distributions. The approximations were obtained by numerically minimising the KL divergence between the approximating distribution and the base distributions. These experiments were conducted on a two-dimensional setting in order to render the computation of the normalising constant of the mGvM and the mvM tractable by numerical integration. The resulting distributions are shown in \[fig:mgvm-comparison\].
-- -- -- -- --
-- -- -- -- --
In \[fig:mgvm-comparison\], the mvM and the multivariate wrapped Gaussian cannot capture the multimodality and asymmetry of the mGvM. Moreover, these distributions approximate the multiple modes by increasing their variance and assigning high probability to the region of low-probability between the modes of the mGvM. On the other hand, when the mvM and the multivariate wrapped Gaussian are approximated by the mGvM, the mGvM is able to approximate well the high-probability zones of the wrapped Gaussian and its unimodality and fully recover the mvM.
We also compared the performance of Gibbs sampling and variational inference for a bivariate GvM. To compare the approximate inference procedures, we analysed the run time for each method and the error it produced in terms of the KL divergence between the true distribution and the approximations on a discretized grid. The Gibbs sampling procedure required a total of 3466 samples and to achieve the same level of error as the variational approximation achieved after . The variational approach was considerably more efficient than Gibbs sampling, and theory suggests this behaviour holds for higher dimensions, see @david_mackay_information_2003.
Regression with the mGvM
------------------------
In this section, we investigate the advantages of employing the mGvM regression model discussed in \[sec:regression\] over two common approaches to handling circular data in machine learning contexts.
The first approach is to ignore the circular nature of the data and fit a non-circular model. This approach is not infrequent as it is reasonable in contexts where angles are constrained to a subset of the unit circle and there is no wrappping. A typical example of the motivation for such models is the use of a first-order Taylor approximation to the rate of change of an angle as can be found in classical aircraft control applications. To represent this approach to modeling, we will fit a one-dimensional GP (1D-GP) to the data sets.
The second approach tries to address the circular behaviour by regressing the sine and cosine of the data. In this approach, the angle can be extracted by taking the arc tangent of the ratio between sine and cosine components. While this approach partially addresses the underlying topology of the data, the uncertainty estimates for a non-circular model can be poorly calibrated. Here, each data point is modeled by a two-dimensional vector with the sine and cosine of each data point using a two-dimensional GP (2D-GP).
Five data sets were used in this evaluation. A toy data set generated by wrapping a Mexican hat function around the unit circle, a dataset consisting Uber ride requests in NYC in April 2014[^2], the tide levels predictions from the UK Hydrographic Office in 2016[^3] as function of the latitude and longitude of a given port, the first side chain angle of aspartate as a function of backbone angles in proteins [@harder_beyond_2010], and yeast cell cycle phase as a function of gene expression [@santos2015].
To assess how well the fitted models approximate the distribution of the data, a subset of the data points was kept for validation and the models scored in terms of the log likelihood of the validation data set. To guarantee fairness in the comparison, the likelihood of the 2D-GP was projected back to the unit circle by marginalising the radial component of the model for each point. This converts the 2D-GP into a one-dimensional projected Gaussian distribution over angles. The results are summarised in \[tab:regression\].
\[3\][>[@[\#1]{}[\#2]{}[\#3]{}]{}c<[@end]{}]{}
[lddd]{} Data set & & &\
Toy & & -1.62 & 8.28\
Uber & & -1.49 & -2.83\
Tides & & -6.46 & -8.41\
Protein & & -3.34 & 1.28\
Yeast & & -1.46 & -1.65\
The results shown in \[tab:regression\] indicate that the mGvM provides a better overall fit than the 1D-GP and the 2D-GP in all experiments. The 1D-GP approach performs poorly in every case studied as it cannot account for the wrapping behaviour of circular data. The 2D-GP performs better than the 1D-GP, however in the Uber, Tides and Yeast datasets its performance is substantially closer to the one presented by the 1D-GP case rather than the mGvM. The toy dataset is examined in \[fig:1d\_regression\], showing the 2D-GP learns a different underlying function and cannot capture bimodality.
![Regression on a toy data set using the mGvM (left) and 2D GP (right): data points are denoted by crosses, the true function by circles and predictions by solid dots.[]{data-label="fig:1d_regression"}](uni_model0_mgvm.pdf "fig:"){width="0.45\columnwidth"}![Regression on a toy data set using the mGvM (left) and 2D GP (right): data points are denoted by crosses, the true function by circles and predictions by solid dots.[]{data-label="fig:1d_regression"}](uni_sincos_gp.pdf "fig:"){width="0.45\columnwidth"}
Dimensionality reduction
------------------------
To demonstrate the dimensionality reduction application, we analysed two datasets: one motion capture dataset comprising marker positions placed on a subject’s arm and captured through a low resolution camera and another set comprising of a noisy simulation of a 4-DOF robot arm under the same motion capture conditions.
We compared the model using point estimates for the matrices $\matA$ and $\matB$, a variational Bayes approach by including ARD priors for $\matA$ and $\matB$, Probabilistic Principal Component Analysis (PPCA) [@tipping_probabilistic_1999] and the Gaussian Process Latent Variable Model (GP-LVM) [@lawrence2004gpml] using a squared exponential kernel and a linear kernel. The models using the mGvM require special attention to initialisation. To initialise the test, we used a greedy clustering algorithm to estimate the matrices $\matA$ and $\matB$. The variational Bayes model was initialised using the learned parameters for the point estimate model.
The performance of each model was assessed by denoising the original dataset corrupted by additional Gaussian noise of 2.5, 5 and 10 pixels and comparing the signal-to-noise ratio (SNR) on a test dataset. The best results after initializing the models at 3 different initial starting points are summarized in \[tab:mocap-results\] and additional experiments for a wider range of noise levels are available in the Supplemental Material. In \[tab:mocap-results\], the point estimate cPCA model performs best and is followed by its variational Bayes version for both datasets (the poor performance of the variational Bayes version is likely to be due to biases that can affect variational methods [@turner-and-sahani:2011a]). In the motion capture dataset, the latent angles are highly concentrated. Under these circumstances, the small-angle approximation for sine and cosine provides good results and the cPCA model degenerates into the PPCA model as shown in the Supplementary Material. This behaviour is reflected in the proximity of the PPCA and cPCA signal to noise ratios in \[tab:mocap-results\]. In the robot dataset, the latent angles are less concentrated. As a result, the behaviour of the PPCA and cPCA models is different which explains the larger gap between the results obtained for these models.
\[3\][>[@[\#1]{}[\#2]{}[\#3]{}]{}c<[@end]{}]{}
------------------------ ---------- ---------- ---------- ---------- ---------- ----------
(lr)[2-4]{}(lr)[5-7]{}
cPCA-Point **29.6** **23.5** **17.6** **33.5** **30.0** **24.9**
cPCA-VB 24.6 21.9 17.6 33.2 29.8 24.8
PPCA 23.6 20.9 17.2 22.3 21.8 20.5
GPLVM-SE 8.6 8.5 8.2 21.8 15.7 15.2
GPLVM-L 11.0 7.5 8.1 24.0 16.6 15.9
------------------------ ---------- ---------- ---------- ---------- ---------- ----------
: Signal-to-noise ratio (dB) of the learned latent structure after denoising corrupted signals with by Gaussian noise.[]{data-label="tab:mocap-results"}
Conclusions
===========
In this paper we have introduced the multivariate Generalised von Mises, a new circular distribution with novel applications in circular regression and circular latent variable modelling in a first attempt to close the gap between circular statistics and the machine learning communities. We provided a brief review of the construction of circular distributions including the connections between the Gaussian distribution and the multivariate Generalised von Mises. We provided a scalable way to perform inference on the mGvM model through the variational free energy framework and demonstrated the advantages of the mGvM over GP and mvM through a series of experiments.
Acknowledgements
================
AKWN thanks CAPES grant BEX 9407-11-1. JF thanks the Danish Council for Independent Research grant 0602-02909B. RET thanks EPSRC grants EP/L000776/1 and EP/M026957/1.
Supplementary Material for\
The Multivariate Generalised von Mises\
Distribution: Inference and application
=======================================
Diagramatic view of circular distributions genesis\[sec:circgenesis\]
---------------------------------------------------------------------
The discussion on circular distributions genesis on the main paper can be diagramatically sumarised as \[fig:circDistSummary\].
![Graphical summary of the genesis of circular distributions through transformations of Euclidean distributions.[]{data-label="fig:circDistSummary"}](tikz-fig.pdf)
Derivation of the Multivariate Generalised von Mises\[sec:genesis\]
-------------------------------------------------------------------
The Multivariate Generalised von Mises distribution can be derived by applying the polar transformation to a $2D$-dimensional multivariate Gaussian distribution and conditioning $D$-pairs to the unit circle. Since order in which we conduct these two operations is interchangeable, we will first condition its pairs to the unit circle and then apply the polar transformation.
More precisely, we assume that $\vxx \in \reals^{2D}$, such that $x_{d}^2 + x_{D + d}^2 = 1$ for $d=1, \ldots, D$. In this case, the polar transformation of $\vxx$ allows us to write $x_{d} = \cos(\arv_{d})$, $x_{D + d} = \sin(\arv_{d})$. Furthermore, without loss of generality, the mean $\vmean$ of a multivariate Gaussian will also be constrained to the unit circle and can be parametrised in terms of angles $\vcmean$ so that $\mean_{d}=\cos(\cmean_n)$, $\mean_{D + d}=\sin(\cmean_{D + d})$.
Now let and $\matW = \cov^{-1}$ be the inverse covariance matrix of a multivariate Gaussian. Using the parametrisation of $\vxx$ and $\vmean$ in terms of $\varv$ and $\vcmean$, we can expand the quadratic in the exponential of the multivariate Gaussian into $$\begin{aligned}
\begin{bmatrix}
\cos(\varv)-\cos(\vcmean)\\
\sin(\varv)-\sin(\vcmean)
\end{bmatrix}^{\top}
&\matW
\begin{bmatrix}
\cos(\varv)-\cos(\vcmean) \\
\sin(\varv)-\sin(\vcmean)
\end{bmatrix}
\label{eq:expanded_prod1}
\\
&= \sum_{d=1}^{D} w_{d,d}(\cos(\arv_{d})-\cos(\cmean_{d}))^2
+ w_{D+n,D+d}(\sin(\arv_{d})-\sin(\cmean_{d}))^2
\notag
\\
&+ \sum_{d=1}^{D}\sum_{\jj=1}^{d-1}
w_{d, \jj}(\cos(\arv_{d})-\cos(\cmean_{d}))(\cos(\arv_{\jj})-\cos(\cmean_{\jj}))
\notag
\\
&+ \sum_{d=1}^{D}\sum_{\jj=1}^{D}
w_{d, D+\jj}(\cos(\arv_{d})-\cos(\cmean_{d}))(\sin(\arv_{\jj})-\sin(\cmean_{\jj}))
\notag
\\
&+ \sum_{d=1}^{D}\sum_{\jj=1}^{D}
w_{D+d,\jj}(\sin(\arv_{d})-\sin(\cmean_{d}))(\cos(\arv_{\jj})-\cos(\cmean_{\jj}))
\notag
\\
&+ \sum_{d=1}^{D}\sum_{\jj=1}^{d-1}
w_{D+d,D+\jj}(\sin(\arv_{d})-\sin(\cmean_{d}))(\sin(\arv_{\jj})-\sin(\cmean_{\jj}))
\notag.\end{aligned}$$
The sums on the RHS in Equation can be expanded into $$\begin{aligned}
&\sum_{d=1}^{D} w_{d,d}(\cos(\arv_{d})^2 - 2\cos(\arv_{d})\cos(\cmean_{d}) -\cos(\cmean_{d})^2)
\label{eq:expanded_prod2}
\\
&+ \sum_{d=1}^{D}w_{D+n,D+d}(\sin(\arv_{d})^2 - 2\sin(\arv_{d})\sin(\cmean_{d}) -\sin(\cmean_{d})^2)
\notag
\\
&+2\sum_{d=1}^{D}\sum_{\jj=1}^{d-1}
w_{d, \jj}(\cos(\arv_{d})\cos(\arv_{\jj})-\cos(\cmean_{d})\cos(\arv_{\jj})
-\cos(\arv_{d})\cos(\cmean_{\jj})+\cos(\cmean_{d})\cos(\cmean_{\jj}))
\notag
\\
&+2\sum_{d=1}^{D}\sum_{\jj=1}^{d-1}
w_{D+d, D+\jj}(\sin(\arv_{d})\sin(\arv_{\jj})-\sin(\cmean_{d})\cos(\arv_{\jj})
-\sin(\arv_{d})\sin(\cmean_{\jj})+\sin(\cmean_{d})\cos(\cmean_{\jj}))
\notag
\\
&+ \sum_{d=1}^{D}\sum_{\jj=1}^{D}
w_{d,D+\jj}(\cos(\arv_{d})\sin(\arv_{\jj})-\cos(\cmean_{d})\sin(\arv_{\jj})-\cos(\arv_{d})\sin(\cmean_{\jj})+\cos(\cmean_{d})\sin(\cmean_{\jj}))
\notag
\\
&+ \sum_{d=1}^{D}\sum_{\jj=1}^{D}
w_{D+d, \jj}(\sin(\arv_{d})\cos(\arv_{\jj})-\sin(\cmean_{d})\cos(\arv_{\jj})-\sin(\arv_{d})\cos(\cmean_{\jj})+\sin(\cmean_{d})\cos(\cmean_{\jj}))
\notag.\end{aligned}$$
By aggregating all terms that are independent of $\arv$ and rearranging terms, Equation becomes $$\begin{aligned}
&\sum_{d=1}^{D}\sum_{\jj=1}^{D}
w_{d, \jj}\cos(\arv_{d})\cos(\arv_{\jj}) + 2
w_{d,D+\jj}\cos(\arv_{d})\sin(\arv_{\jj}) +
w_{D+d,D+\jj}\sin(\arv_{d})\sin(\arv_{\jj}) \label{eq:aggregated}
\\
&-2\sum_{d=1}^{D}\sum_{\jj=1}^{D}
w_{d,\jj}\cos(\arv_{d})\cos(\cmean_{\jj}) +
w_{d,D+\jj}\cos(\arv_{d})\sin(\cmean_{\jj}) \notag
\\
&-2\sum_{d=1}^{D}\sum_{\jj=1}^{D}
w_{D+d,D+\jj}\sin(\arv_{d})\sin(\cmean_{\jj}) +
w_{D+d,\jj}\sin(\arv_{d})\cos(\cmean_{\jj}) \notag\end{aligned}$$
These sums can be written in matrix notation as $$\begin{aligned}
\vcstd_{c}^{\top} \cos(\varv-\vcmean)
+ \vcstd_{s}^{\top} \sin(\varv-\vcmean)
-\frac{1}{2}
\begin{bmatrix}
\cos(\varv) \\
\sin(\varv)
\end{bmatrix}^{\top}
\begin{bmatrix}
\matW^{cc} & \matW^{cs}\\
(\matW^{cs})^\top & \matW^{ss}
\end{bmatrix}
\begin{bmatrix}
\cos(\varv) \\
\sin(\varv)
\end{bmatrix}
\label{eq:constrained_quadratic}\end{aligned}$$ where $\cstd_{d}=\text{abs}\{z_{d}\}$ and $\cmean_{d}=\text{arg}\{z_{d}\}$ with the real and imaginary parts of $z_{d}$ such that $$\Re\{z_{d}\} = -2\sum_{\jj=1}^{D} w_{d, \jj}\cos(\arv_{d})\cos(\cmean_{\jj}) + w_{d,D+\jj}\cos(\arv_{d})\sin(\cmean_{\jj})$$ and $$\Im\{z_{d}\} = -2\sum_{\jj=1}^{D} w_{D+d, D+\jj}\sin(\arv_{d})\sin(\cmean_{\jj}) + w_{D+d, \jj}\sin(\arv_{d})\cos(\cmean_{\jj}).$$
Therefore, Equation imples that a multivariate Gaussian distribution under radial transformation and conditionning to the unit circle yields the log density $$\begin{aligned}
\log p(\varv)
&= \const +
\vcstd^{\top} \cos(\varv-\vcmean) \notag \\
&-\frac{1}{2} \Big(
\cos(\varv)^{\top}\matW^{cc}\cos(\varv) + 2
\cos(\varv)^{\top}\matW^{cs}\sin(\varv) +
\sin(\varv)^{\top}\matW^{ss}\sin(\varv) \Big)
\label{eq:log_mGvM}\end{aligned}$$ which is the log density of a multivariate Generalised von Mises distribution in overparametrised form.
To obtain the minimal number of parameters for the mGvM, Equation can be further simplified using trigonometric identities to yield the minimal form of the mGvM distribution $$\begin{aligned}
\log p(\varv)
&= \const +
\exp
\Big\{
\vcstd_{1}^{\top}\cos(\varv - \vcmean_{1})
+ \vcstd_{2}^{\top}\cos(2 (\varv - \vcmean_{2})) \notag \\
&
+ \frac{1}{2} \sum_{d=1}^{D}\sum_{j=1}^{D} u_{d,j} \cos(\arv_{d}-\arv_{j} - \alpha_{d,j}) + v_{d,j} \cos(\arv_{d} + \arv_{j} - \beta_{d,j})
\Big\}.
\label{minimal-mgvm}\end{aligned}$$ where $\vcstd_{1}=\vcstd$ and $\vcmean_{1}=\vcmean$ given as before, while $\cstd_{2,d}=\text{abs}\{z_{d}\}$ and $\cmean_{2,d}=0.5\text{arg}\{z_{2,d}\}$ with $$z_{d} = \frac{1}{4}(w_{d, d} - w_{D+d, D+d}) + i\frac{1}{2}(w_{d, D+d})$$ and the cross terms given by $u_{d,j}=\text{abs}\{z^{U}_{d,j}\}$, $\alpha_{d,j}=0.5\text{arg}\{z^{U}_{d,j}\}$, $v_{d,j}=\text{abs}\{z^{V}_{d,j}\}$, $\beta_{d,j}=0.5\text{arg}\{z^{V}_{d,j}\}$ where $$\begin{aligned}
z^{U}_{d,j} &= (w_{d,j} + w_{D+d, D+j}) + i (w_{j,D+d} - w_{d, D+j})\\
z^{V}_{d,j} &= (w_{d,j} - w_{D+d, D+j}) + i (w_{j,D+d} + w_{d, D+j}).\end{aligned}$$
A final point to make about the mGvM derivation is related to the distributions it generalises. @gatto_generalized_2007 discussed that the GvM could be constructed by conditioning a 2D Gaussian the unit circle, but were not aware of multivariate generalisations. @mardia_multivariate_2008 constructed the multivariate mvM, which we show is a submodel of the mGvM, but did not relate it to the a multivariate Gaussian nor to kernels.
Informal argument for mGvM being the maximum entropy distribution on the hyper-torus\[sec:maxent\]
--------------------------------------------------------------------------------------------------
The maximum entropy distribution $p$ subject to specified covariance and the first and second moments is the solution for the problem $$\begin{aligned}
\text{minimize}_{p} & \int p(\vt{x}) \log p(\vt{x}) \del \varv \\
\text{subject to} & \int x_{d}^{m} p(\varv) \del \arv_{d} = \alpha_{d,m}, \quad d = 1, \ldots, 2D; m = 0, \ldots, 2\end{aligned}$$ is the multivariate Gaussian distribution. If we further add the constraints to the maximum entropy problem that the distribution must be under the unit circle, the problem becomes $$\begin{aligned}
\text{minimize}_{p} & \int p(\vt{x}) \log p(\vt{x}) \del \varv \\
\text{subject to} & \int x_{d}^{m} p(\varv) \del \arv_{d} = \alpha_{d,m}, \quad d = 1, \ldots, 2D; m = 0, \ldots, 2 \\
& x_{d}^{2} + x_{d+D}^{2} = 1, \quad d = 1, \ldots, D;\end{aligned}$$ the solution of which is a multivariate Gaussian constrained to the unit hyper-torus, hence, a the mGvM distribution.
Conditionals of the mGvM: derivation and their relationship to inference algorithms\[sec:conditionals\]
-------------------------------------------------------------------------------------------------------
The conditionals of the mGvM can be found by expanding the terms containing cosine of the difference and sum of two circular variables terms using sum-to-product relations. More precisely, if we partition the indexes of mGvM distributed circular vector $\varv$ into two disjoint sets $\mathcal{A}$ and $\mathcal{B}$, $$\begin{aligned}
p(\varv_{\mathcal{A}}|\varv_{\mathcal{B}})
&\propto
\exp
\Big\{
\vcstd_{1,\mathcal{A}}^{\top}\cos(\varv_{\mathcal{A}} - \vcmean_{1})
\vcstd_{1,\mathcal{B}}^{\top}\cos(\varv_{\mathcal{B}} - \vcmean_{1})
\notag \\
&
+ \vcstd_{2,\mathcal{A}}^{\top}\cos(2 (\varv_{\mathcal{A}} - \vcmean_{2}))
+ \vcstd_{2,\mathcal{B}}^{\top}\cos(2 (\varv_{\mathcal{B}} - \vcmean_{2}))
\notag \\
&
+ \frac{1}{2} \sum_{d\in\mathcal{A}}\sum_{j\in\mathcal{A}} u_{d,j} \cos(\arv_{d}-\arv_{j} - \alpha_{d,j}) + v_{d,j} \cos(\arv_{d} + \arv_{j} - \beta_{d,j})
\notag \\
&
+ \frac{1}{2} \sum_{d\in\mathcal{A}}^{D}\sum_{j\in\mathcal{B}}^{D} u_{d,j} \cos(\arv_{d}- \arv_{j} - \alpha_{d,j}) + v_{d,j} \cos(\arv_{d} + \arv_{j} - \beta_{d,j})
\notag \\
&
+ \frac{1}{2} \sum_{d\in\mathcal{B}}^{D}\sum_{j\in\mathcal{A}}^{D} u_{d,j} \cos(\arv_{d}- \arv_{j} - \alpha_{d,j}) + v_{d,j} \cos(\arv_{d} + \arv_{j} - \beta_{d,j})
\notag \\
&
+ \frac{1}{2} \sum_{d\in\mathcal{B}}^{D}\sum_{j\in\mathcal{B}}^{D} u_{d,j} \cos(\arv_{d}- \arv_{j} - \alpha_{d,j}) + v_{d,j} \cos(\arv_{d} + \arv_{j} - \beta_{d,j})
\Big\}.\end{aligned}$$ If we consider that the variables whose indexes are in $\mathcal{B}$ are constant and note that the cross terms between variables in $\mathcal{A}$ and $\mathcal{B}$ have the functional form of $\cstd\cos(\arv_{\mathcal{A}} - \cmean)$, we can rewrite the conditional using phasor arithmetic as $$\begin{aligned}
p(\varv_{\mathcal{A}}|\varv_{\mathcal{B}})
&\propto
\exp
\Big\{
\tilde{\vcstd}_{1}^{\top}\cos(\varv_{\mathcal{A}} - \tilde{\vcmean_{1}})
+ \vcstd_{2,\mathcal{A}}^{\top}\cos(2 (\varv_{\mathcal{A}} - \vcmean_{2}))
\notag \\
&
+ \frac{1}{2} \sum_{d\in\mathcal{A}}\sum_{j\in\mathcal{A}} u_{d,j} \cos(\arv_{d}-\arv_{j} - \alpha_{d,j}) + v_{d,j} \cos(\arv_{d} + \arv_{j} - \beta_{d,j})\end{aligned}$$ where $\tilde{\cstd}_{1,d}=\text{abs}(z_d)$, $\tilde{\vcmean}_{1,d}=\text{arg}(z_d)$ and $$\begin{aligned}
\Re(z_d) &= \cstd_{1,d}\cos(\cmean_{1,d}) + \sum_{j=1} u_{d,j}\cos(\arv_{j} - \alpha_{d,j}) - v_{d,j}\cos(\arv_{j} - \beta_{d,j})\\
\Im(z_d) &= \cstd_{1,d}\sin(\cmean_{1,d}) + \sum_{j=1} u_{d,j}\sin(\arv_{j} - \alpha_{d,j}) - v_{d,j}\sin(\arv_{j} - \beta_{d,j})\end{aligned}$$ with $\Re(z)$ denoting the real part of $z$ and $\Im(z)$ denoting the imaginary part of $z$.
In the particular case of the unidimensional conditional, the covariance term $$+ \frac{1}{2} \sum_{d\in\mathcal{A}}\sum_{j\in\mathcal{A}} u_{d,j} \cos(\arv_{d}-\arv_{j} - \alpha_{d,j}) + v_{d,j} \cos(\arv_{d} + \arv_{j} - \beta_{d,j})$$ will vanish as the diagonals of the parameter matrices $\mat{U}$ and $\mat{V}$ are zero.
### Unidimensional conditionals and Gibbs sampling
When the set $\mathcal{A}$ contains a single index, the expressions in the previous section define how to obtain all unidimensional conditionals of the mGvM. These one-dimensional conditionals are the same mentioned in the main paper for the Gibbs sampling, for which the parametric dependencies can be explicitly written as $$\begin{aligned}
\tilde{\cstd}_{1,d} = \text{abs}(z_d) & \tilde{\cmean}_{1,d} = \text{arg}(z_d)\end{aligned}$$ where $z_{d}=$ is given by $$\begin{aligned}
\Re(z_d) &= \cstd_{1,d}\cos(\cmean_{1,d}) + \sum_{j=1} u_{d,j}\cos(\arv_{j} - \alpha_{d,j}) - v_{d,j}\cos(\arv_{j} - \beta_{d,j})\\
\Im(z_d) &= \cstd_{1,d}\sin(\cmean_{1,d}) + \sum_{j=1} u_{d,j}\sin(\arv_{j} - \alpha_{d,j}) - v_{d,j}\sin(\arv_{j} - \beta_{d,j})\end{aligned}$$ with $\Re(z)$ denoting the real part of $z$ and $\Im(z)$ denoting the imaginary part of $z$.
### Unidimensional conditionals and mean field variational approximation
To find the approximation $q(\arv|\rho)$ for a true posterior $p(\arv|\psi, \theta)$ that minimisesthe Kullback-Leiber divergence from $q$ to $p$ under the variational free energy framework [@jordan1999], can equivalently maximise the variational free energy $\free(\q,\params_{\p})$ by noting
$$\begin{aligned}
\text{KL}(q(\varv)||p(\varv|\vt{\psi}))
&= \int q(\varv|\rho)\log \frac{\q(\varv|\rho)}{p(\varv|\vt{\psi},\theta)}\del \varv \\
&=-\int q(\varv|\rho)\log p(\arv|\vt{\psi},\theta)\del \arv +\int q(\varv|\rho)\log q(\varv|\rho) \del \varv\\
&=-\langle\log \p(\varv,\vt{\psi}|\theta)\rangle_{\q(\varv|\rho)} +\log\p(\vt{\psi}|\theta) - \mathcal{H}(q) \\
&= \log \p(\vt{\psi}|\theta) - \free(q, \theta, \rho).\end{aligned}$$
By assuming a fully factored form for the distribution $q$, i.e., $q(\varv)=\prod_{d=1}^{D} q_{d}(\arv_{d})$, we can use calculus of variations to obtain analytically the functional form of the distributions $q_d$, that is $$\frac{\delta}{\delta \q} \free(\q,\params_{\p}) - \lambda \left(\int \q(\varv)\del \varv - 1 \right) = 0$$ which implies $$\begin{aligned}
\frac{\delta}{\delta \q_{\ell}}
\left[
\expect{\log \p(\arv,\psi)}_{\prod_{\dd}^{\DD}\q_{\dd}(\arv_{\dd})}
-\int \prod_{\dd=1}^{\DD} \q_{\dd}(\arv_{\dd})
\left(\sum_{\dd=1}^{\DD}\log \q_{\dd}(\arv_{\dd})\right)
\del \varv
- \lambda \sum_{\dd=1}^{\DD}
\int \q_{\dd}(\arv_{\dd}) d\arv_{\dd}
\right] = 0\end{aligned}$$ and leads to the factors approximation $$\q_{d}(\arv_{d}) = \frac{1}{\exp(\lambda+1)}\exp
\left\{
\expect{\log \p(\arv,\psi)}_{\prod_{\neq d}^{D}\q_{\neq d}(\arv_{\neq d})}
\right\}
\label{eq:meanFieldEquation}$$ resulting in the set of distributions known as mean field approximation. This equation when applied to the mGvM yields GvM distributions $$\begin{aligned}
q(\arv_{d}|\varv_{\neq d}) = \gvm(\arv_{d};
\bar{\cstd}_{1,d}(\vcstd,\vcmean, \langle e^{i\varv_{\neq d}}\rangle_{q_{\neq d}}, \langle e^{i \varv_{\neq d}} \rangle_{q_{\neq d}}),\cstd_{2,d},
\bar{\cmean}_{1,d}(\vcstd,\vcmean, \langle e^{i\varv_{\neq d}}\rangle_{q_{\neq d}}, \langle e^{i \varv_{\neq d}} \rangle_{q_{\neq d}}), \cmean_{2,d})\end{aligned}$$ where $\langle e^{i n \varv}\rangle= \langle \cos(n \varv)\rangle + i\langle \sin(n \varv)\rangle $, $\bar{\cstd}_{1,d} = \text{abs}(z_d)$ and $\bar{\cmean}_{1,d} = \text{arg}(z_d)$ with $z_{d}=$ given in overparametrised form by $$\begin{aligned}
\Re(z_d) &= \cstd_{d}\cos(\cmean_{1,d}) -\frac{1}{2} \left\langle
\begin{bmatrix}
\cos(\varv_{\neq d}) \\
\sin(\varv_{\neq d})
\end{bmatrix}^{\top}
\begin{bmatrix}
\matW^{cc}_{\neq d, d} & \matW^{cs}_{\neq d, d}\\
(\matW^{cs})_{\neq d, d}^\top & \matW^{ss}_{\neq d, d}
\end{bmatrix}
\begin{bmatrix}
\cos(\varv_d) \\
\sin(\varv_d)
\end{bmatrix}
\right\rangle_{q_{\neq \varv_d}}
\\
\Im(z_d) &= \cstd_{d}\sin(\cmean_{1,d}) -\frac{1}{2} \left\langle
\begin{bmatrix}
\cos(\varv_{\neq d}) \\
\sin(\varv_{\neq d})
\end{bmatrix}^{\top}
\begin{bmatrix}
\matW^{cc}_{\neq d, d} & \matW^{cs}_{\neq d, d}\\
(\matW^{cs})_{\neq d, d}^\top & \matW^{ss}_{\neq d, d}
\end{bmatrix}
\begin{bmatrix}
\cos(\varv_d) \\
\sin(\varv_d)
\end{bmatrix}
\right\rangle_{q_{\neq \varv_d}}\end{aligned}$$ with $\Re(z)$ denoting the real part of $z$ and $\Im(z)$ denoting the imaginary part of $z$.
Higher order mGvM\[sec:higher\]
-------------------------------
As with the higher order GvM, the mGvM can also be expanded to include $T$ cosine harmonics if the Gaussian genesis is cast aside. In this case, a mGvM of order $T$ can be defined as $$\begin{aligned}
\mgvm_{T}(\varv; \vcmean_{1:T}, \vcstd_{1:T}, \mat{U}, \mat{V}, \mat{\alpha}, \mat{\beta})
&\propto
\exp
\Big\{
\sum_{t=1}^{T} \vcstd_{t}^{\top}\cos(t (\varv - \vcmean_{t})) \notag \\
&
+ \frac{1}{2} \sum_{i=1}^{D}\sum_{j=1}^{D} u_{i,j} \cos(\arv_{i}-\arv_{j} - \alpha_{i,j}) + v_{i,j} \cos(\arv_{i}+\arv_{j} - \beta{i,j})
\Big\}
\label{eq:mgvm_2}\end{aligned}$$ which is a distribution whose conditionals allow up to $T$ modes, but bears the same correlation correlation structure of the ‘standard’ order 2 mGvM.$\mathcal{A}$ is a single index, the
Assumptions over the precision matrix of the Gaussian that leads to a mvM
-------------------------------------------------------------------------
In this section the mvM is derived by conditioning a 4-dimensional multivariate Gaussian to highlight the assumptions made regarding the precision matrix of the multivariate Gaussian. We take the 4D Gaussian to be zero mean without loss of generality and, after applying the polar variable transformation and constraining the radial components to unity we obtain the distribution $$\begin{aligned}
p(\phi_1, \phi_2)
&\propto \exp
\left\{
-\frac{1}{2}
\begin{bmatrix}
\cos(\phi_1)\\
\cos(\phi_2)\\
\sin(\phi_1)\\
\sin(\phi_2)
\end{bmatrix}^{\top}
\begin{bmatrix}
a_{1,1} & a_{1,2} & a_{1,3} & a_{1,4} \\
a_{1,2} & a_{2,2} & a_{2,3} & a_{2,4} \\
a_{1,3} & a_{2,3} & a_{3,3} & a_{3,4} \\
a_{1,4} & a_{2,4} & a_{3,4} & a_{4,4} \\
\end{bmatrix}
\begin{bmatrix}
\cos(\phi_1)\\
\cos(\phi_2)\\
\sin(\phi_1)\\
\sin(\phi_2)
\end{bmatrix}
\right\}\end{aligned}$$ which can be expanded into $$\begin{aligned}
p(\phi_1, \phi_2)
&\propto \exp
\Big\{
- \frac{1}{2}
(
a_{1,1} \cos(\phi_1)^2 + a_{2,2} \cos(\phi_2)^2 + a_{3,3} \sin(\phi_1)^2 + a_{4,4} \sin(\phi_2)^2 + \notag \\
& 2 a_{1,2} \cos(\phi_1) \cos(\phi_2) + 2 a_{1,3} \cos(\phi_1) \sin(\phi_1) + 2 a_{1,4} \cos(\phi_1) \sin(\phi_2) + \notag \\
& 2 a_{2,3} \cos(\phi_2) \sin(\phi_1) + 2 a_{2,4} \cos(\phi_2) \sin(\phi_2) + 2 a_{3,4} \sin(\phi_1) \sin(\phi_2)
)
\Big\}.
\notag\end{aligned}$$ Using the fundamental identity and double angle formulas, we can rewrite the last Equation as $$\begin{aligned}
p(\phi_1, \phi_2)
&\propto \exp
\Big\{
- \frac{1}{2}
(
a_{1,1} \cos(\phi_1)^2 + a_{2,2} \cos(\phi_2)^2 + a_{3,3} (1 - \cos(\phi_1)^2) + a_{4,4} (1 - \cos(\phi_2)^2) + \notag \\ & 2 a_{1,2} \cos(\phi_1) \cos(\phi_2) + a_{1,3} \sin(2 \phi_1) + 2 a_{1,4} \cos(\phi_1) \sin(\phi_2) + 2 a_{2,3} \cos(\phi_2) \sin(\phi_1) + \notag \\
& a_{2,4} \sin(2 \phi_2) + 2 a_{3,4} \sin(\phi_1) \sin(\phi_2)) )
\Big\} \notag\end{aligned}$$
Further simplifications arise from $$\begin{aligned}
p(\phi_1, \phi_2)
&\propto \exp
\Big\{
- \frac{1}{2}
(
(a_{1,1} - a_{3,3}) \cos(2 \phi_1 - 2 \nu_1) + (a_{2,2} - a_{4,4}) \cos(2 \phi_2 - 2 \nu_2) + \notag \\ & 2 a_{1,2} \cos(\phi_1) \cos(\phi_2) + 2 a_{1,4} \cos(\phi_1) \sin(\phi_2) +
2 a_{2,3} \cos(\phi_2) \sin(\phi_1) + \notag \\
& 2 a_{3,4} \sin(\phi_1) \sin(\phi_2) + a_{1,3} \sin(2 \phi_1) + a_{2,4} \sin(2 \phi_2) )
\Big\}. \notag\end{aligned}$$ The product of sine and cosines can be also translated using product-to-sum formulas $$\begin{aligned}
p(\phi_1, \phi_2)
&\propto \exp
\Big\{
- \frac{1}{2}
(
(a_{1,1} - a_{3,3}) \cos(2 \phi_1 - 2 \nu_1) + (a_{2,2} - a_{4,4}) \cos(2 \phi_2 - 2 \nu_2) + \notag \\
& a_{1,2} \cos(\phi_1 - \phi_2) + a_{1,2} \cos(\phi_1 + \phi_2) + a_{1,4} \sin(\phi_1 + \phi_2) - a_{1,4} \sin(\phi_1 - \phi_2) + \notag \\ & a_{2,3} \sin(\phi_1 + \phi_2) + a_{2,3} \sin(\phi_1 - \phi_2) + a_{3,4} \cos(\phi_1 - \phi_2) - a_{3,4} \cos(\phi_1 + \phi_2) + \notag \\ & a_{1,3} \sin(2 \phi_1) + a_{2,4} \sin(2 \phi_2))
\Big\} \notag\end{aligned}$$ grouping similar terms $$\begin{aligned}
p(\phi_1, \phi_2)
&\propto \exp
\Big\{
- \frac{1}{2}
(
(a_{1,1} - a_{3,3}) \cos(2 \phi_1 - 2 \nu_1) + (a_{2,2} - a_{4,4}) \cos(2 \phi_2 - 2 \nu_2) + \notag \\
& a_{1,3} \sin(2 \phi_1) + a_{2,4} \sin(2 \phi_2) + \notag \\
& (a_{1,2} + a_{3,4}) \cos(\phi_1 - \phi_2) + (a_{1,2} - a_{3,4}) \cos(\phi_1 + \phi_2) + \notag \\
& (a_{2,3} - a_{1,4}) \sin(\phi_1 - \phi_2) + (a_{1,4} + a_{2,3}) \sin(\phi_1 + \phi_2))
\Big\} \notag\end{aligned}$$ Therefore, we can conclude that since the mvM has only the term $\sin(\phi_1 - \phi_2)$ from the equation above that $a_{1,1}=a_{3,3}$, $a_{2,2}=a_{4,4}$, $a_{1,3}=a_{2,4}=a_{1,2}=a_{3,4}=0$ and $a_{1,4}=-a_{3,2}$. This leads to the precision matrix having the sparsity pattern $$\begin{aligned}
\begin{bmatrix}
a_{1,1} & 0 & 0 & a_{1,4} \\
0 & a_{2,2} & -a_{1,4} & 0 \\
0 & -a_{1,4} & a_{1,1} & 0 \\
a_{1,4} & 0 & 0 & a_{2,2} \\
\end{bmatrix}.
\notag\end{aligned}$$
Dimensionality reduction with the mGvM and Probabilistic Principal Component Analysis\[sec:CPCA\]
-------------------------------------------------------------------------------------------------
In this section we discuss in greater detail dimensionality reduction with the Multivariate Generalised von Mises and its relationship to Probabilistic Principal Component Analysis (PPCA) from [@tipping_probabilistic_1999].
The PCA model is defined as $$\begin{aligned}
p(\vxx) &= \gaussian(\vxx; 0, \id)\\
p(\vyy|\vxx) &= \gaussian(\vyy; \matW\vxx, \stdev^{2} \id)
\end{aligned}$$ where $\matW$ is a matrix that encodes the linear mapping between hidden components $\vxx \in \reals^{D}$ and data $\vyy \in \reals^{\MM}$, with $\MM > D$.
If we impose that each of the latent components $x_{d}$ is sinusoidal and may be parametrised by a hidden angle $\arv_{d}$ plus a phase shift $\varphi_{d}$, we obtain the model $$\begin{aligned}
p(\arv_{d}) &= \gvm(\arv; \cstd_{1, d}, \cstd_{2, d},
\cmean_{1, d}, \cmean_{2, d})\\
p(x_{d}|\arv_{d}) &= \delta(x_{d} - \sin(\arv_{d} + \varphi_{d}))\\
p(\vyy|\vxx) &= \gaussian(\vyy; \matW\vxx, \stdev^{2} \id)
\end{aligned}$$
To obtain the relation directly between the data and the hidden angle, we integrate out the latent components $\vxx$ $$p(\vyy|\varv) = \int \delta(\vxx - \sin(\varv + \vec{\varphi}))
\gaussian(\vyy; \matW\vxx, \stdev^{2} \id) d\vxx$$ which results in the model used in the mGvM dimensionality reduction application.
Alternatively, it is also possible to show the limiting behaviour of the model arising fromthe mGvM dimensionality reduction application becomes the PCA model, for mean angles $\vec{\cmean} \rightarrow 0$ and high concentration parameters. In this regime, the small angle approximation $$\sin\arv \approx \arv, \quad \cos\arv \approx 0$$ is valid and leads to the Generalised Von Mises priors simplification to $$\begin{aligned}
p(\arv) &\propto \exp\left\{
\cstd_{1}\cos(\arv - \cmean_{1})
+ \cstd_{2}\cos(2(\arv - \cmean_{2}))
\right\}\\
&\propto \exp\left\{
-\cstd_{1}\cos(\cmean_{2})\arv^{2}
+ (\cstd_{1}\sin(\cmean_{1})
+ 2 \cstd_{2}\sin(2\cmean_{2}))\arv
\right\}\\
&\propto \exp\left\{
-\cstd_{1}\cos(\cmean_{2})
\left[
\arv
- \frac{\cstd_{1}\sin(\cmean_{1})
+ 2 \cstd_{2}\sin(2\cmean_{2})}{2\cstd_{1}\cos(\cmean_{2})}
\right]^{2}
\right\}
\end{aligned}$$ which is proportional to a Gaussian distribution and shows that under the small angle regime, the coefficient matrix $\matA$ is a good approximation for $\matW$ and the model collapses to PCA.
Another connection between the dimensionality reduction with the mGvM and PCA may be established geometrically. While PCA describes the data in terms of hidden hyperplanes, the lower dimensional description of the data with mGvM occurs in terms of hidden tori, as illustrated in \[fig:doughnut\].
![Plots of the model $x = 2 \cos{\arv_{1}} + \error$, $y = 2 \sin{\arv_{1}} + 2 \cos{\arv_{2}} + \error$, $z = 2 \sin{\arv_{2}} + \error$ where $\arv_{1} \sim \vm(50, \pi/2)$ is a peaked von Mises distribution, $\arv_{2} \sim \vm(0.1, 0)$ is an almost-uniform von Mises distribution and the noise is $\error \sim \gaussian(0, 0.01)$ to exemplify a 3-dimensional Cartesian data set as a function of a 2-dimensional angular space: plot of samples from the model (left), samples on the $z = 0$ plane, which is equivalent to fixing $\arv_{2}=\pm \pi$ (middle), samples on the $x = 0$ plane, which is equivalent to fixing $\arv_{1}=\pm \pi/2$ (right).[]{data-label="fig:doughnut"}](donut){width="75.00000%"}
The effect of priors in this systems is also highlighted by \[fig:doughnut\]. The mean angle and concentration of each prior impacts the distribution of mass along the direction of the angular component on the hyper-torus. High concentration values on the prior leads to dense regions around the mean angle, as presented in the middle graph of \[fig:doughnut\] while low concentration leads to uniform mass distribution, shown in the right graph of \[fig:doughnut\].
An analogy often used to describe this shape of the data in the PCA’s hidden space is a “fuzzy pancake”, as the Gaussian noise induces the shape irregularity (“fuzzyness”), of the hidden plane (“pancake”). Likewise, for dimensionality reduction with the mGvM the corresponding analogy would be a “fuzzy doughnut”, as the Gaussian noise also incur in irregularities over the surface of a “doughnut”, which bears similar shape to a torus.
Supporting graphs and analysis for the experiments
--------------------------------------------------
### Regression experiments
The plots in \[fig:1d\_regression,fig:tides\] help us understand the some reasons why the mGvM provides better regression performance than the other models considered. The mGvM is able to accurately infer where the underlying function wraps, and provides a reasonable estimate for both the expected value of the underlying function and variance on the unit circle. The 1D-GP cannot account for the angular equivalences, therefore, it has assign this phenomenon to noise resulting in flat predictions as shown in \[fig:1d\_regression\]. While the 2D-GP is able to cope with wrapping, it learns a different lengthscale parameter. Furthermore, the 2D-GP cannot learn bimodal errors which can be accounted for by the mGvM as shown in \[fig:tides\].
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
---------- -- --
\[16mm\]
---------- -- --
### Dimensionality reduction experiments
In this section, we provide additional experiments and noise values for the average signal-to-noise ratio for motion capture and the simulation of motion capture of a robot arm. In the motion capture data sets, we applied a colour filter to the resulting images to isolate each marker and then the marker position was found by calculating the centre of mass of each marker as shown in Figure \[fig:mocap\_xp\].
0.2in
![Capturing 2D motion: the datasets was generated by recording the motion of a subject with markers on its body then using a colour threshold algorithm and taking the location of the centre of mass of the filtered region.[]{data-label="fig:mocap_xp"}](mocap4.png){width="0.9\columnwidth"}
-0.2in
Additional experiments are given in \[fig:mocap\_results\]. The conclusions and discussion of these experimental results mirror the discussions presented in the main paper.
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
[^1]: To be precise, Gatto and Jammalamadaka define this to be a Generalised von Mises of order 2, but since higher-order Generalised von Mises distributions are more intractable and consequently have found fewer applications, we use the shorthand throughout.
[^2]: <https://github.com/fivethirtyeight/uber-tlc-foil-response>
[^3]: <http://www.ukho.gov.uk/Easytide/easytide/SelectPort.aspx>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Our planet has one permanently bound satellite –the Moon–, a likely large number of mini-moons or transient irregular natural satellites, and three temporary natural retrograde satellites or quasi-satellites. These quasi-moons –(164207) 2004 GU$_{9}$, (277810) 2006 FV$_{35}$ and 2013 LX$_{28}$– are unbound companions to the Earth. The orbital evolution of quasi-satellites may transform them into temporarily bound satellites of our planet. Here, we study the dynamical evolution of the recently discovered Aten asteroid 2014 OL$_{339}$ to show that it is currently following a quasi-satellite orbit with respect to the Earth. This episode started at least about 775 yr ago and it will end 165 yr from now. The orbit of this object is quite chaotic and together with 164207 are the most unstable of the known Earth quasi-satellites. This group of minor bodies is, dynamically speaking, very heterogeneous but three of them exhibit Kozai-like dynamics: the argument of perihelion of 164207 oscillates around -90, the one of 277810 librates around 180and that of 2013 LX$_{28}$ remains around 0. Asteroid 2014 OL$_{339}$ is not currently engaged in any Kozai-like dynamics.'
date: 'Accepted 2014 September 19. Received 2014 September 15; in original form 2014 September 5'
title: 'Asteroid 2014 OL$_{339}$: yet another Earth quasi-satellite'
---
celestial mechanics – minor planets, asteroids: individual: 2004 GU$_{9}$ – minor planets, asteroids: individual: 2006 FV$_{35}$ – minor planets, asteroids: individual: 2013 LX$_{28}$ – minor planets, asteroids: individual: 2014 OL$_{339}$ – planets and satellites: individual: Earth.
Introduction
============
The term “quasi-satellite” was first used in a scientific publication by Danielsson & Ip (1972) while trying to explain the resonant behaviour of the near-Earth Object (NEO) 1685 Toro (1948 OA). However, this early mention was not directly connected with its current use. It is now generally accepted that the term was first introduced and popularized among the scientific community by Mikkola & Innanen (1997), although the concept behind it was initially studied by Jackson (1913) and the energy balance associated with the resonant state was first explored by Hénon (1969), who coined the term “retrograde” satellites to refer to them. Further analyses were carried out by Szebehely (1967), Broucke (1968), Benest (1976, 1977), Dermott & Murray (1981), Kogan (1989) and Lidov & Vashkov’yak (1993, 1994a,b). Most of this early work was completed within the framework of the restricted elliptic three-body problem. The quasi-satellite dynamical state is a specific configuration of the 1:1 mean motion resonance with a host planet in which the object involved appears to travel around the planet but is not gravitationally bound to it, i.e. the body librates around the longitude of its associated planet but its trajectory is not closed.
The first minor body to be confirmed to pursue a quasi-satellite orbit was 2002 VE$_{68}$ that is companion to Venus (Mikkola et al. 2004). Objects in this dynamical state have been found following Ceres and Vesta (Christou 2000b; Christou & Wiegert 2012), Jupiter (Kinoshita & Nakai 2007; Wajer & Królikowska 2012), Saturn (Gallardo 2006), Neptune (de la Fuente Marcos & de la Fuente Marcos 2012c) and Pluto (de la Fuente Marcos & de la Fuente Marcos 2012a). So far, Jupiter has the largest number of documented quasi-satellites with at least six, including asteroids and comets (Wajer & Królikowska 2012). Our planet comes in second place with three detected quasi-satellite companions: (164207) 2004 GU$_{9}$ (Connors et al. 2004; Mikkola et al. 2006; Wajer 2010), (277810) 2006 FV$_{35}$ (Wiegert et al. 2008; Wajer 2010) and 2013 LX$_{28}$ (Connors 2014). As such, these objects are not real, gravitationally bound satellites but, from Earth’s point of view, they appear to travel in the retrograde direction around it over the course of a year although they actually orbit (in the prograde direction) the Sun. Large amounts of interplanetary dust particles are also temporarily trapped in Earth’s quasi-satellite resonance (Kortenkamp 2013) and our planet hosts a small population of transient irregular natural satellites or mini-moons that may have been quasi-satellites before becoming temporarily bound to the Earth (Granvik, Vaubaillon & Jedicke 2012; Bolin et al. 2014)
Here, we show that the recently discovered asteroid 2014 OL$_{339}$ is a quasi-satellite companion to the Earth. The object was originally selected as a co-orbital candidate because of its small relative semimajor axis, $|a - a_{\rm Earth}| \sim$ 0.0002 au; $N$-body calculations are used to confirm its current quasi-satellite engagement with our planet. This paper is organized as follows. In Section 2, we briefly outline our numerical model. Section 3 focuses on 2014 OL$_{339}$. Section 4 reviews the current dynamical status of 164207, 277810 and 2013 LX$_{28}$, using their latest orbital solutions. Section 5 provides a comparative dynamical analysis between 2014 OL$_{339}$ and the other three Earth quasi-satellites. Our conclusions are summarized in Section 6.
Numerical model
===============
Here, we use $N$-body calculations to study the librational properties of the principal resonant angle of 2014 OL$_{339}$ with the Earth in order to understand its current dynamical status. As an Earth co-orbital candidate, the key object of study is the oscillation of the difference between the mean longitudes of the object and the Earth or relative mean longitude, $\lambda_{\rm r}$. The mean longitude of an object is given by $\lambda$ = $M$ + $\Omega$ + $\omega$, where $M$ is the mean anomaly, $\Omega$ is the longitude of the ascending node and $\omega$ is the argument of perihelion (see e.g. Murray & Dermott 1999). An object is co-orbital to the Earth if $\lambda_{\rm r}$ oscillates (librates) around a constant value; if $\lambda_{\rm r}$ can take any value (circulates), then we have a passing object. If $\lambda_{\rm r}$ librates around 0$^{\circ}$, we have the quasi-satellite state; the minor planet orbits the Sun in an approximate ellipse with the same (mean) period as the Earth. However, when viewed in a frame of reference that corotates with the Earth, the quasi-satellite follows a retrograde path around our planet over the course of an orbital period, the sidereal year. In principle, such motion is stabilized by the host planet. The stability of quasi-satellite orbits has been studied by Mikkola et al. (2006) and Sidorenko et al. (2014).
The numerical simulations presented here were completed using a Hermite integration scheme (Makino 1991; Aarseth 2003). The standard version of this direct $N$-body code is publicly available from the IoA web site[^1]. Our model Solar system includes the perturbations by the eight major planets and treats the Earth and the Moon as two separate objects, it also incorporates the barycentre of the dwarf planet Pluto–Charon system and the ten most massive asteroids of the main belt, namely, (1) Ceres, (2) Pallas, (4) Vesta, (10) Hygiea, (31) Euphrosyne, 704 Interamnia (1910 KU), 511 Davida (1903 LU), 532 Herculina (1904 NY), (15) Eunomia and (3) Juno. Relative errors in the total energy at the end of the simulations are $< 1 \times 10^{-15}$. The equivalent error in the total angular momentum is several orders of magnitude smaller. Additional details can be found in de la Fuente Marcos & de la Fuente Marcos (2012b) which also discusses 2002 VE$_{68}$, the first documented quasi-satellite.
0.12truecm
2014 OL$_{339}$ 2004 GU$_{9}$ 2006 FV$_{35}$ 2013 LX$_{28}$
-------------------------------------------------------- --- -------------------------- ------------------------------- ----------------------------- ---------------------------
Semimajor axis, $a$ (au) = 0.999 388$\pm$0.000 007 1.001 268128$\pm$0.000 000003 1.001 27736$\pm$0.000 00002 1.001 5884$\pm$0.000 0012
Eccentricity, $e$ = 0.460 67 $\pm$0.000 03 0.136 2649$\pm$0.000 0006 0.377 531$\pm$0.000 005 0.452 052$\pm$0.000 011
Inclination, $i$ ($^{\circ}$) = 10.190 65 $\pm$0.000 5 13.648 35$\pm$0.000 05 7.101 62$\pm$0.000 13 49.976 1$\pm$0.000 3
Longitude of the ascending node, $\Omega$ ($^{\circ}$) = 252.223 2$\pm$0.001 1 38.675 83$\pm$0.000 03 179.541 89$\pm$0.000 08 76.681 00$\pm$0.000 02
Argument of perihelion, $\omega$ ($^{\circ}$) = 289.656 4$\pm$0.000 5 280.332 87$\pm$0.000 05 170.845 8$\pm$0.000 2 345.781 8$\pm$0.000 2
Mean anomaly, $M$ ($^{\circ}$) = 215.718 $\pm$0.004 121.353 65$\pm$0.000 05 102.135 8$\pm$0.000 3 28.143 7$\pm$0.000 5
Perihelion, $q$ (au) = 0.539 00 $\pm$0.000 03 0.864 8304$\pm$0.000 0006 0.623 264$\pm$0.000 005 0.548 8184$\pm$0.000 0011
Aphelion, $Q$ (au) = 1.459 778 $\pm$0.000 010 1.137 705816$\pm$0.000 000004 1.379 29108$\pm$0.000 00003 1.454 3585$\pm$0.000 0002
Absolute magnitude, $H$ (mag) = 22.6 21.1 21.7 21.7
\[elements\]
Results in the figures have been obtained using initial conditions (positions and velocities referred to the barycentre of the Solar system) provided by the Jet Propulsion Laboratory (JPL) HORIZONS system (Giorgini et al. 1996; Standish 1998) and relative to the JD 2457000.5 epoch which is the $t$ = 0 instant. In addition to the calculations completed using the nominal orbital elements in Table \[elements\], we have performed 75 control simulations for each object with sets of orbital elements obtained from the nominal ones within the accepted uncertainties (up to 6$\sigma$) that reflect the observational incertitude in astrometry. In any case, the control orbits start very close to the nominal ones as the Gaussian errors are quite small (see Table \[elements\]). The computed set of control orbits follows a normal distribution in the six-dimensional orbital parameter space. The orbital evolution is computed in both directions of time at least for 30 kyr. Integration times are longer for the most dynamically stable objects. For clarity, the figures may display just a fraction of the total simulated time. Only a few representative orbits are displayed in the figures.
![The motion of 2014 OL$_{339}$ over the time range (-150, 150) yr is displayed projected onto the ecliptic plane in a coordinate system rotating with the Earth (nominal orbit in Table \[elements\]). The orbit and position of our planet are also indicated. All the investigated control orbits ($\pm6\sigma$) exhibit the same behaviour within this timeframe. []{data-label="qs"}](forbit.eps){width="\linewidth"}
Asteroid 2014 OL$_{339}$, an Aten quasi-satellite
=================================================
Asteroid 2014 OL$_{339}$ was serendipitously discovered by O. Vaduvescu, F. Char, V. Tudor, T. Mocnik, V. Dhillon and D. Sahman observing for EURONEAR (Vaduvescu et al. 2008) from La Palma on 2014 July 29 (Vaduvescu et al. 2014). The object was first detected using the 2.5 m Isaac Newton Telescope at an apparent $R$ magnitude of 21.8. The intended target of the programme was the Apollo asteroid 2013 VQ$_{4}$ but 2014 OL$_{339}$ was visible as a streak near the edge of the observed field. With a value of the semimajor axis, $a$, equal to 0.9994 au, very close to that of our planet (0.9992 au), this Aten asteroid is an NEO moving in an eccentric, $e$ = 0.46, and moderately inclined, $i = 10\fdg19$, orbit that makes it an Earth and Venus crosser, and a Mars grazer. Therefore, its orbit is different from those of the three previously known Earth quasi-satellites (see Table \[elements\]): (164207) 2004 GU$_{9}$, (277810) 2006 FV$_{35}$ and 2013 LX$_{28}$. It is an Aten, not an Apollo, and its eccentricity is the highest of the group which implies that it has the shortest perihelion and the farthest aphelion distances. The source of the Heliocentric Keplerian osculating orbital elements and uncertainties in Table \[elements\] is the JPL Small-Body Database.[^2]
Its very small relative semimajor axis, $|a - a_{\rm Earth}| \sim$ 0.000 197$\pm$0.000 007 au (the smallest found so far), makes this object a clear candidate to be an Earth co-orbital. It completes one orbit around the Sun in 364.92 d or 1.00 yr. Its current orbit is based on 27 observations with a data-arc span of 36 d. As expected of a recent discovery, the quality of the orbit of 2014 OL$_{339}$ is at present lower than that of the other minor bodies in Table \[elements\]. However, it is similar or even better than the one available when the other objects were recognized as unbound companions to our planet. Asteroid 2014 OL$_{339}$ has $H$ = 22.6 mag (assumed $G$ = 0.15) or a diameter of 90 to 200 m for an assumed albedo in the range 0.20–0.04. It is, therefore, smaller than the previously known Earth quasi-satellites (see Table \[elements\]).
The motion of 2014 OL$_{339}$ over the time range (-150, 150) yr as seen in a coordinate system rotating with the Earth projected onto the ecliptic plane is plotted in Fig. \[qs\] (nominal orbit in Table \[elements\]). This minor body is an Earth co-orbital currently following a quasi-satellite orbit around our planet (see Mikkola et al. 2006; Sidorenko et al. 2014). Due to its significant eccentricity and in accordance to theoretical predictions (Namouni, Christou & Murray 1999; Namouni & Murray 2000), the libration angle is rather large. The libration centre corresponds to our planet. Asteroid 2014 OL$_{339}$ appears to pursue a precessing kidney-shaped retrograde path when viewed from our planet over the course of a sidereal year. All the investigated control orbits ($\pm6\sigma$) exhibit the same behaviour within the timeframe mentioned above.
{width="\linewidth"}
{width="\linewidth"}
All the integrated control orbits for 2014 OL$_{339}$ exhibit quasi-satellite libration ($\lambda_{\rm r}$ oscillates around 0) with respect to the Earth at $t$ = 0; the object is a quasi-satellite to our planet at a confidence level $>$ 99.99 per cent (see Figs \[control\] and \[control2\]). This co-orbital episode started at least 775 yr ago and it will end 165 yr from now; the duration of the entire quasi-satellite resonance is, in average, approximately 1 kyr (and certainly less than 2.5 kyr), i.e. its current dynamical status is only temporary. The eviction at 165 yr from now coincides with a relatively distant close encounter with our planet at 0.13 au. Prior to the current quasi-satellite episode, the object was probably also co-orbital with our planet, an L$_4$ or L$_5$ Trojan ($\sim$70 per cent) or a horseshoe ($\sim$20 per cent) or, perhaps, a passing object ($\sim$10 per cent) but still in the immediate neighbourhood of Earth’s co-orbital region. After leaving its current state, it may become an L$_5$ Trojan ($\sim$10 per cent) or, more likely, a horseshoe librator ($\sim$90 per cent). Due to its significant eccentricity and in accordance to theoretical predictions (Namouni, Christou & Murray 1999; Namouni & Murray 2000), the libration angle as Trojan is greater than the usual value of $\pm$60, i.e. the libration centre is displaced from the typical equilateral location.
The overall evolution of all the control orbits within the time interval (-775, 165) yr is virtually identical but beyond those time boundaries, the past and future orbital evolution of this object becomes difficult to predict although it remains in the neighbourhood of Earth’s co-orbital region for thousands of years. As an example, Fig. \[control\] displays the short-term dynamical evolution of an orbit arbitrarily close to the nominal one (central panels) and those of two representative worst orbits which are different from the nominal one. The orbit labelled as ‘-3$\sigma$’ (left-hand panels) has been obtained by subtracting thrice the uncertainty from the orbital parameters (the six elements) in Table \[elements\]. It has the lowest values of $a$, $e$ and $i$ at the 3$\sigma$ level. In contrast, the orbit labelled as ‘+3$\sigma$’ (right-hand panels) was computed by adding three times the value of the uncertainty to the orbital elements in Table \[elements\]. This trajectory has the largest values of $a$, $e$ and $i$ (within 3$\sigma$). Asteroid 2014 OL$_{339}$ was considerably more stable in the past. It may remain as a co-orbital to our planet switching between the various co-orbital states for many kyr. The values of its semi-major axis (C-panels), eccentricity (D-panels) and inclination (E-panels) remain fairly constant during the entire co-orbital evolution and the object stays well beyond the Hill sphere of our planet (A-panels). The value of its argument of perihelion circulates (F-panels). The results of our calculations show that the true phase-space trajectory followed by this object will diverge exponentially from that obtained from the nominal orbital elements in Table \[elements\] within a relatively short time-scale; its e-folding time is of the order of 1 kyr. An additional test for consistency is given in Fig. \[control2\] where the orbital elements have been further modified at the $\pm6\sigma$ level. The short-term dynamical evolution is still consistent with that in Fig. \[control\] although the object was not co-orbital with our planet a few thousand years into the past. We can certainly state that the probability of this object being a currently active quasi-satellite of our planet is 0.9999966.
Earth quasi-satellites: a review
================================
The subject of currently active Earth quasi-satellites has not been revisited recently even if the orbits of those objects recognized as such have been significantly improved in recent times. Here, we provide a brief review of the current dynamical status of (164207) 2004 GU$_{9}$, (277810) 2006 FV$_{35}$ and 2013 LX$_{28}$, using their latest orbital solutions (see Table \[elements\]).
(164207) 2004 GU$_{9}$
----------------------
Asteroid 164207 was discovered by M. Blythe, F. Shelly, M. Bezpalko, R. Huber, L. Manguso, D. Torres, R. Kracke, M. McCleary, H. Stange and A. Milner observing for the Lincoln Near-Earth Asteroid Research (LINEAR) project from Socorro, New Mexico, on 2004 April 13 (Kornos et al. 2004) with the 1.0 m LINEAR telescope. At discovery time its apparent magnitude was 19.6. The orbit of this Potentially Hazardous Asteroid (PHA) of the Apollo class has a value of the semimajor axis $a$ = 1.0013 au. Its orbital eccentricity and inclination are moderate, $e$ = 0.14 and $i$ = 136. With such an orbit, 164207 always remains in the neighbourhood of the orbit of the Earth–Moon system; no close encounters with other inner planets are possible (see Table \[elements\]). Its current orbit is based on 175 observations with a data-arc span of 4718 d. Besides its orbit, little else is known about 164207: its absolute magnitude has a value of 21.1 mag and its albedo is 0.219 with a diameter of 163 m (Mainzer et al. 2011).
{width="\linewidth"}
Asteroid 164207 was recognized as a relatively long-lived quasi-satellite companion to the Earth by Connors et al. (2004) and its dynamics was further studied by Mikkola et al. (2006) and Wajer (2010). With the orbit available at the time, these studies concluded that the minor body would remain a quasi-satellite of our planet for several hundred years. Prior to its current dynamical state, 164207 had been a horseshoe librator to the Earth for many thousands of years, its $\lambda_{\rm r}$ oscillating around 180. Using the latest orbital solution, all the integrated control orbits for 164207 (within 6$\sigma$) exhibit quasi-satellite libration with respect to the Earth at $t$ = 0 (see Fig. \[gu9\]). The historical and future evolution of all the control orbits computed coincide in painting an evolutionary track dotted by multiple quasi-satellite resonant episodes of relatively short-duration, just a few kyr or less (see B-panels, Fig. \[gu9\]). Most of the time, the object is a horseshoe librator to our planet. Transitions between the two resonant states are not triggered by particularly close encounters with the Earth–Moon system but by the persistent action of other mean motion resonances. Asteroid 164207 orbits the Sun in a near 13:8 resonance with Venus so this planet completes 13 orbits around the Sun in the same amount of time the asteroid completes 8. This fact was already pointed out by Wajer (2010). The timings of the transitions depend strongly on the initial conditions. The orbit of this object cannot be predicted with enough certainty beyond a few thousand years. Its present co-orbital episode started about 450 yr ago and it will end nearly 570 yr from now; the duration of the entire quasi-satellite resonance is, in average, nearly 1 kyr with very little dispersion, i.e. like 2014 OL$_{339}$ its current dynamical status is only temporary. Prior to its current engagement as quasi-satellite, this minor body was very probably a horseshoe ($\sim$100 per cent). After leaving its current state, it will return to be a horseshoe librator ($\sim$100 per cent). For this object, horseshoe episodes last in average 4 to 6 kyr. These results are consistent with those from previous studies.
(277810) 2006 FV$_{35}$
-----------------------
Asteroid 277810 was discovered by J. V. Scotti observing from the Steward Observatory at Kitt Peak for the Spacewatch project on 2006 March 29 (Gilmore et al. 2006). The object was detected using a 0.9 m telescope at an apparent magnitude of 21.0. With a value of the semimajor axis $a$ = 1.0013 au, this Apollo asteroid is an NEO moving in an eccentric, $e$ = 0.38, and slightly inclined, $i$ = 710, orbit that makes it cross the orbits of Venus and the Earth–Moon system (see Table \[elements\]). Its current orbit is based on 94 observations with a data-arc span of 6931 d. Although the object has been observed for almost two decades (the first known pre-discovery observations were made on 1995 April 1), little else besides the orbit is known about 277810; its absolute magnitude, $H$ = 21.7 (assumed $G$ = 0.15), indicates a diameter in the range 130–300 m for an assumed albedo in the range 0.20–0.04.
{width="\linewidth"}
Asteroid 277810 was first reported to be a quasi-satellite of our planet by Wiegert et al. (2008). Its dynamics was further studied by Wajer (2010) who found that it will remain in its present quasi-satellite state for more than 10 kyr. Our calculations (see Fig. \[fv35\]) confirm that 277810 is experiencing at present a quasi-satellite resonant episode. The object is currently far more stable than 164207. In contrast with the previous object, 277810 rarely follows a horseshoe path and Trojan episodes are far more common. In Fig. \[fv35\], B-panels, we observe that the relative mean longitude can librate around 60$^{\circ}$, then the object is called an $L_4$ Trojan, or around -60(or 300), then it is an $L_5$ Trojan. In this case, the timings of the transitions coincide with relatively distant –beyond the Hill radius of our planet (0.0098 au)– close encounters with the Earth–Moon system. Its current co-orbital episode started at least 8 kyr ago and it will end about 3 kyr from now; the duration of the entire quasi-satellite resonance is, in average, approximately 18 kyr (and certainly less than 22 kyr), i.e. its current dynamical status is still temporary. Prior to the current quasi-satellite episode, the object was probably also co-orbital with our planet, a horseshoe ($\sim$90 per cent) or, perhaps, a passing object ($\sim$10 per cent) but still very close to Earth’s co-orbital region. After leaving its current state, it may become a passing object ($\sim$10 per cent) or, more likely, an $L_5$ Trojan ($\sim$90 per cent).
2013 LX$_{28}$
--------------
Asteroid 2013 LX$_{28}$ was discovered by N. Primak, A. Schultz, S. Watters and T. Goggia observing for the Pan-STARRS 1 project from Haleakala on 2013 June 12 (Bressi et al. 2013). The object was first observed using a 1.8 m Ritchey-Chretien telescope at an apparent magnitude of 20.7. With a value of the semimajor axis $a$ = 1.0016 au, this Apollo asteroid is an NEO moving in a rather eccentric, $e$ = 0.45, and highly inclined, $i$ = 50.0$^{\circ}$, orbit that makes it cross the orbits of Venus and the Earth–Moon system, grazing that of Mars and almost that of Mercury. Its current orbit is based on 26 observations with a data-arc span of 349 d. As a recent discovery, little else besides its orbit is known about this object; its absolute magnitude, $H$ = 21.7 (assumed $G$ = 0.15), suggests a diameter in the range 130–300 m for an assumed albedo in the range 0.20–0.04. The object was proposed as a Kozai-resonating Earth quasi-satellite by Connors (2014), who pointed out its remarkable stability.
{width="\linewidth"}
Once an object is trapped in a 1:1 mean motion resonance and depending on its relative energy with respect to the host planet (Hénon 1969), it can describe any of the three main orbit types: quasi-satellite, tadpole or horseshoe. Compound states are also possible in which the object may librate around 0$^{\circ}$ with an amplitude $>$ 180$^{\circ}$ encompassing L$_4$ and L$_5$ (compound quasi-satellite-tadpole orbit), asymmetric horseshoe orbits (horseshoe-quasi-satellite orbiters) in which the libration amplitude $>$ 270$^{\circ}$, encompassing the planet, and a few other combinations (see, e.g. Namouni 1999; Namouni et al. 1999). These are typical of objects moving in high-eccentricity, high-inclination orbits and this is what is observed in Fig. \[lx28\]. Although 2013 LX$_{28}$ is neither a quasi-satellite (see Fig. \[lx28\], B-panels) nor a Kozai-resonating body (see Fig. \[lx28\], F-panels) in strict sense; $\lambda_{\rm r}$ librates around 0with amplitude $>$ 120and $\omega$ does not librate (or, at least, does not complete a Kozai cycle) just remains relatively close to 0during the entire compound quasi-satellite-tadpole episode. When 2013 LX$_{28}$ leaves the 1:1 mean motion resonance or before entering it, its argument of perihelion is no longer close to 0. This behaviour is fully consistent across the set of simulations. Its current co-orbital episode with our planet began at least 5.5 kyr ago and it will end about 16–30 kyr from now; the duration of the entire compound quasi-satellite resonance is, in average, approximately 35 kyr (and certainly less than 45 kyr), i.e. its current dynamical status is also temporary. However, its compound resonant state changes a few times during that timeframe, the libration amplitude varies although $\lambda_{\rm r}$ still librates around 0. Prior to the current quasi-satellite episode, the object was probably also co-orbital with our planet, an $L_4$ Trojan ($\sim$50 per cent) or, perhaps, a passing object ($\sim$50 per cent) but still very close to Earth’s co-orbital region. After leaving its current state, it may become a passing object ($\sim$50 per cent) or an $L_5$ Trojan ($\sim$50 per cent). These Trojan episodes last nearly 10 kyr and are rather asymmetric due to the high-eccentricity, high-inclination orbit.
Comparative dynamical evolution of known Earth quasi-satellites
===============================================================
Figure \[all\] displays the comparative evolution of the osculating orbital elements and other parameters of interest of all the known Earth quasi-satellites (nominal orbits in Table \[elements\]). It is clear that this group of objects is, dynamically speaking, very heterogeneous. In particular, three objects exhibit Kozai-like dynamics (see F-panels), see Kozai (1962) and Namouni (1999) for technical details: the argument of perihelion of (164207) 2004 GU$_{9}$ oscillates around -90, the one of (277810) 2006 FV$_{35}$ librates around 180, and that of 2013 LX$_{28}$ remains around 0. The argument of perihelion of 2014 OL$_{339}$ circulates. Some Venus co-orbitals (see e.g. de la Fuente Marcos & de la Fuente Marcos 2013a; de la Fuente Marcos & de la Fuente Marcos 2014) also exhibit Kozai-like dynamics (see fig. 4, F-panels, in de la Fuente Marcos & de la Fuente Marcos 2014). In particular, the value of the argument of perihelion of 2002 VE$_{68}$ (also a quasi-satellite) remains close to zero during its entire quasi-satellite evolution. For eccentric co-orbitals, this type of resonance provides a temporary effective protection mechanism against close encounters with the host planet: the Earth for 2013 LX$_{28}$ and Venus for 2002 VE$_{68}$. In this case, the nodes are located at perihelion and at aphelion, i.e. away from the host planet (see e.g. Milani et al. 1989).
Asteroid 2014 OL$_{339}$ is an Aten, the other three confirmed quasi-satellites are Apollos although the reason for the absence of the Kozai resonance in the case of 2014 OL$_{339}$ is not this but its relatively large eccentricity. For an object following an inclined path, close encounters with major planets are only possible in the vicinity of the nodes. The distance between the Sun and the nodes is given by $r = a (1 - e^2) / (1 \pm e \cos \omega)$, where the ‘+’ sign is for the ascending node and the ‘-’ sign is for the descending node. The positions of the nodes are plotted in the G-panels of Fig. \[all\]. The descending node of 2014 OL$_{339}$ is close to the orbit of the Earth, its ascending node is near Venus. In contrast, both nodes of 164207 are currently near the Earth, the ascending node of 277810 is perturbed by Mars and the descending one is relatively free from perturbations by Venus; the descending node of 2013 LX$_{28}$ is perturbed by Mars and its ascending one is also relatively free from perturbations by Venus. The Kozai resonance is effective in protecting the paths of 277810 and 2013 LX$_{28}$ against close encounters with the Earth–Moon system as their nodes are away from it, stabilizing their orbits but makes the orbit of 164207 rather unstable. Here, the libration occurs at $\omega$ = -90, not 0or 180. Under these circumstances, aphelion and perihelion always occur away from the ecliptic plane. A common feature of the orbital evolutions of 164207 and 2014 OL$_{339}$ is in their enhanced instability when compared to the other two. This translates into relatively frequent episodes in which we observe switching between resonant states. Transfers between tadpole, horseshoe and quasi-satellite orbits are triggered by close encounters with the inner planets and those are the result of the libration of the nodes (Wiegert, Innanen & Mikkola 1998). Asteroids 277810 and 2013 LX$_{28}$ do not exhibit Kozai-like dynamics outside the timeframe in which they are quasi-satellites.
Although there are no two Earth quasi-satellites alike, the closest dynamical relative to 2014 OL$_{339}$ is 164207. It also stays as an Earth quasi-satellite for about 1 kyr (see Fig. \[all\], second column, panel G) which is consistent with previous results presented by Wajer (2010). It was a horseshoe librator prior to its capture as quasi-satellite and it will return to that resonant state after its eviction. Figure \[all\] shows that only the Earth-Moon system plays a significant role in destabilizing its orbit; contrary to previous results in Wajer (2010), Venus does not appear to play a significant role in the current dynamical evolution of this object. The orbits of 164207 and 2014 OL$_{339}$ can only be accurately calculated for a few hundred years forward and backward in time. In sharp contrast, 277810 remains in the quasi-satellite state for a long period of time. Our calculations agree reasonably well with those of Wajer (2010), the object has remained in its current state for more than 15 kyr and it will remain there for a few thousand more years. Discrepancies with Wajer (2010) could be the result of using updated orbits and different physical models. Chaotic orbits are not only sensitive to changes in the initial conditions but also to different dynamical models. Although Connors (2014) classifies 2013 LX$_{28}$ as quasi-satellite, this is incorrect in strict sense because its orbit is hybrid. It is a persistent co-orbital companion to the Earth that follows a compound quasi-satellite-tadpole orbit that encloses Earth’s Lagrangian points L$_5$ and L$_4$, as well as the Earth itself (see e.g. Namouni 1999; Namouni et al. 1999). In principle, close encounters are possible with Mercury, Venus, the Earth and Mars but the asteroid is temporarily protected against close approaches by a Kozai-like resonance with Jupiter. Its dynamics is somewhat similar to that of the well studied Apollo asteroid 10563 Izhdubar (1993 WD) although the argument of perihelion of this object librates around 90(see Christou 2000a) not 0. Even if not strictly a quasi-satellite, 2013 LX$_{28}$ is the most stable of the group with a most probable duration of its current state in the range of 35 to 45 kyr.
{width="\linewidth"}
Conclusions
===========
In this paper, we have identified yet another Earth quasi-satellite, 2014 OL$_{339}$. Its dynamical status is temporary and it is not expected to last more than 1 to 2 kyr as this object is one of the most unstable known Earth quasi-satellites; its e-folding time is $\sim$1 kyr. In the Solar system and among the terrestrial planets, the Earth has the largest number of detected quasi-satellites with four; this is likely to be the result of observational bias, though. A comparative analysis of the short-term dynamical evolution of these objects shows that they are, dynamically speaking, very heterogeneous although three objects exhibit Kozai-like dynamics. The identification of Kozai librators among members of the NEO population is not new (see e.g. Michel & Thomas 1996; de la Fuente Marcos & de la Fuente Marcos 2013b). This indicates that the Kozai resonance plays a significant role in the orbital evolution of many Earth quasi-satellites and also in the chain of events that drives them into this particular resonance and away from it. In this context, 2014 OL$_{339}$ is an outlier as it is the only currently known Earth quasi-satellite not to be submitted to a Kozai resonance. Given their current orbits, none of the objects discussed here may impact our planet within the next few hundred –(164207) 2004 GU$_{9}$ and 2014 OL$_{339}$– or even several thousand –(277810) 2006 FV$_{35}$ and 2013 LX$_{28}$– years. Although these four objects are currently experiencing quasi-satellite episodes within the 1:1 mean motion resonance with the Earth, their dynamical contexts are quite different hinting at a richer picture of the quasi-satellite state than conventionally portrayed, with multiple pathways to the same resonant phase. The diverse dynamical histories found for the members of this group make a common origin for any pair of them rather unlikely.
In this work, relativistic terms and the role of the Yarkovsky and Yarkovsky–O’Keefe–Radzievskii–Paddack (YORP) effects (see e.g. Bottke et al. 2006) have been ignored. The non-inclusion of these effects has no impact on the evaluation of the present dynamical status of the minor bodies studied here but may affect predictions regarding their future evolution and dynamical history. In particular, the Yarkovsky effect may have a role on the medium- and long-term evolution of objects as small as the minor bodies discussed here. Proper modelling of the Yarkovsky force requires knowledge on the physical properties of the objects involved (for example, rotation rate, albedo, bulk density, surface conductivity, emissivity) which is not the case for these minor bodies. Perturbational effects arising from the co-orbital evolution with our planet may render these non-gravitational effects negligible, though.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank the referee for his/her prompt, to-the-point report and S. J. Aarseth for providing the code used in this research. This work was partially supported by the Spanish ‘Comunidad de Madrid’ under grant CAM S2009/ESP-1496. We thank M. J. Fernández-Figueroa, M. Rego Fernández and the Department of Astrophysics of the Universidad Complutense de Madrid (UCM) for providing computing facilities. Part of the calculations and the data analysis were completed on the ‘Servidor Central de Cálculo’ of the UCM and we thank S. Cano Alsúa for his help during this stage. In preparation of this paper, we made use of the NASA Astrophysics Data System, the ASTRO-PH e-print server, the MPC data server and the NEODyS information service.
Aarseth S. J., 2003, Gravitational $N$-body Simulations. Cambridge Univ. Press, Cambridge, p. 27 Benest D., 1976, Celest. Mech., 13, 203 Benest D., 1977, A&A, 54, 563 Bolin B. et al., 2014, Icarus, 241, 280 Bottke W. F., Jr., Vokrouhlický D., Rubincam D. P., Nesvorný D., 2006, Annu. Rev. Earth Planet. Sci., 34, 157 Bressi T. H. et al., 2013, MPEC Circ., MPEC 2013-L72 Broucke R. A., 1968, Technical Report 32-1168, Periodic Orbits in the Restricted Three-Body Problem with Earth-Moon Masses. Jet Propulsion Laboratory, California Inst. Technol., Pasadena, CA, p. 32 Christou A. A., 2000a, Icarus, 144, 1 Christou A. A., 2000b, A&A, 356, L71 Christou A. A., Wiegert P., 2012, Icarus, 217, 27 Connors M., 2014, MNRAS, 437, L85 Connors M., Veillet C., Brasser R., Wiegert P., Chodas P., Mikkola S., Innanen K., 2004, Meteoritics Planet. Sci., 39, 1251 Danielsson L., Ip W.-H., 1972, in Evlius A., ed, From Plasma to Planet. Wiley Interscience Division, New York, p. 35 de la Fuente Marcos C., de la Fuente Marcos R., 2012a, MNRAS, 427, L85 de la Fuente Marcos C., de la Fuente Marcos R., 2012b, MNRAS, 427, 728 de la Fuente Marcos C., de la Fuente Marcos R., 2012c, A&A, 545, L9 de la Fuente Marcos C., de la Fuente Marcos R., 2013a, MNRAS, 432, 886 de la Fuente Marcos C., de la Fuente Marcos R., 2013b, MNRAS, 434, L1 de la Fuente Marcos C., de la Fuente Marcos R., 2014, MNRAS, 439, 2970 Dermott S. F., Murray C. D., 1981, Icarus, 48, 1 Gallardo T., 2006, Icarus, 184, 29 Gilmore A. C., Scotti J. V., Hug G., Spahr T. B., 2006, MPEC Circ., MPEC 2006-F58 Giorgini J. D. et al., 1996, BAAS, 28, 1158 Granvik M., Vaubaillon J., Jedicke R., 2012, Icarus, 218, 262 Hénon M., 1969, A&A, 1, 223 Jackson J., 1913, MNRAS, 74, 62 Kinoshita H., Nakai H., 2007, Celest. Mech. Dyn. Astron., 98, 181 Kogan A. Y., 1989, Cosm. Res., 24, 705 Kornos L. et al., 2004, MPEC Circ., MPEC 2004-G31 Kortenkamp S. J., 2013, Icarus, 226, 1550 Kozai Y., 1962, AJ, 67, 591 Lidov M. L., Vashkov’yak M. A., 1993, Cosm. Res., 31, 187 Lidov M. L., Vashkov’yak M. A., 1994a, Astron. Lett., 20, 188 Lidov M. L., Vashkov’yak M. A., 1994b, Astron. Lett., 20, 676 Mainzer A. et al., 2011, ApJ, 743, 156 Makino J., 1991, ApJ, 369, 200 Michel P., Thomas F., 1996, A&A, 307, 310 Mikkola S., Innanen K., 1997, in Dvorak R., Henrard J., eds, The Dynamical Behaviour of our Planetary System. Kluwer, Dordrecht, p. 345 Mikkola S., Brasser R., Wiegert P., Innanen K., 2004, MNRAS, 351, L63 Mikkola S., Innanen K., Wiegert P., Connors M., Brasser R., 2006, MNRAS, 369, 15 Milani A., Carpino M., Hahn G., Nobili A. M., 1989, Icarus, 78, 212 Murray C. D., Dermott S. F., 1999, Solar system Dynamics. Cambridge Univ. Press, Cambridge, p. 97 Namouni F., 1999, Icarus, 137, 293 Namouni F., Murray C. D., 2000, Celest. Mech. Dyn. Astron., 76, 131 Namouni F., Christou A. A., Murray C. D., 1999, Phys. Rev. Lett., 83, 2506 Sidorenko V. V., Neishtadt A. I., Artemyev A. V., Zelenyi L. M., 2014, Celest. Mech. Dyn. Astron., 120, 131 Standish E. M., 1998, JPL Planetary and Lunar Ephemerides, DE405/LE405. Interoffice Memo. 312.F-98-048, NASA JPL Szebehely V. G., 1967, Theory of Orbits. The Restricted Problem of Three Bodies, Academic Press, New York Vaduvescu O., Birlan M., Colas F., Sonka A., Nedelcu A., 2008, Planet. Space Sci., 56, 1913 Vaduvescu O. et al., 2014, MPEC Circ., MPEC 2014-P23 Wajer P., 2010, Icarus, 209, 488 Wajer P., Królikowska M., 2012, Acta Astronomica, 62, 113 Wiegert P. A., Innanen K. A., Mikkola S., 1998, AJ, 115, 2604 Wiegert P. A., DeBoer R., Brasser R., Connors M., 2008, J. R. Astron. Soc. Can., 102, 52
[^1]: http://www.ast.cam.ac.uk/$\sim$sverre/web/pages/nbody.htm
[^2]: http://ssd.jpl.nasa.gov/sbdb.cgi
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The theory of the electron-phonon interaction ($EPI$) with strong forward scattering peak ($FSP$) in an extreme delta-peak limit [@Kulic-Zeyher1-2]-[@Dan-Dol-Kul-Oud] is recently applied in [@Lee-Interfacial-supplement]-[@Johnston1-2] for the explanation of high $T_{c}(\sim100$ $K)$ in a monolayer $FeSe$ grown on $SrTiO_{3}$ [@Lee-Interfacial-supplement] and $TiO_{2}$ [@Rebec-FeSe-TiO2] substrates. The $EPI$ is due to a long-range dipolar electric field created by the high-energy oxygen vibrations ($\Omega\sim90$ $meV$) at the interface [@Lee-Interfacial-supplement]-[@Johnston1-2]. We show that in leading order (with respect to $T_{c0}/\Omega$) the mean-field critical temperature $T_{c0}=\left\langle V_{epi}(q)\right\rangle _{q}/4)$ $\sim$ $(aq_{c})^{2}V_{epi}(0)$ and the gap $\Delta_{0}=2T_{c0\text{ }}$ are due to an interplay between the maximal $EPI$ pairing potential $V_{epi}(0)$ and the $FSP$-width $q_{c}$. For $T_{c0}\sim100$ $K$ one has $\Delta_{0}\sim16$ $meV$ in a satisfactory agreement with $ARPES$ experiments. We find that in leading order $T_{c0}$ is * mass-independent* and a very small oxygen isotope effect is expected in next to leading order. In clean systems $T_{c0}$ for $s$-wave and $d$-wave pairing is degenerate but both are affected by non-magnetic impurities, which are *pair-weakening* in the $s$-channel and *pair-breaking* in the $d$-channel.
The self-energy and replica bands at $T=0$ and at the Fermi surface are calculated and compared with the corresponding results at $T>0$ [@Johnston1-2]. The $EPI$ coupling constant $\lambda_{m}=\left\langle
V_{epi}(q)\right\rangle _{q}/2\Omega$, which enters the self-energy $\Sigma(k,\omega)$, is *mass-dependent* ($M^{1/2}$) which at $\omega(\ll\Omega)$ makes the slope of $\ \Sigma(k,\omega)(\approx-\lambda
_{m}\omega)$ and the replica intensities $A_{i}(\sim\lambda_{m})$ *mass-dependent*. This result, overlooked in the literature, is contrary to the prediction of its mass-independence in the standard Migdal-Eliashberg theory for $EPI$. The small oxygen isotope effect in $T_{c0}$ and pronounced isotope effect in $\Sigma(k,\omega)$ and $ARPES$ spectra $A_{i}$ of replica bands in $FeSe$ films on $SrTiO_{3}$ and $TiO_{2}$ is a smoking-gun experiment for testing an applicability of the $EPI-FSP$ theory to these systems. The $EPI-FSP$ theory predicts a large number of low-laying pairing states (above the ground state) thus causing internal pair fluctuations. The latter reduce $T_{c0}$ additionally, by creating a pseudogap state for $T_{c}<T<T_{c0}$.
Possibilities to increase $T_{c0}$, by designing novel structures are discussed in the framework of the $EPI-FSP$ theory.
address: |
$^{1}$Institute for Theoretical Physics, Goethe-University D-60438 Frankfurt am Main, Germany\
$^{2}$Institute of Physics, Pregrevica 118, 11080 Belgrade (Zemun), Serbia\
$^{3}$Max-Planck-Institut für Festkörperphysik,70569 Stuttgart, Germany
author:
- '$^{1,2}$M. L. Kulić, $^{3}$O. V. Dolgov'
title: 'The electron-phonon interaction with forward scattering peak is a relevant approach to high $T_{c}$ superconductivity in $FeSe$ films on $SrTiO_{3}$ and $TiO_{2}$ '
---
Introduction
============
The scientific race in reaching high temperature superconductivity ($HTSC$) started by the famous Ginzburg’s proposal of an *excitonic mechanism of pairing* *in metallic-semiconducting sandwich-structures* [@Ginzburg-excitons]. In such a system an electron from the metal tunnels into the semiconducting material and virtually excites high-energy exciton, which is absorbed by another electron, thus making an effective attractive interaction and Cooper pairing. However, this beautiful idea has not been realized experimentally until now. In that sense V. L. Ginzburg founded a theoretical group of outstanding and talented physicists, who studied at that time almost all imaginable pairing mechanisms. In this group an important role has played the Ginzburg’s collaborator E. G. Maksimov, who was an “inveterate enemy” of almost all other mechanisms of pairing in $HTSC$ but for the electron-phonon one - see his arguments in [@Dolg-Kirzh-Maks]. It seems that the recent discovery of superconductivity in a $Fe$-based material made of one monolayer film of the iron-selenide $FeSe$ grown on the $SrTiO_{3}$ substrate - further called $1ML$ $FeSe/SrTiO_{3}$, with the critical temperature $T_{c}\sim(50-100)$ $K$ [@Wang; @QJ-1-Lee; @JJ], as well as grown on the rutile $TiO_{2}$ (100) substrate with $T_{c}\sim65$ $K$ [@Rebec-FeSe-TiO2] - further called $1ML$ $FeSe/TiO_{2}$, in some sense reconciles the credence of these two outstanding physicists. Namely, $HTSC$ is realized in a sandwich-structure but the pairing is due to an high-energy ($\sim90-100$ $meV$) *oxygen optical phonon*. This (experimental) discovery will certainly revive discussions on the role of the electron-phonon interaction ($EPI$) in $HTSC$ cuprates and in bulk materials of the $Fe$-pnictides (with the basic unit $Fe-As$) and Fe-chalcogenides (with the basic unit $Fe-Se$ or $Te$, $S$). As a digression, we point out that after the discovery of high $T_{c}$ in $Fe$-pnictides a non-phononic pairing mechanism was proposed immediately, which is due to: $(i$) nesting properties of the electron- and hole-Fermi surfaces and ($ii$) an enhanced (due to ($i$)) spin exchange interaction ($SFI$) between electrons and holes [@Hirschfeld-review]. This mechanism is called the *nesting SFI pairing*. However, the discovery of alkaline iron selenides $K_{x}Fe_{2-y}Se_{2}$ with $T_{c}\sim30$ $K$, and intercalated compounds $Li_{x}(C_{2}H_{8}N_{2})Fe_{2-y}Se_{2}$, $Li_{x}(NH_{2})_{y}(NH_{3})_{1-y}Fe_{2}Se_{2}$, which contain *only electron-like Fermi surfaces*, rules out the nesting pairing mechanism as a common pairing mechanism in Fe-based materials. In order to overcome this inadequacy of the $SFI$ nesting mechanism a pure phenomenological “strong coupling” $SFI$ pairing is proposed in the framework of the so called $J_{1}-J_{2}$ Heisenberg-like Hamiltonian, which may describe the $s$-wave superconductivity, too. However, this approach is questionable since the LDA calculations cannot be mapped onto a Heisenberg model and there is a need to introduce further terms in form of biquadratic exchange [@Hirschfeld-review]. It is interesting, that immediately after the discovery of high $T_{c}$ in pnictides the electron-phonon pairing mechanism was rather uncritically discarded. This attitude was exclusively based on the $LDA$ band structure calculations of the electron-phonon coupling constant [@Boeri-Dolgov], which in this approach turns out to be rather small $\lambda<0.2$, thus giving $T_{c}<1$ $K$.
In the past there were only few publications trying to argue that the $EPI$ pairing mechanism is an important (pairing) ingredient in the $Fe$-based superconductors [@Kulic-Dolgov-Drechsler]-[@Kulic-Ginzburg; @Conf-2012]. One of the theoretical arguments for it, may be ilustrated in the case of 2-band superconductivity. In the weak-coupling limit $T_{c}$ is given by $T_{c}=1.2\omega_{c}\exp\{-1/\lambda_{\max}\},$ where $\lambda_{\max}=(\lambda_{11}+\lambda_{22}+\sqrt{(\lambda_{11}-\lambda_{22})^{2}+4\lambda_{12}\lambda_{21}})/2$. In the nesting $SFI$ pairing mechanism one assumes a dominance of the repulsive inter-band pairing ( $\lambda_{12}$,$\lambda_{21}<0$), i.e. $\left\vert \lambda_{12},\lambda_{21}\right\vert
\gg\left\vert \lambda_{11},\lambda_{22}\right\vert $. Since the intra-band pairing depends on $\lambda_{ii}=\lambda_{ii}^{epi}-\mu_{ii}^{\ast}$, where $\lambda_{ii}^{epi}$ is the intra-band $EPI$ coupling constant and $\mu
_{ii}^{\ast}>0$ is an screened intra-band Coulomb repulsion, then in order to maximize $T_{c}$ the intra-band $EPI$ coupling it is wishful that $\lambda_{ii}^{epi}$ at least compensate negative effects of $\mu_{ii}^{\ast}$ (on $T_{c}$), i.e. $\lambda_{ii}^{epi}\geq\mu_{ii}^{\ast}$. Since in a narrow band one expects rather large screened Coulomb repulsion $\mu_{ii}^{\ast}$ ($\sim0.2$) then the intra-band $EPI$ coupling should be also appreciable. Moreover, from the experimental side the Raman measurements in Fe-pnictides [@Keimer-A1g] give strong evidence for a large phonon line-width of some $A_{1g}$ modes (where the $As$ vibration along the c-axis dominates). They are almost $10$ times larger than the $LDA$ band structure calculations predict. In [@Kulic-Haghighirad] a model was proposed where high electronic polarizability of $As$ ($\alpha_{As^{3-}}\sim12$ Å$^{3}$) ions screens the Hubbard repulsion and also give rise to a strong $EPI$ with A$_{1g}$ (mainly $As$) modes. An appreciable $As$ isotope effect in $T_{c0}$ was proposed in [@Kulic-Haghighirad], where the stable $^{75}As$ should be replaced by unstable $^{73}As$ - with the life-time of $80$ days, quite enough for performing relevant experiments. The situation is similar with $Fe-Se$ compounds, where an appreciable $EPI$ is expected, since $^{78}Se$ is also highly polarizable ($\alpha_{Se^{2-}}\sim7.5$ Å$^{3}$) and can be replaced by a long-living $^{73}Se$ isotope - the half-time $120$ days. Unfortunately these experiments were never performed.
We end up this digression by paying attention to some known facts, that the $LDA$ band structure calculations are unreliable in treating most high $T_{c}$ superconductors, since as a rule $LDA$ *underestimates non-local exchange-correlation effects and overestimates charge screening effects* - both effects contribute significantly to the $EPI$ coupling constant. As a result, $LDA$ strongly underestimates the $EPI$ coupling in a number of superconductors, especially in those near a metal-isolator transition. The classical examples for this claim are: (*i*) the $(BaK)BiO_{3\text{ }}$superconductor with $T_{c}>30$ $K$ which is $K$-doped from the parent isolating state $BaKBiO_{3}$. Here, $LDA$ predicts $\lambda<0.3$ and $T_{c}\sim1$ $K$, while the theories with an appropriate non-local exchange-correlation potential [@Yin-Kutepov-Kotliar-PRB] predict $\lambda_{epi}\approx1$ and $T_{c}\sim31$ $K$; (*ii*) The high temperature superconductors, for instance $YBaCu_{3}O_{7}$ with $T_{c}\sim100$ $K$, whose parent compound $YBaCu_{3}O_{6}$ is the Mott-insulator [@Kulic-PhysRep], [@Ma-Ku-Do-Advances].
After this digression we consider the main subject of the paper - *the role of the* $EPI$* with forward scattering peak (*$FSP$*) in pairing mechanism* of the $1ML$ $FeSe/SrTiO_{3}$ (and also $1ML$ $FeSe/TiO_{2}$) superconductor(s) with high critical temperatures $T_{c}\sim(50-100)$ $K$. In that respect, numerous experiments on $1ML$ $FeSe/SrTiO_{3}$ (and also on$1ML$ $FeSe/TiO_{2}$), combined with the fact that the $FeSe$ film on the graphene substrate has rather small $T_{c}\approx8$ $K$ (like in the bulk $FeSe$), give strong evidence that interface effects, due to $SrTiO_{3}$ (and $TiO_{2}$), are most probably responsible for high $T_{c}$. It turns out, that the most important results in $1ML$ $FeSe/SrTiO_{3}$ (and also $1ML$ $FeSe/TiO_{2}$), related to the existence of quasi-particle *replica bands* - which are identical to the main quasiparticle band [@Lee-Interfacial-supplement], [@Rebec-FeSe-TiO2], [@Replica-bands], can be coherently described by the $EPI-FSP$* *theory. This approach was proposed in seminal papers [@Lee-Interfacial-supplement]-[@Johnston1-2]. The beauty of these papers lies in the fact that they have recognized sharp replica bands in the $\acute{A}RPES$ spectra and related them to a sharp forward scattering peak in the $EPI$. (This is a very good example for a constructive cooperation of experimentalists and theoreticians.) Let us mention, that the $EPI-FSP$ theory was first studied in a connection with $HTSC$ cuprates [@Kulic-Zeyher1-2], while the extreme case of the $EPI-FSP$ pairing mechanism with delta-peak is elaborated in [@Dan-Dol-Kul-Oud] - see a review in [@Kulic-PhysRep]. Physically, this (in some sense exotic) interaction means that in some specific materials (for instance in cuprates and in $1ML$ $FeSe/SrTiO_{3}$) electron pairs exchange virtual phonons with small (transfer) momenta $q<q_{c}\ll k_{F}$ only, and as a result the effective pairing potential becomes long-ranged in real space [@Kulic-PhysRep]. It turns out that this kind of pairing can in some cases give rise to higher $T_{c}$ than in the standard (Migdal-Eliashberg) $BCS$-like theory. Namely, in the $EPI-FSP$ pairing mechanism one has $T_{c}^{(FSP)}=\left\langle V_{epi}(q)\right\rangle
_{q}/4\sim\lambda_{epi}^{(FSP)}/N(E_{F})$ [@Dan-Dol-Kul-Oud] - see below, instead of the $BCS$ dependence $T_{c}^{(BCS)}\sim\Omega e^{-1/\lambda
_{epi}^{(BCS)}}$. Here, $\lambda_{epi}^{(FSP)}$ and $\lambda_{epi}^{(BCS)}$ are the corresponding *mass-independent* $EPI$ coupling constants, where $\Omega$ - is the phonon energy, $N(E_{F})$ - the electronic density of states (per spin) at the Fermi surface. So, even for small $\lambda
_{epi}^{(FSP)}\ll\lambda_{epi}^{(BCS)}$ the case $T_{c}^{(FSP)}>T_{c}^{(BCS)}$ can be in principle realized. We inform the reader in advance, that the $EPI-FSP$ theory predicts also that $T_{c}^{(FSP)}\sim(q_{c}/k_{F})^{d}V_{epi}(0)$, ($d=1,2,3$ is the dimensionality of the system), which means that when $q_{c}\ll k_{F}$ high $T_{c}^{(FSP)}$ is hardly possible in $3D$ systems. However, the detrimental effect of the phase-volume factor $(q_{c}/k_{F})^{d}$ on $T_{c}^{(FSP)}$ can be compensated by its linear dependence on the pairing potential $V_{epi}(0)$. In some favorable materials this competition may lead even to an increase of $T_{c}$. We stress that properties of the superconductors with the $EPI-FSP$ mechanism of pairing are in many respects very different from the standard (BCS-like) superconductors, and it is completely justified to speak about *exotic superconductors*. For instance, the $EPI-FSP$ theory [@Kulic-PhysRep]-[@Dan-Dol-Kul-Oud] predicts, that in superconductors with the $EPI-FSP$ pairing *the isotope effect should be small* in leading order, i.e. $\alpha\ll1/2$ [@Kulic-PhysRep]-[@Dan-Dol-Kul-Oud] - see discussion in the following. This result is contrary to the case of the isotropic $EPI$ theory in standard metallic superconductors, where $\alpha$ is maximal, $\alpha=1/2$ (for $\mu^{\ast}=0$). We point out, that the $EPI-FSP$ pairing mechanism in strongly correlated systems is rather strange in comparison with the corresponding one in standard metals with good electronic screening, where the large transfer momenta dominate and the pairing interaction is, therefore, short-range. As a result, an important consequence of the $EPI-FSP$ pairing mechanism in case of $HTSC$-cuprates is that $T_{c}$ in the $d$-wave channel is of the same order as in the $s$-wave one. Since the residual repulsion is larger in the $s$- than in the $d$-channel ($\mu_{d}^{\ast}\ll\mu_{s}^{\ast}$) this result opens a door for $d$-wave pairing in $HTSC$-cuprates, in spite of the fact of the $EPI$ dominance [@Kulic-Zeyher1-2]-[@Kulic-PhysRep].
In the following, we study the superconductivity in $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$ [@Rebec-FeSe-TiO2]) in the framework of a semi-microscopic model of $EPI$ first proposed in seminal papers [@Lee-Interfacial-supplement]-[@Johnston1-2]. Namely, due to oxygen vacancies: $(i)$ an electronic doping of the $FeSe$ monolayer is realized, which gives rise to electronic-like bands centered at the $M$-points in the Brillouin zone, while the top of the hole-bands are at around $60$ $meV$ below the electronic-like Fermi surface; $(ii)$ the formed charge in the interface orders dipoles in the nearby $TiO_{2}$ layer; $(iii)$ the free charges in $SrTiO_{3}$ screen the dipolar field in the bulk, thus leaving the $TiO_{2}$ dipolar layer near the interface as an important source for the EPI. The oxygen ions in the $TiO_{2}$ dipolar layer vibrate with high-energy $\Omega\approx90$ $meV$, thus making a long-range dipolar electric field acting on metallic electrons in the $FeSe$ monolayer. This gives rise to a long-ranged $EPI$ [@Lee-Interfacial-supplement], [@Johnston1-2], which in the momentum space gives a forward scattering peak - the $EPI-FSP$* pairing mechanism*.
In this paper we make some analytical calculations in the framework of the $EPI-FSP$ theory with a very narrow $\delta$-peak, with the width $q_{c}\ll
k_{F}$, wher $k_{F}$ is the Fermi momentum [@Dan-Dol-Kul-Oud]. Here, we enumerate the obtained results, only: ($\mathbf{1}$) in leading order the *critical temperature* is linearly dependent on the pairing potential $V_{epi}(\mathbf{q})$, i.e. $T_{c0}\approx\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/4$. In order to obtain $T_{c0}\sim100$ $K$ we set the range of semi-microscopic parameters ($\varepsilon_{\parallel}^{eff}$, $\varepsilon_{\perp}^{eff}$, $q_{eff}$, $h_{0}$, $n_{d}$ - see below) entering $\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}$. Furthermore, since $\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}$ is independent of the the oxygen ($O$) mass, then $T_{c0}$ is *mass-independent* in leading order with respect to $T_{c0}/\Omega$. This means, that in $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$) one expects very small $O$-isotope effect ($\alpha_{O}\ll1/2$). Note, in [@Johnston1-2] large $\alpha_{O}=1/2$ is found; ($\mathbf{2}$) the *self-energy* $\Sigma(\mathbf{k},\omega)$ at $T=0$ is calculated analytically which gives: ($i$) the positions and spectral weights of the replica and quasiparticle bands at $T=0$ - all this quantities are *mass-dependent*; ($ii$) the slope of the quasiparticle self-energy for $\omega\ll\Omega$ ($\Sigma(\omega)\approx-\lambda_{m}\omega$) is *mass-dependent*, since $\sim\lambda_{m}\sim M_{O}^{1/2}$; ($\mathbf{3}$) in the $EPI-FSP$ model (without other interactions) the critical temperature for $s$-wave and $d$-wave pairing is degenerate, i.e. $T_{c0}^{(s)}=T_{c0}^{(d)}$. The presence of non-magnetic impurities (with the parameter $\Gamma=\pi n_{i}N(E_{F})u^{2}$) lifts this degeneracy. It is shown, that even the $s$-wave pairing (in the $EPI-FSP$ pairing mechanism) is sensitive to non-magnetic impurities, which are *pair-weakening* for it, i.e. $T_{c0}^{(s)}$ is decreased for large $\Gamma$, but never vanishes. It is also shown that for $d$-wave pairing $T_{c0}^{(d)}$ strongly depends on impurities, which are *pair-breaking*. The curiosity is that in the presence of non-magnetic impurities $T_{c0}^{(d)}$ in the $EPI-FSP$ pairing mechanism is more robust than the corresponding one in the $BCS$ model; ($\mathbf{4}$) the long-range $EPI-FSP$ pairing potential in real space makes a short-range potential in the momentum space. The latter gives rise to numerous low-laying excitation energy (above the ground-state) of pairs, thus leading to strong internal pair fluctuations which reduce $T_{c0}$. At $T_{c}<T<T_{c0}$ a pseudogap behavior is expected.
The structure of the paper is following: in *Section II* we calculate the $EPI-FSP$ pairing potential as a function of semi-microscopic parameters in the model of a dipolar layer $TiO_{2}$ with vibrations qf the oxygen ions [@Lee-Interfacial-supplement]-[@Johnston1-2]. In *Section III* the self-energy effects, such as replica bands and their intensities at $T=0$, are studied. The critical temperature $T_{c0}$ is calculated in *Section IV* in terms of the semi-microscopic parameters ($\varepsilon
_{\parallel}^{eff}$, $\varepsilon_{\perp}^{eff}$, $q_{eff}$, $h_{0}$, $n_{d}$). The range of of these parameters, for which one has $T_{c0}\sim100$ $K$, is estimated, too. In *Section V* the effect of nonmagnetic impurities on $T_{c0}$ are studied, while the effects of internal fluctuations of Cooper pairs are briefly discussed in *Section VI*. Summary of results are presented in *Section VII*.
$EPI-FSP$ pairing potential due to dipolar oxygen vibrations in the $TiO_{2}$layer
==================================================================================
It is important to point out that in $1ML$ $FeSe/SrTiO_{3}$ material, with $1$-monolayer of $FeSe$ grown on the $SrTiO_{3}$ substrate - mainly on the $(0,0,1)$ plane, the Fermi surface in the $FeSe$ monolayer is *electron-like *and* *centered at four M-points in the Brillouin zone - see more in [@Rebec-FeSe-TiO2], [@Sadovskii1-2]. The absence of the (nested) hole-bands on the Fermi surface rules out all $SFI$ nesting theories of pairing. Even the pairing between an electron- and incipient hole-band [@Hirschfeld-incipient] is ineffective since: ($1$) in the $FeSe$ monolayer the top of the hole band lies below the Fermi level around $60-80$ $meV$; ($2$) because of ($1$) the SFI coupling constant is (much) smaller than in the nesting case. This brings into play the interface interaction effects. The existence of *sharp replica bands* in the $ARPES$ spectra at energies of the order of optical phonons with $\Omega
\sim90$ $meV$, implies inevitably that the dominant interaction in $1ML$ $FeSe/SrTiO_{3}$ (and also in $1ML$ $FeSe/TiO_{2}$ [@Rebec-FeSe-TiO2]) is due to $EPI$ with strong forward scattering peak [@Lee-Interfacial-supplement]-[@Johnston1-2]. The physical mechanism for $EPI-FSP$ is material dependent and the basic physical quantities such as the width of the $FSP$, phonon frequencies and bare $EPI$ coupling can vary significantly from material to material. For instance, in $HTSC$-cuprates the effective $EPI-FSP$ potential $V_{epi}(q)$ is strongly renormalized by strong correlations, which is a synonym for large repulsion of two electrons on the $Cu$ ions - the doubly occupancy is forbidden. In that case the approximative $q$-dependence of $V_{epi}(q)$ is given by $V_{epi}(q)\approx\lbrack
1+(q/q_{c})^{2}]^{-2}V_{0,epi}(q)$, $q_{c}\sim\delta/a$, where $V_{0,epi}(q)$ is the bare (without strong correlations) coupling constant, $\delta(\ll1)$ is the hole concentration and $a$ is the $Cu-O$ distance [@Kulic-Zeyher1-2]-[@Kulic-PhysRep]. The prefactor is a vertex correction due to strong correlations and it means a new kind of (anti)screening in strongly correlated materials.
![*Left*: The microscopic structure of the $FeSe/SrTiO_{3}$ interface of the $1ML$ $FeSe/SrTiO_{3}$ structure. $h_{0}$ - distance between the $FeSe$ monolayer and $TiO_{2}$ dipolar layer. $\delta h$ is amplitude of the oxygen vibration in the dipolar layer. *Right*: The perpendicular view on the $FeSe-SrTiO_{3}$ interface in the model with one $TiO_{2}$ dipolar layer at the end of the $SrTiO_{3}$ substrate. The anisotropy of the effective dielectric constant $varepsilon$ is shown. The similar schema holds also for the $1ML$ $FeSe/TiO_{2}$ structure[]{data-label="1"}](fig1.eps){width="8cm"}
The interface in $1ML$ $FeSe/SrTiO_{3}$ can be considered as highly anisotropic material with the parallel and perpendicular (to the $FeSe$ plane) dielectric constants $\varepsilon_{\parallel}^{eff}\gg\varepsilon_{\perp
}^{eff}$. It is assumed [@Lee-Interfacial-supplement]-[@Johnston1-2] that the oxygen from the $TiO_{2}$ dipolar layer - placed at height ($-h_{0}$) from the $FeSe$ plane, vibrate and make dipolar moments $\delta p_{z}=q_{eff}\delta h(x,y,-h_{0})$ perpendicular to the $FeSe$ ($x-y$) plane - see $Fig.1$. It gives rise to a dipolar electric potential $\Phi_{dip}(x,y,-h_{0}-\delta h)$ acting on electrons in the $FeSe$ ($x-y$ plane). Here, $q_{eff}$ is an effective charge per dipole and $\delta h$ is the polar (dominantly oxygen) displacement along the $z$-axis [@Lee-Interfacial-supplement]. Due to some confusion in the literature on the form of $\Phi_{dip}$ [@Lee-Interfacial-supplement] we recalculate it here, in order to know its explicite dependence on the semi-microscopic parameters $\varepsilon_{\parallel}^{eff}$, $\varepsilon_{\perp}^{eff}$, $q_{eff}$, $h_{0}$, $n_{d}$. An elementary electrodynamics approach [@Landau-Electrodynamics] gives for the dipolar potential $\Phi
_{dip}(x,y,-h_{0}-\delta h)$ $$\Phi_{dip}(x,y,-h_{0}-\delta h)=\frac{(\varepsilon_{\parallel}^{eff})^{1/2}}{(\varepsilon_{\perp}^{eff})^{3/2}}(n_{d}q_{eff}h_{0})\times$$$${\displaystyle\iint}
\frac{dx^{\prime}dy^{\prime}\delta h(x^{\prime},y^{\prime},-h_{0})}{\left(
\frac{\varepsilon_{\parallel}^{eff}}{\varepsilon_{\perp}^{eff}}h_{0}^{2}+(x-x^{\prime})^{2}+(y-y^{\prime})^{2}\right) ^{\frac{3}{2}}},
\label{Phi}$$ where $n_{d}$ is the number of the oscillating $Ti-O$ dipoles per unit $FeSe$ surface. The coefficient in front of the integral is different from that in [@Lee-Interfacial-supplement] - where it is $\varepsilon_{\parallel}^{eff}/(\varepsilon_{\perp}^{eff})^{3/2}(q_{eff}h_{0})$ (probably typos?) and with missed dipole density $n_{d}$. This coefficient does not fulfill the condition $\Phi\sim\varepsilon^{-1}$ in the isotropic case, while $Eq.$(\[Phi\]) does. By introducing $g_{epi}(\mathbf{q})=e\Phi(\mathbf{q})$ the electron-phonon interaction Hamiltonian $H_{epi}=\sum_{\mathbf{q}}e\hat{\Phi}(\mathbf{q})\hat{\rho}(\mathbf{q})$ is rewritten in the form $H_{epi}=\sum_{\mathbf{k},\mathbf{q}}g_{epi}(\mathbf{q})(\hat{b}_{\mathbf{q}}+\hat{b}_{-\mathbf{q}}^{\dagger})\hat{c}_{\mathbf{k}+\mathbf{q}}^{\dagger
}\hat{c}_{\mathbf{k}}$, where $\hat{b}_{-\mathbf{q}}^{\dagger},\hat
{c}_{\mathbf{k}+\mathbf{q}}^{\dagger}$ are boson and fermion creation operators, respectively. The Fourier transformed potential $g_{epi}(\mathbf{q})(=(g_{0}/\sqrt{N})e^{-q/q_{c}})$ is given by $$g_{epi}(\mathbf{q})=\frac{2\pi n_{d}eq_{eff}}{\varepsilon_{\perp}^{eff}}\sqrt{\frac{\hbar}{M\Omega N}}e^{-q/q_{c}}, \label{g-epi}$$ $g_{0}=(2\pi n_{d}eq_{eff}/\varepsilon_{\perp}^{eff})(\hbar/M\Omega)^{1/2}$, $e$ is the electronic charge. Here, the screening momentum $q_{c}=(\varepsilon_{\perp}^{eff}/\varepsilon_{\parallel}^{eff})^{1/2}h_{0}^{-1}$ characterizes the range of the $EPI$ potential, i.e. for $q_{c}\ll k_{F}$ ($k_{F}$ is the Fermi momentum) the $EPI$ is sharply peaked at $\mathbf{q}=0$ - the forward scattering peak ($FSP$), and the potential in real space is long-ranged, while for $q_{c}\sim k_{F}$ it is short-ranged, like in the standard $EPI$ theory. Since we are interested in the $T_{c}$ dependence on the effective parameters $\varepsilon_{\parallel}^{eff},\varepsilon_{\perp
}^{eff},q_{eff},h_{0}$, then an explicit dependence of the potential is important. We shall see below, that in order that this approach is applicable to $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$) $\varepsilon_{\parallel
}^{eff},\varepsilon_{\perp}^{eff}$ must be very different from the bulk values of $\varepsilon$ in the bulk $SrTiO_{3}$ - where $\varepsilon\sim500-10^{4}$, or in the rutile $TiO_{2}$ structure where $\varepsilon<260$ [@Rebec-FeSe-TiO2].
Self-energy effects and $ARPES$ replica bands
=============================================
The general self-energy $\Sigma_{epi}(\mathbf{k}_{F},\omega)$ at $T=0$ in the extreme $FSP$ $\delta$-peak limit with the width $q_{c}\ll k_{F}$) is given by (see Appendix)$$\Sigma_{epi}(\mathbf{k}_{F},\omega)\approx-\lambda_{m}\frac{\omega}{1-(\omega/\Omega)^{2}}, \label{Sigma-epi}$$ where $\lambda_{m}=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/2\Omega$ is the *mass-dependent coupling constant*. Here, the average $EPI$ potential is given by $\left\langle V_{epi}(\mathbf{q})\right\rangle
_{q}=Ns_{c}(2\pi)^{-2}\int d^{2}qV_{epi}(\mathbf{q},0)\approx(1/4\pi
)(aq_{c})^{2}V_{epi}^{0}$, where $s_{c}=2a^{2}$ is the surface of the $FeSe$ unit cell and $a$ is the Fe-Fe distance, and the bare pairing $EPI$ potential is $V_{epi}^{0}=2g_{0}^{2}/\Omega$. The coupling constant $\lambda_{m}$ corresponds to $\lambda_{m}$ used in [@Johnston1-2], where the self-energy effects are studied at $T>0$. It is important to point out that $\lambda_{m}$ is (oxygen) *mass-dependent*, contrary to [@Johnston1-2]. Since $\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}$ is *mass-independent* then $\lambda_{m}\sim\Omega^{-1}\sim M^{1/2}$. In the following we discuss the case when $\mathbf{k}=\mathbf{k}_{F}$, i.e. $\xi(\mathbf{k})=0$. For $\omega\ll\Omega$ one has $\Sigma_{epi}(\mathbf{k},\omega)=-\lambda_{m}\omega$ which means that the slope of $\Sigma_{epi}(\mathbf{k},\omega)$ is *mass-dependent*. The latter property can be measured by $ARPES$ and thus the $EPI-FSP$ theory can be tested. Note, that in the $EPI-FSP$ theory the critical temperature $T_{c0}(=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/4)$ - see details below, is *mass-independent*. Both these results are *opposite to the standard Migdal-Eliashberg theory*, where the self-energy slope is mass-independent and $T_{c0}$ is mass-dependent.
The *quasiparticle and replica bands* at $T=0$ are obtained from $\omega-\Sigma_{epi}(\omega)=0$. In the following we make calculations at $T=0$ and at the Fermi surface $\xi(\mathbf{k}_{F})=0$. The solutions are: ($1$) $\omega_{1}=0$ - the *quasiparticle band*; ($2$) $\omega
_{2}=-\Omega\sqrt{1+\lambda_{m}}$ is the $ARPES$ *replica band*; ($3$) the inverse $ARPES$ replica band $\omega_{3}=\Omega\sqrt{1+\lambda_{m}}$. The single-particle spectral function is $A(\mathbf{k}_{F},\omega,T=0)=\sum
_{i=1}^{3}(A_{i}/\pi)\delta(\omega-\omega_{i})$, where $A_{i}/\pi$ are the spectral weights. For the quasiparticle band $\omega_{1}$ one obtains $A_{1}=(1+\lambda_{m})^{-1}$, while for the replica bands at $\omega_{2}$ and $\omega_{3}$ one has $A_{2}=A_{3}=(\lambda_{m}/2)(1+\lambda_{m})^{-1}$. The ratio of the intensities at $T=0$ of the $\omega_{2}$ replica band and quasiparticle band $\omega_{1}$ is given by$$\frac{A_{2}(\mathbf{k}_{F},\omega,T=0)}{A_{1}(\mathbf{k}_{F},\omega
,T=0)}=\frac{\lambda_{m}}{2}. \label{ARPES-inten}$$ It is necessary to mention that at finite $T(>0)$ this ratio is changed as found in [@Johnston1-2]. In that case $\Sigma_{epi}^{(T)}(\mathbf{k\approx
k}_{F},\omega,T\neq0)\approx\lambda_{m}/(\omega+\Omega)$ which gives the quasiparticle and replica band $\omega_{1}^{(T)}=\Omega(-1+\sqrt
{1+4\lambda_{m}})/2$ and $\omega_{2}^{(T)}=-\Omega(1+\sqrt{1+4\lambda_{m}})/2$ and $(A_{2}/A_{1})_{T}=\lambda_{m}$ [@Steve-Kulic]. This intriguing difference of the $T=0$ and $T\neq0$ results for $(A_{2}/A_{1})$ in the $EPI-FSP$ theory, $(A_{2}/A_{1})_{T}=2(A_{2}/A_{1})_{0}$, is due to the sharpness of the Fermi function $n_{F}(\xi_{\mathbf{k+q}})$ entering in $\Sigma_{epi}(\mathbf{k},\omega)\ $- see $Eq.$(\[Sigma-integ\]) in Appendix [@Steve-Kulic].
We stress that, the $ARPES$ measurements of $A_{2}/A_{1}$ in $1ML$ $FeSe/SrTiO_{3}$ were done at finite temperatures ($T\neq0$) and in the $k=0$ point with $\xi(k=0)\sim-50$ $meV$ which gives $(A_{2}/A_{1})_{T}\approx0.15-0.2$ [@Lee-Interfacial-supplement], [@Replica-bands]. According to the theory in [@Lee-Interfacial-supplement], [@Johnston1-2] one obtains $\lambda_{m}^{(ARPES)}\approx0.15-0.2$. Below we show, that $\lambda_{m}$ can be also extracted from the formula $Eq.$(\[Tco\] ) for $T_{c0}\approx100$ $K$, which gives $\lambda
_{m}^{(T_{c0})}\approx0.18$. The latter value is in a good agreement with $\lambda_{m}^{(ARPES)}$ from $ARPES$ [@Steve-Kulic]. If we put this value in $Eq.$(\[ARPES-inten\]) one obtains that at $T=0$ and at $k=k_{F}$ one has $(A_{2}/A_{1})\sim0.1$. From this analysis we conclude that the ARPES measurements at $k_{F}$ should give the similar ratio as at $k=0$. The calculated $ARPES$ spectra at $T=0$ $K$ and at $k=k_{F}$ give $\Delta
\omega=\left\vert \omega_{2}-\omega_{1}\right\vert =\Omega\sqrt{1+\lambda_{m}}$ while the experimental value is $\Delta\omega\approx100$ $meV$, which for $\lambda_{m}^{(ARPES)}\approx0.2$ gives the optical phonon energy of the order of $\Omega\approx90$ $meV$.
The superconducting critical temperature $T_{c0}$ and gap $\Delta
_{0}$
=================================================================
In the weak coupling limit ($\lambda_{m}\ll1$) of the Eliashberg equations with $q_{c}v_{F}<\pi T_{c0}$ ($v_{F}$ is the Fermi velocity) the linearized gap equation (near $T_{c0}$) is given by $$\Delta(\mathbf{k},\omega_{n})=T_{c0}\sum_{\mathbf{q}}\sum_{m}\frac
{V_{epi}(q,\omega_{n}-\omega_{m})\Delta(\mathbf{k}+\mathbf{q},\omega_{m})}{\omega_{m}^{2}+\xi^{2}(\mathbf{k}+\mathbf{q})}, \label{Delta}$$ where $\omega_{n}=\pi T_{c0}(2n+1)$, $V_{epi}(\mathbf{q},\Omega_{n})=g_{epi}^{2}(\mathbf{q})(2\Omega/(\Omega_{n}^{2}+\Omega^{2}))$, $\Omega
_{n}=2\pi T_{c0}\cdot n$. For $\Omega\gg\pi T_{c0}$ one has $V_{epi}(\mathbf{q},\Omega_{n})\approx V_{epi}(\mathbf{q},0)=2g_{epi}^{2}(\mathbf{q})/\hbar\Omega$. In the strong $FSP$ limit when $(q_{c}v_{F})\ll(\pi
T_{c0})^{2}$ the highest value of $\Delta(\mathbf{k},\omega_{n})$ is reached at $\mathbf{k}=\mathbf{k}_{F}$ ($\xi(\mathbf{k}_{F})=0$ in $Eq.$(\[Delta\]). The solution $\Delta(\mathbf{k},\omega_{n})$ is searched in the standard BCS-like *square-well* approximation $\Delta(\mathbf{k}_{F},\omega
_{n})\approx\Delta_{0}=const$. In leading order with respect to $(T_{c0}/\Omega)\ll1$ one obtains $T_{c0}$ [@Kulic-PhysRep]-[@Dan-Dol-Kul-Oud] $$T_{c0}\approx\frac{1}{\pi^{2}}\left\langle V_{epi}(\mathbf{q})\right\rangle
_{q}\sum_{-\Omega/\pi T_{c0}}^{\Omega/\pi T_{c0}}\frac{1}{(2m+1)^{2}}
\label{Tco}$$ For $\Omega\gg\pi T_{c0}$ this gives $T_{c0}=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/4\approx(1/16\pi)(aq_{c})^{2}(2g_{0}^{2}/\Omega)$, where $a$ is the $Fe-Fe$ distance. Note, that $T_{c0}$ is *mass-independent (*$\alpha_{O}=0$) - note $\alpha_{O}=1/2$ is found in [@Johnston1-2]. The small isotope-effect can be a smoking-gun experiment for the $EPI-FSP$ pairing mechanism in $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$). From $Eq.$(\[Delta-n-FSP\]) in the Appendix it is straightforward to obtain the *energy gap* $\Delta_{0}=2T_{c0}$. Note, that $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$) is a $2D$ system and $T_{c0}\sim q_{c}{}^{2}$, while in the $d$-dimensional space one has $T_{c0}\sim q_{c}{}^{d}$. This means that the $EPI-FSP$ mechanism of superconductivity is more *favorable in low-dimensional systems* ($d=1,2$) than in the $3D$ one. Since high $T_{c}$ cuprates are also quasi-$2D$ systems, where strong correlations make a long-ranged $EPI$, it means that the $EPI-FSP$ mechanism of pairing may be also operative in cuprates [@Kulic-PhysRep]. Note, that in estimating some semi-microscopic parameters we shall use as a reper-value $T_{c0}\approx100$ $K$, while in real systems $T_{c}\sim(60-80)$ $K<T_{c0}$ is realized. However, $T_{c0}$ is the mean-field value obtained in the Migdal-Eliashberg theory, while in $2D$ systems it is significantly reduced by the phase fluctuations - to the Berezinski-Kosterliz-Thouless value. There is an additional reduction of $T_{c0}$ (which might be also appreciable) in the $EPI-FSP$ systems, which is due to internal pair-fluctuations - see discussion below.
One can estimate the coupling constant $\lambda_{m}=\left\langle
V_{epi}(\mathbf{q})\right\rangle _{q}/2\Omega$ in $1ML$ $FeSe/SrTiO_{3}$ from the value of $T_{c0}$. Then for the reper-value $T_{c0}\sim100$ $K$ one has $\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}\approx33$ $meV$ and $\lambda_{m}^{(T_{c0})}\approx$ $0.18$. Since, $\lambda_{m}^{(ARPES)}\approx\lambda_{m}^{(T_{c0})}$ the consistency of the theory is satisfactory. Note, if one includes the wave-function renormalization effects (contained in $Z(i\omega_{n})>1$) then in the case $(T_{c0}/\Omega)\ll1$ and for the square-well solution $T_{c0}$ is lowered to $T_{c0}^{(Z)}=T_{c0}/Z^{2}(0)$, where $Z(0)\approx1+\lambda_{m}$ [@Dan-Dol-Kul-Oud]. This means, that the nonlinear corrections (with respect to $\lambda_{m})$ in $T_{c0}$ and $\Delta_{0}$ [@Johnston1-2], [@Murta] should be inevitably renormalized by the $Z$-renormalization.
Let us estimate the parameters ($\varepsilon_{\parallel},\varepsilon_{\perp
},q_{eff},h_{0}$) which enter in $T_{c0}$. In order to reach $T_{c0}\sim100$ $K$ (and $\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}=4T_{c0}\sim400$ $K\approx33$ $meV$) then for $aq_{c}\approx0.2$ and $\Omega\approx90$ $meV$ one obtains $g_{0}\approx0.7$ $eV$. Having in mind that $g_{0}=(2\pi
n_{d}eq_{eff}/\varepsilon_{\perp}^{eff})(\hbar/M\Omega)^{1/2}$ and that the zero-motion oxygen amplitude is $(\hbar/M\Omega)^{1/2}\approx0.05$ Å and by assuming that $n_{d}\approx\alpha/s_{c}$, $s_{c}=\tilde{a}^{2}$, $\tilde
{a}=\sqrt{2}a\approx4$ Å, $q_{eff}\sim2e$, $\alpha\gtrsim1$, then in order to obtain $g_{0}\approx0.7$ $eV$ $\varepsilon_{\perp}^{eff}$ must be small, i.e. $\varepsilon_{\perp}^{eff}\sim1$. Since $aq_{c}=(a/h_{0})\sqrt
{\varepsilon_{\perp}^{eff}/\varepsilon_{\parallel}^{eff}}\sim0.2$ and for $(a/h_{0})\sim1$ it follows $\varepsilon_{\parallel}^{eff}\sim30$. Note, that in $SrTiO_{3\text{ }}$the bulk $\varepsilon$ is large, $\varepsilon
\sim500-10^{4}$. So, if $T_{c0}$ in $1ML$ $FeSe/SrTiO_{3}$ is due solely to the $EPI-FSP$ mechanism, then in the model where the oxygen vibrations in the single dipolar monolayer $TiO_{2\text{ }}$are responsible for the pairing potential the effective dielectric constants $\varepsilon_{\perp}^{eff}$, $\varepsilon_{\parallel}^{eff}$ are very *different from the bulk values* in $SrTiO_{3}$ (in $1ML$ $FeSe/TiO_{2}$ one has $\varepsilon\leq260$ [@Rebec-FeSe-TiO2]). This is physically plausible since for the nearest (to the $FeSe$ monolayer) $TiO_{2}$ dipolar monolayer there is almost nothing to screen in the direction perpendicular to $FeSe$, thus making $\varepsilon
_{\perp}^{eff}\ll$ $\varepsilon^{bulk}$. Note, that for the parameters assumed in this analysis and for $T_{c0}\approx100$ $K$ one obtains rather large bare pairing potential $V_{epi}^{0}\approx10$ $eV$ . This means that in the absence of the $FSP$ in $EPI$ and for the density of states of the order $N(E_{F})\sim0.5$ $(eV)^{-1}$ (typical for Fe-based superconductors) the bare coupling constant $\lambda_{epi}^{0}=N(E_{F})V_{epi}^{0}$ would be large, $\lambda_{epi}^{0}\sim5$. We stress that the above theory is also applicable to recently discovered $1ML$ $FeSe/TiO_{2}$ [@Rebec-FeSe-TiO2]. To conclude, the high $T_{c0}$ in $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$) is obtained on the expense of the large maximal $EPI$ coupling $V_{epi}^{0}$ which compensates smallness of the (detrimental) phase-volume factor $(aq_{c})^{2}$.
Effects of impurities on $T_{c0}$
=================================
In clean systems with the $EPI-FSP$ mechanism of superconductivity $T_{c0}$ is degenerate - it is equal in $s$- and $d$-channels. In the following we show, that the $s$-wave superconductivity is also affected by isotropic non-magnetic impurities, i.e. $T_{c0}$ is reduced and the Anderson theorem is violated. This may have serious repercussions on the *s-wave superconductivity* in $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$) where $T_{c0}$ may depend on chemistry. Then by using equation $Eq.$(\[Tc-s-wave\]) from Appendix one obtains $T_{c}^{(s)}$$$\frac{T_{c}^{(s)}}{T_{c0}}=\frac{4}{\pi^{2}\rho}\psi(\frac{1}{2}+\frac{\rho
}{2})-\psi(\frac{1}{2}),$$ where $\rho=\Gamma/\pi T_{c}$, $\Gamma=\pi n_{i}N(E_{F})u^{2}$, $n_{i}$ is the impurity concentration and $u$ is the impurity potential. Let us consider some limiting cases: ($1$) for $\Gamma\ll\pi T_{c}^{(s)}$one has $T_{c}^{(s)}\approx T_{c0}[1-7\zeta(3)\Gamma/\pi^{3}T_{c0}]$; ($2$) for $\Gamma
\gg\pi T_{c}$ one has $T_{c}^{(s)}\approx(\Gamma/2\pi)\exp(-\pi\Gamma
/4T_{c0})$, i.e. $T_{c}^{(s)}$ never vanishes. This means that in the $EPI-FSP$ systems the non-magnetic impurity scattering is *pair-weakening* for the s-wave superconductivity.
In the case of *d-wave superconductivity* the solution of $Eq.$(\[Tc-d-wave\]) in limiting cases is: ($1$) $T_{c}^{(d)}\approx
T_{c0}[1-2\Gamma/\pi T_{c0}]$ for $\Gamma\ll\pi T_{c}^{(d)}$. We point out that the slope $-dT_{c}^{(d)}/d(\Gamma)=2/\pi$ is smaller than the slope for the standard $d$-wave pairing, where $-dT_{c}/d(\Gamma)=\pi/4$. ($2$) For $\Gamma>\Gamma_{cr}^{(FSP)}\approx(4/\pi)T_{c0}$ one has $T_{c}^{(d)}=0$, i.e. the effect of non-magnetic impurities is *pair-breaking*. Note, that $\Gamma_{cr}^{(FSP)}>\Gamma_{cr}(=(2/\pi)T_{c0})$. These two results mean that in the presence of non-magnetic impurities the $d$-wave superconductivity which is due to the $EPI-FSP$ pairing is more robust than in the case of the standard $d$-wave pairing. We stress, that the $T_{c}$ dependence on non-magnetic impurities in $1ML$ $FeSe/SrTiO_{3}$ (as well as in$1ML$ $FeSe/TiO_{2}$) might be an important test for the $EPI-FSP$ pairing in this material.
Finally, it is worth of mentioning, that the real isotope effect in $T_{c0}$ of $1ML$ $FeSe/SrTiO_{3}$ (and in $1ML$ $FeSe/TiO_{2}$) might depend on the type of non-magnetic impurities. If their potential is also long-ranged (for instance due to oxygen deffects in the $TiO_{2}$ dipole layer), then there is $FSP$ in the scattering potential, i.e. $u_{imp}^{2}(q)\approx u^{2}\delta(\mathbf{q})$. Then, such impurities affect in the same way $s$- and $d$-wave pairing and they are pair weakening, as shown in [@Dan-Dol-Kul-Oud]. Naimly, one has $(a)$ $T_{c}^{(s,d)}\approx
T_{c0}[1-4\Gamma_{F}/49T_{c0}]$ for $\Gamma_{F}\ll\pi T_{c}$, where $\Gamma_{F}=\sqrt{n_{i}}u$; (b) $T_{c}^{(s,d)}\approx0.88$ $\Omega\exp
(-\pi\Gamma_{F}/4T_{c0})$, for $\Gamma_{F}\gg\pi T_{c}$. There are two important results: $(1)$ There is a nonanalicity in $\Gamma_{F}\sim\sqrt
{n_{i}}$; $(2)$ there is a full isotope effect in the “dirty” limit $\Gamma_{F}\gg\pi T_{c}$, i.e. $\alpha_{O}=1/2$, since $T_{c}^{(s,d)}\sim\Omega\sim M^{-1/2}$. We stress, that if the full isotope effect would be realized experimentally in $1ML$ $FeSe/SrTiO_{3}$ (and in $1ML$ $FeSe/TiO_{2}$), then this does not automatically exclude the $FSP-EPI$ mechanism of pairing, since it may be due to impurity effects. In that case the nonanalicity of $\Gamma_{F}$ in $n_{i}$ might be a smoking-gun effect.
Internal pair fluctuations reduce $T_{c0}$
==========================================
The $EPI-FSP$ theory, which predicts a long-range force between paired electrons, opens a possibility for a pseudogap behavior in $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$). As we have discussed above, the $EPI-FSP$ theory predicts a non-$BCS$ dependence of the critical temperature $T_{c0}$, i.e. $T_{c0}=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/4$. However, this mean-field ($MFA$) value is inevitably reduced by the *phase* and *internal Cooper pair fluctuations* - which are present in systems with long-range attractive forces. Namely, in the $MFA$ the order parameter $\Delta(\mathbf{x},\mathbf{\mathbf{x}}^{\prime})(=V(\mathbf{x-\mathbf{x}}^{\prime})\langle\mathbf{\psi}_{\downarrow
}(\mathbf{x}^{\prime})\mathbf{\psi}_{\uparrow}(\mathbf{x})\rangle)$ depends on the relative (internal) coordinate $\mathbf{r}=\mathbf{x}-\mathbf{\mathbf{x}}^{\prime}$ and the center of mass $\mathbf{R}=(\mathbf{x}+\mathbf{\mathbf{x}}^{\prime}\mathbf{\mathbf{)/2}}$, i.e. $\Delta(\mathbf{x},\mathbf{\mathbf{x}}^{\prime})=\Delta(\mathbf{r},\mathbf{\mathbf{R}})$. In usual superconductors with short-range pairing potential one has $V_{sr}(\mathbf{x-\mathbf{x}}^{\prime})\approx V_{0}\delta(\mathbf{x-\mathbf{x}}^{\prime})$ and $\Delta(\mathbf{r},\mathbf{\mathbf{R}})=\Delta(\mathbf{\mathbf{R}})$. Therefore only the spatial ($\mathbf{\mathbf{R}}$-dependent) fluctuations of the order parameter are important. In case of a long-range pairing potential there are additional pair-fluctuations due to the dependence of $\Delta
(\mathbf{r},\mathbf{\mathbf{R}})$ on internal degrees of freedom (on $\mathbf{r}$). The interesting problem of fluctuations in systems with long-range attractive forces in $3D$ systems was studied in [@Yang] and we sketch it briefly, because it shows that standard and $EPI-FSP$ superconductors belong to different universality classes. The best way to see importance of the internal pair-fluctuations is to rewrite the pairing Hamiltonian in terms of pseudospin operators (in this approximation first done by P. Anderson the single particle excitations are not included) $$\hat{H}=\sum_{\mathbf{k}\sigma}2\xi_{\mathbf{k}}\hat{S}_{\mathbf{k}\sigma}^{z}-(1/2)\sum_{\mathbf{k},\mathbf{k}^{\prime}}V_{\mathbf{k}-\mathbf{k}^{\prime}}(\hat{S}_{\mathbf{k}}^{+}\hat{S}_{\mathbf{k}^{\prime}}^{-}+\hat
{S}_{\mathbf{k}^{\prime}}^{+}\hat{S}_{\mathbf{k}}^{-}), \label{H-pseudo}$$ where $S_{\mathbf{k}\sigma}^{z}=(\hat{c}_{\mathbf{k}\uparrow}^{\dagger}\hat
{c}_{\mathbf{k}\uparrow}-\hat{c}_{-\mathbf{k}\downarrow}^{\dagger}\hat
{c}_{-\mathbf{k}\downarrow}-1)/2$, $\hat{S}_{\mathbf{k}\sigma}^{+}=\hat
{c}_{\mathbf{k}\uparrow}^{\dagger}\hat{c}_{-\mathbf{k}\downarrow}^{\dagger}$ [@Yang]. This is a Heisenberg-like Hamiltonian in the momentum space. In case of the $s$-wave superconductivity with short-range forces $V_{sr}(\mathbf{x-\mathbf{x}}^{\prime})\approx V_{0}\delta
(\mathbf{x-\mathbf{x}}^{\prime})$ one has $V_{\mathbf{k}-\mathbf{k}^{\prime}}=const$ and the pairing potential is *long-ranged in the momentum space*. In that case it is justified to use *the mean-field approximation* $\hat{H}\rightarrow\hat{H}_{mf}=-\sum_{k}\mathbf{h}_{\mathbf{k}}\mathbf{\hat{S}}_{\mathbf{k}}$ with the mean-field $\mathbf{h}_{\mathbf{k}}=-2\xi_{\mathbf{k}}\mathbf{z+}\sum_{\mathbf{k}^{\prime}}V_{\mathbf{k}-\mathbf{k}^{\prime}}\langle S_{\mathbf{k}^{\prime}}^{x}\mathbf{x+}S_{\mathbf{k}^{\prime}}^{y}\mathbf{y\rangle}$ . The excitation spectrum (with respect to the ground state) in this system have a gap, i.e. $E(\mathbf{k})=2\sqrt{\xi_{\mathbf{k}}^{2}+\Delta_{\mathbf{k}}^{2}}$ where the gap $\Delta_{\mathbf{k}}$ is the mean-field order parameter defined by $\Delta_{\mathbf{k}}=\sum_{\mathbf{k}^{\prime}}V_{\mathbf{k}-\mathbf{k}^{\prime}}\langle\hat{S}_{\mathbf{k}^{\prime}}^{x}\rangle$. In case of the $EPI-FSP$ pairing mechanism the pairing potential is long-ranged in real space and short-ranged in the momentum space. For instance, in $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$) one has $V_{\mathbf{k}-\mathbf{k}^{\prime}}=V_{0}\exp\{-\left\vert \mathbf{k}-\mathbf{k}^{\prime}\right\vert /q_{c}\}$ with $q_{c}\ll k_{F}$, and the excitation spectrum is boson-like $0<E(\mathbf{k})<2\sqrt{\xi_{\mathbf{k}}^{2}+\Delta_{\mathbf{k}}^{2}}$ (like in the Heisenberg model) with large number of low-laying excitations (around the ground state). This means, that there are many low-laying pairing states above the ground-state in which pairs are sitting. This, so called internal fluctuations effect, reduces $T_{c0}$ to $T_{c}$. For instance, tin $3D$ systems with $q_{c}\xi_{0}\ll1$ [@Yang] one has $T_{c}\approx(q_{c}\xi
_{0})T_{c0}$, where the coherence length $\xi_{0}=v_{F}/\pi\Delta_{0}$ and $\Delta_{0}=2T_{c0}$. It is expected, that in the region $T_{c}<T<T_{c0}$ the pseudogap ($PG$) phase is realized. However, in $2D$ systems, like $1ML$ $FeSe/SrTiO_{3}$ (and $1ML$ $FeSe/TiO_{2}$), there are additionally phase fluctuations reducing $T_{c}$ further to the Berezinskii-Kosterliz-Thouless value. We stress, that recent measurements of $T_{c}$ in $1ML$ $FeSe/SrTiO_{3}$ by the Meissner effect and resistivity ($\rho(T)$) give that $T_{c}^{(\rho
)}<T_{c}^{(M)}$ what may be partly due to these internal fluctuations of Cooper pairs. It would be interesting to study theoretically these two kind of fluctuations in $2D$ systems, such as $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$.
Summary and Discussion
=======================
In the paper we study the superconductivity in the $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$ sandwitch-structure, which contains one metallic $FeSe$ monolayer grown on the substrate $SrTiO_{3}$, or rutile $TiO_{2}^{(100)}$. It turns out that in such a structure the *Fermi surface is electron-like* and the bands are pockets around the $M$-point in the Brillouin zone. The bottom of the electron-like bands is around $(50-60)$ $meV$ below the Fermi surface at $E_{F}$. The top of the hole-like band at the point $\Gamma$ lies $80$ $meV$ below $E_{F}$ which means that pairing mechanisms based on the electron-hole nesting are ruled out. This holds also for the pairing with hole-incipient bands (very interesting proposal) [@Hirschfeld-review]. The superconductivity in $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$ is realized in the $FeSe$ monolayer with $T_{c}\sim(60-100)$ $K$. The decisive fact for making a theory is that the $ARPES$ spectra show s*harp replica bands* around $100$ $meV$ below the quasiparticle band, what is approximately the energy of the oxygen optical phonon $\Omega\approx90$ $meV$. The analysis of superconductivity is based on the semi-microscopic model - first proposed in [@Lee-Interfacial-supplement], [@Johnston1-2], where it is assumed that a $TiO_{2}$ dipolar layer is formed just near the interface. In that model the oxygen vibrations create a dipolar electric potential, which acts on electrons in the $FeSe$ monolayer, thus making the $EPI$ interaction long-ranged. In the momentum space a forward scattering peak ($FSP$) appears, i.e. $EPI$ is peaked at small transfer momenta ($q<q_{c}\ll
k_{F}$) with $g_{epi}(\mathbf{q})=g_{0}\exp\{-q/q_{c}\}$. Here, this is called the $EPI-FSP$ pairing mechanism. The $EPI-FSP$ theory is formulated first in [@Kulic-Zeyher1-2] for strongly correlated systems, while its extreme case with delta-peak is elaborated in [@Dan-Dol-Kul-Oud] - see also [@Kulic-PhysRep]. This limiting (delta-peak) case makes not only analytical calculations easier, but it makes also a good fit to experimental results [@Lee-Interfacial-supplement], [@Johnston1-2]. In the following, we summarize the main obtained results of the $EPI-FSP$ theory and its relation to the $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$ sandwitch-structures.
$(1)$ - The mean-field critical temperature $T_{c0}$ in the $s$-wave and $d$-wave pairing channels is* degenerate* and given by $T_{c0}=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/4\approx(1/16\pi
)(aq_{c})^{2}V_{epi}^{0}$, where $V_{epi}(\mathbf{q})=2g_{epi}^{2}(\mathbf{q})/\hbar\Omega$ and the maximal pairing potential $V_{epi}^{0}(\equiv V_{epi}(q=0))=(2g_{0}^{2}/\Omega)$. On the first glance this linear dependence of $T_{c0}$ on $V_{epi}^{(0)}$ seems to be favorable for reaching high $T_{c0}$ - note in the $BCS$ theory $T_{c0}$ is exponentially dependent on $V_{epi}^{(0)}$ and very small for small $N(E_{F})V_{epi}^{0}$. However, for non-singular $g_{epi}(\mathbf{q})$ when $g_{epi}(\mathbf{q}=0)$ is finite, $T_{c0}$ is limited by the smallness of the phase-volume effect, which is in $2D$ systems (such as $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$) proportional to $(aq_{c})^{2}\ll1$. In that sense optimistic claims that the $EPI-FSP$ mechanism leads inevitably to higher $T_{c0}$ - than the one in the standard Migdal-Eliashberg theory, are not well founded. This holds especially for $3D$ systems, where $T_{c0}\sim$ $(aq_{c})^{3}$ and $T_{c0}^{(3d)}\ll T_{c0}^{(2d)}$ for the same value of $V_{epi}^{0}$. However, higher $T_{c0}$ (with respect to to the $BCS$ case) can be reached by fine tuning of $aq_{c}$ and $V_{epi}^{0}$. This is probably realized in $HTSC$ cuprates and with certainty in $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$. The weak-coupling theory predicts the superconducting gap to be $\Delta
_{0}=2T_{c0}$ and for $T_{c0}\sim100$ $K$ one has $\Delta_{0}\sim16$ $meV$ what fits well the $ARPES$ experimental values [@Lee-Interfacial-supplement], [@Rebec-FeSe-TiO2]. Note, in order to reach $T_{c0}=100$ $K$ for $aq_{c}\approx0.2$ a very large maximal $EPI$ coupling $V_{epi}^{0}\approx10$ $eV$ is necessary. For $N(E_{F})\sim0.5$ $(eV)^{-1}$ the maximal coupling constant would be rather large, i.e. $\lambda_{epi}^{0}(=N(E_{F})V_{epi}^{0})\approx5$. Note, that $V_{epi}^{0}$ is almost as large as in the metallic hydrogen under high pressure $p\sim20$ $Mbar$, where $T_{c0}\approx600$ $K$ with large $EPI$ coupling constant $\lambda_{epi}^{0}\approx7$ - this important prediction is given in [@Maksimov-Savrasov]. In real $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$ materials the contribution of another pairing mechanism, which exists in the $FeSe$ film in absence of the substrate and is pronounced in the *s-wave channel* with $T_{c0}\approx8$ $K$, triggers the whole pairing to be $s$-wave. The latter only moderately decreases the contribution of the $EPI-FSP$ pairing mechanism. The existence of sharp replica bands in $1ML$ $FeSe/SrTiO_{3}$ and $1UCFeSe/TiO_{2}$ and large value of $V_{epi}^{0}$ imply inevitably that the $EPI-FSP$ pairing mechanism is the main candidate to explain superconductivity in these materials. We stress, that in $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$ high $T_{c0}$ is obtained on the expense of the large maximal $EPI$ coupling $V_{epi}^{0}$, which compensates the small (detrimental) phase volume factor $(aq_{c})^{2}$.
$(2)$ - The semi-microscopic model proposed in [@Lee-Interfacial-supplement], [@Johnston1-2], and refined slightly in this paper, contains phenomenological parameters, such as $n_{d}$ - the number of dipoles per unit cell, $q_{eff}$ - the effective dipole charge, $\varepsilon_{\parallel}^{eff}$, $\varepsilon_{\perp}^{eff}$ - effective parallel and perpendicular dielectric constanty in $SrTiO_{3}$ (and $TiO_{2}$) near the interface, respectively. For $T_{c0}\approx100$ $K$ and by assuming $aq_{c}\approx0.2,$ $q_{eff}\approx2e$, $n_{d}\sim2$ $/unit-cell$ makes $\varepsilon_{\parallel}^{eff}\sim30$, $\varepsilon_{\perp}^{eff}\sim1$. These values, which are physically plausible, are very far from $\varepsilon_{bulk}$ in the bulk $SrTiO_{3}$, where $\varepsilon_{bulk}\sim500-10^{4}$ (and $\varepsilon_{bulk}\leq260$ in the rutile $TiO_{2}$). We point out that our estimation of these parameters is based on the effective microscopic model where the bulk $SrTiO_{3}$ is truncated by a monolayer ($1ML$) made of $TiO_{2}$ [@Lee-Interfacial-supplement]. In reality it may happen that the bulk $SrTiO_{3}$ is truncated by two monolayers ($2ML$) of $TiO_{2}$, as it is claimed to be seen in the synchrotron $x$ray diffraction [@Zou-Bozovic].This finding is confirmed by the $LDA$ calculations in [@Zou-Bozovic], which show that for the $2ML$ $TiO_{2}$ structure: ($\emph{i}$) the electrons are much easier transferred to the $FeSe$ metallic monolayer and ($\emph{ii}$) the top of the hole band is shifted far below the electronic Fermi surface than in the $1ML$ model. If the $2ML$ of $TiO_{2}$ is realized it could be even more favorable for the $EPI-FSP$ pairing, since some parameters can be changed in a favorable way. For instance, the effective charge could be increased, i.e. $q_{eff}^{(2ML)}>q_{eff}^{(1ML)}$and since $T_{c0}\sim
q_{eff}^{2}$ the $2ML$ model may gives rise to higher critical temperature.
$(3)$ - The* isotope effect* in $T_{c0}$* *should be small ($\alpha_{O}\ll1/2$) since in leading order one has $T_{c0}\sim V_{epi}^{0}$, where $V_{epi}^{0}$ is *mass-independent*. This is contrary to [@Johnston1-2] where $\alpha_{O}=1/2$. The next leading order gives $\alpha_{O}\sim(T_{c0}/\Omega)<0.09$. We stress that the small isotope-effect maybe a smoke-gun experiment for the $EPI-FSP$ pairing mechanism.
$(4)$ - In the $EPI$-$FSP$ pairing theory* *the *non-magnetic impurities* affect both $s$-wave and $d$-wave pairing. In the case of $s$-wave* *they are* pair-weakening*, while *for* $d$-wave are* pair-breaking*. However, the non-magnetic impurities with forward scattering peak give in the “dirty” limit ($\Gamma_{F}\gg\pi
T_{c}$) the full isotope effect $\alpha_{O}=1/2$, since $T_{c}^{(s,d)}\sim\Omega\sim M^{-1/2}$. In that case, the nonanalicity of $T_{c}^{(s,d)}$ with respect to the impurity concentration $n_{i}$, would resolve the question - what kind of pairing is realized in $1ML$ $FeSe/SrTiO_{3}$ and $1UCFeSe/TiO_{2}$ - the $EPI-FSP$ or the standard $EPI$.
$(5)$ - In the case of the $EPI$-$FSP$ pairing the superconducting order parameter depends strongly on the internal pair coordinate and of center of mass, i.e. $\Delta=\Delta(\mathbf{r},\mathbf{R})$. The *internal pair fluctuations* reduce additionally the mean-field critical temperature so that in the interval $T_{c}<T<T_{c0}$ a pseudogap behavior is expected.
$(6)$ - The *EPI self-energy* in the normal state at $T=0$ and $\xi(\mathbf{k}_{F})=0$ is given by $\Sigma_{epi}(\mathbf{k},\omega
)\approx-\lambda_{m}\omega/(1-(\omega/\Omega)^{2})$, where $\lambda
_{m}=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/2\Omega$, which for $G^{-1}(\mathbf{k},\omega)=0$ gives the dispersion energy of the quasiparticle band $\omega_{1}=0$ and the replica bands $\omega_{2}$ and $\omega_{3}$. The ratio of the ARPES intensities of the replica band $\omega_{2}$ and the quasiparticle band $\omega_{1}$ at $T=0$ and at the Fermi surface ($k=k_{F}$) is given by $R(T=0,k_{F})=(A_{2}/A_{1})=\lambda_{m}/2$. This means, that for $\lambda_{m}\sim0.2$ the experimental value of $R(T=0,k_{F})$ should be $(A_{2}/A_{1})\approx0.1$. This ratio is slightly smaller than the experimental value $R(T\neq0,k=0)$ measured in [@Lee-Interfacial-supplement], [@Replica-bands].
$(7$) Since the coupling constant $\lambda_{m}$ is *mass-dependent*, $\lambda_{m}\sim M^{1/2}$ then the isotope effect in various quantities, in $1ML$ $FeSe/SrTiO_{3}$ and $1ML$ $FeSe/TiO_{2}$ systems, may be a smoke-gun experiment in favour of the $EPI-FPS$ theory. To remind the reader: $(i)$ $T_{c0}$ is almost *mass-independent*; $(ii)$ the self-energy slope at $\omega\ll\Omega$ is *mass-dependent*, $(-d\Sigma/d\omega)\sim M^{1/2}$; $(iii)$ the $ARPES$ ratio $R$ of the replica band intensities is *mass-dependent*, $R\sim M^{1/2}$.
Concerning the role of $EPI$ in explaining superconductivity in $1ML$ $FeSe/SrTiO_{3}$ there were other interesting theoretical proposals. In [@Gorkov1-2] the $EPI$ is due to the interaction with longitudinal optical phonons and since $\Omega>E_{F}$ the problem is studied in anti-adiabatic limit, where $T_{c}$ is also weakly dependent on the oxygen mass. In [@Cohen] the substrate gives rise to an antiferromagnetic structure in $FeSe$, which opens new channels in the $EPI$ coupling in the $FeSe$ monolayer, thus giving rise for high $T_{c}$. In [@Zi] the intrinsic pairing mechanism is assumed to be due to $J_{2}$-type spin fluctuations, or antiferro orbital fluctuation, or nematic fluctuations. The extrinsic pairing is assumed to be due to interface effects and the $EPI-FSP$ interaction. The problem is studied by the sign-free Monte-Carlo simulations and it is found that $EPI-FSP$ is an important ingredient for high $T_{c}$ superconductivity in this system.
Finally, we would like to comment some possibilities for designing new and complex structures based on $1ML$ $FeSe/SrTiO_{3}$ (or $1ML$ $FeSe/TiO_{2}$) as a basic unit. The first nontrivial one is when a double-sandwich structure with two interfaces is formed, i.e. $SrTiO_{3}/1ML$ $FeSe/SrTiO_{3}$ (or $TiO_{2}/1ML$ $FeSe/TiO_{2}$). Naively thinking in the framework of the $EPI-FSP$ pairing mechanism one expects in an “ideal” case doubling of $T_{c0}$, since phonons at two interfaces are independent. However, this would only happen when the electron-like bands on the Fermi surface due to the two substrates were similar and if the condition $q_{c}v_{F}<\pi T_{c0}$ is kept in order to deal with a sharp $FSP$. However, many complications in the process of growing, such structures may drastically change properties, leading even to a reduction of $T_{c0}$. It needs very delicate technology to control the concentration of oxygen vacancies and appropriate charge transfer at both interfaces. However, eventual solutions of these problems might give impetus for superconductors with exotic properties. For instance, having in mind the above exposed results on effects of non-magnetic impurities on $T_{c0}$, then by controlling and manipulating their presence at both interfaces one can design superconducting materials with wishful properties.
**Acknowledgment** M.L.K. is thankful to Radoš Gajić for useful discussions, comments and advises related to the experimental situation in the field. M.L.K. highly appreciates fruitful discussions with Steve Johnston and Yan Wang on ARPES of the replica bands at finite temperature, and on microscopic parameters of the theory.
Appendix
========
Migdal-Eliashberg equations in superconductors
----------------------------------------------
In the paper we study superconductivity with the $EPI-FSP$ mechanism of pairing by including effects of non-magnetic impurities, too. The full set of Migdal-Eliashberg equations is given for that case. The normal and anomalous Green’s functions are $G_{n}(\mathbf{k},\omega_{n})=$ $-[i\omega_{n}Z_{n}(\mathbf{k})+\bar{\xi}_{n}(\mathbf{k})]/D_{n}(\mathbf{k})$, $F_{n}(\mathbf{k})=-Z_{n}(\mathbf{k})\Delta_{n}(\mathbf{k})/D_{n}(\mathbf{k})$, respectively where $D_{n}(\mathbf{k})=[\omega_{n}Z_{n}(\mathbf{k})]^{2}+\bar{\xi}_{n}^{2}(\mathbf{k})+[Z_{n}(\mathbf{k})\Delta_{n}(\mathbf{k})]^{2}$ ($\omega_{n}=\pi T(2n+1)$). Here, $Z_{n}(\mathbf{k})$ is the wave-function renormalization defined by $i\omega_{n}(1-Z_{n}(\mathbf{k}))=(\Sigma(\mathbf{k},\omega_{n})-\Sigma(\mathbf{k},-\omega
_{n}))/2$, where the self-energy $\Sigma(\mathbf{k},\omega_{n})=\Sigma
_{epi}(\mathbf{k},\omega_{n})+\Sigma_{imp}(\mathbf{k},\omega_{n})$ describes the $EPI$ and impurity scattering, respectively. The energy renormalization is $\bar{\xi}_{n}(\mathbf{k})=\xi(\mathbf{k})+\chi_{n}(\mathbf{k})$, $\chi
_{n}(\mathbf{k})=(\Sigma(\mathbf{k},\omega_{n})+\Sigma(\mathbf{k},-\omega
_{n}))/2$ and $\Delta_{n}(\mathbf{k})$ is the superconducting order parameter.$$Z_{n}(\mathbf{k})=1+\frac{T}{\omega_{n}}\sum_{\mathbf{k}^{\prime},n^{\prime}}\frac{V_{eff}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})\omega_{n^{\prime}}Z_{n^{\prime}}(\mathbf{k}^{\prime})}{D_{n^{\prime}}(\mathbf{k}^{\prime})}
\label{Z-nk}$$$$\bar{\xi}_{n}(\mathbf{k})=\xi(\mathbf{k})-T\sum_{\mathbf{k}^{\prime},n^{\prime}}\frac{V_{eff}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})\bar
{\xi}_{n^{\prime}}(\mathbf{k}^{\prime})}{D_{n^{\prime}}(\mathbf{k}^{\prime})}
\label{Ksi-nk}$$$$Z_{n}(\mathbf{k})\Delta_{n}(\mathbf{k})=T\sum_{\mathbf{k}^{\prime},n^{\prime}}\frac{V_{eff}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})Z_{n^{\prime}}(\mathbf{k}^{\prime})\Delta_{n^{\prime}}(\mathbf{k}^{\prime})}{D_{n^{\prime}}(\mathbf{k}^{\prime})}, \label{Delta-nk}$$ where $V_{eff}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})=V_{epi}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})+V_{imp}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})$, $V_{epi}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})=-g_{epi}^{2}(\mathbf{k}-\mathbf{k}^{\prime})\mathcal{D}_{ph}(\mathbf{k}-\mathbf{k}^{\prime},\omega_{n}-\omega_{n^{\prime
}})$ and $V_{imp}(n-n^{\prime},\mathbf{k}-\mathbf{k}^{\prime})=\delta
_{nn^{\prime}}n_{imp}u^{2}(\mathbf{k}-\mathbf{k}^{\prime})/T$. Here, the phonon Green’s function in the Einstein model with the single frequency $\Omega$ is given by $\mathcal{D}_{ph}(\mathbf{k}-\mathbf{k}^{\prime},\omega_{n}-\omega_{n^{\prime}})=-2\Omega/(\Omega^{2}+(\omega_{n}-\omega_{n^{\prime}})^{2})$ while the impurity scattering is described in the Born-approximation. Here, $n_{imp}$ is the impurity concentration and $u(\mathbf{k}-\mathbf{k}^{\prime})$ is the impurity potential. To these three equations one should add the equation for the chemical potential $\mu$, i.e. $N=\sum G_{n}(\mathbf{k},\omega_{n};\mu)=const$. However, in the following we study only problems where the (small) change of $\mu$ due to $EPI$ and impurity scattering does not change the physics of the problem. For instance we do not study problems such as $BCS-BEC$ transition, where the equation for $\mu$ plays important role, etc.
Note, that in the case of systems with very *large Fermi energy* $E_{F}$ and with an *isotropic* $EPI$ ($Z_{n}(\mathbf{k})\equiv Z_{n}$, $\bar{\xi}_{n}(\mathbf{k})\rightarrow0$) one integrates over the energy $\xi_{\mathbf{k}^{\prime}}$ by introducing the density of states at the Fermi surface $N(0)$, i.e. $\sum_{\mathbf{k}^{\prime}}(...)\Rightarrow
N(0)\int_{-\infty}^{\infty}(...)d\xi_{\mathbf{k}^{\prime}}$. This leads to standard Migdal-Eliashberg equations.$$Z_{n}=1+\frac{\pi T}{\omega_{n}}\sum_{n^{\prime}}\frac{N(0)V_{eff}(n-n^{\prime})\omega_{n^{\prime}}Z_{n^{\prime}}}{\sqrt{(\omega_{n^{\prime}}Z_{n^{\prime}})^{2}+\Delta_{n^{\prime}}^{2}}} \label{Zn}$$$$Z_{n}\Delta_{n}=\pi T\sum_{n^{\prime}}\frac{N(0)V_{eff}(n-n^{\prime
},Z_{n^{\prime}}\Delta_{n^{\prime}}}{\sqrt{(\omega_{n^{\prime}}Z_{n^{\prime}})^{2}+\Delta_{n^{\prime}}^{2}}} \label{Delta-n}$$ In the case of *strongly momentum-dependent* $EPI-FSP$,$\ $where $V_{epi}(n-n^{\prime},\mathbf{q})$ is finite for $\left\vert \mathbf{q}\right\vert <q_{c}\ll k_{F}$, the Migdal-Eliashberg equations are given by$$Z_{n}(\xi)=1+\frac{T}{\omega_{n}}\sum_{m}\left\langle V_{epi}(n-m,\mathbf{q})\right\rangle _{q}\frac{\omega_{m}Z_{m}(\xi)}{D_{m}(\xi)}, \label{Zn-FSP}$$ $$\bar{\xi}_{n}(\xi)=\xi(\vec{k})-T\sum_{m}\frac{\left\langle V_{epi}(n-m,\mathbf{q})\right\rangle _{q}}{D_{m}(\xi)}\bar{\xi}_{m}(\xi),
\label{Ksi-FSP}$$$$Z_{n}(\xi)\Delta_{n}(\xi)=T\sum_{m}\left\langle V_{epi}(n-m,\mathbf{q})\right\rangle _{q}\frac{Z_{m}(\xi)\Delta_{m}(\xi)}{D_{m}(\xi)}.
\label{Delta-n-FSP}$$
Effects of non-magnetic impurities on $T_{c0}$ in the $EPI-FSP$ theory
----------------------------------------------------------------------
In this paper we study the superconductivity which is due to $EPI-FSP$ of the Einstein phonon with $\Omega^{2}\ll(2\pi T_{c0})^{2}$. In that case $\left\langle V_{epi}(n-m,\mathbf{q})\right\rangle _{q}\approx\left\langle
V_{epi}(0,\mathbf{q})\right\rangle _{q}=\left\langle 2g_{epi}^{2}(\mathbf{q})\right\rangle _{q}/\Omega$ and the contribution to $Z_{n}(\xi)$ is $\sim\lambda_{m}=\left\langle V_{epi}(0,\mathbf{q})\right\rangle _{q}/2\Omega
$. Since in the weak coupling limit one has $\lambda_{m}\ll1$ then we neglect this contribution. Also the non-Migdal corrections can be neglected in this case - see [@Johnston1-2] The effects of non-magnetic impurities on $T_{c0}$ is studied in the standard model with weakly momentum dependent impurity potential $u(\mathbf{k}-\mathbf{k}^{\prime})\approx const$ . In that case $Z_{n}(\xi)$ contains the impurity term only. After the integration of the impurity part over the energy $\xi^{\prime}$ in $Eqs.$(\[Z-nk\]-\[Delta-nk\]) -see [@Allen-Mitrovic], and putting $\xi=0$ (since in that case $\Delta_{n}(\xi=0)$ is maximal) one obtains ($D_{m}(\xi
)\approx\omega_{n}^{2}Z_{n}^{2}$) for the *s-wave pairing* ($\Delta=const$) $$Z_{n}=1+\frac{\Gamma}{\left\vert \omega_{n}\right\vert } \label{Zn-imp}$$$$Z_{n}\Delta_{n}=T_{c}\sum_{m}\left\langle V_{epi}(n-m,\mathbf{q})\right\rangle
_{q}\frac{Z_{m}\Delta_{m}}{D_{m}(0)}+\frac{\Gamma}{\left\vert \omega
_{n}\right\vert }\Delta_{n}. \label{Delta-n-imp}$$ Note, the the second term on the right side cancels the same term on the left side. In the approximation $\Delta_{n}(\xi)\approx\Delta$ one obtains the equation for impurity dependence of $T_{c}^{(s)}(\Gamma)$ for a *s-wave superconductor* $$1=T_{c}^{(s)}\left\langle V_{epi}(0,\mathbf{q})\right\rangle _{q}\sum_{m}\frac{1}{\omega_{n}^{2}Z_{n}}. \label{Tc-s-wave}$$ We point out that in the case of *d-wave superconductivity* $\Delta=\Delta(\varphi)$ is angle dependent on the Fermi surface and changes sign. In that case the last term in $Eq.$(\[Delta-n-imp\]) $\Delta$ should be replaced by $\left\langle \Delta(\varphi\right\rangle )=0$ giving equation for $T_{c}^{(d)}$$$1=T_{c}^{(d)}\left\langle V_{epi}(0,\mathbf{q})\right\rangle _{q}\sum_{m}\frac{1}{\omega_{n}^{2}Z_{n}^{2}}. \label{Tc-d-wave}$$ Note, $Z_{n}$ vs $Z_{n}^{2}$ renormalization for the $s$-wave and $d$-wave superconductivity, respectively.
$EPI-FSP$ self-energy in the normal state
-----------------------------------------
We shall calculate the self-energy at $T=0$. The leading order self-energy (on the Matsubara axis) in the Migdal-Eliashberg theory of $EPI$ is given by $$\Sigma_{epi}(\mathbf{k},\omega_{n})=-T\sum_{\mathbf{q},\Omega_{m}}g_{epi}^{2}(\mathbf{q})\mathcal{D}_{ph}(\mathbf{q},\Omega_{m})G(\mathbf{k}+\mathbf{q},\omega_{n}-\Omega_{m}), \label{Sigma-ME}$$ where $\omega_{n}=\pi T(2n+1)$ and $\Omega_{m}=2\pi mT$, $\mathcal{D}_{ph}(\mathbf{q},\Omega_{m})=-2\Omega/(\Omega^{2}+\Omega_{m}^{2})$, $G(\mathbf{k},\omega_{n})=1/(i\omega_{n}-\xi_{\mathbf{k}})$. By defining $V_{epi}(\mathbf{q},0)=2g_{epi}^{2}(\mathbf{q})/\Omega$ and after summation over $\Omega_{m}$ in $Eq.$(\[Sigma-ME\]) one obtains (note that $V_{epi}(\mathbf{q},0)=V_{epi}(-\mathbf{q},0)$ $$\begin{aligned}
\Sigma_{epi}(\mathbf{k},\omega_{n}) & =\frac{\Omega}{2}\sum_{\mathbf{q}}V_{epi}(\mathbf{q},0)\label{Sigma-integ}\\
& \times\left[ \frac{n_{F}(\xi_{\mathbf{k+q}})}{i\omega_{n}-\xi
_{\mathbf{k+q}}+\Omega}+\frac{1-n_{F}(\xi_{\mathbf{k+q}})}{i\omega_{n}-\xi_{\mathbf{k+q}}-\Omega}\right] .\nonumber\end{aligned}$$ Let us calculate $\Sigma_{epi}$ at $\mathbf{k}_{F}$. Since $\mathbf{qv}_{F}=qv_{F}\cos\theta$ and by taking into account that $n_{F}(\xi
_{\mathbf{k}_{F}\mathbf{+q}})=1$ for $\cos\theta<0$, $n_{F}(\xi_{\mathbf{k}_{F}\mathbf{+q}})=0$ for $\cos\theta>0$ one obtains for $q_{c}v_{F}\ll\Omega$ $$\Sigma_{epi}(\mathbf{k},\omega_{n})=-\lambda_{m}\frac{i\omega_{n}}{1-(\frac{i\omega_{n}}{\Omega})^{2}}, \label{Sigma-final}$$ where $\lambda_{m}=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/2\Omega$ and $\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}=Ns_{c}(2\pi)^{-2}\int
d^{2}qV_{epi}(\mathbf{q},0)$, $s_{c}$ is the surface of the $FeSe$ unit cell. Note, that $V_{epi}(\mathbf{q},0)=2g_{epi}^{2}(\mathbf{q})/\Omega$ and $g_{epi}(\mathbf{q})(=(g_{0}/\sqrt{N})e^{-q/q_{c}})$ so that $N$ disappears from $\lambda_{m}(=\left\langle V_{epi}(\mathbf{q})\right\rangle _{q}/2\Omega)$.
[99]{}
M. L. Kulić, R. Zeyher, Phys. Rev. [**B** 49]{}, 4395 (1994); R. Zeyher, M. L. Kulić, Phys. Rev. [**B** 53]{}, 285 (1996)
M. L. Kulić, Phys. Reports [**3**38]{}, 1-264 (2000); M. L. Kulić, O. V. Dolgov, phys. stat. sol. (b) **242**, 151 (2005)
O. V. Danylenko, O. V. Dolgov, M. L. Kulić, V. Oudovenko, Europ. Phys. Jour. [**B**9]{} - Cond. Matter, 201 (1999)
J. J. Lee, F. T. Schmitt, R. G. Moore, S. Johnston, Y. T. Cui, W. Li, M. Yi, Z. K. Liu, Y. Zhang, D. H. Lu, T. P. Devereaux, D.-H. Lee, Z. X. Shen, Nature **515**, 245 (2014)
L. Rademaker, Y. Wang, T. Berlijn, S. Johnston, New J. Phys. **18**, 022001 (2016); Y. Wang, K. Nakatsukasa, L. Rademaker, T. Berlijn, S. Johnston, Supercond. Sci. Technol. **29**, 054009 (2016)
S. N. Rebec, T. Jia, C. Zhang, M. Hashimoto, D. -H. Lu, R. G. Moore, Z. -X. Shen, arXiv: 1606.09358v1
V. L. Ginzburg, Uspekhi Fiz. Nauk **95**, 91 (1968); **101**, 185 (1970); **118**, 316 (1976)
O. V. Dolgov, D. A. Kirzhnits, E. G. Maksimov, Rev. Mod. Phys. **53**, 81 (1981)
Wang Quing-Yan, Li Zhi, Zhang Wen-Hao, Zhang Zuo-Cheng, Zhang Jin-Song, Li Wei, Ding Hao, Ou Yun-Bo, Deng Peng, Ghang Kai, Wen Jing, Song Can-Li, He Ke, Jia Jin-Feng, Ja Shuai-Hua, Wang Ya-Yu, Wang Li-Li, Chan Xi, Ma Xu-Cun, Xue Qi-Kun, Chin. Phys. Lett. **29**, 037402 (2012)
P. J. Hirschfeld, M. M.Korshunov, I. I. Mazin, Rep. Prog. Phys. **74**, 124508 (2011); P. J. Hirschfeld, arXiv: 1510.01386v1
L. Boeri, O. V. Dolgov, A. A. Golubov, Phys. Rev. Lett. , **101**, 026403 (2008)
M. L. Kulić, S. -L. Drechsler, O. V. Dolgov, EPL **85**, 47008 (2009)
M. L. Kulić, A. A. Haghighirad, EPL **87**, 17007 (2009)
M. L. Kulić, Ginzburg Conference on Physics, May 28-June 2, 2012, Moscov, Online Proceedings, http://gc.lpi.ru/proceedings/kulic.pdf
M. Rahlenbeck, G. L. Sun, D. L. Sun, C.T. Lin, B. Keimer, C. Ulrich, Phys. Rev. B **80**, 064509 (2009)
Z. P. Yin, A. Kutepov, G. Kotliar, Phys. Rev. X **3**, 021011 (2013)
E. G. Maksimov, M. L. Kulić, O. V. Dolgov, Adv. Cond. Mat. Phys. **2010**, 423725 (2010)
R. Peng, H. C. Xu, S. Y. Tan, H. Y. Cao, M. Xia, X. P. Shen, Z. C. Huang, C. H. Wen, Q. Song, T. Zhang, B. P. Xie, X. G. Gong, D. L. Fang, Nat. Commun. **5**, 5044 (2014)
M. V. Sadovskii, arXiv: 1605.04426v2; M. V. Sadovskii, E. Z. Kuchinskii, I. A. Nekrasov, JMMM **324**, 3481 (2010); I. A. Nekrasov, M. V. Sadovskii, Pisma v Zhetf **99**, 687 (2014)
A. Linscheid, S. Maiti, Y. Wang, S. Johnston, P. J. Hirschfeld, arXiv: 1603.03739v1; X. Chen, S. Maiti, A. Linscheid, P. J. Hirschfeld, arXiv: 1508.04782v1;
L. D. Landau, E. M. Lifshitz, Electrodynamics of Continuous Media (Pergamon, Oxford, 1989)
It turns out that in the first arXiv version of our paper we *overlooked* the fact that in [@Lee-Interfacial-supplement], [@Johnston1-2] they study the problem at finite temperature $T\neq0$ while our study is limited to $T=0$. This means that both, the $T\neq0$ and $T=0$, results are correct.
B. Murta, A. M. Garcia-Garcia, arXiv:1607.00055v1
K. Yang, S. L. Sondhi, Phys. Rev. B **62**, 11 778 (2000)
E. G. Maksimov, D. Yu. Savrasov, Solid State Comm. **119**, 569 (2011)
Ke Zou, S. Mandal, S. A. Albright, R. Peng, Yu. Pu, D. Kumah, C. Lau, G. H. Simon, O. E. Dagdeviren, Xi He, I. Božović, U. D. Schwarz, E. I. Altman, D. Feng, F. J. Walker, S. I. Beigi, C. H. Ahn, Phys. Rev. B **93**, 180506(R) (2016)
L. P. Gor’kov, Phys. Rev. B **93**, 060507 (2016); ibid **93**, 054517 (2016)
S. Coh, M. L. Cohen, S. G. Louie, New J. Phys. **17**, 073027 (2015)
Zi-Xiang Li, Fa Wang, Hong Yao, Dung-Hai Lee, arXiv:1512.206179v4
P. B. Allen, B. Mitrović, in Solid State Physics, Vol. 37, eds. H. Ehrenreich, F. Seitz, D. Turnbull, (Academic Press, New York, 1982), pp. 1-82
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe a new version of continuous variables quantum holographic teleportation of optical images. Unlike the previously proposed [@Sokolov01; @Gatti04] scheme, it is based on the continuous variables quantum entanglement between the light fields of different frequencies and allows for the wavelength conversion between the original and the teleported images. The frequency tunable holographic teleportation protocol can be used as a part of light-matter interface in quantum [*parallel*]{} information processing and [*parallel quantum memory*]{}.'
author:
- 'Liubov V. Magdenko, Ivan V. Sokolov'
- 'Mikhail I. Kolobov'
title: Quantum teleportation of optical images with frequency conversion
---
quant-ph/0602200
Ø
Continuous variables quantum teleportation allows to transfer an arbitrary quantum state of the electromagnetic field between two spatially separated systems via an exchange of classical information in combination with quantum entanglement shared by these systems. Initially the protocol of continuous variable teleportation was proposed theoretically [@Vaidman94; @Braunstein98a] and realized experimentally[@Furusawa98; @Bowen03] for a single spatial mode of the electromagnetic field. Recently this protocol was generalized theoretically to spatially multimode electromagnetic fields [@Sokolov01; @Gatti04]. The important feature of spatially multimode teleportation scheme is that it allows for simultaneous parallel teleportation of optical images (still or time-varying) containing large number of elements or pixels. The proposed scheme was called [*holographic teleportation*]{} because it resembles the conventional holography with an important difference that the reconstructed image is a quantum copy of the original one with fidelity that can be made close to unity. It is impossible to achieve high fidelity in holographic teleportation without sharing multimode quantum entanglement by the original and the target systems.
One of possible applications of quantum teleportation is the light-matter interface for quantum memory[@Polzik04] for light which allows to record and story a quantum state of light on that of an atomic ensemble. The generalized spatially multimode teleportation scheme opens new perspectives for creation of a [*parallel quantum memory*]{} for parallel processing of quantum information. For an efficient interaction between the light wave and the atomic medium it is desirable to have a possibility for tuning the optical frequency of light to that of the atomic transition without changing the quantum state of the electromagnetic field. Let us mention a proposal of quantum frequency conversion[@Kumar92] that was the first demonstration of quantum state transfer into different frequency. In recent experiment[@Tanzilli05] quantum transfer of qubits was successfully demonstrated between photons of wavelength 1310 nm and 710 nm.
In this letter we describe a new version of holographic teleportation of optical images with [*frequency conversion*]{}. This means that an input image, carried by a light wave with optical frequency $\omega_1$, is teleported to the output image at different optical frequency $\omega_2$ with preservation of its original quantum state. Similar to the frequency-preserving quantum holographic teleportation, discussed in Refs. [@Sokolov01; @Gatti04], to obtain high fidelity in the teleported image one needs to create spatially multimode entanglement between the original and the target systems. In our case the source of spatially multimode entanglement between the input image at frequency $\omega_1$ and the teleported image at frequency $\omega_2$ is a type-I traveling-wave non-degenerate optical parametric amplifier (OPA).
The optical scheme of holographic teleportation with frequency conversion if shown in Fig. 1. An input image which is to be teleported from Alice to Bob, is described by a slowly-varying field operator ${\hat A}^{\rm in}_1(\r,t)$, where $\r=(x,y)$ is a two-dimensional transverse coordinate. An input electromagnetic wave carrying this image is splitted into two secondary waves by a 50/50 beam splitter BS$_1$, and two quadrature components of these secondary waves are homodyne detected by two difference detectors using two local oscillators LO$_x$ and LO$_y$, two 50/50 beam splitters BS$_2$ and BS$_3$, and four efficient CCD cameras with appropriate spatial resolution.
{width="7cm"}
The local oscillator waves have the same frequency $\omega_1$ as the carrier frequency of the input image. The slow-varying difference photocurrent densities from these CCD cameras, containing information about spatio-temporal quantum fluctuations of the input image, are transmitted from Alice to Bob and are used for preparation of the output field ${\hat A}_2(\r,t)$. The multi-channel modulators M$_x$ and M$_y$ perform spatio-temporal modulation of a coherent wave with different carrier frequency $\omega_2$. This part of the scheme can be viewed as a classic non-stationary holography, where the reconstructing wave has different wavelength.
A key ingredient of our teleportation scheme is a pair of spatially-multimode EPR fields ${\hat E}_n(\r,t)$, $n=1,2$. Since an entanglement between different carrier frequencies is needed, the fields are created by a type-I traveling-wave [*non-degenerate*]{} OPA. The dichroic mirror M$_1$ reflects the wave at frequency $\omega_1$ and transmits that at frequency $\omega_2$. In the case of perfect reflectivity there are no frequency matched vacuum fluctuations entering from the open ports of M$_1$ into the corresponding fields.
Let us introduce slowly-varying spatio-temporal annihilation and creation operators ${\hat E}_n(\r,t)$ and ${\hat
E}^{\dag}_n(\r,t)$, $n=1,2$ of the electromagnetic waves with central frequencies $\omega_1$ and $\omega_2$ at the output of the non-degenerate OPA. The frequencies $\omega_1$ and $\omega_2$ obey the condition of energy conservation, $\omega_1+\omega_2 =
\omega_p$, where $\omega_p$ is the frequency of the pump wave. The field operators are normalized so that $\langle {\hat
E}_n^{\dag}({\r},t) {\hat E}_n({\r},t) \rangle$ gives the mean value of the irradiance, expressed in photons per ${\rm cm}^2$ per second.
The transformation of the input fields ${\hat A}_n(\r,t)$ of the non-degenerate OPA in the vacuum state into the output fields ${\hat
E}_n(\r,t)$ in the broadband multimode squeezed state is described in terms of the Fourier components of these operators in frequency and spatial-frequency domain, ${\hat E}(\r,t) \rightarrow {\hat
e}(\q,\Omega)$. The squeezing transformation performed by a non-degenerate OPA, can be written as follows: \_n(,) = U\_n(,) [a]{}\_n(,) + V\_n(,) [a]{}\_[n’]{}\^(-,-), \[squeezing\] where $n=1,2$, $n\neq n'$, the coefficients $U_n(\q,\Omega)$ and $V_n(\q,\Omega)$ depend on the amplitude of the pump field, nonlinear susceptibility and the phase-matching condition. The explicit form of $U_n(\q,\Omega)$ and $V_n(\q,\Omega)$ can be found, for example, in Ref. [@Brambilla04]. The EPR correlations between the fields ${\hat E}_1(\r,t)$ and ${\hat
E}_2(\r,t)$ are determined by two parameters, namely, the orientation angle $\psi_n(\q,\Omega)$ of the major axis of the squeezing ellipse [@Kolobov99] and the degree of squeezing $r_n(\q,\Omega)$, \_n(,) = {U\_n(,) V\_[n’]{}(-,-)}, \[psi\] = |U\_n(,)| |V\_[n’]{} (,)|. \[exp\_r\] It should be noted that in Fig. 1 we have chosen the input field at the frequency $\omega_1$ and the output field at the frequency $\omega_2$, that corresponds to $n=1$ and $n'=2$ in Eqs. (\[psi\]), (\[exp\_r\]), but with the same source of entangled beams the teleportation with frequency conversion can be performed from $\omega_2$ to $\omega_1$ as well.
It is instructive for better understanding of the scheme to give a simple physical explanation for the continuous variables quantum entanglement arising from squeezing in [*non-degenerate*]{} parametric down-conversion. For simplicity we shall omit the spatial dependence of the fields. Let us introduce the sum of two slow amplitudes, ${\cal E}(t) \sim E_1(t)+
E_2(t)$ and consider a contribution to ${\cal E}(t)$ from a pair of Fourier amplitudes, (t) \~e\_1(-) e\^[it]{} + e\_2() e\^[-it]{}. \[pair\_amplitudes\] One can show [@Magdenko06], that the discussed above squeezing ellipse for given $\Omega$ represents exactly the quantum uncertainty region for this contribution. For perfect squeezing this automatically implies a perfect phase matching beyond the classical limit between the slow amplitudes in the right side of (\[pair\_amplitudes\]), as illustrated in Fig. 2a. Let us remind that these amplitudes were defined with respect to the different carrier frequencies $\omega_1$ and $\omega_2$. Hence the corresponding beams can be split by a dichroic mirror, preserving their quantum entanglement.
{width="7.5cm"}
The first slow amplitude is imprinted into the photocurrents by the homodyne detection and into the reconstructing field at different frequency $\omega_2$ by the modulation. The second one [*ab initio*]{} is related to this frequency. A proper matching of the relative amplitudes and phases in quantum and classical channels leads to the quantum noise cancellation and effective teleportation.
Next important point is that the quantum modulation of the output field at a given frequency is now due to two independent pairs of entangled amplitudes, the first pair is shown in (\[pair\_amplitudes\]), another one corresponds to $\Omega \rightarrow -\Omega$. Therefore, in order to achieve optimal results one has to take care of a proper phase matching of two independent squeezing ellipses (at $\Omega$ and $-\Omega$). The frequency dependence of squeezing within the frequency range of effective non-degenerate parametric down-conversion is graphically shown in Fig. 2b.
Apart from these new features, introduced by the frequency non-degenerate nature of the process, the fields evolution in the scheme is very similar to that described in Refs. [@Sokolov01; @Gatti04]. The input field ${\hat A}^{\rm in}_1(\r,t)$ is mixed with one EPR beam ${\hat E}_1(\r,t)$ at the 50/50 beam splitter BS$_1$. The secondary waves ${\hat B}_x$ and ${\hat B}_y$, \_[x,y]{}(,t) = (\^[in]{}\_1(,t) + [E]{}\_1(,t)), with the $+(-)$ sign corresponding to $x(y)$ channel, are photodetected by means of two balanced homodyne detectors. The difference photocurrents $I_x(\r,t)$ and $I_y(\r,t)$ collected from individual pixels of these matrices, carry the information about the spatio-temporal quantum fluctuations of the quadrature components of ${\hat B}_x(\r,t)$ and ${\hat B}_y(\r,t)$, phase matched with LO$_x$ and LO$_y$. These photocurrents are sent from Alice to Bob via two multichannel classical communication lines and are used by Bob for spatio-temporal modulation of an external coherent wave with the frequency $\omega_2$, by means of two multichannel modulators $M_x$ and $M_y$. The teleported field ${\hat A}^{\rm out}_2(\r,t)$ is created by mixing of this modulated wave with a second EPR wave ${\hat E}_2(\r,t)$ at the mirror M with high reflectivity. By choosing appropriately the mirror transmission and the modulation depth, the teleported field ${\hat A}^{\rm out}_2(\r,t)$ can be obtained as \^[out]{}\_2(,t) = [A]{}\^[in]{}\_1(,t)+[ F]{}(,t), \[field\_out\] where ${\hat F}(\r,t)$ is an operator describing the noise added to the teleported field, (,t) = [E]{}\_2(,t) + [E]{}\_1\^(,t). \[noise\] It can be demonstrated that the added noise can be considered as classical, thus preserving the commutation relations for the output field.
An ideal teleportation from $\omega_1$ to $\omega_2$ would correspond to the situation when the output field operator at different spatial points $\r$ and at different time moments $t$ is an exact copy of the input field operator, ${\hat A}^{\rm
out}_2(\r,t) = {\hat A}^{\rm in}_1(\r,t)$. However, as explained in Refs. [@Sokolov01; @Gatti04], this would require an infinite energy of the EPR beams and therefore could never be realized in practice. In real experimental situation the teleportation process will never take place “point-to-point”, but on average within some finite spatial area and within some finite time interval. In order to quantitatively describe the performance of our teleportation scheme we introduce a coarse-grained description of the input and output fields. Precisely, we consider the averaged field operator over a square pixel $S_j$ of area $S=\Delta^2$ and over a time interval $T_i$ of duration $T$: \^[out]{}\_2(j,i)= \_[S\_j]{}d \_[T\_i]{} dt [A]{}\^[out]{}\_2(,t), \[averaged\_field\] with analogous definitions for the input field. The averaged field operators obey standard commutation relations of discrete field oscillators. The quadrature components of the averaged field operators at the output are \^[out]{}\_(j,i) &=& [A]{}\_2\^[out]{}(j,i) e\^[-i]{} + h.c.,\
[Y]{}\^[out]{}\_(j,i) &=& -i [A]{}\_2\^[out]{}(j,i) e\^[-i]{} + h.c., \[averaged\_quadratures\] and similarly for the input field operators.
To characterize the noise added in the teleportation process we compare the correlation functions of the input and output quadrature components defined in Eq. (\[averaged\_quadratures\]). As follows from Eq. (\[field\_out\]), the relation between these correlation functions is \_\^[out]{} (j,i)\_\^[out]{}(j’,i’)= \[correlation\] $$\langle \delta {\hat X}_\varphi^{\rm in} (j,i)\,
\delta {\hat X}_\varphi^{\rm in} (j',i') \rangle
+ {\cal C} (j,j'; i,i'),$$ where ${\cal C} (j,j'; i,i')$ is the covariance matrix of the added noise. The explicit expression of this covariance matrix is ( j,j’ ; i, i’) = 2 B\_() B\_T () \[covariance\] $$\cos{ \left[ \q\cdot(\r_{j} - \r_{j'}) -\Omega (t_i -t_{i'}) \right] }
G(\q,\Omega),$$ where $\r_j=\{x_j,y_j\}$ is the center of the j-th pixel, and $t_i$ is the center of the i-th time interval. The functions $B_\Delta(\q\,)$ and $B_T(\Omega)$ arise from the coarse-graining operation, and for a square pixel of size $\Delta$ they read, B\_() &=& ( ) [sinc\^2]{} ( ),\
B\_T(Ø) &=& ( ). \[B\_delta\_T\] The Green function $G(\q,\Omega)$ expressed in terms of the orientation angle $\psi_2(\q,\Omega)$ and the degree of squeezing $r_2(\q,\Omega)$ from Eq. (\[psi\]) looks like G(,) = e\^[2r\_2(,)]{} \^2\_2(,) + e\^[-2r\_2(,)]{} \^2\_2(,). \[green\] As follows from Eqs. (\[covariance\])-(\[green\]), the covariance matrix ${\cal C} (j,j'; i,i')$ is independent of the phase $\varphi$ in Eq. (\[averaged\_quadratures\]). Thus, the added noise is the same for any quadrature.
In the absence of the EPR correlations, i. e. when $r_2(\q,\Omega)=0$, we obtain the classical limit of teleportation with the covariance matrix ${\cal C}_{\rm cl} (j,j'; i,i') =
2\delta_{jj'}\delta_{ii'}.$ In this limit two units of vacuum noise are added at each pixel exactly as for a single-mode teleportation and for the teleportation of images without frequency conversion.
Choosing the phase matching condition in the OPA such that $\psi_2(0,0)=\pi/2$, we obtain reduction of the quantum noise below the standard quantum level within some bandwidth of spatial frequencies $\q$ and temporal frequencies $\Omega$ determined by the phase-matching conditions in the crystal. When the pixel size $\Delta$ and the time window $T$ are much larger than the characteristic coherence length $l_c$ and the coherence time $T_c$ of the OPA, we obtain \_[ , T ]{} [C]{}\_[cl]{} (j,j’; i,i’)=2\_[jj’]{}\_[ii’]{} . \[quant\_limit\] In Fig. 3 we illustrate the role of the pixel size $\Delta$ and of the integration time $T$ in the teleportation process with frequency conversion. This figure shows the diagonal elements ${\cal C} (j,j; i,i)$ of the covariance matrix as a function of the relative pixel size $D=\Delta/l_c$ for three different observation times $T$ equal to $10T_c$, $T_c$, and $0.1T_c$. The coherence time $T_c$ for a frequency non-degenerate OPA is typically estimated as the time delay at the crystal length $l$ between two wave packets centered at the frequencies $\omega_1$ and $\omega_2$, $T_c \sim l |1/v_1 - 1/v_2|$, arising due to the difference of the group velocities $v_n$. Since a typical frequency spectrum of parametric down-conversion is fairly broad, a wave-packet spread due to the group velocity dispersion can also have an effect on squeezing and entanglement in our teleportation scheme, as we shall discuss elsewhere [@Magdenko06].
{width="7cm"}
The frequency dependence of squeezing orientation angle, arising in a parametric crystal and illustrated in Fig. 2b, is a pure phase effect and can be compensated by propagation in a linear medium with a properly chosen frequency dependence of the refraction index. It is worth noting that in many interference experiments with twin photons, a need for similar compensation has been realized some time ago. Such a compensation can be applied to the plots shown in Fig. 3 and will be illustrated in the forthcoming publication [@Magdenko06].
The coherence time for a non-degenerate OPA is usually much larger than for a degenerate one. Therefore, the assumption $T\gg T_c$ used in Refs. [@Sokolov01; @Gatti04] is not necessarily true anymore. The role of the observation time $T$ is as crucial in the teleportation process as the role of the pixel size $\Delta$. Indeed, as follows from Fig. 3, the diagonal elements of the covariance matrix which characterize the added noise power, decrease with growing pixel size $\Delta$ from the classical limit ${\cal C} (j,j; i,i)=2$ to the EPR limit ${\cal C} (j,j;
i,i)=2\exp[-2r(0,0)]$ when the observation time is large compared with the coherence time $T_c$. However, when the observation time $T$ becomes comparable with the coherence time $T_c$ or smaller, the diagonal elements ${\cal C} (j,j; i,i)$ never reach the EPR limit but remain close to the classical limit even for large pixel size.
This results allow to estimate an effective number of spatio-temporal degrees of freedom for an input non-stationary image that can be teleported in our scheme.
In conclusion we have proposed a new scheme of quantum holographic teleportation of optical images from one optical frequency to another. We have presented preliminary analysis of the performance of our scheme and have revealed the possibilities to achieve high quality teleportation. More detailed calculations including evaluation of multi-pixel fidelity for this scheme are in progress and will be published elsewhere. This work was supported by the Network QUANTIM (IST-2000-26019) of the European Union and by the INTAS under Project No. 2001-2097. The research was performed within the framework of GDRE “Lasers et techniques optiques de l’information”.
I. V. Sokolov, M. I. Kolobov, A. Gatti, and L. A. Lugiato, Optics Communications, [**193**]{}, 175 (2001).
A. Gatti, I. V. Sokolov, M. I. Kolobov, and L. A. Lugiato, Eur. Phys. J. D [**30**]{}, 123 (2004).
L. Vaidman, Phys. Rev. A [**49**]{}, 1473 (1994).
S. L. Braunstein and H. J. Kimble, Phys. Rev. Lett. [**80**]{}, 869 (1998).
A. Furusawa, J. L. Sorensen, S. L. Braunstein, C. A. Fuchs, H. J. Kimble, and E. S. Polzik, Science, [**282**]{}, 706 (1998).
W. P. Bowen, N. Treps, B. C. Buchler, R. Schnabel, T. C. Ralph, H.-A. Bachor, T. Symul, and P. K. Lam, Phys. Rev. A [**67**]{}, 032302 (2003).
B. Julsgaard, J. Sherson, J. Fiurasek, J. I. Cirac, and E. S. Polzik, Nature [**432**]{}, 482 (2004).
J. Huang and P. Kumar, Phys. Rev. Lett. [**68**]{}, 2153 (1992).
S. Tanzilli, W. Tillel, M. Halder, O. Alibart, P. Baldi, N. Gisin, and H. Zbinden, Nature [**437**]{}, 116 (2005).
M. I. Kolobov, Rev. Mod. Phys. [**71**]{}, 1539 (1999).
E. Brambilla, A. Gatti, M. Bache, and L. A. Lugiato, Phys. Rev. A [**69**]{}, 023802 (2004).
L. V. Magdenko, M. I. Kolobov, and I. V. Sokolov (in preparation).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a family of second-order parabolic systems with bounded measurable, rapidly oscillating and time-dependent periodic coefficients, we investigate the sharp convergence rates of weak solutions in $L^2$. Both initial-Dirichlet and initial-Neumann problems are studied.'
author:
- 'Jun Geng[^1]Zhongwei Shen[^2]'
bibliography:
- 'convergence.bib'
title: |
Convergence Rates in Parabolic Homogenization\
with Time-Dependent Periodic Coefficients
---
**Introduction**
================
The primary purpose of this paper is to investigate the sharp convergence rates in $L^2$ for a family of second-order parabolic operators $\partial_t+\mathcal{L}_\varepsilon$ with bounded measurable, rapidly oscillating and time-dependent periodic coefficients. Both the initial-Dirichlet and initial-Neumann boundary value problems are studied. Specifically, we consider $$\label{elliptic operator}
\mathcal{L}_\varepsilon=-\text{div}\left(A\big({x}/{\varepsilon},{t}/{\varepsilon^2}\big)\nabla
\right),$$ where $\e>0$ and $A(y,s)= \big(a_{ij}^{\alpha\beta} (y,s)\big)$ with $1\le i, j\le d$ and $1\le \alpha, \beta \le m$. Throughout this paper we will assume that the coefficient matrix $A(y,s)$ is real, bounded measurable, and satisfies the ellipticity condition, $$\label{ellipticity}
\mu |\xi|^2\le a^{\alpha\beta}_{ij}(y,s)\xi_i^\alpha
\xi^\beta_j\leqslant \frac{1}{\mu}|\xi|^2
\quad \text{ for any }
\xi=(\xi_i^\alpha ) \in \mathbb{R}^{m\times d} \text{ and a.e. } (y,s)\in \mathbb{R}^{d+1},$$ where $\mu >0$, and the periodicity condition, $$\label{periodicity}
A(y+z,s+t)=A(y,s)~~~\text{ for }(z,t)\in \mathbb{Z}^{d+1}\text{ and a.e. }(y,s)\in \mathbb{R}^{d+1}.$$ No additional smoothness condition will be imposed on $A$.
Let $\Omega\subset\mathbb{R}^d$ be a bounded domain and $0<T<\infty$. We are interested in the initial-Dirichlet problem, $$\label{IDP}
\left\{\aligned
(\partial_t +\mathcal{L}_\e) u_\e &= F &\quad &\text { in } \Omega \times (0, T),\\
u_\e & = g & \quad & \text{ on } \partial\Omega \times (0, T),\\
u_\e &=h &\quad &\text{ on } \Omega \times \{ t=0\},
\endaligned
\right.$$ and the initial-Neumann problem, $$\label{INP}
\left\{\aligned
(\partial_t +\mathcal{L}_\e) u_\e &= F &\quad &\text { in } \Omega \times (0, T),\\
\frac{\partial u_\e}{\partial \nu_\e} & = g & \quad & \text{ on } \partial\Omega \times (0, T),\\
u_\e &=h &\quad &\text{ on } \Omega \times \{ t=0\},
\endaligned
\right.$$ where $\displaystyle \left(\frac{\partial u_\e}{\partial \nu_\e}\right)^\alpha
=n_i a_{ij}^{\alpha\beta} (x/\e, t/\e^2)\frac{\partial u_\e^\beta}{\partial x_j}$ denotes the conormal derivative of $u_\e$ associated with $\mathcal{L}_\e$ and $n=(n_1, \dots, n_d)$ is the outward normal to $\partial\Omega$. Under suitable conditions on $F$, $g$, $h$ and $\Omega$, it is known that the weak solution $u_\e$ of (\[IDP\]) converges weakly in $L^2(0, T; H^1(\Omega))$ and strongly in $L^2(\Omega_T)$ to $u_0$, where $\Omega_T =\Omega\times (0, T)$. Furthermore, the function $u_0$ is the weak solution of the (homogenized) initial-Dirichlet problem, $$\label{IDP-0}
\left\{\aligned
(\partial_t +\mathcal{L}_0) u_0 &= F &\quad &\text { in } \Omega \times (0, T),\\
u_0 & = g & \quad & \text{ on } \partial\Omega \times (0, T),\\
u_0 &=h &\quad &\text{ on } \Omega \times \{ t=0\}.
\endaligned
\right.$$ Similarly, the weak solution $u_\e$ of (\[INP\]) converges weakly in $L^2(0, T; H^1(\Omega))$ and strongly in $L^2(\Omega_T)$ to the weak solution of the (homogenized) initial-Neumann problem, $$\label{INP-0}
\left\{\aligned
(\partial_t +\mathcal{L}_0) u_0 &= F &\quad &\text { in } \Omega \times (0, T),\\
\frac{\partial u_0}{\partial \nu_0} & = g & \quad & \text{ on } \partial\Omega \times (0, T),\\
u_0 &=h &\quad &\text{ on } \Omega \times \{ t=0\}.
\endaligned
\right.$$ The operator $\mathcal{L}_0$ in (\[IDP-0\]) and (\[INP-0\]), called the homogenized operator, is a second-order elliptic operator with constant coefficients [@bensoussan-1978].
The following are the main results of the paper, which establish the sharp $O(\e)$ convergence rates in $L^2(\Omega_T)$ for both the initial-Dirichlet and the initial-Neumann problems.
\[main-theorem-1\] Suppose that the coefficient matrix $A$ satisfies (\[ellipticity\]) and (\[periodicity\]). Let $\Omega$ be a bounded $C^{1,1}$ domain in $\R^d$. Let $u_\e, u_0\in L^2(0, T; H^1(\Omega))$ be weak solutions of (\[IDP\]) and (\[IDP-0\]), respectively, for some $F\in L^2(\Omega_T)$. Assume that $u_0\in L^2(0, T; H^2(\Omega))$. Then $$\label{main-estimate-1}
\aligned
& \| u_\e -u_0\|_{L^2(\Omega_T)}\\
& \le C \e \left\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\| F\|_{L^2(\Omega_T)}
+\sup_{\e^2 <t<T} \left(\frac{1}{\e} \int_{t-\e^2}^t \int_\Omega |\nabla u_0|^2\right)^{1/2} \right\},
\endaligned$$ where $C$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$.
\[main-theorem-2\] Let $u_\e\in L^2(0, T; H^1(\Omega))$ be a weak solution of (\[INP\]) for some $F\in L^2(\Omega_T)$ and $u_0\in L^2(0, T; H^1(\Omega))$ the weak solution of the homogenized problem (\[INP-0\]). Under the same assumptions as in Theorem \[main-theorem-1\], the estimate (\[main-estimate-1\]) holds.
\[remark-1.1\] [ In Theorems \[main-theorem-1\] and \[main-theorem-2\] we do not specify the conditions directly on $g$ and $h$, but rather require $u_0 \in L^2(0, T; H^2(\Omega))$. In the case that either $u_\e=u_0=0$ or $\displaystyle \frac{\partial u_\e}{\partial \nu_\e}=\frac{\partial u_0}{\partial \nu_0}=0$ on $\partial\Omega \times (0, T)$, i.e. $g=0$, the third term in the r.h.s. of (\[main-estimate-1\]) may be bounded by $$C\big\{ \|\partial_t u_0\|_{L^2(\Omega_T)} + \| F\|_{L^2(\Omega_T)} +\| h\|_{L^2(\Omega)}\big\}.$$ See (\[3.10-2\]). As a result, we obtain $$\label{main-estimate-2}
\| u_\e -u_0\|_{L^2(\Omega_T)}
\le C\, \e \Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\| F\|_{L^2(\Omega_T)} +\| h\|_{L^2(\Omega)} \Big\},$$ where $C$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$. In particular, if $g=0$ and $h=0$, then $$\| u_0\|_{L^2(0, T; H^2(\Omega))} \le C \| F\|_{L^2(\Omega_T)}$$ (see (\[3.10-5\])). It follows that $$\label{main-estimate-3}
\| u_\e -u_0\|_{L^2(\Omega_T)}
\le C\, \e \| F\|_{L^2(\Omega_T)}.$$ Also, in the case that $g=0$ on $\partial\Omega\times (0, T)$ and $h\in H^1(\Omega)$, it is known that if $\mathcal{L}_0^* =\mathcal{L}_0$, then $$\| u_0\|_{L^2(0, T; H^2(\Omega))}
\le C \Big\{ \| F\|_{L^2(\Omega_T)} +\| h\|_{H^1(\Omega)} \Big\}$$ [@Lady]. This gives $$\label{main-estimate-4}
\| u_\e -u_0\|_{L^2(\Omega_T)}
\le C\, \e \Big\{ \| F\|_{L^2(\Omega_T)} +\| h\|_{H^1(\Omega)} \Big\},$$ where $C$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$. ]{}
The sharp convergence rate is one of the central issues in quantitative homogenization and has been studied extensively in the various settings. For elliptic equations and systems in divergence form with periodic coefficients, related results may be found in the recent work [@Suslina-2012; @Suslina-2013; @KLS2; @KLS3; @KLS4; @Shen-Boundary-2015; @Gu-2015; @SZ-2015] (also see [@bensoussan-1978; @Jikov-1994; @Griso-2004; @Griso-2006; @Onofrei-2007] for references on earlier work). In particular, the order sharp estimate $$\label{elliptic-estimate}
\| u_\e -u_0\|_{L^2(\Omega)} \le C\,\e \| F\|_{L^2(\Omega)},$$ holds, if $\mathcal{L}_\e (u_\e)=\mathcal{L}_0 (u_0)=F$ in $\Omega$ and $u_\e =u_0=0$ or $\frac{\partial u_\e}{\partial\nu_\e}=\frac{\partial u_0}{\partial \nu_0}=0$ on $\partial\Omega$ (see [@Suslina-2012; @Suslina-2013; @Gu-2015; @SZ-2015] for $C^{1,1}$ domains and [@KLS2; @KLS4; @Shen-Boundary-2015] for Lipschitz domains). For parabolic equations and systems various results are known in the case where the coefficients are time-independent [@Jikov-1994; @Suslina-2004; @Zhikov-2006; @Suslina-2015]. We note that in this case, using the partial Fourier transform in the $t$ variable, it is possible to represent the solution of the parabolic system as an integral of the resolvent of the elliptic operator $\mathcal{L}_\e$ and apply the elliptic estimates.
Very few results are known if the coefficients are time-dependent. In fact, to the authors’ best knowledge, the only known estimate in this case is $$\label{parabolic-max}
\| u_\e -u_0\|_{L^\infty(\Omega_T)} \le C \e,$$ obtained by the use of the maximum principle, where $C$ depends on $u_0$ and coefficients are assumed to be smooth [@bensoussan-1978]. Our order sharp estimates (\[main-estimate-2\])-(\[main-estimate-4\]), which extend (\[elliptic-estimate\]) to the parabolic setting, seem to be the first work in this area beyond the rough estimate (\[parabolic-max\]).
We now describe some of key ideas in the proof of Theorems \[main-theorem-1\] and \[main-theorem-2\]. Although it is not clear how to reduce parabolic systems with time-dependent coefficients to elliptic systems by some simple transformations, our general approach to the estimate (\[main-estimate-1\]) is inspired by the work on elliptic systems mentioned above. We consider the function $$\label{w-1}
w_\e =u_\e (x, t)-u_0 (x, t)
-\e \chi (x/\e, t/\e^2) K_\e (\nabla u_0)
-\e^2 \phi (x/\e, t/\e^2) \nabla K_\e (\nabla u_0),$$ where $\chi (y, s)$ and $\phi(y,s)$ are correctors and dual correctors for the family of operators $\partial_t +\mathcal{L}_\e$, $\e>0$ (see Section 2 for their definitions). In (\[w-1\]) the operator $K_\e: L^2(\Omega_T) \to C_0^\infty(\Omega_T)$ is a parabolic smoothing operator at scale $\e$. We note that in the elliptic case [@Suslina-2012; @Suslina-2013; @Shen-Boundary-2015; @SZ-2015], only the first three terms in the r.h.s. of (\[w-1\]) are used. By computing $ (\partial_t +\mathcal{L}_\e) w_\e$, we are able to show that $$\label{1.1-dual}
\aligned
&\Big|\int_0^T \langle \partial_t w_\e, \psi\rangle +\iint_{\Omega_T}
A^\e\nabla w_\e \cdot \nabla \psi \Big|\\
& \le C \Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2}\| \nabla u_0\|_{L^2(\Omega_{T, \e})} \Big\}\\
& \qquad \qquad \qquad \qquad
\cdot \Big\{ \e \| \nabla \psi\|_{L^2(\Omega_T)} +\e^{1/2}
\|\nabla \psi\|_{L^2(\Omega_{T, \e})} \Big\}
\endaligned$$ for any $\psi \in L^2(0, T; H^1_0(\Omega))$ in the case of Dirichlet condition (\[IDP\]), and for any $\psi \in L^2(0, T; H^1(\Omega))$ in the case of the Neumann condition (\[INP\]), where $\Omega_{T, \e}$ denotes the set of points in $\Omega_T$ whose (parabolic) distances to the boundary of $\Omega_T$ are less than $\e$ (see Section 3 for details). By taking $\psi =w_\e$ in (\[1.1-dual\]) we obtain an $O(\sqrt{\e})$ error estimate in $L^2(0, T; H^1(\Omega))$, $$\label{H-1}
\| \nabla w_\e\|_{L^2(\Omega_T)}
\le C \sqrt{\e}
\Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2} \| \nabla u_0\|_{L^2(\Omega_{T, \e})} \Big\},$$ which is more or less sharp, for both the initial-Dirichlet and the initial-Neumann problems. Finally, with (\[1.1-dual\]) at our disposal, we give the proof of Theorems \[main-theorem-1\] and \[main-theorem-2\] in Section 4. This is done by a dual argument, inspired by [@Suslina-2012; @Suslina-2013].
We point out that results on convergence rates are useful in the study of regularity estimates that are uniform in $\e>0$ [@Armstrong-Smart-2014; @Armstrong-Shen-2016; @Shen-Boundary-2015]. For solutions of $(\partial_t +\mathcal{L}_\e)u_\e=F$, the uniform boundary Hölder and interior Lipschitz estimates were proved in [@Geng-Shen-2015] by a compactness method, introduced to the study of homogenization problems in [@AL-1987]. The results obtained in this paper should allow us to establish the boundary Lipschitz estimates as well as Rellich estimates at large scale for parabolic systems in a manner similar to that in [@Shen-Boundary-2015] for elliptic systems of linear elasticity. We plan to carry this out in a separate study.
We end this section with some notations that will be used throughout the paper. A function $h=h(y,s)$ in $\R^{d+1}$ is said to be $1$-periodic if $h$ is periodic with respect to $\mathbb{Z}^{d+1}$. We will use the notation $$h^\e (x,t)= h (x/\e, t/\e^2)$$ for $\e>0$, and the summation convention that the repeated indices are summed. Finally, we use $C$ to denote constants that depend at most on $d$, $m$, $\mu$, $T$ and $\Omega$, but never on $\e$.
**Correctors and dual correctors**
==================================
Let $\mathcal{L}_\varepsilon=-\text{div}\left(A^\e(x,t) \nabla\right)$, where $A^\e (x,t)=A(x/\e, t/\e^2)$ and $A(y, s)$ is 1-periodic and satisfies the ellipticity condition (\[ellipticity\]). For $1\leq j\leq d$ and $1\le \beta\le m$, the corrector $\chi_j^\beta=\chi_j^\beta (y,s)=(\chi_{j}^{\alpha\beta} (y, s))$ is defined as the weak solution of the following cell problem: $$\label{corrector}
\begin{cases}
\big(\partial_s +\mathcal{L}_1\big) (\chi_j^\beta)
=-\mathcal{L}_1(P_j^\beta ) ~~~\text{in}~~Y, \\
\chi_j^\beta =\chi^\beta_j(y,s)~~ \text{is } \text{1-periodic in } (y,s),\\
\int_{Y} \chi_j^\beta = 0,
\end{cases}$$ where $Y=[0,1)^{d+1}$, $P_j^\beta (y)=y_j e^\beta$, and $e^\beta=(0, \dots, 1, \dots, 0 )$ with $1$ in the $\beta^{th}$ position. Note that $$(\partial_s+\mathcal{L}_1)(\chi_j^\beta +P_j^\beta )=0~~~\text{in}~~\mathbb{R}^{d+1}.$$ By the rescaling property of $\partial_t +\mathcal{L}_\e$, one obtains that $$(\partial_t+\mathcal{L}_\varepsilon)\left\{\varepsilon\chi_j^\beta(x/\varepsilon,t/\varepsilon^2)+P^\beta_j(x)\right\}
=0~~~\text{in}~~\mathbb{R}^{d+1}.$$
Let $\widehat{A}=(\widehat{a}^{\alpha\beta}_{ij})$, where $1\leq i,j\leq d$, $1\le \alpha, \beta\le m$, and $$\begin{aligned}
\label{A}
\widehat{a}^{\alpha\beta}_{ij}=\dashint_{Y}\left[a^{\alpha\beta}_{ij}+a^{\alpha\gamma}
_{i k}\frac{\partial}{\partial y_k}\chi^{\gamma\beta}_{j}\right];\end{aligned}$$ that is $$\widehat{A}=\dashint_Y \Big\{ A +A\nabla \chi \Big\}.$$ It is known that the constant matrix $\widehat{A}$ satisfies the ellipticity condition, $$\mu |\xi|^2 \le \widehat{a}_{ij}^{\alpha\beta} \xi_i^\alpha \xi_j^\beta
\le \mu_1 |\xi|^2 \qquad \text{ for any } \xi=(\xi_j^\beta) \in \R^{m\times d},$$ where $\mu_1>0$ depends only on $d$, $m$ and $\mu$ [@bensoussan-1978]. Denote $\mathcal{L}_0=-\text{div}(\widehat{A}\nabla)$. Then $\partial_t+\mathcal{L}_0$ is the homogenized operator for the family of parabolic operators $\partial_t+\mathcal{L}_\varepsilon$, $\e>0$.
To introduce the dual correctors, we consider the 1-periodic matrix-valued function $$\label{B}
B= A + A\nabla \chi -\widehat{A}.$$ More precisely, $B=B(y, s)= \big( b_{ij}^{\alpha\beta}\big)$, where $1\le i, j\le d$, $1\le \alpha, \beta\le m$, and $$\label{b}
b_{ij}^{\alpha\beta}=a_{ij}^{\alpha\beta}+a_{ik}^{\alpha\gamma}
\frac{\partial \chi^{\gamma\beta}_j}{\partial y_k}-\widehat{a}^{\alpha\beta}_{ij}.$$
\[1.15\] Let $1\leq j\leq d$ and $1\le \alpha, \beta\le m$. Then there exist 1-periodic functions $\phi_{kij}^{\alpha\beta}(y,s)$ in $\R^{d+1}$ such that $\phi_{kij}^{\alpha\beta}\in H^1(Y)$, $$\label{1.10}
b_{ij}^{\alpha\beta}=\frac{\partial}{\partial y_k}(\phi^{\alpha\beta}_{kij})~~\text{ and }~~\phi^{\alpha\beta}_{kij}
=-\phi_{ikj}^{\alpha\beta},$$ where $1\le k, i\le d+1$, $b_{ij}^{\alpha\beta}$ is defined by (\[b\]) for $1\le i\le d$, $b_{(d+1)j}^{\alpha\beta}=-\chi_j^{\alpha\beta}$, and we have used the notation $y_{d+1}=s$.
Observe that by (\[corrector\]) and (\[A\]), $b_{ij}^{\alpha\beta} \in L^2 (Y)$ and $$\label{1.8-1}
\int_{Y} b_{ij}^{\alpha \beta}=0$$ for $1\le i\le d+1$. It follows that there exist $f_{ij}^{\alpha\beta} \in H^2 (Y)$ such that $$\label{1.8-2}
\begin{cases}
\Delta_{d+1} f_{ij}^{\alpha\beta}=b^{\alpha\beta}_{ij}~~~~\text{ in } \R^{d+1}, \\
f_{ij}^{\alpha\beta} ~~\text{is 1-periodic}~~~\text{ in } \R^{d+1},
\end{cases}$$ where $\Delta_{d+1}$ denotes the Laplacian in $\R^{d+1}$. Write $$\begin{aligned}
\label{1.11}
b_{ij}^{\alpha\beta}=
\frac{\partial}{\partial y_k}\left\{\frac{\partial}{\partial y_k}f_{ij}^{\alpha\beta}
-\frac{\partial}{\partial y_i}f_{kj}^{\alpha\beta}\right\}+\frac{\partial}{\partial y_i}\left\{\frac{\partial}{\partial y_k}f_{kj}^{\alpha\beta}\right\},\end{aligned}$$ where the index $k$ is summed from $1$ to $d+1$. Note that by (\[corrector\]), $$\label{div}
\sum_{i=1}^{d+1} \frac{\partial b^{\alpha\beta}_{ij}}{\partial y_i}
=\sum_{i=1}^d \frac{\partial}{\partial y_i}b^{\alpha\beta}_{ij}-\frac{\partial}{\partial s}\chi_j^{\alpha\beta}=0.$$ In view of (\[1.8-2\]) this implies that $$\sum_{i=1}^{d+1}\frac{\partial }{\partial y_i} f_{ij}^{\alpha\beta}$$ is harmonic in $\R^{d+1}$. Since it is 1-periodic, it must be constant. Consequently, by (\[1.11\]), we obtain $$\begin{aligned}
\label{1.13}
b_{ij}^{\alpha\beta}=\frac{\partial}{\partial y_k}(\phi_{kij}^{\alpha\beta}),\end{aligned}$$ where $$\begin{aligned}
\label{1.14}
\phi_{kij}^{\alpha\beta}
=\frac{\partial}{\partial y_k}f_{ij}^{\alpha\beta}
-\frac{\partial}{\partial y_i}f_{kj}^{\alpha\beta}\end{aligned}$$ is 1-periodic and belongs to $H^1(Y)$. It is easy to see that $\phi_{kij}^{\alpha\beta}=-\phi_{ikj}^{\alpha\beta}$. This completes the proof.
The 1-periodic functions $(\phi_{kij}^{\alpha\beta})$ given by Lemma \[1.15\] are called dual correctors for the family of parabolic operators $\partial_t +\mathcal{L}_\e$, $\e>0$. As in the elliptic case [@Jikov-1994; @KLS2], they play an important role in the study of the problem of convergence rates. Indeed, to establish the main results of this paper, we shall consider the function $w_\e = ( w_\e^\alpha)$, where $$\label{w}
\aligned
w_\e ^\alpha (x, t) = u_\e^\alpha (x, t) -u_0^\alpha (x, t) & -\e \chi_j^{\alpha\beta} (x/\e, t/\e^2)
K_\e \left(\frac{\partial u_0^\beta}{\partial x_j}\right)\\
&-\e^2 \phi_{(d+1) ij}^{\alpha\beta} (x/\e, t/\e^2)
\frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0^\beta}{\partial x_j}\right),
\endaligned$$ and $K_\e : L^2(\Omega_T) \to C_0^\infty(\Omega_T)$ is a linear operator to be chosen later. The repeated indices $i, j$ in (\[w\]) are summed from $1$ to $d$.
\[Theorem-2.1\] Let $\Omega$ be a bounded Lipschitz domain in $\R^d$ and $0<T<\infty$. Let $u_\e \in L^2(0, T; H^1(\Omega))$ and $u_0\in L^2(0, T; H^2(\Omega))$ be solutions of the initial-Dirichlet problems (\[IDP\]) and (\[IDP-0\]), respectively. Let $w_\e$ be defined by (\[w\]). Then for any $\psi \in L^2(0, T; H^1_0(\Omega))$, $$\label{L-w}
\aligned
\int_0^T &\big\langle \partial_t w_\e, \psi\big\rangle_{H^{-1}(\Omega)\times H^1_0(\Omega)}
+\iint_{\Omega_T} A^\e \nabla w_\e \cdot \nabla \psi\\
=& \iint_{\Omega_T}
(\widehat{a}_{ij} -a_{ij}^\e )
\left( \frac{\partial u_0}{\partial x_j} - K_\e \left( \frac{\partial u_0}{\partial x_j}\right) \right) \frac{\partial \psi}{\partial x_i}\\
& - \e \iint_{\Omega_T}
a_{ij}^\e \cdot \chi_k^\e \cdot \frac{\partial }{\partial x_j}
K_\e \left(\frac{\partial u_0}{\partial x_k} \right) \cdot \frac{\partial \psi}{\partial x_i} \\
& -\e \iint_{\Omega_T}
\phi_{kij}^\e \cdot \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j}\right)
\cdot \frac{\partial \psi}{\partial x_k}\\
&-\e^2 \iint_{\Omega_T}
\phi_{k(d+1)j}^\e \cdot \partial_t K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\cdot\frac{\partial \psi}{\partial x_k} \\
&+ \e \iint_{\Omega_T}
a_{ij}^\e \cdot \left(\frac{\partial}{\partial x_j} ( \phi_{(d+1) \ell k} ) \right)^\e \cdot \frac{\partial}{\partial x_\ell}
K_\e \left(\frac{\partial u_0}{\partial x_k} \right) \cdot \frac{\partial \psi}{\partial x_i} \\
&+\e^2 \iint_{\Omega_T}
a_{ij}^\e \cdot \phi_{(d+1) \ell k } ^\e \cdot \frac{\partial^2 }{\partial x_j \partial x_\ell}
K_\e \left(\frac{\partial u_0}{\partial x_k}\right) \cdot \frac{\partial \psi}{\partial x_i},
\endaligned$$ where we have suppressed superscripts $\alpha, \beta$ for the simplicity of presentation. The repeated indices $i, j, k, \ell$ are summed from $1$ to $d$.
Using (\[IDP\]) and (\[IDP-0\]), we see that $$\aligned
\big(\partial_t +\mathcal{L}_\e) w_\e
&=(\mathcal{L}_0 -\mathcal{L}_\e) u_0
-(\partial_t +\mathcal{L}_\e) \left\{ \e \chi_j^\e K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\right\}\\
&\qquad \qquad \qquad -(\partial_t +\mathcal{L}_\e )
\left\{ \e^2 \phi_{(d+1) ij}^\e \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\}\\
&=-\frac{\partial}{\partial x_i}
\left\{ (\widehat{a}_{ij} -a_{ij}^\e )
\left( \frac{\partial u_0}{\partial x_j} - K_\e \left( \frac{\partial u_0}{\partial x_j}\right) \right) \right\}\\
&\qquad -\frac{\partial}{\partial x_i}
\left\{ (\widehat{a}_{ij} -a_{ij}^\e) K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\}
-(\partial_t +\mathcal{L}_\e) \left\{ \e \chi_j^\e K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\right\}\\
&\qquad \qquad \qquad -(\partial_t +\mathcal{L}_\e )
\left\{ \e^2 \phi_{(d+1) ij}^\e \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\}.
\endaligned$$ By computing the third term in the r.h.s. of the equalities above and using (\[b\]), we obtain $$\aligned
\big(\partial_t +\mathcal{L}_\e) w_\e
&=-\frac{\partial}{\partial x_i}
\left\{ (\widehat{a}_{ij} -a_{ij}^\e )
\left( \frac{\partial u_0}{\partial x_j} - K_\e \left( \frac{\partial u_0}{\partial x_j}\right) \right) \right\}\\
&\qquad +\frac{\partial}{\partial x_i}
\left\{ b_{ij}^\e K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\}
+\e \frac{\partial}{\partial x_i}
\left\{ a_{ij}^\e \cdot \chi_k^\e \cdot \frac{\partial}{\partial x_j} K_\e \left(\frac{\partial u_0}{\partial x_k}\right) \right\}\\
&\qquad - \e \partial_t \left\{ \chi_j^\e K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\right\}
-(\partial_t +\mathcal{L}_\e )
\left\{ \e^2 \phi_{(d+1) ij}^\e \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\}.
\endaligned$$ In view of (\[div\]) this gives $$\label{2.2-10}
\aligned
\big(\partial_t +\mathcal{L}_\e) w_\e
&=-\frac{\partial}{\partial x_i}
\left\{ (\widehat{a}_{ij} -a_{ij}^\e )
\left( \frac{\partial u_0}{\partial x_j} - K_\e \left( \frac{\partial u_0}{\partial x_j}\right) \right) \right\}\\
& \qquad +\e \frac{\partial}{\partial x_i}
\left\{ a_{ij}^\e \cdot \chi_k^\e \cdot \frac{\partial}{\partial x_j} K_\e \left(\frac{\partial u_0}{\partial x_k}\right) \right\}
+b_{ij}^\e\cdot \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\\
&\qquad -\e \chi_j^\e \partial_t K_\e \left(\frac{\partial u_0}{\partial x_j}\right)
-(\partial_t +\mathcal{L}_\e )
\left\{ \e^2 \phi_{(d+1) ij}^\e \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\}.
\endaligned$$
Next, by Lemma \[1.15\], we may write $$\aligned
& b_{ij}^\e\cdot \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)
-\e \chi_j^\e \partial_t K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\\
& \qquad =\e \frac{\partial}{\partial x_k} \Big(\phi_{kij}^\e\Big)\cdot \frac{\partial}{\partial x_i}
K_\e \left(\frac{\partial u_0}{\partial x_j}\right)
+\e^2 \partial_t \Big(\phi_{(d+1) ij}^\e \Big) \cdot \frac{\partial}{\partial x_i} K_\e
\left(\frac{\partial u_0}{\partial x_j} \right)\\
&\qquad \qquad +\e^2 \frac{\partial}{\partial x_k} \Big( \phi_{k (d+1) j }^\e\Big )\cdot
\partial_t K_\e \left(\frac{\partial u_0}{\partial x_j} \right),
\endaligned$$ where we have also used the fact $\phi_{(d+1)(d+1) j}=0$. Furthermore, by the skew-symmetry in (\[1.10\]), we see that $$\aligned
& b_{ij}^\e\cdot \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)
-\e \chi_j^\e \partial_t K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\\
& \qquad =\e \frac{\partial}{\partial x_k} \left\{ \phi_{kij}^\e \cdot \frac{\partial}{\partial x_i}
K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\right\}
+\e^2 \partial_t \left\{ \phi_{(d+1) ij}^\e \cdot \frac{\partial}{\partial x_i} K_\e
\left(\frac{\partial u_0}{\partial x_j} \right)\right\} \\
&\qquad \qquad +\e^2 \frac{\partial}{\partial x_k} \left\{ \phi_{k (d+1) j }^\e\cdot
\partial_t K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\}.
\endaligned$$ This, combined with (\[2.2-10\]), gives the desired equation (\[L-w\]).
The next theorem is concerned with the initial-Neumann problem.
\[Theorem-2.2\] Let $\Omega$ be a bounded Lipschitz domain in $\R^d$ and $0<T<\infty$. Let $u_\e \in L^2(0, T; H^1(\Omega))$ and $u_0\in L^2(0, T; H^2(\Omega))$ be solutions of the initial-Neumann problems (\[INP\]) and (\[INP-0\]), respectively. Let $w_\e$ be defined by (\[w\]). Then the equation (\[L-w\]) holds for any $\psi \in L^2(0, T; H^1(\Omega))$, if $\langle, \rangle$ in its l.h.s. denotes the pairing between $H^1(\Omega)$ and its dual.
It follows from (\[INP\]) and (\[INP-0\]) that $$\int_0^T \big\langle \partial_t u_\e, \psi \big\rangle +\iint_{\Omega_T} A^\e \nabla u_\e \cdot \nabla \psi
=\int_0^T \big\langle \partial_t u_0, \psi\big \rangle +\iint_{\Omega_T} \widehat{A} \nabla u_0 \cdot \nabla \psi$$ for any $\psi \in L^2(0, T; H^1(\Omega))$. This gives $$\aligned
\int_0^T &\big \langle \partial_t w_\e, \psi\big\rangle +\iint_{\Omega_T} A^\e \nabla w_\e \cdot \nabla \psi\\
&=\iint_{\Omega_T} (\widehat{A}-A^\e )\nabla u_0 \cdot \nabla \psi
-\int_0^T \Big\langle (\partial_t +\mathcal{L}_\e) \left\{ \e \chi_j^\e K_\e \left(\frac{\partial u_0}{\partial x_j}\right)\right\},
\psi \Big\rangle \\
& \qquad -\int_0^T \Big\langle
(\partial_t +\mathcal{L}_\e )
\left\{ \e^2 \phi_{(d+1) ij}^\e \frac{\partial}{\partial x_i} K_\e \left(\frac{\partial u_0}{\partial x_j} \right)\right\},
\psi \Big\rangle,
\endaligned$$ where we have used the fact $K_\e (\nabla u_0)\in C_0^\infty(\Omega_T)$. The rest of the proof is similar to that of Theorem \[Theorem-2.1\]. We omit the details.
Error estimates in $L^2(0, T; H^1(\Omega))$
===========================================
We begin by introducing a parabolic smoothing operator. Fix a nonnegative function $\theta=\theta (y,s)
\in C_0^\infty(B(0,1))$ such that $\int_{\mathbb{R}^{d+1}}\theta=1$. Define $$\label{1.18}
\aligned
S_\varepsilon(f)(x,t)&=\frac{1}{\varepsilon^{d+2}}\int_{\mathbb{R}^{d+1}}f(x-y,t-s)
\theta(y/\varepsilon,s/\varepsilon^2)\, dyds
\\&=\int_{\mathbb{R}^{d+1}}f(x-\varepsilon y,t-\varepsilon^2 s)\theta(y,s)\, dyds.
\endaligned$$
\[lemma-S-1\] Let $S_\varepsilon$ be defined as in (\[1.18\]). Then $$\begin{aligned}
\label{1.19-2}
\| S_\varepsilon (f)\|_{L^2({\mathbb{R}^{d+1}})}\leq \| f \|_{L^2({\mathbb{R}^{d+1}})},\end{aligned}$$ $$\begin{aligned}
\label{1.19-3}
\e\, \| \nabla S_\varepsilon (f)\|_{L^2({\mathbb{R}^{d+1}})}
+\e^2 \|\nabla^2 S_\e (f)\|_{L^2(\R^{d+1})} \leq C\, \| f \|_{L^2({\mathbb{R}^{d+1}})},\end{aligned}$$ $$\begin{aligned}
\label{1.19-4}
\e^2 \| \partial_t S_\varepsilon (f)\|_{L^2({\mathbb{R}^{d+1}})}\leq C\, \| f \|_{L^2({\mathbb{R}^{d+1}})},\end{aligned}$$ where $C$ depends only on $d$.
This follows easily from the Plancherel Theorem.
\[lemma-S-2\] Let $S_\e$ be defined as in (\[1.18\]). Then $$\label{S-approx}
\| \nabla S_\e (f) -\nabla f \|_{L^2(\R^{d+1})}
\le C \e \Big\{ \| \nabla^2 f \|_{L^2(\R^{d+1})}
+\| \partial_t f \|_{L^2(\R^{d+1})} \Big\},$$ where $C$ depends only on $d$.
By the Plancherel Theorem it suffices to show that $$|\xi_i \widehat{\theta} (\e \xi^\prime, \e^2 \xi_{d+1}) -\xi_i \widehat{\theta}(0, 0)|
\le C\e \big\{ |\xi^\prime|^2 +|\xi_{d+1}| \big\},$$ where $1\le i\le d$ and $\xi^\prime=(\xi_1, \dots, \xi_d)\in \R^d$. Furthermore, by a change of variables, one may assume that $\e=1$. In this case, if $|\xi^\prime|\ge 1$, then $$|\xi_i \widehat{\theta} ( \xi^\prime, \xi_{d+1}) -\xi_i \widehat{\theta}(0, 0)|
\le C |\xi^\prime|\le C (|\xi^\prime|^2 +|\xi_{d+1}|).$$ If $|\xi^\prime|\le 1$, we have $$|\xi_i \widehat{\theta} ( \xi^\prime, \xi_{d+1}) -\xi_i \widehat{\theta}(0, 0)|
\le C |\xi^\prime| ( |\xi^\prime| +|\xi_{d+1}|)
\le C ( |\xi^\prime|^2 +|\xi_{d+1}|).$$ This completes the proof.
\[lemma-S-3\] Let $g=g(y,s)$ be a 1-periodic function in $(y,s)$. Then $$\begin{aligned}
\label{1.22}
\| g^\e S_\varepsilon (f)\|_{L^p({\mathbb{R}^{d+1}})}\leq C\,
\| g \|_{L^p(Y)} \| f \|_{L^p({\mathbb{R}^{d+1}})}\end{aligned}$$ for any $1\le p<\infty$, where $g^\e (x, t)= g(x/\e, t/\e^2)$ and $C$ depends only on $d$ and $p$.
Note that $S_\e (f) (x, t)= S_1 (f_\e ) (\e^{-1} x, \e^{-2} t)$, where $f_\e (x, t)= f(\e x , \e^2 t)$. As a result, by a change of variables, it suffices to consider the case $\e=1$. In this case we first use $\int_{\R^{d+1}} \theta =1$ and Hölder’s inequality to obtain $$|S_1 (f) (x, t)|^p
\le \int_{\R^{d+1}} |f(y, s)|^p \, \theta (x-y, t-s)\, dyds.$$ It follows by Fubini’s Theorem that $$\aligned
\int_{\R^{d+1}} |g(x,t)|^p | S_1 (f)(x, t)|^p\, dx dt
&\le \sup_{(y, s)\in \R^{d+1}}
\int_{B((y,s), 1)} | g(x, t)|^p \, dx dt \int_{\R^{d+1}} | f(y, s)|^p\, dyds\\
&\le C\, \| g \|^p_{L^p(Y)}
\| f \|_{L^p(\R^{d+1})}^p,
\endaligned$$ where $C$ depends only on $d$. This gives (\[1.22\]) for the case $\e=1$.
\[remark-S\] [The same argument as in the proof of Lemma \[lemma-S-3\] also shows that $$\label{S-remark-1}
\aligned
\| g^\e \nabla S_\e (f)\|_{L^p(\R^{d+1})}
& \le C \e^{-1} \| g\|_{L^p(Y)} \| f\|_{L^p(\R^{d+1})},\\
\| g^\e \partial_t S_\e (f)\|_{L^p(\R^{d+1})}
& \le C \e^{-2} \| g\|_{L^p(Y)} \| f\|_{L^p(\R^{d+1})}
\endaligned$$ ]{} for $1\le p<\infty$, where $C$ depends only on $d$ and $p$.
Let $\delta \in (2\e, 20\e)$. Choose $\eta_1 \in C_0^\infty(\Omega)$ such that $0\le \eta_1\le 1$, $\eta_1 (x)=1$ if dist$(x, \partial\Omega)\ge 2\delta$, $\eta_1 (x)=0$ if dist$(x, \partial\Omega)\le \delta$, and $|\nabla_x \eta_1|\le C \delta^{-1}$. Similarly, we choose $\eta_2\in C_0^\infty(0, T)$ such that $0\le \eta_2\le 1$, $\eta_2 (t) =1$ if $2\delta^2 \le t\le T-2\delta^2 $, $\eta_2 (t)=0$ if $t\le \delta^2 $ or $t>T-\delta^2 $, and $| \eta_2^\prime (t)|\le C \delta^{-2}$. We define the operator $K_\e=K_{\e, \delta}: L^2(\Omega_T) \to C_0^\infty (\Omega_T)$ by $$\label{K}
K_\e (f) (x, t) = S_\e ( \eta_1 \eta_2 f ) (x, t).$$
\[main-lemma-3.1\] Let $\Omega$ be a bounded Lipschitz domain in $\R^d$ and $0<T<\infty$. Let $u_\varepsilon, u_0\in L^2(0,T;H^1(\Omega))$ be weak solutions of (\[IDP\]) and (\[IDP-0\]), respectively, for some $F\in L^2(\Omega_T)$. We further assume that $u_0\in L^2(0, T; H^2(\Omega))$ and $\partial_t u_0 \in L^2(\Omega_T)$. Let $w_\e$ be defined by (\[w\]), where the operator $K_\e$ is given by (\[K\]). Then for any $\psi \in L^2(0, T; H_0^1(\Omega))$, $$\label{main-estimate-3.1}
\aligned
& \Big| \int_0^T \big\langle (\partial_t +\mathcal{L}_\e) w_\e,
\psi \big\rangle_{H^{-1}(\Omega) \times H^1_0(\Omega)} \, dt \Big| \\
&\le
C
\Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0 \|_{L^2(\Omega_T)}
+\e^{-1/2} \| \nabla u_0\|_{L^2(\Omega_{T, 3\delta})}
\Big\}\\
&\qquad \qquad
\cdot \Big\{ \e \|\nabla \psi\|_{L^2(\Omega_T)}
+\e^{1/2} \|\nabla \psi\|_{L^2(\Omega_{T, 3\delta})} \Big\},
\endaligned$$ where $$\label{o-e}
\Omega_{T, \delta} =\left( \big\{ x\in \Omega: \, \text{\rm dist}(x, \partial\Omega)\le \delta \big\}
\times (0, T) \right)\cup \left( \Omega \times (0, \delta^2)\right) \cup \left(\Omega\times (T-\delta^2, T)\right),$$ and $C>0$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$.
Using Theorem \[Theorem-2.1\], it is not hard to see that the l.h.s. of (\[main-estimate-3.1\]) is bounded by $$\label{3.4-1}
\aligned
&C \iint_{\Omega_T} |\nabla u_0 -K_\e (\nabla u_0)| |\nabla \psi|\\
&\qquad
+C \e \iint_{\Omega_T} \Big\{ |\chi^\e| +|\phi^\e| +|(\nabla \phi)^\e|\Big\} |\nabla K_\e (\nabla u_0)||\nabla \psi|\\
&\qquad
+ C \e^2 \iint_{\Omega_T}
|\phi^\e| \Big\{ |\partial_t K_\e (\nabla u_0)| +|\nabla^2 K_\e (\nabla u_0)|\Big\} |\nabla \psi|\\
&=I_1 +I_2 +I_3,
\endaligned$$ where $C$ depends only on $d$, $m$ and $\mu$. To estimate $I_2$, we note that $$\label{3.4-2}
\nabla K_\e (\nabla u_0)
=\nabla S_\e (\eta_1\eta_2(\nabla u_0))
=S_\e (\nabla (\eta_1 \eta_2) (\nabla u_0))
+S_\e (\eta_1\eta_2(\nabla^2 u_0)).$$ It follows by the Cauchy inequality and Lemma \[lemma-S-3\] that $$\aligned
I_2
\le &C \e \left(\iint_{\Omega_T}
| \Big\{ |\chi^\e| +|\phi^\e| +|(\nabla \phi)^\e|\Big\} S_\e (\nabla (\eta_1\eta_2)(\nabla u_0))|^2 \right)^{1/2}
\left(\iint_{\Omega_{T, 3\delta}} |\nabla \psi|^2 \right)^{1/2}\\
&\quad +C
\e \left(\iint_{\Omega_T}
| \Big\{ |\chi^\e| +|\phi^\e| +|(\nabla \phi)^\e|\Big\} S_\e ( \eta_1\eta_2(\nabla^2 u_0))|^2 \right)^{1/2}
\left(\iint_{\Omega_T} |\nabla \psi|^2 \right)^{1/2}\\
&\le C \left(\iint_{\Omega_{T, 3\delta}} |\nabla u_0|^2\right)^{1/2}
\left(\iint_{\Omega_{T, 3\delta}} |\nabla \psi|^2\right)^{1/2}\\
&\qquad+C \e \left(\iint_{\Omega_{T}} |\nabla^2 u_0|^2\right)^{1/2}
\left(\iint_{\Omega_T} |\nabla \psi|^2\right)^{1/2},
\endaligned$$ where we also have used the observation that $S_\e (\nabla (\eta_1\eta_2)(\nabla u_0))$ is supported in $\Omega_{T, 3\delta}$. This shows that $I_2$ is bounded by the r.h.s. of (\[main-estimate-3.1\]).
Next, to handle the term $I_3$, we note that $$\aligned
\partial_t K_\e (\nabla u_0)
&=\partial_t S_\e (\eta_1\eta_2(\nabla u_0))
=S_\e (\partial_t (\eta_1\eta_2) \nabla u_0)
+S_\e (\eta_1\eta_2(\nabla \partial_t u_0))\\
& =S_\e (\partial_t (\eta_1\eta_2) \nabla u_0)
+\nabla S_\e (\eta_1\eta_2 (\partial_t u_0))
-S_\e (\nabla (\eta_1\eta_2) (\partial_t u_0)),
\endaligned$$ and $$\nabla^2 K_\e (\nabla u_0)
=\nabla S_\e (\nabla (\eta_1\eta_2) (\nabla u_0))
+\nabla S_\e (\eta_1 \eta_2( \nabla^2 u_0)).$$ As in the case of $I_2$, by the Cauchy inequality and Remark \[remark-S\] , this gives $$\aligned
I_3 \le &C \left(\iint_{\Omega_{T, 3\delta}} |\nabla u_0|^2\right)^{1/2}
\left(\iint_{\Omega_{T, 3\delta}} |\nabla \psi|^2 \right)^{1/2}\\
&+ C\e \left(\iint_{\Omega_{T}} |\partial_t u_0|^2\right)^{1/2}
\left(\iint_{\Omega_{T}} |\nabla \psi|^2 \right)^{1/2}\\
&+C\e \left(\iint_{\Omega_{T}} |\nabla^2 u_0|^2\right)^{1/2}
\left(\iint_{\Omega_{T}} |\nabla \psi|^2 \right)^{1/2},
\endaligned$$ which is bounded by the r.h.s. of (\[main-estimate-3.1\]).
Finally, to estimate $I_1$, we observe that $$\label{3.4-3}
\aligned
I_1
\le & C\iint_{\Omega_{T, 2\delta}}\Big\{ |\nabla u_0| +S_\e (\eta_1\eta_2 |\nabla u_0|)\Big\} |\nabla \psi|
+C\iint_{\Omega_T\setminus \Omega_{T, 2\delta}}
| (\nabla u_0 -S_\e (\nabla u_0)) | |\nabla \psi|\\
&\le C \left(\iint_{\Omega_{T, 3\delta}} |\nabla u_0|^2\right)^{1/2}
\left(\iint_{\Omega_{T, 3\delta}} |\nabla \psi|^2 \right)^{1/2}\\
&\qquad
+ C \left(\iint_{\Omega\setminus \Omega_{T, 2\delta}} |\nabla u_0 -S_\e (\nabla u_0)|^2\right)^{1/2}
\left(\iint_{\Omega_T} |\nabla \psi|^2 \right)^{1/2}.
\endaligned$$ To treat the second term in the r.h.s. of (\[3.4-3\]), we extend $u_0$ to a function $\widetilde{u}_0$ in $\R^{d+1}$ such that $$\left(\iint_{\R^{d+1}} |\nabla^2 \widetilde{u}_0|^2\right)^{1/2}
+\left(\iint_{\R^{d+1}} |\partial_t \widetilde{u}_0|^2\right)^{1/2}
\le C \Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0\|_{L^2(\Omega_T)} \Big\},$$ using the Calderón’s extension theorem. It follows that $$\aligned
\left(\iint_{\Omega\setminus \Omega_{T, 2\delta}} |\nabla u_0 -S_\e (\nabla u_0)|^2\right)^{1/2}
& \le \left(\iint_{\R^{d+1}}
|\nabla \widetilde{u}_0 -S_\e (\nabla \widetilde{u}_0)|^2\right)^{1/2}\\
&\le C\e \Big\{ \|\nabla^2 \widetilde{u}_0\|_{L^2(\R^{d+1})}
+\|\partial_t \widetilde{u}_0\|_{L^2(\R^{d+1})}\Big\} \\
&\le C\e \Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0\|_{L^2(\Omega_T)} \Big\},
\endaligned$$ where we have used Lemma \[lemma-S-2\] for the second inequality. As a result, we see that $I_1$ is also bounded by the r.h.s. of (\[main-estimate-3.1\]). This completes the proof.
\[remark-u-0\] [ Let $\Omega^\delta=\big\{ x\in \Omega: \text{dist}(x, \partial\Omega)< \delta \big\}$. Then $$\label{boundary-estimate}
\int_{\Omega^{\delta}} |\nabla u_0|^2
\le C \delta \| \nabla u_0\|^2_{H^1(\Omega)}$$ (see e.g. [@SZ-2015] for a proof). It follows that $$\aligned
\|\nabla u_0\|_{L^2(\Omega_{T, 3\delta})}
&\le \left(\int_0^T \int_{\Omega^{3\delta}} |\nabla u_0|^2 \right)^{1/2}
+\left(\int_0^{c\e^2} \int_\Omega |\nabla u_0|^2 \right)^{1/2}
+\left(\int_{T-c\e^2}^T \int_\Omega |\nabla u_0|^2\right)^{1/2}\\
&\le C \e^{1/2}
\left\{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +
\sup_{\e^2<t<T}
\left(\frac{1}{\e} \int_{t-\e^2}^t \int_\Omega |\nabla u_0|^2 \right)^{1/2} \right\}.
\endaligned$$ ]{}
The next theorem provides an $O(\sqrt{\e})$ error estimate in $L^2(0, T; H^1_0(\Omega))$ for the initial-Dirichlet problem (\[IDP\]).
\[IDP-H-1\] Let $w_\e$ be defined by (\[w\]). Under the same assumptions as in Lemma \[main-lemma-3.1\], we have $$\label{estimate-IDP-H-1}
\aligned
&\|\nabla w_\e\|_{L^2(\Omega_T)}\\
&\le C \sqrt{\e}
\left\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0\|_{L^2(\Omega_T)}
+\sup_{\e^2<t<T}
\left(\frac{1}{\e} \int_{t-\e^2}^t \int_\Omega |\nabla u_0|^2 \right)^{1/2} \right\},
\endaligned$$ where $C$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$.
Note that $w_\e \in L^2(0, T; H^1_0(\Omega))$ and $w_\e =0$ on $\Omega \times \{ t=0\}$. It follows that $$\aligned
&\mu \iint_{\Omega_T} |\nabla w_\e|^2
\le \int_0^T \big \langle (\partial_t +\mathcal{L}_\e) w_\e, w_\e\big\rangle_{H^{-1}(\Omega) \times H^1_0(\Omega) }\\
&\le C \sqrt{\e} \|\nabla w_\e\|_{L^2(\Omega_T)}
\Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0 \|_{L^2(\Omega_T)}
+\e^{-1/2} \| \nabla u_0\|_{L^2(\Omega_{T, 60\e})}
\Big\},
\endaligned$$ where we have used Lemma \[main-lemma-3.1\] for the last step. This, together with Remark \[remark-u-0\], gives (\[estimate-IDP-H-1\]).
Next we consider the initial-Neumann problem (\[INP\]).
\[main-lemma-3.2\] Let $\Omega$ be a bounded Lipschitz domain in $\R^d$ and $0<T<\infty$. Let $u_\varepsilon, u_0\in L^2(0,T;H^1(\Omega))$ be weak solutions of the initial-Neumann problems (\[INP\]) and (\[INP-0\]), respectively, for some $F\in L^2(\Omega_T)$. We further assume that $u_0\in L^2(0, T; H^2(\Omega))$ and that $\partial_t u_0 \in L^2(\Omega_T)$. Let $w_\e$ be defined by (\[w\]), where the operator $K_\e$ is given by (\[K\]). Then for any $\psi \in L^2(0, T; H^1(\Omega))$, $$\label{main-estimate-3.2}
\aligned
& \Big| \int_0^T \big\langle \partial_t w_\e, \psi\big \rangle
+\iint_{\Omega_T} A^\e \nabla w_\e \cdot \nabla \psi \Big| \\
&\le
C
\Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0 \|_{L^2(\Omega_T)}
+\e^{-1/2} \| \nabla u_0\|_{L^2(\Omega_{T, 3\delta})}
\Big\}\\
&\qquad \qquad
\cdot \Big\{ \e \|\nabla \psi\|_{L^2(\Omega_T)}
+\e^{1/2} \|\nabla \psi\|_{L^2(\Omega_{T, 3\delta})} \Big\},
\endaligned$$ where $\langle, \rangle$ denotes the pairing between $H^1(\Omega)$ and its dual. The constant $C>0$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$.
This follows from Theorem \[Theorem-2.2\] by the same argument as in the proof of Lemma \[main-lemma-3.1\].
\[INP-H-1\] Let $w_\e$ be defined by (\[w\]). Under the same assumptions as in Lemma \[main-lemma-3.2\], we have $$\label{estimate-INP-H-1}
\aligned
&\|\nabla w_\e\|_{L^2(\Omega_T)}\\
&\le C \sqrt{\e}
\left\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0\|_{L^2(\Omega_T)}
+\sup_{\e^2<t<T}
\left(\frac{1}{\e} \int_{t-\e^2}^t \int_\Omega |\nabla u_0|^2 \right)^{1/2} \right\},
\endaligned$$ where $C$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$.
As in the proof of Theorem \[IDP-H-1\], this follows from Lemma \[main-lemma-3.2\] by letting $\psi =w_\e$.
\[remark-3.10\] [In the case of $u_\e =u_0=0$ or $\frac{\partial u_\e}{\partial \nu_\e}=\frac{\partial u_0}{\partial \nu_0} =0$ on $\partial\Omega\times (0,T)$, we may bound the third term in the r.h.s. of (\[estimate-INP-H-1\]) as follows. Note that $$\label{3.10-1}
\int_\Omega \widehat{A}\nabla u_0 \cdot \nabla u_0
=-\int_\Omega \partial_t u_0 \cdot u_0 +\int_\Omega F\cdot u_0.$$ It follows that $$\aligned
\mu \int_{t-\e^2}^t \int_\Omega
|\nabla u_0|^2
&\le \int_{t-\e^2}^t \int_\Omega |\partial_t u_0||u_0| +
\int_{t-\e^2}^t \int_\Omega |F||u_0| \\
& \le \Big\{ \| \partial_t u_0\|_{L^2(\Omega_T)} +\| F\|_{L^2(\Omega_T)} \Big\}
\left(\int_{t-\e^2}^t \int_\Omega |u_0|^2\right)^{1/2}\\
&\le \e \Big\{ \| \partial_t u_0\|_{L^2(\Omega_T)} +\| F\|_{L^2(\Omega_T)} \Big\}
\sup_{0<t<T} \| u_0(\cdot, t)\|_{L^2(\Omega)}.
\endaligned$$ This, together with the standard energy estimates, gives $$\label{3.10-2}
\sup_{\e^2<t<T}
\left(\frac{1}{\e} \int_{t-\e^2}^t \int_\Omega |\nabla u_0|^2 \right)^{1/2}
\le C \Big\{
\|\partial_t u_0\|_{L^2(\Omega_T)}
+\| F\|_{L^2(\Omega_T)} + \| h \|_{L^2(\Omega)} \Big\},$$ where $C$ depends only on $d$, $m$, $\mu$ and $\Omega$. As a result, for both the initial-Dirichlet problem (\[IDP\]) and the initial-Neumann problem (\[INP\]), if $g=0$ on $\partial\Omega \times (0, T)$, then $$\label{3.10-3}
\| \nabla w_\e\|_{L^2(\Omega_T)}
\le C \sqrt{\e}
\Big\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\| F\|_{L^2(\Omega_T)}
+\| h\|_{L^2(\Omega)} \Big\},$$ where we have used the fact $$\|\partial _t u_0\|_{L^2(\Omega_T)}
\le C \left\{ \|\nabla^2 u_0\|_{L^2(\Omega_T)} +\| F\|_{L^2(\Omega_T)} \right\}.$$ In particular, if $\Omega$ is $C^{1,1}$, $g=0$ on $\partial\Omega\times (0, T)$ and $h=0$ on $\Omega$, then $$\label{3.10-4}
\| \nabla w_\e\|_{L^2(\Omega_T)}
\le C \sqrt{\e} \| F \|_{L^2(\Omega_T)}.$$ To see this, we use the well-known estimate $$\label{3.10-5}
\| u_0\|_{L^2(0, T; H^2(\Omega))} \le C\, \| F\|_{L^2(\Omega_T)},$$ which may be proved by using the partial Fourier transform in the $t$ variable and reducing the problem to the $H^2$ estimate for the elliptic operator $\mathcal{L}_0$ in $C^{1,1}$ domains. We also note that in the case that $g=0$ on $\partial\Omega\times (0, T)$ and $h\in H^1(\Omega; \R^m)$, if $\mathcal{L}^*_0=\mathcal{L}_0$ and $\Omega$ is $C^{1,1}$, then $$\label{3.10-10}
\| u_0\|_{L^2(0, T; H^2(\Omega))}
\le C \Big\{ \| F\|_{L^2(\Omega_T)} +\| h\|_{H^1(\Omega)} \Big\}.$$ This may be proved by using integration by parts as well as $H^2$ estimates for $\mathcal{L}_0$ [@Lady]. ]{}
Proof of Theorems \[main-theorem-1\] and \[main-theorem-2\]
===========================================================
In this section we study the convergence rates in $L^2(\Omega_T)$ and give the proof of Theorems \[main-theorem-1\] and \[main-theorem-2\]. Throughout the section we will assume that $\Omega$ is a bounded $C^{1,1}$ domain in $\R^d$.
We first consider the initial-Dirichlet problem. Let $A^*$ denote the adjoint of $A$; i.e., $A^*=( a^{*\alpha\beta}_{ij})$ with $a_{ij}^{*\alpha\beta} (y, s)=a_{ji}^{\beta\alpha}(y,s)$. For $G\in L^2(\Omega_T)$, let $v_\e$ be the weak solution to $$\label{IDP-dual}
\left\{
\aligned
(-\partial_t +\mathcal{L}^*_\e) v_\e & = G &\quad & \text{ in } \Omega \times (0, T),\\
v_\e &=0 &\quad &\text{ on } \partial\Omega \times (0, T),\\
v_\e & =0& \quad & \text{ on } \Omega \times \{ t=T\},
\endaligned
\right.$$ where $\mathcal{L}^*_\e =-\text{\rm div} (A^{*\e} (x, t)\nabla )$ denotes the adjoint of $\mathcal{L}_\e$, and $v_0$ the weak solution to $$\label{IDP-0-dual}
\left\{
\aligned
(-\partial_t +\mathcal{L}^*_0) v_0 & = G &\quad & \text{ in } \Omega \times (0, T),\\
v_0 &=0 &\quad &\text{ on } \partial\Omega \times (0, T),\\
v_0 & =0& \quad & \text{ on } \Omega \times \{ t=T\},
\endaligned
\right.$$ where $\mathcal{L}_0^*=-\text{\rm div} (\widehat{A}^* \nabla )$. Observe that $v_\e (x, T-t)$ and $v_0 (x, T-t)$ are solutions of the initial-Dirichlet problems of (\[IDP\]) and (\[IDP-0\]), respectively, with coefficient matrix $A(x/\e,t/\e^2)$ replaced by $A^*(x/\e, (T-t)/\e^2)$, and with $g=0$ and $h=0$. Also note that $A^* (y, T-s)$ satisfies the same ellipticity and periodicity conditions as $A(y,s)$.
\[lemma-4.0\] Let $v_0$ be the weak solution to (\[IDP-0-dual\]). Then $$\label{4.0-0}
\| \nabla v_0\|_{L^2(\Omega_T)} +\delta^{-1/2} \|\nabla v_0\|_{L^2(\Omega_{T, \delta})}
\le C \| G \|_{L^2(\Omega_T)},$$ where $\delta \in (2\e, 20\e)$ and $C$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$.
The estimate for $\|\nabla v_0\|_{L^2(\Omega_T)}$ follows directly from the energy estimate, while the estimate for $\delta^{-1/2} \|\nabla v_0\|_{L^2(\Omega_{T, \delta})}$ is proved in Remarks \[remark-u-0\] and \[remark-3.10\].
Let $$\label{z}
\aligned
z_\e (x, t)= v_\e (x, T-t) -v_0 (x, T-t)- & \e \chi_{T, j}^{*\e} S_\e \left(\widetilde{\eta}(x,t) \frac{\partial v_0}{\partial x_j}
(x, T-t)\right) \\
- &\e^2 \phi_{T, (d+1)ij}^{*\e} \frac{\partial}{\partial x_i}
S_\e \left( \widetilde{\eta}(x,t) \frac{\partial v_0}{\partial x_j} (x, T-t)\right),
\endaligned$$ where $\chi_T^*$ and $\phi_T^*$ denote the correctors and dual correctors, respectively, for the family of parabolic operators $\partial_t +\text{div} (A^*(x/\e, (T-t)/\e^2)\nabla )$, $\e>0$. The cut-off function $\widetilde{\eta}$ in (\[z\]) is chosen so that $\widetilde{\eta} (x,t)=0$ if $(x, t)\in \Omega_{T, 10\e}$, $\eta(x, t)=1$ if $(x, t)\in \Omega_T \setminus \Omega_{T, 15\e}$, $|\nabla \widetilde{\eta} |\le C \e^{-1}$ and $|\partial_t \widetilde{\eta}|\le C \e^{-2}$.
\[lemma-4.1\] Let $z_\e$ be defined by (\[z\]). Then $$\label{4.1-0}
\| \nabla z_\e\|_{L^2(\Omega_T)} \le C \sqrt{\e} \| G\|_{L^2(\Omega_T)},$$ where $C$ depends at most on $d$, $m$, $\mu$, $T$ and $\Omega$.
Since $A^* (y, T-s)$ satisfies the same ellipticity and periodicity conditions as $A(y,s)$ and $\Omega$ is $C^{1,1}$, this follows from the estimate (\[3.10-4\]).
We are in a position to give the proof of Theorem \[main-theorem-1\].
Let $u_\e\in L^2(0, T; H^1(\Omega))$ and $u_0\in L^2(0, T; H^2(\Omega))$ be solutions of (\[IDP\]) and (\[IDP-0\]), respectively. Let $G\in L^2(\Omega_T)$. By duality it suffices to show that $$\label{4.3-1}
\aligned
& \Big| \iint_{\Omega_T} (u_\e -u_0) \cdot G \Big|\\
&\le
C \e \| G\|_{L^2(\Omega_T)}
\left\{ \| u_0\|_{L^2(0, T; H^2(\Omega))}
+\|\partial_t u_0\|_{L^2(\Omega_T)}
+\sup_{\e^2<t<T}
\left(\frac{1}{\e} \int_{t-\e^2}^t \int_\Omega |\nabla u_0|^2 \right)^{1/2} \right\}.
\endaligned$$ Let $w_\e$ be defined by (\[w\]), with $\delta=2\e$. Since $$\| \chi^\e K_\e (\nabla u_0)\|_{L^2(\Omega_T)}
+\e \| \phi^\e \nabla K_\e (\nabla u_0)\|_{L^2(\Omega_T)}
\le C \|\nabla u_0\|_{L^2(\Omega_T)},$$ we only need to prove that $\displaystyle | \iint_{\Omega_T} w_\e \cdot G|$ is bounded by the r.h.s. of (\[4.3-1\]).
To this end we write $$\label{4.3-2}
\aligned
&\iint_{\Omega_T} w_\e \cdot G
= \int_0^T \big\langle \partial_t w_\e, v_\e \big\rangle +\iint_{\Omega_T} A^\e \nabla w_\e \cdot \nabla v_\e\\
&= \left\{ \int_0^T \big\langle \partial_t w_\e, z_\e (\cdot, T-t)\big\rangle
+\iint_{\Omega_T} A^\e \nabla w_\e \cdot \nabla z_\e (x, T-t)\right\}\\
&\quad +\left\{ \int_0^T\big \langle \partial_t w_\e, v_0 \big\rangle
+\iint_{\Omega_T} A^\e \nabla w _\e\cdot \nabla v_0\right\} \\
&\quad +\left\{ \int_0^T\big \langle \partial_t w_\e, v_\e -v_0 -z_\e (\cdot, T-t)\big \rangle
+ \iint_{\Omega_T} A^\e \nabla w_\e \cdot
\big\{ v_\e -v_0 - z_\e (\cdot, T-t) \big\}\right\}\\
&=J_1 +J_2 +J_3,
\endaligned$$ where $\langle, \rangle$ denotes the pairing between $H^1_0(\Omega)$ and its dual $H^{-1}(\Omega)$. We shall use Lemma \[main-lemma-3.1\] to bound $J_1$, $J_2$ and $J_3$.
For the term $J_1$, it follows by Lemma \[main-lemma-3.1\] that $$\label{4.3-3}
\aligned
|J_1| & \le C \sqrt{\e}
\Big \{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2} \|\nabla u_0\|_{L^2(\Omega_{T, 6\e})} \Big\}
\| \nabla z_\e\|_{L^2(\Omega_T)}\\
& \le C \e \Big \{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2} \|\nabla u_0\|_{L^2(\Omega_{T, 6\e})} \Big\} \| G\|_{L^2(\Omega_T)},
\endaligned$$ where we have used Lemma \[lemma-4.1\] for the last step.
Next, for $J_2$, we obtain $$\label{4.3-4}
\aligned
|J_2|
&\le C
\Big \{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2} \|\nabla u_0\|_{L^2(\Omega_{T, 6\e})} \Big\}\\
&\qquad\qquad \qquad\qquad
\cdot
\Big\{ \e \| \nabla v_0\|_{L^2(\Omega)} +\e^{1/2} \|\nabla v_0\|_{L^2(\Omega_{T, 6\e})} \Big\}\\
&\le C \e \Big \{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2} \|\nabla u_0\|_{L^2(\Omega_{T, 6\e})} \Big\}\| G\|_{L^2(\Omega_T)},
\endaligned$$ where we have used Lemma \[lemma-4.0\] for the last inequality.
To estimate $J_3$, we note that $v_\e -v_0 -z_\e(x, T-t)$ is supported in $\Omega_T\setminus \Omega_{T, 10\e}$ and in view of (\[z\]) and Lemmas \[lemma-S-1\] and \[lemma-S-3\], $$\|\nabla (v_\e -v_0 -z_\e (x, T-t))\|_{L^2(\Omega_T)}
\le C \| \nabla v_0\|_{L^2(\Omega_T)}\le C \| G\|_{L^2(\Omega_T)}.$$ It follows by Lemma \[main-lemma-3.1\] that $$|J_3|
\le C \e \Big \{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2} \|\nabla u_0\|_{L^2(\Omega_{T, 6\e})} \Big\} \| G\|_{L^2(\Omega_T)}.$$ This, together with (\[4.3-3\]) and (\[4.3-4\]), shows that $$\aligned
&\Big| \iint_{\Omega_T} w_\e \cdot G\Big|\\
& \le C \e\Big \{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\|\partial_t u_0\|_{L^2(\Omega_T)}
+\e^{-1/2} \|\nabla u_0\|_{L^2(\Omega_{T, 6\e})} \Big\} \| G\|_{L^2(\Omega_T)}\\
& \le C \e\left \{ \| u_0\|_{L^2(0, T; H^2(\Omega))} +\|\partial_t u_0\|_{L^2(\Omega_T)}
+\sup_{\e^2<t<T}
\left(\frac{1}{\e} \int_{t-\e^2}^t \int_\Omega |\nabla u_0|^2 \right)^{1/2}
\right\} \| G\|_{L^2(\Omega_T)},
\endaligned$$ which completes the proof.
Finally, we give the proof of Theorem \[main-theorem-2\]
The proof of Theorem \[main-theorem-2\] is similar to that of Theorem \[main-theorem-1\]. Indeed, let $u_\e\in L^2(0, T; H^1(\Omega))$ and $u_0\in L^2(0,T; H^2(\Omega))$ be solutions of (\[INP\]) and (\[INP-0\]), respectively. Let $w_\e$ be defined as in (\[w\]), with $\delta=2\e$. To estimate $\| u_\e -u_0\|_{L^2(\Omega_T)}$, we consider $\displaystyle \iint_{\Omega_T} w_\e \cdot G$, where $G\in L^2(\Omega_T)$. Let $v_\e$ be the weak solution to $$\label{INP-dual}
\left\{
\aligned
(-\partial_t +\mathcal{L}^*_\e) v_\e & = G &\quad & \text{ in } \Omega \times (0, T),\\
\frac{\partial v_\e}{\partial \nu^*_\e} &=0 &\quad &\text{ on } \partial\Omega \times (0, T),\\
v_\e & =0& \quad & \text{ on } \Omega \times \{ t=T\},
\endaligned
\right.$$ and $v_0$ the weak solution to $$\label{INP-dual-0}
\left\{
\aligned
(-\partial_t +\mathcal{L}^*_0) v_0 & = G &\quad & \text{ in } \Omega \times (0, T),\\
\frac{\partial v_0}{\partial \nu^*_0} &=0 &\quad &\text{ on } \partial\Omega \times (0, T),\\
v_0 & =0& \quad & \text{ on } \Omega \times \{ t=T\},
\endaligned
\right.$$ where $\frac{\partial v_\e}{\partial\nu^*_\e}$ and $\frac{\partial v_0}{\partial\nu_0^*}$ denote the conormal derivatives associated with the operators $\mathcal{L}_\e^*$ and $\mathcal{L}_0^*$, respectively. Let $z_\e$ be defined as before. Note that estimates in Lemmas \[lemma-4.0\] and \[lemma-4.1\] continue to hold. Moreover, by (\[INP-dual\]), we have $$\iint_{\Omega_T} w_\e \cdot G
=\int_0^T \big\langle \partial_t w_\e, v_\e\big\rangle
+\iint_{\Omega_T} A^\e \nabla w_\e \cdot \nabla v_\e,$$ where $\langle, \rangle$ denotes the pairing between $H^1(\Omega)$ and its dual. With Lemma \[main-lemma-3.2\] at our disposal, the rest of the proof is exactly the same as that of Theorem \[main-theorem-1\]. We omit the details.
Jun Geng, School of Mathematics and Statistics, Lanzhou University, Lanzhou, P.R. China.
E-mail:[email protected]
Zhongwei Shen, Department of Mathematics, University of Kentucky, Lexington, Kentucky 40506, USA.
E-mail: [email protected]
[^1]: Supported in part by the NNSF of China (11571152) and Fundamental Research Funds for the Central Universities (LZUJBKY-2015-72).
[^2]: Supported in part by NSF grant DMS-1161154.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Reliable anticipation of pedestrian trajectory is imperative for the operation of autonomous vehicles and can significantly enhance the functionality of advanced driver assistance systems. While significant progress has been made in the field of pedestrian detection, forecasting pedestrian trajectories remains a challenging problem due to the unpredictable nature of pedestrians and the huge space of potentially useful features. In this work, we present a deep learning approach for pedestrian trajectory forecasting using a single vehicle-mounted camera. Deep learning models that have revolutionized other areas in computer vision have seen limited application to trajectory forecasting, in part due to the lack of richly annotated training data. We address the lack of training data by introducing a scalable machine annotation scheme that enables our model to be trained using a large dataset without human annotation. In addition, we propose Dynamic Trajectory Predictor (DTP), a model for forecasting pedestrian trajectory up to one second into the future. DTP is trained using both human and machine-annotated data, and anticipates dynamic motion that is not captured by linear models. Experimental evaluation confirms the benefits of the proposed model.'
author:
- 'Olly Styles$^{1}$, Arun Ross$^{2}$ and Victor Sanchez$^{1}$[^1][^2]'
bibliography:
- 'ivbib.bib'
title: '**Forecasting Pedestrian Trajectory with Machine-Annotated Training Data** '
---
INTRODUCTION
============
Interacting with humans in complex urban environments remains a challenging problem for autonomous vehicles (AVs). Unlike highways with well-defined rules for traffic, urban environments necessitate that vehicles interact with other road users, such as pedestrians and cyclists, in a more nuanced manner. For an AV to navigate effectively in such environments, the vehicle must be able to locate and react to pedestrians in order to avoid collisions. The first component of such a navigation system, detecting pedestrians, has seen a tremendous amount of research effort in the past decade [@tenyearsofdetection]. If current trends continue, performance will soon match and even surpass human-level performance on standard evaluation benchmarks [@howfarfromsolving]. The rapid advancements in this area have led to real-world implementations of advanced driver assistance systems (ADAS) to aid drivers in critical situations. Such systems are capable of providing warnings or initiating braking if a pedestrian is detected in front of the vehicle, but are less reliable in the anticipation of potentially dangerous events before a pedestrian steps into the roadway.
As vehicles move towards ever greater autonomy, the need for accurate pedestrian trajectory forecasting also grows. With a human driver in the loop, ADAS may be designed conservatively as false negatives can be tolerated. For an AV, however, the reliable anticipation of pedestrian *intent* is a critical safety feature but a complex challenge. Although driven by long-term motion goals such as reaching a specific destination [@goaldirected], pedestrian motion is highly dynamic and may change at a moment’s notice, such as a child running rapidly into the street. To deal with this uncertainty, human drivers use heuristics such as pedestrian head pose, gait, and scene dynamics to reason about intent [@agreeing]. Without these cues, for example, human drivers find it more challenging to predict if a pedestrian is about to cross the road [@atthekerb].
![We propose a model and training regime for pedestrian trajectory forecasting. Due to a lack of annotated training data, our model is trained jointly with human-annotated and machine-annotated pedestrian bounding boxes generated by a pedestrian detection and tracking algorithm.[]{data-label="fig:pull"}](./figs/pullv2.png){width="0.92\linewidth"}
Modern vehicles equipped with sensors such as LiDAR and Radar can build an accurate representation of the surrounding environment [@perception-survey]. Both LiDAR and Radar, however, lack the capability for extracting high-resolution features and are, thus, commonly supplemented with visible spectrum cameras. Manual annotation of features such as pedestrian head pose and body language cues from camera data is challenging and time-consuming. Furthermore, pedestrian behavior varies across different cultures and driving environments [@culturechapter]. A model trained to anticipate pedestrian behavior in California, USA is unlikely to perform well on the streets of Mumbai, India. Without a practical method for learning from unlabelled data, it is likely that large quantities of data must be manually annotated for deployment in each environment.
Based on the above observations, we present a system for pedestrian trajectory forecasting capable of learning from unlabeled data. The two main contributions of this work are as follows:
1. Dynamic Trajectory Predictor (DTP), a pedestrian trajectory forecasting deep learning model based on motion features from optical flow.
2. A machine annotation scheme for training trajectory forecasting models in the absence of labeled data.
RELATED WORK
============
Our proposed approach builds on the substantial progress made in pedestrian detection and human action recognition. However, in this section, we concentrate on literature more directly relevant to our contributions, that are focused on (a) pedestrian trajectory forecasting and (b) alternative supervision methods for training models in the absence of large-scale human annotated datasets. For pedestrian detection, see recent surveys such as [@tenyearsofdetection; @towardshuman]. For action recognition, see recent surveys such as [@ar-survey-1; @ar-survey-2].
Pedestrian Trajectory Forecasting
---------------------------------
**Dynamic Systems Approach.** Given the absence of large pedestrian trajectory datasets, previous works have modeled the dynamic motion of pedestrian’s using linear dynamic systems (LDS) that combine the assumptions of constant velocity (CV) or constant acceleration (CA) with a filtering algorithm such as the Kalman filter [@pathpredictionwithbayesian]. To model non-linear, dynamic motion, a switching linear dynamic system (SLDS) uses a discrete Markov chain to select between multiple LDS at each timestep based on past observations. However, the SLDS is limited to *reacting* to pedestrian motion rather than *anticipating* a change in dynamics. To address this issue, existing works [@context-based; @gavrilaijcv] focus on additional cues such as pedestrian head pose, motion state, and road scene context or use a non-linear filtering algorithm such as the unscented Kalman filter [@ukf].
**Data-Driven Approach.** Data-driven approaches for trajectory forecasting have gained attention in recent years resulting from the success of deep learning models for related problems such as image classification, action recognition, and pedestrian detection. In particular, deep learning models have been applied to trajectory forecasting in a surveillance setting with a fixed overhead camera on datasets such as UCY [@ucy] and ETH [@eth], or forecasting vehicle trajectories [@egocentric]. In [@behavior-cnn], pedestrian trajectory is forecast by encoding pedestrian location as a sparse vector which is used directly as input to a convolutional neural network (CNN). In [@sociallstm], pedestrian trajectory is forecast from a static, overhead camera using a long short-term memory (LSTM) network. The authors introduce social pooling, which models the social interactions between multiple pedestrians.
Trajectory forecasting is considered from a first-person perspective in [@fpl]. The authors propose a model combining features from the pedestrian’s pose, estimated ego-motion, and past location information. Similarly, in [@longterm], an LSTM is used to predict the future location of pedestrian bounding boxes by first estimating future ego-motion and then using these estimates with observed bounding boxes to forecast the location of future bounding boxes. All data-driven approaches, however, are limited by the lack of available training data.
Alternative supervision
-----------------------
Supervised learning has been a prominent learning paradigm requiring accurate annotation of datasets, which is commonly completed manually through painstaking human effort. Due to the massive quantities of data necessary to effectively train state-of-the-art models, several alternative means of supervision have been proposed. Pre-training neural network models on the large Imagenet dataset [@imagenet], before fine-tuning on a target dataset, has become the de facto standard in settings where annotated data is limited. Alternative means of building large annotated datasets for pre-training such as mining social media websites [@instahashtags] have also been proposed. An alternative learning paradigm is self-supervision. In self-supervision, some subset of a dataset is withheld during the training process, and a model is trained to predict the withheld data. In this way, a model may exploit large-scale datasets without expensive annotation. For example, the authors in [@colorization] convert color images to greyscale and train a model to perform the inverse operation, and the authors in [@patch-prediction] predict the location of image patches in relation to another patch. In an intelligent vehicle setting, existing works have used data collected by one sensor (such as a camera) to predict the data collected by another sensor (such as an inertial measurement unit) [@predict-ego; @end-to-end-driving]. Self-supervision avoids the expensive human annotation component of supervised learning and is, therefore, well-suited to address problems with limited annotated data. Our proposed machine annotation scheme enables us to leverage the power of self-supervision for pedestrian trajectory forecasting.
PROPOSED METHOD {#sec:3}
===============
Problem formulation and baseline {#formulation}
--------------------------------
Consider a pedestrian localized in a video with an associated set of bounding box coordinates for the current and past $m$ frames, such as in Fig \[fig:pull\]. Our goal is to predict the centroid of future bounding boxes with coordinates $x_t$ and $y_t$ as to anticipate potentially dangerous events, such as a pedestrian stepping into the roadway. The horizontal and vertical components of velocity, $v^x_t$ and $v^y_t$ respectively, at time $t$ of a pedestrian relative to the vehicle in the 2D projection obtained by a camera can be estimated by taking the first order derivative of the past centroids:
$$v^x_t = \frac{x_t - x_{t-m}}{m} \quad , \quad v^y_t = \frac{y_t - y_{t-m}}{m}$$
As a baseline, we consider that the pedestrian maintains their average velocity of the previous $m$ timesteps in the future $n$ timesteps:
$$\widetilde{x}_{t+n} = x_t + v^x_t \cdot n \quad , \quad \widetilde{y}_{t+n} = y_t + v^y_t \cdot n$$
We denote a pedestrian’s centoid location at time $t$ as $L_t$, comprised of coordinates $x_t$ and $y_t$. Similarly, we denote velocity at time $t$ as $v_t$, comprised of vertical and lateral velocities $v^x_t$ and $v^y_t$. The predicted location of the centroid at time $t+n$ following the constant velocity assumption is denoted as:
$$\label{cv}
\widetilde{L}_{t+n} = L_t + v_t \cdot n$$
We focus here on predicting the centroid in the 2D coordinate space obtained by a camera, rather than the 3D world coordinates required for full localization by an AV. In practical applications, 2D object detections may be associated with 3D world coordinates using a depth estimation method such as [@stereomatching].
Dynamic Trajectory Predictor {#sec:dynamic-motion}
----------------------------
![DTP forecasts pedestrian trajectory reletive a constant velocity baseline. We use ResNet [@resnet] with modified input and output layers to compute features from past optical flow. See Section \[sec:model\] for details.[]{data-label="fig:model"}](./figs/model.png){width=".75\linewidth"}
In many scenarios, such as when a pedestrian is stationary or walking at a constant speed, the constant velocity assumption is a reasonable predictor of future location. Challenging situations are instances that deviate significantly from this assumption. An effective model must anticipate a change in velocity and adjusts predictions accordingly. The error resulting from the constant velocity assumption is denoted by:
$$\widetilde{e}_{t+n} = |L_{t+n} - \widetilde{L}_{t+n}|$$
Rather than directly predicting a location $\hat{L}_{t+n}$ directly, existing works [@fpl; @longterm] output the location relative to the last observed timestep, $\Delta L_{t+n} = L_{t+n} - L_{t}$. In contrast, we propose to output a compensation term, $C_t = - \widetilde{e}_t$, which corrects for errors in the constant velocity assumption. In this way, our model is first initialised to a strong baseline (in the case where $C_t = 0$, the model’s predictions equal constant velocity) and then fine-tunes predictions on training examples for which the constant velocity assumption results in errors. The final predicted coordinates in the original 2D image projection, $\hat{L}_{t+n}$, are then recovered as follows:
$$\hat{L}_{t+n}=\tilde{L}_{t+n} + \hat{C}_{t+n}$$
Inspired by effective action recognition models [@twostream; @tsn], DTP uses a stack of optical flow frames as input to a CNN that extracts a compact representation of human motion. From this feature vector, a fully connected layer outputs a prediction $\hat{C}_{t+n}$ representing the estimated correction factor. A vector of large magnitude indicates that the pedestrian velocity will increase or decrease, whereas a vector of magnitude close to 0 indicates that the pedestrian will maintain their current velocity. We use ResNet [@resnet] as our backbone network, owing to its consistently good performance on many vision tasks. A high-level diagram of our model is shown in Fig. \[fig:model\]. Further details of the architecture modifications are outlined in Section \[sec:model\].
Machine annotation {#sec:unlabelled}
------------------
The training of trajectory forecasting models in a supervised learning setting requires dense (per-frame) bounding box annotation of pedestrians, which are expensive to obtain by hand. For this reason, the number and size of datasets with densely annotated pedestrian bounding boxes is limited. The size of existing datasets [@context-based; @jaad] is prohibitive for the training of high-capacity deep learning models, which rely on large quantities of data to learn an effective feature representation. To overcome this issue, we propose to learn from unlabeled data by using an automated pedestrian detection and tracking algorithm to generate bounding boxes without human labor.
Given an input video sequence, pedestrian detection algorithms obtain an estimate of the location $L_t$ for each pedestrian, and a tracking method then links these estimated locations across each timestep $t$. Given a set of such detections, we adopt the self-supervision learning paradigm by training our model to predict future pedestrian locations, $L_{t+n}$, given only the current and past locations, viz., $L_{t-m} \ldots L_t$.
A similar annotation process is proposed in [@fpl], in which pedestrians are detected and tracked using [@openpose]. However, automated detectors do not perform on par with human annotators, and make different errors to humans, such as false positive detections of vertical structures [@howfarfromsolving]. Due to this, it is not evident that models trained on machine-annotated data will generalize across datasets and to human-annotated data. To verify our proposed machine-annotation regime, we validate the performance of our model on a human-annotated dataset. We adopt the conventional methodology of pre-training on a large dataset before fine-tuning on a smaller target dataset, intending to improve generalizability on the target dataset [@howtransferable].
EXPERIMENTS {#sec:4}
===========
Datasets
--------
We use two datasets in our experiments, JAAD [@jaad] and BDD-100K [@bdd100k]. Both datasets consist of videos captured by a front-facing camera mounted behind a windshield collected by cars driving on public roads in Europe and North America. The JAAD dataset contains dense pedestrian bounding box annotation, that is, annotations are provided for each frame. BDD-100K, however, contains sparse bounding box annotation. Only one frame per video is annotated. We do not use the sparely annotated bounding boxes. Due to the huge number of videos in BDD-100K, we use only the first 10,000 videos. This subset is henceforth referred to as BDD-10K. Example pedestrian images from both datasets are shown in Fig. \[fig:flow\] (first row). Videos from the JAAD dataset are downsampled with bilinear interpolation to match the BDD-10K dataset resolution of $1280 \times 720$. The frame rate of both datasets is downsampled from 30 to 15 frames per second to reduce redundancy between consecutive frames.
![Example pedestrians with associated optical flow obtained using Flownet2-CSS. Left 3 images are human-annotated pedestrians from the JAAD dataset, right 3 images are pedestrians detected on the BDD-100k dataset using YOLOv3. Optical flow captures motion resulting from both the camera and pedestrian.[]{data-label="fig:flow"}](./figs/flow.png){width=".90\linewidth"}
{height="9.9cm" width="0.85\linewidth"}
Dynamic Trajectory Predictor {#sec:model}
----------------------------
**Implementation.** To evaluate DTP, we use the JAAD dataset. Pedestrians smaller than 50 pixels in height, occluded pedestrians, and tracks shorter than 25 frames are discarded. Optical flow is extracted from cropped pedestrians using the provided human-annotated bounding boxes with the Flownet2-CSS algorithm [@FlowNet2]. Pixel displacements are clipped at $\pm 50$ and scaled to the range $[0,1]$. Example pedestrian flow images are shown in Fig. \[fig:flow\] (second row).
We use a stack of $m$ horizontal and $m$ vertical optical flow frames at timesteps $t-m$ to $t$. Features are computed from the $2m$ input channels using the ResNet-18 CNN architecture [@resnet]. We modify the first convolutional layer to use $2m$ input channels rather than 3, keeping other dimensions the same. We replace the 1000-D softmax output layer with a 30-D fully connected later which produces predictions for the $x$ and $y$ coordinates of the 15 future bounding box centroids. We use cross-modality pre-training and partial batch normalization [@tsn] to initialize our CNN with ImageNet weights. The model is optimized to minimize the $\mathcal{L}_2$ loss between the true and predicted future locations, $L_{t+1} \ldots L_{t+n}$ and $\hat{L}_{t+1} \ldots \hat{L}_{t+n}$, and is trained until convergence using the Adam [@adam] optimizer with an initial learning rate of $10^{-5}$, which is reduced to $10^{-6}$ once performance saturates. We use a batch size of 64 and a weight decay of $10^{-2}$. Each pedestrian is resized to $256 \times 256$ pixels. For data augmentation, a randomly cropped sub-image of size $224 \times 224$ is taken.
We split the JAAD dataset into training (videos 0-250) and testing (videos 251-346) sets. We perform 5 fold cross-validation on the JAAD training set to tune hyperparameters. Once hyperparameters are fixed, we obtain an estimate of the model’s generalizability by training on each of the 5 folds until performance on the respective validation set saturates. We then evaluate the model on the test set. We report the mean performance on the test set with associated 95% confidence intervals for the 5 folds.
**Evaluation.** We use two metrics to evaluate model performance, mean squared error (MSE) and displacement error (DE@$t$) at timesteps up to 15, following [@longterm; @fpl]. The MSE is the mean of the squared errors of the predicted centroid in pixels from all timesteps 1 to $n$ and across all samples in the test set. The DE@$t$ is the mean Euclidean distance in pixels of the predicted and ground truth centroid for timestep $t$ only. Both metrics are relative to an image resolution of $1280 \times 720$.
We evaluate our proposed approach with 4 different inputs: a single RGB frame at time $t$, a single optical flow frame at time $t$, a stack of 5 optical flow frames at times $t-4$ to $t$, and a stack of 9 optical flow frames at times $t-8$ to $t$. We use 9 as our maximum value of $m$ rather the then 10 frames commonly used for action recognition [@twostream; @tsn] for a fair comparison with Future Person Localization (FPL) [@fpl], which uses 10 frames as input. As each optical flow frame requires two consecutive RGB frames to be computed, using 10 input frames results in 9 optical flow frames. Following prior works [@fpl; @sociallstm] we adopt constant velocity (CV) and constant acceleration (CA) as baselines. For the CV baseline, we compute the average velocity in the image space using the previous locations and predict the future location assuming the pedestrian maintains a linear velocity. Similarly, for the CA baseline, we compute the average acceleration using the previous locations and extrapolate these values into the future timesteps assuming linear acceleration. Using 4 previous locations resulted in the best cross-validation performance.
**Results.** Table \[tab:inputs\] shows the performance of each model with different input modalities in comparison with the CV and CA baselines. Due to the relatively poor performance of the RGB input, we do not fuse RGB and optical flow models as in the two-stream model [@twostream]. Example outputs of our model using a stack of 9 optical flow frames compared to baselines are shown in Fig. \[fig:qualitative\]. DTP performs particularly well in situations where a pedestrian first begins walking, and when the ego-vehicle begins to turn sharply (top two rows). DTP performs less well under conditions of significant background motion such as those due to other vehicles (bottom left image) or upper body motion in the counter walking direction (bottom right image).
We compare our method using a stack of 9 optical flow frames with linear baselines and FPL [@fpl] in Table \[tab:model\]. We modify FPL to output 15 timesteps into the future rather than the 10 as in the original architecture and use optical flow for ego-motion estimation as described in [@fpl]. Both DTP and FPL see a reduction in error with our proposed CV correction term $C_t$ (rather than directly predicting the location displacement $\Delta L_{t+n}$). DTP attains the best performance.
**Input modality** **MSE** **DE@5** **DE@10** **DE@15**
--------------------------- ---------------- ---------------- ----------------- -----------------
CA $1426$ 15.3 28.3 $52.8$
CV $1148$ 16.0 26.4 $47.5$
RGB frame $1042$ 11.6 24.9 $45.2$
Optical flow frame $873$ 11.1 23.0 $41.2$
5 optical flow frames $651$ 9.4 19.3 $35.6$
**9 optical flow frames** $\mathbf{610}$ $\mathbf{9.2}$ $\mathbf{18.7}$ $\mathbf{34.6}$
: Input modality comparision.
\[tab:inputs\]
**Model** **MSE** **DE@15**
------------ -- ----------------------- -------------------------
FPL [@fpl] $1405 \pm 182$ $49.5 \pm 2.9 $
FPL [@fpl] $881 \pm 44$ $41.3 \pm 1.2$
DTP $1404 \pm 94$ $54.6 \pm 2.6 $
**DTP** $\mathbf{610 \pm 21}$ $\mathbf{34.6 \pm 0.5}$
: Model comparision.
\[tab:model\]
Machine annotation {#machine-annotation}
------------------
**Implementation.** We annotate pedestrian bounding boxes in the BDD-10K dataset using two popular off-the-shelf object detectors, YOLOv3 [@yolo3] and Faster-RCNN [@faster]. Although the detectors are capable of detecting a variety of objects, we use the pedestrian class only. Our aim here is to evaluate the robustness of our proposed system to multiple automated detectors, rather than to compare detector performance directly. Nonetheless, for consistency, we train both detectors on the same dataset (MS-COCO [@coco]) and threshold confidence scores at $0.6$.
Once frame-wise detections are obtained, detections are associated across frames using the Deepsort [@deepsort] tracking-by-detection algorithm resulting in a series of bounding boxes and tracking identifiers. We use the same setup as the JAAD dataset and discard detections with height fewer than 50 pixels, and tracks shorter than 25 frames. Using this annotation scheme, we find a total of 16,900 valid non-overlapping pedestrian tracks using YOLOv3 and 13,200 using Faster-RCNN.
**Evaluation.** We use an 80%-20% training-validation split for BDD-10K. We pre-train DTP on BDD-10K using the same hyperparameters as outlined in Section \[sec:model\]. Once performance on the validation set saturates, the model is fine-tuned on the JAAD training set. We evaluate the trajectory forecasting performance with and without pre-training rather than the pedestrian detection quality, owing to the lack of human-annotated bounding boxes.
**Results**. The impact of machine-annotated pre-training using the YOLOv3 detector before fine-tuning on the human-annotated JAAD dataset is shown in Table \[tab:pre-training\].
**Model** **MSE** **DE@15**
------------ -- ----------------------- -------------------------
FPL [@fpl] $881 \pm 44$ $41.3 \pm 1.2 $
FPL [@fpl] $805 \pm 46$ $40.1 \pm 1.2$
DTP $610 \pm 21$ $34.6 \pm 0.5$
**DTP** $\mathbf{539 \pm 13}$ $\mathbf{32.7 \pm 0.4}$
: Impact of pre-training on BDD-10K with YOLOv3.
\[tab:pre-training\]
We evaluate the impact of pre-training dataset size and pedestrian detector by training on subsets of BDD-10K ranging from 20% to 100% of the total dataset size. Fig. \[fig:pre-training\] shows the MSE on the JAAD test set for both YOLOv3 and Faster-RCNN. In general, the error on the JAAD test set reduces as larger subsets of our machine-annotated dataset, BDD-10k, is used for pre-training. The reduction in error may be due to the model’s ability to learn the motion patterns of under-represented classes, such as children or the elderly, from a larger dataset.
![Impact of pre-training dataset size and pedestrian detection algorithm on the performance on JAAD test set. Shaded areas show the 95% confidence interval.[]{data-label="fig:pre-training"}](./figs/pretraining.png){width="0.95\linewidth" height="5.6cm"}
CONCLUSION
==========
We have presented a model and complementary machine annotation scheme for pedestrian trajectory forecasting from onboard a moving vehicle. Our model, DTP, forecasts trajectory for time horizons up to one second by anticipating a change in velocity using optical flow information. By introducing a method for annotating data without human labor, DTP and other similar models may leverage large-scale datasets for learning effective feature representations.
ACKNOWLEDGMENT {#acknowledgment .unnumbered}
==============
This work is funded by the UK EPSRC (grant no. EP/L016400/1) and the EU Horizon 2020 project IDENTITY (Project No. 690907). Portions of this work were done when Styles was at MSU. Our thanks to NVIDIA for supporting this research with their generous hardware donation.
[^1]: $^{1}$Olly Styles and Victor Sanchez are with the Department of Computer Science, University of Warwick, Coventry, UK [{o.c.styles | v.f.sanchez-silva}@warwick.ac.uk]{}
[^2]: $^{2}$Arun Ross is with the Department of Computer Science and Engineering, Michigan State University, East Lansing, USA [[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Word embeddings capture semantic relationships based on contextual information and are the basis for a wide variety of natural language processing applications. Notably these relationships are solely learned from the data and subsequently the data composition impacts the semantic of embeddings — which arguably can lead to biased word vectors. Given qualitatively different data subsets, we aim to align the influence of single subsets on the resulting word vectors, while retaining their quality. In this regard we propose a criteria to measure the shift towards a single data subset and develop approaches to meet both objectives. We find that a weighted average of the two subset embeddings balances the influence of those subsets while word similarity performance decreases. We further propose a promising optimization approach to balance influences and quality of word embeddings.'
author:
- |
Stephanie Brandl, David Lassner, Maximilian Alber\
`[stephanie.brandl, lassner, maximilian.alber]@tu-berlin.de`\
\
Technische Universität Berlin, 10623 Berlin, Germany\
Machine Learning Group\
bibliography:
- 'main.bib'
title:
- 'Instructions for [EMNLP-IJCNLP 2019]{} Proceedings'
- |
Balancing the composition of word embeddings\
across heterogenous data sets
---
Introduction
============
The advent of word embeddings [@word2vec; @pennington2014] has shifted the entire field of Natural Language Processing (NLP) from sparse representations, such as Bag-of-Words, to dense, vectorial representations that have proven to be capable of capturing meaningful syntactic and semantic concepts. Word embeddings are widely used in, e.g., text classification [@joulin2016] and machine translations [@mikolov2013b]. Subsequently word embeddings have a crucial impact on downstream applications and, moreover, such models inherit (hidden) assumptions and properties of the data.
Text corpora for training word embeddings are typically composed of subsets with different properties. Properties can manifest, e.g., as U.K./U.S. English, but can also be induced by the authors, e.g., texts written by different genders, in different periods of time or in different contexts such as arts and politics. While it is the intention in the first place to capture semantic and syntactic information from the data in the best possible way — ideally by learning from as much data as possible, we argue that on second thought it is desirable to influence the composition of data (sub)sets.
Given a corpus where one category outnumbers the other, joint word embeddings will expose a bias towards the former — yet this might not reflect (actual) word semantics appropriately or can be simply undesired.
The desideratum would be for example that in a transfer learning setting word embeddings trained on a large data set are to be fine-tuned on a small task-specific data set. Or in order to achieve semantics of cultural diversity, several smaller newspaper data sets with different foci could be added to a large base newspaper data set with a Euro/US centered focus.
The problem of bias increases for word embeddings as they are often used as a starting point in e.g. downstream tasks. Those methods usually work in a black box manner whose decision making is difficult to see through.
Typical state-of-the-art embedding learning algorithms do not distinguish between different data subsets and thus merge their properties in an incidental manner. A notable exception is the work of @goikoetxea2016single that shows how text-based and wordnet-based [@Miller1995] embeddings can be combined to improve the embedding quality, yet does not align the contribution of the individual data sets. For more details on related work we refer to Appendix \[appendix:rw\].
In this contribution we research if and how the influence of individual subsets can be aligned, while retaining embedding quality w.r.t. word vectors learned on all the data. For this aim we propose a measure for the retained semantics of a subset in the final embedding and compare a total of 9 different combination methods (1-9) which are explained in detail in Section \[sec:methods\]. The combinations vary in that they are (1) trained on the complete data set, (2-4) created *without* [@goikoetxea2016single], and (5-9) *with* consideration of the data distribution (our approaches).
Related work {#appendix:rw}
============
Various authors combined text-based word embeddings with additional resources, as for instance wordnet-based information, embeddings trained by different algorithms or additional data sets [@goikoetxea2016single; @rothe2017autoextend; @speer2016ensemble; @Henriksson2014]. The main goal in those articles is to improve the quality of word embeddings overall.\
However, to the best of our knowledge, so far no one adressed the influence subsets have on a combined embedding systematically in order to balance the impact of different data sets after their composition, while retaining the quality of the word embeddings.
Evaluating the influence of data subsets on word embeddings {#sec:evaluation}
===========================================================
-------------------- -------------------- -------------------- ---------------------- --------------------- ---------- ----------- ----------- ---------------------- --------------------- ---------------------- --------------------- ----------- ----------- -----------
$\mathcal{J}^{90}$ $\mathcal{J}^{00}$ $\mathcal{J}^\Delta$ $\bar{\mathcal{J}}$ n=1 n=5 n=10 $\mathcal{J}^{arts}$ $\mathcal{J}^{pol}$ $\mathcal{J}^\Delta$ $\bar{\mathcal{J}}$ n=1 n=5 n=10
\(a) $U_{s}$ 1.00 0.15 0.85 **0.57** 1.61 6.82 9.33 1.00 0.20 0.80 **0.60** 7.47 29.05 36.69
\(b) $U_{l}$ 0.15 1.00 -0.85 **0.57** 1.55 9.66 13.29 0.20 1.00 -0.80 **0.60** 19.37 53.54 62.12
\(1) $U_{s/l}$ 0.20 0.42 -0.21 0.31 *3.82* *16.05* *20.91* 0.27 0.55 -0.27 0.41 **21.21** *60.58* **69.13**
\(2) AVG 0.24 0.44 -0.21 *0.34* *2.67* 11.24 15.31 0.32 0.54 -0.22 *0.43* *19.21* *53.15* *62.20*
\(3) CON 0.30 0.39 -0.09 *0.34* 1.61 *12.16* *16.03* 0.39 0.45 -0.06 0.42 13.70 52.78 61.71
\(4) PCA 0.28 0.36 -0.08 0.32 1.94 9.94 13.52 0.36 0.43 -0.07 0.40 17.08 48.44 57.10
\(5) SAMP 0.22 0.30 -0.08 0.26 **4.29** **18.46** **23.48** 0.30 0.45 -0.15 0.38 *20.36* **60.86** **69.13**
\(6) WAVG 0.31 0.31 **-0.01** *0.31* 2.76 10.74 14.10 0.41 0.42 **-0.01** *0.42* 16.79 48.26 57.56
\(7) $\tau_{10}$ 0.21 0.37 -0.16 0.29 2.57 14.32 18.69 0.28 0.48 -0.20 0.38 14.71 53.94 63.01
\(8) $\tau_{100}$ 0.23 0.37 -0.14 0.30 2.35 12.32 16.28 0.29 0.47 -0.19 0.38 14.85 52.95 62.25
\(9) $\tau_{1000}$ 0.30 0.32 **-0.01** *0.31* 2.67 10.94 14.20 0.38 0.42 -0.04 0.40 15.33 47.87 57.38
-------------------- -------------------- -------------------- ---------------------- --------------------- ---------- ----------- ----------- ---------------------- --------------------- ---------------------- --------------------- ----------- ----------- -----------
Considering how embeddings encode word contexts, we illustrate the influence of data subsets on the final embedding on two real world data sets.
*New York Times 1990-2016:* The New York Times dataset[^1] (NYT) contains headlines and lead texts of news articles published online and offline in the New York Times between 1990 and 2016 with a total of 99.872 documents. Political offices as well as sports teams are very closely discussed based on their representatives players, hot topics and their current score. Their context changes over time. As word embeddings are mainly based on the context of a word, their connotation and vectorial representation are influenced by those changes. We investigate the influence of these changes on common word embeddings by splitting this data set in subsets, the first one reaching from 1990-1999 (33.383 articles) and a second one from 2000-2016 (62.058 articles)
*English Wikipedia:* The Wikipedia data set (Wiki) contains articles from the English Wikipedia snapshot from April 1st, 2019. We select 12.236 articles from the category *Arts* as well as 24.473 articles from the category *Politics* to analyse the individual influence of those 2 fields on joint word embeddings.\
As a first example, we consider the word *shooting* whose nearest neighbors (NNs) in both category groups of the Wiki data set are shown in Fig.\[fig:tsne\]. Clearly, within Politics, *shooting* refers mostly to the firing of a gun, for Arts, *shooting* rather relates to a photo or movie shooting. When we train embeddings on the joint data set, the new vector reflects both realities, but is biased towards Politics due to the increased number of articles (23/100 and 51/100 common neighbors with the embedding from Arts and Politics, respectively).
![2D tSNE embeddings of the word *shooting* with its NN in different embeddings trained on Wiki: (a), (b), (8) in red, blue and orange respectively.[]{data-label="fig:tsne"}](tsne_embed_word_shooting_tau.png){width="\columnwidth"}
Given this intuition, we would like to quantify the retained influence of the data subset (a) and (b) on embeddings $U$. Inspired by the Jaccard index we compare the neighboring words of a given embedding trained on a subset and those of the composed embeddings $U$. In more detail, given the sets of $n$ NNs, $\mathcal{N}_n(u^i)$ and $\mathcal{N}_n(v^i)$ for two embeddings $u^i$ and $v^i$ of a word $i$, the ratio of shared nearest words is: $$\mathcal{J}_n(u^i, v^i) = \frac{|\mathcal{N}_n(u^i) \cap \mathcal{N}_n(v^i)|}{n}$$ and we denote the average over all words as $\mathcal{J}_n(U, V)$. For instance $\mathcal{J}_n(U, V) = 0.6$ would mean that words in $U$ and $V$ share on average $60\%$ of their $n$ NNs. We will use $\mathcal{J}^{s}(U) = \mathcal{J}_n(U_{s}, U)$ and $\mathcal{J}^{l}(U) = \mathcal{J}_n(U_{l}, U)$ to indicate the retained influence of the according subsets on a resulting embedding. We use the cosine similarity to compute NNs for neighborhoods of different sizes $n$.
Methods {#sec:methods}
=======
We use a number of different embeddings that can be divided into three groups: merging the data *before* learning the embeddings, *static* merging algorithms, and *dynamic* merging approaches.
*[Baselines - (a), (b), (1)]{}* As baselines we train word embeddings with GloVe [@pennington2014] on NYT on articles from (a) 1990-1999 and (b) 2000-2016. The resulting embeddings learned on (a), (b), and (1) are denoted as $U_{90}$, $U_{00}$, and $U_{90/00}$. We further trained word embeddings with GloVe on Wiki for
\(a) *Arts* ($U_{arts}$), (b) *Politics* ($U_{pol}$) and (1) the merged data ($U_{a/p}$).\
GloVe embeddings are trained as 50-dimensional word embeddings on both NYT and Wiki with $x_{\max}=100$, $\alpha=0.75$. We choose a context window size of 15 for NYT and 5 for Wiki as the data set is considerably larger than NYT. We select one vocabulary for each data set and consider only words that occur at least 40 (NYT) and 250 (Wiki) times in the whole data set which leads to vocabularies of size 21398 (NYT) and 19936 (Wiki).
*[Static merging - (2), (3), (4)]{}* In constrast to (1) — merging before learning — the following approaches merge trained embeddings. They were proposed by @goikoetxea2016single. Given the embeddings $U_{s}$ and $U_{l}$ of the subsets, method (2) is to average them, i.e. $0.5 \cdot (U_{s}+U_{l})$, (3) is to concatenate them to a 100-dimensional embedding, and (4) extends (3) by extracting the 50 most informative dimensions using PCA. (3) and (4) obtained good results in @goikoetxea2016single.
*[Dynamic merging - (5), (6), (7), (8), (9)]{}* We found that previously presented embeddings are biased towards the larger subset: $\mathcal{J}_{s} << \mathcal{J}_{l}$. To alleviate this we propose the following approaches. A first attempt (5) is to upsample the smaller subset to the same size of the larger set. This leads to embeddings with a high score in analogy tests but a decrease in $\bar{\mathcal{J}}$. We further intent to balance the impact of the subsets by taking an average that is weighted by their inverse proportions (6): $U_{wavg} = 0.65 \cdot U_{s}+ 0.35 \cdot U_{l}$.\
Unfortunately, we found that this approach results in embeddings with inferior quality. We define an optimization problem that on one hand optimizes the GloVe loss to obtain qualitative good embeddings and on the other hand tries to balance the influence of the respective subsets by regularizing the distance of the solution to the weighted embeddings $U_{wavg}$. Given the co-occurence matrix $Y$ and the GloVe weighting function $f(Y)$ [@pennington2014] the embeddings $U$ are created by optimizing: $$\begin{aligned}
\label{thequation}
\operatorname*{argmin}_A ~~ & ~~ f(Y) \odot {\left\lVertU U^\top - \log Y\right\rVert_F}^2 \\
&~~+~\tau ~ {\left\lVertU - U_{wavg}\right\rVert_F}^2 \tag{2} \\[2ex]
\text{where}~~ U &= A \odot U_{s/l} + (1-A) \odot U_{wavg} \\
A_{i, j} &\in [a_{min}, a_{max}], 0 \leq a_{min} < a_{max} \leq 1\end{aligned}$$ and $\odot$ denotes a point-wise multiplication. The regularization parameter $\tau$ allows to trade-off between embedding quality and a balanced influence. We restrict the solution space to the “rectangle” between $U_{s/l}$ and $U_{wavg}$ and leave exploring an unconstrained version to future work. We optimize Eq. \[thequation\] with gradient descent. We therefore use Adam with a learning rate of $1\mathrm{e}{-3}$ and default values for $\beta$. The optimization is stopped after $10 000$ steps. We have implemented this in PyTorch.
[max width=]{}
\(a) 90 90/00 W-AVG \(b) 00 90/00 W-AVG
-------------- ------- ------- ------------- ------- -------
**war** 0 0 **war** 0 0
**vietnam** 3 1 **ii** 2 2
persian 5 3 irag 1 6
gulf 9 4 **vietnam** 3 1
era 17 5 fight 4 10
bidding - 7 combat 13 11
**ii** 2 2 wag - 21
veteran - 8 terrorism 18 15
cold 16 9 enemy 21 22
confrontaion - - hero - -
capture - 16 invasion 12 -
: 10 NNs of the word *war* are displayed for $U_{90}$ and $U_{00}$ in column 1 and 4 (NYT). In column 2 and 3 one can see the position the respective word gets after merging for $U_{90/00}$ and W-AVG. On the right side of the table we did the same for $U_{00}$. []{data-label="war"}
Results
=======
![ Values for $\mathcal{J}_n^{arts}$ and $\mathcal{J}_n^{pol}$ for $n \in \{5, 10, 25, 50\}$ for different weighting parameters of\
$x \cdot$ W-AVG$ + (1-x) \cdot U_{a/p}$ []{data-label="figurebig"}](average_intersections_wiki_arts.png){width="0.83\linewidth"}
We evaluate the quality of the obtained embeddings $U$ by measuring their performance on analogy tests [@mikolov2013b] and how the influence of the subset is balanced by measuring the number of common neighbors $\mathcal{J}_n^{s}(U)$, $\mathcal{J}_n^{l}(U)$, their average $\bar{\mathcal{J}}$ and their difference $\mathcal{J}^{\Delta} = \mathcal{J}^{s}(U)-\mathcal{J}^{l}(U)$ (see Section \[sec:evaluation\]). $J^{s}$ and $J^{l}$ indicate the respective average over $n \in \{5, 10, 25, 50\}$ for slices $s$(mall) and $l$(arge). Results for all methods, evaluated and averaged over the entire vocabulary, are summarized in Table \[tablebig\].
*Balanced influence:* First we note that embeddings of the subsets (a) $U_{s}$ and (b) $U_{l}$ have only few NNs in common. Furthermore, when trained on both subsets the embeddings (1) show a clear shift towards the larger subset (b). Qualitatively this can also be observed in Table \[war\] where we depict the $10$ NNs for the word “war” in subset (a) and subset (b); and the position of the word in the ranked neighbors of (1) in the column “90/00”. We observe that most of the NNs of (a) are not present in the first $10$ NNs of (1), while for (b) the set of 4 NNs is identical with (1). Moreover, we note that also the static merging approaches (2), (3) & (4) exhibit the same shift (see Table \[tablebig\]).\
We try to increase the influence of $U_s$ by upsampling (5) the data set to the size of $U_l$ before training GloVe embeddings. This leads to the same (or even better) quality of the word embeddings as (1) but also results in a decreased $\bar{\mathcal{J}}$. To alleviate this we propose a weighted average (6) in order to consider the subset proportions. The results in Table \[tablebig\] indicate that this simple approach indeed yields, in terms of our measure, balanced embeddings. This can also be observed exemplary in Table \[war\] where NNs of (6) correlate much more with the NNs of the respective subsets.
Unfortunately, we will see that the embedding quality suffers when performing a weighted average. With the aim to align both desiderata — balanced influence of the subsets and quality of the embeddings — we proposed an optimization procedure (7-9). From Table \[tablebig\] we read that the resulting embeddings for different regularization strengths $\tau$ are balanced, but surprisingly the influence of the respective subsets decreases in comparison to (1). As a control experiment we consider the embeddings given by a weighted average between (1) and (6) (Figure \[figurebig\]), where this drop of influence cannot be noted. Yet none of the such averaged embeddings yields good performance and balancing; which justifies the application of an optimization procedure.
*Embedding quality:* We measure the embedding quality by means of analogy tests. The embeddings trained on all the data (1) perform best in this context — hinting that it is beneficial to leverage as much information from data as possible. The statically merged embeddings (2), (3), (4) do not perform as well on our task, in contrast to the results of @goikoetxea2016single.\
Furthermore, we note that the weighted average (6) results also in a decrease in embedding quality. In contrast, we find that our optimization approach is able to capture both, embedding quality and balances the influence of the subsets.
Discussion
==========
Considering that text corpora are often composed of subsets, embedding learners merge them in *incidental* manner — either by merging the text before or the word vectors after training. We argue this can lead to undesired shifts in the embedded semantics and propose a measure for this shift as well as approaches to balance the composition of the subsets.
Our preliminary results show that one can indeed level the impact of different subsets. A weighted average of the subset embeddings yields balanced word embeddings, yet their quality decreases. The proposed optimization routine results in word vectors with good quality and balanced, yet decreased influence of the subsets.
As future work we aim to extend our empirical results and investigate the proposed optimization routine in more detail, e.g., by removing the constraints. As additional experiments we would like to investigate the influence of the different combination methods on downstream tasks, such as classification of sub-categories of the Wikipedia articles. This will further our understanding of the workings of the combination methods in comparison to the analogy tests that are not data slice specific. As alternative to the current regularization — that minimizes the distance to another, presumably balanced embedding — we would like to develop a (differentiable) regularization term that is closer related to our measure $\mathcal{J}_n^{s}(U)-\mathcal{J}_n^{l}(U)$. Adapting the work of @Berman2018TheLL, which proposes surrogate losses for the Jaccard index, seems to be a promising direction for this goal.
An interesting question posed by our results is how merging of data subsets impacts the resulting embedding semantics — considering that many NNs of $U_{s/l}$ are not NNs for the subset embeddings $U_{s}$ and $U_{l}$.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the Federal Ministry of Education and Research (BMBF) for the Berlin Big Data Center BBDC (01IS14013A) and for the MALT III project (01IS17058). We thank L. Ruff, T. Schnake, O. Eberle and S. Dogadov for fruitful discussions. We also thank the reviewers from ACL and the Workshop on Ethical, Social and Governance Issues in AI at NeurIPS 2018 for their valuable comments.
[^1]: <https://sites.google.com/site/zijunyaorutgers/>
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'L. Mancini'
- 'J. Southworth'
- 'S. Ciceri'
- 'J. J. Fortney'
- 'C. V. Morley'
- 'J. A. Dittmann'
- 'J. Tregloan-Reed'
- 'I. Bruni'
- 'M. Barbieri'
- 'D. F. Evans'
- 'G. D’Ago'
- 'N. Nikolov'
- 'Th. Henning'
date: 'Received ; Accepted'
title: 'A lower radius and mass for the transiting extrasolar planet HAT-P-8b'
---
[The extrasolar planet HAT-P-8b was thought to be one of the more inflated transiting hot Jupiters.]{} [By using new and existing photometric data, we computed precise estimates of the physical properties of the system.]{} [We present photometric observations comprising eleven light curves covering six transit events, obtained using five medium-class telescopes and telescope-defocussing technique. One transit was simultaneously obtained through four optical filters, and two transits were followed contemporaneously from two observatories. We modelled these and seven published datasets using the [[jktebop]{}]{} code. The physical parameters of the system were obtained from these results and from published spectroscopic measurements. In addition, we investigated the theoretically-predicted variation of the apparent planetary radius as a function of wavelength, covering the range $330$–$960$nm.]{} [We find that HAT-P-8b has a significantly lower radius ($1.321
\pm 0.037 \, R_{\mathrm{Jup}}$) and mass ($1.275 \pm 0.053 \,
M_{\mathrm{Jup}}$) compared to previous estimates ($1.50_{-0.06}^{+0.08} \, R_{\mathrm{Jup}}$ and $1.52_{-0.16}^{+0.18} \, M_{\mathrm{Jup}}$ respectively). We also detect a radius variation in the optical bands that, when compared with synthetic spectra of the planet, may indicate the presence of a strong optical absorber, perhaps TiO and VO gases, near the terminator of HAT-P-8b.]{} [These new results imply that HAT-P-8b is not significantly inflated, and that its position in the planetary mass–radius diagram is congruent with those of many other transiting extrasolar planets.]{}
Introduction {#sec_1}
============
Transiting extrasolar planetary systems are of interest and importance, because precise measurements of their physical properties can be achieved using spectroscopic and photometric observations. Atomic and molecular absorption within the atmosphere of transiting extrasolar planets (TEPs) can also be investigated through transmission spectroscopy (e.g.[@swain2008; @sing2009; @fossati2010]) and simultaneous multi-colour photometry (e.g.[@ballester2007; @southworth2012b]) of the transits. High-quality photometric observations not only enable the measurement of the masses and radii of TEPs to accuracies of a few percent (e.g. [@torres2008; @southworth09]), but also the detection of transit anomalies due to stellar pulsations [@cameron2010], tidal distortion [@li2010; @leconte2011], additional bodies (moons, planets) [@kipping2009; @tusnski2011], gravity darkening [@barnes2009; @szabo2011] and star spots [@pont2007; @rabus2009; @desert2011]. Besides the Rossiter-McLaughlin effect [@queloz00; @gaudi2007], the photometric follow-up on consecutive/close nights of transits of planets over parent-star starspots represents another fascinating method [@sanchis2011; @sanchis2011b; @sanchis2012; @tregloan2012] to measure the sky-projected spin-orbit alignment.
The increasing number of TEPs discovered every year is progressively revealing a remarkabel diversity. The improving statistical weight of this sample is useful for establishing the correct theoretical framework of planet formation and evolution. Accurate estimates of the planet properties (mass, radius, orbital semi-major axis, etc.) are vital for this purpose, and photometric follow-up of known TEPs can dramatically improve our knowledge of the planet’s characteristics. HAT-P-8b is a transiting hot Jupiter found by the HATNet team [@latham2009], orbiting with a period of $\sim3.07$ days around a star of spectral type F8 [@jones2010] or F5 [@bergfors2012]. At the time of its discovery it was labelled as one of the most inflated transiting giant planets, with a measured mass and radius of $M_{\rm b} =
1.52^{+0.18}_{-0.16}\,M_{\mathrm{Jup}}$ and $R_{\rm b} =
1.50^{+0.08}_{-0.06}\,R_{\mathrm{Jup}}$, respectively. These values differ by 2–3$\sigma$ from the theoretical predictions of @fortney2007.
The Rossiter-McLaughlin effect has been detected in the HAT-P-8 system using radial velocity observations from the SOPHIE and FIES spectrographs. @simpson2011 found a sky-projected orbital obliquity of $\lambda = -9.7^{+9.0}_{-7.7}$degrees and @moutou2011 found $\lambda = -17^{+9.2}_{-11.5}$degrees; both values are consistent with alignment between the orbital axis of the planet and the rotational axis of the star. The two studies between them suggested lower values for $M_{\rm b}$ and $R_{\rm
b}$, but neither calculated the physical properties of the system.
@bergfors2012 have found a faint companion to the HAT-P-8 system using *lucky imaging* observations with the AstraLux Norte instrument at the Calar Alto 2.2m telescope. The companion, a likely M2-4 dwarf, is at an angular distance of $1.027\pm 0.011$arcsec and is fainter in the SDSS $i^\prime$ and $z^\prime$ passbands by $\Delta i^\prime = 7.34 \pm 0.10$mag and $\Delta z^\prime = 6.68 \pm 0.07$mag. The faintness of this star means that it has a negligible effect on optical observations of HAT-P-8.
In this work we present eleven new follow-up light curves six transits in the HAT-P-8 system, obtained using five 1.2–2.5m telescopes. We augment these data with previously published observations of seven transits, and measure the physical properties of the system. We find a substantially lower mass and radius for the planet, removing its outlier status and relegating it to a more well-populated part of the planetary mass-radius diagram.
Our paper is structured as follows. In Sect.\[sec\_2\] we describe the observations and data reduction. In Sect.\[sec\_3\] we analyse the data, and in Sect.\[sec\_4\] we obtain refined orbital ephemerides and physical properties of the HAT-P-8 system. In Sect.\[sec\_5\] we investigate the variation of the planetary radius as function of wavelength. Several anomalies detected in the light curves are discussed in Sect.\[sec\_6\], whereas in Sect.\[sec\_7\] we summarize the results and draw our conclusions.
Observations and data reduction {#sec_2}
===============================
{width="18.cm"}
Six transits of HAT-P-8b were monitored by five different telescopes between 2009 and 2012. Two transits were followed simultaneously by two of the telescopes, and one was simultaneously observed through four optical filters. Except the first, all our transit observations were performed with the telescope defocussing method, in order to minimise the effect of Poisson, scintillation and flat-fielding noise [@southworth2009a]. We used autoguiding during all the observations, and the stars incurred a drift of 10 pixels or less over each observing sequence. The corresponding night logs are reported in Table\[Table\_1\], and the differential photometry is tabulated in Table\[Table\_2\] and plotted in Figs.\[Fig\_01\] and\[Fig\_02\].
Telescope Date Start/End times $N_{\mathrm{obs}}$ $t_{\rm exp}$ Filter Airmass Moon Aperture Scatter PSF $\beta$
----------- ------------ --------------------------- -------------------- --------------- --------------- ------------------------------------------ -------- ------------ --------- ------ ---------
Kuiper 2009 10 22 $04:03 \rightarrow 09:10$ 1083 5 Cousins $I$ $1.00 \rightarrow 1.9$ $25\%$ 10, 20, 25 1.54
Loiano 2009 11 12 $16:52 \rightarrow 22:09$ 103 118 Gunn $r$ $1.02 \rightarrow 1.01 \rightarrow 1.55$ $23\%$ 25, 45, 60 1.15 1963 1.4
INT 2010 08 28 $20:44 \rightarrow 04:02$ 131 140 Strömgren $y$ $2.04 \rightarrow 1.00 \rightarrow 1.18$ $84\%$ 45, 65, 90 0.47 3217 1.7
Loiano 2011 10 05 $18:01 \rightarrow 02:29$ 469 50 Gunn $i$ $1.12 \rightarrow 1.01 \rightarrow 1.45$ $61\%$ 19, 40, 60 1.15 908 1.4
CA 1.23m 2011 10 05 $19:27 \rightarrow 03:07$ 166 70 Johnson $R$ $1.08 \rightarrow 1.05 \rightarrow 1.54$ $61\%$ 26, 38, 55 0.83 1662 1.0
CA 2.2m 2012 08 26 $20:31 \rightarrow 04:48$ 160 80 sdss $u$ $1.52 \rightarrow 1.00 \rightarrow 1.55$ $77\%$ 25, 55, 75 2.37 2827 1.0
CA 2.2m 2012 08 26 $20:31 \rightarrow 04:48$ 157 80 Gunn $g$ $1.52 \rightarrow 1.00 \rightarrow 1.55$ $77\%$ 30, 50, 70 0.97 2463 1.2
CA 2.2m 2012 08 26 $20:31 \rightarrow 04:48$ 159 80 Gunn $r$ $1.52 \rightarrow 1.00 \rightarrow 1.55$ $77\%$ 30, 60, 90 0.71 2642 1.5
CA 2.2m 2012 08 26 $20:31 \rightarrow 04:48$ 162 80 Gunn $z$ $1.52 \rightarrow 1.00 \rightarrow 1.55$ $77\%$ 28, 55, 80 0.92 1662 1.1
Loiano 2012 10 02 $17:45 \rightarrow 02:32$ 232 70 Johnson $I$ $1.03 \rightarrow 1.00 \rightarrow 1.92$ $92\%$ 23, 40, 65 0.79 2521 1.0
CA 1.23m 2012 10 02 $20:07 \rightarrow 03:35$ 373 50 Johnson $I$ $1.11 \rightarrow 1.00 \rightarrow 2.16$ $92\%$ 35, 60, 90 1.69 2290 1.8
Kuiper 1.55m telescope {#sec_2.1}
-----------------------
We observed one transit of HAT-P-8 in October 2009 using the University of Arizona’s 1.55m Kuiper telescope on Mt. Bigelow, Arizona. We used the Mont4k CCD, binned 3$\times$3 to 0.43$^{\prime\prime}$/pixel, for a total field of view (FOV) of $9.7^{\prime} \times 9.7^{\prime}$, and an Arizona-$I$ filter[^1]. The observations were conducted with autoguiding and the telescope focussed. Due to the bright nature of the star we used 5s exposure times, which resulted in an observing cadence of $\approx$$30$s.
Systematic effects were minimised by autoguiding, keeping star wander to less than 5 pixels (2.15) over the course of the night. The resulting images were bias-subtracted, flat-fielded, and bad pixel-cleaned in the usual manner. Aperture photometry was performed using an [idl]{}[^2] pipeline utilising the [find]{} and [aper]{} [@stetson1987] tasks available in the NASA Astronomy User’s Library[^3]. The size of the aperture was chosen to minimise scatter in the data and was 10 pixels (4.3) in radius. Several combinations of reference stars were considered, and the one which gave the lowest scatter in the final light curve was adopted. The 1083 original datapoints were binned to yield 216 final datapoints.
Cassini 1.52m telescope {#sec_2.2}
-----------------------
One transit of HAT-P-8 was observed in November 2009, two in October 2011 and one in October 2012, using the 1.52m Cassini Telescope at the Astronomical Observatory of Bologna in Loiano (Italy). We have previously used this telescope several times to observe planetary transits [e.g. @southworth2012a], with the BFOSC (Bologna Faint Object Spectrograph & Camera) instrument operated in imaging mode.
The CCD was used unbinned, giving a plate scale of $0.58^{\prime\prime}/\rm{pixel}$, for a total FOV of $13^{\prime}
\times 12.6^{\prime}$, and the telescope was autoguided and defocussed. The first transit was observed through a Gunn $r$ filter, the two 2011 transits through a Gunn $i$ filter, and the last one through a Johnson $I$ filter. The first 2011 transit suffered from systematic noise due to a bad pixel in the aperture of the target star, so we did not use these data in our analysis. The transit of 2012 was disturbed by clouds, which affected the photometry particularly at the end of the transit. We removed the points compromised by the clouds.
The observations were analysed using the [idl]{} pipeline from @southworth2009a. The images were debiased and flat-fielded using standard methods, then subjected to aperture photometry using the [aper]{} task. Pointing variations were followed by cross-correlating each image against a reference image. We chose the aperture sizes and comparison stars which yielded the lowest scatter in the final differential-photometry light curve. The relative weights of the comparison stars were optimised simultaneously with fitting a second-order polynomial to the outside-transit observations in order to normalise them to unit flux.
Isaac Newton Telescope {#sec_2.3}
----------------------
One transit of HAT-P-8 was monitored using the 2.5m Isaac Newton Telescope (INT), La Palma (Spain), equipped with the Wide Field Camera (WFC) at prime focus. We used only one of the four CCDs, unbinned and with a plate scale of $0.33^{\prime\prime}/\rm{pixel}$, for a total FOV of $12.6^{\prime} \times 11.3^{\prime}$. The telescope was defocussed and autoguided, and the observations were obtained through a Strömgren $y$ filter. A few datapoints at the start of the observing sequence were rejected as they are affected by systematic noise (due probably to high airmass $>$2). The observations were reduced in the same way as those from the Cassini Telescope (Sect.\[sec\_2.2\]).
![Phased BUSCA light curves of HAT-P-8 compared to the best [jktebop]{} fits using the quadratic LD law (left-hand panel). They are ordered according to central wavelength of the filter used (sdss $u$, Gunn $g$, Gunn $r$, Gunn $z$). The residuals of the fits are plotted at the base of the figure, offset from zero.[]{data-label="Fig_02"}](Fig02.eps){width="9.cm"}
Calar Alto 1.23m telescope {#sec_2.4}
--------------------------
Three transits of HAT-P-8 b, two in October 2011 and one in October 2012, were obtained using the 1.23m telescope at the German-Spanish Calar Alto Observatory (CAHA) near Almería, Spain. Autoguiding was used. During the 2011 observations, we used the 2k$\times$2k SITE\#2b otical CCD, which had[^4] a FOV of $16^{\prime} \times
16^{\prime}$ and a pixel scale of $0.5^{\prime\prime}$ per pixel. We defocussed the telescope, read out only a small window in order to limit the dead time between exposures, and observed through a Johnson $R$ filter. The first transit was incomplete due to clouds, so was not included in our analysis. The 2012 transit was obtained through a Johnson $I$ filter with the new DLR-MKIII camera, which is equipped with an e2v CCD231-84-NIMO-BI-DD sensor with 4k$\times$4k pixels and a FOV of $21^{\prime} \times
21^{\prime}$ at $0.3^{\prime\prime}$ per pixel. Unfortunately, the operator did not included appropriate reference stars in the field of view, and the scatter of the resulting light curve is higher than those of 2011. All the observations were reduced as for the Cassini Telescope (Sect.\[sec\_2.2\]).
Telescope Filter BJD (TDB) Diff. mag. Uncertainty
----------- -------- ---------------- ------------ -------------
Loiano $r$ 2455148.210781 0.0009980 0.0018456
Loiano $r$ 2455148.214867 0.0000050 0.0017916
Loiano $i$ 2455840.256418 0.0000785 0.0016988
Loiano $i$ 2455840.258316 0.0006256 0.0016925
CAHA $R$ 2455840.319832 0.0007127 0.0013972
CAHA $R$ 2455840.329288 -0.0019752 0.0013798
INT $y$ 2455437.412356 0.0052976 0.0009779
INT $y$ 2455437.414575 0.0056092 0.0009998
Kuiper $I$ 2455126.66906 -0.99507 $-$
Kuiper $I$ 2455126.66930 -0.99959 $-$
: Excerpts of the light curves of HAT-P-8: this table will be made available at the CDS. A portion is shown here for guidance regarding its form and content.[]{data-label="Table_2"}
Calar Alto 2.2m telescope {#sec_2.5}
-------------------------
We observed one full transit of HAT-P-8 on the night of 2012 August 26, using the 2.2m telescope and BUSCA imager at CAHA. BUSCA is designed for simultaneous four-colour photometry: the light is split into four wavelength bands from UV to visual IR using three dichroics. In the four corresponding focal planes the same area of the sky is imaged onto 4k$\times$4k $15\mu$m pixel CCDs. For our observations we chose to have the SDSS $u$ filter in the bluest arm and standard Calar Alto Gunn $g$, $r$ and $z$ filters in the other three arms. This choice led to a reduced field of view (from $12^{\prime} \times 12^{\prime}$ to a circle of $6^{\prime}$ in diameter), but had the advantage of a much better throughput in $grz$ compared to the default Strömgren filters. We defocussed BUSCA in such a way as to have as much signal as possible in the $u$ band whilst remaining in the linear regime in the other passbands. The CCDs were binned $2 \times 2$ to shorten the readout time. The autoguider was operated in-focus. The observations were reduced in the same way as those from the Cassini Telescope (Sect.\[sec\_2.2\]) and the resulting light curves are plotted in Fig.\[Fig\_02\].
Light curve analysis {#sec_3}
====================
Sect.\[sec\_2\] introduced eleven transit light curves which were suitable for detailed analysis. These were each fitted using the [jktebop]{}[^5] code [@southworth2004], which represents the star and planet as biaxial spheroids. The primary parameters of the fit were the orbital inclination, $i$, the sum and ratio of the fractional radii of the star and planet, $r_{\mathrm{A}}+r_{\mathrm{b}}$ and $k = r_{\mathrm{b}}/r_{\mathrm{A}}$, defined as $r_{\mathrm{A}} =
R_{\mathrm{A}}/a$ and $r_{\mathrm{b}} = R_{\mathrm{b}}/a$, and transit midpoint, $T_0$. The orbital semimajor axis is $a$ and, $R_{\mathrm{A}}$ and $R_{\mathrm{b}}$ are the absolute radii of the two celestial bodies.
Once a fit was available for each dataset, we rescaled the errorbars to give a reduced $\chi^{2}$ of $\chi_{\nu}^{2}=1$. This step is necessary because the [aper]{} aperture photometry procedure has a tendency to underestimate the measurement errors. Then, in order to take systematic noise into account, we inflated the errorbars further using the $\beta$ approach (e.g.@gillon2006 [@winn2008; @winn2009; @gibson2008; @nikolov2012]). We calculated $\beta$ values for between two and ten datapoints for each light cuurve, and adopted the largest $\beta$ value.
Limb darkening (LD) was accounted for using a quadratic law. The linear LD coefficient was fitted, whereas the non-linear one was fixed at a theoretically predicted value [@claret2004], but perturbed by $\pm$0.1 during the process of error estimating. The atmospheric parameters of the star assumed for deriving the limb-darkening coefficients were: $T_{\mathrm{eff}}=6130$, $\log
g=4.15$, \[Fe/H\]=+0.01, $V_{\mathrm{micro}}=2\,$km s$^{-1}$. Uncertainties in the fitted parameters from each solution were calculated in two ways: from 1000 Monte Carlo simulations and with a residual-permutation algorithm (see @southworth08). The larger of the two possible error bars was retained in each case. The light curves and their best-fitting models are shown in Figs.\[Fig\_01\] and\[Fig\_02\], whereas the parameters of each fit are reported in Table\[Table\_3\].
We also attempted to fit all light curves simultaneously using the [tap]{} @gazak code, but were unsuccessful. Analyses of individual light curves, however, yielded similar results and uncertainties as to those from [jktebop]{}.
Source Filter $r_{\mathrm{A}}+r_{\mathrm{b}}$ $k$ $i^{\circ}$ $r_{\mathrm{A}}$ $r_{\mathrm{b}}$
--------------- -------------------- --------------------------------- -------------------------------- --------------------------- ------------------------------ ---------------------------------
Loiano \#1 Gunn $r$ $0.177 \pm 0.013 $ $0.0921 \pm 0.0034 $ $87.3 \pm 2.3$ $0.162 \pm 0.012 $ $0.0149 \pm 0.0014 $
Loiano \#2 Gunn $i$ $0.1695 \pm 0.0030 $ $0.0943 \pm 0.0022 $ $89.2 \pm 1.0$ $0.1460 \pm 0.0026$ $0.01460 \pm 0.00046$
Loiano \#3 Johnson $I$ $0.1674 \pm 0.0035 $ $0.0884 \pm 0.0010 $ $88.5 \pm 1.0$ $0.1538 \pm 0.0032$ $0.01360 \pm 0.00034$
CA 1.23m \#1 Johnson $R$ $0.218 \pm 0.012 $ $0.1006 \pm 0.0021 $ $82.7 \pm 1.1$ $0.198 \pm 0.010 $ $0.0199 \pm 0.0012 $
CA 1.23m \#2 Johnson $R$ $0.1685 \pm 0.0061 $ $0.0888 \pm 0.0032 $ $88.7 \pm 1.6$ $0.1548 \pm 0.0058$ $0.01374 \pm 0.00084 $
CA 1.23m \#3 Johnson $I$ $0.208 \pm 0.018 $ $0.1056 \pm 0.0042 $ $83.3 \pm 1.6$ $0.188 \pm 0.016 $ $0.0199 \pm 0.0018 $
CA 2.2m sdss $u$ $0.199 \pm 0.024 $ $0.0993 \pm 0.0039 $ $84.1 \pm 2.5$ $0.181 \pm 0.021 $ $0.0180 \pm 0.0026 $
CA 2.2m Gunn $g$ $0.1720 \pm 0.0054 $ $0.0926 \pm 0.0024 $ $89.9 \pm 1.4$ $0.157 \pm 0.050 $ $0.01458 \pm 0.00057 $
CA 2.2m Gunn $r$ $0.1692 \pm 0.0051 $ $0.0959 \pm 0.0018 $ $88.7 \pm 1.4$ $0.1544 \pm 0.0046$ $0.01480 \pm 0.00059$
CA 2.2m Gunn $z$ $0.1695 \pm 0.0059 $ $0.0917 \pm 0.0016 $ $88.3 \pm 1.5$ $0.1552 \pm 0.0053$ $0.01424 \pm 0.00058$
INT Strömgren $y$ $0.1699 \pm 0.0052 $ $0.0886 \pm 0.0016 $ $87.9 \pm 1.2$ $0.1560 \pm 0.0047$ $0.01383 \pm 0.00056$
Kuiper Cousins $I$ $0.175 \pm 0.012 $ $0.0871 \pm 0.0026 $ $86.4 \pm 1.6$ $0.161 \pm 0.012 $ $0.0140 \pm 0.0010 $
UDEM Cousins $I$ $0.220 \pm 0.032 $ $0.0836 \pm 0.0039 $ $81.7 \pm 2.6$ $0.203 \pm 0.030 $ $0.0170 \pm 0.0021 $
Adagio Cousins $V$ $0.192 \pm 0.017 $ $0.0956 \pm 0.0034 $ $85.3 \pm 2.1$ $0.175 \pm 0.015 $ $0.0167 \pm 0.0018 $
OHP Cousins $V$ $0.172 \pm 0.031 $ $0.0977 \pm 0.0088 $ $87.4 \pm 3.9$ $0.157 \pm 0.027 $ $0.0153 \pm 0.0031 $
Quarryview Red filter $0.169 \pm 0.011 $ $0.0901 \pm 0.0024 $ $87.5 \pm 2.1$ $0.155 \pm 0.010 $ $0.0140 \pm 0.0011 $
Oversky Sloan $r^{\prime}$ $0.171 \pm 0.016 $ $0.0888 \pm 0.0033 $ $86.9 \pm 2.6$ $0.157 \pm 0.015 $ $0.0140 \pm 0.0016 $
KeplerCam Sloan $z$ $0.1781 \pm 0.0081 $ $0.0955 \pm 0.0013 $ $86.9 \pm 1.3$ $0.1626 \pm 0.0073$ $0.01553 \pm 0.00084$
Final results $\mathbf{0.09208 \pm 0.00049}$ $\mathbf{87.08 \pm 0.36}$ $\mathbf{0.1590 \pm 0.0014}$ $\mathbf{0.01468 \pm 0.00017}$
@latham2009 $0.1725_{-0.00048}^{+0.00094}$ $0.0953 \pm 0.0009$ $87.5_{-0.9}^{+1.9}$ $0.1575_{-0.0042}^{+0.0084}$ $0.01501_{-0.00054}^{+0.00095}$
@simpson2011 $0.1783_{-0.00063}^{+0.00060}$ $0.09135 \pm 0.00089$ $87.80_{-0.77}^{+0.75}$ $0.1634_{-0.0056}^{+0.0053}$ $0.01493_{-0.00066}^{+0.00063}$
Datasets taken from literature {#sec_3.1}
------------------------------
@latham2009 reported three $z$-band transits (1451 points in total), only one of which was observed in its entirety, obtained with the 1.2m telescope and KeplerCam at the F. L.Whipple Observatory, US. We converted the timestamps into orbital phase (using their ephemeris), sorted them and binned them into 154 points.
@moutou2011 presented five datasets, two of them obtained by the Adagio Association with an 82cm telescope ($V$ band; 720 points), one from the 1.20m telescope at Observatoire de Haute-Provence ($V$ band; 256 points), one from a 35cm telescope at the Oversky observatory ($R$ band; 569 points binned in 113 points), and one from a 32cm telescope at the Quarryview Observatory ($R$ band, 157 points). The two Adagio datasets were phased-binned into 144 points before analysis, to save computing time. The Oversky data were also binned, into 113 points.
One transit was observed with the 36cm Universidad de Monterrey Observatory (UDEM) telescope ($I$ band; 639 points). These data were reported by @todorov2012. Finally, one very good light curve was obtained by F. Hormuth using the CAHA 1.23m telescope ($R$ band; 176 points). These data are available from TRESCA web site and were already studied by @simpson2011.
All of the light curves in this Section were fitted using [jktebop]{} in the same manner as our own data.
Final photometric parameters {#sec:finphot}
----------------------------
All the light curves we fitted are shown in Figs.\[Fig\_01\] and \[Fig\_02\]. The parameters of the fits are given in Table\[Table\_3\]. In order to calculate the final photometric parameters we took the weighted mean of the individual values and uncertainties. This process highlighted the poor agreement between the different datasets. The $\chi_{\nu}^2$ of the parameters of the individual light curves with respect to the final weighted means are $1.5$ for $r_{\rm A}$, $2.4$ for $i$ and $2.7$ for $r_{\rm b}$. The worst agreement is found for $k$, where $\chi_{\nu}^2=6.9$. Analagous situations have been found many times in the course of the [*Homogeneous Studies*]{} project [@southworth08; @southworth09; @southworth10; @southworth11; @southworth12], but the agreement is rarely this poor.
Inspection of Fig.\[Fig\_01\] gives a clue to this problem. The residuals of the fits to almost all of the available light curves, including our own, exhibit systematic deviations from zero. This correlated noise means the amount of information available in the observational data is lower than suggested by the number of datapoints and their uncertainties. We have already increased our errorbars due to correlated noise using the $\beta$ approach (Sect.\[sec\_3\]), but this was not enough to solve the problem. Fortunately, we have 18 separate light curves which have been analysed independently. Making the assumption that correlated noise in data taken on different nights using different telescopes is independent, we can account for it by inflating the errorbars on the final photometric parameters by $\sqrt{\chi_{\nu}^2}$ to give $\chi_{\nu}^2 = 1.0$ for each parameter. Because our results are based on 18 different datasets, we are confident that our analysis has yielded reliable parameters. The light curve anomalies are discussed further in Sect.\[sec\_5\].
Orbital period determination {#sec_3.2}
----------------------------
---------------------------- ------- ---------- ---------------------------------
Time of minimum Cycle Residual Reference
BJD(TDB)$-2400000$ no. (JD)
$54437.67461 \pm 0.00044 $ 0 0.00046 @latham2009
$55034.4824 \pm 0.0031 $ 194 -0.00284 Srdoc (AXA)
$55034.4881 \pm 0.0017 $ 194 0.00286 Kučáková H. (TRESCA)
$55046.7904 \pm 0.0008 $ 198 -0.00022 Norby (AXA)
$55071.400 \pm 0.002 $ 206 -0.00138 Ayiomamitis (AXA)
$55071.4017 \pm 0.0009 $ 206 -0.00032 Srdoc (AXA)
$55071.4056 \pm 0.0012 $ 206 0.00422 Kučáková H. (TRESCA)
$55071.4132 \pm 0.0013 $ 206 0.01182 Trnka J. (TRESCA)
$55074.4752 \pm 0.0013 $ 207 -0.00253 Srdoc (AXA)
$55123.7021 \pm 0.0018 $ 223 0.00284 Tieman B. (TRESCA)
$55126.77446 \pm 0.00079 $ 224 -0.00115 This work (Kuiper 155cm)
$55148.3140 \pm 0.0019 $ 231 0.00397 This work (Loiano 152cm)
$55409.7992 \pm 0.0012 $ 316 -0.00022 [@todorov2012] (Udem 36cm)
$55434.4108 \pm 0.0023 $ 324 0.00061 [@moutou2011] (OHP 120cm)
$55437.48610 \pm 0.00043 $ 325 -0.00043 This work (INT 250cm)
$55437.4834 \pm 0.0011 $ 325 -0.00313 [@moutou2011] (Oversky 35cm)
$55437.48434 \pm 0.00088 $ 325 -0.00219 Hormuth F. (CA 123cm)
$55437.4895 \pm 0.0010 $ 325 0.00297 Ruocco N. (TRESCA)
$55440.5646 \pm 0.0012 $ 326 0.00171 Marino G. (TRESCA)
$55449.79074 \pm 0.00099 $ 329 -0.00118 [@moutou2011] (Hose 32cm)
$55452.86716 \pm 0.00079 $ 330 -0.00110 Hose K. (TRESCA)
$55797.4152 \pm 0.0018 $ 442 -0.00379 Dřevěný R., Kuchtak B. (TRESCA)
$55797.4186 \pm 0.0015 $ 442 -0.00039 Brát L. (TRESCA)
$55800.4916 \pm 0.0013 $ 443 -0.00374 Brát L. (TRESCA)
$55800.4942 \pm 0.0011 $ 443 -0.00114 Trnka J. (TRESCA)
$55800.4968 \pm 0.0025 $ 443 0.00146 Zibar M. (TRESCA)
$55800.4996 \pm 0.0010 $ 443 0.00426 Marek P. (TRESCA)
$55812.80021 \pm 0.00085 $ 447 -0.00051 Shadic S. (TRESCA)
$55834.3372 \pm 0.0019 $ 454 0.00206 Brát L. (TRESCA)
$55837.4136 \pm 0.0025 $ 455 0.00211 Trnka J. (TRESCA)
$55840.48811 \pm 0.00063 $ 456 0.00028 This work (CA 123cm)
$55840.48845 \pm 0.00049 $ 456 0.00062 This work (Loiano 152cm)
$55886.6345 \pm 0.0021 $ 471 0.00148 Dvorak S. (TRESCA)
$56166.5776 \pm 0.0011 $ 562 -0.00289 This work (CA 220cm BUSCA-$u$)
$56166.58178 \pm 0.00057 $ 562 0.00129 This work (CA 220cm BUSCA-$g$)
$56166.58162 \pm 0.00051 $ 562 0.00113 This work (CA 220cm BUSCA-$r$)
$56166.57924 \pm 0.00054 $ 562 -0.00124 This work (CA 220cm BUSCA-$z$)
$56175.81123 \pm 0.00059 $ 565 0.00170 Hose K. (TRESCA)
$56203.49662 \pm 0.00031 $ 574 -0.00002 This work (Loiano 152cm)
$56203.49682 \pm 0.00083 $ 574 0.00018 This work (CA 123cm)
---------------------------- ------- ---------- ---------------------------------
The transit time for each dataset was obtained using [jktebop]{}, and uncertainties were estimated by Monte Carlo simulations. In the determination of the orbital period of the HAT-P-8 system, we also considered 23 timings measured by amateur astronomers and available on the ETD[^6] website (see Table\[Table\_4\]).
We excluded from the analysis the incomplete ETD light curves and those with a Data Quality index higher than 3. All timings were placed on BJD(TDB) time system. The resulting measurements of transit midpoints were fitted with a straight line to obtain a new orbital ephemeris: [$$T_{0} = \mathrm{BJD(TDB)} 2\,454\,437.6742(14) + 3.0763458(24)\,E,$$ ]{} where $E$ is the number of orbital cycles after the reference epoch (which we take to be the midpoint of the first transit observed by @latham2009) and quantities in brackets denote the uncertainty in the final digit of the preceding number. The fit has $\chi_{\nu}^2=5.30$, and the uncertainties given above have been increased to account for this. The large $\chi_{\nu}^2$ indicates that the uncertainties in the various $T_0$ measurements are too small, most probably due to the systematic differences between the light curves and their best fits as noted in Sect.\[sec:finphot\]. A plot of the residuals around the fit is shown in Fig.\[Fig\_03\] and does not indicate any clear systematic deviation from the predicted transit times. We therefore take the conservative option of not interpreting the large $\chi_{\nu}^2$ as a suggestion of transit timing variations.
{width="16.cm"}
The physical properties of HAT-P-8 {#sec_4}
==================================
Following @southworth09, we estimated the physical properties of the HAT-P-8 system from the photometric parameters measured in Sect.\[sec\_3\], published spectroscopic results, and theoretical stellar models. The orbital eccentricity, velocity amplitude and metallicity of the star, ($e=0$, $K_{\mathrm{A}}=153.1 \pm 3.9$, \[Fe/H\]$=+0.01 \pm 0.08$) were taken from @latham2009, and for the parent-star effective temperature ($T_{\mathrm{eff}}=6130 \pm 80$) we used that measured by @knutson2010.
An initial estimate of the velocity amplitude of the planet ($K_{\rm b}$) was iteratively refined by calculating the system properties using standard formulae (e.g. [@hilditch2001]), and comparing the observed $T_{\mathrm{eff}}$ and $r_{\mathrm{A}}$ with values of $T_{\mathrm{eff}}$ and $R_{\mathrm{A}}/a$ predicted by theoretical models for the calculated mass of the star. This calculation was performed over a grid of ages and for five different sets of stellar models (see @southworth10). Statistical errors were propagated by a perturbation analysis. The resulting estimates of the physical properties are given in Table\[Table\_5\]. The final set of physical properties was obtained by taking the unweighted mean of the five sets of values found from the different stellar models, which also allowed us to obtain an estimate of the systematic errors inherent in the use of stellar theory. The results of this process are given in Table\[Table\_6\].
Finally, following the method delineated by @enoch2010 and improved by @southworth11, we used empirical measurements of stars in detached eclipsing binary (dEB) systems to calibrate the parent star of the HAT-P-8 system. This allowed us to measure the physical properties of the system without using stellar models, thus avoiding the dependence on stellar theory. These results are also given in Table\[Table\_5\].
--------------------------------------------- ----------------------- ----------------------- ----------------------- ----------------------- ----------------------- -----------------------
This work This work This work This work This work This work
(dEB constraint) ([Claret]{} models) ([Y$^2$]{} models) ([Teramo]{} models) ([VRSS]{} models) ([DSEP]{} models)
$K_{\mathrm{b}}$ (kms$^{-1}$) $159.1 \pm 3.9$ $154.8 \pm 2.3$ $156.7 \pm 1.9$ $153.4 \pm 1.7$ $154.2 \pm 2.6$ $155.4 \pm 2.13$
$M_{\mathrm{A}}$ ($M_{\sun}$) $1.292 \pm 0.094$ $1.189 \pm 0.054$ $1.234 \pm 0.044$ $1.158 \pm 0.037$ $1.175 \pm 0.061$ $1.202 \pm 0.053$
$R_{\mathrm{A}}$ ($R_{\sun}$) $1.521 \pm 0.041$ $1.480 \pm 0.032$ $1.470 \pm 0.026$ $1.467 \pm 0.023$ $1.474 \pm 0.031$ $1.485 \pm 0.026$
$\log g_{\mathrm{A}}$ (cgs) $4.1852 \pm 0.0137$ $4.1733 \pm 0.0088$ $4.1954 \pm 0.0103$ $4.1694 \pm 0.0105$ $4.1715 \pm 0.0106$ $4.1748 \pm 0.0115$
$M_{\mathrm{b}}$ ($M_{\mathrm{jup}}$) $1.345 \pm 0.071$ $1.273 \pm 0.048$ $1.305 \pm 0.041$ $1.251 \pm 0.037$ $1.263 \pm 0.053$ $1.282 \pm 0.047$
$R_{\mathrm{b}}$ ($R_{\mathrm{jup}}$) $1.357 \pm 0.043$ $1.320 \pm 0.034$ $1.337 \pm 0.032$ $1.308 \pm 0.030$ $1.315 \pm 0.037$ $1.325 \pm 0.034$
$\rho_{\mathrm{b}}$ ($\rho_{\mathrm{jup}}$) $0.503 \pm 0.034$ $0.517 \pm 0.034$ $0.511 \pm 0.033$ $0.522 \pm 0.034$ $0.519 \pm 0.035$ $0.516 \pm 0.034$
$\Theta$ $0.0691 \pm 0.0026$ $0.0711 \pm 0.0023$ $0.0702 \pm 0.0022$ $0.0717 \pm 0.0022$ $0.0714 \pm 0.0025$ $0.0708 \pm 0.0023$
$a$ (AU) $0.04510 \pm 0.00109$ $0.04387 \pm 0.00066$ $0.04442 \pm 0.00053$ $0.04348 \pm 0.00047$ $0.04370 \pm 0.00075$ $0.04403 \pm 0.00065$
Age $-$ $4.7_{-0.5}^{+1.4}$ $3.5_{-0.7}^{+0.4}$ $4.7_{-0.5}^{+1.0}$ $4.2_{-0.4}^{+1.5}$ $4.2_{-1.4}^{+0.5}$
--------------------------------------------- ----------------------- ----------------------- ----------------------- ----------------------- ----------------------- -----------------------
This work (final) @latham2009 @moutou2011 Latham (private comm.)
--------------------------------------------- ----------------------------------- ------------------------ --------------------- ------------------------------
$M_{\mathrm{A}}$ ($M_{\sun}$) $1.192 \pm 0.061 \pm 0.043$ $1.28 \pm 0.04$ $-$ $1.28_{-0.06}^{+0.04}$
$R_{\mathrm{A}}$ ($R_{\sun}$) $1.475 \pm 0.032 \pm 0.010$ $1.58^{+0.08}_{-0.06}$ $-$ $1.57 \pm 0.07$
$\log g_{\mathrm{A}}$ (cgs) $4.177 \pm 0.011 \pm 0.019$ $4.15 \pm 0.03$ $-$ $4.15 \pm 0.03$
$\rho_{\mathrm{A}}$ ($\rho_{\sun}$) $0.371 \pm 0.013 \pm 0.018$ $-$ $-$ $0.46 \pm 0.05$
$M_{\mathrm{b}}$ ($M_{\mathrm{jup}}$) $1.275 \pm 0.053 \pm 0.030$ $1.52^{+0.18}_{-0.16}$ $1.34 \pm 0.05$ $1.38 \pm 0.05$
$R_{\mathrm{b}}$ ($R_{\mathrm{jup}}$) $1.321 \pm 0.037 \pm 0.016$ $1.50^{+0.08}_{-0.06}$ $-$ $1.40 \pm 0.08$
$g_{\mathrm{b}}$ ($\mathrm{ms^{-1}}$) $18.11 \pm 0.82$ $16.98 \pm 1.17$ $-$ $1.38 \pm 0.05$
$\rho_{\mathrm{b}}$ ($\rho_{\mathrm{jup}}$) $0.517 \pm 0.035 \pm 0.006$ $0.568 \pm 0048$ $-$ $0.62 \pm 0.09$
$T_{\mathrm{eq}}$ ($\mathrm{K}$) $1713 \pm 24 \pm 13 $ $1700 \pm 35$ $-$ $1771 \pm 39 $
$\Theta$ $0.0710 \pm 0.0025 \pm 0.0008$ $0.061 \pm 0.003$ $-$ $0.069 \pm 0.004$
$a$ (AU) $0.04390 \pm 0.00075 \pm 0.00052$ $0.0487 \pm 0.0026$ $0.0449 \pm 0.0007$ $0.0450_{-0.0007}^{+0.0004}$
Age (Gyr) $4.3_{-1.4\,-0.5}^{+1.5\,+0.8}$ $3.4 \pm 1.0$ $-$ $3.3_{-0.3}^{+0.7}$
Table\[Table\_6\] also contains a comparison between our own results and those found by @latham2009. We find smaller masses and radii for both the planet and the host star. This is surprising because the values of the fractional radius of the star, which furnishes the vital constraint on the stellar density, are very similar between the two studies (we find $r_{\rm A} =
0.1590 \pm 0.0014$ versus $r_{\rm A} = 0.1575^{+0.0041}_{-0.0089}$ for @latham2009). One small difference is that we adopted $T_{\mathrm{eff}} = 6130 \pm 80$K [@knutson2010] as opposed to the value of $6200 \pm80$K used by @latham2009. Our revised physical properties move HAT-P-8b into a region of parameter space which is comparatively well-supplied with transiting planets. Fig.\[Fig\_04\] shows the change in position in the planet mass–radius plot.
![Plot of the masses and radii of the known TEPs. The orange symbols denote values from the [*Homogeneous Studies*]{} project and the blue symbols results for the other known TEPs. HAT-P-8b is shown in red [@latham2009] and green (this work). Dotted lines show where density is (from bottom to top) 1.0, 0.5, 0.25 and 0.1 $\rho_{\mathrm{Jup}}$.[]{data-label="Fig_04"}](Fig04.eps){width="9.cm"}
We contacted D. Latham in order to check the veracity of the final values of the parameters that they reported in their discovery paper. D. Latham confirmed that some of these results are indeed correct, and has kindly supplied revised values. These are from a re-analysis carried out by J. Hartman on a slight updated dataset, and are given in the final column of Table\[Table\_6\]. They correspond to a smaller planetary mass and radius ($M_{\mathrm{b}}=1.38\,M_{\mathrm{jup}}$, $R_{\mathrm{b}}=1.40\,R_{\mathrm{jup}}$) than given in @latham2009, in agreement with our own findings.
Variation of planetary radius with wavelength {#sec_5}
=============================================
One of the factors that plays a principal role in determining the atmospheric properties of hot-Jupiter planets is the amount of stellar flux incident on the planet’s surface. Variations in this irradiation cause different planets to have different atmospheric chemical mixing ratios and atmospheric opacities. This could potentially lead to divide into classes of hot Jupiters. An initial suggestion was that of @fortney2008 to distinguished pM- and pL-class planets, depending on the presence of strong absorbers such as gaseous titanium oxide (TiO) and vanadium oxide (VO) in their atmospheres. By observing a planetary transit at different wavelengths, it is possible to detect a variation in the value of the radius measured as a function of the wavelength, and probe chemistry and wavelength dependent opacity at the planet’s terminator. This dependence on the wavelength can be therefore used to probe the atmospheric composition of TEPs, being a complementary method to the observations of secondary eclipses.
As an additional possibility offered by the BUSCA data, we made an attempt to investigate possible variations of the radius of HAT-P-8b in different optical passbands. Receiving from its parent star an incident flux of $2.22\pm0.20\times10^9$ ergs$^{-1}$cm$^{-2}$ (Latham, private communication), HAT-P-8b should belong to the pM class of planets. The theoretical models of @fortney2010 predict that its radius should be few percent lower at $350$–$400$nm and $800$–$950$nm versus $500$–$750$nm. Following the strategy used by @southworth2012b, we fitted the BUSCA light curves with all parameters fixed to the final values reported in Table\[Table\_4\], with the exception of $k$ and the LD coefficients. The errors were estimated by a residual-permutation algorithm. The results are exhibited in Fig.\[Fig\_05\], where the points show the data, the vertical bars represent the relative errors in the measurements and the horizontal bars show the full widths at half maximum transmission of the passbands used. As is apparent from Fig.\[Fig\_02\], the BUSCA-$z$ light curve shows some structure in its residuals from phase 0 to 0.02, which suggests that the $z$-band data suffer correlated noise. The result coming from this band should be therefore considered with caution. The optical region not covered by the BUSCA data was investigated by using the $i$-band light curve from Loiano, which is the best one that we have at this band (see Fig.\[Fig\_01\]).
A variation of $r_{\mathrm{b}}$ along the five passbands is clearly visible in Fig.\[Fig\_05\]. We investigated this variation with the help of model atmospheres. We used detailed non-gray atmosphere codes to model the temperature structure and transmission spectrum of the planet. We computed 1D model atmosphere of HAT-P-8b, using the atmosphere code described in @fortney2005 [@fortney2008]. Pressure-temperature profiles are derived that either include or exclude the opacity of TiO and VO molecules. The fully non-gray model uses the chemical equilibrium abundances of @Lodders02 and the opacity database described in @fr08. The atmospheric pressure-temperature profiles simulate planet-wide average conditions or day-side average conditions. We computed the transmission spectrum of the models using the methods described in @fortney2010 and @shabram2011.
Excluding TiO and VO from the opacity calculation, in the upper panel of Fig.5 we compare the transmissions spectrum of the 1D planet-wide average profile (in blue) with a warmer day-side average model (in green), to experimental data. In comparison, the bottom panel of Fig. 5 shows analogous calculations in red and yellow respectively, which include both the TiO and VO species. Coloured boxes indicate the predicted values for these models integrated over the bandpasses of the observations. Models without TiO and VO have optical transmission spectra that are dominated by Rayleigh scattering in the blue, and pressure-broadened neutral atomic lines of Na an K at 589 nm and 770 nm, respectively. Models with TiO and VO absorption show significant optical absorption (a much large transit radius), that broadly peaks around 700 nm, with a sharp falloff in the blue, and a shallower falloff in the red.
Comparing the panels of Fig.\[Fig\_05\], it is readily apparent that the model that gives the best match to the data is the one with TiO and VO opacity in the atmosphere of HAT-P-8b. The increased optical radius is somewhat larger than the model prediction. Taken at face value, the observations are in general agreement with the @fortney2008’s hot-Jupiter classification based on stellar irradiation. We suggest that HAT-P-8b should be an important target for followup studies to confirm or refute these suggestive observations. The clear detection of TiO/VO, or other strong optical absorbers, would be an important step in characterizing hot Jupiter atmospheres, as such absorbers are thought to cause temperature inversions in these atmospheres (e.g. [@fortney2008; @burrows2008]). Previously, @desert2008 suggested a detection of TiO and VO at strongly depleted levels in the red optical spectrum of HD 209845b, using Hubble STIS.
![Variation of the fractional planetary radius $r_{\mathrm{b}}=R_{\mathrm{b}}/a$ with wavelength. The points shown in the plot are from the Calar Alto 2.2m ($u$, $g$, $r$ and $z$ bands) and Loiano ($i$-band) telescopes. The vertical bars represent the errors in the measurements and the horizontal bars show the full widths at half maximum transmission of the passbands used. The experimental points are compared with four models. These use planet-wide average pressure-temperature profiles (in blue and red) or warmer day-side average profiles (in green and yellow). Synthetic spectra in the top panel do not include TiO and VO opacity, while spectra in the bottom do, based on equilibrium chemistry. Coloured squares represent band-averaged model radii over the bandpasses of the observations.[]{data-label="Fig_05"}](Fig05.eps){width="9.cm"}
Light-curve anomalies {#sec_6}
=====================
![An example of the consistency of follow-up observations made of the same transit on the night of 2010/08/28 by different telescopes. This is an example of a “regular” transit. Lighter yellow dots are for the data taken at the INT (this work), whereas the darker brown ones for those taken by F. Hormuth (ETD) at the CAHA 1.23m. The agreement between the two datasets is not very good. The difference in depth between the two datasets is partly due to different filters used, Strömgren $y$ and Johnson $R$, respectively. The vertical line represents the expected transit minimum time.[]{data-label="Fig_06"}](Fig06.eps){width="9.cm"}
While one of our follow-up light curves, obtained at the INT, displays a regular transit shape, the others all show anomalies of a similar shape. They display an asymmetry with respect the line of minimum transit time. It is important to clarify if the anomalies that we detected have an astrophysical nature or are of instrumental or environmental origin. One way to check this is to have independent measures of the same transit event, obtained using multiple telescopes located at different sites. In this way, if both the telescopes reported the same anomaly, it less likely that they are caused by instrumental or Earth-atmosphere effects. This already happened inadvertently in the case of several follow-up observations of WASP-33 carried out by amateur astronomers and reported in @kovacs2012. It also occurred by chance for the transit observations of HAT-P-8 on the night of 2010/08/28, which was observed by ourselves at the INT and contemporaneously at CAHA by F. Hormuth. Fig.\[Fig\_06\] shows the two light curves in the same plot; we note that both have a regular transit shape but disagree over the transit depth. Some fraction of this disagreement is due to the different LD characteristics, as the INT data were obtained with a bluer filter than the CAHA data.
In 2011, we aimed to observe transits in the HAT-P-8 system from two different observatories and this goal was successfully achieved on the night of 2011/10/05 when a transit was simultaneously observed using the 1.52m Cassini and CAHA 1.23m telescopes. The datasets show partial but not complete agreement about the transit shape anomalies, as well as slight differences due to the different LD in the $i$ and $R$ passbands. Details of the two light curves are displayed in the lower panel of Fig.\[Fig\_07\], while the upper panel shows another transit (2009, Loiano) which presents a similar anomaly. The anomalies cannot be removed by choosing different comparison stars for the differential photometry, but are not completely consistent between different datasets for the same transit. We have therefore treated them as correlated noise in our analysis (see Sect.\[sec\_3\]).
![Upper panel: an example of a discrepant transit. The green points are for the data taken at the Cassini telescope (this work) through a Gunn $r$ filter. Lower panel: an example of the consistency of followup observations made on the same date of 2011/10/05 by different observers. Lighter blue dots are for the data taken at the Cassini telescope (this work), whereas the red ones for those taken at CAHA 1.23m (this work). Some of the difference in depth between the two datasets is due to different filters used, Gunn $i$ and Johnson $R$, respectively. The vertical line represents the expected transit minimum time.[]{data-label="Fig_07"}](Fig07.eps){width="9.cm"}
If the recurrent anomalies that we noticed in the HAT-P-8 light curves are not of instrumental origin, is there a reasonable astrophysical explanation for these features? We can easily exclude a gravity darkening effect (e.g. @barnes2009) because the rotation rate of the star is low ([@latham2009; @moutou2011]). The presence of a moon is an unlikely possibility, and in any case is difficult to constrain with so few observations.
Another possibility is that the anomalies are due to single or a belt of stellar spots on the photosphere of the parent star. @knutson2010 estimated that HAT-P-8 has an activity index of $\log(R_{HK}^{\prime})= -4.985$, which indicates that the star has a moderate chromospheric activity. This is related to the star’s magnetic structure and therefore to the presence of photospheric features, such as spots, that modulate luminosity. It is thus possible that, during the transit, dark starspots are occulted by the planet. Similar features were indeed already observed in several cases, e.g. HD189733 [@pont2007], TrES-1 [@rabus2009], Kepler-17 [@desert2011], HAT-P-11 [@sanchis2011], and WASP-19 [@tregloan2012]. One argument against this is the relatively high $T_{\rm eff}$ of the star, $6130 \pm 80$K, which makes starspots less likely to occur. Moreover, spots would only affect small parts of the light curve, while here both light curves show trends for the *full* duration of the transit.
Another possible explanation is that the faint M-dwarf companion of HAT-P-8 is a *flare star*, which could emit giant flares bright enough to significantly affect the light curves. However, the magnitude difference between the two objects ($\Delta
i^{\prime}=7.34 \pm 0.10$, [@bergfors2012]), is such that this star would have to generate a superflare in order to make a noticeable difference. One more explanation for this behaviour is that a significant fraction of the comparison stars display intrinsic variability which is insufficient to detect in individual stars but, when combined, is enough to modify the transit shape in the HAT-P-8 system.
Having exhausted all possibilities, we conclude that the apparent transit shape *distortion* shown in Fig.\[Fig\_07\] could be caused by differential color extinction, pathological variability in the comparison stars, or other low-frequency noise of atmospheric or astrophysical origin. Further investigations of these hypotheses would, inevitably, require substantial further data and is beyond the scope of the current work.
Summary and Conclusions {#sec_7}
=======================
We have reported observations of six transits of the HAT-P-8 system performed using five different medium-class telescopes, for a total of eleven new light curves. All but one of these transits were obtained using the defocussed-photometry technique, achieving a photometric precision of $0.47$–$1.7$mmag per observation. Four of them were observed on the same nights from two different telescopes. In one of these two nights, both the light curves show an anomaly which is probably caused by systematic noise of atmospheric or astrophysical origin. Another transit was simultaneously observed through four optical passbands in a wavelength range between 330 and 960 nm.
We modelled our new and seven published datasets using the [jktebop]{} code, and used the results to determine the physical properties of the planet and its host star. Compared to the discovery paper [@latham2009], we find a significantly smaller radius and mass for HAT-P-8b. We obtain $R_{\rm b} =
1.321 \pm 0.037 \, R_{\mathrm{Jup}}$ versus $1.50_{-0.06}^{+0.08}
\, R_{\mathrm{Jup}}$, and $M_{\rm b} = 1.275 \pm 0.053 \,
M_{\mathrm{Jup}}$ versus $1.52_{-0.16}^{+0.18} \,
M_{\mathrm{Jup}}$. The theoretical radius calculated by @fortney2007 for a core-free planet at age 4.5Gyr and distance 0.045AU is $1.107$–$1.108 \, R_{\mathrm{Jup}}$ for a planet of mass in the range $1.0$–$1.46 \, M_{\mathrm{Jup}}$, which is still significantly smaller than the radius we find.
@latham2009 found that HAT-P-8b was a comparatively highly inflated planet. Instead, our results place it firmly in a well-populated part of the mass–radius diagram, removing its outlier status. Whilst the existing transit light curves of the HAT-P-8 system all show some systematic deviations from the best fits found using simple geometric models, the large number of available datasets means that our overall results are reliable. HAT-P-8 is another system where extensive follow-up photometry has been necessary to determine robust physical properties for a planetary system.
Finally, thanks to the ability of BUSCA to measure stellar flux simultaneously through different filters, covering a quite large range of optical window, we probed the composition of the atmosphere of HAT-P-8b by investigated how vary its radius in these wavelengths. In fact, the presence of strong optical absorbers in the atmosphere of the planet should produce larger transit radius at optical wavelengths than in the near UV or near infrared. We measured a variation of the radius of HAT-P-8b along five passbands, corresponding to a wavelength coverage of $330-960$ nm. This result was then theoretically investigated by using several synthetic spectra based on isothermal model atmospheres in chemical equilibrium. The comparison between the models and the experimental points suggests the presence of molecular gas that strongly absorbs in the optical, potentially composed of TiO and VO gases, in the atmosphere of HAT-P-8b.
Based on observations collected at the Centro Astronómico Hispano Alemán (CAHA) at Calar Alto, Spain, operated jointly by the Max-Planck Institut für Astronomie and the Instituto de Astrofísica de Andalucía (CSIC); observations obtained with the 1.52m Cassini telescope at OAB Loiano Observatory, Italy; data collected at the Isaac Newton Telescope, operated on the island of La Palma, by the Isaac Newton Group in the Spanish Observatorio del Roque de los Muchachos. The reduced light curves presented in this work will be made available at the CDS (http://cdsweb.u-strasbg.fr/). We thank Kamen Todorov and Rodrigo F. Díaz for supplying photometric data. JS acknowledges financial support from STFC in the form of an Advanced Fellowship. We thank Ulli Thiele and Roberto Gualandi for their technical assistance at the CA 2.2m telescope and Cassini telescope, respectively. We thank the anonymous referee for their useful criticisms and suggestions that helped us to improve the quality of the present paper. The following internet-based resources were used in research for this paper: the ESO Digitized Sky Survey; the NASA Astrophysics Data System; the SIMBAD data base operated at CDS, Strasbourg, France; and the arXiv scientific paper preprint service operated by Cornell University. LM thanks the HESS site in Namibia for the kind hospitality.
Ballester, G. E., Sing, D. K., & Herbert, F. 2007, , 445, 511 Barnes, J. W. 2009, , 705, 683 Bergfors, C., Brandner, W., Daemgen, S., et al. 2012, submitted Burrows A., Budaj, J., Hubeny, I. 2008, ApJ, 678, 1436 Claret A., 2004, , 428, 1001 Collier Cameron, A., Guenther, E., Smalley, B., Mcdonald, I. 2010, , 407, 507 Désert, J.-M., Vidal-Madjar, A., Lecavelier Des Etangs, A., et al. 2008, A&A, 492, 585 Désert, J.-M., Charbonneau, D., Demory, B.-O., et al. 2011, , 197, 14 Gazak, J. Z., Johnson, J. A., Tonry, J., et al. 2012, Advances in Astronomy, 2012, 697967 Enoch B., Collier Cameron A., Parley N. R., Hebb L., 2010, , 516, A33 Fossati, L., Haswell, C. A., Froning, C. S., et al. 2010, , 717, L222 Fortney, J. J., Marley, M. S., Lodders, K., Saumon, D., & Freedman, R. 2005, , 627, L69 Fortney, J. J., Marley, M. S., & Barneset, J. W. 2007, , 659, 1661 Fortney, J. J., Lodders, K., Marley, M. S., & Freedman, R. S. 2008, , 678, 1419 Fortney, J. J., Shabram, M., Showman, A. P., et al. 2010, , 709, 1396 Freedman R. S., Marley M. S., Lodders K. 2008, ApJ, 174, 513 Gaudi, B. S. & Winn, J. N. 2007, , 655, 550 Gibson, N. P., Pollacco, D., Simpson, E. K., et al. 2008, , 492, 603 Gillon, M., Pont, F., Moutou, C., et al. 2006, , 459, 249 Hilditch, R. W. 2001, An Introduction to Close Binary Stars, Cambridge University Press, Cambridge, UK Jones B. W., Sleep P. N. 2010, , 407, 1259 Kipping, D. M., Fossey, S. J., Campanella, G. 2009, , 400, 398 Knutson, H. A., Howard, A. W., Isaacson, H. 2010, , 720, 1569 Kovács, G., Kovács, T., Hartman, J. D., Bakos, G. Á., Bieryla, A., Latham, D., Noyes, R. W., Regály, Zs., Esquerdo, G. A. 2012, submitted Latham, D. W., Bakos, G. Á., Torres, G., et al. 2009, , 704, 1107 Leconte, J., Lai, D., Chabrier, G. 2011, , 528, A41 Li, S.-L., Miller, N. Lin, D. N. C., Fortney, J. J. 2010, , 463, 1054 Lodders, K., & Fegley, B. 2002, Icarus, 155, 393 Moutou, C., Díaz, R. F., Udry, S., et al. 2011, , 533, A113 Nikolov, N., Henning, Th., Koppenhoefer, J., et al. 2012, , 539, 159 Pont, F., Gilliland, R. L., Moutou, C., er al. 2007, , 476, 1347 Queloz, D., Eggenberger, A., Mayor, M., Perrier, C., Beuzit, J.-L., Naef, D., Sivan, J.-P., Udry, S. 2000, , 359, L13 Rabus, M., Alonso, R., Belmonte, J. A, et al. 2009, , 494, 391 Sanchis-Ojeda, R., Winn, J. N. 2011, , 743, 61 Sanchis-Ojeda, R., Winn, J. N., Holman, M. J., et al. 2011, , 733, 127 Sanchis-Ojeda, R., Fabrycky, D. C., Winn, J. N., et al. 2012, , 487, 449 Shabram, M., Fortney, J. J., Greene, T. P., Freedman, R. S. 2011, , 727, 65 Sing, D. K., Désert, J.-M., Lecavelier des Etangs, A., et al. 2009, , 505, 891 Simpson, E. K., Pollacco, D., Collier Cameron, A., et al. 2011, , 414, 3023 Southworth, J. 2008, , 386, 1644 Southworth, J. 2009, , 394, 272 Southworth, J. 2010, , 408, 1689 Southworth, J. 2011, , 417, 2166 Southworth, J. 2012, , 426, 1291 Southworth, J., Maxted, P. F. L., Smalley, B., 2004, , 351, 1277 Southworth, J., Hinse, T. C., J[ø]{}rgensen, U. G., et al., 2009a, , 396, 1023 Southworth, J., Bruni, I., Mancini, L., & J. Gregorio 2012a, , 420, 2580 Southworth, J., Mancini, L., Maxted, P. F. L., et al. 2012b, , 422, 3099 Stetson, P. B., 1987, , 99, 191 Swain, M. R., Vasisht, G., & Tinetti, G. 2008, , 452, 329 Szabó, Gy. M., Szabó, R., Benkõ, J. M. et al. 2011, , 736, L4 Todorov, K. O., Deming, D., Knutson, H. A., et al. 2012, , 746, 111 Torres G., Winn J. N., Holman M. W., 2008, , 677, 1324 Tregloan-Reed, J., Southworth, J., Tappert, C. 2012, arXiv:1211.0864 Tusnski, L. R. M., & Valio, A. 2011, , 743, 97 Winn, J. N., Holman, M. J., Torres, G., et al. 2008, , 683, 1076 Winn, J. N., Holman, M. J., Carter, J. A., et al. 2009, , 137, 3826
[^1]: The transmission curve for this filter is shown at: james.as.arizona.edu/[\~[psmith]{}]{}/61inch/FILTERS/harris.jpg
[^2]: The acronym [idl]{} stands for Interactive Data Language and is a trademark of ITT Visual Information Solutions. For further details see [http://www.ittvis.com/ProductServices/IDL.aspx]{}.
[^3]: The IDL Astronomy User’s Library (ASTROLIB) is available at http://idlastro.gsfc.nasa.gov/
[^4]: This CCD has been decommissioned.
[^5]: <span style="font-variant:small-caps;">jktebop</span> is written in FORTRAN77 and the source code is available at http://www.astro.keele.ac.uk/[\~[jkt]{}]{}/
[^6]: The Exoplanet Transit Database (ETD) website can be found at http://var2.astro.cz/ETD
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop an analytic model for the power spectra of polarized filamentary structures as a way to study the Galactic polarization foreground to the Cosmic Microwave Background. Our approach is akin to the cosmological halo-model framework, and reproduces the main features of the Planck 353 GHz power spectra. We model the foreground as randomly-oriented, three-dimensional, spheroidal filaments, accounting for their projection onto the sky. The main tunable parameters are the distribution of filament sizes, the filament physical aspect ratio, and the dispersion of the filament axis around the local magnetic field direction. The abundance and properties of filaments as a function of size determine the slopes of the foreground power spectra, as we show via scaling arguments. The filament aspect ratio determines the ratio of $B$-mode power to $E$-mode power, and specifically reproduces the Planck-observed dust ratio of one-half when the short axis is roughly one-fourth the length of the long axis. Filament misalignment to the local magnetic field determines the $TE$ cross-correlation, and to reproduce Planck measurements, we need a (three-dimensional) misalignment angle with a root mean squared dispersion of about 50 degrees. These parameters are not sensitive to the particular filament density profile. By artificially skewing the distribution of the misalignment angle, this model can reproduce the Planck-observed (and parity-violating) $TB$ correlation. The skewing of the misalignment angle necessary to explain $TB$ will cause a yet-unobserved, positive $EB$ dust correlation, a possible target for future experiments.'
author:
- 'Kevin M. Huffenberger'
- Aditya Rotti
- 'David C. Collins'
bibliography:
- 'ref.bib'
title: 'The Power Spectra of Polarized, Dusty Filaments'
---
Introduction
============
Polarized Galactic microwave emission poses a challenge to the search for primordial $B$-mode polarization in the Cosmic Microwave Background (CMB). The $B$-mode signal could provide direct constraints on the energy scale of inflation, but the Milky Way foregrounds may outshine it at all frequencies, everywhere on the sky . We have a limited understanding of this foreground, and we must learn more to ensure the reliability of future $B$-mode measurements.
The foreground emission involves the turbulent interplay of gas, dust, and magnetic fields in the Galaxy’s interstellar medium (ISM). The magnetic fields organize the flow and control the precession of grains that give rise to the polarized dust signal. Although Planck’s 353 GHz polarization channel has given us a first look, several features of the dust polarization remain without physical explanations.
For example, the amplitude of dust polarization $B$-mode power is approximately half of $E$-mode power, $A_{BB}/A_{EE} = 0.53 \pm 0.01$, when fit on a large portion of the sky (for $f_{\rm sky}^{\rm eff}=0.52$–$0.71$). Smaller patches also show the same mean value $ A_{BB}/A_{EE} = 0.51$, with small patch-to-patch dispersion $\sigma_{BB/EE} = 0.18$ . This observation defied pre-Planck expectations. Random polarization orientations, or coherent orientations overlaying random polarization intensity fluctuations, both yield equal amounts of $E$ and $B$ [@2001PhRvD..64j3001Z; @2014PhRvL.113s1303K].
We have some understanding of dust physics and its relationship to polarization modes. The amplitude and orientation of the dust signal is set by the integrated column density and magnetic field orientation. For $E$ to have more power than $B$ qualitatively means that density fluctuations (structures in the ISM density field) must prefer orientations parallel or perpendicular to the local magnetic field [@2018arXiv180711940R]. This picture is borne out by measurements of the magnetic field orientation in individual, bright, filamentary structures in the Planck 353 GHz data . This is further validated by the observations that linear structures in neutral hydrogen emission, highlighted by a Rolling Hough Transformation, also correlate with the magnetic field direction indicated by Planck dust polarization [@2014ApJ...789...82C; @2015PhRvL.115x1302C].
We do not know if such filamentary structures are the dominant contribution to the polarization foreground. There is certainly evidence for filamentary structure in data as well as in simulations of the interstellar medium , but there is no clear consensus on their origin or evolution . The purpose of this paper is to explore what polarization power spectra are possible for filaments, and what the observed power spectra can tell us about their physical properties.
Other aspects of the dust polarization also need physical explanations. Both $E$-mode and $B$-mode spectra follow power laws ($C_\ell \propto \ell^\alpha$), with approximately the same slope, $\alpha_{BB} = -2.42 \pm 0.02$ and $\alpha_{EE} = -2.45 \pm 0.03$. There is a positive correlation between dust intensity and $E$-mode polarization [noted by @2017ApJ...839...91C], with correlation coefficient $r_{TE}=0.357 \pm 0.003$ [@2018arXiv180104945P] and significant scatter depending on the sky area but little evidence for scale dependence. Perhaps more intriguing is a parity-violating, positive $TB$ correlation [@2018arXiv180104945P]. Finally, the amplitude of dust polarization power correlates to intensity in patches, roughly as $\langle I\rangle_{\rm patch}^{1.9}$, for both $E$ and $B$ .
A few works have already tried to address these observations. @2017ApJ...839...91C examined the dust polarization power spectra of slow, fast, and Alfvén MHD waves in terms of two parameters: the ratio of gas to magnetic pressure, and the anisotropy of the MHD modes around the background field direction. They found two regions of parameter space that can account for the $E$ to $B$ ratio and positive $TE$ correlation but judged that these scenarios are unlikely due to the uniformity of the polarization power spectrum across the sky, and instead suggested that Planck may be seeing large scale displacements that are driving the turbulent ISM, rather than the turbulence itself.
On the other hand, @2017MNRAS.472L..10K argued with a similar analysis that the observed $E/B$ power ratio can be realized in an MHD model, so long at the turbulent flow is sub-Alfvénic. @2018MNRAS.478..530K extended this analysis to examine the $TE$ correlation and synchrotron emission.
Other works approach the problem using MHD simulations. For the most part, the ISM is filled with trans- and super-sonic flows, which are non-linear . Both @2018PhRvL.121b1104K and @2019arXiv190107079K made MHD simulations of the ISM, and modeled the dust polarization signals. Both works find slopes and power ratios that are reasonably close to the observed values, but the slopes are especially sensitive to the masking procedure. What MHD simulations do not provide is a straightforward and direct way to understand why these polarization properties arise.
Here we seek to gain physical intuition with very simple models of polarized filaments. We compute their temperature and polarization power spectra using a method akin to the cosmological halo model [e.g. @2000MNRAS.318..203S; @2002PhR...372....1C]. However, instead of spherical halos, we use magnetized, prolate-spheroidal filaments as the basic ingredients, and integrate over their population.
We organize this paper so that, in Section \[sec:method\], we describe our formalism for characterizing the filament signal and for computing the power spectra. In Section \[sec:results\], we show the power spectra and discuss how the parameters of the filament population affect them. In Section \[sec:discussion\], we conclude and discuss the implications and possible future directions. An appendix describes how the distributions of filament and magnetic field orientations in three dimensions appear when projected onto the plane of the sky.
Method {#sec:method}
======
We define a projected filament profile, $f(\bm{x})$, upon which we paint the temperature (i.e. intensity) and polarization signals. Thus the temperature profile is: $$T(\bm{x}) = T_0 f(\bm{x}),$$ for sky position $\bm{x}$. We model the polarization with an overall polarization fraction and polarization direction. In terms of the Stokes parameters, the polarization for a filament is $$\begin{aligned}
X(\bm{x}) &=& (Q+iU)(\bm{x}) \\ \nonumber
&=& f_{\rm pol} \exp(2i \psi_{\rm pol}) T_0 f(\bm{x}). $$ In the HEALPix polarization convention [@2005ApJ...622..759G], the $+x$-axis points south and the polarization angle $\psi_{\rm pol}$ increases east of south. Because we will integrate over angles in our computation of the power spectrum (and because $E$ and $B$ fields are coordinate independent) we can analyze a filament that has its long axis aligned (in projection) with the $x$-axis without loss of generality. For simplicity, we assume that the intrinsic, microphysical contribution to $f_{\rm pol}$ is common to all filaments, although we will account for geometrical and projection effects in this work. If the long axis of that filament were aligned with the local magnetic field, the precession of the dust grains would cause the polarization angle to be $\psi_{\rm pol} = 90^\circ$, perpendicular to the filament axis.
![Stokes parameters and scalar polarization quantities for idealized filaments as we alter the aspect ratio. The magnetic field (thin arrow) is vertical, so the polarization direction is horizontal (thick line), making $Q<0$ and $U=0$. In sky convention, north is to the top and east is to the left.\[fig:filament\] The filaments are 2, 5, and 20 times longer than they are wide (axis ratio $\epsilon = 0.5, 0.2,0.05$). Scalars $E$ and $B$ are on the same color scale, which has half the range of the $Q$ scale. The $T$ scale differs from $Q$ by an arbitrary polarization fraction.](figs/filamentTQEB.pdf){width="\columnwidth"}
Working in the flat sky approximation, the Fourier components of the scalar polarization modes are: $$(E+iB)({\bm{\ell}}) = \exp(-2i \phi_{{\bm{\ell}}}) X({\bm{\ell}}).$$ Fig. \[fig:filament\] shows the Stokes $T,Q,U$ and scalar $E,B$ quantities on the sky for sample, north–south filaments with $\psi_{\rm pol}=90^\circ$, so the magnetic field is parallel to the filament direction and the polarization is perpendicular. (Our choice of coordinates implies that Stokes $U$ is zero in these cases.) When the magnetic field aligns with the filament direction, @2018arXiv180711940R pointed out that the real-space kernels for the $E/B$ signals show immediately that the $E$-type polarization is positive along the filament, regardless of its orientation. Since the temperature signal is also strong there, such filaments naturally yields a strong and positive $TE$ cross-correlation, as observed in the Planck data. The same work showed that the $B$ signal is concentrated at the ends of the filament, so filaments with long and thin aspect ratios will have less $B$ power relative to $E$ power than more squat ones.
![Like Fig. \[fig:filament\], but showing the polarization quantities as we alter the magnetic field direction (and hence the polarization angle), for a fixed filament orientation. The $Q$ and $U$ Stokes parameters follow HEALPix polarization convention for a north–south filament, but the $E$ and $B$ fields are coordinate independent and appropriate for any orientation. Note that filaments with aligned magnetic fields ($\psi_{\rm pol} = 90^\circ$) have zero $TB$ correlation. Filaments with relative polarization angles $90^\circ < \psi_{\rm pol} < 180^\circ$ have positive $TB$ correlations, as depicted, while those with $0^\circ < \psi_{\rm pol} < 90^\circ$ have negative $TB$ correlations (not shown).[]{data-label="fig:pol_angle"}](figs/filament_anglerotation.pdf){width="\columnwidth"}
By parity symmetry, the $TB$ and $EB$ cross-correlations are zero when the polarization is perpendicular to the filament (i.e. $\psi_{\rm pol} =90^\circ$). In Fig. \[fig:pol\_angle\] we show how the $E$ and $B$ patterns transform into each other (and change sign) as $\psi_{\rm pol}$ varies away from $90^\circ$. In these cases, the $TB$ and $EB$ correlation can be non-zero for individual filaments, but so long as the average $\langle \psi_{\rm pol} \rangle =90^\circ$, there will be no overall cross-correlation for the whole population.
Projection on the sky
---------------------
We next discuss the projection of a three-dimensional filament onto the plane of the sky. Many important quantities depend on the angle to the line of sight of (1) the long axis of the filament ($\theta_L$) and (2) the magnetic field vector ($\theta_H$).[^1] Another important quantity is the the plane-of-sky projection of the angle between these vectors ($\psi_{LH}$), which controls the polarization angle and the amounts of $E$/$B$ polarization present. We depict these angles in Fig. \[fig:anglefig\]. If the magnetic field direction aligns somewhat with the filament direction, as is the case in strong-field MHD, all these angles will be correlated.
![Geometry of the filament direction and the magnetic field. The long axis of the filament $\mathbf{L}$ points in the $x$–$z$ plane at an angle $\theta_L$ from the line of sight. The magnetic field $\mathbf{H}$ has an angle from the line of sight of $\theta_H$. The angular misalignment between the field and the filament is $\theta_{LH}$. The front view gives the projection of the misalignment onto the plane of the sky, $\psi_{LH}$.[]{data-label="fig:anglefig"}](figs/geometry.pdf){width="\columnwidth"}
We assume that on average, the filaments align with the local magnetic field. In the appendix, we use simple geometry to compute the distribution of the magnetic field projection angle $\theta_H$ and relative orientation angle $\psi_{LH}$ as a function of $\theta_L$. We base the distribution on the assumption of a Gaussian distribution for the angle ($\theta_{LH}$) between the filament and the magnetic field in three dimensions, characterized by the dispersion ${\rm RMS}(\theta_{LH})$.
The field angle $\theta_H$ is correlated with $\psi_{LH}$, so our numerical procedure yields the tabulated joint distribution, $$p(\psi_{LH},\theta_H | \theta_L).$$ This distribution centers on aligned filaments ($\psi_{LH} = 0^\circ,\theta_H = \theta_L$), and the distribution for $\psi_{LH}$ broadens for filaments nearly along the line of sight. Its precise form is not vital for this discussion and is plotted in the appendix in Fig. \[fig:psiLH\_thetaH\_dist\].
On the other hand, the probability distribution for the line-of-sight angle of randomly oriented filaments is determined purely by geometry, $$p(\theta_L) = \sin \theta_L,$$ for $\theta_L \in [0,180^\circ]$.
These quantities relate immediately to the polarization. Although the dust polarization fraction depends on the microphysical details of the emission, it has a geometric dependence like $f_{\rm pol} \propto \sin^2 \theta_H$ [@2000ApJ...544..830F]. Meanwhile, the polarization angle for a filament projected along the $x$-axis is $\psi_{\rm pol} = \psi_{LH} + 90^\circ$.
![Total intensity, polarization fraction, and polarization amplitude dependence on the filament orientation. Filament oriented along the line of sight have $\theta_L = 0^\circ$, while filaments in the plane of the sky have $\theta_L = 90^\circ$. The filament is modeled with an axis ratio $\epsilon = 0.25$, and the filament direction ($\theta_L$) and the magnetic field direction ($\theta_H$) are either perfectly aligned or stochastically misaligned in three-dimensions. []{data-label="fig:pol_vs_thetaL"}](figs/pol_vs_orientation.pdf){width="\columnwidth"}
We model the filament as a prolate spheroid, and label the major axis as $L_a$ and the minor axis as $L_b$. The axis ratio is thus $\epsilon = L_b/L_a < 1$. The column density (and therefore the surface brightness and ultimately the observed temperature perturbation) is proportional to the density and the line of sight distance through the filament, and so (approximately) $$\begin{aligned}
T_0 &\propto& \rho_0 \left( L_a^2 \cos^2 \theta_L + L_b^2 \sin^2 \theta_L \right)^{1/2} \\ \nonumber
&\propto& \rho_0 L_a \left( \cos^2 \theta_L + \epsilon^2 \sin^2 \theta_L \right)^{1/2}.
\label{eq:column_density}\end{aligned}$$ where $\rho_0$ is a characteristic density for the filament. So all else being equal, a filament that lies along the line of sight will have the greatest column density and the brightest temperature signal. On the other hand, $f_{\rm pol} \propto \sin^2 \theta_H$, so if the magnetic field lies along the line-of-sight, there is no polarization. The polarization fraction is maximum when the magnetic field is perpendicular to the line of sight.
Fig. \[fig:pol\_vs\_thetaL\] relates the column density and polarization fraction effects of the line-of-sight angle. It also shows that since the polarized amplitude depends on the product of these two, the filaments with the brightest polarization are inclined, but not perpendicular, to the line of sight. The polarization maximum depends on axis ratio through its impact on the column density. Nearly round filaments have the polarization maximum when oriented near 90$^\circ$ to the line of sight, while in the limit of thin filaments ($\epsilon \rightarrow 0$) the polarization maximum orientation approaches $\theta_L = 45^\circ$ for perfect magnetic field alignment. If there is significant misalignment of the magnetic field and filament directions, the situation can become more complicated, depending on the particular combination of axis ratio and misalignment dispersion. In such cases, filaments along the line of sight can have significant polarization. Still, the typical line-of-sight orientation angle for maximum polarization, averaging over the magnetic field directions, is around $\theta_L = 45^\circ$.
We compute the filament’s projected angular sizes along its two axes as if it were a cylinder. These depend its distance $R$ and are: $$\begin{aligned}
\Theta_a &=& \left( L_a^2 \sin^2 \theta_L + L_b^2 \cos^2 \theta_L \right)^{1/2} / R \\ \nonumber
&=& \left( \sin^2 \theta_L + \epsilon^2 \cos^2 \theta_L \right)^{1/2} L_a / R \\ \nonumber
\Theta_b &=& L_b/R = \epsilon L_a/R.\end{aligned}$$ Thus the projected axis ratio is $$\epsilon_\Theta = \Theta_b/\Theta_a = \frac{\epsilon}{\left( \sin^2 \theta_L + \epsilon^2 \cos^2 \theta_L \right)^{1/2}}$$ which goes to unity for filaments along the line of sight, and to the true value ($\epsilon$) for filaments perpendicular to the line of sight.
Filaments in Fourier space
--------------------------
For several terms in our power spectrum calculation, we need the Fourier transform of the projected filament profile: $$f({\bm{\ell}}) = \int d^2x\ f(\bm{x}) \exp(-i {\bm{\ell}}\cdot \bm{x}).$$ Rather than project rays through a 3-dimensional model to obtain the filament profile, we make a simplifying assumption for computational efficiency. From the size and orientation of a filament, we take the angular dimensions and compute under the assumption that the profile is a distortion from an axisymmetric function $g$: $$\begin{aligned}
f(x,y) =& g(x/\Theta_a , y/\Theta_b)\ & \\ \nonumber =& g(x^*, y^*) \quad & = \ g(r)\end{aligned}$$ where $\Theta_a,\Theta_b$ are the semi-major and semi-minor axis of the elliptical distortion. As we stated, by convention and without loss of generality, we orient the long axis of the filament along the $x$-axis.
Then the transform of $f$ is simply related to the transform of $g$: $$\begin{aligned}
f({\bm{\ell}}) &=& \Theta_a \Theta_b \int d^2x^*\ g(\bm{x}^*) \exp(-i (\Theta_a\ell_x x^* + \Theta_b \ell_y y^*)) \nonumber \\
&=& \Theta_a \Theta_b\, g( \ell^* )\end{aligned}$$ where $ \ell^*({\bm{\ell}}) = (\Theta_a^2 \ell_x^2 + \Theta_b^2\ell_y^2)^{1/2}$ and $$\begin{aligned}
g(\ell^*) = \int d^2x^* \ g(\bm{x}^*) \exp(i {\bm{\ell}}^* \cdot \bm{x}^*) \\ \nonumber
= 2\pi \int dr \ r\ g(r) J_0(\ell^* r).\end{aligned}$$ The input profile $g(\mathbf{x}^*)$ is real and even, and so the Fourier transform is too. The power spectra we find are not very sensitive to the profile that we use. In this work we have used an exponential for the basic filament profile ($g(r) = \exp(-r)$), but we have checked our best-fitting power spectrum model with a Gaussian profile ($g(r) = \exp(-r^2/2)$) and a Plummer profile ($g(r) = (1+r^2)^{-5/2}$), and find the same results.
Parameters and one-filament term
--------------------------------
For a parameterized set of filament properties, $$\alpha = ( L_a, L_b, \psi_{LH}, \theta_L, \theta_H,R,\dots),$$ we can write the number density distribution $n(\alpha)$, so that the average number of filaments in a realization of the sky is $$\langle N \rangle = \int d\Omega\ d\alpha\ n(\alpha)$$ where the integral is over $$d\alpha = dL_a dL_b d\psi_{LH} d\theta_H d\theta_L dR.$$
Expressed another way, $n(\alpha) = \langle N \rangle p(\alpha)$, where the normalized probability distribution of the filament population is $$p(\alpha) = p(L_a,L_b)p(\psi_{LH},\theta_H | \theta_L) p(\theta_L)p(R)$$
This integral over the population is at least six dimensional. For a screen at a distance $R$, it is five dimensional integral. Since the angular power spectrum for foregrounds is a power law, if we can reproduce it on a single screen, putting that screen at different distances will maintain the same power spectrum. If we further fix the physical aspect ratio of the filaments, it is a four dimensional integral, over $L_a, \psi_{LH}, \theta_L, \theta_H$. (The projected aspect ratio will still vary with the line-of-sight angle $\theta_L$.)
The power spectrum contributions from filaments correlated with themselves are: $$\begin{aligned}
C_\ell^{TT} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ | T({\bm{\ell}},\alpha)|^2, \\ \nonumber
C_\ell^{EE} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ | E({\bm{\ell}},\alpha)|^2, \\ \nonumber
C_\ell^{BB} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ | B({\bm{\ell}},\alpha)|^2, \\ \nonumber
C_\ell^{TE} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ T({\bm{\ell}},\alpha) E({\bm{\ell}},\alpha)^*.\end{aligned}$$ Similar expressions hold for the other cross correlations, but these vanish if the orientations of the filaments are random. These power spectra computations are directly analogous to the 1-halo term in the cosmological halo model [@2000MNRAS.318..203S].
Results {#sec:results}
=======
There are clear relationships between the physical properties of the filaments and the temperature and polarization power spectra that they produce. The slopes of the power spectra are determined primarily by the size distribution of filaments, with other effects responsible for the smaller differences between the components. The ratio of $BB/EE$ power is determined mostly by the aspect ratio of the filaments and somewhat by the misalignment of the filament directions to the background magnetic field. These same factors also determine the cross correlation $r^{TE}_\ell$, but here misalignment is much more important. They also affect the $TE/EE$ power ratio, but this quantity is more directly affected by the overall polarization fraction.
Power spectrum shape
--------------------
We can relate the slope of a power law spectrum to scaling relations for parameters in the filament profiles. This allows us to place constraints on the distribution of filament sizes and the scaling of other parameters. For a generic parameter $\alpha_0$, if the filament’s contribution to the power spectrum scales as $$C_\ell \propto \int d\alpha_0\ n(\alpha_0)\times \alpha_0^q \,F( \alpha_0^r\, \ell) \label{eqn:shape_scaling}$$ for any function $F$, and furthermore if the weighting distribution for the parameter is a power law, $n(\alpha_0) \propto \alpha_0^p$, then we can rescale the integration with a straightforward substitution, $u=\alpha_0 \ell^{1/r}$: $$\begin{aligned}
C_\ell &\propto& \ell^{-(p+q+1)/r} \times \\ && \nonumber \quad \int d(\alpha_0 \ell^{1/r})\ (\alpha_0 \ell^{1/r})^p \times (\alpha_0 \ell^{1/r})^q \,F( ( \alpha_0\, \ell^{1/r})^r) \\ \nonumber
&\propto& \ell^{-(p+q+1)/r} \times \int du\ u^p u^q \,F( u^r).\end{aligned}$$ The integral no longer has any multipole dependence and evaluates to some constant value, whatever the details of $F$. Thus we are left with a powerlaw power spectrum with $$C_\ell \propto \ell^{-(p+q+1)/r}$$ This argument holds not just for filaments, but for any signal with a power-law power spectrum that is built from a set of objects that are similarly related to each other, so long as they are weighted by powerlaw scalings and distributions. So if we observe a powerlaw spectrum with $C_\ell \propto \ell^s$, it implies that the parameter distribution’s index is $p = -rs - q - 1$, regardless of the objects’ profiles.
We walk through this scaling argument for a simple (and unrealistic) case—with plane-of-sky filaments with identical surface brightnesses ($T_0$ is the same for all filaments) and a constant projected-axis-ratio ($\Theta_b = \epsilon \Theta_a$)—and analyze the distribution for $\Theta_a$, the angular size of filaments. For the filament Fourier transform, we have $f({\bm{\ell}}) \propto \Theta_a^2 g({\bm{\ell}}^*)$ with ${\bm{\ell}}^* \propto \Theta_a$. The power spectrum contribution is proportional to $f^2$, so comparing the scaling for angular size parameter $\Theta_a$ to equation \[eqn:shape\_scaling\], we find $q=4$ and $r=1$. In the polarization case, to reproduce $C_\ell \propto \ell^{-2.4}$ (meaning $s=-2.4$), the number density distribution of such objects on the sky must approximately scale like $n(\Theta_a) \propto \Theta_a^p$ where $p = 2.4 - 4 - 1 = -2.6$. Indeed this yields the desired power spectrum slope when calculated in our model.
![Power spectra slopes of temperature and polarization are set chiefly by the distribution of filament lengths, which has been chosen here so that the slopes match the dotted line, $C_\ell \propto \ell^{-2.4}$. The ratios between the power spectra are set by the overall polarization fraction, the aspect ratio of filaments, and the misalignment between the filaments and their local magnetic field. When maximum and minimum sizes truncate the distribution of the filaments, the power falls below the targeted slope (semi-transparent colors).[]{data-label="fig:slope"}](figs/power_spectrum.pdf){width="\columnwidth"}
In a more realistic case, with three-dimensional filaments, we can make a similar argument to deduce the distribution of filament lengths. Surface brightness depends on column density, which is proportional to length (after integrating out any distribution of axis ratios—compare equation \[eq:column\_density\]—and assuming a density normalization independent of length). The solid angle scales like length squared. After squaring those three powers during the computation of the power spectrum, the overall scaling is $q=6$. The multipole $\ell$ scaling should also go like length, so $r = 1$, the same as the plane-of-sky case above. So with no other dependence on length, we should have distribution of lengths $n(L_a) \propto (L_a)^p$ where $p = 2.4 - 6 - 1 = -4.6$.[^2]
We have verified that this distribution produces the proper slope in Fig. \[fig:slope\]. All the temperature and polarization spectra have the specified slope in common. The complications of the modeling of the three dimensional orientation are not important to the slope, only the weighting and distribution of filament size.
Other than the slope, there are not clear features in the Planck-measured spectra. We note that features in the distribution of filament sizes would break the powerlaw behavior of the resulting spectra. For example, if we impose a maximum filament size (semi-transparent lines in Fig. \[fig:slope\]), it causes the low-$\ell$ behavior of $C_\ell$ to deviate: at scales much larger than the filament, the temperature spectrum adopts the flat, white, Poisson spectrum of point sources. The $EE$ and $BB$ spectra flatten the same way, and on scales large compared to the filaments, the aspect ratios of the filaments become unimportant and the amount of power in $EE$ and $BB$ equalize. The $TE$ cross-correlation falls off at large scales, possibly because the positive and negative contributions to $E$ are being averaged over. If, on the other hand, we impose a minimum filament size, it causes the high-$\ell$ behavior to deviate: the spectrum will drop off with increasing $\ell$, where there is no more contribution to the power.
In yet more realistic cases, we can make the slopes of all the temperature and polarization spectra differ. For example, this can happen if there is another effect that changes the size scaling of the polarization relative to the temperature. For example, to make polarization slope shallower than the temperature slope, one could make smaller objects more polarized than large objects, or if the aspect ratio of filaments changes as a function of size.
Considering the Planck data, it is not immediately clear what conclusion to draw. @2018arXiv180104945P quote $EE, BB, TE$ slopes for the dust foreground, but not the $TT$ slope. Using our own tools, we have computed the $TT$ power spectrum based on the Planck data, and find a $TT$ slope that is about $-2.6$, somewhat steeper than the polarization spectra at $-2.4$ to $-2.5$. Like @2018arXiv180104945P, for the mask we used the LR71 polarization mask supplemented with a point source mask (based on intensity maps), resulting in a mask with effective $f_{\rm sky} \simeq 0.6$. We can approximately reproduce a $-2.6$ temperature slope and $-2.4$ polarization slope with $n(L_a) \propto L_a^{-4.4}$ and $f_{\rm pol} \propto L_a^{-0.08}$. This argues that in the unmasked region, the polarization is higher in smaller filaments.
However, this conclusion may not be correct because it depends sensitively on the mask. We reasoned that small filaments, oriented along the line of sight, might look like a point source and be included in the masked area. These end-on filaments could also have low polarization (note Fig. \[fig:pol\_vs\_thetaL\]), and excluding them might not have much effect on the polarization results. Thus for comparison, we recomputed the $TT$ spectrum with different masks. When we use only the polarization LR71 mask without removing the additional point sources from the intensity map, we get a shallower $TT$ spectrum with slope $-2.5$. When we use a mask that keeps the same large scale features but does not mask any point sources (Planck’s publicly available GAL70 mask) we find a $TT$ slope of $-2.1$, notably shallower than the polarization spectra. The polarization spectra change somewhat between these masks, but the changes in the polarization slopes are small compared to the change in the TT slopes. Some of the masked sources are extragalactic, so this slope with all point sources unmasked is probably too shallow to describe the ISM component, but can serve as a bound. The upshot is that we are not certain whether the spectrum for all filaments is steeper or shallower in temperature than polarization, and so it is difficult to draw conclusions on the size dependence of the polarization fraction.
Another feature of the Planck data is the differing slopes in $E$ and $B$. We can reproduce this feature by varying the aspect ratio as a function of filament size. For the LR71 mask in @2018arXiv180104945P, the slopes for $(BB, EE, TE)$ are roughly $(-2.5, -2.4, -2.5)$ and we found a $TT$ slope of $-2.6$. So $BB$ is steeper than $EE$, which should happen if smaller filaments are proportionally thinner than longer ones. Modifying the aspect ratio in this way also affects the $TT$ slope, breaking the simple relation that we saw earlier in this section. By trial and error, we found that this set of slopes are approximately reproduced with the following parameter dependence: $\epsilon \propto L_a^{0.1}$, $n(L_a) \propto L_a^{-4.45}$, and $f_{\rm pol} \propto L_a^{-0.1}$. Here we are simply exploring what is possible, but the relationship between the measured slopes and these filament parameters should be made more systematic and quantitative.
In light of these complications and uncertainties, in what follows we keep a common slope of $-2.4$ for all the temperature and polarization components while we explore their other parameter dependences.
BB/EE power ratio
-----------------
The aspect ratio of the filaments is the major factor determining the ratio of $B$-mode power to $E$-mode power. In Fig. \[fig:EEBBrat\], we plot the power ratio against the aspect ratio for varying degrees of filament–magnetic field misalignment. To reproduce the Planck-observed ratio of $\sim 0.5$, filaments need to have an aspect ratio $\epsilon$ slightly less than 0.26, so filaments must be slightly less than four times longer than they are wide. If the model deviates too much from this ratio, the required magnetic field misalignment is made so large that the model has trouble fitting the $TE$ correlation.
It is difficult to compare this result quantitatively to the aspect ratios of observed filaments from the literature without making a detailed accounting of the filament selection function. Projection effects will also tend to lower observed aspect ratios. The stacked filaments in Fig. 7 of appear to have axis ratios not so far from what we are finding here. The filaments identified by the Rolling Hough transformation in @2014ApJ...789...82C on HI maps tend to be longer and thinner than this.
![Ratios of $BB$ to $EE$ power as a function of the physical filament axis ratio. Long and thin filaments ($\epsilon$ small) have less $B$-mode power than $E$-mode power. Short and squat filaments (aspect ratio $\epsilon$ close to unity) have $B$ power close to the $E$ power. An aspect ratio of about $\epsilon = 0.26$ can reproduce the Planck-observed ratio of about one half, but this can be traded off against a slight dependence with the dispersion in the misalignment angle $\theta_{LH}$ between the filament direction and the magnetic field direction in three dimensions.[]{data-label="fig:EEBBrat"}](figs/BErat_vs_eps.pdf){width="\columnwidth"}
![Correlation coefficient of $TE$ power as a function of the axis ratio. Thinner filaments (small $\epsilon$) have stronger $TE$ correlations for various magnetic field–filament misalignments. Since the $B$-to-$E$ power ratio requires $\epsilon \approx 0.26$ (Figure \[fig:EEBBrat\]), the $TE$ correlation is diagnostic of the misalignment necessary to produce the Planck-observed $r^{TE} \approx 0.35$, which needs ${\rm RMS}(\theta_{LH}) \approx 50^\circ$.[]{data-label="fig:rTE"}](figs/rTE_vs_eps.pdf){width="\columnwidth"}
TE cross-correlation
--------------------
In the context of the filament model, we find that the level of the $TE$ cross correlation implies that filaments cannot be precisely aligned to their local magnetic field direction. The correlation coefficient is defined as $$r_\ell^{TE} = C_\ell^{TE}/\sqrt{C_\ell^{EE} C_\ell^{TT}},$$ and perfect alignment of the filaments and the fields causes far too much $TE$ correlation compared to the Planck observations.
The Planck dust data show $r_\ell^{TE} \approx 0.35$ with little scale dependence [@2018arXiv180104945P]. Fig. \[fig:rTE\] shows that to match this, the field misalignment angle $\theta_{LH}$ must have an RMS dispersion of nearly $50^\circ$, while maintaining the axis ratio $\epsilon \approx 0.26$ needed to reproduce the $BB/EE$ power ratio. If the misalignment dispersion is independent of filament size, as in our modeling, it causes no scale dependence: $r^{TE}_\ell$ is constant.
[Projection effects cause the distribution of the projected angle $\psi_{LH}$ to have a positive kurtosis (see appendix, Fig. \[fig:psiLH\_marg\]), and so we can describe its dispersion in a few ways. For the ${\rm RMS}(\theta_{LH}) = 50^\circ$ case, 68 percent of the probability is bounded by $|\psi_{LH}| < 45^\circ$. Alternatively, $[{\rm Var}(\psi_{LH})]^{1/2} = 48^\circ$.]{} For comparison, fit a Gaussian a with $19^\circ$ dispersion (1$\sigma$) to the projected field–projected filament histogram of relative orientations for the filaments they found. Again this comparison is not direct because of selection effects. Their Hessian-based selection of filaments would disfavor filaments with small projected aspect ratios (close to the line of sight), and such filaments can have the largest differences in the projected orientation.
The overall level of $C_\ell^{TE}$ (and the polarization spectra) depends on the polarization fraction. The relative power ratio has a dependence like $$C_\ell^{TE}/C_\ell^{EE} \propto \langle f_{\rm pol} \rangle/\langle f_{\rm pol}^2 \rangle,$$ while for the cross correlation it is $$r_\ell^{TE} \propto \langle f_{\rm pol} \rangle/\langle f_{\rm pol}^2 \rangle^{1/2}.$$ Thus a purely multiplicative rescaling of the polarization fraction affects the ratios of the power in $TT/TE/EE$ but not the correlation coefficient $r^{TE}$.
To reproduce the Planck-measure ratio $C_\ell^{TE}/C_\ell^{EE} \sim 2.7$ (in our case that already fits $C_\ell^{BB}/C_\ell^{EE}$ and $r_\ell^{TE}$) requires $f_{\rm pol} = 0.15\, \sin^2 \theta_H$. We have only modeled the polarization fraction amplitude and the geometric dependence on the magnetic field orientation, but in addition, the polarization fraction depends on grain geometry and small-scale turbulence [@2000ApJ...544..830F], and filaments need not in reality have all the same intrinsic polarization fraction.
Our other tests have shown that the $TE$ correlations differ in their sensitive to intrinsic dispersion in the polarization fraction. For example, the correlation $r^{TE}$ is not very sensitive to the maximum polarization fraction, but the power ratio is very sensitive to it: decreasing the maximum polarization fraction decreases $C_\ell^{TE}$ but decreases the denominator $C_\ell^{EE}$ more, and so raises the ratio.
Parity violation: TB and EB
---------------------------
One surprising finding in the @2018arXiv180104945P dust spectra is a non-zero $TB$ correlation, with $r_\ell^{TB} \approx 0.05$. Like $r^{TE}$, the observed $r^{TB}$ correlation has little scale dependence (up to multipoles of several hundred).
Because non-zero $TB$ and $EB$ are parity-violating correlations, our model cannot reproduce them for randomly-oriented filaments. To get a positive $TB$ correlation we would need to favor polarization angles in the range $\psi_{\rm pol} \in [90^\circ,180^\circ]$ relative to the filament direction (compare Fig. \[fig:pol\_angle\]). Equivalently, this corresponds to projected field angles in the range $\psi_{LH} \in [0^\circ,90^\circ]$. Such an effect may be due to some large scale feature in the Galaxy’s magnetic field or differential gas flow .
We can determine how far away from random this correlation is by artificially weighting the distribution of the projected misalignment angles, favoring the $\psi_{LH} > 0$ portion of the distribution of $p(\psi_{LH},\theta_H | \theta_L)$ over the $\psi_{LH} < 0$ portion, while keeping the same functional form. We find that we can approximately reproduce the Planck measured $TB$ correlation by giving the preferred $\psi_{LH}$ directions about 55 percent of the total weight, rather than the 50 percent than comes naturally from randomly oriented filaments. Similar to the $r^{TE}$ correlation (also set by field–filament misalignment), this effect is not scale dependent, and so $r^{TB}$ is constant to high $\ell$ in this model.
Our modeling comes from the 1-halo term only, and shows that the $TB$ correlation can be explained if filaments orientations in projection are slightly twisted counterclockwise from the projected local magnetic field. Our model does not address the structure of that underlying field, but we may speculate that some differential, shearing hydrodynamic forcing could preferentially twist the filaments, according to our point of view, from the global mean field direction of the Milky way.
, argue that the observed $TE$ and $TB$ correlations may be features of the large scale structure of the Galactic magnetic field. They show that a helical component can create such correlations, but in their modeling, the correlation show a strong scale dependence, with $r^{TE,TB}_\ell$ falling substantially already by $\ell = 22$. The Planck spectra have much flatter $r^{TE,TB}_\ell$ correlations, consistent with the filament modeling here.
Planck did not detect an $EB$ correlation, but since $E$ and $B$ both have a factor of the polarization fraction, we would naturally expect this correlation to be smaller. It may be there, hidden in the noise. In the presence of a positive $TB$ correlation, in the context of the filament modeling, we would expect a positive $EB$ correlation too (including at high-$\ell$), and it should be a target for future experiments. Both $TB$ and $EB$ dust correlations can potentially interfere with sky-calibration of the polarization angles of CMB-instruments [@2016MNRAS.457.1796A] or with CMB lensing reconstruction [e.g. @2012JCAP...12..017F; @2018JCAP...04..018C].
Conclusions {#sec:discussion}
===========
We do not know how much of the dusty microwave polarization foreground is due to filamentary structure, but if it is a substantial portion, we can discern details of the filament population from the foreground power spectra. We showed that the slopes of the power spectra relate to the distribution of lengths. We showed that the $BB$/$EE$ power ratio relates to the filament axis ratio. We showed that the $TE$ cross-correlation relates to the axis ratio and the RMS misalignment of filaments to the magnetic field. We showed that $TB$ correlations could be caused by a slight preference for one handedness in the misalignment between the magnetic field and the filament orientation.
Despite its relative success in reproducing the features of the dust polarization power spectrum, this formalism lacks some essential features for modeling the real sky. Foremost, this formalism includes only the one-halo term in the power spectra. This is obviously an approximation, for the Planck data have shown that the Galaxy’s projected magnetic field has coherent, large-scale features, and the HI-identified filaments are clearly correlated with it and with starlight polarization measurements [@2015PhRvL.115x1302C]. On the other hand, the transition from one-halo-dominated to two-halo-dominated scales often leaves a mark on the power spectrum. Since in the dust polarization spectra there are not clear features, like a break in the slope, we may speculate that the two-halo component may not be necessary to describe the main properties of the power spectra. Inclusion of a proper two-halo formalism is complicated by the correlated direction dependence of the filaments. We may be able to import some of the techniques developed to describe galaxy intrinsic alignments [e.g @2010MNRAS.402.2127S], since the mathematical description of the problem is similar.
We have not tried to systematically probe the parameter degeneracies or place proper uncertainties on any of the parameters of this filament model. We can do this straightforwardly by interfacing the model with a Monte Carlo Markov Chain and developing a likelihood based on the Planck dust spectra. We plan to pursue this in further work.
By looking at observations and simulations of the filamentary ISM, we could attempt to verify some of the population statistics for filaments. For example, we could compare the length distribution of actual or simulated filaments to that implied by the slope of the power spectra.
Because this filament model is non-Gaussian, we may be able to use it to design novel diagnostics to probe for residual foregrounds in surveys that aim for the primordial $B$-modes [in the spirit of @2014PhRvL.113s1303K; @2016JCAP...09..034R; @2018MNRAS.479.5577P; @2019arXiv190104515C]. Similarly, we could use this model to compute the four-point contributions to polarized CMB lensing estimators. This could help place constraints on potential foreground contamination. Such statistics may be sensitive to the internal density structure of the filaments is a way that the power spectrum is not.
Due to its flexibility, its ability to model the Planck dust polarization data, its ease of computation, and its straightforward interpretation, this filament model may become a useful tool in the study of CMB polarization foregrounds.
Acknowledgments {#acknowledgments .unnumbered}
===============
KMH, AR, and DCC acknowledge support from the NASA ATP program under grant NNX17AF87G. KMH acknowledges support from the NSF AAG program under grant 1815887. AR acknowledges support from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 725456, CMBSPEC). We thank J. Colin Hill, Susan Clark, Marc Kamionkowski, and Carlos Hervias-Caimapo for useful conversations. We thank Francois Boulanger for providing access to masks used in the @2018arXiv180104945P analysis.
Distributions of angles
=======================
We describe the directions of (the long-axis of) the filament and the magnetic field with the coordinates in Fig. \[fig:anglefig\]. The line of sight lies along the $z$-axis, while $x$-axis is down, and the $y$-axis is left. We assume that the filament $(\mathbf{L})$ lies in the $x$–$z$ plane, making angle $\theta_L$ to the line-of-sight. Equivalently, the directions can be expressed as a rotation around the $y$-axis: $$\hat {\mathbf L} = {\mathbf R}_{\hat {\mathbf y}} (\theta_L) \hat {\mathbf z}.$$ The magnetic field $(\mathbf{H})$ we describe with respect to the filament axis, using spherical-polar coordinates $(\theta_{LH},\phi_{LH})$: $$\hat {\mathbf H} = {\mathbf R}_{\hat {\mathbf L}} (\phi_{LH}) {\mathbf R}_{\hat {\mathbf y}} (\theta_{LH}) \hat {\mathbf L}$$ When we numerically generate realizations of these directions, we use the Rodrigues rotation formula to rotate the vectors around the proper axis. We use the distribution of $\theta_{LH}$ values to statistically characterize the misalignment in three dimensions between the filaments and their local magnetic fields.
The important quantities for computing the polarization of this filament are the angle that the magnetic field makes with the line-of-sight, expressed as: $$\cos \theta_{H} = \hat {\mathbf H} \cdot \hat {\mathbf z}.$$ When projected onto the plane of the sky, the filament and the field are separated by a misalignment angle $\psi_{LH}$, computed as: $$\tan \psi_{LH} = \hat H_y/ \hat H_x$$ in the proper quadrant. For dust, the polarization angle relative to the filament direction is $\psi_{\rm pol} = \psi_{LH} + \pi/2$.
For a particular filament angle $\theta_L$ and field–filament misalignment $\theta_{LH}$, changing the angle $\phi_{LH}$ rotates the field $\mathbf{H}$ to sweep out a cone around the filament direction. We can use the fact that $p(\phi_{LH})$ is uniform on $[0,2\pi]$ to numerically accumulate the joint distribution of the field angle and projected field–filament misalignment: $$p(\theta_H, \psi_{LH} | \theta_{L}, \theta_{LH}).$$ Because of the projections, this makes a loop of probability in the $(\theta_H, \psi_{LH})$ parameter space. When the separation between the filament and field $\theta_{LH}$ is small, the loop centers tightly around $\theta_H = \theta_L$, $\psi_{LH} = 0$, and when $\theta_{LH}$ is larger, the loop is larger and more distorted.
![The joint distribution of the line-of-sight angle of the magnetic field ($\theta_H$) and the projected angle between the magnetic field and the long-axis of the filament ($\psi_{LH}$), under the assumption that the magnetic field direction has a Gaussian random distribution around the filament direction. The contours mark lines of constant probability density, and the number records the integrated probability outside the contour. The peak of the distribution is at $\psi_{LH}=0^\circ$, $\theta_H = \theta_L$, corresponding to a filament aligned with the local magnetic field. The distribution is symmetric in the projected separation so we only show the half with $\psi_{LH} > 0^\circ$. In the left column, the field and filament are more closely aligned (${\rm RMS}(\theta_{LH}) = 10^\circ$) than in the right column (${\rm RMS}(\theta_{LH}) = 50^\circ$). In the top row, the filament is perpendicular to the line-of-sight ($\theta_L = 90^\circ$) and so is in the plane of the sky. In the bottom row, the filament aligns nearly along the line of sight ($\theta_L = 10^\circ$). For filaments along the line of sight (small $\theta_L$), even a small misalignment with magnetic field can cause the projected angle ($\psi_{LH}$) to vary widely, and so the distribution of the projected angle is broad. []{data-label="fig:psiLH_thetaH_dist"}](figs/psiLH_thetaH_dist.pdf){width="0.75\columnwidth"}
![(Left) Distribution of the magnetic field angle, for three different filament angles, assuming a Gaussian distribution for the misalignment of the field and filament angle in three-dimensions. Each distribution peaks at the filament direction (for aligned filaments). (Right) Distribution of the projected misalignment between the filament and the magnetic field. The distribution is symmetric about $\psi_{LH} = 0^\circ$.[]{data-label="fig:marginal_psiLH_thetaH_dist"}](figs/psiLH_dist.pdf "fig:"){width="0.49\columnwidth"} ![(Left) Distribution of the magnetic field angle, for three different filament angles, assuming a Gaussian distribution for the misalignment of the field and filament angle in three-dimensions. Each distribution peaks at the filament direction (for aligned filaments). (Right) Distribution of the projected misalignment between the filament and the magnetic field. The distribution is symmetric about $\psi_{LH} = 0^\circ$.[]{data-label="fig:marginal_psiLH_thetaH_dist"}](figs/thetaH_dist.pdf "fig:"){width="0.49\columnwidth"}
![The distribution of the projected filament–field misalignment angle for randomly oriented filaments, under the assumption that the three-dimensional misalignment angle is Gaussian distributed, for various misalignment dispersions. The projected distribution is symmetric around $\psi_{LH} = 0$.[]{data-label="fig:psiLH_marg"}](figs/psiLH_dist_marg_LH_L.pdf)
If we make an assumption for the distribution of the field–filament misalignment angle $\theta_{LH}$, we can marginalize over it. $$p(\theta_H, \psi_{LH} | \theta_L) = \int d\theta_{LH}\, p(\theta_H, \psi_{LH} | \theta_{L}, \theta_{LH}) p(\theta_{LH}) .$$ \[eq:thetaH\_psiLH\] In Fig. \[fig:psiLH\_thetaH\_dist\], we show this joint distribution for several filament directions, under the assumption that the misalignment angle is Gaussian distributed. For filaments along the line of sight (small $\theta_L$), the distribution in projected angle ($\psi_{LH}$) is broad, which makes sense as variations in the angle $\phi_{LH}$ cause a wide variety of $\psi_{LH}$ angles. In Fig. \[fig:marginal\_psiLH\_thetaH\_dist\], we also examine the marginal distributions $p(\psi_{LH}|\theta_L)$ and $p(\theta_{H}|\theta_L)$.[^3] In Fig. \[fig:psiLH\_marg\], we further marginalize over the filament angles to get the distribution of the projected field–filament misalignment: $$p(\psi_{LH}) = \int d\theta_{L}\, d\theta_{H}\, p( \psi_{LH} | \theta_L) p(\theta_{L}).$$ These marginalized distribution have positive kurtosis and are more sharply peaked than the Gaussian distribution from which they are derived. It is important to note that the polarized flux in practice will depend both on the polarization fraction (dependent on $\sin^2 \theta_H$) and the optical depth (dependent on the filament physical size, aspect ratio, and orientation $\theta_L$), and so the observed distribution of misalignment above some signal-to-noise cut will differ, and must be computed from the multidimensional distribution accounting for survey characteristics.
[^1]: Elsewhere in the ISM literature, the angles are often given with reference to the plane of the sky, e.g. $\gamma_H = 90^\circ - \theta_H$ and so on.
[^2]: If the column density normalization depends on length, this procedure yields a net distribution that is a product of the size distribution and the density squared distribution (both as a function of length). Such a case could arise, for example, if small filaments are collapsed versions of large ones and have higher density.
[^3]: At the beginning we fixed the plane-of-sky orientation of the filament, but we could have favored the magnetic field instead, and by symmetry we should have $p(\theta_{H}|\theta_L) = p(\theta_{L}|\theta_H)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
During the last twenty years or so a wide range of realizability interpretations of classical analysis have been developed. In many cases, these are achieved by extending the base interpreting system of primitive recursive functionals with some form of bar recursion, which realizes the negative translation of either countable or countable dependent choice. In this work we present the many variants of bar recursion used in this context as instantiations of a general, parametrised recursor, and give a uniform proof that under certain conditions this recursor realizes a corresponding family of parametrised dependent choice principles. From this proof, the soundness of most of the existing bar recursive realizability interpretations of choice, including those based on the Berardi-Bezem-Coquand functional, modified realizability and the more recent products of selection functions of Escardó and Oliva, follows as a simple corollary. We achieve not only a uniform framework in which familiar realizability interpretations of choice can be compared, but show that these represent just simple instances of a large family of potential interpretations of dependent choice principles.
**Keywords:** Dependent choice, Modified realizability, Open induction, Bar recursion, Continuous functionals.
author:
- |
Thomas Powell\
Institute of Computer Science, University of Innsbruck[^1]
bibliography:
- '/home/thomas/Documents/tp.bib'
title: 'Parametrised bar recursion: A unifying framework for realizability interpretations of classical dependent choice'
---
Introduction {#sec-introduction}
============
One of the central problems of mathematical logic and computer science is to understand the constructive meaning of non-constructive proofs. From the early 20th century onwards a rich variety of constructive interpretations of classical logic have been developed, together with techniques for extracting computational content from proofs. The majority of these techniques initially deal with proofs in formal systems of classical predicate logic or weak subsystems of mathematics such as Peano arithmetic. In order to interpret stronger non-constructive proofs from subsystems of mathematical analysis which contain principles such as the axiom of dependent choice, $$\DC \; \colon \; \forall n,x^\rho\exists y^\rho A_n(x,y)\to\exists f^{\NN\to \rho}\forall n A_n(f(n),f(n+1)),$$ these techniques must typically be adapted and extended - a process which tends to be highly non-trivial.
In this paper we focus on just one method of giving a computational interpretation to proofs: namely classical modified realizability. A long established way of extending this technique to deal with choice principles is to introduce some kind of bar recursion to the usual interpreting system of primitive recursive functionals. Bar recursive interpretations of choice principles are among the most widely known and studied methods for giving a computational interpretation to classical analysis, and numerous instances of this method can be found in the literature. However, both the particular variant of bar recursion used and the form of choice that it realizes differs from case to case, and it is not well-understood how these variants compare as realizers. This is compounded by the fact that each variant was typically devised in a distinct setting, and both their existence and correctness often established using a slightly different method of proof, obscuring the inherent similarities between them all.
The main contribution of this paper is to construct a general bar recursive term which contains several free parameters. We show that whenever these parameters obey certain conditions, the term can be used to realize the negative translation of a correponding variant of countable dependent choice, and that moreover essentially all of the known variants of bar recursion used to extend modified realizability to classical analysis appear as simple instantiations of the parameters.
Our motivation for this is twofold. Firstly, we obtain a clear, unifying framework in which the differences between existing variants of bar recursion and their correctness proofs are essentially reduced to simple structures on the natural numbers, and thus their behaviour as realizers is more easily compared. Secondly, in doing this we greatly generalise the interpretation of classical analysis by giving not one but a whole class of realizers of choice principles, which can be freely chosen between to suit the problem at hand. Having a such a wide range of options in front of us means that in concrete instances of program extraction we are given the potential to tailor our realizer to the situation at hand and extract computational content which is more efficient and semantically meaningful.
Realizability interpretations of classical analysis: A brief overview {#sec-introduction-history}
---------------------------------------------------------------------
This work is designed to be as self-contained as possible, and in particular no prior knowledge of bar recursion is assumed. However, because we are partly motivated by the desire to unify existing interpretations, we provide here a very short summary of some of the best known variants of bar recursion that we aim to bring together. We simply state without explanation their defining equations: the purpose here is not to treat in any detail the exact meaning of these objects, but to allow the reader to see certain key features of the recursion that will be dealt with in a more general setting later.
The idea of interpreting countable choice with bar recursion was first established not for realizability interpretations, but for G[ö]{}del’s slightly more intricate Dialectica interpretation, in a fundamental paper of Spector in 1962 [@Spector(1962.0)]. Here, bar recursion was given as the schema $$\label{eqn-intro-SBR}\SBR(Y,G,H,s^{X^\ast})=\begin{cases}G(s) & \mbox{if $Y(\ext{s})<|s|$}\\ H(s,\lambda x\; . \; \SBR(Y,G,H,s\ast x)) & \mbox{otherwise}.\end{cases}$$ The key details here are that $s\colon X^\ast$ is a finite sequence, $s\ast x$ denotes the extension of $s$ with the object $x$, $\ext{s}$ its canonical extension as an infinite sequence, and the output type of $Y$ is a natural number. Spector’s bar recursion is therefore a form of ‘backward recursion’ in which recursive calls are made on extensions $s\ast x$ of the argument $s$. The parameter $Y$ is responsible for terminating the recursion, based on the assumption that the underlying tree barred by sequences satisfying $Y(\ext{t})<|t|$ is well-founded - a fact that is true in all continuous models and also in non-continuous structures such as the strongly majorizable functionals [@Bezem(1985.0)]. What this means is that for any sequence $s\ast x_1,s\ast\pair{x_1,x_2},\ldots$ of recursive calls made in the computation of $\SBR(Y,G,H,s)$, there is always some $N$ such that $Y(\widehat{s\ast\pair{x_1,\ldots,x_N}})<|s|+N$, and thus $\SBR(Y,G,H,s)$ is a well-defined value.
An early adaptation of this idea to realizability was given by Berardi et al. [@BBC(1998.0)], in which an interpretation for dependent choice broadly similar but somewhat simpler than Spector’s variant of bar recursion was established. This realizer was later reformulated in the more standard framework of modified realizability in [@BergOli(2005.0)], where it was given its now familiar name *modified bar recursion* and was defined as $$\label{eqn-intro-MBR}\MBR(Y,H,s)=_\NN Y(s\at\lambda n^\NN.H(s,\lambda x\; . \; \MBR(Y,H,s\ast x))).$$ Here $s$ is again a finite sequence, $s\at\alpha$ denotes the infinite sequence $\alpha$ overwritten by $s$, and the outcome type of $Y$ is a natural number. This restriction on the type of $Y$ ensures that in continuous models $Y$ only looks at a finite amount of information from its input sequence. In other words, given a sequence of recursive calls $s\ast x_1,s\ast\pair{x_1,x_2},\ldots$ in the computation of $\MBR(Y,H,s)$, there is some point $N$ such that the value of $q(s\ast \pair{x_1,\ldots,x_N}\oplus\ldots)$ is determined based only on $s\ast \pair{x_1,\ldots,x_N}$, and so no further recursive calls are necessary. Thus, like Spector’s variant of bar recursion (\[eqn-intro-SBR\]), modified bar recursion is carried out over some underlying well-founded tree, although unlike Spector’s bar recursion, this tree is not computable but it *implicitly* well-founded by some kind of continuity argument.[^2]
In addition to what later became modified bar recursion, [@BBC(1998.0)] contains a striking realizer for (the negative translation of) countable choice, which took from Spector only the basic idea of backward recursion, replacing the sequential bar recursive calls $s\mapsto s\ast x$ of (\[eqn-intro-SBR\]) and (\[eqn-intro-MBR\]) with a symmetric updating $u\mapsto\update{u}{n}{x}$ of finite partial functions. This realizer was again put into a standard realizability framework in [@Berger(2002.0); @Berger(2004.0)], and will be referred to here as the Berardi-Bezem-Coquand, or BBC-functional. Its defining equation (following [@Berger(2002.0)]) is given by $$\label{eqn-intro-BBC}\BBC(Y,H,u)=_\NN Y(u\at\lambda n.H(n,\lambda x\; . \; \BBC(\update{u}{n}{x}))).$$ Here, $u$ is a finite partial function and $\update{u}{n}{x}$ is the domain-theoretic extension of $u$ with value $x$ at input $n$. As with modified bar recursion, $\BBC$ terminates by continuity of $Y$, although proving this is somewhat technical and as shown in [@Berger(2004.0)] is most elegantly done with Zorn’s lemma in the form of open induction. It has been argued that (\[eqn-intro-BBC\]) provides a computational interpretation of countable choice that is superior to standard bar recursion in that it is ‘demand driven’, in the sense that to compute an approximation for the $n$th point in a choice sequence it is not automatically necessary to compute approximations for $1,\ldots,n-1$ first [@BBC(1998.0)].
More recently, a family of new variants of bar recursion known as *products of selection functions* were developed and explored by Escard[ó]{} and Oliva, beginning in [@EscOli(2010.0)]. One of these, the so-called *implicit product of selection functions*, was shown in [@EscOli(2012.0)] to realize not only the negative translation of countable dependent choice but more generally a dependent version of the so called $J$-shift arising from the Pierce translation. The product of selection functions is distinguished by the fact that it incorporates course-of-values recursion into its bar recursive calls. It can be formulated as $$\label{eqn-intro-IPS}\IPS(Y,H,s)=_\NN Y(s\at\lambda n. H(t_n,\lambda x\; . \; \IPS(Y,H,t_n\ast x)))$$ where $t_n$ is the sequence of length $n$ primitive recursively defined by $$(t_n)_i=s_i\mbox{ if $i<|s|$, else $H(t_i,\lambda x\; . \; \IPS(Y,H,t_i\ast x))$ for $i<n$}.$$ In addition to solving the modified realizability interpretation of choice, this form of bar recursion has deep links to game theory as a functional that computes optimal strategies in a class of higher-type, continuously well-founded sequential games [@EscOli(2011.0)]. This provides a highly illuminating bridge between the computational content of the axiom of choice and the world of game theory, and is exploited in e.g. [@OliPow(2012.1)] to give a game-theoretic constructive interpretation of Ramsey’s theorem.
Of course, this list is by no means exhaustive, and several further variants of (\[eqn-intro-MBR\])-(\[eqn-intro-IPS\]) have been devised for realizing choice principles. For example, a realizer of the refined $A$-translation of a seqential variant of dependent choice is given in [@Seisenberger(2008.0)], and is used in [@Seisenberger(2003.0)] to extract a realizer for Higman’s lemma. Similarly, forms of bar recursion closely related to (\[eqn-intro-MBR\]) have been used for realizing dependent choice in a range of settings, including Parigot’s $\lambda\mu$-calculus [@BloRib(2013.0)] and in the context of realizability toposes [@Streicher(2014.0)].
Nevertheless, all of these recursors share two important features in common:
1. they take as input some partial object $t$, and make recursive calls over extensions of this object;
2. they terminate because the value of $Y(\alpha)$ only depends on a finite portion of the input $\alpha$.
In fact, one could go as far as to say that (\[eqn-intro-MBR\])-(\[eqn-intro-IPS\]) along with their many other variants are essentially the same realizer, with the single exception that recursive calls are made using slightly different patterns. In this paper we make this idea precise, and show that the basic recipe used to form the realizers mentioned above can be used to construct, in a completely uniform way, an infinite class of bar recursive functionals, each one of which can be used to interpret a particular set of variants of the axiom of dependent choice.
Note that we restrict our attention to interpretations of analysis based on modified realizability, and in particular bar recursors of the form (\[eqn-intro-SBR\]) which arise from the Dialectica interpretation do not fit into our framework. Nevertheless, we believe that the basic ideas behind this work could be readily lifted to the Dialectica interpretation (see for example the author’s recent article with P. Oliva [@OliPow(2014.0)] on a symmetric form of (\[eqn-intro-SBR\]) which updates partial funtions similarly to the BBC-functional), and may also be helpful in linking proof interpretations to more direct computational interpretations of analysis, such as the learning-based realizabilities of e.g. [@Aschieri(2011.1); @Avigad(2002.0)].
Outline of the paper {#sec-introduction-outline}
--------------------
After setting up our basic formal systems in Section \[sec-formal\], we define in Section \[sec-open\] a general principle of backward induction which will be used throughout the paper in order to prove the correctness of our realizers. Backward induction replaces principles such as bar induction or dependent choice which are more typically used on the meta-level to prove correctness of realizers, and is preferred here due to its far greater flexibility. We also define an analogous schema of backward recursion which will be used to construct our parametrised realizers. This section is strongly influenced by Berger [@Berger(2004.0)], and our formulation of backward induction is similar to but slightly more general than his update induction.
Sections \[sec-countable\] and \[sec-dependent\] form the core of the paper, in which we define our parametrised realizer and use it to realize dependent choice. In Section \[sec-countable\] we restrict ourselves to the simpler double negation shift: While this section’s main result, Theorem \[thm-main-AC\], is eventually subsumed by the later Theorem \[thm-main-DC\], it is instructive to first focus on the double negation shift so that the main ideas of Section \[sec-dependent\] can be appreciated. In the sequel we introduce a family of dependent choice principles parametrised by a well-founded ordering $\lhd$ on the natural numbers, and show that the negative translation of each of these priniciples can be realized by our parametrised bar recursor under certain conditions. Theorem \[thm-main-DC\] gives essentially all the results listed in Section \[sec-introduction-history\] as a Corollary, and vastly generalises all of them.
We conclude on a more informal level, and in Section \[sec-semantics\] discuss a potential semantic interpretation of our parametrised realizer, inspired by the game-theoretic semantic reading of the Berardi-Bezem-Coquand functional in [@BBC(1998.0)].
Preliminaries {#sec-formal}
=============
Throughout this paper we work in variant of extensional Heyting/Peano arithmetic in all finite types, which will be essentially the $\EHAomega$/$\EPAomega$ as defined in e.g. [@Kohlenbach(2008.0); @Troelstra(1973.0)], but based on a slightly richer type system.
The finite types {#sec-formal-types}
----------------
For us, the finite types consist of base types $\N$ and $\B$ for natural numbers and booleans, and are build from the formation of function types $\rho\to\tau$, product types $\rho\times\tau$, and finite sequence types $\rho^\ast$. We sometimes use the abbreviation $\tau^\rho$ for $\rho\to\tau$.
A *discrete* type is any type $\tau$ which can be encoded in $\N$, and so in all standard models $\B$, $\N$, $\B\times \N$, $\N^\ast$ etc. are discrete types but $\N\to \N$ is not.
The theory $\EHAomega$ {#sec-formal-theory}
----------------------
Terms of $\EHAomega$ are typed lambda terms formed by application and abstraction, and include variables for each type, the usual constructors and deconstructors for product and sequence types, the arithmetic constants $0\colon \N$ and $s\colon \N\to \N$, and finally recursors $\R_\rho\colon\rho\to (\N\to\rho\to\rho)\to\N\to\rho$ of each type.
Equations in $\EHAomega$ are formed using basic symbols $=_\B$ and $=_\N$ for equality of type $\B$ and $\N$, while formulas of $\EHAomega$ are built from $=_\N$ and $=_\B$ together with the usual logical connectives and quantifiers for all types. Equality at arbitrary types is defined inductively, so for example $f=_{\rho\to\tau} g$ abbreviates $\forall x^\rho(fx=_\tau gx)$.
The axioms and rules of $\EHAomega$ are the standard axioms and rules of classical logic in all finite types, along with those of the typed lambda calculus, defining equations for all the constants, induction for arbitrary formulas, and finally full extensionality: $$\forall f^{\rho\to\tau}\forall x^\rho,y^\rho(x=_\rho y\to fx=_\tau fy).$$ In what follows we often consider extensions of $\EHAomega$ with new recursively defined functionals $\F$, in which case by $\EHAomega+\F$ we mean the extension of $\EHAomega$ with new constants $\F$ and their associated defining axioms.
We use, throughout, the following notation and abbreviations.
- $0_\rho$ is the inductively defined $0$ term of type $\rho$.
- $\pi_i$ and $\pair{\;,\;}$ denote the projection and pairing operations. For a sequence $\alpha\colon (\rho_0\times\rho_1)^\N$ we sometimes write $\alpha_i\colon\rho^\N$ for $\lambda n.\pi_i(\alpha(n))$.
- For a finite sequence $s\colon\rho^\ast$, $|s|$ denotes the length of $s$, while $s\ast t:=\pair{s_0,\ldots,s_{m-1},t_0,\ldots,t_{n-1}}$ denotes the concatenation $s$ and $t$. We also use $s\ast x$ to denote $s\ast\pair{x}$, and $s\ast\alpha$ for the concatenation of $s$ with the infinite sequence $\alpha$.
- $s\at\alpha\colon\rho^\N$ denotes the overwriting of $\alpha$ with the finite sequence $s$ i.e. $$(s\at\alpha)(n):=\begin{cases}s_n & \mbox{if $n<|s|$}\\ \alpha(n) & \mbox{if $n\geq |s|$}\end{cases}.$$
- $\initSeg{\alpha}{n}:=\pair{\alpha(0),\ldots,\alpha(n-1)}$ denotes the finite initial segment of length $n$ of $\alpha\colon\rho^\N$, while $\ext{s}$ denotes the canonical extension $s\oplus \zero_{\rho^\N}$ of the finite sequence $s\colon\rho^\ast$.
- For a decidable predicate $P(x)$, the term $`y^\rho\mbox{ if $P(x)$}'$ of type $\rho$ is shorthand for $$\begin{cases}y & \mbox{if $P(x)$}\\ 0_\rho & \mbox{otherwise}\end{cases}.$$
- For a decidable predicate $P(n)$ on $\N$, the term $\least i\leq n.P(i)\colon\N$ is the least $i\leq n$ satisfying $P(i)$, and just $n$ if no such $i$ exists.
In addition to all this, throughout the article we work with a type of partial sequences with potentially infinite domain. These can be encoded, for example, by total sequences of type $\N\to\Bool\times\rho$, where defined values are represented as $(1,x)$, and undefined values as $(0,0_\rho)$. Accordingly, for an object $u$ of this type we say that $n$ is in the domain of $u$, or $n\in\dom(u)$, if $\pi_0 u(n)=1$, and $n\notin\dom(u)$ otherwise. Membership of $\dom(u)$ is a decidable property.
We imagine $\Bool\times\rho$ as simulating a type $\ps{\rho}\equiv\rho+\one$, and we write $u(n)=_{\ps{\rho}}x$ instead of $u(n)=_{\Bool\times\rho}(1,x)$, and $u(n)=_{\ps{\rho}}\bot$ whenever $n$ is not in the domain of $u$. Similarly, $u=_{\ps{\rho}} v$ if for all $n$ we have either $\pi_0u(n)=\pi_1u(n)=0$ or $\pi_0u(n),\pi_0v(n)=1$ and $\pi_1u(n)=\pi_1v(n)$.
- We use the symbol $\emptyset$ to denote the partial function with empty domain.
- We extend the overwrite operation $\at$ given for finite sequences above to partial functions by defining $u\at v\colon\ps{\rho}^\N$ by $$\label{eqn-overwrite}(u\at v)(n)=\begin{cases}u(n) & \mbox{if $n\in\dom(u)$}\\ v(n) & \mbox{if $n\notin\dom(u)$}.\end{cases}$$ We define $u\at\alpha\colon\rho^\N$ for total sequence $\alpha^\N$ analogously. It will always be clear from the context which types the operator $\at$ takes as input.
- We isolate as a special case the addition of a single value to $u$: For $n^\N$ and $x^\rho$ we define $\update{u}{n}{x}=u\at (n,x)$ where $(n,x)$ is the partial function taking defined value $x$ at point $n$, and undefined elsewhere. When $n\notin\dom(u)$ we say that $\update{u}{n}{x}$ is an *update* of $u$.
- Finally, we write $u\sqsupseteq v$ whenever $\forall i\in\dom(v)(u(i)=v(i))$. Note that we have $u\at v\sqsupseteq u$ for any $v$, and $u\at v\sqsupset u$ whenever there is some $n\notin\dom(u)$ such that $n\in\dom(v)$. We also have $\update{u}{n}{x}\sqsupset u$ when $n\notin\dom(u)$.
For us, the system $\EHAomega$ together with its classical variant $\EPAomega$ acts as a standard lambda calculus equipped with a robust meta-theory for reasoning about terms. However, the exact details of our formal system are not particularly important as everything which follows can be easily lifted to alternative settings. For example, a slightly different approach would be to work a weaker, quantifier-free term calculus $\sf T$ and to do all the reasoning in an unspecified meta-theory, as in [@BBC(1998.0)]. Alternatively we could work in a theory of partial continuous functionals as in [@Berger(2002.0)], taking our base type to represent the flat domain $\N_{\bot}$.
Models of $\EHAomega$ {#sec-formal-models}
---------------------
In order to prove both the existence and correctness of bar recursive realizers, it is typically necessary to work in a constructive interpretation $T:\equiv (T_\rho)$ of $\EHAomega$ which satisfies some form of the following two properties:
1. $(T_\rho)$ contains *arbitrary* choice sequences, in other words $T_{\N\to\rho}$ contains all sequences $\NN\to T_\rho$ and so in particular $(T_\rho)$ validates dependent choice;
2. Whenever $\tau$ is a discrete type, functionals of type $F\colon\rho^\N\to \tau$ satisfy the following continuity principle: $$\CONT \; \colon \; \forall\alpha^{\rho^\N}\exists n\forall\beta(\initSeg{\alpha}{n}=_{\rho^\ast}\initSeg{\beta}{n}\to F\alpha=_\tau F\beta).$$
Both of these principles are satisfied automatically by the Kleene-Kreisel continuous functionals $\modcont$ [@Kleene(1959.0); @Kreisel(1959.0)], whereas for term models such as the theory $\mathcal{P}$ of [@BBC(1998.0)], (1) is obtained by adding infinite choice sequences explicitly. Here we do not choose any particular intepretation of $\EHAomega$, rather we simply add principles such as $\CONT$ and dependent choice to our meta-theory whenever they are required.
Backward induction and recursion {#sec-open}
================================
We now develop some of the crucial background theory that will be required in order to prove our main results. In particular, we formulate a general principle of backward induction and define an associated backward recursor, both of which will be used to construct and verify the correctness of our parametrised realizer. In simple terms, backward induction is induction over domain-theoretic extensions of partial sequences, and should be seen as a generalisation of bar induction. Analogously, backward recursion is a generalisation of the implicit forms of bar recursion used to give realizability interpretations to countable choice. This section is largely inspired by [@Berger(2004.0)] in that we formulate backward induction as an instance of the still more general principle of open induction.
Open induction {#sec-open-open}
--------------
Open induction, first considered by Raoult in [@Raoult(1988.0)], is an extension of well-founded (or Noetherian) induction to chain-complete partial orders. Recall that a partial order $(X,\leq)$ is chain-complete if every non-empty chain $\gamma$ in $X$ has a least upper bound $\bigvee \gamma$. A predicate $B$ on $X$ is *open* if it satisfies the property $$B(\bigvee \gamma)\to\exists x\in\gamma B(x)$$ for every non-empty chain $\gamma$ in $X$, and the principle of open induction over $X$ is given by $$\OI_{(X,\leq)} \; \colon \; \forall x(\forall y>x B(y)\to B(x))\to\forall x B(x)$$ where $B$ ranges over open predicates. Note that open induction implies well-founded induction since whenever $>$ is well-founded $(X,\leq)$ is trivially chain-complete and all predicates are automatically open. However, in general $>$ need not be well-founded, in which case openness becomes a non-trivial property.
\[thm-Zorn-OI\]Any chain-complete partial order satisfies open induction.
This is a direct consequence of Zorn’s lemma. Suppose that the open predicate $B$ satisfies the premise of open induction, which is classically equivalent to $$\label{eqn-Zorn-OI}\forall x(\neg B(x)\to \exists y>x\neg B(y)),$$ and suppose for contradiction that the set $$S=\{x\in X \; | \; \neg B(x)\}$$ is non-empty. We show that every chain in $S$ has an upper bound in $S$. For the empty chain this is trivial since $S$ contains at least one element. On the other hand, if $\gamma$ is non-empty, then it has an upper bound $\bigvee\gamma$ in $X$ by completeness, and moreover $\bigvee\gamma\in S$ since $\forall x\in\gamma\neg B(x)\to\neg B(\bigvee\gamma)$ by openness of $B$. Therefore by Zorn’s lemma $S$ contains a maximal element, which contradicts (\[eqn-Zorn-OI\]). Thus $S$ must be empty.
Backward induction {#sec-open-backward}
------------------
We are now ready to define backward induction, which we simply take to be open induction over the partial order $(\N\to\ps{\rho},\sqsubseteq)$, where $\sqsubseteq$ is the extension relation on partial sequences defined in Section \[sec-formal-theory\]. In other words, backward induction is the schema $$\backI \ \colon \ \forall u^{\N\to\ps{\rho}}(\forall v\sqsupset u B(v)\to B(u))\to\forall u B(u)$$ where $B$ ranges over open predicates. This formulation of backward induction makes sense and is valid in any interpretation $T$ of $\EPAomega$ that admits arbitrary sequences, since for any chain $\gamma$ in $T_{\N\to\ps{\rho}}$, $$(\bigvee\gamma)(n):=\begin{cases}u(n) & \mbox{if $n\in\dom(u)$ for some $u\in\gamma$}\\ \bot & \mbox{otherwise}.\end{cases}$$ is a perfectly well-defined object of $\NN\to T_{\ps{\rho}}\equiv T_{\N\to\ps{\rho}}$ and is the least upper bound of $\gamma$ with respect to $\sqsubseteq$. It will be convenient to isolate the following syntactic notion of an open formula, which will be sufficient for everything that follows.
\[prop-openp\]Suppose that the formula $B(u)$ on partial sequences is of the form $$B(u):\equiv \forall n[n\in\dom(u)\to A(n,u)]\to \exists nP(\initSeg{u}{n})$$ where $P$ is an arbitrary predicate on $\ps{\rho}^\ast$, and $A$ is monotone in the following sense $$\label{eqn-openp-mon}n\in\dom(u)\wedge u\sqsubseteq v\wedge A(n,u)\to A(n,v).$$ Then $B(u)$ is open with respect to $\sqsubseteq$.
Given some chain $\gamma$ let $v=\bigvee\gamma$ and assume that $B(v)$ holds. We must show that $B(u)$ holds for some $u\in\gamma$. If $B(v)$ holds, then by classical logic we either have $P(\initSeg{v}{n})$ for some $n$, in which case $B(u)$ holds for e.g. the least $u\in\gamma$ satisfying $\initSeg{u}{n}=\initSeg{v}{n}$, or we have $$n\in\dom(v)\wedge\neg A(n,v)$$ for some $n$. In this latter case pick the least $u\in\gamma$ satisfying $n\in \dom(u)$. Then since $u\sqsubseteq v$, we have $A(n,u)\to A(n,v)$ by the monotonicity condition, which is a contradiction. Therefore $\neg A(n,u)$ and $B(u)$ holds.
Note that Berger [@Berger(2004.0)] generally studies formulas which are open with respect to the *lexicographic ordering* on infinite sequences, which are not necessarily open with respect to $\sqsubseteq$ (which is strictly contained in a lexicographic ordering). However, any formula equivalent to one of the form $B(u):\equiv \exists n P(\initSeg{u}{n})$ is open in both senses.
Some additional remarks on backward induction {#sec-open-backward}
---------------------------------------------
We take the opportunity to explore backward induction in more detail and relate it to other well-known principles in logic, including the minimal bad sequence argument and bar induction. This section is not strictly necessary for the remainder of the paper, so if the reader prefers they can proceed directly to the definition of backward recursion given in Setion \[sec-open-recursion\].
We first point out that, analogously to [@Berger(2004.0)], backward induction does not require the full strength of Zorn’s lemma, and is provable from just dependent choice, using a version of the minimal-bad-sequence argument due to Nash-Williams [@NashWilliams(1963.0)].
The principle of backward induction is provable in $\EPAomega+\DC$.
Take some open formula $B(u)$, which for simplicity we assume is open in the sense of Proposition \[prop-openp\]. Suppose for contradiction that we have $\forall u(\neg B(u)\to\exists v\sqsupset u\neg B(v))$ but there exists some partial sequence $u_0$ such that $\neg B(u_0)$. Using dependent choice construct the sequence $(u_n)$ as follows: Supposing that we have already constructed $\pair{u_0,\ldots,u_n}$ for $n\geq 0$, define
(i) \[item-MBSi\] $u_{n+1}:=w$ if $n\notin\dom(u_n)$ and $w$ is such that the following four properties are satisfied: $\neg B(w)$; $w\sqsupset u_n$; $\initSeg{w}{n}=\initSeg{u_n}{n}$ and $n\in\dom(w)$,
(ii) \[item-MBSii\] $u_{n+1}:=u_n$ if either $n\in\dom(u_n)$ or $n\notin\dom(u_n)$ and no $w$ in the sense of (\[item-MBSi\]) exists.
First, it is clear by a simple induction that for all $n$ we have
(a) \[item-MBSpi\] $\neg B(u_n)$,
(b) \[item-MBSpii\]$\initSeg{u_n}{n}=\initSeg{u_{n+1}}{n}$, and
(c) \[item-MBSpiii\]$u_n\sqsubseteq u_{n+1}$. Define $\tilde u:=\lambda n.u_{n+1}(n)$. Then it follows that for all $n$ we have \[[\[(a)\]]{}\]
(d) \[item-MBSpiv\] $\initSeg{\tilde u}{n}=\initSeg{u_n}{n}$ and
(e) \[item-MBSpv\] $u_n\sqsubseteq \tilde u$.
The first of these is done by a simple induction using (\[item-MBSpii\]). For the latter, take $i\in\dom(u_n)$. Then either $i< n$ in which it is clear by (\[item-MBSpv\]) that $\tilde u(i)=u_n(i)$, or $i\geq n$ and we obtain $\tilde u(i)=u_{i+1}(i)=u_n(i)$ by $u_n\sqsubseteq\ldots\sqsubseteq u_{i+1}$.
Now we prove that $\neg B(\tilde u)$, which is classically equivalent to $$\forall i([i\in\dom(u)\to A(i,u)]\wedge\neg P(\initSeg{u}{i})).$$ Taking some arbitrary $n$, and setting $u=u_{n+1}$ and $i=n$ we get, by $\neg B(u_{n+1})$ (true by (\[item-MBSpi\])), $$(\ast) \ \ [n\in\dom(u_{n+1})\to A(n,u_{n+1})]\wedge \neg P(\initSeg{u_{n+1}}{n}).$$ But $\neg P(\initSeg{u_{n+1}}{n})\to\neg P(\initSeg{\tilde u}{n})$ by (\[item-MBSpii\]) and (\[item-MBSpiv\]), and furthermore $n\in\dom(\tilde u)$ is equivalent to $n\in\dom(u_{n+1})$ and hence by $(\ast)$ implies $A(n,u_{n+1})$, so using monotonicity of $A$ and the fact that $u_{n+1}\sqsubseteq \tilde u$ we have $A(n,\tilde u)$ and have therefore established $$n\in\dom(\tilde u)\to A(n,\tilde u).$$ Taken together and bearing in mind that $n$ is arbitrary, this implies $\neg B(\tilde u)$. But now we know by the backward induction hypothesis that there exists some $v\sqsupset\tilde u$ such that $\neg B(v)$ holds.
We can show that this contradicts the construction of $(u_n)$, and therefore there cannot exists any initial sequence $u_0$ satisfying $\neg B(u_0)$, and we’re done. Let $m$ be the least point such that $m\in\dom(v)$ but $m\notin\dom(\tilde u)$. Then firstly by (\[item-MBSpv\]) we have $u_m\sqsubseteq\tilde u\sqsubset v$ and thus $u_m\sqsubset v$, and secondly $\initSeg{v}{m}=\initSeg{\tilde u}{m}=\initSeg{u_m}{m}$, the first equality by minimality of $m$ and the second by (\[item-MBSpiv\]), and therefore $v$ satisfies the required properties of $w$ in (\[item-MBSi\]) at point $m$. In addition we know that $m\notin\dom(u_{m})$ else we’d have $m\in\dom(\tilde u)$ by (\[item-MBSpv\]), and therefore $u_{m+1}$ must be constructed using (\[item-MBSi\]) and thus $m\in\dom(u_{m+1})\subseteq \dom(\tilde u)$, contradicting the assumption that $m\notin\dom(\tilde u)$.
As we will see in Sections \[sec-countable\] and \[sec-dependent\], one of the key ideas in this paper is construct forms of recursion based on restricted, or relativised variants of backward recursion which take as input partial functions that are downward closed with respect to some relation on $\N$. These will be closely related to the following form of relativised backward induction.
Let $\lhd$ be some decidable relation on $\N$, and define the predicate $u\in D_\lhd$ by $$u\in D_\lhd:\equiv\forall n\in\dom(u)[\forall i\lhd n(i\in\dom(u))].$$ Equivalently, we say that $\dom(u)$ is $\lhd$-closed. Then for any relation $\sqsubset'$ on $\ps{\rho}^\N$ such that $u\sqsubset' v\to u\sqsubset v$, the following principle of relativised backward induction is provable from $\backI$: $$\forall u\in D_\lhd(\forall v\sqsupset' u [v\in D_\lhd\to B(v)]\to B(u))\to\forall u\in D_\lhd \; B(u).$$
First note that $u\in D_\lhd$ is of the form $\forall n\in\dom(u) D_0(n,u)$ with $D_0(n,u)$ monotone in the sense of (\[eqn-openp-mon\]), therefore the predicate $B'(u):\equiv u\in D_\lhd\to B(u)$ is open for any open $B(u)$. Thus we obtain $$\forall u(\forall v\sqsupset' u B'(v)\to B'(u))\to \forall u(\forall v\sqsupset u B'(v)\to B'(u))\to\forall u B'(u)$$ the first implication following from the inclusion $\sqsubset'\subseteq\sqsubset$ and the second from normal backward induction applied to $B'(u)$. Rearranging this gives us relativised backward induction.
We can now instantiate $\lhd$ and $\sqsubset'$ to obtain certain well-known instances of backward induction.
If we define $u\sqsubset' v$ iff $v$ is an update of $u$, and let $\lhd$ just be the empty relation, then relativised backward induction just becomes update induction in the sense of [@Berger(2004.0)].
\[ex-bar-ind\]Now consider the case $\lhd=<$. By classical logic, if $u\in D_<$ then either $u=\alpha$ for some total sequence $\alpha$ or $u=\ext{s}$ where $s$ is some finite sequence and $\ext{s}$ its embedding as a partial function. Therefore relativised backward induction is equivalent in $\EPAomega$ to $$\label{eqn-bar-ind0}\forall \alpha B(\alpha)\wedge \forall s(\forall v\sqsupset' \ext{s}[v\in D_<\to B(v)]\to B(\ext{s}))\to\forall s B(\ext{s})$$ In addition, if we define $u\sqsubset' v$ iff $v$ is an update of $u$, then (\[eqn-bar-ind0\]) becomes equivalent to $$\label{eqn-bar-ind1}\forall \alpha B(\alpha)\wedge \forall s(\forall x B(\ext{s\ast x})\to B(\ext{s}))\to\forall s B(\ext{s})$$ which is just a variant of bar induction. It is not too difficult to show that this is equivalent over $\EPAomega$ to the more standard formulations of bar induction found in e.g. [@Troelstra(1973.0)].
Backward recursion {#sec-open-recursion}
------------------
The purpose of introducing backward induction was to give us a way to reason about backward recursion, which we define and discuss in this section. In the same way that backward induction is a special case of open induction, backward recursion is closely related (and in fact definable from) open recursion as defined in [@Berger(2004.0)].
To begin with, in order to motivate what follows let us consider as a comparison the entirely standard concept of well-founded recursion over some decidable well-founded relation $\prec$ on $\rho$. Assuming we are working in a structure such as PCF or the Scott continuous functionals, we can define a well-founded recusor $\wR_{\prec}$ as the fixpoint of the following recursive equation $$\wR_{\prec}^\psi(x)=_\sigma \psi_x(\lambda y\; . \; \wR_{\prec}^g(y)\mbox{ if $y\prec x$}),$$ and prove that the recursor defines a *total* functional for any outcome type $\sigma$ using well-founded induction over $\prec$: $$\wI \; \colon \; \forall x(\forall y\prec x A(y)\to A(x))\to\forall x A(x).$$ We want to define a backward recursor in a similar way - although we have two problems: firstly the relation $u\sqsubset v$ is not decidable, and secondly backward induction is only valid for open formulas. We avoid these issues by defining backward recursion to be the fixpoint of the following recursive equation $$\BR_{\rho,\tau}^\psi(u)=_\tau\psi_u(\lambda n,v\;. \; \BR^\psi(u\at v)\mbox{ if $n\in\dom(v)\backslash\dom(u)$}),$$ where $\tau$ is restricted to being a discrete type, while $n\in\dom(v)\backslash\dom(u)$ denotes the decidable predicate $n\in\dom(v)\wedge n\notin\dom(u)$. Observe that any $w\sqsupset u$ is of the form $w=u\at v$ for some $v$ which is defined at at least one point $n\notin\dom(u)$, and so $\BR$ makes recursive calls over all $w\sqsupset u$, although crucially it must always have access to a point $n\in\dom(w)$ such that $n\notin\dom(u)$. The necessity of the restriction on $\tau$ is to ensure that totality of $\BR^\psi(u)$ is an open property on total input $u$ (here we are referring to total elements of the *model* of partial continuous functionals as opposed to the type $\ps{\rho}$). Indeed for discrete $\tau$ and total $u$ we have $$\BR^\psi(u)\mbox{ is total}\leftrightarrow\exists n\forall w(\BR^\psi(\initSeg{u}{n}\at w)\mbox{ is total})$$ assuming sequential continuity $\CONT$ for functionals with total output. Therefore totality of $\BR$ is provable using backward induction. Note that alternatively, a direct proof via Zorn’s lemma that $\BR$ exists as a total element of the Scott partial continuous functionals can be carried out using the same manner as the proof of totality of the Berardi-Bezem-Coquand functional in [@Berger(2002.0)].
Alternatively, one can justify the existence of backward recursion in continuous models by showing that backward recursion is definable from the slightly more general schema of open recursion on the lexicographic ordering considered in [@Berger(2004.0)]. Open recursion is defined to be the fixpoint of the following recursive equation: $$\OR_{\rho,\tau}^{\psi}(u)=_\tau \psi_u(\lambda n,v\; . \; \OR^\psi(\initSeg{u}{n}\at v)\mbox{ if $n\in\dom(v)\backslash\dom(u)$})$$ where once again $\tau$ is discrete, and lexicographically open recursive functionals of the above form are shown to be total in [@Berger(2004.0) Proposition 5.1].
\[def-open-back\]$\BR$ is instance-wise primitive recursively definable from $\OR$.
Primitive recursively define $$m_{n,u,v}:=\mbox{least $i\leq n$ s.t. $i\in\dom(v)\backslash\dom(u)$, else $n$},$$ and set $\BR_{\rho,\tau}^\psi(u)=\OR_{\rho,\tau}^{\tilde\psi}(u)$ where $$\tilde\psi_u(f^{\N\times\ps{\rho}^\N\to\tau}):=\psi_u(\lambda n,v\; . \; f(m_{n,u,v},u\at v)\mbox{ if $n\in\dom(v)\backslash\dom(u)$}).$$ Then expanding definitions we have $$\begin{aligned}\BR^\psi(u)&=\tilde\psi_u(\lambda n,v\; . \; \BR^{\tilde\psi}(\initSeg{u}{n}\at v)\mbox{ if $n\in\dom(v)\backslash\dom(u)$})\\
&\stackrel{(a)}{=}\psi_u(\lambda n,v\; . \; \BR(\initSeg{u}{m_{n,u,v}}\at (u\at v))\mbox{ if $n\in\dom(v)\backslash\dom(u)$})\\
&\stackrel{(b)}{=}\psi_u(\lambda n,v\; . \; \BR(u\at v)\mbox{ if $n\in\dom(v)\backslash\dom(u)$})\end{aligned}$$ where for $(a)$ we use $n\in\dom(v)\backslash\dom(u)\to m_{n,u,v}\in\dom(u\at v)\backslash \dom(u)$, and $(b)$ follows by minimality of $m_{n,u,v}$.
\[ex-update-rec\]It is a fairly easy observation that update recursion as defined in [@Berger(2004.0)] - $$\UR^H(u)=H_u(\lambda n,x^\rho\; . \; \UR^H(\update{u}{n}{x})\mbox{ if $n\notin\dom(u)$}),$$ is a simple instance of backward recursion obtained by setting $$\psi_u(f^{\N\times\ps{\rho}^\N\to\tau}):=H_u(\lambda n,x\; . \; f(n,\update{u}{n}{x})\mbox{ if $n\notin\dom(u)$}).$$
\[ex-bar-rec\]For those readers interested in the computability theory of bar recursion, it might be instructive to pause for a moment to consider a natural instance of bar recursion that arises from backward recursion. Let us define $\IBR(H,s^{\rho^\ast}):=\BR^\psi(\ext{s})$ where $$\psi_u(f^{\N\times\ps{\rho}^\N\to\tau}):=_{\tau} H(u,\lambda t\; . \; f(|t|-1,\ext{t})).$$ Then it is not too hard to show that $\IBR$ satisfies $$\IBR(H,s)=H(\ext{s},\lambda t\;.\;\IBR(H,s\at t)\mbox{ if $|t|>|s|$}),$$ and this can be viewed as a ‘implicitly well-founded’ variant of Spector’s bar recursion (\[eqn-intro-SBR\]). The reason we highlight this is that while several such implicit variants of Spector’s so-called ‘special’ instance of bar recursion[^3] have been studied, including both modified bar recursion and the implicit product of selection functions, constructing a direct analogue to the general form is more complicated (for example, an implicit form of the so-called product of quantifiers is known not to exist [@EscOli(2011.0)]).
The subtle reason for this is that such variants of bar recursion must not be allowed to access the length of the input sequence $s$. For example, no object $\Phi$ can satisfy the slightly altered equation $$\Phi(H,s)=H(s,\lambda t\;.\;\Phi(H,s\at t)\mbox{ if $|t|>|s|$})$$ even for discrete output type, since we could just take $\tau=\N$ and define $H(s,g):=1+f(s\ast 0)$, and then $\Phi(\pair{})=n+1+\Phi(H,\initSeg{0}{n+1})>n$ for all $n$, which cannot hold in any model of arithmetic. Indeed, trying to define this from $\BR$ with continuous $\psi_u$ is impossible, since we’d require a non-continuous unbounded search (and thus totality of the underlying instance of $\BR$ would no longer be an open predicate). Thus we overcome the difficulty with implicit variants of bar recursion by removing access to the length of the input. Note that this problem is not a feature of modified bar recursion and implicit products of selection functions (or indeed any of the realizers we define in the following sections), since these are defined ‘pointwise’, and make recursive calls only when we are accessing points already greater than the length of the input sequence.
A computational interpretation of the double negation shift {#sec-countable}
===========================================================
Now that we have completed the mathematical groundwork we come to the core of the paper. In this section we give a new, general realizability interpretation to the double negation shift. Ultimately, this will form a special case of the interpretation of full dependent choice given in the next section. However, by focusing first on the double negation shift we have an opportunity to present our main ideas in a slightly simplified setting, then the extension to full dependent choice will mostly be a matter of taking care of some additional technical details.
Modified realizability intepretation of extensions of $\PAomega$ {#sec-countable-mr}
----------------------------------------------------------------
We begin by very briefly recalling how Kreisel’s modified realizability can be used in conjunction with the so-called Friedman trick to extract programs from classical proofs of $\Pi^0_2$-formulas. This is all completely standard, so we omit most of the details. For every formula in the language of $\HAomega$ the realizability relation $x\mr A$ is inductively defined by $$\begin{aligned}()\mr A &\equiv A\mbox{ if $A$ is an atomic formula},\\
x,y\mr (A\wedge B)&\equiv x\mr A\wedge y\mr B,\\
n^\N,x,y\mr (A\vee B)&\equiv (n=0\to x\mr A)\wedge (n\neq 0\to y\mr B),\\
f\mr (A\to B)&\equiv \forall x(x\mr A\to fx\mr B),\\
x\mr \forall zA(z)&\equiv\forall z(xz\mr A(z)),\\
x,y\mr\exists zA(x)&\equiv y\mr A(x).\end{aligned}$$ It is well-known that whenever $\HAomega\vdash A$ then $\HAomega\vdash t\mr A$ where $t$ is some primitive recursive term extracted from the proof of $A$. The interpretation of classical logic, on the other hand, is more subtle. A simple combination of the negative translation with modified realizability fails to work since the atomic formula $\bot$ is realized by $()$ and therefore all negated formulas are trivially interpreted. In particular, this method gives us no way of extracting realizers from $\Pi^0_2$-formulas $\forall x^\N\exists y^\N A(x,y)$.
One well established way of overcoming this problem is to slightly alter the definition of modified realizability by regarding $x^\N\mr\bot$ as an uninterpreted formula. Then, as discussed in e.g. [@BergOli(2005.0); @BergSch(1995.0)], from a classical derivation $\PAomega\vdash\forall y^\N\exists x^\N A(y,x)$ one can extract a term $t$ such that $\HAomega\vdash\forall y A(y,ty)$, utilising the aforementioned Friedman trick of replacing the formula $x\mr \bot$ by the quantifier-free formula $A(y,x)$. This idea can be smoothly expanded to extensions of $\PAomega$ with some additional axiom(s) $\Gamma$. Provided that $\HAomega+\Delta\vdash\Phi\mr \Gamma^{\negt}$, where $\Gamma^\negt$ denotes the negative translation of $\Gamma$, $\Delta$ is some set of axioms satisfying some natural closure properties with respect to $\bot$, and $\Phi$ is some closed term in the language of $\HAomega+\Delta$, then from a classical proof $\PAomega+\Gamma\vdash \forall y\exists x^\N A(y,x)$ one can extract a term $t$ in $\Phi$ such that $\HAomega+\Delta\vdash \forall y A(y,ty)$.
Thus we have a method that allows us to extract realizers for $\Pi^0_2$ formulas from any extension $\Gamma$ of Peano arithmetic whenever we can realize the negative interpretation $\Gamma^{\negt}$ of $\Gamma$. In the remainder of this paper we develop this idea and focus on constructing terms $\Phi$ such that $\Phi\mr\Gamma^{\negt}$ where $\Gamma$ is either countable or countable dependent choice, and $\Gamma^{\negt}$ is the adapted realizability interpretation which treats $x\mr\bot$ as a new predicate in $x$. In fact, following [@EscOli(2012.0)] we generalise slightly and replace $\bot$ by some arbitrary formula $R$ whose type of realizers is a discrete type, emphasising the fact that $\bot$ acts as some undefined object to be realized. However, if the reader prefers they can just treat this as a relabelling and imagine $R=\bot$ throughout.
The $J$-shift and its variants {#sec-countable-real}
------------------------------
The axiom of countable choice is given by $$\AC \; \colon \; \forall n\exists x^\rho B_n(x)\to\exists \alpha^{\N\to\rho}\forall n B_n(\alpha(n)).$$ It is well-known that the negative translation of $\AC$, $$\forall n((\exists x B^{\negt}_n(x)\to R)\to R)\to (\exists\alpha\forall n B^{\negt}_n(\alpha(n))\to R)\to R,$$ (here with an arbitrary discretely-realized $R$ in place of $\bot$) is provable using the (trivially realized) intuitionistic axiom of choice from the simpler double-negation shift, $$\DNS \; \colon \; \forall n((A(n)\to R)\to R)\to (\forall n A(n)\to R)\to R,$$ by setting $A(n):=\exists x B^{\negt}_n(x)$. Thus a realizability interpretation of countable choice follows directly from that of $\DNS$. Note that this version of $\DNS$ for arbitrary $R$ is also called the $K$-shift in [@EscOli(2012.0)].
In order to successfully realize $\DNS$ one typically relies on a term $h$ realizing ex-falso-quodlibet in the form $\forall n(R\to A(n))$, and so in practice one must work with a restricted class of formulas $A(n)$ that admit such a realizer, such as any formula in the image of the negative translation (for which one can trivially construct such a $h$ even uniformly in $n$). This need for additional realizers and a corresonding restriction on formulas can seem slightly inelegant, and so a reformulation of $\DNS$ is given in [@EscOli(2012.0)] which, rather than separately assuming $R\to A(n)$, adds this positive information directly to the premise of $\DNS$, yielding $$\forall n((A(n)\to R)\to A(n))\to (\forall n A(n)\to R)\to R.$$ In [@EscOli(2012.0)] this is actually written in a equivalent form called the $J$-shift: $$\mbox{$J$-shift} \; \colon \; \forall n((A(n)\to R)\to A(n))\to (\forall n A(n)\to R)\to \forall nA(n),$$ and this is given a realizability interpretation using the product of selection functions, an interpretation which is valid for *arbitrary* formulas $A(n)$. Then, in the case that $R\to A(n)$ is realizable, one easily reobtains an interpretation of the double negation shift:
$J$-shift implies $\DNS$ over minimal logic, whenever $R\to A(n)$ holds.
Here, we adopt the convention of [@EscOli(2012.0)] in adding the positive information we need directly to the premise of the double negation shift, so that our interpretation is valid for all $A(n)$, and as in [@EscOli(2012.0)], and as we show in our examples in Section \[sec-dependent-examples\], we can always convert our realizer to one of $\DNS$ for negated formulas $A(n)$, which in turn is sufficient to realize the axiom of countable choice.
For notational reasons we interpret a pair of syntactically more flexible variants of the $J$-shift, designed to match the family of realizers we construct.
We define the $J^\ast_i$-shifts for $i=1,2$ by $$\begin{aligned}\mbox{$J^\ast_1$-shift} \; &\colon \; \forall m,n((A(m)\to R)\to A(n))\to (\forall n A(n)\to R)\to R \\
\mbox{$J^\ast_2$-shift} \; &\colon \; \forall m,n((A(m)\to R)\to A(n))\to (\forall n A(n)\to R)\to \forall nA(n)\end{aligned}$$ where $R$ has discrete realizing type.
The following result confirms that our $J^\ast_i$-shift principles are nothing more than simple rephrasings of the original $J$-shift.
$\mbox{$J^\ast_1$-shift}\Leftrightarrow \mbox{$J^\ast_2$-shift}\Leftrightarrow\mbox{$J$-shift}$ over minimal logic.
$\mbox{$J^\ast_1$-shift}\Rightarrow \mbox{$J^\ast_2$-shift}$ follows from the observation that $$\forall m,n((A(m)\to R)\to A(n))\to (R\to\forall nA(n)),$$ which is true because for arbitrary $n$ we have $$R\to (A(n)\to R)$$ and thus $$((A(n)\to R)\to A(n))\to (R\to A(n)).$$ The implication $\mbox{$J^\ast_2$-shift}\Rightarrow \mbox{$J$-shift}$ follows from $$\forall n((A(n)\to R)\to A(n))\to \forall m,n((A(m)\to R)\to A(n)),$$ which is true because firstly $$((A(m)\to R)\to A(m))\to ((A(m)\to R)\to R),$$ and since $R\to (A(n)\to R)$ this yields $$((A(m)\to R)\to A(m))\to ((A(m)\to R)\to (A(n)\to R))$$ and so finally $$((A(n)\to R)\to A(n))\wedge((A(m)\to R)\to A(m))\to ((A(m)\to R)\to A(n))$$ The remaining direction $\mbox{$J$-shift}\Rightarrow \mbox{$J^\ast_1$-shift}$ is straightforward.
The reasons that we highlight these variants of the $J$-shift is that we want versions of the $J$-shift that are directly realized by our variants of bar recursion. As we will see, in most cases each variant will only use the premise of $J^\ast_i$-shift for $(m,n)\in I\subseteq\N\times\N$ for some $I$.
Realizing the $J_i^\ast$-shift {#sec-countable-real}
------------------------------
We focus on constructing a realizer for the $J_1^\ast$-shift, then a realizer of the $J^\ast_2$-shift comes out immediately Suppose that the realizing types of $A(n)$ and $R$ are $\rho$ and $\tau$ respectively, where we assume that $\tau$ is discrete. The $J^\ast_1$-shift is realized by a term $\Phi$ of type $(\N\to \N\to (\rho\to\tau)\to\rho)\to (\rho^\N\to\tau)\to\tau$, which, given terms $\varepsilon\colon\N\to\N\to (\rho\to\tau)\to\rho$ and $q\colon\rho^\N\to\tau$ that satisfy $$\label{eqn-varepsilon-q}\begin{aligned}&\forall m,n,p^{\rho\to\tau}(\forall x^\rho(x \mr A(m)\to p(x)\mr R)\to\varepsilon_{m,n}(p)\mr A (n))\\
&\forall\alpha^{\rho^\N}(\forall n(\alpha(n)\mr A(n))\to q(\alpha)\mr R)\end{aligned}$$ returns a term $\Phi\varepsilon q\colon\tau$ satisfying $\Phi\varepsilon q\mr R$.
The basic idea that unites all such existing realizers of double-negation shift principles is to form an auxiliary functional $\Psi$ which performs a backward recursive loop, which builds increasingly large partial realizers of $\forall nA(n)$.
More precisely, suppose that $\varepsilon$ and $q$ satisfy the premise of the $J_1^\ast$ shift as above, and imagine we are given a partial realizer $u\colon \ps{\rho}^\N$ which satisfies $\forall n\in\dom(u)(u(n)\mr A(n))$. Then let us somewhat informally define $\Psi^{\varepsilon,q}u:=q(u\at c(u))$, where $c$ is some as yet unspecified function on partial sequences, but the aim is that it forms a completion of the partial realizer of $u$ and that therefore $\Psi^{\varepsilon,q}u\mr R$ for all $u$. This completion $c(u)$ can be constructed by backward recursion using $\varepsilon$. For $n\notin\dom(u)$ we define $$c(u)(n):=\varepsilon_{mnu,n}(\lambda x\; . \; \Psi^{\varepsilon ,q}(\update{vnu}{m}{x}))$$ for some index $mnu$ and partial realizer $vnu$ of $\forall nA(n)$. Again, both $mnu$ and $vnu$ are left unspecified for now, but note that if at the very least
(i) $vnu$ is a partial realizer of $\forall nA(n)$ satisfying $u\sqsubseteq vnu$ and
(ii) $mnu\notin\dom(u)$ whenever $n\notin\dom(u)$,
then $u\sqsubset \update{vnu}{mnu}{x}$ and $\update{vnu}{mnu}{x}$ is a partial realizer whenever $x\mr A(mnu)$. Therefore $\lambda x.\Psi^{\varepsilon,q}(\update{vnu}{mnu}{x})$ is a realizer of $A(mnu)\to R$ under the assumption that $\Psi$ is correct for all *extensions* of $u$. The idea now is that we can admit this assumption as a backward induction hypothesis, and so by backward induction we can prove that $\Psi^{\varepsilon,q}u\mr R$ for all $u$. Then setting $\Phi\varepsilon q:=\Psi^{\varepsilon,q}\emptyset$ gives us a realizer for the $J^\ast_1$-shift, since $\emptyset$ is trivially a partial realizer of $\forall n A(n)$.
What remains is to formalise this idea and make some sensible choice of $mnu$ and $vnu$ satisfying (i) and (ii) above. The most natural might be to set $mnu=n$ and $vnu=u$ - in this case, as we will see below, is precisely the idea behind the Berardi-Bezem-Coquand functional of [@BBC(1998.0)]. However, more intricate choices lead to other realizers, including modified bar recursion and the product of selection functions. Our aim now is to make this intuition precise, and provide a sufficiently rich general construction of $m$ and $v$ which captures all of these realizers and much more.
\[prop-main-AC\]Suppose that $\rho$ and $\tau$ are the realizing types of $A(n)$ and $R$, with $\tau$ discrete, and that were are given a computable relation $\prec\colon\ps{\rho}^\N\to (\N\times\N\to\Bool)$ and an index $m\colon\N\times \ps{\rho}^\N\to\N$ such that
(i) \[item-AC-conda\] $\prec_{u}$ is well-founded,
(ii) \[item-AC-condb\] $n\notin\dom(u)\to m{nu}\notin\dom(u)\cup\{k\; | \; k\prec_u n\}$
for all $u$ and $n$. Then there is a term $\Psi_{(\prec,m)}^{\varepsilon ,q}\colon\ps{\rho}^\N\to\tau$ with parameters of type $\varepsilon\colon\N\to\N\to (\rho\to\tau)\to\rho$ and $q\colon\rho^\N\to\tau$ which is primitive recursively definable in $\BR+\lambda u.\wR_{\prec_{u}}$, and satisfies the recursive equation $$\Psi^{\varepsilon,q}(u)=q(\underbrace{u\at\lambda n\; . \; \varepsilon_{m{nu},n}(\lambda x\;. \; \Psi^{\varepsilon,q}(u\at\update{\initSegsss{\alpha_u}{n}}{m{nu}}{x})}_{\alpha_u})),$$ where $\alpha_u\colon\rho^\NN$ denotes the argument of $q$ as indicated above, and $\initSegsss{\alpha_u}{n}\colon\ps{\rho}^\NN$ is defined to be $$\lambda k.(\alpha_u(k)\mbox{ if $k\prec_{u} n$ else $\bot$}).$$
\[rem-wR\]Technically speaking the term $\lambda u.\wR_{\prec_{u}}$ is not properly defined in $\EHAomega$ - we are simply assuming here that there exists a function $F$ definable in $\EHAomega$ such that for all $u$, $F(u)$ satisfies the defining equation of $\wR_{\prec_{u}}$ i.e. well-founded recursion over $\prec_u$ as considered in Section \[sec-open-recursion\]. However, in all of the concrete examples we consider, $\prec$ will not depend on $u$ and $\wR_{\prec}$ will always be trivially definable in $\EHAomega$, so this rather casual definition will not be problematic.
Proposition \[prop-main-AC\] above is a special case of Proposition \[prop-main-DC\] in the next section, whose proof can be found in the appendix, and so we omit a proof of Proposition \[prop-main-AC\] here. However, on an informal level, it is not too difficult to see that $\Psi$ is well-defined. First we note that the definition of $\initSegsss{\alpha_u}{n}$ is not circular, since it is only used to define $\alpha_u(n)$, and thus $\alpha_u$ as a whole is constructed using the well-founded recursor $\wR_{\prec_u}$. Then one observes that the whole expression is a well-defined backward recursive functional since to compute $\Psi(u)$ we only call $\Psi$ on arguments of the form $u\at \update{\initSegsss{\alpha_u}{n}}{m}{x}$ for $m\notin\dom(u)$ by condition (\[item-AC-condb\]), which are always strict extensions of $u$.
\[thm-main-AC\]Suppose that $\Psi_{(\prec,m)}$ is defined as in Proposition \[prop-main-AC\] for $\prec$ and $m$ satisfying conditions (\[item-AC-conda\]) and (\[item-AC-condb\]). Then the term $\Phi_{(\prec,m)}:=\lambda \varepsilon,q\; . \; \Psi_{(\prec,m)}^{\varepsilon,q}(\emptyset)$ realizes the $J^\ast_1$-shift, provably in $\EHAomega+\CONT+\backI+\BR+(\wI_{\prec_{u}})+\lambda u.\wR_{\prec_u}$.
\[rem-wI\]Here $(\wI_{\prec_{u}})$ denotes the collection of well-founded induction schemata over the well-founded relations $\prec_u$. As with the corresponding modes of recursion discussed in Remark \[rem-wR\], in practise $\prec_u$ will typically not depend on $u$ and $\wI_{\prec}$ will be easily provable in $\EHAomega$.
Assume that $\varepsilon$ and $q$ realize the premise of the $J^\ast_1$-shift. We prove that $\Phi_{(\prec,m)}\varepsilon q\mr R$ using a main backward induction and an auxiliary well-founded induction. Let us define $$B(u):\equiv \forall n\in\dom(u)(u(n)\mr A(n))\to\Psi^{\varepsilon,q}(u)\mr R.$$ This is an open formula in the sense of Proposition \[prop-openp\] since by $\CONT$ we have $$\Psi(u)\mr R\leftrightarrow \exists n\forall w\Psi(\initSeg{u}{n}\at w)\mr R.$$ Now, to prove the backward induction step for $B$, assume that $u$ is a partial realizer of $\forall n A(n)$ (i.e. the premise of $B(u)$ holds) and suppose that $B(v)$ holds for all $v\sqsupset u$. We want to show that $\Psi^{\varepsilon,q}(u)\mr R$.
We do this by first proving using $\wI_{\prec_u}$ that $\alpha_u$ (as defined in Proposition \[prop-main-AC\]) realizes $\forall n A(n)$. Fix $n$ and assume as an auxiliary induction hypothesis that $\alpha_u(k)\mr A(k)$ for all $k\prec_u n$. If $n\in\dom(u)$ we trivially have $\alpha_u(n)=u(n)\mr A(n)$, so assume that $n\notin\dom(u)$. In this case, first observe that $$x\mr A(mnu)\to \forall i\in\dom(v)(v(i)\mr A(i))\mbox{\; \; and \; \;}\forall i\in\dom(v)(v(i)\mr A(i)) \to \Psi^{\varepsilon,q}(v)\mr R$$ for $v:=u\at\update{\initSegsss{\alpha_u}{n}}{mnu}{x}$. The first step is clear by the auxiliary induction hypothesis, which implies that $u\at\initSegsss{\alpha_u}{n}$ is a partial realizer of $\forall nA(n)$, while the second step follows from the main hypothesis $B(v)$, since by condition (\[item-AC-condb\]) we know that $n\notin\dom(u)$ implies that $mnu\notin\dom(u)$ and thus $v\sqsupset u$. Putting this together we see that $\lambda x.\Psi(u\at\update{\initSegsss{\alpha_u}{n}}{mnu}{x})\mr (A(mnu)\to R)$ and thus $$\alpha_u(n)=\varepsilon_{mnu,n}(\lambda x.\Psi(u\at\update{\initSegsss{\alpha_u}{n}}{mnu}{x}))\mr A(n).$$ by correctness of $\varepsilon$. This completes the auxiliary well-founded induction, giving us $\forall n(\alpha_u(n)\mr A(n))$, and therefore $\Psi(u)=q(\alpha_u)\mr R$, which completes the main backward induction step. Finally, then, we obtain $\forall u B(u)$ by $\backI$, and so in particular by $B(\emptyset)$ we have $\Psi^{\varepsilon,q}(\emptyset)\mr R$, which completes the proof.
\[cor-main-AC\]There is a term $\tilde\Phi_{({\prec},{m})}$ primitive recursive in $\BR+\lambda u.\wR_{\prec_{u}}$ which realizes the $J^\ast_2$-shift, provably in $\EHAomega+\CONT+\backI+\BR+(\wI_{\prec_{u}})+\lambda u.\wR_{\prec_u}$.
Keeping all the notation of Theorem \[thm-main-AC\], define $\tilde\Psi^{\varepsilon,q}(u)=\alpha_u$ so that it satisfies the recursive equation $$\tilde\Psi^{\varepsilon,q}(u)=u\at\lambda n\; . \; \varepsilon_{m{nu},n}(\lambda x\;. \; q(\tilde\Psi^{\varepsilon,q}(u\at\update{\initSegsss{\alpha_u}{mnu}}{m{nu}}{x}))).$$ Define $\tilde\Phi\varepsilon q:=\tilde\Psi^{\varepsilon,q}(\emptyset)$. Then it follows immediately from $\forall u B(u)$ and $\prec_\emptyset$-induction, as in the proof of Theorem \[thm-main-AC\], that $\tilde\Phi\varepsilon q\mr \forall nA(n)$.
Let us now briefly consider some specific instantiations of the parameters of Theorem \[thm-main-AC\] (a more detailed discussion will be given in Section \[sec-dependent-examples\]). Firstly, setting $\Psi_0=\Psi_{(\prec,m)}$ in the simple case that $\prec_u=\emptyset$ and $mnu=n$ for all $u$, we have $\initSegsss{\alpha_u}{n}=\emptyset$ and, using the abbreviation $\varepsilon_n$ for $\varepsilon_{n,n}$ our realizer becomes $$\Psi_0^{\varepsilon,q}(u)=q(u\at\lambda n\; . \; \varepsilon_n(\lambda x\; . \; \Psi_0(\update{u}{n}{x})))$$ which is nothing more than a simple variant of the Berardi-Bezem-Coquand realizer of countable choice given in [@BBC(1998.0)] and discussed in [@Berger(2004.0)]. On the other hand, suppose that we still keep $\prec_u=\emptyset$ for all $u$, but define $mnu:=\least i\leq n(i\notin\dom(u))$, where recall that $\least$ is the bounded search operator which in this case returns the least $i\leq n$ satisfying $i\notin\dom(u)$, and $n$ if no such $i$ exists. Then finite input for this variant of $\Psi$ will be of the form $\ext{s}$ for some sequences $s\colon\rho^\ast$. Defining $\Psi_1(s):=\Psi_{(\emptyset,m)}(\ext{s})$ and observing that for $n\geq |s|$ we have $mn\ext{s}=|s|$, we obtain $$\Psi^{\varepsilon,q}_1(s):=q(\ext{s}\at\lambda n\; . \; \varepsilon_{|s|,n}(\lambda x\; . \; \Psi_1(s\ast x)))$$ which is just a (non-dependent) form of modified bar recursion. Finally, let us define $\prec_u=<$ and $mnu=n$ for all $u$. Then setting $\Psi_2(s):=\Psi_{(<,m)}(\ext{s})$, abbreviating $\varepsilon_{n,n}$ by $\varepsilon_n$ and observing that for $n\geq |s|$ we have $\ext{s}\at \initSegsss{\alpha_u}{n}=\initSeg{\alpha_u}{n}$ we obtain a realizing term satisfying $$\Psi_2^{\varepsilon,q}(s)=q(\underbrace{\ext{s}\at\lambda n\; . \; \varepsilon_n(\lambda x\; . \; \Psi_2(\initSeg{\alpha_u}{n}\ast x))}_{\alpha_u}).$$ The corresponding variant $\tilde\Psi_2$ which realizes $J^\ast_2$-shift is exactly the simple implicit product of selection functions of [@EscOli(2010.1)].
Thus three completely different modified realizability intepretations of countable choice appear as simple instances of Theorem \[thm-main-AC\]. Moreover, in each instance we only require a restricted form of backward induction and well-founded recursion which corresponds exactly to the soundness proofs used in the original papers: for $\Psi_0$ Theorem \[thm-main-AC\] is reduced to the proof of the double negation shift using update recursion given in [@Berger(2004.0)], while $\Psi_1$ and $\Psi_2$ require backward induction relativised to downward closed partial functions, which is entirely equivalent to the variants of bar induction used to prove their correctness in [@BergOli(2005.0)] and [@EscOli(2010.1)] respectively. Thus Theorem \[thm-main-AC\] doesn’t simply provide a parametrised framework with which different realizers can be compared, but also a framework in which their correctness proofs can be viewed in a uniform way as relativisations of backward induction.
Of course the construction of such a framework is only partially motivated by the desire to compare existing interpretations. Theorem \[thm-main-AC\] generalises existing work in that one can use an arbitrary parameters to define new realizers of the $J^\ast_i$-shift that are automatically correct, giving an additional level of flexibility and power when it comes to extracting computational content from proofs in practise.
However, we do not discuss this in any more detail here, instead proceeding straight to the generalisation of Theorem \[thm-main-AC\] to full dependent choice.
A computational interpretation of dependent choice {#sec-dependent}
==================================================
We now give a parametrised realizability interpretation to the principle of countable dependent choice. Our formulation of dependent choice will be slightly more general that the usual sequential variants treated in e.g. [@BergOli(2005.0); @EscOli(2012.0); @Seisenberger(2008.0)], in the sense that we parametrise the principle itself by a decidable well-founded relation $\lhd$ on $\N$ which dictates the underlying dependency of the choice sequence.
To be more precise, given a decidable strict well-founded partial order $\lhd$ let us extend our type system with types $\dc{\rho}{\lhd}$ which represent the set $\bigcup_{n\in\N} \dc{\rho}{\lhd_n}$ where $$\dc{\rho}{\lhd_n}:\equiv \{m \; | \; m\lhd n\}\to\rho.$$ We tacitly assume that the types $\dc{\rho}{\lhd}$ can be smoothly incorporated into our system, and come equipped with a length function $|\cdot|\colon\dc{\rho}{\lhd}\to \N$ returning for each $t\colon\dc{\rho}{\lhd}$ a unique index $|t|$ such that $t\in \dc{\rho}{\lhd_{|t|}}$. For $\lhd=<$, the type $\dc{\rho}{<}$ is isomorphic to the type $\rho^\ast$ of finite sequences over $\rho$, objects of type $$\{m \; | \; m<n\}\to\rho$$ representing finite sequences of length $n$. In fact $\dc{\rho}{\lhd}$ is essentially a generalisation of the finite sequence type to arbitrary $\lhd$-closed partial functions (which need not have finite domain, though). Note that when $\lhd=\emptyset$ is the empty relation, the type $\dc{\rho}{\emptyset}$ is isomorphic to $\N$, and as we will see, in this case our parametrised dependent choice principle collapses to normal countable choice.
Now, the principle of $\oDC{\lhd}$ is given by the schema $$\forall s^{\dc{\rho}{\lhd}}(\forall i\lhd |s|A_i(\initSegss{s}{i},s(i))\to \exists x^\rho A_{|s|}(s,x))\to\exists \alpha^{\N\to\rho}\forall n A_n(\initSegss{\alpha}{n},\alpha(n))$$ where $\initSegss{\alpha}{n}\colon \dc{\rho}{\lhd n}$ is just the $\lhd$-initial segment of $\alpha$ i.e. $\lambda m\lhd n.\alpha(m)$, and analogously for $s$ (note that $\initSegss{s}{i}$ is well-defined for $i\lhd |s|$ by transitivity of $\lhd$). Note that $\oDC{\lhd}$ follows from the full axiom of choice together with well-founded recursion over $\lhd$ as follows: by classical logic and full choice we have $$\forall s(\forall i\lhd |s|A_i(\initSegss{s}{i},s(i))\to \exists x A_{|s|}(s,x))\to \exists \Theta\forall s(\forall i\lhd |s|A_i(\initSegss{s}{i},s(i))\to A_{|s|}(s,\Theta s)).$$ Now, recursively defining $\alpha(n):=\Theta(\initSegss{\alpha}{n})$, we prove $\forall n A_n(\initSegss{\alpha}{n},\alpha(n))$ by $\wI_{\lhd}$, since from the assumption that $\forall i\lhd n A_i(\initSegss{\alpha}{i},\alpha(i))$ we obtain $A_n(\initSegss{\alpha}{n},\alpha(n))$ using that fact that $\initSegss{\initSegss{\alpha}{n}}{i}=\initSegss{\alpha}{i}$ for $i\lhd n$.
In Section \[sec-dependent-examples\] below we discuss more well-known instances dependent choice, including its canonical formulation as $$\forall n,x^X\exists y A_n(x,y)\to\exists f\forall n A_n(f(n),f(n+1)),$$ which is easily provable from $\oDC{<}$. However, here we take advantage of the fact that our setting allows us to interpret dependent choice in the non-standard, but very general form $\oDC{\lhd}$.
Realizing $\oDC{\lhd}^{J^\ast_i}$ {#sec-dependent-real}
---------------------------------
As in the previous section, we realize a positive form of the negative translation of $\oDC{\lhd}$ which is somewhat analogous to the dependent $J$-shift of [@EscOli(2012.0)]. As with the previous section, we can easily derive a realizer for the standard negative translation of dependent choice principles from a realizer of our shift principles - and we illustrate this in Section \[sec-dependent-examples\].
\[defn-Nshift-dep\] We define the translated principle $\oDC{\lhd}^{J^\ast_i}$ by $$\begin{aligned}\oDC{\lhd}^{J^\ast_1} \; &\colon \; \begin{cases}\forall s,r(\forall i\lhd |s| A_i(\initSegss{s}{i},s(i))\to (\exists x A_{|s|}(s,x)\to R)\to \exists x A_{|r|}(r,x)) \\
\to(\exists \alpha\forall n A_n(\initSegss{\alpha}{n},\alpha(n))\to R)\to R\end{cases}\\
\oDC{\lhd}^{J^\ast_2} \; &\colon \; \begin{cases}\forall s,r(\forall i\lhd |s| A_i(\initSegss{s}{i},s(i))\to (\exists x A_{|s|}(s,x)\to R)\to \exists x A_{|r|}(r,x)) \\
\to(\exists \alpha\forall n A_n(\initSegss{\alpha}{n},\alpha(n))\to R)\to \exists \alpha\forall n A_n(\initSegss{\alpha}{n},\alpha(n))\end{cases}\end{aligned}$$ where $A$ is arbitrary formula over $\N\times\dc{\rho}{\lhd}\times \rho$ and the realizing type of $R$ is restricted to being discrete.
Again, these variants are nothing more than convenient syntactical rephrasings of the dependent shift principle given in [@EscOli(2012.0)], this time extended to arbitrary partial orderings $\lhd$.
Now, if $\sigma$ and $\tau$ are the realizing types of $A_n(s,x)$ and $R$ respectively, then $\oDC{\lhd}^{J^\ast_1}$ is realized by a term $\Phi$ of type $(\dc{(\rho\times\sigma)}{\lhd}\to\dc{\rho}{\lhd}\to (\rho\times\sigma\to\tau)\to\rho\times\sigma)\to ((\rho\times\sigma)^\N\to\tau)\to\tau$ which given input $\varepsilon\colon\dc{(\rho\times\sigma)}{\lhd}\to\dc{\rho}{\lhd}\to (\rho\times\sigma\to\tau)\to\rho\times\sigma$ and $q\colon (\rho\times\sigma)^\N\to\tau$ that satisfies $$\begin{aligned}&\forall s^{\dc{(\rho\times\sigma)}{\lhd}},r^{\dc{\rho}{\lhd}},p^{\rho\times\sigma\to\tau}\begin{cases}\forall i\lhd |s|(s(i)_1\mr A_i(\initSegss{s_0}{i},s(i)_0))\to \\ (\forall x^{\rho\times\sigma}(x_1\mr A_{|s|}(s_0,x_0)\to p(x)\mr R)\to \varepsilon_{s,r}(p)_1\mr A_{|r|}(r,\varepsilon_{s,r}(p)_0))\end{cases}\\
&\forall\alpha^{(\rho\times\sigma)^\N}(\forall n \; \alpha(n)_1\mr A_n(\initSegss{\alpha_0}{n},\alpha(n)_0)\to q(\alpha)\mr R)\end{aligned}$$ returns a term $\Phi\varepsilon q\colon\tau$ satisfying $\Phi\varepsilon q\mr R$. In the formulae above and results that follows, $s_0\colon\dc{\rho}{\lhd}$ denotes the first projection of the term $s\colon\dc{(\rho\times\sigma)}{\lhd}$ i.e. $s_0(m):=s(m)_0$, and similarly for infinite sequences $\alpha$. As in the previous section, we construct a family of realizing terms which follow the same basic principle of backward recursion on partial functions, and which give a computational interpretation to $\oDC{\lhd}^{J^\ast_1}$. However, this time our choices of $\prec_u$ and $m{nu}$ require some additional conditions to ensure they are now compatible with $\lhd$.
\[prop-main-DC\]Suppose that $\sigma$ and $\tau$ are the realizing types of $A(n)$ and $R$, with $\tau$ discrete, and that $\prec\colon\wideps{(\rho\times\sigma)}^\N\to (\N\times\N\to\Bool)$ and $m\colon\N\times \wideps{(\rho\times\sigma)}^\N\to\N$ are such that
(i) \[item-DC-conda\] $\forall k(k\prec_u n\vee k\lhd n\vee k\lhd mnu\to k\prec'_u n)$ for some well-founded relation $\prec'_u$,
(ii) \[item-DC-condb\] $n\notin\dom(u)\to m{nu}\notin\dom(u)\cup\{k\; | \; k\prec_u n\}$,
for all $u$ and $n$. Then there is a term $\Psi^{\varepsilon,q}_{(\lhd,\prec,m)}\colon\overline{(\rho\times\sigma)}^\N\to\tau$ with parameters of type $\varepsilon\colon(\dc{(\rho\times\sigma)}{\lhd}\to\dc{\rho}{\lhd}\to (\rho\times\sigma\to\tau)\to\rho\times\sigma)$ and $q\colon((\rho\times\sigma)^\N\to\tau)$ which is primitive recursively definable in $\BR+\lambda u.\wR{\prec'_u}$, and satisfies the recursive equation $$\Psi^{\varepsilon,q}(u)=q(\underbrace{u\at\lambda n\;. \; \varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(\lambda x\; . \; \Psi^{\varepsilon,q}(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}))}_{\alpha_u})$$ where $\alpha_u\colon (\rho\times\sigma)^\N$ denotes the argument of $q$ as indicated above and $\initSegsss{\alpha_u}{n}\colon\wideps{(\rho\times\sigma)}^\N$ is defined to be $$\lambda k.(\alpha_u(k)\mbox{ if $k\prec_u n$ else $\bot$}).$$
Regarding the somewhat casual definition of the term $\lambda u.\wR{\prec'_u}$ see Remark \[rem-wR\] - again, in all concrete cases discussed later $\prec_u$ will not even depend on $u$ and $\wR{\prec'_u}$ will be trivially definable in $\EHAomega$. As with Proposition \[prop-main-AC\], on an intuitive level it is not too hard to see that $\Psi$ is well-defined. Firstly $\alpha_u$ is well-defined by $\prec'_u$-recursion since in order to compute $\alpha_u(n)$ we require $\alpha_u(k)$ for either $k\prec_u n$, $k\lhd n$ or $k\lhd mnu$, which by condition (\[item-DC-conda\]) implies that $k\prec'_u n$. Then by (\[item-DC-condb\]) for $n\notin\dom(u)$ we have $mnu\notin\dom(u)$ and thus $u\sqsubset u\at \update{\initSegsss{\alpha}{n}}{mnu}{x}$ and so recursive calls on $\Psi$ are always made on strict extensions $u$. A formal proof that $\Psi$ is definable in $\BR+\lambda u.\wR{\prec_u}$ is given in Appendix \[sec-app\].
\[thm-main-DC\]Suppose that $\Psi_{(\lhd,\prec,m)}$ is defined as in Proposition \[prop-main-DC\] for $\lhd$, $\prec$ and $m$ satisfying conditions (\[item-DC-conda\]) and (\[item-DC-condb\]), and additionally
(i) \[item-DC-condc\] if $\dom(u)$ is $\lhd$-closed then so are $\dom(u)\cup\{k\; | \; k\prec_u n\}$ and $\dom(u)\cup\{k\; | \; k\prec_u n\}\cup \{m{nu}\}$,
Then the term $\Phi_{(\lhd,\prec,m)}:=\lambda\varepsilon,q\;.\; \Psi^{\varepsilon,q}_{(\lhd,\prec,m)}(\emptyset)$ realizes $\oDC{\lhd}^{J^\ast_1}$ provably in $\T+\CONT+\backI+\BR+(\wI_{\prec'_{u}})+\lambda u.\wR_{\prec'_u}$.
While the proof of this Theorem is fairly lengthy, it is little more than a straightforward variant of the much simpler proof of Theorem \[thm-main-AC\]. The additional technical details are simply required to deal with the dependency $\lhd$.
All terms $s\colon\dc{\rho}{\lhd}$, when embedded as partial functions, are $\lhd$-closed by transitivity of $\lhd$, but the converse is not necessarily true.
Assume that $\varepsilon$ and $q$ realize the premise of $\oDC{\lhd}^{J^\ast_1}$. Analogous to the proof of Theorem \[thm-main-AC\], we use a main backward induction along with an auxiliary well-founded induction, this time on the slightly more complex formula $$B(u):\equiv \bar{A}(u)\to\Psi^{\varepsilon,q}(u)\mr R.$$ for $$\bar{A}(u):\equiv\forall n\in\dom(u)(\forall k\lhd n(k\in\dom(u))\wedge u(n)_1\mr A(\initSegss{{u_0}}{n},u(n)_0)),$$ where now as our premise we require that $u$ is a $\lhd$-closed partial realizer of the dependent choice sequence. As before, $B(u)$ is open in the sense of Proposition \[prop-openp\] by $\CONT$. To prove the backward induction step, let’s assume that $\bar{A}(u)$ holds and $B(v)$ is true for all $v\sqsupset u$ and try to derive $\Psi(u)\mr R$.
This time we use $\prec_u'$ induction to show that $\forall n C(n)$ where $$C(n):\equiv \alpha_u(n)_1\mr A_n(\initSegss{(\alpha_u)_0}{n},\alpha_u(n)_0).$$ Then we are done since this would imply that $\Psi(u)=q(\alpha_u)\mr R$. So fix $n$ and assume that $\forall k\prec_u' n\; C(k)$. There are two cases to consider. If $n\in\dom(u)$ and then by $\lhd$-closedness of $u$, $C(n)$ becomes $u(n)_1\mr A_n(\initSegss{u_0}{n},u(n)_0)$ which is true by assumption. Otherwise, if $n\notin\dom(u)$ then $C(n)$ becomes $$(\ast) \ \ \varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(p)_1\mr A_n(\initSegss{(\alpha_u)_0}{n},\varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(p)_0)$$ where $$p:=\lambda x\; . \; \Psi^{\varepsilon,q}(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}).$$ We now prove that $$\begin{aligned}&x_1\mr A_{mnu}(\initSegss{(\alpha_u)_0}{mnu},x_0)\to \bar{A}(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x})\mbox{\; and \;}\\
&\bar{A}(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x})\to\Psi(u\at\update{\initSegsss{\alpha_u}{n}}{mnu}{x})\mr R.\end{aligned}$$ The second implication follows from the main backward induction hypothesis since $u\sqsubset u\at\update{\initSegsss{\alpha_u}{n}}{mnu}{x}$ by (ii). To prove the first implication, let’s abbreviate $v:=u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}$. Then $\dom(v)=\dom(u)\cup\{k\; | \; k\prec_u n\}\cup\{mnu\}$, and since by $\bar{A}(u)$ we know that $\dom(u)$ is $\lhd$-closed, so is $\dom(v)$ by condition (iii). Thus to obtain $\bar A(v)$ it remains to show that $$(\ast\ast) \ \ \forall k\in\dom(v)(v(k)_1\mr A(\initSegss{v_0}{k},v(k)_0)).$$ There are three possibilities. If $k\in\dom(u)$ then $(\ast\ast)$ follows from $\bar{A}(u)$, while if $k=mnu$ then $(\ast\ast)$ becomes $x_1\mr A(\initSegss{(\alpha_u)_0}{mnu},x_0)$ by (iii) which is our premise. Otherwise, for $k\prec_u n$, $(\ast\ast)$ becomes $\alpha_u(k)_1\mr A(\initSegss{(\alpha_u)_0}{k},\alpha_u(k)_0)$ by condition (iii), and this we know is true by the $\prec'_u$-induction hypothesis.
Therefore, putting everything together we obtain $$x_1\mr A_{mnu}(\initSegss{(\alpha_u)_0}{mnu},x_0)\to p(x)\mr R.$$ Since in addition we have $\forall i\lhd mnu(\alpha_u(i)_1\mr A_i(\initSegss{(\alpha_u)_0}{i},\alpha_u(i)_0)$, a fact which also follows from the $\prec'_u$-induction hypothesis along with condition (i), then by correctness of $\varepsilon$ we obtain $(\ast)$.
Thus we have shown that $\forall k\prec_u' n\; C(k)\to C(n)$ and therefore $\forall n C(n)$ follows by induction. Since this in turn implies $\Psi(u)\mr R$, we have shown $\forall v\sqsupset u B(v)\to B(u)$, and thus $\forall u B(u)$ by backward induction. Finally, then, since $\bar{A}(\emptyset)$ is trivially satisfied, $B(\emptyset)$ implies $\Psi^{\varepsilon,q}(\emptyset)\mr R$, which completes the proof.
\[cor-main-DC\]There is a term $\tilde\Phi_{(\lhd,{\prec},{m})}$ primitive recursive in $\BR+\lambda u.\wR_{\prec_{u}}$ which realizes the $\oDC{\lhd}^{J^\ast_2}$, provably in $\EHAomega+\CONT+\backI+\BR+(\wI_{\prec_{u}})+\lambda u.\wR_{\prec_u}$.
Just as in the previous section, and keeping the notation of Theorem \[thm-main-DC\], define $\tilde\Psi^{\varepsilon,q}(u)=\alpha_u$ so that it satisfies the recursive equation $$\tilde\Psi^{\varepsilon,q}(u)=u\at\lambda n\;. \; \varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(\lambda x\; . \; q(\tilde\Psi^{\varepsilon,q}(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}))).$$ Define $\tilde\Phi\varepsilon q:=\tilde\Psi^{\varepsilon,q}(\emptyset)$. Then it follows immediately from $\forall u B(u)$ and $\prec_\emptyset'$-induction that $\tilde\Phi\varepsilon q=\alpha_{\emptyset}\mr \exists\alpha\forall nA(n)$ since $\forall n\; \alpha_\emptyset(n)_1\mr A_n(\initSegss{(\alpha_{\emptyset})_0}{n},\alpha_\emptyset(n)_1)$.
Examples {#sec-dependent-examples}
--------
We now show that essentially all of the solutions to the modified realizability interpretation choice principles given across e.g. [@BBC(1998.0); @BergOli(2005.0); @EscOli(2012.0); @Seisenberger(2008.0)] appear as special cases of Theorem \[thm-main-DC\], given suitable instantiations of $\lhd$, $\prec$, $m$ and the choice formula $A_n(s,x)$.
### The Berardi-Bezem-Coquand functional ($\lhd=\prec_u=\emptyset$) {#sec-dependent-examples-BBC}
It is easy to see that the principle $\oDC{\emptyset}$ is just countable choice, since $\dc{\rho}{\emptyset_n}$ is just a singleton object $\{\bullet\}$ indexed by $n$, and so $\dc{\rho}{\emptyset}$ is isomorphic to $\N$. Setting $A_n(\bullet,x):=B_n(x)$ we obtain $\AC$ as defined in Section \[sec-countable\]. In fact, Theorem \[thm-main-DC\] completely reduces to Theorem \[thm-main-AC\] for the formula $A(n):=\exists x B_n(x)$ once we eliminate $\lhd$. Therefore as expected there is a direct corresondence between the realizers of the two theorems in this case.
For any function $m$ satisfying $n\notin\dom(u)\to mnu\notin\dom(u)$, we can define a generalised version of the Berardi-Bezem-Coquand functional as $\BBC_{(m)}:=\Psi_{(\emptyset,\emptyset,m)}$, which satisfies the defining equation $$\BBC_{(m)}^{\varepsilon,q}(u)=q(u\at\lambda n\; . \; \varepsilon_{mnu,n}(\lambda x\; . \; \BBC_{(m)}^{\varepsilon,q}(\update{u}{mnu}{x}))).$$ Regardless of the choice of $m$, this functional always gives a computational interpretation to countable choice. Now suppose that we move back into the more conventional setting of the double negation translation of countable choice, setting $R=\bot$ and assuming $B_n(x)$ is a negated formula, so that there exists a term $h$ satisfying $\forall n,x\; (h\mr (\bot\to B_n(x))($. Then if $\phi$ satisfies $$\phi\mr\forall m((\exists x B_m(x)\to \bot)\to\bot)$$ then $\varepsilon^\phi_{m,n}(p):=_{\rho\times\sigma}\pair{0_\rho,h(\phi_m(p))}$ realizes the premise of $\oDC{\emptyset}^{J^\ast_1}$, and thus defining $\BBC_{(m),1}^{\phi,q}:=\BBC _{(m)}^{\varepsilon^\phi,q}$ - which satisfies the equation $$\BBC_{(m),1}^{\phi,q}(u)=q(u\at\lambda n\; . \; \pair{0,h(\phi_{mnu}(\lambda x\; . \; \BBC_{(m),1}^{\varepsilon,q}(\update{u}{mnu}{x})))}).$$ - we obtain a term which realizes the negative translation of $\AC$. In particular, setting $mnu=n$ we obtain $$\BBC_{2}^{\phi,q}(u)=q(u\at\lambda n\; . \; \pair{0,h(\phi_{n}(\lambda x\; . \; \BBC_{2}^{\varepsilon,q}(\update{u}{n}{x})))})$$ which is just the BBC functional of [@BBC(1998.0)] (more precisely, the variant of the Berardi-Bezem-Coquand functional for input with arbitrary domain considered in [@Berger(2004.0)]).
### Modified bar recursion ($\prec_u=\emptyset$) {#sec-dependent-examples-MBR}
Suppose that we retain the simplification $\prec_u=\emptyset$, but now allow $\lhd$ to range over arbitrary decidable partial orders. Then we obtain a realizer which makes recursive calls over updates of its input, just like the BBC functional, but now $m$ must not only satisfy conditions (\[item-DC-conda\]) and (\[item-DC-condb\]) but also (\[item-DC-condc\]): $$\mbox{$u$ is $\lhd$-closed }\to\mbox{ $u\cup\{mnu\}$ is $\lhd$-closed}$$ In this case, the term $\MBR^{\varepsilon,q}_{(\lhd,m)}:=\Psi^{\varepsilon,q}_{(\lhd,\emptyset,m)}$, which has defining equation $$\MBR^{\varepsilon,q}_{(\lhd,m)}(u)=q(u\at\lambda n\;. \; \varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(\lambda x\; . \; \MBR^{\varepsilon,q}_{(\lhd,m)}(\update{u}{mnu}{x})))$$ forms a realizer for $\oDC{\lhd}$. This can be viewed as a generalisation of modified bar recursion as first defined in [@BergOli(2005.0)]. All existing variants of $\MBR$ occur when $\lhd$ is the usual ordering on $\N$, but $\MBR$ is perfectly well-defined in more unusual cases. For instance, suppose that we have a bijective encoding $c\colon\N\to\Bool^\ast$, and that $$m\lhd n:=\mbox{$c(m)$ is a proper prefix of $c(n)$}.$$ Then $\lhd$-closed partial functions are precisely partial functions whose domain is a binary tree, and in this case there are many valid choices for $m$, a canonical one being $$mnu:=\mbox{$n$ if $n\in\dom(u)$ else $i$ where $c(i)$ is the least prefix of $c(n)$ not in $\dom(u)$}.$$ This variant of $\MBR$ yields an intuitive realizer for dependent choice over binary trees.
Now, suppose that we do indeed have $\lhd=<$. Then $<$-closed partial functions are either total or of the form $\ext{s}$ where $s$ is a finite sequence. Observing that in order to evaluate $\MBR(\emptyset)$ we can restrict ourselves to input with finite domain we can redefine our realizer in this case as $\MBR_1^{\varepsilon,q}(s^{\rho^\ast}):=\MBR_{(<,m)}^{\varepsilon,q}(\ext{s})$, setting $mnu:=\least i\leq n(i\notin\dom(u))$. Clearly such an $m$ satisfies (\[item-DC-condc\]), and in particular $mn\ext{s}=|s|$ for $n\geq |s|$. Therefore $\MBR_1$ has defining equation $$\MBR^{\varepsilon,q}_{1}(s)=q(\ext{s}\at\lambda n\;. \; \varepsilon_{s,\initSegss{\alpha_0}{n}}(\lambda x\; . \; \MBR^{\varepsilon,q}_{1}(s\ast x)))$$ This directly realizes $\oDC{<}$, which is isormorphic to $$\forall s^{\rho^\ast}(\forall i<|s| A_i(\initSeg{s}{i},s(i))\to\exists x A_{|s|}(s,x))\to\exists\alpha\forall n A_n(\initSeg{\alpha}{n},\alpha(n)).$$ We can now easily rederive various concrete instances of $\MBR_1$ found in the literature which arise from setting $R=\bot$ and instantiating $A_i(s,x)$ by the correct formula. First, note that we immediately derive $\AC$ from $\oDC{<}$ by setting $A_n(s,x):=B_n(x)$, and a corresponding realizer for $\AC$ by defining $\MBR_2^{\phi,q}:=\MBR_1^{\varepsilon^\phi,q}$ for $\varepsilon_{s,r}^\phi(p)=\pair{0_\rho,h(\phi_{|s|}(p))}$, where $\phi$ and $h$ are as in Section \[sec-dependent-examples-BBC\]. This has defining equation $$\MBR^{\phi,q}_{2}(s)=q(\ext{s}\at\lambda n\;. \; \pair{0,h(\phi_{|s|}(\lambda x\; . \; \MBR^{\phi,q}_{2}(s\ast x)))})$$ which is exactly the realizer of countable choice constructed in [@BergOli(2005.0)]. Note that the same realizer could have been constructed from $\BBC_{(m),1}$ for suitable $m$.
Now suppose that $A_0(s,x):=B_0(x_0,x)$ and $A_n(s,x):=B_n(s_{|s|-1},x)$ for $n>0$. Then $\oDC{<}$ immediately implies the following, standard formulation of dependent choice: $$\label{eqn-DC-BergOli}\DC \; \colon \; \forall n,y\exists x B_n(y,x)\to\forall x_0\exists \alpha(\alpha(0)=x_0\wedge\forall n B_n(\alpha(n),\alpha(n+1)).$$ The challenge for realizing $\DC$ is as follows: we must construct a realizer of $\bot$, given realizers $$\begin{aligned}\phi^{\N\to\rho\to(\rho\times\sigma)\to\tau)\to\tau} &\mr \forall n,y((\exists x B_n(y,x)\to \bot)\to \bot)\\
Y^{(\rho\times\sigma)^\N\to\tau}&\mr \exists\alpha(\alpha(0)=x_0\wedge\forall nB_n(\alpha(n),\alpha(n+1)))\to \bot\end{aligned}$$ and in addition assuming that $B$ is negated, thus guaranteeing the existence of a realizer $h$ of ex-falso quodlibet satisfying $\forall n,y,x\; h\mr (\bot\to B_n(y,x))$. But in this case, we can easily define realizers of the premise of $\oDC{<}$ as $$\begin{aligned}\varepsilon^\phi_{s,r}(p)&:=\pair{0,h(\phi_{|s|,(x_0\ast s_0)_{|s|}}(p))}\\
q^Y(\alpha)&:= Y_{x_0}(\alpha):= Y(\pair{x_0\ast\alpha_0,\alpha_1})\end{aligned}$$ and defining $\MBR_3^{\phi,Y}:=\MBR_1^{\varepsilon^\phi,q^Y}$ yields a realizer for $\DC$ satisfying $$\MBR^{\phi,Y}_{3}(s)=Y_{x_0}(\ext{s}\at\lambda n\;. \; \pair{0,h(\phi_{|s|,(x_0\ast s_0)_{|s|}}(\lambda x\; . \; \MBR^{\phi,Y}_{3}(s\ast x)))})$$ which is this time exactly the realizer of $\DC$ given in [@BergOli(2005.0)]. In an entirely analogous way, the bar recursive solution to the sequential variant of dependent choice considered in [@Seisenberger(2008.0)]: $$\label{eqn-DC-Seis}B(\pair{})\to\forall s^{\rho^\ast}(B(s)\to\exists xB(s\ast x))\to\exists \alpha\forall n B(\initSeg{\alpha}{n})$$ can be defined in terms of $\MBR_1$, since (\[eqn-DC-Seis\]) is easily implied by $\oDC{<}$ for $A_n(s,x):=B(s\ast x)$. Given realizers $$\begin{aligned}a_0^\rho&\mr B(\pair{})\\
\phi^{\rho^\ast\to\sigma\to (\rho\times\sigma\to\tau)\to\tau}&\mr \forall s(B(s)\to (\exists xB(s\ast x)\to \bot)\to \bot)\\
Y^{(\rho\times\sigma)^\N\to\tau}&\mr \exists \alpha\forall n B(\initSeg{\alpha}{n})\to \bot\end{aligned}$$ and assuming the existence of a realizer $\forall s\; h\mr (R\to B(s))$, it is easy to see that $$\begin{aligned}\varepsilon_{s,r}^{a_0,\phi}(p)&:=\pair{0,h(\phi_{s_0,(a_0\ast s)_{|s|}}(p)}\\
q^{a_0,Y}(\alpha)&:= Y_{a_0}(\alpha):= Y(\pair{\alpha_0,G_0\ast\alpha_1})\end{aligned}$$ realize the premise of $\oDC{<}$, and therefore the realizer we obtain is $\MBR^{a_0,\phi,Y}_4:=\MBR_1^{\varepsilon_{s,r}^{a_0,\phi},q^{a_0,Y}}$ which satisfies $$\MBR^{a_0,\phi,Y}_4(s)=Y_{a_0}(\ext{s}\at\lambda n\;. \; \pair{0,h(\phi_{s_0,(a_0\ast s)_{|s|}}(\lambda x\; . \; \MBR^{a_0,\phi,Y}_4(s\ast x)))})$$ which is exactly the term used to interpret dependent choice in [@Seisenberger(2008.0)] (and also extract an algorithm from Higman’s lemma in [@Seisenberger(2003.0)]).
### Products of selection functions ($mnu=n$, $\lhd$ total) {#sec-dependent-examples-IPS}
We finally consider the case in which the functional $m$ is defined to be the projection function $mnu=n$. Clearly, the conditions (i)-(iii) are satsified whenever $\prec_u=\lhd$, in which case the realizer simplifies to $$\Psi^{\varepsilon,q}(u)=q(\underbrace{u\at\lambda n\;. \; \varepsilon_{\initSegss{\alpha}{n}}(\lambda x\; . \; \Psi^{\varepsilon,q}(u\at \update{\initSegss{\alpha}{n}}{n}{x}))}_{\alpha_u})$$ where we abbreviate $\varepsilon_{\initSegss{\alpha}{n}}=\varepsilon_{\initSegss{\alpha}{n},\initSegss{\alpha_0}{n}}$ and $\initSegss{\alpha}{n}$ is now treated as a partial function. Now, suppose that $\lhd$ is a *total* order, and therefore constitutes an encoding of some countable ordinal $\xi$. It is always the case that we can evaluate $\Psi(\emptyset)$ by restricting the input to being $\lhd$-closed. However, when $\lhd$ is total then the set of (domain-theoretically) non-total $\lhd$-closed partial sequences is isomorphic to $\dc{\rho}{\lhd}$, since if $u$ is $\lhd$-closed then $\dom(u)=\{k \; | \; k\lhd n\}$ where $n$ is the least element of the set of undefined elements of $u$. Let us suppose that $\lhd$ comes equipped with a computable ordinal successor function $\suc\colon\N\to\N$ i.e. for each number $n$, $\suc(n)$ is the least number greater than $n$ with respect to $\lhd$. Define $\PS_{(\lhd)}^{\varepsilon,q}\colon\dc{\rho}{\lhd}\to\rho^\N$ by $\PS^{\varepsilon,q}(s):=\tilde\Psi^{\varepsilon,q}(\ext{s})$ (where $\tilde\Psi$ is the term defined in Corollary \[cor-main-DC\]). Then $\PS$ satisfies the equation $$\PS_{(\lhd)}^{\varepsilon,q}(s)=\ext{s}\at\lambda n\;. \; \varepsilon_{\initSegss{\alpha}{n}}(\lambda x\; . \; q(\PS_{(\lhd)}^{\varepsilon,q}(\initSegss{\alpha}{n}\star x))).$$ where $\initSegss{\alpha}{n}\star x\colon \dc{\rho}{\lhd_{\suc(n)}}$ is defined by $$(\initSegss{\alpha}{n}\star x)(m):=\begin{cases}\alpha(m) & \mbox{if $m\lhd n$}\\ x & \mbox{if $m=n$}.\end{cases}$$ This functional is a generalisation of the (implicitly well-founded) product of selection functions of Escard[ó]{} and Oliva to arbitrary recursive ordinals, and is not only a realizer $\oDC{\lhd}^{J^\ast_2}$, but a direct realizer of the following transfinite $J$-shift principle: $$\begin{cases}\forall s(\forall i\lhd |s| A_i(\initSegss{s}{i},s(i))\to (\exists x A_{|s|}(s,x)\to R)\to \exists x A_{|s|}(s,x)) \\
\to(\exists \alpha\forall n A_n(\initSegss{\alpha}{n},\alpha(n))\to R)\to \exists \alpha\forall n A_n(\initSegss{\alpha}{n},\alpha(n)).\end{cases}$$ For the particular case that $\lhd$ is the normal ordering on $\N$ and $A_n(s,x):=B_n(s\ast x)$ this becomes $$\begin{cases}\forall s^{\rho^\ast}(\forall i< |s| B_i(\initSeg{s}{i+1})\to (\exists x B_{|s|}(s\ast x)\to R)\to \exists x B_{|s|}(s\ast x)) \\
\to(\exists \alpha\forall n B_n(\initSeg{\alpha}{n+1})\to R)\to \exists \alpha\forall n B_n(\initSeg{\alpha}{n+1}).\end{cases}$$ which is precisely the dependent $J$-shift of [@EscOli(2012.0)], and our realizer becomes $$\PS_{(<)}^{\varepsilon,q}(s)=\ext{s}\at\lambda n\;. \; \varepsilon_{\initSeg{\alpha}{n}}(\lambda x\; . \; q(\PS_{(<)}^{\varepsilon,q}(\initSeg{\alpha}{n}\ast x)))$$ which is isomorphic to the dependent product of selection functions of [@EscOli(2012.0)]. Setting $A_n(s,x):=B_n(x)$ as in Section \[sec-dependent-examples-MBR\] we see that the dependent $J$-shift implies the non-dependent $J$-shift: $$\forall n(\exists x B_{n}(x)\to R)\to \exists x B_{n}(x))
\to(\exists \alpha\forall n B_n(\alpha(n))\to R)\to \exists \alpha\forall n B_n(\alpha(n))$$ and our realizer can be simplified to $$\PS_{(<)}^{\varepsilon,q}(s)=\ext{s}\at\lambda n\;. \; \varepsilon_{n}(\lambda x\; . \; q(\PS_{(<)}^{\varepsilon,q}(\initSeg{\alpha}{n}\ast x)))$$ which is now isomorphic to the non-dependent product of selection functions of [@EscOli(2012.0)].
### Summary {#sec-dependent-examples-summary .unnumbered}
We have demonstrated that the many variants of bar recursion used to intepret choice principles are just instances of the same basic cominatorial idea. Theorem \[thm-main-DC\] gives us a completely uniform framework in which to understand the variety of different realizers currently in the literature, and moreover in each case we are able to provide generalisations of these realizers to more complex orderings on $\N$. We summarise all this in the table below (here $m_1nu:=\least i\leq n(i\notin\dom(u))$ and $m_2nu=n$).\
$m$ $\prec_u$ $\lhd$ $A_n(s,x)$ **realizer**
------- ------------- ------------------ -------------------- -----------------------------------------------------------------------------
$m_1$ $\emptyset$ $\emptyset$ $B_n(x)$ Berardi-Bezem-Coquand functional [@BBC(1998.0)]
$m_2$ $\emptyset$ $\emptyset\;/<$ $B_n(x)$ simple modified bar recursion [@BergOli(2005.0)]
$m_2$ $\emptyset$ $<$ $B_n(s_{|s|-1},x)$ dependent modified bar recursion [@BergOli(2005.0)]
$m_2$ $\emptyset$ $<$ $B(s\ast x)$ dependent modified bar recursion [@Seisenberger(2008.0)]
$m_1$ $<$ $\emptyset \;/<$ $B_n(x)$ simple product of selection functions [@EscOli(2010.1); @EscOli(2012.0)]
$m_1$ $<$ $<$ $B_n(s\ast x)$ dependent product of selection functions [@EscOli(2010.1); @EscOli(2012.0)]
\
However, these examples consitute only an extremely limited range of possibilities for $\Psi_{(m,\lhd,<)}$ based on very simple instantiations of its parameters. Theorem \[thm-main-DC\] gives us far more than just a unifying perspective for existing realizers: it gives us a very general recipe for devising new realizers for choice principles that can be tailored to the situation at hand. Thus, rather than simply giving a fixed interpretation of choice and extracting programs relative to this realizer, we have an additional level of flexibility which should allow us to extract much more efficient and meaningful programs.
Understanding the parametrised realizer {#sec-semantics}
=======================================
Having completed the main theoretic work of this paper, the purpose of this section is to give a somewhat informal graphical representation of the structure of our realizer and the proof of Theorem \[thm-main-DC\]. In doing so we suggest a potential semantic interpretation of the realizer.
Much work has been done in the last few decades on providing a computational interpretation of classical logic that can be understood on an intuitive level terms of *learning* - the basic idea being that realizers of classical principles typically carry out some kind of ‘learning procedure’ in order to construct an approximation to that principle.
In [@BBC(1998.0)] a connection is suggested between realizers of negative translated formulas and winning strategies related to the Novikoff interpretation of classical formulas [@Coquand(1994.0); @Novikoff(1943.0)]. In particular, an illuminating semantic interpretation of what we have called the Berardi-Bezem-Coquand functional is given, in which the functional represents a strategy for building an approximation of a choice functional which wins against any continuous opponent. We attempt to extend this interpretation to our generalised realizer of choice, and argue that the our parameters can be giving a clear meaning in this context.
Let us first very briefly recall the basic ideas of [@BBC(1998.0); @Coquand(1994.0)] (the reader is referred to these papers for a proper treatment). In the Novikoff calculus, the formula $\exists x\forall y A(x,y)$ is mapped to the propositional formula $$\bigvee_x\bigwedge_y A(x,y).$$ The truth of such a formula is debated by two players $\Eloise$ and $\Abelard$, who support truth and falsity respectively. First, $\Eloise$ selects some value $x_0$ for which she claims that $\bigwedge_y A(x_0,y)$ is true, and $\Abelard$ follows this by choosing some $y_0$ in an attempt to falsify $A(x_0,y_0)$. The formula as a whole is intuitionistically valid iff $\Eloise$ has a winning strategy regardless of any choices made by $\Abelard$.
Now, it is clear that such a correspondence does not work in the case of classical logic, since there are $\Sigma_2$-formulas classically true but for which there is no effective strategy for $\bigvee_x\bigwedge_y A(x,y)$. In this case, validity of $\bigvee_x\bigwedge_y A(x,y)$ is interpreted as the existence of some $x_0$ such that $$A(x_0,y')\vee\bigvee_x\bigwedge_y A(x,y)$$ is valid for all $y'$. The idea here is that $\Eloise$ picks a potential witness $x_0$, which is followed by an attempt at a counterexample $y_0$ from $\Abelard$, and the game becomes $A(x_0,y_0)\vee\bigvee_x\bigwedge_y A(x,y)$. In other words, either $A(x_0,y_0)$ is true, in which case $\Eloise$ wins, or it is false, and $\Eloise$ can backtrack and start again, this time using falsity of $A(x_0,y_0)$ as constructive information. In this way, $\Eloise$ is allowed to ‘learn’ from $\Abelard$’s choices. Moreover, as demonstrated in [@BBC(1998.0)], this notion of learning and backtracking is captured by the recursive functionals which realize the negative translation of the classical formula.
Let us now take this basic idea and consider how the truth of countable choice can be interpreted as a dialogue between $\Eloise$ and $\Abelard$ in which $\Eloise$ eventually wins. Countable choice can be written as the disjuction $$\label{eqn-AC-dis}\exists n\forall x\neg A_n(x)\vee\exists f\forall n A_n(f(n)).$$ We give this formula a rough interpretation along the lines of [@BBC(1998.0)] as follows. $\Eloise$ begins by attempting to realize the conclusion of $\AC$ with some default function $f_0=\lambda n.0$ whose ‘domain’ of genuine constructive information is empty, and $\Abelard$ responds by selecting some point $n_0$ such that this attempt fails i.e. we cannot provide a realizer for $A_{n_0}(0)$. $\Eloise$ responds by now attempting to falsify the premise at some $m_0$, to which $\Abelard$ responds with a point $x_0$ and a realizer for $A_{m_0}(x_0)$.
$\Eloise$ has now been given some constructive information by $\Abelard$, so she starts the whole process again, this time with the function $f_1:=f_0[m_0\mapsto x_0]$, which now has a domain of $\{m_0\}$. This time, $\Abelard$ picks a point $n_1$ at which $f_1$ fails. Either he picks $n_1=m_0$ in which case he loses since by his own admission $A_{m_0}(x_0)$ is true (see [@BBC(1998.0)] for a formal translation of this logic into Novikoff strategies), or he chooses $n_1\notin\{m_0\}$. Then $\Eloise$ responds with some $m_1\notin\{m_0\}$ which falsifies the premise of $\AC$, and again $\Abelard$ responds with some $x_1$ and a realizer for $A_{m_1}(x_1)$. $\Eloise$ now updates her approximation again to some $f_2$ which includes this information, and continues as before. Roughly speaking, such a strategy should eventually result in success for $\Eloise$ whenever $\Abelard$ is ‘continuous’, because he will eventually be forced to pick $n_i$ in the domain of $f_i$.
We have been vague as to two details in this strategy: firstly how $\Eloise$ decides on which $m_i$ to choose in light of $\Abelard$’s original choice $n_i$, and secondly in how she chooses to update her approximation each time. In the first instance it is clear that if $\Abelard$ has a ‘good’ choice for $n_i$ which is not in the domain $f_i$ then $\Eloise$ must also respond with $m_i$ not in the domain of $f_i$ if she has any chance of falsifying the premise of $\AC$. Then, when it comes to updating $f_i$ with the information $(m_i,x_i)$, she could either just add this directly to $f_i$ and define $f_{i+1}:=f_i[m_i\to x_i]$, or she could potentially erase some of the existing elements in the domain of $f_i$.
The most natural choice is of course to pick $m_i=n_i$ each time and update directly without any erasing, and this is precisely the strategy given in [@BBC(1998.0)]. However, in the case of *dependent choice* where the elements of the choice sequences are related in some way, it is crucial for $\Eloise$ to update her approximation in a manner which is coherent with the underlying dependency required for the choice sequence.
Let us now move on to our realizer for the negative translation of countable dependent choice (in the form of the $\oDC{\lhd}^{J^\ast_1}$-shift), and show how the parameters of the realizer correspond to the choices in $\Eloise$’s strategy we have just discussed. In Figure \[fig-a\] below we give a (very informal) diagrammatic representation of the proof of Theorem \[thm-main-DC\]. We retain all our notation, in particular we write $\bar A(u)$ to denote that $u$ is a partial realizer. For simplicity we now work in a concrete realizability setting with $R=\bot$ and $\varepsilon_{s,r}(p)=\pair{0,h(\phi_s(p))}$ where $$\phi\mr \forall s((\exists x A_{|s|}(s,x)\to\bot)\to\bot).$$ and $\forall n,r,x (h\mr (\bot\to A_n(r,x)))$, so in particular $\varepsilon$ realizes the premise of $\oDC{\lhd}^{J^\ast_1}$.
$$\tiny
\xymatrix{ & A \; \fbox{$\bar A(u)\vdash \Psi(u)\mr \bot$}\ar@/^8.0pc/@{<.}[dddd]!R^{ u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}\sqsupset u}& \\
& \fbox{$\bar A(u)\vdash \alpha_u\mr\exists\alpha\forall n A_n(\initSegss{\alpha}{n},\alpha(n))$} \; B\ar@{<-}[u]_{\mbox{\scriptsize $q\mr \exists\alpha\forall n A_n(\initSegss{\alpha}{n},\alpha(n))\to\bot$}} & \\
& \fbox{$\bar A(u), n\notin\dom(u)\vdash \alpha_u(n)_1\mr A_n(\initSegss{(\alpha_u)_0}{n},0)$} \; C \ar@{<-}[u]_{\mbox{\scriptsize $\Abelard$ chooses $n$}} & \\
& \fbox{$\bar A(u),n\notin\dom(u)\vdash \phi_{\initSegss{\alpha_u}{mnu}}(\lambda x.\Psi(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}))\mr\bot$ } \; D \ar@{<-}[u]_{\mbox{\scriptsize $h\mr (\bot\to A_n(\initSegss{(\alpha_u)_0}{n},0))$}} & \\
& E \; \fbox{$\bar A(u),n\notin\dom(u),x_1\mr A_{mnu}(\initSegss{\alpha_0}{mnu},x_0)\vdash\Psi(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x})\mr \bot$} \ar@{<-}[u]_{\mbox{\scriptsize $\Abelard$ chooses $x$, $\forall s\neg\neg \exists xA_{|s|}(s,x)$}} & \\
& \fbox{$\bar A(u),n\notin\dom(u),x_1\mr A_{mnu}(\initSegss{\alpha_0}{mnu},x_0)\vdash \bar A(u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x})$} \; F \ar@{<.}[u]\ar@/^11.5pc/@{->}[uuu]!L^{k\prec_u n} &
}$$
Let us now run through the diagram, step by step. Roughly speaking, each box represents a stage in a game between $\Eloise$ and $\Abelard$, and an arrow represents a reverse implication in the proof of Theorem \[thm-main-DC\]. We start at $A$ with the assumtion that $\Eloise$ has already computed a partial realizer $u$ that is correct wherever it’s defined. At step $B$, $\Eloise$ plays an approximation $(\alpha_u)_0:=u_0\at 0$ to a choice sequence with a corresponding sequence $(\alpha_u)_1$ of realizers, and in response, $\Abelard$ selects some $n$ and challenges $\Eloise$ to realise $A_n(\initSegss{(\alpha_u)_0}{n},(\alpha_u)_0)$. If $n\in\dom(u)$ then $\Eloise$ wins, so we assume the contrary and move on to step $C$. $\Eloise$ responds to the challenge in the next step $D$ by claiming that the premise of choice principle is false at point $mnu$. $\Abelard$ is now forced to produce $x$ such that $x_1\mr A_{mnu}(\initSegss{(\alpha_u)_0}{mnu},x_0)$ - if he fails then $\Eloise$ wins, and if he succeeds then $\Eloise$ is given constructive information and takes this as an assumption.
We now come to the subtle part: the manner in which $\Eloise$ updates $(\alpha_u)_0$ to reflect this new information and repeat the loop. From a proof theoretic perspective, $E$ is implied by $A$ and $F$ by the cut rule. We can interpret this semantically as follows: $\Eloise$ forms the updated function $(\alpha_u)_0[mnu\mapsto x_0]$, but states that if in future $\Abelard$ queries this realizer for any $k\prec_u n$ with $k\notin\dom(u)$, she will ignore any subsequent information received and revert to stage $C$ as if $\Abelard$ had chosen $k$ instead of $n$. This is reflected in the definition of $\Psi$, since it makes a recursive call on the partial function $u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}$, but the fact that this is a partial realizer for $k\prec_u n$ relies on nested recursive calls of the form $\Psi(u\at\update{\initSegsss{\alpha}{k}}{mku}{y})$, which in particular forgets the value of $x$ at $mnu$.
Therefore at stage $E$ there are two possibilities. Either at some point in the future $\Abelard$ does query $k\prec_u n$ for $k\notin\dom(u)$, in which case all subsequent information is deemed irrelevant and we make an auxiliary $\prec_u$-recursive loop back to $C$, or $\Abelard$ never queries $k\prec_u n$, in which case we can treat $u\at \update{\initSegsss{\alpha_u}{n}}{mnu}{x}$ as a partial realizer and make a backward recursive loop back to stage $A$. In this way, the game corresponding to dependent choice can be seen as a path through Figure \[fig-a\]. By the combination of $\prec_u$-induction and backward induction, using the fact that $\prec_u$ is well-founded and $\Abelard$’s choice of $n$ at step $B$ is based on a continuous strategy, the proof of Theorem \[thm-main-DC\] essentially says that there is no infinite path starting from $\emptyset$, and therefore $\Eloise$ has a winning strategy in the game as a whole.
Of course, we have not made our proposed link between Theorem \[thm-main-DC\] and the world of Novikoff games precise! But at the very least we hope to have provided some insight into how our somewhat syntactic construction in the previous sections could be viewed from an intuitive, semantic perspective.
Let us now think about our concrete examples. For the Berardi-Bezem-Coquand functional we have $\initSegsss{\alpha_u}{n}=\emptyset$ and so the ‘forgetful’ subloop is completely avoided. In this case $\Eloise$ simply responds to $\Abelard$’s choice of $n$ directly, and so when $\Abelard$ plays $(\pair{n_i,x_i})=\pair{1,x_0},\pair{4,x_1},\pair{3,x_2},\ldots$ we get the following sequence of updates $$\mapsto [0,x_0,0,0,0]\mapsto [0,x_0,0,0,x_1]\mapsto [0,x_0,0,x_2,x_1]\mapsto\ldots$$ This semantic interpretation of the BBC-functional is of course completely analogous to the one given in [@BBC(1998.0)]. For modified bar recursion, whenever $\Abelard$ makes a sensible choice $n\notin\dom(u)$, $\Eloise$ switches instead the least element not already in the domain. Again we avoid the forgetful subloop, but this time $\Eloise$’s updates are done in sequence - the same three choices from $\Abelard$ results in the following response: $$\mapsto [x_0,0,0,0,0]\mapsto [x_0,x_1,0,0,0]\mapsto [x_0,x_1,x_2,0,0]\mapsto\ldots$$ Finally, for the product of selection functions, $\Eloise$ never changes $\Abelard$’s choice of $n$, but now assumes a policy of forgetting everthing above the point being updated. This time the result is $$\mapsto [0,x_0,0,0,0]\mapsto [0,x_0,0,0,x_1]\mapsto [0,x_0,0,x_2,0]\mapsto\ldots$$ the point $x_1$ being forgotten as a result of $\Abelard$ choosing $3<4$ as his third move.
The strategy related to the BBC-functional fails for dependent choice because it is assumed that $\Abelard$ always picks $x_i$ such that $A_m(\initSeg{u\at 0}{m},x_i)$ holds, and so e.g. we have $A_4([0,x_0,0,0],x_1)$. However, if ever in the future he gives $\Eloise$ some new information $x_{i+j}$ for point $k<m$ in the sequence, then $x_i$ is no longer valid - for instance we would require $A_4([0,x_0,0,x_2],x_1)$ to hold, which is not necessarily true.
Modified bar recursion and the product of selection functions represent two methods of overcoming this. For the former, $\Eloise$ ensures that $\Abelard$ always gives her constructive information in sequence, while for the latter she is happy to take information for any point $\Abelard$ chooses, but whenever updating she erases everthing above, relying on future moves to regain this information.
Concluding remarks {#sec-concluding}
==================
There are several directions in which the work presented here could be developed. We have already demonstrated that Theorem \[thm-main-DC\] provides a uniform soundness proof through which most of the existing variants of backward recursion used to interpret dependent choice can be derived. The most obvious next step would be to explore the use of new variants of bar recursion which arise from giving more interesting values to the parameters of $\Psi_{(m,\prec,\lhd)}$.
Take, for example, the following classical statement:
> For any function $f\colon\Bool^\ast\to \N$, there exists a function $g\colon\Bool^\ast\to\N$ such that for any branch $s\colon\Bool^\ast$ we have $f(g(s))\leq f(t)$ whenever $t$ has $s$ as a prefix.
This is a direct consequence of the axiom of countable choice, and so any instantiations of $(m,\prec)$ satisfying the conditions of Theorem \[thm-main-AC\] will yield a realizer capable of building an approximation to the choice function $g$. However, it would seem most sensible to choose the parameters so that the recursion is carried out over the natural tree structure $s\lhd t$ iff $s$ is a prefix of $t$ (relative to some encoding of $\Bool^\ast$ into $\N$) - using, for example, the variant of modified bar recursion sketched out the beginning of Section \[sec-dependent-examples-MBR\]. Then if an approximation to $g$ is defined at $s$ then this information could be used to extend the approximation to extensions $t$ of $s$.
It would be interesting to examine in general how such realizers, tailored to the situation at hand, compare to those built from the existing forms of bar recursion. One would expect an advantage in terms of both algorithmic efficiency and the syntactic expressiveness of the extracted program.
Another interesting application of our parametrised form of bar recursion would be to extend the work of Escard[ó]{} and Oliva on Nash-equilibria of unbounded games to the transfinite case. In [@EscOli(2011.0); @EscOli(2012.0)] it is shown that the infinite products of selection functions corresponding to the functional $\PS_{(<)}$ defined in Section \[sec-dependent-examples-IPS\] computes optimal strategies in infinite sequential games over the natural numbers. However, $\PS_{(\lhd)}$ is well-defined for any computable well ordering on $\N$, and in particular it is not too difficult to show that in these cases Spector’s equations: $$\begin{aligned}\alpha(n)&=\varepsilon_{\initSegss{\alpha}{n}}(p_n)\\ p_n(\varepsilon_{\initSegss{\alpha}{n}}(p_n))&=q(\alpha)\end{aligned}$$ can be solved in $\varepsilon$ and $q$ for arbitrary $\lhd$ by setting $\alpha=\PS^{\varepsilon,q}(\emptyset)$ and $p_n(x):=q(\PS^{\varepsilon,q}(\initSegss{\alpha}{n}\star x))$, entirely analogously to the normal ordering on $\N$. However, it would be useful to formalise this properly and to investigate whether there are any interesting applications of higher-type *transfinite* games.
Finally, to the author’s knowledge all known realizability interpretations of choice are restricted to countable or dependent choice principles, and it is not known how to extend these to choice over function spaces, for example: $$\forall f^{\N\to\N}\exists x^\rho B_f(x)\to\exists F^{(\N\to\N)\to\rho}\forall f B_f(F(f)).$$ One possibility would be to build an approximation to $F$ over some countable basis $(c_i)_{i\in\N}$ for the space $\N\to\N$. However, in this case updates to the approximation must be made in a coherent way, since the values of two distinct elements $c_1$ and $c_2$ must be compatible with their intersection. Therefore the ideas behind our parametrised realizer could be helpful here, the aim being to assign an ordering to the basis and make sure that the updates respect this ordering. However, in the author’s opinion it is likely that some additional ingenuity would be required here, as the challege posed by giving a computational interpretation non-countable choice seems to be a significant one.
**Funding.** This work was supported by a LabEx CARMIN postdoctoral research fellowship, and also the Austrian Science Fund (FWF) project “Automated Complexity Analysis via Transformations” (project number P25781).
Appendix {#sec-app}
========
We define $\Psi_{(\lhd,\prec,m)}^{\varepsilon,q}(u):=\BR^{\psi^{\varepsilon,q}}(u)$ where $$\psi^{\varepsilon,q}_u(f^{\N\times \overline{(\rho\times\sigma)}^\N\to\tau}):=_\tau q(\alpha^{\varepsilon,f}_u)$$ where $\alpha^{\varepsilon,f}_u\colon (\rho\times\sigma)^\N$ is recusively defined over $\prec'_u$ by $$\alpha^{\varepsilon,f}_u(n):=\begin{cases}u(n) & \mbox{if $n\in\dom(u)$}\\
\varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(\lambda x\; . \; f(mnu,u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x})) & \mbox{otherwise}\end{cases}$$ where $\initSegsss{\alpha_u}{n}:=\lambda k.(\alpha_u^{\varepsilon,f}(k)\mbox{ if $k\prec_u n$})$. By condition (\[item-DC-conda\]) of the proposition, $\alpha_u^{\varepsilon,f}$ is well-founded. Now, unwinding definitions, we obtain $$\begin{aligned}\Psi^{\varepsilon,q}(u)&=\psi_u(\underbrace{\lambda n,v\; . \; \Psi^{\varepsilon,q}(u\at v)\mbox{ if $n\in\dom(v)\backslash\dom(u)$}}_{f})\\
&=q(\alpha_u)\\
&=q(u\at\lambda n\; . \;\varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(\lambda x\; . \; f(mnu,u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x})))\\
&\stackrel{(\ast)}{=}q(u\at\lambda n\; . \;\varepsilon_{\initSegss{\alpha}{mnu},\initSegss{\alpha_0}{n}}(\lambda x\; . \; \Psi(u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x}))) \end{aligned}$$ where step $(\ast)$ follows from $$\begin{aligned}f(mnu,u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x})&=\Psi(u\oplus(u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x}))\mbox{ if $mnu\in\dom(u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x})\backslash\dom(u)$}\\
&=\Psi(u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x})\mbox{ if $mnu\notin\dom(u)$}\\
&=\Psi(u\oplus \update{\initSegsss{\alpha_u}{n}}{mnu}{x})\end{aligned}$$ the last step following from condition (\[item-DC-condb\]).
[^1]: **Email:** [email protected] **Tel:** +43 512 507 53293
[^2]: Indeed, a fact closely related to this observation, proved in [@BergOli(2006.0)], is that (\[eqn-intro-MBR\]) is not S1-S9 computable in the total continuous functionals and is therefore strictly stronger than (\[eqn-intro-SBR\]).
[^3]: See [@OliPow(2012.2)] for the distinction between the special and generals forms of Spector’s bar recursor.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We perform a model-independent analysis of kinetically driven inflation (KDI) which (partially) includes Generalized G-inflation and Ghost inflation. We evaluate the background evolution splitting into the inflationary attractor and the perturbation around it. We also consider the quantum fluctuation of scalar mode with a usual scaling and derive the spectral index ignoring the contribution from the second order products of slow-roll parameters. Using these formalisms, we find that within our generic framework, the models of KDI which possess the shift symmetry of scalar field cannot create the quantum fluctuation consistent to the observation. Breaking the shift symmetry, we obtain a few essential conditions for viable models of KDI associated with the graceful exit.'
author:
- Rio Saitou
title: Generic analysis of kinetically driven inflation
---
Introduction {#sec1}
============
Inflation is an elegant scenario which can resolve the problems in the big bang universe by accelerating the early universe [@Guth:1980zm; @Starobinsky:1980te; @Sato:1980yn]. The quantum fluctuations produced in the inflationary universe become the seed of large-scale structure, and their nearly scale-invariant, Gaussian features [@Mukhanov:1981xt; @Starobinsky:1982ee; @Hawking:1982cz; @Guth:1982ec] are compatible with the precise observation of cosmic microwave background anisotropies [@Planck:2013jfk; @Ade:2015lrj]. Now, it is not too much say that inflation is the standard model of the early universe, while we still have some alternatives [@Steinhardt:2001st; @Mukohyama:2009gg; @Creminelli:2010ba]. In the inflationary models, slow-roll inflation, in which a scalar field rolls down on its potential slowly, is known as a common class of the models [@Albrecht:1982wi; @Linde:1983gd; @Steinhardt:1984jj; @Martin:2013tda]. In that class, the potential energy of scalar field plays a role of the vacuum energy to accelerate the universe. Typically, slow-roll inflation ends naturally and changes to the big bang universe after the scalar field reaches to a smaller value than the Planck mass. Taking a few stringent constraints from the observation of cosmic microwave background, the form of scalar potential has been narrowed down until now.
To begin with, however, are there any reasons why inflation is driven by the scalar potential? In reality, we do have another huge class of accelerating the universe, so-called kinetically driven inflation (KDI) [@ArmendarizPicon:1999rj; @Garriga:1999vw; @ArkaniHamed:2003uz; @Kobayashi:2010cm; @Burrage:2010cu; @Kobayashi:2011nu]. In this class, the kinetic terms of scalar field play an important role to inflate the universe. KDI can also transit to the big bang universe by tuning functions of the scalar field in the models. After the first model of KDI [@ArmendarizPicon:1999rj] was proposed, many models of this class have been united in the frame of Horndeski theory [@Kobayashi:2011nu; @Horndeski:1974wa]. Although the quantum fluctuation behaves differently from the Horndeski theory, ghost inflation [@ArkaniHamed:2003uz] holds the same mechanism for inflating the universe as KDI. Some models of KDI exhibit features in sharp contrast with slow-roll inflation [@Garriga:1999vw; @ArkaniHamed:2003uz; @Kobayashi:2010cm; @Burrage:2010cu; @Babich:2004gb; @Chen:2006nt; @Kobayashi:2011pc; @Gao:2011qe; @DeFelice:2011uc; @DeFelice:2013ar], and thus it is expected that the precise observations in the future can distinguish which class has much concordance with the observational data.
So far, however, any unified analytical methods for describing the whole evolution of KDI have not been developed. The main purpose of this article is to reformulate the previous courses of analyses and develop a unified formulation for evaluating the behavior of KDI systematically. We perform a generic analysis of background evolution splitting into the inflationary attractor and linear perturbations around it explicitly. We derive the equation of motion for the linear perturbations with an ansatz where all derivatives with respect to the scalar field $\phi$ are sufficiently small to be treated as perturbations. Further, we consider the scalar quantum fluctuation around the background with a usual scaling dimension. Using a well-known approximation where we keep the first order slow-roll parameters only, we classify the asymptotic form of the mode function by the rank of Hankel function. Using those formalisms, we analyze the behavior of KDI for the shift symmetric case and $\phi$-dependent case separately. Then, we find a few model-independent constraints with which viable models of KDI should be satisfied.
This article is organized as follows. In the next section, we provide a theoretical framework in a model-independent manner and show the inflationary attractor of KDI with an explicit example. In Sec. \[3\], we analyze the background evolution and the quantum fluctuation perturbatively. We apply those formalisms to the shift symmetric case of KDI in Sec. \[4\] and to $\phi$-dependent case in Sec. \[5\]. Sec. \[6\] is devoted for the conclusion. We add Appendix A and B, where technical details are provided. In Appendix \[C\], we shortly review a non-attractor solution, so-called ultra slow-roll solution [@Kinney:2005vj; @Hirano:2016gmv], which the theory can hold potentially.
Attractor solution of kinetically driven inflation {#2}
==================================================
In this article, we intend to derive model-independent features of kinetically driven inflation. Thus, instead of specifying a model, we just impose the following conditions to a theory;
- The theory preserves a symmetry under 4-dimensional diffeomorphism.
- The action of theory consists of a single scalar field $\phi$ and a metric field $g_{\mu\nu}$, i.e., $S[\phi,\, g_{\mu\nu}]$.
- The theory contains one scalar mode and two tensor modes only in the focused energy scale.
- The theory has the flat Friedmann-Robertson-Walker(FRW) background solution.
We give the flat FRW metric as $ds^2 = -N(t)^2dt^2 + a(t)^2d\vec{x}^2$ where $N(t)$ is the lapse function. By variation of the action with respect to $N(t)$ and the background field $\phi=\phi(t)$, we can always get the following form of constraint equation for the background $$\label{Feq}
\mathcal{E}(\phi,\,\dot{\phi},\, H,\, \cdots)=0 \ ,$$ and the equation of motion for $\phi(t)$ $$\begin{aligned}
\label{phi}
\dot{J}+3HJ=P_{,\phi} \ , \end{aligned}$$ where $$\begin{aligned}
\label{j}
J&=J(\phi,\,\dot{\phi},\, H,\, \cdots), \nonumber \\
P &= P(\phi,\,\dot{\phi},\, \ddot{\phi},\, H,\, \dot{H},\,\cdots) \ .\end{aligned}$$ The dot denotes the derivative with respect to time $t$ and $ H\equiv\dot{a}/a$ is the Hubble parameter. The ellipses imply contributions from the higher order time derivatives of $\phi$ and $H$, which are assumed to be negligible throughout this article. The functions ${\cal E},\ J$ and $P$ are supposed to be regular for their arguments. For simplicity, we restrict ${\cal E}$ and $J$ as they include up to the first order time derivatives of $\phi$ and $a$ as the valid terms. In this article, a comma followed with subscript(s) denotes a partial derivative with respect to the subscript(s), e.g., $P_{,\phi}= \partial P/\partial\phi$. The constraint equation (\[Feq\]) is equivalent to the Friedmann equation. For the derivation of these equations, see Appendix \[A\]. If the derivative with respect to $\phi$ in the right hand side of Eq. (\[phi\]) are sufficiently small to exclude it from a leading evolution, or, the theory has a shift symmetry under $\phi\rightarrow \phi +c$ where $c$ is a constant parameter, the equation of motion for $\phi$ is reduced to $$\label{att}
\dot{J} + 3HJ \simeq 0\ .$$ A general solution for $J$ is $$\label{J}
J \simeq c_Ja^{-3}\ ,$$ where $c_J$ is a constant of integration. When the universe expands, $J$ approaches to $0$ in the asymptotic future. If the field velocity $\dot{\phi}$ approaches to a nonzero value as $J\rightarrow 0$, we could obtain a non-trivial quasi-de Sitter attractor solution. Assume that the simultaneous equation system ${\cal E}=0$ and $J=0$ has at least one root solution as $$\label{atts}
\dot{\phi}_0 = f(\phi_0)\neq 0,\quad H_0 = H_0(\phi_0) \ ,$$ where $f$ is a function of $\phi$ and the subscript $0$ stands for the leading order approximation. This root solution can be regarded as an attractor since the motion of background scalar field is determined by the first order differential equation. If the Hubble parameter varies sufficiently slowly with respect to $\phi_0$, the root solution can become a quasi-de Sitter attractor. Inflation induced by this mechanism is referred to as kinetically driven inflation. As an explicit example, we consider the following model [@ArmendarizPicon:1999rj] $$\label{exa}
S = \int d^4x \sqrt{-g}\left[ \frac{M_{\rm Pl}^2}{2}R - K(\phi)X + X^2\right] ,$$ where $M_{\rm Pl}$ is the Planck mass, $R$ is the scalar curvature and $X:= -\frac{1}{2}g^{\mu\nu}\partial_\mu\phi\partial_\nu\phi$. For this action, we get the constraint equation $$\label{ }
{\cal E} = -KX+3X^2 - 3M_{\rm Pl}^2H^2=0\ ,$$ and the equation of motion for $\phi$ $$\begin{aligned}
\label{ }
&\dot{J} + 3HJ = -K_{,\phi} X, \nonumber \\
&J = \dot{\phi}(-K+2X) \ .\end{aligned}$$ When $K>0$ and its derivative term with respect to $\phi$ is enough small to ignore, we obtain an attractor solution $$\label{ }
\dot{\phi}_0 = \pm \sqrt{K(\phi_0)},\quad H_0= \frac{K(\phi_0)}{\sqrt{3}M_{\rm Pl}}\ .$$ In the region where the function $K$ varies sufficiently slowly, the solution becomes a quasi-de Sitter attractor.
As we mentioned, the attractor (\[atts\]) is the leading order approximation of full solution, and the actual motion of scalar field deviates from the attractor. In the following sections, we evaluate the deviation perturbatively.
Perturbations around the KDI attractor {#3}
======================================
To derive the post-attractor solution, we consider isotropic and homogeneous perturbations around the attractor solution. We derive equations of motion for the perturbations using the full equations (\[Feq\]) and (\[phi\]). Further, we consider the quantum fluctuations around the background solution.
Perturbative solution of background
-----------------------------------
We expand the background scalar field and the Hubble parameter as $$\label{pert}
\phi(t) = \phi_0(t) + \phi_1(t),\quad H(t)= H_0(t) + h_1(t) \ .$$ $\phi_1$ and $h_1$ are the linear perturbations around the attractor solution. Introducing a small parameter $|\xi|\ll1$, we make an ansatz that *all* of derivatives with respect to $\phi$ are as small as the linear perturbations: $$\begin{aligned}
\label{ansatz}
&\left\{\phi_1,\,h_1,\,A_{,\phi}\right\} = O(\xi), \\
&A= \sum^\infty_{n=0}\left(\frac{\partial^n}{\partial \phi^n}\right)A_n, \nonumber\end{aligned}$$ where $\{A_n\}$ are arbitrary functions of the background fields without differentiation with respect to $\phi$. Then, the constraint equation (\[Feq\]) is expanded up to $O(\xi)$ as $$\begin{aligned}
\label{Fex}
\mathcal{E}&= \mathcal{E}_0 + \left(\mathcal{E}_{,\dot{\phi}}\right)_0\dot{\phi}_1+ \left(\mathcal{E}_{,H}\right)_0h_1 = 0,\\
{\cal E}_0 &:= {\cal E}(\phi_0, \dot{\phi}_0, H_0, \cdots), \nonumber \end{aligned}$$ where we ignored contributions from the higher order time derivative terms of $\phi$ and $H$. The bracket with subscript $0$ implies that we substitute the leading values of background fields into the derivative in the bracket. Since the leading part ${\cal E}_0$ vanishes by itself, we obtain a relation between the perturbations $$\label{h1}
h_1= -\left(\frac{\mathcal{E}_{,\dot{\phi}}}{\mathcal{E}_{,H}}\right)_0\dot{\phi}_1 .$$ Similarly, the quantity $J$ is expanded up to $O(\xi)$ as $$\begin{aligned}
\label{}
J&=J_0 + \left(J_{,\dot{\phi}}\right)_0\dot{\phi}_1 + \left(J_{,H}\right)_0h_1, \\
J_0&:=J(\phi_0\,,\dot{\phi}_0\,,H_0\,,\cdots) . \nonumber\end{aligned}$$ The leading term $J_0$ vanishes, so we get $$\label{j1}
J= \left\{ \left(J_{,\dot{\phi}}\right)_0 - \left(J_{,H}\frac{\mathcal{E}_{,\dot{\phi}}}{\mathcal{E}_{,H}}\right)_0\right\}\dot{\phi}_1,$$ where we used Eq. (\[h1\]). Substituting Eq. (\[j1\]) into Eq. (\[phi\]), we obtain the equation of motion for $\phi_1$ as $$\begin{aligned}
\label{phi1}
&\ \ \ddot{\phi}_1 + 3H_0\dot{\phi}_1 = \left(\widetilde{P_{,\phi}}\right)_0, \\
&\widetilde{P_{,\phi}}:=\frac{P_{,\phi}}{\left(J_{,\dot{\phi}}\right)_0 - \left(J_{,H}\frac{\mathcal{E}_{,\dot{\phi}}}{\mathcal{E}_{,H}}\right)_0} .\nonumber\end{aligned}$$ This equation is similar to the equation of motion for a canonical scalar field in the FRW spacetime. Note that to make use of Eq. (\[phi1\]), we must impose a condition $$\label{small}
|\dot{\phi}_0|\gg|\dot{\phi}_1|\ ,$$ otherwise we can not treat $\phi_1$ as the linear perturbation around the KDI attractor (\[atts\]).
Before solving the equation of motion for $\phi_1$, we consider the quantum fluctuation of scalar mode around the background in the next subsection.
Quantum fluctuation
-------------------
We assume that after solving all of the constraints the theory has, we obtain the second order action for the scalar fluctuation $\zeta$ as $$\begin{aligned}
\label{s2}
S_2^{(S)}&= \int dtd^3xa^3\left[ {\cal G}_s\dot{\zeta}^2 - \frac{{\cal F}_s}{a^2}(\vec{\nabla} \zeta)^2 + \cdots\right] \ , \\
{{\cal F}_s}&={\cal F}_s(\phi,\,\dot{\phi},\, \ddot{\phi},\, H\,, \dot{H},\,\cdots), \nonumber \\
{{\cal G}_s}&= {\cal G}_s(\phi,\,\dot{\phi},\,\ddot{\phi},\, H,\,\dot{H},\,\cdots) \ . \end{aligned}$$ We may regard $\zeta$ as the comoving curvature perturbation. The ellipses in the functions ${{\cal F}_s}$ and ${{\cal G}_s}$ denote the negligible higher derivative terms of $\phi$ and $H$, while the ellipsis in the action (\[s2\]) represents contributions from the higher derivative terms of $\zeta$. If the higher derivative terms of $\zeta$ become relevant, the scaling of quantum fluctuation would change significantly[^1], which requires individual analysis for each model. Thus, we do not take such cases into account. The generic action (\[s2\]) covers, at least, all classes of Horndeski theory and DHOST theory [@Gleyzes:2013ooa; @Gleyzes:2014dya; @Gleyzes:2014qga; @Gleyzes:2014rba; @Langlois:2017mxy]. For the derivation of Eq. (\[s2\]), see Appendix. \[B\] and [@Tanaka:2012wi] also. Using the background equations (\[h1\]) and (\[phi1\]), we expand ${{\cal F}_s}$ and ${{\cal G}_s}$ up to $O(\xi)$ as $$\begin{aligned}
\label{fs}
{{\cal F}_s}&={{{\cal F}_s}}_0 + \left({{{\cal F}_s}}_{,\dot{\phi}}\right)_0\dot{\phi}_1 + \left({{{\cal F}_s}}_{,\ddot{\phi}}\right)_0\ddot{\phi}_1 \nonumber \\
&\quad + \left({{{\cal F}_s}}_{,H}\right)_0h_1 + \left({{{\cal F}_s}}_{,\dot{H}}\right)_0\dot{h_1}
\nonumber \\
&= {{{\cal F}_s}}_0 + 3H_0\left({{{\cal F}_s}}_1-{{{\cal F}_s}}_2\right)\dot{\phi}_1 + {{{\cal F}_s}}_2\left(\widetilde{P_{,\phi}}\right)_0, \\
\label{gs}
{{\cal G}_s}&= {{{\cal G}_s}}_0 + \left({{{\cal G}_s}}_{,\dot{\phi}}\right)_0\dot{\phi}_1 + \left({{{\cal G}_s}}_{,\ddot{\phi}}\right)_0\ddot{\phi}_1 \nonumber \\
&\quad + \left({{{\cal G}_s}}_{,H}\right)_0h_1 + \left({{{\cal G}_s}}_{,\dot{H}}\right)_0\dot{h_1}
\nonumber \\
&= {{{\cal G}_s}}_0 + 3H_0\left({{{\cal G}_s}}_1-{{{\cal G}_s}}_2\right)\dot{\phi}_1 + {{{\cal G}_s}}_2\left(\widetilde{P_{,\phi}}\right)_0,\end{aligned}$$ where $$\begin{aligned}
\label{ }
&{{{\cal F}_s}}_0:= {{\cal F}_s}(\phi_0,\,\dot{\phi}_0,\,H_0,\,\cdots), \nonumber \\
&{{{\cal F}_s}}_1:= \frac{1}{3H_0}\left\{\left({{{\cal F}_s}}_{,\dot{\phi}}\right)_0 - \left({{{\cal F}_s}}_{,H}
\frac{\mathcal{E}_{,\dot{\phi}}}{\mathcal{E}_{,H}}\right)_0\right\}, \nonumber \\
&{{{\cal F}_s}}_2:= \left({{{\cal F}_s}}_{,\ddot{\phi}}\right)_0 - \left({{{\cal F}_s}}_{,\dot{H}}
\frac{\mathcal{E}_{,\dot{\phi}}}{\mathcal{E}_{,H}}\right)_0, \nonumber \\
&{{{\cal G}_s}}_0:= {{\cal G}_s}(\phi_0,\,\dot{\phi}_0,\,H_0,\,\cdots), \nonumber \\
& {{{\cal G}_s}}_1:= \frac{1}{3H_0}\left\{\left({{{\cal G}_s}}_{,\dot{\phi}}\right)_0
- \left({{{\cal G}_s}}_{,H}\frac{\mathcal{E}_{, \dot{\phi}}}{\mathcal{E}_{,H}}\right)_0
\right\}, \nonumber \\
&{{{\cal G}_s}}_2:= \left({{{\cal G}_s}}_{,\ddot{\phi}}\right)_0 -
\left({{{\cal G}_s}}_{,\dot{H}}\frac{\mathcal{E}_{,\dot{\phi}}}
{\mathcal{E}_{,H}}\right)_0.\end{aligned}$$ The square of speed of sound for $\zeta$ is defined as $$\begin{aligned}
\label{ }
c_s^2 &:= \frac{{{\cal F}_s}}{{{\cal G}_s}} \nonumber \\
&= \frac{{{{\cal F}_s}}_0 + 3H_0\left({{{\cal F}_s}}_1-{{{\cal F}_s}}_2\right)\dot{\phi}_1 + {{{\cal F}_s}}_2\left(\widetilde{P_{,\phi}}\right)_0}{{{{\cal G}_s}}_0 + 3H_0\left({{{\cal G}_s}}_1-{{{\cal G}_s}}_2\right)\dot{\phi}_1 + {{{\cal G}_s}}_2\left(\widetilde{P_{,\phi}}\right)_0}\ .\end{aligned}$$ For the fluctuation being stable, ${{\cal F}_s}$ and ${{\cal G}_s}$ should be in the range $$\label{stability}
{{\cal F}_s}\geq 0, \quad {{\cal G}_s}>0,\quad {{\cal G}_s}\geq {{\cal F}_s},$$ which leads to $0\leq c_s^2 \leq 1$. Then, we define slow-roll parameters as $$\begin{aligned}
\label{sr}
\epsilon_1&:= -\frac{\dot{H}}{H^2},\quad \epsilon_2:= \frac{\dot{\epsilon}_1}{H\epsilon_1}, \nonumber \\
f_{s1}&:= \frac{\dot{{{\cal F}_s}}}{H{{\cal F}_s}},\quad f_{s2}:= \frac{\dot{f}_{s1}}{Hf_{s1}}, \\
g_{s1}&:= \frac{\dot{{{\cal G}_s}}}{H{{\cal G}_s}}, \quad g_{s2}:= \frac{\dot{g}_{s1}}{Hg_{s1}} ,\nonumber\end{aligned}$$ and generalized conformal time as $$\label{ }
d\tau_s := \frac{c_s}{a}dt .$$ The derivative of $c_s/aH$ with respect to $\tau_s$ yeilds $$\label{csahd}
\frac{d}{d\tau_s}\left(\frac{c_s}{aH}\right) = -1+\epsilon_1+\frac{1}{2}(f_{s1}-g_{s1}).$$ If we regard the slow-roll parameters $\epsilon_1$, $f_{s1}$ and $g_{s1}$ as constants, the integral of Eq. (\[csahd\]) reduces to $$\label{csah}
\frac{c_s}{aH}\simeq -\left\{1- \epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right\}\tau_s .$$ This approximation implies practically that we ignore contributions from the second order products of slow-roll parameters $\epsilon_1\epsilon_2$, $f_{s1}f_{s2}$ and $g_{s1}g_{s2}$. In fact, from Eq. (\[csahd\]), we obtain $$\begin{aligned}
\label{tau}
\tau_s &= -\int^{\tau_s} \frac{\frac{d}{d\tau_s'}\left(\frac{c_s}{aH}\right)}{\left\{1-
\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right\}}d\tau_s' \nonumber \\
&= -\frac{c_s}{aH}\frac{1}{\left\{1- \epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right\}} \nonumber \\
&\quad + \int^{\tau_s}\frac{\epsilon_1\epsilon_2
+ \frac{1}{2}(f_{s1}f_{s2}-g_{s1}g_{s2})}{\left\{1- \epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right\}^2}d\tau_s' \ .\end{aligned}$$ To get the relation (\[csah\]), we need to discard the integral consisting of $\epsilon_1\epsilon_2$, $f_{s1}f_{s2}$ and $g_{s1}g_{s2}$ in the last line of Eq. (\[tau\]). We always employ this approximation. In addition, we impose a condition $$\label{taut}
1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1}) > 0$$ which enforces $\tau_s$ to run from $-\infty$ to $0$ when the universe expands.
Discarding the contribution from $\epsilon_1\epsilon_2$, $f_{s1}f_{s2}$ and $g_{s1}g_{s2}$, we obtain the equation of motion for the scalar fluctuation in Fourier space as $$\begin{aligned}
\label{zeta}
&\partial_{\tau_s}^2\zeta_k - \frac{2\nu_s -1}{\tau_s} \partial_{\tau_s}\zeta_k+ k^2\zeta_k=0 , \\
\label{nus}
&\nu_s:= \frac{3-\epsilon_1+ g_{s1}}{2-2\epsilon_1-f_{s1}+g_{s1}} .\end{aligned}$$ The linearly independent solutions of Eq. (\[zeta\]) are $$\begin{aligned}
\label{}
\zeta_k= &\frac{1}{2}\sqrt{\frac{\pi}{2}}(-k\tau_s)^{3/2}\frac{H\left(1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right){{\cal G}_s}^{1/4}}{k^{3/2}{{\cal F}_s}^{3/4}} \nonumber \\
&\times H_{\nu_s}^{(1)}(-k\tau_s) \end{aligned}$$ and its complex conjugate, where $H_{\nu_s}^{(1)}$ is the Hankel function of the first kind. We normalized the solution so that deep inside the horizon, the mode function $\zeta_k$ behaves as in Minkowski spacetime: $$\label{ }
\zeta_k\approx \frac{1}{\sqrt{2}a({{\cal F}_s}{{\cal G}_s})^{1/4}}\frac{\mathrm{e}^{-ik\tau_s}}{\sqrt{2k}} \qquad \text{for}\quad |k\tau_s|\rightarrow \infty.$$ On superhorizon scales where $|k\tau_s| \ll 1$, the asymtotic form of $\zeta_k$ for $\nu_s>0$ is $$\begin{aligned}
\label{ }
|\zeta_k| \simeq &\frac{1}{2}(-k\tau_s)^{3/2-\nu_s}2^{\nu_s-3/2}\frac{\Gamma(\nu_s)}{\Gamma(3/2)}\nonumber \\
&\times \frac{H\left(1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right){{\cal G}_s}^{1/4}}{k^{3/2}{{\cal F}_s}^{3/4}} .\end{aligned}$$ The scalar spectrum and the spectral index are obtained as $$\begin{aligned}
\label{ }
&P_\zeta := \frac{k^3}{2\pi^2}|\zeta_k|^2 = \frac{\gamma_s}{2}(-k\tau_s)^{3-2\nu_s}\frac{{{\cal G}_s}^{1/2}}{{{\cal F}_s}^{3/2}}\frac{H^2}{4\pi^2} , \\
&n_s-1:= \frac{d\text{ln}P_\zeta}{d\text{ln}k} = 3-2\nu_s,\end{aligned}$$ where $\gamma_s:= 2^{2\nu_s-3}|\Gamma(\nu_s)/\Gamma(\frac{3}{2})|^2\left(1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right)^2$. For $\nu_s<0$, the asymptotic form of $\zeta_k$ on superhorizon scales is $$\begin{aligned}
\label{ }
|\zeta_k| &\simeq \frac{|A|}{2}(-k\tau_s)^{3/2+\nu_s}\frac{H\left(1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right){{\cal G}_s}^{1/4}}{k^{3/2}{{\cal F}_s}^{3/4}}, \\
A &:= \sqrt{\frac{\pi}{2}}2^{-\nu_s}\left[ \frac{1}{\Gamma(1+\nu_s)} + \frac{{\rm cos}(\pi\nu_s)\Gamma(-\nu_s)}{\pi i}\right] \nonumber .\end{aligned}$$ In this case, the scalar spectrum and the spectral index are obtained as $$\begin{aligned}
\label{}
&P_\zeta = \frac{|A|^2\left(1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right)^2}{2}(-k\tau_s)^{3+2\nu_s}\nonumber \\
&\qquad\ \times \frac{{{\cal G}_s}^{1/2}} {{{\cal F}_s}^{3/2}}\frac{H^2}{4\pi^2} ,\\
\label{nm}
&n_s-1 = 3+2\nu_s .\end{aligned}$$ For $\nu_s=0$, the asymptotic form of $\zeta_k$ on superhorizon scales becomes $$\begin{aligned}
\label{}
|\zeta_k|\simeq &\sqrt{\frac{1}{2\pi}}(-k\tau_s)^{3/2}{\rm ln}(-k\tau_s)\nonumber \\
&\times \frac{H\left(1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right){{\cal G}_s}^{1/4}}{k^{3/2}{{\cal F}_s}^{3/4}} .\end{aligned}$$ The scalar spectrum and the spectral index are obtained as $$\begin{aligned}
\label{}
P_\zeta =& \frac{\left(1-\epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right)^2}{\pi}(-k\tau_s)^{3}\left\{{\rm ln}(-k\tau_s)\right\}^2\nonumber \\
&\times \frac{{{\cal G}_s}^{1/2}} {{{\cal F}_s}^{3/2}}\frac{H^2}{4\pi^2}, \\
n_s-1 &= 3 + \frac{2}{{\rm ln}(-k\tau_s)} \simeq 3 .\end{aligned}$$ We note that if $\nu_s\leq0$ or, equivalently, $3-\epsilon + g_{s1}\leq0$, the scalar fluctuation grows even on superhorizon scales. The two independent solutions of $\zeta$ in real space on superhorizon scales are obtained as $$\begin{aligned}
\label{}
\zeta= \text{const.} \quad \text{and} \quad \int^t \frac{dt'}{{{\cal G}_s}a^3}.\end{aligned}$$ We denote the latter solution as the decaying mode. The decaying mode can be expressed as $$\begin{aligned}
\label{}
\zeta_{\rm decay}\propto\begin{cases}
& a^{-(3-\epsilon + g_{s1})}\quad \text{for}\quad 3-\epsilon + g_{s1}\neq0 \\
& {\rm ln}\,a \ \,\qquad\qquad \text{for} \quad 3-\epsilon + g_{s1}=0\ .
\end{cases} \end{aligned}$$ Thus, the decaying mode does decay if $3-\epsilon_1+g_{s1}>0$, while it grows actually if $3-\epsilon_1+g_{s1}\leq0$. Using these formalisms, we evaluate the behavior of background and quantum fluctuation dividing into two cases in the following sections.
Case I: shift symmetric system {#4}
==============================
First, we consider a simpler case that the action has a shift symmetry under $\phi \rightarrow \phi+c$, where $c$ is a constant parameter. Provided at least one root solution of the KDI attractor, we get a $\phi$-*independent* attractor solution $$\label{}
\dot{\phi}_0= c_*= \text{const} \neq 0,\quad H_0 = \text{const} > 0,$$ which reflects the shift symmetry of the action. Since $\phi$-dependence gets lost by the shift symmetry, Eq. (\[phi1\]) reduces to $$\ddot{\phi}_1 + 3H_0\dot{\phi}_1= 0 .$$ The solution is $$\label{phi1a}
\dot{\phi}_1 \simeq \frac{c_1}{a^3},$$ where $c_1$ is a constant of integration. The condition (\[small\]) can be easily satisfied by choosing $c_1$ properly. The behavior of perturbation $\phi_1$ is almost the same as the scalar field in the original ultra slow-roll inflation [@Kinney:2005vj] (See Appendix.). Therefore, when the theory possesses the shift symmetry, the KDI attractor becomes an exact de Sitter solution, and the linear perturbation around it behaves like ultra slow-roll inflation. Then, we consider features of the quantum fluctuation for the shift symmetric system. Using Eq. (\[phi1a\]), we obtain the functions (\[fs\]) and (\[gs\]) as $$\begin{aligned}
\label{}
{{\cal F}_s}& \simeq {{{\cal F}_s}}_0 + 3H_0c_1\left({{{\cal F}_s}}_1-{{{\cal F}_s}}_2\right)a^{-3} \ , \\
{{\cal G}_s}& \simeq {{{\cal G}_s}}_0 +3H_0c_1\left({{{\cal G}_s}}_1-{{{\cal G}_s}}_2\right)a^{-3} \ ,\end{aligned}$$ and the speed of sound as $$\begin{aligned}
\label{}
c_s^2 \simeq \frac{{{{\cal F}_s}}_0 + 3H_0c_1\left({{{\cal F}_s}}_1-{{{\cal F}_s}}_2\right)a^{-3}}{{{{\cal G}_s}}_0 + 3H_0c_1\left({{{\cal G}_s}}_1-{{{\cal G}_s}}_2\right)a^{-3}} \ .\end{aligned}$$ We note that for the shift symmetric system, all the functions but $a$ in $c_s^2$ become constants. We can classify the shift symmetric system into three types depending on the values of ${{{\cal F}_s}}_0$ and ${{{\cal G}_s}}_0$: (i) ${{{\cal G}_s}}_0\geq{{{\cal F}_s}}_0 > 0$ (ii) ${{{\cal F}_s}}_0=0$ and ${{{\cal G}_s}}_0 > 0$ (iii) ${{{\cal F}_s}}_0={{{\cal G}_s}}_0=0$. We do not consider any other types since they do not satisfy the stability condition (\[stability\]). For the type (i), after several e-folds, we eventually reach to a phase in which $$\begin{aligned}
\label{ }
{{{\cal F}_s}}_0 &\gg 3H_0c_1\left({{{\cal F}_s}}_1-{{{\cal F}_s}}_2\right)a^{-3}, \nonumber \\
{{{\cal G}_s}}_0 &\gg 3H_0c_1\left({{{\cal G}_s}}_1-{{{\cal G}_s}}_2\right)a^{-3} \ .\end{aligned}$$ Then, the slow-roll parameters (\[sr\]) almost behave as $$\begin{aligned}
\label{}
\epsilon_1\propto f_{s1}\propto g_{s1} \propto a^{-3} , \quad \epsilon_2 \simeq f_{s2}\simeq g_{s2}\simeq -3 \ . \end{aligned}$$ We obtain $O(1)$ values for $\epsilon_2,\ f_{s2}$ and $g_{s2}$. In this case, the contribution from the second order products $\epsilon_1\epsilon_2$, $f_{s1}f_{s2}$ and $g_{s1}g_{s2}$ in Eq. (\[tau\]), which are discarded in our formalism, becomes comparable with the contribution from $\epsilon_1$, $f_{s1}$ and $g_{s1}$: $$\begin{aligned}
\label{lead}
&\int^{\tau_s}\frac{\epsilon_1\epsilon_2
+ \frac{1}{2}(f_{s1}f_{s2}-g_{s1}g_{s2})}{\left\{1- \epsilon_1-\frac{1}{2}(f_{s1}-g_{s1})\right\}^2}d\tau_s' \nonumber \\
\simeq &\int^{\tau_s}\left\{-3\epsilon_1- \frac{3}{2}(f_{s1}-g_{s1})\right\}d\tau_s' \nonumber \\
\simeq &\frac{3c_s}{4aH}\left\{\epsilon_1 +\frac{1}{2}(f_{s1}-g_{s1})\right\}\ ,\end{aligned}$$ where we picked up the leading part of integral only. Therefore, we must dismiss $\epsilon_1$, $f_{s1}$ and $g_{s1}$ also from Eq. (\[tau\]). This results in $$\label{ }
\quad\nu_s = \frac{3}{2},\quad n_s-1 = 0.$$ Thus, we have no scalar spectral tilt for the type (i) theories[^2] as well as the original ultra slow-roll inflation [@Kinney:2005vj]. In the case of type (i) theories, however, the scalar fluctuation does never grow on superhorizon scales unlike ultra slow-roll inflation since $\nu_s$ for the type (i) theories is always positive, while $\nu_s$ for ultra slow-roll inflation would be negative (See Appendix \[C\]).
For the type (ii), after several e-folds, we get $$\label{ }
{{{\cal F}_s}}_0 = 0,\quad {{{\cal G}_s}}_0 \gg 3H_0c_1\left({{{\cal G}_s}}_1-{{{\cal G}_s}}_2\right)a^{-3} \ .$$ The slow-roll parameters almost behave as $$\begin{aligned}
\label{}
\epsilon_1 \propto g_{s1}\propto a^{-3} , \quad \epsilon_2 \simeq f_{s1} \simeq g_{s2} = -3, \quad f_{s2}= 0 . \end{aligned}$$ Similarly to the type (i), we must dismiss $\epsilon_1$ and $g_{s1}$ from Eq. (\[tau\]) while keeping $f_{s1}$. This leads to the following spectral tilt $$\label{ii}
\nu_s = \frac{3}{5},\quad n_s-1 = \frac{9}{5}.$$ Thus, we have the blue spectrum for the type (ii) theories. Lastly, for the type (iii) where ${{{\cal F}_s}}_0 = {{{\cal G}_s}}_0 = 0$, we get the following slow-roll parameters $$\begin{aligned}
\label{}
\epsilon_1 \propto a^{-3} , \quad \epsilon_2 \simeq f_{s1} = g_{s1} = -3,\quad f_{s2}=g_{s2}= 0 .\end{aligned}$$ We dismiss $\epsilon_1$ from Eq. (\[tau\]) while keeping $f_{s1}$ and $g_{s1}$. The spectral tilt reduces to $$\label{ }
\nu_s = 0,\quad n_s-1 = 3 ,$$ and we have the blue spectrum for the type (iii) theories also. Since $\nu_s=0$, the scalar fluctuation of the type (iii) theories grows logarithmically in terms of $a$ on superhorizon scales. Given these results, we find that the shift symmetric models of KDI within our generic framework cannot create the scalar fluctuation satisfied with $n_s\sim0.968$ if the higher order derivatives of $\zeta$ in the action remain irrelevant or do not exist. For example, the action of scalar fluctuation of Horndeski theory is described as Eq. (\[s2\]) [@Kobayashi:2011nu], so that all the shift symmetric models of KDI in Horndeski theory can not explain the observation. To obtain viable models of KDI, we need to introduce other sources for the scalar fluctuation, or, break the shift symmetry [@ArmendarizPicon:1999rj]. In the next section, we consider the case where the shift symmetry is broken, but the leading background evolution is still illustrated by the KDI attractor .
Case II: $\phi$-dependent system {#5}
================================
We break the shift symmetry and accept the system to depend on $\phi$. This case has been considered in the previous models to a limited extent [@ArmendarizPicon:1999rj; @ArkaniHamed:2003uz; @Kobayashi:2010cm; @Burrage:2010cu]. In this section, we perform a model-independent analysis for this case.
Provided at least one root solution of the KDI attractor depending on $\phi_0$, we obtain a leading solution (\[atts\]) and the eqution of motion for $\phi_1$ (\[phi1\]): $$\begin{aligned}
\label{}
&\dot{\phi}_0= f(\phi_0),\quad H_0=H_0(\phi_0), \nonumber \\
&\ddot{\phi}_1+3H_0\dot{\phi}_1=(\widetilde{P_{,\phi}})_0 \ . \nonumber\end{aligned}$$ If $|\ddot{\phi_1}| \ll |(\widetilde{P_{,\phi}})_0|$, we obtain an approximate solution as $$\label{ }
\dot{\phi}_1 \simeq \frac{(\widetilde{P_{,\phi}})_0}{3H_0} \ .$$ This indicates that the linear perturbation of $\phi$-dependent system “slow-rolls" around the quasi-de Sitter KDI attractor. Thus, if we can neglect the second order time derivatives of $\phi_1$, the whole evolution of background is determined by $\phi$-dependence of $f$, $H_0$ and $(\widetilde{P_{,\phi}})_0$ similarly to slow-roll inflation. The slow-roll condition for the perturbation follows from $|\ddot{\phi_1}| \ll |(\widetilde{P_{,\phi}})_0|$ up to $O(\xi)$ as $$\begin{aligned}
\label{eta}
\eta_\phi:= \frac{f(P_{,\phi\phi})_0}{3H_0(P_{,\phi})_0},\quad |\eta_\phi|\ll 1 \ .\end{aligned}$$ By imposing $|\dot{\phi}_0|\gg|\dot{\phi}_1|$, we obtain another condition as $$\begin{aligned}
\label{ephi}
\epsilon_\phi:= \frac{(P_{,\phi})_0}{3fH_0\left\{ (J_{,\dot{\phi}})_0 - \left(J_{,H}\frac{\mathcal{E}_{,\dot{\phi}}}{\mathcal{E}_{,H}}\right)_0\right\}},\quad |\epsilon_\phi| \ll 1 . \end{aligned}$$ The condition $|\epsilon_\phi|\ll 1$ generalizes the equivalent conditions in k-inflation and Ghost inflation. Further, to obtain a quasi-de Sitter expansion, we need the following condition $$\begin{aligned}
\label{e1}
\epsilon_1 &\simeq -\frac{f}{H_0^2}\left\{ {H_0}_{,\phi_0} - 3H_0\left(\frac{\mathcal{E}_{,\dot{\phi}}}{\mathcal{E}_{,H}}\right)_0\epsilon_\phi\eta_\phi \right\}\nonumber \\
&\simeq -\frac{f{H_0}_{,\phi_0}}{H_0^2},\nonumber \\
|\epsilon_1|&\ll 1 \ ,\end{aligned}$$ where we ignored $\epsilon_\phi\eta_\phi$ since it is the second order product of small parameters. For the $\phi$-dependent system of KDI, the above three conditions are essential to realizing a quasi-de Sitter expansion mimicking slow-roll inflation.
Due to the $\phi$-dependence of system, we could stop the acceleration dynamically if the slow-roll parameter $\epsilon_1$ varies to unity $$\label{e11}
|\epsilon_1|\simeq \left|\frac{f{H_0}_{,\phi_0}}{H_0^2}\right| \sim 1 \ .$$ Note that Eq. (\[e11\]) is a *necessary* condition to quit the acceleration and reach to the graceful exit since depending on the functional form of $H_0$, $|\epsilon_1|$ could return to a much less value than unity again. Thus, for ending up with the graceful exit, it is essential to restrict $H_0$ so as not to inflate the universe again. Then, we turn to consider the quantum fluctuation. In the $\phi$-dependent case, the coefficient functions ${{\cal F}_s}$ and ${{\cal G}_s}$ are given as $$\begin{aligned}
\label{}
{{\cal F}_s}&\simeq {{{\cal F}_s}}_0 + {{{\cal F}_s}}_1(\widetilde{P_{,\phi}})_0 , \nonumber \\
{{\cal G}_s}&\simeq {{{\cal G}_s}}_0 + {{{\cal G}_s}}_1(\widetilde{P_{,\phi}})_0 .\end{aligned}$$ We derive the slow-roll parameter $f_{s1}$ up to $O(\xi)$ as $$\begin{aligned}
\label{fs1}
f_{s1}&\simeq \frac{f\left( {{{\cal F}_s}}_0' + 9{{{\cal F}_s}}_1H_0^2\epsilon_\phi\eta_\phi\right)}{H\left({{{\cal F}_s}}_0 + {{{\cal F}_s}}_1(\widetilde{P_{,\phi}})_0\right)} \nonumber \\
&\simeq\begin{cases}
\frac{f{{{\cal F}_s}}'_0}{H_0{{{\cal F}_s}}_0} \qquad \text{for}\quad {{{\cal F}_s}}_0 \gg {{{\cal F}_s}}_1(\widetilde{P_{,\phi}})_0, \\
3\eta_\phi\qquad \quad\text{for} \quad {{{\cal F}_s}}_0 = 0.
\end{cases} \end{aligned}$$ We do not need to consider any other cases as long as we impose the ansatz (\[ansatz\]). The prime on ${{{\cal F}_s}}_0$ denotes the total derivative with respect to $\phi_0$, which is written up to $O(\xi)$ as $$\begin{aligned}
\label{ }
{{{\cal F}_s}}_0' :&= \frac{d{{{\cal F}_s}}_0}{d\phi_0} \nonumber \\
&= {{{{\cal F}_s}}_0}_{,\phi_0} + {{{{\cal F}_s}}_0}_{,\dot{\phi}_0}f_{,\phi_0} + {{{{\cal F}_s}}_0}_{,\ddot{\phi}_0}ff_{,\phi_0\phi_0} \nonumber \\
&+ {{{{\cal F}_s}}_0}_{,H_0}{H_0}_{,\phi_0} + {{{{\cal F}_s}}_0}_{,\dot{H}_0}f{H_0}_{,\phi_0\phi_0} \ .\end{aligned}$$ $g_{s1}$ is derived in the same way as $f_{s1}$ with replacing ${{{\cal F}_s}}_0$ and ${{{\cal F}_s}}_1$ to ${{{\cal G}_s}}_0$ and ${{{\cal G}_s}}_1$. Thus, if we associate with the ansatz (\[ansatz\]) and the condition (\[eta\]), $f_{s1}$ and $g_{s1}$ remain at smaller values than unity. Then, to create the spectral tilt consistent to the observations, we must keep at least one of the slow-roll parameters in Eq. (\[tau\]) while we get rid of the integral consisting of the second order products of slow-roll parameters. We do not specify what conditions we need for realizing this situation since it highly depends on the models. In general, however, the conditions for getting rid of the second order products of slow-roll parameters reduce to constraints for the third or higher order derivatives with respect to $\phi$, e.g. $P_{,\phi\phi\phi}$, while Eq. (\[eta\])-(\[e1\]) constrain up to the second order derivatives with respect to $\phi$. Thus, we must impose those constraints independently.
In addition to the above conditions, we must impose the stability condition (\[stability\]) and a few observational constraints to the spectral index $n_s$, tensor-to-scalar ratio $r$, non-gaussianity parameter $f_{NL}$ etc. to obtain viable models of KDI. It might seem that models of KDI are bounded so tightly that there remain only a few regions for successful models. Practically, however, we can satisfy all of the constraints by tuning a few functions of $\phi$ in the action. In the case of action (\[exa\]), for instance, all the above constraints reduce to constraints for the function $K(\phi)$ since, in the action, only $K(\phi)$ depends on $\phi$. If we introduce more functions of $\phi$ into the action, easier it would become to satisfy the constraints.
Summary {#6}
=======
We have constructed the model-independent analytical method for kinetically driven inflation (KDI). In the analysis, we have evaluated the cosmological background splitting into the inflationary attractor and the linear perturbations around it. Establishing the ansatz that all of the derivatives with respect to the scalar field are so small as to be treated as the perturbations, we have derived the generic equation of motion for the linear perturbation of scalar field. We have also considered the quantum fluctuation around the background, which has a usual scaling. Given the generic action of fluctuation, we have classified the scalar mode function to three cases by the rank of Hankel function under the approximation discarding the integral of the second order products of slow-roll parameters.
Using these formalisms, we have investigated the behaviors of background and quantum fluctuation dividing into two cases. First, we have considered the case that the system has the scalar shift symmetry. In this case, we have obtained a description of the background evolution that the perturbation of scalar field behaves almost as ultra slow-roll inflation around the exact de Sitter attractor. However, the ultra slow-roll motion of perturbation causes rapid changes of the slow-roll parameters, so that we must dismiss all or a few slow-roll parameters from the spectral index to apply our formalisms. As a result, it turns out that all the shift symmetric models of KDI, which are incorporated in our generic framework, cannot create the scalar spectral tilt consistent to the observations. Therefore, to obtain viable models of KDI within our framework, we need to add another source of the scalar fluctuation, or, break the shift symmetry. Then, we have considered the shift breaking case. In this case, the background evolution is described as the perturbation of scalar field “slow-rolls" around the quasi-de Sitter attractor. Imposing a few theoretical requirements, we have derived the four essential conditions for the shift breaking models of KDI associated with the graceful exit. These conditions generalize the corresponding conditions of the previous models.
We have obtained the model-independent properties of KDI only with a few assumptions. If new models of KDI based on new scalar-tensor theories are constructed in the future, we can easily discriminate which models are viable or not by using our formalisms as far as the new theories belong in our generic framework.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work was supported partially by the National Natural Science Foundation of China under Grant Nos. 11475065 and 11690021.
Derivation of the general form of background equations {#A}
======================================================
We start from the action $S[\phi,g_{\mu\nu}]$. Taking the flat FRW background, we obtain the action for the background as $$\begin{aligned}
\label{}
S_{\rm bg}= \int dtd^3xNa^3P\left[\phi,\,\dot{\phi},\, \ddot{\phi}, \, H,\, \dot{H},\,N,\, \cdots\right] \ ,\end{aligned}$$ where the ellipsis denotes the higher order time derivative terms. Differentiating this with respect to $N$, we obtain a Euler-Lagrange equation. That equation must be a constraint equation corresponding to the Fridmann equation, otherwise we have extra degrees of freedom since we cannot determine the evolution of the background with a initial condition for ($\phi,\,\dot{\phi})$ alone. Then, taking the flat FRW solution $N=1$, we obtain the general form of constraint equation as Eq. (1).
Next, we consider the variation of action with respect to $\phi$. The functional $P$ varies with respect to $\phi$ as $$\begin{aligned}
\label{}
\delta_\phi P= \frac{\partial P}{\partial \phi}\delta \phi + \sum_{n=1}\frac{\partial P}{\partial \phi^{(n)}}\left(\frac{d}{dt}\right)^n\delta \phi \ ,\end{aligned}$$ where $\phi^{(n)}$ implies the n-th order time derivative of $\phi$. The variation of action with respect to $\phi$ can be reduced after a integration by parts to $$\begin{aligned}
\label{}
\delta_\phi S_{\rm bg}= &\int dt d^3x\left\{ Na^3\frac{\partial P}{\partial \phi} \right.\nonumber \\
&+ \left. \sum_{n=1}(-1)^n\left(\frac{d}{dt}\right)^n\left(Na^3\frac{\partial P}{\partial \phi^{(n)}}\right)\right\}\delta \phi \ .\end{aligned}$$ Differentiating the action with respect to $\phi$ and taking $N=1$, we obtain $$\begin{aligned}
\label{}
&\frac{d}{dt}\left( a^3J\right)= P_{,\phi} \ , \\
&J:= \sum_{n=1}\frac{(-1)^{n-1}}{a^3}\left(\frac{d}{dt}\right)^{n-1}\left(a^3\frac{\partial P}{\partial \phi^{(n)}}\right)\ .\end{aligned}$$ Thus, we can always get the form of Eq. (2) if we have the flat FRW background solution.
Derivation of the massless action for the scalar fluctuation {#B}
============================================================
In this appendix, we show how to get the massless action (\[s2\]) for the single field inflation. We consider only the case we can use the scalar field $\phi$ as a clock field and take the unitary gauge where $$\begin{aligned}
\label{}
\phi = \phi(t) + \delta \phi,\quad \delta \phi = 0 \ . \end{aligned}$$ We can always take this gauge within the finite time if on the inflationary background, the theory has 2 tensor modes and 1 scalar mode only in the focused energy range. Using the ADM variables [@Arnowitt:1962hi] $$\label{ }
ds^2 = -N^2dt^2 + h_{ij}(N^idt + dx^i)(N^jdt + dx^j) \ ,$$ we can reduce the original action $S[\phi, g_{\mu\nu}]$ to the unitary gauge action $$\begin{aligned}
\label{sug}
S_{\rm u.g}= \int dtd^3x \sqrt{h}\,{\cal L}_{\rm u.g}[N,\,K_{ij},\,{}^3R_{ij},\,h^{ij},\,t,\,\cdots] \ . \end{aligned}$$ $K_{ij}$ and ${}^3R_{ij}$ are the extrinsic curvature and the intrinsic curvature respectively, and the ellipses imply irrelevant terms involved with higher order derivatives. At this point, the unitary gauge action lost the general diff-invariance but still preserves the spatial diff-invariance.
Conversely, when we have a clock field, we can recover the general diff-invariance by identifying the timelike unit normal $n_\mu$ to the product of clock field [@Cheung:2007st; @Gao:2014soa] as $$\begin{aligned}
\label{trick}
n_\mu&= -\frac{\partial_\mu\phi}{\sqrt{2X}} \ .\end{aligned}$$ This procedure is regarded as the Stückelberg trick. If we re-introduce the scalar field as described above, we can replace all of the ingredients in the unitary gauge action to the products of $\phi$ and $g_{\mu\nu}$, for example, $$\begin{aligned}
\label{}
h_{ij}\rightarrow g_{\mu\nu} + n_\mu n_\nu = g_{\mu\nu} +\frac{\partial_\mu\phi\partial_\nu\phi}{2X} \ .\end{aligned}$$ Using the Stückelberg trick, we find that we should treat the first order derivatives of $N$ in the unitary gauge action as irrelevant operators. First, we consider the acceleration $a_\mu:= n^\nu\nabla_\nu n_\mu = (0,\, \partial_i{\rm ln}N)$. If we recover the general diff-invariance by the trick, the terms consisting of acceleration are converted to the higher order derivative terms, for instance, as $$\begin{aligned}
\label{}
a_\mu a^\mu = \frac{\nabla_\mu X\nabla^\mu X}{(2X)^2} + \frac{(\partial^\mu\phi\nabla_\mu X)^2}{(2X)^3} \ .\end{aligned}$$ It is well-known that in general, those higher order terms introduce extra degrees of freedom which cause the ghost instability and break the unitarity [@Woodard:2006nt]. Therefore, if we prohibit any accidental cancellations between the higher order derivative terms originated from the acceleration and the other higher order derivative terms, we should ignore all the terms involved with the acceleration from the unitary gauge action as irrelevant operators. For the same reason, we should ignore the time derivative terms of $N$ also, such as $N_\bot:= \dot{N}-N^i\partial_iN$. Thus, we regard the unitary gauge action provided in Eq. (\[sug\]) as the most generic action for the single field inflation, which we can obtain naturally. We will discuss later about the case where the accidental cancellations does occur.
To fix the residual gauge degrees of freedom, we take the comoving gauge as follows: $$\begin{aligned}
\label{com}
h_{ij}&= {\rm e}^{2(\rho + \zeta)}[{\rm e}^\mathbf{\gamma}]_{ij}, \nonumber \\
{\rm e}^\rho &:= a(t),\quad \partial_i\gamma_{ij} = 0=\gamma_{ii}\ ,\end{aligned}$$ where $\zeta$ is the comoving curvature perturbation and $\gamma_{ij}$ is the tensor fluctuation. Hereafter, we ignore $\gamma_{ij}$ and set $[{\rm e}^\mathbf{\gamma}]_{ij}$ to $\delta_{ij}$ since we intend to derive the quadratic action of $\zeta$. Then, the action for scalar fluctuations reduces to $$\begin{aligned}
\label{sc}
S&= \int dtd^3x\,{\rm e}^{3(\rho+\zeta)}{\cal L}_c[N,\,K^i_{\ j},\,{}^3R^i_{\ j},\, \delta^i_j,\,t,\,\cdots] \ ,\end{aligned}$$ where we rewrote the ingredients of Lagrangian density so as not to include $h^{ij}$ for later convenience. The extrinsic and intrinsic curvatures in the comoving gauge are given as $$\begin{aligned}
\label{K}
K^i_{\ j}&= \frac{h^{ik}}{2N}\left(\dot{h}_{kj} - 2D_{(k}N_{j)}\right) \nonumber \\
&=\frac{1}{N}\left[ (H+\dot{\zeta})\delta_{ij} \right.\nonumber \\
&\quad\left. -{\rm e}^{-2(\rho+\zeta)}\left(\partial_{(i}N_{j)}
-2\partial_{(i}\zeta N_{j)} + \partial_k\zeta N_k\delta_{ij}\right)\right] \ , \\
\label{R}
{}^3R^i_{\ j}&= -{\rm e}^{-2(\rho+\zeta)}\left(\partial_i\partial_j \zeta +\partial^2\zeta\delta_{ij}
-\partial_i\zeta\partial_j\zeta + (\partial\zeta)^2\delta_{ij}\right) \ .
$$ We do not distinguish $\delta^i_j$, $\delta^{ij}$ and $\delta_{ij}$, and we raise or lower the indices of spatial derivatives by $\delta_{ij}$. $\partial^2$ denotes $\partial_i\partial_i$. We note that in the action (\[sc\]), there does not exist the quadratic term of $N_i$ without spatial differentiation since the unitary gauge action (\[sug\]) preserves the spatial diff-invariance.
Differentiating the action (\[sc\]) with respect to $N$ and $N_i$, we obtain the Hamiltonian constraint and the momentum constraint respectively as $$\begin{aligned}
\label{conN}
\frac{\delta S}{\delta N}& = {\rm e}^{3(\rho+\zeta)}\frac{\partial {\cal L}_c}{\partial N} = 0 \ , \\
\label{mom}
\frac{\delta S}{\delta N_i}& = \sqrt{h}h^{(ik}
D_j\left( \frac{1}{N}\frac{\partial {\cal L}_c}{\partial K^k_{\ j)}}\right) \nonumber \\
&= {\rm e}^{(\rho+\zeta)}
D_j\left( \frac{1}{N}\frac{\partial {\cal L}_c}{\partial K^{(i}_{\,\ j)}}\right) = 0\ .\end{aligned}$$ Based on the structure of indices, we can decompose the momentum constraint (\[mom\]) as $$\begin{aligned}
\label{mom2}
&{\rm e}^{(\rho+\zeta)}D_j\left(F\delta^j_i + F_{(i}^{\,\ j)}\right) = 0 \ , \\
& F=F[N,\,K^i_{\ j},\,{}^3R^i_{\ j},\, \delta^i_j,\,t] ,\nonumber \\
&F^{\ j}_i = F^{\ j}_i[N,\,K^i_{\ j},\,{}^3R^i_{\ j},\, \delta^i_j,\,t] \ ,\nonumber \end{aligned}$$ where $F$ is a scalar and $F^{\ j}_i$ is a tensor on the spatial hypersurface. To obtain the quadratic action, we need to solve the constraints only up to the first order of fields. Expanding the Hamiltonian constraint (\[conN\]) up to the first order of fields, we obtain $$\begin{aligned}
\label{ham}
{\rm e}^{3(\rho+\zeta)}\frac{\partial {\cal L}_c}{\partial N} = {\rm e}^{3\rho}(1+3\zeta)\left.\frac{\partial {\cal L}_c}{\partial N}\right|_{\rm 0} + {\rm e}^{3\rho}\left.\frac{\partial {\cal L}_c}{\partial N}\right|_1 = 0 \ ,\end{aligned}$$ where the subscripts denote the order of fields. The zeroth order part, however, corresponds to the constraint equation of background, so it vanishes. Then, from the remaining term, we obtain a generic form of Hamiltonian constraint as $$\label{ham2}
c_1N_1+ c_2\partial_iN_i + c_3\dot{\zeta} + c_4{\rm e}^{-2\rho}\partial^2\zeta= 0 \ ,$$ where $c_{1\sim4}$ are functions of time, and $N = 1 + N_1$. Note that from Eq. (\[K\]) and (\[R\]), each spatial derivative and each $N_i$ must be followed with a factor ${\rm e}^{-\rho}$. We write the factor separately only for the spatial derivative of $\zeta$. The momentum constraint (\[mom2\]), up to the first order of fields, can be reduced to $$\label{ }
\partial_i F + \partial_jF^{\ j)}_{(i} = 0 \ .$$ From this, we obtain a generic form of momentum constraint as $$\begin{aligned}
\label{mom3}
&\partial_i\left(c_2N_1 + c_5\partial_jN_j + c_6\dot{\zeta} + c_7{\rm e}^{-2\rho}\partial^2\zeta\right) \nonumber \\
&+ \partial_j\left(c_8\partial_{(i}N_{j)}\right)= 0\ , \end{aligned}$$ where $c_{5\sim8}$ are functions of time. Since we have already fixed the gauge sufficiently[^3] at Eq. (\[com\]), we can solve the constraints (\[ham2\]) and (\[mom3\]) for $N_1$ and $\partial_iN_i$. If we take a “physical" solution which is held apart from the spatial infinity, we get $$\begin{aligned}
\label{solution}
N_1&=N_1[\dot{\zeta},\,{\rm e}^{-2\rho}\partial^2\zeta] \ , \nonumber \\
\partial_iN_i&=(\partial_iN_i)[\dot{\zeta},\,{\rm e}^{-2\rho}\partial^2\zeta] \ .\end{aligned}$$ On the other hand, expanding the action (\[sc\]) up to the second order of the fields, we obtain the quadratic action consisting of $N$, $N_i$ and $\zeta$. We do *not* have, however, any terms made of $N\zeta$, $\zeta\partial_iN_i$ and $\partial_i\zeta N_i$ in the quadratic action since the generic constraint equations (\[ham2\]) and (\[mom3\]) do not include any terms proportional to $\zeta$ without differentiation. Therefore, associated with Eq. (\[K\]) and (\[R\]), we can reduce the ingredients of quadratic action to $$\begin{aligned}
\label{ss2}
S_2^{(S)} &= \int dtd^3x{\rm e}^{3\rho} \nonumber \\
&\quad\times{\cal L}_2[N_1,\,\partial_iN_j,\,\dot{\zeta},\, {\rm e}^{-2\rho}\partial_i\partial_j\zeta,\, {\rm e}^{-2\rho}(\partial\zeta)^2,\,\delta_{ij},\,t] \ .
$$ Here, ${\rm e}^{3\rho}{\cal L}_2$ implies the second order part of Lagrangian density. Substituting the solution (\[solution\]), we can find that the classical quadratic action for $\zeta$ becomes massless; $$\label{ }
S_2^{(S)} = \int dt d^3x {\rm e}^{3\rho}{\cal L}_2[\dot{\zeta},\, {\rm e}^{-\rho}\partial_i\zeta,\,{\rm e}^{-\rho}\partial_i,\,t\,] \ .$$ If we consider the usual scaling case where $\omega\propto k$, we can restrict the generic form of quadratic action of $\zeta$ to Eq. (\[s2\]).
We may extend the same procedure to the non-linear fluctuations [@Tanaka:2012wi]. If we can solve the constraints (\[conN\]) and (\[mom\]) perturbatively to arbitrary order of the fields and we do never encounter any strong couplings among the fields, we can conclude that at least barely, the non-linear comoving curvature perturbation remains massless for the system described by Eq. (\[sug\]).
Lastly, while it is an unnatural situation, we discuss about the case where the first order derivatives of Lapse function exist in the unitary gauge action as relevant operators. As we mentioned, when we recover the general diff-invariance by the Stückelberg trick, we need the accidental cancellations between the higher order derivative terms originated from the derivatives of Lapse function and the other higher order derivative terms to avoid the ghost instability. Such accidental cancellations indeed occur in the kinetic matrix of a class of scalar-tensor theories, so called DHOST theory [@Gleyzes:2013ooa; @Gleyzes:2014dya; @Gleyzes:2014qga; @Gleyzes:2014rba; @Langlois:2017mxy]. Although DHOST theory includes the first order derivatives of Lapse function in the unitary gauge action, it is showed that the resulting quadratic action for the scalar mode has the same form as Eq. (19) [@Langlois:2017mxy]. We have not identified all classes of DHOST theory yet, so that Eq. (19) could cover the known and unknown classes of DHOST theory potentially.
Ultra slow-roll solution {#C}
========================
We consider the shift symmetric system which do *not* have any root solutions for $J=0$ and ${\cal E}=0$. Even for this case, we may obtain a quasi-de Sitter solution if the action includes a positive constant potential. When the evolution of the universe is dominated by the potential, we may obtain the following background evolutions: $$\begin{aligned}
\label{FU}
&H^2\simeq \frac{\alpha V_0}{M_{\rm Pl}^2} = \text{const.}, \\
&J(\dot{\phi},\, H,\,\cdots)= c_Ja^{-3},\end{aligned}$$ where $\alpha$ is a constant and $V_0$ is the constant potential. Since it is supposed that there is no root solution, the behavior of scalar field differs from that of KDI crucially. Assuming that $J$ is expanded by a positive power series of $\dot{\phi}$ and the dominant term of $J$ at a certain time is given by the $p$-th power of $\dot{\phi}$, we obtain $$\begin{aligned}
\label{Jp}
J&\sim c_p\dot{\phi}^p \ , \\
\label{usr}
\dot{\phi}& \propto a^{-\frac{3}{p}},\end{aligned}$$ where $p>0$ and $c_p$ is a constant. Since the field velocity decreases with the quasi-de Sitter expansion obtained by (\[FU\]), after several e-folds, $J$ is eventually governed by the lowest power of $\dot{\phi}$ [@Hirano:2016gmv]. In this mechanics, unlike KDI, the field velocity always decreases towards $0$ in negative powers of $a$, and the scalar field moves quite slowly on the constant potential. This “ultra" slow motion of scalar field is encoded into the decreasing time derivative of Hubble parameter, which would be written as $$\label{ }
\dot{H}\propto \dot{\phi}^{q} \propto a^{-\frac{3q}{p}},$$ where $q$ is a positive constant. Thus, we still have a slight deviation from an exact de Sitter expansion involved with the quite small field velocity. Inflation induced by this mechanism is referred to as ultra slow-roll inflation [@Kinney:2005vj; @Hirano:2016gmv]. We note that the ultra slow-roll solution cannot become an attractor solution since from Eq. (\[usr\]), the second order time derivative of scalar field balances with the first order time derivative of scalar field unlike the KDI attractor.
Then, we turn to consider the quantum fluctuation. In the shift symmetric case, all the quantities should be expressed by constants and the scale factor $a$ only. We assume that at a certain time, ${{\cal F}_s}$ and ${{\cal G}_s}$ are written as $$\begin{aligned}
\label{}
{{\cal F}_s}&\sim f_0a^{-3p_f},\quad {{\cal G}_s}\sim g_0a^{-3p_g},\nonumber \\
c_s^2&\sim \frac{f_0}{g_0}a^{-3(p_f -p_g)}\ ,\end{aligned}$$ where all the characters are assumed to be positive constants. The slow-roll parameters behave as $$\begin{aligned}
\label{}
\epsilon_1&\propto a^{-\frac{3q}{p}}, \quad f_{s1}= -3p_f,\quad g_{s1} = -3p_g , \nonumber \\
\epsilon_2& \simeq -\frac{3q}{p},\quad f_{s2}= g_{s2}= 0 \ . \end{aligned}$$ Similarly to the shift symmetric case of KDI, $\epsilon_1$ decreases towards $0$, while $\epsilon_2$ remains almost as a constant. Thus, to apply the approximation where we discard $\epsilon_1\epsilon_2$, we must dismiss $\epsilon_1$ also from the spectral index. If $p_g\geq1$, the scalar fluctuation grows even on superhorizon scales since the parameter $\nu_s$ (\[nus\]) becomes zero or less than zero. In the original model of ultra slow-roll inflation [@Kinney:2005vj] where the action is given by $$\label{ }
S = \int d^4x\left[ \frac{M_{\rm Pl}^2}{2}R -X -V_0\right] ,$$ the parameters are derived as $$\begin{aligned}
\label{ }
p&= 1,\quad q=2, \quad f_0=g_0, \quad p_f=p_g =2, \nonumber \\
\nu_s &= -\frac{3}{2}.\end{aligned}$$ Using Eq. (\[nm\]) for this model, we have no spectral tilt as well as the type (i) theories of KDI. The scalar fluctuation for this model, however, grows on superhorizon scales since $\nu_s<0$.
[99]{}
A. H. Guth, Phys. Rev. D [**23**]{}, 347 (1981). doi:10.1103/PhysRevD.23.347
A. A. Starobinsky, Phys. Lett. [**91B**]{}, 99 (1980). doi:10.1016/0370-2693(80)90670-X
K. Sato, Mon. Not. Roy. Astron. Soc. [**195**]{}, 467 (1981).
V. F. Mukhanov and G. V. Chibisov, JETP Lett. [**33**]{}, 532 (1981) \[Pisma Zh. Eksp. Teor. Fiz. [**33**]{}, 549 (1981)\].
A. A. Starobinsky, Phys. Lett. [**117B**]{}, 175 (1982). doi:10.1016/0370-2693(82)90541-X
S. W. Hawking, Phys. Lett. [**115B**]{}, 295 (1982). doi:10.1016/0370-2693(82)90373-2
A. H. Guth and S. Y. Pi, Phys. Rev. Lett. [**49**]{}, 1110 (1982). doi:10.1103/PhysRevLett.49.1110
A. Albrecht and P. J. Steinhardt, Phys. Rev. Lett. [**48**]{}, 1220 (1982). doi:10.1103/PhysRevLett.48.1220
A. D. Linde, Phys. Lett. [**129B**]{}, 177 (1983). doi:10.1016/0370-2693(83)90837-7
P. J. Steinhardt and M. S. Turner, Phys. Rev. D [**29**]{}, 2162 (1984). doi:10.1103/PhysRevD.29.2162
J. Martin, C. Ringeval and V. Vennin, Phys. Dark Univ. [**5-6**]{}, 75 (2014) doi:10.1016/j.dark.2014.01.003 \[arXiv:1303.3787 \[astro-ph.CO\]\].
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], Astron. Astrophys. [**571**]{}, A22 (2014) doi:10.1051/0004-6361/201321569 \[arXiv:1303.5082 \[astro-ph.CO\]\].
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], Astron. Astrophys. [**594**]{}, A20 (2016) doi:10.1051/0004-6361/201525898 \[arXiv:1502.02114 \[astro-ph.CO\]\].
P. J. Steinhardt and N. Turok, Phys. Rev. D [**65**]{}, 126003 (2002) doi:10.1103/PhysRevD.65.126003 \[hep-th/0111098\].
S. Mukohyama, JCAP [**0906**]{}, 001 (2009) doi:10.1088/1475-7516/2009/06/001 \[arXiv:0904.2190 \[hep-th\]\].
P. Creminelli, A. Nicolis and E. Trincherini, JCAP [**1011**]{}, 021 (2010) doi:10.1088/1475-7516/2010/11/021 \[arXiv:1007.0027 \[hep-th\]\].
C. Armendariz-Picon, T. Damour and V. F. Mukhanov, Phys. Lett. B [**458**]{}, 209 (1999) doi:10.1016/S0370-2693(99)00603-6 \[hep-th/9904075\].
J. Garriga and V. F. Mukhanov, Phys. Lett. B [**458**]{}, 219 (1999) doi:10.1016/S0370-2693(99)00602-4 \[hep-th/9904176\].
N. Arkani-Hamed, P. Creminelli, S. Mukohyama and M. Zaldarriaga, JCAP [**0404**]{}, 001 (2004) doi:10.1088/1475-7516/2004/04/001 \[hep-th/0312100\].
T. Kobayashi, M. Yamaguchi and J. Yokoyama, Phys. Rev. Lett. [**105**]{}, 231302 (2010) doi:10.1103/PhysRevLett.105.231302 \[arXiv:1008.0603 \[hep-th\]\].
C. Burrage, C. de Rham, D. Seery and A. J. Tolley, JCAP [**1101**]{}, 014 (2011) doi:10.1088/1475-7516/2011/01/014 \[arXiv:1009.2497 \[hep-th\]\].
T. Kobayashi, M. Yamaguchi and J. Yokoyama, Prog. Theor. Phys. [**126**]{}, 511 (2011) doi:10.1143/PTP.126.511 \[arXiv:1105.5723 \[hep-th\]\].
G. W. Horndeski, Int. J. Theor. Phys. [**10**]{}, 363 (1974). doi:10.1007/BF01807638
D. Babich, P. Creminelli and M. Zaldarriaga, JCAP [**0408**]{}, 009 (2004) doi:10.1088/1475-7516/2004/08/009 \[astro-ph/0405356\].
X. Chen, M. x. Huang, S. Kachru and G. Shiu, JCAP [**0701**]{}, 002 (2007) doi:10.1088/1475-7516/2007/01/002 \[hep-th/0605045\].
T. Kobayashi, M. Yamaguchi and J. Yokoyama, Phys. Rev. D [**83**]{}, 103524 (2011) doi:10.1103/PhysRevD.83.103524 \[arXiv:1103.1740 \[hep-th\]\].
X. Gao and D. A. Steer, JCAP [**1112**]{}, 019 (2011) doi:10.1088/1475-7516/2011/12/019 \[arXiv:1107.2642 \[astro-ph.CO\]\].
A. De Felice and S. Tsujikawa, Phys. Rev. D [**84**]{}, 083504 (2011) doi:10.1103/PhysRevD.84.083504 \[arXiv:1107.3917 \[gr-qc\]\].
A. De Felice and S. Tsujikawa, JCAP [**1303**]{}, 030 (2013) doi:10.1088/1475-7516/2013/03/030 \[arXiv:1301.5721 \[hep-th\]\].
W. H. Kinney, Phys. Rev. D [**72**]{}, 023515 (2005) doi:10.1103/PhysRevD.72.023515 \[gr-qc/0503017\].
S. Hirano, T. Kobayashi and S. Yokoyama, Phys. Rev. D [**94**]{}, no. 10, 103515 (2016) doi:10.1103/PhysRevD.94.103515 \[arXiv:1604.00141 \[astro-ph.CO\]\].
J. Gleyzes, D. Langlois, F. Piazza and F. Vernizzi, JCAP [**1308**]{}, 025 (2013) doi:10.1088/1475-7516/2013/08/025 \[arXiv:1304.4840 \[hep-th\]\].
J. Gleyzes, D. Langlois, F. Piazza and F. Vernizzi, Phys. Rev. Lett. [**114**]{}, no. 21, 211101 (2015) doi:10.1103/PhysRevLett.114.211101 \[arXiv:1404.6495 \[hep-th\]\].
J. Gleyzes, D. Langlois, F. Piazza and F. Vernizzi, JCAP [**1502**]{}, 018 (2015) doi:10.1088/1475-7516/2015/02/018 \[arXiv:1408.1952 \[astro-ph.CO\]\].
J. Gleyzes, D. Langlois and F. Vernizzi, Int. J. Mod. Phys. D [**23**]{}, no. 13, 1443010 (2015) doi:10.1142/S021827181443010X \[arXiv:1411.3712 \[hep-th\]\].
D. Langlois, M. Mancarella, K. Noui and F. Vernizzi, JCAP [**1705**]{}, no. 05, 033 (2017) doi:10.1088/1475-7516/2017/05/033 \[arXiv:1703.03797 \[hep-th\]\].
T. Tanaka and Y. Urakawa, PTEP [**2013**]{}, 083E01 (2013) doi:10.1093/ptep/ptt057 \[arXiv:1209.1914 \[hep-th\]\].
R. L. Arnowitt, S. Deser and C. W. Misner, Gen. Rel. Grav. [**40**]{}, 1997 (2008) doi:10.1007/s10714-008-0661-1 \[gr-qc/0405109\].
R. P. Woodard, Lect. Notes Phys. [**720**]{}, 403 (2007) doi:10.1007/978-3-540-71013-4\_14 \[astro-ph/0601672\].
C. Cheung, P. Creminelli, A. L. Fitzpatrick, J. Kaplan and L. Senatore, JHEP [**0803**]{}, 014 (2008) doi:10.1088/1126-6708/2008/03/014 \[arXiv:0709.0293 \[hep-th\]\].
X. Gao, Phys. Rev. D [**90**]{}, 081501 (2014) doi:10.1103/PhysRevD.90.081501 \[arXiv:1406.0822 \[gr-qc\]\].
[^1]: Ghost inflation [@ArkaniHamed:2003uz] is included in this case. Thus, by using our formalisms, we can analyze the background evolution of ghost inflation, but cannot the evolution of quantum fluctuation of the model.
[^2]: In [@Hirano:2016gmv], the spectral tilt of a Galileon model included into the type (i) theories is derived as proportional to $\epsilon_1$. However, we should ignore the tilt if we employ the approximation where we discard $\epsilon_1\epsilon_2\simeq -3\epsilon_1$.
[^3]: We still have gauge symmetries such as the spatial rotation, $x^i\rightarrow x^i + {\epsilon}^i_jx^j$. These residual gauge symmetries, however, does never affect to solvability of the constraint equations.
| {
"pile_set_name": "ArXiv"
} |
[**Bubble nucleation in disordered Landau-Ginzburg model**]{}\
\
Centro Brasileiro de Pesquisas Físicas - CBPF\
Rua Dr. Xavier Sigaud, 150, Rio de Janeiro - RJ, 22290-180, Brazil\
[*C.A.D. Zarro*]{}\
Universidade Federal do Rio de Janeiro, Instituto de Física\
Av. Athos da Silveira Ramos, 149, Rio de Janeiro - RJ, 21941-909, Brazil
\
Abstract {#abstract .unnumbered}
--------
In this paper we investigate bubble nucleation in a disordered Landau-Ginzburg model. First we adopt the standard procedure to average over the disordered free energy. This quantity is represented as a series of the replica partition functions of the system. Using the saddle-point equations in each replica partition function, we discuss the presence of a spontaneous symmetry breaking mechanism. The leading term of the series is given by a large-$N$ Euclidean replica field theory. Next, we consider finite temperature effects. Below some critical temperature, there are $N$ real instantons-like solutions in the model. The transition from the false to the true vacuum for each replica field is given by the nucleation of a bubble of the true vacuum. In order to describe these irreversible processes of multiple nucleation, going beyond the diluted instanton approximation, an effective model is constructed, with one single mode of a bosonic field interacting with a reservoir of $N$ identical two-level systems.
[keywords:]{} disordered systems; free energy; bubble nucleation.
PACS numbers: 05.20.-y,75.10.Nr
.18in
Introduction {#intro}
============
The critical behavior of disordered systems has been discussed since the $70'$s in the literature. Two concepts that are of fundamental importance in such systems are respectively frustration and quenched disorder. Frustration was introduced to describe properties of spin-glasses with many different ground states [@and]. The free energy landscape of these systems have a multivalley structure. In quenched disordered systems, the disorder is in static equilibrium and therefore these systems are spatially random. The study of quenched disordered systems leads to new universality classes in critical regions and also the possibility of a large number of metastable states in free energy landscape. In such systems defined in the continuum with quenched disorder, it is a hard task to perform a perturbative expansion in any model, since these systems are intrinsically inhomogeneous. One way to circumvent such problem is to average over the ensemble of all realizations of the disorder quantities of interest. For example, average the free energy functional with respect to the probability distribution of the disordered field. In these disordered systems, the replica symmetry breaking with its physical consequences, has been intensely discussed by the physical community [@edwards; @SerKir; @RBS1; @RBS2; @livro1; @livro3; @livro4].
Recently, it was proposed a new method to average the disorder dependent free energy [@distributional; @distributional2]. Physical consequences of this approach were investigated in Refs. [@polymer; @zarro]. The motivation of this paper are the following. First is to stress the main differences between perturbative expansions in field theories without or in the presence of disorder fields, discussing cluster properties of disordered average $n$-point correlation functions. The second one is to discuss the physical consequences of the results obtained in Ref. [@zarro]. Finally, going beyond the above discussed results, we introduce an effective model to describe false-true vacuum transitions of replica fields. Specifically, we are interested to describe the phase transitions present in the continuous version of the $d$-dimensional random field Ising model [@larkin; @livro2; @naterman]. Using the approach discussed above, the structure of the replica space is investigated using the saddle-point equations obtained from each replica field theory. Assuming the replica symmetric ansatz, we investigate the spontaneous symmetry breaking mechanism in some replica partition functions. Our approach reveals the existence of replica instantons-like solutions (real or complex) in this model [@dotsenko4; @Maxin; @Allan]. For the case of real instantons-like solutions our methodology produced the following scenario.
Vacuum decay in this theory with $N$ replicas can be described in the following way. For low temperatures, there is a critical temperature where each replica field has two non-degenerate vacuum states. Consequently, for each replica field there will be a transition from the false vacuum to the true one with nucleation of a bubble of the true vacuum. This first-order phase transition, in the low temperature limit, was investigated in Refs. [@aharonytricritical; @aharonyetal]. The crucial question here is which tools we can use to describe these irreversible processes, i.e., the nucleation of bubbles of true vacuum in the false vacuum environment.
We shall now be concerning with the description of the collective behavior of the $N$ replica fields. The point of departure is given by Ref. [@Leggett]. There, the authors emphasize that it is possible to represent $N$ structures with a false and a true vacuum using two-level systems. The situation where the nucleation of bubbles occurs, decreasing the free energy of the system is characteristic of an open system. To go further describing this multiple nucleation, i.e., the collective nucleation of bubbles in the disordered system, an effective model is constructed using the functional integral formalism developed to study phase transitions in quantum optics systems by Popov and Fedotov [@Popov1; @Popov2; @Popov3]. In the large $N$ limit, the functional integral describes an ensemble of $N$ two-level systems interacting with one single bosonic mode, instead of the usual situation of a the countable infinitely set of field modes.
The justifications for introducing the bosonic mode are the following: this bosonic mode is connecting the two-level systems and also it makes possible the decay for each replica field from the false vacuum to the true one. In this scenario, it is possible to show the existence of a temperature where the free energy is non-analytic. The equivalence between these two quite distinct physical models can be justified using the following argument. In the disordered Landau-Ginzburg model, the leading replica partition function in the series representation for the free energy shows that all the replica fields, with a false and true vacuum states are strongly correlated. This is exactly the situation discussed by Popov and Fetotov, where only one single mode is resonant with the two-level atoms. All the two-level atoms interact coherently with this single mode.
The organization of this paper is the following. In Sec. \[sec:disoderedLG\], we discuss a $d$-dimensional disordered Landau-Ginzburg model. In Sec. \[sec:distributionalzeta\], in a generic replica partition function we discuss the structure of the replica space using the saddle-point equations of the model. In Sec. \[sec:replicainstantons\], we demonstrate at low temperatures the emergence of $N$ instantons-like solutions in the leading replica partition function of the model. In Sec. \[sec:effectivemodel\], to describe the bubble nucleation in the disordered model, an effective model is constructed using the formalism developed by Popov and Fedotov. Conclusions are given in Sec. \[sec:conclusions\]. We use the units $\hbar=c=k_{B}=1$.
A disordered Landau-Ginzburg model {#sec:disoderedLG}
==================================
In magnetic materials with disorder, in principle there are two kind of systems. The first set is one where the disorder is related to the local spin interaction. In this case the disorder generates multiple disordered ground states, the spin-glass phase. The second, is one where the disorder is a random external perturbation. One disordered model that belongs to this second set is the random field Ising model. This model has been studied intensively from the theoretical and experimental point of view and used to describe many systems in nature. One is the case of diluted antiferromagnetic in a homogeneous external field [@anti1; @anti2] and also binary fluids in porous media. For instance, in order to model binary fluids confined in porous media, when the pore surfaces couple differently to the two components of a phase-separating mixture, the random field has been used by the literature. These systems can develop a second or a first-order phase transition [@fluids1; @fluids2; @fluids3]. The random field Ising model in a hypercubic lattice in $d$-dimensions is described by the Hamiltonian
$$H=-J\sum_{(i,j)}^{N}\,S_{i}S_{j}-\sum_{i}\,h_{i}S_{i}, \label{1}$$
where $(i,j)$ indicates that the sum is performed over nearest neighbour pairs and $S_{i}=\pm 1$. In the above equation $N$ is the total number of Ising spins. Periodic boundary conditions can be used and the thermodynamic limit must be used in the end. The partition function is $Z=Tr\, e^{-\beta H}$. In Eq. (\[1\]) the $h_{i}$’s are the quenched random variables totally uncorrelated on different sites. The average free energy is defined by $F=-\frac{1}{\beta}\,\mathbb{E}[\ln Z]$, where $\mathbb{E}\,[...]$ means the average over the ensemble of all the realizations of the quenched disorder. Here we consider a Gaussian distribution defined by
$$P(h_{i})=\frac{1}{\sqrt{2\pi h_{0}^{2}}}\exp\bigg(\frac{-h_{i}^{2}}{2h_{0}^{2}}\biggr). \label{no}$$
The probability distribution of such quenched random variables has zero mean-value, $\mathbb{E}\,[h_{i}]=0$, and correlation functions given by $\mathbb{E}\,[h_{i}h_{j}]=h_{0}^{2}\delta_{ij}$. Here we are interested in the small disordered limit, i.e., $h_{0}^{2} << 1$.
The properties of the phase transition of the random field Ising model is still under debate [@mezard1; @mezard2; @dotsenko; @orland; @dotsenko2; @dotsenko3; @sherington1; @sherington]. The question of the lower critical dimension, bellow which long-range order is absence and the upper critical dimension, above which the model presents mean-field behavior independent of the dimension has been a matter of controversy. Imry and Ma obtained that the model with nearest neighbor interaction presents spontaneous magnetization only for $d\geq \,3$ [@ma]. This result is in contradiction with the dimensional reduction argument [@sourlas; @parisi1]. The controversy was solved by Bricmont and Kupiainen, who proved that there is a phase transition in the random field Ising model for $d\geq\,3$ [@bric1; @bric2], and Aizenman and Wehr, that showed the absence of phase transition for $d=2$ in the model [@wehr1].
The behavior of systems defined in a lattice near the critical point can be modeled by continuous statistical field theories. This can be achieved replacing the lattice structure by a continuum where the order parameter can be obtained averaging with respect to a statistical weight a random continuous field. For instance, the effective $O(n)$ Landau-Ginzburg model is defined by $\varphi_{i}(x)$, a $n$-component field. This model is able to describe several universality classes. For the case $n\rightarrow 0$ can describe self-avoided polymers [@gennes; @gaspari; @schafer]. For $n=1$ it describes the critical behavior of the Ising model. For $n=2$ the critical behavior of the $XY$ model and also the two-dimensional Coulomb gas are described. For $n=3$ the Heisenberg model [@livrozinnjustin] and low energy dynamics of QCD can be modeled by $n=4$. Finally for $n\rightarrow \infty$ it is possible to solve exactly the model. In this paper we study the critical properties of the random field Ising model, by means of a continuous scalar field theory defined in $\mathbb{R}^{d}$ with symmetry $Z_{2}$ (the $n=1$ case).
We assume that the critical behavior of the random field Ising model in $\mathbb{Z}^{d}$ can be described by a continuous disordered Landau-Ginzburg model. Firstly, let us briefly discuss the model without disorder. We are following Ref. [@JAC]. The Landau-Ginzburg functional, $i.e.$, the Hamiltonian $H(\varphi)$ for the scalar field is given by
$$H=\int d^{d}x\Biggl(\frac{1}{2}\varphi(x)\left(-\Delta+m_{0}^{2}\right)\varphi(x)
+\frac{\lambda_{0}}{4!}\varphi^{4}(x)\Biggr), \label{9}$$
where the symbol $\Delta$ denotes the Laplacian in $\mathbb{R}^{d}$ and $\lambda_{0}$ and $m_{0}^{2}$ are analytic functions of the temperature. Actually, $m_{0}^{2}$ is the inverse of the correlation length. Doing a parallel with Euclidean field theory we call them respectively the bare coupling constant and the squared mass of the model. For high temperatures, away from the critical point, the correlation functions of the model are short-ranged. Near the critical point, the correlation functions becomes long ranged, where the characteristic length scale, the correlation length $\xi$ has a power law behavior, exactly as in Euclidean field theory. The partition function of the model is defined by the functional integral
$$\begin{aligned}
Z=\int_{\partial\Omega} [d\varphi]\,\, &\exp\bigl(-H(\varphi)\bigr), \label{88}\end{aligned}$$
where $[d\varphi]$ is a formal Lebesgue measure, given by $[d\varphi]=\prod_{x} d\varphi(x)$, and $\partial\Omega$ in the functional integral means that the field $\varphi(x)$ satisfies some boundary condition in the boundary $\partial\Omega$ of some bounded domain, i.e., connected open set $\Omega \subset \mathbb{R}^{d}$. Periodic boundary conditions can be imposed to preserve translational invariance, replacing $\mathbb{R}^{d}$ by the torus $\mathbb{T}^{d}$. To remove the ultraviolet divergences, in the Fourier decomposition of the field a cut-off must be introduced. This cut-off is related with a elementary length scale, the lattice spacing of the original model. Since in all the discussions of this paper we need no more than the tree-level calculations, this technical remark is immaterial for the results presented in the paper.
The question now arises is the cluster properties of correlation functions for the model with or without disorder. Therefore, let us start briefly discussing these quantities. Averaging with respect to the Boltzmann weight we get the $n$-point correlation functions of the model
$$\label{eq:correlationpure}
\langle\varphi(x_{1})...\varphi(x_{n})\rangle=
\frac{1}{Z}\int [d\varphi]\prod_{i=1}^{n}\varphi(x_{i}) \exp\bigl(-H(\varphi)\bigr).$$
Introducing a fictitious source $j(x)$ we can define $Z(j)$, the generating functional of all $n$-point correlation functions as [@livro5; @lk]
$$Z(j)=\int_{\partial\Omega} [d\varphi]\,\, \exp\left(-H(\varphi)+\int d^{d}x\, j(x)\varphi(x)\right).
\label{eq:generatingfunctional}$$
Taking functional derivatives with respect to the source and setting to zero in the end, we obtain the $n$-point correlations functions of the model
$$\label{eq:correlationfunction2}
\langle\varphi(x_{1})..\varphi(x_{k})\rangle=
Z^{-1}(j)\left.\frac{\delta^{k} Z(j)}{\delta j(x_{1})...\delta j(x_{k})}\right|_{j=0}.$$
Notice that these $n$-point correlation functions are given by the sum of all diagrams with $n$-external legs, including the disconnected ones. Next, using the linked cluster theorem, it is possible to define the generating functional of connected correlation functions given by $W(j)=\ln Z(j)$. The order parameter of the model without disorder $\langle\varphi(x)\rangle$ is given by
$$\langle\varphi(x)\rangle=Z^{-1}(j)\left.\frac{\delta Z(j)}{\delta j(x)}\right|_{j=0}.$$
Before continue, we would like to clarify terminology. Although streaking speaking, the order parameter is defined by the above equation, throughout this paper we may call local order parameter for the continuous field $\varphi(x)$ defined for $x \in \mathbb{R}^{d}$. Applying two functional derivatives on the generating functional of connected correlation functions we get
$$\begin{aligned}
\nonumber
\langle\varphi(x_{1})\varphi(x_{2})\rangle_{connected} & =
\biggl[\frac{1}{Z(j)}\frac{\delta^{2}Z(j)}{\delta j(x_{1})\delta j(x_{2})} \\
&\left.-\frac{1}{Z(j)^{2}}\frac{\delta Z(j)}{\delta j(x_{1})}\frac{\delta Z(j)}{\delta j(x_{2})}\biggr]\right|_{j=0}.\end{aligned}$$
The large distance decay properties of these connected correlation functions are called cluster properties. These correlation functions goes to zero for $|x_{1}-x_{2}|\rightarrow \infty$. In an Euclidean field theory the cluster properties of the Schwinger functions are equivalent to the uniqueness of the vacuum.
We briefly present the basic tools that we need to discuss disordered systems [@malivro]. The continuum version for the $d$-dimensional random field Ising model, is given by a $d$-dimensional Landau-Ginzburg scalar $\lambda\varphi^{4}$ model in the presence of a disorder field linearly coupled to the scalar field. The Hamiltonian in the presence of disorder is given by
$$H(\varphi,h)=H(\varphi)+ \int d^{d}x\,h(x)\varphi(x),$$
where $H(\varphi)$ is the Landau-Ginzburg Hamiltonian, defined in Eq. (\[9\]), and $h(x)$ is a quenched disorder field. The disordered functional integral $Z(h)$ is defined by
$$Z(h)=\int_{\partial\Omega} [d\varphi]\,\, \exp\bigl(-H(\varphi,h)\bigr).
\label{8}$$
Eq. (\[8\]) defines the partition function associated with the scalar field for a given disorder configuration. The $n$-point correlation functions for one specific realization of the disorder field reads
$$\label{eq:disordercorrelationfunction}
\langle\varphi(x_{1})..\varphi(x_{n})\rangle_{h}=\frac{1}{Z(h)}\int [d\varphi]\prod_{i=1}^{n}\varphi(x_{i})
\exp\bigl(-H(\varphi,h)\bigr).$$
To introduce a generating functional for one realization of the disorder field, $Z(h;j)$, we again use a fictitious source $j(x)$:
$$Z(h;j)=\int_{\partial\Omega} [d\varphi]\,\, \exp\biggl(-H(\varphi,h)+\int d^{d}x \,j(x)\varphi(x)\biggr).
\label{eq:disorderedgeneratingfunctional}$$
For a particular realization of the disorder field, $Z(h;j)$ can be used to obtain the $n$-point correlation function given by Eq. (\[eq:disordercorrelationfunction\]) by means of functional derivatives. With these correlation functions, one can compute the disorder-averaged correlation functions given by
$$\mathbb{E}\bigl[\langle\varphi(x_{1})...\varphi(x_{n})\rangle_{h}\bigr]=\int\,[dh]P(h)\langle\varphi(x_{1})...
\varphi(x_{n})\rangle_{h},$$
where $\langle\varphi(x_{1})..\varphi(x_{n})\rangle_{h}$ is given by Eq. (\[eq:disordercorrelationfunction\]) and $[dh]=\prod_{x} dh(x)$ is again a formal Lebesgue measure. As in the pure system case, one can define a generating functional for one disorder realization, $W_{1}(h;j)=\ln Z(h;j)$. We take the disorder-average of this generating functional, $W_{2}(j)=\mathbb{E}[W_{1}(h;j)]$. We have
$$W_{2}(j)=\int\,[dh]P(h)\ln Z(h;j).
\label{eq:disorderedfreeenergy}$$
Taking the functional derivative of $W_{2}(j)$ with respect to $j(x)$, we get
$$\left.\frac{\delta W_{2}(j)}{\delta j(x)}\right|_{j=0}=\int\,[dh]P(h)\biggl[\frac{1}{Z(h;j)}\left.\frac{\delta
Z(h;j)}{\delta j(x)}\biggr]\right|_{j=0}.$$
Since $\langle\varphi(x)\rangle_{h}$ is the average of the field for a given configuration of the disorder in the disorderd Landau-Ginzburg model the above quantity $\mathbb{E}\bigl[\langle\varphi(x_{1})\rangle_{h}\bigr]$ is the order parameter of the model [@parisi1]. The second functional derivative of $W_{2}(j)$ with respect to $j(x)$ gives $G(x_{1}-x_{2})$. We have
$$\begin{aligned}
\left.\frac{\delta^{2}W_{2}(j)}{\delta j(x_{1})\delta j(x_{2})}\right|
_{j=0}&=\mathbb{E}\left[\langle\varphi(x_{1})\varphi(x_{2})\rangle_{h}\right]\nonumber \\
& -\mathbb{E}\left[\langle\varphi(x_{1})\rangle_{h}\langle\varphi(x_{2})\rangle_{h}\right].
\label{eq:two-point}\end{aligned}$$
Notice that, in general, the following quantities are not equal, i.e.,
$$\label{eq:difference}
\mathbb{E}\left[\langle\varphi(x_{1})\rangle_{h}\langle\varphi(x_{2})\rangle_{h}\right]\neq
\mathbb{E}\left[\langle\varphi(x_{1})\rangle_{h}\right]\mathbb{E}\left[\langle\varphi(x_{2})\rangle_{h}\right].$$
Therefore the Eq. (\[eq:two-point\]) is not the disordered average two-point connected correlation function. To proceed let us define the following averaged quantity
$$\chi(x_{1}-x_{2})=\mathbb{E}\left[\langle\varphi(x_{1})\rangle_{h}\langle\varphi(x_{2})\rangle_{h}\right].$$
This above disconnected correlation function can be different from zero even if the order parameter of the model is zero. The decay of these two-point correlation functions $G(x_{1}-x_{2})$ and $\chi(x_{1}-x_{2})$ at critical region defines two critical exponents $\eta$ and $\eta'$ [@rieger]. We have
$$G(x_{1}-x_{2})\approx |x_{1}-x_{2}|^{-(d-2+\eta)}.$$
and
$$\chi(x_{1}-x_{2})\approx |x_{1}-x_{2}|^{-(d-4+\eta')}.$$
In a pure system, taking functional derivatives of $W(j)$ we get the connected correlation functions, that satisfies clustering property. Applying two functional derivatives, the disordered average functional $W_{2}(j)=\mathbb{E}
[W_{1}(h;j)]$ does not generate the disordered average two-point connected correlation functions of the model. This can be generalized to the $n$-point correlation functions, or being more precise, investigating cluster properties of disordered average $n$-point correlation functions. The fundamental problem is the fact that since there are many minima [@Parisi88; @aharony; @fytas] in these systems, we can not expand around only one specific minimum, hence a non-perturbative scenario emerges. The non-perturbative scenario can not be studied neither using the renormalization group equations nor the composite operator formalism [@jackiw; @ananos; @gap]. Composite operator formalism is a way to use resummation methods (sum of infinite series of diagrams) to avoid the infrared divergences of a massless theory. These methods can not reveal the vacuum structure of the disordered system.
One possible way to proceed is the following. In the presence of these metastable states one must identify clustering states, i.e., the states where the connected correlation functions vanishes for large distances, and introduce an order parameter that characterize such domain [@PRL]. We do not expect that this program can be implemented in a straightforward way. To deal with this above discussed problem, the first step is to identify the metastable states, i.e., show the presence of many local minima in the free energy landscape. In other words, this fundamental difficulty may point that a local approach of field theory based in the correlation functions must be substituted, at least in the beginning by another more promising procedure. Instead of concentrate our efforts to define local objects, we may study only global quantities, such as, the averaged free energy. As we expected, here we will show the presence of a large number of metastable states in the disordered system.
For instance, for free fields without disorder the spectral zeta-function technique [@seeley; @ray; @hawking; @dowker; @fulling], which is a way to regularize the determinant of Laplace operator, can be used to compute the free energy of this pure system. In the next section, we show how this approach can be used to access the non-perturbative landscape of the disordered system. Here, we proceed as follows. We are interested to compute $W_{2}(j)|_{j=0}=\mathbb{E}[W_{1}(h;j)]|_{j=0}$, namely the disorder-averaged free energy.
Distributional zeta-function approach {#sec:distributionalzeta}
=====================================
In order to circumvented the problem of many local minima that the perturbative expansion fail to take into account, Lancaster *et al.* [@lancaster] discussed a model where many solutions of the mean field equations obtained from each realization of the disorder are weighted by Boltzmann factors. In the following we show that it is possible to investigate a non-perturbative scenario using the distributional zeta-function approach [@distributional; @distributional2]. This approach has similarities with the above discussed method.
Here, we do not give details of the derivation but only the essential steps of the mathematical rigorous procedure that allow to use the replica partition functions in order to compute the disorder-averaged free energy. For a given probability distribution of the disorder, one is mainly interested in averaging the disorder dependent free energy functional which reads
$$F=-\frac{1}{\beta}\int\,[dh]P(h)\ln Z(h),
\label{sa2}$$
where $\beta^{-1}=T$, where $T$ is the temperature of the system. This averaged free energy represents, in an Euclidean field theory, the connected vacuum-to-vacuum diagrams in the disordered system. For a general disorder probability distribution, using the disordered functional integral $Z(h)$ given by Eq. (\[8\]), the distributional zeta-function, $\Phi(s)$, is defined as
$$\Phi(s)=\int [dh]P(h)\frac{1}{Z(h)^{s}},
\label{pro1}$$
for $s\in \mathbb{C}$, this function being defined in the region where the above integral converges. The above equation is a natural generalization of the families of zeta-functions [@riem; @riem2; @landau; @fro; @primes0; @primes; @voros]. The average free energy can be written as
$$F=(d/ds)\Phi(s)|_{s=0^{+}}, \,\,\,\,\,\,\,\,\,\, \Re(s) \geq 0,$$
where one defines the complex exponential $n^{-s}=\exp(-s\log n)$, with $\log n\in\mathbb{R}$. Using analytic tools, the average free energy can be represented as
$$F=\frac{1}{\beta}\Biggl[\sum_{k=1}^\infty \frac{(-1)^{k}a^{k}}{k}\,\mathbb{E}\,[Z^{\,k}]+\ln(a)+\gamma-R(a)\Biggr]
\label{m23e}$$
where $a$ is a dimensionless arbitrary constant, $\gamma$ is the Euler-Mascheroni constant, and, for large $a$, $|R(a)|$ is quite small, therefore, the dominant contribution to the average free energy is given by the replica partition functions of the model. For simplicity we write $\mathbb{E}\left[Z(h)^{k}\right]\equiv\mathbb{E}\left[Z^{k}\right]$. Note that a $\frac{1}{k!}$ factor was absorbed in $\mathbb{E}\,[Z^{\,k}]$. To proceed, we assume that the probability distribution of the disorder is written as $[dh]\,P(h)$, where
$$P(h)=p\,\exp\Biggl(-\frac{1}{2\,\sigma}\int\,d^{d}x(h(x))^{2}\Biggr).
\label{dis2}$$
The quantity $\sigma$ is a positive parameter associated with the disorder and $p$ is a normalization constant. In this case we have a delta correlated disorder field, i.e., $\mathbb{E}[{h(x)h(y)}]=\sigma\delta^{d}(x-y)$. As it was stressed by many authors, it is important to clarify the behavior of the model for small values of $\sigma$. After integrating over the disorder we get that each replica partition function $\mathbb{E}\,[Z^{\,k}]$ can be written as
$$\mathbb{E}\,[Z^{\,k}]=
\frac{1}{k!}\int\,\prod_{i=1}^{k}[d\varphi_{i}]\,\exp\Bigl(-H_{\textrm{eff}}(\varphi_{i})\Bigr),
\label{aa11}$$
where the effective Hamiltonian $H_{\textrm{eff}}(\varphi_{i})$ describing the field theory with $k$-replica field components is given by
$$\begin{aligned}
H_{\textrm{eff}}(\varphi_{i})&=
\int d^{\,d}x\Biggl[\sum_{i=1}^{k}\biggl(\frac{1}{2}\varphi_{i}(x)\bigl(-\Delta+m_{0}^{2}\bigr)\varphi_{i}(x)
\nonumber\\
&+\frac{\lambda_{0}}{4!}\varphi_{i}^{4}(x)\biggr)
-\frac{\sigma}{2}\sum_{i,j=1}^{k}\varphi_{i}(x)\varphi_{j}(x)\Biggr].
\label{Seff1}\end{aligned}$$
In the original Landau mean-field theory to discuss second-order phase transitions, an expansion for the free energy near the critical temperature as a power series of the order parameter is introduced. It is important to keep in mind that in the framework discussed by us the same idea is introduced. Nevertheless, by the presence of the disorder field, instead of a series in the order parameter we get a series in the replica partition functions of the model to define the averaged free energy.
The mean-field theory corresponds to a saddle-point approximation in each replica partition function. A perturbative approach gives us the fluctuation corrections to mean-field theory. Hence, to implement a perturbative scheme, it is necessary to investigate fluctuations around the mean-field equations. From each replica field theory, let us investigate the solutions of the saddle-point equations which are given by
$$\Bigl(-\Delta\,+m_{0}^{2}\Bigr)\varphi_{i}(x)
+\frac{\lambda}{3!}\varphi^{3}_{i}(x)=\sigma\sum_{j=1}^{k}\varphi_{j}(x).
\label{sp}$$
Imposing the replica symmetric ansatz, $i.e.$, $\varphi_{i}(x)=\varphi(x)$, the saddle-point equation, in each replica partition function, reads
$$\Bigl(-\Delta\,+m_{0}^{2}-k\sigma\Bigr)\varphi(x)+\frac{\lambda_{0}}{3!}\varphi^{3}(x)=0.
\label{sp1}$$
At this stage it is easy to understand why the original replica method has problems, at least in this model. In this method, the average free energy is obtained using the formula
$$\mathbb{E}\,[{\,\ln Z(h)}]=\lim_{n\rightarrow 0}\frac{\partial}{\partial n}\mathbb{E}\,[ Z(h)^{n}].$$
The $n\rightarrow 0$ limit in Eqs. (\[aa11\]), (\[Seff1\]) is translated to a field theory with the dimension of the order parameter going to zero. Therefore, we would like to briefly discuss the limit $n\rightarrow 0$ in the $O(n)$ Landau-Ginzburg model. It is well known that the self-avoiding random walk can be used as a mathematical model for polymers chains, where effects of excluded volume must be modeled [@fisher; @dhar]. Since it represents a non- Markovian stochastic process, there are many open questions in the literature, as, for instance, how many walks there are between two points. In the case of the self-avoiding random walk problem, the probability of finding the particle at $y$ at time $t$ if the particle was released in point $x$ at $t=0$, is a sum of diagrams that are exactly those for the correlation function of the $O(n)$ Landau-Ginzburg model for $n\rightarrow 0$.
In the original replica method although one work with a replica field theory where the number of replicas must go to zero, the situation is quite different from the above discussed cases. The average free energy involves derivation of the integer moments of the partition function. One consequence of this fact is that using the simplest possible replica symmetric ansatz in each replica partition function reduce the equations to the saddle-point equations of systems without disorder. Therefore, the replica symmetry breaking is introduced as a necessary condition to recover information from the disorder field in the theory.
Using the distributional zeta-function method we can go further, since we have obtained analytic expression for the average free energy that does not involve derivation of such integer moments. Notice that, in principle, we have to consider all terms in Eq. (\[m23e\]), since all values of $k$ are allowed. However, we have a constraint as the squared mass, $m_{0}^{2}-k\sigma$, must be positive definite to describe a well-defined physical theory. In this case, one has a critical value of $k$, namely, $k_{c}=\left\lfloor m_{0}^{2}/\sigma\right\rfloor$, above which one would obtain a negative squared mass, where $\left\lfloor x \right\rfloor$ means the integer part of $x$. For $k<k_{c}$, the replica fields fluctuate around the zero value. For $k>k_{c}$, we have to shift these replica fields since the zero value is not a stable equilibrium state. The last situation represents a spontaneous symmetry breaking mechanism.
In the framework of distributional zeta-function method, defining $v=\left(\frac{6(\sigma N-m_{0}^{2})}{\lambda_{0}}\right)^{1/2}$, the simplest choice of the replica space is given by
$$\begin{cases}
\varphi_{i}^{(l)}(x)=\varphi(x) \;\;\; \hbox{for $l=1,\cdots,k_{c}$ and $i=1,\cdots,l$} \\
\varphi_{i}^{(l)}(x)=\phi(x)+v \;\;\hbox{for $l=k_{c}+1,\cdots,N$ and $i=1,\cdots,l$} \\
\varphi_{i}^{(l)}(x)=0 \quad \,\,\,\,\,\hbox{for $l>N$.}
\end{cases}
\label{RSB2}$$
Notice that we find a positive squared mass with self-interactions terms $\phi(x)^{3}$ and $\phi(x)^{4}$. From Eq. (\[RSB2\]) and for $a$ and $N$ very large, the average free energy can be written as
$$F=\frac{1}{\beta}\sum_{k=1}^{N} \frac{(-1)^{k}a^{k}}{k}\,\mathbb{E}\,[Z^{\,k}],
\label{KMenergy}$$
which has its leading term for $k=N$. Therefore, in the large-$N$ limit, the expression for disorder-averaged free energy is reduced to the contribution of only one replica partition function, consisting in a large $N$-component replica fields. In the context of a large-$N$ scenario, we introduce two ’t Hooft couplings, namely, $f_{0}=\sigma N$ and $g_{0}=\lambda_{0} N$. These parameters are finite in $N\rightarrow \infty$ although $\lambda\rightarrow 0$ and $\sigma\rightarrow 0$.
Replica instantons-like solutions in the disordered system {#sec:replicainstantons}
==========================================================
The mean-field approach approach is used to analyze the phase diagram of our model. First, we consider that $m_{0}^{2}$ is a regular function of temperature. This situation is more complex than in a ordered system. We find three regions of interest. The first occurs for $m_{0}^{2}\geq\sigma N$. In this case, all the replica fields oscillate around $\varphi=0$, the trivial vaccum. For $a\gg N$, a very large $N$ limit is represented by only one replica partition function with $N$ ($N$ even) replica fields $\phi_{i}$. The $N$ replica fields has the symmetry $[\mathbb{Z}_{2}\times\mathbb{Z}_{2}\cdots\times\mathbb{Z}_{2}]$. There is also a critical temperature $T^{(1)}_{c}$, where $m_{0}^{2}=N\sigma$. The $[\mathbb{Z}_{2}\times\mathbb{Z}_{2}\cdots\times\mathbb{Z}_{2}]$ symmetry is broken below $T^{(1)}_{c}$. For the second region, $\sigma \leq m_{0}^{2}<\sigma N$, replica fields in some partition functions oscillates around the trivial vaccum, whereas fields in other replica partition functions now oscillates around the non-trivial vacuum. We are not interested in these ranges of $m_{0}^{2}$, for more details see Ref. [@zarro]. For $m_{0}^{2}<\sigma$, all the replica fields in each replica partition functions are oscillating around the non-trivial vacuum. In this case, for $a \gg N$ and for a very large-$N$ limit ($N$ even), the average free energy reads
$$F=\frac{1}{\beta}\,\mathbb{E}\,[{Z^{\,N}}], \label{LGEnergyCase3}$$
where $a$ is absorbed in normalization of the functional integration and $\mathbb{E}[{Z^{\,N}}]$ is written as
$$\label{E2(Zk)}
\mathbb{E}[{Z^{\,N}}]=\frac{1}{N!}\int\,\prod_{i=1}^{N}\left[d\phi_{j}\right]\,\exp\biggl(-
H_{\textrm{eff}}\left(\phi_{j}\right)\biggr),$$
and the effective Hamiltonian $H_{\textrm{eff}}(\phi_{i})$ is given by
$$\begin{aligned}
H_{\textrm{eff}}(\phi_{i})=&\int d^{\,d}x\Biggl[\sum_{i=1}^{N}\biggl(\frac{1}{2}
\phi_{i}(x)\Bigl(-\Delta+3f_{0}-2m_{0}^{2}\Bigr)\phi_{i}(x)\nonumber\\
&\,+\Bigl(\frac{f_{0}g_{0}}{3!N}\Bigr)^{\frac{1}{2}}\left(1-\frac{m_{0}^{2}}{f_{0}}\right)^{\frac{1}
{2}}\phi_{i}^{3}(x)+\frac{g_{0}}{4!N}\phi_{i}^{4}(x)\biggr)\nonumber \\
&-\frac{f_{0}}{2N}\sum_{i,j=1}^{N}\phi_{i}(x)\phi_{j}(x)\Biggr].
\label{Seff(2)}\end{aligned}$$
Notice that the symmetry $[\mathbb{Z}_{2}\times\mathbb{Z}_{2}\cdots\times\mathbb{Z}_{2}]$ for $N$ replica fields is broken. A relevant question in the random field Ising model concerns the existence of an upper critical dimension, which, above it, the mean field approximation is exact. Since we have a cubic term in the action, the upper critical dimension is obtained from the relation $\frac{3}{2}(d-2)=d$, where the coupling constant becomes dimensionless, therefore the critical dimension is $d=6$. This result was discussed by Imry and Ma [@ma] and more recently in Ref. [@Ahrens].
Our fundamental result is the following. To describe critical phenomena for systems without disorder it is introduced an order parameter that describes second-order phase transition where for low temperatures a state of reduced symmetry appears. In the disordered system the order parameter is now a $N$-vector field. Our aim is to describe bubble nucleation in the disordered model at low temperatures. A representation similar to the strong-coupling expansion in field theory [@SKD; @RMDHS; @CBFC; @sce] or the linked cluster expansion [@lusher1; @lusher2; @lusher3; @reisz1; @reisz2] can be used to represent a replica field theory. Rather than the usual case, which relies upon a gradient-free action, now the replicas become connected after applying a functional differential operator on a well-defined replica partition function. Here we would like to stress that the use of the linked cluster expansion in the Ising model was introduced in the literature by Englert [@Englert].
To proceed, an external source $\mathcal{J}_{i}(x)$ in replica space linearly coupled with each replica is introduced. Defining $R(x-y)=\sigma\delta^{d}(x-y)$, each replica partition function, $\mathbb{E}\,
[Z^{N}]=\mathcal{Z}(\mathcal{J})$, is written as a functional differential operator applied on $Q_{0}(\mathcal{J})$. Hence
$$\begin{aligned}
&\mathcal{Z}(J)=\nonumber \\
&\exp\biggr[-\frac{1}{2}\sum_{i,j=1}^{N}\int d^{\,d}x\,d^{\,d}y\frac{\delta}{\delta
\mathcal{J}_{i}(x)}R\frac{\delta}{\delta \mathcal{J}_{j}(y)}\biggr]Q_{0}(J).
\label{instanton1}\end{aligned}$$
In the above equation, $Q_{0}(\mathcal{J})$, a modified replica partition function, is written as
$$Q_{0}(\mathcal{J})=\frac{1}{N!}\int\prod_{j=1}^{N}[d\phi_{j}]\,\exp\biggl(-H_{\textrm{eff}}^{(0)}(\phi_{j},
\mathcal{J}_{i})\biggr),
\label{instanton2}$$
where $H_{\textrm{eff}}^{(0)}(\phi_{i},\mathcal{J}_{i})$ is given by
$$\begin{aligned}
&H_{\textrm{eff}}^{(0)}(\phi_{i},\mathcal{J}_{i})=\int d^{\,d}x \sum_{i=1}^{N}\Biggl[\frac{1}{2}\phi_{i}
(x)\Bigl(-\Delta+3f_{0}-2m_{0}^{2}\Bigr)\phi_{i}(x)\nonumber\\
&\Bigl(\frac{f_{0}g_{0}}{3!N}\Bigr)^{\frac{1}{2}}\left(1-\frac{m_{0}^{2}}{f_{0}}\right)^{\frac{1}{2}}\phi_{i}^{3}
(x)+\frac{g_{0}}{4!N}\phi_{i}^{4}(x)+\mathcal{J}_{i}(x)\phi_{i}(x)\Biggr].\nonumber \\
\label{instantonN}\end{aligned}$$
Notice that the above equation does not contain interaction terms between replica fields. It is important to notice that Eqs. (\[instanton2\]) and (\[instantonN\]) fixes all ultraviolet divergences of our model that can be regularized by standard analytic regularization procedures [@bol; @analit1; @analit2; @analit3; @analit5]. The main idea is that in the $\epsilon=(4-d)$ expansion all the primitively divergent correlation functions contain poles. The principal part of the Laurent expansion defines the counterterms that we have to introduce to cancel such polar contributions. Introducing the renormalization constants $Z_{\varphi}$, $Z_{\lambda}$ and $Z_{m}$ the theory becomes finite. This pertubative expansion program with the regularization and renormalization procedures can be straightforwardly implemented. However, we will not follow it further in this analysis. Instead, we will study the vacuum structure of the first factor of Eq. (\[instanton1\]), $i.e.$, $\mathcal{Z}(\mathcal{J})= Q_{0}(\mathcal{J})$. It is possible to define the generating functional of connected correlation functions $\mathcal W(\mathcal{J})=\ln \mathcal{Z}(\mathcal{J})$. For simplicity we assume that we have one replica field. The generating functional of one-particle irreducible correlations (vertex functions), $\Gamma[\overline{\phi}]$, is gotten by taking the Legendre transform of $\mathcal{W}(\mathcal{J})$ [@amit]
$$\Gamma[\overline{\phi}]+\mathcal{W}(\mathcal{J})=\int d^{\,d}x\biggl(\mathcal{J}(x)\overline{\phi}(x)\biggr),$$
where
$$\overline{\phi}(x)=\left.\dfrac{\delta \mathcal{W}(\mathcal{J}) }{\delta \mathcal{J}}\right|_{\mathcal{J}=0}.$$
Now, we assume that the field $\overline{\phi}(x)=\phi$, is uniform. In this case, we can write the effective potential, $V(\phi)$, as
$$\Gamma[\phi]=\int\,d^{d}x \,V(\phi),$$
where $V(\phi)$ takes into account the fluctuations in the model. From above discussion it is possible to write the tree-level effective potential for each replica field in the leading replica partition function. We have $V_{tree}(\phi)=U(\phi)$ where
$$U(\phi)=\frac{1}{2}(3f_{0}-2m_{0}^{2})\phi^{2}+\frac{\lambda_{0} v}{3!}\phi^{3}+\frac{\lambda_{0}}{4!}\phi^{4},
\label{potential}$$
where $v=\sqrt{6(f_{0}-m_{0}^{2})/\lambda_{0}}$ and the replica symmetric ansatz was evoked. The false and the true vacuum states $\phi_{(\pm)}$ can be obtained
$$\phi_{(\pm)}=-\frac{3v}{2}\pm3\sqrt{-\frac{f_{0}}{2\lambda_{0}}-\frac{m^{2}_{0}}{6\lambda_{0}}}.
\label{raises}$$
Therefore, we get the following interesting result: *there are instantons-like solutions in our model*. The first term in the series representation for the functional differential operator is the diluted instanton approximation, i.e., $N$ non-interacting instantons-like solutions. For $f_{0}>m_{0}^{2}>-3f_{0}$, the system develops a spontaneous symmetry breaking in the leading replica partition function. In this case, all $N$ instantons- like solutions are complex. On the other hand, for $m_{0}^{2}<-3f_{0}$ we get a similar situation as before, however all the instantons-like solutions are real.
Vacuum transition in this theory with $N$ replicas can be described in the following way. Lowering the temperature each replica field has two non-degenerate vacuum states. The transition from the false vacuum to the true one will nucleate bubbles of the true vacuum. This first-order phase transition, in the low temperature limit, was investigated in Refs. [@aharonytricritical; @aharonyetal]. The crucial question here is which tools we can use to describe the nucleation of bubbles.
Bubble nucleation and the fermionic Dicke model {#sec:effectivemodel}
===============================================
In this section, we introduce a quite simple model to study the collective nucleation of bubbles in the disordered system. Our aim is to transform the original problem substituting by one that is technically treatable where the physical essence of the original problem is maintained. Let us remind the reader that one fundamental problem in quantum optics is the description of spontaneous emission of atoms [@Einstein; @Agarwal; @Fonda]. In fluorescence situation, in the decay by spontaneous emission the atoms tend to decay independently. However, other regime also happens when the atoms act together. Superradiance is exactly this collective behavior when $N$ excited atoms in a cavity or in the free space where they are close together, with some characteristic length, decay spontaneously [@andreev; @gross]. The Dicke model was introduced to describe such collective behavior [@Dicke; @TBrandes; @Barry]. In this model it is assumed that the system is composed by an ensemble of two-level atoms, all of them in the excited state initially. Furthermore one assume that the two-level atoms are trapped in a high-Q cavity, then effectively one single mode in the countable infinitely set of field modes trapped by the cavity interact with the atoms. Other possibility is to assume that the two-level atoms interact with the free space continuum of field modes, but all the atoms are confined in a region with a characteristic length small compare with the wavelength of the resonant field mode. Both situations can describe a collective effect of emission, the superradiance, although irreversibility occurs only in the second situation, since the high-Q cavity makes the first situation time-invertible. In conclusion, this spin-boson model, even in the case of a single mode, is able to describes a phase transition from the fluorescent to superradiant phase, characterized by the fact that atoms in quite special conditions behaves cooperatively. They start to radiate spontaneously much faster and strongly than the emission of independent atoms.
From the multimode Dicke model, with spatially varying coupling between the two-level atoms and the bosonic modes, a spin-glass behavior is obtained after integrating out the bosonic field [@Sarang; @Sachev; @Rotondo1; @Rotondo2]. What firstly comes to mind is the feasibility of the reverse situation, $i.e.$, starting from the random field Landau-Ginzburg model, a particular disordered statistical field theory model defined in the continuum, to use the Dicke model to describe the phase transitions of the system.
Let us start, discussing first the decay of one replica field from the false to the true vacuum state. Suppose that each replica field $\phi^{i}(x)$ is in the metastable state $\phi_{(+)}^{i}$. Let us assume that the free energy gap per unit volume between the metastable state $\phi^{i}_{(+)}$ and the state $\phi^{i}_{(-)}$ is $\omega_{i}$. With the bubble formation of radius $R_{i}$ the free energy decreases by $\frac{4\pi}{3} R^{3}_{i}\omega_{i}$ inside the bubble. The interface makes the free energy increases by $4\pi R_{i}^{2}\eta_{i}$ where $\eta_{i}$ is the interface free energy per unit area for each replica field. The contribution for each replica field to the free energy $\Delta F_{i}$ is $4\pi R_{i}^{2}\eta_{i}-\frac{4\pi}{3} R^{3}_{i}\omega_{i}$. There is a critical radius $R_{c}$ where for $R>R_{c}$ the nucleation of bubbles occurs. For finite temperature we have thermal nucleation of bubbles. In the case where $\beta\rightarrow \infty$ there is a quantum nucleation of bubbles. There is a standard procedure to find the decay rate in a Euclidean scalar theory [@SColeman1; @SColeman2; @Flores]. This formalism is not able to describe the collective behavior, i.e., the nucleation of $N$ bubbles. Since we would like to describe the nucleation of $N$ bubbles, we discuss here an alternative approach where the description of a cooperative behavior of two-level systems was presented.
Going back to the disordered model, lowering the temperature, each replica field has two non-degenerate vacuum states. The transition from the false vacuum to the true one will nucleate bubbles of the true vacuum. Our aim is to obtain an collective effective model to deal with a gas of $N$ real interacting instantons-like solutions (see, $e.g.$ Eq. (\[Seff(2)\])). We claim that the qualitative features of the disordered system at very low temperatures can be described by the generalized Dicke model with only one single bosonic mode. In the Dicke model there is a mean-field type phase transition with a critical temperature below which the system is in a superradiant state. Some seminal papers discussing the phase transition in such model are Refs. [@Hepp1; @Hepp2; @Wang; @Hioe].
Following Ref. [@Leggett], it is possible to represent $N$ structures with a false and true vacuum by $N$ two- level systems. Referring to Eqs. (\[instanton1\]), (\[instanton2\]) and (\[instantonN\]), we are modeling the effect of considering more terms of the series, $i.e.$, going beyond the diluted instanton approximation, as a bosonic mode interacting with all the two-level systems. The effective bosonic mode was introduced to play a two-fold role: is an effective mode that allows the interactions between the two-level systems and also to make the decay $\phi^{i}_{(+)}\rightarrow \phi^{i}_{(-)}$, possible. Note that we have actually an open system. In conclusion, the situations where nucleation of bubbles occurs, decreasing the free energy of the system will be substituted by an effective model. It is important to point out that we have assumed that going beyond the diluted instanton approximation, the vacuum structure associated to each replica field is not modified. If the inclusion of more terms of the series defined by Eq. (\[instanton1\]) increase number of false vacuum states for each replica field, it is necessary to generalize the Dicke model using intermediate statistics [@Greenberg1; @Greenberg2].
In order to achieve the effective description discussed above, let us introduce, following Popov and Fedotov, the fermionic generalized Dicke model. See also Refs. [@Aparicio1; @Aparicio2; @Aparicio3; @Aparicio4]. To proceed, let us define an auxiliary model to be called the fermionic full Dicke model in terms of fermionic raising and lowering operators $\alpha^{\dagger}_{i}$, $\alpha_{i}$, $\beta^{\dagger}_{i}$ and $\beta_{i}$, that satisfy the anti- commutator relations $\alpha_{i}\alpha^{\dagger}_{j}+\alpha^{\dagger}_{j}\alpha_{i}=\delta_{ij}$ and $\beta_{i}\beta^{\dagger}_{j}+\beta^{\dagger}_{j}\beta_{i}=\delta_{ij}$. We can also define the following bilinear combination of fermionic operators, $\alpha^{\dagger}_{i}\alpha_{i} -\beta^{\dagger}_{i}\beta_{i}$, $\alpha^{\dagger}_{i}\beta_{i}$ and $\beta^{\dagger}_{i}\alpha_{i}$ which obey the same commutation relations as the pseudo-spin operators $\sigma^z_{(\,i)}$, $\sigma^+_{(\,i)}$ and $\sigma^-_{(\,i)}$.
$$\sigma_{i}^{z}\longrightarrow \alpha_{i}^{\dagger}\alpha_{i}-\beta_{i}^{\dagger}\beta_{i}\, , \label{34}$$
$$\sigma_{i}^{+}\longrightarrow \alpha_{i}^{\dagger}\beta_{i}\, ,
\label{35}$$
and finally
$$\sigma_{i}^{-}\longrightarrow \beta_{i}^{\dagger}\alpha_{i}\, .
\label{36}$$
The Hamiltonian $ H_F$ of the auxiliary fermionic full Dicke model is
$$H_F=\frac{\Omega}{2}\,\sum_{i=1}^{N}(\alpha_{i}^{\dagger}\alpha_{i}-\beta_{i}^{\dagger}\beta_{i})
+\,\omega_0\,b^{\dagger}\,b\,+\frac{g_1}{\sqrt{N}}\sum_{i=1}^{N}
\left(b\,\alpha_{i}^{\dagger}\beta_{i}+b^{\dagger}\,\beta_{i}^{\dagger}\alpha_{i}\right)
\,+\frac{g_2}{\sqrt{N}}\sum_{i=1}^{N}
\left(b\,\beta_{i}^{\dagger}\alpha_{i}+b^{\dagger}\,\alpha_{i}^{\dagger}\beta_{i}\right),
\label{HFdipole}$$
where $\Omega$ is a known function of $m_{0}$, $\lambda_{0}$ and $f_{0}$. It is related to the energy gap between the false and the true vacuum for each replica field. See Eq. (\[raises\]). On the other hand, $\omega_{0}$, $g_{1}$ and $g_{2}$ are phenomenological quantities that are related to the physical parameters $m_{0}$, $\lambda_{0}$ and $f_{0}$ of the disordered model. In this situation, the Euclidean action $S$ associated to the fermionic Dicke model is given by
$$S=\int_0^{\beta} d\tau \left(b^*(\tau)\,\partial_{\tau}b(\tau)+ \sum_{i=1}^{N}
\Bigl(\alpha^*_i(\tau)\,\partial_{\tau}\alpha_i(\tau)
+\beta^*_i (\tau)\,\partial_{\tau}\beta_i(\tau)\Bigr)\right) -\int_0^{\beta}d\tau H_{F}(\tau)\,,
\label{66}$$
where the Hamiltonian density $H_{F}(x)$ is obtained from Eq. (\[HFdipole\]). In order to define the partition function, the functional integrals have to be done in the space of complex functions $b^*(\tau)$ and $b(\tau)$ and Grassmann variables $\alpha_i^*(\tau)$, $\alpha_i(\tau)$, $\beta_i^*(\tau)$ and $\beta_i(\tau)$. Since we use thermal equilibrium boundary conditions in the Euclidean time, the integration variables obey periodic boundary conditions for the Bose field, $i.e.$, $b(0)=b(\beta)$ and anti-periodic boundary conditions for Grassmann variables, i.e., $\alpha_i(\beta)=-\alpha_i(0)$ and $ \beta_i(\beta)=-\beta_i(0)$ [@kubo; @martin].
To proceed, let us define the formal quotient of two functional integrals, i.e., the partition function of the generalized fermionic Dicke model and the partition function of the free fermionic Dicke model. Therefore we are interested in calculating the following quantity
$$\frac{Z_{F}}{Z_{F_{0}}}=\frac{\int [d\eta]\,e^{\,S}}{\int[d\eta]\,e^{\,S_{0}}}\, , \label{65}$$
where $S$ is the Euclidean action of the generalized fermionic Dicke model given by Eq. (\[66\]), $S_0$ is the free Euclidean action for the free single bosonic mode and the free two-level systems. In the above equation $[d\eta]$ is the standard functional measure for the fermionic and bosonic degrees of freedom. The free action for the single mode bosonic field $S_{0}(b)$ is given by
$$S_{0}(b, b^{*}) = \int_{0}^{\beta} d\tau \biggl(b^{*}(\tau)\frac{\partial b(\tau)}{\partial \tau}
-\omega_{0}\,b^{*}(\tau)b(\tau)\biggr)\, . \label{67}$$
Then we can write the action $S$ of the generalized fermionic Dicke model, given by Eq. (\[66\]), using the free action for the single mode bosonic field $S_{0}(b, b^{*})$, given by Eq. (\[67\]), plus an additional term that can be expressed in a matrix form. For more details see the Refs. [@Aparicio1; @Aparicio2]. Performing straightforward calculations it is possible to show that the critical temperature $T_{c}$ where $T^{-1}=\beta$, is
$$\beta_{c}=\frac{2}{\Omega}\text{arctanh}\left[\frac{\omega_{0}\,\Omega}{(g_1+g_2)^2}\right]\,.
\label{eqfase11}$$
Notice that it is possible to have a quantum phase transition when $\omega_{0}\,\Omega=(g_1+g_2)^2$. The experimental realization of the Dicke superradiance in cold atoms in optical cavities was presented in Ref. [@bauman].
In the disordered system, this situation discussed above corresponds to the quantum nucleation of bubbles. We would like to stress that this scenario, where these bubble nucleations are a collective effect in the system, is a oversimplification of the exact full model. At this point we would like to comment the similarities between these two physical systems, the $N$ two-level systems trapped in a cavity and the random field Landau-Ginzburg model. In the first case, the ensemble of two-level atoms interacts effectively with one bosonic field mode present in the cavity. There are strong correlations between the two-level systems. In the disordered Landau-Ginzburg model the Gaussian disorder is able to make the same effect of the cavity. All the replicas are strongly correlated. See Eq. (\[Seff(2)\]). All the replicas are under the effect of the background generated by the other replicas.
Conclusions {#sec:conclusions}
===========
In this work we discuss the phase transitions in the continuous version of the $d$-dimensional random field Ising model. First we adopt the general strategy to average over the disordered free energy. Recently it was proposed a new method to average the disorder dependent free energy in systems defined in the continuum. Using this technique, the free energy is represented as a series of the replica partition functions of the system. The structure of the replica space was investigated using the saddle-point equations obtained from each replica field theory. We discuss the presence of a spontaneous symmetry breaking mechanism in some replica partition functions. For very low temperatures there are $N$ replica instantons-like solutions (real or complex) in this model. For the case of real instantons-like solutions, each replica field has two non-degenerate vacua. The transition from the false vacuum to the true one for each replica field corresponds to the nucleation of bubble of the true vacuum.
As we discussed, it is possible to obtain a spin-glass behavior from the multimode Dicke model of quantum optics, integrating out the bosonic field. This spin-boson model describes a phase transition from the fluorescent to superradiant phase. We show that the reverse situation is also feasible. To describe the phase transition in the disordered statistical field theory model we use the one mode Dicke model. The similarities between these two physical systems, the $N$ two-level systems trapped in a cavity and the random field Landau-Ginzburg model are evident.
The ensemble of two-level atoms interact effectively with one bosonic field mode present in the cavity. This fact generates strong correlations between the two-level atoms. In the disordered Landau-Ginzburg model, the Gaussian disorder is able to make the same effect, since all the replicas are strongly correlated. All the replicas are under the effect of the background generated by the other replicas.
Using the formalism developed by Popov and Fedotov the critical temperature is found. This temperature can be characterized by a non-analytical behavior of the thermodynamic quantities as a function of the temperature. At this temperature the free energy of the system is non-analytic, and the system present a transition to the normal to the superradiant phase.
A crucial question is the size of the bubbles in the disordered model. In scalar models in field theory with compactification in one spatial direction, the mass can depend upon the periodicity length in the compact direction [@LHFord; @Toms; @Toms2; @Denardo; @ford]. This situation allow that topological effects play a role in the breaking and restoration of symmetries in different models. We believe that using the formalism discussed in this section and the above discussed mechanism, it is possible to to predict the size of the nucleating bubbles.
Another natural continuation of our investigations still using the distributional zeta-function method in disordered field theory models, consists in studying the nature of phase transitions in the disordered (random temperature) $d$- dimensional Ising ferromagnet, which can be described by a statistical field theory model with quenched disorder, i.e., the $d$-dimensional random temperature Landau-Ginzburg model.
As we discussed in Sec. \[intro\], two concepts that are of fundamental importance in disordered systems are respectively quenched disorder and frustration. The presence of frustration in some disordered systems, as for example the spin glasses suggests that there are many different ground states in such systems. At low temperatures, in the spin-glass there are domains where the spins becomes frozen in space. This randomness in space that characterize the spin-glass phase corresponds to the fact that the free energy landscape of the system has a multivalley structure. Some authors discussed the possibility of a existence of this multivalley structure of the spin-glass phase in the random temperature Landau-Ginzburg model [@ma1; @targus].
Our aim is to investigate the possibility of found a multivalley structure in the average free energy of the random temperature Landau-Ginzburg model using the distributional zeta-function approach. This subject is under investigation by the authors.
Acknowledgments
===============
We would like to thank G. Krein, S. Queirós G. Menezes and M. Aparicio Alcalde for useful discussions. This paper was partially supported by Conselho Nacional de Desenvolvimento Científico e Tecnol[ó]{}gico (CNPq, Brazil).
[99]{}
P. W. Anderson, Journal of Les-Common Metals, [**62**]{}, 291 (1978). S. F. Edwards and P. W. Anderson, J. Phys. F [**5**]{}, 965 (1975). T. R. Kirkpatrick and D. Sherrington, Phys. Rev. B [**17**]{}, 4384 (1978). G. Parisi, Phys. Rev. Lett. [**43**]{}, 1754 (1979). G. Parisi, Jour. Phys. A [**13**]{}, L115 (1980). M. Mezard, G. Parisi and M. Virasoro, [*[“Spin-Glass Theory and Beyond"]{}*]{}, World Scientific, (1987). V. Dotsenko, [*[“Introduction to the Replica Theory in Disordered Statistical Systems"]{}*]{}, Cambridge University Press (2001). C. De Dominicis and I. Giardina, [*[“Random Fields and Spin Glass"]{}*]{}, Cambridge University Press (2006). B. F. Svaiter and N. F. Svaiter, Int. Jour. Mod. Phys. A [**31**]{}, 1650144 (2016). B. F. Svaiter and N. F. Svaiter, arXiv:math-phys 1606.04854 (2016). R. Acosta Diaz, C. D. Rodríguez-Camargo and N. F. Svaiter, arXiv: cond-mat 1609.07084 (2016). R. Acosta Diaz, G. Menezes, N. F. Svaiter and C. A. D. Zarro, Phys. Rev. D [**96**]{}, 065012 (2017). A. Larkin, Sov. Phys. JETP [**31**]{}, 784 (1970). T. Nattterman, in [*[“Spin-Glasses and Random Fields"]{}*]{}, A. P. Young (Editor), World Scientific (1988). T. Nattermann and P. Rujan, Int. Jour. Mod. Phys. B [**3**]{}, 1597 (1989). G. Parisi and V. Dotsenko, J. Phys. A [**25**]{}, 3143 (1992). M. Dzero, J. Schmalian and P. G. Wolynes, Phys. Rev. B [**72**]{}, 100201 (2005). A. Adams, T. Anous, J. Lee and S. Yaida, Phys. Rev. E [**91**]{}, 032148 (2015). A. Aharony, Phys. Rev. B [**18**]{}, 3318 (1977). M. Gofman, J. Adler, A. Aharony, A. B. Harris and M. Schwartz, Phys. Rev. B [**53**]{}, 6362 (1996). A. J. Leggett, C. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg, and W. Zwerger, Rev. Mod. Phys. [**59**]{}, 1 (1987). V. N. Popov and S. A. Fedotov, Theo. Math. Phys. [**51**]{}, 363 (1982). V. N. Popov and S. A. Fedotov, Sov. Phys.-JETP [**67**]{}, 535 (1988). V. N. Popov, [*[Functional Integrals and Collective Excitations]{}*]{} (Cambridge University Press, Cambridge,1987). S. Fishman and A. Aharony, J. Phys. C, [**12**]{}, L729 (1979). J. Cardy, Phys. Rev. B, [**24**]{} 505 (1985). P. G. de Gennes, J. Phys. Chem [**88**]{}, 6469 (1984). T. MacFarland, G. T. Barkema and J. F. Marko, Phys. Rev. B [**53**]{}, 148 (1996). R. L. C. Vink, K. Binder and H. Löwen, Phys. Rev. Lett. [**97**]{}, 230603 (2006). M. Mézard and A. P. Young, Europhys. Lett. [**18**]{}, 653 (1992). M. Mezard and R. Monasson, Phys. Rev. [**B50**]{}, 7199 (1994). V. Dotsenko, A. B. Harris, D. Sherington and R. B. Stinchcombe, J. Phys. A [**28**]{}, 3093 (1995). C. De Dominicis, H. Orland and T. Tenesvari, J. Phys. I [**5**]{}, 987 (1995). V. Dotsenko and M. Mézard, J. Phys. A [**30**]{}, 3363 (1997). V. Dotsenko, J. Phys. A [**32**]{}, 2949 (1999). F. Krzakala, F. Ricci-Tersengui and L. Zdeboravá, Phys. Rev. Lett. [**104**]{}, 207208 (2010). F. Krzakala, F. Ricci-Tersengui, D. Sherington and L. Zdeboravá, J. Phys. A [**14**]{}, 042003 (2011). Y. Imry and S. -K. Ma, Phys. Rev. Lett. [**35**]{}, 1399 (1975). G. Parisi and N. Sourlas, Phys. Rev. Lett. [**43**]{}, 744 (1979). G. Parisi, [*[“An introduction to the statistical mechanics of amorphous systems"]{}*]{}, in [*[“Field Theory, Disorder and Simulations"]{}*]{},Word Scientific, Singapore (1992). J. Bricmont and A. Kupiainen, Phys. Rev. Lett. [**59**]{}, 1829 (1987). J. Bricmont and A. Kupiainen, Comm. Math. Phys. [**116**]{}, 539 (1988). M. Aizenman and T. Wehr, Phys. Rev. Lett. [**62**]{}, 2503, (1989). P. G. de Gennes, Phys. Lett. A **38**, 339 (1972). G. Gaspari and J. Rudnick, Phys. Rev. B [**33**]{}, 3295 (1986). L. Schäfer, Phys. Rev. B [**35**]{}, 100201 (1987). J. Zinn-Justin, [*[Quantum Field Theory and Critical Phenomena]{}*]{} (Oxford Science Publications, Oxford, 1993). J. Iliopoulos, C. Itzykson and A. Martin, Rev. Mod. Phys. [**47**]{}, 165 (1975). R. J. Rivers, [*[Path Integral Methods in Quantum Field Theory]{}*]{} (Cambridge University Press, Cambridge,1988). J. R. Klauder, [*[Beyond Conventional Quantization]{}*]{} (Cambridge University Press, Cambridge, 2000). S. -K. Ma, [*[Modern Theory of Critical Phenomena]{}*]{} (Perseus Publishing, Cambridge, 1976). H. Rieger, Phys. Rev. B [**52**]{}, 6659 (1995). G. Parisi, in I. A. Batalin (Ed.) *et al.*: *Quantum Field Theory and Quantum Statistics*, Vol. 1, 381-392 (1988). O. Aharony, Z. Komargodski and S. Yankielowicz, JHEP [**04**]{} (2016). N. G. Fytas, V. Martin-Mayor, P. Picco and S. Sourlas, Phys. Rev. E [**95**]{}, 042117 (2017). J. M. Corwall, R. Jackiw and E. Tomboulis, Phys. Rev. D [**15**]{}, 2428 (1974). G. N. J. Añaños, A. P. C. Malbouisson and N. F. Svaiter, Nucl. Phys. B [**547**]{}, 221 (1999). N. F. Svaiter, Physica A [**285**]{}, 493 (2000). P. Contucci, C. Giardinà, C. Gilberti, G. Parisi and C. Vernia, Phys. Rev. Lett. [**103**]{}, 017201 (2009). R. T. Seeley, Proc. Symp. Pure Math. Chicago III, 288, Am. Math. Soc. (1966). D. B. Ray and I. M. Singer, Advances in Math. [**7**]{}, 145 (1971). S. W. Hawking, Comm. Math. Phys. [**55**]{}, 133 (1974). J. S. Dowker and R. Crichley, Phys. Rev. D [**13**]{}, 224 (1976). S. A. Fulling, J. Phys. A [**36**]{}, 6857 (2003). D. Lancaster, E. Marinari and G. Parisi, J. Phys. A [**28**]{}, 3359 (1995). B. Riemann, Monatsberichte der Berliner Akademie, 671 (1859). A. E. Ingham, [*“The Distribution of Prime Numbers"*]{}, Cambridge University Press, Cambridge 1990). V. E. Landau and A. Walfisz, Rend. Circ. Mat. Palermo, [**44**]{}, 82 (1920). C. E. Fröberg, BIT [**8**]{}, 187 (1968). G. Menezes and N. F. Svaiter, arXiv: 1211.5198,[*[“Quantum field theory and prime numbers spectrum"]{}*]{}, (2011). G. Menezes, B. F. Svaiter and N. F. Svaiter, Int. Jour. Mod. Phys. A [**28**]{}, 1350128 (2013). A. Voros, [*[“Zeta Functions over Zeros of Zeta Functions"]{}*]{}, Springer Verlag, Berlin, Heidelberg (2010). M. E. Fisher and M. F. Sykes, Phys. Rev. [**114**]{}, 45 (1959). D. Dhar, J. Math. Phys. [**19**]{}, 5 (1978). B. Ahrens and A. K. Hartmann, Phys. Rev. B [**83**]{}, 014205 (2011). S. Kövesi-Domokos, IL Nouvo Cimento [**33**]{}, 769 (1976). R. Menikoff and D. R. Sharp, Jour. Math. Phys. [**19**]{}, 135 (1977). C. M. Bender, F. Cooper G. S. Guralnik and D. H. Sharp, Phys. Rev. D [**19**]{}, 1865 (1979). N. F. Svaiter, Physica A [**345**]{}, 517 (2005). M. Lüscher and P. Weisz, Nucl. Phys. B 300 \[FS22\] (1988) 325. M. Lüscher and P. Weisz, Nucl, Phys. B 290 \[FS20I (1987) 25. M. Lüscher and P. Weisz, Nucl. Phys. B 295 \[FS21\] (1988) 65. T. Reisz, Nucl. Phys. B 450 \[FS\], 569 (1995). H. Meyer-Ortmanns and T. Reisz, Eur. Phys. Jour. B [**27**]{}, 549 (2002). F. Englert, Phys. Rev. [**129**]{}, 567 (1963). C. G. Bollini, J. J. Giambiagi and A. Gonzáles Dominguez, Il Nuovo Cim. [**31**]{}, 550 (1964). C. G. Bollini and J. J. Giambiagi, Nuovo Cim. B [**12**]{}, 20 (1972). J. F. Ashmore, Nuovo Cim. Lett. [**4**]{}, 289 (1972). G. ’t Hooft and M. Veltman, Nucl. Phys. B [**44**]{}, 189 (1972). G. Leibrandt, Rev. Mod. Phys. [**47**]{}, 849 (1975). D. Amit, [*[Field Theory, the Renormalization Group and Critical Phenomena]{}*]{} (McGraw-Hill, New York, 1978). A. Einstein, Physikalische Zeitschrift **18**, 121 (1917). G. S. Agarwal, *“Quantum Statistical Theories of Spontaneous Emission and their Relation to Other Approaches”* (Springer-Verlag, Berlin, 1974). L. Fonda, G. C. Ghirardi and A. Rimini, Rep. Prog. Phys. **41**, 587 (1978). A. V. Andreev, V. I. Emel’yanov and Y. A. Il’nskii, Sov. Phys. Usp. [**23**]{}, 493 (1980). M. Gross and S. Haroche, Phys. Rep. [**93**]{}, 301 (1982). R. H. Dicke, Phys. Rev. [**93**]{}, 99 (1954). T. Brandes, Phys. Rep. [**408**]{}, 315 (2005). B. M. Garraway, Phil. Trans. R. Soc. A [**369**]{}, 1137 (2011). S. Gopalakrishnan, B. L. Lev and P. M. Gouldbart, Phys. Rev. Lett. [**107**]{}, 277201 (2011) P. Stack and S. Sachdev, Phys. Rev. Lett. [**107**]{}, 277202 (2011). P. Rotondo, E. Tesio and S. Cracciolo, Phys. Rev. B [**91**]{}, 014415 (2015). P. Rotondo, M. C. Lagomarsino and G. Viola, Phys. Rev. Lett. [**114**]{}, 143601 (2015). S. Coleman, Phys. Rev. D [**15**]{}, 2929 (1977). S. Coleman and F. De Luccia, Phys. Rev. D [**21**]{}, 3305 (1980). G. H. Flores, R. Ramos and N. F. Svaiter, Int.Jour. Mod. Phys. A [**14**]{}, 3715 (1999). K. Hepp and E. H. Lieb, Ann. Phys. [**76**]{}, 360 (1973). K. Hepp and E. H. Lieb, Phys. Rev. A [**8**]{}, 2517 (1973). Y. K. Wang and F. T. Hioe, Phys. Rev. A [**7**]{}, 931 (1973). F. T. Hioe, Phys. Rev. A [**8**]{}, 1440 (1973). O. W. Greenberg and A. M. L. Messiah, Phys. Rev. [**138**]{}, B1155 (1965). O. W. Greenberg, Phys, Rev. Lett. [**64**]{}, 705 (1990). M. A. Alcalde, A. L. L. de Lemos and N. F. Svaiter, J. Phys. A [**40**]{}, 11961 (2007). M. A. Alcalde, R. Kullock and N. F. Svaiter, J. Math. Phys. [**50**]{}, 013511 (2009). M. A. Alcalde, A. H. Cardenas, N. F. Svaiter and V. B. Bezerra, Phys. Rev. A [**81**]{}, 032335 (2010). M. A. Alcalde, J. Stephany and N. F. Svaiter, J. Phys. A [**44**]{}, 505301 (2011). R. Kubo, J. Phys. Soc. Jap. [**12**]{}, 570 (1957). P. Martin and J. Schwinger, Phys. Rev. [**115**]{}, 1342 (1959). K. Bauman, R. Mottl, F. Brennecke and T. Esslinger, Phys. Rev. Lett. [**107**]{}, 140402 (2011). L. H. Ford and T. Yoshimura, Phys. Lett. A [**70**]{}, 89 (1979). D. J. Toms, Phys. Rev. D [**21**]{}, 928 (1980). D. J. Toms, Phys. Rev. D [**21**]{}, 2805 (1980). G. Denardo and E. Spalucci, Nucl. Phys. B [**169**]{}, 514 (1980). L. H. Ford and N. F. Svaiter, Phys. Rev. D [**51**]{}, 6981 (1995). S. K. Ma and I. Rudnick, Phys. Rev. Lett. [**40**]{}, 589 (1978). G. Targus and V. Dotsenko, J. Phys. [**A35**]{}, 1627 (2001).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We precisely measured branching ratios of and hypernuclei produced via reaction. Using these $\pi^0$ branching ratios with the branching ratios and the lifetimes, we obtained the decay widths and the non-mesonic weak decay widths at high statistics with the accuracy of $\sim$5%(stat) for both hypernuclei.'
author:
- |
S. Okada$^a$, S. Ajimura$^b$, K. Aoki$^c$, A. Banu$^d$, H. C. Bhang$^e$, T. Fukuda$^f$, O. Hashimoto$^g$, J. I. Hwang$^e$, S. Kameoka$^g$, B. H. Kang$^e$, E. H. Kim$^e$, J. H. Kim$^e$, M. J. Kim$^e$, T. Maruta$^h$, Y. Miura$^g$, Y. Miyake$^b$, T. Nagae$^c$, M. Nakamura$^h$, S. N. Nakamura$^g$, H. Noumi$^c$, Y. Okayasu$^g$, H. Outa$^i$, H. Park$^j$, P. K. Saha$^f$, Y. Sato$^c$, M. Sekimoto$^c$, T. Takahashi$^g$, H. Tamura$^g$, K. Tanida$^i$, A. Toyoda$^c$, K. Tsukada$^g$, T. Watanabe$^g$, H. J. Yim$^e$\
$^a$Department of Physics, Tokyo Institute of Technology, Tokyo, 152-8551, Japan\
$^b$Department of Physics, Osaka University, Osaka, 560-0043, Japan\
$^c$High Energy Accelerator Research Organization (KEK), Ibaraki, 305-0801, Japan\
$^d$Gesellschaft f$\ddot{\mbox{u}}$r Schwerionenforschung mbH (GSI), Darmstadt, 64291, Germany\
$^e$Department of Physics, Seoul National University, Seoul, 151-742, Korea\
$^f$Osaka Electro-Communication University, Osaka, 572-8530, Japan\
$^g$Department of Physics, Tohoku University, Miyagi, 980-8578, Japan\
$^h$Department of Physics, University of Tokyo, Tokyo, 113-0033, Japan\
$^i$The Institute of Physical and Chemical Research (RIKEN), Saitama, 351-0198, Japan\
$^j$Korea Reserch Institute of Standards and Science (KRISS), Daejeon, 305-600, Korea\
title: decay branching ratios of and hypernuclei
---
Introduction
============
It is well known that a $\Lambda$ hyperon in free space decays into a nucleon associated with a pion (: $\Lambda \to p\pi^{-}$, : $\Lambda \to n \pi^{0}$). A bound $\Lambda$ in a nucleus ($\Lambda$ hypernucleus) also decays via the decay process, called the mesonic weak decay of $\Lambda$ hypernucleus. The relevant momentum transfer is only $\sim$100 MeV/$c$ which is not necessarily enough high to exceed the Fermi momentum. The decay rate is therefore suppressed due to the Pauli blocking effect on the outgoing nucleon. Especially in light hypernuclei, it is sensitive to overlap of the $\Lambda$ wave function with the nucleus. Thus, the mesonic weak decay width of light hypernuclei gives significant information to investigate $\Lambda$-nucleus potential shape.
It is believed that the folding potential between $\Lambda$ and light (s-shell) nuclei has a central repulsive core, which is calculated from the commonly used YNG $\Lambda N$ interaction. However, it has not been confirmed so far experimentally. For comparison of $\alpha$-$\Lambda$ potential, Motoba calculated the mesonic decay widths of for two different potentials [@Mot94N]. One is derived from the YNG interaction, and the other is a simply attractive potential derived from one-range-gaussian two-body interaction called “ORG”, which are determined to reproduce the $\Lambda$ binding energy of . According to their calculation, the difference in the decay widths between them is $\sim$20% as shown in Table \[table1\]. However, existent experimental data cannot distinguish the two due to the large error. In the present experiment, we precisely measured both mesonic decay widths, [@Kam03] and . On the other hand, the bound $\Lambda$ in a nucleus can interact with a neighboring nucleon (: $\Lambda p \to np$, : $\Lambda n \to nn$), called the non-mesonic weak decay of $\Lambda$ hypernucleus. It gives unique opportunity to study baryon-baryon weak interaction which is hidden by strong interaction in normal nuclei. The total non-mesonic weak decay width () is one of the most important observables for the study. It is difficult to measure the (= $+$) directly due to experimental difficulties such as final state interaction effect. Thus, the is usually obtained by subtracting the mesonic weak decay widths from the total decay width (inverse of the lifetime), as $=$ $-$$-$. Table \[table1\] shows the latest experimental data for and . The errors of for both hypernuclei mainly come from those of , so that the precise measurement of is awaited.
In the present paper, we concentrate the measurement of $\pi^{0}$ branching ratios, and s and s for and are derived from the $\pi^-$ branching ratios and the lifetimes.
Refs. / / / /
--------- ------------------ --------------- ----------------- ----------------- -----------------
(exp.) [@Szy91] 1.03$\pm$0.08 0.44$\pm$0.11 0.18$\pm$0.20 0.41$\pm$0.14
(ORG) [@Mot94N] 0.321 0.177
(YNG) [@Mot94N] 0.393 0.215
(exp.) [@Sat03][@Sak91] 1.14$\pm$0.08 0.113$\pm$0.015 0.200$\pm$0.068 0.828$\pm$0.087
: Previous experimental results and theoretical calculations for the and . []{data-label="table1"}
\
Experimental Method
===================
The present experiments (E462/E508) were performed at the K6 beam line of the KEK 12-GeV proton synchrotron (KEK-PS). Hypernuclei of $^5_{\Lambda}$He and $^{12}_{\Lambda}$C were produced by the ($\pi^+$,$K^+$) reaction at 1.05 GeV/$c$ on $^6$Li and $^{12}$C (active) targets. The hypernuclear mass spectra were calculated by reconstructing momenta of incoming $\pi^+$ and outgoing $K^+$ using the beam line spectrometer (QQDQQ) and the SKS spectrometer, respectively.
The schematic view of the decay counter system is shown in Ref [@Out03]. Neutral decay particles were detected by the T4 counter arrays comprising 6 layers of 5$cm$-thick plastic scintillators. Charged decay particles were vetoed by thin plastic scintillators installed just before T4 counter arrays. $\pi^0$ from NMWD was identified by detecting high energy $\gamma$ ray, because the energy of this $\gamma$ ray is about 70 MeV, and that from other decay process is about a few MeV. The $\gamma$ rays were separated from neutrons by means of time-of-flight technique between the start timing counter of incident beam and the T4 counter.
Analysis and Results
====================
The formations of each hypernucleus were identified by gating the ground state region in the excitation energy spectra of and as shown in Figure \[fig:Pi0ExcitationEnergy\] (a). Neutral particles from the decay were detected at T4 counter with 2 MeVee (MeV electron equivalent) threshold. The 1/$\beta$ spectrum for $^{12}_{\Lambda}$C is shown in Figure \[fig:Pi0BetainvE508\] (a), which shows good $\gamma$/$n$ separation. The $\gamma$ gate corresponds to 0 $\le$ 1/$\beta \le$ 2. Using the yields below the $\gamma$ peak (1/$\beta <$ 0), the accidental background within the $\gamma$ gate was estimated as good as $\sim$ 2 %.
In order to estimate the efficiency of the detector setup, the GEANT-based Monte Carlo simulation was performed. The efficiency depends on the energy of $\pi^0$. We assumed mono-energetic (104.9 MeV/$c$) $\pi^0$ for ( $\to \pi^0 + ^5$He (g.s.)), and we used the $\pi^0$ distribution for given by Motoba [@Mot94P]. Figure \[fig:Pi0ADCsumE462\] shows the $\gamma$ energy spectra for . The points with error bars are the experimental data, and the shaded one is the simulation. To select $\gamma$-ray shower clearly, we applied the multiplicity cut for the identification. Upper figure shows the spectrum with applying multiplicity $M \ge 1$, and lower figure shows that for $M \ge 2$. There is a low energy background in the spectrum for the $M \ge 1$ condition, whereas the background disappear in that for $M \ge 2$ condition. In order to remove the low energy background completely, we determined the cut condition as “$M \ge 2$” and “ADC sum $\ge 20$ MeVee”. The 1/$\beta$ spectrum applied this cut condition is shown in Figure \[fig:Pi0BetainvE508\] (b). In this figure, the $\gamma$ ray from $\pi^0$ decay more clearly separated from neutron. For the efficiency estimation, the same cut condition was applied in the simulation. The detection efficiency (including the detector acceptance) is estimated to be $\varepsilon \sim$ 10.5 %. The good agreement of the energy spectra between data and simulation in Figure \[fig:Pi0ADCsumE462\] (2) shows that we can count for the efficiency estimation.
Figure \[fig:Pi0ExcitationEnergy\] (b) shows the excitation energy of and with the cut condition. The branching ratio is represented by $b_{\pi^{0}} = N_{\pi^{0}} / N_{inc} / \varepsilon$, where $N_{inc}$ and $N_{\pi^{0}}$ are the numbers gated for the ground state regions shown in the Figure \[fig:Pi0ExcitationEnergy\] (a) and (b), respectively. Consequently, the branching ratios for and were determined to be $b_{\pi^-}$ = 0.212 $\pm$ 0.008 and 0.133 $\pm$ 0.005, respectively (statistical error only), though preliminary yet.
s and s for and were derived from our results of the lifetimes and the $\pi^{-}$ branching ratios[@Kam03] as shown in Table \[table2\]. The result of for is located in between those of ORG- and YNG-based calculations, which is consistent with the result [@Kam03]. It indicates that the $\alpha$-$\Lambda$ overlapping is larger than that of the YNG-based calculation.
The statistical errors of obtained s for and were much improved as 34% $\to$ 5% for and 11% $\to$ 5% for . The theoretical calculations of non-mesonic weak decay are required to meet these $\Gamma_{nm}$ results and our $\Gamma_n/\Gamma_p$ results[@Out03] simultaneously.
/ / / /
--- -------------------------- ---------------------------------- ------------------------- -------------------------
0.947$\pm$0.038 [@Kam03] 0.340$\pm$0.016 [@Kam03] [**0.201$\pm$0.011**]{} [**0.406$\pm$0.020**]{}
1.242$\pm$0.042 [@Kam03] 0.123$\pm$0.015 [@Sat03][@Kam03] [**0.165$\pm$0.008**]{} [**0.953$\pm$0.032**]{}
: Summary of present preliminary results for the and .[]{data-label="table2"}
\
[9]{} T. Motoba [*et al.*]{}, [*Nucl. Phys.*]{} [**A534**]{}, 597 (1994). S. Kameoka , in this proceedings. J. J. Szymanski , [*Phys. Rev.*]{} [**C43**]{}, 849 (1991). Y. Sato , to be published in [*Phys. Rev.*]{} [**C**]{}. A. Sakaguchi , [*Phys. Rev.*]{} [**C43**]{}, 73 (1991). T. Motoba and K. Itonaga, [*Prog. Theor. Phys. Suppl.*]{} [**No. 117**]{} 447 (1994). H. Outa , in this proceedings.
![Excitation energy spectra of (left figure) and (right figure). (a) for inclusive, (b) with the $\pi^{0}$ cut condition.[]{data-label="fig:Pi0ExcitationEnergy"}](Pi0ExcitationEnergyE462.eps){width="1.00\linewidth"}
![Excitation energy spectra of (left figure) and (right figure). (a) for inclusive, (b) with the $\pi^{0}$ cut condition.[]{data-label="fig:Pi0ExcitationEnergy"}](Pi0ExcitationEnergyE508.eps){width="1.00\linewidth"}
![$\gamma$ energy (ADC sum) spectra from decay of (point with error bar) are compared with the simulation (shaded one). (1) layer multiplicity $M \ge 1$, (2) $M \ge 2$.[]{data-label="fig:Pi0ADCsumE462"}](Pi0BetainvE508.eps){width="0.95\linewidth"}
![$\gamma$ energy (ADC sum) spectra from decay of (point with error bar) are compared with the simulation (shaded one). (1) layer multiplicity $M \ge 1$, (2) $M \ge 2$.[]{data-label="fig:Pi0ADCsumE462"}](Pi0ADCsumE462.eps){width="0.95\linewidth"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The thermodynamic equilibrium states of a static thin ring shell in a (2+1)-dimensional flat spacetime is analyzed. Inside the ring the spacetime is flat, whereas outside it is conical flat. The first law of thermodynamics applied to the thin shell leads to a shell’s entropy which is a function of its mass alone. Two simple forms for this mass function are given leading to two different expressions for the entropy. The equations of thermodynamic stability are analyzed resulting in certain allowed regions for the free parameters. Contrary to the usual (3+1)-dimensional case this shell’s entropy is purely classic, as the only fundamental constant that enters into the problem is the (2+1)-dimensional gravitational constant $G_3$, which has units of inverse mass.'
author:
- 'José P. S. Lemos'
- 'Gonçalo M. Quinta'
title: 'Thermodynamics, entropy, and stability of thin shells in 2+1 flat spacetimes'
---
Introduction
============
The study of thermodynamics of gravitational systems has proved important in many respects, such as the understanding that these systems can have negative specific heat, and that the gravitational object par excellence, the black hole, has definite temperature and entropy associated with it (see, e.g., [@york1]), which in turn indicates that a quantum theory of gravity might be in sight. In particular, a special kind of gravitating system, a thin shell with its surrounding spacetime, is prone to a direct attack of its thermodynamic properties. Indeed, Davies, Ford and Page [@daviesfordpage] and Hiscock [@hiscock], have shown the usefulness of studying thin shells in (3+1)-dimensional general relativistic spacetimes from a thermodynamic viewpoint. Further progress was achieved by Martinez [@Mart], in which several thermodynamic quantities are discussed and a stability analysis of thin shells is performed using the formalism set in [@callen]. For related studies of thermodynamics of gravitating matter, especially on the verge of becoming a black hole, see [@lemoszaslavski1; @lemoszaslavskii2].
Although (3+1)-dimensional spacetimes are of the greatest interest, it is also pertinent to study thin shells and their properties in (2+1)-dimensional spacetimes. The relevance in the study of three-dimensional gravity started with the work of Deser, Jackiw, and ’t Hooft [@djt], where it was shown that, though the corresponding vacuum solution is trivial since it consists of Minkowski spacetime, a point particle distorts it into a conical space with the particle being located at the vertex of the cone, and moreover, moving point particles display nontrivial dynamics. The next simplest object, beyond a point a particle in 2+1 dimensions, is a thin shell, i.e, a ring dividing two vacua regions, the interior and exterior to the ring itself. A detailed analysis of the dynamics of such a ring, collapsing or expanding was performed in [@mannoh], where the usual junction condition formalism [@Israel] was used.
To further understand the thermodynamic properties of gravitational systems we propose to study the thermodynamics of a static thin ring shell in (2+1)-dimensional general relativity. By making use of the appropriate junction conditions for general relativity [@Israel], one can determine the pressure and rest mass of the shell in order for it to be static with interior and exterior spacetimes both flat. Using then the formalism developed by Martinez [@Mart], with the thermodynamic theory as presented in [@callen], one can find generic expressions for the shell’s entropy, which upon some minimal assumptions about the structure of the matter fields making the system, i.e., an ansatz for the shell temperature in terms of the gravitational quantities that characterize the system, yields a definite expression for the entropy of the shell, and permits a stability analysis.
The paper is organized as follows. In Sec. \[thinsh\], we will compute the components of the extrinsic curvature of the shell that will lead to the shell’s linear density and pressure. In Sec. \[thermo\] we will use these results by directly inserting them in the first law of thermodynamics to obtain the differential of the entropy, which will naturally have a degree of freedom parametrized by an arbitrary function associated with the matter fields that make up the shell. Two phenomenological expressions will be considered for the arbitrary function, namely a simple power law of the rest mass and a power law of a quadratic expression of the rest mass, which will lead to two different expressions for the entropy. In Sec. \[eqstate\] we will analyze the thermodynamic stability of the system for each entropy function by calculating the allowed intervals of the free parameters in order for the shell to remain thermodynamically stable. In Sec. \[conc\] we conclude.
The thin shell spacetime {#thinsh}
========================
In 2+1 dimensions Einstein’s equation takes the form $$\label{efe}
G_{ab} = 8\pi G_3 T_{ab}\,,$$ where $G_{ab}$ is the (2+1)-dimensional Einstein tensor, $T_{ab}$ is the stress-energy tensor, and $G_3$ is the gravitational constant in 2+1 dimensions. $G_3$ has units of inverse mass. The speed of light is taken to be equal to $1$.
To find the solution for a thin shell in a (2+1)-dimensional spacetime, we follow [@Israel] and start by considering a one-dimensional timelike hypersurface $\Sigma$ that partitions spacetime into two spherically symmetric regions: an inner region $\mathcal{V}_-$ and an outer region $\mathcal{V}_+$.
Inside the hypersurface we will use flat-polar coordinates $(t,r,\theta)$ for a flat metric with line element $$\label{LEI}
ds^2 = -dt^2 + dr^2 + r^2 d\theta^2\,, \quad r<R\,,$$ where $r=R$ is the radius of the thin-shell hypersurface.
On the outside of the shell, the spacetime is again flat but the presence of matter justifies the use of conical-polar coordinates $(t,r,\theta)$, allowing the line element to be written as $ds^2 =
-\beta^2 dt^2 + dr^2 + r^2 {\alpha}^2 d\,\theta^2$, for $r>R$, and some constant ${\alpha}$. It is preferable to make the change ${\alpha}\,r\to r$, and without loss of generality one can put $\beta=\alpha$, so that the metric takes the form $$\label{LEO}
ds^2 = - \alpha^2
dt^2 + \frac{dr^2}{{\alpha}^2} + r^2
d\,\theta^2\,, \quad r>R\,.$$ The metric (\[LEO\]) has a conical singularity at $r=0$ if the thin-shell hypersurface has a radius $R(\tau) \to 0$, i.e., it turns into a point particle.
The parametric equations of the thin shell hypersurface $\Sigma$ are described by $r=R(\tau)$, $t = T(\tau)$, where $\tau$ is the proper time on the thin shell hypersurface. Choosing coordinates $(\tau,\theta)$, we have an induced metric $h_{ab}$ given by $$ds_{\Sigma}^2 = -d\tau^2 + R^2(\tau) d\theta^2.$$
The first junction condition states that in order for $\mathcal{V}_+$ and $\mathcal{V}_-$ to be joined smoothly at $\Sigma$, the induced metric seen from both sides must satisfy $[h_{ab}]=0$, where the parentheses symbolize the jump in the metric across the hypersurface. This condition leads to the relations $$\label{J1}
{\alpha}^2 \dot{T}^2 - \frac{\dot{R}^2}{{\alpha}^2} = \dot{T}^2 - \dot{R}^2 = 1\,$$ where the dot denotes differentiation with respect to $\tau$.
The second junction condition involves the extrinsic curvature $K^{a}{}_{b}$ defined as $$K^{a}{}_{b} = h^{ac}h^{d}{}_{b}\nabla_{c} n_{d}\,$$ where $\nabla_{c}$ denotes covariant derivation. In general, when there is a thin matter shell, the boundary stress-energy tensor $S^{a}{}_{b}$ is related to the jump in extrinsic curvature by $$S^{a}{}_{b}=-\frac{1}{8 \pi G_3}\left([K^{a}{}_{b}]-[K]h^{a}{}_{b}\right)$$ where $K = h^{b}{}_{a} K^{a}{}_{b}$ and $G_3$ is the gravitational constant in 2+1 dimensions. From the line elements given in Eqs. (\[LEI\]) and (\[LEO\]) and using Eq. (\[J1\]), one can compute the nonzero components of $K^{a}{}_{b}$, $$\begin{aligned}
K^{\tau}_{+}{}_{\tau} &= \frac{\ddot{R}}{\sqrt{{\alpha}^2+\dot{R}^2}}\,, \\
K^{\tau}_{-}{}_{\tau} &= \frac{\ddot{R}}{\sqrt{1+\dot{R}^2}}\,, \\
K^{\theta}_{+}{}_{\theta} &= \frac{1}{R}\sqrt{{\alpha}^2+\dot{R}^2}\,, \\
K^{\theta}_{-}{}_{\theta} &= \frac{1}{R}\sqrt{1+\dot{R}^2}\,,\end{aligned}$$ which, in turn, lead to the nonzero components of the linear boundary stress-energy tensor $$\begin{aligned}
S^{\tau}{}_{\tau} & = \frac{1}{8 \pi G_3}\frac{\sqrt{{\alpha}^2
+ \dot{R}^2}-\sqrt{1+\dot{R}^2}}{R} \label{S1} \\
S^{\theta}{}_{\theta} & = \frac{1}{8 \pi G_3}
\left(\frac{\ddot{R}}{\sqrt{{\alpha}^2 +\dot{R}^2}}-
\frac{\ddot{R}}{\sqrt{1+\dot{R}^2}}\right). \label{S2}\end{aligned}$$
Assuming the shell is a perfect fluid, the specific physical form of the linear boundary tensor is $$S^{a}{}_{b} = (\lambda +p) u^a u_b + p
h^{a}{}_{b}\,,
\label{perffluid}$$ where $p$ is the linear pressure and $\lambda$ is the linear mass density of the shell. So for a perfect fluid $S^{\tau}{}_{\tau} = -\lambda$ and $S^{\theta}{}_{\theta} = p$. Thus, from Eqs. (\[S1\]) and (\[S2\]), it follows that $$\begin{aligned}
\lambda & = \frac{1}{8 \pi G_3}\frac{\sqrt{1
+ \dot{R}^2}-\sqrt{{\alpha}^2+\dot{R}^2}}{R} \\
p & = \frac{1}{8 \pi G_3}\left(\frac{\ddot{R}}{
\sqrt{{\alpha}^2 +\dot{R}^2}}-\frac{\ddot{R}}{\sqrt{1+\dot{R}^2}}\right)\end{aligned}$$ This set of equations has been previously found and studied in [@mannoh] in a 2+1 dynamical shell collapsing setting. Taking the static limit $\dot{R} = \ddot{R} = 0$ in the above equations and using the definition of the shell’s rest mass $$M = 2 \pi
R \lambda\,,
\label{massdef}$$ we have immediately $$\begin{aligned}
M & = \frac{1-{\alpha}}{4 G_3} \label{m}\, \\
p & = 0. \label{m2}\end{aligned}$$ To have a properly defined shell radius one has to impose ${\alpha}>0$. Imposing positive mass it follows $0<{\alpha}<1$, whereas negative masses appear in the range $1<{\alpha}<\infty$. From Eq. (\[m2\]) we see that in order for the shell to be static in a (2+1)-dimensional spacetime, its linear pressure must vanish.
We now want to study this static thin shell spacetime [@mannoh] from a thermodynamics point of view. Through the formalism presented in [@york1] and developed by Martinez [@Mart] for the thermodynamics of thin shells, we study the thermodynamics and find the entropy of these thin shells in (2+1) flat dimensions.
Thermodynamics, entropy equation for the shell, and stability {#thermo}
=============================================================
We assume that the entropy of the shell can be expressed in terms of its characteristics, namely, its rest mass $M$ and radius $R$. It is more useful, though equivalent, to work with $M$ and $A$ [@Mart], where $A$ is the area (here circumference) of the shell, $A \equiv 2 \pi R$. Thus we write, $$S = S(M,A)\,.$$ Assuming further that the shell is at some local temperature $T$, an explicit expression for $S$ can then be found by directly integrating the first law of thermodynamics $$TdS = dM + p\, dA\,.$$ By putting $p=0$ as given in Eq. (\[m2\]), the first law for the (2+1)-dimensional thin shell spacetime simplifies to $$\label{1L}
T dS = dM\,.$$ Equation (\[1L\]) can be integrated provided that the integrability conditions are satisfied. In this case there is only one condition. It states that $T=T(M)$, i.e., the temperature $T$ is a function of the mass $M$ alone. This simple dependence stems from the fact that $p=0$ as given in Eq. (\[m2\]). Then, the most general expression for $S$ is $$\label{SM}
S(M) = \int {\beta}(M)dM + S_0$$ where $\beta \equiv 1/T$ is the local inverse temperature of the shell at the equilibrium position $r = R$, and $S_0$ is an integration constant. Note that ${\beta}(M)$ is an arbitrary function of the mass that can be specified once the specific matter fields that constitute the shell are known.
Local intrinsic stability of the shell can also be studied. The former is guaranteed as long as all the following inequalities are verified $$\label{C1}
\left(\frac{{\partial}^2 S}{{\partial}M^2}\right)_A \leq 0\,,$$ $$\label{C2}
\left(\frac{{\partial}^2 S}{{\partial}A^2}\right)_M \leq 0\,,$$ $$\label{C3}
\left(\frac{{\partial}^2 S}{{\partial}M^2}\right)\left(\frac{{\partial}^2 S}{{\partial}A^2}\right)
- \left(\frac{{\partial}^2 S}{{\partial}M {\partial}A}\right)^2 \geq 0\,,$$ where the formalism developed in [@callen] is being followed.
Two specific equations of state for the thin shell matter: Entropy and stability {#eqstate}
================================================================================
The simplest equation of state
------------------------------
The simplest form that can be considered for ${\beta}(M)$ is a power law, $$\label{f}
{\beta}(M) = {\gamma}\, G_3^{(1+u)}M^u\,$$ where we are assuming $M\geq0$, ${\gamma}$ and $u$ are free parameters with ${\gamma}>0$ to guarantee positive temperature, the factor $G_3$ must be present for dimensional reasons, and Boltzmann’s constant is set to $1$. In 3+1 dimensions, Planck’s length $l_{\rm p}$, with $l_{\rm p}=\sqrt{G_4\,h}$ ($G_4$ being the gravitational constant in four-dimensional spacetime and $h$ Planck’s constant), appears naturally in the temperature of a thin shell, since for a given mass there is always an intrinsic length associated with it (the gravitational radius of the system), and so to have the correct units for the entropy one must resort to $l_{\rm p}$. However, here in flat 2+1 dimensions there is no intrinsic spacetime radius, and so Planck’s length does not appear in this analysis at all. This problem is thus purely classic and $G_3$, with units of inverse mass, suffices to set the scale.
When ${\beta}(M)$ has the form given in Eq. (\[f\]), one can substitute this in (\[SM\]) to get $$\label{S11}
S(M) = \frac{{\gamma}}{u+1}\left(G_3\,M\right)^{(1+u)} + S_0\,,
\,\quad {\rm for} \; u \neq -1\,,$$ and $$\label{S12}
S(M) = {\gamma}\ln(G_3\,M) + S_0\,,
\,\quad {\rm for} \; u = -1\,.$$
Although the values of the parameters $\gamma$ and $u$ cannot be calculated without first specifying the nature of the matter fields, it is possible to constrain them such that physical equilibrium states of the shell are possible. Starting with $S_0$, it is natural to assume that a zero mass shell should have zero entropy, i.e., $ S(M
\to 0)=\int {\beta}(M)dM + S_0 \to 0$. It is seen directly from Eq. (\[f\]) that the entropy diverges when $M\to0$ for $u\leq-1$. Therefore the above normalization condition can only be satisfied for $u > -1$ and $S_0 = 0$.
In relation to stability, it is seen that conditions (\[C2\]) and (\[C3\]) are automatically satisfied for any $u$. On the other hand, one can find that condition (\[C1\]) can only be satisfied provided that $u\leq0$. Thus we conclude that assuming a power law equation of the form (\[f\]), stability of the shell is possible for any $M\geq0$ as long as the parameter values of $u$ are restricted to $$\label{u}
-1 < u \leq 0\,.$$
One can also consider negative values of $M$, i.e., $\alpha>1$. The relation (\[f\]) would be of the same form with the proviso that one takes the absolute values of $M$. The same results would follow.
A more contrived equation of state
----------------------------------
Another possibility for ${\beta}(M)$ could be a quadratic function in $M$, of the form $ {\beta}(M) =
{\delta}\, G_3^{(1+a)}(M+C M^2)^a
$, where ${\delta}$ and $a$ are some parameters with ${\delta}>0$ to guarantee positive temperature, and $C$ is some constant. The constant $C$, however, has a natural connection to Eq. (\[m\]). Indeed, defining $m \equiv M + C M^2$ we can solve for $M$, obtaining the physical solution $
M = \frac{-1\\+\sqrt{1+4 C m}}{2 C}.
$ Comparing with Eq. (\[m\]), we can make the association $C = -2 G_3$ and so ${\alpha}= \sqrt{1+4Cm} = \sqrt{1-8 G_3 m}$, thus arriving at the natural quadratic expression for ${\beta}$ $$\label{b}
{\beta}(M) = {\delta}\,G_3^{(1+a)}(M - 2 G_3 M^2)^a.$$ Equation (\[b\]) is most easily integrated in the variable $m$. Changing from $M$ to $m$ in Eq. (\[SM\]), defining the parameter $\eta = {\delta}/{\alpha}$, and changing back again to $M$, we obtain the entropy $$\begin{aligned}
\label{S21}
S(M) = \frac{\eta}{a+1}\, G_3^{(1+a)} (M - 2 G_3 & M^2)^{(1+a)} + S_0
\,,\nonumber\\
\,& \quad {\rm for} \; a \neq -1\end{aligned}$$ and $$\label{S22}
S(M) = \eta \ln \left[G_3\left(M - 2 G_3\, M^2\right)\right] + S_0
\,,
\,\quad {\rm for} \; a = -1\,.$$
Again, although the values of the parameters $\eta$ and $a$ cannot be calculated without first specifying the nature of the matter fields, it is possible to constrain them such that physical equilibrium states of the shell are possible. Starting with $S_0$, it is natural to assume that a zero mass shell should have zero entropy, i.e., $ S(M
\to 0)=\int {\beta}(M)dM + S_0 \to 0$. It is seen directly from (\[b\]) that the entropy diverges when $M\to0$ for $a > -1$. Therefore the above normalization condition can only be satisfied for $a > -1$ and $S_0 = 0$.
From Eq. (\[b\]), the stability equations (\[C2\]) and (\[C3\]) are automatically satisfied since the entropy does not depend on $A$. Also, as mentioned above, ${\alpha}>0$ to have a physical acceptable solution. Considering Eq. (\[C1\]), it is possible to show that it implies the inequality $$(2a+1){\alpha}^2 -1 \leq 0\,.
\label{aaa}$$ We can study the two cases, $M>0$ and $M<0$.
For $M>0$, ${\alpha}$ is in the range $0 < {\alpha}\leq 1$. In this case Eq. (\[aaa\]) is automatically satisfied if the exponent $a$ obeys $a \leq 0$. It is also satisfied for $a
> 0$ but only if $0<{\alpha}\leq \sqrt{\frac{1}{2a+1}}$. Equivalently, this means that the rest mass $M$ of the shell must be within the range $$\frac{1-\sqrt{1-2a/(2a+1)}}{4 G_3} < M \leq \frac{1}{4 G_3}\,,$$ for $a> 0$.
For $M<0$, we know that ${\alpha}>1$, and Eq. (\[aaa\]) requires additionally that ${\alpha}\leq \sqrt{\frac{1}{2a+1}}$. In terms of rest mass $M$, this represents the range $$\frac{1-\sqrt{1-2a/(2a+1)}}{4 G_3} < M < 0.$$ However, since $\sqrt{\frac{1}{2a+1}} > 1$ and $2a +1 >0$, we see that the analysis for $M<0$ is only valid for parameter values $-1/2 < a < 0$.
Conclusions {#conc}
===========
General relativity in a (2+1)-dimensional spacetime has no curvature in empty space but in matter distributions curvature may still exist. Einstein’s equation thus still plays a role in determining the required pressure and energy of a static thin shell (or ring in this (2+1)-dimensional setting). Indeed, we have seen that in this situation the pressure must be zero and the rest mass of the shell must satisfy Eq. (\[m\]).
Upon using the first law of thermodynamics we have found a specific differential equation for the entropy of the ring that contained a degree of freedom encoded in the inverse temperature $\beta$. We have chosen the two simple ansatz for the inverse temperature, a power law on the shell’s rest mass and a quadratic form of it, obtaining two distinct expressions for the ring’s entropy. This shell’s entropy is purely classic, as the only fundamental constant that enters into the problem is the (2+1)-dimensional gravitational constant $G_3$, which has units of inverse mass. This entropy could perhaps be explained quantically if the shell’s mass is given in terms of the mass of its elementary constituents and a fundamental theory for the mass of those constituents particles is at hand. A thermodynamic stability analysis yielded the range for the allowed parameters, revealing that the shell’s rest mass must be confined to a given interval if the shell is to be stable.
Our results are of importance if one uses a concrete model for the shell, e.g., a model involving fundamental scalar fields. One would then extract from the model the specific equation of state with its precise values for the exponents and constants, and could immediately determine through the above equations whether it would be thermodynamic stable.
We also note that these ring shell spacetimes when extended into 3+1 dimensions, through the use of a trivial coordinate $z$ say, represent infinite cylinders. Thus this thermodynamic study also holds in 3+1 general relativity for those cylindrical thin shells.
We thank FCT-Portugal for financial support through Projects No. PTDC/FIS/098962/2008 and No. PEst-OE/FIS/UI0099/2011.
[99]{}
H. Braden, J. D. Brown, B. F. Whiting, and J. W. York, Phys. Rev. D **42**, 3376 (1990).
P. C. W. Davies, L. H. Ford, and D. N. Page, Phys. Rev. D **34**, 1700 (1986).
W. A. Hiscock, Phys. Rev. D **53**, 7062 (1996).
E. A. Martinez, Phys. Rev. D **40**, 1336 (1989).
H. B. Callen, *Thermodynamics and an Introduction to Thermostatistics* (Wiley, New York, 1985).
J. P. S. Lemos and O. B. Zaslavskii, Phys. Rev. D **81**, 064012 (2010).
J. P. S. Lemos and O. B. Zaslavskii, Phys. Lett. B **695**, 37 (2011)
S. Deser, R. Jackiw, and G. ’t Hooft, Ann. Phys. (N.Y.) **152**, 220 (1984).
R. B. Mann and J. J. Oh, Phys. Rev. D **74**, 124016 (2006).
W. Israel, Nuovo Cimento B **44**, 1 (1966).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we present the results of measurements of the thermal conductivity of Cu$_2$Te$_2$O$_5$Br$_2$, a compound where tetrahedra of Cu$^{2+}$ ions carrying $S=1/2$ spins form chains along the $c$-axis of the tetragonal crystal structure. The thermal conductivity $\kappa$ was measured along both the $c$- and the a-direction as a function of temperature between 3 and 300 K and in external magnetic fields $H$ up to 69 kOe, oriented both parallel and perpendicular to the $c$-axis. Distinct features of $\kappa(T)$ were observed in the vicinity of $T_{N}=11.4$ K in zero magnetic field. These features are unaltered in external fields which are parallel to the $c$-axis, but are more pronounced when a field is applied perpendicularly to the $c$-axis. The transition temperature increases upon enhancing the external field, but only if the field is oriented along the $a$-axis.'
author:
- 'A.V. Sologubenko, R. Dell’Amore and H.R. Ott'
- 'P. Millet'
title: 'Anisotropic field dependence of the magnetic transition in Cu$_2$Te$_2$O$_5$Br$_2$'
---
epsf
Introduction
============
Thermal transport in low-dimensional quantum spin systems has recently been investigated in detail, both experimentally and theoretically. Considerable progress has been made in the theoretical understanding of heat transport in idealized one-dimensional (1D) and two-dimensional (2D) model spin systems. For 1D systems that are dominated by antiferromagnetic (AFM) couplings of the spins, the integrability of the corresponding model Hamiltonians leads to interesting and nontrivial results for various transport properties, including spin-, charge-, and energy-transport [@Castella95; @Zotos97; @Saito96; @Narozhny98; @Naef98; @Kluemper02; @Alvarez02_Ano; @HeidrichMeisner02; @Saito03]. For the description of heat transport in real materials, more realistic models, considering perturbations such as spin-lattice coupling, defects and three-dimensional interactions have to be considered. Only a small amount of theoretical work along these lines is available in the literature. For example, the spin-phonon coupling in Heisenberg AFM $S=1/2$ spin chains has been considered and calculations including the interaction of spins with defects in AFM $S=1/2$ spin chains and ladders were made [@Shimshoni03; @Orignac03].
Experimental investigations treating quasi-2D spin systems concentrated on measurements of thermal transport in layered cuprates and vanadates [@Nakamura91; @Cohn95; @Hess03; @Hofmann03; @Sun03; @Sales02] and in the Shastry-Sutherland spin-lattice compound SrCu$_2$(BO$_3$)$_2$ [@Vasilev01; @Hofmann01; @Kudo01_The]. For quasi-1D systems, experimental results were reported for $S=1/2$ spin ladders [@Sologubenko00_lad; @Hess01] and $S=1/2$ Heisenberg spin chains [@Sologubenko01; @Sologubenko03_Uni; @Sologubenko03_Dif; @Markina03], including the inorganic spin-Peierls compound CuGeO$_3$ [@Vasilev97; @Ando98; @Salce98; @Vasilev98; @Takeya00; @Takeya00a; @Hofmann02; @Takeya01]. The common feature of all these compounds is that the heat transport is dominated by phonons, except along the directions of strong spin-spin interactions, i.e., along the chains in 1D systems and in the plains in 2D systems. For these cases, significant heat transport carried by spin excitations is observed in limited temperature intervals. In most materials, however, the spin system simply acts as a source of phonon scattering and cannot be regarded as a channel of significant energy transport.
The material studied in the present work may be viewed as a quasi-zero-dimensional spin system. The essential structural subunits are weakly interacting spin tetrahedra. The ground state and the excited states of noninteracting spin tetrahedra are well understood [@Johnsson00; @Lemmens01]. Various anisotropic and frustrated interactions between spin tetrahedra, even if relatively weak, lead to interesting and non-trivial ground states and quantum phase transitions. An acceptable physical realization of the spin-tetrahedra model was recently found in compounds of the type Cu$_2$Te$_2$O$_5X_2$ with $X$=Cl or Br [@Johnsson00; @Lemmens01], for which the tetragonal crystal structure of Cu$_2$Te$_2$O$_5X_2$ is formed by distorted tetrahedra of Cu$^{2+}$ ions aligned along the $c$-axis [@Johnsson00]. Two types of AFM bonds within the tetrahedra are associated with exchange integrals $J_1$ and $J_2$. The magnetic susceptibility $\chi(T)$ exhibits a peak and subsequently decreases exponentially with decreasing $T$ at low temperatures. This suggests that the spins are dimerized and the corresponding energy gap $\Delta/k_B$, separating the ground state from excited states, is about 40 K [@Johnsson00; @Lemmens01]. The analysis of the susceptibility data, assuming that $r \equiv J_2/J_1 =1$, results in $J_1/k_B=J_2/k_B= 38.5$ and 43 K for $X=$ Cl and Br, respectively [@Johnsson00; @Lemmens01]. A mean-field (MF) type analysis of results of Raman scattering measurements on the Br-compound suggests that $r=0.66$, and $J_1/k_B=47 {\rm ~K}$. The inter-tetrahedral coupling parameter $J_c = 0.85 J_1$ [@Gros03] leads to phase transitions to an AFM ordered state at $T_N = 18.2$ and 11.4 K for $X=$ Cl and Br, respectively [@Lemmens01]. The application of an external magnetic field $H$ reduces $T_N$ for Cu$_2$Te$_2$O$_5$Cl$_2$. For Cu$_2$Te$_2$O$_5$Br$_2$, however, an unusual increase of the transition temperature with increasing $H$ was observed. It was argued [@Lemmens01; @Gros03] that the latter anomalous behavior is caused by the vicinity of a quantum critical transition, which is expected if $J_c=0.75 J_1$. Several theoretical models have since been put forward to explain the magnetically ordered state and the related excitation spectrum of Cu$_2$Te$_2$O$_5$Br$_2$. The essential inputs were based on invoking, e.g., anisotropic inter-tetrahedral and Dzyaloshinsky-Moriya-type interactions [@Brenig01; @Totsuka02; @Jensen03; @Kotov04cm], but the nature of the low-temperature phase still remains largely unexplained. Very recent neutron diffraction experiments [@Zaharko04cm] were interpreted as to indicate the formation of an incommensurate long-range magnetic order for both $X$=Cl and Br.
Results on the temperature dependences of the thermal conductivity $\kappa(T)$ of Cu$_2$Te$_2$O$_5X_2$ for both $X$=Br and $X$=Cl in zero magnetic field were recently reported in Ref. [@Prester04]. The authors observed a strong anomaly in $\kappa(T)$ near $T_N$ for Cu$_2$Te$_2$O$_5$Cl$_2$ and attributed it to an unexpectedly large spin-lattice coupling in this compound. In contrast, no anomaly was observed for the isomorphic Cu$_2$Te$_2$O$_5$Br$_2$, which was ascribed to an intrinsically weak spin-lattice coupling in this material. The present work includes measurements of the low-temperature thermal conductivity $\kappa(T,H)$ of Cu$_2$Te$_2$O$_5$Br$_2$ with a special emphasis on investigating the influence of external magnetic fields $H$ with different orientations. In the vicinity of the ordering transition, we observe a pronounced anomaly of $\kappa(T)$, which is remarkably sensitive to the strength and the orientation of the external magnetic field. The main result is the observation of an anomalous increase of $T_N$ with increasing $H$, but only if the field orientation is perpendicular to the $c$-axis, at least up to 6 T. The implications of this observation are discussed in view of recent theoretical suggestions for the cause of the magnetic ordering transition in Cu$_2$Te$_2$O$_5$Br$_2$.
Samples and experiment
======================
The samples for this investigation were cut from a large single crystal of Cu$_2$Te$_2$O$_5$Br$_2$, grown as described in Ref. [@Johnsson00]. Two bar-shaped samples with approximate dimensions of $0.5\times 0.5\times 2$ mm$^{3}$ were cut in such a manner that the longest direction was, for one sample, along the $c$-direction, and perpendicular to the $c$-direction for the other specimen. The thermal conductivity was measured in the temperature region between 2 and 300 K by using the standard method of uniaxial heat flow as described in Ref. [@Sologubenko03_Dif]. The magnetic fields were oriented along either the $c$- or the $a$-axis of the crystal structure. Complementary measurements of the magnetic susceptibility were made with a commercial SQUID magnetometer at temperatures between 2 and 300 K.
Results
=======
The temperature dependences of the thermal conductivities, $\kappa(T)$, along two crystallographic orientations in zero magnetic field and in $H=60$ kOe are shown in Fig. \[KvsT\_H0\].
=1
The general features of $\kappa(T)$ along the two heat flux directions are essentially the same for $H=0$, especially above approximately 7 K, where the data for the two samples differ by practically a constant factor, such that $\kappa_{\parallel c}/\kappa_{\perp c} \approx 1.6$. At lower temperatures, this ratio is gradually reduced to about 0.6 at 3 K. Each $\kappa(T)$ curve in Fig. \[KvsT\_H0\] exhibits a maximum between 3.5 and 4.5 K and a distinct feature around $T_{N}=11.4 {\rm ~K}$. This type of low-temperature maximum of $\kappa(T)$ is typical for insulators and, with increasing temperature, reflects the gradual change from the dominant boundary scattering to enhanced phonon-phonon scattering of the itinerant lattice excitations. A sharp feature of $\kappa(T)$ is usually related to some kind of phase transition, in the present case to magnetic ordering. Applying an external magnetic field well above $T_{N}$ leads to only a slight decrease of $\kappa$, almost independent of the field orientation. However, a significant and $H$-orientation dependent reduction of the thermal conductivity by magnetic field is observed in the vicinity and below $T_N$ (see the inset in Fig. \[KvsT\_H0\]).
In view of the following discussion, we concentrate on the anomalous features of $\kappa(T)$ in the vicinity of $T_N$. In Fig. \[KcTn\] a and b, we display the data for $\kappa(T)$ along the $c$-axis at temperatures between 10 and 13 K and for different values and orientations of the magnetic field. In order to emphasize the change of the slope of $\kappa(T)$ at the transition, we show the corresponding temperature derivatives $\partial\kappa /\partial T$ in Fig. \[KcTn\] c and d.
=1
The qualitative difference in the behavior for the two field orientations is obvious. While the transition, reflected in the sudden drop of $\partial\kappa /\partial
T$ vs $T$ with decreasing $T$ occurs at the same temperature $T_{N}=11.4 {\rm ~K}$ for $H\parallel c \leq 60 {\rm kOe}$, the drop of $\partial\kappa /\partial T$ vs. $T$ for $H \perp c \geq 20 {\rm kOe}$ is preceded by an initial increase, thus forming a narrow peak. With increasing $H$, the peak shifts to higher temperatures, obviously reflecting the anomalous $T_{N}(H)$-enhancement reported in the literature [@Lemmens01]. The absence of any variation of $T_N$ for $H\parallel c$ has not been claimed before.
Our data clearly demonstrate a feature in $\kappa(T)$ at $T_N$, at variance with the results of Ref. [@Prester04], where similar effects were observed for the Cl-compound only. We suspect that the absence of an anomaly in $\kappa(T)$ at $T_N$ for the Br-compound in the data of Ref. [@Prester04] may be due to a dominating influence of defects in that Cu$_2$Te$_2$O$_5$Br$_2$ sample, masking the influence of the intrinsic scattering mechanisms that are related to the magnetic ordering. Because of essentially the same features of $\kappa(T)$ at $T_N$ for the two compounds, we question the reasoning in Ref. [@Prester04] which suggests a drastic and intrinsic difference in the spin-phonon coupling between the two compounds. The same crystal structure and the only slightly different size of the unit cell make the conjecture of Ref. [@Prester04] rather unlikely. Without presenting the taken data for the magnetic susceptibility $\chi(T)$ in the temperature region between 2 and 300 K, we note that they exhibit all the characteristic features reported for $\chi(T)$ of the Cu$_2$Te$_2$O$_5$Br$_2$ single-crystal in Ref. [@Prester04]. These features include a maximum of $\chi(T)$ at about 30 K, slightly higher $\chi$ values for $H\parallel c$ than $H \perp c$, the saturation to constant values below about 5 K, and also the field-dependence of $\chi$ at low temperatures reported in Ref. [@Gros03].
Discussion
==========
In magnetic insulators, energy may be transported in both the crystal lattice and the spin system. In those cases where the approach of invoking excitations or quasiparticles is applicable, the total thermal conductivity can be represented as the sum of a phonon contribution $\kappa_{\rm ph}$ and a contribution of spin excitations (magnons, spinons etc.) $\kappa_s$. Each contribution $\kappa_i \propto C_i v_i \ell_i$ is given by the specific heat $C_i$ of the corresponding subsystem, the velocity of the related quasiparticles $v_i$ and their mean free path $\ell_i$. For each type of quasiparticle, the relaxation rate $\tau_i=\ell_i/v_i$ depends on the mutual interaction of the quasiparticles and the influence on their motion by various imperfections, such as point defects, dislocations, grain and domain boundaries etc. In many cases, $\tau_i^{-1}= \sum_j \tau_{i,j}^{-1}$, where $j$ corresponds to a particular type of scatterers. Among the various scattering processes, the spin-phonon interaction is of paramount importance in magnetic materials. Apart from influencing the magnitude and shape of $\kappa(T,H)$, the spin-phonon interaction provides, in a standard experimental arrangement of thermal-conductivity measurements, the necessary channel of heat transfer from the lattice to the spin system [@Sanders77].
Since the magnon band in Cu$_2$Te$_2$O$_5$Br$_2$ is separated from the ground state by an energy gap $\Delta/k_B \approx 40 {\rm ~K}$, $\kappa_s(T)$ is expected to be negligibly small at $T \ll \Delta$ and to increase exponentially at $T \leq \Delta$. Any anisotropy of the spin interaction is expected to lead to an anisotropic increase in $\kappa_s(T)$. This expectation was confirmed in previous investigations of a number of quasi-1D and -2D magnetic systems, where pronounced changes from a weakly temperature-dependent anisotropy (of phononic origin) at low temperatures to a strongly $T$-dependent anisotropy at higher temperatures were observed and interpreted as evidence for the onset of $\kappa_s$ (see discussion in Ref. [@Sologubenko01] and references therein). In Cu$_2$Te$_2$O$_5$Br$_2$, no change in the ratio $\kappa_a/\kappa_c$ is observed above the ordering transition (see Fig. \[KvsT\_H0\]), in spite of the anisotropy of the magnon spectrum along these directions [@Brenig03]. This suggest that the spin contribution in Cu$_2$Te$_2$O$_5$Br$_2$ is negligibly small in comparison with the phonon contribution.
The field-induced reduction of phonon transport is most pronounced and most sensitive to $H$ in the vicinity of the ordering transition. In magnetic materials, phonon-magnon scattering occurs via single-ion-lattice and magnetostrictive interactions [@Kawasaki63; @Stern65]. The intensity of the latter type of scattering is proportional to the magnetic specific heat $C_s(T)$ which exhibits an anomaly at $T_N$. If near the transition temperature, $C_s(T)$ is dominated by a discontinuity, usually leading to a peak-shaped anomaly, and if the phonons are predominantly scattered by the spin excitations, then a sharp dip in $\kappa(T)$ or, equivalently, a discontinuity in $\partial\kappa /\partial T$ vs $T$ is expected at $T_N$ [@Kawasaki63]. This is rarely the case for real materials where various $T$-dependent phonon-scattering mechanisms involving defects, boundaries, and the phonon-phonon interaction are stronger or at least of similar strength as the phonon-magnon scattering. If the anomaly in $C_s(T)$ is broadened for some reason, $\kappa(T)$ exhibits a broadly distributed reduction rather than the sharp dip mentioned above.
Specific heat data by Lemmens [*et al.*]{} [@Lemmens01], taken on a powder sample of Cu$_2$Te$_2$O$_5$Br$_2$, reveal, upon the application of an external magnetic field, the growth and shift of the broad peak of $C(T)$ at $T_N$ to higher temperatures. This correlates with the enhanced reduction of $\kappa(T)$ and the shift of the related anomaly to higher temperatures, which we observe in our experiments for $H \parallel a$, see Fig. \[KcTn\] (a,c). It is to be noted, however, that no such effect is observed for $H \parallel c$. The transition temperature $T_N(H)$, calculated from the $\partial\kappa /\partial T$ vs $T$ data, is shown in Fig. \[Tn\].
=0.9
As mentioned, $T_N$ increases with $H\parallel a$, in agreement with the quoted earlier observations [@Lemmens01]; the critical temperature does not change in the investigated field region for $H \parallel c$, however.
An increasing $T_N$ with increasing field cannot simply be explained by the classical theory of antiferromagnetism. Nonetheless, such behavior has been observed for several AFM compounds [@Oliveira79; @Clark75; @Hijmans78; @Honda98; @Honda01]. It was argued that this behavior is related to the fact that, generally, a lower spin dimensionality leads to higher values of the critical field. A magnetic field oriented along particular crystallographic directions may lead to a reduction of the effective spin dimensionality [@Fisher74; @Fisher75]. However, if the magnetic field is oriented in such a way that it does not change the spin dimensionality, such as along the easy axis of an Ising model system or perpendicular to the easy plane of a planar model, an MF theory-consistent reduction of $T_N$ with increasing field is expected. For 3D AFM spin systems, the dimensionality-driven enhancement of $T_N$ is, at most, of the order of 0.1% [@Oliveira79], but quantum effects in spin compounds containing structural elements with lower dimensionality, particularly in spin-chain compounds [@imry75], lead to much stronger field-induced changes of $T_N$ [@Clark75; @Hijmans78; @Honda98; @Honda01]. If the same type of arguments is valid for the anisotropic shifts of $T_N(H)$ observed in this work for tetragonal Cu$_2$Te$_2$O$_5$Br$_2$, it may be concluded that the magnetic order is characterized by an easy-axis moment orientation along the $c$-axis.
This conclusion is in agreement with the results of calculations of Jensen [*et al.*]{} [@Jensen03], which include a Dzyaloshinskii-Moriya type anisotropy term in the Hamiltonian describing a model of dimerized interacting tetrahedra [@Gros03]. In relation with the Br-compound, the model calculations predict an AFM ordered state below $T_N$ with staggered moments aligned along the $c$-axis. The calculations also predict $T_N$ to increase with $H \perp c$ for all values of $H$, consistent with our result, but also a weak initial decrease of $T_N(H)$ for $H \parallel c$, intercepted by a spin-flop transition at about 37 kOe. The latter transition is not reflected in our data, but the calculated value of the spin-flop field is parameter-dependent. It is certainly not inconceivable that in reality, this field is higher than the calculated value of 37 kOe. An increase of $T_N$ with the mentioned field configuration is also consistent with another analysis of a model of coupled spin tetrahedra taking into account a Dzyaloshinskii-Moriya-type interaction [@Kotov04cm]. These calculations also predict a decrease of $T_N(H)$ at almost the same rate, if a field is applied in the $c$-direction. This expectation is not supported by our experiments, however.
Summary
=======
In this work the thermal conductivity of the spin-tetrahedral compound Cu$_2$Te$_2$O$_5$Br$_2$ has been studied. The results clearly indicate that phonons dominate the heat transport in this compound. A feature in $\kappa(T)$ at $T_N$, close to 11.4 K, is associated with a magnetic ordering transition. The transition temperature and the amplitude of the associated $\kappa(T)$ anomaly are affected by external magnetic fields only if they are oriented along the $a$-axis. This $T_N(H)$ anisotropy is qualitatively consistent with recent theoretical predictions.
This work was financially supported in part by the Schweizerische Nationalfonds zur Förderung der Wissenschaftlichen Forschung.
[10]{}
H. Castella, X. Zotos, and P. Prelovšek, Phys. Rev. Lett. [**74**]{}, 972 (1995).
X. Zotos, F. Naef, and P. Prelovšek, Phys. Rev. B [**55**]{}, 11029 (1997).
K. Saito, S. Takesue, and S. Miyashita, Phys. Rev. E [**54**]{}, 2404 (1996).
B. N. Narozhny, A. J. Millis, and N. Andrei, Phys. Rev. B [**58**]{}, R2921 (1998).
F. Naef and X. Zotos, J. Phys.: Condens. Matter [**10**]{}, L183 (1998).
A. Klümper and K. Sakai, J. Phys. A [**35**]{}, 2173 (2002).
J. V. Alvarez and C. Gros, Phys. Rev. Lett. [**89**]{}, 156603 (2002).
F. Heidrich-Meisner, A. Honecker, D. C. Cabra, and W. Brenig, Phys. Rev. B [ **66**]{}, 140406(R) (2002).
K. Saito, Phys. Rev. B [**67**]{}, 064410 (2002).
E. Shimshoni, N. Andrei, and A. Rosch, Phys. Rev. B [**68**]{}, 104401 (2003).
E. Orignac, R. Chitra, and R. Citro, Phys. Rev. B [**67**]{}, 134426 (2003).
Y. Nakamura [*et al.*]{}, Physica C [**2**]{}, 1409 (1991).
J. L. Cohn, C. K. Lowe-Ma, and T. A. Vanderah, Phys. Rev. B [**52**]{}, R13134 (1995).
C. Hess [*et al.*]{}, Phys. Rev. Lett. [**90**]{}, 197002 (2003).
M. Hofmann [*et al.*]{}, Phys. Rev. B [**67**]{}, 184502 (2003).
X. F. Sun, J. Takeya, S. Komiya, and Y. Ando, Phys. Rev. B [**67**]{}, 104503 (2003).
B. C. Sales [*et al.*]{}, Phys. Rev. Lett. [**88**]{}, 095901 (2002).
A. N. Vasil’ev, M. M. Markina, A. V. Inyushkin, and H. Kageyama, JETP Lett. [**73**]{}, 633 (2001).
M. Hofmann [*et al.*]{}, Phys. Rev. Lett. [**87**]{}, 047202 (2001).
H. Kudo [*et al.*]{}, J. Phys. Soc. Jpn. [**70**]{}, 1448 (2001).
A. V. Sologubenko [*et al.*]{}, Phys. Rev. Lett. [**84**]{}, 2714 (2000).
C. Hess [*et al.*]{}, Phys. Rev. B [**64**]{}, 184305 (2001).
A. V. Sologubenko [*et al.*]{}, Phys. Rev. B [**64**]{}, 054412 (2001).
A. V. Sologubenko, H. R. Ott, G. Dhalenne, and A. Revcolevschi, Europhys. Lett. [**62**]{}, 540 (2003).
A. V. Sologubenko [*et al.*]{}, Phys. Rev. B [**68**]{}, 94432 (2003).
M. Markina [*et al.*]{}, J. Mag. Mag. Mater. [**259**]{}, 398 (2003).
A. M. Vasil’ev [*et al.*]{}, JETP Lett. [**66**]{}, 868 (1997).
Y. Ando [*et al.*]{}, Phys. Rev. B [**58**]{}, R2913 (1998).
B. Salce [*et al.*]{}, Phys. Lett. A [**245**]{}, 127 (1998).
A. N. Vasil’ev [*et al.*]{}, Phys. Rev. Lett. [**81**]{}, 1949 (1998).
J. Takeya [*et al.*]{}, Phys. Rev. B [**61**]{}, 14700 (2000).
J. Takeya [*et al.*]{}, Phys. Rev. B [**62**]{}, R9260 (2000).
M. Hofmann [*et al.*]{}, Physica B [**312-313**]{}, 597 (2002).
J. Takeya [*et al.*]{}, Phys. Rev. B [**63**]{}, 214407 (2001).
M. Johnsson, K. W. T[ö]{}rnroos, F. Mila, and P. Millet, Chem. Mater. [**12**]{}, 2853 (2000).
P. Lemmens [*et al.*]{}, Phys. Rev. Lett. [**87**]{}, 227201 (2001).
C. Gros [*et al.*]{}, Phys. Rev. B [**67**]{}, 174405 (2003).
W. Brenig and K. W. Becker, Phys. Rev. B [**64**]{}, 214413 (2001).
K. Totsuka and H. J. Mikeska, Phys. Rev. B [**66**]{}, 054435 (2002).
J. Jensen, P. Lemmens, and C. Gros, Europhys. Letters [**64**]{}, 689 (2003).
V. N. Kotov, M. E. Zhitomirsky, M. Elhajal, and F. Mila, cond-mat/0404674 (unpublished).
O. Zaharko [*et al.*]{}, cond-mat/0405513 (unpublished).
M. Prester [*et al.*]{}, Phys. Rev. B [**69**]{}, 180401(R) (2004).
D. J. Sanders and D. Walton, Phys. Rev. B [**15**]{}, 1489 (1977).
W. Brenig, Phys. Rev. B [**67**]{}, 64402 (2003).
K. Kawasaki, Progr. Theor. Phys. [**29**]{}, 801 (1963).
H. Stern, J. Phys. Chem. Solids [**26**]{}, 153 (1965).
J. Oliveira, N. F. and Y. Shapira, J. Appl. Phys. [**50**]{}, 1790 (1979).
W. G. Clark, L. J. Azevedo, and E. O. McLean, in [*Proceedings of the 14th International Conference on Low Temperature Physics*]{} (North-Holland, Amsterdam, Netherlands, 1975), pp. 369–71.
J. P. A. M. Hijmans, K. Kopinga, F. Boersma, and W. J. M. de Jonge, Phys. Rev. Lett. [**40**]{}, 1108 (1978).
Z. Honda, H. Asakawa, and K. Katsumata, Phys. Rev. Lett. [**81**]{}, 2566 (1998).
Z. Honda, K. Katsumata, Y. Nishiyama, and I. Harada, Phys. Rev. B [**63**]{}, 064420 (2001).
M. E. Fisher and D. R. Nelson, Phys. Rev. Lett. [**32**]{}, 1350 (1974).
M. E. Fisher, Phys. Rev. Lett. [**34**]{}, 1634 (1975).
Y. Imry, P. Pincus, and D. Scalapino, Phys. Rev. B [**12**]{}, 1978 (1975).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have developed a general model for determining density-dependent effective dielectronic recombination (DR) rate coefficients in order to explore finite-density effects on the ionization balance of plasmas. Our model consists of multiplying by a suppression factor those highly-accurate total zero-density DR rate coefficients which have been produced from state-of-the-art theoretical calculations and which have been benchmarked by experiment. The suppression factor is based-upon earlier detailed collision-radiative calculations which were made for a wide range of ions at various densities and temperatures, but used a simplified treatment of DR. A general suppression formula is then developed as a function of isoelectronic sequence, charge, density, and temperature. These density-dependent effective DR rate coefficients are then used in the plasma simulation code Cloudy to compute ionization balance curves for both collisionally ionized and photoionized plasmas at very low ($n_{\rm{e}}=1~\rm{cm}^{-3}$) and finite ($n_{\rm{e}}=10^{10}~\rm{cm}^{-3}$) densities. We find that the denser case is significantly more ionized due to suppression of DR, warranting further studies of density effects on DR by detailed collisional-radiative calculations which utilize state-of-the-art partial DR rate coefficients. This is expected to impact the predictions of the ionization balance in denser cosmic gases such as those found in nova and supernova shells, accretion disks, and the broad emission line regions in active galactic nuclei.'
author:
- 'D. Nikolić, T. W. Gorczyca, K. T. Korista'
- 'G. J. Ferland'
- 'N. R. Badnell'
title: Suppression of Dielectronic Recombination Due to Finite Density Effects
---
Introduction {#intro}
============
Astronomical emission or absorption sources have an enormous range of densities. Two examples include the intergalactic medium, with $n_{\rm{e}} \sim 10^{-4}\ {\rm cm}^{-3}$, and the broad emission-line regions of Active Galactic Nuclei, with $n_{\rm{e}} \sim 10^{10}\ {\rm cm}^{-3}$. The gas producing the spectrum is not in thermodynamic equilibrium [@osterbrockferland], so microphysical processes determine the physical conditions.
The two common cases encountered for ionization are photoionization and collisional (e.g., electron-impact) ionization. In both cases, ions are recombined by dielectronic and radiative recombination, with dielectronic recombination (DR) usually the dominant process for elements heavier than helium. Databases give ionization and recombination rates that are the sum of several contributing processes. Examples include @voronov for electron impact ionization, [@verner] for photoionization, and the DR project [@drproject] for dielectronic recombination and [@rrproject] for radiative recombination; it is these latter data [^1] which will be of primary interest to us in the present study.
The collisional ionization and recombination rate coefficients used in astrophysics are frequently assumed to depend on temperature but to have no density dependence. The rigorous treatment of density dependent ionization and recombination rate coefficients is via collisional-radiative modeling. This was introduced by @bates for radiative recombination only and extended to treat the much more complex case of dielectronic recombination by @burgsum. Summers applied their techniques to determine density dependent ionization and recombination rate coefficients, and the consequential ionization balance for collisional plasmas, for H-like thru Ar-like ions. Graphical results were presented for the elements C, O and Ne [@summers1] and then N, Mg and Si [@summers2]. Reduced temperatures and densities were used so as to enable easy interpolation for other elements. Tables of such recombination rate coefficients were made available only via a Laboratory Report — @summersRAL — due to their voluminous nature at that point in history. The ‘difficulty’ in utilizing this pioneering data led to some modelers attempting to develop simplified approaches. For example, @jordan used an approach which was based on truncating the zero-density DR sum over Rydberg states using a simple density dependent cut-off which itself was based on early collisional-radiative calculations by @burgsum; a suppression factor was formed from its ratio to the zero-density value and then used more generally. Also, @davidson simplified the collisional-radiative approach of @burgsum and, using hydrogenic atomic data, determined suppression factors for Li-like and . New calculations for were made by @badnell utilizing more advanced (generalized) collisional-radiative modeling [@sumhoop] and much improved atomic data at collisional plasma temperatures (see the references in @badnell).
All of the above works were for electron collisional plasmas and used rather basic DR data (excluding @badnell) as epitomized in the @burgess General Formula, viz. a common dipole transition for dielectronic capture, autoionization, and radiative stabilization. The purpose of the present paper is to explore density suppression of DR in photoionized plasmas, and within collisional plasmas, using state-of-the art DR data which takes account of a myriad of pathways not feasible in the early works above, but which has been shown to be necessary by comparison with experiment. We wish to gain a broad overview utilizing the large test-suite maintained by the plasma simulation code Cloudy. We utilize an approach to DR suppression which is motivated initially by the detailed collisional-radiative results given in @badnell for at $T=10^5$ K, along with known scalings to all temperatures, charges, and densities. Using these results as a guideline, a more general suppression formula is then determined by fitting to suppression results from extensive detailed collisional-radiative calculations [@summersRAL] for a wide range of ions at several densities and (high) temperatures. Additional modifications are then introduced to account for low temperature DR.
The outline of the rest of the paper is as follows: in the next section we describe the DR suppression model we use; we then apply this suppression to the zero-density DR data, and use the resultant density-dependent DR data in Cloudy to determine the ionization distribution produced under photoionized and collisional ionization equilibrium at low and moderate densities.
Generalized Density Suppression Model {#Sec:Suppress}
=====================================
We use the following approach, detailed more fully in the subsections below. First, the high-temperature collisional-radiative modeling results of @badnell for DR suppression in are parameterized by a pseudo-Voigt profile to study the qualitative behavior of suppression as a function of density and temperature. Next, this formulation is then used as a guideline for developing a more comprehensive suppression formula which is obtained by fitting to collisional radiative data for various isoelectronic sequences, ionic charges, densities, and temperatures [@summersRAL]. Lastly, the suppression formulation is extended to low-temperatures according to the nature of the sequence-specific DR.
High-Temperature Suppression for Li-like
-----------------------------------------
We begin by considering DR of Li-like , for which the density dependent total DR rate coefficient, and therefore the suppression factor, has been computed rigorously within a collisional-radiative modeling approach [@badnell].
In the electron collisional ionization case, because of the consequential high temperature of peak abundance, dielectronic recombination occurs mainly through energetically high-lying autoionizing states (via dipole core-excitations) for which radiative stabilization is by the core electron into final states just below the ionization limit: $$\begin{aligned}
e^- + 1s^22s \rightarrow &
1s^22pnl \rightarrow & 1s^22snl + h\nu\ .
\label{eq1}\end{aligned}$$ In the zero-density limit, the intermediate $1s^22snl$ states can only decay further via radiative cascading until the $1s^22s^2$ final recombined ground state is reached, thereby completing the DR process: $$\begin{aligned}
1s^22snl \longrightarrow & 1s^22sn^\prime l^\prime+ h\nu_1 \rightarrow ... \rightarrow & 1s^22s^2 + h\nu_1+h\nu_2+... \label{eqradstab}\end{aligned}$$ For finite electron densities $n_{\rm{e}}$, on the other hand, there is also the possibility for reionization via electron impact, either directly or stepwise, $$\begin{aligned}
e^- + 1s^22snl \longrightarrow & 1s^2 2sn^\prime l^\prime+ e^- \rightarrow ... \rightarrow & 1s^22s + e^- + e^- \ ,\label{eqreion}\end{aligned}$$ and the probability of the latter pathway is proportional to the electron density $n_{\rm{e}}$. Because of this alternative reionization pathway at finite densities, the [*effective*]{} DR rate coefficient $\alpha^{\rm{eff}}_{\rm{DR}}(n_{\rm{e}},T)$ is thus [*suppressed*]{} from the zero-density value $\alpha_{\rm{DR}}(T)$ by a density-dependent suppression factor $S(n_{\rm{e}},T)$: $$\begin{aligned}
\alpha^{\rm{eff}}_{\rm{DR}}(n_{\rm{e}},T) & \equiv & S(n_{\rm{e}},T)\alpha_{\rm{DR}}(T)\ .
\label{eqsuppress}\end{aligned}$$
From the earlier detailed studies of @davidson and @badnell, the suppression factor is found to remain unity, corresponding to zero suppression, at lower densities until a certain [*activation*]{} density $n_{\rm{e},\rm{a}}$ is reached, beyond which this factor decreases exponentially from unity with increasing density. We have found that this suppression factor, as a function of the dimensionless log density parameter $x=\log_{10}n_{\rm{e}}$, can be modeled quite effectively by a pseudo-Voigt profile [@Wertheim:1974] — a weighted mixture $\mu$ of Lorentzian and Gaussian profiles of widths $w$ for densities above the activation density $x_{\rm{a}}=\log_{10}n_{\rm{e},\rm{a}}$: $$S(x;x_{\rm{a}}) = \left\{
\begin{matrix}
1 & x \le x_{\rm{a}} \\
\mu \left[\frac{1}{1+(\frac{x-x_{\rm{a}}}{w})^2}\right]
+ (1 - \mu)\left[{\rm e}^{-(\frac{x-x_{\rm{a}}}{w/\sqrt{\ln 2}})^{2}}\right] & x \geq x_{\rm{a}}
\end{matrix}\right. \ .
\label{eqsuppression}$$ Fitting this expression to the suppression factor of @badnell for (which was computed at $T=10^5$ K) yielded the values $\mu=0.372$, $w=4.969$, and $x_{\rm{a}}=0.608$, and this parameterization formula is found to be accurate to within 5% for all densities considered (see Fig. \[fig1\]).
![\[fig1\] Pseudo-Voigt fit of the suppression factor for , as given in Eq. \[eqsuppression\] with a scaled activation density as given by Eq. \[eqxa\], shown for two different temperatures. The red solid curve shows that the parameterization for $T=1\times10^{5}$ K, corresponding to an activation density of $x_{\rm{a}}=0.608$ (with $\mu=0.372$ and $w=4.969$), is in close agreement with the actual data of [@badnell] (open circles). The blue dashed curve is the parameterization for $T=1.5\times10^{4}$ K, using instead an activation density of $x_{\rm{a}}=0.196$ (and the same $\mu$ and $w$), and giving satisfactory agreement with the data of [@davidson] (solid circles). ](Fig1.eps){width="\textwidth"}
Generalized High-Temperature Suppression Formula {#BurgessHT}
------------------------------------------------
Given the suppression formula for Li-like , corresponding to ionic charge $q_0=3$ and temperature $T_0=10^5$ K, we wish to generalize this expression to other Li-like ions of charge $q$ and (high) $T$ according to the following qualitative guidelines. It is well known that density effects scale as $q^7$ — see @bates and @burgsum. The activation density is attained when the reionization rate in Eq. \[eqreion\], which depends linearly on the density, becomes comparable to the radiative stabilization rate in Eq. \[eqradstab\]. The radiative rate is independent of density and temperature, but scales with charge as $A_r\sim q^4$, whereas the electron-impact ionization rate depends on all three, viz. $n_{\rm{e}}\alpha_{\rm{eII}}\sim n_{\rm{e}} q^{-3}T^{-1/2}$. An initial suggestion is that the activation density is attained when these two are approximately equal, i.e., $$\begin{aligned}
n_{\rm{e},\rm{a}}q^{-3}T^{-1/2}\sim q^4\ ,\end{aligned}$$ indicating that the activation density should scale as $ n_{\rm{e},\rm{a}}\sim q^7T^{1/2}$, if the above qualitative discussion holds. The log activation density for all $q$ and $T$ might therefore be expected to obey the scaling relationship $$\begin{aligned}
x_{\rm{a}}(q,T) & = & x_{\rm{a}}(q_0,T_0) + log_{10}\left[\left(\frac{q}{q_0}\right)^7\left(\frac{T}{T_0}\right)^{1/2}\right]\ , \label{eqxa}\end{aligned}$$ where $x_{\rm{a}}(q_0,T_0)=0.608$, $q_0=3$, and $T_0=10^5$ K are the (log) activation density, the charge, and the temperature for the case treated by [@badnell]. We note that this expression, when applied to Li-like , gives an increase in the activation density by a factor of $(5/3)^7 = 35.7$, in agreement with the approximate factor of 40 found by @davidson. Furthermore, when scaled in temperature, the formula gives fairly good agreement with the suppression results of @davidson for at $T=1.5\times 10^4$ K (see Fig. \[fig1\]).
### Fit to the Collisional Radiative Data {#cr}
The preceding treatment reasonably extends the suppression factor at $10^5$ K to other high temperatures and to other Li-like ions. However, we need suppression factors applicable to all ionization stages of all elements up to at least Zn for a general implementation within Cloudy. Unfortunately, detailed collisional-radiative modeling data with state-of-the-art DR data is still rather limited. However, extensive tables of effective recombination rate coefficients have been computed by [@summersRAL] for a wide variety of isoelectronic sequences, charge-states, temperatures, and densities. The treatment of DR there is somewhat simplified, but we only require the [*ratio*]{} of finite- to zero-density rate coefficients to determine the suppression factor. We then combine this ratio with our state-of-the-art zero density DR rate coefficients again for use within Cloudy. This ratio is much less sensitive to the specific treatment of DR.
The rather simplistic scaling formula in Eq. \[eqxa\] was found to be inadequate when applied to the extensive tabulation of suppression factors found in [@summersRAL]. Instead, in order to fit the data accurately, a more generalized formula was arrived at, where a pseudo-Gaussian, corresponding to $\mu=0$ in Eq. \[eqsuppression\], was more appropriate, $$S^N(x;q,T) = \left\{
\begin{matrix}
1 & x \le x_a(q,T,N)\\
{\rm e}^{-(\frac{x-x_a(q,T,N)}{w/\sqrt{\ln 2}})^{2}} & x \geq x_a(q,T,N)
\end{matrix}\right. \ .
\label{eqsuppression2}$$ Furthermore, the activation density was found to be best represented by the function $$\begin{aligned}
x_a(q,T,N) & = & x_{a}^{0} + log_{10}\left[\left(\frac{q}{q_0(q,N)}\right)^{7}\left(\frac{T}{T_0(q,N)}\right)^{1/2}\right]\ , \label{eqxanew}\end{aligned}$$ where the variables $q_0(q,N)$ and $T_0(q,N)$ are taken to be functions of the charge $q$ and the isoelectronic sequence, labeled by $N$. A fit of the suppression factors of @summersRAL for all ions yielded a global (log) activation density $x_{a}^{0}=10.1821$ and more complicated expressions for the zero-point temperature $T_0$ and charge $q_0$. These were found to depend on both the ionic charge $q$ and the isoelectronic sequence $N$ viz. $$\begin{aligned}
T_0(q,N) &= &5\times10^{4}\,[q_0(q,N)]^{2} \label{t0}\end{aligned}$$ and $$\begin{aligned}
q_0(q,N) & = & (1 - \sqrt{2/3q})A(N)/\sqrt{q}\ , \label{q0}\end{aligned}$$ where $$\begin{aligned}
A(N) & = & 12 + 10N_{1} + \frac{10N_{1}-2N_{2}}{N_{1}-N_{2}}(N-N_{1}) \label{AN}\end{aligned}$$ depends on the isoelectronic sequence in the periodic table according to the specification of the parameters $$\begin{aligned}
(N_{1},N_{2}) & = &
\begin{pmatrix}
(3,10) & N\in 2^{nd} \, \rm{row} & & (37,54) & N\in 5^{th} \, \rm{row} \\
(11,18) & N\in 3^{rd} \, \rm{row} & & (55,86) & N\in 6^{th} \, \rm{row} \\
(19,36) & N\in 4^{th} \, \rm{row} & & (87,118) & N\in 7^{th} \, \rm{row} \\
\end{pmatrix}\ .
\label{N1N2}\end{aligned}$$ However, even this rather complicated parameterization was inadequate for the lower isoelectronic sequences $N\le 5$, and for these we explicitly list the optimal values for $A(N)$ in Table \[TableXX\]. Furthermore, at electron temperatures and/or ionic charges for which the $q$-scaled temperature $\theta \equiv T/q^2$ was very low ($\theta\leq 2.5\times10^{4}$ K), a further modification to the coefficients $A(N)$ for $N\le 5$ is necessary in that the values in Table \[TableXX\] should be multiplied by a factor of two.
[rcc]{} H-like & 1 & 16\
He-like & 2 & 18\
Li-like & 3 & 66\
Be-like & 4 & 66\
B-like & 5 & 52\
\
The above final formulation, which consists of the use of Eq. \[eqsuppression2\], with $\mu=0$, $w=5.64548$, and a rather complicated activation density given by Eqs. \[eqxanew\], \[t0\], \[q0\], \[AN\], and \[N1N2\], with $x_{a}^{0}=10.1821$, has been found to model the entire database of ions, temperatures, and densities considered in the @summersRAL data fairly well. To illustrate the general level of agreement over a large range of ions and environments, we compare our parameterized model formulation to the actual suppression data from that report [@summersRAL] for a few selected cases in Fig. \[figsurvey\].
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ A comparison between the present parameterized suppression factor and the collisional radiative results of @summersRAL for a sample of ions and temperatures, as a function of density. \[figsurvey\] ](Fig2a.eps "fig:"){width="3.2in"} ![ A comparison between the present parameterized suppression factor and the collisional radiative results of @summersRAL for a sample of ions and temperatures, as a function of density. \[figsurvey\] ](Fig2b.eps "fig:"){width="3.2in"}
![ A comparison between the present parameterized suppression factor and the collisional radiative results of @summersRAL for a sample of ions and temperatures, as a function of density. \[figsurvey\] ](Fig2c.eps "fig:"){width="3.2in"} ![ A comparison between the present parameterized suppression factor and the collisional radiative results of @summersRAL for a sample of ions and temperatures, as a function of density. \[figsurvey\] ](Fig2d.eps "fig:"){width="3.2in"}
![ A comparison between the present parameterized suppression factor and the collisional radiative results of @summersRAL for a sample of ions and temperatures, as a function of density. \[figsurvey\] ](Fig2e.eps "fig:"){width="3.2in"} ![ A comparison between the present parameterized suppression factor and the collisional radiative results of @summersRAL for a sample of ions and temperatures, as a function of density. \[figsurvey\] ](Fig2f.eps "fig:"){width="3.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
In order to quantify more fully the extent of agreement, we focus on the case of iron ions, for which we study density effects on ionization balance determination in the next section. A comparison is shown in Fig. \[figiron\] between our predicted suppression factors and the data from the @summersRAL report.
![\[figiron\] Agreement between the suppression curve of Eq. \[eqxanew\] and the @summersRAL data for all iron ions Fe$^{q+}$, $q=9-19$. The upper panel shows the detailed level of agreement of the two end cases, Fe$^{9+}$ and Fe$^{19+}$. The lower panel shows the $2-\sigma$ (95.4%) confidence level as a function of charge state; this means that 95.4% of all density data points in the @summersRAL data, for the given charge and temperature, are within that percentage of the prediction from Eq. \[eqsuppression2\]. The symbols denote different values of the scaled temperature $\theta=T/q^2$. ](Fig3a.eps "fig:"){width="70.00000%"} ![\[figiron\] Agreement between the suppression curve of Eq. \[eqxanew\] and the @summersRAL data for all iron ions Fe$^{q+}$, $q=9-19$. The upper panel shows the detailed level of agreement of the two end cases, Fe$^{9+}$ and Fe$^{19+}$. The lower panel shows the $2-\sigma$ (95.4%) confidence level as a function of charge state; this means that 95.4% of all density data points in the @summersRAL data, for the given charge and temperature, are within that percentage of the prediction from Eq. \[eqsuppression2\]. The symbols denote different values of the scaled temperature $\theta=T/q^2$. ](Fig3b.eps "fig:"){width="70.00000%"}
It is seen that our model fits that data to within 21% for all densities, temperatures, and ionic stages reported [@summersRAL]. More broadly, we have applied a similar $2-\sigma$ analysis to [*all*]{} ions in that report, at all temperatures and densities, and find the same agreement (20-26% confidence level).
Lastly, it is of interest to investigate how our final suppression factor in Eq. \[eqsuppression2\] compares to our original, motivating, formulation of Eq. \[eqsuppression2\] for , shown in Fig. \[figcivnew\].
![\[figcivnew\] A comparison between the final suppression factor of Eq. \[eqsuppression2\] (solid line), corresponding to a pseudo-Gaussian profile with activation density $x_a=0.8314$ ($q_0=40.284$) and width $w=5.64548$, the @summersRAL data points (solid diamonds), and the original formulation of Eq. \[eqsuppression\] (dashed line), corresponding to a pseudo-Voigt profile with activation density $x_a=0.608$, width $w=4.696$, and mixture coefficient $\mu=0.372$. The temperature $T=1\times10^{5}$ K is the same as in Fig. \[fig1\]. ](Fig4.eps){width="\textwidth"}
There is generally good qualitative agreement. However, it is seen that the original formulation, based on the @badnell results, shows a somewhat stronger suppression effect up to $x\approx 11$. This is likely due to the more accurate treatment of the partial DR data of @badnell entering the collisional-radiative modeling, although some difference due to the collisional-radiative modeling itself may also be present. This indicates that even collisional plasmas require collisional-radiative modeling with state-of-the-art DR data. The stronger suppression tails-off at $x\gtrsim 11$ as three-body recombination starts to become relevant and which, at even higher densities (not shown), causes the suppression factor to rise (since it is a ratio of effective recombination rate coefficients, i.e. includes three-body recombination.)
Suppression Formula at Low Temperatures. {#NSFSLT}
----------------------------------------
The preceding formulation was based on the suppression factor found by @summersRAL for electron collisionally ionized plasmas, i.e., at higher temperatures, where DR is dominated by high-$n$ resonances attached to a dipole-allowed core excited state. In photoionization equilibrium, however, the temperature at which a given ion forms is substantially smaller than that found in the electron collisional case. Due to the lower kinetic temperatures, DR occurs mainly through energetically low-lying autoionizing states, often via non-dipole core-excitations for which radiative stabilization is by the (outer) Rydberg electron. These states are not, in general, as susceptible to density suppression as their high-$n$ counterparts, and so it may be necessary to modify the preceding suppression formulation.
We first consider sequences with partially-occupied $p$-subshells in the ground state, which includes the B-like $2p(^2P_{1/2,3/2})$, C-like $2p^2(^3P_{0,1,2})$, O-like $2p^4(^3P_{0,1,2})$, F-like $2p^5(^2P_{3/2,1/2})$, Al-like $3p(^2P_{1/2,3/2})$, Si-like $3p^2(^3P_{0,1,2})$, S-like $3p^4(^3P_{0,1,2})$, and Cl-like $3p^5(^2P_{3/2,1/2})$ systems. For these sequences, there is fine-structure splitting in the ground state and a correspondingly small excitation energy, $\epsilon_N$, giving dielectronic capture into high principal quantum numbers (because of the Rydberg relation $q^2/n^2\le \epsilon_N$). Stabilization is via $n\rightarrow n'$ transitions and the recombined final state is built upon an excited parent. Ultimately, it is the strength of collisional coupling of this final state with the continuum which determines whether recombination or ionization prevails. As the density increases, collisional LTE extends further down the energy spectrum. It is difficult to give a general statement about the position of such final states relative to the ionization limit. So, we assume a worst case scenario, i.e., that such states are subject to suppression, and we use the preceding suppression formula. If density effects are found to be small in photoionized plasmas then this is likely sufficient. If they appear to be significant then a more detailed treatment based on collisional-radiative modeling will be needed. Thus, for these systems, we retain the same suppression formula developed above, that is, $S^N(x,q,T)=S(x,x_a(q,T))$ for $N=\left\{5,6,8,9,13,14,16,17\right\}$, and for all $q$ and $T$.
For the hydrogenic and the closed-shell He-like and Ne-like cases, on the other hand, the excitations proceed via an increase in core principal quantum number — $1s\rightarrow 2s$ or $\left\{2s,2p\right\}\rightarrow\left\{3s,3p,3d\right\}$ — giving the dominant dielectronic capture into the low-$n<10$ resonances. Even following core radiative stabilization, these low-lying states are impervious to collisional reionization for the range of densities $x\le 10$, and thus we set $S^N(x,q,T)=1.0$ for $N=\left\{1,2,10\right\}$. However, at densities $x> 10$, the @summersRAL data for these three isoelectronic sequence show suppression factors that are fit well by the usual Eq. \[eqsuppression2\], so we do not modify $S^N(x,q,T)$ for these cases.
Lastly, we consider the intermediate isoelectronic sequences for which excitation arises from neither a fine-structure splitting of the ground state nor a change in principal quantum number of the core. These include the Li-like $2s\rightarrow 2p$, Be-like $2s^2\rightarrow 2s2p$, N-like $2s^22p^3(^4S)\rightarrow 2s2p^4(^4P)$, Na-like $3s\rightarrow 3p$, Mg-like $3s^2\rightarrow 3s3p$, and P-like $3s^23p^3(^4S)\rightarrow 3s3p^4(^4P)$ cases up through the third row sequences. Any large low-temperature DR contribution arising from near threshold resonances is to low-lying states, for which suppression is negligible, i.e. the high-temperature suppression factor must be switched-off ($S^N \rightarrow 1$) at low-$T$.
To illustrate the general demarcation between low-$T$ and high-$T$ DR, we first consider DR of , an overview of which is depicted in Fig. \[fig4\]. The DR cross section, shown in the inset, is dominated by two features. The first is the $n\rightarrow \infty$ accumulation of resonances at the $\epsilon=8$ eV series limit — those which can be treated in the usual high-$T$ fashion [@burgess; @burgsum] and are therefore susceptible to suppression according to our formulation above. However, there is a second strong contribution from the lowest accessible resonances just above the threshold electron energy, which, according to the Rydberg consideration $3^2/n^2\approx \epsilon_3=0.6$ Ryd, occur here for $n=4$. More generally, these low-lying states are typical of the low-lying DR spectrum [@nussbaumer][^2]. The $1s^22p4l$ resonances decay predominantly to the $1s^22s2p$, $1s^22p^2$ and $1s^22s4l$ states. These states lie well below the ionization limit and so are not susceptible to further reionization. Since there should be no density suppression then, we seek a modified suppression factor which tends toward unity (i.e., no suppression) at lower temperatures.
![\[fig4\] DR of . The inset shows the (zero-density) DR cross section convoluted with a 0.1 eV FWHM Gaussian. The spectrum is dominated by two features: the $n=4$ DR resonance manifold below 1.0 eV and the $n\rightarrow\infty$ Rydberg resonances accumulating at the $2s\rightarrow 2p$ series limit $\epsilon_{3}(q_0)\approx 8$ eV. The main figure shows the effective DR rate coefficient for several densities. Our modified suppression formulation for $x>0$, using Eqs. \[eqsuppression2\] and \[eqsmod\], ensures that the high-$T$ peak, corresponding to the $n\rightarrow\infty$ Rydberg series of resonances, is suppressed but the low-$T$ peak, corresponding to the $n=4$ resonances, is not suppressed. ](Fig5.eps){width="\textwidth"}
In order to make a smooth transition from the high-$T$ suppression factor $S\left(x;q,T\right)$ given in Eq. \[eqsuppression2\], which is appropriate for the high-$T$ peak region $kT\approx kT_{max}=2\epsilon_N/3$, to the low-$T$ region, where $S^N \rightarrow 1$, we use the modified factor $$\begin{aligned}
S^N(x;q,T) & = & 1 - \left[ 1 - S\left(x;q,T\right) \right]\, \exp\left(-\frac{\epsilon_N(q)}{10kT}\right) \ ,
\label{eqsmod}\end{aligned}$$ where $\epsilon_N(q)=8$ eV for the particular case of ($N=3$ and $q=3$). As seen in Fig. \[fig4\], the density-dependent effective DR rate coefficient, $\alpha_{DR}^{eff}(n_e,T)$, indeed satisfies the requirement that the high-$T$ peak is suppressed according to the formulation of [@badnell] whereas suppression is totally turned off for the lower-$T$ peak.
We have investigated the application of Eq. \[eqsmod\] for all ions that exhibit these same low-$T$ resonances features, namely, all isoelectronic sequences $N=\left\{3,4,7,11,12,15\right\}$, and we have found that the correct transitioning from suppression at the high-$T$-peak to no suppression at low-$T$ is indeed satisfied, provided, of course, that the appropriate dipole-allowed excitation energy $\epsilon_N(q)$ is employed. For efficient representation, the excitation energies along each isoelectronic sequence are parameterized by the expression $$\begin{aligned}
\epsilon_N(q) & = & \sum_{j=0}^5 p_{N,j} \left(\frac{q}{10}\right)^j\ .
\label{eqepsilon}\end{aligned}$$ These parameters, which are determined by fitting the above expression to the available NIST excitation energies [@nist], are listed in Table \[table1\].
[lcrrrrrr]{} Li-like & 3 & 1.963\[+0\] & 2.030\[+1\] & -9.710\[-1\] & 8.545\[-1\] & 1.355\[-1\] & 2.401\[-2\]\
Be-like & 4 & 5.789\[+0\] & 3.408\[+1\] & 1.517\[+0\] & -1.212\[+0\] & 7.756\[-1\] & -4.100\[-3\]\
N-like & 7 & 1.137\[+1\] & 3.622\[+1\] & 7.084\[+0\] & -5.168\[+0\] & 2.451\[+0\] & -1.696\[-1\]\
Na-like & 11 & 2.248\[+0\] & 2.228\[+1\] & -1.123\[+0\] & 9.027\[-1\] & -3.860\[-2\] & 1.468\[-2\]\
Mg-like & 12 & 2.745\[+0\] & 1.919\[+1\] & -5.432\[-1\] & 7.868\[-1\] & -4.249\[-2\] & 1.357\[-2\]\
P-like & 15 & 1.428\[+0\] & 3.908\[+0\] & 7.312\[-1\] & -1.914\[+0\] & 1.051\[+0\] & -8.992\[-2\]\
H-, He-, Ne-like & 1,2,10 & $\infty^{\dag}$ & 0.0 & 0.0 & 0.0 & 0.0 & 0.0\
B-, C-, O-, F-like & 5,6,8,9 & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 & 0.0\
Al-, Si-, S-, Cl-like & 13,14,16,17 & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 & 0.0\
& $\ge 18$ & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 & 0.0\
& &\
We note that all isoelectronic sequences and ionization stages are now included in this prescription — our final comprehensive model for treating DR suppression, albeit in a simplified fashion. For those ions with fine-structure splitting in the ground state, we have $\epsilon_N(q)\approx 0$, so that $S^N(x;q,T)=S(x;q,T)$. (We apply this generally also for Ar-like sequences and above ($N\ge 18$), based-on the density of states — see, for example, [@badnell3pq].) For the closed-shell cases, on the other hand, we have $\epsilon_N(q)\rightarrow \infty$. Thus, $S^N(x;q,T)=1$ for hydrogenic and closed-shell systems, i.e., there is no suppression (for $x\le 10$). Lastly, for the intermediate cases, the suppression factor is gradually increased toward unity at lower temperatures and begins to admit low-$n$ DR resonances.
Results
=======
The suppression factors derived here have been applied to the state-of-the-art total DR rate coefficients taken from the most recent DR database.[^3] These modified data have been incorporated into version C13 of the plasma simulation code Cloudy, most recently described by [@cloudy13]. Cloudy can do simulations of both photoionized and collisionally ionized cases, and we show the effects of collisional suppression on both.
Figure \[figcollision\] shows the ionization distribution of iron for the collisional ionization case. Figure \[figphoto\] shows a similar calculation for photoionization equilibrium. Both show two hydrogen densities, 1 cm$^{-3}$, where collisional suppression of DR should be negligible, and 10$^{10}$ cm$^{-3}$, where collisional suppression should greatly affect the rates for lower charges and temperatures. The upper panel shows the ionization fractions themselves, for these two densities, while the lower panel shows the ratio of the high to low density abundances.
![\[figcollision\] Upper panel: collisional ionization fractional abundance vs. electron temperature for all ionization stages of Fe. The solid curves correspond to a density of 1 cm$^{-3}$ and the dashed curves correspond to a density of $10^{10}~\rm{cm}^{-3}$. From left to right, the curves range from to . Lower panel: ratio of the calculated fractional abundances for the two densities. ](Fig6.eps){width="\textwidth"}
![\[figphoto\] Upper panel: photoionization fractional abundance vs. the ionization parameter $U$ for all ionization stages of Fe. The solid curves correspond to a density of 1 cm$^{-3}$ and the dashed curves correspond to a density of $10^{10}~\rm{cm}^{-3}$. From left to right, the curves range from to . Lower panel: ratio of the calculated fractional abundances for the two densities. ](Fig7.eps){width="\textwidth"}
Cloudy’s assumptions in computing collisional ionization equilibrium, as shown in Figure \[figcollision\], have been described by @lykins. It is determined by the balance between collisional ionization from the ground state and recombination by radiative, dielectronic, and three body recombination to all levels of the recombined species.
The photoionization case shown in Figure \[figphoto\] depicts the Active Galactic Nucleus spectral energy distribution (SED), described by @mathews, as a function of the ionization parameter $$\begin{aligned}
U & \equiv & \frac{\Phi_H}{n_H\;c}\ ,\end{aligned}$$ where $\Phi_H$ is the hydrogen-ionizing photon flux, $n_H$ is the density of hydrogen, and $c$ is the speed of light. There is only an indirect relationship between the gas kinetic temperature and the ionization of the gas in this case. Here, the level of ionization is determined by a balance between photoionization by the energetic continuum and the total recombination rate.
The lower panels of Figs. \[figcollision\] and \[figphoto\] show that the amount that the ionization increases due to DR suppression can be large — the ratio can easily exceed 1 dex. Clearly, these results demonstrate that density effects on the ionization balance need to be considered more precisely.
Conclusion
==========
We have investigated the effects of finite densities on the effective DR rate coefficients by developing a suppression factor model, which was motivated by the early work of @badnell for and extended to all other ions using physically-motivated scaling considerations, and more precise fitting of collisional-radiative data [@summersRAL]. Accurate zero-density DR rate coefficients were then multiplied by this suppression factor and introduced into Cloudy to study the finite-density effects on computed ionization balances of both collisionally ionized and photoionized plasmas. It is found that the difference in ionization balance between the near-zero and finite-density cases is substantial, and thus there is sufficient justification for further studies of collisional suppression from generalized collisional-radiative calculations. This is expected to impact the predictions of the ionization balance in denser cosmic gases such as those found in nova and supernova shells, accretion disks, and the broad emission line regions in active galactic nuclei.
The present results are intended to be preliminary, and to demonstrate the importance of density effects on dielectronic recombination in astrophysical plasmas. Given the approximations adopted, we suggest that their incorporation into models (e.g., via Cloudy) be used with a little caution. For example, one might run models with and without the effects of suppression at finite density, especially in modeling higher density plasmas (e.g., the broad emission line region in quasars). Nevertheless, it is nearly half a century since [@burgsum] demonstrated significant density effects on DR, and it is time that some representation exists within astrophysical modeling codes to assess its impact on the much more rigorous demands made by modern day modeling, especially given its routine incorporation by magnetic fusion plasma modeling codes. In the longer term, we intend to present results based on detailed collisional-radiative calculations using state-of-the-art state-specific DR rate coefficients.
Acknowledgments
===============
DN, TWG, and KTK acknowledge support by NASA (NNX11AF32G). GJF acknowledges support by NSF (1108928; and 1109061), NASA (10-ATP10-0053, 10-ADAP10-0073, and NNX12AH73G), and STScI (HST-AR-12125.01, GO-12560, and HST-GO-12309). UK undergraduates Mitchell Martin and Terry Yun assisted in coding the DR routines used here. NRB acknowledges support by STFC (ST/J000892/1).
[23]{} natexlab\#1[\#1]{}
, N. R. 2006, , 167, 334
, N. R. 2006, , 651, L73
, N. R., [O’Mullane]{}, M. G., [Summers]{}, H. P., [et al.]{} 2003, , 406, 1151
, N. R., [Pindzola]{}, M. S., [Dickson]{}, W. J., [et al.]{} 1993, , 407, L91
, D. R., [Kingston]{}, A. E., & [McWhirter]{}, R. W. P. 1962, Proc. R. Soc. Lond. A, 267, 297
, A. 1965, , 141, 1588
, A. & [Summers]{}, H. P. 1969, , 157, 1007
, K. 1975, , 195, 285
, G. J., [Porter]{}, R. L., [van Hoof]{}, P. A. M., [et al.]{} 2013, ArXiv e-prints
, C. 1969, , 142, 501
, M., [Ferland]{}, G. J., [Porter]{}, R. L., [et al.]{} 2012, , submitted
, W. G. & [Ferland]{}, G. J. 1987, , 323, 456
, H. & [Storey]{}, P. J. 1984, , 56, 293
, D. E. & [Ferland]{}, G. J. 2006, Astrophysics of Gaseous Nebulae and Active Galactic Nuclei, 2nd edn. (Sausalito, CA: University Science Books)
, M. S., [Loch]{}, S. D., & [Robicheaux]{}, F. 2011, , 83, 042705
, Y., [Kramida]{}, A. E., [Reader]{}, J., & [NIST ASD Team]{}. 2011, National Institute of Standards and Technology, http://physics.nist.gov/asd
, H. P. 1972, , 158, 255
, H. P. 1974, , 169, 633
, H. P. 1974 & 1979, Appleton Laboratory Internal Memorandum IM367 & re-issued with improvements as AL-R-5
, H. P. & [Hooper]{}, M. B. 1983, Plasma Physics, 25, 1311
, D. A. & [Yakovlev]{}, D. G. 1995, A&AS, 109, 125
, G. S. 1997, Atomic Data and Nuclear Data Tables, 65, 1
, G. K., [Butler]{}, M. A., [West]{}, K. W., & [Buchanan]{}, D. N. E. 1974, Rev. Sci. Instrum., 45, 1369
[^1]: http://amdpp.phys.strath.ac.uk/tamoc/DATA/
[^2]: We note that the $n=4$ resonance manifold has been the subject of further near-threshold density-dependent effects [@c4].
[^3]: http://amdpp.phys.strath.ac.uk/tamoc/DATA/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The results of experimental study of the magnetoresistivity, the Hall and Shubnikov-de Haas effects for the heterostructure with HgTe quantum well of $20.2$ nm width are reported. The measurements were performed on the gated samples over the wide range of electron and hole densities including vicinity of a charge neutrality point. Analyzing the data we conclude that the energy spectrum is drastically different from that calculated in framework of $kP$-model. So, the hole effective mass is equal to approximately $0.2\,m_0$ and practically independent of the quasimomentum ($k$) up to $k^2\gtrsim 0.7\times
10^{12}$ cm$^{-2}$, while the theory predicts negative (electron-like) effective mass up to $k^2=6\times 10^{12}$ cm$^{-2}$. The experimental effective mass near $k=0$, where the hole energy spectrum is electron-like, is close to $-0.005 m_0$, whereas the theoretical value is about $-0.1\, m_0$.
author:
- 'G. M. Minkov'
- 'A. V. Germanenko'
- 'O. E. Rut'
- 'A. A. Sherstobitov'
- 'S. A. Dvoretski'
- 'N. N. Mikhailov'
title: 'Two-dimensional semimetal in a wide HgTe quantum well: magnetotransport and energy spectrum '
---
Introduction {#sec:intr}
============
Two-dimensional (2D) systems based on gapless semiconductors such as HgTe represent unique object. A great variety of two-dimensional electron and hole systems based on this materials can be realized depending on the quantum well width ($d$) and content of cadmium in the well and barriers Hg$_{1-x}$Cd$_x$Te. It is well established now that the energy spectrum in single CdTe/HgTe/CdTe quantum well at $d=d_c\simeq 6.5$ nm is gapless[@Gerchikov90] and close to the linear Dirac-like spectrum at small quasimomentum ($k$).[@Bernevig06] When thickness $d<d_c$ (i.e., when the HgTe quantum well is narrow) the energy spectrum is analogous to the spatially quantized spectrum of narrow gap semiconductor such as InSb. For thick HgTe layer, $d>d_c$, the quantum well is in inverted regime when the main electron subband of spatial quantization is formed at $k=0$ from the heavy hole states.[@Dyak82e]
The energy spectrum and transport phenomena of 2D carriers in HgTe based structures were studied intensively last decade both experimentally[@Landwehr00; @Zhang02; @Ortner02; @Zhang04; @Koenig07; @Gusev11; @Kvon11] and theoretically.[@Bernevig06; @Tkachov11; @Nicklas11; @Ostrovsky12] The experimental data on the energy distance between the different 2D subbands at zero quasimomentum are in satisfactory agreement with the theory. Electron energy spectrum, electron effective mass and their dependence on the quantum well width are in agreement with the calculations also. As regards to the experimental data on the valence band energy spectrum, namely the value of bands overlapping, role of strain, effective masses at $k=0$ and at large quasimomentum, they are discrepant and call for further investigation.
In this paper, we present the results of experimental study of the transport properties of the heterostructure with the HgTe quantum well with the inverted energy spectrum. The measurements were performed over wide range of electron and hole densities including the vicinity of the charge neutrality point (CNP) with nearly equal electron and hole densities. Analysis of experimental data brings us to the picture of the energy spectrum, which drastically differs from the commonly accepted one.
Experimental {#sec:expdet}
============
Our HgTe quantum wells were realized on the basis of HgTe/Hg$_{1-x}$Cd$_{x}$Te ($x=0.58$) heterostructure grown by means of MBE on GaAs substrate with the (013) surface orientation.[@Mikhailov06] The nominal width of the quantum well was $d=20.2$ nm. The samples were mesa etched into standard Hall bars. To change and control the electron and hole densities ($n$ and $p$, respectively) in the quantum well, the field-effect transistors were fabricated with the parylene as an insulator and aluminium as a gate electrode. The measurements were performed at the temperature of liquid helium in the magnetic field up to $8$ T. All the data will be presented for $T=1.35$ K, unless otherwise specified. The architecture and the energy diagram of the structure investigated is shown in Fig. \[f1\](a) and Fig. \[f1\](b), respectively. The sketch of energy spectrum calculated within framework of the isotropic six-band $kP$-model with taking into account the lattice mismatch is presented in Fig. \[f1\](c). One can see that h1-to-h2 interband distance at $k=0$ is about $5$ meV, the dispersion $E(k)$ in the valence band is non-monotonic.
![(Color online) Architecture (a) and energy diagram (b) of the structure under investigation. (c) – The energy spectrum calculated within the framework of the isotropic six-band $kP$-model.[]{data-label="f1"}](f1.eps){width="\linewidth"}
Results and discussion {#sec:res}
======================
An overview of the magnetic field dependences of a transverse ($\rho_{xy}$) and longitudinal ($\rho_{xx}$) resistivity for different gate voltages ($V_g$) is presented in Fig. \[f2\]. It is seen that well defined quantum Hall plateaus of $\rho_{xy}$ and minimum of $\rho_{xx}$ are observed at electron ($V_g>2$ V) and hole ($V_g<-1$ V) conductivity. Two peculiarities of these dependences should be pointed out. First, some minimum on the $\rho_{xy}$ versus $B$ and $\rho_{xx}$ versus $B$ dependences at $B=(4-6)$ T \[marked by arrows in Fig. \[f2\](a) and in the inset in Fig. \[f2\](b)\] is observed. Its position only slightly depends on $V_g$ within the gate voltage range from $3.0$ V to $0.8$ V.
![(Color online) The magnetic field dependences of $\rho_{xy}$ (a) and $\rho_{xx}$ (b, c) measured for the different gate voltages. The minimum resulting from the crossing between the (h1, $n=0$) and (h2, $n=2$) Landau levels is marked by arrow (for more details, see Section \[sec:cnp\]). The inset in (b) illustrates the weak sensitivity of the minimum position to the gate voltage near the charge neutrality point, $V_g\simeq 1.8$ V.[]{data-label="f2"}](f2.eps){width="\linewidth"}
![(Color online) The low-magnetic-field dependences of $\rho_{xy}$ (a) and $\rho_{xx}$ (b) measured for $B<1$ T at different gate voltages near the charge neutrality point.[]{data-label="f3"}](f3.eps){width="0.8\linewidth"}
Second, the alternative sign Hall resistivity $\rho_{xy}$, negative at low magnetic fields $B<(0.1-0.5)$ T and positive at higher ones \[see Fig. \[f3\](a)\], is observed within the gate voltage range $(-3\ldots
+1.8)$ V. Such behavior of $\rho_{xy}$ accompanied by strong positive magnetoresistivity \[Fig. \[f3\](b)\] shows that two types of carriers, electrons and holes, take part in the conductivity within this gate voltage range. The Hall densities of electrons and holes found as $1/eR_H(B)$ at $B=0.05\text{ T}$ and $B=2\text{ T}$, respectively, are plotted against the gate voltage in Fig. \[f4\]. Excepting the gate voltage range from $0$ V to $2$ V, the data points fall on a straight line with the slope $-5.5\times 10^{10}$ cm$^{-2}$V$^{-1}$, which is close to $-1/eC$, where $C=9.1$ nF/cm$^2$ is the capacity between the 2D gas and gate electrode measured on the same structure.[^1] So, beyond the range $V_g\simeq (0\ldots2)$ V, $-1/eR_H(0.05\text{ T})$ and $1/eR_H(2\text{ T})$ give the electron and hole densities, $n$ and $p$, respectively. The gate voltage $V_g=1.8$ V, at which the straight line crosses zero, corresponds to CNP.
![(Color online) The gate voltage dependence of $1/eR_H(0.05\text{~T})$ and $1/eR_H(2\text{~T})$. The line is the charge density in the quantum well calculated as $C(1.8\text{~V}-V_g)/e$, where $C$ is the capacity between the gate electrode and quantum well measured experimentally, $C=9.1$ nF/cm$^2$.[]{data-label="f4"}](f4.eps){width="0.8\linewidth"}
It is clear that both the peculiarities may result from specific features of the energy spectrum of 2D carries in the structures under study. There are several papers[@Landwehr00; @Zhang01; @Zhang02; @Ortner02; @Kvon11; @Orlita11] on the energy spectrum of electrons and holes in HgTe quantum wells with approximately the same width of the well. However, the energy spectrum, especially of the valence band, is not understood up to now. Therefore, let us start from analysis of the Shubnikov-de Haas (SdH) oscillations.
The positions of the oscillation minima are plotted in the $(B,V_g)$ plane in Fig. \[f5\]. This figure resembles a fan-chart showing the energies of Landau levels as a function of magnetic field. However, it should be noted that the gate voltage, rather than the energy is plotted in the vertical axis. Only in the case when the carrier effective mass does not depend on the energy, the energy varies in direct proportion with $V_g$, $E\propto C V_g/e\nu$, where $\nu$ is density of states. It is clearly seen that the points laying above the dashed line fall on the straight lines, which are extrapolated to $V_g=(1.8\pm 0.1)$ V when $B\to0$ corresponding to CNP. Just such the behavior should be observed when the variation of density of carriers with $V_g$ is determined by the geometrical capacity only.[@Note1] The behavior of $\rho_{xx}$ minima near the dashed line will be discussed below.
![(Color online) The fan-chart diagram showing the positions of the minima in $\rho_{xx}$ versus $B$ dependence. Symbols are the experimental results, the squares correspond to the minima labeled in Fig. \[f2\] by arrows. The dashed line is the $B$ dependence of the gate voltage corresponding to a crossing of the Landau level (h2, $n=2$) with the Fermi level. Solid lines show the expected minima positions found as $V_g^N=\pm eBN/5.5\times 10^{10}h+1.8$ V (above dashed line) and $V_g^N=-eB(N+1)/5.5\times 10^{10}h+1.8$ V (below dashed line). The inset is schematic dispersion of Landau levels. []{data-label="f5"}](f5.eps){width="\linewidth"}
At hole density higher than $10^{11}$ cm$^{-2}$ that corresponds to $V_g< 0$ V, one can find the range of low magnetic field, where the spin-unsplit SdH oscillations are observed \[for example see Fig. \[f6\](b)\]. Fitting the temperature dependence of oscillation amplitude to the Lifshitz-Kosevich formula,[@LifKos55] we have found the hole effective mass. We succeeded in such analysis within the density range $(1\ldots4)\times10^{11}$ cm$^{-2}$. The results are plotted in Fig. \[f6\](a). One can see that the hole effective mass is equal to $m_h=(0.2\pm 0.05)\,m_0$ and only slightly increases with the increasing hole density. Note that the hole density found from SdH oscillations is close to that found as $1/eR_H(2\text{ T})$.
Let us compare this result with the result of the conventional calculation performed within framework of isotropic envelop function approximation based on six-band $kP$-Hamiltonian. Because the valence band spectrum noticeably depends on the strain, we present in Fig. \[f6\] the $E$ versus $k$ and $m_h$ versus $k$ dependences calculated for two cases: with and without taking into account the strain caused by the HgTe and Hg$_{1-x}$Cd$_x$Te lattice mismatch. The strain effect is characterized by the quantity $2\Delta_\epsilon$, which is the splitting of $\Gamma_8$ band at $k=0$ in the bulk HgTe. The estimate for our case, $x=0.58$, gives $2\Delta_\epsilon\simeq
10$ meV. It is seen that the theory for both cases predicts so called “Mexican hat” hole energy spectrum, characterized by the electron-like dispersion $E(k)$ with the positive curvature near $k=0$. It is significant that the hole effective mass calculated theoretically is negative up to $k^2\simeq 2\times 10^{12}$ cm$^{-2}$ or $\simeq6\times 10^{12}$ cm$^{-2}$ depending on the strain. Experimentally, $m_h$ is positive when $k^2\gtrsim0.7\times
10^{12}$ cm$^{-2}$. This conclusion follows immediately from the fact that the oscillation minima at $V_g<0$ shift to the more negative $V_g$ with the growing magnetic field (Fig. \[f5\]). Note that found values of $m_h$ are close to those found from cyclotron resonance.[@Kvon11-1] Thus, experimentally found hole effective mass at low $k$ differs drastically from the calculated one to the extend that they are different in sign.
![(Color online) (a) – The hole effective mass plotted against the $k^2$ value as obtained experimentally (symbols) and calculated theoretically from six-band $kP$-model with and without taking into account the lattice mismatch (the sold and dashed curves, respectively). (b) – An example of the SdH oscillations measured for $V_g=0.25$ V ($p=1.05\times 10^{11}$ cm$^{-2}$) at different temperatures. (c) – The dispersion for the upper hole subband h2 calculated from six-band $kP$-model with and without taking into account the lattice mismatch (the solid and dashed curves, respectively). []{data-label="f6"}](f6.eps){width="\linewidth"}
The electron effective mass $m_e$ measured by the same way for $n=(0.6\ldots 1.5)\times 10^{11}$ cm$^{-2}$ is equal to $(0.02\pm
0.005)\,m_0$ that also coincides with the result obtained in the cyclotron resonance experiments.[@Kvon11-1] Such the value of $m_e$ is something less than the calculated one, which is equal to $0.028\,m_0$ and practically independent of the density up to $n=3\times 10^{11}$ cm$^{-2}$.
The results discussed above do not give information on the energy spectrum at small $k$ values, $k^2\lesssim 0.7\times
10^{12}$ cm$^{-2}$, and on overlapping value of the conduction and valence bands. Such information can be obtained from analysis of the low-field magnetoresistivity and Hall effect near CNP.
![(Color online) The magnetic field dependences of $\sigma_{xy}$ (a) and $\sigma_{xx}$ (b) for several gate voltages near CNP.[]{data-label="f7"}](f7.eps){width="\linewidth"}
Transport near CNP. Two types of charge carriers {#sec:cnp}
================================================
The detailed measurements at low magnetic field show, that over the gate voltage range $(-6 \ldots 1.7)$ V when the electron density is less than $(1.0\ldots 1.5)\times 10^{10}$ cm$^{-2}$ and the hole density is less than $5\times 10^{11}$ cm$^{-2}$, the Hall resistivity $\rho_{xy}$ is strongly nonlinear in the magnetic field at $B<(0.1\ldots0.5)$ T insofar that it changes the sign \[see Fig. \[f3\](a)\]. The large positive magnetoresistivity is observed within this magnetic field range \[Fig. \[f3\](b)\]. Both these facts strongly suggest that two types of carriers, electrons and holes, take part in transport, i.e., the system is in two type carrier conductivity (TTCC) regime. It is more instructive for this case to plot the magnetic field dependence of $\sigma_{xy}$ and $\sigma_{xx}$ instead of $\rho_{xx}$ and $\rho_{xy}$ because just the conductivity tensor components are additive. These dependences for the different gate voltages are plotted in Fig. \[f7\]. It is worth noting that at $V_g=1.7$ V $\sigma_{xy}$ changes the sign from electron to hole one at $B=2.2$ T therewith the last electron plateau of $\sigma_{xy}$ (with $\sigma_{xy}=e^2/h$) is observed at lower magnetic field, $B=
(0.5-1.5)$ T. Such unusual quantum Hall effect determined by minority carriers, which are electrons in this situation, is observed down to $V_g=1.4$ V.
Let us analyze the results relating to the gate voltage range where two types of carriers take part in the transport in more detail. There are several physical reasons for such regime: (i) the existence of the edge states,[@Bernevig06; @Koenig07; @Roth09] which give the electron contribution to the conductivity, while the 2D gas is of hole type; (ii) existence of the electron and hole drops due to potential fluctuations; (iii) the overlapping between the conduction and valence bands following from the standard $kP$-model.[@Kvon11]
If the edge states result in the conductivity by two types of carriers, their relative contribution can be estimated from the value of step-like drop $\Delta\sigma=1/\rho_{xx}(0.5\text{ T})-1/\rho_{xx}(0)$ evident in the $1/\rho_{xx}$ versus $B$ dependence (see inset in Fig. \[f8\]). Because this drop should be inversely proportional to the channel width for this mechanism, we have measured magnetoresistance of wide ($0.5$ mm) and narrow ($0.1$ mm) parts of the Hall bar sketched in Fig. \[f8\]. As seen from Fig. \[f8\] the relative drop values, $\Delta\sigma/\sigma$, are practically identical for the wide and narrow parts of the bar. It means that the edge states in the structures investigated do not responsible for the two types carriers conductivity regime.
![(Color online) The relative value of the drop in the $1/\rho_{xx}$ dependence plotted as a function of the gate voltage for the wide (open symbols) and narrow (solid symbols) parts of the samples shown in the sketch. The inset is $1/\rho_{xx}$ plotted as a function of the magnetic field. Arrow indicates the drop $\Delta\sigma$.[]{data-label="f8"}](f8.eps){width="0.78\linewidth"}
Another possible reason of two type charge carrier conductivity is that the large long-range fluctuation potential leads to existence of $n$ (or $p$) type droplets in the $p$ (or $n$) type matrix. Due to large transparency of $p$-$n$ junction in the structure with inverted energy spectrum, such medium will demonstrate features typical for systems with two types of carriers. The upper limit of the fluctuation potential amplitude can be estimated from the value of the Dingle temperature ($T_D$). As follows from analysis of the SdH oscillations its value does not exceed $1$ meV. So, this mechanism may account for two type carrier conductivity within narrow gate voltage range only, which can be estimated as $\Delta V_g\simeq \nu_h T_D
(dn/dV_g)^{-1}\simeq 1$ V ($\nu_h=m_h/\pi\hbar^2$ is the hole density of states, $m_h=0.2m_0$). Experimentally, this regime occurs within much wider range of $V_g$: from $-6$ V to $+1.7$ V.
![(Color online) The magnetic field dependences of $\rho_{xx}$ (a) è $R_H$ (b). The solid curves are measured experimentally, the dashed lines are the results of the best fit to the classical formula for TTCC regime. []{data-label="f9"}](f9.eps){width="\linewidth"}
Thus, we assume that TTCC regime arises from bands overlapping due to non-monotonic “Mexican hat” energy spectrum of the upper valence subband and thus the data can be analyzed in the framework of model of laterally homogeneous 2D gas. It is naturally in this case to use the classical hand-book formulae for two types of carriers to fit the magnetic field dependences of $\rho_{xx}$ and $R_H$ (see, e.g., Ref. ). Such the fitting procedure has been performed at low magnetic field, $B<0.3$ T, with the use of densities and mobilities of electrons and holes as the fitting parameters under assumption that they are independent of magnetic field. The results of the best for $V_g=1.1$ V and $V_g=-2$ V are shown in Fig. \[f9\]. It is seen that this simple model quite well describes both dependences, $\rho_{xx}(B)$ and $R_H(B)$. The gate voltage dependence of electron density found by this manner at $V_g<1.7$ V is plotted in Fig. \[f10\](a) \[points at $V_g>1.8$ V are obtained as $1/e|R_H(0.03\text{~T})|$\]. It is necessary to stress that electron contribution to the conductivity occurs down to the very large negative gate voltages, $V_g\simeq -6$ V, when the hole density becomes about $5\times 10^{11}$ cm$^{-2}$ (see Fig. \[f4\]).
![(Color online) (a) and (b) – The gate voltage dependences of the electron density and mobility, respectively, obtained at $V_g<1.7$ V within framework of standard TTCC model. (c) – The partial electron and hole conductivities plotted against the gate voltage. The parameters of electrons for $V_g>1.8$ V for all the panels are obtained as follows: $n=1/e|R_H(0.03\text{~T})|$, $\mu_e=|R_H(0.03\text{~T})|\sigma$.[]{data-label="f10"}](f10.eps){width="\linewidth"}
Let us consider the $V_g$ dependence of density in greater detail. It is rather complicated as evident from Fig. \[f10\](a). There are three $V_g$ intervals distinguished by the slopes. (i) At $V_g>1.8$ V, when the Fermi level lies in the conduction band, the electron density decreases with decreasing $V_g$ with the rate $\Delta n/\Delta V_g$ of about $5.5\times 10^{10}$ cm$^{-2}$ V$^{-1}$. (ii) At $V_g=(1.0\ldots
1.7)$ V, the rate is about ten times less, than that at $V_g>1.8$ V, $\Delta n/\Delta V_g \simeq 6.0\times 10^{9}$ cm$^{-2}$ V$^{-1}$ \[see inset in Fig. \[f10\](a)\]. Quite apparently this feature results from the fact that the holes appear at $V_g\simeq 1.8$ V, and the rate of decrease of the Fermi energy is already determined by the hole density of states, which is ten times larger than the electron one: $dE_F/dV_g|_{V_g>1.8\text{~V}}=(1+m_h/m_e)\,
dE_F/dV_g|_{V_g<1.8\text{~V}}$. Thus, extrapolating the $n$ versus $V_g$ plot to $n=0$ one obtains that the electrons of the conduction band have to disappear at $V_g\approx0$ V \[see the solid line in Fig. \[f10\](a) and the inset in it\]. (iii) At $V_g=(-6\ldots
+1)$ V, the electron contribution to the conductivity remains essential, as it follows from analysis of the magnetic field dependences of $R_H$ and $\rho_{xx}$, despite the fact that the electrons in the conduction band are expected to disappear at $V_g\approx0$ V. The data points in this range fall on the straight line with the slope $\Delta n/\Delta V_g\simeq 1.5\times
10^{9}$ cm$^{-2}$ V$^{-1}$.
We focus now attention on the gate voltage dependence of the electron mobility \[Fig. \[f10\](b)\]. As clearly seen it is nonmonotonic; sharp minimum is evident near $V_g\simeq 1.8$ V. At $V_g> 1.8$ V, when the conductivity is determined by the conduction band electrons and the holes in the valence band are absent, the mobility decreases with the decreasing gate voltage. This is natural because the decrease of the electron density and hence the electron energy leads to the increase of scattering probability independently of that short- or long-range scattering potential determines the mobility. The increase of electron mobility with decreasing $V_g$ at $V_g< 1.8$ V, where the electron density carries on decreasing, seems strange at first sight. However it can be explained by holes appeared at these gate voltages, which screen the potential of scatterers effectively due to the large effective mass. The authors of Ref. who observed analogous behavior of the electron mobility come to the same conclusion.
As was discussed above, the contribution of the conduction band electrons to the conductivity has to disappear at $V_g\approx 0$ V so presence of electron contribution up to the large negative $V_g$ seems surprising. Broadly speaking, shunting conductivity channels of technological nature can make such a type of contribution. They can be located either above the quantum well or below it. If such channels are situated above the well, i.e., between the quantum well and the gate electrode, they should reveal themselves in the capacitance measurements. However, no peculiarities in the voltage-capacity characteristics caused by the depletion of these channels were observed in our experiments. Besides, the contribution of the channels to the conductivity has to be enhanced with the increasing positive gate voltage due to the increase of electron density in these channels. However, as Fig. \[f10\](c) shows the electron contribution to the conductivity has a maximum at $V_g\simeq 0$ V and decreases with the increasing positive gate voltage. If the shunting channels are located below the well, their contribution should disappear at the negative voltage applied to the back gate electrode. Our measurements performed on the back-gated samples show that the voltage applied to the back- and top gate electrodes changes the magnetic field dependences of $R_H$ and $\rho_{xx}$ analogously. Thus no conducting channels exist above or below the quantum well.
It remains to assume that the electron contribution at $V_g<0$ V is caused by carriers in the well, namely, by the holes with electron-like energy spectrum characterized by the negative effective mass in vicinity of $k=0$. Thus, such data interpretation leads to the energy spectrum sketched in Fig. \[f11\]. Qualitatively, it is close to the calculated one, however quantitative difference is significant. The conduction and valence bands are overlapped. The value of overlapping ($\Delta E_\text{ovrl}$) can be estimated from the value of the electron density at the gate voltage corresponding to appearance of the hole contribution to the conductivity, $V_g\simeq 1.8$ V (see Fig. \[f10\]). As seen from this figure $n\sim 1\times
10^{10}$ cm$^{-2}$ at $V_g\simeq 1.8$ V that gives $\Delta
E_\text{ovrl} \simeq n/\nu_e\sim 1$ meV if one uses the experimental value of electron effective mass $m_e=0.02m_0$ to calculate the electron density of states $\nu_e$.
Experimentally, the hole energy spectrum is monotonic at $p\gtrsim
10^{11}$ cm$^{-2}$ that corresponds to $k=(2\pi p)^{1/2}\gtrsim
8\times10^5$ cm$^{-1}$. This conclusion follows immediately from the two facts. First, the hole Hall density $p=1/eR_H$ is close to that found from SdH oscillations. Second, the peaks of SdH oscillations are shifted to the higher magnetic field when the gate voltage becomes more negative (see Fig. \[f5\]). For $k<(7-8)\times 10^5$ cm$^{-1}$, the hole energy spectrum is electron-like; there is narrow minimum in the dispersion $E(k)$. The depth of this minimum can be easily estimated from the hole density wherein the electron contribution disappears: $\Delta=p|_{n=0}/\nu_h$, where $p|_{n=0}$ is the hole density for the gate voltage $V_g\simeq -6$ V \[see Fig. \[f10\](a)\]. With the use of experimental value of the hole effective mass, $m_h=0.2m_0$, and $p|_{n=0}=p(V_g=-6\text{ V})\simeq (4 \ldots 5)\times
10^{11}$ cm$^{-2}$ (see Fig. \[f4\]), one obtains $\Delta\simeq
5$ meV. The effective mass of electron-like states of the valence band is about $m_{e-l}\simeq -0.005\,m_0$. This estimate is obtained from the depth of minimum $\Delta$ and from the amount of electron-like states, $\simeq1\times 10^{10}$ cm$^{-2}$, estimated as $n$ in the region $V_g=(0\ldots 1)$ V \[see Fig. \[f10\](a)\]: $|m_{e-l}|=\pi
\hbar^2 n(V_g=1\text{ V})/\Delta$. Since three types of carriers take part in the transport at $V_g=(0\ldots 1)$ V, the $n$ values at these gate voltages are obtained with considerably low accuracy. Therefore, this estimate for $m_{e-l}$ should be considered as rough enough. In Fig. \[f11\], we reconstruct the energy spectrum using the parameters $m_e$, $m_h$, $m_{e-l}$, $\Delta$ and $\Delta E_\text{ovrl}$ and compare it with the spectrum calculated in framework of the $kP$-model. One can see that $E$ versus $k$ dependence in conduction band is close to the theoretical one. The difference between the dispersion curves $E(k)$ for the valence band is crucial.
![(Color online) The dispersion $E(k)$ reconstructed from the data analysis as described in the text (solid lines) and calculated within framework of isotropic six-band $kP$-model (dashed lines). []{data-label="f11"}](f11.eps){width="0.8\linewidth"}
We perceive that strong difference between the calculated and reconstructed spectra indicates that discussion of alternative approaches to the interpretation of the data is required. This will be done below. Now let us examine the additional arguments in favor of the energy spectrum depicted in Fig. \[f11\]. They can be obtained from analysis of the data obtained in strong magnetic field.
As noted above, the experimental dependences $\rho_{xy}(B)$ and $\rho_{xx}(B)$ exhibit some peculiarities at the gate voltages near CNP \[in Figs. \[f2\](a) and \[f2\](b), they are marked by the arrows\]. The position of this minimum in the $(B,V_g)$ coordinates is shown in Fig. \[f5\] by squares. Its origin directly relates to specific features of the magnetic field quantization of the energy spectrum in the quantum well with inverted spectrum.[@Schultz98] There are two singular Landau levels responsible for that \[see Fig. 1 in Ref. and the inset in Fig. \[f5\]\]. One of them is the lowest level (h1, $n=0$) of the conduction band. Due to complex nature of the band structure, its energy linearly decreases with increasing magnetic field, while the other Landau levels of the conduction band increase their energies as it takes place in usual system. The second singular Landau level is the singular level (h2, $n=2$) of the valence band, which starts at $B=0$ from the energy of h2 branch at $k=0$ and shifts upward with the increasing $B$. Unusual behavior of these levels leads to a crossing of the conduction- and valence band states at the critical value of the magnetic field, $B=B_c$. The behavior of the crossing point with the change of the quantum well width is investigated in Ref. . If one extrapolates the calculation results[@Koenig07] to $d=20$ nm, we obtain $B_c\simeq (4-5)$ T, which is close to position of the peculiarities evident in the $\rho_{xx}$- and $\rho_{xy}$ versus $B$ dependences.
Thus, for the case when $n$ is slightly larger than $p$, the Fermi level lies in the Landau level (h1, $n=0$) at $B\lesssim B_c$, while at higher magnetic field, $B\gtrsim B_c$, it occurs in the Landau level (h2, $n=2$) resulting in switching of the electron ground state and in the peculiarity of $\rho_{xx}$ and $\rho_{xy}$. Existence of such switching was discussed in recent paper.[@Raichev12]
The level (h2, $n=2$) reveals itself not only at low charge carrier density. Really, inspection of Fig. \[f5\] shows that the data points located above the dashed line fall on the straight lines, which are well extrapolated to the common point $V_g\simeq 1.8$ V at $B=0$. Their positions can be estimate from simple relation $n,p(V_g)=\pm ó BN/h$, where $N$ is the number of occupied Landau levels. Because experimentally $n+p\simeq -5.5\times 10^{10}(V_g-1.8\text{
V})$, cm$^{-2}$ (see Fig. \[f4\]), we obtain for $V_g^N$: $V_g^N=\pm
eBN/5.5\times 10^{10}h+1.8$ V. The solid lines in Fig. \[f5\] are drawn according to this equation. It is seen that the separation between the data points corresponding to $N=2$ and $3$ is well resolved while the points with $N=4$ and $5$, $N=6$ and $7$, and so forth, are merged. The points below the dashed line turn out to be shifted from the lines. The reason is clear. When the level (h2, $n=2$) increasing its energy with the growing magnetic field crosses the Fermi level at magnetic field $B=B^\star$ (see inset in Fig. \[f5\]), the number of Landau levels occupied by holes becomes larger on $1$ so that the positions of minima below the dashed line in Fig. \[f5\] should be described by $V_g^N=-eB(N+1)/5.5\times 10^{10}h+1.8$ V and, thus, the oscillations turn out to be shifted (see solid lines in Fig. \[f5\]). Thus, by and large this model describes the behavior of the oscillation minima rather well. Some discrepancy is not surprising because this simple model is valid when the overlapping between the Landau levels is small, while the experimental data were obtained within the wide magnetic field range involving both the Shubnikov-de Haas oscillations and the quantum Hall ranges.
The main conclusion, which follows from this consideration and from Fig. \[f5\] is that the Landau level (h2, $n=2$) at $B\to 0$ crosses the Fermi level at $V_g\simeq -6$ V, when the hole density is about $4.5\times 10^{11}$ cm$^{-2}$. It means that the Fermi level lies below the top of the valence band for a distance of $\sim 5$ meV. This value is consistent with that estimated from the hole density at the gate voltage corresponding to disappearance of the electron contribution to the conductivity, that supports the energy spectrum presented in Fig. \[f11\] as well.
Let us now discuss alternative interpretations of data. Analogous heterostructures were investigated in Refs. . Partially, our results and results reported in these papers are overlapping. It concerns in particular the value of the hole effective mass,[@Kozlov11] existence of electron contribution to the conductivity up to the hole density $(4-5)\times
10^{11}$ cm$^{-2}$.[@Kvon11] To interpret the results, the authors of Refs. suggest that there is an overlap by about $5$ meV of the conduction band minimum at $k=0$ with two symmetrical maxima of the valence band located at $k\neq 0$ in the direction $[031]$ characterized by the effective masses close to $0.2\,m_0$. This picture agrees well with the value of electron density $n\simeq (4-5)\times 10^{10}$ cm$^{-2}$ (it is four-to-five times higher than that in our case), at which the hole contribution to the conductivity appears in samples investigated in Ref. . However this model of the energy spectrum is in conflict with the fact that the hole density found from the period of SdH oscillations coincides with that obtained from the Hall effect. Because the valence band has two maxima in the spectrum, the hole density obtained from the SdH experiments should be twice as small as the Hall density. To resolve this contradiction the authors[@Kvon11] assume that the spectrum is split by spin-orbit interaction due to asymmetry of the quantum well. It is, however, strange that this effect does not reveal itself in SdH oscillations. The results[@Kvon08; @Olshanitsky09] can be understood in framework of our model, where the lines of constant energy are (nearly) circle centered at $k=0$. However our results cannot be interpreted within framework of the model suggested in Refs. . First, the hole contribution to the conductivity in our samples appears at the electron density of about $1\times 10^{10}$ cm$^{-2}$ instead of $5 \times
10^{10}$ cm$^{-2}$ in Ref. . This corresponds to the overlap value of about $1$ meV instead of $5$ meV. Second, the spin-orbit splitting in our samples is evident in SdH effect at $n\gtrsim 2\times 10^{11}$ cm$^{-2}$ and $p\gtrsim 4\times
10^{11}$ cm$^{-2}$. At lower densities its value is estimated as $1$ meV or less, i.e., the valence band can be considered as unsplit one under our experimental conditions.
Thus, our results and their interpretation lead us to the conclusions on the energy spectrum of the HgTe quantum well which are inconsistent with that obtained in framework of traditional $kP$-model. A most surprising result is existence of narrow electron-like pit in the center of the valence band with the depth of about $5$ meV characterized by the very low effective mass $|m_{e-l}|\simeq
0.005\,m_0$.
Conclusion {#sec:concl}
==========
We have studied the transport phenomena in HgTe single quantum well with inverted energy spectrum. Consistent analysis of the magnetic field dependences of the magnetoresistivity, the Hall coefficient, and the SdH effect in the gated samples carried out over the wide range of the electron and hole densities including the charge neutrality point leads us to the conclusion that the structure of the top of the valence band is drastically different from that predicted in framework of standard $kP$-approach. We obtain that the hole effective mass is equal to approximately $0.2\,m_0$ at $k\gtrsim 0.7\times 10^{12}$ cm$^{-2}$ and practically independent of the hole density, while the theory predicts negative (electron-like) effective mass up to $k^2\simeq
6\times 10^{12}$ cm$^{-2}$. The experimentally obtained effective mass near $k=0$, where the spectrum is electron-like, is close to $-0.005\,m_0$, whereas the theory predicts the value less than $-0.1\,
m_0$. All this indicates that the further experimental and theoretical investigations are needed to find the answer to the question of whether the standard $kP$ model adequately describes the energy spectrum of wide HgTe based single quantum well.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work has been supported in part by the RFBR (Grant Nos. 10-02-91336, 10-02-00481, and 12-02-00098).
[27]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} [****, ()](\doibase 10.1126/science.1133734) @noop [****, ()]{}, @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.65.045324) [****, ()](\doibase
10.1103/PhysRevB.66.075322) [****, ()](\doibase 10.1103/PhysRevB.69.115340) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevB.84.121302) [****, ()](\doibase
10.1103/PhysRevB.83.193304) [****, ()](\doibase 10.1103/PhysRevB.84.035444) [****, ()](\doibase 10.1103/PhysRevB.84.121308) [****, ()](\doibase 10.1103/PhysRevB.86.125323) [****, ()](\doibase 10.1504/IJNT.2006.008725) [****, ()](\doibase
10.1103/PhysRevB.63.245305) [****, ()](\doibase
10.1103/PhysRevB.83.115307) @noop [****, ()]{}, @noop [****, ()]{}, @noop [****, ()]{} @noop [**]{} () p. @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.57.14772) [****, ()](\doibase
10.1103/PhysRevB.86.155320) @noop [****, ()]{}, @noop [****, ()]{}, @noop [****, ()]{},
[^1]: The capacity measured is practically independent of the gate voltage over the entire $V_g$ range. Only one percent variation of $C$ resulting from the finite value of compressibility of the electron gas is evident at $V_g\approx 1.8$ V.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Insulator-metal transition is investigated self-consistently on the frustrated Shastry-Sutherland lattice in the framework of Slave-Boson mean-field theory. Due to the presence of quasi-flat band structure characteristic, the system displays a spin-density-wave (SDW) insulating phase at the weak doping levels, which is robust against frustration, and it will be transited into an SDW metallic phase at high doping levels. As further increasing the doping, the temperature or the frustration on the diagonal linking bonds, the magnetic order $m$ will be monotonically suppressed, resulting in the appearance of a paramagnetic metallic phase. Although the Fermi surface of the SDW metallic phase may be immersed by temperature, the number of mobile charges is robust against temperature at weak doping levels.'
author:
- 'Huai-Xiang Huang'
- Jing Chen
- Wei Ren
- Yi Gao
- Wei Li
- Yan Chen
title: 'Insulator-metal transition and quasi-flat-band of Shastry-Sutherland lattice'
---
Introduction
============
Geometrically frustrated lattices [@1; @2; @3; @4; @hxh; @5; @14prb; @15prb] have exotic quantum states and rich phase diagrams since the antiparallel alignment of adjacent spins cannot be fully satisfied due to the energy competing. Among of them the Shastry-Sutherland lattice (SSL) is one of the simplest systems, which alternates diagonal links located on a square lattice. Compounds with topology equivalent to SSL, such as $\mathrm{SrCu_2(BO_3)_2}$, $\mathrm{Yb_2Pt_2Pb}$, [@akoga; @whz1; @16; @18; @19; @20; @mk1; @mk2; @su] have been synthesized and provided an excellent platform to study the effects of frustration on the correlated electron systems.
Various analytical methods and numerical techniques have been employed to study the SSL. [@dca; @chc; @ala; @dav; @sel; @sha; @haidi] From the viewpoint of localized two dimensional Heisenberg Hamiltonian, the original paper gave an exactly analytical solvable ground state [@13], which is the production of valence-bond dimer singlets on the disjointed diagonal links when the ratio of exchange coupling on diagonal bonds and that on $x$-axes $\alpha=J_{diag}/J_{x}>1.5$. [@akoga; @whz1] While for small $\alpha$, the Néel states will become the ground state since a SSL tends to degenerate to square lattice. Between the dimer-singlet state and the Néel state, many intermediate phases [@mal; @wzh2; @pco; @hxh2; @zhen] and the magnetization plateaus [@ssmag; @ssmag1; @ssmag2; @Trinh] have also been reported.
The lately realized frustrated Lieb lattice [@fla1; @fla2; @fla3; @fla20; @zhi] and the paradigmatic frustrated Kagomé lattice have unpredicted properties owing to the flat-band structure. [@fla18; @ssmag; @fla1] Moreover, the frustrated graphene sheets lead the system displaying a Mott-like insulator due to the strong electron correlation. [@yuan] SSL can be constructed as a quasi-flat-banded, frustrated and correlated electron system based on $t_1$-$t_2$-$J_1$-$J_2$ model and it motivates us to study the intriguing metal-insulator transition systematically.
Considering the itinerant electron behaviors, the $t$-$J$ model has been used to study the possible superconducting phase on SSL, [@hxh2; @chunghouchung] and the Hubbard model has been used to clarify the metal-insulator transition at around the half-filling. [@haidi] However, insulator-metal phase transition on SSL at finite doping and the finite temperature is still lacking investigation theoretically as far as we have known.
In the present paper, we mainly focus on the study of intriguing insulator-metal phase transition on SSL by using a $t_1$-$t_2$-$J_1$-$J_2$ model in the framework of Slave-Boson (SB) approach. [@ale; @cts; @kot; @yuanqsh] Once the mean-field order parameters are self-consistently calculated, the band structure and Fermi surface topology can be evaluated straightforwardly. Based on the linear-response theory, the temperature dependent Drude weight proportionally to the electrical conductivity is also investigated. The effect of the finite third nearest-neighbor (n.n.) hopping term $t_3$ on the flat band structure will also be addressed in the appendix.
Besides the unfrustrated square lattice, lots of unpredicted properties will arise from the disjointed diagonal frustrated parameters $t_2$ and $J_2$ on the SSL. Strong electron-electron interaction will separate the energy bands into two subspaces at weak dopings, and the system displays an insulating state with SDW order (SDW-Ins) and absence of Fermi surface, which is robust against $t_2$, $J_2$ with tiny Drude weight. At high doping levels, the system enters into a metallic phase, which is separated as SDW metallic (SDW-M) phase and paramagnetic metallic (PM-M) phase depending on whether the staggered magnetic order is finite.
The rest of this paper is organized as follows. In Sec. \[model\], we will introduce the theoretical model and the corresponding analytical formalism. The detailed calculated phase diagram is shown in Sec. \[phasediagram\]. In Sec. \[temperature\_effects\], we will turn to discuss the temperature dependent effects on the calculated physical quantities. At last, a summary is given in Sec. \[summary\].
Model Hamiltonian and formula {#model}
=============================
![(Color Online) (a) Schematic illustration of SSL, $i=1..4$ denoting the four different sublattices. Order parameters on the dark links belong to the same unit cell. Note that $1\rightarrow2$ and $2\rightarrow\textbf{\emph{1}}$ are two different links in one unit cell. $t_1$ , $J_1$ are on the n.n. links denoted by the red solid line, $t_2$, $J_2$ are on the diagonal links denoted by the blue dotted line, and $t_3$ on the third n.n. is denoted by the green dashed line. (b) The first Brillouin zone of SSL. $\Gamma$, M, K are the three high symmetry $\mathbf{k}$-points. (c) The energy band dispersion of the tight-binding model along the high symmetric line of M-K-$\Gamma$-M for the parameters of $t_1=\pm 1$, $t_2=0.3$, $t_3=0$ and $\delta=0.02$.[]{data-label="fig1"}](F1.eps){width="8.5cm"}
We consider the $t_1$-$t_2$-$t_3$-$J_1$-$J_2$ phenomenological model to systematically study the frustrated SSL by using SB mean-field theory, where $t_3$ is set to zero without explicitly specified. The crystal structure of SSL is illustrated in Fig. \[fig1\](a). Within a unit cell, there are four inequivalent sites labeled by $1-4$, the hopping integral $t_1$ and spin exchange coupling $J_1$ are on the n.n. links, and $t_2$ and $J_2$ are on the diagonal links. Generally, the $t$-$J$ model can be derived analytically from the Hubbard model in the strong coupling limit with $J=4t^2/U$, where $U$ is the on-site Coulomb interaction in the Hubbard model. The generalized $t$-$J$ model with independent parameters of $t$ and $J$ can be used to describe the more complex systems, like SSL from the phenomenological viewpoint. In real space the Hamiltonian reads $$\begin{aligned}
\label{Hamreal}
&H&=H_t+H_J-\mu_0\sum_i n_i=-t_1\sum_{\langle ij \rangle,\sigma}(\hat{c}^{\dag}_{i\sigma }\hat{c}_{j\sigma }+h.c.)\nonumber\\
&&-t_2\sum_{\langle ij \rangle_2,\sigma}(\hat{c}^{\dag}_{i\sigma }\hat{c}_{j\sigma }+h.c.)-t_3\sum_{\langle ij \rangle_3,\sigma}(\hat{c}^{\dag}_{i\sigma }\hat{c}_{j\sigma }+h.c.)\nonumber\\
&+&J_{1}\sum_{ \langle ij \rangle} (\hat{S}_i \cdot \hat{S}_j-\frac{1}{4}\hat{n}_i\hat{n}_j)+J_{2}\sum_{ \langle ij \rangle_2 } (\hat{S}_i \cdot \hat{S}_j-\frac{1}{4}\hat{n}_i\hat{n}_j)\nonumber\\
&-&\mu_0\sum_i n_i,\end{aligned}$$ where $\sigma$ denotes spin, $\langle \rangle$, $\langle \rangle_2$, $\langle \rangle_3$ represent the n.n., second n.n. and third n.n. sites, respectively, as shown in Fig.1(a), with $\hat{S}_i=\frac{1}{2}(\hat{c}^{\dag}_{i\alpha}\mathbf{\sigma}_{\alpha\beta}\hat{c}_{i\beta})$ denoting the spin operator. For a strong coupling limit and hole-doped case, $\hat{c}_{j\sigma }=c_{j\sigma }(1-\mathfrak{n}_{j\bar{\sigma}})$ means annihilating an electron on a single occupied site with $(1-\mathfrak{n}_{j\bar{\sigma}})\simeq1$, and $\hat{c}^{\dag}_{i\sigma }=(1-\mathfrak{n}_{i\bar{\sigma}})c^{\dag}_{i\sigma }$ represents creating an electron on an empty site with $(1-\mathfrak{n}_{i\bar{\sigma}})\simeq \delta$, $\mathfrak{n}$ representing the occupation number of real electrons and $\delta$ describing the hole-doped concentration. SB [@ale; @cts; @kot; @yuanqsh] approach allows a physical description of the electron correlated effects, by writing $\hat{c}_{i\sigma}=b^{\dag}_if_{i\sigma}$ with $b_i$ Boson holon operator and $f_{i\sigma}$ the Fermionic spinon operator. Without double occupancy constraint, $b_i^{\dag} b_i+\Sigma_{\sigma} f_{i\sigma}^{\dag}f_{i\sigma}=1$ has to be satisfied and is imposed on the Hamiltonian through a Lagrange multiplier, thus $\mu_0$ in Eq. (\[Hamreal\]) changes to $\mu$ hereafter, which is the chemical potential and controls the electronic concentrations $n_i=\sum_{\sigma}n_{i\sigma}=\sum_{\sigma}f^{\dag}_{i\sigma}f_{i\sigma}$. After defining the staggered magnetic order $m=(-1)^{i}\langle n_{i\uparrow}-n_{i\downarrow}\rangle/2$, we obtain the relation $\langle n_{i\sigma}\rangle=(1-\delta)/2+\sigma(-1)^i m$. In the static SB approximation, the boson condensation is assumed $\langle b_i^{\dag} b_i\rangle=\delta$ since the bosonic fluctuations are suppressed.
Introducing the mean-field bond order $\chi^{\sigma}_{ij}=\langle f^{\
\dag}_{i\sigma}f_{j\sigma}\rangle$, spin coupling interaction in mean-field level reads $$\begin{aligned}
\label{Hamj}
H_J&=&-\Sigma_{\langle ij \rangle,\langle ij \rangle_2,\sigma} \frac{J_{ij}}{2}[c^{\dag}_{i\sigma}c^{\dag}_{j\bar{\sigma}}c_{i\bar{\sigma}}c_{j\sigma}-c^{\dag}_{i\sigma}c^{\dag}_{j\bar{\sigma}}c_{i\sigma}c_{j\bar{\sigma}}]\nonumber\\
&=&-\Sigma_{\langle ij \rangle,\langle ij \rangle_2, \sigma} \frac{J_{ij}}{2}[\langle n_{i\sigma}\rangle f^{\dag}_{j\bar{\sigma}}f_{j\bar{\sigma}} +\langle n_{j\sigma}\rangle f^{\dag}_{i\bar{\sigma}}f_{i\bar{\sigma}}\nonumber\\
&+&\chi^{\sigma}_{ij}f^{\dag}_{j\bar{\sigma}}f_{i\bar{\sigma}}+\chi^{\sigma}_{ji}f^{\dag}_{i\bar{\sigma}}f_{j\bar{\sigma}}],\end{aligned}$$ where $J_{ij}$ equals $J_1$ on the n.n. links and $J_2$ on diagonal links, otherwise it is zero, shown in Fig. \[fig1\](a). Since we are interested in the insulator-metal transition, the superconducting order is ignored for simplify. [@pwl] In the momentum $\mathbf{k}$ space, the Hamiltonian is expressed as $\psi^{\dag} [H_k-\mu]\psi$ after dropping a constant term, where $\psi^{\dag}=[f^{\dag}_{1\uparrow}(k)...f^{\dag}_{4\uparrow}(k),f^{\dag}_{1\downarrow}(k)...f^{\dag}_{4\downarrow}(k)]$, and $$\begin{aligned}
\label{1}
H_k&=&\left(
\begin{array}{cc}
H_{t,k}+ H^{\uparrow}_{J,k} & 0 \\
0 & H_{t,k}+H^{\downarrow}_{J,k} \\
\end{array}
\right),\\
H_{t,k}&=&\delta \left(
\begin{array}{cccc}
t_3a_3 &-2t_1\cos{k_x} & -t_2 e^{-\mathrm{i}(k_x+k_y)} & -2t_1\cos{k_y} \\
-2t_1\cos{k_x} & t_3a_3 & -2t_1\cos{k_y} & -t_2 e^{-\mathrm{i}(k_x-k_y)} \\
-t_2 e^{\mathrm{i}(k_x+k_y)} & -2t_1\cos{k_y} & t_3a_3 & -2t_1 \cos{k_x} \\
-2t_1\cos{k_y} & -t_2 e^{\mathrm{i}(k_x-k_y)} & -2t_1\cos{k_x} & t_3a_3 \\
\end{array}
\right), \nonumber\end{aligned}$$ with $a_3=-2(\cos{2k_x}+\cos{2k_y})$, and $\mathbf{k}$ is restricted in the first Brillouin zone as shown in Fig. \[fig1\](b). In the deduction $\langle b^{\dag}_ib_j \rangle$ is approximated as $\delta$, and $b_ib_jb^{\dag}_ib^{\dag}_j$ is replaced by unity. In the following $|t_1|$ and the distance between the n.n. site are set as energy unit and length unit, respectively. It is easy to prove that $\pm t_1$ will give the same energy band dispersions for the tight-binding model $H_{t,k}$. For the special range $k_x=k_y$, one of the band dispersion of $H_{t,k}$ has a simply analytical expression as $$\begin{aligned}
\label{2}
\epsilon=\delta( t_3a_3+t_2),\end{aligned}$$ which is a constant for $t_3=0$, indicating a quasi-flat band represented by the flat segment from K to $\Gamma$ shown in Fig. \[fig1\](c).
Interaction Hamiltonian $ H^{\sigma=\uparrow,\downarrow}_{J,k}$ depends on spin indexes, for example $H^{\sigma}_{J,k}(1,2)=-0.5J_1 (\chi^{\bar{\sigma}}_{21}e^{\mathrm{i}k_x}+\chi^{\bar{\sigma}}_{2\textbf{\emph{1}}}e^{-\mathrm{i}k_x} )$, $H^{\sigma}_J(2,4)=-0.5J_2 \chi^{\bar{\sigma}}_{42}e^{-\mathrm{i}(k_x-k_y)}$, where the $\chi^{\sigma}_{21}$ and $\chi^{\sigma}_{2\textbf{\emph{1}}}$ are on different links shown in the Fig. \[fig1\](a), and the diagonal element $H^{\sigma}_{J,k}(1,1)$ reads $-J_1(1-\delta-\sigma2m)-0.25J_2(1-\delta+\sigma2m)$. In the Hamiltonian $m$, $\mu$ as well as the $20$ mean-field $\chi^{\sigma}_{ij}$ are self-consistently calculated.
In the presence of a slowly varying vector potential $\textbf{A}$ along $x$-direction, the associated Peierls phase is $ c^{\dag}_{i\sigma}c_{j\sigma}\exp{\mathrm{i}\frac{e}{\hbar
c}\int^{r_i}_{r_j}\textbf{A}(\textbf{r},t)\cdot \mathrm{d} \textbf{r}}$, and the charge current density can be decomposed into diamagnetic and paramagnetic part $J_x(r_i)=eJ^{p}_x(r_i)+e^2K_x(r_i)\textbf{A}_x(r_i,t)$, where $$\begin{aligned}
\label{curr}
K_x(r_i)&=&-\delta\sum_{\sigma j}t_{ij}(f^{\dag}_{i\sigma}f_{j,\sigma}+H.c.),\\
J^P_x(r_i)&=&-\mathrm{i}\delta\sum_{\sigma j}t_{ij}(f^{\dag}_{i\sigma}f_{j,\sigma}-H.c.),\end{aligned}$$ with $j=i+\vec{x}, i+\vec{x}\pm \vec{y}$. Applying the linear response theory, the Drude weight $\frac{D}{\pi e^2}=\frac{1}{N}\Pi_{xx}(\textbf{q}=0,\omega\rightarrow 0)-\langle K_x \rangle_0$, a measurement of the ratio of density of mobile chargers to their mass, can be evaluated straightforwardly. [@1d; @2d; @3d; @4d; @5d] The current-current correlation reads $\Pi_{xx}(\textbf{q},\tau)=-\langle T_{\tau} J^P_x(\textbf{q},\tau)J^P_x(-\textbf{q},0)\rangle_0$ with $J^P_x(\textbf{q},\tau)=e^{\tau H}J^P_x(\textbf{q}) e^{-\tau H}$, $\mathrm{T}_{\tau}$ time ordering operator, $\tau$ imaginary time, and $J^P_x(\textbf{q})=\sum_i e^{-\mathrm{i}\textbf{q}\cdot r_i}J^P_x(r_i)$. Here it should be noted that for an insulating phase $D$ is close to zero, whereas $D$ is finite for a metallic phase.
The diamagnetic current $K_x$ is easy to derive since it needs to be calculated to zeroth order of $\textbf{A}$. The current correlated function is expressed as $\Pi_{xx}(q,\tau)=\frac{1}{\beta}\sum_{n}\Pi_{xx}(q,\mathrm{i}\omega_n)e^{-\mathrm{i}\omega_n\tau}$ in the Matsubara formalism. It can be expressed as $$\begin{aligned}
\label{para}
\Pi_{xx}(\textbf{q},\mathrm{i}\omega)
\!\!=\!\!\!\!\!\sum_{{\bf{k}}m_1m_2} \!\! \frac{ Y_{m_1m_2}(k)Y_{m_2m_1}(k)(\mathbb{F}(E_{{\bf{k}},m_1})-\mathbb{F}(E_{{\bf{k}},m_2}) ) }{\mathrm{i}\omega+(E_{{\bf{k}},m_1}-E_{{\bf{k}},m_2})},\end{aligned}$$ by using the equation of motion, where $\mathbb{F}$ is the Fermi distribution function, $Y_{m_1m_2}(k)$ is a lengthy straightforward function of transformation matrix $\mathbb{T}$, obtained in the process of diagonalizing $H_k$. Through analytic continuation of $\Pi_{xx}(q,\mathrm{i}\omega)|_{i\omega=\omega+i\eta}$, $\Pi_{xx}(\textbf{q},\omega)$ appearing in the expression of $D$ is obtained.
Throughout the this paper, we use the parameters of $t_1=\pm1$ and $J_1=0.3$, by following the previous studies on the square lattice, [@yuanqsh; @tto; @tkl; @pwl] and these choices do not affect on the physical discussions. The number of unit cell in the self-consistent calculation is $N=128\times 128$ with the accuracy less than $10^{-4}$, while in the calculations of density of state (DOS), Drude weight $D$ and the band structure are used $N=640\times 640$. Since there are $16$ n.n. bond orders, which can be divided into two groups with $ \chi^{\uparrow}_{4\textbf{\emph{1}}}=\chi^{\uparrow}_{3\textbf{\emph{2}}}=\chi^{\uparrow}_{3\textbf{\emph{4}}}=\chi^{\uparrow}_{2\textbf{\emph{1}}}=
\chi^{\downarrow}_{43}=\chi^{\downarrow}_{23}=\chi^{\downarrow}_{12}=\chi^{\downarrow}_{14}$ and $ \chi^{\downarrow}_{4\textbf{\emph{1}}}=\chi^{\downarrow}_{3\textbf{\emph{2}}}=\chi^{\downarrow}_{3\textbf{\emph{4}}}=\chi^{\downarrow}_{2\textbf{\emph{1}}}=
\chi^{\uparrow}_{43}=\chi^{\uparrow}_{23}=\chi^{\uparrow}_{12}=\chi^{\uparrow}_{14}$ having a negligible difference, we use the average of them as parameter $\chi_{x,y}=\frac{1}{16}\sum_{\sigma i}\chi^{\sigma}_{ij=i+\vec{x},\vec{y}}, i=1..4$ for simplicity. Similarly, the average bond order on the diagonal links is denoted as $\chi_d=\frac{1}{4}\sum_{ \sigma}\chi^{\sigma}_{2,4}+\chi^{\sigma}_{3,\textbf{\emph{1}}}$. Although we use $\chi_{x,y}$ and $\chi_d$ as parameters in the plots, in all calculations the exact $\chi^{\sigma}_{ij}$ are used.
phase diagram at $T=0.001$ {#phasediagram}
==========================
![(Color Online) Doping dependent order parameters for fixed $t_2=0.3$ and four different $J_2$. (a) $\chi_{d}$ as function of $\delta$ for $t_1=\pm1.0$. (b) $\chi_{x,y}$ as function of $\delta$, the upper lines are for $t_1=1$ and the lower lines are for $t_1=-1.0$. (c) $|m|$ as function of $\delta$ for $t_1=\pm1.0$. []{data-label="fig4"}](F2.eps){width="8.5cm"}
![(Color Online) (a) Phase diagram of $t_1$-$t_2$-$J_1$-$J_2$ model on the $J_2/J_1$-$\delta$ plane for $t_2=0.3$. The blue solid line corresponds to the first order phase transition. (b) Phase diagram for $t_2=0.6$. The grey dashed line in $t_2=0.3(0.6)$ is the boundary of SDW-Ins phase for $t_2=0.6 (0.3)$. (c) DOS of the point $J_2=0.9J_1,\delta=0.1$ for $t_2=0.3,0.6$, respectively. (d) Energy dispersion of the quasi-flat band in the lower subset along the line of M-K-$\Gamma$-M for $ t_2=0.3$, $J_2=0.2J_1$. From down to up the curves corresponds to $\delta=0.01,0.02,0.04$, respectively.[]{data-label="fig5"}](F3a.eps "fig:"){width="4cm"} ![(Color Online) (a) Phase diagram of $t_1$-$t_2$-$J_1$-$J_2$ model on the $J_2/J_1$-$\delta$ plane for $t_2=0.3$. The blue solid line corresponds to the first order phase transition. (b) Phase diagram for $t_2=0.6$. The grey dashed line in $t_2=0.3(0.6)$ is the boundary of SDW-Ins phase for $t_2=0.6 (0.3)$. (c) DOS of the point $J_2=0.9J_1,\delta=0.1$ for $t_2=0.3,0.6$, respectively. (d) Energy dispersion of the quasi-flat band in the lower subset along the line of M-K-$\Gamma$-M for $ t_2=0.3$, $J_2=0.2J_1$. From down to up the curves corresponds to $\delta=0.01,0.02,0.04$, respectively.[]{data-label="fig5"}](F3b.eps "fig:"){width="4cm"} ![(Color Online) (a) Phase diagram of $t_1$-$t_2$-$J_1$-$J_2$ model on the $J_2/J_1$-$\delta$ plane for $t_2=0.3$. The blue solid line corresponds to the first order phase transition. (b) Phase diagram for $t_2=0.6$. The grey dashed line in $t_2=0.3(0.6)$ is the boundary of SDW-Ins phase for $t_2=0.6 (0.3)$. (c) DOS of the point $J_2=0.9J_1,\delta=0.1$ for $t_2=0.3,0.6$, respectively. (d) Energy dispersion of the quasi-flat band in the lower subset along the line of M-K-$\Gamma$-M for $ t_2=0.3$, $J_2=0.2J_1$. From down to up the curves corresponds to $\delta=0.01,0.02,0.04$, respectively.[]{data-label="fig5"}](F3c.eps "fig:"){width="4cm"} ![(Color Online) (a) Phase diagram of $t_1$-$t_2$-$J_1$-$J_2$ model on the $J_2/J_1$-$\delta$ plane for $t_2=0.3$. The blue solid line corresponds to the first order phase transition. (b) Phase diagram for $t_2=0.6$. The grey dashed line in $t_2=0.3(0.6)$ is the boundary of SDW-Ins phase for $t_2=0.6 (0.3)$. (c) DOS of the point $J_2=0.9J_1,\delta=0.1$ for $t_2=0.3,0.6$, respectively. (d) Energy dispersion of the quasi-flat band in the lower subset along the line of M-K-$\Gamma$-M for $ t_2=0.3$, $J_2=0.2J_1$. From down to up the curves corresponds to $\delta=0.01,0.02,0.04$, respectively.[]{data-label="fig5"}](F3d.eps "fig:"){width="4cm"}
In this section we will discuss the phase diagram of $t_1$-$t_2$-$J_1$-$J_2$ model at $T=0.001$. With fixed $t_2=0.3$, Fig. \[fig4\] depicts the doping dependent mean-field order parameters for four different values of $J_2$. It shows that the diagonal bond order $\chi_{d}$ and the magnetic order $|m|$ are equal for $t_1=\pm1$, whereas $\chi_{x,y}$ is equal in magnitude but has plus sign for $t_1=1$ and minus sign for $t_1=-1$. The sign change of the mean-field $\chi_{x,y}$ along the short bond order under a hopping parameter switch $t_1\rightarrow-t_1$ shows that the band structure remains unchanged for $t_1=\pm 1$, therefore the sign of $t_1$ is an irrelevant quantity in our investigations. The large $J_2$ prefers to enhance $|\chi|$ but suppress $|m|$, which is ascribed to the presence of competition between kinetic energy and magnetic order. From Fig. \[fig4\](c) we can see that for $J_2=0$, $|m|$ versus $\delta$ is a smooth curve, while for $J_2>0.45J_1$, $|m|$ drops to zero abruptly and corresponds to the occurrence of the first order phase transition. At zero doping all kinds of $\chi=0$, finite doping leads to finite $\chi$ ascribing to the characteristic of itinerate electrons.
![(Color Online) For $t_2=0.3$, the energy band dispersions along the high symmetric line M-K-$\Gamma$-M for the four selected set of $J_2$ and doping. The grey dashed lines denote the Fermi level. The inset of (c) is the corresponding Fermi surface topology. []{data-label="fig6"}](F4a.eps "fig:"){width="7.2cm"} ![(Color Online) For $t_2=0.3$, the energy band dispersions along the high symmetric line M-K-$\Gamma$-M for the four selected set of $J_2$ and doping. The grey dashed lines denote the Fermi level. The inset of (c) is the corresponding Fermi surface topology. []{data-label="fig6"}](F4b.eps "fig:"){width="7.2cm"}
Fig. \[fig5\](a) and (b) show the phase diagram on the $J_2/J_1$-$\delta$ plane for fixed $t_2=0.3$ and $0.6$, respectively. The remarkable feature is that the area of SDW-Ins phase almost remains the same for different values of $t_2$, while the area of SDW-M phase shrinks as increasing the $t_2$. The point $J_2=0.9J_1,\delta=0.1$ denoted by the hollow square locates in the SDW-M phase for $t_2=0.3$, where the corresponding DOS $\rho$ is finite at $\omega=0$ with two pronounced peaks located at the edge of the spin gap \[blue dotted line in Fig. \[fig5\](c)\], while it locates in PM-M phase for $t_2=0.6$, with a pronounced in-gap resonance peak arising \[red solid line in Fig. \[fig5\](c)\], suggesting a highly DOS due to the vanishing $|m|$.
The strong electron-electron interaction separates the energy bands into two subspaces at small dopings. For $J_2=0.2J_1$ and $t_2=0.3$, Fig. \[fig5\](d) shows the band dispersion along the line of high symmetric M-K-$\Gamma$-M for the quasi-flat band $E_{4,k}$, which is the top band in the lower subspace. We find that in the SDW-Ins phase, a finite distance between the low-energy bands and the Fermi level exists, and $E_{4,k}$ is much closer to the Fermi level for higher doping. Numerical calculations show that on the flat segment electron concentration is $\langle n_p(k)\rangle=0.968, 0.881, 0.579$ for $\delta=0.01, 0.02, 0.04$, which equals to the Fermi distribution function.
![(Color Online) The map of spectral weight of $\omega=0$ for $t_2=0.3$ and $J_2=0.9J_1$ in the $\mathbf{k}$ space with (a) corresponding to SDM-M phase at $\delta=0.06$ and (b) corresponding to PM-M phase at $\delta=0.14$. []{data-label="ferm1"}](F5.eps){width="8cm"}
Taking $t_2=0.3$ as an example, we show the energy band structure in the whole energy range in Fig. \[fig6\]. For an undoped case, the system displays an insulating phase. For $J_2=0$ and $\delta=0.02$, the system is in SDW-Ins phase with band structure looking like two straight lines. At $\delta=0.06$ \[see panel (b)\], the system is in SDW-M phase, and the eight bands are well separated with the flat segment of $E_{4,k}$ slightly above the Fermi level. As shown in Fig. \[fig6\] (b) and (c), the larger frustration $J_2$ will gradually suppress the height of the upper bands and diminish the SDW gap. The inset of Fig. \[fig6\] (c) shows the corresponding Fermi surface of the SDW-M phase. For $J_2=0.9J_1$ and $\delta=0.14$, the low and high energy bands will be touched, and the system will be driven into the PM-M phase with the eight bands degenerated into four bands. In all cases, the flat segment from M to $\Gamma$ remains the same features.
The distribution function of occupation numbers in the $\mathbf{k}$ space can be expressed as $n_k=\int d\omega A(k,\omega)f(\omega)$, where the spectral function $A(k,\omega)=-\frac{\delta}{\pi}\mathrm{Im} \frac{1}{4}\sum_i\mathcal{G}_i(\textbf{k},\mathrm{i}\omega)|_{\mathrm{i}\omega=\omega+\mathrm{i}0^{+}}$ is the imaginary part of the single-particle Green’s function multiplied by $-1/\pi$, giving the correlation of the electron creation and annihilation operations. For each sublattice $\mathcal{G}_i(\textbf{k},\mathrm{i}\omega)=\int^{\beta}_0 -T_{\tau}\langle f_{i k}(\tau)f^{\dag}_{i k}(0) \rangle e^{\mathrm{i}\omega\tau}$ is a function of transformation matrix and $\mathcal{G}^0_n(\textbf{k},\mathrm{i}\omega)=1/(\mathrm{i}\omega-E_{n,k}))$. For an interaction system single-particle Green’s function or $A(k,\omega)$ carries information about the underlying potential and no longer concentrated at single energy like that in non-interaction systems. Experimentally, the angle-resolved photoemission spectroscopy (ARPES) experiment is often used to detect the single-particle information and the Fermi surface topologies. To qualitatively compare with the ARPES measurements, the spectral weight is evaluated, which is obtained by integrating the spectral function $A(k,\omega)$ times the Fermi distribution function over an energy interval of $[-0.01,0.01]$ around $\omega$. The corresponding Fermi surface topology is set $\omega=0$ and shown in Fig. \[ferm1\]. In the SDW-Ins phase, the map of spectral weight is invisible, which is a typical feature of insulating state. For $t_2=0.3$, $J_2=0.9J_1$, and $\delta=0.06$, the system locates in the SDW-M phase, the corresponding map of spectral weight is shown in Fig. \[ferm1\](a). As doping increases to $\delta=0.14$, the system locates in the PM-M phase, the corresponding spectral weight is depicted in Fig. \[ferm1\](b), which is quite different to that of Fig. \[ferm1\](a).
At the end of this section, two remarks should be noted that the phase diagram on the plane of $t_2$-$\delta$ is similar to that on the plane of $J_2/J_1$-$\delta$ shown in Fig. \[fig5\]. For a finite hopping $t_3$, the area of SDW-Ins phase will shrink remarkably, and the detailed results are presented in the appendix.
temperature dependence calculations {#temperature_effects}
===================================
![(Color Online) For $t_2=0.3$ and $J_2=0$, the energy band dispersions along the high symmetric $\mathbf{k}$-point line M-K-$\Gamma$-M for different $T$ at $\delta=0.02$ and $\delta=0.06$, respectively. The violet dotted lines are the corresponding DOS for $\delta=0.02$, and the grey dashed lines denote the Fermi level.[]{data-label="fig8"}](F6a.eps "fig:"){width="8cm"} ![(Color Online) For $t_2=0.3$ and $J_2=0$, the energy band dispersions along the high symmetric $\mathbf{k}$-point line M-K-$\Gamma$-M for different $T$ at $\delta=0.02$ and $\delta=0.06$, respectively. The violet dotted lines are the corresponding DOS for $\delta=0.02$, and the grey dashed lines denote the Fermi level.[]{data-label="fig8"}](F6b.eps "fig:"){width="8cm"}
We then turn to discuss the temperature dependent properties using the parameters of $J_2=0$ and $t_2=0.3$. At sufficiently high temperature, $|m|$ will vanish, both the SDW-Ins and SDW-M phases will transit into the PM-M phase accompanying with the disappearance of $\Delta_{SDW}$ and $\Delta_{Fm}$. We find that all bands move downward as $T$ is increased from $0.001$ to the critical temperature of magnetic order $T_c(m)$, accompanying with enlargement of $|\Delta_{Fm}|$. Seen from Fig. \[fig8\], the lower subbands of SDW-Ins phase is deep-lying the Fermi level by increasing the temperature for $x=0.02$. For $\delta=0.06$, it is interesting to point out that the distinct Fermi surface of $T=0.001$ disappears ascribing to the increasing temperature. Besides the strong evidence from the presence of Fermi surface for metal, the electrical conductivity $\sigma_{xx}=D\delta(\omega)$ for zero frequency electric field is also evaluated numerically to solidify the nature of a metal, where $\delta(\omega)$ is the delta function.
![(Color Online) For $t_2=0.3$ and $J_2=0$, (a) the Drude weight $D(\delta)$ as function of $\delta$ for $T=0.001$ and $0.1$, respectively. (b) The Drude weight $D(T)$ as function of $T$ with fixed $\delta=0.02$.[]{data-label="fig10"}](F7.eps){width="7.0cm"}
For a single band system, the diamagnetic term determines the Drude weight. While for a multi-band system, the interplay among the bands below and above the Fermi level will induce a finite $\Pi_{xx}$ even at zero $T$ and is named as a geometric contribution. [@fla2] Fig. \[fig10\](a) depicts the $D(\delta)$ as function of doping for $T=0.001$ and $T=0.1$, respectively. In both cases the $D(\delta)$ monotonously increases as a function of doping, the higher doping levels correspond to the larger $D(\delta)$ and the more mobile charges. The Drude weight increases insignificantly at weak dopings, for $\delta=0.02$ it is $0.004$, and for $\delta=0.04$ it is $0.016$. For considerable large doping levels, the $|m|$ vanishes along with zero paramagnetic current due to the touching of energy subspaces, and the $D$ has a linear behavior versus doping. It is worth pointing out that the Drude weight is independent on the temperature at weak dopings since the two curves of $T=0.001$ and $0.1$ overlap for a small value of $\delta$.
Fig. \[fig10\](b) shows the $D(T)$ as function of $T$ for $\delta=0.02$. In the SDW-Ins phase, the mobile charge remains unchanged value at various **$T<0.15$**. As $T$ increasing up to $0.15$, the $D(T)$ jumps discontinuously and corresponds to a PM-M phase. Except for the half-filling, the $D$ is not exactly zero, instead of by a tiny value in the SDW-Ins phase.
![(Color Online) (a) The phase diagram in $T$-$\delta$ plane for the parameters of $t_2=0.3$ and $J_2=0$. States in the left side of the grey dashed line denote the absence of Fermi surface. (b) Critical temperature of bond order $T_c(\chi)$ in $T$-$\delta$ plane. (c) $\Delta_{SDW}$ and $|\Delta_{Fm}|$ as a function of $T$ at the high symmetric point $M$ of the energy band $E_{4,k}$ for $\delta=0.02$.[]{data-label="fig7"}](F7a.eps "fig:"){width="4.0cm"} ![(Color Online) (a) The phase diagram in $T$-$\delta$ plane for the parameters of $t_2=0.3$ and $J_2=0$. States in the left side of the grey dashed line denote the absence of Fermi surface. (b) Critical temperature of bond order $T_c(\chi)$ in $T$-$\delta$ plane. (c) $\Delta_{SDW}$ and $|\Delta_{Fm}|$ as a function of $T$ at the high symmetric point $M$ of the energy band $E_{4,k}$ for $\delta=0.02$.[]{data-label="fig7"}](F7b.eps "fig:"){width="4.0cm"}
Phase diagram in the plane of $T$ vs $\delta$ is shown in Fig. \[fig7\](a). As increasing the temperature $T$ from $0.001$, $|m|$ is obviously suppressed and approached zero at $T_c(m)$, whereas $\chi$ has negligible change in the range of \[$0.001$,$T_c(m)$\]. The system locates in the SDW-Ins phase for $x\leq0.04$ and $T<0.16$, with finite $\Delta_{SDW}$ and finite $\Delta_{Fm}$ as well as a small number of mobile charge. It will be transited into the PM-M phase when the temperature is beyond $T_c(m)$. If the doping level is larger than $0.04$, the system favors to stay in the SDW-M phase at low $T$, and then it will enter into the PM-M phase as the temperature is increased beyond $T_c(m)$, with large $\delta$ corresponding to low $T_c(m)$. In the SDW-M phase, the states located at the left side of the grey dashed line denote the absence of Fermi surface.
Fig. \[fig7\](b) plots the critical temperature of $\chi$ denoted by $T_c(\chi)$ in the plane of T and $\delta$, above the curve all $\chi^{\sigma}_{ij}$ are zero, which will not influence on the division of the phase diagram in panel (a), since $T_c(\chi)$ is much higher than that of $m$. The movement of boson bond order $\langle b^{\dag}_{i}b_{j}\rangle$ follows that of fermion bond order $\chi^{\sigma}_{ij}$, [@kp1; @kp2; @kp3] thus the Bosonic fluctuations only appear on a much higher energy scale and the Bonson condensation is assumed in our discussions.
Fig. \[fig7\](c) depicts the $\Delta_{SDW}(M)$ and $|\Delta_{Fm}(M)|$ as function of $T$ at the high symmetric point $M$ on the flat band $E_{4,k}$ with fixed $\delta=0.02$. One can see that the $\Delta_{SDW}(M)$ monotonously decreases as $T$ increases, while $|\Delta_{Fm}(M)|$ increases from $0.007$, and then drops as the system approaches the boundary of PM-M phase.
summary
=======
SSL can be realized in a group of synthesized compound, where the exhibiting antiferromagnetic metallic phase has been reported within a specified range of parameters, [@haidi] but the SDW-Ins phase defined in our discussions has rarely been mentioned since the flat-banded system has triggered intensive interests only recently by the realization of Lieb lattice. Based on SB mean-field theory we investigate the intriguing insulator-metal phase transition on SSL by using $t_1$-$t_2$-$J_1$-$J_2$ model. For the strongly correlated electron interactions, the lower bands and upper bands are well separated. SDW-Ins phase is resulted from the quasi-flat band in the range of $T=[0.001,0.15)$, $\delta\leq0.04$, displaying a finite $\Delta_{SDW}$ gap and a very tiny $D$ as well as the absence of Fermi surface. Only at half-filling, the Drude weight is exactly zero. The increasing of frustrated hopping $t_2$ and interaction $J_2$ almost are unaffected to the SDW-Ins phase.
The appearance of PM-M phase indicates the diminishing of magnetic order. The larger $t_2$ and $J_2$ as well as the larger doping will shrink the range of SDW-M phase since magnetic order is suppressed gradually. As $T$ is increased, the SDW-Ins phase and SDW-M phase will transit into the PM-M phase at the critical temperature. Although the Fermi surface of SDW-M phase will be immersed by higher $T$, due to the presence of the flat-band features, the Drude weight is robust against $T$.
Ascribing to the presence of the frustration and the special geometry of the lattice, SSL has a quasi-flat band, which localizes the electrons and leads to the appearance of SDW-Ins phase at low doping levels. The effects of doping, temperature, hopping and exchange coupling on the phase diagram of SSL are systemically studied, which provide a useful theoretical guidance for the understanding the nature of flat band and correlated insulating materials.
acknowledgements
================
We thank Prof. Yongping Zhang for discussions. This work was supported by the State Key Programs of China (Grant Nos. 2017YFA0304204 and 2016YFA0300504), the National Natural Science Foundation of China (Grant Nos. 51672171, 11625416, and 11774218), and the Natural Science Foundation from Jiangsu Province of China (Grant No. BK20160094). W.L. also acknowledges the start-up funding from Fudan University. The fund of the State Key Laboratory of Solidification Processing in NWPU (SKLSP201703), the supercomputing services from AM-HPC, and Fok Ying Tung education foundation are also acknowledged.
finite $t_3$ destroys SDW-Ins phase
====================================
![(Color Online) For $T=0.001$ and $t_2=0.3$, (a) the phase diagram on $J_2/J_1$-$\delta$ plane for $t_3=-0.1$. (b) Variation of the top band of the low-energy space along the high symmetric $\mathbf{k}$-point line M-K-$\Gamma$-M for $J_2=0$ and $\delta=0.02$, where we take the different $t_3=-0.01,-0.03,-0.1$.[]{data-label="fig11"}](FA2a.eps "fig:"){width="4cm"} ![(Color Online) For $T=0.001$ and $t_2=0.3$, (a) the phase diagram on $J_2/J_1$-$\delta$ plane for $t_3=-0.1$. (b) Variation of the top band of the low-energy space along the high symmetric $\mathbf{k}$-point line M-K-$\Gamma$-M for $J_2=0$ and $\delta=0.02$, where we take the different $t_3=-0.01,-0.03,-0.1$.[]{data-label="fig11"}](FA2b.eps "fig:"){width="4cm"}
![(Color Online) The map of spectral weight of $\omega=0$ in the $\mathbf{k}$ space for $T=0.001, t_2=0.3, J_2=0.9J_1, \delta=0.06$ with (a) $t_3=-0.1$ and (b) $t_3=0.1$.[]{data-label="fa3"}](FA3.eps){width="9cm"}
Since a finite $t_3$ changes the curvature of the quasi-flat band, it will affect on the SDW-Ins phase dramatically. Here we find that SDW-Ins phase almost vanishes as $t_3$ increased to $|t_3|=0.1$ at $T=0.001$, shown in Fig. \[fig11\](a). Fig. \[fig11\](b) shows the topmost band in the lower subspace for various $t_3$ at $\delta=0.02$. At a small value of $t_3=-0.01$, the flat segment from K to $\Gamma$ is sloped, and the symmetry between M-K and $\Gamma$-M is broken. As further increasing $|t_3|$, the $E_{4,k}$ band will be dispersive significantly. Thus we conclude that a tiny value of $t_3$ will shrink the area of SDW-Ins phase remarkably due to the disturbing flat band.
Although both the positive and minus $t_3$ will change the curvature of the flat bands, they have a different topology of Fermi surface. For the minus value of $t_3=-0.1$, the pockets of Fermi surface locate at around the center point of $\Gamma$, while for the positive value of $t_3=0.1$, the Fermi surface becomes disconnected sheets at around the four points of $(\pm \pi/2,\pm \pi/2)$ in the first Brillouin zone. The corresponding maps of the spectral weight also have quite different patterns, which are shown in Fig. \[fa3\].
electronic doping
==================
General speaking, SB theory is widely used in the hole-doped case where there is no double occupied site. In fact after particle-hole transformation, it can also be applied in electron-doped case.
Deviated from half-filling in electron-doped case, an electron can propagate from a double occupied site to a single occupied site $n_{i\bar{\sigma}}c^{\dag}_{i\sigma}c_{j\sigma}n_{j\bar{\sigma}}$, here $n_{j\bar{\sigma}}\approx \delta$ and $n_{i\bar{\sigma}}\approx 1$. After particle-hole transformation $c^{\dag}_{i\sigma}=c_{i\sigma}\xi_i$, the above expression can be rewritten as $-(1-n_{i\bar{\sigma}})c^{\dag}_{j\sigma}c_{i\sigma}(1-n_{j\bar{\sigma}})\xi_i\xi_j$, which looks like the form of hole-doped case. The sign of $\xi_i=\pm1$ depends on which sublattice $i$ belongs to. The four particle interaction Hamiltonian has the same form in the particle-hole transformation. Thus we can deal the electron- and hole- doped cases in the same manner as we choosing a special $\xi_i$.
On the mean-field described $t_1$-$t_2$-$J_1$-$J_2$ SSL, after the particle-hole transformation of $\xi_i=(-1)^i, i=1...4$, and taking $t_1=-t_1$ simultaneously, kinetic Hamiltonian transforms as $H_{t,k}\rightarrow -H^*_{t,k}$. The interaction Hamiltonian can also be changed into $ H^{\sigma}_{J,k}\rightarrow-H^{*\sigma}_{J,k}$, as long as mean-field $\chi$ changes as $\chi^{\sigma}_{i,j}\rightarrow\chi^{\sigma}_{i,j}$ on diagonal links, and $\chi^{\sigma}_{i,j}\rightarrow-\chi^{\sigma}_{i,j}$ on n.n. links. In addition, it can be proved that $H_k=H_k^*$ by changing $k_x\rightarrow-k_x,k_y\rightarrow-k_y$ in Eq.(\[1\]). Therefore particle-hole transformations along with minus sign of $t_1$ will transform $H_k\rightarrow -H_k$ in our discussions. Thus doping electrons actually means doping holes in our treatment, with the energy band being inverted $E_n\rightarrow -E_n$.
[99]{}
M. Indergand, C. Honerkamp, A. Läuchli, D. Poilblanc, and M. Sigrist 2007 Phys. Rev. B [**75**]{} 045105
H. Aoki, T. Sakakibara, H. Shishido, R. Settai, Y. Ōnuki, P Miranović, and K Machida 2004 J. Phys.: Condens. Matter [**16**]{} L13
K. Takada, H. Sakurai, E. Takayama-Muromachi, F. Lzumi, R. A. Dilanina, and T. Sasaki 2003 Nature [**422**]{} 53
R. Coldea, D. A. Tennant, A. M. Tsvelik, and Z. Tylczynski 2001 Phys. Rev. Lett. [**86**]{} 1335
H. X. Huang, Y. Q. Li, J. Y. Gan, Y. Chen, and F. C. Zhang 2007 Phys. Rev. B [**75**]{} 184523
S. Yonezawa, Y. Muraoka, Y. Matsushita, and Z. Hiroi 2004 J. Phys.: condens. Matter [**75**]{} L9
A. A. Lopes, B. A. Z. António, and R. G. Dias 2014 Phys. Rev. B [**89**]{} 235418
J. Ma, J. H. Lee, S. E. Hahn, Tao Hong, H. B. Cao, A. A. Aczel, Z. L. Dun, M. B. Stone, W. Tian, Y. Qiu, J. R. D. Copley, H. D. Zhou, R. S. Fishman, and M. Matsuda 2015 Phys. Rev. B [**91**]{} 020407(R)
A. Koga and N. Kawakami 2000 Phys. Rev. Lett. [**84**]{} 4461
W. Zheng, J. Oitmaa, and C. J. Hamer 2001 Phys. Rev. B [**65**]{} 014408
E. Müller-Hartmann, R. R. P. Singh, C. Knetter, and G. S. Uhrig 2000 Phys. Rev. Lett. [**84**]{} 1808
H. Kageyama, K. Yoshimura, R. Stern, N. V. Mushnikov, K. Onizuka, M. Kato, K. Kosuge, C. P. Slichter, T. Goto, and Y. Ueda 1999 Phys. Rev. Lett. [**82**]{} 3168
B. D. Gaulin, S. H. Lee, S. Haravifard, J. P. Castellan, A. J. Berlinsky, H. A. Dabkowska, Y. Qiu, and J. R. D. Copley 2004 Phys. Rev. lett. [**93**]{} 267202
M. Takigawa, S. Matsubara, M. Horvatić, C. Berthier, H. Kageyama, and Y. Ueda 2008 Phys. Rev. Lett. [**101**]{} 037202
M. S. Kim, M.C. Bennett, and M.C. Aronson 2008 Phys. Rev. B [**77**]{} 144425
M. S. Kim and M.C. Aronson 2013 Phys. Rev. Lett. [**110**]{} 017201
L. Su and P. Sengupta 2015 Phys. Rev. B [**92**]{} 165431
D. Carpentier and L. Balents 2001 Phys. Rev. B [**65**]{} 024427
C. H. Chung, J. B. Marston, and S. Sachdev 2001 Phys. Rev. B [**64**]{} 134407
A. Läuchli, S. Wessel, and M. Sigrist 2002 Phys. Rev. B [**66**]{} 014401
D. C. Ronquillo and M. R. Peterson 2014 Phys. Rev. B [**90**]{} 201108(R)
S. El Shawish and J. Bonca 2006 Phys. Rev. B [**74**]{} 174420
S. Haravifard, S. R. Dunsiger, S. El Shawish, B. D. Gaulin, H. A. Dabkowska, M. T. F. Telling, T. G. Perring, and J. Bonca 2006 Phys. Rev. Lett. [**97**]{} 247206
H.-D. Liu, Y.-H. Chen, H.-F. Lin, H.-S. Tao, and W.-M. Liu 2014 Sci. Rep. [**4**]{} 4829
B. S. Shastry and B. Sutherland 1981 Physica B+C [**108**]{} 1069
W. Zheng, J. Oitmaa, and C. J. Hamer 2001 Phys. Rev. B [**65**]{} 014408
M. Albrecht and F. Mila 1996 Europhysics Lett. [**34**]{} 145
P. Corboz and F. Mila 2013 Phys. Rev. B [**87**]{} 115144
Z. Wang and C. D. Batista 2018 arXiv:1802.09428
H. X. Huang, Y. Chen, Y. Gao, and G. H. Yang 2016 Physica C [**525-526**]{} 1
P. Farkašovský 2018 Eur. Phys. J. B. [**91:**]{} 74
H. Kageyama, K. Yoshimura, R. Stern, N. V. Mushnikov, K. Onizuka, M. Kato, K. Kosuge, C. P. Slichter, T. Goto, and Y. Ueda 1999 Phys. Rev. Lett. [**82**]{} 3168
K. Siemensmeyer, E. Wulf, H. J. Mikeska, K. Flachbart, S. Gabani, S. Matas, P. Priputen, A. Efdokimova, and N. Shitsevalova 2008 Phys. Rev. Lett. [**101**]{} 177201
J. Trinh, S. Mitra, C. Panagopoulos, T. Kong, P. C. Canfield, and A. P. Ramirez 2018 Phys. Rev. Lett. [**121**]{} 167203
C.-H. Chung and Y. B. Kim 2004 Phys. Rev. Lett. [**93**]{} 207004
A. Julku, S. Peotta, T. I. Vanhala, D.-H. Kim, and P. Törmä 2016 Phys. Rev. Lett. [**117**]{} 045303
L. Liang, T. I. Vanhala, S. Peotta, T. Siro, A. Harju, and P. Törmä 2017 Phys. Rev. B [**95**]{} 024515
H. Tasaki 1998 Prog. Theor. Phys. [**99**]{} 489
L. Santos, M. A. Baranov, J. I. Cirac, H. U. Everts, H. Fehrmann, and M. Lewenstein 2004 Phys. Rev. Lett. [**93**]{} 030601
Z. Li, J. Zhuang, L. Chen, L. Wang, H. Feng, X. Xu, X. Wang, C. Zhang, K. Wu, S. X. Dou, Z. Hu, and Y. Du 2017 arXiv:1708.04448
V. J. Kauppila, F. Aikebaier, and T. T. Heikkilä 2016 Phys. Rev. B [**93**]{} 214505
Y. Cao, V. Fatemi, A. Demir, S. Fang, S. L. Tomarken, J. Y. Luo, J. D. Sanchez-Yamagishi, K. Watanabe, T. Taniguchi, E. Kaxiras, R. C. Ashoori, and P. Jarillo-Herrero 2018 Nature [**556**]{} 80-84
A. Mezio, R. H. McKenzie 2017 Phys. Rev. B [**96**]{} 035121
C. T. Shih, T. K. Lee, R. Eder, C.-Y. Mou, and Y. C. Chen 2004 Phys. Rev. Lett. [**92**]{} 227002
G. Kotliar and A. E. Ruckenstein 1986 Phys. Rev. Lett. [**57**]{} 1362
Q. Yuan, Y. Chen, T. K. Lee, and C. S. Ting 2004 Phys. Rev. B [**69**]{} 214523
P. W. Leung and Y. F. Cheng 2004 Phys. Rev. B [**69**]{} 180403(R)
D. J. Scalapino, S. R. White, and S. C. Zhang 1993 Phys. Rev. B [**47**]{} 7995
D. J. Scalapino, S. R. White, and S. C. Zhang 1992 Phys. Rev. Lett. [**68**]{} 2830
H. Huang, Y. Gao, J.-X. Zhu, and C. S. Ting 2012 Phys. Rev. Lett. [**109**]{} 187007
T. Das, J. X. Zhu, and M. J. Graf 2011 Phys. Rev. B [**84**]{} 134510
T. Xiang and J. M. Wheatley 1995 Phys. Rev. B [**51**]{} 11721
T. Tohyama and S. Maekawa 1994 Phys. Rev. B [**49**]{} 3596
T. K. Lee, Chang-Ming Ho, and N. Nagaosa 2003 Phys. Rev. Lett [**90**]{} 067001
M. Inaba, H. Matsukawa, M. Saitoh, and H. Fukuyama 1996 Physica C [**257**]{} 299
M. U. Ubbens and P. A. Lee 1992 Phys. Rev. B [**46**]{} 8434
T.-H. Gimm, S.-S. Lee, S.-P. Hong, and S.-H. Suck Salk 1999 Phys. Rev. B [**60**]{} 6324
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'I. Pérez, J.A.L. Aguerri'
- 'J. Méndez-Abreu'
bibliography:
- 'patternhighz.bib'
date:
nocite:
- '[@barway2011]'
- '[@elmegreen2004; @vanderbergh2002]'
- '[@sheth2008]'
- '[@das2003; @buta2010]'
- '[@tremaine1984]'
- '[@weinberg1985]'
- '[@adelman2007]'
- '[@faure2008; @mobasher2007; @ilbert2006; @lilly2007]'
- '[@athanassoula1982]'
- '[@athanassoula2010]'
title: Bar pattern speed evolution over the last 7 Gyr
---
[The tumbling pattern of a bar is the main parameter characterising its dynamics. From numerical simulations, its evolution since bar formation is tightly linked to the dark halo in which the bar is formed through dynamical friction and angular momentum exchange. Observational measurements of the bar pattern speed with redshift can restrict models of galaxy formation and bar evolution. ]{} [We aim to determine, for the first time, the bar pattern speed evolution with redshift based on morphological measurements.]{} [We have selected a sample of 44 low inclination ringed galaxies from the SDSS and COSMOS surveys covering the redshift range 0$<z<$0.8 to investigate the evolution of the bar pattern speed. We have derived morphological ratios between the deprojected outer ring radius ($R_{ring}$) and the bar size ($R_{bar}$). This quantity is related to the parameter ${\cal R}=R_{\rm CR}/R_{bar}$ used for classifiying bars in slow and fast rotators, and allow us to investigate possible differences with redshift.]{} [We obtain a similar distribution of $R$ at all redshifts. We do not find any systematic effect that could be forcing this result.]{} [The results obtained here are compatible with both, the bulk of the bar population ($\sim 70\%$) being fast-rotators and no evolution of the pattern speed with redshift. We argue that if bars are long-lasting structures, the results presented here imply that there has not been a substantial angular momentum exchange between the bar and halo, as predicted by numerical simulations. In consequence, this might imply that the discs of these high surface-brightness galaxies are maximal.]{}
Introduction
============
Stellar bars are thought to be the main internal mechanism driving the dynamical and secular evolution of disc galaxies. They are able to modify the central structure and morphology of galaxies, mostly because they are responsible for a substantial redistribution of mass and angular momentum in the discs [e.g., @sellwood1981; @sellwood1993; @pfenniger1991; @athanassoula2003; @debattista2006]. In the last decade, numerical simulations have addressed the importance of the transference of angular momentum between baryonic and dark matter components [e.g., @debattista1998; @debattista2000]. The amount of angular momentum exchanged is related to the specific properties of the galaxies, such as the bar mass, halo density, and halo velocity dispersion [@debattista1998; @athanassoula2003; @sellwood2006] and it takes place mainly at the disc resonances [@athanassoula2003; @martinezvalpuesta2006]. Recent works have shown that gas fraction can also play an important role in the bar-halo interplay and therefore in the bar evolution [@bournaud2002; @romanodiaz2009; @villavargas2010]. Moreover, bars are efficient at funneling material toward the galaxy centre and possibly they influence the building of the stellar bulge [e.g., @kormendy2004] and the central black hole [e.g., @shlosman1989; @regan1999; @corsini2003]. Peanut/boxy bulges in galaxies are also thought to be associated with bending instabilities and bar vertical resonances [@bureaufreeman1999; @debattista2004; @debattista2006; @athanassoula2005; @martinezvalpuesta2006; @mendezabreu2008].
As already mentioned, the bar formation and destruction rate, as well as the morphological and dynamical changes suffered by the discs during their lifetime are strongly affected by the angular momentum exchange. Therefore, the cosmological evolution of the bar fraction can also depend on this effect. Observations show that bars in low redshift galaxies are ubiquitous, with a fraction of $\sim$45% at visual wavelengths [e.g., @marinova2007; @reese2007; @barazza2008; @aguerri2009; @masters2011] and rising to $\sim$70% in the near-infrared [@knapen2000; @eskridge2000; @menendezdelmestre2007]. The bar fraction depends on morphological type, being lower in lenticular galaxies than in spirals (Marinova & Jogee 2007; Aguerri et al. 2009; Nair & Abraham 2010; Barway et al. 2011; but see also Masters et al. 2011). Some recent results show that bar fraction is a strong function of galaxy mass [@mendezabreu2010; @nair2010] and color [@hoyle2011]. In contrast, bar fraction is only barely affected by the environment [@aguerri2009; @li2009; @mendezabreu2010].
The evolution of the bar fraction with redshift is still a matter of debate. @abraham1999 found that the fraction of barred galaxies at $z>0.5$ is lower than the local fraction. However, other authors claim that this may be the consequence of selection effects, due to the high angular resolution needed to find bars (Elmegreen et al. 2004; but see van den Bergh 2002). To deal with the angular resolution problem, several studies have carried out this analysis using the Advanced Camera for Surveys (ACS) on the Hubble Space Telescope (HST). Thus, Elmegreen et al. (2004) and Jogee et al. (2004) found the same bar fraction ($\sim$40%) at redshift $z=1.1$ as in the local Universe, suggesting that the bar dissolution cannot be common during a Hubble time unless the bar formation rate is comparable to the bar destruction rate. On the contrary, Sheth et al. (2008), in a recent study using images from the Cosmological Evolution Survey [COSMOS; @scoville2007] and using a larger sample than previous studies, found that the bar fraction at $z=0.84$ is one-third of the present-day value. They also found a much stronger evolution for low mass galaxies and late-type morphological types. Part of the differences may be due to the selection effects and other systematic effects that still need to be investigated further.
In any case, these results show that bars have been common structural components of the discs of galaxies during the last 8 Gyrs. The study of their origin and evolution could be crucial for understanding the galaxy evolution since $z=1$. This study can be done by analysing three parameters that characterise the bars: length, strength and pattern speed. Several methods and techniques have been proposed in order to measure these bar parameters. The bar length has been obtained directly by visual inspection on galaxy images [@kormendy1979; @martin1995; @mendezabreu2010; @nair2010; @masters2011], searching for the maximum ellipticity of the galaxy isophotes [@wozniak1995; @laine2002; @marinova2007; @aguerri2009], locating variations of the isophotal position angle [@sheth2003; @erwin2005], analyzing Fourier moments [@quillen1994; @aguerri2000; @aguerri2003], or by photometric decomposition of the surface-brightness profiles of galaxies [@prieto1997; @prieto2001; @aguerri2005; @laurikainen2005; @gadotti2008; @weinzirl2009; @gadotti2011]. The resulting studies reported that the typical bar length is about 3-4 kpc, and strongly correlates with the disc scale-length [@aguerri2005; @perez2005; @erwin2005; @marinova2007; @laurikainen2007]. Bar length is also a function of galaxy size, morphology and color [@aguerri2009; @hoyle2011].
The bar strength has been determined by measuring bar torques [@buta2001], isophotal ellipticity [@martinet1997; @aguerri1999; @whyte2002; @marinova2007], or Fourier modes [@otha1990; @aguerri2000; @laurikainen2005; @athanassoula2002b]. This parameter depends on galaxy morphology. Bars in lenticular galaxies are generally weaker than in spirals (Das et al. 2003; Laurikainen et al. 2007; Barazza et al. 2008; Aguerri et a. 2009; Buta et al. 2010).
The bar pattern speed, $\Omega_{\rm b}$, is the main kinematic observable and describes the dynamics of the bar. This tumbling pattern determines the position of the resonances in the disc and it is most usefully parametrised by a distance independent parameter ${\cal R}=R_{\rm CR}/R_{\rm bar}$, where $R_{\rm CR}$ is the Lagrangian/corotation radius, where the gravitational and centrifugal forces cancel out in the rest frame of the bar, and $R_{\rm bar}$ is the bar semi-major axis. Therefore, bars that end near corotation (1$<{\cal R}<$1.4) are considered fast, while shorter bars (${\cal
R}>$1.4) are commonly called slow. If ${\cal R}<1.0$ then orbits are elongated perpendicular to the bar, and self consistent bars cannot exist in this regime [@contopoulos1980]. The most reliable method for obtaining the location of corotation was that proposed by Tremaine & Weinberg (1984, hereafter TW method) which uses a set of simple kinematic measurements to derive the bar pattern speed assuming that the tracer obeys the continuity equation, that the discs are flat and that there is one well defined pattern speed. However, large integration times are required in medium-size telescopes to reach the high signal-to-noise required to apply the TW method. This limits its application to a small number of candidates. Despite the difficulties in obtaining bar pattern speeds, a reasonable number of nearby galaxies have been investigated [@merrifield1995; @debattista2002; @aguerri2003; @corsini2007] finding that all bars end near corotation. Some of these assumptions are not applicable for galaxies with nested bars, and there is now a simple extension of the TW method to multiple pattern speeds [@maciejwski2006; @corsini2003; @meidt2009] and the fact that some authors have shown that the TW method can be applied to CO [@rand2004; @zimmer2004] and H$_{\alpha}$ velocity fields [@hernandez2005; @emsellem2006; @fathi2007; @chemin2009; @gabbasov2009; @fathi2009] opens a new window to these studies.
Some indirect ways to derive the bar pattern speed include methods based on numerical modelling: generating either self-consistent models or models using potentials derived from the light distributions [@duval1983; @lindblad1996; @laine1998; @weiner2001; @perez2004; @zanmar2008] and then matching numerical experiments with the observed velocity fields; or by matching numerical simulations to the galaxy morphology [@hunter1988; @england1989; @laine1998; @aguerri2001; @rautiainen2005]. Other indirect methods to derive the bar pattern speed include identifying morphological or kinematic features with resonances: using a variety of features [@elmegreen1990]; the shape of dust lanes [@athanassoula1992]; the sign inversion of the radial streaming motion across corotation [@canzian1993]; rings as resonance indicators [@buta1986; @buta1995]; phase-shift between the potential and density wave patterns [@zhang2007]; location of minimum of star formation [@cepa1990; @aguerri2000]; or comparison of the behaviour of the phase Fourier angle in blue and near-infrared images [@puerari1997; @aguerri1998]. Although possibly the most accurate indirect method to calculate pattern speeds is the the comparison of gas velocities to those obtained in numerical simulations that use a potential obtained from optical or near-infrared light, it is also very time consuming and can only be applied to a relatively small number of objects.
The technique to determine the bar pattern speed based on connecting the location of rings to orbital resonances was introduced by Buta (1986). It is based on the theoretical work presented by Schwarz in a series of papers [@schwarz1981; @schwarz1984a; @schwarz1984b] showing how these ring structures appear near the dynamical Lindblad resonances due to a bar–like perturbation. To directly apply this method to find the specific value of the pattern speed not only the location of the ring and the association to a resonance is required, but some kinematic information is also needed. However, we can use the ${\cal R}$ parametrisation of the bar introduced previously, and determine the ratio between the outer ring radius (linked to the outer Lindblad resonance, OLR) and the bar length. In this way, we can indirectly determine, not the pattern speed, but whether the bars measured are in the slow or fast regime.
The bar parameters discussed above have been analysed in local galaxy samples. There are no previous studies in the literature about the evolution of the length, strength and pattern speed of bars. In this article, we study for the first time, the dynamical evolution of bars with redshift, going from the local Universe to $z\sim0.8$. We use a well selected sample of barred galaxies with outer rings to exploit the power of this method. The study of the dynamical evolution of bars is critical to constrain the angular momentum exchange between the disc and the halo and their subsequent evolution. Weinberg (1985) predicted that a bar would lose angular momentum due to a massive dark matter halo through dynamical friction, slowing down in the process. This prediction was further confirmed in numerical simulations [@debattista1998; @debattista2000; @athanassoula2003; @sellwood2006] where they found that bars are slowed efficiently if a substantial density of dark matter is present in the region of the bar. On the other hand, if the mass distribution is dominated by the stellar disc, then the bar remains rapidly rotating for a long time. We show in this work that bars do not show a systematic change in their dynamical state in the last $\sim7$ Gyrs.
The article is organised as follows: we present the sample selection and morphology discussion in Sect. \[sample\]. We describe the method followed to measure the ring and bar radius in Sect. \[olr\]. The results are presented in Sect. \[results\] and we discuss their implications in Sect. \[discussion\]. Conclusions are provided in Sect. \[conclusions\]. Throughout the paper the cosmological parameters used are: $H_{0}=70$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\Lambda}=0.7$, and $\Omega_{m}=0.3$.
Sample selection {#sample}
================
The galaxy samples studied in this article were extracted from two different surveys: low redshift galaxies were taken from the Sloan Digital Sky Survey (SDSS; $0.01 < z < 0.04$), and high redshift galaxies were selected from COSMOS ($0.125 < z < 0.75$).
Two caveats must be discussed before the samples are described in detail: first, it is worth noticing that the galaxy samples are not meant to be complete in any sense, however, the selection criteria make the two samples fully comparable. Second, in Aguerri et al. (2009) we studied how the resolution of the SDSS images can affect our detection of bars. We worked out, using artificial galaxies, that the shortest bars that we are able to resolve have a length of $\sim9$ pixels. Considering a mean PSF in our SDSS images with a FWHM of 109 (2.77 pixel), we conclude that we resolve bars larger than $\sim3\times$ FWHM, or equivalently, $\sim0.5$ kpc at $z=0.01$ and $\sim2$ kpc at $z=0.04$. The COSMOS sample was selected using the ACS data in the F814W filter. The images were processed to a resolution of 005 pixel$^{-1}$ with an averaged PSF FWHM of 0097 [@scoville2007; @koekemoer2007]. Based on the previous considerations, we will resolve bars larger than $\sim3\times$FWHM which corresponds to $\sim0.6$ kpc at $z=0.125$ and $\sim2.2$ kpc at $z=0.75$, matching perfectly the SDSS spatial resolution in the low redshift range.
Outer Ring Morphological Classification {#OLRmorph}
---------------------------------------
The ring morphological classification used in this study is based on the work of [@butacrocker1991]. They divide the outer rings in three main morphological classes resembling the rings developed in numerical simulations near the OLR [@schwarz1981]. The first class, called R$_{1}'$, is characterised by a 180$^{\circ}$ winding of the spiral arms with respect to the ends of a bar. The second type is known as an R$_{2}'$ ring. It is defined by a 270$^{\circ}$ winding of the outer arms with respect to the bar ends, so that in two opposing quadrants the arm pattern is doubled. The R$_{1}'$ and R$_{2}'$ morphologies were predicted by Schwarz (1981) as the kind of patterns that would be expected near the OLR in a barred galaxy. The third class is referred to in Buta & Crocker (1991) as the R$_{1}$R$_{2}'$ morphology, where the outer arms break not from the ends of the bar, but from an R$_{1}'$-type ring. The existence of this combined type, which may be linked to the population of both main families of OLR periodic orbits [@schwarz1981], provides some of the clearest evidence of the OLR in barred galaxy morphology. Some examples of this classification, taken from our sample of low and high redshift galaxies, are shown in Fig. \[fig:ringtype\]. Buta et al. (1995) derived the distribution of intrinsic axis ratios for the outer rings using the Catalog of Southern Ringed Galaxies. They found that outer rings present in barred galaxies are intrinsically elliptical with an axis ratio $\sim0.82\pm0.07$, and that the intrinsic ellipticity varies from the R$_{1}'$ ($\sim0.74\pm0.08$) to the R$_{2}'$ ($\sim0.87\pm0.08$). The intrinsic shape of the rings plays an important role when deprojecting distances such as the bar length and the ring radius, thus, more intrinsically elliptical rings will increase the uncertainties in the measurements. We decided to remove from our samples the R$_{1}'$ type of rings, and keep only the R$_{2}'$ types since they are intrinsically rounder. In fact, their intrinsic shape is very similar to that of typical discs [@fasano1993; @ryden2004].
Low redshift {#lowz}
------------
The barred ringed galaxies at low redshift were obtained from the galaxy sample analysed in Aguerri et al. (2009). They selected a volume limited sample of galaxies from the spectroscopic catalogue of the SDSS Data Release 5 (SDSS-DR5, Adelman-McCarthy et al. 2007). This sample covers the redshift range $0.01<z<0.04$, down to an absolute magnitude of $M_{r}<-20$, and with low inclination $i<60^{\circ}$. The full sample consist of 3060 galaxies with a morphological mix of 26% ellipticals, 29% lenticulars, 20% early type spirals, and 25% late type spirals. Galaxies were classified in barred and unbarred systems by searching for absolute maxima in the ellipticity radial profiles of their isophotes (see Aguerri et al. 2009 for details). From the barred sample we visually inspected the SDSS galaxy images in order to look for the presence of outer rings of type R$_{2}'$.
We obtained a total of 18 barred galaxies with suitable outer rings features. Table \[samplesdss.tab\] shows the main properties of the bars and rings features measured in these galaxies.
High redshift
-------------
As for the low redshift sample, we have chosen a number of low inclination galaxies from the third release of the COSMOS HST survey [@scoville2007]. We first downloaded all the 81 image tiles from the COSMOS/ACS fields, observed using the F814W ($I$-band) filter, from the Multimission Archive at STScI (MAST[^1]) for visual inspection. These observations cover $\sim$ 2 deg$^{2}$ with a pixel scale (for the drizzled data) of 005 pixel$^{-1}$.
We visually scanned the COSMOS fields to look for clearly ringed barred galaxies. After a preliminary list was created, we correlated the positions with the spectroscopic redshifts from the Very Large Telescope (VLT) and Magellan COSMOS spectroscopic surveys (zCOSMOS Survey; Lilly et al. 2007) to search for candidates with reliable redshifts. We searched also for photometric redshifts for the remaining candidates. We used photometric redshifts determined by Faure et al. (2008) using the Le Phare photometric redshift estimation code (Ilbert et al. 2006), details concerning the multi-wavelength photometry can be found in Mobasher et al. (2007). Faure et al. (2008) used 1095 spectroscopic redshifts from the zCOSMOS Survey (Lilly et al. 2006) to calibrate the ground-based photometric zero points. Using eight bands, this method achieves a photometric redshift accuracy of $\sigma_{\Delta
z}$/(1+$z_{s}$)=0.031.
From this sample, only galaxies showing type R$_{2}'$ ring were included in the sample. Finally the high redshift sample consists of 26 galaxies. The main properties of the bars and rings features are shown in Table \[samplez.tab\]. The sample covers the redshift range $0.125<z<0.75$. Figure \[fig:histo\] shows the redshift distribution.
------------------------- ------------------- ------------------- ---------------- ----------------------- ------------------ ---------------- -------
Name R$_{\rm bar-min}$ R$_{\rm bar-max}$ PA$_{\rm bar}$ $\epsilon_{\rm ring}$ PA$_{\rm ring} $ R$_{\rm ring}$ $z$
(kpc) (kpc) (degrees) (degrees) (kpc)
(1) (2) (3) (4) (5) (6) (7) (8)
SDSSJ104924.86-002547.5 6.6 5.1 121.6$\pm$0.9 0.402$\pm$0.004 142.2$\pm$0.8 13.0 0.039
SDSSJ102543.29+393846.9 7.3 5.1 153.4$\pm$0.2 0.075$\pm$0.005 74.8$\pm$2.3 13.0 0.023
SDSSJ122529.23+471623.4 5.7 4.7 11.6$\pm$0.1 0.174$\pm$0.005 30.1$\pm$14.3 10.8 0.025
SDSSJ130235.73+411924.1 4.8 3.7 165.4$\pm$0.7 0.186$\pm$0.005 45.4$\pm$1.5 9.9 0.028
SDSSJ120732.62+324846.7 7.1 4.7 54.4$\pm$0.5 0.084$\pm$0.007 74.1$\pm$20.0 11.6 0.026
SDSSJ133259.13+321913.6 5.7 3.8 177.3$\pm$0.3 0.044$\pm$0.007 106.7$\pm$8.6 8.7 0.035
SDSSJ012858.63-005656.3 8.0 6.3 6.2$\pm$0.3 0.341$\pm$0.002 83.6$\pm$0.4 16.4 0.018
SDSSJ083220.43+412132.0 3.5 2.4 52.8$\pm$0.5 0.048$\pm$0.005 111.3$\pm$10.9 7.9 0.025
SDSSJ083630.84+040215.6 6.6 4.5 94.2$\pm$0.2 0.164$\pm$0.009 37.0$\pm$3.8 12.7 0.029
SDSSJ091426.23+360644.1 6.3 5.0 164.4$\pm$0.0 0.163$\pm$0.031 156.1$\pm$1.0 10.1 0.022
SDSSJ123234.57+492312.2 5.3 3.4 176.1$\pm$0.5 0.074$\pm$0.012 114.1$\pm$10.7 8.7 0.040
SDSSJ142412.12+350846.0 3.6 2.7 45.3$\pm$1.0 0.238$\pm$0.006 52.9$\pm$26.7 9.3 0.029
SDSSJ153619.30+493428.3 5.3 2.9 17.8$\pm$2.1 0.211$\pm$0.006 73.1$\pm$0.7 9.7 0.038
SDSSJ160331.62+492017.3 8.7 5.2 70.9$\pm$0.4 0.255$\pm$0.005 32.3$\pm$7.2 16.9 0.020
SDSSJ172721.89+593837.6 5.9 4.0 171.3$\pm$0.1 0.202$\pm$0.004 148.8$\pm$1.1 11.4 0.028
SDSSJ123313.69+121449.2 3.9 3.1 118.5$\pm$0.9 0.023$\pm$0.004 65.3$\pm$11.5 6.9 0.026
SDSSJ120609.11-025653.2 5.3 3.9 61.5$\pm$0.7 0.149$\pm$0.009 17.3$\pm$4.6 11.2 0.026
SDSSJ111044.88+043039.0 6.8 5.0 97.6$\pm$0.2 0.062$\pm$0.007 64.6$\pm$9.2 12.5 0.029
------------------------- ------------------- ------------------- ---------------- ----------------------- ------------------ ---------------- -------
NOTE. Col. (1): Galaxy name from SDSS; Col. (2): bar radius calculated using the position of the minimum ellipticity; Col. (3): bar radius calculated using the position of the maximum ellipticity; Col. (4): position angle of the bar; Col. (5): ring ellipticity; Col. (6): position angle of the ring; Col. (7): ring radius; Col. (8): spectroscopic redshift from SDSS
--------------------- ------------------- ------------------- ---------------- ----------------------- ----------------- ---------------- -------------
Name R$_{\rm bar-min}$ R$_{\rm bar-max}$ PA$_{\rm bar}$ $\epsilon_{\rm ring}$ PA$_{\rm ring}$ R$_{\rm ring}$ $z$
(kpc) (kpc) (degrees) (degrees) (kpc)
(1) (2) (3) (4) (5) (6) (7) (8)
812947 5.4 4.1 76.3$\pm$0.1 0.068$\pm$0.010 120.5$\pm$14.8 9.8 0.125 ($s$)
816960 5.5 3.4 174.8$\pm$1.2 0.059$\pm$0.035 85.2$\pm$22.6 7.5 0.311 ($s$)
J095928.30+020109.0 5.3 3.8 63.2$\pm$0.6 0.088$\pm$0.045 141.3$\pm$41.8 9.1 0.530 ($p$)
817887 5.6 4.6 116.2$\pm$2.8 0.056$\pm$0.028 88.4$\pm$17.5 15.1 0.672 ($s$)
823705 3.6 2.7 46.4$\pm$0.7 0.165$\pm$0.028 74.4$\pm$17.2 6.3 0.491 ($s$)
824759 3.7 2.9 136.1$\pm$1.8 0.161$\pm$0.023 158.1$\pm$ 7.0 7.0 0.751 ($s$)
825492 5.1 3.6 108.4$\pm$2.4 0.152$\pm$0.057 108.1$\pm$22.2 8.7 0.736 ($s$)
833039 4.3 3.3 100.6$\pm$1.1 0.203$\pm$0.056 32.8$\pm$ 4.8 8.8 0.360 ($s$)
J100233.98+022524.3 4.7 2.2 172.7$\pm$0.9 0.167$\pm$0.015 63.1$\pm$ 4.4 9.4 0.720 ($p$)
J095938.81+020658.7 10.3 7.6 111.7$\pm$0.4 0.071$\pm$0.035 149.6$\pm$75.9 16.6 0.409 ($p$)
J095935.08+020127.2 4.3 3.0 176.8$\pm$79.2 0.158$\pm$0.044 26.3$\pm$36.9 10.3 0.357 ($p$)
J100204.95+022739.7 5.8 3.7 124.4$\pm$1.0 0.116$\pm$0.019 122.7$\pm$ 6.6 10.5 0.507 ($p$)
841055 5.7 3.4 134.9$\pm$0.6 0.188$\pm$0.020 178.0$\pm$56.2 13.7 0.376 ($s$)
J095759.45+022810.5 4.4 2.6 20.6$\pm$0.9 0.042$\pm$0.013 131.0$\pm$23.0 9.2 0.119 ($s$)
851598 6.4 4.4 9.6$\pm$0.5 0.080$\pm$0.047 102.2$\pm$30.7 10.0 0.346 ($s$)
852495 6.8 5.7 96.7 $\pm$3.1 0.265$\pm$0.015 21.9$\pm$ 1.5 12.2 0.705 ($s$)
852636 7.1 4.9 0.7$\pm$80.0 0.040$\pm$0.026 125.6$\pm$50.7 12.5 0.345 ($s$)
852155 7.7 5.6 173.0$\pm$0.3 0.306$\pm$0.009 160.4$\pm$ 0.8 15.3 0.305 ($s$)
J100254.88+024645.8 4.4 3.2 18.7$\pm$1.5 0.065$\pm$0.036 128.0$\pm$45.7 8.5 0.468 ($p$)
840577 5.1 3.5 68.3$\pm$2.4 0.085$\pm$0.023 89.5$\pm$15.0 7.9 0.539 ($s$)
838743 4.4 2.8 48.7$\pm$0.5 0.144$\pm$0.025 60.0$\pm$3.8 8.9 0.126 ($s$)
830974 9.6 5.3 11.5$\pm$1.4 0.149$\pm$0.025 57.7$\pm$7.0 13.9 0.695 ($s$)
811921 4.9 3.1 25.2$\pm$1.6 0.068$\pm$0.027 89.5$\pm$30.7 8.2 0.371 ($s$)
813153 7.5 6.0 60.0$\pm$0.7 0.159$\pm$0.069 47.3$\pm$28.1 13.2 0.529 ($s$)
831775 3.6 1.8 125.1$\pm$0.1 0.140$\pm$0.028 156.8$\pm$76.5 6.2 0.381 ($s$)
J100217.12+023024.1 4.9 3.4 157.8$\pm$0.8 0.083$\pm$0.022 97.7$\pm$15.6 8.1 0.379 ($p$)
--------------------- ------------------- ------------------- ---------------- ----------------------- ----------------- ---------------- -------------
NOTE. Col. (1): Galaxy name from COSMOS; Col. (2): bar radius calculated using the position of the minimum ellipticity; Col. (3): bar radius calculated using the position of the maximum ellipticity; Col. (4): position angle of the bar; Col. (5): ring ellipticity; Col. (6): position angle of the ring; Col. (7): ring radius; Col. (8): ($s$): spectroscopic redshift from zCOSMOS; ($p$): photometric redshift from Faure et al. (2008).
Outer ring radius and bar size definition {#olr}
=========================================
Our approach to quantifying the dynamical state of the bars in our sample of ringed galaxies is based on the measurements of both the bar and ring radius. We deproject bar size in the plane of the galaxy using $i$, the galaxy inclination, and $\theta$, the position angle of the galaxy component (bar or ring). We assume that the outer ring reflects the properties of the disc, and therefore that the ellipticity and position angle of the ring and disc are the same. In Sect. \[OLRmorph\] we have discussed that both components are intrinsically similar and possible differences will only affect our results by introducing a large scatter. Under this hypothesis, the ring radius does not need to be deprojected, since it is measured along the major axis of the galaxy, and the galaxy inclination can be derived simply by $i=\arccos{(1-\epsilon_{\rm ring})}$.
Ellipse fitting {#ellipse.fitting}
----------------
The low redshift sample measurements were derived by using the ellipticity and position angle radial profiles extracted from the symmetrised images. This approach allows us to clean the images from spurious sources. It works as follows: each image is rotated 180$^{\circ}$ with respect to the galaxy centre. Then, we subtract the rotated frame from the original one. The residual image was sigma-clipped to identify all the pixels with a number of counts lower than $1\sigma$, where $\sigma$ is the r.m.s. of the image background. The value of the deviant pixels was set to zero. Finally, the cleaned image was subtracted from the original one to get the symmetrised image. The ellipses were then fitted to the isophotes of the symmetrised images of the galaxies using the IRAF[^2] task [ELLIPSE]{} [@jedrzejewski1987]. We used an iterative wrapped procedure which runs the ellipse fitting several times, changing the trial values at each fit iteration, until a good fit at all radii is obtained. At each fixed semi-major axis length, the coordinates of the centre of the fitting ellipse were kept fixed. This centre was identified with the position of the central intensity peak. The trial values for the ellipticity and position angle were randomly chosen between 0 and 1 and between $-90^\circ$ and $90^\circ$, respectively. The fitting procedure stopped when either convergence was reached or after 100 iterations.
The high redshift sample ellipticity and position angle profiles were derived using the same wrapping procedure to maximise the goodness of the ellipse fitting. However, in this case we preferred not to symmetrise the images but apply a $2\times2$ pixels box smoothing. The image symmetrisation was not needed since the sample galaxies were not contaminated by other sources within their projected surface and the smoothing provided better radial profiles by improving the signal-to-noise ratio of the images.
Bar length {#bar.length}
----------
The ellipticity radial profile of a typical barred galaxy reaches a minimum at the centre, because of either seeing effects or a spherical bulge. Then, it usually shows a global increase to a local maximum, and then suddenly decreases towards a minimum at the location where the isophotes become axisymmetric in the face-on case. The position angle profile is also characteristic in barred galaxies, being constant in the bar region and then changing to fit the outer disc orientation [e.g., @wozniak1995; @aguerri2000b]. These characteristic profiles are produced by the shape and orientation of the stellar orbits of the bar [see @contopoulos1989; @athanassoula1992]. Different methods have been used to measure the bar length based on the ellipticity and position angle radial profiles [see @athanassoula2002; @micheldansac2006]. However, the solution is always ambiguous and it can lead to misleading results. In order to remove these uncertainties we decided to measure the bar length as the midpoint between the radius of the maximum and minimum ellipticity. These two different measurements of the bar length represent the extreme cases [@micheldansac2006] and therefore they represent an upper limit of our errors in the bar radius measurements. We preferred this solution not to bias our conclusions. The position angle of the bar, which is needed to deproject the bar length, was measured at the position of the maximum ellipticity, so we avoid problems related to position angle variation in the bar-disc region. An example of this method applied to three of our low and high redshift galaxies is shown in Figs. \[fig:method1\], \[fig:method2\], respectively.
Ring radius
-----------
As for the bar component, the ring radius was derived based on the ellipticity and position angle radial profiles. In the ring region, we expect that the ellipticity and position angle radial profiles will remain constant due to the stellar orbits in the ring. Therefore, we identify the region of the profile where the ring is present and we measured the ring radius as the position where the ellipticity and position angle become constants. The ring ellipticity and position angle also needed for deprojecting, were derived as a mean of these constant values. The error in the ring radius has been calculated by comparing the estimated ring radius with the radius at which the ellipticity varies more than 3 times the standard deviation of the disc ellipticity. Figures \[appendix1\] and \[appendix2\] from the Appendix show all the galaxies with the ring radius overplotted.
Results
=======
Table \[samplesdss.tab\] and Table \[samplez.tab\] shows the obtained parameters for the ring radius, ellipticity, position angle and the bar semi-major axis, as derived in Sect. \[ellipse.fitting\]. Most galaxy inclinations lie below $i<40^{\circ}$. The bar size range, using the maximum ellipticity, covers from 2.5 to 6.3 kpc. Most of the bars in the local Universe (about 70$\%$, see Aguerri et al. 2009) are within this bar size range. Similar values of the bar size range are found for our high redshift galaxies. The mean bar radius of our low and high redshift galaxies are 4.5$\pm$1.04 and 3.5$\pm$1.33 kpc, respectively. This means that within the errors both galaxy samples have similar bars according to their lengths and similar to local samples of barred galaxies (see Aguerri et al. 2009). The average bar size, using the minimum ellipticity, for our low and high redshift galaxies are also similar: 5.9$\pm$1.43 and 5.3$\pm$1.70 kpc, respectively. Thus, both samples of galaxies show similar bar sizes independent of the method used for determining the bar length. It has been argued [@micheldansac2006] that the sizes calculated using the minimum of the ellipticity correlate well with the position of corotation, giving a more physically significant size than measurements obtained with the maximum of the ellipticity, which clearly underestimates the true bar size. To avoid problems related to the bar size calculation, as explained in Sect. \[bar.length\], we have opted for using the mid-point and to take into account the values of R$_{\rm bar}$ using both methods to obtain the errors.
We have determined the strength of the bars for the low and high redshift galaxies by using the maximum ellipticity of the bar (see Aguerri et al. 2009). Both samples cover the same range of bar strengths. Thus, the mean values of the bar strength of our low and high redshift samples are: 0.20$\pm$0.07, and 0.17$\pm$0.05. These values are similar to the mean strength of bars in the local Universe (0.20$\pm$0.07; see Aguerri et al. 2009). We can conclude that according to the size and strength of the bars, our low and high redshift galaxy samples have similar bars as those found in a complete local sample of barred galaxies (see Aguerri et al. 2009).
To determine whether our galaxies are in the [*fast*]{} or [*slow*]{} range (see Sect.\[introduction\]) we define the ratio ${\cal R}_{\rm ring}$= R$_{\rm ring}$/R$_{\rm bar}$, where R$_{\rm ring}$ is the ring radius and R$_{bar}$, is the bar semi-major axis, as characterised in Sect. \[olr\]. Because we cover this ratio for galaxies with redshifts between $0 < z < 0.8$, we can study possible changes of this ratio with redshift. Figure \[fig:rbar.rolr\] shows the distribution of ring radii (R$_{\rm ring}$) vs. the bar semi-major axis (R$_{\rm bar}$) for the whole sample. We consider a fast bar those bars for which the R$_{\rm CR}$/R$_{\rm bar}$ ratio lies between 1.0 and 1.4. This ratio has been plotted in Fig. \[fig:rbar.rolr\] for both values and is calculated using linear resonance theory and a flat rotation curve (Athanassoula et al. 1982), in this case the position of the OLR (i.e., the ring radius) and the CR are related in the following way:
$$(\frac{R_{\rm ring}}{R_{\rm CR}})^\delta= 1+(1-\frac{1}{2}\delta)^{1/2}
%\frac{r_{\rm sky}}{\cos{i}}\, \sqrt{\sin^2{\theta} + \cos^2{\theta} \cos^2{i}},
\label{eq:depro}$$
where $\delta$ lies between 0.7 and 1.0 for early type discs, see Athanassoula et al. (1982). We take $\delta$=1.0 in Fig. \[fig:rbar.rolr\] for simplicity but this choice does not alter the results. It is clear from Fig. \[fig:rbar.rolr\] that all the galaxies, independent of their redshift bin, fall into the fast-bar category.
We have investigated the influence of the inclination of the galaxies on this result. Thus, the average values for ${\cal R}_{\rm ring}$ at different inclinations are the following: i) for the low redshift sample and b/a $>$ 0.9, the average ${\cal R}_{\rm ring}$ = 0.51$\pm$ 0.06, for b/a $<$ 0.9, the average ${\cal R}_{\rm ring}$ = 0.50 $\pm$ 0.08 ii) for the high-redshift sample; for b/a $>$ 0.9, the average ${\cal R}_{\rm ring}$ = 0.55 $\pm$ 0.09 and for b/a $<$ 0.9, the average ${\cal R}_{\rm ring}$ = 0.52 $\pm$ 0.12. All the values, independently of redshift and inclination are comparable; and therefore, we do not see changes of this ratio with redshift.
Discussion \[discussion\]
=========================
Possible caveats
----------------
The tightness of the results shown in Fig. \[fig:rbar.rolr\] is somewhat unexpected considering the intrinsic uncertainties inherent to the measurements used in this work. For instance, we have assumed that the outer ring are perfectly circular, which is critical for the deprojection of the ring and bar lengths. In Sect. \[olr\] we justified the assumption of roundness for the R$_{2}'$ and R$_{1}$R$_{2}'$ ring morphologies. In addition, we have taken into account the limits of intrinsic axis ratios given in [@buta1995] to calculate the errors in the projected sizes, and we have demonstrated in the previous section that our result does not depend on the inclination of the galaxies. Therefore we conclude that projection effects are not biasing our results.
Another possible caveat to our result might be that the choice of ring galaxies biases the sample towards a certain pattern speed domain. However, numerical simulations [@byrd1994] have shown that resonant outer rings can be present in both fast and slow bars. All types of ring morphologies are found at different pattern speeds. In the same work, all types of ring morphologies also developed for different bar strengths. Athanassoula et al. (2010), who presented a new theory for ring and spiral formation, argues that there is a connection between the bar strength and the morphology of the rings. Nevertheless, R$_{2}'$ rings, as are those selected in this work, are located in barred galaxies with similar bar strengths as our galaxies (see Athanassoula et al. 2010). In addition, they show that R$_{2}'$ type rings can be formed in galaxies with fast and slow bars.
Although the bar size as measured in different rest-frame band passes could be different, it has been recently shown (Gadotti 2011) that in fact the difference in bar size is negligible and therefore we are not introducing a bias by measuring the bars a high-redshift near the $g$-band rest-frame while the low-redshift bar sizes are measured from the SDSS $r-$band.
Our low and high redshift samples are by no means complete. Therefore, it is customary to investigate whether this fact is affecting the results presented in Fig. \[fig:rbar.rolr\]. Since our low and high redshift barred galaxies are similar in size and strength we could be biasing the resulting pattern speeds toward a particular regime. In other words, the fact that we have not observed evolution in the pattern speed could be just due to the selection of similar fast bars. However, we know from a study of a complete sample of local barred galaxies (Aguerri et al. 2009), that only 30$\%$ of the local bars show larger lengths than our ringed barred galaxies. Studies of high redshift bars, 0.4 $<$ z $>$0.8, [@jogee2004; @barazza2009] have shown that the bar size distribution is similar to that of local galaxies and; therefore, as discussed before, similar to the bars size range of the bars presented in this work. Therefore, we do not seem to be looking at any special type of bar by analysing ringed galaxies. As previously mentioned, from numerical models, bars get longer and slower as they age. We can then set a 30$\%$ upper limit to the bars that could have suffered a change in their pattern speed in the last 7 Gyr, assuming that nearby long bars are the end-products of the evolution of fast bars. The remaining 70$\%$ of bars did not substantially lose angular momentum to the halo, maintaining their pattern speed. This discussion might be related to an implicit morphological bias, since it remains, even for local galaxies, to derive the pattern speed of bars in very late-type gaseous rich spirals which might suffer an intrinsically different evolution (e.g., Bournaud and Combes 2002).
Comparison with the results from numerical modelling
----------------------------------------------------
A recent numerical work [@villavargas2010] shows that the evolution of the pattern speed and bar-growth of a bar embedded in a live dark matter halo depends strongly on the gas content. In their simulations, a fixed fraction of the total mass was converted to gas mass, and the evolution of the bar parameters is then followed in time. The presence of gas changes the evolution of both the bar growth and the pattern speed evolution, the addition of gas can stop, or even speed-up, the pattern speed of the bar with time. The bar size is anti-correlated with the disc gas fractions. These gas-rich galaxies would be related to early-type galaxies because the gas leads to larger central mass concentration and therefore larger bulges. The results we present in this paper could be in agreement with these gas-rich models. However, the full picture is still unclear since it is observed that longer bars reside in late-type galaxies (e.g., Erwin 2005) which is against the model predictions. Furthermore, we should then explain why all the galaxies should have started with similar gas fractions in their discs.
There is also the possibility that the bars that we see at $z\sim0.8$ do not survive till the present and therefore, we do not see evolution because the time-scales involved in the formation and destruction are too short. It has been discussed [see @pfenniger1990; @bournaud2002; @bournaud2005] that gas-rich bars, i.e., late-type spirals, are short lived, with lifetimes of 1-2 Gyr. This short time scale would mimic a lack of evolution of long-lived bars; however, the galaxies in the sample show morphologies typical of early type spirals and there is evidence, from stellar population studies [@perez2009; @sanchezblazquez2011], that bars in early type galaxies are long-lived. If this is the case, and most of the ring galaxies we observe present long-lasting bars, it would imply that bars cannot have grown in time and kept being in the fast speed regime without increasing significantly in size. Therefore, the fact that we see the ring radius and the bar size covering the same size range at all redshift, and moreover large bars at high-redshift, implies that bars do not grow significantly in size with time.
The result shown in Fig. \[fig:rbar.rolr\] implies that bars have not evolved considerably, neither in size nor in pattern speed, since around the time when the Universe was half its present age. Most numerical simulations obtain bars that evolve with time, getting longer and stronger while slowing down (e.g., Debattista &Sellwood 1998; Athanassoula 2003). This effect is mostly due to the angular momentum exchange of the bar-disc system with the dark matter halo. Thus, the fact that bars are compatible with fast rotators at all redshifts indicates that the angular momentum exchange between the bar and halo has not been important enough in the last 7 Gyr to slow down bars. If the pattern speed can be used to set constrains to the halo-to-disc mass ratio, these results might imply that the discs in the high surface-brightness galaxies of our sample are maximal.
Summary and conclusions {#conclusions}
=======================
We have analysed 44 low inclination ringed galaxies spanning a redshift range between $0 < z < 0.8$ to study the possible evolution of the pattern speed in the last 7 Gyrs.
We calculated for each galaxy a morphological parameter indicative of the dynamical state of their bars. In particular, we derived whether they are fast or slow rotators. We find that the bar pattern speed does not seem to change with redshift and that all bars are compatible with being fast bars.
If the bars analysed are long-lasting, their size and bar strength have not significantly changed in time. The fact that, independent of the redshift, the bars are fast rotators and their size has not significantly changed in time could have also large implication for bar evolution models that mostly predict a bar growth with time. It has been argued that the exchange of angular momentum with a centrally dense halo causes the bar to evolve; however the present results might imply that the disc in the high surface-brightness galaxies is maximal and the central mass density is dominated by the stellar component which would lower the angular momentum exchange between the disc and the halo and slow down the bar evolution (e.g., Debattista & Sellwood, 2000, but for a different conclusion, see Athanassoula, 2003).
This is the first time that the pattern speed evolution has been investigated from the observational point of view. The results presented here place strong constrains on the bar evolution models.
We thank the referee for his/her useful comments. We would like to thank Victor Debattista for the careful reading of the manuscript. I.P. was supported by the Spanish Ministry of Science and Innovation (MICINN) (via grants AYA2010-21322-C03-02, AYA2010-21322-C03-03, AYA2007-67625-C02-02 and Consolider-Ingenio CSD2010-00064) and by the Junta de Andalucía (FQM-108). JALA and JMA were supported by the projects AYA2010-21887-C04-04 and by the Consolider-Ingenio 2010 Program grant CSD2006-00070.
Sample galaxies
===============
[^1]: Based on observations made with the NASA/ESA Hubble Space Telescope, obtained from the data archive at the Space Telescope Science Institute. STScI is operated by the Association of Universities for Research in Astronomy, Inc. under NASA contract NAS 5-26555.
[^2]: IRAF is distributed by NOAO, which is operated by AURA Inc., under contract with the National Science Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We challenge claims that the principle of maximum entropy production produces physical phenomenological relations between conjugate currents and forces, even beyond the linear regime, and that currents in networks arrange themselves to maximize entropy production as the system approaches the steady state. In particular: (1) we show that Ziegler’s principle of thermodynamic orthogonality leads to stringent reciprocal relations for higher order response coefficients, and in the framework of stochastic thermodynamics, we exhibit a simple explicit model that does not satisfy them; (2) on a network, enforcing Kirchhoff’s current law, we show that maximization of the entropy production prescribes reciprocal relations between coarse-grained observables, but is not responsible for the onset of the steady state, which is, rather, due to the minimum entropy production principle.'
author:
- 'Matteo Polettini, University of Luxembourg [^1]'
title: 'Fact-Checking Ziegler’s Maximum Entropy Production Principle beyond the Linear Regime and towards Steady States'
---
[^1]: Address: Campus Limpertsberg, 162a avenue de la Faïencerie, L-1511 Luxembourg. E-Mail: [email protected].
| {
"pile_set_name": "ArXiv"
} |
${}$\
[*S. Jordan* ]{} and [*R. Loll*]{}
Radboud University Nijmegen,\
Institute for Mathematics, Astrophysics and Particle Physics,\
Heyendaalseweg 135, NL-6525 AJ Nijmegen, The Netherlands.\
\
[**Abstract**]{}
We present a detailed analysis of a recently introduced version of Causal Dynamical Triangulations (CDT) that does not rely on a distinguished time slicing. Focussing on the case of 2+1 spacetime dimensions, we analyze its geometric and causal properties, present details of the numerical set-up and explain how to extract “volume profiles". Extensive Monte Carlo measurements of the system show the emergence of a de Sitter universe on large scales from the underlying quantum ensemble, similar to what was observed previously in standard CDT quantum gravity. This provides evidence that the distinguished time slicing of the latter is not an essential part of its kinematical set-up.
Introduction
============
First attempts in the 1990s to define quantum gravity nonperturbatively with the help of Dynamical Triangulations (DT) were based on an intrinsically Euclidean path integral, whose configuration space consists of four-dimensional, curved Riemannian spaces. DT works with a regularized version of this space in terms of triangulations, piecewise flat spaces of positive definite metric signature. Its elementary building block is a four-simplex, a generalization to four dimensions of a triangle (two-simplex) and a tetrahedron (three-simplex). An individual building block is a piece of flat four-dimensional Euclidean space and therefore does not carry any curvature. However, numerical investigations of the nonperturbative dynamics of DT quantum gravity found that it has neither a large-scale limit compatible with general relativity [@Ambjorn:1991pq; @Agishtein:1992xx; @Catterall:1994pg], nor a second-order phase transition allowing for a continuum limit in the sense of lattice quantum field theory [@Bialas:1996wu; @deBakker:1996zx]. Despite the negative nature of this result, the fact that the model contains criteria which could be used for its falsification should be appreciated. For many other candidate theories of quantum gravity this is not obviously the case.
Causal Dynamical Triangulations (CDT) were introduced in [@al] in an attempt to overcome these problems. The key new idea of CDT is to incorporate aspects of the [*causal structure*]{} of classical general relativity at a more fundamental level into the nonperturbative gravitational path integral.[^1] The elementary building blocks of CDT quantum gravity are flat four-simplices of [*Lorentzian*]{} signature, that is, pieces of Minkowski space. The carrier space of the corresponding path integral consists of piecewise flat simplicial manifolds assembled from these building blocks. In addition, each path integral history has a distinguished discrete foliation and an associated notion of (discrete) proper time, which ensures the presence of a well-defined causal structure globally (see [@cdtreviews; @physrep] for details on motivation, construction and results in CDT). Numerical simulations in 3+1 dimensions have shown that these modifications lead to a completely different quantum dynamics, compared to the earlier Euclidean DT model: CDT quantum gravity in 3+1 dimensions contains a phase whose nonperturbative ground state of geometry is extended, macroscopically four-dimensional and on large scales can be matched to a de Sitter universe [@cdt4d; @desitter]. Moreover, the theory has recently been shown to possess a second-order phase transition, which according to standard arguments is a prerequisite for the existence of a continuum limit [@trans].
The causal structure of CDT is realized by putting together its simplicial building blocks such that each CDT configuration has a product structure, not just at the level of topology – usually chosen as $[0,1]\times{}^{(3)}\Sigma$, for fixed three-topology ${}^{(3)}\Sigma$ – but [*as triangulations*]{}. A (3+1)-dimensional CDT geometry consists of a sequence of slabs or layers, each of thickness 1, which may be thought of as a single unit of proper time. The orientation of the light cones of all four-simplices in a given slab is consistent with this notion of time. In this way the causal structure becomes linked to a preferred discrete foliation of spacetime.
To understand better how the preferred time foliation on the one hand and the causal structure on the other contribute to the evidence of a good classical limit – the key distinguishing feature of CDT quantum gravity – it would be highly desirable to disentangle these two elements of “background structure". We recently proposed a modification of standard CDT which does exactly that [@jl]. The main idea, applicable in any spacetime dimension $d$, is to enlarge the set of $d$-simplices by new types of simplicial building blocks, also pieces of $d$-dimensional Minkowski space, but with different link type assignments and therefore a different orientation of the light cone relative to the boundaries of the simplex. Including the new building blocks is in general not compatible with the preferred foliation of CDT. Nevertheless, with a suitable choice of gluing rules one can still obtain Lorentzian simplicial manifolds with a well-defined causal structure, at least locally. In this way the issue of causality becomes dissociated from the notion of a preferred foliation. The interesting question is then whether the quantum-gravitational model using “nonfoliated CDT" can reproduce the results of the standard formulation, especially those concerning the large-scale properties of quantum spacetime. The main conclusion of the present paper, previously announced in [@jl], is that it [*can*]{}, at least in space-time dimension 2+1. At least in higher dimensions, this result appears to weaken the potential link of CDT quantum gravity with Hořava-Lifshitz gravity [@hl; @phasediagram], where the presence of a preferred foliation is a key ingredient.[^2]
To summarize, our intention is to get rid of the [*distinguished*]{} foliation (and associated discrete time label $t$) of CDT, whose leaves for integer $t$ coincide with simplicial spatial hypermanifolds consisting entirely of [*space*]{}like subsimplices of codimension 1. This does not mean that the nonfoliated CDT configurations cannot in principle be foliated with respect to some continuous notion of time, but only that there is in general no canonical way of doing this in terms of a distinguished substructure of the triangulated spacetimes.
Of course, having [*a*]{} notion of time at the level of the regularized geometries is important. In the standard formulation of CDT, we get a notion of (discrete proper) time $t$ for free, simply by counting consecutive “slabs", as described earlier. The presence of this time label allows us to define a transfer matrix and an associated propagator with the correct behaviour under composition (whose continuum limit in dimension 1+1 can be computed analytically [@al]), and prove a version of reflection positivity [@3d4d]. Yet another advantage of having an explicit time variable is that we can construct observables like the volume profile, which measures the distribution of spatial volume as a function of time. The analysis of these volume profiles in 3+1 dimensions has been crucial in relating the large-scale behaviour of CDT to a de Sitter cosmology in the continuum [@desitter; @semiclassical].[^3]
By contrast, in the enlarged CDT set-up we will be considering, typical triangulations will be more complicated, in the sense that the purely spatial subsimplices of codimension 1 will no longer align themselves into a neat sequence of simplicial hyper[*manifolds*]{}, but instead will form branching structures, as will be explained in more detail below. This also implies that we no longer have a distinguished time variable at our disposal. Nevertheless, as we shall demonstrate for the nontrivial case of 2+1 dimensions, it is possible to construct a meaningful time variable, whose restriction to standard, nonbranching CDT configurations agrees with the usual proper time label $t$. This will enable us to extract volume profiles from the numerical simulations and compare their dynamics with that of the standard formulation.
There are a number of good reasons for beginning our investigation in 2+1 dimensions, as we are doing in the present work. In standard CDT, the large-scale properties of the quantum universe generated by the nonperturbative quantum dynamics are qualitatively very similar to those in 3+1 dimensions, and are well described by a (three-dimensional) Euclidean de Sitter universe [@3dcdt; @benedettihenson]. Furthermore, as we will describe in Sec. \[sec:numsetup\], the nonfoliated CDT model requires new Monte Carlo moves, which are significantly more difficult to implement than the generalized Pachner moves used in standard CDT. Writing the simulation software for the new model becomes a very challenging task already in 2+1 dimensions. At the same time, the increased complexity of the simulation software leads to longer simulation times. It is not even clear currently whether analogous simulations in 3+1 dimensions could be performed with acceptable running times using contemporary simulation hardware.
Before embarking on our exploration of the dynamics of nonfoliated CDT, let us comment briefly on prior work which considered explicitly a possible relaxation of CDT’s strict time slicing.[^4] A soft way of relaxing the foliation in 1+1 dimensions was studied in [@markopoulousmolin], where under certain conditions the timelike links were allowed to have varying length. In this approach the foliation is still present, since the connectivity of the underlying triangulation is unchanged, but the individual leaves of the foliation are not placed at equidistant intervals. The authors argued that this should not affect the continuum limit of the model. A similar idea in 2+1 dimensions was considered in [@konopka], where it was also suggested to add new elementary building blocks to CDT, which extend over two slabs of the foliation instead of one. This study did not include details of how the path integral of the corresponding generalization of CDT quantum gravity should be formulated or simulated.
The remainder of this article is structured as follows. We begin with a brief review of Causal Dynamical Triangulations in Sec. \[cdt\]. In Sec. \[sec:ncdt\_2d\] we discuss aspects of CDT without preferred foliation in 1+1 dimensions, to illustrate the basic geometric idea behind the enlarged model. Sec. \[sec:nfct21\] contains a detailed study of the kinematical aspects of nonfoliated CDT in 2+1 dimensions. Sec. \[sec:actions\] deals with actions and the Wick rotation, and Sec. \[sec:numsetup\] summarizes the new numerical set-up. This includes an overview of the Monte Carlo moves in 2+1 dimensions, a prescription for how a notion of time can be defined on the quantum ensemble, and how the corresponding volume profiles can be extracted. In Sec. \[sec:explo\] we explore the phase diagram of the model numerically. In Sec. \[sec:tetdist\] we study distributions of tetrahedra as a function of the coupling constants, which allows us to understand how foliated (in the sense of regular CDT) typical configurations are. The results of our analysis of the volume profiles and their matching to a de Sitter universe are presented in Sec. \[sec:voldist\], and our conclusions in Sec. \[sec:conclusions\]. – Further details on the numerical implementation as well as a documentation of the relevant software can be found in [@thesis].
Review of Causal Dynamical Triangulations {#cdt}
=========================================
To set the stage for our subsequent generalization, we recall in the following some elements of Causal Dynamical Triangulations, mainly based on the original literature on CDT in 1+1 [@al], 2+1 [@3dcdt] and 3+1 dimensions [@cdt4d; @desitter]. For more extensive reviews and lecture notes on CDT we refer the interested reader to [@cdtreviews; @physrep; @cdtlectures].
The central object of interest in the CDT approach to quantum gravity is the gravitational path integral, which in the continuum can be written formally as $$\label{eq:pigrav}
Z(G,\Lambda)=\hspace{-.6cm}\mathop{\int}_{\mathrm{geometries\, [g]}}\hspace{-.7cm}{\mathcal{D}[g]}\exp(iS_{\mathrm{EH}}[g]),$$ where $S_{\mathrm{EH}}[g]$ is the Einstein-Hilbert action, written as functional of the metric $g$, $\mathcal{D}[g]$ is a measure on the space of geometries (the space of equivalence classes $[g]$ of metrics under the action of the diffeomorphism group), $G$ is Newton’s constant and $\Lambda$ is the cosmological constant. To define the path integral properly it needs to be regularized, which in CDT is done by performing the “sum over histories" (\[eq:pigrav\]) over a set of piecewise flat, simplicial Lorentzian geometries – in other words, triangulations – effectively discretizing the curvature degrees of freedom of spacetime. The way in which triangulations encode curvature is illustrated best in two dimensions. In the Euclidean plane, consider a flat disc consisting of six equilateral triangles which share a central vertex, and remove one of the triangles (Fig. \[fig:reggecurvature\], left). By identifying the opposite sides of the gap thus created, the piecewise flat disc acquires nontrivial (positive Gaussian) curvature, whose magnitude is equal to the deficit angle $\pi/3$ at the vertex (Fig. \[fig:reggecurvature\], right). This also coincides with the rotation angle undergone by a two-dimensional vector parallel-transported around the vertex, and is therefore an [*intrinsic*]{} property of the two-dimensional disc. The principle of encoding curvature through deficit angles (located at subsimplices of dimension $d-2$ in a $d$-dimensional triangulation) works in any dimension and for any metric signature.
Let us review the difference between the (Euclidean) DT and the (Lorentzian) CDT path integral, again for simplicity in dimension two. Recall that the geometric properties of a flat triangle (or, in higher dimensions, a flat simplex) are completely determined by the lengths of its edges.[^5], including the [*signature*]{} of the flat metric in the building block’s interior. In the Euclidean DT approach in two dimensions one works with a single type of building block, an equilateral triangle of Euclidean signature, all of whose edges have some fixed spacelike length, the same length for all triangles in the triangulation. As we have seen above, the number of such triangles around each interior vertex of the triangulation characterizes the local curvature at that vertex. In this way the formal integral over curved geometries in the continuum path integral (\[eq:pigrav\]) is turned into a sum over triangulations, typically subject to some manifold conditions, which ensure that the triangulation looks like a two-dimensional space everywhere.
By contrast, the flat triangles used in CDT quantum gravity have Lorentzian signature, something that cannot be achieved for equilateral triangles. The standard choice of an elementary CDT building block in 1+1 dimensions is one whose base has squared length $\ell_s^2>0$ (and therefore is spacelike) and whose remaining two edges both have squared length $\ell_t^2<0$ (and therefore are timelike). From the point of view of triangulating Lorentzian spacetimes, one could in principle have chosen $\ell_t^2$ to be space- or lightlike, but then CDT’s prescription of a Wick rotation – to be described below – would no longer be applicable.
Fig. \[fig:dt2d\] illustrates the fundamental building blocks of two-dimensional DT (left) and CDT (right), and how they are put together to obtain piecewise flat manifolds of Euclidean and Lorentzian signature. The two-dimensional graphs correctly represent the neighbourhood relations of adjacent triangles, but not the actual length assignments, since it is impossible to flatten out a curved surface and preserve the edge lengths at the same time. Note that in the Lorentzian case the spacelike edges form a sequence of one-dimensional simplicial submanifolds, which can be interpreted as hypermanifolds of constant time $t=0,1,2, ...$, endowing each triangulation with a distinguished notion of (proper) time. This time can be extended continuously to the interior of all triangles [@dl].
The ensemble of spacetimes forming the carrier space of the CDT path integral are all triangulations which consist of a fixed number $t_{tot}$ of triangulated strips $\Delta t=1$, where each strip is an arbitrary sequence of up- and down-triangles between times $t$ and $t+1$. The topology of space (usually chosen to be $S^1$) is not allowed to change in time, that is, branchings into multiple $S^1$-universes are forbidden. It was shown in [@benedettihensonmatrix] that the global foliation of a 1+1 dimensional CDT spacetime into such strips can be understood as consequence of a [*local*]{} regularity condition, namely, that precisely two spacelike edges be incident on any vertex. Note that these geometries are causally well behaved and obey a piecewise linear analogue of global hyperbolicity.
As already described in the introduction, the idea of creating individual path integral configurations with a well-behaved causal structure by imposing a preferred foliation on the underlying simplicial manifold is also realized in CDT in higher dimensions. Fig. \[fig:foliation\] shows part of a 2+1 dimensional CDT spacetime. Each leaf at integer-$t$ of the foliation forms a two-dimensional triangulation of the same fixed topology ${}^{(2)}\Sigma$, and consists of equilateral spacelike triangles with link length $\ell_s$. Adjacent triangulated spatial hypermanifolds are connected using Lorentzian tetrahedra to form a 2+1-dimensional simplicial manifold with spacetime topology $[0,1]\times {}^{(2)}\Sigma$, which again is a sequence of $t_{tot}$ triangulated “slabs" $\Delta t=1$. Each such geometry comes with a time label $t$ which can be defined continuously throughout the triangulation, and for integer values coincides with the discrete labeling of the simplicial leaves just described [@dl]. For reasons of simplicity, in the simulations time is often compactified. The topology then becomes $S^1\times {}^{(2)}\Sigma$.
The fundamental building blocks of three-dimensional CDT quantum gravity are flat Minkowskian tetrahedra, whose geometric properties are determined by their edge lengths. As in 1+1 dimensions, one considers two different link lengths, one spacelike with squared length $\ell_s^2$ and one timelike with $\ell_t^2=-\alpha \ell_s^2$. Without loss of generality we have introduced here a positive constant $\alpha$, which quantifies the relative magnitude of space- and timelike edge lengths and in what follows will be referred to as the *asymmetry parameter*.
CDT path integral configurations are assembled from two different tetrahedron types, which can be distinguished by their orientation with respect to the preferred foliation. As can be seen from Fig. \[fig:foliation\], spacelike edges of a tetrahedron are always contained in a spatial submanifold of integer $t$, whereas timelike edges always connect different spatial slices. The (3,1)-tetrahedron has three spacelike links forming a spacelike triangle, while the (2,2)-tetrahedron contains only two spacelike links. The notation $(i,j)$ indicates that $i$ vertices of the tetrahedron are located on one spatial slice and the remaining $j$ vertices on an adjacent one.
The kinematical set-up of CDT quantum gravity in 3+1 dimensions can be defined in a similar way. The leaves of the preferred foliation are three-dimensional Euclidean triangulations of fixed topology, and neighbouring slices are connected using Minkowskian four-simplices. Path integral configurations are simplicial manifolds assembled from two types of these building blocks, denoted by (4,1) and (3,2), depending on how their vertices are distributed among adjacent spatial slices. By labeling the foliation with increasing integers we again get a time variable for free, with every vertex being assigned a definite discrete time label.
Another ingredient that needs to be specified to make the model complete is an implementation on piecewise flat geometries of the Einstein-Hilbert action in the path integral (\[eq:pigrav\]), which in CDT quantum gravity is done following Regge’s prescription [@regge]. One should point out that models of the type we are studying tend to be very robust with respect to changes in the precise form of the action (which obviously is subject to discretization ambiguities) and of the configuration space, in the sense that a wide range of different regularizations and kinematical ingredients will lead to the same continuum physics, if the latter can be defined meaningfully. An exception to this is of course the imposition of causality constraints, which distinguishes CDT from DT quantum gravity and leads in all dimensions studied so far to genuinely different continuum results. In two dimensions, this can be demonstrated exactly, for example, by comparing specific observables and critical exponents, since the CDT model can be solved analytically [@al]. In higher dimensions, information about the behaviour of observables comes primarily from numerical simulations: three-dimensional CDT quantum gravity has only been solved partially and for restricted classes of triangulations [@blz], while in four dimensions analytical methods are mostly unavailable and one must resort to Monte Carlo simulations to extract physical results.
In order to analyze the dynamics of CDT quantum gravity using such simulations, a Wick rotation must be performed to convert the complex path integral amplitudes to real Boltzmann weights. This can be achieved by performing an analytic continuation of the asymmetry parameter, by rotating it in the lower half of the complex plane such that $\alpha$ is mapped to $-\alpha$ [@3d4d]. As a consequence, the gravitational path integral becomes a statistical partition function of the form $$\label{partition}
Z=\sum_{T\in \cal{C}}\frac{1}{C(T)}\exp(-S_{\mathrm{Regge}}^{\mathrm{eucl}}(T)),$$ where $\cal{C}$ is the space of all causal, Lorentzian triangulations $T$, $S_{\mathrm{Regge}}^{\mathrm{eucl}}$ the Euclideanized Regge action and $1/C(T)$ the discrete analog of the path integral measure, with $C(T)$ denoting the order of the automorphism group of $T$. In Sec. \[sec:actions\] below we will derive and discuss the explicit functional form of the Regge implementation of the three-dimensional Einstein-Hilbert action in terms of the triangulation data and the coupling constants of the nonfoliated CDT model.
Relaxing the foliation: 1+1 dimensions {#sec:ncdt_2d}
======================================
As a warm-up for the three-dimensional case, we will in this section illustrate our general strategy for relaxing the distinguished foliation by discussing the situation in 1+1 dimensions. The key idea is to add new elementary Minkowskian building blocks, while sticking to two types of links, one spacelike and one timelike, where we will continue to use the notation $\ell_s^2$ and $\ell_t^2\! =\!-\alpha \ell_s^2$ for their squared lengths. Fig. \[fig:fourtriangles\] shows the four types of triangles which can be built from these two link types. By calculating the metric inside the triangles one finds that there are exactly two which have Lorentzian signature $(-+)$, the two at the centre of the figure. We will call them the $sst$- and the $tts$-triangle respectively, in reference to the spacelike ($s$) and timelike ($t$) edges they contain. Note that in standard CDT in two dimensions only the $tts$-triangle is used.
The use of Lorentzian building blocks is not a sufficient condition for the triangulation to have a well-defined causal structure locally, we also need to check that we obtain well-defined light cones in points where triangles are glued together. If the gluing happens according to the standard rule of only identifying links of the same type (spacelike with spacelike, timelike with timelike), the only local causality violations[^6] can occur at the vertices of the triangulation. Counting past and future light cones separately, the point is that one may obtain more or fewer than the required two light cones at a vertex, as illustrated by the local neighbourhoods depicted in Fig. \[fig:causality2d\]. Local causality implies crossing exactly two light cones when going around a vertex or, equivalently, crossing exactly four lightlike lines emanating radially from the vertex.
As a next step, we will consider the global causal structure of individual triangulated manifolds satisfying the local vertex causality condition everywhere. This global structure will in general depend on the chosen topology. For simplicity, we will restrict ourselves to the cases $[0,1]\times S^1$ (space is a compact circle) and $[0,1]\times [0,1]$ (space is a closed interval), where the initial and final boundary are assumed to be spacelike, and any other boundaries (in the second case) timelike. We will call such a spacetime globally causal if it does not contain any closed timelike curves. A “timelike curve" for our purposes will be a sequence of oriented, timelike links in a triangulation. Since the interior of every Minkowskian triangle has a well-defined light cone structure, a choice of orientation (i.e. a choice of which one is the past and which one the future light cone) induces an orientation on its timelike edges, which can be captured by drawing a future-pointing arrow onto the edge. Conversely, these arrow assignments fix the orientation of the triangle uniquely. To follow the “flow of time" in a triangulation it is convenient to also associate a future-pointing arrow with each spacelike edge, which is drawn perpendicular to the edge, see Fig. \[fig:flowoftime2d\].
Choosing a consistent orientation for all building blocks in standard CDT in this way is completely straightforward: each triangle sits in a strip between discrete times $t$ and $t+1$, which fixes its orientation uniquely. Independent of the spatial boundary conditions, there are no closed timelike curves (unless we impose periodic boundary conditions [*in time*]{}, which trivially makes any timelike curve closed, a situation we are not considering here).
By contrast, the situation in nonfoliated CDT is slightly more involved. Given a time-oriented triangle, the orientation of a neighbouring triangle that shares an edge with the first one is uniquely determined by consistency. It is easy to see that when vertex causality is violated (like in the example of Fig. \[fig:causality2d\], left), inductively assigning orientations in this way will fail – i.e. lead to contradictions – even for a local vertex neighbourhood. If vertex causality is satisfied, one can show that for noncompact spatial topology there are no closed timelike curves [@hoekzema]. For compact spatial slices, where the spacetime topology is that of a cylinder, one can construct explicit geometries which exhibit noncontractible, closed timelike curves. Of course, we do not know a priori whether the presence of closed timelike curves in individual path integral configurations has any influence on the continuum limit of the model, and perhaps leads to undesirable continuum properties. It appears that in the context of our present investigation this issue is largely circumvented. Although the causality conditions we impose are of a local nature, and may admit the presence of closed timelike curves, it turns out that the geometries dominating the sum over histories dynamically retain a weak degree of foliation (see Sec. \[sec:tetdist\] below), which suggests that such curves are certainly not abundant. We have not seen closed timelike curves in random samples, but have not systematically tested for their presence either.
Anticipating the choice of boundary conditions we will make in 2+1 dimensions, we may relax the local causality constraint slightly by allowing for the presence of an isolated “source" and “sink" of time. By this we mean a vertex where only timelike links meet, all of them either time-oriented away from the vertex (source) or toward it (sink), as illustrated by Fig. \[fig:sourcesink2d\]. For compact spatial boundary conditions, choosing a source and a sink as initial and final (degenerate spatial) boundaries will convert the cylinder into a spherical $S^2$-spacetime topology. A similar choice of boundary conditions in 2+1 dimensions will lead to a $S^3$-spacetime topology, with the source and sink forming the south and north pole of the sphere, as we will see later.
There is a particular substructure of the triangulations, called a [*bubble*]{} [@hoekzema], which involves the newly added building blocks and is helpful in analyzing the geometry of nonfoliated CDT. In 1+1 dimensions it is simply a pair of $sst$-triangles with a chain of $tts$-triangles in between. This is the general structure of a two-dimensional connected region bounded by a closed loop of spacelike links, and whose interior contains only timelike links, as shown in Fig. \[fig:bubblestrip\], left (we are assuming that vertex causality is satisfied everywhere). This should be contrasted with the structure of a strip, which likewise denotes a two-dimensional piece of triangulations bounded by spacelike links and without spacelike links in its interior, but whose boundary is disconnected (Fig. \[fig:bubblestrip\], right). CDT quantum gravity in 1+1 dimensions has only strips, whereas the version without distinguished foliation has both strips and bubbles. Analogous structures will play a role in our analysis in 2+1 dimensions too, where also the interior structure of a bubble can become more complicated.
Relaxing the foliation: Kinematics in 2+1 dimensions
====================================================
### Local causality conditions {#sec:nfct21 .unnumbered}
Following an analogous procedure in 2+1 dimensions to arrive at a model without distinguished simplicial hypermanifolds, we first must determine which flat tetrahedra – again only built from two types of edge lengths – give rise to Minkowskian building blocks of the correct Lorentzian signature $(-++)$. Fig. \[fig:tetrahedra\] shows all types of tetrahedra which can be constructed using space- and timelike links with fixed squared lengths $\ell_s^2$ and ${\ell_t^2\! =\! -\alpha \ell_s^2}$ respectively. In the rest of this document we will set $\ell_s\! =\! 1$. By calculating the metric in the interior of each tetrahedron type, one finds that only $T_2$, $T_3$, $T_5$ and $T_9$ have the required signature for all values $\alpha\! >\! 0$ of the asymmetry parameter, and type $T_7$ only for $0\! <\! \alpha\! <\! 1$. Note that standard CDT quantum gravity only uses the tetrahedra $T_5$ and $T_9$, in Sec. \[cdt\] referred to as (3,1)- and (2,2)-tetrahedra respectively.
In the present work we will for reasons of simplicity investigate the version of the model where causal spacetimes are assembled from the tetrahedral types $T_2$, $T_3$, $T_5$ and $T_9$ (without $T_7$). As we will see, this already serves our purpose of breaking up the fixed foliated structure.
In 2+1 dimensions violations of local causality – which should therefore be forbidden by the gluing rules – can in principle occur at the links and the vertices of a triangulation. To check whether the light cone structure at a given link is well-behaved, it is sufficient to consider the geometry of a two-dimensional piecewise flat surface orthogonal to the link at its midpoint. This geometry is completely characterized by the set of tetrahedra sharing the link, the so-called star of the link (Fig. \[fig:causality3d\], centre). Each tetrahedron in the star contributes a dihedral angle, defined by the intersection of the tetrahedron with the plane perpendicular to the link (Fig. \[fig:causality3d\], left). The plane segments spanned by all the dihedral angles associated to the given link form a new plane[^7], as shown in Fig. \[fig:causality3d\] (right).
We can distinguish between two cases. If the link at the centre of the star is timelike, the metric of the plane has Euclidean signature, all dihedral angles are Euclidean, and there are no further causality conditions to be satisfied. If on the other hand the link is spacelike, the orthogonal plane is Lorentzian, and so are the dihedral angles. Like in the 1+1 dimensional case discussed in the previous section, we must then require that there is exactly one pair of light cones at the central vertex, and that we encounter exactly four lightlike directions when circling around it. We say that the triangulation satisfies *link causality* if this condition is satisfied for every spacelike link.
Link causality guarantees that light cones everywhere in the triangulation are regular, except possibly at vertices. Intersecting the light cone(s) at a vertex $V$ with the surface of a unit ball around $V$, we obtain two disconnected circles if and only if local causality holds at $V$, see Fig. \[fig:vertexcausality\] (left). In terms of the triangulated surface $\cal S$ of the unit neighbourhood around $V$, vertex causality can be characterized as follows. Mark the end of a timelike link between $V$ and $\cal S$ by a red dot and that of a spacelike one by a blue dot. In addition, whenever the light cone through $V$ crosses a link on $\cal S$, mark the link with a green dot. Recalling the situation depicted in Fig. \[fig:fourtriangles\], it is clear that a green dot will always occur on a surface link which connects a red and a blue dot. If we cut all links that are marked with a green dot, the surface triangulation will break up into a number of connected components. If two of the components thus obtained contain red vertices and one component contains blue vertices, we say that *vertex causality* holds at $V$. If this is true for all vertices, we say that the triangulation satisfies vertex causality. We have not found any Monte Carlo moves which destroy vertex causality but maintain link causality. Also, we have not been able to explicitly construct a triangulation that satisfies link causality and violates vertex causality, but we do not currently have a proof that link causality implies vertex causality.
In order to compute the explicit action for the generalized CDT model of 2+1 dimensional quantum gravity, we will need the values of all dihedral angles. As usual, we will use Sorkin’s complex angle prescription [@sorkin] for the latter, which conveniently keeps track of both Euclidean and Lorentzian angles. The analytic expressions for the cosines and sines of the dihedral angles are listed in Table \[tab:tetraangles\], from which the angles can be computed uniquely. Closer inspection of the geometry of the tetrahedra reveals that a dihedral angle contains a light cone crossing whenever the two triangles bounding the angle are a pair of a spacelike (Euclidean) and a Lorentzian triangle. Local link causality therefore implies that the triangle type changes exactly four times between spacelike and non-spacelike when we circle around a spacelike link once. The number of light cone crossings in the case of a Lorentzian angle is also contained in the table.
tetrahedron links $\cos(\Theta)$ $\sin(\Theta)$ light cone crossings
------------- ----------- ----------------------------------------------------- ------------------------------------------------------------------------ ----------------------
$T_2$ $1,3,5,6$ $\frac{i}{\sqrt{3}}\sqrt{\frac{\alpha}{4+\alpha}}$ $\frac{2}{\sqrt{3}}\frac{\sqrt{3+\alpha}}{\sqrt{4+\alpha}}$ $1$
$2$ $1+\frac{2\alpha}{3}$ $\frac{2i\sqrt{\alpha(3+\alpha)}}{3}$ $0$
$4$ $\frac{2+\alpha}{4+\alpha}$ $\frac{2\sqrt{3+\alpha}}{4+\alpha}$ $-$
$T_3$ $1$ $\frac{i(1+2\alpha)}{\sqrt{3+12\alpha}}$ $2\sqrt{\frac{1+\alpha(4+\alpha)}{3+12\alpha}}$ $1$
$2,3$ $\frac{2+\alpha}{\sqrt{3}\sqrt{-\alpha(4+\alpha)}}$ $\frac{2}{\sqrt{3}}\sqrt{\frac{1+\alpha(4+\alpha)}{\alpha(4+\alpha)}}$ $1$
$4,5$ $\frac{1}{\sqrt{17+\frac{4}{\alpha}+4\alpha}}$ $\frac{2}{\sqrt{4+\frac{\alpha}{1+\alpha(4+\alpha)}}}$ $-$
$6$ $\frac{2+\alpha(4+\alpha)}{\alpha(4+\alpha)}$ $-\frac{2i\sqrt{1+\alpha(4+\alpha)}}{\alpha(4+\alpha)}$ $0$
$T_5$ $1,2,3$ $-\frac{i}{\sqrt{3}\sqrt{1+4\alpha}}$ $\frac{2}{\sqrt{3}}\frac{\sqrt{1+3\alpha}}{\sqrt{1+4\alpha}}$ $1$
$4,5,6$ $\frac{1+2\alpha}{1+4\alpha}$ $\frac{2\sqrt{\alpha(1+3\alpha)}}{1+4\alpha}$ $-$
$T_9$ $1,6$ $\frac{3+4\alpha}{1+4\alpha}$ $-\frac{2i\sqrt{2+4\alpha}}{1+4\alpha}$ $0$
$2,3,4,5$ $-\frac{1}{1+4\alpha}$ $\frac{2\sqrt{2}\sqrt{\alpha(1+2\alpha)}}{1+4\alpha}$ $-$
: Dihedral angles $\Theta$ for all tetrahedra types, given in terms of their trigonometric functions. The link numbers refer to the numbering given in Fig. \[fig:tetrahedra\]. For Lorentzian angles, also the number of light cone crossings is given.[]{data-label="tab:tetraangles"}
Just like in 1+1 dimensions, choosing a time-orientation for a tetrahedron induces an orientation on its timelike links, as well as on the normal to any of its spacelike triangles. It is operationally convenient to keep track of these data in terms of future-oriented arrow assignments, as illustrated by Fig. \[fig:flowoftime3d\]. Again, the local causality conditions do not guarantee that the time orientation can be extended to the full triangulation. In addition to these conditions, we will therefore require (and enforce by way of our computer algorithm) that the complete triangulation can be time-oriented consistently.
### Simplicial substructures {#simplicial-substructures .unnumbered}
In trying to understand the local geometry of nonfoliated CDT configurations and how it is affected by the Monte Carlo moves defined in Sec. \[sec:numsetup\] below, it is useful to isolate specific local substructures built from the fundamental tetrahedra of Fig. \[fig:flowoftime3d\]. To start with, note that only tetrahedra of type $T_2$ and $T_3$ contain triangles with exactly two spacelike edges. Furthermore, both tetrahedra have exactly two such triangles. If we glue two of them together along such a triangle, the resulting simplicial complex again has two such triangles on its boundary. Iterating this gluing procedure we end up with a chain of tetrahedra of type $T_2$ and $T_3$. We conclude that in a triangulation without boundary, the set of all tetrahedra of type $T_2$ and $T_3$ necessarily organizes itself into a collection of closed rings. In a triangulation with boundary also open chains are possible.
Using these simplicial substructures, we can construct three-dimensional analogues of the “bubbles" of Sec. \[sec:ncdt\_2d\] above, by which we will mean connected pieces of triangulation enclosed by a surface made of only spacelike triangles, with no such triangles in its interior. If a ring only contains tetrahedra of type $T_2$, we get a simple bubble, consisting of two spatial discs with identical structure and a timelike link in its interior (Fig. \[fig:rings\]a.) By inserting $T_3$-tetrahedra into a ring of $T_2$-tetrahedra we can form more complicated bubbles, as illustrated by Fig. \[fig:rings\]b. More general bubbles consist of an outer ring of $T_2$- and $T_3$-tetrahedra, enclosing one or more tetrahedra of the other two types, like the one shown in Fig. \[fig:rings\]c.
We can also consider a ring of $T_3$-tetrahedra, as depicted in Fig. \[fig:rings\]d. The spacelike triangles marked in yellow form a spatial disc, with a similar spacelike disc just below. Both discs meet in a single vertex, their respective centres, which we will refer to as a [*pinching*]{} at that vertex. This situation can be generalized by inserting tetrahedra of type $T_2$ into the $T_3$-ring, as shown in Fig. \[fig:rings\]e. The effect is that the two spatial discs now intersect in a link rather than a vertex. Bubbles and pinchings can occur in combination to create even more complicated structures, an example of which is shown in Fig. \[fig:rings\]f. A feature of bubbles which we have not yet mentioned is that they can self-overlap, in the sense that the spherical (or possibly higher-genus) surface bounding a bubble may touch itself along some subset of the surface triangulation. As explained below, we will exclude one kind of self-overlapping bubbles from our simulations, namely, those that wrap nontrivially around the spatial two-sphere.
### Kinematical constraints {#sec:constraints .unnumbered}
The simplest information one can extract from a triangulation is the number of its subsimplices of a particular type. We will use the following counting variables for the four fundamental tetrahedra and the lower-dimensional subsimplex types, as well as their sums in each dimension: $$\begin{aligned}
\label{countvar}
N_0&=\textrm{number of vertices}\nonumber \\
N_1^s&=\textrm{number of spacelike links}\nonumber \\
N_1^t&=\textrm{number of timelike links}\nonumber \\
N_1&:=N_1^s+N_1^t \nonumber\\
N_2^{sss}&=\textrm{number of triangles with three spacelike links}\nonumber \\
N_2^{sst}&=\textrm{number of triangles with two spacelike links}\nonumber \\
N_2^{tts}&=\textrm{number of triangles with one spacelike link} \\
N_2&:=N_2^{sss}+N_2^{sst}+N_2^{tts} \nonumber \\
N_3^{T_2}&=\textrm{number of tetrahedra of type}\ T_2 \nonumber\\
N_3^{T_3}&=\textrm{number of tetrahedra of type}\ T_3 \nonumber \\
N_3^{T_5}&=\textrm{number of tetrahedra of type}\ T_5 \nonumber \\
N_3^{T_9}&=\textrm{number of tetrahedra of type}\ T_9 \nonumber\\
N_3&:=N_3^{T_2}+N_3^{T_3}+N_3^{T_5}+N_3^{T_9}.\nonumber\end{aligned}$$
There exist linear identities among these numbers, which for CDT have been described in [@3d4d]. Here we will repeat the analysis for the extended ensemble, including the new tetrahedral building blocks. The first identity, $$\label{id1}
N_0-N_1+N_2-N_3=\chi,$$ involves the Euler characteristic $\chi$ of the simplicial spacetime manifold. Since every tetrahedron contains four triangles and every triangle is shared by two tetrahedra, we also have the constraint $$\label{id2}
N_2=2 N_3 .$$ Both relations (\[id1\]) and (\[id2\]) are shared by Euclidean DT and standard CDT. In the latter we also have the foliation constraint $2 N_2^{sss}\! =\! N_3^{T_5}$, which expresses the fact that in CDT every spacelike triangle is shared by two tetrahedra of type $T_5$, while every such tetrahedron contains exactly one spacelike triangle. In the present case, the analogous relation is $$\label{id3}
2 N_2^{sss}=2 N_3^{T_2}+N_3^{T_3}+N_3^{T_5} .$$ This is easily understood by counting all spacelike faces in the triangulation – the right-hand side of (\[id3\]) – and noting that the number of spacelike triangles is half the number of spacelike faces.
In CDT we have two more constraints which explicitly involve the leaves of the preferred foliation. One is the Euler constraint $N_0-N_1^s+N_2^{sss}= t_{tot} \tilde{\chi}$, where $t_{tot}$ counts the number of leaves (for periodic boundary conditions in time), and $\tilde{\chi}$ is the Euler characteristic of a spatial section. This constraint no longer exists in the generalized model, since we have Monte Carlo moves which change the quantity $N_0-N_1^s+N_2^{sss}$. Furthermore, in standard CDT every spacelike triangle has three spacelike links and every spacelike link is shared by two spacelike triangles, yielding the relation $N_1^s\! =\! 3 N_2^{sss}/2$. As shown in [@thesis], this can be generalized to the case at hand, leading to the linear relation $$N_1^s=\frac{1}{2}(3N_2^{sss}-N_3^{T_2}) .$$ Lastly, a constraint which does not have a counterpart in foliated CDT follows directly from our earlier observation of $T_2$- and $T_3$-tetrahedra forming closed rings (assuming compact spatial topology), namely, $$N_2^{sst}=N_3^{T_2}+N_3^{T_3} .$$
We have checked that the Monte Carlo moves for the nonfoliated CDT model, described in Sec. \[sec:numsetup\] below, are not compatible with the existence of other linear relations among the counting variables (\[countvar\]). This means that we have a total of 5 such relations for 10 variables, compared with 5 relations for 7 counting variables for standard CDT quantum gravity in 2+1 dimensions. In the next section, we will express the gravitational action as function of the five remaining independent counting variables.
Action and Wick rotation {#sec:actions}
========================
The gravitational path integral (\[eq:pigrav\]) assigns to every spacetime geometry $[g]$ a complex amplitude $\exp(iS[g])$, where $S[g]$ is its classical action. As already noted, we will use the same Regge implementation of the Einstein-Hilbert action in 2+1 dimensions as previous work on CDT quantum gravity [@3d4d], namely, $$\begin{aligned}
\label{eq:reggeaction3d}
S_{\mathrm{Regge}}\!=\! k\!\!\! \sum_{\substack{\mathrm{spacelike}\\ l}} \!\!\! V(l)\frac{1}{i}
\left(2\pi- \!\!\!\!\! \sum_{\substack{\mathrm{tetrahedra}\\ \mathrm{at}\, l}} \!\!\!\! \Theta\right)+
k\!\!\! \sum_{\substack{\mathrm{timelike}\\ l}} \!\!\! V (l)\left(2\pi-\!\!\!\!\!\sum_{\substack{\mathrm{tetrahedra}\\ \mathrm{at}\, l}}
\!\!\!\! \Theta\right)
\! -\!\lambda\!\!\!\!\!\! \sum_{\substack{ \mathrm{tetrahedra}\, T }}\!\!\!\!\!\! V(T),\end{aligned}$$ where $k$ and $\lambda$ are the gravitational and cosmological couplings (up to rescaling), $V(l)$ and $V(T)$ the volumes of a link $l$ and a tetrahedron $T$, and $\sum\Theta$ denotes the sum over dihedral angles contributed by the tetrahedra sharing a link $l$. It was shown in [@3d4d] that an analytic continuation $\alpha\!\mapsto\! -\alpha$ in the asymmetry parameter through the lower-half complex plane defines a nonperturbative Wick rotation which converts the amplitudes $\exp(iS_{\mathrm{Regge}})$ to real weights $\exp(-S_{\mathrm{Regge}}^\mathrm{eucl})$, and thereby makes it possible to analyze the path integral with the help of Monte Carlo simulations. Maintaining the relation $\ell_t^2\! =\! - \alpha \ell_s^2$ between time- and spacelike length assignments, this implies that timelike links acquire a [*positive*]{} squared length after the Wick rotation and therefore effectively become spacelike. The requirement that the full set of link lengths correspond to a proper triangulation [*after*]{} the Wick rotation means that they have to obey triangle inequalities, which in turn puts a restriction on the value of $\alpha$ before the Wick rotation, which for CDT in 2+1 dimensions takes the form $\alpha\! >\! 1/2$.
Let us study how the enlargement of the ensemble of configurations in nonfoliated CDT affects the behaviour of the Regge action (\[eq:reggeaction3d\]) under the map $\alpha\!\mapsto\! -\alpha$. In the first term, $V(l)\! =\! 1$, because the link is spacelike. The plane orthogonal to the link has Lorentzian signature. Because we have imposed link causality, we will cross the light cone four times when circling around the link in the plane. According to our complex angle prescription, each crossing adds a real contribution $\pi/2$ to the total dihedral angle, such that the deficit angle – the expression $(2\pi\! -\! \sum\Theta)$ inside the parentheses – becomes purely imaginary, like in usual CDT, and under the analytic continuation becomes a real deficit angle.
tetrahedron volume Wick rotation condition
------------- ----------------------------------------- ------------------------------------
$T_2$ $\frac{1}{12}\sqrt{\alpha(3+\alpha)}$ $0 < \alpha < 3$
$T_3$ $\frac{1}{12}\sqrt{1+4\alpha+\alpha^2}$ $2-\sqrt{3} < \alpha < 2+\sqrt{3}$
$T_5$ $\frac{1}{12}\sqrt{1+3\alpha}$ $\frac{1}{3} < \alpha $
$T_9$ $\frac{1}{6}\sqrt{\frac{1}{2}+\alpha}$ $ \frac{1}{2} < \alpha$
: Volumes of the four elementary tetrahedra and conditions on the asymmetry parameter $\alpha$, which ensure that the building blocks after the Wick rotation are well defined.[]{data-label="tab:tetravolumes"}
In the second term, the plane orthogonal to the timelike link is Euclidean and remains so after the Wick rotation. On the other hand, we have $V(l)\! =\!\sqrt{\alpha}$, which acquires a factor of $-i$ under the analytic continuation. This implies that the second term changes from real to purely imaginary, as it should. To evaluate the third term we need the volumes of the tetrahedra, which are shown in Table \[tab:tetravolumes\]. The three-volumes as functions of $\alpha$ are useful quantities to look at. In the Lorentzian sector $(\alpha \! >\! 0)$, they are all real and positive. After Wick rotation, a vanishing of the volume $V(T)$ signals a geometric degeneracy of the underlying (Euclidean) tetrahedron $T$, associated with a violation of the triangle inequalities. In addition, note that for the Wick-rotated expressions to give the correct contributions to the Euclidean action, the arguments of the square roots in the second column have to be negative after the substitution $\alpha$ by $-\alpha$, leading to the restrictions on the original $\alpha$-values displayed in the third column of the table. Since all of these constraints have to be satisfied simultaneously, we conclude that in nonfoliated CDT in three dimensions we need $$\frac{1}{2} < \alpha < 3$$ in order for the usual Wick rotation to be well defined, which is stronger than the corresponding condition $1/2 < \alpha$ in CDT, where only the building blocks $T_5$ and $T_9$ are used.
Evaluating the Regge action (\[eq:reggeaction3d\]) with the help of the expressions in Tables \[tab:tetraangles\] and \[tab:tetravolumes\], and applying the Wick rotation leads to $$\label{eq:linearaction2}
S^\mathrm{eucl}=\widetilde{c_1} N_0 + \widetilde{c_2} N_3 + \widetilde{c_3} N_3^{T_2} +
\widetilde{c_4} N_3^{T_3} + \widetilde{c_5} N_3^{T_5}$$ for the Euclideanized Regge action, as function of a specific linearly independent subset of the counting variables (\[countvar\]), where the explicit functional form of the coefficients $\widetilde{c_i}=\widetilde{c_i}(k,\lambda,\alpha)$ has been derived in [@thesis].
For the special case $\alpha\! =\! -1$ and after some manipulations using the kinematical constraints, the action (\[eq:linearaction2\]) can be written as $$\label{eq:action_edt}
S^\mathrm{eucl}=-2 k \pi N_1 + \left(\frac{\lambda}{6\sqrt{2}}+6 k \arccos\frac{1}{3}\right)N_3,$$ which coincides with the action of Euclidean dynamically triangulated gravity in three dimensions. We have also checked that by setting $N_3^{T_2}\! =\! N_3^{T_3}\! =\! 0$ the action (\[eq:linearaction2\]) can be rewritten to precisely match the Regge action for standard CDT quantum gravity in 2+1 dimensions given in [@3d4d].
Numerical set-up {#sec:numsetup}
================
### Monte Carlo moves {#monte-carlo-moves .unnumbered}
To set up numerical simulations of nonfoliated CDT quantum gravity in 2+1 dimensions, we need to define a set of Monte Carlo moves. In this section, we will present a compact description of the moves, which fall into two groups; further details can be found in [@thesis]. The first group contains generalizations of the moves that were already used for CDT simulations in 2+1 dimensions [@3d4d], and which in turn are adapted versions of the original Pachner moves for Euclidean DT [@pachner; @gross].
Fig. \[fig:genpach\] shows the three adapted Pachner moves in 2+1 dimensions. They all change the interior of a small compact region of the simplicial manifold, while leaving its boundary invariant. For CDT triangulations, once the location of a link to be added has been fixed, its type (timelike or spacelike) is also fixed. This is no longer true in the nonfoliated CDT model, where each of these moves comes in several “flavours". A move of this kind is called a “$m$-$n$ move” if $m$ and $n$ are the numbers of tetrahedra in the local simplicial neighbourhood before and after the move is executed. The 2-6 move is the only generalized Pachner move which creates a new vertex.
The Monte Carlo moves in the second group are new compared to standard CDT. Three of them implement the collapse of a link, and only differ in the types of links and the local neighbourhood involved, as illustrated by Fig. \[fig:newmoves\]. They can be seen as special cases of the most general link collapse move, of which we currently do not know whether and how it can be implemented efficiently.
The bubble move (Fig. \[fig:newmoves\], top left) operates on a ring of $T_2$-tetrahedra with a single timelike link in its interior, forming a “bubble" according to the definition given in Sec. \[sec:nfct21\]. It collapses the timelike link to a single vertex and simultaneously collapses the bubble to a spatial disc. The pinching move (Fig. \[fig:newmoves\], bottom left) operates on a pair of spatial discs whose centres are connected by a timelike link. It collapses this link, leading to a configuration where the discs touch in a single vertex, thereby forming a “pinching" as described in Sec. \[sec:nfct21\] (c.f. Fig. \[fig:rings\]d).
We have also implemented a move which collapses a spacelike link (Fig. \[fig:newmoves\], top right). Note that in the configuration before the collapse the link types of the upper and lower disc do not necessarily have to match. In the special cases when they do, we call this move [*symmetric*]{}. It means that during the collapse, only links of the same type get identified pairwise. To keep the complexity of the implementation at a manageable level, we have restricted ourselves to the symmetric version of this move.
Lastly, recall our introduction in Sec. \[sec:ncdt\_2d\] of an isolated source and sink of time in 1+1 dimensions (Fig. \[fig:sourcesink2d\]). We will use a straightforward generalization to 2+1 dimensions of these local (causality-violating) configurations as our boundary conditions. The polar move operates on the neighbourhood of such a source or sink of time, and moves it around. Fig. \[fig:newmoves\] (bottom, right) illustrates the situation for a time source, initially located at the top of the single tetrahedron on the left. The move subdivides the tetrahedron into four, with the newly created vertex at the centre becoming the new source of time.
An important feature of a set of Monte Carlo moves is that it should be ergodic, that is, any element of the configuration space can be reached in a finite number of moves. In our case, the configuration space $\tilde{\mathcal{C}}$ consists of all locally causal gluings of the elementary building blocks $T_2$, $T_3$, $T_5$ and $T_9$ that can be time-oriented consistently, and satisfy further regularity conditions specified in the next subsection. We have made the standard choice of a direct product $[0,1]\!\times\! S^2$ for the spacetime topology. It is possible that the moves described here are ergodic in this configuration space; in fact, the original motivation for introducing additional building blocks was to let us move around in the space of triangulations more efficiently. However, we do not have a proof of ergodicity, and suspect this could be rather nontrivial, given the nonlocal character of part of the causality conditions.
### Defining the ensemble {#defining-the-ensemble .unnumbered}
As already mentioned in Sec. \[cdt\], previous simulations of CDT in 2+1 dimensions have worked with a fixed spacetime topology of direct-product form $[0,1]\!\times\! {}^{(2)}\Sigma$, or $S^1\!\times\! {}^{(2)}\Sigma$ if time is compactified. The standard, simplest choice[^8] for the spatial topology – which we will also employ in the present work – is the sphere, ${}^{(2)}\Sigma\! =\! S^2$. A posteriori, the choice of compactifying time in this case does not appear to make much of a difference, because it turns out that the [*dynamics*]{} of 2+1 CDT quantum gravity (for sufficiently large time extension $t_{tot}$ of the configurations) drives the shape of the universe towards a de Sitter space with $S^3$-topology [@3dcdt].
As we will now go on to explain, the most convenient choice of boundary conditions for nonfoliated CDT is that of a direct-product spacetime $[0,1] \times S^2$, where the beginning and end of time are allowed to degenerate to a point, leading effectively to an $S^3$-topology. Recall that in simulations of CDT quantum gravity, the number of time steps $t_{tot}$ is fixed. Since genuine foliated CDT triangulations form a subset of the present ensemble $\tilde{\cal C}$, the question arises whether it is possible to go from one strictly foliated configuration to another one with a different number $t'_{tot}$ of time steps via nonfoliated configurations and using the Monte Carlo moves described in the last subsection. As explained in detail in [@thesis], the answer is yes. It follows that if there is a region in the phase diagram where the configurations are close to foliated, the standard CDT notion of the “number of time steps" will also make sense approximately and one can ask which equilibrium value for this quantity is found after thermalization. During early test simulations in the ensemble $\tilde{\cal C}$ with compactified time we did find configurations that were approximately foliated, but the number of time steps would not thermalize properly. We have been able to circumvent this problem by not compactifying time, and adding a source and a sink of time as the two poles of a three-sphere.
Another technical issue which appeared during early test runs was that the simulations would often end up in “frozen" states where virtually no progress could be made using the implemented Monte Carlo moves. The problem could be traced back to the presence of globally self-overlapping bubbles, winding once or multiple times around the spatial sphere (c.f. our discussion in Sec.\[sec:nfct21\]). Since we were unable to overcome this problem by finding additional moves, we looked for a mechanism to prevent the globally self-overlapping bubbles from appearing. We found that these problematic structures do not form when we forbid all moves which merge or split bubbles. The moves of Fig. \[fig:newmoves\] are essentially unaffected by these restrictions (see [@thesis] for details). The simulations on the reduced ensemble behaved much better after this alteration, although there are still phase space regions where they do not thermalize sufficiently well, as we will discuss later.
Finally, we will use local regularity conditions for the gluings that make the triangulations into simplicial manifolds, which means that each (interior) vertex has a ball-like neighbourhood whose surface is a triangulated two-sphere. This is the choice made in most of the work on higher-dimensional CDT quantum gravity, and will allow for a better comparison of results.
### Re-introducing time {#re-introducing-time .unnumbered}
The distribution of spatial volume as a function of time is an important large-scale observable, and its analysis has been instrumental in relating CDT quantum gravity to a de Sitter minisuperspace cosmology in 2+1 and 3+1 dimensions [@3dcdt; @desitter]. In order to perform a similar analysis also in nonfoliated CDT, we need to define a time coordinate on its generalized configurations. As explained in the introduction, the fact that spatial slices will generally branch and form “bubbles" means that we can no longer use them to define a distinguished time variable.
To explain our alternative prescription of “time", consider a time-oriented member of the configuration space $\tilde{\cal C}$. Given a vertex $V$, consider the set of all future-oriented paths connecting $V$ with the north pole. The number of links in each path defines a distance between $V$ and the north pole. By averaging this quantity over all paths we obtain an average distance $d_f$. Repeating the procedure for past-oriented paths, connecting $v$ to the south pole, gives another average distance $d_p$. The time coordinate of $V$ is then defined as $t\! =\! d_f-d_p$. Note that for foliated CDT configurations, this coincides with the usual discrete proper time, up to a trivial factor. We have experimented with other notions of time, including that of shortest distance to the poles; they generally lead to a “washing out" of the tetrahedron distributions described below. It is possible that alternative notions of time are more appropriate or practical for observables different from the ones studied here.
Since the number of oriented paths between a vertex and a pole can become very large, in practice we used a modified algorithm, which calculates $t$ in an approximate fashion. For each vertex we constructed a fixed number of future-oriented paths, using a random process which jumps iteratively from a vertex to a randomly chosen future neighbour until the north pole is reached. We then repeated this process for past-oriented paths and finally calculated the time coordinate as the difference of both average distances.
Given this new notion of time, we can now also assign an (average) time to spatial slices. We define a [*spatial slice*]{} in nonfoliated CDT – with boundary conditions as specified above – as any subset of spatial triangles forming a two-sphere, such that by cutting along the sphere the spacetime triangulation decomposes into two disconnected parts, with time flowing consistently from one side of the cut to the other. In other words, the future-pointing arrows introduced in Fig. \[fig:flowoftime3d\] are all lined up to point in the same direction away from the slice. The time coordinate we assign to such a slice is the average of all time coordinates of its vertices. Note that unlike in standard CDT, where different spatial slices are always disjoint, spatial slices here can have any amount of overlap.
We now have all the ingredients to measure the desired volume profiles. Since the number of spatial slices of an individual path integral configuration can become very large, we use a statistical method to generate a subset of spatial slices which is evenly distributed along the time direction. In order to perform an ensemble average of the volume distribution we use a nontrivial averaging algorithm, details of which are described in [@thesis].
Exploring the phase diagram {#sec:explo}
===========================
We have developed the necessary Monte Carlo simulation software, using C`++` as programming language and taking advantage of object-oriented design principles to incorporate modularity and flexibility into the software. We had anticipated that the software would be more complex than for the CDT simulations, but in the event its complexity even surpassed our expectations. An extended discussion of the details of the software implementation, with special emphasis on the Monte Carlo moves, is given in [@thesis]. In what follows, we will present the results obtained with the simulation software, beginning with an exploration of the phase diagram of nonfoliated CDT quantum gravity in 2+1 dimensions.
The Regge form (\[eq:reggeaction3d\]) of the gravitational action contains two couplings, $k$ and $\lambda$, which are proportional to the inverse bare Newton’s constant and the bare cosmological constant respectively. When evaluating the action on causal triangulations, a third parameter – the asymmetry $\alpha$ – naturally appears because of the distinction between space- and timelike links. Together they span a three-dimensional space of bare actions. Of course, from the way $\alpha$ is introduced in the regularized theory, there is no a priori reason why it should play the role of a coupling constant. Different $\alpha$-values should lead to the same continuum gravity theory. This expectation is consistent with the dynamical results found below.[^9]
As usual in dynamically triangulated systems, we do simulations at fixed system size $N_3$ and then perform a finite-size scaling analysis to extrapolate to the limit of infinite size. This means that the phase diagram of the model is spanned by the parameters $k$ and $\alpha$. As we have derived in Sec. \[sec:actions\], the existence of a Wick rotation limits the allowed values for $\alpha$ to the region $1/2\! <\! \alpha\! <\! 3$. Sticking with the notation “$\alpha$" for this parameter also after the analytic continuation, the Wick rotation maps this region to the range $-3\! <\! \alpha\! <\! -1/2$. The phase diagram of our generalized CDT model is therefore a strip bounded by these two $\alpha$-values, as illustrated by Fig. \[fig:pd\_lcdt\].
In the following we will analyze the dynamics of our model at a range of points along the two axes drawn in the figure. While the simulations work well on the axis defined by constant $\alpha\! =\! -1$, we encountered difficulties when exploring the axis of constant $k\! =\! 0$ in the region $-3\! <\! \alpha\! <\! -1$. As one moves away from $\alpha\! =\! -1$ towards $\alpha\! =\! -3$, the number of accepted Monte Carlo moves goes down significantly and the thermalization time increases rapidly. A closer analysis revealed that the severity of the problems correlates with the presence of bubbles with a complicated internal structure. These problems imply that we currently must concentrate our investigation on the region $-1\! <\! \alpha\! <\! -1/2$.
### Bounds on the vertex density {#bounds-on-the-vertex-density .unnumbered}
In 2+1 dimensions there are kinematical bounds on the ratios of certain counting variables, like the vertex density $N_0/N_3$ and the link density $N_1/N_3$. In the case of CDT the link density satisfies $1\! \leq\! N_1/N_3\!\leq\! 5/4$ [@3d4d], which should be compared with the weaker bound $1\!\leq\! N_1/N_3 \!\leq\! 4/3$ for DT. Using the linear relations (\[id1\]) and (\[id2\]), one easily derives $N_0/N_3=N_1/N_3-1$ in the infinite-volume limit, which means that we can translate the link density bounds into the vertex density bounds $0\!\leq\! N_0/N_3\!\leq\! 1/4$ for CDT and $0\! \leq\! N_0/N_3 \!\leq\! 1/3$ for Euclidean triangulations.
The derivation of the link density bound in CDT involves the spatial Euler constraint, which is not present in the ensemble $\tilde{\cal C}$ we have specified in Sec. \[sec:numsetup\] above. To find the analogous bound for nonfoliated CDT configurations we follow [@3d4d] in considering all Monte Carlo moves that create a vertex. Since all of them change the number of tetrahedra by some amount $\Delta N_3$, the strategy is to select those moves for which $\Delta N_3$ is minimal. Starting with a minimal triangulation and repeatedly applying only the selected moves, the vertex density and thus also the link density will be maximized, and the corresponding bounds follow upon taking the infinite-volume limit.
In the case at hand we have two Monte Carlo moves which create one vertex and three tetrahedra, namely, the bubble move and the polar move described in Sec. \[sec:numsetup\] above. Both are unconstrained moves which can always be executed. We conclude that in nonfoliated CDT quantum gravity the vertex and link densities satisfy the bounds $$0\leq \frac{N_0}{N_3} \leq \frac{1}{3} \qquad \mathrm{and} \qquad
1\leq \frac{N_1}{N_3} \leq \frac{4}{3} .$$ These relations agree with those for Euclidean DT, but the configurations which saturate them differ substantially, as we will see. Note that a relaxation (in the sense of [@thor]) of the local regularity conditions for a simplicial manifold would weaken these bounds, since then bubbles with fewer than three tetrahedra can occur.
Fig. \[fig:n0n3\_k\] shows the measurements of the average vertex densities for various values of the coupling $k$, from simulations with $\alpha\! =\! -1$ and system size $N_3\! =\! 40.000$. The vertex density increases monotonically with $k$, which is not surprising since (at fixed $N_3$) larger values of $k$ favour the creation of vertices. This can be seen easily by rewriting eq. (\[eq:action\_edt\]) with the help of the kinematical constraints, yielding $S^\mathrm{eucl}\! =\! -2 k \pi N_0$ plus a term proportional to $N_3$. As expected, the measured curve in Fig. \[fig:n0n3\_k\] approaches the upper kinematical bound of $1/3$ for large values of $k$. We also see a clear signal of a phase transition between $k\! =\! 0.24$ and $k\! =\! 0.28$, from a phase of low to one of high vertex density, reminiscent of the first-order transitions in the inverse gravitational coupling found in both DT [@edt3d] and CDT [@3dcdt]. Analogous measurements for fixed $k\! =\! 0$ and varying $\alpha$ show that the vertex densities are approximately constant at low values, without any sign of a phase transition. Of course, since we are only investigating the region $-1\! <\!\alpha\! <\! -1/2$ of the phase diagram, we cannot exclude the presence of further phase transitions in the complementary region.
### Emergence of foliated triangulations {#emergence-of-foliated-triangulations .unnumbered}
Foliated CDT geometries form a subset of the ensemble $\tilde{\cal C}$, characterized by the condition $N_3^{T_2}\! =\! N_3^{T_3}\! =\! 0$.[^10] By plotting the number of tetrahedra of these two types as function of the couplings we can therefore look for regions in phase space where foliated triangulations emerge dynamically. Fig. \[fig:n3\_k\] shows the numbers of all tetrahedral types used, averaged over the configurations sampled from $\tilde{\cal C}$, as function of the coupling constant $k$. In the phase with low vertex density on the left, although the building blocks of standard CDT dominate, also the other two types appear in significant numbers, from which we deduce that the triangulations along the line $\alpha\! =\! -1$ apparently are not foliated.
Conversely, Fig. \[fig:n3\_alpha\] shows the expectation values of the numbers of tetrahedra at fixed $k\! =\! 0$, as function of $\alpha$. Note that the phase boundary $\alpha\! =\! -0.5$ does not belong to the phase diagram, since the Wick rotation is not defined there. The measurements corresponding to the rightmost data points in the figure have been performed at $\alpha\! =\! -0.52$. We find that both $N_3^{T_2}$ and $N_3^{T_3}$ approach zero as we move towards the phase boundary. At $\alpha\! =\! -0.52$ we have measured $\langle N_3^{T_2}\rangle\!\approx\! 2.9$ and $\langle N_3^{T_3}\rangle\!\approx\! 14.3$, which means that in the entire system consisting of 40.000 tetrahedra almost none of the building blocks belong to the new types $T_2$ and $T_3$.
We conclude that the configurations appearing close to $\alpha\! =\! -0.5$ are almost perfectly foliated and belong to the phase with low vertex density. Effectively, the dynamics should therefore be very close to the known dynamics of 2+1 dimensional CDT in the extended phase [@3dcdt], and we would expect the geometries to be macroscopically extended with a characteristic blob-shaped volume distribution. These expectations will be confirmed later on.
Tetrahedron distributions {#sec:tetdist}
=========================
We have seen in the last section that foliated configurations emerge close to the boundary of the phase diagram. As we move away from the boundary, the configurations become less foliated. A strict foliation is attained whenever $N_3^{T_2}\! =\! N_3^{T_3}\! =\! 0$, but it is unclear how to translate nonzero values into a measure of foliatedness of a triangulation. We would like to have a more refined observable which tells us how foliated a triangulation is. In the following we will use tetrahedron distributions based on the time coordinate introduced in Sec. \[sec:numsetup\] as a qualitative tool to investigate the degree of foliatedness of a triangulation.
To start with, let us assume that all vertices have been assigned a time coordinate using the algorithm described in Sec. \[sec:numsetup\]. For each tetrahedron we then calculate the sum of the time coordinates of its four vertices and round this value to the nearest integer. By definition, this value gets assigned to the tetrahedron as its new time coordinate. This “tetrahedron time" clearly has a different relative normalization compared to the “vertex time" from which it was derived, but this does not matter as long as we do not use both time coordinates simultaneously.
In a given configuration, we can now count the number of tetrahedra that share the same value of (tetrahedron) time, and plot these numbers as a function of time to generate a tetrahedron distribution. Fig. \[fig:prof\_cdt\] (left) shows such a distribution for a strictly foliated CDT geometry, which we have generated by running our simulation with foliation constraint enabled. We observe that the distribution appears as a superposition of two blob-shaped distributions. One can show that one of them consists of $T_5$-tetrahedra and the other one of $T_9$-tetrahedra [@3dcdt]. An enlarged version of the central part of the distribution is shown in Fig. \[fig:prof\_cdt\] (right), which illustrates that the peaks are organized in groups of three. One can show that every such group corresponds to a “thick slice”, which is the triangulation enclosed between two adjacent simplicial spatial hypermanifolds [@thesis], the higher-dimensional analogue of a “strip" in 1+1 dimensions. Note that for pure CDT configurations the tetrahedron time is effectively a refinement of the number of time steps associated with the preferred foliation, similar to what was considered in [@semicl] to produce finer-grained volume distributions in 3+1 dimensions.
Let us return to the more general setting of nonfoliated CDT quantum gravity and focus on triangulations which are a little further away from the phase boundary. Fig. \[fig:prof\_k0.0a-0.7\] shows the tetrahedron distribution of a single triangulation extracted from a simulation at $k\! =\! 0,\ \alpha\! =\! -0.7$, with volume $N_3=40.000$. We observe a sequence of peaks, with some remnants of the three-peak structure exhibited by Fig. \[fig:prof\_cdt\]. The tendency of these structures to become blurred most likely depends both on changes in the actual triangulation and on the precise algorithm used to define the time coordinate and the tetrahedron distribution. Consequently, the relevant information lies not so much in the precise structure of the peaks, but in the overall pattern formed by the succession of all the peaks. Comparison of the two configurations suggests that every peak in Fig. \[fig:prof\_k0.0a-0.7\] corresponds to a group of three peaks in Fig. \[fig:prof\_cdt\] and describes a structure which resembles a thick slice in a foliated triangulation. Another aspect in which the two configurations differ is the fact that the gaps between each group of three peaks in Fig. \[fig:prof\_cdt\] (right) – marking the location of spatial triangulated hypermanifolds – are getting filled in when the foliation is relaxed. This can be interpreted as a “decoration" of the spatial slices by the creation of bubbles.
Based on these findings we can interpret the pattern shown in Fig. \[fig:prof\_k0.0a-0.7\] as a triangulation where decorated spatial slices alternate with modified thick slices. A triangulation exhibiting such a structure will be called *weakly foliated*. This is obviously not a sharp definition since we are not providing a sharp criterion for when a weakly foliated triangulation changes into a truly nonfoliated one. We will take an operational point of view here and consider a triangulation to be weakly foliated whenever the tetrahedron distribution shows the characteristic alternating pattern.
The next important step is to understand how the foliatedness of a triangulation changes as one moves around in phase space. Fig. \[fig:prof\_alpha\] shows a sequence of typical tetrahedron distributions of single triangulations extracted from simulations at $k\! =\! 0$, for various choices of $\alpha$. From an almost strict foliation at $\alpha\! =\! -0.52$ the signal – although remaining distinctly visible – gradually weakens as we move towards $\alpha\! =\! -1$. It would be interesting to follow the development of this pattern beyond this point towards the other phase boundary, but technical issues currently prevent us from doing so, as we have discussed earlier. We have performed a similar analysis on the line of constant $\alpha\! =\! -1$, and have observed that the alternating pattern remains visible, but becomes less pronounced when one moves from $k\! =\! 0$ towards $k\! =\! -1$, indicating a further weakening of the foliation. When going from $k\! =\! 0$ in the other direction towards the phase transition, the data quality decreases significantly, to such an extent that an interpretation based on the tetrahedron distribution becomes unreliable. To summarize, it appears that all investigated configurations in the phase of low vertex density exhibit some kind of (weak) foliation, whose degree varies significantly, from an almost strict foliation near the phase boundary at $(k,\alpha)\! =\! (0,-0.5)$ to a much less pronounced one for larger $| \alpha |$.
Volume distributions {#sec:voldist}
====================
### A phase of extended geometry {#a-phase-of-extended-geometry .unnumbered}
In Sec. \[sec:numsetup\] we introduced the notions of [*time*]{} and [*spatial slices*]{} for a general, nonfoliated CDT geometry. The presence of these ingredients allows us to measure volume distributions – also called [*volume profiles*]{} – just like in standard CDT quantum gravity. In the following we will present the results of our numerical investigations. Fig. \[fig:vp\_k\] shows the expectation value of the measured volume distributions for various values of the coupling $k$, for fixed $\alpha\! =\! -1$. In all cases the average geometry is macroscopically extended and the average volume profile has a characteristic blob shape, strongly reminiscent of what is found in CDT in the physically interesting phase [@3dcdt]. We will report later in this section on a quantitative analysis of the average volume distributions.
Fig. \[fig:vp\_k\] illustrates that with increasing $k$ the time extension of the average geometry also increases. In addition, as one approaches the phase transition, the emergent geometry develops a “tail" at both ends of the volume profile, by which we mean a region of small, approximately constant spatial volume. Since this structure is not resolved in Fig. \[fig:vp\_k\], we have replotted the distributions close to the phase transition (at $k\! =\! 0,\, 0.2,\, 0.24$) in Fig. \[fig:vp\_k\_3\], with an enlarged scale for the time axis and a small upward shift of the distribution curves. This tail looks similar to the stalk observed in simulations of CDT, but is not necessarily related because of the different choices of boundary conditions. In CDT, its presence is enforced by the fact that the simulations are run at a fixed total time extent (equal to the number of spatial slices in the foliation), [*and*]{} that the two-volume is not allowed to vanish, but is bounded below by a minimum of four triangles, set by the manifold conditions. In the present case, we employ the same regularity condition, but the time extension of the geometry is dynamical and the stalk develops spontaneously as we move from $k\! =\! 0$ towards the phase transition. Anticipating our interpretation below of the volume profiles in terms of de Sitter universes, the appearance of the tails could be related to quantum corrections to the underlying effective minisuperspace action near the phase transition.
Consider now a volume distribution on the line $\alpha\! =\! -1$ [*beyond*]{} the transition, that is, in the phase of high vertex density. Fig. \[fig:vp\_tube\] shows a volume distribution of a typical path integral configuration from a simulation at $(k,\alpha)\! =\! (0.4,-1)$ with 40.000 tetrahedra. The qualitative picture in this phase is completely different: the vast majority of spatial slices have (almost) minimal size $N_2(t)$, and the triangulation forms a very long stalk with minimal spatial extent almost everywhere. At this phase space point, we have checked that the time extension of the stalk scales linearly with the system size. In the infinite-volume limit, it would therefore appear that the “universe" becomes a one-dimensional timelike string.
We can now summarize our findings. At all phase space points investigated we have found average geometries which are macroscopically extended and whose volume profile has a characteristic blob-like shape. The time extension of the average geometry increases with increasing $k$, and near the phase transition the geometry starts to develop tails. On the other side of the transition the geometries degenerate into long tubes, unrelated to any 2+1 dimensional classical geometry.
### Evidence for three-dimensionality from finite-size scaling {#evidence-for-three-dimensionality-from-finite-size-scaling .unnumbered}
We will investigate next whether we can assign a macroscopic dimensionality to the extended structure of the volume profiles found in the phase of low vertex density, by performing a systematic finite-size scaling analysis. To this end, we have run another extended series of simulations, taking data at six points along the axis of constant $\alpha\! =\! -1$, ranging from $k\! =\! -1.0$ to $k\! =\! 0.0$, and at six points along the axis of constant $k\! =\! 0$, ranging from $\alpha\! =\! -0.52$ to $\alpha\! =\! -1.0$. At each point we have performed four simulations with different system sizes, namely, $N_3\! =\! 40$, 80, 120 and 160$k$.
Fig. \[fig:vpfss\_k0.0\] (left) shows the measurements of the expectation value of the volume distributions for the four different system sizes at $(k,\alpha)\! =\! (0.0,-1.0)$. Following the strategy of [@desitter] for foliated CDT triangulations in 3+1 dimensions, we will use finite-size scaling to achieve a best overlap of these curves. Assuming that the average geometry has macroscopic dimension $d$, we expect time intervals to scale like $N_3^{1/d}$ and spatial volumes like $N_3^{(d-1)/d}$. When plotting the distributions with axes rescaled accordingly, the measured curves should fall on top of each other.
To find an estimate for $d$ we have run an algorithm that scans through an interval of $d$-values in steps of $\Delta d=0.005$, which for each $d$-value measures how well the volume profiles overlap. We have employed a standard least-squares measure with appropriate normalization to quantify the quality of the overlap. The value of $d$ which minimizes this measure is taken as an estimate for the macroscopic dimension. For the case at hand, the algorithm yields a best estimate of $d\! =\! 2.91$. The plot in Fig. \[fig:vpfss\_k0.0\] (right) shows all four distributions with axes rescaled using this value for the dimension, resulting in a virtually perfect overlap.
We have repeated the same analysis for the other points in the phase diagram. Fig. \[fig:d\] summarizes the calculated estimates for the macroscopic dimension $d$, for fixed $\alpha\! =\! -1$ (left) and fixed $k\! =\! 0$ (right). The large dots indicate measurements with an overlap of excellent quality, the small dots those of a somewhat lesser quality. We observe that all six high-quality measurements yield macroscopic dimensions between $d\! =\! 2.85$ and $d\! =\! 3.00$, while the values from the remaining six measurements have a larger spread.
We have not included error bars in the plots of Fig. \[fig:d\] and the values obtained for the dimension $d$, because they are dominated by systematic errors we currently cannot estimate, one possible source being algorithmic dependences. The calculation of the dimension observable is highly nontrivial and involves several algorithmic choices which potentially affect the final result. Recall that we first had to define a time coordinate, which we did by calculating the average distance between a vertex and the poles of the three-sphere. Secondly, we assigned a time coordinate to the spatial slices, by averaging over the time coordinates of the vertices in the slice. Finally, we ran a rather sophisticated averaging algorithm to produce the final distributions. The nonuniqueness of this entire process is likely to lead to systematic errors not captured by standard error algorithms such as the bootstrap method.
It is clear from this discussion that a single dimension measurement does not provide sufficient evidence to support the $d\! =\! 3$ hypothesis, which would imply compatibility with the standard CDT result. On the other hand, all twelve results together suggest strongly that the average geometries in the phase of low vertex density are three-dimensional. The results on the functional form of the volume profiles presented below will strengthen this preliminary conclusion even further.
### Comparison with the three-sphere {#comparison-with-the-three-sphere .unnumbered}
A crucial piece of evidence that CDT quantum gravity has a well-defined classical limit comes from matching the average distributions of spatial volumes with those of a Wick-rotated version of a solution to the classical Einstein equations, namely, a de Sitter universe [@desitter]. More specifically, the distributions coming from the simulations have been compared to a volume profile of the form $V_3(t)=a \cos^3(b t)$, where $t$ is by assumption proportional to proper time. This is the volume profile of Euclidean de Sitter space (equivalently, the round four-sphere), where the two free parameters $a$ and $b$ depend on the overall size of the universe and a finite relative scaling between spacelike and timelike directions. The measured volume profiles in 3+1 dimensional CDT can be fitted with high accuracy to the analytical $\cos^3$-expression [@desitter; @semicl], with the exception of the region very close to the end points of the curve, which cannot be resolved with sufficient precision and is obscured by the regularity condition $\langle N_2(t)\rangle\! \geq\! 4$, as we have discussed earlier.
We will perform an analogous analysis of nonfoliated CDT in 2+1 dimensions, using the average volume distributions from our simulations. The volume profile of the corresponding continuum de Sitter universe in three dimensions has the functional form $V_2(t)\! =\! a \cos^2(b t)$, where $a$ and $b$ are constants. To extract an optimal fit to this two-parameter family of curves from our Monte Carlo data, we have selected only those points in the phase diagram where the rescaled average volume profiles overlap with excellent quality, and where we have a well-defined curve to compare to.
Fig. \[fig:fit\_k0.0\] shows the outcome of this comparison at the point $(k,\alpha)\! =\! (0.0,-1.0)$. Obviously, the only relevant part of the fit function is the region between the two zeros of the $\cos^2$-function. We see that the functional ansatz fits the average volume distributions almost perfectly, except at the two ends, where the simulation data show a small tail which is not present in the fit function once we cut away the parts outside the two minima. We have already commented earlier on the appearance of such tails in the vicinity of the phase transition (see also Fig. \[fig:vp\_k\_3\]); in the context of the de Sitter interpretation of our universe they indeed seem to be related to small-scale deviations from the classically expected result.
From this point of view it is interesting to understand how the situation changes when one repeats the comparison at a point further away from the phase transition. Fig. \[fig:fit\_k-0.8\] shows the result of the same analysis performed at the point $(k,\alpha)\! =\! (-0.8,-1.0)$. We again observe an almost almost perfect fit in the region where the spatial volume $N_2(t)$ is nonminimal. Remarkably, now even the total time extension of the dynamically generated universe and the de Sitter fit function between the two minima agrees, and we get an almost perfect semiclassical matching. The quality of the fit becomes slightly reduced towards both ends, which is not surprising because discretization effects become large when the spatial volumes become small.
Summary and conclusions {#sec:conclusions}
=======================
We begun our investigation with the aim to isolate and understand the role of the preferred time slicing in standard CDT quantum gravity, while maintaining causality of the individual path integral histories. In this article we have presented many details of the kinematical and dynamical properties of the new, “nonfoliated" CDT model in 2+1 dimensions, which implements the dissociation of the causal structure and the preferred notion of time. Due to the presence of new elementary building blocks, the foliation in terms of equally spaced triangulated spatial hypermanifolds is broken up in this extended version of CDT, acquiring novel simplicial substructures such as “bubbles" and “pinchings".
Gravitational dynamics in the new model is implemented in terms of the standard Regge action, defined as a linear function on the space of independent counting variables, which is five-dimensional, compared to CDT’s two dimensions. After fixing the total system size, there are two coupling constants spanning the phase space of the model, the bare inverse Newton coupling $k$ and the coupling $\alpha$, which quantifies the anisotropy between space- and timelike length assignments in the regularized theory. This asymmetry parameter has to satisfy the inequalities $1/2\! <\! | \alpha | \! <\! 3$ for the Wick rotation to exist, which is necessary to be able to probe the nonperturbative properties of the model with the help of Monte Carlo simulations. This introduces two boundaries in the phase diagram.
The presence of thermalization problems, preventing the effective implementation of the Monte Carlo algorithm, led us to eliminate certain global simplicial substructures. This allowed us to investigate the region $-1\!\lesssim\! \alpha\! <\! -1/2$ of the phase diagram (in terms of the analytically continued $\alpha$), while we still observed severe thermalization problems in the complementary region. We ran the simulations with the spacetime topology of a three-sphere with a source and a sink of (Euclidean) time at the two poles.
In terms of results, we have found two phases of geometry with low and high vertex density, for $k$-values below and above some critical value $k_c$ of the inverse gravitational coupling respectively. The analysis of the tetrahedron distributions revealed that the triangulations remain weakly foliated throughout the investigated phase space region of low vertex density, but that the strength of this signal varies significantly as function of the bare couplings. In addition, we observed the emergence of almost perfectly foliated simplicial geometries close to the boundary $\alpha\! =\! -1/2$ of the phase diagram. We constructed a volume distribution observable and an averaging procedure to study the expectation value of the volume profiles of the emergent geometries in the weakly foliated phase. A finite-size scaling analysis provided strong evidence that the extended geometries are macroscopically three-dimensional. Additional support for this came from fitting the measured profiles to a $\cos^2$-ansatz corresponding to a classical de Sitter universe, which found an almost perfect agreement. We have repeated the analysis for various points in the phase diagram, giving consistent results.
These results provide compelling evidence that the phases of low vertex density of both foliated and nonfoliated CDT quantum gravity have the same large-scale properties in the continuum limit and lie in the same universality class. Since apart from removing the [*distinguished*]{} time slicing we essentially left all other ingredients of the kinematics intact, this would imply that the presence or absence of a preferred foliation in CDT is not a relevant ingredient. As remarked already in the introduction, the same is not true for Hořava-Lifshitz gravity [@hl] – to the extent that our nonperturbative, coordinate-free set-up can be compared with this continuum formulation – where a fixed spatial foliation is essential. It does not mean that CDT, or suitable extensions like that studied in [@cdthorava], cannot provide a framework suitable for studying anisotropic gravity models. Our results also conform with the expectation that in 2+1 dimensions the value of the parameter $\alpha$ is irrelevant from the point of view of the continuum theory.
Because of the strong similarities of the large-scale properties of CDT quantum gravity in 2+1 and 3+1 dimensions, it is plausible to conjecture that also in 3+1 dimensions the presence or absence of a direct-product structure of the triangulations does not influence the final outcome. If this is the case, one may want to stick with the simpler “standard" formulation of Causal Dynamical Triangulations as a matter of convenience and computational simplicity, as we have already pointed out elsewhere [@jl].
[**Acknowledgements.**]{} We thank J. Ambjørn for a critical reading of the manuscript. The authors’ contributions are part of the research programme of the Foundation for Fundamental Research on Matter (FOM), financially supported by the Netherlands Organisation for Scientific Research (NWO). The work was also sponsored by NWO Exacte Wetenschappen (Physical Sciences) for the use of supercomputer facilities, with financial support from NWO.
[99]{}
J. Ambjørn and J. Jurkiewicz: [*Four-dimensional simplicial quantum gravity*]{}, Phys. Lett. B 278 (1992) 42. M.E. Agishtein and A.A. Migdal: [*Critical behavior of dynamically triangulated quantum gravity in four-dimensions*]{}, Nucl. Phys. B 385 (1992) 395 \[hep-lat/9204004\]. S. Catterall, J.B. Kogut and R. Renken: [*Phase structure of four-dimensional simplicial quantum gravity*]{}, Phys. Lett. B 328 (1994) 277 \[hep-lat/9401026\]. P. Bialas, Z. Burda, A. Krzywicki and B. Petersson: [*Focusing on the fixed point of 4-D simplicial gravity*]{}, Nucl. Phys. B 472 (1996) 293-308 \[hep-lat/9601024\]. B.V. de Bakker: [*Further evidence that the transition of 4-D dynamical triangulation is first order*]{}, Phys. Lett. B 389 (1996) 238-242 \[hep-lat/9603024\]. J. Ambjørn and R. Loll, [*Non-perturbative Lorentzian quantum gravity, causality and topology change*]{}, Nucl. Phys. B [536]{} (1998) 407-434 \[hep-th/9805108\]. R. Loll: [*Discrete Lorentzian quantum gravity*]{}, Nucl. Phys. Proc. Suppl. [94]{} (2001) 96 \[hep-th/0011194\];\
[*A discrete history of the Lorentzian path integral*]{}, Lect. Notes Phys. [631]{} (2003) 137 \[hep-th/0212340\];\
J. Ambjørn, J. Jurkiewicz and R. Loll: [*Quantum gravity, or The art of building spacetime*]{}, in [*Approaches to quantum gravity*]{}, ed. D. Oriti (Cambridge University Press, Cambridge, UK, 2009) 341-359 \[hep-th/0604212\];\
R. Loll: [*The emergence of spacetime or quantum gravity on your desktop*]{}, Class. Quant. Grav. [25]{} (2008) 114006 \[arXiv: 0711.0273, gr-qc\];\
J. Ambjørn, A. Görlich, J. Jurkiewicz and R. Loll: [*The emergence of (Euclidean) de Sitter space-time*]{}, in [*Path integrals - New trends and perspectives*]{}, eds. W. Janke and A. Pelster (World Scientific, Singapore, 2008) 191-198;\
[*The quantum universe*]{}, Acta Phys. Polon. B [39]{} (2008) 3309;\
J. Ambjørn, J. Jurkiewicz and R. Loll: [*Deriving spacetime from first principles*]{}, Annalen Phys. [19]{} (2010) 186;\
[*Causal dynamical triangulations and the quest for quantum gravity*]{}, in [*Foundations of space and time*]{}, eds. G. Ellis, J. Murugan and A. Weltman (Cambridge University Press, Cambridge, UK, 2012) \[arXiv: 1004.0352, hep-th\];\
[*Lattice quantum gravity - an update*]{}, PoS LATTICE [2010]{} (2010) 014 \[arXiv: 1105.5582, hep-lat\]. J. Ambjørn, A. Görlich, J. Jurkiewicz and R. Loll: [*Nonperturbative quantum gravity*]{}, Phys. Rept. 519 (2012) 127 \[arXiv: 1203.3591, hep-th\]. J. Ambjørn, J. Jurkiewicz and R. Loll: [*Emergence of a 4D world from causal quantum gravity*]{}, Phys. Rev. Lett. 93 (2004) 131301 \[hep-th/0404156\],\
[*Reconstructing the universe*]{}, Phys. Rev. D 72 (2005) 064014 \[hep-th/0505154\]. J. Ambjørn, A. Görlich, J. Jurkiewicz and R. Loll: [*Planckian birth of the quantum de Sitter universe*]{}, Phys. Rev. Lett. 100 (2008) 091304 \[arXiv: 0712.2485, hep-th\],\
[*The nonperturbative quantum de Sitter universe*]{}, Phys. Rev. D 78 (2008) 063544 \[arXiv: 0807.4481, hep-th\]. J. Ambjørn, S. Jordan, J. Jurkiewicz and R. Loll: [*A second-order phase transition in CDT*]{}, Phys. Rev. Lett. 107 (2011) 211303 \[arXiv: 1108.3932, hep-th\],\
[*Second- and first-order phase transitions in CDT*]{}, Phys. Rev. D 85 (2012) 124044 \[arXiv: 1205.1229, hep-th\]. S. Jordan and R. Loll: [*Causal Dynamical Triangulations without preferred foliation*]{}, Phys. Lett. B 724 (2013) 155-159 \[arXiv: 1305.4582, hep-th\]. J. Ambjørn, L. Glaser, Y. Sato and Y. Watabiki: [*2d CDT is 2d Hořava-Lifshitz quantum gravity*]{}, Phys. Lett. B 722 (2013) 172-175 \[arXiv:1302.6359, hep-th\]. P. Hořava: [*Quantum gravity at a Lifshitz point,*]{} Phys. Rev. D 79 (2009) 084008 \[arXiv: 0901.3775, hep-th\],\
[*General covariance in gravity at a Lifshitz point*]{}, Class. Quant. Grav. 28 (2011) 114012 \[arXiv: 1101.1081, hep-th\];\
P. Hořava and C.M. Melby-Thompson: [*General covariance in quantum gravity at a Lifshitz point,*]{} Phys. Rev. D [82]{} (2010) 064027 \[arXiv: 1007.2410, hep-th\].
J. Ambjørn, A. Görlich, S. Jordan, J. Jurkiewicz and R. Loll: [*CDT meets Hořava-Lifshitz gravity*]{}, Phys. Lett. B 690 (2010) 413-419 \[arXiv: 1002.3298, hep-th\]. J. Ambjørn, J. Jurkiewicz and R. Loll: [*Dynamically triangulating Lorentzian quantum gravity,*]{} Nucl. Phys. B [610]{} (2001) 347-382 \[hep-th/0105267\];\
[*A non-perturbative Lorentzian path integral for gravity,*]{} Phys. Rev. Lett. [85]{} (2000) 924 \[hep-th/0002050\] J. Ambjørn, A. Görlich, J. Jurkiewicz, R. Loll, J. Gizbert-Studnicki and T. Trzesniewski: [*The semiclassical limit of causal dynamical triangulations,*]{} Nucl. Phys. B [849]{} (2011) 144 \[arXiv: 1102.3929, hep-th\]. J. Ambjørn, J. Jurkiewicz and R. Loll: [*Nonperturbative 3-d Lorentzian quantum gravity*]{}, Phys. Rev. D 64 (2001) 044011 \[hep-th/0011276\]. D. Benedetti and J. Henson: [*Spectral geometry as a probe of quantum spacetime*]{}, Phys. Rev. D 80 (2009) 124036 \[arXiv: 0911.0401, hep-th\]. B. Durhuus and C.W.H. Lee, [*A string bit Hamiltonian approach to two-dimensional quantum gravity*]{}, Nucl. Phys. B [623]{} (2002) 201-219 \[hep-th/0108149\]. F. Markopoulou and L. Smolin: [*Gauge fixing in causal dynamical triangulations*]{}, Nucl. Phys. B 739 (2006) 120 \[hep-th/0409057\]. T. Konopka: [*Foliations and 2+1 causal dynamical triangulation models*]{}, Phys. Rev. D 73 (2006) 024023 \[hep-th/0505004\]. S. Jordan: [*Globally and locally causal dynamical triangulations*]{}, doctoral thesis, Radboud University Nijmegen (2013), to be published.
J. Ambjørn, J. Jurkiewicz and R. Loll: [*Lorentzian and Euclidean quantum gravity: Analytical and numerical results*]{}, in [*Proceedings of M-theory and quantum geometry*]{}, 1999 NATO Advanced Study Institute, Akureyri, Island, eds. L. Thorlacius et al. (Kluwer, 2000) 382-449 \[hep-th/0001124\];\
[*Quantum gravity as sum over spacetimes*]{}, Lect. Notes Phys. 807 (2010) 59 \[arXiv: 0906.3947, gr-qc\];\
J. Ambjørn, A. Görlich, J. Jurkiewicz and R. Loll: [*CDT - an entropic theory of quantum gravity*]{} \[arXiv: 1007.2560, hep-th\]. B. Dittrich and R. Loll: [*Counting a black hole in Lorentzian product triangulations*]{}, Class. Quant. Grav. 23 (2006) 3849-3878 \[gr-qc/0506035\]. D. Benedetti and J. Henson: [*Imposing causality on a matrix model*]{}, Phys. Lett. B 678 (2009) 222-226 \[arXiv: 0812.4261, hep-th\]. T. Regge: [*General relativity without coordinates,*]{} Nuovo Cim. 19 (1961) 558-571. D. Benedetti, R. Loll and F. Zamponi: [*(2+1)-dimensional quantum gravity as the continuum limit of Causal Dynamical Triangulations*]{}, Phys. Rev. D 76 (2007) 104022 \[arXiv: 0704.3214, hep-th\]. R. Hoekzema: [*Generalized Causal Dynamical Triangulations in two dimensions*]{}, Master Thesis, U. Utrecht (2012).
R. Sorkin: [*Time evolution problem in Regge calculus*]{}, Phys. Rev. D 12 (1975) 385-396 \[Erratum-ibid. D 23 (1981) 565\]. U. Pachner: [*P.L. homeomorphic manifolds are equivalent by elementary shellings*]{}, Eur. J. Comb. 12 (1991) 129-145.
M. Gross and S. Varsted: [*Elementary moves and ergodicity in D-dimensional simplicial quantum gravity*]{}, Nucl. Phys. B 378 (1992) 367-380. T.G. Budd and R. Loll: [*Exploring torus universes in Causal Dynamical Triangulations*]{}, Phys. Rev. D 88 (2013) 024015 \[arXiv: 1305.4702, hep-th\] G. Thorleifsson: [*Three-dimensional simplicial gravity and degenerate triangulations*]{}, Nucl. Phys. B 538 (1999) 278-294 \[hep-lat/9807026\]. J. Ambjørn, D.V. Boulatov, A. Krzywicki and S. Varsted: [*The vacuum in three-dimensional simplicial quantum gravity*]{}, Phys. Lett. B 276 (1992) 432-436;\
J. Ambjørn and S. Varsted: [*Three-dimensional simplicial quantum gravity*]{}, Nucl. Phys. B [373]{} (1992) 557-580. J. Ambjørn, A. Görlich, J. Jurkiewicz, R. Loll, J. Gizbert-Studnicki and T. Trzesniewski: [*The semiclassical limit of Causal Dynamical Triangulations*]{}, Nucl. Phys. B 849 (2011) 144 \[arXiv: 1102.3929, hep-th\]. C. Anderson, S.J. Carlip, J.H. Cooperman, P. Hořava, R.K. Kommu and P.R. Zulkowski: [*Quantizing Hořava-Lifshitz gravity via Causal Dynamical Triangulations*]{}, Phys. Rev. D 85 (2012) 044027 \[arXiv: 1111.6634, hep-th\].
[^1]: By contrast, Euclidean gravity does not distinguish between space and time, and therefore has no causal structure and no notion of causality.
[^2]: This differs from 1+1 dimensions, where CDT is a quantum-mechanical system of a single length variable and has been shown to coincide with projectable Hořava-Lifshitz gravity [@2dcdthl].
[^3]: One does not know a priori whether $t$ or any other time label of the regularized model will assume a distinguished physical role in the continuum theory; this can only be determined by studying suitable continuum [*observables*]{}, like the volume profiles.
[^4]: In CDT in 1+1 dimensions, one can rederive the exact continuum propagator without invoking an explicit proper-time slicing (see, for example, [@durhuuslee]). It is unclear how to generalize such a construction to higher dimensions.
[^5]: To treat space-, time- and lightlike edges on the same footing, it is convenient to work with the [*squared*]{} edge lengths.
[^6]: Whenever we talk about geometries being [*causal*]{}, what we have in mind is that they possess a well-behaved [*causal structure*]{}. This should not be confused with the notion of causality in standard (quantum) field theory, which refers to the behaviour of matter fields on a given background that typically already comes with a fixed causal structure.
[^7]: This is a slight misnomer; in general, this “plane" will not be flat, because the vertex at the centre will carry a nonvanishing deficit angle.
[^8]: For a recent investigation with toroidal slices, see [@bl].
[^9]: In 3+1 dimensions the status of the analogous parameter $\alpha$ is more involved [@trans; @physrep].
[^10]: With periodic boundary conditions in time one could in principle construct a triangulation obeying $N_3^{T_2}\! =\! N_3^{T_3}\! =\! 0$ consisting of a single bubble winding around both space and time, which clearly is not foliated. However, such configurations do not lie in $\tilde{\cal C}$ because in our case time is not compactified and we do not allow for globally self-overlapping bubbles.
| {
"pile_set_name": "ArXiv"
} |
**AN INTUITIVE PROOF OF THE DATA PROCESSING INEQUALITY**
NORMAND J. BEAUDRY AND RENATO RENNER
*Institute for Theoretical Physics, ETH Zurich, Wolfgang-Pauli-Str. 27*
*8093 Zurich, Switzerland*
Introduction
============
The data processing inequality (DPI) has an intuitive interpretation: the information content in a quantum system cannot increase by performing local data processing on that system. It is an extremely useful property that is used extensively in quantum information [@nielsen00]. The DPI is known to hold for different entropy measures, and is stated generally as $$\label{eq:DPI}
\bar{H}(A|BC)_{\rho} \leq \bar{H}(A|B)_{\rho},$$ where $\bar{H}(A|B)_{\rho}$ is a conditional entropic information measure of the state $\rho_{AB}$. Conditional entropy measures characterize the uncertainty about a system $A$ given a system $B$. The DPI is typically stated for the case where the local operation is a partial trace (i.e. a joint system $(B,C)$ is reduced to the system $B$), but this can be generalized to any physical operation.[^1]
In particular, the DPI holds for one of the most widely used entropy measures: the conditional von Neumann entropy, $H(A|B)_{\rho}$ [@vonneumann55]. It is defined for normalized density operators acting on a bipartite Hilbert space $\mathcal{H}_{AB}$, $\rho \in S_{=}(\mathcal{H}_{AB})$ (where $S_{=}(\mathcal{H}):=\{\rho\in \mathcal{P}(\mathcal{H})|,{\mathrm{Tr}}(\rho)=1\}$ and $\mathcal{P}(\mathcal{H})$ is the set of positive semi-definite operators on $\mathcal{H}$), as $H(A|B)_{\rho} := H(AB)_{\rho} - H(B)_{\rho}$, where $H(A)_{\rho} := - {\mathrm{Tr}}(\rho_{A} \log \rho_{A})$ (all logarithms are taken to the base 2). For simplicity, we will not place the labels on density operators to denote which space they act on when it is clear from the context. Also, Eq. \[eq:DPI\] for the von Neumann entropy is equivalent to its strong subadditivity: $H(ABC)_{\rho}+H(B)_{\rho}\leq H(AB)_{\rho}+H(BC)_{\rho}$.
The first proofs of the DPI for the von Neumann entropy relied on abstract operator properties [@lieb73; @lieb73b; @simon79]. Recently these proofs have been simplified [@nielsen05; @petz86; @ruskai07]. Other approaches have used the operational meaning of the von Neumann entropy [@horodecki06; @horodecki05], Minkowski inequalities [@carlen99; @carlen08], or holographic gravity theory [@headrick07; @hirata07]. There has also been recent interest in the structure of states where there is equality in the DPI [@hayden04; @herbut04; @jencova10]. Our approach provides a new perspective by decomposing the proof of the DPI into a simple proof of a more fundamental property, followed by a specialization. It also provides a new approach to teaching the DPI.
Most precisely, we first prove the DPI for a different entropy: the smooth min-entropy (Theorem \[thm:subaddsme\]). This proof is almost trivial and only involves the partial trace applied to the definition of the smooth min-entropy [@koenig08]. Then we can specialize the smooth min-entropy to the von Neumann entropy by the quantum asymptotic equipartition property (QAEP) (Theorem \[thm:QAEP\]) [@tomamichel08]. Here we provide a short proof that omits the analysis of the rate of convergence of this specialization, as apposed to [@tomamichel08]. We therefore obtain a self-contained proof for the von Neumann entropy DPI (Theorem \[thm:subaddvN\]).
We begin by introducing the smooth min-entropy (Section 2). This is followed by a high level proof of the data processing inequality for the von Neumann entropy (Section 3). Section 4 provides a proof of the QAEP. Finally Section 5 contains lemmas needed for the proofs in the previous sections.
Smooth Min-Entropy
==================
It has become apparent in recent works [@koenig08; @tomamichel08; @datta08; @renner05] that smooth min-entropy is a relevant quantity for measuring quantum information. It characterizes operational tasks in information processing such as data compression and physics in the general one-shot setting, such as in statistical mechanics. Note that the one-shot setting does not make assumptions about the structure of relevant states, for example that they have product form. Since the von Neumann entropy also has an operational significance under certain additional assumptions, it could be expected that the von Neumann entropy can be obtained from smooth entropies as a special case. This is indeed true: the von Neumann entropy can be seen as an “averaged" smooth entropy via the QAEP. We introduce a particular entropy, the min-entropy[^2] $$H_{\min}(A|B)_{\rho}:=\max_{\lambda}\{\lambda\in\mathbb{R} \mid \exists \; \sigma_{B}\in S_{=}(\mathcal{H}_{B}) \text{ s.t. } \rho_{AB} \leq 2^{-\lambda} \mathbbm{1}_{A} \otimes \sigma_{B}\},$$ which leads to the smooth min-entropy, defined as $$H_{\min}^{\epsilon}(A|B)_{\rho} := \max_{\rho_{AB}' \in \mathcal{B}^{\epsilon}(\rho_{AB})} H_{\min}(A|B)_{\rho'}. \label{eq:Hmin1}$$ The state $\sigma_{B}$ is chosen from the set of normalized states $S_{=}(\mathcal{H}_{B})$ in the Hilbert space $\mathcal{H}_{B}$. The state $\rho'_{AB}$ is chosen from the set of subnormalized states in the Hilbert space $\mathcal{H}_{AB}$ that are also close to the state $\rho_{AB}$: $\mathcal{B}^{\epsilon}(\rho_{AB}) := \left\{ \rho'_{AB} | \rho'_{AB} \in S_{\leq}(\mathcal{H}_{AB}) , P(\rho_{AB},\rho'_{AB}) \leq \epsilon \right\}$. To specify this $\epsilon$-ball around a state $\rho$, we use the purified distance [@tomamichel09] $P(\rho,\sigma) := \sqrt{1-F(\rho,\sigma)^2}$ (where $F(\rho,\sigma):={\left\| \sqrt{\rho}\sqrt{\sigma} \right\|}_{1}$ and ${\left\| \rho \right\|}_{1}:={\mathrm{Tr}}\sqrt{\rho \rho^{\dag}}$).[^3]
Data processing inequality
==========================
We are now ready to state our main result and provide a high-level proof. If the entropies of interest are interpreted operationally then Theorem \[thm:subaddsme\] below deals with data processing in the one-shot scenario: a local physical operation is performed on a tri-partite quantum system *once*, and a statement is made about the information content of such a system. Theorem \[thm:subaddvN\] can be interpreted as an average scenario: a statement is made about the information content *on average* after applying a local physical operation to a tri-partite quantum state.
It is important to note that our proof of the DPI for the smooth min-entropy (Theorem \[thm:subaddsme\], below) applies to infinite- and finite-dimensional systems (see [@furrer10]), while our proof of the DPI for the von Neumann entropy (Theorem \[thm:subaddvN\], below) only applies to finite dimensions.
General Data Processing Inequality
----------------------------------
\[[@renner05; @tomamichel09; @koenig08] Smooth min-entropy DPI\] \[thm:subaddsme\] Let $\rho \in S_{=}(\mathcal{H}_{ABC})$. Then $$\label{eq:dpi}
H_{\min}^{\epsilon}(A|BC)_{\rho} \leq H_{\min}^{\epsilon}(A|B)_{\rho}.$$
Specialized Data Processing Inequality
--------------------------------------
Now we have completed the proof of the DPI in the most general case, and the only remaining difficulty is to specialize Theorem \[thm:subaddsme\] to the DPI for the von Neumann entropy. This specialization is achieved by using the limit of many i.i.d. copies of a state, called the QAEP.
\[thm:QAEP\] Let $\rho\in S_{=}(\mathcal{H}_{AB})$. Then $$\lim_{\epsilon\to 0} \lim_{n\to\infty}\frac{1}{n}H^{\epsilon}_{\min}(A^n|B^n)_{\rho^{\otimes n}} = H(A|B)_{\rho}.$$
This directly reduces Theorem \[thm:subaddsme\] to the DPI for the von Neumann entropy.
\[thm:subaddvN\] Let $\rho \in S_{=}(\mathcal{H}_{ABC})$. Then $$H(A|BC)_{\rho} \leq H(A|B)_{\rho}.$$
However, in order to have a self contained proof of the data processing inequality for the von Neumann entropy we provide an alternative, shorter proof of the QAEP than that of [@tomamichel08].
Quantum Asymptotic Equipartition Property
=========================================
In order to prove Theorem \[thm:QAEP\], we upper and lower bound $\lim_{\epsilon\to 0}\lim_{n\to\infty}H^{\epsilon}_{\min}(A^n|B^n)_{\rho^{\otimes n}}$ by $H(A|B)_{\rho}$. These bounds rely on basic properties of smooth entropies, which will be proved in Section 5. The lower bound (Lemma \[lemma:lower\]) is obtained by applying a chain rule to the conditional smooth min-entropy such that it is bounded by a difference of non-conditional smooth entropies (Lemma \[lemma:chain\]). The i.i.d. limit of non-conditional smooth entropies can then be taken (Lemmas \[lemma:lowerboundHminlimit\] and \[lemma:upperboundH0limit\]). The upper bound (Lemma \[lemma:upper\]) can be obtained by bounding the smooth min entropy by the von Neumann entropy of a nearby state (Lemma \[lemma:upperbound\]), and then using the continuity of the von Neumann entropy when the i.i.d. limit is taken (Lemma \[lemma:limitH\]).
For these proofs we will need the smooth $0^{\text{th}}$ order Rényi entropy, which is defined as $H_{0}^{\epsilon}(A)_{\rho} := \min_{\rho' \in \mathcal{B}^{\epsilon}(\rho)} H_{0}(A)_{\rho'}$, where $H_{0}(A)_{\rho} := \log {\mathrm{rank}}\rho_{A}$. In addition, we will need the non-conditional smooth min-entropy defined as $H^{\epsilon}_{\min}(A)_{\rho}:=\max_{\rho'\in{\mathcal{B}^{\epsilon} (\rho)}} H_{\min}(A)_{\rho'}$, where $H_{\min}(A)_{\rho} := -\log {\left\| \rho_{A} \right\|}_{\infty}$. The infinity norm is defined as ${\left\| \rho \right\|}_{\infty} := \max_{i} \{ | \lambda_{i} | \}$, where $\lambda_{i}$ are the eigenvalues of $\rho$. In addition, note that $H^{\epsilon}_{\min}(A|B)_{\rho}$ reduces to $H_{\min}(A)$ in the case that $B$ is trivial and $\epsilon=0$.
\[lemma:lower\] Let $\rho\in S_{=}(\mathcal{H}_{AB})$. Then $$H(A|B)_{\rho} \leq \lim_{\epsilon\to 0}\lim_{n\to\infty} \frac{1}{n} H^{\epsilon}_{\min}(A^{n}|B^{n})_{\rho^{\otimes n}}.$$
\[lemma:upper\] Let $\rho\in S_{=}(\mathcal{H}_{AB})$. Then $$\lim_{\epsilon\to 0}\lim_{n\to\infty} \frac{1}{n} H^{\epsilon}_{\min}(A^{n}|B^{n})_{\rho^{\otimes n}} \leq H(A|B)_{\rho}.$$
General Properties of Smooth Entropies
======================================
The following are properties of smooth entropies used to prove Lemmas \[lemma:lower\], and \[lemma:upper\]. In particular, we bound the smooth min-entropy and smooth $0^{th}$-order Rényi entropy in order to perform the i.i.d. limit of $\epsilon\to 0$, $n\to \infty$. The proofs rely on certain basic properties of the von Neumann entropy and distance measures, which are provided in the appendices.
\[lemma:chain\] Let $\rho \in S_{=}(\mathcal{H}_{AB})$. Then $$H_{\min}^{\epsilon}(AB)_{\rho} - H_{0}^{\epsilon}(B)_{\rho} \leq H_{\min}^{3 \epsilon}(A|B)_{\rho}.$$
Now we provide some bounds on non-conditional smooth Rényi entropies by non-conditional Rényi entropy (Lemmas \[lemma:lowerboundHmin\] and \[lemma:upperboundH0\]). We then use these bounds to show one direction of the non-conditional QAEP (Lemmas \[lemma:lowerboundHminlimit\] and \[lemma:upperboundH0limit\]). Note that the non-conditional QAEP is known, and is sometimes referred to as Schumacher compression [@schumacher95]. It can be proved by using projectors onto a typical set. It can also be essentially reduced to a classical problem that can be shown using the law of large numbers [@cover91]. We provide our proofs below since they provide an alternative proof using bounds on smooth entropies in terms of Rényi entropies, and these bounds may be of general interest in quantum information theory.
\[lemma:lowerboundHmin\] Let $\rho\in S_{=}(\mathcal{H}_{A})$, $\alpha>1$, and $\epsilon\in (0,1]$. Then $$H_{\alpha}(A)_{\rho} + \frac{\log (1-\sqrt{1-\epsilon^2})}{\alpha-1} \leq H_{\min}^{\epsilon}(A)_{\rho}.
\label{eq:limitpart2}$$
\[lemma:upperboundH0\] Let $\rho\in S_{=}(\mathcal{H}_{A})$, $1/2<\alpha<1$, and $\epsilon\in [0,1)$. Then $$H_{0}^{\epsilon}(A)_{\rho} \leq H_{\alpha}(A)_{\rho}+\frac{1}{\alpha-1}\log \sqrt{1-\epsilon}.$$
\[lemma:lowerboundHminlimit\] Let $\rho\in S_{=}(\mathcal{H}_{A})$. Then $$H(A)_{\rho} \leq \lim_{\epsilon \to 0} \lim_{n \to \infty} \frac{1}{n}H_{\min}^{\epsilon}(A^n)_{\rho^{\otimes n}}.$$
\[lemma:upperboundH0limit\] Let $\rho\in S_{=}(\mathcal{H}_{A})$. Then $$H(A)_{\rho} \geq \lim_{\epsilon \to 0} \lim_{n \to \infty} \frac{1}{n}H_{0}^{\epsilon}(A^n)_{\rho^{\otimes n}}.$$
\[lemma:upperbound\] Let $\rho \in S_{=}(\mathcal{H}_{AB})$. Then $\exists \; \tilde{\rho} \in {\mathcal{B}^{\epsilon} (\rho)}$ such that $$H^{\epsilon}_{\min}(A|B)_{\rho} \leq H(A|B)_{\tilde{\rho}}.$$
The authors acknowledge support from the European Research Council (grant no. 258932), and the Swiss National Science Foundation (grant no. 200020-135048).
[10]{}
Michael A Nielsen and I Chuang. . Cambridge University Press, 2000.
John [Von Neumann]{}. . Berlin: Springer, 1955.
Elliott H. Lieb. . , 11:267–288, 1973.
Elliott H. Lieb and Mary B. Ruskai. . , 14:1938–1941, 1973.
Barry Simon. . Cambridge University Press, Cambridge, 1979.
Michael A. Nielsen and Dénes Petz. . , 5(6):517–513, 2005.
Dénes Petz. . , 23:57–65, 1984.
Mary B. Ruskai. . , 60:1–12, 2007.
Micha[ł]{} Horodecki, Jonathan Oppenheim, and Andreas Winter. . , 269(1):107–136, October 2006.
Micha[ł]{} Horodecki, Jonathan Oppenheim, and Andreas Winter. , 436(7051):673–6, August 2005.
Eric A. Carlen and Elliott H. Lieb. . , 189(2):59–69, January 1999.
Eric A. Carlen and Elliott H. Lieb. . , 83(2):107–126, 2008.
Matthew Headrick and Tadashi Takayanagi. . , 76(10):9, November 2007.
Tomoyoshi Hirata and Tadashi Takayanagi. . , 2007(02):042–042, February 2007.
Patrick Hayden, Richard Jozsa, Denes Petz, and Andreas Winter. . , 246(2):359–374, 2004.
Fedor Herbut. . , 37:3535–3542, 2004.
Anna Jencova and Mary B. Ruskai. . , 22(9):1099–1121, March 2010.
Robert König, Renato Renner, and Christian Schaffner. . , 55(9):4337–4347, 2009.
Marco Tomamichel, Roger Colbeck, and Renato Renner. . , 55(12):5840–5847, 2009.
Nilanjana Datta. . , 55, 2816-2826, 2009.
Renato Renner. . Phd thesis, ETH Zürich, 2005.
Fabian Furrer, Johan Aberg, and Renato Renner. . , 306(1):165–186, April 2011.
Marco Tomamichel, Roger Colbeck, and Renato Renner. . , 56, 4674-4681, 2010.
Benjamin Schumacher. . , 51(4):2738–2747, April 1995.
Thomas M. Cover and Joy A. Thomas. . Wiley, 1991.
Christopher A. Fuchs and Jeroen van de Graaf. . , 45(4):1216–1227, 1999.
Mark Fannes. . , 31:291–294, 1973.
Armin Uhlmann. . , 9(2):273–279, September 1976.
Mario Berta, Matthias Christandl, Roger Colbeck, Joseph M. Renes, and Renato Renner. . , 6(9):659–662, July 2010.
R. A. Horn and C. R. Johnson .
The following are known properties used in the proof of Theorem \[thm:QAEP\], which we include here for completeness.
\[thm:uhlmann\] Let $\rho,\sigma \in S_{=}(\mathcal{H})$. Then $$F(\rho,\sigma)=\max_{{| \psi \rangle},{| \phi \rangle}} |{\langle \psi |} \phi \rangle| = \max_{{| \phi \rangle}} |{\langle \psi |} \phi \rangle|,$$ where ${| \phi \rangle},{| \psi \rangle}$ are purifications of $\rho$ and $\sigma$ respectively.
\[lemma:ballcpm\] Let $\mathcal{E}$ be a trace non-increasing map, and $\rho,\sigma \in S_{\leq}(\mathcal{H})$. Then $$P(\mathcal{E}(\rho),\mathcal{E}(\sigma)) \leq P(\rho,\sigma).$$
This can be proven by using the fact that the generalized fidelity cannot decrease under completely positive trace non-increasing maps.
\[lemma:eig\] Let $\rho,\sigma\in S_{=}(\mathcal{H})$, and let $r_{i}$ and $s_{i}$ be their eigenvalues respectively in non-increasing order ($r_{i+1}\leq r_{i}$ and $s_{i+1}\leq s_{i}$ $\forall i$). Also, define $\tilde{\sigma}:=\sum_{i}s_{i}{| i \rangle}{\langle i |}$, where ${| i \rangle}$ are the eigenvalues of $\rho$. Then $$P(\rho,\sigma)\geq P(\rho,\tilde{\sigma})$$
\[lemma:limitH\] Let $\rho \in S_{=}(\mathcal{H}_{AB})$ and $\sigma_{n} \in {\mathcal{B}^{\epsilon} (\rho^{\otimes n})}$. Then $$\label{eq:limiteq2}
\lim_{\epsilon \to 0} \lim_{n \to\infty} \frac{1}{n}H(A^n|B^n)_{\sigma_{n}} = H(A|B)_{\rho}.$$
\[lemma:alphato1\] Let $\rho \in S_{=}(\mathcal{H}_{A})$. Then $$\lim_{\alpha \to 1} H_{\alpha}(A)_{\rho} = H(A)_{\rho}.$$
[^1]: The Stinespring dilation allows for any completely positive trace preserving (CPTP) map to be decomposed into a unitary followed by a partial trace. Since entropy measures are generally invariant under unitaries, the DPI applies to any CPTP map applied to the system $BC$.
[^2]: It is sufficient to take the maximum over $\lambda$ if a finite dimensional system is considered. However, in infinite dimensions it is necessary to take a supremum [@furrer10].
[^3]: If $\rho$ and $\sigma$ are not normalized, then the generalized fidelity is used: $\bar{F}(\rho,\sigma) := {\left\| \sqrt{\rho \oplus (1-{\mathrm{Tr}}\rho)}\sqrt{\sigma \oplus (1-{\mathrm{Tr}}\sigma)} \right\|}_{1}$. If either $\rho$ or $\sigma$ is normalized, then the generalized fidelity reduces to the standard fidelity.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'While there exist now formulations of initial boundary value problems for Einstein’s field equations which are well posed and preserve constraints and gauge conditions, the question of geometric uniqueness remains unresolved. For two different approaches we discuss how this difficulty arises under general assumptions. So far it is not known whether it can be overcome without imposing conditions on the geometry of the boundary. We point out a natural and important class of initial boundary value problems which may offer possibilities to arrive at a fully covariant formulation.'
author:
- |
Helmut Friedrich\
Max-Planck-Institut für Gravitationsphysik\
Am Mühlenberg 1\
14476 Golm, Germany
title: 'Initial boundary value problems for Einstein’s field equations and geometric uniqueness'
---
Introduction
============
This article is concerned with the general principle of geometric uniqueness in the context of the initial boundary value problem for Einstein’s field equations. Because Jürgen Ehlers had always been attracted by problems of principle and he took a certain interest in the initial boundary value problem ([@ehlers:kind]) I devote this article to his memory.
The Cauchy problem for Einstein’s field equations has a long history ([@chrusciel:friedrich]). It was put on a solid basis when Choquet-Bruhat proved the first local existence result ([@foures-bruhat]) and its basic principles are now considered as well understood. In contrast, the systematic and general study of the initial boundary value problem for Einstein’s field equations started only recently and there are still basic open problems. There are two reasons for this.
The PDE theory of initial boundary value problems is technically more involved than that of the Cauchy problem and it has been developed in sufficient generality only in the last few decades (cf. [@benzoni-gavage:serre], [@kreiss:lorenz] for references). On the other hand, there seem to be only a few applications in general relativity whose geometry asks naturally for an analysis in terms of initial boundary value problems. Examples for this are the solutions of Anti-de Sitter type. They can be studied by prescribing for Einstein’s vacuum field equations with negative cosmological constant data at space-like and null infinity, which is represented in terms of the conformally rescaled and extended space-time by a time-like hypersurface ${\cal J}$, and by prescribing Cauchy data on a space-like initial hypersurface ${\cal S}$ which extends to ${\cal J}$. The study of this problem in [@friedrich:AdS] gives the first analysis of an initial boundary value problem for Einstein’s field equations which is general in the sense that no symmetries are required. Remarkably, the problem to be addressed in the following does not occur.
The main interest in initial boundary value problems in general relativity stems from numerical relativity. The need to perform numerical calculations on finite grids suggests to introduce artificial time-like hypersurfaces as boundaries of the grid. The general picture is then that the metric $g$ is calculated on a manifold ${\cal M}$ with boundary ${\cal S} \cup {\cal T}$ and edge $\Sigma = {\cal S} \cap {\cal T}$. Here ${\cal S}$ and ${\cal T}$ are hypersurfaces which are space- and time-like respectively for $g$ and intersect in the space-like surface $\Sigma$, the set ${\cal M} \setminus {\cal S}$ is lying in the future of ${\cal S}$. Such a space-time $({\cal M}, g)$ will be referred to as an $ST$-space-time and it will be called an $ST$-vacuum-space-time if it solves Einstein’s vacuum field equations $R_{\mu \nu} = 0$. In some of the following remarks one will have to assume that $({\cal M}, g)$ satisfies some suitable version of ‘global hyperbolicity’. We will not discuss this any further here.
To simplify the discussion we assume all fields to be smooth. In the applications $\Sigma$ will usually be compact but this will not be important in the following. Because hyperbolic problems can be localized we can focus our attention on some neighbourhood of a given point $p \in \Sigma$. We shall mainly be concerned with evolutions local in time that cover some neighbourhood of ${\cal S}$ (which includes a neighbourhood of $\Sigma$ in ${\cal T}$). Problems arising in long term evolutions will be addressed only if they shed light on the need to control geometric uniqueness.
Given an $ST$-vacuum-space-time, one may ask [*which data need to be given on ${\cal S}$ and ${\cal T}$ to reconstruct the space-time uniquely by solving the field equations*]{}. A first answer to this question is given in [@friedrich:nagy], where it is shown that Einstein’s field equations admit well-posed initial boundary value problems. One of the main difficulties overcome in this work is to ensure that gauge conditions and/or constraints are preserved. For reasons discussed below this task is more subtle here than in the case of the Cauchy problem.
In the subsequent years there was a certain activity concerned with ‘constraint preserving’ formulations of initial boundary value problems in numerical relativity (cf. [@alcubierre] for some of the references). In the present article we shall consider besides the formulation of [@friedrich:nagy] only the one proposed recently by Kreiss, Reula, Sarbach, and Winicour (cf. ([@kreiss:et:al:2009] and references to earlier work given there). It is based on a general and systematic analysis of well posed initial boundary value problems for systems of wave equations. Though the application to Einstein’s field equations is not completely analysed yet, the approach appears to be very flexible and it may offer chances to clarify a question which I consider a main open problem of the subject. While being pointed out at various occasions ([@friedrich:dpg], [@friedrich:nagy]) it has largely been ignored so far. Analysing in local evolution problems the preservation of constraints and gauge conditions is obviously a task which needs to be solved before geometric uniqueness can be addressed.
The space-time structures underlying the general theory of relativity are isometry classes of space-times. A general discussion of boundary value problems of some sort for Einstein’s field equations should thus produce solutions $g_{\mu \nu}$ which, considered as geometric objects, should not depend on any extraneous structures such as background metrics, frame fields, gauge source functions, gauge conditions, coordinates, etc. which are introduced to formulate suitable PDE problems. It should depend only on the ‘isometry class of the data’.
In the case of the Cauchy problem this is guaranteed by a [*geometric uniqueness theorem*]{} which shows that Cauchy data which are isometric in a suitable sense develop into solutions of Einstein’s field equations which are unique up to isometries if the solutions are required to be globally hyperbolic and maximal with this property ([@choquet-bruhat:geroch], [@ringstroem]). In concrete cases it may not be easy to decide whether two Cauchy data sets are isometric but at least there is a well defined concept available. In the case of the initial boundary value problem the situation is not so clear.
It should perhaps be emphasized that geometric uniqueness is not just a problem of academic interest. As long as it has not been clarified it may not be clear what do in analytical studies of initial boundary value problems if the gauge threatens to break down in the course of a development and requires a redefinition. Also, by just comparing their initial boundary data two numerical relativists may not be able to decide, before they start their calculations, whether they can expect to obtain isometric solutions. As shown below the comparison may require the knowledge of the complete development in time of their solutions near the boundary and the subsequent calculation of a gauge transformation on the boundary.
It is the purpose of this article to illustrate the difficulties to arrive at an appropriate concept of geometric uniqueness in the context of general initial boundary value problems. Whether they will find a natural resolution or whether we will have to live with them remains an open question. At the end of the article we discuss a natural and important class of initial boundary value problems which seem to offer a chance for a positive answer.
Some observations
=================
The following [*flat, linear model problem*]{} is the prototype of an initial boundary value problem for a hyperbolic equation. In terms of coordinates $x^{\mu}$ on Minkowski space in which $g_{\mu \nu} = \eta_{\mu \nu} \equiv diag(-1, 1, 1, 1)$ we consider the wave equation $$g^{\mu \nu}\,\phi_{,\,\mu \nu} = f \quad \mbox{on} \quad {\cal M} = \{t \ge 0, |x| \le R\},$$ where $t = x^0$, $R = const. > 0$, and $|x| = \sqrt{\sum_{a = 1}^3 (x^a)^2}$. We prescribe initial conditions $$\phi = \phi_0, \quad \partial_t\,\phi = \phi_1 \quad \mbox{on} \quad {\cal S} = \{t = 0, |x| \le R\},$$ and boundary conditions $$\label{boundary-cond}
B^{\mu}\,\phi_{,\,\mu} = h \quad \mbox{on} \quad {\cal T} = \{t \ge 0, |x| = R\},$$ which are defined in terms of the vector field $$\label{M-def}
B = (1 + e)\,T_* + (1 - e)\,N,$$ on ${\cal T}$, where $T_*$ denotes the future directed time-like unit vector field $\partial_t$ tangential to ${\cal T}$, $N$ the outward directed unit normal to ${\cal T}$, and $$\label{e-def}
e \in C^{\infty}({\cal T}, \mathbb{R}), \quad \quad |e| \le 1,$$ is a given function. The initial data $\phi_0$, $\phi_1$, the boundary data $h$, and $f$ are given real-valued functions which are assumed to be smooth.
The authors of [@kreiss:et:al:2009] consider more general boundary conditions. They require $|e| < 1$ in (\[boundary-cond\]) but admit arbitrary future directed, unit time-like vector fields $T$ tangential to ${\cal T}$ instead of $T_*$ and allow for right hand sides of much more general form. The problem posed above suffices to illustrate the points we wish to discuss.
Clearly, it can only have a solution which is smooth if the data satisfy certain [*consistency conditions*]{} along the edge $$\Sigma = \{t = 0, |x| = R\} = {\cal S} \cap {\cal T}.$$ The wave equation and the initial data imply on ${\cal S}$ the relations $$\label{A-consist-cond}
\partial_t^{2\,k}\,\phi= \Delta^k\,\phi_0 +
\sum_{l = 0}^{k - 1} \Delta^l\,\partial_t^{2(k - 1 + l)} f|_{t = 0},$$ $$\partial_t^{2\,k + 1}\phi = \Delta^k\,\phi_1 +
\sum_{l = 0}^{k - 1} \Delta^l\,\partial_t^{2(k - 1 + l) + 1} f|_{t = 0},
\quad \quad k = 0, 1, 2, \ldots,$$ where $\Delta$ denotes the flat Laplacian. The boundary condition implies on $\Sigma$ $$\label{B-consist-cond}
\sum_{j = 0}^k {k \choose j}\left\{
\partial_t^jB^t\,\partial^{k - j + 1}_t\,\phi +
\sum_{a = 1}^3\partial_t^jB^a\,(\partial^{k - j}_t\,\phi)_{,\,a}\right\}
= \partial^{k}_t\,h,
\quad \quad k = 0, 1, 2, \ldots.$$ For our initial boundary value problem to admit smooth solutions, conditions (\[A-consist-cond\]) and (\[B-consist-cond\]), read for given function $f$ as conditions on the data $\phi_0$, $\phi_1$, and $h$, must be satisfied on $\Sigma$. We consider two special situations.
i\) The intended applications requires a fixed boundary datum $h$ and thus fixed values of $\partial^k_t \,h$ on $\Sigma$. Inserting the time derivatives of $\phi$ given by (\[A-consist-cond\]) in the relations (\[B-consist-cond\]) then results in a sequence of conditions on the space derivatives of the data $\phi_0$ and $\phi_1$ on $\Sigma$. Depending on the function $e$ the construction of such data may be fairly complicated.
ii\) The application requires us to use some prescribed initial data $\phi_0$, $\phi_1$. Inserting the time derivatives of $\phi$ given by (\[A-consist-cond\]) in (\[B-consist-cond\]) determines then the values of $\partial^k_t\,h$ on $\Sigma$. Finding functions $h$ on ${\cal T}$ whose time derivatives assume on $\Sigma$ these values leaves a large freedom for $h$ in the future of $\Sigma$. If, for a given function $f$, this procedure if it is carried out for all possible data $\phi_0$, $\phi_1$, all admissible initial and boundary data are obtained.
If the conditions above are satisfied our initial boundary value problem admits a unique smooth solution. Obviously, this result needs to be largely generalized before it can be applied to Einstein’s field equations. Two quite different such applications will be discussed in the next chapters. Here we use the example above to draw some general conclusions concerning initial boundary value problems for Einstein’s vacuum field equations.
To formulate such problems one has to provide in a first step suitable initial data $h_{ab}$, $\kappa_{ab}$ on a space-like initial hypersurface ${\cal S}$ with boundary $\Sigma = \partial {\cal S}$ for Einstein’s vacuum field equations. The fields $h_{ab}$ and $\kappa_{ab}$ are supposed to be isometric to the intrinsic metric and the second fundamental form induced by the prospective space-time solution $g$ on the embedded hypersurface ${\cal S}$. They will thus have to satisfy the constraints implied by the field equations on space-like hypersurfaces. There are known now fairly general methods to construct solutions to the vacuum constraints ([@bartnik:isenberg], [@holst:nagy:2008], [@maxwell]).
The consistency conditions to be satisfied on $\Sigma$ may create difficulties though. If the intended application requires us to follow the procedure (i), the data will need to satisfy differential relations at any order on $\Sigma$ (the order will be finite if the differentiability requirements on the solutions are relaxed). Such data cannot be constructed by the conformal standard method, in which the problem of solving the constraints is immediately reduced to one of solving elliptic equations. More recent methods show more flexibility in exploiting the underdeterminedness of the constraints ([@chrusciel:delay:2003], [@corvino], [@corvino:schoen]). But even if they turn out to be useful in analytical studies of the present problem there may remain challenges for the numerical relativist.
If the intended application allows us to follow the procedure (ii), it turns out that things are easier also in the case of Einstein’s field equations. In the formulation discussed in the next chapter there arises no problem at all, in the second formulation this question has apparently not been analysed in detail but it also appears to be simpler.
If we want to analyse the freedom to prescribe boundary data for Einstein’s field equations on the boundary ${\cal T} \sim \mathbb{R}_0^+ \times \Sigma$, the first observation to be made in the model problem above is that the freedom to prescribe data on the boundary is only half as large as the freedom to prescribe initial data. Let us consider Einstein’s equations as equations of second order for the metric coefficients. In [*wave coordinates*]{}, characterized by the conditions $$\label{harm-coord}
0 = \Box_g\,x^{\mu} = - \Gamma^{\mu} = - g^{\lambda \rho}\,\Gamma_{\lambda}\,^{\mu}\,_{\rho},$$ the field equations take the form of a system of wave equations (referred to as the ‘reduced system’) for which only ten boundary conditions can be given. This leads to various complications. To get an idea what has to be achieved by the initial and the boundary conditions, we do some simple function counting.
On the space-like initial hypersurface $S = \{t \equiv x^0 = 0\}$ the Cauchy data $g_{\mu \nu}$ and $\partial_t g_{\mu \nu}$ have to be given. The functions $g_{0 \nu}$ and $\partial_t g_{0 \nu}$ are used to remove the gauge freedom and to satisfy the gauge conditions. The remaining metric coefficients can be identified with the metric $h_{ab}$ and the remaining time derivatives are related to the second fundamental form $\kappa_{ab} = \kappa^*_{ab} + \frac{1}{3}\,\kappa\,h_{ab}$ with $h$-trace free part $\kappa^*_{ab}$ and mean extrinsic curvature $\kappa$. Three of the six functions $h_{ab}$ can locally be disposed of by removing the freedom to prescribe the coordinates $x^a$ and a further condition is imposed by the Hamiltonian constraint. Of the six functions $\kappa_{ab}$ the mean extrinsic curvature can be thought of as fixing the shape of the embedding of ${\cal S}$ into the solution space-time while the trace free part $\kappa^*_{ab}$ is restricted by the three momentum constraints. In both cases there remains a freedom of two functions which corresponds to the two degrees of freedom of the gravitational field.
While these considerations gives some insight into the structure and the role of the data it should be said that they are in fact very crude. For instance, it is known that a prescribed mean extrinsic curvature (and possibly a space-like surface $\Sigma$) can be used to determine an embedded space-like hypersurface (with boundary $\Sigma$) in a space-time which is given ([@bartnik:1984], [@bartnik:1988]). In an initial value problem the embedding of the initial hypersurface is constructed, however, simultaneously with the space-time and the roles of the different parts of the data and the constraints cannot be seperated from each other so clearly. Nevertheless, we shall discuss the boundary data in a similar way to illustrate what has to be achieved.
The boundary conditions and data must control the evolution of the boundary. In [@friedrich:nagy] this is achieved by prescribing on the boundary ${\cal T}$ its prospective mean extrinsic curvature and on the edge $\Sigma$ a direction tangential to ${\cal T}$ and transverse to $\Sigma$. The coordinate $x^3$ is then defined so that ${\cal T} = \{x^3 = 0\}$. In [@kreiss:et:al:2009] this has not been completely analysed yet. We note that an understanding of the mechanism which controls the boundary becomes important if the boundary developes a tendency to form cusps or selfintersections.
The boundary conditions have to be given such that constraints and/or gauge conditions are preserved. This requirement, which is taken care of in quite different ways in [@friedrich:nagy] and [@kreiss:et:al:2009] poses considerable difficulties. In an approach based on wave coordinates an obvious choice would be to require $$\label{gamma=0-on-T}
\Gamma^{\mu} = 0 \quad \mbox{on} \quad {\cal T}.$$ In such an approach equations (\[harm-coord\]) are implicit in the reduced evolution system. Thus three of the boundary conditions for the reduced system must be given so that they comprise suitable boundary condition for the wave equations (\[harm-coord\]) which govern the development of the coordinates. This leaves us with the freedom to specify two conditions which control the two degrees of freedom in the gravitational field. Because of the vagueness of the concept of ‘gravitational degrees of freedom’ it is far from obvious how this to be done. The main difficulty, however, is how these requirements can be met so that they result in a well-posed initial boundary value problem for the reduced equations. The operators acting on the $g_{\mu \nu}$ in (\[gamma=0-on-T\]) are already fairly complicated and impose severe restrictions on the remaining choices.
We close this discussion by pointing out two basic differences between initial and boundary data.
i\) The choice of the local coordinates on the initial hypersurface ${\cal S}$ is rather arbitrary and of little consequence for the space-time development local in time. In contrast, the gauge along the boundary ${\cal T}$ is tightly related to the evolution process.
ii\) Together with the reduced equations the Cauchy data on ${\cal S}$ give us control on the geometry at all orders on ${\cal S}$. In contrast, only very little direct information on the geometry of the boundary ${\cal T}$ is provided by the boundary data and conditions. In general, neither the induced metric nor the second fundamental form is available there before the solution has been determined.
The difficulties of controlling the preservation of constraints and gauge conditions which result from these properties have been overcome. There remains, however, the problem of geometric uniqueness.
An approach based on the Bianchi equation
=========================================
To illustrate the nature of the difficulties with geometric uniqueness we need to consider the initial boundary value problem for Einstein’s vacuum equation in some detail. We concentrate on those features which are specific to the treatment of the boundary.
Formulation of the PDE problem
------------------------------
In the approach of [@friedrich:nagy] the Einstein equations are expressed in terms of the following unknows
$-$ the coefficients $e^{\mu}\,_k$ of a frame $\{e_k\}_{k= 0, 1, 2, 3}$ in suitable coordinates $x^{\mu}$, $\mu = 0, 1, 2, 3$,\
where the frame satisfies[^1] $g(e_j, e_k) = \eta_{jk} = diag(-1, 1, 1, 1)$,
$-$ the coefficients $\Gamma_k\,^i\,_j$ of the Levi-Civita connection in the frame $e_k$, which satisfy with\
the covariant derivative operator $\nabla$ defined by $g$ $$\nabla_k\,e_j \equiv \nabla_{e_k}\,e_j = \Gamma_k\,^i\,_j\,e_i,$$ $-$ a tensor $C^{i}\,_{jkl}$ with the algebraic properties of a conformal Weyl tensor, which is given\
in the frame $e_k$.
The field equations are represented by the first structural equation $$\label{first-structural}
[e_i,e_j] = (\Gamma_i\,^k\,_j - \Gamma_j\,^k\,_i)\,e_k,$$ the second structural equation with the assumption that the Ricci tensor vanishes $$\label{second-structural}
e_{[k}(\Gamma_{l]}\,^i\,_j) +
\Gamma_{[k}\,^i\,_{|m|} \Gamma_{l]}\,^m\,_{j}
- \Gamma_{[k}\,^m\,_{l]}\,\Gamma_m\,^i\,_j = 1/2\,\,C^i\,_{jkl},$$ and the Bianchi equation $$\label{bianchi}
\nabla_{i}\,C^{i}\,_{jkl} = 0,$$ which must be satisfied by the conformal Weyl tensor of a vacuum solution.
The coordinates and the frame field have to be restricted by gauge conditions to obtain a useful PDE problem. For simplicity we assume in this chapter the normal to ${\cal S}$ to be tangential to ${\cal T}$ (the general analysis is found in [@friedrich:nagy]). We focus our attention to some neighbourhood of a given point $p \in \Sigma$. A type of gauge as described below will be referred to as [*admissible*]{}
It holds $x^0 = 0$ on ${\cal S}$ and $x^1$, $x^2$, $x^3$ are local coordinates near $p$ with $x^3 \ge 0$ and $x^3 = 0$ on $\partial {\cal S} = \Sigma$. The fields $e_A$, $A = 1, 2$, are tangential to ${\cal S}_c = {\cal S} \cap \{x^3 = c =const. \ge 0\}$.
On ${\cal M} \sim [0, T[ \times {\cal S}$ the time-like unit vector field $e_0$ is future directed, tangential to ${\cal T} = [0, T[ \times \partial {\cal S}$, and orthogonal to $S_c$. The coordinate $x^0$ is a natural parameter on the integral curves of $e_0$ and the other coordinates are dragged along with $e_0$ so that $e_0(x^{\mu}) = \delta^{\mu}\,_0$. The fields $e_A$ are $D$-Fermi-transported in direction of $e_0$ on ${\cal T}_c = \{x^3 = c\}$, where $D$ denotes the covariant derivative defined by the metric induced on ${\cal T}_c$. The field $e_3$ is normal to ${\cal T}_c$ and inward pointing on ${\cal T} = {\cal T}_0$.
In this gauge we have $$e^{\mu}\,_0 = \delta^{\mu}\,_0, \quad
D_a\,e_b = \Gamma_a\,^c\,_b\,e_c, \quad
\chi_{ab} \equiv \Gamma_a\,^3\,_b = \Gamma_b\,^3\,_a, \quad
\chi \equiv g^{ab}\,\chi_{ab}, \quad a, b, c = 0, 1, 2,$$ with $\chi_{ab}$ the second fundamental form and $\chi$ the mean extrinsic curvature on ${\cal T}_c$, $$\label{e0-eA-transport}
\Gamma_0\,^A\,_B = 0, \quad
D_{e_0}\,e_0 = \Gamma_0\,^A\,_0\,e_A, \quad
D_{e_0}\,e_A = - g_{AB}\,\Gamma_0\,^B\,_0\,e_0, \quad A, B = 1, 2,$$ where the summation convention applies to both groups of indices.
As discussed in [@friedrich:nagy], the three functions $\Gamma_0\,^A\,_B$, $\Gamma_0\,^A\,_0$ play the role of gauge source functions on ${\cal M}$ (cf. [@friedrich:2hyp; @red]) and can be chosen arbitrarily. The first of these functions has been disposed of here in a convenient way, choosing the other two functions represents the usual gauge problem in the interior of ${\cal M}$ but it is delicate task on the boundary. The function $\chi$, which can also be prescribed, plays the role of a gauge source function in the interior of ${\cal M}$ and the role of a boundary datum on ${\cal T}$.
With this gauge it is easy to extracted from the complete, overdetermined system (\[first-structural\]), (\[second-structural\]), (\[bianchi\]) a ‘reduced system’ for the unknowns (that are not gauge source functions) which is symmetric hyperbolic and for which well-posed initial boundary value problems can be formulated. But in general the resulting evolution will not preserve the constraints so that not all of equations (\[first-structural\]), (\[second-structural\]), (\[bianchi\]) will be satisfied.
The main theorem of [@friedrich:nagy] shows, however, that there do exist ‘reduced systems’ for which well-posed initial boundary value problems can be formulated whose solutions do satisfy the complete system (\[first-structural\]), (\[second-structural\]), (\[bianchi\]). The following discussion refers to this result.
It is well understood how to provide standard Cauchy data for Einstein’s vacuum field equations on the space-like initial hypersurface ${\cal S}$. The Weyl curvature is then derived from these data ${\cal S}$ by using the Gauss-Codazzi equations. We shall not repeat the details here and concentrate instead on the structure of the boundary conditions and the boundary data on ${\cal T}$, which are critical for our discussion.
It is convenient to use on ${\cal T}$ a double null frame $l$, $k$, $m$, $\bar{m}$ so that, with $i^2 = - 1$, $$l = \frac{1}{\sqrt{2}}(e_0 + e_3),\quad
k = \frac{1}{\sqrt{2}}(e_0 - e_3),\quad
m = \frac{1}{\sqrt{2}}(e_1 + i\,e_2).$$ Since the fields $e_1$, $e_2$ have been fixed only up to rotations $m \rightarrow e^{i\,\phi}\,m$ with functions $\phi \in C^{\infty}({\Sigma}, \mathbb{R})$, there is a corresponding freedom in the data. The ‘spin-weights’ given below refer to the phase factors picked up under these rotations by the various quantities.
In [@friedrich:nagy] the following data are prescribed on ${\cal T}$:
$-$ The mean extrinsic curvature $\chi \in C^{\infty}({\cal T}, \mathbb{R})$,
$-$ functions $q, \, \alpha, \, \beta, \, \Gamma \equiv \Gamma_0\,^1\,_0 + i\, \Gamma_0\,^2\,_0
\in C^{\infty}({\cal T}, \mathbb{C})$ which are of spin-weight $-2, -4, 0, 1$\
respectively and so that $$\left[
\begin{array}{cc}
Re(\bar{\alpha}\,\beta) - \frac{1}{2}(1 - |\alpha|^2 - |\beta|^2)
& Im(\bar{\alpha}\,\beta) \\
Im(\bar{\alpha}\,\beta) &
- Re(\bar{\alpha}\,\beta) - \frac{1}{2}(1 - |\alpha|^2 - |\beta|^2)
\\
\end{array}
\right] \le 0,$$ in the sense of quadratic forms.
The following boundary condition is required on ${\cal T}$: $$\label{bdry-cond-*}
q = - \Psi_4 + \alpha\,\Psi_0 +
\beta\,\bar{\Psi}_0,$$ where $$\quad\mbox{with}\quad
\Psi_0 = C_{\mu \nu \lambda \rho}
\,l^{\mu}\,m^{\nu}\,l^{\lambda}\,m^{\rho},\,\,\,\,\,\,\,\,
\Psi_4 = C_{\mu \nu \lambda \rho}
\,\bar{m}^{\mu}\,k^{\nu}\,\bar{m}^{\lambda}\,k^{\rho}.$$
Properties of the setting and the main problem
----------------------------------------------
If the gauge source functions and the functions $\alpha$, $\beta$ have been prescribed, the symmetric hyperbolic reduced system and the Cauchy data allows us to determine for all unknowns their time derivatives at of any order on ${\cal S}$. The boundary condition (\[bdry-cond-\*\]) then gives us restrictions on the function $q$ on $\Sigma$. If $q$ is chosen accordingly, the consistency conditions will be satisfied.
Because the gauge conditions of the present setting are explicit, an analogue of (\[gamma=0-on-T\]) is not required. Instead of controlling the gauge condition one has to control constraints. This task more or less motivated the way the problem has been arranged in [@friedrich:nagy].
The functions $\alpha$ and $\beta$ do not carry critical information but they allow us to obtain physically or geometrically convenient formulations of the problem. As an example we note that special choices of $\alpha$ and $\beta$ lead to expressions for (\[bdry-cond-\*\]) which only involve certain components of either the electric or the magnetic part of the conformal Weyl tensor with respect to $e_0$ or $e_3$.
The freedom which remains in the definition of the admissible gauge would allow us to choose $\Gamma = 0$ near $\Sigma$. This implies, however, that the field $e_0$ is geodesic with respect to the metric induced on ${\cal T}$, that its flow lines may develop caustics, and that the gauge may break down after some finite time. Finding good choices of $\Gamma$ appears to correspond to the usual gauge problem in long term evolutions which also occur in the Cauchy problem. As seen below, there is much more to it.
A priori it cannot be excluced that parts of the boundary converge towards each other with a tendency to form selfintersections so that the boundary would stop being diffeomorphic to $\mathbb{R}_0^+ \times \Sigma$. One would try to avoid this by choosing $\chi$ suitably. This is different from a change of gauge because it may lead to changes of the geometry. It is not only that the manifold underlying the metric may change its ‘size’ because the boundary evolves in a different way but the metric itself may change in essential ways because the function $q$, given now on a different boundary hypersurface, will acquire a new meaning with respect to the curvature.
It is, of course, understood that the ‘free’ functions $\chi$ and $\Gamma_0\,^A\,_0$ extend smoothly into the interior of ${\cal M}$ as gauge source functions. The question whether the solution is independent of the choice of extension has been discussed in detail in [@friedrich:nagy]. A certain problem was left open there but I expect that it can be resolved.
Irrespective of this problem the results of [@friedrich:nagy] guarantee the existence of (unique) solutions locally near ${\cal S}$ for given initial-boundary data. In particular, given an $ST$–vacuum-space-time in one of the standard gauges, we can prescribe functions $\alpha$ and $\beta$ on ${\cal T}$ with the appropriate spin weights and read off the boundary datum $q$ satisfying (\[bdry-cond-\*\]). Given now the implied boundary data on ${\cal T}$ and the Cauchy data on ${\cal S}$, the solution can be reconstructed uniquely by solving the field equations.
Let some $ST$–vacuum-space-time $({\cal M}, g)$ be given in two different admissible gauges which coincide on ${\cal S}$ but which are such that their respective time-like unit vector fields $e_0$ and $e'_0$ do not coincide on ${\cal T}$ in the future of $\Sigma$. Corresponding to the two gauges choose functions $\alpha$, $\beta$ and $\alpha'$, $\beta'$ respectively in (\[bdry-cond-\*\]). If the initial data induced on ${\cal S}$ and the two sets of data $(\chi, q, \Gamma)$ and $(\chi', q', \Gamma')$ induced on ${\cal T}$ in are read off in these different gauges and the corresponding gauge source functions are use together with the field equations to reconstruct the space-times, we can conclude by PDE uniqueness that the two resulting solutions are isometric because we know them to represent $({\cal M}, g)$ in two different gauges.
We can now state our main question:
[*What could be said about the relation between the solutions determined by the two sets of initial boundary data without having the information that the data have been derived from the same solution by employing two different admissible gauges ?*]{}
Given two sets of Cauchy data one can ask the analogous question whether the corresponding maximal, globally hyperbolic developments in time admit an isometry which maps the respective emdedded data hypersurfaces onto each other. This may be quite difficult to decide in a concrete case but in principle there exists a clear criterion which only involves the two Cauchy data sets: they must be isometric. Unfortunately, such a straightforward answer seems not to be available in the general situation considered above.
Both the functions $\chi$ and $\chi'$ represent the mean extrinsic curvature of the hypersurface ${\cal T}$. They must be given, however, in the form $\chi = \chi(x^{\alpha})$ respectively $\chi' = \chi'(x^{\alpha'})$ with coordinates $x^{\alpha}$ and $x^{\alpha'}$, $\alpha, \alpha' = 0, 1, 2$, which are different from each other on ${\cal T} \setminus \Sigma$ because $e_0 \neq e'_0$ there. To see that $\chi$ and $\chi'$ represent the same object the coordinate transformation $x^{\alpha} = x^{\alpha}(x^{\beta'})$ needs to be known. One might think of deriving this transformation by using the data $\Gamma$, $\Gamma'$ in the last two equations of (\[e0-eA-transport\]). But the boundary data do not contain any information on the metric induced on ${\cal T}$. The operators $D$ and $D'$ are thus only available after the full solution has been determined on ${\cal T}$.
The functions $q$ and $q'$ are difficult to compare for still another reason. The frames $e_k$, $e'_j$ corresponding to the two gauges are related on ${\cal T}$ by a point dependent Lorentz transformation which leaves $e_3$ invariant. In Newman-Penrose notation one finds that the the components $\Psi_l$ and $\Psi'_k$, of the conformal Weyl tensor in the two frames are related by a transformation $$\Psi_k \rightarrow \Psi'_k = \Psi_l\,s^l\,_k
\quad \mbox{with} \quad s^l\,_k \neq 0,\,\,\,\,l, k = 0, \ldots, 4.$$ This implies a relation $$q' = - \Psi'_4 + \alpha'\,\Psi'_0 + \beta'\,\bar{\Psi}'_0
= \eta\,q + \xi\,\bar{q} + \sum_{k = 0}^3\eta_k\,\Psi_k
+ \sum_{k = 0}^3 \xi_k\,\bar{\Psi}_k$$ with coeffficients $\eta_k$, $\xi_k$ which do not all vanish. The calculation shows that one cannot have $\eta_2 = 0$, $\eta_3 = 0$ simultaneously near $\Sigma$ in ${\cal T} \setminus \Sigma$ if $e'_0 \neq e_0$ there. The relation above thus involves components of the conformal Weyl tensor which are not provided by the given boundary data and which will become available only after the development of the solutions in time has been determined.
This suggests that [*under general assumption there does not exist a reasonable concept of a ‘diffeomorphism class of initial boundary data’*]{}. The boundary data sets $(\chi, q, \Gamma)$ and $(\chi', q', \Gamma')$ cannot be compared by operations on ${\cal T}$ which only involve these data sets and possibly interior equations induced by the field equations on ${\cal T}$. The comparison requires knowledge of the development in time of the two initial boundary data sets.
It should be emphasized that rotations of the frame $e_A$ tangential to ${\cal S}_c$ are reflected in the boundary data by simple phase transformations. The problems of the present approach arise from the need to single out a time-like vector field tangential to the boundary ${\cal T}$. In general, there does not exist a distinguished choice.
While this is also true in the case of solutions of Anti-de Sitter type studied in [@friedrich:AdS], geometric uniqueness can be shown. It follows because the initial boundary data are given by standard Cauchy data on ${\cal S}$ and a Lorentzian conformal structure on the conformal boundary ${\cal J}$ which satisfy the appropriate consistency conditions. The possibility to arrive at such a geometric statement, which has no need for a time-like vector field, is related to the fact that the boundary is not simply ‘put in by hand’ but a consequence of the compatibility of the field equations with the requirement of the existence of a conformal boundary ${\cal J}$.
As a consequence, the geometric fields induced on the boundary have the following special properties. There exists a gauge based on the conformal structure of $g$ which reduces on ${\cal J}$ to an analogous gauge based on the induced conformal structure on ${\cal J}$. With a suitable choice of $\alpha$ and $\beta$ a certain component of the ‘boundary-magnetic part’ of the rescaled Weyl tensor represents the free datum $q$. This component is related to the Coton tensor of the intrinsic conformal structure on ${\cal J}$. Via this relation it gives precisely the information needed in the structural equations of the normal conformal Cartan connection to determine from $q$ the conformal structure induced on ${\cal J}$ and vice versa.
An approach based on wave equations
===================================
To see how the features pointed out above are reflected in a different formulation, we give an outline of the approach of [@kreiss:et:al:2009], in which Einstein’s vacuum field equations are understood as equations of second order for the metric coefficients.
Setting up the PDE problem
--------------------------
The authors of [@kreiss:et:al:2009] consider a fixed background metric $\hat{g}_{\mu \nu}$ on ${\cal M} = [0, t_*[ \times {\cal S}$ with some $t_* > 0$ so that the slices $\{t\} \times {\cal S}$, with $0 \le t < t_*$, ${\cal S} \equiv \{0\} \times {\cal S}$, are space-like and the boundary ${\cal T} = [0, t_*[ \times \partial {\cal S}$ is time-like with respect to $\hat{g}_{\mu \nu}$. They impose on ${\cal M}$ the gauge condition $$\label{non-harm-map}
0 = C^{\mu} \equiv g^{\lambda \rho}\,(\Gamma_{\lambda}\,^{\mu}\,_{\rho}
- \hat{\Gamma}_{\lambda}\,^{\mu}\,_{\rho}) - H^{\mu}
= g^{\lambda \rho}\,g^{\mu \nu}\,(
\hat{\nabla}_{\lambda}\,h_{\rho\nu}
- 1/2\,\,\hat{\nabla}_{\nu}\,h_{\lambda \rho}
) - H^{\mu},$$ where $H^{\mu}$ denotes a given vector field, $\Gamma_{\lambda}\,^{\mu}\,_{\rho}$, $\hat{\Gamma}_{\lambda}\,^{\mu}\,_{\rho}$ denote the Christoffel symbols of $g_{\mu \nu}$, $\hat{g}_{\mu \nu}$ respectively, $\nabla$, $\hat{\nabla}$ denote the covariant operators defined by $g$ and $\hat{g}$, and $$h_{\mu \nu} = g_{\mu \nu} - \hat{g}_{\mu \nu}.$$ In this gauge the reduced vacuum field equations take the form of a system of wave equations $$\label{red-wave-equ}
g^{\lambda \rho}\,\hat{\nabla}_{\lambda}\,\hat{\nabla}_{\rho}\,h_{\mu \nu}
= F_{\mu \nu}(h,\, \hat{\nabla}\,h,\,g,\,Riem[\hat{g}])
+ \nabla_{(\mu}\,H_{\nu)},$$ with a certain polynomial function $F_{\mu \nu}$.
Boundary conditions are imposed as follows. With respect to g let $T$ denote a future directed time-like unit vector field tangential to ${\cal T}$, $N$ the outward directed unit normal to ${\cal T}$, and define the real null vectors $K = T + N,\quad L = T - N$. Let $Q$ be a complex linear combination of vectors orthogonal to $K$ and $L$ which satisfies together with its complex conjugate $\bar{Q}$ the normalization conditions $g(Q, Q) = 0$, $g(Q, \bar{Q}) = 2$. The metric can then be written $$\label{K-L-Q-metric-expr}
g_{\mu \nu} = 1/2\,\left(Q_{\mu}\,\bar{Q}_{\nu} + \bar{Q}_{\mu}\,Q_{\nu}
- K_{\mu}\,L_{\nu} - L_{\mu}\,K_{\nu}\right).$$ With suitable functions $$q_{KK}, \, \, q_{Q\bar{Q}} \in C^{\infty}({\cal T}, \mathbb{R}), \quad \quad
q_{KQ}, \,\, q_{QQ} \in C^{\infty}({\cal T}, \mathbb{C}),$$ as boundary data, the gauge requirements $$\left(K^{\mu}\,C_{\mu}\right)_{\cal T} = 0, \quad
\left(L^{\mu}\,C_{\mu}\right)_{\cal T} = 0, \quad
\left(Q^{\mu}\,C_{\mu}\right)_{\cal T} = 0,$$ on ${\cal T}$, where the metric on the right hand side of (\[non-harm-map\]) is written as in (\[K-L-Q-metric-expr\]), are supplemented by the boundary conditions $$\label{a-bdry-con}
\left(K^{\mu}\,K^{\nu}\,K^{\rho}\,\hat{\nabla}_{\mu}\,h_{\nu \rho}
+ \frac{2}{r}\,K^{\nu}\,K^{\rho}\,h_{\nu \rho}\right)_{\cal T} = - q_{KK},$$ $$\label{b-bdry-con}
\left(K^{\mu}\,K^{\nu}\,L^{\rho}\,\hat{\nabla}_{\mu}\,h_{\nu \rho}
+ \frac{1}{r}\,K^{\nu}\,L^{\rho}\,h_{\nu \rho}
+ \frac{1}{r}\,Q^{\nu}\,\bar{Q}^{\rho}\,h_{\nu \rho}\right)_{\cal T}
= - q_{Q\bar{Q}},$$ $$\label{c-bdry-con}
\left(K^{\mu}\,K^{\nu}\,Q^{\rho}\,\hat{\nabla}_{\mu}\,h_{\nu \rho}
+ \frac{2}{r}\,K^{\nu}\,Q^{\rho}\,h_{\nu \rho}\right)_{\cal T} = - q_{KQ},$$ $$\label{d-bdry-con}
\left(K^{\mu}\,Q^{\nu}\,Q^{\rho}\,\hat{\nabla}_{\mu}\,h_{\nu \rho}
- Q^{\mu}\,Q^{\nu}\,K^{\rho}\,\hat{\nabla}_{\mu}\,h_{\nu \rho}\right)_{\cal T}
= - q_{QQ},$$ where the function $r$ denotes the areal radius of the cross section $\{ t \} \times \partial S$ with respect to the background metric.
Some features of the PDE problem
--------------------------------
No detailed specification of the frame $K$, $L$, $Q$, $\bar{Q}$ has been made in [@kreiss:et:al:2009]. To obtain a well defined PDE problem one has to decide, however, on a definit prescription. This choice enters the consistency conditions, it determines the evolution of the boundary and the gauge, and it affects the isometry class of the solution if the right hand sides of (\[a-bdry-con\]) - (\[d-bdry-con\]) are given.
Since the metric is not available on ${\cal T}$ at this stage, one has to give an abstract prescription for the frame field. Assuming $\partial_0$ to be time-like and $\partial_a$, $a = 1, 2, 3$, to be space-like on $\Sigma = {\cal S} \cap {\cal T}$, the specification of the frame can then be made in terms of a Gram-Schmidt orthonormalization of the coordinate frame, which gives the frame $K$, $L$, $Q$, $\bar{Q}$ in terms of the yet unknown metric coefficients $g_{\mu \nu}$ near $\Sigma$. Note that there is a large freedom in doing this and that it will not be clear a priori for how long into the future a particular choice will be well behaved.
Again it is the choice of the time-like unit vector field $T$ which is decisive here. The dependence of the boundary data on the complex vector fields $Q$ and $\bar{Q}$ orthogonal to $T$ and $N$ is well controlled because the boundary data pick up phase factors $e^{i\, k \,\psi}$, with certain $k \in \mathbb{Z}$, under rotations $Q^{\mu} \rightarrow e^{i\,\psi} \,Q^{\mu}$ with $\psi \in C^{\infty}({\cal T}, \mathbb{R})$. In the following a definite prescription of the frame will be assumed.
The meaning of the boundary conditions has been discussed to some extent in [@kreiss:et:al:2009], but it is not easy to analyse and probably not fully understood yet. They affect in particular the gauge defined by (\[non-harm-map\]). The tensorial nature of this gauge condition allows one to change the coordinates conveniently without affecting the hyperbolicity of the reduced equations. The fields $H^{\mu}$ reflect the usual freedom to prescribe four gauge source functions. Their use allows us to generalize the maps of $({\cal M}, g)$ onto $({\cal M}, \hat{g})$ considered below to [*general wave maps*]{}.
To simplify the discussion we consider here only the case where $H^{\mu} = 0$. If $g$ exists on ${\cal M}$, relation (\[non-harm-map\]) then tells us that the identity map of ${\cal M}$ defines a [*wave map*]{} of $({\cal M}, g)$ onto $({\cal M}, \hat{g})$. More generally, a map $\Phi: {\cal M} \rightarrow {\cal M}$ is a wave map for $g$ and $\hat{g}$ if it satisfies the variational principle $\delta \int tr_{g}(\Phi^* \hat{g})\,d\mu_g = 0$. Let $(U, x^{\mu'})$ denote a coordinate patch on ${\cal M}$, $x^{\mu}$ coordinates defined on $\Phi (U)$, and $\Phi^{\mu}(x^{\mu'})$ the local representation of $\Phi$. The Euler-Lagrange equations of this principle are then equivalent to the system of wave equations $$\label{harm-map-equ}
g^{\mu' \nu'}\left(\Phi^{\rho}\,_{, \mu' \nu'} - \Gamma_{\mu'}\,^{\lambda'}\,_{\nu'}\,\Phi^{\rho}\,_{, \lambda'}
+ \hat{\Gamma}_{\mu}\,^{\rho}\,_{\nu}\,\Phi^{\mu}\,_{, \mu'}\,\Phi^{\nu}\,_{, \nu'}
\right) = 0,$$ where $ \Gamma_{\mu'}\,^{\lambda'}\,_{\nu'}$ and $\hat{\Gamma}_{\mu}\,^{\rho}\,_{\nu}$ denote the Christoffel symbols of the metrics $g_{\mu' \nu'}$ and $\hat{g}_{\mu \nu}$ respectively and $\hat{\Gamma}_{\mu}\,^{\rho}\,_{\nu}$ is taken at $\Phi^{\mu}(x^{\mu'})$.
Such maps may be constructed by solving initial boundary value problems for (\[harm-map-equ\]) with Cauchy data on ${\cal S}$ and boundary conditions resp. data on ${\cal T}$. To allow them to define diffeomorphisms ${\cal M} \rightarrow {\cal M}$ of the desired type, the data should given such that the maps induce diffeomorphisms of ${\cal S}$ and ${\cal T}$ onto themselves respectively and such that the tangent maps $T_q \Phi$ have maximal rank at points $q$ in ${\cal S}$ or ${\cal T}$. A solution will then map some neighbourhood of ${\cal S}$ in ${\cal M}$ diffeomorphically onto another such neighbourhood. Depending on the prescribed data and the metrics $g$ and $\hat{g}$, its life time as a diffeomorphism may be limited, however.
Assume such a map $\Phi$ to be given, denote its inverse by $\Psi$ and the pull back of $g$ under $\Psi$ by $g' = \Psi^*g$. Writing the left hand side of (\[harm-map-equ\]) in terms of the argument $x^{\mu}$, one finds that the equations can be rewritten in the form $$\label{harm-map-transf}
g'^{\lambda \rho}\,(\Gamma'_{\lambda}\,^{\mu}\,_{\rho}
- \hat{\Gamma}_{\lambda}\,^{\mu}\,_{\rho}) = 0.$$ This is (\[non-harm-map\]) with $g$ and its Christoffel symbols replaced by $g'$ and its derived fields. Condition (\[non-harm-map\]) thus remains unchanged if we allow for pull backs of $g$ by inverses of wave maps. This freedom is removed by the initial boundary conditions for (\[harm-map-equ\]) which are implicit in the initial boundary conditions for (\[red-wave-equ\]).
It is not easy to see whether the boundary conditions (\[a-bdry-con\]) - (\[d-bdry-con\]) contain information which can be related directly to the behaviour of the geometry defined by $g$ near the boundary. We shall ignore that point here, though the question becomes important when the geometry develops a tendency to collapse near the boundary.
On the boundary ${\cal T}$ we have the freedom to choose the time-like vector field $T$ and the related frame and the data $q = (q_{KK}, \, q_{Q\bar{Q}}, \,q_{KQ}, \,q_{QQ})$. The question arises what happens under a transition $$(T, q) \rightarrow (T', q') \quad \mbox{with} \quad T \neq T'.$$ It could be that it is just implied by a change of the gauge and the frame but it could also correspond to a transition to a different isometry class of metrics. Again it turns out to be impossible to decide this only on the basis of the data given on ${\cal T}$.
The boundary conditions (\[a-bdry-con\]) - (\[d-bdry-con\]) are covariant with respect to coordinate transformations but they have a complicated behaviour under gauge transformations. It is natural to accompany the gauge transformation $g \rightarrow \Psi^* g$ leading to (\[harm-map-transf\]) by the push forward of the frame based on the harmonic map $\Phi$. The defining properties of the frame will be preserved and some calculations will be simplified. The transformation laws of the functions comprised by $q$ follow from the transformation laws of the left hand sides of the boundary conditions. Because the background metric $\hat{g}$ and thus the operatore $\hat{\nabla}$ is kept fixed, the field $h = g - \hat{g}$ transforms into $\Psi^* g - \hat{g}$ and the transformation of $q$ will involve derivatives of $\Psi$ up to second order. Without knowing the transformed solution near ${\cal T}$ these derivatives cannot be determined on ${\cal T}$.
Even if the transition$(T, q) \rightarrow (T', q')$ above would result from a simple redefinition of the frame we would not be able to recognize that. The transformation formula for $q$ under transformations of the frame which leave the normal vector $N$ fixed requires information on the unkowns which is only available when the solutions are known near ${\cal T}$.
We conclude that in both approaches, [@friedrich:nagy] and [@kreiss:et:al:2009], the problem with geometric uniqueness is related to the frame dependence of the boundary data. There is an ‘[*inner frame dependence*]{}’, which refers, depending on the method, to the coordinates in which the boundary data are given or to the inverse wave map acting on the metric, and an ‘[*outer frame dependence*]{}’ which refers to the need to perform in a transformation linear combinations of some of the given data with other data which are not available. In both cases it is the choice of the time-like vector field tangential to ${\cal T}$ which is critical. In general there does not exist a distinguished one. Because it is based on a very general analysis of initial boundary value problems for systems of wave equations, the approach of [@kreiss:et:al:2009] and of previous articles by the same authors may offer more flexibility than that of [@friedrich:nagy] and it may offer new and unexpected possibilities to address the problem of geometric uniqueness.
Covariant boundary data and distinguished time-like vector fields
=================================================================
To avoid problems arising from the frame dependence one may wish to find formulations in which the boundary data which do not serve to control the gauge are prescribed in terms of frame independent fields. In an approach which employs gauge conditions such as (\[gamma=0-on-T\]) and uses wave equations as reduced equations they should be given in terms of the first and second fundamental form $$k_{\mu \nu} = g_{\mu \nu} - N_{\mu}\,N_{\nu}, \quad
\chi_{\mu \nu} = k_{\mu}\,^{\lambda}\,k_{\nu}\,^{\rho}\,\nabla_{\lambda}\,N_{\rho},$$ induced on ${\cal T}$. Here $N$ denotes the outward pointing unit normal to ${\cal T}$ with respect to $g$. The first invariant associated with these fields in the mean extrinsic curvature $\chi = k^{\mu \nu}\,\chi_{\mu \nu}$. Two further invariants are supplied by the eigenvalues of the trace free part $$\chi^*_{\mu \nu} = \chi_{\mu \nu} - \frac{\chi}{3}\,k_{\mu \nu},$$ of the second fundamental form. In local coordinates $x^{\alpha}$ on ${\cal T}$ in which the induced metric takes at a given point $p' \in {\cal T}$ the standard form $k_{\alpha \beta} = \eta_{\alpha \beta} \equiv diag(-1, 1, 1)$ the field $\chi^{* \mu}\,_{\nu} = k^{\mu \rho}\,\chi^*_{\rho \nu} $ is represented at $p'$, possibly after a rotation of the coordinates which leaves the $x^0$-axis fixed, by a matrix of the form $$\chi^{*\alpha}\,_{\beta} =
\left(\begin{array}{ccc}
b + d & - c^1 & - c^2\\
c^1 & - b & 0\\
c^2 & 0 & - d
\end{array}\right).$$ The eigenvalues $\lambda_i$, $i = 0, 1, 2$, of $\chi^*_{\alpha \beta}$, which satisfy $\lambda_0 + \lambda_1 + \lambda_2 = 0$, are then the roots of the equation $$0 = \det(\chi^{*\alpha}\,_{\beta} - \lambda\,\delta^{\alpha}\,_{\beta}) = - \lambda^3
+ 1/2\,\chi^{*\alpha}\,_{\beta}\,\chi^{*\beta}\,_{\alpha}\,\lambda + \det(\chi^{*\alpha}\,_{\beta}).$$ They are functions of the coefficients of this equation and one might try to use directly the invariants $$\label{chi-invariants}
\chi, \quad \,\, \chi^{*\alpha}\,_{\beta}\,\chi^{*\beta}\,_{\alpha}, \quad \,\, \det(\chi^{*\alpha}\,_{\beta}),$$ as boundary data. Even if they could be used for that purpose there remains the problem that the way they must be given may depend on the time-like vector field $T$.
One of the eigenvalues of $\chi^{*\alpha}\,_{\beta}$ is necessarily real and we may ask whether the real eigenvectors of $\chi^{*\alpha}\,_{\beta}$ can be of any use for us. If $b = d = 0$ and $c^A \neq 0$ two of the eigenvalues are complex conjugates of each other and there is one real eigenvector which turns out to be space-like. This remains true if $|c| = \sqrt{(c^1)^2 + (c^2)^2}$ is much larger than $|b|$ and $|d|$. If $c^A = 0$ there exists an orthonormal frame of eigenvectors of $\chi^{*\alpha}\,_{\beta}$ but the time-like eigenvector need not be unique. In fact, if $b = - 2\,d$ or $d = - 2\,b$ there exists a 2-dimensional, time-like subspace of eigenvectors.
A view at the space-time setting underlying the flat, linear model problem considered in the beginning may suggest a reasonable condition under which $\chi^{*\alpha}\,_{\beta}$ can be expected to admit a unique time-like eigenvector. In that case we find $\chi = 2/R > 0$ and $\chi^{*\alpha}\,_{\beta} = \frac{1}{R}\,(\frac{1}{3}\,k_{\mu \nu} + T_{* \mu}\,T_{*\nu})$. It follows that $\chi^{*\alpha}\,_{\beta}$ admits $T_* = \partial_t$ as its unique future directed, time-like unit eigenvector. The property that $T_*$ is orthogonal to the hypersurfaces $\{t = const.\}$ cannot be expected to extend to the curved case but the fact that the pull-back of $\chi^*_{\alpha \beta}$ to the plane orthogonal to $T_*$ is (positive) definite suggests a useful generalization.
Back to the general case, suppose that $\chi^{*\alpha}\,_{ \beta}$ has a time-like eigenvector $T \neq 0$ tangential to ${\cal T}$ at $p'$ with eigenvalue $\lambda_0$. With a suitable scaling of $T$ we can assume, possibly after a Lorentz-Transformation, which leaves the form $k_{\alpha \beta} = \eta_{\alpha \beta}$ unchanged, that $T^{\alpha} = \delta^{\alpha}_0$. It follows that $$\chi^*_{\alpha0} = - \lambda_0\,\delta^0_{\alpha}.$$ By a rotation about the $x^0$-axis, which leaves the form of $k_{\alpha \beta}$ and $T^{\alpha}$ unchanged, the symmetric trace free tensor $ \chi^*_{\alpha \beta}$ can then be brought into the diagonal form $$\chi^*_{\alpha \beta} = (\lambda_1 + \lambda_2)\,\delta^0_{\alpha}\,\delta^0_{\beta}
+ \lambda_1\,\delta^1_{\alpha}\,\delta^1_{\beta}
+ \lambda_2\,\delta^2_{\alpha}\,\delta^2_{\beta}.$$
Cases in which the time-like eigendirection is not unique are excluded if we assume that the pull- back of $\chi^*_{\alpha \beta}$ to the hyperplane orthogonal to $T$ is positive definite (negative definite, with corresponding changes below, might also be considered). The representation above then implies
$(*) \quad$ [*the quadratic form $\chi^*_{\alpha \beta}\,X^{\alpha}\,X^{\beta}$ on the tangent space $T_{p'} {\cal T}$ is positive definite*]{}.
Conversely, consider on $T_{p'} {\cal T}$ the functions $h = k_{\alpha \beta}\,X^{\alpha}\,X^{\beta}$, $f = \chi^*_{\alpha \beta}\,X^{\alpha}\,X^{\beta}$, and the set $H = \{X \in T_p {\cal T}|\,\,h(X)= -1\}$. Assuming $(*)$, we conclude that $f(X) \rightarrow \infty$ on $H$ if the direction of $X$ approaches the null cone of $k$. The restriction of $f$ to $H$ thus assumes a minimum at some point $X_* \in H$ and by Lagrange’s method of underdetermined multipliers there exists a real constant $\lambda$ with $$\chi^*_{\alpha \beta}\,X_*^{\beta} = \lambda \,k_{\alpha \beta}\,X_*^{\beta},$$ so that $X_*$ is a time-like eigenvector of $\chi^*_{\alpha \beta}$. Because the restriction of the form $\chi^*_{\alpha \beta}\,X^{\alpha}\,X^{\beta}$ to the plane orthogonal to $X_*$ is positive definite, we are again in the situation which led to $(*)$.
We complement the assumption above by
$(**) \quad$ [*the mean extrinsic curvature $\chi$ is positive*]{}.
Conditions $(*)$ and $(**)$ then imply that the set ${\cal M}$ is locally convex at ${\cal T}$ in the following sense. Suppose that $({\cal M'}, g')$ is a smooth extension of the space-time $({\cal M}, g)$ for which ${\cal T}$ is an interior hypersurface. Any geodesics $\gamma(\tau)$ in this extension which is tangential to ${\cal T}$ at the point $\gamma(0) \in {\cal T}$ will only be tangential to ${\cal T}$ at first order and remain outside ${\cal M}$ for $0 < |\tau| << 1$ if its tangent vector $\gamma'(0)$ belongs to the set $\{X \in T_{ \gamma(0)}{\cal T}| \,k(X, X) \ge 0\}$ or is sufficiently close to it.
Assumptions $(*)$, $(**)$ are quite natural if we want to pose an initial boundary value problem for an interior part of an asymptotically flat solution whose boundary is sufficiently close to space-like infinity. They single out a future directed, time-like unit vector field tangential to ${\cal T}$ which is distinguished by the geometry of the problem. The question about the dependence of the invariants (\[chi-invariants\]) on the frame does not arise any longer. Moreover, conditions $(*)$, $(**)$ are preserved under small perturbations. If $\chi$ and the eigenvalues could be prescribed as boundary data the conditions could in fact be ensured during the development in time.
Whether the invariants (\[chi-invariants\]) can be used to encode the two degrees of freedom of the gravitational fields is not obvious. As mentioned before, the right hand side of the boundary condition (\[bdry-cond-\*\]) can be expressed completely in terms of the $e_3$-magnetic part of the conformal Weyl tensor. By the Codazzi equations the latter is given by certain covariant derivatives of $\chi_{\alpha \beta}$ on ${\cal T}$ so that the function $q$ is related to $\chi$ and the eigenvalues. Because the covariant derivatives involve the connection defined by $k_{\alpha \beta}$ it is not easy to see that the information encoded in $q$ can be extracted from the $\lambda_i$ and $\chi$. It may be that the situation is more easily analysed in the setting of [@kreiss:et:al:2009].
There is a test which may shed some light on this question. Consider one of the two approaches above and assume that the initial and boundary data are given such that the solutions will coincide for $0 \le t < \epsilon$ with the space-time setting of the flat, linear model case for some small $ \epsilon > 0$. Assume for $t \ge \epsilon$ the free functions entering the gauge condition and the specification of the boundary evolution to be given such that the gauge and the boundary reduce to that of the flat model case if the boundary data which refer to the gravitational degrees of freedom are given such as to imply the setting of the flat model case. Consider now a solution which is determined by some given but unspecified boundary data. We can ask then whether the additional requirement that the eigenvalues of $\chi_{\alpha \beta}$ coincide with those of the flat model case implies that the solution must be flat. A positive answer can be expected to indicate that the eigenvalues constitute suitable boundary data. Moreover, the argument which leads to this answer may give some insight into how initial boundary value problems which include these data must be formulated.
There remains, of course, the complicated question whether a time-like eigenvector or even an eigenframe of the second fundamental form can be implemented together with a condition of the type (\[gamma=0-on-T\]) and possibly a prescription of the eigenvalues of the second fundamental form in a formulation of a well posed initial boundary value problem. The answer requires a detailed analysis which will not be attempted here.
Conclusions
===========
The formulations of the boundary conditions considered in [@friedrich:nagy] and [@kreiss:et:al:2009] require the choice of a future directed, time-like unit vector field $T$ tangential to the boundary. With the resulting initial boundary conditions and data one arrives at well posed PDE problems which determine $ST$-vacuum-space-times locally in time near the initial hypersurface ${\cal S}$. The latter are unique apart from possible extensions into the future. Moreover, any $ST$-vacuum-space-time which is given in one of the gauges considered above can be constructed in this way locally in time.
The vector field $T$, for which no natural choice exists in general, is characterized indirectly and becomes explicitly available only after solving the equations. Problems arise if one wants to compare solutions pertaining to boundary conditions based on different choices of $T$ and on different boundary data.
The boundary conditions and data contain only very little direct information on the geometry on the boundary and the meaning of the boundary data is related to the choice of $T$. As a consequence, the question whether solutions determined by two different sets of boundary conditions and data are isometric can not be answered in terms of the boundary conditions and data alone. The complete solutions must be available along the boundary to perform a comparison.
This situation leads to awkward practical problems if gauge transformations need to be considered in the course of an evolution. It is an open question whether this is an intrinsic problem of the initial boundary value problem for Einstein’s field equations or whether there can be formulated, under general assumptions, initial boundary value problems for Einstein’s field equations which avoid these difficulties.
We pointed out a class of initial boundary value problem for which a time-like vector field $T$ is distinguished by the geometry of the boundary. Its defining property is stable under perturbations and the class is fairly large and quite important from the point of view of applications. If there existed formulations of well posed initial boundary problems based on this vector field and possibly on the invariants considered above the problem with geometric uniqueness arising in more general situations would not be present.
[**Acknowledgement.**]{} This article was inspired by seminars and discussions during the program on ‘Geometry, Analysis, and General Relativity’ at the Institut Mittag-Leffler. I should like to thank the institute for hospitality, financial support, and a most stimulating environment.
[11]{}
M. Alcubierre. Introduction to 3 + 1 Numerical Relativity. Oxford, Oxford University Press, 2008
R. Bartnik. Existence of maximal surfaces in asymptotically flat space-times. 94 (1984) 155 - 175.
R. Bartnik. Regularity of variational maximal surfaces. 161 (1988) 145 - 181.
R. Bartnik, J. Isenberg. The constraint equations. In: P. T. Chruściel, H. Friedrich (eds.): [*The Einstein equations and the large scale behaviour of gravitational fields.*]{} Birkhäuser, Basel, 2004.
S. Benzoni-Gavage, D. Serre. Multidimensional Hyperbolic Partial Differential Equations. Clarendon Press, Oxford, 2007.
Y. Choquet-Bruhat, R. Geroch. Global aspects of the Cauchy problem in general relativity. 14 (1969) 329 - 335.
P. T. Chruściel, E. Delay. On mapping properties of the general relativistic constraints operator in weighted function spaces, with application. , Marseille, 2003.
P. T. Chruściel, H. Friedrich (eds.). Birkhäuser, Basel, 2004.
J. Corvino. Scalar curvature deformation and a gluing construction for the Einstein constraint equations. 214 (2000) 137–189.
J. Corvino, R. Schoen On the Asymptotics for the Vacuum Einstein Constraint Equations. http://xxx.lanl.gov/abs/gr-qc/0301071
J. Ehlers, S. Kind. Initial-boundary value problem for the spherically symmetric Einstein equations for a perfect fluid. 10 (1993) 2123 - 2136.
Fourès-Bruhat, Y. Théorème d’existence pour certains systèmes d’équations aux dérivées partielles non linéaires. Acta Mathematica 88 (1952) 141 - 225.
H. Friedrich. . 13 (1996) 1451 - 1469.
H. Friedrich. . , 17 (1995) 125–184.
H. Friedrich. Is general relativity ‘essentially understood’ ? 15 (2006) 84 - 108.
H. Friedrich, G. Nagy. . 201 (1999) 619 - 655.
M. Holst, G. Nagy, G. Tsogtgerel. Far-from-constant mean curvature solutions of Einstein’s constraint equations with positive Yamabe metrics. 100 (2008) 161101.
H.-O. Kreiss, J. Lorenz. Initial-boundary value problems and the Navier-Stokes equations. Academic Presss, Boston, 1989.
H.-O. Kreiss, O. Reula, O. Sarbach, J. Winicour. Boundary conditions for coupled quasilinear wave equations with application to isolated systems. , 2009, to appear.
D. Maxwell. A class of solutions of the vacuum Einstein constraint equations with freely specified mean curvature. arXiv: 0804.0874 \[gr-qc\]
H. Ringström. The Cauchy Problem in General Relativity. ESI Lectures in Mathematics and Physics. European Mathematical Society. 2009.
[^1]: In [@friedrich:nagy] the signature $(1, -1, -1, -1)$ is used.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper investigates modifying an existing neural network architecture for static saliency prediction using two types of recurrences that integrate information from the temporal domain. The first modification is the addition of a ConvLSTM within the architecture, while the second is a conceptually simple exponential moving average of an internal convolutional state. We use weights pre-trained on the SALICON dataset and fine-tune our model on DHF1K. Our results show that both modifications achieve state-of-the-art results and produce similar saliency maps. Source code is available at <https://git.io/fjPiB>.'
bibliography:
- 'bmvc\_final.bib'
title: Simple vs complex temporal recurrences for video saliency prediction
---
Introduction {#sec:intro}
============
Visual saliency pertains to how an object or any piece of information may stand out from its surroundings. Detecting saliency is an integral part of how sentient organisms process information. We live in a world where the visual data we receive on a daily basis is immense and cluttered with noise; therefore, the brain has evolved in such a way that allows living organisms to focus their attention on the most relevant information, so as to function efficiently. Efforts in the computer vision community have been ongoing for many years to simulate this biological process artificially leading to the development of large-scale static gaze datasets, (e.g. SALICON [@SALICON]) and, more recently, dynamic gaze datasets (e.g. DHF1K [@Wang2018a]). Based on these datasets, model-driven approaches tackle the task of saliency prediction by estimating heatmaps of probabilities, where every probability corresponds to how likely it is that the corresponding pixel will attract human attention. Thanks to the availability of large-scale datasets, deep learning architectures have managed to significantly improve the accuracy achievable in this task (e.g. [@Wang2018a; @Pan2017; @gorji2018going; @SALICON; @pan2016shallow]).
Most scientific interest has so far been focused on image-based saliency models, with video saliency prediction gaining more traction in recent years with the introduction of large-scale video saliency datasets ([@Wang2018a; @Hollywood_UCF]). When it comes to extracting visual information from the temporal domain, ConvLSTMs have become increasingly popular, achieving state-of-the-art results in various computer vision tasks (e.g. [@CLSTM; @Wang2018a; @xu2018youtube]). In this work we augment a state-of-the-art architecture for image saliency [@Pan2017] by adding a ConvLSTM module within its internal structure, similar to [@Wang2018a; @gorji2018going]. More interestingly, we also test a much simpler method for temporal stability. We wrap a convolutional layer with a temporal exponential moving average (EMA) [@EMA] operation. Using this recurrence, the output will always be a smoothed average of its previous states. This method is already used in gradient descent with momentum [@momentumDL] to speed up convergence, replacing the current gradient with the exponential moving average of current and past gradients, derived from mini-batches of the data. To the best of our knowledge, this is the first time that this method has been applied within the architecture of a neural network.
Ablation studies are commonly used to better understand the performance impact of added components. Whilst this has merit, we propose that simple functions should also be used to investigate the necessity of complex modifications. To this end, in this work we consider both an elaborate ConvLSTM recurrence and a much simpler weighted average recurrence, and show that the simpler approach competes with the ConvLSTM on the task of video saliency.
Related Work
============
Video saliency prediction with deep neural networks has basically adapted to this task the architectures proposed for video action recognition. A first popular option are two-stream networks [@simonyan2014two], in which the motion information is encoded by a pre-computation of the optical flow and adding it in a separate tower from the RGB channels. This is the approach adopted by STSConvNet [@bak2017spatio]. This solution presents two important limitations: the computation overhead that is necessary to compute the optical flow, and the lack of temporal perspective further than the pairs of consecutive frames typically considered when computing optical flow. These shortcomings are partially addressed with the neural architectures where the temporal relation across frames is computed by a recurrent neural network (RNN) [@donahue2015long]. RNN-based deep models for saliency prediction have already been explored [@bazzani2016recurrent; @jiang2017predicting; @Wang2018a; @gorji2018going] and are the core of the state of the art solutions. Similarly to [@montes2016temporal] for activity detection, RMDN [@bazzani2016recurrent] combined the short-term memory encoded by C3D spatio-temporal convolutions [@tran2015learning] with a long short-term memory encoded by a plain LSTM. However, most current works have adopted a ConvLSTM layer as temporal recurrence, so that the recurrent layer would have a notion of space at a local scale. The OM-CNN model proposed in [@jiang2017predicting] fuses the RGB and optical flow from two-stream architecture with two ConvLSTMs. The authors of the largest dataset for video saliency prediction, the DHF1K (Dynamic Human Fixation 1K) dataset[@Wang2018a], trained a deep neural model based on ConvLSTM layers with attention (ACLNet). The authors of [@gorji2018going] exploit an existing model pre-trained for static saliency prediction, but with a more complex architecture composed of four branches fused with a ConvLSTM. Our model outperforms the presented state of the art with a simple architecture that only considers RGB frames as input. As in some of the referred works, we exploit a model pre-trained with static images and study its enhancement with two types a temporal recurrence.
Architecture
============
The adopted neural architecture follows an encoder-decoder scheme that processes the temporal recurrence in the bottleneck. The topology of both encoder-decoder is adopted from SalGAN [@Pan2017], the current top performing static saliency model on the DHF1K saliency benchmark. SalGAN encoder corresponded to the popular VGG-16 convolutional network [@VGG] designed and trained to solve an image classification task. At the decoder side, SalGAN used the same layers as in the encoder in reverse order, and interspersed by upsampling instead of pooling operations. The original SalGAN model was trained using a combination of adversarial and binary cross entropy (BCE) loss. Here, for simplicity, we use only BCE and term the resulting architecture *SalBCE*.
![Architecture of the our model. A frame is input to the model at each time step. Information encoded from the past frames persists via our recurrence that is located deeper in the network. The output is a per-frame saliency map.[]{data-label="fig:model_arch"}](TemporalEDmodel.png){width="65.00000%"}
We introduce a temporally aware component into the SalBCE network. This is either the addition of a ConvLSTM layer or an exponential moving average (EMA) applied on a pre-existing convolutional layer. Figure \[fig:model\_arch\] presents a schematic of our architecture.
ConvLSTM
--------
An LSTM is an autoregressive architecture that controls the flow of information in the network using 3 gates: update, forget, and output (Figure \[figEMA\_LSTM\], *left*). In ConvLSTMs [@CLSTM], the operations at each gate are convolutions. Temporal information is preserved in the cell state *C~t~* upon which gated element-wise operations are performed by the update and forget gate. The hidden state *H~t~* is concatenated with the input at each step and propagated through linear and non-linear operations at the gates. At each gate the current state *S~t~* of the model is passed through the ConvLSTM gates and the cell state *C~t~* and hidden state *H~t~* are updated. In the following equations ‘$\circ$’ represents the element-wise product, ‘$\ast$’ a convolution operation, ‘$\sigma$’ the sigmoid logistic function and ‘$\tanh$’ the hyperbolic tangent. The **update**, **forget**, and **output** gates can be written as: $$\begin{aligned}
u_t &= \sigma(W_u^S\ast S_t + W_u^H\ast H_{t-1}+W_u^C\circ C_{t-1} + b_u) \\
f_t &= \sigma(W_f^S \ast S_t + W_f^H \ast H_{t-1} + W_f^C \circ C_{t-1} + b_f) \\
o_t &= \sigma(W_o^S \ast S_t + W_o^H \ast H_{t-1} + W_o^C \circ C_{t-1} + b_o)\end{aligned}$$ and the new cell state $C_t$ and hidden state $H_t$ are then given by: $$\begin{aligned}
C_t &= f_t \circ C_{t-1} + u_t \circ \tanh(W_C^S \ast S_t + W_C^H \ast H_{t-1} + b_C) \\
H_t &= o_t \circ \tanh(C_t)\end{aligned}$$ where $W_*^*$ and $b_*$ are the model parameters. We added the ConvLSTM architecture at the bottleneck of our model, so that the input to the ConvLSTM is an encoded representation of the frame at time $t$. The output cell state $C_t$ is fed to the decoder for further processing that results in a saliency map. To obtain the saliency map, a $1\times1$ convolution is used at the final layer of the decoder, so as to filter out all channels but one. We sequentially pass video frames to the model as input and get a sequence of time-correlated saliency maps in the output. The ConvLSTM component learns to leverage the temporal features during training. The name we gave to this type of model is *SalCLSTM*.
![*(Left)* LSTM recurrence. Parametric operations are highlighted in yellow. *(Right)* EMA recurrence[]{data-label="figEMA_LSTM"}](lstm.png "fig:"){width="40.00000%"} ![*(Left)* LSTM recurrence. Parametric operations are highlighted in yellow. *(Right)* EMA recurrence[]{data-label="figEMA_LSTM"}](ema.png "fig:"){width="40.00000%"}
Exponential Moving Average
--------------------------
As an alternative approach, the exponential moving average (EMA) recurrence [@EMA] is added on a specified layer so that at time $t$ the convolutional state of this layer will be a decaying weighted average of the current and all previous states (Figure \[figEMA\_LSTM\], *right*). At time $t$ the convolutional layer $S_t$ outputs a state that is fed to the exponential weighted average. The output $E_t$ is then propagated further in the model. Note that there is a hyperparameter $\alpha$ that affects the impact of previous states on the current time step (the lower the value the higher the impact). $$E_t = \alpha S_t + (1-\alpha)E_{t-1}$$
This recurrence is straightforward to implement, especially compared to the ConvLSTM. We experimented with the placement of the EMA function at several different layers with $\alpha=0.1$. We name our model *SalEMA*. On the initial step, where there is no past information, the model runs like a static saliency map predictor.
Training
========
The parameters of SalCLSTM and SalEMA were estimated by backpropagating a pixel-wise content loss that compared the value of each pixel in the predicted saliency map with its corresponding pixel in the ground truth map. The total binary cross entropy loss was computed as the average of the individual binary cross entropies (BCE) over all pixels: $$L\textsubscript{\it BCE} = -\frac{1}{N}\sum_{\it n=1}^{N}P\textsubscript{\it n} \log(Q\textsubscript{\it n})+(1-P\textsubscript{\it n}) \log(1-Q\textsubscript{\it n})$$ where *P* represents the predicted saliency map and [*Q*]{} the ground truth saliency map.
SalCLSTM and SalEMA were not trained from scratch though, as the parameters of the encoder-decoder convolutional layers were adopted from SalBCE. SalBCE was trained for 27 epochs over the SALICON [@SALICON] dataset of still images using only the same BCE loss. We also utilized data augmentation techniques (mirroring and rotation of frames) which resulted in improved performance.
Our next step was adding recurrence that uses the intrinsic temporal information of video datasets and train it with the DHF1K dataset [@Wang2018a]. The DHF1K dataset [@Wang2018a] contains 700 annotated videos at 640$\times$360 resolution. We extracted frames at their original 30 fps rate, and resized them to 192$\times$256 resolution. We loaded them using a batch size of 10 frames from a single video at a time. By backpropagating the loss through time up to a maximum of 10 frames, we avoid exceeding memory capacity and potential vanishing or exploding gradients. We found it was necessary to initialize the ConvLSTM recurrence with the Xavier initialization method [@Xavier_Initialization], otherwise this model would converge to black images rather than saliency maps. This was likely due to oversaturation of the sigmoid activation layer. We trained all our models for 7 epochs, where we observed the loss reaching a plateau on our baseline. We used the Adam optimizer [@Adam] with a learning rate of $10\textsuperscript{-7}$.
Evaluation
==========
[Xclllll]{} & tuned on DHF1K & AUC-J &s-AUC & NSS & CC & SIM\
SalBCE (Baseline) && 0.874 & 0.724 & 2.047 & 0.382 & 0.268\
SalBCE && 0.880 & 0.632 & 2.285 & 0.420 & 0.339\
SalEMA && 0.883 & **0.734** & 2.144 & 0.400 & 0.276\
SalEMA && 0.883 & 0.685 & **2.402** & **0.435** & **0.349**\
SalCLSTM && **0.887** & 0.693 & 2.364 & **0.435** & 0.322\
The effect of temporal recurrences proposed for SalEMA and SalCLSTM was assessed with five different visual saliency metrics: Normalized Scanpath Saliency (NSS), Similarity Metric (SIM), Linear Correlation Coefficient (CC), AUC-Judd (AUC-J), and shuffled AUC (s-AUC). In all cases, a higher value corresponds to a better performance. The reader is referred to [@bylinskii2019different] for a detailed description of these metrics. The reported figures correspond to an average per video, that is, we first compute the metric on each frame, then average across all frames of each video, and we finally average across all videos.
We train and evaluate our models on three video saliency datasets, namely DHF1K [@Wang2018a], Hollywood-2 and UCF-sports [@Hollywood_UCF]. DHF1K is a large scale dataset with a high diversity of contents and variable length (from 400 frames to 1200 frames at 30fps). It includes 1000 videos, out of which 700 are publicly annotated, and 300 are withheld for testing purposes. In contrast to DHF1K, Hollywood-2 [@Hollywood-Origins] and UCF-sports [@soomro2014action] are limited to human actions and can be categorized as task-driven, given that the observers were explicitly asked to identify actions and scene context. These datasets were originally formed for the task of action recognition and were later adopted as a video saliency benchmark. Furthermore, both datasets have been divided into separate shots, so that no scene change occurs in the sequences that are fed into the models. Hollywood-2 is split into a training set of 3100 clips and a test set of 3559 clips, while UCF-sports has been split to a training set of 104 clips and a test set of 48 clips. These shots are much smaller in size than a DHF1K video sample, ranging from 40 frames to just a single frame per shot. We also use SALICON [@SALICON], a large-scale image saliency database, to set a baseline. DHF1K is used for experimenting with variations over the proposed models, as well as for comparison with the state of the art together with Hollywood-2 and UCF-sports.
The results in Table \[tab:SalEMAvsSalCLSTM\] indicate that the simple addition of EMA even without extra training does almost as well as a sophisticated ConvLSTM recurrence, and even improves it after being fine-tuned with the DHF1K training partition. EMA essentially performs a smoothing over the frames of the video by averaging. A possible explanation for why this boosts performance in video saliency is that saliency tends to be relatively consistent across frames, with the exception of rapid movements.
Encouraged by the positive results of our EMA modification at the bottleneck (layer 30), we explored more possible locations of the EMA function. In particular we tested its placement on: output (layer 61), decoder (layer 56), encoder (layer 7). We also implemented a variation that integrates EMA at two separate layers simultaneously, one in the encoder (7) and one in the decoder (56). In that case we set $\alpha$ to 0.3 at each location so as to not have an over-smoothing effect that would result in a significant lag at adapting to changes in the scene. Furthermore, in a video there can be spontaneous scene changes. In such instances, it would be optimal to have the EMA reset and forget all the previous states. However, EMA is not adaptive in this way, so we experimented with a skip connection that allows information to bypass this layer instead [@ResNet]. We also applied a second type of regularization, the dropout technique [@Dropout], at the convolutional layer right before the EMA layer. Dropout essentially turns off neurons with a preassigned probability (0.5) at each training step. This mitigates co-adaptation of neurons during training, allowing for clusters of neurons to learn independently. This way, at test time, we get the average from an ensemble of layers at location 30. The average of this ensemble pertains to spatial information, but since we are also using EMA, we get the moving average across the temporal dimension as well. The results reported in Table \[tab:SalEMA\] do not show a clear winning configuration across the five metrics metrics but, as NSS and CC are considered as the most appropriate ones to capture viewing behavior [@bylinskii2019different], we adopted SalEMA30 with dropout as our best configuration.
[Xclllll]{} **Model** & tuned on DHF1K &**AUC-J** &**s-AUC** & **NSS** & **CC** & **SIM**\
SalEMA30 && 0.883 & 0.734 & 2.144 & 0.400 & 0.276\
SalEMA30 && 0.883 & 0.685 & 2.402 & 0.435 & 0.349\
SalEMA30 (dropout)&& 0.886& 0.690& **2.495**& **0.450**& **0.360**\
SalEMA30 (residual)&& 0.875 &0.670 &2.274 &0.415 &0.339\
SalEMA61 && 0.884& **0.737**& 2.133& 0.399& 0.270\
SalEMA61 && **0.888** & 0.681 & 2.394 & 0.438 & 0.354\
SalEMA54 && 0.883& 0.734& 2.149& 0.401& 0.276\
SalEMA7 &&0.872 &0.656 &2.217 &0.409 &0.318\
SalEMA7&54 && 0.828& 0.561& 1.403& 0.366& 0.344\
Furthermore, we evaluated our two models on Hollywood-2 and UCF-sports [@Hollywood_UCF]. We compare our models to the current state-of-the-art as evaluated on the test split of the corresponding datasets by Wang [@Wang2018a]. Like ACLNet [@Wang2018a], our models were trained first for DHF1K in all cases, and later fine-tuned for the specific Hollywood-2 or UCF-Sports dataset. Table \[tab:SoA\] shows how, for DHF1K, SalEMA achieves the best performance compared to other models in the current benchmark across all metrics but s-AUC. On the other hand, SalCLSTM obtains the best results on all metrics for UCF-Sports and leads the performance on AUC-J, NSS and CC for Hollywood-2.
[lXlllll]{} **Dataset**&**Model** & **AUC-J** &**s-AUC** & **NSS** & **CC** & **SIM**\
&SalEMA & **0.890** & 0.667 & **2.573** & **0.449** & **0.465**\
&SalCLSTM & 0.887 & 0.693 & 2.364 & 0.435 & 0.322\
&ACLnet [@Wang2018a]& **0.890** & 0.601 & 2.354 & 0.434 & 0.315\
&SalGAN [@Pan2017] & 0.866 & **0.709** & 2.043 & 0.370 & 0.262\
&DVA [@DVA]& 0.860 & 0.595 & 2.013 & 0.358 & 0.262\
&SalEMA & 0.919 & 0.708 & 3.186 & 0.613 & 0.487\
&SalCLSTM & **0.933** & 0.715 & **3.499** & **0.672** & 0.530\
&ACLnet [@Wang2018a] & 0.913 & **0.757** & 3.086 & 0.623 & **0.542**\
&OM-CNN [@jiang2017predicting] & 0.887 & 0.693 & 2.313 & 0.446 & 0.356\
&DVA [@DVA]& 0.860 & 0.727 & 2.459 & 0.482 & 0.372\
&SalEMA & 0.906 & 0.740 & 2.638 & 0.544 & 0.431\
&SalCLSTM &**0.914** &**0.782** & **3.063** & **0.611** & **0.477**\
&ACLnet [@Wang2018a] &0.897 & 0.744 & 2.567& 0.51& 0.406\
&DVA [@DVA]& 0.872 & 0.725 & 2.311 & 0.439 & 0.339\
&OM-CNN [@jiang2017predicting] & 0.870 & 0.691 & 2.089 & 0.405 & 0.321\
A more detailed analysis between SalEMA and SalCLSTM was obtained by plotting the difference in their NSS and CC performance per video in the DHF1K validation set (100 videos). Concretely, we subtracted the metric value achieved by the SalCLSTM from that of SalEMA in each video and display the results in Figure \[NSS-comparison\]. This way, we can assess whether the two configurations end up producing similar results. In this case we would expect the variance to be low and the NSS difference to be close to zero most of the time. However, the results are sparse and diverge from video to video. This observation serves as evidence that the function approximated by the ConvLSTM is differs from that of an exponential moving average, despite its similar overall effectiveness.
![Per-video comparison between SalEMA and SalCLSTM using the NSS and CC metric on the DHF1K validation set. The values represent the margin by which a model’s performance differs from the other. []{data-label="NSS-comparison"}](NSS-DHF1K.png "fig:"){width="45.00000%"} ![Per-video comparison between SalEMA and SalCLSTM using the NSS and CC metric on the DHF1K validation set. The values represent the margin by which a model’s performance differs from the other. []{data-label="NSS-comparison"}](CC.png "fig:"){width="45.00000%"}
![We picked two samples that showed high divergence in performance between the two methods and visualized the predictions. SalEMA did much better than SalCLSTM on the sample displayed on the left side, while it was the opposite for the other sample. The images displayed correspond to intervals of 100 frames in the video.[]{data-label="QResults"}](QResults_crop.png){width="\textwidth"}
[XXXXXl]{} ****$\alpha$**** & **AUC-J** &**s-AUC** & **NSS** & **CC** & **SIM**\
*0.05* & **0.886** & 0.687 & 2.470 & 0.448 & 0.358\
*0.1* & **0.886** & **0.690** & **2.495** & **0.450** & **0.360**\
*0.2* & 0.885 & 0.688 & 2.476 & 0.446 & 0.358\
*0.3* & 0.884 & 0.685 & 2.451 & 0.442 & 0.356\
We also delved deeper into the Hollywood-2 dataset for potential clues that would explain the difference in performance. This dataset consists of very small shots, including even single-frame shots. In these cases we found that the ConvLSTM does much better than the EMA (by a margin of around $4$ NSS points). We also noticed, however, that in these cases the ground truths for the saliency maps correspond to a central Gaussian, despite the fact that other salient objects are present in other locations of the frame. Figure \[figHollywoodQ\] shows two examples in which the provided ground truth focuses in the center, although different faces appear in the image. In these examples, SalEMA captures these salient objects better, while SalCLSTM seems to focus on the center.
![Predictions from two Hollywood outliers where SalEMA performed particularly bad. The order corresponds to: SalEMA (left image), SalCLSTM (middle image), ground truth (right image). The ground truth appears aberrant, as it completely ignores human faces that are well-known to be salient objects.[]{data-label="figHollywoodQ"}](sample-1.png "fig:"){width="\textwidth"} ![Predictions from two Hollywood outliers where SalEMA performed particularly bad. The order corresponds to: SalEMA (left image), SalCLSTM (middle image), ground truth (right image). The ground truth appears aberrant, as it completely ignores human faces that are well-known to be salient objects.[]{data-label="figHollywoodQ"}](sample-2.png "fig:"){width="\textwidth"}
Finally, we experimented with the $\alpha$ hyperparameter by varying its value and also by making it trainable. Table \[tabAlpha\] shows relatively stable performance despite the variations on the value. We also had our model learn alpha on its own by introducing a trainable parameter $p$. To ensure that the resulting update equation represents a convex combination of the current features and previous state, $p$ is passed through a sigmoid so that the final value is constrained to $[0,1]$. The resulting recurrence is: $$E_t = \sigma(p) S_t + (1-\sigma(p))E_{t-1}$$
Whereas all other parameters of the model are set to a learning rate of $10\textsuperscript{-7}$, the learning rate of alpha was set to 0.1 and was trained separately for 3 epochs on SalEMA pretrained with $\alpha=0.1$. We set $\sigma(p)$ to 0.5 at the start of this tuning and by the end, it converges to 0.1477. The final performance was found to be approximately the same as the best model in Table \[tabAlpha\].
Conclusions
===========
This work has presented SalEMA and SalCLSTM, two variations of a convolutional neural network for video saliency prediction. Their main difference is how temporal recurrence is modelled, whether with a simple yet effective exponential moving average with a single parameter, or a convolutional LSTM that despite being adopted for many video sequence processing tasks, seems needlessly complex for this specific task of video saliency prediction. This indicates that, in some cases, components of more sophisticated models may just learn to approximate much simpler functions. It is likely that similar methods can be conceived of in other types of tasks as well.
On another note, ablation studies are a common practice for evaluating the contribution that an added component has on a model’s performance. We argue that there should be a more detailed effort in analyzing the behavior of deep architectures. Using predefined functions like the one presented in this work may shed more light on the necessity of a complex architecture.
#### Acknowledgements
This publication has emanated from research conducted with the financial support of Science Foundation Ireland (SFI) under grant number SFI/15/SIRG/3283 and SFI/12/RC/2289. This work has been developed in the framework of project TEC2016-75976-R, funded by the Spanish Ministerio de Economia y Competitividad and the European Regional Development Fund (ERDF).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper focuses on estimating probability distributions over the set of 3D rotations ($SO(3)$) using deep neural networks. Learning to regress models to the set of rotations is inherently difficult due to differences in topology between $\mathbb{R}^N$ and $SO(3)$. We overcome this issue by using a neural network to output the parameters for a matrix Fisher distribution since these parameters are homeomorphic to $\mathbb{R}^9$. By using a negative log likelihood loss for this distribution we get a loss which is convex with respect to the network outputs. By optimizing this loss we improve state-of-the-art on several challenging applicable datasets, namely Pascal3D+, ModelNet10-$SO(3)$ and UPNA head pose. All code used for this paper is available [online](https://github.com/Davmo049/Public_prob_orientation_estimation_with_matrix_fisher_distributions)'
author:
- |
David Mohlin\
KTH/Tobii\
`[email protected]`\
Gerald Bianchi\
Tobii\
`[email protected]`\
Josephine Sullivan\
KTH\
`[email protected]`\
bibliography:
- 'refs.bib'
title: Probabilistic orientation estimation with matrix Fisher distributions
---
Conclusion & Future work
========================
In this paper we have introduced a way to use neural networks to output probability distributions over the set of rotations with the matrix Fisher distribution. We show when optimizing the negative log likelihood of this distribution we end up with a convex loss. When applying this method on several datasets we get state-of-the-art performance. Our ablation studies show the relative robustness of the approach. Since the matrix Fisher distribution is unimodal it poorly models classes which have rotational symmetries. It could be interesting to try to create a loss supporting multimodal distributions while keeping the nice optimization properties of our loss. It could be possible to use these estimated probabilities for time tracking filters such as the one described in [@lee2018bayesian].
Broader Impact {#broader-impact .unnumbered}
==============
The methods described in this paper has obvious applications in fields which some consider ethically questionable such as for surveillance and military systems. One example could be determining heading for ships or airplanes for tactical planning. That being said, the orientation of objects is a fundamental property of objects in the real world and being able to accurately estimate this property should be helpful for many applications of either an ethically desirable or undesirable nature. In our opinion improving the techniques used for orientation estimation has a similar societal impact as improving the techniques used for classification or object detection.
The persons in the UPNA dataset are unlikely to be sampled from a uniform distribution of people across the world, for this reason one can not expect the reported performance to be accurate for the world population in general, that being said due to the small test size this reported performance might not reflect the average performance for any population. We do not believe this is an issue since models for predicting head pose which are deployed on a wider scale are very unlikely to use this dataset due to its small size and non-commercial licence. The method itself is not reliant on any population specific feature.
The main author is an industrial PhD student at Tobii. This research is funded by scholarship from Wallenberg AI, Autonomous systems and Software Program (WASP). Computational resources was provided by KTH.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper is essentially an exercise in studying the minima of a certain least squares optimization using the second partial derivative test. The motivation is to gain insight into an optimization-based solution to the problem of tracking human limbs using IMU sensors.'
author:
- Jad Nohra
title: Uniqueness of Minima of a Certain Least Squares Problem
---
Introduction
============
We study[^1] the minima of a specific least squares problem using the second partial derivative test. The problem’s origin is an optimization-based solution proposed in (Seel, Schauer, and Raisch 2012) to enable robust tracking of human limbs using IMU sensors. The original problem works with 6 dof rigid-body limbs in three dimensional space, but we shall instead work on a planar version of it to simplify the analysis. This is harmless given that our purpose is to study the uniqueness of minima: if the minima are not unique in the planar case problem, they are also not so for the spatial one.
Analysis
========
Problem Statement and Notation
------------------------------
We consider the nonlinear optimization problem of minimizing the sum of square errors objective function $O_n(\theta_1,\theta_2)$, where $n$ is the number of samples and $\theta_i$ are angles. Each sample $s$ is a six dimensional vector that determines the sample’s error and is denoted by[^2] $(\tensor[^s]{w}{_{11}},\tensor[^s]{w}{_{12}},\tensor[^s]{w}{_{13}},\tensor[^s]{w}{_{21}},\tensor[^s]{w}{_{22}},\tensor[^s]{w}{_{23}})$. Given the above, $O_n$ is given by
$$\begin{aligned}
O_n(\theta_1,\theta_2) &= \displaystyle \sum_{s=1}^{n} (\tensor[^s]{d}{}(\theta_1,\theta_2))^2, \\
\tensor[^s]{d}{}(\theta_1,\theta_2) &= p(\theta_1, \tensor[^s]{w}{_{1}}) - p(\theta_2, \tensor[^s]{w}{_{2}}), \\
p(\theta_i, \tensor[^s]{w}{_{i}}) &= [\tensor[^s]{w}{_{i1}} \sin(\theta_i) - \tensor[^s]{w}{_{i3}} \cos(\theta_i)]^2 + (\tensor[^s]{w}{_{i2}})^2.\end{aligned}$$
This corresponds to equation (1) in (Seel, Schauer, and Raisch 2012) up to variable names after the switching to a two-dimensional planar hinge problem where only a single angle has to be determined, moving to spherical coordinates with
$$j_i = \left( \begin{array}{c} \cos(\theta_i) \\ 0 \\ \sin(\theta_i) \end{array} \right),$$
and application of trigonometric identities.
We shall mostly skip the dependent variables for functions within expressions and move any indices that need to be retained under the letter. Per example we shall write $p_1$ to mean $p(\theta_1, w_1)$. Additionally, we shall employ Newton’s dot even for a partial derivative when there is no ambiguity.
We now procede to finding the stationary points and characterizing them using the second partial derivative test.
Insight by Computer
-------------------
Despite the simplicity of the problem, a direct attempt using symbolic mathematics software is undermined by the fact that the symbolic expressions generated for the relevant quantities for even the single-sample problem $O_1(\theta_1, \theta_2)$ are unworkable for a human as one can judge from their length and form in the appendix.
Analytical Solution of the Single-Sample Problem
------------------------------------------------
### Preliminaries
It is a fact that the sum of two equal period sinusoids is another sinusoid with the same period. Such sinusoids are known in Physics as $\textit{phasors}$, and the fact can be proved using trigonometric identities and is expressed by:
$$A \sin(x) - B \cos(x) = \sqrt{A^2+B^2} \cos(x + \text{atan}[{\frac{A}{B}}]).$$
Given this, let us define a number of functions that will prove helpful when applying the chain rule during differentiation:
$$\begin{aligned}
r(\theta_i, w_i) &= (\theta_i + \text{atan}[\frac{w_{i1}}{w_{i3}}]). \\
s(w_i) &= (w_{i1}^2 + w_{i3}^2). \\
t(\theta_i, w_i) &= w_{i1} \sin(\theta_i) - w_{i3} \cos(\theta_i), \\
&= \sqrt{s(w_i)} \cos(r(\theta_i, w_i)).\\
p(\theta_i, w_i) &= t^2(\theta_i, w_i) + (w_{i2})^2.\end{aligned}$$
Due to the abundance of $\sin(k\,r_i)$ and $\cos(k\,r_i)$ where $k$ is an integer in what follows, we shall shorten such terms in this manner:
$$\begin{aligned}
\cos(k\,r_i) &\twoheadrightarrow \left[{k \atop\text{co}_i}\right],\nonumber \\
\cos(k_1\,r_i)\,\cos(k_2r_j) &\twoheadrightarrow \left[{k_1 \atop\text{co}_i}{k_2 \atop\text{co}_j}\right],\nonumber\end{aligned}$$
and similarly for the sine function. In this notation, the known half-angle trigonometric identities are
$$\begin{aligned}
\left[{k \atop\text{co}_i}{k \atop\text{si}_i}\right] &= \frac{1}{2} \left[{2k \atop\text{si}_i}\right].\nonumber\\
\left[{k \atop\text{co}_i}\right]^2 &= \frac{1}{2} + \frac{1}{2} \left[{2k \atop\text{co}_i}\right].\nonumber\end{aligned}$$
### Partial Derivatives and Hessian Determinant
#### First Partial Derivatives
$$\begin{aligned}
\frac{\partial r_i}{\partial \theta_i} &= 1. \nonumber\\
\frac{\partial t_i}{\partial \theta_i} &= \sqrt{s_i}\,\dot{\cos}(r_i)\nonumber\\
&= \sqrt{s_i}\,[-\sin(r_i)]\,\dot{r_i}\nonumber\\
&= (-\sqrt{s_i})\,\sin(r_i). \nonumber\\
\frac{\partial p_i}{\partial \theta_i} &= 2\,t_i\,\dot{t_i}\nonumber \\
&= 2\,[\sqrt{s_i}\,\cos(r_i)]\,[(-\sqrt{s_i})\,\sin(r_i)]\nonumber \\
&= (-s_1)\,\sin(2r_i). \nonumber\\
\frac{\partial d}{\partial \theta_1} &= \frac{\partial (p_1-p_2)}{\partial \theta_1}\nonumber \nonumber\\
&= (-s_1)\,\sin(2r_1). \\
\frac{\partial d}{\partial \theta_2} &= \frac{\partial (p_1-p_2)}{\partial \theta_2}\nonumber \nonumber\\
&= (s_2)\,\sin(2r_2).\nonumber \\
\frac{\partial O_1}{\partial \theta_1} &= 2\,d\,\frac{\partial d}{\partial \theta_1}\nonumber \\
&= -2s_1\,d\,\sin(2r_1). \\
\frac{\partial O_1}{\partial \theta_2} &= 2\,d\,\frac{\partial d}{\partial \theta_2}\nonumber \\
&= 2s_2\,d\,\sin(2r_2).\end{aligned}$$
We are now in a position to write the Jacobian as
$$\text{Jac }{O_1} = \left(-2s_1\,d\left[{2\atop\text{si}_1}\right],\,\, 2s_2\,d\left[{2\atop\text{si}_2}\right] \right).$$
Let us additionally expand $\dfrac{\partial O}{\partial \theta_i}$ in terms of trigonometric functions for later use.
$$\begin{aligned}
d &= (t_1^2+w_{12}^2) - (t_2^2+w_{22}^2)\nonumber \\
&= \frac{s_1}{2} + \frac{s_1}{2} \left[2\atop\text{co}_1\right] + w_{12}^2 - \frac{s_2}{2} - \frac{s_2}{2} \left[2\atop\text{co}_2\right] - w_{22}^2\nonumber \\
&= \frac{1}{2} \left( c + s_1\left[2\atop\text{co}_1\right] - s_2\left[2\atop\text{co}_2\right] \right).\end{aligned}$$
where
$$c(w_1,w_2) = [ s_1+2w_{12}^2-s_2-2w_{22}^2 ].$$
Hence,
$$\begin{aligned}
\frac{\partial O_1}{\partial \theta_1}
&= (-2\,s_1)\left[\frac{1}{2} \left( c + s_1\left[2\atop\text{co}_1\right] - s_2\left[2\atop\text{co}_2\right] \right)\right]\left[2\atop\text{si}_1\right]\nonumber \\
&= (-s_1)\left[2\atop\text{si}_1\right]\left( c + s_1\left[2\atop\text{co}_1\right] - s_2\left[2\atop\text{co}_2\right]\right), \label{eq:note39}\\
&= (-s_1)\left( c\left[2\atop\text{si}_1\right] + \frac{1}{2}s_1\left[{4\atop\text{si}_1}\right] - s_2\left[{2\atop\text{si}_1}{2\atop\text{co}_2}\right]\right).\end{aligned}$$
Similarly, we have for $\dfrac{\partial O_1}{\partial \theta_2}$:
$$\begin{aligned}
\frac{\partial O_1}{\partial \theta_2}
&= (s_2)\left[2\atop\text{si}_2\right]\left( c + s_1\left[2\atop\text{co}_1\right] - s_2\left[2\atop\text{co}_2\right]\right),\\
&= (s_2)\left( c\left[2\atop\text{si}_2\right] +
s_1\left[{2\atop\text{co}_1}{2\atop\text{si}_2}\right] - \frac{1}{2}s_2\left[{4\atop\text{si}_2}\right] \right).\end{aligned}$$
#### Second Partial Derivatives
Making use of the various derivations from the last section we have:
$$\begin{aligned}
\frac{\partial^2 O_1}{\partial \theta_1^2}
&= \frac{\partial \left(-2s_1\,d\,\sin(2r_1)\right)}{\partial \theta_1}, \nonumber \\
&= -2s_1 \frac{\partial(d\,\sin(2r_1))}{\partial \theta_1}, \nonumber \\
&= -2s_1 \left[ d(2\cos(2r1) + (-s_1\sin(2r_1)(\sin(2r_1) \right], \nonumber \\
&= (-2s_1) \left(2d\left[2\atop\text{co}_1\right] -s_1\left[2\atop\text{si}_1\right]^2\right), \nonumber \\
&= (-2s_1) \left(2d\left[2\atop\text{co}_1\right] -s_1\left(1-\frac{1}{2}-\frac{1}{2}\left[4\atop\text{co}_1\right]\right)\right), \nonumber \\
&= (-2s_1) \left( \frac{-s_1}{2} + 2d\left[2\atop\text{co}_1\right] + \frac{s_1}{2}\left[4\atop\text{co}_1\right]\right).\end{aligned}$$
A full expansion of the expression along with the half-angle identity gives the alternative form
$$\frac{\partial^2 O_1}{\partial \theta_1^2}
= (-2s_1) \left( c\left[2\atop\text{co}_1\right] + s_1\left[4\atop\text{co}_1\right] - s_2\left[{2\atop\text{co}_1}{2\atop\text{co}_2}\right] \right).$$
In a similar vein we obtain the derivative relative to $\theta_2$:
$$\begin{aligned}
\frac{\partial^2 O_1}{\partial \theta_2^2} = (2s_2) \left( \frac{s_2}{2} + 2d\left[2\atop\text{co}_2\right] - \frac{s_2}{2}\left[4\atop\text{co}_2\right]\right), \\
= (2s_2) \left( c\left[2\atop\text{co}_2\right] - s_2\left[4\atop\text{co}_2\right] + s_1\left[{2\atop\text{co}_1}{2\atop\text{co}_2}\right] \right).\end{aligned}$$
It is obvious that the order of differentiation does not matter when considering the partial derivative with respect to $\partial\theta_1\partial\theta_2$. This derivative is given by
$$\begin{aligned}
\frac{\partial^2 O_1}{\partial\theta_2\partial\theta_1}
&= \frac{\partial \left( -2s_1\,d\sin(2r_1) \right)}{\partial\theta_2}, \nonumber \\
&= \left( -2s_1\sin(2r_1) \right) \frac{\partial d}{\partial\theta_2}, \nonumber \\
&= -2{s_1}{s_2} \left[{2\atop\text{si}_1}{2\atop\text{si}_2}\right].\end{aligned}$$
Finally, we denote the problem’s Hessian’s determinant by
$$\begin{aligned}
\lvert{H}\rvert
&= \underbrace{ \frac{\partial^2 O_1}{\partial \theta_1^2}\frac{\partial^2 O_1}{\partial \theta_2^2} }_{{ \lvert{\tensor*[_ {L}]{H}{}}\rvert }} - \underbrace{ {\left[\frac{\partial^2 O_1}{\partial \theta_1\partial \theta_2}\right]}^2 }_{{ \lvert{\tensor*[_ {R}]{H}{}}\rvert }}, \\
&= { \lvert{\tensor*[_ {L}]{H}{}}\rvert } - 4 \left( s_1 \left[2\atop\text{si}_1\right] \right)^2 \left( s_2 \left[2\atop\text{si}_2\right] \right)^2.\end{aligned}$$
### Stationary Points
#### Zeros of First Partials
The expressions of the first partials are the product of three terms, the values of $\theta_i$ at which we obtain zeros are the union of three sets. The first terms $(-s_1)$ and $(s_2)$ are not a function of $\theta_i$ so we consider the case $(s_1 = s_2 = 0)$ as degenerate going forward and ignore it; since $s_i$ is non-negative by definition, we assume from now on that
$$s_i > 0.$$
This leaves us with two sets per partial. Paying close attention to similarities and differences between the two partials, we denote the sets as
$$\begin{aligned}
{ \tensor*[^{\theta_{i}}_{1}]{Z}{} } &: \{\theta_i : \left[2\atop\text{si}_i\right] = 0 \}, \\
{ \tensor*[^{\theta_{}}_{2}]{Z}{} } &: \{(\theta_1,\theta_2) : \left( c + s_1\left[2\atop\text{co}_1\right] - s_2\left[2\atop\text{co}_2\right]\right) = 0 \}.\end{aligned}$$
As we shall see, the sets are more tersely described if we focus on $2r_i$ instead of on $\theta_i$ and this is harmless since the former are merely offset and scaled functions of the latter. Out of terseness as well we let
$$\begin{aligned}
\alpha_i(w_i) &= \text{atan}[\frac{w_{i1}}{w_{i3}}], \\
\beta(w) &= 2w_{12}^2-2w_{22}^2.\end{aligned}$$
${ \tensor*[^{\theta_{i}}_{1}]{Z}{} }$ has the simple solution
$$\begin{aligned}
\sin(2r_i) &= 0, \\
2r_i &= k\pi,\ (k \in \mathbb{Z}).\end{aligned}$$
#### First Stationary Set
The intersection of ${ \tensor*[^{\theta_{i}}_{1}]{Z}{} }$ directly gives us a first set of stationary points in the $(\theta_1,\theta_2)$ plane as
$$\begin{aligned}
{ \tensor*[^{}_{1}]{S}{} }
&: \{ ( - \alpha_1 + [\frac{k_1\pi}{2}], - \alpha_2 + [\frac{k_2\pi}{2}] ) : k_1,k_2 \in \mathbb{Z} \}.\end{aligned}$$
Let us conduct the second partial derivative test on this set. We first find the cases to consider. Since $2r_i = k\pi$, all sines involved are zero. On the other hand the cosines fall in the set $\{-1,1\}$ and we need to consider both cases. The first one is
$$\begin{aligned}
\cos(2r_i) &= -1, \\
2r_i = 2k_i\pi+\pi, \\
r_i = k_i\pi+\frac{\pi}{2}.\end{aligned}$$
The second is similar and we find that we need to consider the even and odd values of $k_i$ separately. This gives a total combination of four cases as follows:
$$\begin{dcases}
k_1 \text{ even }, k_2 \text{ even}, \\
k_1 \text{ odd }, k_2 \text{ even}, \\
k_1 \text{ even }, k_2 \text{ odd}, \\
k_1 \text{ odd }, k_2 \text{ odd}.
\end{dcases}$$
Treating all four cases simultaneously with the aid of an unorthodox but straightforward notation, we have that ${ \tensor[^{{ \tensor*[^{}_{1}]{S}{} }}]{\lvert{\tensor*[_ {L}]{H}{}}}{} \rvert }$ equals to[^3]
$$\begin{aligned}
& -s_1 ( 2c{\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } + 2s_1 {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{1}\\\scriptscriptstyle{1}\\\scriptscriptstyle{1}\\\scriptscriptstyle{1}\end{array} \kern-1.5ex\right]\kern-0.5ex } -2s_2 {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } )
s_2( 2c {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } - 2s_2 {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{1}\\\scriptscriptstyle{1}\\\scriptscriptstyle{1}\\\scriptscriptstyle{1}\end{array} \kern-1.5ex\right]\kern-0.5ex } +2s_1 {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } ) \\
&=
(-s_1 s_2) ( 2s_1 + 2c {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } -2s_2 {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{+1}\end{array} \kern-1.5ex\right]\kern-0.5ex } ) ( -2s_2 + 2c {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{-1}\end{array} \kern-1.5ex\right]\kern-0.5ex } + 2s_1 {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{+1}\end{array} \kern-1.5ex\right]\kern-0.5ex } )\end{aligned}$$
Expanding $c$ we get after a few steps that
$$\begin{aligned}
{ \tensor[^{{ \tensor*[^{}_{1}]{S}{} }}]{\lvert{\tensor*[_ {L}]{H}{}}}{} \rvert }
&= (-s_1 s_2) ({\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{4s_1}\\\scriptscriptstyle{0}\\\scriptscriptstyle{4s_1}\\\scriptscriptstyle{0}\end{array} \kern-1.5ex\right]\kern-0.5ex } + {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{-4s_2}\\\scriptscriptstyle{4s_2}\\\scriptscriptstyle{0}\\\scriptscriptstyle{0}\end{array} \kern-1.5ex\right]\kern-0.5ex } + {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{2\beta}\\\scriptscriptstyle{-2\beta}\\\scriptscriptstyle{2\beta}\\\scriptscriptstyle{-2\beta}\end{array} \kern-1.5ex\right]\kern-0.5ex }) ({\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{4s_1}\\\scriptscriptstyle{0}\\\scriptscriptstyle{-4s_1}\\\scriptscriptstyle{0}\end{array} \kern-1.5ex\right]\kern-0.5ex } + {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{-4s_2}\\\scriptscriptstyle{-4s_2}\\\scriptscriptstyle{0}\\\scriptscriptstyle{0}\end{array} \kern-1.5ex\right]\kern-0.5ex } + {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{2\beta}\\\scriptscriptstyle{2\beta}\\\scriptscriptstyle{-2\beta}\\\scriptscriptstyle{-2\beta}\end{array} \kern-1.5ex\right]\kern-0.5ex }), \nonumber \\
&= \begin{drcases} (- s_1 s_2) {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{+1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{-1}\\\scriptscriptstyle{+1}\end{array} \kern-1.5ex\right]\kern-0.5ex } {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{(4s_1-4s_2+2\beta)^2}\\\scriptscriptstyle{(4s_2-2\beta)^2}\\\scriptscriptstyle{(4s_1+2\beta)^2}\\\scriptscriptstyle{(-2\beta)^2}\end{array} \kern-1.5ex\right]\kern-0.5ex } \quad \end{drcases} \begin{array}{c}{< 0}\\{> 0}\\{> 0}\\{< 0}\end{array}.\end{aligned}$$
Since it is clear that ${ \tensor[^{{ \tensor*[^{}_{1}]{S}{} }}]{\lvert{\tensor*[_ {R}]{H}{}}}{} \rvert }$ amounts to zero unconditionally, we reach the result that ${ \tensor[^{{ \tensor*[^{}_{1}]{S}{} }}]{\lvert{\tensor*[_ {}]{H}{}}}{} \rvert }$ does not depend on the sample’s data. Ignoring degenerate cases, we see that we have saddle points for the first and fourth cases, and extrema for the second and third.
We continue the test and determine the nature of the extrema by examining the sign of $\frac{\partial^2 O_1}{\partial \theta_1^2}$, the expression of which we already worked out during the previous calculation and which is
$$\tensor[^{{ \tensor*[^{}_{1}]{S}{} }}]{\frac{\partial^2 O_1}{\partial \theta_1^2} }{} =
(- s_1) {\kern-0.5ex\left[\kern-1.5ex \begin{array}{c}\scriptscriptstyle{4s_1-4s_2+2\beta}\\\scriptscriptstyle{4s_2-2\beta}\\\scriptscriptstyle{4s_1+2\beta}\\\scriptscriptstyle{-2\beta}\end{array} \kern-1.5ex\right]\kern-0.5ex }.$$
Let us derive for the extremal cases the conditions under which we obtain maxima, which is the relevant case as we shall see later. For the second case we require:
$$\begin{aligned}
(-s_1) (4s_2-2 [2w_{12}^2 - 2w_{22}^2] ) < 0, \\
w_{21}^2+w_{23}^2 - w_{12}^2 + w_{22}^2 > 0, \\
w_{12}^2 < w_{21}^2+w_{23}^2+w_{22}^2.\end{aligned}$$
We obtain a similar result for the third case and we conclude that the extrema are maxima under the elegant condition
$$w_{12}^2 < \|w_2\|^2 \text{ and } w_{21}^2 < \|w_1\|^2. \label{eqn:note49}$$
Pictorially, we obtain a simple grid of saddle points and maxima for ${ \tensor*[^{}_{1}]{S}{} }$ as shown in figure \[fig:note50\].
; ; iin [0,...,3]{} [ (-1,i) – (-0.4,i); (-0.4,i) – (5.4,i); (5.4,i) – (6,i); ]{} iin [0,...,5]{} [ (i,-1) – (i,-0.4); (i,-0.4) – (i,3.4); (i,3.4) – (i,4); ]{}
in [0,2]{} [ in [0,2,4]{} [ at (,) ; ]{} in [1,3,5]{} [ at (,) ; ]{} ]{} (6.2,0) – (6.2,1); at (6.7,0.5) [$\frac{\pi}{2}$]{};
(4,-1.3) – (5,-1.3); at (4.5,-2.0) [$\frac{\pi}{2}$]{};
in [1,3]{} [ in [1,3,5]{} [ at (,) ; ]{} in [0,2,4]{} [ at (,) ; ]{} ]{}
at (8,3)
[l]{}
at (0,0) ;
\
at (0,0) ;
;
#### Second Stationary Set
Unlike the discrete set above, the second set of zeros already implicitly links the two variables into a curve which we shall study. By this we have that ${ \tensor*[^{}_{2}]{S}{} }$ is identical to ${ \tensor*[^{\theta_{}}_{2}]{Z}{} }$ and is characterized by
$$\cos(2r_1) = \frac{-c}{s_1} + \frac{s_2}{s_1}\cos(2r_2). \label{eq:np55}$$
Let us first study the set’s existence, which is obviously determined by the condition that the right-hand-side of the equation above is within $[-1,1]$.
We have that
$$\begin{aligned}
\frac{-c}{s_1}
&= \frac{ -s_1-2w_{12}^2+s_2+2w_{22}^2 }{s_1} \\
&= -1 + \frac{s_2}{s_1} - \frac{\beta}{s_1},\end{aligned}$$
and hence
$$\begin{aligned}
\underbrace{\left(\frac{-c}{s_1}\right) + \underbrace{ \left(\frac{s_2}{s_1}\right) \underbrace{\cos(2r_2)}_{[-1,1]}}_{\left[\dfrac{-s_2}{s_1},\dfrac{s_2}{s_1}\right]}}_{}\quad \nonumber \\
\left[-1-\frac{\beta}{s_1}, -1-\frac{\beta}{s_1}+\frac{2s_2}{s_1}\right].\end{aligned}$$
By this, we have a non-empty solution set if and only if
$$(-1-\frac{\beta}{s_1} >= -1) \text{ and } (-1-\frac{\beta}{s_1}+\frac{2s_2}{s_1} <= 1).$$
For the first predicate we require
$$\begin{aligned}
-1 - \frac{2w_{12}^2-2w_{22}^2}{s_1} \leq 1, \\
-w_{12}^2+w_{22}^2 \leq s_1, \\
w_{22}^2 \leq w_{11}^2+w_{13}^2+w_{12}^2.\end{aligned}$$
The second predicate leads to a similar calculation and we obtain, (harmlessly in our context) making the inequalities strict, here again exactly the elegant condition (\[eqn:note49\]). By this we see that for the second and third cases of ${ \tensor*[^{}_{1}]{S}{} }$, if the set contained minima it would also have to contain their related maxima, an altogether ‘degenerate’ situation compared to the minima coming from the curves generated by ${ \tensor*[^{}_{2}]{S}{} }$. It is because of this that we spared ourselves conducting the full second partial derivative test for that case. Condition (\[eqn:note49\]) is therefore a test for ‘correct’ and ‘useful’ samples.
In fact there is a more direct explanation for the condition. It is easy to see that $p_i$ has its minimum at $w_{i2}^2$ when $\cos(r_i)$ is zero and its maximum at $\|w_i\|^2$ when the cosine is at an extremum. At the same time, the error function $d$ can only be zero when there are $\theta_i$ such that $p_1=p_2$. But considering the ranges above, this is equivalent to our condition. In other words, samples that do not satisfy it are guaranteed to come from inexact measurements. It goes without saying that the converse is not necessarily true.
What the condition and its relation to ${ \tensor*[^{}_{1}]{S}{} }$ also tells us is that the ${ \tensor*[^{}_{2}]{S}{} }$ must be a set of minima. Despite that, we proceed to check this fact in a direct manner.
To do this, we feed the solution set into the relevant expressions by doing the substitution
$$s_2\cos(2r_2) = c + s_1\cos(2r_1).$$
By this we obtain for $\dfrac{\partial^2 O_1}{\partial \theta_1^2}$:
$$\begin{aligned}
& (-s_1) [2c { { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } + 2s_1 { { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } -2{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } (c + s_1{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } ) ] \\
=& (-s_1) [2s_1{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } -2s_1 (\frac{1}{2} + \frac{1}{2}{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } )] \\
=& (-s_1^2)({ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }-1).\end{aligned}$$
While for $\dfrac{\partial^2 O_1}{\partial \theta_2^2}$ we have:
$$\begin{aligned}
& 2c [s_2{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } }] - 2s_2[s_2{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } }] + 2s_1{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } [s_2{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } }] \\
=& \underbrace{2c^2 + 4c s_1{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }}_{A} + \underbrace{2{s_1}^2 { { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }^2}_{B} - \underbrace{2s_2 [s_2 { { \left[\kern-0.5ex{4}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } }]}_{C}.\end{aligned}$$
With
$$\begin{aligned}
B &= 2{s_1}^2(\frac{1}{2}+\frac{1}{2}{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }) \\
&= {s_1}^2 + {s_1}^2{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } },\end{aligned}$$
and
$$\begin{aligned}
C &= 2(2{s_2} { { \left[\kern-0.5ex{2}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } } {s_2} { { \left[\kern-0.5ex{2}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } } - {s_2}^2) \\
&= 2(2 [c+s_1 { { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }]^2 - {s_2}^2) \\
&= 4c^2 + 2B + 8c s_1 { { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } - 2 {s_2}^2 \\
&= 4c^2 + 2{s_1}^2+2{s_1}^2 { { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }+8 c s_1 { { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } - 2 {s_2}^2.\end{aligned}$$
We then reach for $\dfrac{\partial^2 O_1}{\partial \theta_2^2}$ the expression:
$$(-2c^2-s_1^2+2s_2^2) - (-s_1) (4c{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }+s_1{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }),$$
and consequently that
$$\begin{aligned}
{ \tensor[^{{ \tensor*[^{}_{2}]{S}{} }}]{\lvert{\tensor*[_ {L}]{H}{}}}{} \rvert } =& [(-s_1^2)({ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }-1)] \\ &\quad[(-2c^2-s_1^2+2s_2^2) - (-s_1) (4c{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }+s_1{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } })].\end{aligned}$$
For ${ \tensor[^{{ \tensor*[^{}_{2}]{S}{} }}]{\lvert{\tensor*[_ {R}]{H}{}}}{} \rvert }$ on the other hand, we obtain while skipping a number steps the expression
$$\begin{aligned}
& 4(s_1{ { \left[\kern-0.5ex{2}\atop{{\text{si}}_{1}}\kern-0.5ex\right] } })^2(s_2{ { \left[\kern-0.5ex{2}\atop{{\text{si}}_{2}}\kern-0.5ex\right] } })^2 \\
=& 2 (s_1{ { \left[\kern-0.5ex{2}\atop{{\text{si}}_{1}}\kern-0.5ex\right] } })^2 (2s_2^2-s_2^2-s_2^2{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } }) \\
=& s_1^2(1-{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }) s_2^2 (1-{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } }) \\
=& s_1^2[1-{\kern-0.5ex}{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }{\kern-0.2ex}] [(-2c^2-{\kern-0.5ex}s_1^2+{\kern-0.5ex}2s_2^2) -{\kern-0.5ex}s_1(4c{\kern-0.5ex}{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }{\kern-0.8ex}-s_1{\kern-0.7ex}{ { \left[\kern-0.5ex{4}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } })].\end{aligned}$$
A swift comparison show that in fact this is the negative of ${ \tensor[^{{ \tensor*[^{}_{2}]{S}{} }}]{\lvert{\tensor*[_ {R}]{H}{}}}{} \rvert }$ and therefore
$${ \tensor[^{{ \tensor*[^{}_{2}]{S}{} }}]{\lvert{\tensor*[_ {}]{H}{}}}{} \rvert } = 0.$$
The test being inconclusive for our second stationary point set, we seek a different path. Since this is a problem with a non-negative objective functions, let us check if there is a relation between the zeros of the objective function and the set at hand. The zeros of $O_1$ occur at
$$\begin{aligned}
\left( [s_1{ { \left[\kern-0.5ex{1}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } }^2 + w_{12}^2] - [s_2{ { \left[\kern-0.5ex{1}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } }^2 + w_{22}^2] \right)^2 = 0, \nonumber \\
s_1(\frac{1}{2}+\frac{1}{2}{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } + w_{12}^2 = s_2(\frac{1}{2}+\frac{1}{2}{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } } + w_{22}^2, \nonumber \\
\frac{s_1}{2}{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{1}}\kern-0.5ex\right] } } - \frac{s_2}{2}{ { \left[\kern-0.5ex{2}\atop{{\text{co}}_{2}}\kern-0.5ex\right] } } = -\frac{s_1}{2}-w_{12}^2+\frac{s_2}{2}+w_{22}^2.\end{aligned}$$
Multiplying the above by two gives exactly ${ \tensor*[^{}_{2}]{S}{} }$ and this confirms that this is a set of minima despite the inconclusive test.
Having done this, we proceed to a short study of the solution set’s curve. Since the cosine function is even, we notice that the curve must be symmetric about the origin of the $(2r_1,2r_2)$ plane as well as self-repeating with a period of $2\pi$. Therefore to study it we only need to focus on the $[0,\pi] \times [0,\pi]$ patch of the aforementioned plane. Within this patch, when a solution set exists, it is characterized by the monotone nature of $\arccos$ since we then have
$$2r_1 = \arccos\left(\frac{-c}{s_1} + \frac{s_2}{s_1}\cos(2r_2)\right).$$
A canonical form of (\[eq:np55\]) is
$$\cos{x} = u\cos(y)+v,\quad u > 0. \label{eq:s2canon}$$
Since the function is monotone it is sufficient for a qualitative but still exhaustive understanding of it to study its intersections with the lines that delimit the patch. They are the lines $x=0$, $x=\pi$, $y=0$ and $y=\pi$.
For $x=0$ we have an intersection if and only if $[1 = u\cos(y)+v]$ has solutions which leads to the condition $$\frac{\lvert{1-v}\rvert}{u} \leq 1.$$ A similar calculation gives the following set of four individual conditions:
$$\begin{aligned}
\frac{\lvert{1-v}\rvert}{u} &\leq 1, \\
\frac{\lvert{-1-v}\rvert}{u} &\leq 1, \\
\lvert{u+v}\rvert &\leq 1, \\
\lvert{-u+v}\rvert &\leq 1. \\\end{aligned}$$
It is not very difficult to see that these conditions can be independent but finding out which combinations of intersections are possible points towards a dull path. We created a computer script (using Python) to do the work. With the above conditions labelled as A,C,D,F, the results of the script are that 9 combinations are possible out of the total 16. This makes sense since this is the effect of the fact that $b$ is positive, allowing only half of the combinations, ignoring the last ‘degenerate’ one. The possible combinations along with witness parameters for $b$ and $c$ and witness curves are provided in figure (\[fig:script1\])
Finally note that a calculus analysis can provide the same result by showing that the slope of the curve within the patch is always positive.
Uniqueness of Solutions
-----------------------
Having obtained decent understanding of the single-sample problem and its characterizing solution curves we note that, in the absence of noise, insight into multi-sample problem can be gained by considering superpositions of the curves in the $(\theta_1,\theta_2)$ plane for each of the samples. We have seen that each ${ \tensor*[^{}_{2}]{S}{} }$ is symmetric about one of the ${ \tensor*[^{}_{1}]{S}{} }$ points. Due to this it becomes relevant to pose the question of existence of samples that have unequal data, equal ${ \tensor*[^{}_{1}]{S}{} }$ grids, but different ${ \tensor*[^{}_{2}]{S}{} }$ curves. When such samples intersect within one of the four sectors of a patch around a maximum, they necessarily intersect within all others creating three false minima for each true one. If additionally the intersection point can be quite arbitrary, the distances between the so obtained minima would be arbitrarily small and this would make the problem ill-posed. We proceed to find such two samples.
For any two samples $A$ and $B$, looking first at ${ \tensor*[^{}_{1}]{S}{} }$, we need $\sin(2{ \,{ \tensor[^{\scriptscriptstyle{A}}]{r}{_{i}} } }) = 0$ if and only if $\sin(2{ \,{ \tensor[^{\scriptscriptstyle{B}}]{r}{_{i}} } }) = 0$, which reduces to
$${\text{atan}[{\frac{\tensor[^{\scriptscriptstyle{A}}]{w}{_{i{1}}}}{\tensor[^{\scriptscriptstyle{A}}]{w}{_{i{3}}}} }] } = {\text{atan}[{\frac{\tensor[^{\scriptscriptstyle{B}}]{w}{_{i{1}}}}{\tensor[^{\scriptscriptstyle{B}}]{w}{_{i{3}}}} }] },\label{eqn:taneq}$$
Since we write $\text{atan}$ when we really mean $\text{atan2}$, this is equivalent to
$${\left({ \tensor[^{\scriptscriptstyle{A}}]{w}{_{i{1}}} }\atop{ \tensor[^{\scriptscriptstyle{A}}]{w}{_{i{3}}} }\right)} = \lambda_i{\left({ \tensor[^{\scriptscriptstyle{B}}]{w}{_{i{1}}} }\atop{ \tensor[^{\scriptscriptstyle{B}}]{w}{_{i{3}}} }\right)} + 2k_i\pi,\quad \lambda_i > 0. \label{eqn:lbdaeq}$$
This is then the condition for two samples to have identical ${ \tensor*[^{}_{1}]{S}{} }$ grids. Having obtained this, we turn to ${ \tensor*[^{}_{2}]{S}{} }$ looking for a family of curves passing through a specific point which we force to lie on the $x=y$ line to simplify finding an explicit example. Using the canonical form (\[eq:s2canon\]) we have at the common point $(t,t)$ that
$$\begin{aligned}
\cos(t) &= u\cos(t)+v, \\
\cos(t) &= \frac{v}{1-u}.\end{aligned}$$
Since the point is common to all curves its cosine must be equal in all of them so we have for two samples $A$ and $B$ that
$$\begin{aligned}
\frac{{ \tensor[^{\scriptscriptstyle{A}}]{v}{_{}} }}{1-{ \tensor[^{\scriptscriptstyle{A}}]{u}{_{}} }} &= \frac{{ \tensor[^{\scriptscriptstyle{B}}]{v}{_{}} }}{1-{ \tensor[^{\scriptscriptstyle{B}}]{u}{_{}} }}. \nonumber \\
{ \tensor[^{\scriptscriptstyle{B}}]{u}{_{}} } &= 1-{ \tensor[^{\scriptscriptstyle{B}}]{v}{_{}} } \left[ \frac{1+{ \tensor[^{\scriptscriptstyle{A}}]{u}{_{}} }}{{ \tensor[^{\scriptscriptstyle{A}}]{v}{_{}} }} \right] \label{eqn:note122pp}.\end{aligned}$$
Going back to the original form (\[eq:np55\]) and baking in the condition (\[eqn:lbdaeq\]) (while dropping the $2\pi$ period since we are working within one patch) we have the two relations:
$$\begin{aligned}
{ \tensor[^{\scriptscriptstyle{A}}]{v}{_{}} } &= -1 + \left(\frac{{\lambda_2}^2}{{\lambda_1}^2}\right)\frac{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{2}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }} - \frac{{ \tensor[^{\scriptscriptstyle{A}}]{\beta}{_{}} }}{({\lambda_1}^2){ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }},\\
{ \tensor[^{\scriptscriptstyle{A}}]{u}{_{}} } &= \left(\frac{{\lambda_2}^2}{{\lambda_1}^2}\right) \frac{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{2}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}.\end{aligned}$$
By (\[eqn:note122pp\]) we then require the follwing relation between the samples:
$$\frac{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{2}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }} = 1-
\left[\frac
{-1 + \frac{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{2}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }} - \frac{{ \tensor[^{\scriptscriptstyle{B}}]{\beta}{_{}} }}{{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}}}
{-1 + \left(\frac{{\lambda_2}^2}{{\lambda_1}^2}\right)\frac{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{2}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }} - \frac{{ \tensor[^{\scriptscriptstyle{A}}]{\beta}{_{}} }}{({\lambda_1}^2){ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}}
\right]
\left[
1 + \left(\frac{{\lambda_2}^2}{{\lambda_1}^2}\right) \frac{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{2}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}
\right] \label{eqn:note123}$$
We denote
$$\begin{aligned}
Q &= \frac{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{2}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}, \\
R &= \frac{{\lambda_2}^2}{{\lambda_1}^2},\end{aligned}$$
and tentatively set
$$R = \frac{1}{Q}$$
which simplifies (\[eqn:note123\]) to
$$R = 1-2\left[R-1-\frac{{ \tensor[^{\scriptscriptstyle{B}}]{\beta}{_{}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}\right] \left[\frac{{(\lambda_1}^2){ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}{-{ \tensor[^{\scriptscriptstyle{A}}]{\beta}{_{}} }}\right]$$
Again tentatively setting
$$R = 2$$
we obtain
$$\begin{aligned}
\frac{{(\lambda_1}^2){ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}{-{ \tensor[^{\scriptscriptstyle{A}}]{\beta}{_{}} }} &= 2\left[1-\frac{{ \tensor[^{\scriptscriptstyle{B}}]{\beta}{_{}} }}{{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} }}\right], \\
{ \tensor[^{\scriptscriptstyle{B}}]{\beta}{_{}} } &= { \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} } - \frac{1}{2{\lambda_1}^2}{ \tensor[^{\scriptscriptstyle{A}}]{\beta}{_{}} }.\end{aligned}$$
To finally obtain our example we set
$$\begin{aligned}
{ \tensor[^{\scriptscriptstyle{B}}]{s}{_{1}} } &= 4,\\
{\lambda_1}^2 &= 4,\\
{ \tensor[^{\scriptscriptstyle{A}}]{\beta}{_{}} } &= 4,\\
{\lambda_2}^2 &= 2,\\\end{aligned}$$
This determines the two samples almost completely and by inspection we set
$${ \tensor[^{\scriptscriptstyle{A}}]{w}{_{22}} } = { \tensor[^{\scriptscriptstyle{B}}]{w}{_{22}} } = 0,$$
to make the samples valid viz. satisfy (\[eqn:note49\]). The samples are then
$$\begin{aligned}
{ \tensor[^{\scriptscriptstyle{A}}]{w}{_{1}} },{ \tensor[^{\scriptscriptstyle{A}}]{w}{_{2}} } &= (\sqrt{12},\sqrt{2},2),\,(\sqrt{2},0,\sqrt{14}), \\
{ \tensor[^{\scriptscriptstyle{B}}]{w}{_{1}} },{ \tensor[^{\scriptscriptstyle{B}}]{w}{_{2}} } &= (\sqrt{3},\sqrt{\frac{7}{4}},1),\,(1,0,\sqrt{7}), \\\end{aligned}$$
and the ensuing skinned cat figure (\[fig:script2\]) confirms the result.
Despite the barbarism of the attempt, we easily found our example. This indicates that a continuum of such curves almost certainly exists but we leave a proof of this as subject for further work. Thence, we have shown that the problem is from the point of view of unique solutions, ill-posed or more exactly, only conditionally well-posed.
Appendix
========
Objective Function Intuition
----------------------------
Very few sentences are used in (Seel, Schauer, and Raisch 2012) to explain the derivation of the objective function. We here attempt an explanation with the following intuitive perspective.
Consider a canonical hinge setup. It naturally leads to conceptually thinking of the whole three-dimensional space as the union of two. The ‘hinge plane’ that is orthogonal to the hinge axis and in which the constrained bodies canonically lie, and it’s complement. The part of the angular velocities that determine the planar rotation within the hinge plane is given by projecting them onto the hinge axis. This exhausts the hinge’s single degree of freedom. The projection rests (also called rejections) must therefore be equal. Another way to think about this is that any difference in the rejections would ‘break’ the plane for the two canonical bodies would rotate such that they would not anymore lie in a common plane. By known vector projection and rejection formulas, with $a$ being the unit hinge axis, $\theta$ the hinge angle and $w_i$ being the angular velocities, we have:
$$\begin{split}
w_i = \underbrace{(w_i \cdot a) a}_{{ \tensor[_ {\parallel}]{w}{_{i}} }} + \underbrace{a \times (w_i \times a)}_{{ \tensor[_ {\perp}]{w}{_{i}} }}. \\
{ \tensor[_ {\parallel}]{w}{_{2}} } - { \tensor[_ {\parallel}]{w}{_{1}} } = \dot{\theta}a. \\
{ \tensor[_ {\perp}]{w}{_{2}} } - { \tensor[_ {\perp}]{w}{_{1}} } = 0.
\end{split}$$
All of this holding in any space, our objective function simply expresses the equality of rejections in two unknown spaces. Since the spaces are unknown, the magnitude of both vectors is taken and the spaces that make the equality hold determine the sought for transformations.
Expressions Generated by SymPy
------------------------------
$$\begin{split}
\frac{\partial O_1}{\partial \theta_1} &= (2 (w_{11} \sin(\theta_1)-w_{13} \cos(\theta_1)) (2 w_{11} \cos(\theta_1)\\
&\quad +2 w_{13} \sin(\theta_1)) (w_{12}^2-w_{22}^2+(w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1))^2-(w_{21} \sin(\theta_2)-w_{23} \cos(th2))^2)) \\
\frac{\partial^2 O_1}{\partial \theta_1^2} &=(2 (-2 w_{11} \sin(\theta_1)+2 w_{13} \cos(\theta_1)) (w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1)) (w_{12}^2-w_{22}^2+(w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1))^2-(w_{21} \sin(th2)-w_{23} \cos(th2)\\
&\quad )^2)+2 (w_{11} \sin(\theta_1)-w_{13} \cos(\theta_1))^2 (2 w_{11} \cos(\theta_1)\\
&\quad +2 w_{13} \sin(\theta_1))^2+2 (w_{11} \cos(\theta_1)+w_{13} \sin(\theta_1)\\
&\quad ) (2 w_{11} \cos(\theta_1)+2 w_{13} \sin(\theta_1)) (w_{12}^2-w_{22}^2+(w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1))^2-(w_{21} \sin(th2)-w_{23} \cos(th2))^2))
\\
|\text{Hess }{O_1}| &= (-4 (w_{11} \sin(\theta_1)-w_{13} \cos(\theta_1))^2 (2 w_{11} \cos(\theta_1)\\
&\quad +2 w_{13} \sin(\theta_1))^2 (w_{21} \sin(\theta_2)-w_{23} \cos(\theta_2)\\
&\quad )^2 (2 w_{21} \cos(\theta_2)+2 w_{23} \sin(\theta_2))^2+(2 (-2 w_{11} \sin(\theta_1)\\
&\quad +2 w_{13} \cos(\theta_1)) (w_{11} \sin(\theta_1)-w_{13} \cos(\theta_1)\\
&\quad ) (w_{12}^2-w_{22}^2+(w_{11} \sin(\theta_1)-w_{13} \cos(\theta_1)\\
&\quad )^2-(w_{21} \sin(\theta_2)-w_{23} \cos(\theta_2))^2)+2 (w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1))^2 (2 w_{11} \cos(\theta_1)+2 w_{13} \sin(\theta_1)\\
&\quad )^2+2 (w_{11} \cos(\theta_1)+w_{13} \sin(\theta_1)) (2 w_{11} \cos(\theta_1)\\
&\quad +2 w_{13} \sin(\theta_1)) (w_{12}^2-w_{22}^2+(w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1))^2-(w_{21} \sin(\theta_2)-w_{23} \cos(\theta_2)\\
&\quad )^2)) (-2 (-2 w_{21} \sin(\theta_2)+2 w_{23} \cos(\theta_2)\\
&\quad ) (w_{21} \sin(\theta_2)-w_{23} \cos(\theta_2)) (w_{12}^2-w_{22}^2+(w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1))^2-(w_{21} \sin(\theta_2)-w_{23} \cos(\theta_2)\\
&\quad )^2)+2 (w_{21} \sin(\theta_2)-w_{23} \cos(\theta_2))^2 (2 w_{21} \cos(\theta_2)\\
&\quad +2 w_{23} \sin(\theta_2))^2-2 (w_{21} \cos(\theta_2)+w_{23} \sin(\theta_2)\\
&\quad ) (2 w_{21} \cos(\theta_2)+2 w_{23} \sin(\theta_2)) (w_{12}^2-w_{22}^2+(w_{11} \sin(\theta_1)\\
&\quad -w_{13} \cos(\theta_1))^2-(w_{21} \sin(\theta_2)-w_{23} \cos(\theta_2))^2)))
\end{split}$$
References {#references .unnumbered}
==========
Seel, Thomas, Thomas Schauer, and J[ö]{}rg Raisch. 2012. “Joint Axis and Position Estimation from Inertial Measurement Data by Exploiting Kinematic Constraints.” In *Control Applications (CCA), 2012 IEEE International Conference on*, 45–49. IEEE.
SymPy Development Team. 2014. *SymPy: Python Library for Symbolic Mathematics*. <http://www.sympy.org>.
[^1]: As suggested to us by Teodor Cioacă and Horea Cărămizaru, whom we both thank.
[^2]: We shall freely drop the sample index when convenient.
[^3]: By ${ \tensor[^{{ \tensor*[^{}_{1}]{S}{} }}]{\lvert{\tensor*[_ {L}]{H}{}}}{} \rvert }$ We mean ${ \tensor[^{}]{\lvert{\tensor*[_ {L}]{H}{}}}{} \rvert }$ evaluated at points in ${ \tensor*[^{}_{1}]{S}{} }$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The chiral $p$-wave superconducting state is comprised of spin triplet Cooper pairs carrying a finite orbital angular momentum. For the case of a periodic lattice, calculating the net magnetisation arising from this orbital component presents a challenge as the circulation operator $\hat{\bf{r}} \times \hat{\bf{p}}$ is not well defined in the Bloch representation. This difficulty has been overcome in the normal state, for which a modern theory is firmly established. Here, we derive the extension of this normal state approach, generating a theory which is valid for a general superconducting state, and go on to perform model calculations for a chiral $p$-wave state in . The results suggest that the magnitude of the elusive edge current in is finite, but lies below experimental resolution. This provides a possible solution to the long-standing controversy concerning the gap symmetry of the superconducting state in this material.'
author:
- Joshua Robbins
- 'James F. Annett'
- Martin Gradhand
bibliography:
- 'orb.bib'
title: Modern Theory for the Orbital Moment in a Superconductor
---
An unconventional superconducting state exhibits a lower order of symmetry than the $s$-wave singlet pairing observed in conventional BCS superconductors. An example of this is the chiral $p$-wave paired state, which arises in conjunction with a breaking of time-reversal symmetry at the superconducting transition [@kal1]. Such a state consists of spin triplet Cooper pairs carrying a finite orbital angular momentum. The symmetry breaking associated with this pairing theoretically facilitates a number of new and exotic phenomena, such as the Kerr effect [@xia1; @xia2] and edge currents [@mats; @stone].
Of major significance in the study of this class of materials is the topological nature of superconducting states with chiral symmetry [@qi1; @sato1]. A chiral edge mode in a topological superconducting state would support a protected Majorana bound state confined to the edges of the sample [@sato; @leij]. The existence of these bound states is inextricably linked to the orbital moment of the spin triplet Cooper pairs, as both phenomena arise from the chiral nature of the superconducting order parameter.
Given this interest, it is surprising that there currently exists no general framework with which to calculate the total orbital magnetic moment in a superconducting state. The orbital angular momentum carried by the Cooper pairs should, in principle, lead directly to an orbital magnetisation in the superconducting lattice. Contributions to the magnetic moment are expected from edge currents [@mats; @stone], while bulk contributions are also predicted in multi-orbital systems [@ann1]. The goal of this letter is to present a general approach to this problem.
A rigorous theory for the orbital magnetisation in a normal state periodic lattice has been defined previously [@thon1; @cer1]. Obtaining a formalism of this nature had been an outstanding issue due to the problem of evaluating the circulation operator ($\hat{\bf{r}} \times \hat{\bf{p}}$) in a Bloch representation. In an infinite lattice, the position operator ($\hat{\bf{r}}$) is unbound and the cell-periodic Bloch functions ($u_{\bf{k}} (\bf{r})$) are not localised. The coexistence of these two factors means that the position expectation values of Bloch wavefunctions cannot be evaluated directly. The normal state theory was developed by reformulating the problem in a localised basis, the Wannier representation [@cer1; @mar]. Here, we extend this formalism to the orbital magnetisation in the superconducting state.
The new theory for the orbital moment in an infinite periodic lattice has previously been applied to cases of insulators and metals, for both single-band and multi-band configurations [@cer1]. The derivation introduced two distinct contributions to the total moment, referred to as the “local" and “itinerant" circulations. The terms correspond to orbital moments generated by the movement of the centres of mass of orbital wavefunctions (itinerant), and the moment due to self-rotation about their centres of mass (local).
Extending this theory to a general superconducting state, we obtain equivalent expressions for the local and itinerant contributions. We further break down the local contribution by performing a tight-binding expansion, extracting the purely on-site component defined previously [@ann1]. The formalism developed here will then be applied to a multi-band tight-binding model of .
We begin our analysis by giving an outline of the derivation of the orbital moment in the superconducting state. In second quantised form, the operator for the total orbital angular momentum in an arbitrary state is given by:
$$\label{eq1}
\hat{L}_z = \int d {\bf{r}} \, {\hat{a}}^{\dagger} \! ({\bf{r}}) \, \hat{l}_z \, {\hat{a}} ({\bf{r}})$$
where $a^\dagger$, $a$ are Fermi creation and annihilation operators, respectively, and $\hat{l}_z = [\hat{\bf{r}} \times \hat{\bf{p}}]_{z}$. The total orbital magnetic moment is then given by $\gamma \langle \hat{L}_z \rangle$, where $\gamma = -e/(2m_e)$.
In order to obtain a second quantised operator valid for a gapped state, we perform the Bogoliubov-Valatin transformation on the creation and annihilation operators [@kett]:
\[eq1a\]
$$\begin{aligned}
{\hat{a}} &= \sum_{{n{\bf{k}}}} {\theta}_{{n{\bf{k}}}} ({\bf{r}}) \hat{\gamma}_{{n{\bf{k}}}} + {\chi}_{{n{\bf{k}}}}^{*} ({\bf{r}}) \hat{\gamma}_{{n{\bf{k}}}}^{\dagger} \\
{\hat{a}}^{\dagger} &= \sum_{{n{\bf{k}}}} {\theta}_{{n{\bf{k}}}}^{*} ({\bf{r}}) \hat{\gamma}_{{n{\bf{k}}}}^{\dagger} + {\chi}_{{n{\bf{k}}}} ({\bf{r}}) \hat{\gamma}_{{n{\bf{k}}}}\end{aligned}$$
where $n$ is the number of spin-resolved bands, ${\bf{k}}$ is the Bloch wavevector and $\gamma^{\dagger},\gamma$ are quasiparticle creation and annihilation operators. The functions ${\theta}$, ${\chi}$ are, respectively, electron and hole components of a Bloch-type wavefunction $\psi$. This transformation recasts the equation into an expression for the orbital moment arising from Bogoliubov quasiparticles which appear as excitations in a superconductor.
To obtain the total orbital moment in an arbitrary superconducting state, we compute the expectation value of the transformed operator by applying the following relations:
\[eq1b\]
$$\begin{aligned}
\langle \hat{\gamma}_{{n{\bf{k}}}}^{\dagger} \hat{\gamma}_{n'{\bf{k}}'} \rangle &= \delta_{nn'} \delta_{{\bf{k}}{\bf{k}}'} f_{{n{\bf{k}}}} \\
\langle \hat{\gamma}_{{n{\bf{k}}}} \hat{\gamma}_{n'{\bf{k}}'}^{\dagger} \rangle &= \delta_{nn'} \delta_{{\bf{k}}{\bf{k}}'} (1-f_{{n{\bf{k}}}}) \\
\langle \hat{\gamma}_{{n{\bf{k}}}} \hat{\gamma}_{n'{\bf{k}}'} \rangle &= \langle \hat{\gamma}_{{n{\bf{k}}}}^{\dagger} \hat{\gamma}_{n'{\bf{k}}'}^{\dagger} \rangle = 0\end{aligned}$$
where $f$ is the Fermi-Dirac function. The transformed equation and it’s associated operators then take the following form:
$$\label{eq2}
\langle \hat{L}_z \rangle = \sum_{n{\bf{k}}} \int d {\bf{r}} \, {\psi}_{n{\bf{k}}}^{\dagger} ({\bf{r}}) \, {\bf{{L}}}_z \, {\psi}_{n{\bf{k}}} ({\bf{r}})$$
$${\bf{{L}}}_z=\begin{pmatrix} \hat{l}_z f_{n{\bf{k}}} & 0 \\ 0 & -\hat{l}_{z}^{*} (1-f_{n{\bf{k}}}) \end{pmatrix} , \, {\psi}_{n{\bf{k}}} ({\bf{r}})= \begin{pmatrix} {\theta}_{n{\bf{k}}} ({\bf{r}}) \\ {\chi}_{n{\bf{k}}} ({\bf{r}}) \end{pmatrix} \nonumber$$
At this point, we can defer to the derivation laid out for the normal state in terms of Wannier orbitals [@thon1; @cer1], where we now consider two-component Wannier wavefunctions containing electron and hole amplitudes in correspondence with the Bloch-type eigenfunctions. We also introduce the cell-periodic components of the Bloch wavefunctions, $(u_{n{\bf{k}}} ({\bf{r}}),v_{n{\bf{k}}} ({\bf{r}}))={\textrm{e}}^{-i {\bf{k}}\cdot{\bf{r}}}({\theta}_{n{\bf{k}}} ({\bf{r}}),{\chi}_{n{\bf{k}}} ({\bf{r}}))$.
Following the steps of this derivation, we are able to remove the dependence of Eq. (\[eq2\]) on the problematic operators $\hat{\bf{r}}$ and $\hat{\bf{v}}$. Performing a Fourier transform on the real space expressions obtained via this approach, we obtain two reciprocal space expressions which generate the orbital magnetisation via Brillouin zone integrals:
$$\label{eq4}
{\bf{M}}_{\textrm{LC}}=-\gamma \textrm{Im}\left[ \int_{BZ} \frac{d {\bf{k}}}{(2 \pi)^3} \sum_{n} \Big( \langle \partial_{\bf{k}} u_{n{\bf{k}}} | \times \hat{H}_{\bf{k}} | \partial_{\bf{k}} u_{n{\bf{k}}} \rangle f_{n{\bf{k}}} - \langle \partial_{\bf{k}} v_{n{\bf{k}}} | \times \hat{H}_{\bf{k}}^{*} | \partial_{\bf{k}} v_{n{\bf{k}}} \rangle (1-f_{n{\bf{k}}}) \Big) \right]$$
$$\label{eq5}
{\bf{M}}_{\textrm{IC}}=\gamma \textrm{Im}\left[ \int_{BZ} \frac{d {\bf{k}}}{(2 \pi)^3} \sum_{n} E_{n{\bf{k}}} \Big( \langle \partial_{\bf{k}} u_{n{\bf{k}}} | \times | \partial_{\bf{k}} u_{n{\bf{k}}} \rangle f_{n{\bf{k}}} + \langle \partial_{\bf{k}} v_{n{\bf{k}}} | \times | \partial_{\bf{k}} v_{n{\bf{k}}} \rangle (1-f_{n{\bf{k}}}) \Big) \right]$$
where LC and IC refer to local and itinerant circulations, as defined previously [@thon1], and the total magnetisation is given by ${\bf{M}} = {\bf{M}}_{\textrm{LC}} + {\bf{M}}_{\textrm{IC}}$. We have divided by the unit cell volume, to convert from the magnetic moment to magnetisation, and also introduced Dirac notation where, crucially, the expectation values taken in equations (\[eq4\]) and (\[eq5\]) are now evaluated for the unit cell only. These equations constitute our central result: a comprehensive framework for computing the total orbital magnetisation in a general bulk superconducting state.
The cell-periodic functions are obtained through self-consistent calculation of the Bogoliubov-de Gennes (BdG) equation:
$$\label{eq3}
\begin{pmatrix} \hat{H}_{\bf{k}} ({\bf{r}}) & \Delta ({\bf{r}}) \\ \Delta^{\dagger} ({\bf{r}}) & -\hat{H}_{-\bf{k}}^{*} ({\bf{r}}) \end{pmatrix} \begin{pmatrix} u_{n{\bf{k}}} ({\bf{r}}) \\ v_{n{\bf{k}}} ({\bf{r}}) \end{pmatrix} = E_{n{\bf{k}}} \begin{pmatrix} u_{n{\bf{k}}} ({\bf{r}}) \\ v_{n{\bf{k}}} ({\bf{r}}) \end{pmatrix}$$
where $\hat{H}_{\bf{k}}$ is the ${\bf{k}}$-dependent normal state Hamiltonian [@grad1]. The gap function ($\Delta$) enforces the symmetry of the superconducting state in question.
In order to perform model calculations, we must recast the Bloch equations into a tight-binding representation. Performing the $k$-derivatives in (\[eq4\]) and (\[eq5\]) and expanding in terms of the Bloch wavefunctions, we obtain:
\[eq9\]
$$\begin{aligned}
\partial_{\bf{k}} u_{n{\bf{k}}} ({\bf{r}}) &= \textrm{e}^{-i {\bf{k}} \cdot {\bf{r}}} \big(\partial_{\bf{k}} \theta_{n{\bf{k}}} ({\bf{r}}) - i {\bf{r}} \theta_{n{\bf{k}}} ({\bf{r}}) \big) \\ \partial_{\bf{k}} v_{n{\bf{k}}} ({\bf{r}}) &= \textrm{e}^{-i {\bf{k}} \cdot {\bf{r}}} \big(\partial_{\bf{k}} \chi_{n{\bf{k}}} ({\bf{r}}) - i {\bf{r}} \chi_{n{\bf{k}}} ({\bf{r}}) \big)\end{aligned}$$
Substituting Eqs. (\[eq9\]) into (\[eq5\]), we find one term containing ${\bf{r}} \times {\bf{r}}$, which will vanish. For the local component, however, this does not occur and we can split the equation into two parts of the form $\partial_{\bf{k}} \theta_{n{\bf{k}}}^{*} \times \hat{H} \partial_{\bf{k}} \theta_{n{\bf{k}}}$ and $\theta_{n{\bf{k}}}^{*} [ {\bf{r}} \times \hat{H} {\bf{r}} ] \theta_{n{\bf{k}}}$ respectively. Using the standard definition of the velocity operator, ${\bf{r}} \times \hat{H} {\bf{r}}$ can be re-written as $-i \hat{l}_z$.
Having re-written Eq. (\[eq5\]) in terms of $\theta$, $\chi$ we can subsequently apply a general tight-binding expansion of the Bloch wavefunction via:
$$\label{eq12}
\begin{pmatrix}
{\theta}_{n{\bf{k}}} ({\bf{r}}) \\ {\chi}_{n{\bf{k}}} ({\bf{r}}) \end{pmatrix} = \sum_{L,{\bf{R}}} \textrm{e}^{i {\bf{k}} \cdot {\bf{R}}} \begin{pmatrix} u_{nL} ({\bf{k}}) \\ v_{nL} ({\bf{k}}) \end{pmatrix} \phi_L ({\bf{r}}-{\bf{R}})$$
where $L$ is the orbital index and $\phi_L$ is the corresponding orbital wavefunction. Substituting Eq. (\[eq12\]) into (\[eq4\]), we obtain the following terms:
$$\label{eq10}
{\bf{M}}_{\textrm{LC}}^{\textrm{(1)}}=-\gamma \textrm{Im} \bigg[\sum_{nLL'} \int_{BZ} \frac{d {\bf{k}}}{(2 \pi)^3} \Big( \partial_{\bf{k}} u_{nL}^{*}({\bf{k}}) \times \hat{H}_{LL'}({\bf{k}}) \partial_{\bf{k}} u_{nL'}({\bf{k}}) f_{n{\bf{k}}} - \partial_{\bf{k}} v_{nL}^{*}({\bf{k}}) \times \hat{H}_{LL'}^{*}({\bf{k}}) \partial_{\bf{k}} v_{nL'}({\bf{k}}) (1-f_{n{\bf{k}}}) \Big) \bigg]$$
$$\label{eq11}
{\bf{M}}_{\textrm{LC}}^{\textrm{(2)}}=\gamma \textrm{Re} \Bigg[\sum_{nLL'} \int_{BZ} \frac{d {\bf{k}}}{(2 \pi)^3} \Big( u_{nL}^{*}({\bf{k}}) \left( \hat{l}_{z,LL'} \right) u_{nL'}({\bf{k}}) f_{n{\bf{k}}} + v_{nL}^{*}({\bf{k}}) \left( \hat{l}_{z,LL'}^{*} \right) v_{nL'}({\bf{k}}) (1-f_{n{\bf{k}}}) \Big) \Bigg]$$
The eigenvectors ($u_{nL}$, $v_{nL}$) are computed by solving Eq. (\[eq3\]) self-consistently in the tight-binding basis. The terms $\hat{H}_{LL'}$ represent the matrix elements of the tight-binding Hamiltonian. Similarly, the matrix elements $\hat{l}_{z,LL'}$ correspond to the orbital angular momentum expectation values of the orbitals contained in the tight-binding basis. These elements can be calculated by direct consideration of the spherical harmonics of the basis.
The second term, , is identical to the purely on-site orbital moment computed previously [@ann1]. We therefore label 2 as the “on-site" component, and continue to refer to as the local contribution.
Now that the framework for calculating the magnetic moment has been set up, we briefly outline the model for that will be used to perform the calculations. The superconducting state of is widely believed to exhibit chiral $p$-wave superconductivity below its transition temperature of 1.5 K [@rev1; @rev2], such that the superconducting order parameter is given by ${\bf{d}} \! \sim \! (\sin k_x \pm i \sin k_y){\bf{\hat{z}}}$. This hypothesis is supported by measurements of spin susceptibility [@ish2; @duf1] and indirect observations of time-reversal symmetry breaking at T$_c$ [@luke1]. In addition, a finite Kerr shift has been measured in this material [@xia1], providing direct evidence of a macroscopic orbital magnetisation in the bulk superconducting state.
The classification of as a $p$-wave superconductor remains a point of controversy, however, as phenomenological and quasiclassical approaches have predicted that large edge currents should accompany the single-band chiral superconducting state [@sig1; @mats; @hua]. Such currents have remained elusive despite years of intensive experimental work [@kirt; @hicks1; @curr]. A large surface-based current would provide a significant contribution to the total orbital magnetisation. By generating a full theoretical description of the orbital magnetic moment and its various sources in such a state, we provide a vital avenue through which we can attempt to reconcile these observations with theory.
We have constructed a three-dimensional tight-binding Hamiltonian consisting of three Ru 4$d$ orbitals ($d_{xy}$, $d_{xz}$ and $d_{yz}$) contributing to the normal state Fermi surface, resulting in a 2D band (denoted $\gamma$) and two quasi-1D bands ($\alpha$ and $\beta$). In many approaches to modelling , the model is formulated such that superconductivity arises primarily on $\gamma$, with accompanying gaps on $\alpha$ and $\beta$ arising only through proximity effects. Here, we treat all three bands on an equal footing, resulting in a fully multi-band superconductivity picture. The included 1D bands exhibit horizontal line nodes, leading to the experimentally observed power law for the specific heat. This model has been covered in more detail previously [@ann2; @grad1].
The distinct contributions to the magnetisation in this model are plotted in Fig. \[fig1\]. It should be noted that the contributions and diverge to plus and minus infinity respectively as $T$ approaches $T_c$. This problem arises due to the fact that these components are not separately gauge-invariant, and thus we must take the sum of the two. It has been shown previously that gauge-invariant forms of the normal state equations for and can be obtained [@sou1]. However, this requires an absolute distinction between the occupied and unoccupied states in the electron bandstructure. The Bogoliubov transformation enforces mixing of the electron and hole states. This mixing is essential to recover the quasiparticle bandstructure of the superconducting state, but prevents any attempts to project excitations onto occupied states and thus our expressions cannot be converted into separately gauge-invariant forms.
Comparison of the itinerant contributions and the on-site component reveal that they are of similar orders of magnitude. This corresponds to a magnetic field of the order of $\mu$G, which is below the resolution of the most recent attempts to experimentally identify an edge current in via magnetometry measurements ($\sim 2.5$ mG [@curr]). We can therefore surmise that it is possible the edge currents accompanying the chiral state in have remained elusive because their magnitude is significantly smaller than has been previously hypothesised.
The reason for this suppression in the orbital moment in comparison to other theoretical approaches likely lies in the multi-band, nodal nature of our tight-binding model and gap structure. Significantly, this result agrees with other experimental and theoretical observations which support the idea that multi-band superconductivity is prevalent in this material. It has been shown previously that inter-orbital transitions are necessary in order for the Kerr effect to arise intrinsically in the superconducting state [@grad1; @robb1]. In order to see the effect in a single band picture, extrinsic mechanisms such as skew scattering must be considered [@gor2]. The inclusion of the additional 1D, line nodal bands also leads to the correct specific heat below $T_c$ [@robb1]. The nodeless 2D band would not produce the experimentally observed power laws in heat capacity [@nish] or NMR spin relaxation rate [@ish3].
We also wish to assess the influence of spin-orbit coupling (SOC) on the magnetic moment in the chiral state. To do this, we compare results using a tight-binding model with an additional spin-orbit Hamiltonian derived in an on-site approximation. As was shown previously [@robb1], a model including spin-orbit coupling with coupling parameter $\lambda=12.5$ meV is able to replicate experimental features such as the Fermi surface, bandwidth and heat capacity. In the following, we compare the non-SOC case ($\lambda=0$) to the case with SOC ($\lambda=12.5$ meV).
The equivalent plots for the model including SOC are displayed in Fig. \[fig4\]. We also compare the total magnetisation for the cases with and without SOC (see Fig. \[fig2a\]), noting that the addition of SOC results in a clear suppression of the total orbital moment. We observe a significant quantitative reduction in all contributions, without any qualitative differences in the temperature-dependence displayed. This suppression is also of similar order to that seen in the Kerr effect under the influence of SOC as reported previously [@robb1].
![On-site moment 2 alongside the sum of the itinerant and local components + for the unperturbed model.[]{data-label="fig1"}](model.pdf)
![On-site moment 2 alongside the sum of the itinerant and local components + for the model including spin-orbit coupling.[]{data-label="fig4"}](modelsoc.pdf)
In order to fully assess the influence of SOC, it is informative to also compute the spin moment of the chiral state. To do this we start with the equation for the spin expectation value in the orbital basis:
$$\label{eq6}
\langle \hat{S}_z \rangle = \sum_{mm'\sigma\sigma'} \langle m \sigma | \frac{\hbar}{2} {\mathbf{\sigma_z}} | m' \sigma' \rangle n_{mm'}^{\sigma\sigma'}$$
where $m,\sigma$ are the orbital and spin degrees of freedom respectively and $n$ are the single particle density matrices.
The density matrix can be evaluated in terms of solutions to the BdG equation, while the $\sigma_z$ matrix elements are $\pm 1$ for $\sigma=\sigma'=\pm 1$ and $m=m'$. The final expression is then:
$$\label{eq7}
\langle \hat{S}_z \rangle = \sum_{m} \frac{\hbar}{2} \left( n_{mm}^{\uparrow\uparrow} - n_{mm}^{\downarrow\downarrow} \right)$$
$$\label{eq8}
n_{mm}^{\sigma\sigma}=\frac{1}{N} \sum_{n\bf{k}} |u_{n{\bf{k}}}^{\sigma}|^2 f(E_{n{\bf{k}}}) + |v_{n{\bf{k}}}^{\sigma}|^2 \left(1-f(E_{n{\bf{k}}})\right)$$
![Total orbital magnetic moment for the model with and without SOC.[]{data-label="fig2a"}](total.pdf)
![Spin moment in the model including SOC alongside the difference in orbital moment between the two models.[]{data-label="fig2"}](spin.pdf)
The spin magnetic moment is given by $\gamma_s \langle \hat{S}_z \rangle$ where $\gamma_s = -e g/(2 m_e)$ and $g$ is the spin gyromagnetic ratio.
It is interesting to note here that the spin moment in this context becomes non-zero when SOC is included (see Fig. \[fig2\]), but is zero otherwise. The spin moment in the SOC regime is of similar order to the reduction in the total orbital moment induced by the spin-orbit interaction (which we have denoted $\Delta {\bf{M}}$). This would suggest that the spin-orbit interaction mediates a transfer of magnetic moment from the orbital degrees of freedom (where it arises from the chiral order parameter) to the spin degrees (which are otherwise disordered).
This observation provides an interesting insight into the origin of the Kerr effect, a phenomenon which is driven by the anomalous Hall conductivity present in systems with a finite orbital moment. The microscopic origin of this effect in unconventional superconductors has been extensively debated [@gor2; @min2]. The current controversy concerns whether the origin is extrinsic (i.e. arising from disorder [@kim1; @lut2; @gor1]) or an intrinsic mechanism arising from coupling of the pair state to orbital degrees of freedom at the Fermi level [@wyso1; @grad1; @robb1].
In the normal state ferromagnet, the intrinsic mechanism facilitating the Kerr effect is induced by coupling of the ordered spins to the orbital component via SOC. Namely, the symmetry breaking in the spin degree of freedom is transferred to the orbital component via the spin-orbit interaction. This is a clear analogue to the results reported here, where orbital order arises naturally due to the chiral superconducting order parameter, and is then reduced via coupling to the disordered spin component. These results coincide with the observations reported previously, where the magnitude of the Kerr shift in the same chiral superconducting model was also shown to be suppressed by a similar order following the introduction of SOC [@robb1]. Our model is thus able to effectively describe an intrinsic origin of the anomalous phenomena observed in .
This analysis of the influence of SOC is further supported by assessing the regions of the Brillouin zone in which the spin moment arises (see Fig. \[fig3\]). We see here that the spin moment is present in regions of near-degeneracies between the orbital degrees of freedom in the bandstructure. These regions on the Brillouin zone contribute strongly to the Berry curvature, which gives rise to an anomalous Hall conductivity [@grad2]. This implies that these regions contain the highest density of ordered orbital moments, which in turn suggests that the spin magnetisation is arising directly as a result of coupling of the spins to the orbital degree of freedom.
In conclusion, a new formalism for computing the orbital magnetisation in a superconductor has been derived and calculations for the model chiral $p$-wave superconductor have been performed. The results suggest that early estimations of the itinerant magnetisation in this state were too generous and that the magnitude of edge currents may lie well below the resolution of magnetometry based investigations. This same model has been shown to also give a physically reasonable estimate of the observed Kerr effect [@grad1]. An interesting insight into the influence of SOC on a magnetic superconducting state has also been highlighted.
It should be stressed that the general result here is not restricted to the model used. We note that our theory would also apply to other pairing states which have been proposed for , such as the chiral $d$-wave [@zut1], $f$-wave [@has1], or long range $p$-wave states [@sca1]. In addition, the equations presented here could be used to investigate the unconventional pairing symmetries observed in other materials, such as the underdoped cuprates and heavy-fermion compounds.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe detailed simulations of X-ray-emitting populations to evaluate the levels of contamination by both Galactic and extragalactic X-ray sources unrelated to a star-forming region under study. For Galactic contaminations, we consider contribution from main-sequence stars and giants (not including cataclysmic variables and other classes of accretion-driven X-ray binary systems) as they make the dominant contribution at the position of the Carina Nebula. The simulations take into consideration a variety of technical factors involving a Galactic population synthesis model, stellar X-ray luminosity functions, $Chandra$ telescope response, source detection methodology, and possible spatial variations in the X-ray background and absorption through molecular clouds. When applied to the 1.42 square-degree field of the $Chandra$ Carina Complex Project (CCCP), the simulations predict $\sim5000$ contaminating sources (1 source per square arcminute of the survey), evenly distributed across the field. The results of the simulations are further employed in a companion CCCP study to assign membership probabilities to individual sources.'
author:
- 'Konstantin V. Getman, Patrick S. Broos, Eric D. Feigelson, Leisa K. Townsley, Matthew S. Povich, Gordon P. Garmire, Thierry Montmerle, Yoshinori Yonekura, Yasuo Fukui'
title: 'Source Contamination in X-ray Studies of Star-Forming Regions: Application to the [*Chandra*]{} Carina Complex Project'
---
Introduction \[sec\_introduction\]
==================================
Identifying stellar members associated with a star-forming region is important for studies of the stellar initial mass function (IMF), cluster dynamics, and star formation processes. Stars with dusty protoplanetary disks are readily found in infrared (IR) surveys where disk emission is strong, but when the disks have dissipated, the stars are most easily identified by their strong X-ray flaring. High spatial resolution X-ray observations of an active star forming region with the $Chandra$ X-ray Observatory will often reveal hundreds or thousands of pre-main sequence (PMS) stars, as well as OB stars [@Feigelson07].
An important challenge for membership studies is to distinguish young stars formed in a molecular cloud from various types of contaminant sources along the line-of-sight. In X-ray surveys of star-forming regions, stellar X-ray contaminants include main-sequence (MS) stars [e.g. @Schmitt95; @Schmitt97] and some types of giants [e.g. @Pizzolato00; @Gondoin05]. These foreground and background stellar contaminants have much less impact on surveys in X-ray images than IR images because magnetic activity in PMS stars is elevated $10^{1}-10^{4}$ above levels in older stars [e.g. @Preibisch05]. For star forming regions located in the quadrant of the Galactic plane centered on the Galactic Center, contamination by cataclysmic variables should be also considered [e.g. @Muno04]. X-ray surveys also suffer some contamination by extragalactic sources, mainly quasars and other active galactic nuclei (AGN) [e.g. @Brandt05]. These can be seen through the Galactic Plane as faint, absorbed X-ray sources.
Careful simulations of contaminating populations can provide estimates for the number of contaminants, their spatial distribution, and some of their apparent X-ray and IR properties. These estimates can inform efforts to assign individual classifications to the detected sources [@Broos11b]. Such simulation studies have been performed for the Orion Nebula Cluster (ONC) and the Cepheus B (Cep B) star-forming regions [@Getman05; @Getman06]. For the deep $Chandra$ observation of the ONC, $150-200$ AGNs, $15-20$ foreground stars, and zero background stars are predicted, constituting $\sim10-15$% of the $>1600$ X-ray detected sources. For the shallow $Chandra$ observation of Cep B, $20-30$ AGNs, $10-15$ foreground stars, and a few background stars are predicted, again $\sim10$% of the $Chandra$ source population.
The number and observed properties of contaminating populations depend on a number of factors: observatory sensitivity, source detection method, observation exposure time, survey field of view, Galactic direction, distance to the star forming region, and absorption through the local molecular cloud. Therefore, it is ill-advised simply to extrapolate the contamination results for the Cep B and ONC $Chandra$ fields to substantially different X-ray observations.
The purpose of this study is to present our contaminant simulation method and its application to the $Chandra$ Carina Complex Project [CCCP; @Townsley11a]. The simulation methods are refined from those described by @Getman06. The Carina Nebula (NGC 3372) is one of the richest nearby high-mass star-forming complexes, located $2.3$ kpc from the Sun towards the Galactic direction $(l,b)=(287^{\circ}.6, -0^{\circ}.8)$ [@Smith08]. The CCCP combines 22 spatially contiguous $Chandra$-ACIS-I pointings with nominal exposure of $60$ ksec covering $1.42$ square degrees around the chain of famous massive star-forming regions in Carina [@Townsley11a]. A relatively shallow molecular cloud [@Yonekura05] covers roughly 1/3 of the CCCP field (Figure \[fig\_spatial\]c). More than $14000$ CCCP point sources have been detected [@Broos11a].
The simulations of Galactic stellar and extragalactic contaminating populations in CCCP are described in §\[contaminants\_main\_section\]. They show that $\sim5000$ contaminating sources are expected in the CCCP catalog. Basic properties of these simulated contaminating sources (spatial distribution, X-ray median energies, and $J$-band magnitudes) are presented in §\[contamination\_properties\_section\]. The comparison of the properties of the simulated contaminating sources and the real CCCP sources is considered, at a simplistic level, in §\[qualitative\_analysis\_section\]. Based on the results of this study, with additional information on X-ray variability, IR excess, and optical spectral types, individual membership classifications for the CCCP sources are derived by @Broos11b.
Simulations of Cluster Contaminants \[contaminants\_main\_section\]
===================================================================
Extragalactic Sources \[contaminants\_agn\_section\]
----------------------------------------------------
We evaluate the expected contamination by extragalactic X-ray sources using methods similar to the analyses of @Getman05 [@Getman06]. Nominal (corresponding to the the high Galactic latitude fields with low Galactic absorbing column used for extragalactic studies) hard-band ($2-8$ keV) fluxes incident on [*Chandra*]{} are drawn from the $\log N - \log S$ distribution of extragalactic sources described by @Moretti03 assuming a power law source spectrum consistent with flux-dependencies described by @Brandt01. More than 22600 extragalactic sources down to a nominal flux of $\log(F_X) = -16.7$ erg cm$^{-2}$ s$^{-1}$ are predicted for the 1.42 deg$^{2}$ CCCP field. One hundred Monte Carlo simulations are constructed by placing $>22600$ artificial sources randomly across the CCCP field. Source photon indices are drawn from uniform distributions in the ranges $-0.5 < \Gamma < 2$ and $1 < \Gamma < 2$ for nominal fluxes $\log(F_X) < -14.1$ erg cm$^{-2}$ s$^{-1}$ and $\log(F_X) > -14.1$ erg cm$^{-2}$ s$^{-1}$, respectively.
The nominal flux of each simulated source is transformed to apparent (i.e., observed) flux, taking into consideration an extra absorbing column that is estimated as the sum of two components: a uniform H[I]{} column density of $N_{H} \sim1.4 \times 10^{22}$ cm$^{-2}$ through the entire Galactic disk in the direction of Carina [@Dickey90][^1] and spatially variable absorption through the Carina molecular clouds. The cloud component is derived from the velocity-integrated maps of CO line emission obtained with the NANTEN telescope by @Yonekura05. The gas column density outside the dense molecular cores was estimated from the maps of $^{12}$CO emission intensity integrated over the velocity range of $-30$ to $-10$ km s$^{-1}$ ($I(^{12}$CO$)$ \[K km s$^{-1}$\]) using the relationship $N(H_2) = I(^{12}$CO$) \times 1.6 \times 10^{20}$ [@Yonekura05]. Adopting the gas-to-dust ratio of @Vuong03, these gas column densities correspond to visual extinctions up to $A_V \sim3$ mag in the molecular clouds. The dense core column densities are evaluated from the C$^{18}$O measurements and correspond to $A_V \sim5-10$ mag [Table 3 in @Yonekura05].
$Chandra$ source photon spectra were then simulated, and observed count rates and median energies were derived using the [*fakeit*]{} command of the XSPEC software package [@Arnaud96]. This program samples photons from the AGN spectrum, calculating the soft X-ray absorption by the line-of-sight material, and convolving the results with the $Chandra$ mirror and ACIS detector response. Typical on-axis CCCP calibration redistribution matrix and auxiliary response calibration files were used [@Broos11a]. After applying local background levels found in the CCCP data, we then remove very weak extragalactic sources that would have fallen below the CCCP source detection threshold[^2].
Our simulations predict that $\sim2500$ extragalactic sources should be detected in the CCCP observation.
Galactic Sources \[contaminants\_stellar\_section\]
---------------------------------------------------
It has been argued that the point source component of the hard “Galactic ridge emission”, which stretches over 90 degrees in Galactic longitude and only 2 degrees in latitude around the Galactic Center, is mainly composed of cataclysmic variables [e.g @Ebisawa05; @Revnivtsev07; @Hong09]. The current version of our simulation omits the distribution of hard ($\ge3$ keV) X-ray sources in the Galactic plane attributed to cataclysmic variables and other classes of accretion-driven X-ray binary systems. However, this population and its X-ray emission should be unimportant in the Carina star forming region because it is located 72 degrees away from the Galactic Center where the Galactic Ridge emission is weak [see Figure 1 in @Revnivtsev06].
We simulate the stellar contamination from main sequence and giant stars in the Galactic Plane in three stages. First, Monte Carlo simulations of the Galactic stellar population expected within the CCCP field are constructed based on the ‘Besançon model’ of stellar populations [@Robin03]. The simulations are computed by the Web service provided by the Besançon group[^3]. Using models of Galactic structure, stellar mass function, and star formation history, the calculation gives the number distribution of stars along a chosen line-of-sight including individual spectral types, optical and near-infrared magnitudes, ages, distances, and source extinctions. An additional position-dependent absorption representing the Carina molecular clouds was added for stars behind Carina ($D > 2.3$ kpc). Besançon model limitations include a simplistic assumption of uniform extinction throughout the Galactic Plane and omission of spiral arms. We run the Galactic population model with an unrealistically deep $V=28$ mag detection limit, as the faint stars are later omitted by X-ray selection.
Second, X-ray luminosities for each star obtained from a Besançon model simulation are then estimated by Monte Carlo sampling of X-ray luminosity functions (XLFs) of MS stars measured in the solar neighborhood from [*ROSAT*]{} surveys [@Schmitt95; @Schmitt97; @Hunsch99]. Based on these studies, we adopt the following lognormal XLFs with the following means and standard deviations in $\log L_X$ \[erg s$^{-1}$\] in the [*ROSAT*]{} PSPC energy band ($0.1-2.4$ keV): $<28.20> \pm 0.79$ for MS F-type stars; $<27.94> \pm 1.02$ for MS G-type stars; $<27.86> \pm 0.65$ for MS K-type stars; and $<27.44> \pm 0.76$ for MS M-type stars. For giant stars, we adopt $<28.91> \pm 0.86$ [@Pizzolato00]. X-ray spectral shapes were assumed to follow the stellar temperature-luminosity relations of @Gudel98 and @Pizzolato00 for MS and giant stars, respectively. The adopted temperature-luminosity relations in the [*ROSAT*]{} ($0.1-2.4$ keV) band are: $\log(T)$ \[MK\] $=-9.2 + 0.33\log(L_X)$ \[erg s$^{-1}$\] for MS stars and $\log(T)$ \[MK\] $=-5.1 + 0.2\log(L_X)$ \[erg s$^{-1}$\] for giants.
Third, given the X-ray luminosity in the [*ROSAT*]{} ($0.1-2.4$ keV) band, spectrum, distance, and line-of-sight absorption for each of the simulated stars, the count rate in the $Chandra$ ACIS detector in the ($0.5-8.0$ keV) band is obtained using the Portable Interactive Multi-Mission Simulator (PIMMS[^4])[^5]. The detectability of each simulated star is then evaluated by locally-defined thresholds of the source significance statistic, similar to our treatment of extragalactic contaminants (§ \[contaminants\_agn\_section\]). For the stars that satisfy the detection threshold, the XSPEC [*fakeit*]{} command was used to simulate the spectrum to derive X-ray median energies for individual stars.
A typical run of the Besançon Galactic disk population model of the 1.42 deg$^2$ CCCP exposure of the Carina Nebula region predicts $\sim200,000$ foreground (distance $< 2.3$ kpc) MS stars in the following mass distribution: 4,000 F stars, 10,000 G stars, 20,000 K stars, and 170,000 M stars. About 1,000 foreground giants are present. Behind the cloud at distances $2.3-5.0$ kpc, the model predicts $\sim930,000$ background stars: 30,000 F stars, 77,000 G stars, 180,000 K stars, 630,000 M stars, and 9,000 giants. The vast majority of these stars are undetectable in the CCCP X-ray exposures. This is expected from the appearance of star forming regions at distances around $1-3$ kpc in different wavebands: infrared images are typically dominated by dense populations of field stars, while X-ray images are dominated by PMS stars in the young clusters.
A typical Monte Carlo run predicts $\sim1800$ foreground and $\sim900$ background field stars will be detected in the X-ray survey. The foreground detections include: 150 F stars, 450 G stars, 150 K stars, 850 M stars, and 150 giants. Approximately 20%, 50%, and 30% of the simulated foreground detections have distances from the Sun in the ranges $< 0.5$ kpc, $0.5-1.5$ kpc, and $1.5-2.3$ kpc, respectively. The background detections include: 100 F stars, 500 G stars, 10 K stars, 40 M stars, and 250 giants. Approximately 35%, 35%, and 30% of the simulated background detections have distances in the ranges $2.3-3.0$ kpc, $3.0-4.0$ kpc, and $4.0-5.0$ kpc, respectively.
Several caveats pertain to the simulation. First, the line of sight towards the Carina complex is tangent to the Sagittarius-Carina spiral arm, but the Besançon model of the Galactic disk population (which our simulations rely on) lacks spiral arm modeling [@Robin03]. Second, X-ray luminosities of stars with ages $<1$ Gyr might be even higher than those of stars from the solar neighborhood [@Preibisch05]. Roughly $1/4$ of the simulated Besançon stars have ages $<1$ Gyr. Third, due to computational reasons, the simulations of Galactic background stars were truncated at a distance of $5$ kpc. Due to these three factors, we might expect even larger stellar contamination populations. Fourth, on the other hand, a possible decrease in X-ray activity with mass and age in giant stars [@Pizzolato00] is ignored in our simulations. If one considers only the most plausible class of X-ray emitting giants, F and G giants with ages $<1$ Gyr, the expected number of detected giants in the CCCP field could be as small as $\sim$10 in the foreground and $\sim$10 in the background.
Finally, we note that many Galactic field stars will each produce $<3$ photons in the CCCP image, too faint for detection as a point source. The undetected foreground stars produce a total of $\sim10,000$ X-ray photons in the soft ($0.5-2$ keV) band. These photons will appear as a uniform diffuse X-ray component in the CCCP image. These stars likely contribute $\sim$1% of the observed $> 10^6$ count soft X-ray emission that pervades the CCCP field [@Townsley11b].
Properties of Cluster Contaminants \[contamination\_properties\_section\]
=========================================================================
Basic properties of the simulated contaminating populations $-$ spatial distributions, X-ray median energies, and $J$-band magnitudes $-$ are described here. These properties are employed by @Broos11b to establish Carina membership probabilities for each of the CCCP sources.
Spatial Density and Distribution \[spatial\_contamination\_subsection\]
-----------------------------------------------------------------------
Our simulations predict that roughly 5000 contaminants, or approximately 1 source arcmin$^{-2}$ in the 5112 arcmin$^2$ CCCP survey, will be detected in our CCCP observations. About half of these are expected to be extragalactic sources, and half Galactic field stars. With $>14000$ detected sources [@Broos11a], contaminants are expected to comprise $\sim35$% of the detected CCCP source population. For comparison, the source surface densities of the X-ray contaminants detected in the ONC and Cep B observations (§ \[sec\_introduction\]) are 0.8 and 0.2 source arcmin$^{-2}$, respectively. A variation among these numbers is a product of multiple competing factors such as, detection technique, exposure times, thicknesses of molecular clouds and distances to the clouds, Galactic coordinates.
Across most of the Carina molecular cloud, the absorption is low, typically $A_V < 3$ mag, and reaches $\sim$10 mag only in a few small molecular cores [@Yonekura05]. Thus the spatial distributions of the simulated contaminants for all classes of the contaminants are relatively uniform across the field of view. We ignore small differences in the surface density of field stars between the northern (closer to the Galactic plane) and the southern portions of the CCCP field, predicted by the Besançon model: $<2$% and $<5$% for the foreground and background stars, respectively. The spatial distribution of the contaminants is very different from the clustered distribution of the observed CCCP sources, as shown in maps of the source distributions [@Townsley11a; @Feigelson11]. The difference can be quantitatively seen in projection along the right ascension axis (Figure \[fig\_spatial\]). The lower numbers of simulated contaminating sources at the both ends of the right ascension range, compared to that of the central part of the field (Figure \[fig\_spatial\]a), are simply due to the decreasing angular coverage of the CCCP observation near the east and west boundaries (Figure \[fig\_spatial\]c).
X-ray Median Energy \[me\_contamination\_subsection\]
-----------------------------------------------------
The X-ray median energy statistic, $MedE$, is the most robust among the measured photometric quantities used to describe the spectral shapes of weak X-ray sources. $MedE$ is most effective as a surrogate for the absorption column, but if the absorption is independently known, it can give a rough measure of plasma temperature [@Feigelson05; @Getman10]. From the simulated spectra of the Carina contaminants (§ \[contaminants\_main\_section\]), we obtain histograms of $MedE$ for each of the three simulated contaminant classes and compare them with the observed CCCP sources in Figure \[fig\_me\]. Notice that the observed CCCP sources are mixture of objects from all four classes: PMS stars, foreground stars, background stars, and AGNs. A few important findings emerge.
1. The three contaminant classes occupy mostly distinct ranges of $MedE$: 96% of the simulated foreground stars have very soft X-ray spectra with $0.7 < MedE < 1.1$ keV range; 81% of the simulated background stars are within the $1.1 < MedE < 1.5$ keV range; 94% of the simulated extragalactic sources are within the $2.5 < MedE < 4.5$ keV range.
2. The observed CCCP sources span a wide range of $MedE$, from 0.6 to 7 keV, with a skewed distribution peaking at $1.4-1.5$ keV.
3. The histograms of the observed CCCP sources and the simulated foreground stars have similar numbers of sources in the range $0.7 < MedE < 1$ keV, suggesting that nearly all the CCCP sources in this range are likely foreground stars. This also implies that our simulation did not significantly underestimate the foreground population due to an unmodeled population of MS stars in the Carina spiral arm (§ \[contaminants\_stellar\_section\]).
4. The histograms of the observed CCCP sources and the simulated AGNs have similar numbers of sources in the range $2.5 < MedE < 4.5$ keV, suggesting that nearly all the CCCP sources in this energy range can be AGNs.
5. The peak of the simulated background stars coincides with the peak of the observed CCCP sources, but the numbers of the CCCP and simulated background sources are drastically different. It is clearly impossible to distinguish between background stars and Carina PMS members using $MedE$.
We see that the X-ray median energy is an important discriminant for distinguishing some contaminant populations. It can be effective for foreground stars and extragalactic sources, but not background stars which have the same median energies as many Carina PMS stars.
*J*-band Magnitude \[jmag\_contamination\_subsection\]
------------------------------------------------------
Younger PMS stars with protoplanetary disks (infrared Class 0-I-II systems) can be readily distinguished from Galactic field stars by their infrared excess. But often, as in the Carina complex, the majority of stars associated with a star forming region are older, diskless (Class III) systems. X-ray studies are particularly effective in detecting Class III stars which have high levels of magnetic activity. While in some X-ray studies of star forming regions, IR color-color and/or color-magnitude diagrams can be employed to successfully distinguish between [*diskless*]{} PMS and older stars [e.g. @Getman05; @Getman06; @Kuhn10], the small source extinction range (typically only $A_V \sim 1-3$ mag for most parts of the Carina region) and the large range of distances (only with respect to color-magnitude diagrams) for the simulated contaminating stars (§\[contaminants\_stellar\_section\]) blunt this technique. Therefore, here we restrict our consideration to a simple near-IR photometric property, $J$-band magnitude. In addition to $J$-band magnitude, @Broos11b employ some near-IR and mid-IR colors to distinguish between [*disk-bearing*]{} PMS and other stars. Wider usage of IR colors and incorporation of IR-to-X-ray flux ratios (when applicable) will be considered in our future studies of other star forming regions.
Figure \[fig\_jmag\] compares $J$-band magnitude histograms among the $6,182$ CCCP sources with available 2MASS counterparts within the entire CCCP field of view (dashed lines) and the three simulated contamination populations (solid lines). $J$-band magnitudes for the foreground and background field stars are obtained directly from the Besançon model simulations; for the background stars, magnitudes are slightly increased for the absorption from the Carina molecular cloud. For the extragalactic contaminants, we produce an approximate distribution of $J$-band magnitudes using the relationship between the nominal (see §\[contaminants\_agn\_section\]) hard-band X-ray flux and $I$-band magnitude reported for the [*Chandra*]{} Deep Field North sources [Figure 7 in @Alexander01]. The $I$-band magnitudes are transformed to $J$-band magnitudes following Figure 11 in Alexander et al., accounting for an average combined visual absorption of $A_V \sim10-12$ mag through the Galactic disk in the direction of Carina and through the Carina molecular clouds. For better compatibility with the CCCP histograms, we also present (Figure \[fig\_jmag\]b) the histograms of $J$-band magnitudes for simulated foreground and background stars multiplied by the “generic” 2MASS $J$-band detection completeness curve from @Skrutskie06[^6].
Several results emerge from the $J$-band magnitude distributions. First, the histogram of the simulated foreground population has two peaks (red curve). The fainter peak from M-type field dwarfs, constituting $\sim40$% of the simulated foreground population, lies below the 2MASS sensitivity limit of $J \sim 16.5-17$ mag. A bimodal shape is indeed seen in the deeper HAWK-I $J$-band magnitude histogram of the CCCP sources with $MedE < 1.1$ keV [@Preibisch11]. Second, for bright stars with $J \la 16$ mag, the $J$-band magnitude histogram of the simulated foreground population (red curve) agrees well both in number and shape with the magnitude distribution of CCCP sources with median energies $<1.1$ keV (black dashed curve). This indicates, in agreement with the low-$MedE$ distributions discussed in § \[me\_contamination\_subsection\], that most of the observed CCCP sources in this sub-sample can be attributed to foreground stars. Third, roughly 60% of the simulated stellar background population lies below the 2MASS sensitivity limit of $J \ga 16.5$ mag. Fourth, most of the simulated extragalactic objects lie below $J=21$ mag. The HAWK-I survey can detect some of these extragalactic objects near its detection limit [@Preibisch11].
Spatial Distributions Stratified by Median Energy and *J*-band Magnitude \[qualitative\_analysis\_section\]
===========================================================================================================
Here we consider the spatial distribution of CCCP sources stratified by $MedE$ and by $J$-band magnitude that, as seen in § \[contamination\_properties\_section\], are often linked to contaminant classes. We expect that Galactic and extragalactic contaminants will have random locations in the CCCP field, but many Carina members will lie in young stellar clusters. The spatial distributions are shown in Figures \[fig\_aciswith2mass\]-\[fig\_acisno2mass\], with $MedE$ strata informed by Figure \[fig\_me\]. The contours show the distribution of molecular material from NANTEN CO maps [@Yonekura05]. The CCCP sources are further divided into two groups that have or lack 2MASS counterparts, giving two magnitude strata divided approximately at $J \sim 16.5$ mag[^7]. Interpretation of spatial distributions of CCCP sources is confused by the large variations in sensitivity due to overlapping exposures and off-axis angle [@Broos11a]. Our qualitative discussion here is designed to assess the role of contaminants. A more quantitative treatment of spatial groupings from a spatially complete sub-sample of CCCP sources is given by @Feigelson11.
Hard CCCP Sources\[obscured\_section\]
--------------------------------------
In this section we consider all CCCP sources with X-ray median energy $MedE > 2$ keV, equivalent to $A_V \ga 10$ mag [@Getman10]. In panels $d-f$ of Figure \[fig\_aciswith2mass\], $70$% of the $\sim700$ CCCP sources with 2MASS counterparts are spatially concentrated towards the edges of the Carina molecular clouds. Three prominent source groupings are seen: members of the Treasure Chest cluster, stars within the Trumpler 16 cluster, and stars at the interface between the rich Trumpler 14 cluster and C$^{18}$O core \#10. The latter region also has a dense grouping of $MSX$ mid-infrared sources. These groupings of absorbed X-ray sources are likely to be very young stellar objects. In contrast, out of $\sim2800$ hard CCCP sources without 2MASS counterparts (Figure \[fig\_acisno2mass\]$d-f$), $\sim2000$ are evenly distributed across the CCCP field and are likely extragalactic candidates. This number is in satisfactory agreement with the number of simulated AGNs ($2500$ from §\[contaminants\_agn\_section\]).
Soft CCCP Sources\[unobscured\_section\]
----------------------------------------
In this section we consider all CCCP sources with X-ray median energy $MedE < 2$ keV. Figure \[fig\_aciswith2mass\] presents all CCCP sources with available 2MASS counterparts. Panel $a$ ($MedE<1.1$ keV) shows that only 40% of the sources are spatially concentrated in the ridge of famous young stellar clusters in Carina (including Trumpler 15, Trumpler 14, Trumpler 16, Treasure Chest, and Bochum 11 [@Feigelson11])[^8], while the majority of the sources are relatively uniformly distributed across the CCCP field of view[^9] (ignoring the “egg-crate effect”, due to variations in sensitivity with off-axis angle [@Broos11a]). In contrast to panel $a$, 70% of the sources in panel $b$ ($1.1< MedE <1.5$ keV) and 80% of the sources in panel $c$ ($1.5< MedE <2.0$ keV) are seen in projection against the chain of Carina clusters. Figure \[fig\_acisno2mass\] presents all CCCP sources without 2MASS counterparts. Panels $a$ ($MedE <1.1$ keV), $b$ ($1.1< MedE <1.5$ keV), and $c$ ($1.5< MedE <2.0$ keV) show that approximately 80%, 50%, and 50% of sources, respectively, are evenly distributed across the CCCP field.
Thus, based simply on their uniform spatial distributions, as many as $700$ ($+1400$) with $MedE<1.1$ keV, $800$ ($+1100$) with $1.1<MedE<1.5$ keV, and $300$ ($+700$) with $1.5<MedE<2.0$ keV sources with (without) 2MASS counterparts could be potential Galactic stellar contaminants. Further consideration of the expected $MedE$ and $J$-band properties for the simulated stellar contaminants (§\[contamination\_properties\_section\]) suggests that indeed, most of the $700+1400 = 2100$ sources with $MedE<1.1$ keV could be foreground stars, and roughly half of the $800+1100=1900$ sources with $1.1<MedE<1.5$ keV could be background stars. Therefore, this leaves us with at least $\sim300$ ($MedE<1.1$ keV), $\sim1000$ ($1.1<MedE<1.5$ keV), and $\sim1000$ ($1.5<MedE<2.0$ keV) CCCP sources evenly distributed across the field that could be either additional contaminants unaccounted for by the simulations, and/or young stellar objects found outside the famous Carina clusters. In the former case, our underestimation of the Galactic stellar contamination population could be due to un-modeled young MS stars associated with Galactic spiral arm toward Carina, or background field stars with distances $>5$ kpc (§\[contaminants\_stellar\_section\]). In the latter case, we have an indication for the presence of a widely distributed population of young stars in the Carina complex.
Summary {#summary_section}
=======
This work describes simulations of X-ray-emitting source populations that will appear in high-resolution [*Chandra X-ray Observatory*]{} studies along the Galactic Plane. The effort is particularly relevant to discriminating PMS stars in star forming regions from Galactic field star and extragalactic contamination. Our simulations account for two X-ray classes: extragalactic sources (primarily AGNs) seen through the Galactic disk and Galactic field stars (main sequence and giants from types F to M) distributed throughout the disk. The simulations are based on a variety of technical considerations (§ \[contaminants\_main\_section\]): use of a Galactic population synthesis model and measured X-ray luminosity functions to obtain field star X-ray source distributions; convolution of realistic contaminant X-ray intensity and spectral distributions through the telescope response; consideration of spatially varying absorption across the field due to molecular clouds; application of a realistic source detection method including threshold effects due to position-dependent point spread functions and X-ray background. The current version of the simulations omits contributions from accretion binary systems and will thereby underestimate the population at high $MedE$ for fields close to the Galactic Center.
The results of the simulations are applied to the CCCP X-ray survey of the Carina star forming complex [@Townsley11a] to evaluate the levels of non-Carina contaminants among the $>14000$ CCCP point sources. The simulations predict that about 2500 AGNs (§\[contaminants\_agn\_section\]), 1800 foreground stars, and 900 background stars (§\[contaminants\_stellar\_section\]) will be detected. The simulated contaminants are expected to have uniform spatial distributions and exhibit significant differences in their distributions of the X-ray median energy and $J$-band magnitude (§ \[contamination\_properties\_section\]). These properties are thus useful for classifying individual CCCP sources as likely Carina members or contaminants [@Broos11b].
Comparing properties of the simulated contaminating sources and the CCCP sources, at a simplistic level (§\[qualitative\_analysis\_section\]), we find that the number of hard CCCP sources evenly distributed across the field is in satisfactory agreement with our extragalactic contamination prediction, but the number of evenly-distributed soft CCCP sources is a factor of 2 higher than our Galactic contamination prediction (§\[qualitative\_analysis\_section\]). This discrepancy may be due either to unmodeled effects (e.g., stars beyond $D = 5$ kpc, an additional population of MS stars in the Carina spiral arm, or evolution of X-ray activity in older stars), or the presence of a widely distributed population of young stars in the Carina complex. The inference of distributed star formation, or the kinematic drifting of an older generation of Carina stars, is substantiated by more detailed studies of the CCCP sources lying outside the rich clusters [@Feigelson11; @Preibisch11; @Povich11b].
We thank the anonymous referee for his time and many useful comments that improved this work. This work is supported by $Chandra$ GO grant SAO GO8-9131X (L. Townsley, PI) and the $Chandra$ ACIS Team contract SV4-74018 (G. Garmire, PI), issued by the Chandra X-ray Center operated by the Smithsonian Astrophysical Observatory for and on behalf of NASA under contract NAS8-03060. M.S.P. is supported by an NSF Astronomy and Astrophysics Postdoctoral Fellowship under award AST-0901646. This publication makes use of data products from the Two Micron All Sky Survey (a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by NASA and NSF).
Alexander, D. M., Brandt, W. N., Hornschemeier, A. E., Garmire, G. P., Schneider, D. P., Bauer, F. E., & Griffiths, R. E. 2001, , 122, 2156
Arnaud, K. A. 1996, in ASP Conf. Ser. 101, Astronomical Data Analysis Software and Systems V, ed. G. Jacoby & J. Barnes (San Francisco, CA: ASP), 17
Brandt, W. N., et al. 2001, , 122, 2810
Brandt, W. N., & Hasinger, G. 2005, , 43, 827
Broos, P. S., Townsley, L. K., Feigelson, E. D., Getman, K. V., Bauer, F. E., & Garmire, G. P. 2010a, , 714, 1582
Broos, P. S., et al. 2011, , submitted (CCCP Catalog Paper)
Broos, P. S., et al. 2011, , submitted (CCCP Classifier Paper)
Dickey, J. M., & Lockman, F. J. 1990, , 28, 215
Ebisawa, K., et al. 2005, , 635, 214
Feigelson, E. D., et al. 2005, , 160, 379
Feigelson, E., Townsley, L., G[ü]{}del, M., & Stassun, K. 2007, in Protostars and Planets V, ed. B. Reipurth, D. Jewitt, & K. Keil (Tucson, AZ: Univ. Arizona Press), 313
Feigelson, E. D., et al. 2011, , submitted (CCCP Clustering Paper)
Getman, K. V., Feigelson, E. D., Grosso, N., McCaughrean, M. J., Micela, G., Broos, P., Garmire, G., & Townsley, L. 2005, , 160, 353
Getman, K. V., Feigelson, E. D., Townsley, L., Broos, P., Garmire, G., & Tsujimoto, M. 2006, , 163, 306
Getman, K. V., Feigelson, E. D., Broos, P. S., Townsley, L. K., & Garmire, G. P. 2010, , 708, 1760
Gondoin, P. 2005, , 444, 531
G[" u]{}del, M., Guinan, E. F., & Skinner, S. L. 1998, in ASP Conf. Ser. 154, Cool Stars, Stellar Systems and the Sun, ed. R. A. Donahue & J. A. Bookbinder (San Francisco, CA: ASP), 1041
Hong, J. S., van den Berg, M., Grindlay, J. E., & Laycock, S. 2009, , 706, 223
H[" u]{}nsch, M., Schmitt, J. H. M. M., Sterzik, M. F., & Voges, W. 1999, , 135, 319
Kuhn, M. A., Getman, K. V., Feigelson, E. D., Reipurth, B., Rodney, S. A., & Garmire, G. P. 2010, , 725, 2485
Moretti, A., Campana, S., Lazzati, D., & Tagliaferri, G. 2003, , 588, 696
Muno, M. P., et al. 2004, , 613, 1179
Pizzolato, N., Maggio, A., & Sciortino, S. 2000, , 361, 614
Povich, M. S., et al. 2011, , submitted (CCCP IR YSOs Paper)
Preibisch, T., & Feigelson, E. D. 2005, , 160, 390
Preibisch, T., et al. 2011, , submitted (CCCP HAWK-I Paper)
Revnivtsev, M., Sazonov, S., Gilfanov, M., Churazov, E., & Sunyaev, R. 2006, , 452, 169
Revnivtsev, M., Vikhlinin, A., & Sazonov, S. 2007, , 473, 857
Robin, A. C., Reyl[' e]{}, C., Derri[\` e]{}re, S., & Picaud, S. 2003, , 409, 523
Schmitt, J. H. M. M., Fleming, T. A., & Giampapa, M. S. 1995, , 450, 392
Schmitt, J. H. M. M. 1997, , 318, 215
Skrutskie, M. F., et al. 2006, , 131, 1163
Smith, N., & Brooks, K. J. 2008, in Handbook of Star Forming Regions: Vol. II, The Southern Sky, ASP Monograph Publications, Vol. 5, ed. B. Reipurth (San Francisco, CA: ASP), 138
Townsley, L. K., et al. 2011, , submitted (CCCP Intro Paper)
Townsley, L. K., et al. 2011, , submitted (CCCP Diffuse Paper)
Vuong, M. H., Montmerle, T., Grosso, N., Feigelson, E. D., Verstraete, L., & Ozawa, H. 2003, , 408, 581
Yonekura, Y., Asayama, S., Kimura, K., Ogawa, H., Kanai, Y., Yamaguchi, N., Barnes, P. J., & Fukui, Y. 2005, , 634, 476
![Comparison of histogram (a) and cumulative (b) distributions in right ascension among the observed CCCP sources (black), simulated foreground stars (red), background stars (green), and extragalactic objects (blue). (c) Spatial distribution of simulated extragalactic objects (blue) superposed on the $Chandra$ Carina exposure map. The yellow contours show the $^{12}$CO emission and green contours indicate C$^{18}$O cores from @Yonekura05. \[fig\_spatial\]](f1.eps){width="5.5in"}
![Comparison of median energy distributions for the observed CCCP sources (black) and simulated foreground stars (red), background stars (green), and extragalactic objects (blue). \[fig\_me\]](f2.eps){width="6.5in"}
![Panel (a): comparison of $J$-band distributions for 6182 CCCP sources with available 2MASS counterparts (black, magenta, orange, cyan) and simulated foreground stars (red), background stars (green), and extragalactic objects (blue). The CCCP sources are stratified by median energy: $MedE < 1.1$ keV (black), $1.1 < MedE < 1.5$ keV (magenta), $1.5 < MedE < 2.5$ keV (orange), and $MedE > 2.5$ keV (cyan). Panel (b): Histograms of $J$-band magnitudes for simulated foreground (red) and background (green) stars multiplied by the 2MASS $J$-band detection completeness curve from @Skrutskie06. As a reference, the histogram of the CCCP sources with $MedE < 1.1$ keV (black) from panel (a) is also plotted here. \[fig\_jmag\]](f3.eps){width="5.0in"}
![Spatial distributions of all observed CCCP sources with 2MASS counterparts (red circles) superposed on the [*Chandra*]{} Carina exposure map. Distributions are stratified by X-ray median energy, a surrogate for line-of-sight absorption. Panels are stratified by X-ray median energies: (a) $MedE < 1.1$ keV; (b) $1.1< MedE <1.5$ keV; (c) $1.5< MedE <2.0$ keV; (d) $2.0< MedE <2.5$ keV; (e) $2.5< MedE <4.5$ keV; and (f) $MedE > 4.5$ keV. The figure legends also provide numbers of plotted CCCP sources. The yellow contours show the $^{12}$CO emission and green contours indicate C$^{18}$O cores from @Yonekura05. \[fig\_aciswith2mass\]](f4.eps){width="5.0in"}
![Same as in Figure \[fig\_aciswith2mass\], but for CCCP sources without 2MASS counterparts. \[fig\_acisno2mass\]](f5.eps){width="5.0in"}
[^1]: This was obtained from NASA’s HEASARC tool located at <http://heasarc.gsfc.nasa.gov/cgi-bin/Tools/w3nh/w3nh.pl.>
[^2]: Position-dependent detection thresholds corresponding to the CCCP detection process were calculated using local background levels estimated from the CCCP data. The very sophisticated and innovative source detection technique used in CCCP [@Broos10a; @Broos11a] can be approximated by thresholding the source significance statistic, $Signif$, which is calculated by the ACIS Extract (<http://www.astro.psu.edu/xray/docs/TARA/ae_users_guide.html>) package as the ratio of net counts to the uncertainty on that quantity. The distribution of source significance for all CCCP sources peaks at $Signif = 0.8$, has a very sharp decay below 0.8 encompassing only 3% of sources, and has an exponential decay above 0.8. In our contamination simulations we thus choose the source detection threshold of $Signif = 0.8$.
[^3]: <http://model.obs-besancon.fr/>.
[^4]: <http://cxc.harvard.edu/ciao/ahelp/pimms.html>.
[^5]: At this stage of the simulations, it was not feasible to use XSPEC to derive count rates for our very large sample of simulated stars.
[^6]: See also the 2MASS $J$-band detection completeness curves in the 2MASS All Sky Data Release Documentation at <http://www.ipac.caltech.edu/2mass/releases/allsky/doc/sec6_5a1.html>. Notice, however, that this “generic” 2MASS detection completeness curve does not account for the effects specific to the CCCP field, such as the source crowding and IR nebulosity.
[^7]: A $\sim 3$ circular ‘hole’ in the distribution of 2MASS sources is seen in Figure \[fig\_aciswith2mass\]b due to saturation by $\eta$ Car in the 2MASS images.
[^8]: The field of the cluster ridge is defined here as the lowest contour of the X-ray source density shown in Figure 1 of @Feigelson11.
[^9]: In this section, source density estimates of evenly distributed populations are based on average source densities in the four ACIS-I fields outside the chain of the famous clusters. In @Townsley11a these four fields are labelled as E2, E4, SB1, and SB2.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study wave turbulence in shallow water flows in numerical simulations using two different approximations: the shallow water model, and the Boussinesq model with weak dispersion. The equations for both models were solved using periodic grids with up to $2048^2$ points. In all simulations, the Froude number varies between $0.015$ and $0.05$, while the Reynolds number and level of dispersion are varied in a broader range to span different regimes. In all cases, most of the energy in the system remains in the waves, even after integrating the system for very long times. For shallow flows, non-linear waves are non-dispersive and the spectrum of potential energy is compatible with $\sim k^{-2}$ scaling. For deeper (Boussinesq) flows, the non-linear dispersion relation as directly measured from the wave and frequency spectrum (calculated independently) shows signatures of dispersion, and the spectrum of potential energy is compatible with predictions of weak turbulence theory, $\sim k^{-4/3}$. In this latter case, the non-linear dispersion relation differs from the linear one and has two branches, which we explain with a simple qualitative argument. Finally, we study probability density functions of the surface height and find that in all cases the distributions are asymmetric. The probability density function can be approximated by a skewed normal distribution as well as by a Tayfun distribution.'
author:
- 'P. Clark di Leoni'
- 'P. J. Cobelli'
- 'P. D. Mininni'
bibliography:
- 'Tesis.bib'
title: Wave turbulence in shallow water models
---
Introduction
============
Turbulence and non-linear wave interactions in the ocean surface are related to important processes in atmospheric sciences and oceanography, such as the exchange of energy between the atmosphere and the ocean [@iafrati_modulational_2013; @dasaro_enhanced_2011]. This exchange, in turn, plays an important role in the dynamics of the planetary and oceanic boundary layers, with consequences on the transport and mixing of momentum, CO$_2$, and heat [@ivey_density_2008]. The incorrect modeling of these phenomena affects climate evolution predictions [@rose_upper-ocean--atmosphere_2010; @cavaleri_wind_2012]. Ocean surface waves are also of interest in the search of renewable energies [@falnes_review_2007].
There are several ocean surface models which provide an excellent framework for studying *weak turbulence theory* [@hasselmann_non-linear_1962; @hasselmann_non-linear_1963; @hasselmann_non-linear_1963-1; @zakharov_kolmogorov_1992]. This theory was developed to describe the out-of-equilibrium behavior of systems of dispersive and weakly non-linear waves (see, e.g., [@newell_wave_2011; @nazarenko_wave_2011]). Unlike theories of strong turbulence, for waves and under the assumption of weak nonlinearities, the equations for two-point correlations can be closed and exact solutions with constant flux can be found. Besides this assumption, it is also assumed that wave fields are homogeneous, and that free waves are uncorrelated.
Weak turbulence theory has been applied to capillary and gravito-capillary waves [@zakharov_kolmogorov_1992], vibrations on a plate [@during_weak_2006], rotating flows [@galtier_weak_2003], and magnetohydrodynamic waves [@galtier_weak_2000; @nazarenko_wave_2011; @schekochihin_weak_2012]. For some of these systems, the predictions of the theory are compatible with results obtained from experiments or from numerical simulations. For example, see Refs. [@deike_decay_2012; @falcon_capillary_2009; @kolmakov_quasiadiabatic_2004] for capillary waves, [@cobelli_different_2011] for gravitocapillary waves, [@mordant_are_2008; @boudaoud_observation_2008; @cobelli_space-time_2009] for vibrations on a plate, and [@leerink_multiscale_2012; @mininni_energy_2007] for magnetohydrodynamic waves. Although agreement has been found between theory, numerical simulations and experiments, there are also discrepancies. In some of these cases the compatibility is limited to the spectrum of certain fields (see, e.g., [@mininni_energy_2007]), or to specific configurations used to generate the excitations. Moreover, for many systems it is also not clear whether the weak turbulence approximation holds for all times, as the solutions are not homogeneous in wavenumber space and at sufficiently small scales eddies may become faster than waves resulting in strong turbulence [@chen_resonant_2005].
One of the most important applications of weak turbulence is in surface gravity waves. In oceanography, the Phillips’ spectrum [@phillips_equilibrium_1958], derived using dimensional arguments from strong turbulence and considering the coupling between waves, was long considered to be correct. However, different observational and experimental data [@toba_local_1973; @donelan_directional_1985], as well as numerical simulations [@badulin_weakly_2007], suggest that the actual spectrum is closer to that of weak turbulence. In fact, Phillips himself suggested that his spectrum may not be valid in the ocean [@phillips_spectral_1985]. Nonetheless, a scaling compatible with Phillips’ spectrum is still observed in numerical simulations [@korotkevich_simultaneous_2008] when the forcing is strong. This suggests that while weak turbulence provides an elegant theoretical way to study wave turbulence in the ocean, more considerations are necessary to appropriately describe the diversity of regimes found in these flows [@newell_wave_2011].
Most of the work done in ocean surface waves under the weak turbulence approximation concerns deep water flows. But the theory can also be applied to the shallow water case, i.e., for gravity waves whose wavelengths are large compared to the height of the fluid column at rest (see [@zakharov_statistical_1999; @onorato_four-wave_2008]). In this case, the theory leads to the prediction that the energy spectrum follows a $\sim
k^{-4/3}$ behavior. Behavior compatible with this prediction was found both experimentally and observationally [@smith_equilibrium_2003; @kaihatu_asymptotic_2007; @falcon_observation_2011]. It was also found that an inertial range with a $\sim k^{-2}$ dependency can develop in the shallower regions of the fluid. The comparison between different shallow water models, with different degree of shallowness (and of dispersion) is therefore of interest, e.g., for the study of waves in basins with inhomogeneous depth.
In the shallow water regime there are several models that can be considered to describe the ocean surface. There is the linear theory (see, e.g., [@landau_fluid_2004]) which can predict the dispersion relation of small amplitude waves, but which is insufficient to study turbulence. There are also non-linear theories, such as the shallow water model [@pedlosky_geophysical_1987] for non-dispersive waves, as well as the Boussinesq model [@whitham_linear_1974] for weakly dispersive waves which is the one used in some of the most recent works on the subject [@onorato_four-wave_2008]. While the former non-linear model is valid in the strict shallow water limit, the latter can be used in cases in which the wavelengths are closer to (albeit still larger than) the depth of the basin.
In the present work, we study turbulent solutions of the shallow water model and of the Boussinesq equations using direct numerical simulations. Previous numerical studies considered the Hamiltonian equations for a potential flow with a truncated non-linear term, or the kinetic equations resulting from weak turbulence theory at moderate spatial resolution [@dyachenko_weak_2004; @korotkevich_simultaneous_2008] (with the notable exception of [@yokoyama_statistics_2004]). Here, we solve the primitive equations, without truncating the non-linear terms, potentially allowing for the development of vortical motions and of strong interactions between waves, and with spatial resolutions up to $2048^2$ grid points.
The paper is organized as follows. In section \[equations\] we introduce both models, show the assumptions made in order to obtain them, derive their energy balance equations, and briefly discuss the predictions obtained in the framework of weak turbulence theory. In section \[simulations\] we describe the numerical methods employed and the simulations. Then, in section \[results\] we introduce several dimensionless numbers defined to characterize the flows, and present the numerical analysis and results. We present wavenumber energy spectra and fluxes, time resolved spectra (as a function of the wavenumber and frequency), frequency spectra, and probability density functions of the fluid free surface height. Finally, in section \[conclusion\] we present the conclusions. The most important results are: (a) As in previous experimental and observational studies [@smith_equilibrium_2003; @kaihatu_asymptotic_2007] we find now in simulations that different regimes arise depending on the fluid depth and the degree of nonlinearity of the system. (b) We obtain a power spectrum of the surface height compatible (within statistical uncertainties) with $\sim k^{-2}$ in the shallow water (non-dispersive) case, and one compatible with a $\sim k^{-4/3}$ spectrum as the fluid depth is increased using the Boussinesq (weakly dispersive) model. The latter spectrum is also compatible with predictions of weak turbulence theory [@onorato_four-wave_2008]. (c) Dispersion in the Boussinesq model results in more prominent small scale features and the development of rapidly varying waves. (d) In the weakly dispersive regime, the non-linear dispersion relation obtained from the simulations has two branches in a range of wavenumbers, one branch corresponding to non-dispersive waves, and another corresponding to dispersive waves. We interpret this as the result of short wavelength waves seeing an effectively deeper flow resulting from the interaction with waves with very long wavelength. (e) The probability density function of surface height can be approximated by both skewed normal and Tayfun distributions. In the latter case, the parameters of the distribution are compatible with those previously found in observations and experiments [@falcon_observation_2011].
The shallow water and Boussinesq equations {#equations}
==========================================
Let us consider a volume of an incompressible fluid with uniform and constant (in time) density, with its bottom surface in contact with a flat and rigid boundary, and free surface at pressure $p_0$. A sketch illustrating the configuration is shown in Fig. \[diagram\]; $x$ and $y$ are the horizontal coordinates, $z$ is the vertical one, $h$ is the height of the fluid column (i.e., the $z$ value at the free surface), $h_0$ is the height of the fluid column at rest, $L$ is a characteristic horizontal length, gravity acts on the $-\hat{z}$ direction and its value is given by $g$. It is assumed that $L \gg h_0$ as we are interested in shallow flows.
![The shallow water geometry considered in the simulations: $x$ and $y$ are the horizontal coordinates, and $z$ is the vertical coordinate. The surface height is $h$, with $h_0$ the height of the fluid column at rest. The fluid surface is at pressure $p_0$. $L$ is a characteristic horizontal length (assumed to be much larger than $h_0$). Gravity acts on the $-\hat{z}$ direction and its acceleration has a value of $g$. []{data-label="diagram"}](Figure1.eps){width="48.00000%"}
In the inviscid case, both the Euler equation and the incompressibility condition hold in the fluid body, $$\begin{aligned}
\label{euler_basico}
\frac{\partial \mathbf{v}}{\partial t}
+ (\mathbf{v} \cdot {\boldsymbol{\nabla}}) \mathbf{v}
&= - \frac{1}{\rho} {\boldsymbol{\nabla}}p - g \hat{z}, \\
{\boldsymbol{\nabla}}\cdot \mathbf{v} &= 0 .
\label{eq:incompressible}\end{aligned}$$ Under certain assumptions, to be discussed in the following paragraphs, the evolution of the free surface can be adequately described by means of a vector equation for the two horizontal components of the velocity at the interface, plus an equation for the local height of the fluid column.
Linear dispersion relation
--------------------------
Considering the case of very small amplitude waves, one can linearize the system of Eqs. and (see, e.g., [@landau_fluid_2004]). The solutions of the resulting equations are gravity waves with the following dispersion relation $$\label{rel_disp_tot}
\omega^2 = gk \frac{1 - e^{-2kh_0}}{1 + e^{-2kh_0}} .$$
We are interested in the shallow water case, i.e., when $h_0 \ll
L \Rightarrow h_0 k \ll 1 $. In that limit the following dispersion relation results $$\label{rel_disp_sw}
\omega = \sqrt{g h_0} k = c_0 k ,$$ where $c_0 = \sqrt{g h_0}$ is the phase velocity. Note in this case waves are not dispersive, unlike the general case given by Eq. .
Shallow water model
-------------------
It is possible to derive a set of non-linear equations for the surface height and the horizontal velocity at the surface by using the fact that the fluid layer is shallow. Considering the characteristic magnitudes of all quantities in Eq. ($L$, $p_0$, $h_0$, $g$, etc.), and using the fact that in a shallow flow $h_0 k \ll 1$ with $k=2\pi/L$, one obtains hydrostatic balance in the vertical direction (for further details, see [@pedlosky_geophysical_1987]), which results in the pressure profile $$p = \rho g (h-z) + p_0 .$$
As $h$ is not a function of $z$, neither will be the horizontal pressure gradient and the horizontal components of the velocity (as long as they do not depend initially on $z$). In this way, the horizontal components of Eq. can be written as $$\label{eqs_horz}
\frac{\partial \mathbf{u}}{\partial t} = -(\mathbf{u} \cdot {\boldsymbol{\nabla}}) \mathbf{u}
-g {\boldsymbol{\nabla}}h .$$ where $\mathbf{u}(x,y,t) = v_x \hat{x} + v_y \hat{y}$ is the horizontal velocity, and ${\boldsymbol{\nabla}}$ is now the horizontal gradient.
Using the fact that $v_x$ and $v_y$ are independent of $z$ we can integrate the incompressibility condition, obtaining $$v_z (x,y,z,t) = -z \left( \frac{\partial v_x}{\partial x} + \frac{\partial
v_y}{\partial y} \right) + \tilde{v}_z (x,y,t) .
\label{eq:vz}$$
Finally, by taking the appropriate boundary conditions and setting $z=h(x,y,t)$, Eq. provides an equation for the evolution of the height of the fluid column, namely $$\label{eq_h}
\frac{\partial h}{\partial t} + \frac{\partial }{\partial x} (h v_x)
+ \frac{\partial }{\partial y}(h v_y) = 0 .$$ Note that we do not have to assume irrotationality to derive neither Eq. nor Eq. .
If we linearize these equations, we find again the dispersion relation given by Eq. , as expected. In the presence of external forcing $\mathbf{F}$, and viscosity $\nu$, the equations can be written as $$\begin{aligned}
\label{sw_u} \frac{\partial \mathbf{u}}{\partial t} &=
-(\mathbf{u} \cdot {\boldsymbol{\nabla}}) \mathbf{u} -g {\boldsymbol{\nabla}}h + \frac{\nu}{h} {\boldsymbol{\nabla}}\cdot (
h {\boldsymbol{\nabla}}\mathbf{u}) + \mathbf{F}, \\ \label{sw_h} \frac{\partial h}{\partial
t} &= -{\boldsymbol{\nabla}}\cdot (h \mathbf{u}) . \end{aligned}$$ We will refer to this set of equations as the shallow water model, or SW model. In these equations the viscosity $\nu$ was added to the horizontal velocity field $\mathbf{u}$, which behaves as a compressible flow (i.e., it has non-zero divergence, see [@marche_derivation_2007]). This choice of the viscous term ensures conservation of the momentum $h \mathbf{u}$, and also that energy dissipation is always negative, as in Sec. \[energybalancesec\].
Boussinesq model
----------------
As the depth of the fluid increases, dispersion becomes important. There are several models that introduce weak dispersive effects perturbatively, but many are built to study waves propagating in only one direction. The Boussinesq equations for surface waves (see, e.g., [@whitham_linear_1974; @choi_nonlinear_1995]) provide a model to study weakly dispersive waves propagating in any direction. This model not only broadens the range of physical phenomena encompassed by the SW model, but adding dispersive effects also makes it more enticing to weak turbulence theory, for which dispersion effects are of the utmost importance.
Let us take a look at the first terms in the Taylor expansion of the dispersion relation in Eq. , $$\label{taylor_exp}
\omega^2 = c^2_0 k^2 - \frac{1}{3} c^2_0 h^2_0 k^4 + \ldots ,$$ where the first term is the non-dispersive shallow water term. The idea is to add terms to Eqs. and such that the linear dispersion relation of the new system coincides, up to the fourth order, with the expansion in Eq. . This can be done by adding the term $h^2_0 {\nabla^2}\partial_t \mathbf{u}/3$ to Eq. , resulting in the following system, $$\begin{aligned}
\frac{\partial \mathbf{u}}{\partial t} =& -(\mathbf{u} \cdot {\boldsymbol{\nabla}}) \mathbf{u} -g
{\boldsymbol{\nabla}}h + \frac{1}{3} h^2_0 {\nabla^2}\frac{\partial \mathbf{u}}{\partial t} \nonumber \\
{}& + \frac{\nu}{h} {\boldsymbol{\nabla}}\cdot ( h {\boldsymbol{\nabla}}\mathbf{u}) + \mathbf{F}, \label{bq_u} \\
\label{bq_h}
\frac{\partial h}{\partial t} =& -{\boldsymbol{\nabla}}\cdot (h \mathbf{u}) .\end{aligned}$$ We will refer to this system as the Boussinesq model, or BQ model. For $\mathbf{F}=0$ and $\nu=0$, the dispersion relation obtained by linearizing these equations is $$\label{rel_disp_bous}
\omega = \frac{c_0 k }{\sqrt{1 + \frac{h^2_0 k^2}{3}}} ,$$ which, up to the fourth order, coincides with Eq. .
Note that there are other choices for the extra term in Eq. that result in many formulations of the Boussinesq model, all compatible up to fourth order in a Taylor expansion in terms of $h_0 k$ [@whitham_linear_1974]. The formulation we use here was employed in previous studies of wave turbulence [@onorato_four-wave_2008], and is also easy to solve numerically using pseudospectral methods by writing Eq. as $$\frac{\partial \mathbf{u'}}{\partial t} = -(\mathbf{u} \cdot {\boldsymbol{\nabla}}) \mathbf{u} -g
{\boldsymbol{\nabla}}h + \frac{\nu}{h} {\boldsymbol{\nabla}}\cdot ( h {\boldsymbol{\nabla}}\mathbf{u}) +
\mathbf{F},
\label{eq:Helmholtz}$$ where $\mathbf{u'} = {\cal H}\mathbf{u}$, and where ${\cal H} = (1-h_0^2 {\nabla^2}/3)$ is the Helmholtz operator. This operator can be easily inverted in Fourier space [@mininni_numerical_2005; @mininni_numerical_2005-1], and the resulting equations can be efficiently solved by means of pseudospectral codes. It is interesting that the same operator appears in Lagrangian-averaged models [@foias_navier-stokes-alpha_2001]. In these models, and in regularized versions of the shallow water equations [@camassa_integrable_1993], it introduces dispersion that results in an accumulation of energy at small scales [@graham_highly_2007].
Energy balance {#energybalancesec}
--------------
An exact energy balance can be easily derived for the SW model. The equation is useful to verify conservation in pseudospectral codes. By taking the dot product of Eq. and $h \mathbf{u}$, setting ${\bf F} = 0$, and using Eq. , we obtain $$\begin{aligned}
\frac{\partial }{\partial t} \left( \frac{h u^2}{2} + g \frac{h^2}{2}
\right) = &- {\boldsymbol{\nabla}}\cdot \left( \frac{h u^2}{2} \mathbf{u} + g h^2 \mathbf{u}
\right) \\
&+ \nu \mathbf{u} \cdot [ {\boldsymbol{\nabla}}\cdot ( h {\boldsymbol{\nabla}}\mathbf{u}) ] .
\end{aligned}$$ Integrating in $x$ and $y$ over an area $A$ and taking periodic boundary conditions yields $$\label{balance}
\frac{\mathrm{d} E}{\mathrm{d} t} = - 2 \nu Z,$$ where $$E = \frac{1}{A} \iint \left( \frac{h u^2}{2} + g \frac{h^2}{2} \right)
\mathrm{d} x \mathrm{d} y$$ is the mean total energy, and $$Z = \frac{1}{A} \iint \frac{h \lvert {\boldsymbol{\nabla}}\mathbf{u}
\rvert^2}{2} \mathrm{d} x \mathrm{d} y$$ is a mean pseudo-enstrophy, such that $-2 \nu Z$ is the mean energy dissipation rate. As $h$ is always positive, the energy dissipation is always negative. The total energy is conserved when $\nu = 0$.
Now we can define $$U = \frac{1}{A} \iint \label{ec_cin}
\frac{h u^2}{2}\mathrm{d} x \mathrm{d} y$$ as the mean kinetic energy, and $$\label{ec_pot}
V = \frac{1}{A} \iint g \frac{h^2}{2} \mathrm{d} x \mathrm{d} y$$ as the mean potential energy, such that the sum of both gives the mean total energy $E$.
The dispersive term present in the BQ model changes the balance given by Eq. . However, since the extra term is of order $(h_0/L)^2$, as long as we are in a sufficiently shallow flow it will be very small, and therefore, negligible for the conservation of energy. We verified this is the case in our numerical simulations.
\[sec:weak\]Weak Turbulence prediction
--------------------------------------
We briefly present some results obtained in the framework of weak turbulence theory for the BQ model (as the derivation is a bit cumbersome, only a general outline will be given here; please see [@onorato_four-wave_2008] for details). Weak turbulence is studied in the BQ model assuming the fluid is inviscid and irrotational, so that the velocity can be written in terms of a velocity potential. To obtain a statistical description of the wave field, it is also assumed that it is homogeneous and that the free modes are uncorrelated.
At first sight, the quadratic nonlinear terms in Eqs. and indicate modes interact in triads, with the wave vectors of the three interacting modes lying over a triangle, and the three frequencies satisfying the resonant condition (see, e.g., [@nazarenko_wave_2011]) $$\begin{aligned}
\mathbf{k} &= \mathbf{p} + \mathbf{q} \\
\omega(\mathbf{k}) &= \omega(\mathbf{p}) + \omega(\mathbf{q}) .\end{aligned}$$ However, as there are no three wave vectors $\mathbf{k},
\mathbf{p}, \mathbf{q}$ that satisfy these two conditions when the dispersion relation is given by Eq. , three wave interactions are forbidden. Thus, only four wave interactions are present (which do satisfy their corresponding condition).
After a transformation of the fields, it is possible to write an equation for the evolution of the two-point correlator of the transformed fields. This is the so-called kinetic equation, and has the following form $$\begin{gathered}
\begin{aligned}
\frac{\partial N_0}{\partial t} = 4 \pi &\int \lvert T_{0,1,2,3} \rvert^2
N_0 N_1 N_2 N_3 \\
& \left( \frac{1}{N_0} + \frac{1}{N_1}- \frac{1}{N_2}-
\frac{1}{N_3} \right) \\
& \delta (\mathbf{k}_0 + \mathbf{k}_0 - \mathbf{k}_2 - \mathbf{k}_3) \\
&\delta (\omega_0 + \omega_0 - \omega_2 - \omega_3) \textrm{d}
\mathbf{k}_{123},
\label{kinetic_eq}
\end{aligned}\end{gathered}$$ where $N_i = N(\mathbf{k}_i)$ is the wave action spectral density (i.e., the two-point correlator of the wave action, the latter being a quantity proportional to the surface height), the deltas express the fact that interactions are between four wave vectors and their associated frequencies, $T_{0,1,2,3}$ is the coupling coefficient between the four modes, and $\textrm{d} \mathbf{k}_{123} = \textrm{d}
\mathbf{k}_{1} \textrm{d} \mathbf{k}_{2} \textrm{d} \mathbf{k}_{3}$. From this equation, dimensional analysis yields the following expression for the energy spectrum $$\label{predicc_wt}
E(k) \sim k^{-4/3} .$$ From this spectrum and using dimensional analysis, it is easy to show that in the presence of dissipation, the dissipation wavenumber in such a flow is $k_\eta \sim [\epsilon/(h^2_0 \nu^3)]^{1/5}$, where $\epsilon$ is the mean energy injection rate.
A scaling compatible with a $\sim k^{-4/3}$ spectrum was observed in laboratory and field datasets [@smith_equilibrium_2003; @kaihatu_asymptotic_2007], where a spectrum compatible with $\sim k^{-2}$ was also found in shallower regions of the fluid.
The prediction in Eq. applies to the BQ model, when dispersion is not negligible. Before proceeding, we should comment on some peculiarities of the SW model regarding wave turbulence. First, an inspection of its dispersion relation, Eq. , indicates that three wave interactions are possible in this model, and as a result the arguments above for four-wave interactions do not apply. Weak turbulence theory can be used in systems with three-waves interactions (with the case of deep water flows being a paradigmatic one, but see also the case of rotating [@galtier_weak_2003] and of magnetohydronamic [@galtier_weak_2000] flows). However, the SW model is non-dispersive, and as a result the resonance condition is only satisfied for collinear wave vectors. Resonant interactions can then only couple modes that propagate in the same direction (i.e., along the ray of the wave), and non-resonant interactions must be taken into account to consider other couplings. But more importantly, dispersion is crucial in weak turbulence theory to have decorrelation between different waves: without dispersion, all modes propagate with the same velocity, and the modes initially correlated remain correlated for all times (see, e.g., [@lvov_statistical_1997] for a discussion of these effects in the context of acoustic turbulence).
![[*(Color online)*]{} Total energy as a function of time for simulation $A06$ (see Table \[table1\]). As the fluid starts from rest and the forcing is applied, energy increases until it reaches a turbulent steady state (note energy at $t=0$ is different from zero, as the flow potential energy is never zero). All the analysis of the simulations was performed after the simulations reached the turbulent steady state. [*Inset:*]{} Energy balance as a function of time (see Eq. ). Note the balance is satisfied up to the seventh decimal place. []{data-label="eng_cons"}](Figure2.eps){width="48.00000%"}
Numerical simulations {#simulations}
=====================
------------ ----- -------- ------- -------------------- -------------------- ----------- --------------------- ------
Simulation Re Fr $D_s$ $N_l$ $h_0/L_0$ $f_0/U_0$ $[k_{f_1},k_{f_2}]$ $N$
$A01$ 260 0.005 0.27 $1.2\times10^{-5}$ $8.0\times10^{-4}$ 0.76 \[3,8\] 1024
$A02$ 370 0.0059 0.22 $1.6\times10^{-5}$ $6.4\times10^{-4}$ 0.71 \[3,8\] 1024
$A03$ 820 0.0075 0.33 $2.6\times10^{-5}$ $4.9\times10^{-4}$ 0.64 \[3,8\] 2048
$A04$ 760 0.0067 0.36 $2.2\times10^{-5}$ $5.3\times10^{-4}$ 0.69 \[3,8\] 2048
$A05$ 760 0.007 0.33 $2.3\times10^{-5}$ $4.8\times10^{-4}$ 0.69 \[3,8\] 2048
$A06$ 360 0.0066 0.33 $2.1\times10^{-5}$ $4.8\times10^{-4}$ 0.73 \[3,8\] 2048
$A07$ 570 0.0091 0.43 $4.1\times10^{-5}$ $6.4\times10^{-4}$ 0.92 \[3,8\] 2048
$A08$ 350 0.0083 0.43 $3.4\times10^{-5}$ $6.4\times10^{-4}$ 1 \[3,8\] 2048
$A09$ 290 0.0086 0.43 $3.6\times10^{-5}$ $6.4\times10^{-4}$ 0.98 \[3,8\] 2048
$A10$ 420 0.012 0.43 $7.8\times10^{-5}$ $6.4\times10^{-4}$ 1.4 \[3,8\] 2048
------------ ----- -------- ------- -------------------- -------------------- ----------- --------------------- ------
------------ ------ -------- ------- -------------------- -------------------- ----------- --------------------- ------
Simulation Re Fr $D_s$ $N_l$ $h_0/L_0$ $f_0/U_0$ $[k_{f_1},k_{f_2}]$ $N$
$B01$ 5600 0.022 0.14 $2.5\times10^{-4}$ $8.0\times10^{-4}$ 0.45 \[1,5\] 512
$B02$ 3700 0.015 0.14 $1.1\times10^{-4}$ $8.0\times10^{-4}$ 0.34 \[1,5\] 512
$B03$ 5000 0.012 0.14 $7.2\times10^{-5}$ $8.0\times10^{-4}$ 0.29 \[1,5\] 512
$B04$ 7100 0.013 0.11 $9.1\times10^{-5}$ $3.2\times10^{-4}$ 0.24 \[1,5\] 1024
$B05$ 830 0.005 0.11 $1.2\times10^{-5}$ $3.2\times10^{-4}$ 0.8 \[3,8\] 1024
$B06$ 1200 0.011 0.11 $6.2\times10^{-5}$ $3.2\times10^{-4}$ 0.57 \[3,8\] 1024
$B07$ 120 0.0046 0.14 $1.0\times10^{-5}$ $8.0\times10^{-4}$ 0.82 \[3,8\] 512
$B08$ 980 0.012 0.17 $7.9\times10^{-5}$ $2.5\times10^{-4}$ 0.54 \[3,8\] 2048
$B09$ 2500 0.038 0.27 $7.4\times10^{-4}$ $8.0\times10^{-4}$ 0.31 \[3,8\] 1024
$B10$ 670 0.0042 0.17 $8.5\times10^{-6}$ $2.5\times10^{-4}$ 0.79 \[3,8\] 2048
$B_{SW}11$ 100 0.0039 0.14 $7.6\times10^{-6}$ $8.0\times10^{-4}$ 0.96 \[3,8\] 512
$B_{SW}12$ 470 0.013 0.11 $9.0\times10^{-5}$ $3.2\times10^{-4}$ 0.24 \[1,5\] 1024
------------ ------ -------- ------- -------------------- -------------------- ----------- --------------------- ------
------------ ------ -------- ------- -------------------- -------------------- ----------- --------------------- ------
Simulation Re Fr $D_s$ $N_l$ $h_0/L_0$ $f_0/U_0$ $[k_{f_1},k_{f_2}]$ $N$
$C01$ 1400 0.0067 0.27 $1.9\times10^{-5}$ $8.0\times10^{-4}$ 0.56 \[3,8\] 1024
$C02$ 1400 0.0073 0.24 $2.4\times10^{-5}$ $7.0\times10^{-4}$ 0.55 \[3,8\] 1024
$C03$ 1900 0.0092 0.31 $4.1\times10^{-5}$ $4.6\times10^{-4}$ 0.54 \[3,8\] 2048
$C04$ 1400 0.0057 0.43 $1.6\times10^{-5}$ $6.4\times10^{-4}$ 0.74 \[3,8\] 2048
$C05$ 470 0.0067 0.54 $2.2\times10^{-5}$ $8.0\times10^{-4}$ 1.1 \[3,8\] 2048
------------ ------ -------- ------- -------------------- -------------------- ----------- --------------------- ------
We performed several numerical simulations of both the shallow water and the Boussinesq models. These were done using the GHOST code [@gomez_mhd_2005; @gomez_parallel_2005; @mininni_hybrid_2011], which uses a pseudospectral method with periodic boundary conditions on a $L_0 \times L_0 = 2 \pi \times 2 \pi$ sized box (with $L_0$ the box length), the “$2/3$ rule” for the dealiasing [@canuto_spectral_1988], explicit second order Runge-Kutta for time stepping, and is parallelized using MPI and OpenMP. Almost all simulations shown here were done on grids of $N^2=2058^2$ points, with a few on grids of $N^2=1024^2$ or $512^2$ points (with $N$ the linear resolution). As a result of dealiasing, the maximum resolved wavenumber is $$k_\textrm{max}=N/3 .
\label{dealising}$$ Note all magnitudes in the code are dimensionless, with the smallest wavenumber $k_\textrm{min}=2\pi/L_0=1$, and the largest wavenumber $k_\textrm{max}=2\pi/\lambda_\textrm{min}$ being associated with the minimum resolved scale $\lambda_\textrm{min}$.
All runs are direct numerical simulations, with all relevant space and time scales resolved explicitly. The pseudospectral method with the $2/3$ rule is equivalent to a purely spectral method [@canuto_spectral_1988]: it converges exponentially fast, it conserves all quadratic invariants of the equations (i.e., there is no numerical dissipation introduced by the method), and it also has no numerical dispersion. All this was verified explicitly during the development of the code, using several test problems for the SW and BQ equations.
Most previous numerical studies on wave turbulence in gravity waves were done at lower resolutions, with the exception of [@yokoyama_statistics_2004]. But the key difference between previous simulations and the ones presented here (besides the fact that these are for shallow flows, not for deep flows) is that the physical model we use does not assume potential flow, and, more importantly, we do not truncate the non-linear term, thus retaining all high order non-linearities. Another difference is that we do not introduce an artificial dissipation term as it is usually done, but one based on physical grounds. The key motivation for these choices is to be able to compare with experiments in the future, where vortical motions can develop, and where dissipation also plays a non-negligible role. To achieve higher resolutions than the ones studied here becomes increasingly more expensive as the BQ model is dispersive.
All the simulations were started from the fluid at rest. An external mechanical forcing injected energy in the system, allowing it to reach for sufficiently long times an out-of-equilibrium turbulent steady state, after an initial transient. To excite waves, and prevent external injection of energy into vortical motions, the forcing had the following form $$\mathbf{F} = {\boldsymbol{\nabla}}f,$$ where $f$ is a randomly generated scalar function, with a time correlation of one time unit, amplitude $f_0$, and applied in a band of wavenumbers in Fourier space between $k_{f_1}$ and $k_{f_2}$ (see Tables \[table1\], \[table2\], and \[table3\]). Note that having a mechanical forcing in the momentum equation adds an extra term to the right hand side of Eq. , $$\label{balance_full}
\frac{\mathrm{d} E}{\mathrm{d} t} = - 2 \nu Z + \epsilon,$$ where the mean energy injection rate can be computed as $$\label{eps_balance}
\epsilon = \frac{1}{A} \iint\limits_A h \mathbf{u}\cdot \mathbf{f} \mathrm{d}x
\mathrm{d}y .$$
Under the procedure described above, the typical evolution of the energy in a numerical simulation is shown in Fig. \[eng\_cons\]. The energy starts from the value corresponding to the fluid at rest (i.e., all the energy is the potential energy associated with the equilibrium height $h_0$). The total energy then grows under the action of the external mechanical forcing, and after $t \approx 80$ the system reaches a turbulent steady state in which the energy fluctuates around a mean value, and in which the energy injection and dissipation are equilibrated on the average. Even though pseudospectral methods are known to introduce no numerical dissipation, in the inset of Fig. \[eng\_cons\] we also show explicitly that the energy balance (Eq. ) is satisfied with an error of order $10^{-7}$, which remains stable and does not grow even after integrating for very long times.
To ensure that the flow in the simulations remained shallow for all excited wavenumbers, we enforced the following condition $$\begin{gathered}
\frac{h_0}{\lambda_\textrm{min}} = h_0 \frac{k_\textrm{max}}{2
\pi} < 1 \nonumber \\
\Rightarrow h_0 < \frac{6 \pi}{N} .
\label{cond_disp}\end{gathered}$$ where $\lambda_\textrm{min}$ is, as already mentioned, the shortest wavelength resolved by the code in virtue of the condition given by Eq. .
\[results\] Results
===================
Description and classification of the simulations
-------------------------------------------------
The spectral behavior of the flow in the simulations depends on the external parameters. We can independently control the height of the fluid at rest $h_0$, the viscosity $\nu$, the gravity acceleration $g$, the amplitude of the forcing $f_0$, the range of wavenumbers in which the force is applied, and the linear resolution $N$. However, all these parameters can be reduced to a smaller set of dimensionless controlling parameters.
One of these parameters is the Froude number $$\textrm{Fr}=\frac{U_0}{\sqrt{g h_0}},$$ which measures the ratio of inertia to gravity acceleration in the momentum equation, and where $U_0$ is the r.m.s. velocity.
Another dimensionless parameter is the non-linear number, $N_l$. In order to be in the regime of weak turbulence, nonlinearities should be small. The effect of nonlinearities can be measured by how large perturbations in $h$ are compared to $h_0$, so we define $N_l$ as $$N_l=\frac{h_\textrm{rms}-h_0}{h_0},$$ where $h_\textrm{rms}$ is the r.m.s. value of $h$.
The two remaining dimensionless numbers are the Reynolds number, $$\label{reynolds}
\textrm{Re} =\frac{U_0 L_f}{\nu},$$ where $L_f$ is the forcing scale (defined as $2 \pi/k_{f_0}$), and what we will call the dispersivity, $D_s$, defined as $$\label{dispersivity}
D_s = h_0 k_\textrm{max} = \frac{2\pi
h_0}{\lambda_\textrm{min}} = \frac{N h_0}{6 \pi}$$ following Eq. . This last number, only relevant for the Boussinesq model, measures how strong the dispersion is at the smallest scales, and for sufficiently small $D_s$ we can expect the solutions of the Boussinesq model to converge to the solutions of the shallow water model. In fact, it is easy to show from the weak turbulence spectrum in Eq. that when the maximum resolved wavenumber $k_{\textrm{max}}$ is associated with the dissipation wavenumber $k_\eta$, then $$\textrm{Re} \sim \frac{U_0 L_f}{h_0 \epsilon^{1/3}} D^{5/3}_s .
\label{reycurve}$$ Decreasing $D_s$ below the value given by this relation should result in negligible dispersion at all resolved wavenumbers. Note that the level of dispersion in a given Boussinesq run depends on the wavenumber, and $D_s$ actually quantifies the strongest possible dispersion at the smallest scales in the flow.
By qualitatively assessing each run, we can classify them into three sets, $A$, $B$, and $C$. In tables \[table1\], \[table2\], \[table3\] the different dimensionless parameters, along with a few other useful quantities, are given for each simulation in each set, respectively. How and why these three sets differ from each other will be made clear in the following sections, when we discuss the actual results. But, for the moment, it is fruitful to analyze the behavior of the values of $\textrm{Re}$ and $D_s$ in each set, so as to keep them in mind for later on.
The values of $\textrm{Re}$ and $D_s$ for all the Boussinessq runs are shown in Fig. \[phase\_dia\]. As a reference, Fig. \[phase\_dia\] also shows the curve given by Eq. with $U_0 L_f/(h_0 \epsilon^{1/3})$ estimated from the values from the simulations in set $A$. Points below that curve are expected to have non-negligible dispersion. Runs in set $A$ have relatively small $\textrm{Re}$ ($\lesssim 1000$), and $D_s$ varying between $\approx 0.02$ and $\approx 0.05$. In other words, dispersion effects in runs in set $A$ are important. Runs in set $B$ have smaller values of $D_s$ (except for one run with $D_s \approx 0.27$, all other runs have $D_s <0.2$), and $\textrm{Re}$ varying between $\approx 100$ and $\approx 7000$. These runs have small or negligible dispersion, and note all the SW runs we performed belong to this set. The runs in set $C$ are intermediate between these two regimes.
![[*(Color online)*]{} Values of the Reynolds number $\textrm{Re}$ and dispersivity $D_s$ for all the Boussinesq runs, separated into three sets, $A$ (circles in the gray/red region), $B$ (triangles in the dark/blue region), and $C$ (stars in the light/green region), according to their different spectral behavior as discussed in Sec. \[spectra\]. The boundaries separating the three regions are arbitrary. The solid white curve corresponds to $\textrm{Re} \sim D^{5/3}_s$; points below that curve are expected to have non-negligible dispersion (see Eq. ). Note that runs in set $A$ have relatively small $\textrm{Re}$ but larger dispersion, while runs in set $B$ have either small or negligible dispersion.[]{data-label="phase_dia"}](Figure3.eps){width="48.00000%"}
Finally, although the mechanical forcing we use introduces no vorticity in the horizontal velocity field, some vorticity is spontaneously generated as the flow evolves. This is probably also the case in experiments. In order to quantify the presence of vortical structures, we calculated the ratio of vorticity to divergence in the horizontal velocity field $$\frac{\langle | {\boldsymbol{\nabla}}\times \mathbf{u} | \rangle }
{\langle | {\boldsymbol{\nabla}}\cdot \mathbf{u} | \rangle } ,$$ which turns out to be $\approx 0.1$ for all simulations. As a result, although the flow is not perfectly irrotational, the amplitude of vortical modes is small compared with the amplitude of modes associated with the waves.
Energy spectra {#spectra}
--------------
The power spectrum of $h$ (proportional to the spectrum of the potential energy) as a function of the wave number is shown in Fig. \[spectra\_far\] for runs $A06$, $B08$, and $C02$. Figure \[spectra\_zoom\] shows a close-up of the same spectrum in the inertial range. It is clearly seen that runs in each set show a different behavior. On the one hand, the run belonging to group $A$ has an inertial range compatible with $\sim k^{-4/3}$ scaling, which is the spectra predicted by weak turbulence. On the other hand, the run in set $B$ displays an inertial range compatible with $\sim k^{-2}$ dependency. While this spectrum is not predicted by weak turbulence, it was observed before in experiments and observations [@kaihatu_asymptotic_2007]. The run in group $C$ shows a shallower spectrum with no clear inertial range. We think of runs in this set as transitional between the other two.
The other runs in sets $A$, $B$, and $C$ show similar power spectra for $h$. To show this, we present the compensated spectra for the simulations in sets $A$ and $B$ in Figs. \[compensado\_A\] and \[compensado\_B\] respectively (simulations from set $C$ do not have a clearly defined inertial range and are therefore not shown). The simulations from set $A$ are compensated by $h^{2/3}_0
\epsilon^{2/3} k^{-4/3}$ (which is the weak turbulence spectrum, using the height of the fluid column at rest, $h_0$, and the energy injection rate, $\epsilon$, as prefactors), while the ones in set $B$ are compensated by $g h_0
k^{-2}$ (more details on the choice of the prefactor are given below). These figures indicate that, within statistical uncertainties, all spectra in each set collapse to the same power laws, and that the simulations are well converged from the point of view of spatial resolution. Furthermore, we verified that the energy flux is approximately constant in the scales corresponding to the inertial range of each simulation. Within the limitations of spatial resolution and the drop in the flux for large wavenumbers caused by viscous dissipation, an incipient inertial range can be identified in the flux of each simulation. Figure \[flux\] shows the instantaneous energy flux (normalized by the energy injection rate $\overline{\epsilon}$ averaged over time) as a function of $k$ for several simulations in sets $A$ and $B$. The energy flux $\Pi(k)$ was calculated from the energy balance equation in Fourier space, as is usually done for turbulent flows. Figure \[flux\] also shows the normalized energy dissipation rate as a function of time (equivalent to the normalized energy flux as a function of time) for the same runs, to show that this quantity fluctuates around a mean value in the turbulent steady state.
The kinetic energy spectrum is similar to the power spectrum of $h$, and in approximate equipartition with the potential energy spectrum once the system reaches a turbulent steady state. It is interesting to analyse this in the light of the values of the dimensionless parameters in the runs as shown in Fig. \[phase\_dia\]. As was explained in the previous section, set $A$ corresponds to runs with lower Reynolds number and larger dispersivity ($\textrm{Re}\lesssim 1000$, and $D_s$ varying between $\approx
0.02$ and $\approx 0.05$). As a result, these runs can be expected to display weak turbulence behavior as described in Section \[sec:weak\], because the nonlinearities are not so large as to break the weak turbulence hypothesis [@onorato_four-wave_2008], and the dispersion is not so low as to render the higher order terms of Eq. negligible (in which case four-wave interactions would no longer be dominant, and the hypothesis used to derive Eq. would not be satisfied). In contrast, runs in set $B$ have larger $\textrm{Re}$ and lower $D_s$ (except for one run with $D_s \approx
0.27$, all other runs have $D_s <0.2$, and $\textrm{Re}$ varying between $\approx 100$ and $\approx 7000$). In this case dispersion is smaller or negligible, while nonlinearities can be expected to be larger, two conditions that render the derivation resulting in Eq. invalid.
![[*(Color online)*]{} Power spectrum of $h$ (proportional to the spectrum of the potential energy) for runs $A06$ (BQ model, $2048^2$ grid points, $\textrm{Re} = 360$, and $D_s = 0.33$), $B08$ (BQ model, $2048^2$ grid points, $\textrm{Re} = 980$, and $D_s = 0.17$), and $C02$ (BQ model, $1024^2$ grid points, $\textrm{Re} = 1430$, and $D_s = 0.24$). Two power laws, $\sim k^{-4/3}$ and $\sim k^{-2}$, are shown as references.[]{data-label="spectra_far"}](Figure4.eps){width="48.00000%"}
![[*(Color online)*]{} Detail of the three spectra in Fig. \[spectra\_far\] for a subset of wavenumbers to show the inertial range of the runs. Note the scaling of runs $A06$ and $B08$.[]{data-label="spectra_zoom"}](Figure5.eps){width="48.00000%"}
![[*(Color online)*]{} Compensated spectrum of potential energy for several simulations in set $A$. The spectra are compensated by $h^{2/3}_0 \epsilon^{2/3} k^{-4/3}$. The average slope for all the runs is $-1.34\pm0.12$.[]{data-label="compensado_A"}](Figure6.eps){width="48.00000%"}
![[*(Color online)*]{} Compensated spectrum of potential energy for several simulations in set $B$. The spectra are compensated by $g h_0 k^{-2}$. The average slope for all the runs is $-2.18\pm0.29$.[]{data-label="compensado_B"}](Figure7.eps){width="48.00000%"}
![[*(Color online)*]{} (a) Energy flux (normalized by the mean energy injection rate) as a function of $k$. For each simulation, a range of wavenumbers can be identified for which $\Pi(k)$ remains approximately constant, and this range is in reasonably good agreement with the inertial ranges identified in Figs. \[compensado\_A\] and \[compensado\_B\]. (b) Energy dissipation rate (normalized by the averaged in time energy injection rate) as a function of time for the same simulations. []{data-label="flux"}](Figure8a.eps "fig:"){width="48.00000%"}\
![[*(Color online)*]{} (a) Energy flux (normalized by the mean energy injection rate) as a function of $k$. For each simulation, a range of wavenumbers can be identified for which $\Pi(k)$ remains approximately constant, and this range is in reasonably good agreement with the inertial ranges identified in Figs. \[compensado\_A\] and \[compensado\_B\]. (b) Energy dissipation rate (normalized by the averaged in time energy injection rate) as a function of time for the same simulations. []{data-label="flux"}](Figure8b.eps "fig:"){width="48.00000%"}
![[*(Color online)*]{} Power spectrum of $h$ for simulations $B04$ and $B_{SW}12$. The former corresponds to a numerical solution of the BQ model, while the later to a solution of the SW model. A $\sim k^{-2}$ power law is indicated as a reference.[]{data-label="comp_sp"}](Figure9.eps){width="48.00000%"}
Moreover, the $\sim g h_0 k^{-2}$ spectra observed in the simulations in set $B$ include those runs that solve the SW model. Therefore, these spectra cannot be explained by weak turbulence, as SW simulations have no dispersion and the arguments in Section \[sec:weak\] do not apply. Also, note that in the non-dispersive limit, for constant and fixed $h$, the SW equations can be reduced to the two-dimensional Burgers equations, which amplify negative field gradients by strong nonlinearities resulting in sharp fronts in the velocity. Such a field would actually have a spectrum $\sim k^{-2}$ (note this is also the behavior expected for two-dimensional non-dispersive acoustic turbulence [@lvov_statistical_1997], that also develops sharp fronts). The spectrum can be obtained from dimensional analysis and the scaling that results for the energy is equivalent to Phillips’ spectrum [@phillips_equilibrium_1958] but in two dimensions. In the presence of strong nonlinearities, we can assume that the nonlinear and gravity terms are of the same order, $$\mathbf{u} \cdot {\boldsymbol{\nabla}}\mathbf{u} \sim g {\boldsymbol{\nabla}}h .$$ It is also reasonable to assume that the kinetic and potential energies will be of the same order (i.e., in equipartition) in the turbulent steady state. This implies that $g$ is the only dimensional constant the spectra can depend on. This is precisely how Phillips derived his spectrum.
With these assumptions in mind, it is easy to obtain the observed spectra. The energy spectrum has units of energy in the fluid column per unit surface per wavenumber, $E(k) \sim h_0 u^2/k$, and assuming $E(k) \sim g h_0 k^{-\alpha}$, from dimensional analysis the only possible solution is $$E(k) \sim g h_0 k^{-2} .$$ The independence of the spectrum on the energy injection rate suggests that the energy transfer between the different scales must take place by a mechanism such as wave breaking in the case of Phillips’ spectrum, which occurs when the slope of the surface is larger than a critical value, or by nonlinear wave steepening in our case (which is finally regularized by the viscosity). Such a mechanism is independent of the power injected by external forces. Of course, this can only hold in a region of parameter space, as in the presence of weak forcing and dispersion, the solution in Eq. is expected instead.
In summary, based on the numerical results, the simulations with weaker forcing and higher dispersion develop a spectrum compatible with the predictions from weak turbulence theory, while the runs with stronger forcing or with less (or no) dispersion are compatible with dimensional analysis based on strong turbulence arguments.
Comparison between SW and BQ models {#comp_sw_bq}
-----------------------------------
![One dimensional cut of the height $h$ in the turbulent steady state of runs $B04$ and $B_{SW}12$, at the same time. The former run corresponds to a numerical solution of the BQ model, while the latter to a solution of the SW model. While the long length scales show the same behavior in both runs, note the BQ model has larger fluctuations at short length scales. Both runs were computed with a linear resolution of $N=1024$ grid points, and the fast fluctuations are well resolved.[]{data-label="secc20"}](Figure10.eps){width="48.00000%"}
All simulations of the SW model belong to set $B$, as that is the set of runs that has negligible or no dispersion. All other sets have moderate dispersion, and as a result the flow dynamics cannot be captured by the SW model. Note that runs in set $B$ are also the runs with an inertial range compatible with $\sim
k^{-2}$ scaling. However, the BQ and SW runs in set $B$ are not identical. In this subsection we discuss the differences between these runs.
As an example of two runs with and without dispersive effects, the power spectra of $h$ for runs $B04$ and $B_{SW}12$ are shown in Fig. \[comp\_sp\]. Both simulations have the same parameters, except for the viscosity which is larger in the simulation using the SW model. At small wavenumbers, where dispersion is negligible, the spectra of the BQ and SW models coincides. For wavenumbers larger than $\approx 30$, dispersion in the BQ model becomes important and a bump (an accumulation of energy at small scales) develops. This accumulation in the BQ model results in an increased dissipation (as dissipation is proportional to $k^2 E(k)$), thus allowing us to simulate the system with smaller viscosity. This difference at large wavenumbers is the most distinct feature in the two spectra in Fig. \[comp\_sp\].
As a result of the extra power at larger wavenumbers, dispersion in the BQ model results in more prominent small scale features, and in rapidly varying waves. As an example, Fig. \[secc20\] shows a transversal cut in the elevation field for runs $B04$ and $B_{SW}12$. The cuts are taken at the same place and at the same time in both runs. Even though both simulations have the same behavior at large scales, at short length scales the BQ model presents fast fluctuations. These fluctuations are well resolved (the cut corresponds to $1024$ grid points), and there is no indication that resolution is insufficient to resolve the sharp gradients. In the BQ model, while the large scales may correspond to a shallow flow, as long as there is enough scale separation, there will always be a wavenumber where the finite depth effects can be seen. Thus, the Boussinesq equations provide an interesting model to study weakly dispersive waves.
Regarding the accumulation of energy that leads to a flatter spectrum for high wavenumbers in some of the BQ simulations (for several runs in set $B$ as can be seen in Fig. \[compensado\_B\], but specially in the runs in set $C$), such an accumulation has been observed before in turbulent flows. As mentioned above, we verified that this accumulation is not the result of insufficient resolution (e.g., by comparing the runs with different grid points $N$). The accumulation of energy in the spectrum near the dissipative range is often termed “bottleneck”, and bottlenecks can have dissipative [@falkovich_bottleneck_1994] or dispersive [@graham_highly_2007; @krstulovic_dispersive_2011] origins. In the former case, the accumulation results from the viscous damping of the triads at small scales, resulting in a decrease of the energy flux. Such a viscous bottleneck should be visible also in the non-dispersive simulations, and its absence in those runs indicates a dispersive origin. In the latter case, the bottleneck arises from the increasingly harder to satisfy resonant condition for the wave frequencies, as the waves become faster at smaller scales. Models with a field filtered by the Helmholtz operator (as is the case for the BQ model, see Eq. \[eq:Helmholtz\]) tend to develop a bottleneck (see [@graham_highly_2007] for a detailed description of its origin). A qualitative way to explain the tendency towards a flatter spectrum in the BQ model can be obtained by assuming that dispersion is strong enough for the dispersive term to be balanced with the buoyancy and with the non-linear terms in the BQ equations (i.e., all terms are of the same order). Then the energy spectra can depend only on both $g$ and $h_0$, and a possible solution is $E(k)
\sim g h_0^2$. A detailed study of the origin of this bottleneck is left for future work.
At this point it is worth pointing out that when $D_s \approx 1$ and dispersion becomes too strong, the Boussinesq approximation breaks down as more terms in the Taylor expansion in Eq. should be preserved. As a result, the Boussinesq approximation is useful as long as $D_s < 1$ at the smallest excited scales in the system. On the other hand, from Fig. \[phase\_dia\], if $D_s \lesssim 0.15$ the behavior of the system in the inertial range is that of a shallow water flow for all Reynolds numbers studied.
![Power spectrum $E_h(k,\omega)$ for simulation $B04$. The darker regions correspond to larger power density, while the lighter regions correspond to smaller power density. (a) Normalized power spectrum $E_h(k,\omega)/E_h(k)$. (b) Non-normalized power spectrum. The white dashed line appearing in the bottom panel indicates the linear dispersion relation from Eq. . Note that as in this run dispersion is negligible, the dispersion relation is almost that given by Eq. , and non-dispersive.[]{data-label="cubo32"}](Figure11a.eps "fig:"){width="48.00000%"} ![Power spectrum $E_h(k,\omega)$ for simulation $B04$. The darker regions correspond to larger power density, while the lighter regions correspond to smaller power density. (a) Normalized power spectrum $E_h(k,\omega)/E_h(k)$. (b) Non-normalized power spectrum. The white dashed line appearing in the bottom panel indicates the linear dispersion relation from Eq. . Note that as in this run dispersion is negligible, the dispersion relation is almost that given by Eq. , and non-dispersive.[]{data-label="cubo32"}](Figure11b.eps "fig:"){width="48.00000%"}
Time-resolved spectra and non-linear dispersion relations
---------------------------------------------------------
Wavenumber spectra, as the spectra discussed so far, give information of how energy is distributed in spatial scales, but do not provide a quantitative estimate of how much energy in the system is associated with wave motions. A frequency spectrum $E(\omega)$ is often obtained from the wavenumber spectrum $E(k)$ using the dispersion relation (or vice versa). However, in systems that can sustain both wave and vortical motions there is no clear justification to use the dispersion relation to go from one spectrum to the other.
A quantification of the amount of energy in waves, and on whether non-linear effects change the dispersion relation of the system from the linear one, can be directly obtained from the frequency and wavenumber spectrum $E(k, \omega)$ without any assumption. The spectrum $E(k, \omega)$ can be computed by storing the Fourier coefficients of the height $\hat{h}({\bf k}, t)$ as a function of time (as well as the Fourier coefficients of the velocity field), then computing the Fourier transform in time, and finally computing the isotropic power spectrum by averaging in the $(k_x, k_y)$-plane. To this end, several large-scale wave periods and turnover times must be stored (to resolve the slowest frequencies in the system), with sufficient time resolution $\Delta t$ to resolve the fastest frequencies. In the analysis we show below, time series spanning at least three periods of the slowest waves were used, and with time resolution $\Delta t \approx 3 \times 10^{-4}$.
![Power spectrum $E_h(k,\omega)$ for simulation $A02$. The darker regions correspond to larger power density, while the lighter regions correspond to smaller power density. (a) Normalized power spectrum. (b) Non-normalized power spectrum. The white dashed line appearing in the bottom panel indicates the (non-dispersive) linear dispersion relation from Eq. , and the white dash-dotted line indicates the BQ dispersion relation from Eq. .[]{data-label="cubo20"}](Figure12a.eps "fig:"){width="48.00000%"} ![Power spectrum $E_h(k,\omega)$ for simulation $A02$. The darker regions correspond to larger power density, while the lighter regions correspond to smaller power density. (a) Normalized power spectrum. (b) Non-normalized power spectrum. The white dashed line appearing in the bottom panel indicates the (non-dispersive) linear dispersion relation from Eq. , and the white dash-dotted line indicates the BQ dispersion relation from Eq. .[]{data-label="cubo20"}](Figure12b.eps "fig:"){width="48.00000%"}
Figures \[cubo32\] and \[cubo20\] show the power spectrum of the flow height $E_h(k,\omega)$ for simulations $B04$ and $A02$ respectively. The linear dispersion relations for shallow water flows (Eq. \[rel\_disp\_sw\]) and for Boussinesq flows (Eq. \[rel\_disp\_bous\]) are also shown as references, using the parameters from each run. Note both runs present an energy accumulation near the dispersion relation. This indicates most of the energy is in the waves, and remains there as time evolves. As we are not solving the equations for a potential flow, and the system can develop vortical motions, this tells us that the non-linear energy transfer is mostly done between waves, and that the energy injected at large scales in wave motions is mostly transferred towards wave motions at smaller scales and faster frequencies. This is needed for weak turbulence to hold, but is also observed in run $B04$ that has a spectrum compatible with strong turbulence phenomenological arguments. There is also a turbulent broadening of the dispersion relation, also visible in cross sections of the spectrum at different wavenumbers in Fig. \[blnc\]. From this broadening, the characteristic time of non-linear wave interactions can be obtained, as was done in [@miquel_nonlinear_2011].
![Cross sections of $E_h(k,\omega)$ at different (and fixed) values of $k=k^*$ for run $A02$. Note the peaks and surrounding wavenumbers have most of the power. Note also the two peaks for $k=200$, one corresponding to the shallow-water dispersion relation, and the other to the Boussinesq dispersion relation. []{data-label="blnc"}](Figure13.eps){width="48.00000%"}
Some of the most important results in this paper are associated with these two figures. First, note that in run $B04$ most of the energy is concentrated near a dispersion relation that, as dispersion is negligible, corresponds in practice to the non-dispersive shallow-water case (Eq. \[rel\_disp\_sw\]). All runs in set $B$ have the same spectral behavior in $E_h(k,\omega)$, and confirm that the $\sim k^{-2}$ spectrum is observed when dispersion is negligible or absent (i.e., when the flow is sufficiently shallow). Second, note that the spectrum $E_h(k,\omega)$ in run $A02$ presents clear signs of dispersive effects (i.e., most of the energy for large enough $k$ is concentrated over a curve that deviates from a linear relation between $k$ and $\omega$), and this run displays a scaling in $E_h(k)$ compatible with the weak turbulence prediction $\sim
k^{-4/3}$. This behavior was observed in the other runs in set $A$.
However, $E_h(k,\omega)$ for runs in set $A$ presents yet another interesting feature. As expected, for small $k$ the dispersion is negligible and the energy is concentrated over a straight line in $(k,\omega)$ space. At large $k$, as already mentioned, the effective dispersion relation is compatible with that of the linearized Boussinesq equations. But at intermediate wavenumbers two branches of the dispersion relation can be observed, one that is compatible with non-dispersive waves and another compatible with dispersive waves. When both branches are present, their amplitudes are of the same order, as can be seen in Fig. \[blnc\].
At first sight, the existence of these two branches could be attributed to bound waves. Bound waves are small amplitude waves which are [*bounded*]{} to a parent wave of larger amplitude. The waves are bounded in the sense that they follow the parent wave, i.e., they travel with the same phase velocity as the parent, and thus they follow an anomalous dispersion relation (see, e.g., a discussion of bound waves in the context of gravito-capillary waves in [@longuet-higgins_generation_1963; @herbert_observation_2010]). The condition that they have the same phase velocity as the parent wave implies that they must follow a modified dispersion relation which verifies $\Omega (k) = \omega (k_0) k/k_0$, where $k_0$ is the wavenumber of the parent wave. Bound waves result in multiple branches in the $E(k,\omega)$ spectrum (and in multiple peaks in the frequency spectrum). Indeed, it is easy to show that for $k=Nk_0$, $N=2,3,4,\dots$, these multiple branches satisfy $$\Omega_N (Nk_0) = N\omega (k_0)
\label{bound_condition}$$ (see, e.g., the discussion in [@herbert_observation_2010]). Extending the analysis in [@herbert_observation_2010] to our case, bound waves in the BQ model should satisfy the following dispersion relation, $$\label{boundwaves}
\Omega_N (Nk) = \frac{c_0 k}{\sqrt{1 + \frac{h^2_0 k^2}{3 N^2}}} ,$$ which verifies Eq. . However, the second branch in Fig. \[cubo20\] cannot be described by this dispersion relation for any value of $N$ up to 4, and thus they are not bound waves in the sense often used in oceanography.
Another explanation for the existence of these two branches can be given by keeping in mind that at intermediate wavenumbers slight variations in the fluid depth may trigger a transition in the waves from dispersive to non-dispersive (as the level of dispersion depends on the product of the wavenumber with the surface height). Indeed, in the turbulent flow there are waves with short wavelengths which ride over long ones, that have a larger amplitude. For sufficient scale separation, the fast waves see an effective depth that can be larger or smaller than $h_0$ depending on whether the wave is on a crest or a valley of the slow wave, generating in one case dispersive waves, and in the other non-dispersive waves.
We can estimate the variation in the effective dispersion at a given wavenumber $k$. In simulation $A02$, $h_0=4\times 10^{-3}$ and the longer waves have an amplitude $\delta \approx 4\times 10^{-5}$ (as can be estimated, e.g., from the maximum value of the power spectrum of $h$). From the system dispersion relation, $$\omega^2 = c^2_0 k^2 \left(1 - \frac{1}{3} h_0^2 k^2 \right) ,$$ dispersion is controlled by the amplitude of the $h_0^2 k^2/3$ term. Assuming that fast waves experience an effective depth $h_0 \pm \delta$ (where the sign depends on whether they are on a valley or a crest), the variation in the dispersion is proportional to the difference between $(h_0-\delta)^2$ and $(h_0+\delta)^2$. So, for this simulation, the variation is around $4\%$, and when multiplied by $k^2$, it is sufficient to explain the two branches in $E_h(k,\omega)$ for $k$ between $\approx 150$ and $250$.
Time frequency energy spectra
-----------------------------
From the spectra in Figs. \[cubo32\] and \[cubo20\] the frequency spectrum $E_h(\omega)$ can be easily obtained, simply by summing over all wavenumbers, $$E_h(\omega) = \sum_k E_h(k,\omega) .
\label{freqspectra}$$
As already mentioned, in experiments and simulations $E_h(\omega)$ is sometimes estimated instead from $E_h(k)$ by using the dispersion relation in the form $k = k(\omega)$. Figure \[edomega\] shows the power spectrum of $h$ as a function of $\omega$ for simulations $A02$ and $B04$. In both cases, the spectrum was calculated explicitly using Eq. , and also estimated using the dispersion relation. For each run, the two spectra show a very good agreement, which can be expected as most of the energy is in the waves. The behavior of the inertial range in each run is also in good agreement with the one found previously for $E_h(k)$ in Sec. \[spectra\].
![[*(Color online)*]{} Power spectrum of $h$ as a function of the frequency for simulations (a) $A02$ and (b) $B04$. In both cases, the spectrum was calculated by summing over all wavenumbers in the time and space resolved spectrum, $\textstyle{\sum_k} E_h
(k,\omega)$, and also by using the dispersion relation given by Eq. to estimate the frequency spectrum from the wavenumber spectrum $E_h(k)$. As a reference, power laws $\sim \omega^{-4/3}$ and $\sim \omega^{-2}$ are shown in each case. The behavior is in good agreement with the one found for $E_h(k)$.[]{data-label="edomega"}](Figure14a.eps "fig:"){width="48.00000%"} ![[*(Color online)*]{} Power spectrum of $h$ as a function of the frequency for simulations (a) $A02$ and (b) $B04$. In both cases, the spectrum was calculated by summing over all wavenumbers in the time and space resolved spectrum, $\textstyle{\sum_k} E_h
(k,\omega)$, and also by using the dispersion relation given by Eq. to estimate the frequency spectrum from the wavenumber spectrum $E_h(k)$. As a reference, power laws $\sim \omega^{-4/3}$ and $\sim \omega^{-2}$ are shown in each case. The behavior is in good agreement with the one found for $E_h(k)$.[]{data-label="edomega"}](Figure14b.eps "fig:"){width="48.00000%"}
Probability density functions
-----------------------------
We calculated the probability density function (PDF) of the free surface height for different simulations. Figure \[distro\] shows the PDF of $h/\sigma$ for run $A06$, where $\sigma$ is the standard deviation of the surface height. The probability distribution is asymmetric, with a larger probability of measuring large values of $h$ than of small values. The shape can be adjusted by two distributions: We consider a skewed normal distribution [@azzalini_class_1985], $$\label{skew_normal}
f(x) = \frac{2}{\kappa}\phi\left(\frac{x-\xi}{\kappa}\right)
\Phi\left(\alpha\frac{x-\xi}{\kappa}\right),$$ where $\kappa$ is the so-called scale parameter (associated with the variance of the distribution), $\xi$ is the location parameter (associated with the mean value), $\alpha$ is the shape parameter (associated with the skewness), and $$\begin{aligned}
\phi(x) &=\frac{1}{\sqrt{2\pi}}e^{-x^2/2}, \\
\Phi(x) &= \int_{-\infty}^{x} \phi(t)\ {\mathrm{d}}t = \frac{1}{2} \left[ 1 +
\operatorname{erf} \left(\frac{x}{\sqrt{2}}\right)\right]. \end{aligned}$$ We also consider a Tayfun distribution $$p(x) = \int^\infty_0 \frac{e^{-[y^2 + (1-c)^2]/(2 s^2)}}{\pi s c} {\mathrm{d}}y,$$ with $c= \sqrt{1 + 2 s x + y^2}$ and where $s$ is the mean steepness of the waves[@tayfun_narrow-band_1980].
![[*(Color online)*]{} Probability density function of the values of $h$ in simulation $A06$ (solid blue line). The dash-dotted (red) line indicates a maximum likelihood fit using a skewed normal distribution, while the dashed (green) line corresponds to a maximum likelihood fit for the Tayfun distribution.[]{data-label="distro"}](Figure15.eps){width="48.00000%"}
For run $A06$, and from a Maximum Likelihood Estimation method for the skewed normal distribution, the location parameter is $\xi \approx -1.00$, the scale parameter is $k \approx 1.43$, and the shape parameter is $\alpha \approx
1.94$. For the same run, and for the Tayfun distribution, the mean steepness of the waves is $s \approx 0.15$. This latter value is more relevant as the Tayfun distribution is often used in oceanography and in experiments of surface waves. In this context, it is interesting to point out that experiments in [@falcon_observation_2011] found similar values for $s$.
This behavior (a PDF of $h$ described correctly by both a skewed normal distribution and a Tayfun distribution with asymmetry to the left) was observed in all simulations, no matter what set they belonged to.
Conclusions {#conclusion}
===========
We studied wave turbulence in shallow water flows in numerical simulations using the shallow water and Boussinesq models. The equations were solved using grids up to $2048^2$ points, and the parameters were varied to study different regimes, including regimes with larger and smaller Reynolds number, and larger and smaller dispersion, while keeping the Froude number approximately the same. We summarize below the main conclusions following the same ordering as in the introduction:
\(a) As in previous experimental and observational studies [@smith_equilibrium_2003; @kaihatu_asymptotic_2007], we found that the flows can be classified in different sets depending on the value of the Reynolds number (i.e., on the strength of the nonlinearities) and on the level of dispersion (associated with the fluid depth). A first set ($A$) has smaller Reynolds numbers and stronger dispersion, a second set ($B$) has larger Reynolds numbers and weaker or negligible dispersion, and a third set of runs seems to be transitional between the two.
\(b) Runs in sets $A$ and $B$ have different power spectra of the surface height. Runs in set $A$, with stronger dispersion, present a spectrum compatible (within statistical uncertainties) with $E_h(k) \sim k^{-4/3}$. This is the spectrum predicted by weak turbulence theory for the Boussinesq equations [@onorato_four-wave_2008]. Runs in set $B$ with negligible or zero dispersion (i.e., for a shallower flow) show a spectrum compatible within error bars with $E_h(k) \sim k^{-2}$. This spectrum can be obtained from phenomenological arguments coming from strong turbulence [@phillips_equilibrium_1958]. The runs in set $C$ have no discernible inertial range.
\(c) The Boussinesq (dispersive) model tends to develop more power in waves with short wavelengths than the shallow water model. This is associated with the development of a bottleneck for large wavenumbers in the energy spectrum.
\(d) Inspection of the wave and frequency spectrum $E_h(k,\omega)$ confirms that most of the energy is in the waves in all the simulations. In runs in set $B$, most of the energy is concentrated in the vicinity the linear dispersion relation for shallow water waves, which are non-dispersive. In runs in set $A$, the resulting non-linear dispersion relation obtained from $E_h(k,\omega)$ has two branches: one that corresponds to non-dispersive waves, and another corresponding to dispersive waves. The two branches can be explained as the result of the superposition of rapidly varying waves which ride over slowly varying waves, the latter with sufficient amplitude to change whether the former see a shallower or deeper fluid.
\(e) Independently of the differences between the runs, the probability distribution functions of $h$ for the runs in all sets is asymmetric, with larger probabilities of finding larger values of $h$ than smaller values. The probability distribution functions can be approximated by both a skewed normal distribution and a Tayfun distribution [@tayfun_narrow-band_1980]. In the latter case, the only parameter of the distribution, the mean steepness of the waves, has values compatible with those found in observations and experimental studies (see [@falcon_observation_2011]). The obtained probability density functions also indicate limitations in the hypothesis of Gaussianity of the fields assumed in early theories of weak turbulence. However, extensions of the theory to allow for non-Gaussian distributions exist and can be found for example in [@choi_probability_2004] and [@lvov_noisy_2004; @choi_joint_2005].
All the results presented here were obtained solving numerically equations that do not assume that the flow is inviscid or irrotational, and with realistic terms for the viscous dissipation. We believe this approach can be useful to compare with experiments, as in experiments vorticity can develop in the flow, and viscosity cannot be neglected.
The authors would like to thank Prof. Oliver Buhler and the anonymous referees for their useful comments. The authors acknowledge support from grants No. PIP 11220090100825, UBACYT 20020110200359, and PICT 2011-1529 and 2011-1626. PDM and PJC acknowledge support from the Carrera del Investigador Científico of CONICET, and PCdL acknowledges support from CONICET.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Building upon the well-posedness results in [@snse1], in this note we prove the existence of invariant measures for the stochastic Navier-Stokes equations with stable Lévy noise. The crux of our proof relies on the assumption of finite dimensional Lévy noise.'
author:
- Leanne Dong
bibliography:
- 'l''sthesis.bib'
date: ' School of Mathematics and Statistics, The University of Sydney. '
title: 'Invariant measures for the Stochastic Navier-Stokes equation on a 2D rotating sphere with stable Lévy noise'
---
Invariant measures {#sec:im}
==================
In this note we are concerned with the existence of an invariant measure of the solution $u$ to the abstract equation (\[asnse4\]). Let $A$ be the Stokes operator in $H$. Let $B$ is the bilinear operator. Let $C$ be the Coriolis operator in $H$ and $G$ is a bounded linear operator. (See [@snse1] for rigorous definitions of these operators) $$\begin{aligned}
\label{asnse4}
du(t)+Au(t)+B(u(t),u(t))+\mathbf{C}u=fdt+GdL(t),\quad u(0)=u_0,\end{aligned}$$
In our earlier work [@snse1] we proved
- Theorem \[t1u\] on the existence and uniqueness of a weak solution.
- Theorem \[t2u\] on the continuous dependence on initial data.
- Theorem \[t4\] on the existence of a strong solution.
\[t1u\] Suppose that $\alpha\ge 0$, $z\in L^4_{\text{loc}}([0,\infty);\mathbb{L}^4(\mathbb{S}^2)\cap H)$, $ v_0\in H$ and $ f\in V'$. Then there exists ${\mathbb{P}}$-a.s. a unique solution $ u\in D([0,\infty);H)\cap L^2_{\text{loc}}([0,\infty);V)$ of equation (\[asnse4\]). In particular, if $$\sum_{l=1}^\infty|\sigma_l|^{\beta}\lambda^{\beta/2}_l<\infty\,,$$ then the theorem holds.
\[t2u\] Assume that, $$u^0_n\to u\quad\text{in}\quad H$$ and for some $T>0$, $$\begin{aligned}
z_n\to z\quad\text{in}\quad L^4([0,T];\mathbb{L}^4(\mathbb{S}^2)\cap H)\qquad f_n\to f\quad\text{in}\quad L^2(0,T;V').\end{aligned}$$
\[t4\] Assume that $\alpha\ge 0$, $z\in L^4_{\text{loc}}([0,\infty);\mathbb{L}^4(\mathbb{S}^2)\cap H)$, $ f\in H$ and $ v_0\in H.$ Then, there exists ${\mathbb{P}}$-a.s. unique solution of (\[asnse4\]) in the space $D(0,T;H)\cap L^2(0,T;V).$ which belongs to $D(\epsilon,T; V)\cap L^2_{\text{loc}}(\epsilon,T;D(A))$ for all $\epsilon>0.$ and $T>0.$ Moreover, if $ v_0\in V$, then $ u\in D(0,T; V)\cap L^2_{\text{loc}}(0,T;D(A))$ for all $T>0$, $\omega\in\Omega$. Moreover, if $$\sum_{l=1}^\infty|\sigma_l|^{\beta}\lambda^{\beta/2}_l<\infty\,,$$ then the theorem holds.
It is well known (see for instance Chapter 9 of [@MR2356959]) that strong solution implies a weak solution, and the weak solution is equivalent to a mild solution. Hence the three concepts of solutions are equivalent. With the aid to these results, our main aim in this section is to study the large time behaviour of $u$, that is, the law $\mathcal{L}(u(t,x))$ as $t\to\infty.$ In particular, we prove (\[asnse4\]) admits at least one invariant measure Here we consider a general cadlag Markov process, $$\begin{aligned}
(\Omega,\{\mathcal{F}^0_t\}_{t\ge 0},\mathcal{F},\{u^x_t\}_{t\ge 0},({\mathbb{P}}_x)_{x\in H})\end{aligned}$$ whose transition probability is denoted by $\{P(t,x,dy\}_{t\ge 0}$, where $\Omega :=D([0,\infty);H)$ is the space of the càdlàg function from $[0,\infty)$ to $H$ equipped with the Skorokhod topology, $\mathcal{F}^0_t=\sigma\{u_s,0\le s\le t\}$ is the natural filtration. Now denote (resp.) by $C_b(H)$, $B_b(H)$ the space of bounded continuous functions and the space of bounded borel measurable functions on $H.$ That is, $$\begin{aligned}
C_b(H) &:=\{\varphi:H\to{\mathbb{R}}: \varphi\,\,\text{is continuous and bounded}\},\\
B_b(H) &:=\{\varphi:H\to{\mathbb{R}}: \varphi\,\,\text{is bounded and borel measurable}\}.\end{aligned}$$ For all $\varphi\in B_b(H)$, define$$\begin{aligned}
P_t \varphi(x)=\int_H \varphi(y)P(t,x,dy),\quad\forall\,\,t\ge 0,x\in H.\end{aligned}$$ For any $t\ge 0$, $P_t$ is said to be Feller if $$\begin{aligned}
\label{feller}
\varphi\in C_b(H)\to P_t\varphi\in C_b(H),\quad\forall\quad t\ge 0.\end{aligned}$$ $P_t$ is said to be strong Feller if (\[feller\]) holds for a larger class of function: $\varphi\in B_b(H).$ Moreover, $P_t$ is said to be *irreducible* in $H$, if $P_t 1_{A}(x)=P_x(t,A)>0$ for any $x\in H$ and any non-empty open subset $A$ of $H$. If $P_t$ is irreducible then any invariant measure $\mu$ is full, that is, one has $\mu(B(x,r))>0$ for any ball $B(x,r)$ of center $x\in H$ and radius $r$. Indeed, it follows from the definition of invariant measure that $$\begin{aligned}
\mu(B(x,r))&=\int_H P_t 1_{A}(x)\mu(dx)>0.\end{aligned}$$
The main theorem proved in this section is Theorem \[existenceim\] which states below.
\[existenceim\] Assume additionally, that there exists $m>1$ such that $\sigma_l=0$ for all $l\ge m$. Then the solution $u$ to (\[asnse4\]) admits at least one invariant measure.
We claim that the SNSE (\[asnse4\]) has an invariant measure. The key to proving this is to use the Krylov-Bogolyubov Theorem, which guarantees the existence of invariant measures for certain well-defined maps defined on some well-defined space. More precisely, the theorem states that,
Assume $(P_t,t\ge 0)$ is a **Feller** semigroup. If there exists a point $x\in H$ for which the family of probability measure $\{\mu_t(x,\cdot)\}_{t\ge 0}$ is uniformly **tight**, that is, there exists a compact set $K_{\epsilon}\subset H$ such that $\mu(K_{\epsilon})\ge 1-\epsilon$ for any $\mu\in\Lambda$ on $(H,\mathcal{B}(H))$ then there exists at least one invariant measure.
\[kbcor\] If for some $\nu\in\mathcal{P}$ and some sequence $T_n\uparrow +\infty$ the sequence $\{P^*_{T_n}\nu\}$ is tight, then there exists an invariant measure for $(P_t,t\ge 0).$
We shall remark that there are various versions of Krylov-Bogolyubov theorem which conveys the same idea. All that required to be proved are Feller, Markov property of the solution $v$ (and so $u$) and convergence of the family of probability measures $\{\mu_t, t\ge 0\}$ in $H$. This is comparable to the concept of weak convergence of distribution in finite dimension (equivalence to weak convergence of r.v.). However, in infinite dimension, the convergence of distribution is more involved. Hence extra conditions are needed besides the convergence of finite-dimensional distributions.
Note that, it is known that tightness is a necessary condition to prove convergence of probability measure, especially when measure space is infinite dimensional. In this sense the two statements of the theorem is equivalent.
The following inequalities would be used quite often. $$\begin{aligned}
|\hat{ A}^{\sigma}e^{-\hat{ A}t}|&\le C(\sigma)t^{-\sigma},\quad\forall\,\,\sigma>0,\end{aligned}$$
$$\begin{aligned}
\label{bua}
|B( u)|_V= \langle A^{\frac{1}{2}}B( u), A^{\frac{1}{2}}B( u)\rangle =\langle AB( u),B( u)\rangle&=| A^{\frac{1}{2}}B(u)|\\
&\le | u|| u|^{\frac{1}{2}}_V| A u|^{\frac{1}{2}},\end{aligned}$$
$$\begin{aligned}
|B( u)-B( v)|\le C(| u|^2_V+| u|^2_V| v|+| v|^2_V),\end{aligned}$$
$$\begin{aligned}
|B( u)|\le C| u|^{\frac{1}{2}}_V| u|_V|A u|^{\frac{1}{2}}=C| u|^{\frac{3}{2}}_V|A u|^{\frac{1}{2}}.\end{aligned}$$
Transition Semigroup {#markovfellersoln}
--------------------
Let us denote by $u(\cdot,x)$ the solution of (\[asnse4\]). We set $$\begin{aligned}
P_t f(x)={\mathbb{E}}f(u(t,x)),\quad f\in B_b(H), t\ge 0, x\in H.\end{aligned}$$ It follows from uniqueness and time homogeneity of $L$ that the following relation holds, $$\begin{aligned}
P_t\circ P_s=P_{t+s}.\end{aligned}$$ Recall from our Theorem \[t4\] that, we have proved there exists a unique strong solution to (\[asnse4\]) with mild form in the space $D([0,T;H)\cap L^2([0,T];V)$ which belongs to $D([h,T];V)\cap L^2_{\text{loc}}([h,T];D(A))$ for all $h>0$, $T>0$ for every initial condition $ u_0\in H$, $\omega\in\Omega.$ Moreover, if $ u_0\in V$, then $ u\in D([h,T;V])\cap L^2_{\text{loc}}([h,T];D(A)) $ for all $h>0$ and $T>0.$ The solution depends continuously on initial data $ x.$
Let $u(t;x)$ be the solution at $t$ starting from $x$ at time 0. Now suppose we have two solutions, resp. $ u_n$ and $u$ of (\[asnse4\]) started at $ \xi_n$ and $ \xi$, if the conditions in Theorem \[t2u\] satisfied, then it follows that $ u_n(t)\to u(t)$ a.s. for any $t.$ Therefore, $f( u_n(t))\to f( u(t))$ as $f$ is continuous. Thus, invoke Lesbesgue Dominated Convergence Theorem, one has $$\begin{aligned}
{\mathbb{E}}f( u_n(t))\to{\mathbb{E}}f( u(t)).\end{aligned}$$ Whence the equation (\[asnse4\]) defines a Feller Markov process. Then we can define the operator $P_t: C_b(H)\to C_b(H)$ by $$\begin{aligned}
P_t f={\mathbb{E}}f(u(t;x)),\end{aligned}$$ and $P_t$ is said to be a Feller semigroup.
The equation (\[asnse4\]) defines a Markov process in the sense that $$\begin{aligned}
\label{markovian}
{\mathbb{E}}[ f ( u^{ x}_{t+s})|\mathcal{F}_t]=P_s( f )(u^{ x}_t),
\end{aligned}$$ for all $t,s>0$, $ f \in C_b(H)$, where $ u^{ x}_t$ denotes[^1] the solution to (\[asnse4\]) over $[0,\infty]$ starting from the point $ u(0)=x$, $\mathcal{F}_t$ denotes the sigma-algebra generated by $L(\tau)$ for $\tau\le t.$
By uniqueness, $$\begin{aligned}
u^{ x}_{t+s}= u^{ u^{ x}_t}_{t,t+s},\quad ({\mathbb{P}}-a.s.),\end{aligned}$$ where $( u^{\eta}_{t_0,t})_{t\ge t_0}$ denotes the **unique** solution on the time interval $[t_0,\infty)$, with the $\mathcal{F}_{t_0}$-measurable intial condition $ u^{\eta}_{t_0,t_0}=\eta$. To prove (\[markovian\]), it suffices to prove $$\begin{aligned}
{\mathbb{E}}[ f ( u^{\eta}_{t,t+s})|\mathcal{F}_t]=P_s( f )(\eta),\end{aligned}$$ for every $H$-valued $\mathcal{F}_t$-measurable r.v. $\eta.$ Note that (\[markovian\]) holds for all $ f \in C_b(H)$, holds for $ f =1_{\Gamma}$ where $\Gamma$ is an arbitrary Borel set of $H$ and consequently for all $f\in B_b(H).$ Without loss of generality, Let us assume $f\in C_b(H)$. We know that, if $\eta=\eta_i$ ${\mathbb{P}}$ a.s., then the r.v. $u(t+s,t,\eta_i)$ is independent to $\mathcal{F}_t$ and therefore $$\begin{aligned}
{\mathbb{E}}(f( u(t+s,t,\eta_i))|\mathcal{F}_t)={\mathbb{E}}f( u(t+s,t,\eta_i))=P_{t,t+s} f(\eta_i)=P_s f(\eta_i),\quad{\mathbb{P}}\,\,\text{a.s.}\quad.\end{aligned}$$ It suffices to prove (\[markovian\]) holds for every r.v. $\eta$ of the form $$\begin{aligned}
\eta=\sum^N_{i=1}\eta^{(i)}1_{\Gamma^{(i)}},\end{aligned}$$ where $\eta^{(i)}\in H$ and $\Gamma^{(i)}\subset\mathcal{F}_t$ is a partition of $\Omega$, $\eta_i$ are elements of $H.$ Then $$\begin{aligned}
u(t+s,t,\eta_i)=\sum^N_{i=1} u(t+s,t,\eta_i)1_{\Gamma_i},\quad{\mathbb{P}}\,\,\text{a.s.}\quad.\end{aligned}$$ Hence, $$\begin{aligned}
{\mathbb{E}}(f( u(t+s,t,\eta))|\mathcal{F}_{t})=\sum^N_{i=1}{\mathbb{E}}(f( u(t+s,t,\eta_i))1_{\Gamma_i}|\mathcal{F}_t)\quad{\mathbb{P}}-\text{a.s.}\quad.\end{aligned}$$ Take into account the r.v. $ u(t+s,t,\eta_i)$ independent to $\mathcal{F}_t$ and $1_{\Gamma_i}$ are $\mathcal{F}_t$ measurable, $i=1,\cdots, k$, one deduces that $$\begin{aligned}
{\mathbb{E}}[f( u(t+s,t,\eta))|\mathcal{F}_t]&=\sum^N_{i=1} P_s f(\eta_i)1_{\Gamma_i}=P_s f(\eta),\quad{\mathbb{P}}-\text{a.s.}\quad,\end{aligned}$$ and so (\[asnse4\]) defines a Markov process in the above sense for all $f\in C_b(H)$. Now, let $u(t;\eta)$ be the solution of the SNSE (\[asnse4\]) with initial condiction $\eta\in H$.
Let $(P_t, t\ge 0)$ be the Markov Feller semigroup on $C_b(H)$ associated to the SNSE (\[asnse4\]) defined as $$\begin{aligned}
\label{markovsemi}
P_t f(\eta)={\mathbb{E}}[f(u(t;\eta))]=\int_H f(y)P(t,y)dy=\int_H f(y)\mu_{t,s}(dy),\quad f\in C_b(H),\end{aligned}$$ where $P(t,x,dy)$ is the transition probability of $u(t;\eta)$ and $\mu_{t,x}(dy)$ is the law of $u(t;\eta).$ From (\[markovsemi\]), we have $$\begin{aligned}
P_t f(x)=(f,\mu_{t,x})=(P_t f,\mu),\end{aligned}$$ where $\mu$ is the law of the initial data $\eta\in H$. Thus it follows from above that $\mu_{t,\eta}=P^*_t\mu$. If $$\begin{aligned}
P^*_t\mu=\mu\quad\forall\quad t\ge 0,\end{aligned}$$ then a probability measure $\mu$ on $H$ is said to be an invariant measure.
The proof of tightness
-----------------------
We proceed the claim of tightness by first proving the following a priori estimate. The main difficulty is overcome by introducing a simplified auxiliary Ornstein Uhblenbeck process, which enables us to use the classical arguments in the spirit of p.51-150 [@SFS08]. To prove existence of invariant measures for (\[asnse4\]), we write the problem in a slightly different form.
Let $H$, $A: D(A)\subset H\to H$, $V=D(A^{1/2})=D(\widehat{A}^{1/2})$ and $B:V\times V\to V', \,\,C$ be spaces and operators introduced in the previous section. Suppose that there exists a constant $c_B>0$ such that $$\begin{aligned}
\label{buvw0}
\langle B(u,v),w\rangle=|b(u,v,w)|&\le c_B|u|^{1/2}|u|^{1/2}_V|v|^{1/2}|v|^{1/2}_V|w|_V,\quad\forall\quad u,v,z\in V,\end{aligned}$$
$$\begin{aligned}
\langle B(u,v),v\rangle\le c_B|u|^{1/2}|Au|^{1/2}|v|_V|z|,\end{aligned}$$
for all $u\in D(A)$, $v\in V$ and $z\in H.$
In order to prove there exists at least one invariant measure, we use standard method in the spirit of Chapter 15 in [@MR1417491]. However, the analysis of Navier-Stokes equations with additive noise in our case requires some non-trivial consideration, as pointed out in [@MR1305587],. In particular, a critical question arises when analyzing the estimate $\frac{d^+}{dt}|v(t)|^2$, the usual estimates for the nonlinear term $b(v(t),z(t),v(t))$ yields a term $|v(t)|^2 |z(t)|^4_4,$ so we were not able to deduce any bound in $H$ for $|v(t)|^2$ under classical lines. Nevertheless, in light of the method developed in Crauel and Flandoli [@MR1305587], via the usual change of variable and by writing the noise and the Ornstein-Uhlenbeck process as finite sequence of 1D processes, we are able to prove there exists at least one invariant measure to (\[asnse4im\]).
We remark that this fundamental ODE is different from the one used in the proof of existence and uniqueness. Let $f\in H$ and $m>1$ be given. Consider $$\begin{aligned}
\label{asnse4im} du(t)=[-Au(t)-B(u(t),u(t))+\mathbf{C}u(t)+f]dt+\sum^m_{l=1}\sigma_le_ldL_l(t),\end{aligned}$$ where operators $A$, $B$, $C$ are are as defined, $f\in H$, $L_1,\,L_2\,\cdots L_l$ are i.i.d. ${\mathbb{R}}$-valued symmetric $\beta$-stable process on a common probability space $(\Omega,\mathcal{F},{\mathbb{P}})$, $\sigma$ is a bounded sequence of real numbers and $e_l$ is the complete orthonormal system of eigenfunctions on $H$.
### Auxiliary Ornstein-Uhlenback Process
Let $(\tilde{L}(t),t\ge 0)$ be a Lévy process that is an independent copy of $L$. Denote by $\bar{L}$ a Lévy process on the whole real line by $$\begin{aligned}
\bar{L}(t)
\begin{cases}
L(t),\quad t\ge 0\\
\tilde{L}(-t),\quad t<0,
\end{cases}\end{aligned}$$ and by $\bar{\mathcal{F}}_t$ the filtration $$\begin{aligned}
\bar{\mathcal{F}}_t=\sigma(\bar{L}(s),s<t),\quad t\in{\mathbb{R}}.\end{aligned}$$ Let $\alpha>0$ be given; For each $l=1,\cdots,m$, let $z^0_l$ be the stationary (ergodic) solution of the one dimensional equation $$\begin{aligned}
dz^0_l+(\lambda_l+\alpha) z^0_ldt=\sigma_ldL_l(t)\end{aligned}$$ so that $$\begin{aligned}
z^0_l(t)=\int^t_{-\infty}e^{-(\lambda_l+\alpha)(t-s)}\sigma_ldL_l(s)\end{aligned}$$ Note that the integral above is well defined, since for any $p\in(1,\beta)$ with $\beta>1$ we have $$\label{eq_fin}
\begin{aligned}
\mathbb E\left|z^0_l(t)\right|^p&=C_{p,\beta}\int_0^\infty e^{-p(\lambda_l+\alpha)(t-s)}\sigma_l^p\,ds\\
&=\dfrac{C_{p,\beta}\sigma_l^p}{p\left(\alpha+\lambda_l\right)}.
\end{aligned}$$ More precisely, let $$z^0_l(t,s)=\int^t_{s}e^{-(\lambda_l+\alpha)(t-r)}\sigma_ldL_l(r)\,.$$ Then one can show directly evaluating integrals in the same way that $$\lim_{s\to-\infty}z_l(t,s)=z_l(t)$$ exists. Putting $z^0(t)=\sum^m_{l=1}z^0_l(t)e_l$ one has $$\begin{aligned}
\label{auxou}
dz^0+(A+\alpha I) z^0\,dt=GdL(t)\,,\end{aligned}$$ where $Ge_l=\sigma_le_l$, or $$z^0(t)=\int_{-\infty}^t e^{-(t-s)(A+\alpha I)}GdL(s)\,.$$
We have for any $s,t$ such that $-\infty<s<t<\infty$ $$z(t)=\int_{-\infty}^te^{-(t-s)\widehat A}GdL(s)=e^{-(t-s)\widehat A}z(s)+\int_s^te^{-(t-r)\widehat A}GdL(r)\,.$$ We need another lemma.
\[dom\_A\] We have $$\sup_{-1\le t\le 0}|Az(t)|^2<\infty\,.$$
Note first that the process $Z^0=Az^0$ is well defined and satisfies all the assumptions of Lemma 3.1 in [@snse1] with the process $L$ replaced by another Lévy process $AL$. Therefore, we have $$\sup_{-1\le t\le 0}\left|Z^0(t)\right|^2<\infty\,.$$ Since $D(A)=D\left(\widehat A\right)$, one can repeat all arguments from the proof of Lemma 3.1. This yields $$\sup_{-1\le t\le 0}\left|Az(t)\right|^2<\infty\,.$$
Now, using the lemma above and Lemma 3.1 applied with $\delta=\frac12$ we find that the process $z$ is càdàg in $V$ and $$\begin{aligned}
\label{zzvaz}
\sup_{-1\le t\le 0}\left(|z(t)|^2+|z(t)|^2_V+|Az(t)|^2\right)<\infty\quad{\mathbb{P}}\,\,\text{a.s.}\quad .\end{aligned}$$ Using equation (4.12) in [@MR2773026], one can now choose $\alpha>0$ such that $$\begin{aligned}
\label{z10}
4\eta m{\mathbb{E}}|z_1(0)|\le \frac{\lambda_1}{4},\end{aligned}$$ where $\lambda_1$ is the first eigenvalue of $A$, since ${\mathbb{E}}|z_1(0)|^p\to 0$ as $\alpha\to\infty.$
From (\[z10\]) and the Ergodic Theorem we obtain $$\begin{aligned}
\lim_{t_0\to-\infty}\frac{1}{-1-t_0}\int^{-1}_{t_0}4\eta\sum^m_{l=1}|z_l(s)|ds=4\eta m{\mathbb{E}}|z_1(0)|<\frac{\lambda_1}{4}.\end{aligned}$$ Put $\gamma(t)=-\frac{\lambda_1}{2}+4\eta\sum^m_{l=1}|z_l(t)|$, we get $$\begin{aligned}
\label{gp1}
\lim_{t_0\to-\infty}\frac{1}{-1-t_0}\int^{-1}_{t_0}\gamma(s)ds<-\frac{\lambda_1}{4}.\end{aligned}$$ From this fact and by stationarity of $z_l$ we finally obtain $$\begin{aligned}
\label{gp2} \lim_{t_0\to-\infty}e^{\int^{-1}_{t_0}\gamma(s)ds}=0\quad{\mathbb{P}}-\text{a.s.}\quad,\end{aligned}$$
$$\begin{aligned}
\label{gp3} \sup_{t_0<-1}e^{\int^{-1}_{t_0}\gamma(s)ds}|z(t_0)|^2<\infty,\quad{\mathbb{P}}-\text{a.s.}\quad .\end{aligned}$$
$$\begin{aligned}
\label{gp4}
\int^{-1}_{-\infty}e^{\int^{-1}_{\tau}\gamma(s)ds}(1+|z_l(\tau)|^2+|z_l(\tau)|^2_V+|z_l(\tau)|^2|z_l(\tau)|)d\tau<\infty,\quad{\mathbb{P}}-\text{a.s.}\quad. \end{aligned}$$
for all $1\le j$, $l\le m.$ Indeed, note for instance that for $t<0$, $$\begin{aligned}
\frac{z_l(t)}{t}=\frac{z_l(0)}{t}-\frac{1}{t}(\alpha+A_l)\int^0_t z_l(s)ds+\frac{L_l(t)}{t},\end{aligned}$$ whence $\lim_{t\to-\infty}\frac{z_l(t)}{t}=0$ ${\mathbb{P}}$-a.s., which implies (\[gp2\]) and (\[gp3\]). Consider the abstract SNSE $$\begin{aligned}
du+[Au+B(u)+\mathbf{C}u]dt=fdt+GdL(t)\end{aligned}$$ and the Ornstein-Uhlenback equation $$\begin{aligned}
dz+(\hat{A}+\alpha I) z dt=GdL(t),\end{aligned}$$ where $L(t)=\sum^{m}_{l=1}e_l L_l(t)$. We now use the change of variable $v(t)=u(t)-z(t).$ Then, by subtracting the Ornstein-Uhlenback equation from the abstract SNSE, we find that $v$ satisfies the equation $$\begin{aligned}
\label{odeim}
\frac{d^+v}{dt}=-\nu Av(t)-\mathbf{C}v(t)-B(u,u)+f+\alpha z.\end{aligned}$$
Recall the Poincare inequalities $$\begin{aligned}
\label{poincare}
|u|^2_V&\ge\lambda_1|u|^2,\quad\forall\quad u\in V,\\
|Au|^2&\ge\lambda_1|u|^2,\quad\forall\quad u\in D(A).\end{aligned}$$ Let us note that there exists $\eta>0$ such that $$\begin{aligned}
\label{bsum}
|\langle B(u,e_l),u\rangle|\le\eta|u|^2,\quad u\in V, l=1,\cdots,m.\end{aligned}$$ Then the following holds.
\[est1\] Let $\alpha>0$, $ v$ is a mild solution of (\[odeim\]), there exist cosntants $c, c'>0$ depending only on $\lambda_1$ such that $$\begin{aligned}
\label{ineq1}
\frac{1}{2}\frac{d^+}{dt}|v|^2+\frac{1}{2}|v|^2_V\le \left(-\frac{\lambda_1}{4}+2\eta\sum^m_{l=1}|z_l|\right)|v|^2+c|f|^2+c\alpha|z|^2+2\eta|z|^2\sum^m_{l=1}|z_l|.\end{aligned}$$
Let $\alpha> 0$ be given. Denote for simplicity by $z(t)$ the stationary Orstein Uhbleck process, corresponding to $\alpha$, introduced in earlier. Using the classical change of variable $v(t)=u(t)-z(t)$ , the well known identity $\frac{1}{2}\partial_t|v(t)|^2=(v(t),v(t))$, and the antisymmetric term $(Cv,v) =0$ we have $$\begin{aligned}
\label{deq} \frac{1}{2}\frac{d^+}{dt}|v|^2&=-\nu(Av,v)-\langle B(u,z),u\rangle+(\alpha z,v)+\langle f,v\rangle\\
&\le -\nu |v|^2_V-\langle B(u,z),u\rangle+\alpha |z||v|+|f||v|.\end{aligned}$$ By the definition of $z$ and assumptions (\[bsum\]), $$\begin{aligned}
\langle B(u,z),u\rangle&=\sum^m_{l=1}\langle B(u,e_l),u\rangle z_l\le\eta|u|^2\sum^m_{l=1}|z_l|\\
&\le 2\eta|v|^2\sum^m_{l=1}|z_l|+2\eta|z|^2\sum^m_{l=1}|z_l|,\end{aligned}$$ and the inequalities $$\begin{aligned}
\langle \alpha z, v\rangle = c\alpha |z|^2+c'|v|^2,\end{aligned}$$
$$\begin{aligned}
\langle f, v\rangle \le c |f|^2+c'|v|^2.\end{aligned}$$
For simplicity we take $\nu=1$. Then via Young inequality, one can show that there exists $c, c'>0$ depending only on $\lambda_1$ such that $$\begin{aligned}
\frac{1}{2}\frac{d^+}{dt}|v|^2+\frac{1}{2}|v|^2_V\le -\frac{1}{2}|v|^2+2\eta|v|^2\sum^m_{l=1}|z_l|+2\eta |z|^2\sum^m_{l=1}|z_l|+c|f|^2+2c'|v|^2+c\alpha |z|^2+2c|z|^2_V+c'|v|^2_V.\end{aligned}$$ So $$\begin{aligned}
\frac{1}{2}\frac{d^+}{dt}|v|^2+\frac{1}{2}|v|^2_V\le (-\frac{\lambda_1}{4}+2\eta\sum^m_{l=1}|z_l|+2c')|v|^2+c|f|^2+c\alpha|z|^2+2\eta|z|^2\sum^m_{l=1}|z_l|.\end{aligned}$$ Hence one can find a constant $c, c'>0$ depending only on $\lambda_1$ for which the claim follows. Moreover, Let $\gamma(t)$, and $p(t)$ are defined as : $$\begin{aligned}
p(t)=c|f|^2+c\alpha|z|^2+\eta|z|^2\sum^m_{l=1}|z_l(t)|,\end{aligned}$$ $$\begin{aligned}
\gamma(t)=-\frac{\lambda_1}{2}+4\eta\sum^m_{l=1}|z_l(s)|,\end{aligned}$$ we have $$\begin{aligned}
\label{dineq} \frac{1}{2}\frac{d^+}{dt}|v|^2+\frac{1}{2}|v|^2_V\le \frac{1}{2}\gamma(t)|v|^2+p(t).
\end{aligned}$$
Temporarily disregard the $|v(t)|_V$ term, we have $$\begin{aligned}
\frac{d^+}{dt}|v(t)|^2\le\gamma(t)|v(t)|^2+2p(t)\end{aligned}$$ which implies $$\begin{aligned}
\label{ineq}
|v(t)|^2\le |v(\tau)|^2 e^{\int^t_{\tau}\gamma(s)ds}+\int^t_{t_0}e^{\int^t_s\gamma(\xi)d\xi}2p(s)ds.\end{aligned}$$ Now drop out the first term in (\[dineq\]), integrate over $[\tau,t]$ we have $$\begin{aligned}
\label{intineq}
\int^t_{\tau}|v(s)|^2_Vds\le (\sup_{\tau\le s\le t}|v(s)|^2)\int^t_{\tau}\gamma(\xi)d\xi +\int^t_{\tau} 2p(s)ds.\end{aligned}$$ Let us recall, that we proved the existence an uniqueness of solutions to the stochastic Naver-Stokes equation under the assumption that $$\label{eq_v1}
\sum_{l=1}^\infty|\sigma_l|^{\beta}\lambda^{\beta/2}_l<\infty\,,$$ and then the process $z(\cdot)$ has a càdlàg version in $V=D(A^{1/2})$.
We will show that, under the above assumption, there exist at least one invariant measure for SNSE. Let $f\in H$, be given. For an arbitrary real number $s$, $u(t,s)$, $t\ge s$, is the unique solution to the SNSE $$\begin{aligned}
\begin{cases}
d u(t)+A u(t)dt+B( u(t), u(t))dt+\mathbf{C} u(t)dt=fdt+dL(t),\\
u(s)=0
\end{cases}\end{aligned}$$
The space $D( A^{\delta})$ is compactly embedded into the space $H.$
Consequently, if one prove that the process $ u(t,0), t\ge 1$ is bounded in probability as a process with values on $D( A^{\delta})$, one gets immediately the law $\mathcal{L}( u(t,0))$, $t\ge 1$ are tight on $H.$ This suffices the claim of existence of an invariant measure. More precisely, one proves in two steps.
**Step 1** Assuming that holds we will prove an a priori bound in $H$. For any $\alpha\ge 0$, denote by $ z$ the stationary solution of $$\begin{aligned}
d z_{\alpha}+(\hat{A}+\alpha I) z_{\alpha} dt=d\bar L(t),\end{aligned}$$ where $$\begin{aligned}
z_{\alpha}(t)= z(t)+e^{-(\hat{A}+\alpha)(t-s)}(z_{\alpha}-z(s))-\alpha\int^t_s e^{-(\hat{A}+\alpha)(t-s)}z(\sigma)d\sigma.\end{aligned}$$ Let $$\begin{aligned}
v_{\alpha}(t,s)= u(t,s)- z_{\alpha}(t),\quad t\ge s.\end{aligned}$$ Then $ v_{\alpha}(t)= v_{\alpha}(t,s)$, $t\ge s$ is the mild solution to $$\begin{aligned}
\partial_t v_{\alpha}(t)+\nu A v(t)dt+\mathbf{C}(v_{\alpha}(t)+z_{\alpha}(t))&=-B( v_{\alpha}(t)+ z_{\alpha}(t))+f+\alpha z_{\alpha}(t),\quad t\ge s,\\
v(s)=- z_{\alpha}(s)\end{aligned}$$ Following step 1 one has the following proposition
\[vaest\] There exists $\alpha>0$ and a random variable $ \xi$ such that ${\mathbb{P}}$-a.s. $$\begin{aligned}
| v_{\alpha}(t,s)|&\le \xi\quad\forall\,t\in [-1,0]\quad\text{and all}\quad s\le -1,\label{vats}\\
\int^0_{-1}| v_{\alpha}(t,s)|^2_{V}ds&< \xi\quad\forall\,\,t\in [-1,0]\,\,\text{and all}\,\,s\le -1.\label{vatsi}\end{aligned}$$
In view of inequality \[ineq\], one obtains $$\begin{aligned}
\label{va}
| v_{\alpha}(t,s)|^2\le |v(s)|^2 e^{\int^{t}_s -\frac{\lambda_1}{2}+4\eta\sum^m_{l=1}|z_l(\xi)|d\xi}+\int^t_s e^{\int^t_r\gamma(\xi)d\xi}2p(r)dr.\end{aligned}$$ Based on the earlier discussion, the first term is finite; The second term is also finite under the assumption (\[zzvaz\]).
We now use the ergodic properties of $z.$ Since $z_{\alpha}(t)$, $-\infty<t<\infty$, is an ergodic process which is supported by $D(A^{\delta})\subset \mathbb{L}^4(\mathbb{S}^2)$. Then by the Marcinkiewicz strong law of large number, we have ${\mathbb{P}}$ a.s. that and by Prop 8.4 [@MR2584982] that $$\begin{aligned}
\lim_{s\to-\infty}\frac{1}{-1-s}\int^{-1}_{s}4\eta\sum^m_{l=1}|z_l(\sigma)|d\sigma=4\eta m{\mathbb{E}}|z_1(0)|<\frac{\lambda_1}{4}.\end{aligned}$$
The existence and uniqueness of invariant measure for the OU equation driven by Lévy process is well-known [@MR900115].
Let $\mu_{\alpha}$ be the unique invariant measure of Lévy type. It is easy to see that $$\begin{aligned}
\lim_{\alpha\to\infty}\int_{V}4\eta\sum^m_{l=1}|z_l(s)|\mu_{\alpha}(dz)=0.\end{aligned}$$ Then for sufficiently large random $s_0>0$ and $s<-s_0$ $$\begin{aligned}
\label{expb}
e^{\int^{t}_s -\frac{\lambda_1}{2}+4\eta\sum^m_{l=1}|z_l(\xi)|d\xi}\le e^{-\frac{\lambda_1}{4}(t-s)}.\end{aligned}$$
To complete the proof this proposition we need the following Lemma.
\[estz\] Assume that $X$ is a stationary process taking values in a Banach space $B$. Moreover, assume that for a certain $p>0$ we have $${\mathbb{E}}\sup_{t\in[-1,0]}|X(t)|_B^p<\infty\,.$$ Then for every $\kappa>0$ such that $\kappa p>1$ there exists a random variable $ \xi$ such that ${\mathbb{P}}$ a.s. $$\begin{aligned}
|X(t)|_B\le \xi+2^\kappa |t|^{\kappa},
\end{aligned}$$ for all $t\le 0$.
Let $\eta_n=\sup_{-n\le s\le -n+1}|X(s)|_B$, $n=0,1,\ldots$ Then by stationarity $$\begin{aligned}
{\mathbb{E}}\eta^p_n={\mathbb{E}}\eta^p<\infty . &
\end{aligned}$$ Therefore, $$\begin{aligned}
{\mathbb{P}}(\eta_n\ge n^{\kappa})\le\frac{{\mathbb{E}}\eta^p_1}{n^{\kappa p}}.\end{aligned}$$
If $\kappa p>1$, then $\sum^{\infty}_{n=1}{\mathbb{P}}(\eta_n\ge n^{\kappa})<\infty$, and by the Borel Cantelli lemma, ${\mathbb{P}}$-a.s., for any sufficiently large $n$, $$\begin{aligned}
\eta_n\le n^{\kappa}.\end{aligned}$$ That is, for every $\omega$, there exists $N(\omega)$ such that $$\begin{aligned}
\eta_n(\omega)\le n^{\kappa},\quad\text{for}\quad n>N(\omega). \end{aligned}$$ Therefore, for $t\in[-n,-n+1]$ we have $$\begin{aligned}
|X(t,\omega)|_B\le \eta_n(\omega)&\le\eta_n(\omega) I_{n\le n(\omega)}+n^\kappa I_{n>N(\omega)}\\
&\le\eta_n(\omega)I_{n\le N(\omega)}+2^\kappa |t|^\kappa.
\end{aligned}$$ Since ${\mathbb{P}}(N<\infty)=1$ the random variable $$\xi(\omega)=\max_{n\le N(\omega)}\eta_n(\omega)$$ is finite ${\mathbb{P}}$-a.s. and the Lemma follows.
With the aid of this lemma, combine with equations (\[vats\]), (\[va\]) and (\[expb\]). We deduce the claim in Proposition \[vaest\]. Moreover, via an apriori estimate about $\int^T_0| v(t)|^2_Vdt$, that is (\[intineq\]) the inequality (\[vatsi\]) follows.
**Step 2** *Measure support.* We now generalise Proposition \[est1\] by proving regularizing property of equation (\[odeim\]) via deducing a priori estimate in $D( A^{\delta})$ for some $\delta>0.$ This allows us to establish support of invariant measure.
\[est3\] For any $\delta\in [0,\frac{1}{2}]$, there exists $C=C(\delta)$ such that for any mild solution $v(\cdot)$ of (\[odeim\]), one has $$\begin{aligned}
\label{est1spt}
| A^{\delta} v(t)|^2\le e^{C\int^t_0| v(s)|^2| A^{\frac{1}{2}} v(s)|^2ds}|A^{\delta} v(0)|^2+C\int^t_0 e^{C\int^t_{\sigma}| v(s)|^2| A^{\frac{1}{2}} v(s)|^2ds}(| A^{\delta+\frac{1}{2}}f|^2+|z(\sigma)|^2+| A^{\frac{1+2\delta}{4}}z(\sigma)|^4)d\sigma.\end{aligned}$$
Multiply (\[odeim\]) by $ A^{2\delta} v$ and integrating over $\mathbb{S}^2$, one finds that $$\begin{aligned}
\label{eq2}
&\ \frac{1}{2}\partial_t| A^{\delta} v(t)|^2+| A^{\frac{1}{2}+\delta} v(t)|^2+(\mathbf{C} v(t), A^{2\delta} v(t))\notag\\
=&\ -b( v(t)+ z_{\alpha}(t), v(t)+ z_{\alpha}(t), A^{2\delta} v(t))+\alpha( A^{\delta} z_{\alpha}(t), A^{\delta} v(t))+\langle A^{\delta}f, A^{\delta} v(t)).\end{aligned}$$ From Lemma 2.4 in [@snse1] it is clear that $$\begin{aligned}
( \mathbf{C} v, A^{2\delta} v)=0.\end{aligned}$$
To complete the proof we need to estimate the terms $b(v+z,v+z,A^{2\delta}v)$, $\alpha\langle A^{2\delta}v, z\rangle$, $\langle A^{2\delta}v, f\rangle$
Using Young inequality with $ab=\sqrt{\frac{\nu}{10}}a \sqrt{\frac{10}{\nu}}b,\,p=2$, we have
$$\begin{aligned}
\alpha|\langle A^{2\delta}v, z\rangle|\le\frac{\nu}{6}|A^{\delta+\frac{1}{2}}v|^2+\frac{3\alpha^2}{2\nu}|z|^2\end{aligned}$$
$$\begin{aligned}
|\langle A^{2\delta}v, f\rangle|\le\frac{\nu}{6}|A^{\delta+\frac{1}{2}}v|^2+\frac{3}{2\nu}|f|^2\end{aligned}$$
Finally, following the method of deriving (15.4.12) as in [@MR1417491], one can show that, for any $\nu>0$, there exists a $K(\nu)$ such that $$\begin{aligned}
\label{bvzadelv}
|\langle A^{2\delta}v,B( v+ z, v+ z)\rangle|&=|b( v+ z, v+ z, A^{2\delta} v)|\\
&\le \frac{\nu}{6} | A^{\delta+\frac{1}{2}} v|^2+K(\nu)(| v|^2| A^{1/2} v|^2+| A^{\frac{1+2\delta}{4}} z|^4).\end{aligned}$$ Combing the above estimates, we have $$\begin{aligned}
&\ \frac{1}{2}\partial_t| A^{\delta} v(t)|^2+(1-3\frac{\nu}{6})|A^{\frac{1}{2}+\delta} v(t)|^2\\
\le &\ K(\nu)|v(t)|^2| A^{\frac{1}{2}} v(t)|^2+K(\nu)|A^{\frac{1+2\delta}{4}} z(t)|^4+\frac{3\alpha^2}{2\nu}| z(t)|^2+\frac{3}{2\nu}|f|^2.\end{aligned}$$ Therefore, invoking Gronwall, it follows that $$\begin{aligned}
| A^{\delta} v(t)|^2&\le e^{K(\nu)\int^t_0| v(s)|^2| A^{\frac{1}{2}} v(s)|^2 ds}| A^{\delta} v(0)|^2\\&+\int^t_0 e^{K(\nu)\int^t_{\sigma}| v(s)|^2| A^{\frac{1}{2}} v(s)|^2ds}\left(\frac{3\alpha^2}{2\nu}|z|^2+K(\nu)|A^{\frac{1+2\delta}{4}}|^4+\frac{3}{2\nu}|f|^2\right)d\sigma.\end{aligned}$$
To complete the proof of invariant measure. It follows from Proposition \[est3\] that for any $t\le -1\le r\le 0$, $$\begin{aligned}
&\ | A^{\delta} v_{\alpha}(0,t)|^2\\
=&\ e^{K(\nu)\int^0_r| v_{\alpha}(s,t)|^2| A^{\frac{1}{2}} v_{\alpha}(s,t)|^2 ds}| A^{\delta}v_{\alpha}(r,t)|^2\\&+\int^0_re^{K(\nu)\int^0_{\sigma}|v_{\alpha}(s,t)|^2| A^{\frac{1}{2}}v_{\alpha}(s,t)|^2 ds}(\frac{3\alpha^2}{\nu}|z|^2+K(\nu)|A^{\frac{1+2\delta}{4}}|^4+\frac{3}{2\nu}|f|^2)d\sigma\\
\le &\ e^{K(\nu)[\sup_{-1\le s\le 0}| v_{\alpha}(s,t)|^2]\int^0_{-1}| A^{\frac{1}{2}} v_{\alpha}(s,t)|^2 ds}\times\left[| A^{\delta} v_{\alpha}(r,t)|^2+\frac{3\alpha^2}{2\nu}|z|^2+K(\nu)|A^{\frac{1+2\delta}{4}}|^4+\frac{3}{\nu}|f|^2d\sigma\right].\end{aligned}$$ Consequently, integrating the above over the interval $[-1,0]$, one gets for $t\le -1$ that $$\begin{aligned}
&\ | A^{\delta} v_{\alpha}(0,t)|^2\\
\le &\ e^{K(\nu)[\sup_{-1\le s\le 0}| v_{\alpha}(s,t)|^2]\int^0_{-1}| A^{\frac{1}{2}} v_{\alpha}(s,t)|^2 ds}\\&\times\left[| A^{\frac{1}{2}} v_{\alpha}(r,t)|+\frac{3\alpha^2}{2\nu}|z|^2+K(\nu)|A^{\frac{1+2\delta}{4}}|^4+\frac{3}{2\nu}|f|^2d\sigma\right].\end{aligned}$$ By Proposition \[vaest\] there exists a random variable $\eta$ such that ${\mathbb{P}}$ a.s. $$\begin{aligned}
| A^{\delta} v_{\alpha}(0,t)|\le \xi,\quad\forall\,\,t\le -1.\end{aligned}$$ Moreover, $$\begin{aligned}
| A^{\delta} u(0,t)|\le | A^{\delta} v_{\alpha}(0,t)|+| A^{\delta} z_{\alpha}(0)|.\end{aligned}$$ Since $ z_{\alpha}(0)$ takes value in $D( A^{\delta})$ there exists another random variable $\zeta$ such that ${\mathbb{P}}$ a.s. $$\begin{aligned}
\label{adelu}
| A^{\delta} u(0,t)|\le\zeta\quad\forall\,\,t\le -1.\end{aligned}$$ So $u(0,t)$ is bounded in probability in the space $D( A^{\delta})$ for some $\delta>0$ satisfies $\sum_{l\ge 1}|\sigma_l|^{\beta}\lambda^{\beta\delta}_l<\infty$: $$\begin{aligned}
\forall\,\,\epsilon>0\,\,\exists\,\,R>0\,\,\forall\,\,t\ge 0\quad{\mathbb{P}}(| u(0,t,u_0)|\ge R)<\epsilon.\end{aligned}$$ Now Let $ u_0$ be fixed and let $\nu_{t, u_0}$ be the law of $u(t, u_0).$ Set $$\begin{aligned}
\mu_T=\frac{1}{T}\int^T_0\nu_{t, u_0}dt.\end{aligned}$$ Let $B_R=\{ x\in D( A^{\delta});| A^\delta x|\le R\}$, equation (\[adelu\]) implies for $p\in(1,\beta)$ $$\begin{aligned}
\mu_T(B^c_R)&\le \frac{1}{TR^p}\int^T_0{\mathbb{E}}|A^{\delta} u(0,t, u_0)|^p\,dt\\
&\le\frac{1}{TR^p}T\mathbb E\zeta^p=\frac{\mathbb E\zeta^p}{R^p}.\end{aligned}$$ We have that, for any $\epsilon>0$, $\mu_T(B_R)=1-\epsilon$ for sufficient large $R$. Hence $\mu_T$ is tight and its limit is an invariant probability measure of the solution $u$ of equation (\[asnse4\]), by Corollary \[kbcor\]. Moreover, the support of the invariant measure is in $D(A^{1/2})$.
Combine with the markov feller properties proved for $u$ earlier, the solution $u$ to equation (\[asnse4\]) admits at least one invariant measure and is supported in $D(A^{1/2})$. Hence, Theorem \[existenceim\] is proved.
To prove uniqueness of invariant measure, one needs irreducibility and strong feller properties of solution semigroup. However, there is a trade-off between well-posedness of the SNSE and the strong Feller property (see the publication [@Dong:2012uq]).
[^1]: The notation $u^{ x}_t$ is used interchangeably with $u(t;x)$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'J.C. Ibáñez-Mejía $^,$$^,$[^1]'
- 'J. Braithwaite'
bibliography:
- 'new\_Toroidal\_field\_stability.bib'
title: Stability of toroidal magnetic fields in stellar interiors
---
[Magnetic fields play an important role during the formation and evolution of stars. Of particular interest in stellar evolution is what effect they have on the transport angular momentum and mixing of chemical elements along the radial direction in radiative regions. Current theories suggest a dynamo loop as the mechanism responsible for maintaining the magnetic field in the radiative zone. This loop consists of differential rotation on one side and magnetohydrodynamic instability - the so–called Tayler instability - on the other. However, how this might work quantitatively is still an unsettled question, largely because we do not yet understand all the properties of the instability in question. In this paper we explore some properties of the Tayler instability.]{} [We present 3D MHD simulations of purely toroidal and mixed poloidal-toroidal magnetic field configurations to study the behavior of the Tayler instability. For the first time the simultaneous action of rotation and magnetic diffusion are taken into account and the effects of a poloidal field on the dynamic evolution of unstable toroidal magnetic fields is included. ]{} [In the absence of diffusion, fast rotation (rotation rate, $\Omega_{\parallel}$, compared to Alfvén frequency, $\omega_{\rm{A},\phi}$) is able to suppress the instability when the rotation and magnetic axes are aligned and when the radial field strength gradient $p<1.5$ (where $p\equiv\partial\ln B/\partial\ln \varpi$ and $\varpi$ is the cylindrical radius coordinate). When diffusion is included, this system turns unstable for diffusion dominated and marginally diffusive dominated regions. If the magnetic and rotation axes are perpendicular to each other, $\Omega_{\perp}$, the stabilizing effect induced by the Coriolis force is scale dependent and decreases with increasing wavenumber. In toroidal fields with radial field gradients bigger than $p>1.5$, rapid rotation does not suppress the instability but instead introduces a damping factor $\omega_{\rm{A}}/2\Omega_{\parallel}$ to the growth rate, in agreement with the analytic predictions. For the mixed poloidal-toroidal fields we find an unstable axisymmetric mode, not predicted analytically, right at the stability threshold for the non-axisymmetric modes; it has been argued that an axisymmetric mode is necessary for the closure of the Tayler-Spruit dynamo loop. ]{}
Introduction {#sec:introduction}
============
Rotation affects the structure of stars, inducing fluid motions such as circulation, and can enhance the mixing of chemical elements. In stably stratified radiative interiors, however, fluid motions are largely restricted to isobaric surfaces, and strong turbulent fluid motions parallel to the isobars smooth velocity gradients in shells resulting in a shellular-like rotation [@Zahn92]. In order to smooth the radial velocity gradients, hydrodynamic processes appear not to be able to transport enough angular momentum between neighboring shells. Consequently, it is thought that some process involving magnetic fields must account for any angular momentum transport in the radial direction.
Even a relatively weak magnetic field is sufficient to couple different parts of the star and maintain a state of nearly uniform rotation. For the interior of the Sun, for example, a field of less than 1 gauss would be able to transmit the torque exerted by the solar wind through the interior [@Mestel53]. Measurements of the solar core’s rotation rate [@Chaplin01] show a very uniform rotation. This uniform rotation may be due to a magnetic field, but this field’s origin, configuration, and strength are not known. By analogy with the magnetic A stars, one might speculate that a ‘fossil’ magnetic field could exist in the core of the Sun. Since no significant net field is seen at the surface (averaged over the solar cycle), the radial component of such a fossil would, however, have to be weak – on the order of a gauss or less. A field weaker than this will quickly wrap up into a predominantly toroidal field, under the action of the remaining differential rotation in the core. Eventually an instability will set in, limiting the growth of this toroidal field.
Near the end of the stellar lifetime, progenitors of white dwarfs and supernovae go through stages where the core contracts and spins up while the envelope expands and spins down. The degree of coupling between core and envelope by a magnetic field in this stage will affect the rotation rates of the stellar leftovers, pulsars, and white dwarfs. It is then fair to ask whether the rotation rate of the remnant is determined by the initial rotation of the progenitor, or if a secondary process must be responsible .
@Tayler73 showed that any purely toroidal field should be unstable in at least some part(s) of the star. He derived stability conditions that are local in the meridional plane but global in the azimuthal direction; indeed, it is the low azimuthal $m$ modes that become unstable first. Physically, one can think of the axisymmetric mode $m=0$ in the following way. A purely toroidal field $B_{\phi}=B_{\phi} (r,\theta)$, which must be axisymmetric in order to satisfy $\bm{\nabla}\cdot \textbf{B} = 0$, can be thought of as a collection of discrete neighboring circular flux tubes. Since all stars have a very high plasma-$\beta$, magnetically-induced motions are restricted to be almost incompressible, $\bm{\nabla}\cdot \textbf{v} \approx 0$. Now, any flux tube whose volume is held constant can reduce its energy by becoming shorter and fatter, which in the case of a flux tube encircling the axis of symmetry means contracting towards the axis. Obviously there is more energy to be released if the field is stronger, so that if we have two flux tubes (of equal volume) lying at different distances from the symmetry axis, if the outer tube contains a greater magnetic flux than the inner tube, interchanging the two will release energy. In a stellar radiative zone, the stable stratification largely suppresses motion in the radial direction, $\textbf{v}\cdot \textbf{g} \approx 0$. However, to interchange flux tubes some radial motion is necessary; it is impossible for flux tubes confined to the same spherical surface to slip past one another. Given that the unstable $m=0$ mode is axisymmetric, rotation, and thus Coriolis force, has no net effect on the stability conditions or growth rate, as it simply causes the annulus to rotate around the axis of symmetry. The non-axisymmetric modes are also driven by the reduction of energy by letting regions with strong magnetic field, and therefore a strong inwards-pointing curvature force, move inwards, and can be thought of in similar terms to the fluting instability in sunspots.
The growth timescale of the instability is on the order of the time taken for an Alfvén wave to travel around the star on a toroidal field line. This timescale is very short compared to the evolutionary timescale of the star, e.g. about 10 years in a main-sequence star with a field of $1$ kG. This form of instability is likely to be the first to set in as the field strength of the toroidal field is increased [@Spruit99] but it is still uncertain what happens next. According to a scenario developed by @Spruit02 the instability could lead to a regeneration of the poloidal field which will eventually be transformed into a toroidal field as a result of the remaining differential rotation, setting up a hydromagnetic dynamo loop. The balance between wrapping-up by differential rotation on the one hand and the destruction of the toroidal field by Tayler instability on the other determines the strength and configuration of the field. The rate at which angular momentum and chemical elements are transported through the stellar interior depends on the resulting field strength, configuration and the details of the instability. This process has been implemented as a sub-grid model in stellar evolution calculations by @Woosley02 [@Heger03] and @Maeder03 and has been numerically reproduced in, with a very ideal configuration, simulations by @Braithwaite06a. More recently the process has apparently been seen in simulations by @Rudiger2014. [^2]
In the long-term, we would like to determine the type of magnetic field that is maintained by differential rotation in a stably-stratified star, under the action of magnetic instabilities, and to develop a quantitative theory for the transport of angular momentum and chemical elements by magnetic fields in stars. This will require extensive simulations which include differential rotation and have a spherical geometry (Boldt & Braithwaite, in prep.) In the meantime we aim to improve our understanding of the Tayler instability. Progress on this front has been made with local simulations by @Braithwaite06b, and has also been investigated in global simulations by . More recently have looked at various properties of the Tayler instability, in particular the effect of rotation, and find that rotation is essentially unable to suppress the instability; one of our aims in this paper is to confirm this result with a different method and to ascertain the growth rate of the instability in the presence of fast rotation.
Another good reason to study the Tayler instability is to gain a better understanding of the range of magnetostatic equilibria available in radiative and other non-convective stars and zones. Since the discovery by @Babcock47 of a strong magnetic field in the A star 78 Vir, strong, large-scale fields are now known to exist in a subset of early-type stars as well as white dwarfs and neutron stars (see e.g. Braithwaite & Spruit subm. for a review). These fields, known in the literature as fossil fields, are the relics of some previous stage of evolution, and have found their way into a stable equilibrium, evolving only on the very long diffusive timescale. An arbitrary initial magnetic field is known to relax in a non-convective star into such an equilibrium , and there are apparently a large range of equilibria available, including both approximately axisymmetric ’twisted torus’ configurations as well as more complex non-axisymmetric configurations [@Braithwaite08]. Amongst the axisymmetric equilibria, it is an interesting question to ask in what ratio the poloidal and toroidal field components must be present to ensure stability, since both are unstable on their own. It seems that a relatively weak poloidal field can stabilize a toroidal field [@Braithwaite09; @Akgun13] but the detail of how this works is still uncertain. One reason why the ratio of the two is particularly interesting is that a star with a predominantly poloidal field is oblate in shape, but if the field is predominantly toroidal, the star is prolate. Fast-spinning highly-magnetized neutron stars (the so-called millisecond magnetars) with predominantly toroidal magnetic fields might flip over until the magnetic and rotation axes are orthogonal, the damping of free precession reducing the kinetic energy by maximizing the moment of inertia about the rotation axis. Such stars would emit gravitational waves potentially detectable with the next generation of detectors.
In this paper we use numerical methods to investigate the linear properties of the Tayler instability, much of which is still uncertain. We provide a useful check on the analytical results, which may have missed some aspects of the mechanism. We use similar methods to those used by @Braithwaite06b to investigate the Tayler instability, but we investigate different aspects of the instability. In the next section, we look at what is known about the instability, before describing our numerical model in section \[sec:model\], presenting our results in section \[sec:results\] and summarizing in section \[sec:conclusions\].
Stability of toroidal fields: analytic results {#sec:analytics}
==============================================
@Tayler73 derived the stability conditions of purely toroidal fields in radiative stellar interiors, ignoring the effects of viscosity and of thermal and, magnetic diffusion. The occurrence of instability for the adiabatic case is independent of the field strength but only depends on the field configuration. However, when the instability is already present the growth rate depends on some measure of the field strength and the stratification. The field strength is characterized by the Alfvén frequency, which can be thought as the inverse of the timescale taken by an Alfvén wave to go around the star along a toroidal field line, defined as $$\omega_{\rm{A}}=\frac{\textrm{v}_{\rm{A}}}{R_\ast}=\frac{\rm{B}_{\phi}}{R_\ast \sqrt{4\pi\rho}},
\label{eq:tayler_inst}$$ where $\textrm{v}_{\rm{A}}= \rm{B}_\phi /\sqrt{4\pi\rho}$ is the Alfvén velocity in a toroidal field line and $R_\ast$ is the radius of the star. The stratification is characterized by the Brunt-Väisälä or the buoyancy frequency, $N$, and radiative zones are always in the regime $$N \gg \omega_{\rm{A}, \phi}.$$ This is the same as saying that the thermal energy density is much bigger than the magnetic energy density, i.e. a large plasma- $\beta=\rm{P}/8\pi \rm{B}^2$ parameter, which is true in all stars.
Following Tayler, we make the analysis easier by using cylindrical coordinates $(\varpi,\phi,z)$. We now consider an axisymmetric, azimuthal toroidal field independent of height $z$, $\textbf{B}=\textrm{B}(\varpi)\bf{e}_\phi$, and look at the region around the axis. The important parameter is the radial field gradient $p$: $$p\equiv\frac{\rm{d}\log \rm{B}}{{\rm d}\log \varpi} \quad\quad\quad \text{as} \quad\quad\quad \varpi\rightarrow 0.
\label{eq:field_grad}$$
Now, if we perturb a toroidal field in equilibrium, the resulting displacement can be described by $$\xi \sim \it{e}^{i(m\phi+kz)+\sigma t},
\label{eq:displacement}$$ where the variation in the $\varpi$ direction is not included as its length scale is much larger. Although displacements are predominantly horizontal because of the vertical stratification, the incompressible continuity equation (whose use is justified by the very high plasma- $\beta$) shows that a horizontal displacement $\xi_{\rm{h}}$ is necessarily accompanied by a radial displacement on the order of $\xi_r \sim \xi_{h} l_{r} / l_{h}$, where $l_h$ and $l_{r}$ are horizontal and radial length-scales, respectively, so that near the axis $l_{r}=1/k$. The shape of some of the unstable modes is shown in figure \[fig:tayler\_inst\]. Note that the $m=0$ mode is also known as the sausage mode and the $m=1$ mode as the kink mode.
![Physical form of the different unstable azimuthal $m$ modes present in a plasma column with a purely toroidal field. Below each mode the cross section of the co-moving surface is shown along with the cross section of the equilibrium state. The only axisymmetric mode is $m=0$ and the only mode that preserves the column’s cross section is the $m=1$ mode.[]{data-label="fig:tayler_inst"}](tayler_inst.eps)
In this figure we can observe why perturbations like (\[eq:displacement\]) in a purely toroidal field are unstable. Regions where magnetic field is pinched have higher magnetic pressure forcing the fluid to displace opposite to the pressure gradient, leading to a runaway process. Figure \[fig:tayler\_inst\] shows the equilibrium configuration and the shape of the azimuthal $m=0$ and $m=1$ unstable modes. In a stellar interior we often expect a vertical stratification, and vertical displacements experience a damping effect because they have to do work against the stratification. However, given that the $m=1$ mode mostly drives azimuthal displacements, the damping effect of the stable stratification is very weak.
@Tayler57 showed that for the instability to develop the field must satisfy $$p>\frac{m^2}{2}-1 \,\, (m \neq 0) \,\,\, \text{and} \,\,\, p>1 \,\,\, (m=0).
\label{eq:stability_non-rotating}$$ These instability conditions are valid only for non-viscous, non-rotating, non-diffusive, radiative stellar interiors. Examining this stability condition (\[eq:stability\_non-rotating\]) we observe that for a $p=1$ field, the $m=1$ mode is the only unstable mode, the $m=0$ and $2$ modes are marginally stable and the other modes are stable. For a somewhat steeper radial field gradient of $p=2$, modes $m=0, 1, 2$ are expected to be unstable.
To study the growth of the instability we define the local toroidal[^3] Alfvén frequency as $$\omega_{\textrm{A}, \phi} \equiv \textrm{v}_{\textrm{A},\phi}/\varpi = \frac{B_{\phi}} {\varpi \sqrt{4\pi\rho}},
\label{eq:alfven_freq_toroidal}$$ where $\textrm{v}_{\textrm{A},\phi}$ is the toroidal Alfvén speed and $\varpi$ the distance to the axis, we expect the growth rate of the instability to be $$\sigma \sim \omega_{\textrm{A}, \phi}.
\label{eq:ideal_growth}$$ Given that toroidal magnetic fields in stars are buried under the surface, there is no direct observational evidence of their configuration, therefore we have no knowledge of the radial magnetic field gradient. Even less understood is the configuration of a magnetic field inside a differentially rotating star. However we know that towards the magnetic axis the field must tend to zero, and to avoid singularities the radial gradient $p \geq 1$ for $\varpi \rightarrow 0$, and for this reason in this paper we start by studying the $p=1$ case. Since it is expected that there is a qualitative threshold at $p=3/2$ (see section \[subsec:rotation\_effects\]), we then also study the case of a $p=2$ field gradient.
The unstable growth rate in equation \[eq:ideal\_growth\] is valid only for purely toroidal, non-rotating, non-stratified and non-diffusive cases. We now discuss the behavior of the Tayler instability for situations where these very ideal conditions are relaxed and discuss what happens when magnetic diffusion is included, rotation is present and the magnetic field configuration is not purely toroidal anymore. By doing so, the stability conditions and growth rates are modified, and furthermore growth rates develop length scale dependencies. A stable stratification has major effects on the behavior of the instability, this is however outside the scope of this paper and is already discussed in @Braithwaite06b.
Effects of rotation {#subsec:rotation_effects}
-------------------
All stars are rotating and a big fraction rotates at very high speeds [@Collins63; @Ramirez-Agudelo].
### Rotation and magnetic axes aligned
and @Spruit99 showed that both the stability conditions and the unstable growth rate are affected by fast rotation. When the rotation frequency is higher than the Alfvén frequency ($\Omega_{\parallel} > \omega_{\rm{A}, \phi}$) and the rotation and magnetic axes are aligned, the stability condition is modified from (\[eq:stability\_non-rotating\]) to $$p>\frac{m^2}{2}+1 \;\; ( m \neq 0 ) \quad\quad\quad \text{ and } \quad\quad\quad p>1 \;\; (m=0)
\label{eq:stability_rotating}$$ so that the condition of stability against all $m$ modes is $p<3/2$ rather than $p<-1/2$ as before, and furthermore, the growth rate when any unstable mode is present is damped by a factor of $\omega_{\textrm{A}, \phi}/\Omega_{\parallel}$ [see @Acheson78; @Spruit99].
![[*Above:*]{} Mixed poloidal-toroidal field configuration. If the toroidal field is too weak or indeed absent, the poloidal field is subject to an instability. [*Below:*]{} Left, an equilibrium configuration with poloidal field loops threaded by the neutral line, and on the right-hand-side the kink-like instability present around the neutral line. Unstable displacements are restricted to be approximately perpendicular to the stratification. Figure taken from @Braithwaite07. []{data-label="fig:poloidal_field"}](poloidal_field.eps "fig:") ![[*Above:*]{} Mixed poloidal-toroidal field configuration. If the toroidal field is too weak or indeed absent, the poloidal field is subject to an instability. [*Below:*]{} Left, an equilibrium configuration with poloidal field loops threaded by the neutral line, and on the right-hand-side the kink-like instability present around the neutral line. Unstable displacements are restricted to be approximately perpendicular to the stratification. Figure taken from @Braithwaite07. []{data-label="fig:poloidal_field"}](instability_neutral_line.eps "fig:")
### Rotation and magnetic axes not aligned
It is not always the case that the magnetic axis, the rotation axis and the gravity are aligned, as in the cases looked at above. Figure \[fig:poloidal\_field\] (upper panel) shows the poloidal field lines in an axisymmetric configuration, and a toroidal-field-like configuration is observed around the neutral field line around the star, where the poloidal field vanishes. This configuration suffers from the same kind of instability (Tayler instability when considered locally), the difference being that the magnetic axis (in this case the neutral line) is perpendicular to gravity, but because of the switch in geometry, it is now the poloidal field that drives the instability, instead of the toroidal field as before; during local considerations it therefore makes sense to switch the definitions of the two words. Also note that in the presence of rotation, the magnetic and rotation axis wont be alligned somewhere in the star. Kink-like instabilities appear with displacements perpendicular to the direction of stratification, $ \textbf{v} \cdot \textbf{g} \approx 0$, as appreciated in figure \[fig:poloidal\_field\] (lower panel). Displacements are mainly in the direction of the arrows, but, as discussed in section \[sec:analytics\], there are also necessarily smaller displacements in the direction parallel to the neutral line . The ratio between the displacements in these two directions should be equal to the ratio of length scales in the same two directions, as we can see from the constraint $\bm{\nabla} \cdot \textbf{v} \approx 0$, a consequence of the high plasma-$\beta$.
The Coriolis force affects motions perpendicular to the rotation axis; in the case where the magnetic and rotation axes are aligned, unstable displacements are mainly perpendicular to the rotation and magnetic axes, and stabilization in the case $-1/2<p<3/2$ is achieved when $\Omega_{\parallel} \geq \omega_{\textrm{A}, \phi}$. When the magnetic and rotation axes are not aligned, the amplitude of the displacements perpendicular to the rotation axis decrease with increasing inclination angle, and the extreme case occurs when the rotation and magnetic axes are perpendicular to each other. Then naively we think that in order to stabilize the perpendicular rotator, a faster rotation rate is required, and as we assumed that the fluid is almost incompressible, $\bm{\nabla} \cdot \textbf{v}\approx 0$, we derive the stability condition $$\Omega_\perp \geq \omega_{\textrm{A}, \phi} \varpi_{0} k,
\label{eq:critical_rotation_oblique}$$ where $\Omega_{\perp}$ is the perpendicular rotation rate with respect to the magnetic field axis, $\varpi_{0}$ is the typical length scale of the motion perpendicular to the magnetic field and $k$ is the wavenumber in the direction of the magnetic axis (here the neutral line) as defined in (\[eq:displacement\]). Note that the new perpendicular rotating stability conditions now dependent on the wavenumber, unlike in the aligned rotator case.
![Representations of the first two vertical $n$ modes of the unstable $m=1$ mode. The figure on top shows the purely toroidal magnetic field configuration, where the orientation of the magnetic symmetry axis $\hat{B}$ is perpendicular to the rotation axis $\hat{\Omega}_{y}$. It also shows what an $n=1$ and $n=2$ disk looks like as the red and blue box, respectively. The two figures below show the slipping disks for the $n=1$ (left) and $n=2$ (right) cases. The magnitude of the deformation $\Delta l_{\perp}$, and the deformation velocity $\Delta \textrm{v}_{\perp}$, suffered by the disks is different depending on the unstable mode. Small $n$ modes present large deformations, whereas high $n$ modes show small deformations, $\Delta l_{\perp}(n=1) > \Delta l_{\perp} (n=2)$ and $\Delta \textrm{v}_{\perp}(n=1) > \Delta \textrm{v}_{\perp} (n=2)$. []{data-label="fig:oblique_deformation"}](SigmaOmegay.eps)
This effect can be explained if we imagine the plasma column to be a series of disks subject to pressure on top and below. When the disks displace perpendicular to the magnetic axis they suffer a deformation fattening at one side and shrinking at the other. The magnitude of the deformation depends on the size of the disks, as shown in figure \[fig:oblique\_deformation\], where disk of different sizes correspond to different vertical $n$ modes. This is simply a consequence of the constraint ${\bm\nabla}\cdot \textbf{v}\approx0$, which can be written as $k \textrm{v}_z\sim \varpi_0 \textrm{v}_\varpi$. The Coriolis force affects motion in the vertical direction; Greater deformations, in other words small $\textit{n}$ modes, feel a stronger stabilizing effect. Note that the Coriolis force also affects motions perpendicular to both the magnetic and rotation axis, but this motion, at least for the $m=1$ mode, can be minimized by combining left- and right-handed spiral modes with the right phase.
Effects of magnetic diffusion {#subsec:mag_diff_effects}
-----------------------------
Until now we have only considered adiabatic perturbations, that is ignoring the effects of thermal and magnetic diffusion, as well as viscosity. We now relax the flux freezing condition and allow the field lines to slip through the plasma.
and @Spruit99 suggested that in the presence of any diffusive process, thermal or magnetic, the instability condition in fast rotators (\[eq:stability\_rotating\]) would recover its original form given in condition (\[eq:stability\_non-rotating\]) [see @Spruit99 Appendix]. The reason for this can be thought of in terms of energy dissipation: rotation tends to give stability because the Coriolis force, which acts perpendicularly to the velocity and has consequently no effect on the energies, drives the fluid back into its equilibrium position; diffusion damps these epicycle motions, making a return to the equilibrium energetically impossible; as magnetic energy is turned into heat the system moves irreversibly away from its original equilibrium.
Magnetic diffusion is scale dependent, and therefore when included in the analysis, the unstable growth rate of the Tayler instability becomes scale dependent as well [see e.g. @Spruit99]. For a given magnetic diffusivity, $\eta$, the diffusion rate is given by, $\eta k^2$, where $k$ is the wavenumber.At first glance one expects this to impose an upper limit for the unstable wavenumbers that are able to grow; perturbations at a larger wavenumbers should be smoothed away faster than they can grow. We have stability if $$\begin{aligned}
\eta k^2 > \sigma
\label{eq:max_wavenumber}.\end{aligned}$$ Therefore for a given wavenumber we can define a critical diffusivity $$\eta_{\rm crit} = \omega_{A,\phi} /k^2,$$ From now on, we will express the different diffusive environments as fractions of the critical magnetic diffusivity $$n_{\eta} = \frac{\eta}{\eta_{crit}},
\label{eq:magdiff_frac}$$ where one can always move back and forth between fraction of critical magnetic diffusivity and the wavenumber of interest with the following relation $$n_{\eta} = k^2 \left(\frac{\eta}{\omega_{A,\phi}}\right),
\label{eq:n_eta}$$ where $\eta$ is the magnetic diffusivity of the system and $k = 1/l_{z}$ is the vertical wavenumber.
Mixed poloidal-toroidal configurations {#subsec:mixed_field_effects}
--------------------------------------
The addition of a poloidal field is expected to modify or suppress the instability. If one imagines the toroidal and poloidal fields to operate independently of each other (in some vague sense) then some of the energy liberated from the toroidal field must be used to bend the poloidal field lines. If the poloidal field is strong, the instability might be suppressed altogether. However, the effect will depend on the wavenumber of the mode; whilst all wavenumbers release the same energy from the toroidal field, higher wavenumbers have to do more work against the poloidal field. This gives rise to a wavenumber stability threshold. In general, if an instability displaces a fluid element horizontally a distance $\xi_{\textit{h}}$, it does so by means of a force per unit mass of $\textit{F}_{\textit{h}}=\sigma^2\xi_{\textit{h}}$ where $\sigma$ is the growth rate. In the case of the Tayler instability, $\sigma\approx \omega_{\textrm{A}, \phi}$ as discussed at the beginning of section 2 (equation \[eq:tayler\_inst\]). Imagine, in the cylindrical geometry used above, that there is a poloidal field in the $z$ direction. The displacement caused by the instability drags the poloidal field with it, which reacts against this bending with a force per unit mass of $B_z B_\varpi k/\sqrt{4\pi\rho}\approx B_z^2 \xi_{\rm h} k^2/\sqrt{4\pi\rho}$, where $k$ is the radial (vertical) wavenumber. For a perturbed displacement to grow it is required that the instability force is bigger than the restoring force by bending the poloidal field line, which gives a condition on the unstable wavenumbers $$k < \frac{\sigma}{\textrm{v}_{{\rm A},z}},
\label{eq:poloidal_stability}$$ where $\textrm{v}_{{\rm A},z}=B_z/\sqrt{4\pi\rho}$ is the vertical Alfvén speed. See @Spruit02 and @Braithwaite09 for a fuller discussion of this topic.
Model {#sec:model}
=====
To determine the effects of rotation and magnetic diffusion on the Tayler instability, local simulations of a radiative stellar interior are performed. For our local simulations we select a box containing the magnetic axis of the star, located at the center of the box, as shown in figure \[fig:Simulation\_box\]. Global simulations of the stellar radiative region are not suitable at this stage because of resolution limitations to accurately resolve small scales. However, global simulations might eventually be necessary to investigate the non-linear behavior in a differentially rotating star, or rather, global in two dimensions and local in the radial direction.
![Predominantly toroidal magnetic field configuration in the interior of a star. The box shows a zoomed-in region containing the toroidal field lines around the magnetic axis and bold arrows showing the direction of the parallel and perpendicular rotation axes.[]{data-label="fig:Simulation_box"}](Simulation_box.eps)
The equation of state used is the ideal equation of state, in units such that the molar gas constant divided by the molecular mass is unity $$\begin{aligned}
\textit{P}=\rho\textit{T} \;\;\; \text{and} \;\;\; \textit{e}=\frac{P}{\gamma-1}.\end{aligned}$$
To trace the evolution of the system we solve the standard MHD equations, momentum equation: $$\begin{aligned}
\frac{\partial{\rho\textbf{v}}}{\partial t}=-{\bm\nabla}\cdot(\rho\textbf{vv})-{\bm\nabla} P+ \frac{1}{4\pi}{\bm\nabla}\times {\bf B} \times {\bf B} -\textbf{g}\rho+2\rho\textbf{v}\times\Omega,
\label{eq:momentum_eq}\end{aligned}$$ continuity equation, $$\begin{aligned}
\frac{\partial{\rho}}{\partial t}=-\nabla \cdot \rho \textbf{v},
\label{eq:conservation_of_mass}\end{aligned}$$ induction equation, $$\begin{aligned}
\frac{\partial{\textbf{B}}}{\partial t} = \nabla\times \left[ (\textbf{v}\times\textbf{B})-\eta\nabla\times\textbf{B} \right],
\label{eq:induction2}\end{aligned}$$ and the specific internal energy: $$\begin{aligned}
\frac{\partial e}{\partial t}=-\nabla\cdot(\textit{e}\textbf{v}) - P\nabla\cdot\textbf{v}
\label{eq:energy} $$ where $\rho$ is the density, $\textbf{v}$ the velocity, $\textbf{B}$ the magnetic field, $\eta$ the magnetic diffusivity, $\textit{e}$ the internal energy, $\textbf{g}$ the gravity, $\textit{P}$ gas pressure, and $\textit{T}$ gas temperature. The ratio of specific heats $\gamma$ is set to $5/3$. We note that in the momentum equation (\[eq:momentum\_eq\]) we neglected viscosity and the Coriolis term is explicitly included. In the induction equation (\[eq:induction2\]) magnetic diffusion $\eta$ is spatially constant. The energy equation neglects viscous and Ohmic heating and thermal conduction, since we are not interested at this stage in viscous and thermal effects. We do not include the effects of gravity in this study.
Numerical code {#subsec:numerical_code}
--------------
We use the stagger code, a parallel 3D MHD code originally developed by [@stagger], and continuously improved over the years by its user community. The Stagger-code uses a sixth-order explicit finite-difference scheme for numerical derivatives and the corresponding fifth- order interpolation scheme. The solution of the hydrodynamic equations is advanced in time using an explicit third-order Runge-Kutta integration method [@Williamson80]. The code operates in a staggered, Eulerian, Cartesian mesh, which seem to be a better implementation than cylindrical and spherical coordinates, as a Cartesian mesh avoids the coordinate singularity in the axis present in cylindrical and spherical grid-based codes. Another advantage of a Cartesian based code is reflected in the simplicity of the equations with respect to another coordinate system. Cartesian coordinates also have their own disadvantages such as the noise introduced by fitting a cylindrical shape, e.g. torus, in squared boxes, or the extra computational time cost due to the un-used corners of the box.\
For the simulations of purely toroidal fields a resolution of $48^2$ in the azimuthal direction and 32 in the vertical direction was used, for the mixed poloidal-toroidal fields the resolution was $48^3$. Test cases with a resolution of $96^3$ were performed in order to compare the results with the low resolution simulations and no significant differences were observed. Periodic boundaries are used in all simulations; see below.
The code also contains highly localized diffusivities that prevent the growth of wavelengths close to the Nyquist frequency. We however decided to switch off this artificial diffusive terms in order to avoid non-physical effects. This was possible given the clean initial conditions, where the development and growth of the Tayler instability occurs at very early times when no artificial stabilization is required.
Initial conditions {#subsec:initial_conditions}
------------------
The initial conditions correspond to the equilibrium configuration of a plasma containing a magnetic field of the form: $$\begin{aligned}
\textbf{B}=\textrm{B}_\phi(\varpi)\textbf{e}_\phi+\textrm{B}_z\textbf{e}_z.
\label{eq:magnetic_field}\end{aligned}$$ The magnetic configuration in this paper has no radial component, $\textrm{B}_{\varpi}=0$, and is a combination of an azimuthal, $\textrm{B}_{\phi}$, and a vertical, $\textrm{B}_z$, components with no dependence on height $z$ or angle $\phi$.
The azimuthal component is an axisymmetric torus defined by $$\begin{aligned}
\label{eq:toroidal_field}
B_{\phi}(\varpi)=B_{0,\phi}\left(\frac{\varpi}{\varpi_0}\right)^p e^{-(\varpi/\varpi_0)^2} ,
\label{eq:bphi}\end{aligned}$$ where $\varpi_{0}$ is a constant length, set to one quarter of the horizontal size of the box, $\textit{p}$ is the radial field gradient, set to either 1 or 2 in the simulations, and $\textrm{B}_{0,\phi}$ is a constant. The exponential term $\exp(\varpi^2$/$\varpi_0^2)$ shapes the field such that it behaves as $\textrm{B}\propto \varpi^{\textit{p}}$ near the axis and vanishes farther away so as to avoid problems at the boundaries. The local toroidal Alfvén frequency, as defined in (\[eq:alfven\_freq\_toroidal\]), is $\omega_{\textit{A},\phi} = \textrm{B}_\phi/\varpi\sqrt{4\pi\rho}$. The maximum initial field strength in the box corresponds to a minimum plasma beta parameter of $\beta = 10$, i.e. thermal pressure dominates over magnetic pressure.
The initial poloidal field is implemented as a uniform field in the vertical direction $$\begin{aligned}
\label{eq:poloidal_field}
\textrm{B}_{\it{z}} = \textrm{B}_{0,\textit{z}}
\label{eq:bz0}\end{aligned}$$ for this field configuration the vertical Alfvén frequency is also constant everywhere, $\omega_{\textit{A},\textit{z}} = \textrm{v}_{\textit{A,z}}(2\pi/\textit{L}_{\textit{z}})$, where $\textrm{v}_{\it{A,z}}=\textrm{B}_{\it{z}}/\sqrt{4\pi\rho}$ is the vertical Alfvén velocity and $\it{L}_{\it{z}}$ is the vertical extent of the box.
![a) Initial purely toroidal magnetic configuration. The region of interest is between $0 \le \varpi \lesssim \varpi_{0}$ given that the field strength falls off at greater $\varpi$. Colored areas represent surfaces of equal distance to the axis. Field lines are also plotted; their density is related to the increase in the field strength proportional to the distance from the axis. b) Mixed poloidal-toroidal field configuration. The field lines show a spiral morphology, for this particular case (with $p=1$) two spirals turns are completed in the box, i.e. toroidal Alfvén frequency is double the vertical Alfvén frequency $\omega_{\it{A}, \phi} = 2\omega_{\it{A},z}$. []{data-label="fig:schematic_view"}](schematic_view.eps "fig:") ![a) Initial purely toroidal magnetic configuration. The region of interest is between $0 \le \varpi \lesssim \varpi_{0}$ given that the field strength falls off at greater $\varpi$. Colored areas represent surfaces of equal distance to the axis. Field lines are also plotted; their density is related to the increase in the field strength proportional to the distance from the axis. b) Mixed poloidal-toroidal field configuration. The field lines show a spiral morphology, for this particular case (with $p=1$) two spirals turns are completed in the box, i.e. toroidal Alfvén frequency is double the vertical Alfvén frequency $\omega_{\it{A}, \phi} = 2\omega_{\it{A},z}$. []{data-label="fig:schematic_view"}](poloidal_toroidal.eps "fig:")
A visual example of the field configuration, for purely toroidal and mixed poloidal-toroidal fields, is shown in figure \[fig:schematic\_view\]. The initial conditions for the thermodynamic variables are determined by the equilibrium solutions found by setting $\partial/\partial t = \bf{v} = \eta = 0$ in equations (\[eq:momentum\_eq\]) to (\[eq:energy\]), with the initial magnetic field given by equations (\[eq:magnetic\_field\], \[eq:bz0\] and \[eq:bphi\]). No height dependence is considered, therefore pressure, density and temperature depend only on the distance from the axis. For a full derivation of these stability configurations see @Braithwaite06b (section 3.3).
Four different cases for two different field configurations are simulated:
- rotating diffusive star,
- perpendicular rotator,
- rotating star with a quadratic radial field gradient: $p=2$, and
- mixed poloidal-toroidal field.
**Case 1**: Rotating diffusive star with linear radial field gradient, $p=1$. From (\[eq:toroidal\_field\]) we see that the toroidal Alfvén frequency varies as $\omega_{\rm A}\propto \varpi^{p-1}$ at low $\varpi$, so that it is approximately independent of location in the inner part of the computational box. The magnetic and rotation axes are aligned. We explore the parameter space of rotation-only, magnetic diffusion-only and the simultaneous action of rotation and diffusion.\
**Case 2**: Perpendicular rotator with linear radial field gradient, $p=1$. Rotation and magnetic axes are perpendicular. We explore the rotation rate parameter space.\
**Case 3:** Rotating star with a quadratic radial field gradient, $p=2$. The Alfvén frequency increases linearly with the distance to the axis. For this case rotation and magnetic axes are aligned. We explore the rotation rate parameter space.\
**Case 4:** Mixed poloidal-toroidal configuration. The poloidal field is constant in the box, therefore the vertical Alfvén frequency is also constant. The toroidal field has a linear radial field gradient, $p=1$, thus the azimuthal Alfvén frequency is also constant in the inner part of the box. We explore the poloidal-to-toroidal field strength ratio parameter space.\
Given that the initial configurations correspond to equilibrium solutions, an initial perturbation is included to trigger instability.
Initial perturbation {#subsec:initial_perturbation}
--------------------
The initial conditions in section \[subsec:initial\_conditions\] correspond to a perfect hydrostatic equilibrium, in order to see any instability grow we seed tiny perturbations in the velocity field. Given the cylindrical symmetry of this problem, we can perturb the different azimuthal, $m$, and vertical, $n$, modes independently. The shape and description of the perturbation used here are the same as used by @Braithwaite06b [sec. 3.5]. The perturbation has its maximum at $\varpi=0$ and decreases with increasing $\varpi$, almost vanishing at $\varpi=\varpi_{0}$. There is no perturbation to the vertical velocity component.
The initial perturbation has the form $$\begin{aligned}
\textbf{v} &= \sum_{m=0,n} V^{n}_{0} \frac{\varpi}{\varpi_{0}}\text{exp} \left (- \frac{5}{2} \frac{\varpi^2}{\varpi^2_{0} }\right) \text{cos}(k z) \textbf{e}_{\varpi} + \nonumber \\
& \sum_{m \not=0,n} V^{n}_{m} \text{exp}\left(-3\frac{\varpi}{\varpi_0}- \frac{\varpi^2}{\varpi^2_0}\right) [\text{cos}(k z-m\phi) \textbf{e}_{\varpi} + \text{sin}(k z-m\phi)\textbf{e}_{\phi}]. \nonumber
\label{eq:displacement_mmode}\end{aligned}$$
The amplitude coefficients of the perturbed velocity field are normalized to the maximum Alfvén velocity in the box: $\textrm{V}_{\textit{m}}^{\textit{n}}/v_{A,\text{max}}=1\times10^{-8}$. Such a value is large enough to trigger the Tayler instability and small enough to let us follow the linear growth phase for several growth timescales.
Boundary conditions {#subsec:initial_perturbations}
-------------------
The box represents a small volume of a radiative zone in a stellar interior. Figure \[fig:Simulation\_box\] is a cartoon of the box location in star with respect to the magnetic field axis.
The difference in length scales between the instability and the radius of the star are huge. The instability is many orders of magnitude smaller[^4] than the radius of the star $(\it{l}_{\rm{r}} \ll \rm{R}_{*})$, these instabilities are local in the radial direction and global in the azimuthal direction, therefore different shells of the star are in effect disconnected from each other.
Periodic boundaries are the best option for our simulation, given that we want to avoid the selection of special locations in the stellar interior and to prevent additional effects such as current sheets or shear.
Results {#sec:results}
=======
First, test cases without rotation or magnetic diffusion were simulated with field gradient $p=1$. The Tayler instability is properly traced for a number of growth timescales. More physical ingredients are built up each step in order to approach our simple model with the conditions found in a radiative stellar interior.
For the analysis of the simulations we need a formalism for measuring the amplitude of the various azimuthal $m$ and vertical $n$ modes, keeping in mind that $n$ is the dimensionless vertical wavenumber; the $n=1$ mode corresponds to a wavelength equal to the height of the computational box. We extract information from two fields: the change in the magnetic field $\delta\textbf{B}=\textbf{B}-\textbf{B}(t=0)$, and the velocity field **v**. If we imagine the field to be frozen into the fluid, then $\delta \textbf{B}$ reflects the displacement ${\bm\xi}$, the integral of the velocity **v**.
We can now decompose the $\phi$ component of these fields using Fourier series in order to extract the amplitudes of the different azimuthal $m$ modes. For instance with the velocity field $$\begin{aligned}
\textbf{v}(\varpi,\phi, \textit{z},\, t) = \frac{1}{2} \textbf{\textrm{A}}_{0} (\varpi, \textit{z},t) +\sum_{ \textit{m}=1}^{\textit{m}=\infty} \, \Re(\textbf{A}_{m}(\varpi,\textit{z},\textrm{t}) \, \textit{e}^{\textit{im}\phi}),
\label{eq:first_fft}\end{aligned}$$ where $\textbf{A}_{0}$ and $\textbf{A}_{m}$ are the Fourier coefficients given by $$\begin{aligned}
\textbf{A}_{0} (\varpi, \textit{z},t) = \frac{1}{\pi} \int_{-\pi}^{\pi} \textbf{v}(\varpi, \phi, z, t) \, {\rm d}\phi \end{aligned}$$ and $$\begin{aligned}
\textbf{A}_{m} (\varpi, \textit{z},t) = \frac{1}{\pi} \int_{-\pi}^{\pi} \textbf{v} (\varpi, \phi, z, t) \, \cos(m\phi) \, {\rm d}\phi.\end{aligned}$$
We then compute the amplitude of each $\textit{m}$ mode as $$\begin{split}
\textrm{v}\textit{\_Amp}_{\textit{m}}(\rm{t}) = \left( \frac{\int_{0}^{\varpi_{\textrm{max}}} \int_0^{L_z} \textbf{A}_{m}^{*} \, \textbf{A}_{m} \, 2\pi\varpi \, {\rm d}\varpi {\rm d}\rm{z} }{\varpi_{\rm max}^2 L_z \pi}\right)^{1/2},
\label{eq:amplitude_mode}
\end{split}$$ where $\varpi_{\textrm{max}}$ is set to $\varpi_0/2$, as the magnetic field strength in the region between $\varpi=0 $ to $ \varpi_{0}/2$ is approximately proportional to the distance from the axis to the power $\rm{p}$, i.e. $B_\phi \propto \varpi^{\textrm{p}}$. A similar procedure can be applied to the magnetic fields to derive the magnetic displacement amplitude, $\delta B\textrm{\_Amp}_m(t)$.
![Amplitude of the first four $m$ modes as a function of time, extracted from the magnetic field $\delta B(t)$, and normalized to the wavelength corresponding to the $n=1$ vertical mode. The azimuthal $m=1$ mode, solid red line, is the only unstable mode. []{data-label="fig:amplitude_default_case"}](amp_growth_p1.eps)
Figure \[fig:amplitude\_default\_case\] shows the evolution of the amplitudes of the first three $m$ modes as a function of time. The $m=1$ mode is the only unstable mode, as predicted above in (\[eq:stability\_non-rotating\]). When the amplitude of the unstable $m=1$ mode is comparable to the wavelength, $\delta B\_Amp_m \approx \lambda_{n=1}$, the system goes into the non-linear regime. At this point energy can be transferred between modes and $ m\neq 1$ modes begin to grow. This is observed to happen around $t/\tau_{\rm A}\approx 10$ (where $\tau_{\rm A}=1/\omega_{\rm A}$) in figure \[fig:amplitude\_default\_case\].
The displacement amplitude computed using equation (\[eq:amplitude\_mode\]) contains the contribution of all vertical wavelengths. For further analysis, it will be necessary to measure the amplitudes of each individual vertical $n$ modes separately. A second Fourier decomposition, similar to the one applied in equation (\[eq:first\_fft\]), is performed in the vertical direction, $$\begin{aligned}
\delta\textbf{v}(\varpi, \phi, z, t) = \frac{1}{2} \textbf{C}_{m, 0} \, (\varpi, t) + \sum_{n=1}^{n=\infty}\Re(\textbf{C}_{m, n}(\varpi,t) \, \textit{e}^{inz}),
\label{eq:second_fft}\end{aligned}$$ where $\textbf{C}_{m,0}$ and $\textbf{C}_{m, n}$ are the new Fourier coefficients given by: $$\textbf{C}_{m,0} = \frac{1}{\pi L_{z}} \int_{0}^{L_{z}} \int_{-\pi}^{\pi} \delta \textbf{v} \, \cos(m\phi) \, d\phi \, dz$$ and $$\textbf{C}_{m, n} = \frac{1}{\pi L_{z}} \int_{0}^{L_{z}} \int_{-\pi}^{\pi} \delta \textbf{v} \, \cos(m \phi) \, \cos(nz) \, {\rm d}\phi \, {\rm d}z$$ and are used to compute the displacement amplitude for each individual pair of azimuthal $m$ and vertical $n$ modes, $\textrm{v}\textrm{\_Amp}_{\textit{m,n}}(t)$, similar to the procedure applied in equation \[eq:amplitude\_mode\].
Case 1 – rotating diffusive star {#subsec:case1}
--------------------------------
We first simulated the effects of just rotation and just diffusion on their own, for the constant-Alfvén-frequency case, $p=1$. Rotation is parallel to the magnetic axis. We compare our results to those presented by @Braithwaite06b and show that we obtained similar results.
The rotation-only case has a threshold for the rotation rate equal to the Alfvén frequency, $\Omega_{\parallel}=\omega_{{\rm A},\phi}$. Rotation rates faster than that completely suppress the Tayler instability, presumably imposing some epicyclic motion of the perturbed fluid elements around an equilibrium point.
![Amplitude $\delta B(t)$ of the $m=1$, $n=1$ mode as a function of time for various rotation rates. Blue lines correspond to rotation rates below the Alfvén frequency, and red lines correspond to rotation rates above the Alfvén frequency. []{data-label="fig:amp_vs_omega_p1"}](amp_rotating_p1.eps)
Figure \[fig:amp\_vs\_omega\_p1\] shows the amplitude of the $m=1$, $n=1$ mode in simulations with different $\Omega_{\parallel}$, ranging from $\Omega_{\parallel}/\omega_{\textrm{A},\phi} = 0 \, - \, 10$. It is clearly observed how the system is unstable for $\Omega_{\parallel} / \omega_{\textrm{A},\phi} < 1$ (blue lines), and is stabilized for rotation rates $\Omega_{\parallel} / \omega_{\textrm{A},\phi}>1$ (red lines).\
In contrast, the diffusion-only case shows no threshold beyond which the Tayler instability is completely stabilized.
![Growth rate of the Tayler instability for the $m=1$ and $n=1$ mode as a function of the relative magnetic diffusivity $\eta/\eta_{\rm crit}$ measured from the velocity field, $v\_{Amp}_{m, n}(t)$.[]{data-label="fig:sigma_vs_eta"}](sigma_eta_algo.eps)
Figure \[fig:sigma\_vs\_eta\] shows the measured unstable growth rates in terms of $n_{\eta}$. It is observed that even at diffusivities as high as $n_{\eta}=50$; unstable growth is still present.
For the simultaneous action of both rotation and magnetic diffusion a total of 56 simulations were conducted. Figure \[fig:omega\_eta\] summarizes the measured growth rates normalized to the Alfvén frequency, $\sigma/\omega_{\rm{A},\phi}$.
![Contours of mean growth rate plotted as a function of normalized rotation rate $n_{\Omega_{\parallel}} = \Omega_{\parallel}/\omega_{{\rm A},\phi}$ and the relative magnetic diffusion $n_{\eta} = \eta/\eta_{\textrm{crit}}$. The white region represents stability, for at least $100 \tau_{A}$. The color contours correspond to increasing growth rates of the instability, where black corresponds to small and blue to fast growth rates. The yellow dashed line corresponds to the relation $n_{\Omega_{\parallel}} = 2n_{\eta} + 1$ separating the stable and unstable regions.[]{data-label="fig:omega_eta"}](omega_eta_copy.eps)
We consider a system to be stable if no unstable growth is observed for a hundred Alfvén crossing timescales, $100 \tau_A$, corresponding to the white region in figure \[fig:omega\_eta\]. We see that at very low magnetic diffusivities, left side of the figure, the system behaves similarly to the non-diffusive case, the system is stable for rotation rates $\Omega_{\parallel} > \omega_{A,\phi} $. Now if we move from the stable region in a straight horizontal line towards the right side of the figure we go towards higher magnetic diffusive environments and penetrate a region of unstable growth. This is a very interesting result as one would [*prima facie*]{} expect that two stabilizing processes, such as rotation and magnetic diffusion, to collaborate in suppressing the instability, but instead the presence of magnetic diffusion disturbs the stabilization introduced by rotation.
Presumably this happens because the Coriolis force pushes a fluid element onto an epicyclic equilibrium motion; magnetic diffusion constantly removes energy from the magnetic field such that there is no longer sufficient energy to return the fluid element to an equilibrium epicycle, and it will slowly spiral outwards.
The dashed yellow line in figure \[fig:omega\_eta\] is a relation between $n_{\Omega_{\parallel}}$ and $n_{\eta}$. If we assume that the boundary dividing the stable and unstable regimes can be separated by a linear relation we obtain $$n_{\Omega_{\parallel}} \approx 2n_{\eta} + 1.
\label{eq:stabilityrelation}$$ It is not clear why the threshold should be here. The factor of 2 may come from the 2 in the Coriolis force itself.
Alternatively the stability condition can be expressed as $$\omega_{A,\phi} < \Omega_{\parallel} - \eta k^2.
\label{eq:stabilityconditions}$$ This is the stability condition for fast rotating, diffusive systems as a function of the perturbed wavenumber $k$. The growth rate of the instability is in general a function of the Alfvén frequency, rotation rate, and magnetic diffusion rate, $\sigma=\sigma(\omega_{\rm{A}},\Omega_{\parallel},\eta\it{k}^2)$ for the non-stratified case.
Case 2 – perpendicular rotator {#subsec:case2}
------------------------------
We simulated a rotator perpendicular to the magnetic field axis of symmetry. The magnetic field is oriented in the $z$-axis, $\hat{\bf B} = \hat{e}_{z}$, and the rotation axis is oriented along the $\varpi(\phi=0)$-direction ${\bm\Omega}=\Omega_\perp(\phi=0){\bf e}_{\varpi}$. As in the previous section, the field gradient is $p=1$. We expect rotation to have a different effect on different vertical wavelengths, as given by equation \[eq:critical\_rotation\_oblique\]. Figure \[fig:oblique\_mn\] shows the growth rates of the first three vertical $n$ modes (all $m=1$) as a function of the rotation rate normalized to the critical rotation for each $n$ mode. We can observe how the growth rate of the instability is damped with increasing rotation rate and furthermore how this damping affects different vertical modes. It is however important to remember that the horizontal axis of the plot is normalized to the critical rotation rate for each vertical mode. This means that given a rotation rate of e.g. $\Omega_{\perp} = 4 \Omega_{\perp,{\rm crit}}(n=1) = 20 \omega_{{\rm A},\phi}$, the $n=1$ mode is damped by $\sim 90\%$, $n=2$ mode by $\sim 45\%$, and $n=3$ mode by $\sim 25\%$. In agreement with our prediction that perpendicular rotation damps different vertical modes with a different strength.
![Growth rate of the Tayler instability, normalized to the Alfvén frequency, as a function of the perpendicular rotation rate, normalized to the predicted critical rotation rate for each $n$ mode. The lines approximately on top of each other show the dependence of the damping on wavenumber is as predicted.[]{data-label="fig:oblique_mn"}](oblique_mn.eps)
Even if we did not see complete stability in our simulation, figure \[fig:oblique\_mn\] suggests that there is stabilization for each vertical mode at a rotation rates of $\Omega_{\perp} \approx 5 \Omega_{\perp,{\rm crit}}(n)$. Apparently our stability condition (\[eq:critical\_rotation\_oblique\]) is missing a factor of order unity. This is not surprising, given that the damping effect should depend on the precise arrangement of the displacements within the mode.
Case 3 – quadratic radial field gradient {#subsec:case3}
----------------------------------------
We saw above that in the case where $p=1$, rotation can completely stabilize the magnetic field, at least where there is no diffusion. We now investigate the case with a radial field gradient of $p=2$, which is predicted to behave differently. The Alfvén frequency grows proportional to distance from the axis $\omega_{{\rm A},\phi}(\varpi)\propto\varpi$. For this reason we expect that the instability grows faster at larger radii than close to the axis. We introduce a new quantity, $\delta B\textrm{\_Amp}_{m, n}(\varpi,t)$, corresponding to the displacement amplitude, for a given pair of azimuthal and vertical modes, as a function of time and distance to the axis $\varpi$: $$\begin{aligned}
\delta B\textit{\_Amp}_{m, n}(\varpi,\textrm{t})=\left( \frac{{\bf C}_{m,n}^{*}(\varpi, t){\bf C}_{m,n}(\varpi, t) }{\varpi^2 L_z \pi}\right)^{1/2}.
\label{eq:amplitude_mode_wavenumber_varpi}\end{aligned}$$ This new quantity allows us to study the growth rate of the instability at different distances from the axis.
![Growth rate $\sigma$ of the Tayler instability as a function of the solid-body rotation rate $\Omega_{\parallel}$, in the $p=2$ case. Normalized local growth rates are given for three different distances from the axis, $\sigma(\varpi)$. The dashed line corresponds to the predicted growth rate of the Tayler instability for slow rotators, $\sigma = \omega_{A,\phi}$. The dotted-dashed line is the predicted behavior of the instability for fast rotators, $\sigma = \omega_{A,\phi}^2/2\Omega_{\parallel}$.[]{data-label="fig:p2_fit_sigma_omega"}](p2_fit_sigma_omega.eps)
Local growth rates, $\sigma(\varpi)$, are computed from the local amplitude displacements, $\delta B{\rm \_Amp}_{m, n}(\varpi,t)$. This quantity is then divided by the local Alfvén frequency, $\omega_{{\rm A},\phi}(\varpi)$, and the normalized local growth rate is obtained, $\sigma(\varpi)/\omega_{{\rm A},\phi}(\varpi)$.
Figure \[fig:p2\_fit\_sigma\_omega\] shows the behavior of these normalized local growth rates as a function of rotation rate. The first main feature of the figure is that even for the very fast rotators, instability is still present. This is in agreement with the prediction that instability is always present in fast rotating stars containing a purely toroidal field with a radial field gradient $p \gtrsim 3/2$, obtained from the stability conditions (\[eq:stability\_rotating\]).
There are two behaviors observed in this plot separated by a critical rotation rate at $\Omega_{\parallel}\approx\omega_{\rm{A},\phi}$. In the slow rotation regime where $\Omega_{\parallel}/\omega_{{\rm A},\phi}\lesssim1$, the growth rate $\sigma\approx\omega_{\rm{A},\phi}$, whereas in the fast rotation regime where $\Omega_{\parallel}/\omega_{\rm{A},\phi}\gtrsim1$, we have instead $\sigma\approx\omega_{\rm{A},\phi}^2/2\Omega_{\parallel}$. Which is exactly what and @Spruit99 predicted a couple of decades ago.[^5]
Case 4 – mixed poloidal-toroidal field {#subsec:case4}
--------------------------------------
This section is intended to shed some light in the stability of mixed poloidal-toroidal magnetic field configurations in radiative stellar interiors. We focus our analysis on the behavior of the instability known to be present in purely toroidal fields, but now in mixed poloidal-toroidal field configurations. As discussed in section \[subsec:mixed\_field\_effects\] a poloidal field could damp, and if strong enough probably be able to suppress, the instability of a toroidal field. Tayler’s (1973) stability conditions however are not valid for mixed field configurations.
We note that the results presented in this section should be reviewed carefully because of the appearance of an unknown, spurious mode in the simulations growing always at the smallest scale in the vertical direction[^6]. Such a mode was absent in purely poloidal and purely toroidal field simulations. We work under the hypothesis that any well resolved unstable mode growing faster than this unknown perturbation is real, and relevant for our analysis. The nature of the unknown mode will be studied in detail in a forthcoming paper and it must be established whether it is astrophysically relevant.\
We first introduce some important definitions to avoid misunderstandings; the azimuthal Alfvén frequency, $\omega_{{\rm A},\phi} = \textrm{v}_{{\rm A},\phi}/\varpi$, is calculated with the toroidal Alfvén velocity. In the initial conditions, the radial field gradient of the toroidal field is chosen for simplicity to grow linearly with the distance from the axis, $p=1$, therefore the Alfvén frequency is constant throughout the volume of interest. The vertical Alfvén frequency $\omega_{\rm{A},z}$ is calculated with the vertical Alfvén velocity, $\textrm{v}_{\rm{A},z}$ and is given by $$\omega_{{\rm A},z} = \frac{\textrm{v}_{{\rm A},z}}{L_z} = \frac{B_z}{L_z\sqrt{4\pi\rho}}.$$ The poloidal field strength is constant everywhere, $B_z = B_0$, thus the vertical Alfvén velocity and the vertical Alfvén frequency are constant.\
For this field configuration field lines follow spirals as shown in figure \[fig:schematic\_view\]b. The number of spirals a field line completes in the box depends on the relative strengths of the toroidal and poloidal fields. The more complete spirals a field line draws, the stronger the toroidal field is. We call this number of spirals in the box the pitch–number, $P_n$, which is given by $$P_n = \frac{\omega_{{\rm A},\phi}}{2\pi\omega_{{\rm A},z}};$$ re-writing the instability condition (\[eq:poloidal\_stability\]) in terms of the pitch–number and solving for the $\textit{n}=1$ mode, we obtain that instability is expected to be present for $P_n > 1$. We note that $P_n$ is only independent of $\varpi$ in the special case $p=1$, which is what we use here.
![Growth rate of the first two azimuthal $\textit{m}$ modes as a function of the pitch–number. The solid black line corresponds to the $\textit{m}=0$ mode, stable at very high and low pitch–numbers but with a peak in unstable growth near the critical pitch–number, $P_n \approx 1$. The red dashed line corresponds to the $\textit{m}=1$ mode, unstable for very high pitch–numbers and stabilized near the critical pitch–number, $P_{n} \approx 1$.[]{data-label="fig:Bz_sigma"}](Bz_sigma_m0_2.eps)
Figure \[fig:Bz\_sigma\] shows the behavior of the first two azimuthal modes, $m=0$ and $m=1$, as a function of pitch–number. The $m=1$ mode has a growth rate on the order of the Alfvén frequency for a predominantly toroidal field configuration, i.e. large pitch–number, and drops suddenly when the pitch–number reaches a critical value. This drop of the growth rate agrees with the stability conditions (\[eq:poloidal\_stability\]). We note that no complete suppression of the instability is observed, even for very small pitch–numbers. However, these small growth rates are the effect of non-linear interaction with the spurious modes.
The second interesting mode shown in figure \[fig:Bz\_sigma\] is the $m=0$ mode. This mode was marginally stable for the purely toroidal field (\[eq:stability\_non-rotating\]) and it is being perturbed in the mixed configuration. Growth of the $m=0$ mode begins to appear with increasing poloidal field strength, reaching a maximum near the critical poloidal-to-toroidal field ratio, $P_n\approx1$.
The presence of this unstable mode, not predicted analytically, might be important for the hydromagnetic dynamo theory, given that this instability triggers axisymmetric vertical motions which it is often claimed are necessary for closure of the dynamo loop. We will investigate the importance of this mode in a future paper.
Discussion and summary {#sec:conclusions}
======================
Solid-body rotation, magnetic diffusion and field configuration are expected to have a strong influence on the presence and evolution of hydro-magnetic instabilities in toroidal fields in radiative stellar interiors. We present a set of 3D MHD simulations of purely toroidal and of mixed poloidal-toroidal magnetic fields in non-convective stellar interiors. A combination of various radial field strength gradients ($p={\rm d}\ln B/{\rm d}\ln \varpi$), solid body rotation rates ($\Omega$), magnetic diffusivities ($\eta$), and poloidal-to-toroidal field strengths ($P_n$) is scoped.
We observe the presence of the kink type $m = 1$ Tayler instability in purely toroidal fields. We measure its growth rate, and compare our results with the various analytical results for non-rotating and non-diffusive [@Tayler73], fast rotating , highly diffusive [@Spruit99], and rotating-diffusive radiative stellar interiors. We also study the development of instabilities in a mixed poloidal-toroidal field configuration.\
Magnetic diffusion is not able to suppress the occurrence of the instability even at high diffusivity, in accordance with the results of @Braithwaite06a.
[**Diffusion and rotation parallel to magnetic axis:**]{} If the magnetic and rotation axes are parallel, fast rotation ($\Omega_{\parallel}>\omega_{{\rm A},\phi}$) is able to stabilize the configuration (see figure \[fig:amp\_vs\_omega\_p1\] ), where $\omega_{{\rm A},\phi}$ is the azimuthal Alfvén frequency. This is in the case where the field gradient $p=1$. However as soon as magnetic diffusion is included, the stability conditions now depend on a combination of the rotation rate, the magnetic field strength and the magnetic diffusion rate given by equation \[eq:stabilityconditions\], Which suggests that very fast rotation can stabilize the system only if the magnetic diffusion rate does not subtract energy fast enough such that Coriolis force is unable maintain the fluid elements in stable epicyclic motions. [**Perpendicular rotation:**]{} If the magnetic field axis and the rotation axis are perpendicular, the damping strength introduced by rotation is different for different vertical wavelengths. The damping effect is stronger for small vertical wavenumbers $k$ compared to large wavenumbers because of the larger deformation occurring perpendicular to the rotation axis at smaller wavenumbers. It is found that the instability is suppressed if $\Omega_{\perp}\gtrsim 5\omega_{{\rm A},\phi}\varpi_0k$, where $\varpi_0$ is the length scale of the motions in the horizontal direction.\
[**Radial field gradient:**]{} To analyze the effects of the radial field gradient and rotation in the instability we investigated two cases, the linear radial field gradient, $p=1$ and the quadratic radial field gradient, $p=2$. As described above, in the $p=1$ case (without diffusion) stability is reached for fast rotation (meaning faster than the azimuthal Alfvén frequency). In this $p=2$ case, very fast rotation is never able to stabilize the configuration. Linear growth rates behave broadly as predicted by , where for slow rotators the growth rate is roughly equal to the Alfvén frequency, whilst for fast rotators, the growth rate is damped by a factor of $\omega_{\rm{A},\phi}/2\Omega_{\parallel}$. It is probably safe to assume that this is also the case for all values of $p$ above $3/2$, as originally predicted. We have therefore provided a check on the theory, useful in light of the fact that the originators of the analytic results were, we think it is fair to say, less than 100% certain they were correct.
[**Mixed poloidal-toroidal field:**]{} Finally we simulated a mixed poloidal-toroidal field, and in particular the stabilizing influence of a uniform poloidal field. In the $p=1$ case, where in the absence of a poloidal field, only the $m=1$ mode is unstable, we find that the $m=1$ mode is indeed suppressed by the poloidal field, and that the stability criterion is where it is predicted to be. Interestingly, we also find that the $m=0$ mode is unstable in the vicinity of this threshold.\
The presence of the unstable axisymmetric $m=0$ mode in the mixed poloidal-toroidal field configuration may be quite important. In the case where the field gradient is such that it is marginally stable in the purely toroidal case, it becomes dominant in the vicinity of the threshold of stabilization by the poloidal field, which of course is precisely where the dynamo is likely to be. It opens the possibility for axisymmetric turbulent radial motions which affect the mixing and the regeneration of the poloidal field. It is often claimed by various members of the MHD community that a poloidal field cannot be renewed without the $m=0$ mode; whether this claim is correct is, if the $m=0$ mode is present and perhaps even dominant, no longer relevant.
Indeed, as far as the field gradient $p$ is concerned, in the fast-rotating (which is expected to occur much more frequently in nature than the slowly rotating) adiabatic case we can summaries behavior in terms of the threshold: where $p<1$ all modes are stable; where $1<p<3/2$ just the $m=0$ mode is unstable; where $p>3/2$ the $m=1$ mode is also unstable, with thresholds at higher values of $p$ introducing instability at progressively higher values of $m$. However, these thresholds are probably of little practical importance. Whilst it is often said that the Tayler instability is dominated by the $m=1$ mode because it is that mode which is in some sense the first to become unstable, this is misleading: only in the rather special circumstance of $-1/2<p<1$ and $\Omega_{\parallel}<\omega_{\rm A}$ is the $m=1$ mode the only one present. Firstly, the second of those two conditions is rarely fulfilled in nature. Secondly, in the context of the fluctuating field envisaged in the Tayler-Spruit dynamo model, the field gradient $p$ will have values all the way from $p=-\infty$ to $p=+\infty$. It seems perfectly plausible that the dynamo could be dominated by the $m=0$ mode, or indeed by $m=20$. Roughly speaking, the instability is present where $p$ is positive and absent where it is negative: the dynamo can therefore be expected to be intermittent in some sense.
The authors would like to thank Alfio Bonanno and Henk Spruit for fruitful discussions. The analysis and plots presented here were performed using VisIt (Visualization Tool) and Ipython notebook, and the bibliography was put together thanks to the Astrophysical Database ADS and Mendeley.
[^1]: [email protected]
[^2]: There have been claims that this Tayler-Spruit mechanism cannot work because of the lack of an axisymmetric mode; it turns out, however, that this is both misleading and untrue because (a) there is indeed a strong axisymmetric mode, as explained below, and (b) there is no convincing argument that an axisymmetric mode is necessary.
[^3]: For a purely toroidal field the toroidal and total (that is, replacing $B_\phi$ by $B$ in the definition) Alfvén frequencies are the same; however for mixed poloidal-toroidal fields this distinction is necessary.
[^4]: Up to 8 or 9 orders of magnitude assuming a $1 \rm{G}$ toroidal field in the Sun
[^5]: Except that they neglected the factor of 2. At a latitude of $30^{o}$ this factor will go away anyway since the component of $\bm{\Omega}$ perpendicular to the spherical surfaces in which the fluid is moving around is half as much as at the poles.
[^6]: Despite the many vertical resolutions tested: 8, 16, 32, 64 and 96 these perturbations, with vertical wavelength corresponding to that of the Nyquist frequency, was present in the $m = 1$ mode with a growth rate $\sigma_{\rm{unk}} \approx \omega_{\rm A}/2$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate an interfacial spin-transfer torque and $\beta$-term torque with alternating current (AC) parallel to a magnetic interface. We find that both torques are resonantly enhanced as the AC frequency approaches to the exchange splitting energy. We show that this resonance allows us to estimate directly the interfacial exchange interaction strength from the domain wall motion. We also find that the $\beta$-term includes an unconventional contribution which is proportional to the time derivative of the current and exists even in absence of any spin relaxation processes.'
author:
- Junji Fujimoto
- Mamoru Matsuo
bibliography:
- 'reference.bib'
title: 'Alternating Current-induced Interfacial Spin-transfer Torque'
---
#### Introduction.—
A variety of physical phenomena arises near interfaces, such as spin-dependent transports [@baibich1988; @valet1993; @miyazaki1995; @moodera1995; @gould2004; @miwa2017], interfacial magnetic phenomena [@carcia1985; @camley1989; @engel1991; @fert1980; @levy1981; @crepieux1998a], and chiral/topological phenomena [@bode2007; @ferriani2008; @meckler2009; @heinze2011], which have attracted attention from many years ago [@hellman2017]. Among these, the spin-dependent transport has been closely related to the aspect of not only fundamental physics but device application; especially the tunneling magnetoresistance [@valet1993; @miyazaki1995; @moodera1995] impacted upon the invention of the magnetoresistive random access memory [@zhu2006].
The spin-dependent transports near the interfaces are important from the viewpoint of the understanding of recent developments in spintronics, such as the spin pumping effect (SPE) [@mizukami2001; @tserkovnyak2002; @simanek2003; @tserkovnyak2004; @takahashi2010; @ohnuma2014; @tatara2017] and the spin Seebeck effect (SSE) [@uchida2008; @adachi2011], because the mutual dependence between the magnetization dynamics and the spin-dependent transports is the key mechanism in various spin-dependent phenomena. The two effects are the ways of generating spin currents without electric currents, in a bilayer system consisting of a ferromagnet (FM) and a normal metal (NM); the spin precession due to the rf microwave in FM induces the spin current in NM in the case of SPE, and the temperature difference between FM and NM induces that for SSE. Both of effects can be described by the tunnel Hamiltonian method [@adachi2011; @ohnuma2014; @matsuo2018], which also captures tunneling magnetoresistance.
The interfacial exchange interaction between conduction electrons in NM and magnetization in FM plays a crucial role in SPE and SSE, which are proportional to $J_{sd}^2$, where $J_{sd}$ is the interfacial exchange interaction strength [@adachi2011; @ohnuma2014]. In general, the exchange interaction possibly gives rise to an essential contribution to spin-related phenomena near the interfaces, such as the spin Hall magnetoresistance [@nakayama2013]. However, this physically essential parameter $J_{sd}$ has not been directly measured, and the direct method of evaluating it is not yet proposed.
 Schematic description of the configuration, where a ferromagnet (FM) having two magnetic domains with one domain wall (DW) is adjoined by a normal metal (NM) whose conduction electron couples with the magnetization in FM through the interfacial exchange interaction. The alternating current (AC) is applied parallel to the interface due to the electric field with finite frequency. The AC-induced spin torques on the DW leads to the oscillation of the position of the DW center, which allows us to evaluate the interfacial exchange interaction strength. (b) DW configuration, which is described by the corrective coordinates of the DW center $X$ and the angle $\phi_0$. ](setup.pdf){width="0.9\linewidth"}
In this Letter, we present a direct method of evaluating the interfacial exchange interaction strength $J_{sd}$ from the domain wall dynamics in FM adjoined by NM, applying an alternating current (AC) parallel to the interface (Fig. \[fig:1\] (a)). It is a well-known fact that in bulk ferromagnetic metals with noncolinear magnetic textures such as domain walls, the direct current (DC) accompanying with spin polarization exerts spin torques on the magnetization, which leads to its dynamics such as the domain wall motion [@berger1984; @berger1992; @salhi1993; @yamaguchi2004; @zhang2004; @tatara2008a]. We here extend the DC-induced spin torques into the region of an arbitrary frequency of the current, based on the quantum field theoretical approach, and apply this to the interfacial exchange interacting system of the FM-NM bilayer. We consider so thin NM that we focus only on the spin polarized electronic states near the interface due to the interfacial exchange interaction. We find that the AC-induced spin torques consist of corresponding extensions of the spin-transfer torque [@slonczewski1996; @bazaliy1998; @tatara2004; @zhang2004] and the so-called $\beta$-term torque [@thiaville2004; @zhang2004; @tserkovnyak2006; @duine2007; @kohno2007]. However, we also find that the results we obtain include physically a novel contribution to the $\beta$-term torque, which depends on the time derivative of the current density. Our important finding is that both spin torques are proportional to $(1 - \omega^2 \tau_{\mathrm{sd}}^2)^{-1}$ for the case of no spin relaxation processes, where $\omega$ is the AC frequency and $\tau_{\mathrm{sd}} = \hbar / 2 \Delta$ with $2 \Delta$ being the interfacial exchange splitting. The exchange splitting is related to $J_{sd}$ by $\Delta = S J_{sd}$, where $S$ is the localized spin length constructing the magnetization. This dependence suggests that we can evaluate $J_{sd}$ from the magnetization dynamics driven by the spin torques. In the viewpoint of application, the enhancement of the spin torques have an advantage in that less current density is needed to excite the magnetization dynamics.
We then solve the equation of motion of a rigid domain wall (DW) [@tatara2004; @tatara2008a] driven by the obtained spin torques, in the presence of a spin relaxation process. The equation is expressed by the two collective coordinates; the position of the DW center $X$ and the angle $\phi_0$ (Fig. \[fig:1\] (b)), and the spin torques act as the forces to $X$ and $\phi_0$. We find that $X$ and $\phi_0$ oscillate along with the frequency $\omega$ in the region of the small electric current density, and the amplitude of the oscillation of $X$ increases resonantly near $\omega \tau_{sd} \simeq 1$. Hence, we conclude that the dependence of $X$ on the frequency allows us to estimate the interfacial exchange splitting.
This Letter is organized as follows. We first present the total Lagrangian of the magnetization in FM and the conduction electron in NM as well as their interfacial exchange interaction, and introduce the *rotated frame picture* sometimes used in the context of the ferromagnetic spintronics. Then, the AC-induced spin torques are evaluated based on the linear response theory with the thermal Green function method. As an application, we consider the DW dynamics driven by the obtained spin torques.
#### Theory.—
The total Lagrangian that we consider is given by $\mathcal{L} = \mathcal{L}_{m} + \mathcal{L}_{e} - \mathcal{H}_{sd}$, where $\mathcal{L}_{m}$ is the Lagrangian of the magnetization in the FM layer, $\mathcal{L}_{e}$ is that of the conduction electron in the NM layer, and $\mathcal{H}_{sd}$ is the $sd$-like interfacial exchange interaction between them.
Considering that the magnetization is constructed by the localized spins ordering, we express the Lagrangian of the magnetization as that of the localized spin, $\bm{M} = - M_S \bm{m}$ with $\bm{m} = (\sin \theta \cos \phi, \sin \theta \sin \phi, \cos \theta)$, where $M_S$ is the saturated magnetization, and $\theta = \theta (\br, t)$, $\phi = \phi (\br, t)$. Here, $\bm{m}$ does not represent the unit vector of the magnetization, but that of the localized spin, whose signs are opposite. The Lagrangian of the localized spin is defined as $\mathcal{L}_{m} = \int \mathrm{d} \br \,(\hbar S / a^3) \dot{\phi} (\cos \theta - 1) - \mathcal{H}_m$ with $$\begin{aligned}
\mathcal{H}_m
& = \int \frac{\mathrm{d} \br}{a^3} \left[ \frac{ J_{\mathrm{ex}} }{2} S^2 \left( \bm{\nabla} \bm{m} \right)^2 - \frac{K}{2} S^2 m_z^2 + \frac{K_{\perp}}{2} S^2 m_y^2 \right]
,\end{aligned}$$ where $a$ is the lattice constant of FM, $J_{\mathrm{ex}}$ is the exchange interaction between the localized spins, and $K$ and $K_{\perp}$ are easy- and hard-axis magnetic anisotropies, respectively. Note that the saturated magnetization $M_S$ is related to the localized spin length by $M_S = \gamma_e \hbar S / a^3$ with the gyromagnetic ratio $\gamma_{\mathrm{e}}$, and $J_{\mathrm{ex}}$, $K$, and $K_{\perp}$ are all positive.
We show the rest of the Lagrangian, which is written by $\mathcal{L}_{e} - \mathcal{H}_{sd} = \int \mathrm{d} \bm{r} \psi^{\dagger} (\br, t) ( \zi \hbar \partial_t - H_e - H_{sd} ) \psi^{} (\br, t)$, where $\psi^{(\dagger)}$ is the field operator of electrons, $H_e = \bp^2/2 \me + V$ describes the kinetic energy with the electron mass $\me$ and the nonmagnetic and magnetic impurity potentials given by $V = u_{\mathrm{i}} \sum_{i=1}^{N_{\mathrm{i}} } \delta (\br - \bm{R}_i) + u_{\mathrm{s}} \sum_{j=1}^{N_{\mathrm{s}} } ( \bm{S}_{\mathrm{imp}, j} \cdot \bm{\sigma} ) \delta (\br - \bm{R}'_j)$ with the impurity numbers $N_{\mathrm{i}}$ and $N_{\mathrm{s}}$ and with the strengths $u_{\mathrm{i}}$ and $u_{\mathrm{s}}$, and $H_{sd} = - \Delta \bm{m} (\br,t) \cdot \bm{\sigma}$ represents the interfacial exchange interaction with the coupling constant $\Delta > 0 $ with $\bm{\sigma} = (\sigma^x, \sigma^y, \sigma^z)$ being the Pauli matrices. The magnetic impurity spin $\bm{S}_{\mathrm{imp},j}$ is assumed to be quenched.
Then, we transform the Hamiltonian into the ‘rotated frame’ [@korenman1977; @tatara2008a] by using the unitary transformation $U (\br, t)$ defined by $U^{\dagger} (\br, t) \left( \bm{m} (\br, t) \cdot \bm{\sigma} \right) U (\br, t) = \sigma^z$ with $\bar{\psi} = U^{\dagger} \psi$. The physical meaning of the unitary transformation is that the quantization axis of the electron spin is to be reoriented to $\bm{m} (\br, t)$ at each position and time. Hence, we call the frame after the transformation as the rotated frame and denote $\bar{A}$ as the quantity $A$ in the rotated frame. The electron described by $\bar{\psi}^{(\dagger)}$ feels the uniform exchange interaction in the rotated frame. We also express the rotational unitary transformation by using the rotational matrix $\mathcal{R} (\br, t)$ for the three-dimensional vector defined by $U^{\dagger} (\br, t) \, \bm{\sigma} \, U (\br, t) = \mathcal{R} (\br, t) \bm{\sigma}$. This expression of the unitary transformation is useful for the magnetic impurity potential and the spin torques. Note that the relation to the definition of $U$ is $U^{\dagger} (\bm{m} \cdot \bm{\sigma} ) U = \bm{m} \cdot (\mathcal{R} \bm{\sigma}) = (\mathcal{R}^{-1} \bm{m}) \cdot \bm{\sigma} = \sigma^z$, hence $\mathcal{R}^{-1} \bm{m} = \hat{z}$, where $\hat{z}$ is the unit vector along the $z$-axis.
We now look into the equation of motion of the localized spin, which is obtained from the Euler-Lagrange equation with the relaxation function $\mathcal{W}$ [@tatara2008a], $$\begin{aligned}
\frac{ \mathrm{d} }{ \mathrm{d} t } \left( \frac{\delta \mathcal{L}}{\delta \dot{q}} \right) - \frac{\delta \mathcal{L}}{\delta q}
& = - \frac{\delta \mathcal{W}}{\delta \dot{q}}
\label{eq:EL}
,\end{aligned}$$ where $q \in \{\theta, \phi\}$, and $\mathcal{W} = \int \mathrm{d} \br\, (\hbar S \alpha_{\mathrm{G}}/2 a^3) \dot{\bm{m}}^2$ with the Gilbert damping constant $\alpha_{\mathrm{G}}$. Equation (\[eq:EL\]) leads to the Landau-Lifshitz-Gilbert equation, $\dot{\bm{m}}
= \gamma_{\mathrm{e}} \bm{H}_{\mathrm{eff}} \times \bm{m}
+ \alpha_{\mathrm{G}} \dot{ \bm{m} } \times \bm{m}
+ \bm{\tau}_{e}
,$ where $\bm{H}_{\mathrm{eff}}$ is the effective magnetic field defined as $\gamma_{\mathrm{e}} \bm{H}_{\mathrm{eff}} = (1/\hbar S)\, \delta \mathcal{H}_m / \delta \bm{m}$, and $\bm{\tau}_e$ is the spin torque through the interfacial exchange interaction; $$\begin{aligned}
&\gamma_{\mathrm{e}} \bm{H}_{sd}
= \frac{1}{\hbar S} \left\langle \frac{\delta \mathcal{H}_{sd}}{\delta \bm{m}} \right\rangle_{\mathrm{neq}}
, \\
&\bm{\tau}_{e}
= \gamma_{\mathrm{e}} \bm{H}_{sd} \times \bm{m}
= - \frac{\Delta}{\hbar S} \left\langle\bm{s}\right\rangle_{\mathrm{neq}} \times \bm{m}
\label{eq:tau_e}
.\end{aligned}$$ Here, $\bm{s} = \bm{s} (\br, t) = \psi^{\dagger} (\br, t) \bm{\sigma} \psi (\br, t)$ is the spin density operator divided by $\hbar / 2$, and $\langle\,\cdots\rangle_{\mathrm{neq}}$ describes the statistical average in the nonequilibrium.
The spin torque is expressed in the rotated frame as $\bar{\bm{\tau}}_{e} = \mathcal{R}^{-1} \bm{\tau}_{e} = - (\Delta/\hbar S) \langle \bar{\bm{s}} \rangle_{\mathrm{neq}} \times \hat{z}$. We emphasise that, in the rotated frame, the perpendicular components of the nonequilibrium spin polarization $\langle \bar{\bm{s}} \rangle_{\mathrm{neq}}$ to the $\hat{z}$-axis only act as torques.
In this Letter, we evaluate the nonequilibrium spin polarization $\langle \bar{\bm{s}} \rangle_{\mathrm{neq}}$ in the linear response to the electric field $\bm{E} (\omega)$ with the frequency $\omega$ as $$\begin{aligned}
\langle \bar{s}^{\alpha} (\bq, \omega) \rangle_{\mathrm{neq}}
& = \bar{\chi}^{\alpha i} (\bq, \omega) E_i (\omega)
\label{eq:spin_in_rotated_frame}\end{aligned}$$ in the Fourier space ($\alpha = x, y$ and $i = y, z$). From the linear response theory, we can obtain the response coefficient $\bar{\chi}^{\alpha i} (\bq, \omega)$ from $\bar{\chi}^{\alpha i} (\bq; \omega) = (\bar{K}^{\alpha i} (\bq; \omega) - \bar{K}^{\alpha i} (\bq; 0)) / \zi \omega$, where $\bar{K}^{\alpha i} (\bq; \omega)$ can be evaluated from the following spin-current correlation function in the Matsubara form $$\begin{aligned}
\bar{\mathcal{K}}^{\alpha i} (\bq; \zi \omega_{\lambda})
& = \int_0^{1/\kB T} \dd{\tau} e^{\zi \omega_{\lambda} \tau} \left\langle \mathrm{T}_{\tau} \bar{s}^{\alpha} (\bq, \tau) \bar{J}_i (0) \right\rangle_{\mathrm{eq}}
\label{def:correlation_in_Matsubara}\end{aligned}$$ through the analytic continuation $\zi \omega_{\lambda} \to \hbar \omega + \zi 0$; $\bar{K}^{\alpha i} (\bq; \omega) = \bar{\mathcal{K}}^{\alpha i} (\bq; \hbar \omega + \zi 0)$. Here $\omega_{\lambda} = 2 \pi \lambda \kB T$ is the Matsubara frequency of bosons with the temperature $T$ [@fujimoto2019], and the spin and the electric current operator in the rotated frame are given by $$\begin{aligned}
\bar{s}^{\alpha} (\bq)
& = \frac{1}{V} \sum_{\bk} \bar{c}^{\dagger}_{\bk-\bq} \sigma^{\alpha} \bar{c}^{}_{\bk}
\label{eq:spin}
,\end{aligned}$$ $$\begin{aligned}
\bar{J}_i
& = - e \sum_{\bk} \frac{\hbar k_i}{\me} \bar{c}^{\dagger}_{\bk} \bar{c}^{}_{\bk}
- \frac{e \hbar}{2 \me} \sum_{\bq'} A^{\beta}_{i} (\bq') \sum_{\bk} \bar{c}^{\dagger}_{\bk} \sigma^{\beta} \bar{c}^{}_{\bk-\bq'}
\label{eq:current}
,\end{aligned}$$ where $\bar{c}^{(\dagger)}_{\bk}$ is the Fourier transform of the field operator $\bar{\psi}^{(\dagger)} (\br)$. The first term of Eq. (\[eq:current\]) is the normal velocity term and the second is the anomalous velocity term due to the spin gauge field $A_i^{\beta} (\br) = - \zi \tr [ U^{\dagger} \partial_i U \sigma^{\beta} ]$ with $\beta = x, y, z$.
As the detailed calculation will be shown elsewhere, we here sketch out the procedures of the calculation. By substituting Eqs. (\[eq:spin\]) and (\[eq:current\]) into Eq. (\[def:correlation\_in\_Matsubara\]), we rewrite the correlation function by using the thermal Green functions according to Wick’s theorem. We expand the Green function by the spin gauge field up to the first order, and take the statistical average on the impurity positions, then we obtain $$\begin{aligned}
\bar{\mathcal{K}}^{\alpha i} (\bq; \zi \omega_{\lambda})
& = \frac{e \hbar}{4 \me V} A^{\beta}_{j}(\bq) \kB T \sum_{n} \sum_{\bk} \Phi_{i j}^{\alpha \beta} (\bk; \zi \epsilon_n^{+}, \zi \epsilon_n)\end{aligned}$$ where $\epsilon_n^{+} = \epsilon_n + \omega_{\lambda}$, $\epsilon_n = (2 n + 1) \kB T$ is the Matsubara frequency of fermions, and $$\begin{aligned}
\Phi_{i j}^{\alpha \beta} (\bk; \zi \epsilon_n^{+}, \zi \epsilon_n)
& = 2 \delta_{i j} \tr \left[
\Lambda^{\alpha} g_{\bk} (\zi \epsilon_n^{+}) \sigma^{\beta} g_{\bk} (\zi \epsilon_n)
\right]
\nonumber \\ & \hspace{0em}
- \frac{\hbar^2 k_i k_j}{\me}
\tr \left[
\Lambda^{\alpha} g_{\bk} (\zi \epsilon_n^{+}) \sigma^{\beta} g_{\bk} (\zi \epsilon_n^{+}) g_{\bk} (\zi \epsilon_n)
\right]
\nonumber\\ & \hspace{0em}
- \frac{\hbar^2 k_i k_j}{\me}
\tr \left[
\Lambda^{\alpha} g_{\bk} (\zi \epsilon_n^{+}) g_{\bk} (\zi \epsilon_n) \sigma^{\beta} g_{\bk} (\zi \epsilon_n)
\right]
\label{def:Phi}
.\end{aligned}$$ Here, $g_{\bk} (\zi \epsilon_n^{(+)}) = (\zi \epsilon_n^{(+)} + \mu - \hbar^2 k^2 / 2 \me - \Delta \sigma^z - \Sigma (\zi \epsilon_n^{(+)}) )^{-1}$ is the thermal Green function with the self energy within the self-consistent Born approximation $\Sigma ( \zi \epsilon_n) = n_{\mathrm{i}} u_{\mathrm{i}}^2 \sum_{\bk} g_{\bk} (\zi \epsilon_n) + n_{\mathrm{s}} u_s^2 S_{\mathrm{imp}}^2 \sum_{\bk} \sigma^{\gamma} g_{\bk} (\zi \epsilon_n) \sigma^{\gamma}$, where $n_{\mathrm{i}}$ and $n_{\mathrm{s}}$ are the impurity concentrations of nonmagnetic and magnetic impurities, respectively, and we have taken the statistical average on the impurity spins and assume the spherical spins, $\overline{S^{\alpha}_{\mathrm{imp}, i} S^{\beta}_{\mathrm{imp},j}} = (S_{\mathrm{imp}}^2/3) \delta_{i j} \delta^{\alpha \beta}$. In Eq. (\[def:Phi\]), we have evaluated $\Phi_{i j}^{\alpha \beta}$ by assuming $\bq = 0$ since $\bar{\mathcal{K}}^{\alpha i} (\bq; \zi \omega_{\lambda})$ is already in the $q$-linear order because of $A_i^{\beta} (\bq)$. The full vertex of spin $\Lambda^{\sigma} = \Lambda^{\sigma} (\zi \epsilon_n^{+}, \zi \epsilon_n)$ is given by $ \Lambda^{\alpha}
= \sigma^{\alpha}
+ n_{\mathrm{i}} u_{\mathrm{i}}^2 \sum_{\bk} g_{\bk} (\zi \epsilon_n^{+}) \Lambda^{\alpha} g_{\bk} (\zi \epsilon_n)
+ \frac{1}{3} n_{\mathrm{s}} u_{\mathrm{s}}^2 S_{\mathrm{imp}}^2\sum_{\bk} \sigma^{\gamma} g_{\bk} (\zi \epsilon_n^{+}) \Lambda^{\alpha} g_{\bk} (\zi \epsilon_n) \sigma^{\gamma}$. After some straightforward calculation and taking the analytic continuation with the assumption of $T = 0$, we then obtain $$\begin{aligned}
\bar{\chi}^{\alpha i} (\bq, \omega)
& = \frac{\hbar \sigma_{\mathrm{s}} }{e}A^{\beta}_i (\bq) \sum_{\sigma = \pm} \left( \delta^{\alpha \beta} + \zi \sigma \epsilon^{\alpha \beta z} \right)
\frac{ \sigma }{ 2 \sigma \Delta - \hbar \omega + \zi \hbar / \tau_{\mathrm{s}} }
\label{eq:result:chi}
,\end{aligned}$$ where we neglected the higher order contribution of $\hbar / \epsilon_{\mathrm{F} \sigma} \tau_{\sigma}$ with the spin-dependent Fermi energy $\epsilon_{\mathrm{F} \sigma}$ and the momentum lifetime $\tau_{\sigma}$. Here, $\sigma_{\mathrm{s}} = \sigma_{\mathrm{s}} (\omega)$ is the spin conductivity, $ \sigma_{\mathrm{s}} (\omega) = (e^2/\me) \sum_{\sigma = \pm} \sigma n_{\sigma} \tau_{\sigma} / (1 - \zi \omega \tau_{\sigma}) $, with the spin-dependent electron density $n_{\sigma}$ and lifetime $\tau_{\sigma}$ with $\sigma = \pm$, and $\tau_s$ is the relaxation time due to the magnetic impurity scattering defined as $\tau_{\mathrm{s}}^{-1} = (2 \pi / 3 \hbar) n_{\mathrm{s}} u_{\mathrm{s}}^2 S_{\mathrm{imp}}^2 \bigl( \nu_{+} +\nu_{-} \bigr)$, where $\nu_{\pm}$ is the density of states at the Fermi level.
#### Results.—
Here we show the expression of the AC-induced spin-transfer torque and $\beta$-term torque obtained from Eq. (\[eq:result:chi\]) combined with Eq. (\[eq:spin\_in\_rotated\_frame\]) and (\[eq:tau\_e\]), $$\begin{aligned}
\bm{\tau}_{e} (\br, \omega)
& =
\frac{
\bigl( \bm{j}_{\mathrm{s}} \cdot \bm{\nabla} \bigr) \bm{m}
+ (\zi \omega \tau_{sd} + \zeta_{\mathrm{s}}) \bm{m} \times \bigl( \bm{j}_{\mathrm{s}} \cdot \bm{\nabla} \bigr) \bm{m}
}{ 1 + (\zi \omega \tau_{sd} + \zeta_{\mathrm{s}})^2 }
\label{eq:result:torque}\end{aligned}$$ in the laboratory frame, where $\tau_{sd} = \hbar / 2 \Delta$, $\zeta_{\mathrm{s}} = \tau_{sd} / \tau_{\mathrm{s}}$, and we used $\mathcal{R} \bm{A}^{\perp}_i = - \bm{m} \times \partial_i \bm{m}$ and $\mathcal{R} (\hat{z} \times \bm{A}^{\perp}_i) = \partial_i \bm{m}$. The frequency-dependent spin current is denoted by $\bm{j}_{\mathrm{s}} = \bm{j}_{\mathrm{s}} (\omega) = (\hbar/2eS) \sigma_{\mathrm{s}} (\omega) \bm{E} (\omega)$. By taking the static field limit of $\omega \to 0$, we find that the first term proportional to $\bigl( \bm{j}_{\mathrm{s}} \cdot \bm{\nabla} \bigr) \bm{m}$ corresponds to the spin-transfer torque and the second term proportional to $\bm{m} \times \bigl( \bm{j}_{\mathrm{s}} \cdot \bm{\nabla} \bigr) \bm{m}$ coincides with the $\beta$-term torque, and we confirm that our result agrees with that by Zhang and Li [@zhang2004] and by Kohno and Shibata [@kohno2007] for the model of the conduction electron in a ferromagnet, although they are not for the interfacial exchange interaction like our situation. Hence, our result (\[eq:result:torque\]) is an extension of the DC-induced spin torques into the arbitrary frequency. Equation (\[eq:result:torque\]) is the main result of this Letter.
Considering the case of the dilute magnetic-impurity concentration, $\tau_{\mathrm{s}} \to \infty$, so that $\zeta_{\mathrm{s}} \to 0$, we find $$\begin{aligned}
\bm{\tau} (\br, \omega)
& = \frac{ 1 }{ 1 - \omega^2 \tau_{sd}^2 }
\left( ( \bm{j}_{\mathrm{s}} \cdot \bm{\nabla} ) \bm{m}
+ \zi \omega \tau_{sd} \bm{m} \times \bigl( \bm{j}_{\mathrm{s}} \cdot \bm{\nabla} \bigr) \bm{m} \right)
,\end{aligned}$$ which implies that the spin torques increase resonantly as the AC frequency approaches to the $1/\tau_{sd}$. As shown in *Application*, this frequency dependence allows us to determine the magnitude of the interfacial exchange interaction.
We also find that the $\beta$-term toque is present proportional to the frequency $\omega$, without the magnetic impurity which results in a spin relaxation process. The $\beta$-term torques are known to arise from the spin relaxation process [@zhang2004], such as the scattering due to the magnetic impurity potential [@kohno2006] and spin-orbit impurity potential [@tatara2008]. Actually, Eq. (\[eq:result:torque\]) shows that there is also the $\beta$-term proportional to the magnetic impurity concentration, $\zeta_{\mathrm{s}} \sim n_{\mathrm{s}}$. The $\beta$-term torques also arise from *nonadiabaticity*, which stands for the higher order of the derivatives, such as the terms proportional to $\bm{m} \times \partial_t \partial_i \bm{m}$ [@tserkovnyak2006; @thorwart2007]. Note that Eq. (\[eq:result:torque\]) is the Fourier form in the frequency space, so that in the real time space, the $\beta$-term torque is expressed as $\bm{m} \times \bigl( (\mathrm{d} \bm{j}_{\mathrm{s}} / \mathrm{d} t) \cdot \bm{\nabla} \bigr) \bm{m}$, which is the first order derivative for the magnetization, not higher orders. For these reasons, the $\beta$-term torque we obtain is different from the ones which are already known.
It should be discussed that the relation of the spin torques obtained here to the Rashba spin-orbit torques (SOT) [@manchon2008; @obata2008; @manchon2009; @kurebayashi2014] and the spin Hall torques (SHT) [@kim2013b; @garello2013; @pai2015]. Both Rashba SOT and SHT originate from the spin-orbit couplings (SOCs); Rashba SOT comes from the interfacial SOC due to the inversion symmetry breaking and SHT arises from the bulk SOC in NM. We have assumed that these SOCs are weak so that these torques do not contribute much; for instance, that is the case of Cu as a NM and Py as a FM. For the strong SOCs, we have to develop our theory which contains these strong SOCs, but that is out of our focus.
#### Application.—
Now, we focus on the domain wall (DW) motion as an application of the obtained torques. Following Tatara *et al.* [@tatara2008a] and assuming $K_{\perp} \ll K$ and no pinning potentials, we rewrite the Lagrangian $\mathcal{L}_m$ into that of DW, introducing the corrective coordinates of the DW center $X(t)$ and the angle $\phi_0(t)$ (Fig. \[fig:1\] (b)), $$\begin{aligned}
\bm{m}
& = \left(
\frac{\cos \phi_0 (t)}{\cosh \frac{z - X(t)}{\lambda}}, \quad
\frac{\sin \phi_0 (t)}{\cosh \frac{z - X(t)}{\lambda}}, \quad
\tanh \frac{z - X(t)}{\lambda}
\right)
,\end{aligned}$$ where $\lambda = \sqrt{J_{\mathrm{ex}} / K}$ is the DW width. By using $X(t)$ and $\phi_0(t)$, the DW Lagrangian and the dissipation function are written by $\mathcal{L}_{\mathrm{w}} = N_{\mathrm{w}} S \left( \hbar \dot{X} \phi_0 / \lambda - (K_{\perp} S / 2) \sin^2 \phi_0 \right)$ and $\mathcal{W}_{\mathrm{w}} = \frac{\alpha N_{\mathrm{w}} \hbar S}{2} \left( \frac{\dot{X}^2}{\lambda^2} + \dot{\phi}_0^2 \right)$, where $N_{\mathrm{w}} = 2 \lambda A / a^3$ is the number of spins in the wall with $A$ being the cross-sectional area. We have neglected the spin wave excitations. From these, the equation of motion is written as
$$\begin{aligned}
\dot{\phi}_0 + \alpha \frac{\dot{X}}{\lambda}
& = - \tau_{sd} \frac{ \mathrm{d} \mathcal{T} }{ \mathrm{d} t} + \zeta_{\mathrm{s}} \mathcal{T}
\label{eq:EOMa}
, \\
\frac{\dot{X}}{\lambda} - \alpha \dot{\phi}_0
& = \frac{v_{c}}{\lambda} \sin 2 \phi_0
+ \mathcal{T}
\label{eq:EOMb}
,\end{aligned}$$
where $$\begin{aligned}
\mathcal{T} = \frac{a^3}{2 e S \lambda} \frac{ P j_0 e^{-\zi \omega t} }{1 + (\zi \omega \tau_{sd} + \zeta_{\mathrm{s}})^2}
, \qquad
v_c = \frac{K_{\perp} \lambda S}{2 \hbar}\end{aligned}$$
with the electric current density $j_0$ and its polarization $P$. Here, $\mathcal{T}$ is the spin torques that we obtain and act as the forces to $X (t)$ and $\phi_0 (t)$. Solving Eqs. (\[eq:EOMa\]) and (\[eq:EOMb\]) numerically, we find that the DW position $X(t)$ and angle $\phi_0(t)$ oscillate with the period $2 \pi / \omega$ for the low current density $(a^3/2 e S) P j_0 \lesssim v_c$. We also find that the amplitude of the oscillations become larger as $\omega \tau_{sd}$ approaching to unitary (Fig. \[fig:2\]). Figure \[fig:2\] depicts the oscillation amplitude of the DW position for the case of $(a^3/2 e S) P j_0 / v_c = 10^{-4}$ and $v_c \tau_{sd} / \lambda = 0.1$, which are equivalent to the case that, for $v_c \simeq 3 \,[\mathrm{m/s}]$ [@yamaguchi2004] and $a \sim 1.5 \, [\AA]$, $j_0 /S \sim 3 \times 10^{8} \, [\mathrm{A/m^2}]$ for $P = 0.1$ and $\tau_{sd} = 6.7 \times 10^{-10} \,[\mathrm{s}]$ assuming $\lambda = 20 \,[\mathrm{nm}]$. Hence, when observing the DW position as changing the AC frequency, we estimate the exchange interaction strength from the particular frequency in which the oscillation amplitude takes a maximum value. Note that the current density is four order smaller than common one [@yamaguchi2004].
![\[fig:2\]Amplitude of the oscillation of the position $X$ during the period $1 / 2 \pi \omega$ for various $\zeta_{\mathrm{s}}$. It is clear that the amplitude enhances near the resonance point $\omega \tau_{sd} = 1$. ](DWM.pdf){width="0.9\linewidth"}
In conclusion, we have developed a theory of the interfacial spin-transfer torque and $\beta$-term torque, by consider the bilayer structure of the normal metal and the ferromagnet with the spatially-varying magnetic texture, applying alternating current parallel to the interface. We find that both torques are enhanced as the alternating current frequency $\omega$ approaches to $1 / \tau_{sd} = 2 \Delta / \hbar$. We also find that the $\beta$-term torque we obtain here includes a novel contribution which is proportional to the time derivative of the current and exists even in the absence of spin relaxation processes. Evaluating the domain wall motion due to the spin torques, we directly estimate the interfacial exchange interaction strength. We have revealed an aspect of the spin-transfer torque with finite frequency, which are enhanced by the resonance of electronic states. By using this enhancement, less current density is needed to magnetization dynamics, which may lead to low-energy consuming magnetic devices.
We thank S. Maekawa, G. Tatara, and J. Shibata for fruitful discussion. We also thank Y. Nozaki for giving stimulating information.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The goal of an algorithm substitution attack (ASA), also called a subversion attack (SA), is to replace an honest implementation of a cryptographic tool by a subverted one which allows to leak private information while generating output indistinguishable from the honest output. Bellare, Paterson, and Rogaway provided at CRYPTO ’14 a formal security model to capture this kind of attacks and constructed practically implementable ASAs against a large class of *symmetric encryption schemes*. At CCS’15, Ateniese, Magri, and Venturi extended this model to allow the attackers to work in a fully-adaptive and continuous fashion and proposed subversion attacks against *digital signature schemes*. Both papers also showed the impossibility of ASAs in cases where the cryptographic tools are deterministic. Also at CCS’15, Bellare, Jaeger, and Kane strengthened the original model and proposed a universal ASA against sufficiently random encryption schemes. In this paper we analyze ASAs from the perspective of steganography – the well known concept of hiding the presence of secret messages in legal communications. While a close connection between ASAs and steganography is known, this lacks a rigorous treatment. We consider the common computational model for secret-key steganography and prove that successful ASAs correspond to secure stegosystems on certain channels and vice versa. This formal proof allows us to conclude that ASAs are stegosystems and to “rediscover” several results concerning ASAs known in the steganographic literature.'
author:
- Sebastian Berndt
- Maciej Liśkiewicz
title: |
Algorithm Substitution Attacks from a\
Steganographic Perspective
---
Introduction
============
The publication of secret internal documents of the NSA by Edward Snowden (see [e.g.]{}[@ball2013revealed; @greenwald2014no; @perlroth2013nsa]) allowed the cryptographic community a unique insight into some well-kept secrets of one of the world’s largest security agency. Two conclusions may be drawn from these reveals:
- On the one hand, even a large organization such as the NSA seems not to be able to break well established implementations of cryptographic primitives such as RSA or AES.
- On the other hand, the documents clearly show that the NSA develops methods and techniques to *circumvent* the well established security notions by [e.g.]{}manipulating standardization processes ([e.g.]{}issues surrounding the number generator `Dual_EC_DRBG` [@checkoway2014dual; @schneier2007did; @shumow2007back]) or reason about metadata.
This confirms that the security guarantees provided by the cryptographic community are sound, but also indicates that some security definitions are too narrow to evade all possible attacks, including (non-)intentional improper handling of theoretically sound cryptographic protocols. A very realistic attack which goes beyond the common framework is a modification of an appropriate implementation of a secure protocol. The modified implementation should remain indistinguishable from a truthful one and its aim is to allow leakage of secret information during subsequent runs of the subverted protocol. Attacks of this kind are known in the literature [@young1997kleptography; @young1996dark; @bellare2015asa; @bellare2014asa; @ateniese2015sig; @russell2016cliptography] and an overview on this topic is given in the current survey [@schneier2015survey] by @schneier2015survey
A powerful class of such attacks that we will focus on – coined *secretly embedded trapdoor with universal protection (SETUP) attacks* – was presented over twenty years ago by Young and Yung in the *kleptographic* model framework [@young1996dark; @young1997kleptography]. The model is meant to capture a situation where an adversary (or “big brother” as we shall occasionally say) has the opportunity to implement (and, indeed, “mis-implement” or subvert) a basic cryptographic tool. The difficulty in detecting such an attack is based on the hardness of program verification. By using *closed source* software, the user must trust the developers that their implementation of cryptographic primitives is truthful and does not contain any backdoors. This is especially true for hardware-based cryptography [@bellare2014asa]. But it is difficult to verify this property. Even if the software is *open source* – the source code is publicly available – the sheer complexity of cryptographic implementations allows only very specialized experts to be able to judge these implementations. Two of the most prominent bugs of the widely spread cryptographic library `OpenSSL`[^1] – the *Heartbleed bug* and Debian’s faulty implementation of the pseudorandom number generator – remained undiscovered for more than two years [@schneier2015survey].
Inspired by Snowden’s reveals, the recent developments reignited the interest in these kind of attacks. @bellare2014asa named them *algorithm substitution attacks (ASA)* and showed several attacks on certain symmetric encryption schemes [@bellare2014asa]. Note that they defined a very weak model, where the only goal of the attacker was to distinguish between two ciphertexts, but mostly used a stronger scenario with the aim to recover the encryption key. @degabriele2015asa criticized the model of [@bellare2014asa] by pointing out the results crucially rely on the fact that a subverted encryption algorithm always needs to produce valid ciphertexts (the *decryptability assumption*) and proposed a refined security notion [@degabriele2015asa]. The model of algorithm substitution attacks introduced in [@bellare2014asa] was extended to signature schemes by @ateniese2015sig in [@ateniese2015sig]. Simultaneously, @bellare2015asa [@bellare2015asa] strengthened the result of [@bellare2014asa] by enforcing that the attack needs to be stateless.
In this paper we thoroughly analyze (general) ASAs from the *steganographic* point of view. The principle goal of steganography is to hide information in unsuspicious communication such that no observer can distinguish between normal documents and documents that carry additional information. Modern steganography was first made popular due to the prisoners’ problem by Simmons [@simmons1984prisoners] but, interestingly, the model was inspired by detecting the risk of ASAs during development of the SALT2 treaty between the Soviet Union and the United States in the late seventies [@simmons1998history]. This sheds some light on the inherent relationship between these two frameworks which is well known in the literature (see [e.g.]{}[@young1996dark; @young1997kleptography; @russell2016destroying]). A related result showing that so called *decoy password vaults* are very closely related to stegosystems on a certain kind of channels was presented by @pasquini2017decoy in [@pasquini2017decoy].
Our main achievement is providing a strict relationship between secure algorithm substitution attacks and the common computational model for secret-key steganography. Particularly, we prove that successful ASAs correspond to secure stegosystems on certain channels and vice versa. This formal proof allows us to conclude that ASAs are stegosystems and to “rediscover” results of [@bellare2014asa; @bellare2015asa; @ateniese2015sig] concerning ASAs.
The computational model for steganography used in this paper was first presented by Hopper, Langford, and von Ahn [@hopper2002provably; @hopper2009provably] and independently proposed by Katzenbeisser and Petitcolas [@katzenbeisser2002defining]. A *stegosystem* consists of an encoder and a decoder sharing a key. The encoder’s goal is to *embed* a secret message into a sequence of documents which are send via a public *communication channel* ${\mathcal{C}}$ monitored by an adversary (often called the *warden* due to the prisoners problem of Simmons [@simmons1984prisoners]). The warden wants to distinguish documents that carry no secret information from those sent by the encoder. If all polynomial-time (in the security parameter $\kappa$) wardens fail to distinguish these cases, we say that the stegosystem is *secure*. If the decoder is able to reconstruct the secret message from the sequence send by the encoder, the system is called *reliable*.
Our Results {#our-results .unnumbered}
-----------
We first investigate algorithm substitution attacks against symmetric encryption schemes in the framework by @bellare2015asa [@bellare2015asa]. We model encryption schemes as steganographic channels in appropriate way which allows to relate algorithm substitution attacks with steganographic systems and vice versa. This leads to the following result.
\[thm:stego:asa:iff:on:ses\] Assume that $\operatorname{\mathsf{SES}}$ is a symmetric encryption scheme. Then there exists an indistinguishable and reliable algorithm substitution attack against $\operatorname{\mathsf{SES}}$ if and only if there exists a secure and reliable stegosystem on the channel determined by $\operatorname{\mathsf{SES}}$.
The proof of the theorem is constructive in the sense that we give an explicit construction of an algorithm substitution attack against $\operatorname{\mathsf{SES}}$ from a stegosystem and vice versa. As conclusion we provide a generic ASA against *every* symmetric encryption scheme $\operatorname{\mathsf{SES}}$ whose insecurity is negligible if, roughly speaking, $\operatorname{\mathsf{SES}}$ has sufficiently large min-entropy. Our algorithm against $\operatorname{\mathsf{SES}}$ achieves almost the same performance as the construction of @bellare2015asa (see Theorem 4.1 and Theorem 4.2 in [@bellare2015asa] and also our discussion in Section \[Sec:ASA:against:encrypted:as:stego\]).
Next, we generalize our construction and show a generic algorithm substitution attack $\operatorname{\mathsf{ASA}}$ against any (polynomial-time) randomized algorithm $\operatorname{\mathsf{R}}$ which, with hardwired secret $s$, takes inputs $x$ and generates outputs $y$. Algorithm $\operatorname{\mathsf{ASA}}$, using a hidden hardwired random key ${\textit{ak}}$, returns upon the secret $s$ the sequence $\tilde{y}_1,\tilde{y}_2,\ldots$ such that the output is indistinguishable from $\operatorname{\mathsf{R}}(s,x_1), \operatorname{\mathsf{R}}(s,x_2),\ldots$ and $\tilde{y}_1,\tilde{y}_2,\ldots$ embeds the secret $s$. From this result we conclude:
\[thm:generic-attack\] There exists a generic algorithm substitution attack $\operatorname{\mathsf{ASA}}$ that allows an undetectable subversion of any cryptographic primitive of sufficiently large min-entropy.
\[thm:min-entropy\] Let $\Pi$ be a cryptographic primitive consisting with algorithms $(\Pi.A_{1},\Pi.A_{2},\ldots,$ $ \Pi.A_{r})$ such that $\{A_{i}\mid i\in
I\}$ for some $I\subseteq \{1,\ldots,r\}$ are deterministic. Then there is no ASA on $\Pi$ which subverts only algorithms $\{A_{i}\mid i\in I\}$.
As a corollary we obtain the result of @ateniese2015sig (Theorem 1 in [@ateniese2015sig]) that for every *coin-injective* signature scheme, there is a successful algorithm substitution attack of negligible insecurity. Moreover we get (Theorem 2 in [@ateniese2015sig]) that for every *coin-extractable* signature scheme, there is a successful and secure ASA. We can conclude also (Theorem 3 in [@ateniese2015sig]) that *unique signature schemes* are resistant to ASAs fulfilling the *verifiability condition*. Roughly speaking the last property means that each message has exactly one signature and the ASA can only produce valid signatures.
We furthermore introduce the concept of *universal ASAs* that can be used without a detailed description of the implementation of the underlying cryptographic primitive and note that almost all known ASAs belong to this class. Based upon this definition, we prove the following upper bound on the information that can be embedded into a single ciphertext:
\[thm:universal:bound\] No universal ASA is able to embed more than $\mathcal{O}(1)\cdot \log(\kappa)$ bits of information into a single ciphertext.
The paper is organized as follows. Section \[sec:prelim\] contains the basic preliminaries and notations that we use throughout this work, Section \[sec:asa\] presents the formal definitions of algorithm substitution attacks, and Section \[sec:stego\] gives the necessary background on steganography. In order to relate ASAs and steganography, we make use of an appropriate channel for symmetric encryption schemes defined in Section \[sec:channel\]. The proof of Theorem \[thm:stego:asa:iff:on:ses\] is given in Section \[Sec:ASA:against:encrypted:as:stego\], where one direction is contained in Theorem \[thm:asa:on:ses:impl:stego\] and the other direction is given as Theorem \[thm:ses:on:ses:impl:stego\]. We generalize our results to arbitrary randomized algorithms in Section \[sec:general\]. Combining the positive results of Theorem \[thm:generic-attack:against:R\] with the generic stegosystem provided by Theorem \[thm:rejsam:secure\] allows us to conclude Theorem \[thm:generic-attack\]. The negative results of Theorem \[thm:no-attack:against:R\] directly give Theorem \[thm:min-entropy\]. Finally, Section \[sec:bound\] defines universal ASAs and contains the upper bound on the transmission rate of these ASAs via a sequence of lemmata that results in Corollary \[cor:bound\] implying Theorem \[thm:universal:bound\].
Basic Preliminaries and Notations {#sec:prelim}
=================================
We use the following standard notations. A function $f\colon {\mathbb{N}}\to {\mathbb{N}}$ is *negligible*, if for all $c\in {\mathbb{N}}$, there is an $n_{0}\in {\mathbb{N}}$ such that $f(n) < n^{-c}$ for all $n\geq n_{0}$. The set of all strings of length $n$ on an alphabet $\Sigma$ is denoted by $\Sigma^{n}$ and the set of all strings of length at most $n$ is denoted by $\Sigma^{\leq
n}:=\cup_{i=0}^{n}\Sigma^{i}$. If $S$ is a set, $x\gets S$ denotes the uniform random assignment of an element of $S$ to $x$. If $\mathsf{A}$ is a randomized algorithm, $x\gets \mathsf{A}$ denotes the random assignment (with regard to the internal randomness of $\mathsf{A}$) of the output of $\mathsf{A}$ to $x$. The *min-entropy* measures the amount of randomness of a probability distribution $D$ and is defined as ${H_{\infty}}(D)=\inf_{x\in \operatorname{supp}(D)}\{-\log \Pr_{D}(x)\}$, where $\operatorname{supp}(D)$ is the support of $D$. Moreover, PPTM stands for probabilistic polynomial-time Turing machine.
A *symmetric encryption scheme* $\operatorname{\mathsf{SES}}$ is a triple of probabilistic polynomial-time algorithms $({\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}},{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}},{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Dec}}})$ with parameters ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ describing the length of the encrypted message and ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{cl}}}(\kappa)$ describing the length of a generated cipher message. The algorithms have the following properties:
- The *key generator* ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}}$ produces upon input $1^{\kappa}$ a key $k$ with $|k|=\kappa$.
- The *encryption algorithm* ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}$ takes as input the key $k$ and a message $m\in \{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$ of length ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ and produces a *ciphertext* $c\in \{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{cl}}}(\kappa)}$ of length ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{cl}}}(\kappa)$.
- The *decryption algorithm* ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Dec}}}$ takes as input the key $k$ and a ciphertext $c\in \{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{cl}}}(\kappa)}$ and produces a message $m'\in
\{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$.
If the context is clear, we also write $\operatorname{\mathsf{Gen}}$, $\operatorname{\mathsf{Enc}}$, $\operatorname{\mathsf{Dec}}$, $\operatorname{\mathsf{ml}}$ and $\operatorname{\mathsf{cl}}$ without the prefix $\operatorname{\mathsf{SES}}$. We say that $(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Enc}},\operatorname{\mathsf{Dec}})$ is *reliable*, if $\operatorname{\mathsf{Dec}}(k,\operatorname{\mathsf{Enc}}(k,m))=m$ for all $k$ and all $m$.
An *cpa-attacker* $\operatorname{\mathsf{A}}$ against a symmetric encryption scheme is a PPTM that mounts *chosen-plaintext-attacks (cpa)*: It is given a challenging oracle $\operatorname{CH}$ that either equals $\operatorname{\mathsf{Enc}}_{k}$ for a randomly generated key $k$ or produces random bitstrings of length $\operatorname{\mathsf{cl}}(\kappa)$. For an integer $\lambda$, let $\operatorname{RAND}(\lambda)$ be an algorithm that returns uniformly distributed bitstrings of length $\lambda$. The goal of $\operatorname{\mathsf{A}}$ is to distinguish between those settings. Formally, this is defined via the following experiment named $\operatorname{\mathsf{CPA-Dist}}$:
attacker $\operatorname{\mathsf{A}}$, symmetric encryption scheme $\operatorname{\mathsf{SES}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Enc}},\operatorname{\mathsf{Dec}})$
$k\gets \operatorname{\mathsf{Gen}}(1^{\kappa})$; $b\gets \{0,1\}$ $b' \gets \operatorname{\mathsf{A}}^{\operatorname{CH}}(1^{\kappa})$
$b=0$ [**then**]{} [ **return** $\operatorname{\mathsf{Enc}}(k,m)$]{}
A symmetric encryption scheme $\operatorname{\mathsf{SES}}$ is *cpa-secure* if for every attacker $\operatorname{\mathsf{A}}$ there is a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
&\operatorname{\mathbf{Adv}}^{\operatorname{cpa}}_{\operatorname{\mathsf{SES}}}(\kappa) :=
|\Pr[\operatorname{\mathsf{CPA-Dist}}_{\operatorname{\mathsf{A}},\operatorname{\mathsf{SES}}}(\kappa)=\textsf{true}] - 1/2|\leq
\operatorname{\mathsf{negl}}(\kappa).\end{aligned}$$ The maximal advantage of any attacker against $\operatorname{\mathsf{SES}}$ is called the *insecurity* of $\operatorname{\mathsf{SES}}$ and is defined as $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{cpa}}_{\operatorname{\mathsf{SES}}}(\kappa) := \max_{\operatorname{\mathsf{A}}}\{\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{A}},\operatorname{\mathsf{SES}}}^{\operatorname{cpa}}(\kappa)\}.\end{aligned}$$
For a $\operatorname{\mathsf{SES}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Enc}},\operatorname{\mathsf{Dec}})$ we will assume that it has nontrivial randomization measured by the min-entropy ${H_{\infty}}(\operatorname{\mathsf{SES}})$ of ciphertexts that is defined via $$2^{-{H_{\infty}}(\operatorname{\mathsf{SES}})} = \max_{k,m,c} \Pr[{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(k,m)=c].$$
For two numbers $\ell,\ell'\in {\mathbb{N}}$, denote the *set of all function* from $\{0,1\}^{\ell}$ to $\{0,1\}^{\ell'}$ by $\operatorname{\mathsf{Fun}}(\ell,\ell')$. Clearly, in order to specify a random element of $\operatorname{\mathsf{Fun}}(\ell,\ell')$, one needs $2^{\ell}\times \ell'$ bits and we can thus not use completely random functions in an efficient setting. Therefore we will use efficient functions that are indistinguishable from completely random functions. A *pseudorandom function* is a pair of PPTMs $\operatorname{\mathsf{F}}=({\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}},{\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Gen}}})$ such that ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Gen}}}$ upon input $1^{\kappa}$ produces a key $k\in \{0,1\}^{\kappa}$. The keyed function ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}$ takes the key $k\gets {\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$ and a bitstring $x$ of length ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{in}}}(\kappa)$ and produces a string ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}_{k}(x)$ of length ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{out}}}(\kappa)$. An attacker, called *distinguisher* $\operatorname{\mathsf{Dist}}$, is a PPTM that upon input $1^{\kappa}$ gets oracle access to a function that either equals ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}_{k}$ for a randomly chosen key $k$ or is a completely random function $f$. The goal of $\operatorname{\mathsf{Dist}}$ is to distinguish between those cases. A pseudorandom function $\operatorname{\mathsf{F}}$ is secure if for every distinguisher $\operatorname{\mathsf{Dist}}$ there is a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
&\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{Dist}},\operatorname{\mathsf{F}}}^{\operatorname{prf}}(\kappa) \ := \\
& \quad \quad \left| \Pr[\operatorname{\mathsf{Dist}}^{{\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}_{k}}(1^{\kappa})=1] -
\Pr[\operatorname{\mathsf{Dist}}^{f}(1^{\kappa})=1]\right| \leq \operatorname{\mathsf{negl}}(\kappa),\end{aligned}$$ where $k\gets {\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$ and $f\gets
\operatorname{\mathsf{Fun}}({\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{in}}}(\kappa),{\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{out}}}(\kappa))$. If $\operatorname{\mathsf{Dist}}$ outputs $1$, this means that the distinguisher $\operatorname{\mathsf{Dist}}$ believes that he deals with a truly random function.
As usual, the maximal advantage of any distinguisher against $\operatorname{\mathsf{F}}$ is called the *prf-insecurity* $\operatorname{\mathbf{InSec}}_{\operatorname{\mathsf{F}}}^{\operatorname{prf}}(\kappa)$ and defined as $$\begin{aligned}
\operatorname{\mathbf{InSec}}_{\operatorname{\mathsf{F}}}^{\operatorname{prf}}(\kappa) := \max_{\operatorname{\mathsf{Dist}}}\{\operatorname{\mathbf{Adv}}^{\operatorname{prf}}_{\operatorname{\mathsf{Dist}},\operatorname{\mathsf{F}}}(\kappa)\}.\end{aligned}$$
Algorithm Substitution Attacks against Encryption Schemes {#sec:asa}
=========================================================
While it is certainly very useful for an attacker to be able to reconstruct the key, one can also consider situations, where the extractor should be able to extract different information from the ciphertexts or signatures. We will thus generalize the algorithm substitution attacks described in the literature to the setting, where the substituted algorithm also takes a message ${\textit{am}}$ as argument and the goal of the extractor is to derive this message from the produced ciphertext. By always setting ${\textit{am}}:= k$, this is the setting described by @bellare2015asa in [@bellare2015asa]. We thus strengthen the model of [@bellare2014asa] and [@bellare2015asa] in this sense.
Below we give in detail our definitions based upon the model proposed by Bellare et al. in [@bellare2015asa]. If the substitution attack is *stateful*, we allow the distinguisher that tries to identify the attack to also choose this state and observe the internal state of the attack. Every algorithm substitution attack thus needs to be *stateless*, as in the model of Bellare et al. in [@bellare2015asa]. Note that this is a stronger requirement than in [@bellare2014asa] and [@ateniese2015sig], as those works also allowed stateful attacks.
In our setting an *algorithm substitution attack* against a symmetric encryption scheme $\operatorname{\mathsf{SES}}=({\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}},{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}},{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Dec}}})$ is a triple of PPTMs $$\begin{array}{rcl}
\operatorname{\mathsf{ASA}}&=&({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}},{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}},{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Ext}}}) \\
\end{array}$$ with parameter ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ for the *message length* – the length of the attacker message – and the following functionality.
- The *key generator* ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}$ produces upon input $1^{\kappa}$ an attacker key ${\textit{ak}}$ of length $\kappa$.
- The *encryption algorithm* ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}$ takes an attacker key ${\textit{ak}}\in \operatorname{supp}({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa}))$, attacker message ${\textit{am}}$ such that ${\textit{am}}\in \{0,1\}^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, an encryption key $k\in \operatorname{supp}({\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}}$ $(1^{\kappa}))$, an encryption message $m\in \{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, and a state $\sigma\in \{0,1\}^{*}$ and produces a ciphertext $c$ of length ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{cl}}}(\kappa)$ and a new state $\sigma'$.
- The *extraction algorithm* ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Ext}}}$ takes as input an attacker key ${\textit{ak}}\in \operatorname{supp}({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa}))$ and $\ell =
{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)$ a ciphertext $c_{1},\ldots,c_{\ell}$ with $c_{i}\in
\{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{cl}}}(\kappa)}$ and produces an attacker message ${\textit{am}}'$.
An algorithm substitution attack needs (a) to be indistinguishable from the symmetric encryption scheme and (b) should be able to reliably extract the message ${\textit{am}}$ of length ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ from the ciphertexts. Due to information-theoretic reasons, it might be impossible to embed the attacker message ${\textit{am}}$ into a single ciphertext: If ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}$ uses $10$ bits of randomness, at most $10$ bits from ${\textit{am}}$ can be reliably embedded into a ciphertext. Hence, the algorithm substitution attack needs to produce more than one ciphertext in this case. For message $m_{1},\ldots,m_{\ell}$, the complete output, denoted as ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}^{\ell}({\textit{ak}},{\textit{am}},k,m_{1},\ldots,m_{\ell})$ is defined as follows:
1. $\sigma = \varnothing$
2. [**for**]{} $j=1$ to $\ell$ [**do**]{} $(c_{j},\sigma)\gets {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}({\textit{ak}},{\textit{am}},k,m_{j},\sigma)$
3. [**return**]{} $c_{1},\ldots,c_{\ell}$
To formally define the probability that the extractor is able to reliably extract ${\textit{am}}$ from the given ciphertexts $c_{1},\ldots,c_{\ell}$, we define its *reliability*[^2] as $1-\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)$, where the *unreliability* $\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}$ is given as $$\begin{aligned}
&\max \{
\Pr[{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Ext}}}({\textit{ak}},{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}^{\ell}({\textit{ak}},{\textit{am}},k,m_{1},\ldots,m_{\ell}))\neq
{\textit{am}}]\},\end{aligned}$$ with the maximum taken over all ${\textit{ak}}\in \operatorname{supp}({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})),
{\textit{am}}\in\{0,1\}^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, and $m_{i}\in\{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$. The algorithm is *successful*, if there is negligible function $\operatorname{\mathsf{negl}}$ with $\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa) \leq \operatorname{\mathsf{negl}}(\kappa).$
The indistinguishability of an ASA is defined as follows. Call a *watchdog* $\operatorname{\mathsf{Watch}}$ a PPTM that tries to distinguish the output of the attacker encryption algorithm ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}$ from the original encryption algorithm $\operatorname{\mathsf{Enc}}$. The indistinguishability is defined via the game named $\operatorname{\mathsf{ASA-Dist}}$:
watchdog $\operatorname{\mathsf{Watch}}$, algorithm substitution attack $\operatorname{\mathsf{ASA}}=({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}, {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}},$ $ {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Ext}}})$, and encryption scheme $\operatorname{\mathsf{SES}}=({\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}},{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}},{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Dec}}})$
${\textit{ak}}\gets {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$; $b\gets \{0,1\}$ $b' \gets \operatorname{\mathsf{Watch}}^{\operatorname{CH}}(1^{\kappa})$ $b=0$ [**then**]{} $c\gets {\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(k,m)$ $(c,\sigma) \gets {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}({\textit{ak}},{\textit{am}},k,m,\sigma)$
An algorithm substitution attack $\operatorname{\mathsf{ASA}}$ is called *indistinguishable* from the symmetric encryption scheme $\operatorname{\mathsf{SES}}$, if for every watchdog $\operatorname{\mathsf{Watch}}$, there is a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
&\operatorname{\mathbf{Adv}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{Watch}}, \operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa) \
:=\\ &\quad \quad |\Pr[\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)=\textsf{true}] -1/2|
\leq \ \operatorname{\mathsf{negl}}(\kappa).\end{aligned}$$
The maximal advantage of any watchdog distinguishing $\operatorname{\mathsf{ASA}}$ from $\operatorname{\mathsf{SES}}$ is called the *indistinguishability* or *insecurity* of $\operatorname{\mathsf{ASA}}$ and is defined as $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa) := \max_{\operatorname{\mathsf{Watch}}}\{\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}^{\operatorname{enc-watch}}(\kappa)\}.\end{aligned}$$
In [@bellare2014asa], Bellare et al. proposed a (stateless) construction $\operatorname{\mathsf{ASA}}$ against all symmetric encryption schemes $\operatorname{\mathsf{SES}}$. They prove in Theorem 3 that if $\operatorname{\mathsf{SES}}$ is a randomized, stateless, coin-injective symmetric encryption scheme with randomness-length $r$ and if the ASA uses a PRF $\operatorname{\mathsf{F}}$ then for a watchdog $\operatorname{\mathsf{Watch}}$ that makes $q$ queries to its $\operatorname{CH}$ oracle we can construct an adversary $\operatorname{\mathsf{A}}$ such that $ \operatorname{\mathbf{Adv}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{Watch}}, \operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa) \le q/2^{2^r} +
\operatorname{\mathbf{Adv}}^{\operatorname{prf}}_{\operatorname{\mathsf{A}},\operatorname{\mathsf{F}}}(\kappa)$, where $\operatorname{\mathsf{A}}$ makes $q$ oracle queries and its running time is that of $\operatorname{\mathsf{Watch}}$.
Bellare et al. conclude that as long as their scheme uses a non-trivial amount of randomness, for example $r \ge 7$ bits resulting $2^r \ge 128$, Theorem 3 implies that the subversion is undetectable.
Backgrounds of Steganography {#sec:stego}
============================
The definitions of the basic steganography concepts presented in this section are essentially those of [@hopper2009provably] and [@dedic2009upper].
In order to define undetectable hidden communication, we need to introduce a notion of *unsuspicious* communication. We do this via the notion of a *channel* ${\mathcal{C}}$. A channel ${\mathcal{C}}$ on the alphabet $\Sigma$ with maximal document length ${{{\mathcal{C}}\,\!.\!\,n}}$ is a function that maps a string of previously send elements $h\in (\Sigma^{\leq {{{\mathcal{C}}\,\!.\!\,n}}})^{*}$ – the *history* – to a probability distribution upon $\Sigma^{\leq {{{\mathcal{C}}\,\!.\!\,n}}}$. We denote this probability distribution by ${\mathcal{C}}_{h}$. The elements of $\Sigma^{\leq {{{\mathcal{C}}\,\!.\!\,n}}}$ are called *documents*. As usually, we will assume that the sequences of documents are efficiently prefix-free recognizable.
A *stegosystem* $\operatorname{\mathsf{S}}$ on a family of channels $\bm{{\mathcal{C}}}=\{{\mathcal{C}}^{\kappa}\}_{\kappa\in {\mathbb{N}}}$ is a triple of probabilistic polynomial-time (according to the security parameter $\kappa$) algorithms: $$\begin{array}{rcl}
\operatorname{\mathsf{S}}&=&({\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Gen}}},{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}},{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Dec}}})
\end{array}$$ with parameters ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ describing the *message length* of the subliminal (hidden, or attacker) message and ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)$ describing the length of a generated sequence of stego documents to embed the whole hidden message. The algorithms have the following functionality:
- The *key generator* ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Gen}}}$ takes the unary presentation of an integer $\kappa$ – the *security parameter* – and outputs a key (we will call it an attacker key) ${\textit{ak}}\in \{0,1\}^{\kappa}$ of length $\kappa$.
- The *stegoencoder* ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}}$ takes as input the key ${\textit{ak}}$, the attacker (or hidden) message ${\textit{am}}\in
\{0,1\}^{{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, a history $h$, and a state $\sigma$ and outputs a document $d$ from ${\mathcal{C}}^{\kappa}$ such that ${\textit{am}}$ is (partially) embedded in this document and a new state. In order to produce the document, ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}}$ also has sampling access to ${\mathcal{C}}^{\kappa}_{h}$. We denote this by writing ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}}^{{\mathcal{C}}}({\textit{ak}},{\textit{am}},h,\sigma)$.
- The *(history-ignorant) stegodecoder* ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Dec}}}$ takes as input the key ${\textit{ak}}$ and $\ell={\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)$ documents $d_{1},\ldots,d_{\ell}$ and outputs a message ${\textit{am}}'$. A history-ignorant stegodecoder thus has no knowledge of previously sent documents. The stegodecoders of nearly all known systems are history-ignorant.
To improve readability, if the stegosystem is clear from the context, we will omit the prefix $\operatorname{\mathsf{S}}$. If $\bm{{\mathcal{C}}}=\{{\mathcal{C}}^{\kappa}\}_{\kappa\in \mathbb{N}}$ is a family of channels, the *min-entropy* of ${H_{\infty}}(\bm{{\mathcal{C}}},\kappa)$ is defined as ${H_{\infty}}(\bm{{\mathcal{C}}},\kappa) = \min_{h\in
\Sigma^{*}}\{{H_{\infty}}({\mathcal{C}}^{\kappa}_{h})\}$. In order to be useful, the stegodecoder should *reliably* decode the embedded message from the sequence of documents. As in the setting of algorithm substitution attack, the complete output of $\ell$ documents of the stegosystem for the history $h$ on the subliminal message ${\textit{am}}$ of length ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ is denoted as ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}}^{\ell,{\mathcal{C}}}({\textit{ak}},{\textit{am}},h)$ and is defined as follows.
1. $\sigma = \varnothing$
2. [**for**]{} $j=1$ to $\ell$ [**do**]{}
3. $(d_{j},\sigma)\gets{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}}^{{\mathcal{C}}}({\textit{ak}},{\textit{am}},h,\sigma)$; $h = h \mid\mid d_{j}$
4. [**return**]{} $d_{1},\ldots,d_{\ell}$
The *unreliability* $\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)$ of the stegosystem $\operatorname{\mathsf{S}}$ on the channel family $\{{\mathcal{C}}^{\kappa}\}_{\kappa\in {\mathbb{N}}}$ with security parameter $\kappa$ is defined as $$\begin{aligned}
& \operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) := \\
& \max_{{\textit{ak}}, {\textit{am}}} \max_{h}
\{\Pr[{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Dec}}}({\textit{ak}},{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}^{\ell,{\mathcal{C}}}}({\textit{ak}},{\textit{am}},h))\neq {\textit{am}}]\},\end{aligned}$$ where the maximum is taken over all ${\textit{ak}}\in \operatorname{supp}({\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})),
{\textit{am}}\in \{0,1\}^{{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, and $h\in(\Sigma^{n(\kappa)})^{*}$. If there is a negligible function $\operatorname{\mathsf{negl}}$ such that $\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\leq \operatorname{\mathsf{negl}}(\kappa)$, we say that $\operatorname{\mathsf{S}}$ is *reliable* on ${\mathcal{C}}$. Furthermore, the *reboot-reliability* of the stegosystem $\operatorname{\mathsf{S}}$ is defined as $$\begin{aligned}
& \operatorname{\mathbf{UnRel}}^{\star}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) := \\
&\max_{{\textit{ak}}, {\textit{am}}} \max_{\tau}\max_{ h_1,\ldots,h_{\tau}}\max_{\ell_1,\ldots,\ell_{\tau}} \{
\Pr[{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Dec}}}({\textit{ak}},
d_1,d_2,\ldots,d_{\ell}
)\neq {\textit{am}}]\} \end{aligned}$$ where the maxima are taken over all ${\textit{ak}}\in \operatorname{supp}({\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa}))$, ${\textit{am}}\in \{0,1\}^{{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, all positive integers $\tau\le \ell$, all histories $h_{1},\ldots,h_{\tau}$, and all positive integers $\ell_1,\ldots,\ell_{\tau}$ such that $\ell_1 + \ldots + \ell_{\tau}=\ell$. The documents $d_{1},\ldots, d_{\ell}$ are the concatenated output of the runs $$\begin{aligned}
{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}^{\ell_1,{\mathcal{C}}}}({\textit{ak}},{\textit{am}},h_1)\mid\mid
\ldots \mid\mid
{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}^{\ell_\tau,{\mathcal{C}}}}({\textit{ak}},{\textit{am}},h_\tau). \end{aligned}$$ We say that the stegosystem $\operatorname{\mathsf{S}}$ is *reboot-reliable* if $\operatorname{\mathbf{UnRel}}^{\star}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)$ is bounded from above by a negligible function. This corresponds to a situation where the stegoencoder is restarted $\tau$ times, each time with the history $h_i$, and is allowed to generate $\ell_i$ documents. Note that reboot-reliability is a strictly stronger requirement than reliability and we can thus conclude $$\begin{aligned}
\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\leq \operatorname{\mathbf{UnRel}}^{\star}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa).\end{aligned}$$
To define the security of a stegosystem, we first specify the abilities of an attacker: A *warden* $\operatorname{\mathsf{Ward}}$ is a probabilistic polynomial-time algorithm that will have access to a *challenge oracle* $\operatorname{CH}$. This challenge oracle can be called with a message ${\textit{am}}$ and a history $h$ and is either equal to ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}}^{{\mathcal{C}}}({\textit{ak}},{\textit{am}},h,\sigma)$ for a key ${\textit{ak}}\gets {\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$ or equal to random documents of the *channel*.
The goal of the warden is to distinguish between those oracles. It also has access to samples of the channel ${\mathcal{C}}^{\kappa}_{h}$ for a freely chosen history $h$. Formally, the *chosen-hiddentext-attack-advantage* is defined via the following game $\operatorname{\mathsf{SS-CHA-Dist}}$:
warden $\operatorname{\mathsf{Ward}}$, stegosystem $\operatorname{\mathsf{S}}$, channel ${\mathcal{C}}$
${\textit{ak}}\gets {\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$ $b\gets \{0,1\}$ $b' \gets \operatorname{\mathsf{Ward}}^{\operatorname{CH},{\mathcal{C}}}(1^{\kappa})$ $b=0$ [**then**]{} $d \gets {\mathcal{C}}^{\kappa}_{h}$ $(d,\sigma)\gets {\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{Enc}}}({\textit{ak}},{\textit{am}},h,\sigma)$
A stegosystem $\operatorname{\mathsf{S}}$ is called *secure against chosen-hiddentext attacks* if for every warden $\operatorname{\mathsf{Ward}}$, there is a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
\operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}}, \operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
&:= |\Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)=\textsf{true}] -1/2| \\
& \leq \ \operatorname{\mathsf{negl}}(\kappa).\end{aligned}$$
The maximal advantage of any warden against $\operatorname{\mathsf{S}}$ is the *insecurity* $\operatorname{\mathbf{InSec}}^{\operatorname{cha}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)$ and defined as $\max_{\operatorname{\mathsf{Ward}}}\{\operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\}$.
A very common technique in the design of secure stegosystems called *rejection sampling* goes back to an idea of Anderson, presented in [@anderson1996limits]. The basic concept is that the stegoencoder samples from the channel until he finds a document that already encodes the hiddentext. This was first used by Cachin in [@Cachin2004] to construct a secure stegosystem in the information-theoretic sense.
In the following, let $\operatorname{\mathsf{F}}$ be pseudorandom function that maps input strings of length ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{in}}}(\kappa)$ (documents) to strings of length ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{out}}}(\kappa)=\log(\operatorname{\mathsf{ml}}(\kappa))+1$ (message parts). To simplify notation, we treat the output of ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}_{k}$ as a pair $(b,j)$ with $|b|=1$ and $|j|=\log(\operatorname{\mathsf{ml}}(\kappa))$. The encoder of the *rejection sampling stegosystem*, which we denote as $\operatorname{\mathsf{RejSam}}^{\operatorname{\mathsf{F}}}$, is defined as follows:
key ${\textit{ak}}$, message ${\textit{am}}$, history $h$, state $\sigma$
$i := 0$; $d \gets {\mathcal{C}}_{h}$ $i := i+1$ $(b,j) := {\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}_{{\textit{ak}}}(d)$
The key generator ${\operatorname{\mathsf{RejSam}}^{\operatorname{\mathsf{F}}}\!.\!\operatorname{\mathsf{Gen}}}$ is equal to ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Gen}}}$ and the decoder derives ${\textit{am}}$, as long as its input documents contain every bit ${\textit{am}}[j]$, by applying ${\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}_{{\textit{ak}}}$ to these documents. Below we present the description of the decoder. Note that the stegosystem is stateless.
key ${\textit{ak}}$, documents $d_{1},\ldots,d_{{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)}$
let ${\textit{am}}_{j} := \bot$ $(b,j) := {\operatorname{\mathsf{F}}\!.\!\operatorname{\mathsf{Eval}}}_{k}(d_{i})$ let ${\textit{am}}_{j} := b$
In [@hopper2009provably], Hopper et al. were the first to prove the security of this stegosystem in the complexity-theoretic model. Their argument was simplified by Dedi[ć]{} et al. in [@dedic2009upper] and by Backes and Cachin in [@backes2005public]. The version given here is based upon the stateless construction of Dedi[ć]{} et al. and also uses the idea of Bellare et al. in [@bellare2015asa] to apply the *coupon collector’s problem* to completely get rid of the state by randomly choosing an index to embed.
The analysis of the coupon collector’s problem shows that by sending $\operatorname{\mathsf{ml}}(\kappa)\cdot (\ln \operatorname{\mathsf{ml}}(\kappa)+\beta)$ documents – for an appropriate value $\beta$ – one only introduces a term $\exp(-\beta)$ into the unreliability (see [e.g.]{}[@mitzenmacher2005probability] for a proof of this fact), which can be made negligible by setting $\beta\geq \operatorname{\mathsf{ml}}(\kappa)-\ln(\operatorname{\mathsf{ml}}(\kappa))$. The output length on messages of length $\operatorname{\mathsf{ml}}(\kappa)$ will thus be bounded by $\operatorname{\mathsf{ml}}(\kappa)^{2}$.
The security of this system directly follows from the analysis of Dedi[ć]{} et al. in [@dedic2009upper]:
\[thm:rejsam:secure\] For every polynomial $\operatorname{\mathsf{ml}}(\kappa)$, there exists a universal history-ignorant stegosystem $\operatorname{\mathsf{S}}=\operatorname{\mathsf{RejSam}}^{\operatorname{\mathsf{F}}}$ with security parameter $\kappa$ and $s\ge 1$ such that for every channel ${\mathcal{C}}^{\kappa}$ we have
- ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)=\operatorname{\mathsf{ml}}(\kappa)$,
- $\operatorname{\mathbf{InSec}}^{\operatorname{cha}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) \leq
\mathcal{O}(\operatorname{\mathsf{ml}}(\kappa)^{4}\cdot
2^{-{H_{\infty}}({\mathcal{C}}^{\kappa})}+\operatorname{\mathsf{ml}}(\kappa)^{2}\cdot
\exp(-s))+\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}},{\mathcal{C}}}(\kappa)$, and
- $\operatorname{\mathbf{UnRel}}^{\star}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) \leq
\operatorname{\mathsf{ml}}(\kappa)^{2}(2\cdot
\exp(-2^{{H_{\infty}}({\mathcal{C}}^{\kappa})-3}) +
\exp(-2^{-2} s))+
\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}},{\mathcal{C}}}(\kappa)$.
The notation $\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}},{\mathcal{C}}}(\kappa)$ indicates the insecurity of the pseudorandom function $\operatorname{\mathsf{F}}$ *relative to the channel ${\mathcal{C}}$*. Informally, this means that the attacker against $\operatorname{\mathsf{F}}$ also has sampling access to ${\mathcal{C}}$ (for a formal definition, see [@dedic2009upper]). For an *efficiently sampleable* channel ${\mathcal{C}}$ ([i.e.]{}one that can be simulated by a PPTM), it clearly holds that $\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}},{\mathcal{C}}}(\kappa)=\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}}}(\kappa)$. All channels used in this work are efficiently sampleable and we will thus omit the index ${\mathcal{C}}$ from the term $\operatorname{\mathbf{InSec}}$.
Encryption Schemes as Steganographic Channels {#sec:channel}
=============================================
Let $\operatorname{\mathsf{SES}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Enc}},\operatorname{\mathsf{Dec}})$ be a symmetric encryption scheme that encodes messages of length $\operatorname{\mathsf{ml}}(\kappa)$ into ciphertexts of length $\operatorname{\mathsf{cl}}(\kappa) \geq \operatorname{\mathsf{ml}}(\kappa)$ and let $\ell$ be a polynomial of $\kappa$. For $\operatorname{\mathsf{SES}}$ we define a channel family, named ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)$, indexed with parameter $\kappa\in {\mathbb{N}}$, where the documents will correspond to the input of generalized algorithm substitution attack against encryption schemes. The essential idea behind the definition of the channel ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)$ is that for all $k\in \operatorname{supp}(\operatorname{\mathsf{Gen}}(1^{\kappa}))$ and every sequence of messages $m_{1},m_{2},\ldots, m_{\ell(\kappa)}$, with $m_{i}\in\{0,1\}^{\operatorname{\mathsf{ml}}(\kappa)}$, for the history $$h= k \mid\mid m_{1} \mid\mid m_{2} \mid\mid \ldots \mid\mid m_{\ell(\kappa)}$$ the distribution of the sequences of documents $$\begin{aligned}
c_{1} \mid\mid c_{2} \mid\mid \ldots \mid\mid c_{\ell(\kappa)}
\end{aligned}$$ generated by the channel is exactly the same as the distribution for $$\begin{aligned}
\operatorname{\mathsf{Enc}}(k,m_{1}) \mid\mid \operatorname{\mathsf{Enc}}(k,m_{2}) \mid\mid \ldots \mid\mid \operatorname{\mathsf{Enc}}(k,m_{\ell(\kappa)}).
\end{aligned}$$ To give a formal definition of $\{{\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)\}_{\kappa\in {\mathbb{N}}}$ we need to specify the probability distributions for any history $h$. Thus, we define the family, on the alphabet $\{0,1\}$, as follows.
For the empty history $h=\varnothing$, define $$\begin{aligned}
{\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)_{\varnothing} \end{aligned}$$ as the distribution of all keys generated by $\operatorname{\mathsf{Gen}}(1^{\kappa})$. For a key $k\in \operatorname{supp}(\operatorname{\mathsf{Gen}}(1^{\kappa}))$ and a (possibly empty) sequence of messages $m_{1},m_{2},\ldots, m_{r}$, with $m_{i}\in \{0,1\}^{\operatorname{\mathsf{ml}}(\kappa)}$ and $0\leq r\leq \ell(\kappa)-1$, the distribution $$\begin{aligned}
{\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)_{k\mid\mid m_{1}\mid\mid
m_{2}\mid\mid \ldots \mid\mid m_{r}}
\end{aligned}$$ is the uniform distribution on all messages $m_{r+1}\in
\{0,1\}^{\operatorname{\mathsf{ml}}(\kappa)}$. For $k\in \operatorname{supp}(\operatorname{\mathsf{Gen}}(1^{\kappa}))$, a sequence of messages $m_{1},m_{2},\ldots,$ $ m_{\ell(\kappa)}$ with $m_{i}\in
\{0,1\}^{\operatorname{\mathsf{ml}}(\kappa)}$, and a (possibly empty) sequence of ciphertexts $c_{1},\ldots,c_{r}$, with $c_{i}\in \operatorname{supp}(\operatorname{\mathsf{Enc}}(k,m_{((i-1)\bmod
\ell(\kappa))+1})),$ the distribution $$\begin{aligned}
{\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)_{k\mid\mid m_{1}\mid\mid
m_{2}\mid\mid \ldots \mid\mid m_{\ell(\kappa)}\mid\mid
c_{1}\mid\mid c_{2}\mid\mid \ldots \mid\mid
\ldots \mid\mid c_{r}}
\end{aligned}$$ is the distribution of $\operatorname{\mathsf{Enc}}(k,m_{(r \bmod \ell(\kappa))+1})$.
ASAs against Encryption in the Steganographic Model {#Sec:ASA:against:encrypted:as:stego}
===================================================
The main message of our paper is that algorithm substitution attacks against a primitive $\Pi$ are equivalent to the use of steganography on a corresponding channel ${\mathcal{C}}_{\Pi}$ determined by the protocol $\Pi$. Focusing on symmetric encryption schemes as a common cryptographic primitive, we will show in this section exemplary proofs for the general relations between ASAs and steganography.
In the previous section we showed a formal specification of the family of communication channels ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)$ determined by a symmetric encryption scheme $\operatorname{\mathsf{SES}}$. We will now prove that a secure and reliable stegosystem on ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)$ implies the existence of an indistinguishable and successful algorithm substitution attack on $\operatorname{\mathsf{SES}}$. On the other hand, we will also show that the existence of an indistinguishable and successful algorithm substitution attack on $\operatorname{\mathsf{SES}}$ implies a secure and reliable stegosystem on ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(\ell)$.
As a consequence we get a construction of an ASA against any encryption scheme using a generic stegosystem like [e.g.]{}this proposed by Dedi[ć]{} et al. [@dedic2009upper]. Thus, we can conclude Theorem 1 and Theorem 3 proposed by Bellare et al. in [@bellare2014asa] that there exist indistinguishable and successful ASAs against encryption schemes. Moreover we obtain Theorem 4 in [@bellare2014asa] which says that an ASA is impossible for unique ciphertext symmetric encryption schemes.
Steganography implies ASAs
--------------------------
\[thm:asa:on:ses:impl:stego\] Assume $\operatorname{\mathsf{SES}}$ is a symmetric encryption scheme and let $\operatorname{\mathsf{S}}$ be a stegosystem on the channel ${\mathcal{C}}:={\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}({\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa))$ determined by $\operatorname{\mathsf{SES}}$. Then there exists an algorithm substitution attack $\operatorname{\mathsf{ASA}}$ against $\operatorname{\mathsf{SES}}$ of indistinguishability, resp. reliability such that: $$\begin{array}{rcl}
\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa) &\le& \operatorname{\mathbf{InSec}}^{\operatorname{cha}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
\quad \text{and} \\
\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa) &=& \operatorname{\mathbf{UnRel}}^{\star}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa).
\end{array}$$
Let $\operatorname{\mathsf{SES}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Enc}},\operatorname{\mathsf{Dec}})$ be a symmetric encryption scheme and $\operatorname{\mathsf{S}}=(\operatorname{\mathsf{SGen}},\operatorname{\mathsf{SEnc}},\operatorname{\mathsf{SDec}})$ be a stegosystem on the channel ${\mathcal{C}}$. To simplify notation, let $\ell=\ell(\kappa):={\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)$. We will construct the algorithm substitution attack $\operatorname{\mathsf{ASA}}=(\operatorname{\mathsf{AGen}},\operatorname{\mathsf{AEnc}},$ $\operatorname{\mathsf{AExt}})$ on $\operatorname{\mathsf{SES}}$ from the stegosystem $\operatorname{\mathsf{S}}$ and show the indistinguishability and success of $\operatorname{\mathsf{ASA}}$ depending on security and reliability of $\operatorname{\mathsf{S}}$. The components of the $\operatorname{\mathsf{ASA}}$ are defined as follows.
The key generator $\operatorname{\mathsf{AGen}}$ just simulates $\operatorname{\mathsf{SGen}}$ – the key generator of the stegosystem. It will output the attack key ${\textit{ak}}$. The encoding algorithm $\operatorname{\mathsf{AEnc}}$ on input ${\textit{ak}}\in
\operatorname{supp}(\operatorname{\mathsf{AGen}}(1^{\kappa}))$, ${\textit{am}}\in
\{0,1\}^{{\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, $k\in
\operatorname{supp}(\operatorname{\mathsf{Gen}}(1^{\kappa}))$, and $m\in
\{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$ simulates $\operatorname{\mathsf{SEnc}}$ on channel ${\mathcal{C}}$ with input key ${\textit{ak}}$, the message ${\textit{am}}$ and the history $h=k\mid\mid m^{\ell}$, where $m^{\ell}$ is the string of length $\ell\cdot |m|$ containing $\ell$ copies of $m$. Whenever $\operatorname{\mathsf{SEnc}}$ makes a query to its channel oracle, algorithm $\operatorname{\mathsf{AEnc}}$ uses $\operatorname{\mathsf{Enc}}$ on input $k$ and $m$ to produce a corresponding ciphertext and sends it to $\operatorname{\mathsf{SEnc}}$. The encoder $\operatorname{\mathsf{AEnc}}$ then outputs the document $d$ generated by $\operatorname{\mathsf{SEnc}}$. Finally, the extraction algorithm $\operatorname{\mathsf{AExt}}$ on input ${\textit{ak}}\in \operatorname{supp}(\operatorname{\mathsf{AGen}}(1^{\kappa}))$ and documents $d_{1},\ldots,d_{\ell}$ just simulates $\operatorname{\mathsf{SDec}}$ on the same inputs.
As one can see from the definitions, $\operatorname{\mathsf{ASA}}$ is a generalized algorithm substitution attack against $\operatorname{\mathsf{SES}}$. We will now prove that it is indistinguishable from $\operatorname{\mathsf{SES}}$ and that it is successful.
We prove first indistinguishability of the system. Let $\operatorname{\mathsf{Watch}}$ be a watchdog against the above $\operatorname{\mathsf{ASA}}$ with maximal advantage, i.e. $$\begin{aligned}
\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}^{\operatorname{enc-watch}}(\kappa) = \operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa),
\end{aligned}$$ where $\operatorname{\mathbf{Adv}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{Watch}}, \operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)$ is equal to the success probability that $\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)=\textsf{true}$. We will now construct a warden $\operatorname{\mathsf{Ward}}$ from $\operatorname{\mathsf{Watch}}$ such that $$\begin{aligned}
\operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
=
\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}^{\operatorname{enc-watch}}(\kappa). \end{aligned}$$ Thus, we will get that $$\label{ineq:insecasa:insecstego}
\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)
\le
\operatorname{\mathbf{InSec}}^{\operatorname{cha}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa).$$ The warden $\operatorname{\mathsf{Ward}}$ on input $1^{\kappa}$ just simulates the watchdog $\operatorname{\mathsf{Watch}}$ and gives the same output as $\operatorname{\mathsf{Watch}}$ at the end of the simulation. Whenever the watchdog makes a query on input ${\textit{am}}$, $k$, and $m$ to its challenging oracle (that is either equal to $\operatorname{\mathsf{SES}}$’s encryption algorithm $\operatorname{\mathsf{Enc}}(k,m)$ or to $\operatorname{\mathsf{ASA}}$’s encryption $\operatorname{\mathsf{AEnc}}({\textit{ak}},{\textit{am}},k,m,\sigma)$ for ${\textit{ak}}\gets \operatorname{\mathsf{AGen}}(1^{\kappa})$), the warden $\operatorname{\mathsf{Ward}}$ queries its own challenging oracle with message ${\textit{am}}$, state $\sigma$ and history $h=k\mid\mid m^{\ell}$. Note that the challenging oracle of $\operatorname{\mathsf{Ward}}$ is either equal to the channel ${\mathcal{C}}$ or to $\operatorname{\mathsf{SEnc}}({\textit{ak}},{\textit{am}},h,\sigma)$ for ${\textit{ak}}\gets \operatorname{\mathsf{SGen}}(1^{\kappa})$.
If the challenging oracle of $\operatorname{\mathsf{Ward}}$ is equal to the steganographic encoding $\operatorname{\mathsf{SEnc}}({\textit{ak}},{\textit{am}},h,\sigma)$ ([i.e.]{}the bit $b$ in $\operatorname{\mathsf{SS-CHA-Dist}}$ equals $1$, denoted by $\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\langle b=1\rangle$), the answer of $\operatorname{\mathsf{Ward}}$ is the same as the output of the $\operatorname{\mathsf{Watch}}$ in case it queries the ASA’s encoding algorithm $\operatorname{\mathsf{AEnc}}({\textit{ak}},{\textit{am}},k,m)$ by construction. Thus, $$\begin{aligned}
&\Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\langle b=1\rangle=\textsf{true}]\\
&= \ \Pr[\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)\langle b=1\rangle=\textsf{true}].\end{aligned}$$
If the challenging oracle of $\operatorname{\mathsf{Ward}}$ is equal to the channel (the bit $b$ in $\operatorname{\mathsf{SS-CHA-Dist}}$ equals $0$), by the definition of the channel ${\mathcal{C}}$ for the symmetric encryption scheme $\operatorname{\mathsf{SES}}$, the answer of the challenging oracle is equal to the output of $\operatorname{\mathsf{Enc}}(k,m)$. Hence, $$\begin{aligned}
&\Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\langle b=0\rangle=\textsf{true}]\\
& = \
\Pr[\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)\langle b=0\rangle=\textsf{true}].\end{aligned}$$ We thus have $$\begin{array}{rcl}
\operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
& = &
|\Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)=\textsf{true}] - 1/2| \\
& = & |\Pr[\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)=\textsf{true} -1/2| \\
& = &
\operatorname{\mathbf{Adv}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{Watch}}, \operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)
\end{array}$$ which completes the proof of .
We still need to prove that $\operatorname{\mathsf{AExt}}$ is reliably able to extract the attacker message ${\textit{am}}$ from the ciphertext. But, as $\operatorname{\mathsf{AExt}}=\operatorname{\mathsf{SDec}}$, the reboot-reliability of $\operatorname{\mathsf{SDec}}$ directly implies that $\operatorname{\mathsf{AExt}}$ is successful with probability of $1-\operatorname{\mathsf{negl}}(\kappa)$.
By combining Theorem \[thm:asa:on:ses:impl:stego\] and Theorem \[thm:rejsam:secure\], we can conclude the following corollary.
\[coroll:thm:ipl:bellare\] For every symmetric encryption scheme $\operatorname{\mathsf{SES}}$, there exists an algorithm subsection attack $\operatorname{\mathsf{ASA}}$ with message length $\operatorname{\mathsf{ml}}(\kappa)$ and parameter $s\ge 1$ such that $$\begin{array}{rcl}
\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)&\leq&
\mathcal{O}(\operatorname{\mathsf{ml}}(\kappa)^{4}\cdot 2^{-{H_{\infty}}({\mathcal{C}}^{\kappa})})+\\
&&\mathcal{O}(\operatorname{\mathsf{ml}}(\kappa)^{2}\cdot
\exp(-s))+
\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}}}(\kappa),\\[2mm]
\operatorname{\mathbf{UnRel}}^{\star}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) &\leq&
2 \operatorname{\mathsf{ml}}(\kappa)^{2}\cdot
\exp(-2^{{H_{\infty}}({\mathcal{C}}^{\kappa})-3}) +\\
&& \operatorname{\mathsf{ml}}(\kappa)^{2}\cdot \exp(-2^{-2} s)+
\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}}}(\kappa)
\end{array}$$ where ${\mathcal{C}}:={\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}({\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa))$
One can compare this corollary to the construction used in the proof of Theorem 4.1 and Theorem 4.2 in [@bellare2015asa]. We can see that our generic algorithm substitution attack gets almost the same bounds for insecurity and for unreliability.
Note that the protocols in [@bellare2015asa; @ateniese2015sig] and our generic protocol of Corollary \[coroll:thm:ipl:bellare\] have a very bad rate: $
\frac{\operatorname{\mathsf{ml}}}{\operatorname{\mathsf{ml}}\cdot (\ln \operatorname{\mathsf{ml}}+\beta)} = 1/(\ln \operatorname{\mathsf{ml}}+\beta)
$ for an appropriate value $\beta$. One can easily modify the above constructions such that instead of one bit $b$ of a message ${\textit{am}}$ we embed a block of $\log(\operatorname{\mathsf{ml}})$ bits per ciphertext. This improves the rate to $
\frac{\log \operatorname{\mathsf{ml}}}{\ln(\operatorname{\mathsf{ml}}) - \ln\log(\operatorname{\mathsf{ml}}) +\beta} = \Theta(1).
$
ASAs imply Steganography
------------------------
\[thm:ses:on:ses:impl:stego\] Assume $\operatorname{\mathsf{SES}}$ is a symmetric encryption scheme and let $\operatorname{\mathsf{ASA}}$ be an algorithm substitution attack against $\operatorname{\mathsf{SES}}$ of output length ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)$. Then there exists a stegosystem $\operatorname{\mathsf{S}}$ with the output length ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)=2\cdot {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)+1$ on the channel ${\mathcal{C}}:={\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}({\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa))$ determined by $\operatorname{\mathsf{SES}}$ such that $\operatorname{\mathsf{S}}$’s insecurity, resp. its reliability satisfy $$\begin{array}{rcl}
\operatorname{\mathbf{InSec}}^{\operatorname{cha}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) &\le& \operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)
\quad \text{and} \\
\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) &=& \operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa) .
\end{array}$$
Let $\operatorname{\mathsf{SES}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Enc}},\operatorname{\mathsf{Dec}})$ be a symmetric encryption scheme and $\operatorname{\mathsf{ASA}}=(\operatorname{\mathsf{AGen}},\operatorname{\mathsf{AEnc}},\operatorname{\mathsf{AExt}})$ be an algorithm substitution attack against $\operatorname{\mathsf{SES}}$. To simplify notation, let $\ell={\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)$. We construct the stegosystem $\operatorname{\mathsf{S}}=(\operatorname{\mathsf{SGen}},\operatorname{\mathsf{SEnc}},\operatorname{\mathsf{SDec}})$ on ${\mathcal{C}}$ out of the $\operatorname{\mathsf{ASA}}$.
The key generation algorithm $\operatorname{\mathsf{SGen}}$ simply simulates $\operatorname{\mathsf{AGen}}$. It will output the key ${\textit{ak}}$. To encode a message ${\textit{am}}$ using the key ${\textit{ak}}$, the stegoencoding algorithm $\operatorname{\mathsf{SEnc}}$ generates for any history $h$ a sequence of ${\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)=2\ell+1$ documents such that the last $\ell$ documents embed the message ${\textit{am}}$. To describe the algorithm we need to distinguish between different given histories $h$.
In this case, $\operatorname{\mathsf{SEnc}}$ chooses a random key $k\gets {\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$ using the generation algorithm of $\operatorname{\mathsf{SES}}$ and outputs $k$.
Encoder $\operatorname{\mathsf{SEnc}}$ samples a random message $m_{r+1}$ and outputs it.
The stego-encoder $\operatorname{\mathsf{SEnc}}$ simulates $\operatorname{\mathsf{AEnc}}({\textit{ak}},{\textit{am}},k,m_{(r+1) \bmod \ell +1})$ and outputs the generated ciphertext.
Note that by construction, in any case the last $\ell$ documents generated by $\operatorname{\mathsf{SEnc}}^{2\ell+1}$ embed the message ${\textit{am}}$ in the same way as done by $\operatorname{\mathsf{ASA}}^{\ell}$.
If the decoder $\operatorname{\mathsf{SDec}}$ is given documents $d_{1},\ldots,d_{2\ell+1}$, we output $\operatorname{\mathsf{AExt}}({\textit{ak}},d_{\ell+2},\ldots,d_{2\ell+1})$.
As one can see from the definitions, the decoding algorithm of $\operatorname{\mathsf{S}}$ is history-ignorant. We will prove that on the channel ${\mathcal{C}}={\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{SES}}}(2\ell+1)$ the security and reliability of the stegosystem $\operatorname{\mathsf{S}}$ satisfy the stated conditions.
We first analyze the security of the system. Let $\operatorname{\mathsf{Ward}}$ be a warden against $\operatorname{\mathsf{S}}$ on ${\mathcal{C}}$ with maximal advantage, i.e. $$\begin{aligned}
\operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
=
\operatorname{\mathbf{InSec}}^{\operatorname{cha}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa),
\end{aligned}$$ where $ \operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
= \Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)=\textsf{true}].
$ We will construct a watchdog $\operatorname{\mathsf{Watch}}$ against the algorithm substitution attack $\operatorname{\mathsf{ASA}}$ with the same advantage as $\operatorname{\mathsf{Ward}}$: $$\begin{aligned}
\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}^{\operatorname{enc-watch}}(\kappa) =
\operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa).
\end{aligned}$$ This will prove that $$\label{ineq:insecstego:insecasa}
\operatorname{\mathbf{InSec}}^{\operatorname{cha}}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
\le
\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa).$$
The watchdog $\operatorname{\mathsf{Watch}}$ on input $1^{\kappa}$ simply simulates the warden $\operatorname{\mathsf{Ward}}$. Whenever the warden $\operatorname{\mathsf{Ward}}$ makes a query to its channel oracle ${\mathcal{C}}$ with a history $h$, the watchdog $\operatorname{\mathsf{Watch}}$ simulates the oracle response as follows:
- If $h=\varnothing$, the watchdog uses $\operatorname{\mathsf{Gen}}(1^{\kappa})$ to construct a key $k$ and returns $k$ to the warden.
- If $h=k\mid\mid m_{1}\mid\mid \ldots \mid\mid
m_{r}$ with $r < \ell$, the watchdog uniformly chooses a message $m_{r+1}$ from $\{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$ and outputs $m_{r+1}$.
- If $h=k\mid\mid m_{1}\mid\mid \ldots\mid\mid
m_{\ell}\mid\mid c_{1}\mid\mid \ldots\mid\mid c_{r}$ with $r\geq 0$, the watchdog computes $c_{r+1}\gets \operatorname{\mathsf{Enc}}(k,m_{((r+1)\bmod \ell)+1})$ and outputs $c_{r+1}$.
Clearly, this simulates the channel distribution ${\mathcal{C}}$ perfectly. If the warden queries its challenge oracle ${\operatorname{\mathsf{Ward}}\!.\!\operatorname{CH}}$ with chosen message ${\textit{am}}$, state $\sigma$, and history $h$ (that is either equivalent to sampling from ${\mathcal{C}}_{h}$ or to calling $\operatorname{\mathsf{SEnc}}({\textit{ak}},{\textit{am}},h,\sigma)$), the watchdog simulates the response of the oracle ${\operatorname{\mathsf{Ward}}\!.\!\operatorname{CH}}$ as follows:
- If $h=\varnothing$ then $\operatorname{\mathsf{Watch}}$ chooses a random key $k\gets \operatorname{\mathsf{Gen}}(1^{\kappa})$ and outputs it.
- If $h=k\mid\mid m_1\mid\mid m_2\mid\mid \ldots \mid\mid
m_r$ for $0\leq r\leq \ell-1$ then $\operatorname{\mathsf{Watch}}$ samples a random message $m$ and outputs it.
- If $h=k\mid\mid m_{1}\mid\mid m{_2}\mid\mid \ldots
\mid\mid m_{\ell}\mid\mid c_{1}\mid\mid \ldots \mid\mid
c_{r}$ with $r\geq 0$ then $\operatorname{\mathsf{Watch}}$ queries its own oracle on $k$ and $m_{((r+1)\bmod \ell)+1}$.
If ${\operatorname{\mathsf{Watch}}\!.\!\operatorname{CH}}$ is equal to $\operatorname{\mathsf{Enc}}$ of $\operatorname{\mathsf{SES}}$ (the bit $b$ in $\operatorname{\mathsf{ASA-Dist}}$ is set to $0$) the corresponding answer is identically distributed to a sample of the channel ${\mathcal{C}}$. Hence, $$\begin{aligned}
&\Pr[\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)\langle
b=0\rangle=\textsf{true}]=\\
&\Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\langle b=0\rangle=\textsf{true}].
\end{aligned}$$ On the other hand, if ${\operatorname{\mathsf{Watch}}\!.\!\operatorname{CH}}$ is equal to $\operatorname{\mathsf{AEnc}}$ (the bit $b$ in $\operatorname{\mathsf{ASA-Dist}}$ is set to $1$), the corresponding answer is identically distributed to $\operatorname{\mathsf{SEnc}}({\textit{ak}},{\textit{am}},h,\sigma)$ and thus $$\begin{aligned}
& \Pr[\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)\langle b=1\rangle=\textsf{true}]=\\
&\Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)\langle b=1\rangle=\textsf{true}].
\end{aligned}$$ We thus have $$\begin{array}{ll}
& \operatorname{\mathbf{Adv}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)
\ =\\
&\quad\quad |\Pr[\operatorname{\mathsf{ASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)=\textsf{true}] - 1/2| \ =\\
&\quad\quad |\Pr[\operatorname{\mathsf{SS-CHA-Dist}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)=\textsf{true}] - 1/2| \ =\\
&\quad\quad \operatorname{\mathbf{Adv}}^{\operatorname{cha}}_{\operatorname{\mathsf{Ward}},\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa)
\end{array}$$ which proves .
The reliability of $\operatorname{\mathsf{S}}$ is the same as the success probability of $\operatorname{\mathsf{ASA}}$ since $\operatorname{\mathsf{SDec}}$ simply simulates $\operatorname{\mathsf{AExt}}$.
By using the fact that channels with min-entropy $0$ can not be used for steganography (see [e.g.]{}Theorem 6 in [@hopper2009provably]) and observing that channels corresponding to deterministic encryption schemes have min-entropy $0$, we can conclude the following corollary:
For all deterministic encryption schemes $\operatorname{\mathsf{SES}}$ and all algorithm substitution attacks $\operatorname{\mathsf{ASA}}$ against $\operatorname{\mathsf{SES}}$: $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)\geq 1.\end{aligned}$$
Note that this exactly Theorem 4 in [@bellare2014asa].
General Results {#sec:general}
===============
Let $\operatorname{\mathsf{R}}$ be a polynomial-time randomized algorithm with hardwired secret $s$ which takes inputs $x$ and generates outputs $y$. The general task of an algorithm substitution attack against $\operatorname{\mathsf{R}}$ is to construct a subverted algorithm $\operatorname{\mathsf{AR}}_{{\textit{ak}}}$ which using a hidden hardwired random key ${\textit{ak}}$ outputs on the secret $s$ in the sequence of calls $\operatorname{\mathsf{AR}}_{\textit{ak}}(s,x_1), \operatorname{\mathsf{AR}}_{\textit{ak}}(s,x_2),\ldots$ a sequence such that
the output $\operatorname{\mathsf{AR}}_{\textit{ak}}(s,x_1), \operatorname{\mathsf{AR}}_{\textit{ak}}(s,x_2),\ldots$ is indistinguishable from $\operatorname{\mathsf{R}}(s,x_1), \operatorname{\mathsf{R}}(s,x_2),\ldots$ and
$\operatorname{\mathsf{AR}}_{\textit{ak}}(s,x_1), \operatorname{\mathsf{AR}}_{\textit{ak}}(s,x_2),\ldots$ embeds the secret $s$.
In our setting we model the attack on $\operatorname{\mathsf{R}}$ as a stegosystem on a channel determined by $\operatorname{\mathsf{R}}$ and define such a channel.
ASA against a Randomized Algorithm
----------------------------------
In this section we give formal definitions for algorithm substitution attack $\operatorname{\mathsf{AR}}$, its advantage $\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{AR}},\operatorname{\mathsf{R}}}$, etc. Formally, an *algorithm substitution attack against $\operatorname{\mathsf{R}}$* is a triple of efficient algorithms $\operatorname{\mathsf{ASA}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{AR}},\operatorname{\mathsf{Ext}})$, where $\operatorname{\mathsf{Gen}}$ generates the key ${\textit{ak}}$, the algorithm $\operatorname{\mathsf{AR}}$ takes the key ${\textit{ak}}$, a secret $s$ and all inputs $x_{1},x_{2},\ldots$ to $\operatorname{\mathsf{R}}$ and the extractor $\operatorname{\mathsf{Ext}}$ tries to extract $s$ from the outputs of $\operatorname{\mathsf{AR}}$ with the help of ${\textit{ak}}$ (but without knowing $x_{1},x_{2},\ldots$). Similarly to the setting for encryption schemes, $\operatorname{\mathsf{ASA}}$ is called *indistinguishable*, if every PPTM $\operatorname{\mathsf{Watch}}$ – the *watchdog* – is not able to distinguish between $\operatorname{\mathsf{AR}}_{{\textit{ak}}}(s,x_{1}),\operatorname{\mathsf{AR}}_{{\textit{ak}}}(s,x_{2}),\ldots$ and $\operatorname{\mathsf{R}}(x_{1}),\operatorname{\mathsf{R}}(x_{2}),\ldots$ even if he is allowed to choose $s$ and all $x_{i}$. This is defined via the game $\operatorname{\mathsf{RASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}$ defined analogously to $\operatorname{\mathsf{ASA-Dist}}$. The maximal advantage of any watchdog distinguishing $\operatorname{\mathsf{ASA}}$ from $\operatorname{\mathsf{R}}$ is called the *insecurity* or indistinguishability of $\operatorname{\mathsf{ASA}}$ and is formally defined as $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{asa}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}(\kappa) := \max_{\operatorname{\mathsf{Watch}}}\{\operatorname{\mathbf{Adv}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}^{\operatorname{asa}}(\kappa)\},\end{aligned}$$ where $$\begin{aligned}
&\operatorname{\mathbf{Adv}}^{\operatorname{asa}}_{\operatorname{\mathsf{Watch}}, \operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}(\kappa)
:= \\
&\quad\quad |\Pr[\operatorname{\mathsf{RASA-Dist}}_{\operatorname{\mathsf{Watch}},\operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}(\kappa)=\textsf{true}] - 1/2|.\end{aligned}$$
The *unreliability* of $\operatorname{\mathsf{ASA}}$ is also defined like before: $$\begin{aligned}
&\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}(\kappa) \ :=\\ &\max
\{
\Pr[{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{AExt}}}({\textit{ak}},{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{AR}}}({\textit{ak}},{\textit{am}},x_{1},\ldots,x_{\ell}))\neq
{\textit{am}}]\},\end{aligned}$$ where the maximum is taken over all ${\textit{ak}}\in \operatorname{supp}({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})), \allowbreak
{\textit{am}}\in\{0,1\}^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, and $x_{1},\ldots,x_{\ell}$ being inputs to $\operatorname{\mathsf{R}}$.
Known examples which fit into this setting include [e.g.]{}the subversion-resilient signature schemes presented in the work of Ateniese et al. [@ateniese2015sig].
Channel determined by a Randomized Algorithm
--------------------------------------------
Let $\operatorname{\mathsf{R}}$ be a polynomial-time randomized algorithm with parameter $\kappa$. We assume that the secret $s$ is generated by $\operatorname{\mathsf{Gen}}$ and the inputs $x$ to $\operatorname{\mathsf{R}}$ are generated by the randomized polynomial-time algorithm $\operatorname{\mathsf{GenInput}}$, associated with $\operatorname{\mathsf{R}}$ (which may be chosen adversarially as shown in the definition above). Let $\ell$ be a polynomial of $\kappa$. For $\operatorname{\mathsf{R}}$ we define a channel family, named ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{R}}}(\ell)$, indexed with parameter $\kappa\in {\mathbb{N}}$, with documents which correspond to the input of $\operatorname{\mathsf{AR}}$. The essential idea behind the definition of the channel ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{R}}}(\ell)$ is that for all $s\in \operatorname{supp}(\operatorname{\mathsf{Gen}}(1^{\kappa}))$ and every sequence of inputs $x_{1},x_{2},\ldots, x_{\ell(\kappa)}$, with $x_{i}\in\operatorname{supp}(\operatorname{\mathsf{GenInput}}(1^{\kappa}))$, for the history $$h= s \mid\mid x_{1} \mid\mid x_{2} \mid\mid \ldots \mid\mid x_{\ell(\kappa)}$$ the distribution of the sequences of documents $$\begin{aligned}
y_{1} \mid\mid y_{2} \mid\mid \ldots \mid\mid y_{\ell(\kappa)}
\end{aligned}$$ generated by the channel is exactly the same as the distribution for $$\begin{aligned}
\operatorname{\mathsf{R}}(s,x_{1}) \mid\mid \operatorname{\mathsf{R}}(s,x_{2}) \mid\mid \ldots \mid\mid \operatorname{\mathsf{R}}(s,x_{\ell(\kappa)}).
\end{aligned}$$ To give a formal definition of $\{{\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{R}}}(\ell)\}_{\kappa\in {\mathbb{N}}}$ we need to specify the probability distributions for any history $h$. Thus, we define the family, on the alphabet $\{0,1\}$, as follows: For empty history $h=\varnothing$, we define ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{R}}}(\ell)_{\varnothing}$ as the distribution on all possible keys generated by $\operatorname{\mathsf{Gen}}(1^{\kappa})$. For $s\in \operatorname{supp}(\operatorname{\mathsf{Gen}}(1^{\kappa}))$ and a (possibly empty) sequence inputs $x_{1},x_{2},\ldots, x_{r}$ with $x_{i}\in \operatorname{supp}(\operatorname{\mathsf{GenInput}}(1^{\kappa}))$ and $0\leq r\leq \ell(\kappa)-1$, the distribution ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{R}}}(\ell)_{s\mid\mid x_{1}\mid\mid
x_{2}\mid\mid \ldots \mid\mid x_{r}}$ is the distribution on inputs $x_{r+1}\gets \operatorname{\mathsf{GenInput}}(1^{\kappa})$. For $s\in \operatorname{supp}(\operatorname{\mathsf{Gen}}(1^{\kappa}))$, a sequence of inputs $x_{1},x_{2},\ldots, x_{\ell(\kappa)}$ with $x_{i}\in\operatorname{supp}(\operatorname{\mathsf{GenInput}}(1^{\kappa}))$, and a (possibly empty) sequence of $\operatorname{\mathsf{R}}$’s outputs $y_{1},\ldots,y_{r}$ with $y_{i}\in \operatorname{supp}(\operatorname{\mathsf{R}}(s,x_{((i-1)\bmod
\ell(\kappa))+1}))$, the probability distribution of ${\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{R}}}(\ell)_{s\mid\mid x_{1}\mid\mid
x_{2}\mid\mid \ldots \mid\mid x_{\ell(\kappa)}\mid\mid
y_{1}\mid\mid y_{2}\mid\mid \ldots \mid\mid
\ldots \mid\mid y_{r}}$ is the probability distribution of $\operatorname{\mathsf{R}}(s,x_{(r \bmod \ell(\kappa))+1})$.
Results
-------
The theorems proved in the previous section can simply be generalized by using our general construction of the channel ${\mathcal{C}}^{k}_{\operatorname{\mathsf{R}}}(\ell)$ for the randomized algorithm $\operatorname{\mathsf{R}}$ and the generic stegosystem $\operatorname{\mathsf{RejSam}}^{\operatorname{\mathsf{F}}}$ provided by Theorem \[thm:rejsam:secure\].
\[thm:generic-attack:against:R\] For every randomized algorithm $\operatorname{\mathsf{R}}$, there exists a generic algorithm substitution attack $\operatorname{\mathsf{ASA}}$ against $\operatorname{\mathsf{R}}$ such that $$\begin{array}{rcl}
\operatorname{\mathbf{InSec}}^{\operatorname{\operatorname{\mathsf{ASA}}}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}(\kappa)&\leq&
\mathcal{O}(\operatorname{\mathsf{ml}}(\kappa)^{4}\cdot 2^{-{H_{\infty}}({\mathcal{C}}^{\kappa})})+\\
&&\mathcal{O}(\operatorname{\mathsf{ml}}(\kappa)^{2}\cdot
\exp(-s))+
\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}}}(\kappa),\\[2mm]
\operatorname{\mathbf{UnRel}}^{\star}_{\operatorname{\mathsf{S}},{\mathcal{C}}}(\kappa) &\leq&
2 \operatorname{\mathsf{ml}}(\kappa)^{2}\cdot
\exp(-2^{{H_{\infty}}({\mathcal{C}}^{\kappa})-3}) +\\
&& \operatorname{\mathsf{ml}}(\kappa)^{2}\cdot \exp(-2^{-2} s)+
\operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\operatorname{\mathsf{F}}}(\kappa)
\end{array}$$ where ${\mathcal{C}}:={\mathcal{C}}^{\kappa}_{\operatorname{\mathsf{R}}}({\operatorname{\mathsf{S}}\!.\!\operatorname{\mathsf{ol}}}(\kappa))$.
\[thm:no-attack:against:R\] For all deterministic algorithms $\operatorname{\mathsf{R}}$ and all algorithm substitution attacks $\operatorname{\mathsf{ASA}}$ against $\operatorname{\mathsf{R}}$: $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{asa}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{R}}}(\kappa) = 1.\end{aligned}$$
Theorem \[thm:generic-attack\] is thus just a consequence of Theorem \[thm:generic-attack:against:R\] and Theorem \[thm:min-entropy\] is just a consequence of Theorem \[thm:no-attack:against:R\].
These general results also imply several other results from the literature, for example on signature schemes. Ateniese et al. [@ateniese2015sig] study algorithm substitution attacks[^3] on *signature schemes* $\operatorname{\mathsf{SIG}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Sign}},\operatorname{\mathsf{Vrfy}})$, where
- The *key generator* ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Gen}}}$ produces upon input $1^{\kappa}$ a pair $({\textit{pk}},{\textit{sk}})$ of keys with $|{\textit{pk}}|=|{\textit{sk}}|=\kappa$. We call ${\textit{pk}}$ the *public key* and ${\textit{sk}}$ the *secret key*.
- The *signing algorithm* ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Sign}}}$ takes as input the secret key ${\textit{sk}}$ and a message $m\in \{0,1\}^{{\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{ml}}}}$ of length ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ and produces a signature $\sigma\in \{0,1\}^{{\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{sl}}}(\kappa)}$ of length ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{sl}}}(\kappa)$.
- The *verifying algorithm* ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Vrfy}}}$ takes as input the public key ${\textit{pk}}$, the message $m$ and a signature $\sigma$ and outputs a bit $b$.
On the positive side (from the view of an algorithm substitution attack) they show that all randomized *coin-injective* schemes and all *coin-extractable* schemes have ASA. A randomized algorithm $A$ is *coin-injective*, if the function $f_{A}(x,\rho)=A(x;\rho)$ (where $\rho$ denotes the random coins used by $A$) is injective and *coin-extractable* if there is another randomized algorithm $B$ such that $\Pr[B(A(x;\rho))=\rho] \geq 1-\operatorname{\mathsf{negl}}$ for a negligible function $\operatorname{\mathsf{negl}}$. They prove the following theorems:
For every coin-injective signature scheme $\operatorname{\mathsf{SIG}}$, there is a successful algorithm substitution attack $\operatorname{\mathsf{ASA}}$ and a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{asa}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SIG}}}(\kappa)\leq \operatorname{\mathbf{InSec}}^{\operatorname{prf}}_{\mathsf{F}}(\kappa)+\operatorname{\mathsf{negl}}(\kappa)
\end{aligned}$$ for a pseudorandom function $\mathsf{F}$.
For every coin-extractable signature scheme $\operatorname{\mathsf{SIG}}$, there is a successful algorithm substitution attack $\operatorname{\mathsf{ASA}}$ and a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{asa}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SIG}}}(\kappa)\leq \operatorname{\mathsf{negl}}(\kappa).
\end{aligned}$$
Both of these results are easily implied by Theorem \[thm:generic-attack:against:R\].
On the negative side (from the view of an algorithm substitution attack), they show that *unique signature schemes* are resistant to ASAs fulfilling the *verifiability condition*. Informally this means that (a) each message has exactly on signature (for a fixed key-pair) and (b) each signature produced by the ASA must be valid.
For all unique signature schemes $\operatorname{\mathsf{SIG}}$ and all algorithm substitution attacks $\operatorname{\mathsf{ASA}}$ against them that fulfill the verifiability condition, there is a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
\operatorname{\mathbf{InSec}}^{\operatorname{asa}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SIG}}}(\kappa)\geq 1-\operatorname{\mathsf{negl}}(\kappa).
\end{aligned}$$
As unique signature schemes do not provide enough min-entropy for a stegosystem, this results follows from Theorem \[thm:min-entropy\].
A Lower Bound for Universal ASA {#sec:bound}
===============================
A setting similar to steganography, where *universal* stegosystems exist, that can be used for *any* channel of sufficiently large min-entropy, would be quite useful for attackers that plan to launch algorithm substitution attacks. Such a system would allow them to attack any symmetric encryption scheme *without* knowing the internal specification of the encryption algorithm. A closer look at the results in [@bellare2014asa; @bellare2015asa; @ateniese2015sig] reveals that their attacks do indeed go without internal knowledge of the used encryption algorithm. They only manipulate the random coins used in the encryption process. Note that ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(k,m;r)$ (where $r$ denotes the random coins used by $\operatorname{\mathsf{Enc}}$) is a deterministic function, as ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}$ is a PPTM.
We thus define a *universal algorithm substitution attack* as a triple of PPTMs such that for every symmetric encryption scheme $\operatorname{\mathsf{SES}}$, the triple $$\begin{aligned}
\operatorname{\mathsf{ASA}}^{\operatorname{\mathsf{SES}}}=({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}},{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(\cdot,\cdot; \cdot)},{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Ext}}}) \end{aligned}$$ is an ASA against $\operatorname{\mathsf{SES}}$. Hence, ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}$ has only oracle access to the encryption algorithm ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}$ of the encryption scheme: It may thus choose arbitrary values $k$, $m$, and $r$ and receives a ciphertext $$\begin{aligned}
c\gets {\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(k,m; r) \end{aligned}$$ without having a complete description of the encryption schemes.
As noted above, all attacks in [@bellare2014asa; @bellare2015asa; @ateniese2015sig] are universal and @bellare2015asa explicitly state in their work [@bellare2015asa] that their ASA works against any encryption scheme of sufficiently large min-entropy. We also remark that the rejection sampling ASA presented earlier is universal.
For a universal algorithm substitution attack $\operatorname{\mathsf{ASA}}$ and a symmetric encryption scheme $\operatorname{\mathsf{SES}}$, let ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\operatorname{\mathsf{SES}},\kappa,{\textit{ak}},{\textit{am}},k,m_j,\sigma)$ be the expected number of oracle calls that a single call of the substitution encoder ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(\cdot,\cdot;\cdot)}({\textit{ak}},{\textit{am}},k,m_j,\sigma)$ makes to its encryption oracle ${\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}$. We then define $$\begin{aligned}
&{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\operatorname{\mathsf{SES}},\kappa)=\\
&\max_{\substack{{\textit{ak}}\in\operatorname{supp}({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})),\\ {\textit{am}}\in \{0,1\}^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)},\\ k\in \operatorname{supp}({\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})),\\ m\in \{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)},\\ \sigma\in \{0,1\}^{*}}} \{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\operatorname{\mathsf{SES}},\kappa,{\textit{ak}},{\textit{am}},k,m_j,\sigma)\}.\end{aligned}$$ For a family $\mathcal{F}$ of encryption schemes, let ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\mathcal{F},\kappa)$ be the maximal value of ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\operatorname{\mathsf{SES}},\kappa)$ for $\operatorname{\mathsf{SES}}\in \mathcal{F}$.
In the steganographic setting, @dedic2009upper showed in [@dedic2009upper] that (under the cryptographic assumption that one-way functions exist) no universal stegosystem can embed more than $\mathcal{O}(1)\cdot \log(\kappa)$ bits per document and thus proved that the rejection sampling based systems have optimal rate. The needed ingredients of this proof are summarized by two key lemmas based on Lemma 12 and Lemma 13 in [@Berndt2016].
\[lem:secure\_asa\] Let $\operatorname{\mathsf{ASA}}$ be a algorithm substitution attack for the symmetric encryption scheme $\operatorname{\mathsf{SES}}$ such that $\operatorname{\mathsf{ASA}}$ is secure against $\operatorname{\mathsf{SES}}$. Then for all integers $\kappa\in {\mathbb{N}}$, messages $m\in \{0,1\}^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, ciphertexts $c_{1},c_{2},\ldots,c_{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)}\gets
{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}({\textit{ak}},{\textit{am}},k,m,\sigma)$ and all positions $i\in
\{1,\ldots,{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)\}$: $$\begin{aligned}
\Pr_{{\textit{ak}}\gets {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})}[c_{i}\not\in \operatorname{supp}({\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(k,m))]
\leq \operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa).\end{aligned}$$
\[lem:reliable\_asa\] Let $\operatorname{\mathsf{ASA}}$ be a universal and reliable algorithm substitution attack against the symmetric encryption scheme $\operatorname{\mathsf{SES}}$. Then for every $\kappa$, the probability that the encoder ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}$ produces a ciphertext, which was not provided by the encryption oracle, is at least $$\begin{aligned}
1-\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}}(\kappa)-\frac{({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)\cdot {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\operatorname{\mathsf{SES}},\kappa))^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)}}{2^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}}.
\end{aligned}$$
We will now show how one can modify an existing symmetric encryption scheme $\operatorname{\mathsf{SES}}$ with the help of a signature scheme $\operatorname{\mathsf{SIG}}$ into a family of encryption schemes such that no universal ASA can achieve a super-logarithmic rate on all of these encryption schemes. The construction is very similar to the construction used in [@Berndt2016].
A *signature scheme* $\operatorname{\mathsf{SIG}}=({\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Gen}}},{\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Sign}}},{\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Vrfy}}})$ is a triple of probabilistic polynomial-time algorithms with the following properties:
- The *key generator* ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Gen}}}$ produces upon input $1^{\kappa}$ a pair $({\textit{pk}},{\textit{sk}})$ of keys with $|{\textit{pk}}|=|{\textit{sk}}|=\kappa$. We call ${\textit{pk}}$ the *public key* and ${\textit{sk}}$ the *secret key*.
- The *signing algorithm* ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Sign}}}$ takes as input the secret key ${\textit{sk}}$ and a message $m\in \{0,1\}^{{\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{ml}}}}$ of length ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)$ and produces a signature $\sigma\in \{0,1\}^{{\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{sl}}}(\kappa)}$ of length ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{sl}}}(\kappa)$.
- The *verifying algorithm* ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Vrfy}}}$ takes as input the public key ${\textit{pk}}$, the message $m$ and a signature $\sigma$ and outputs a bit $b$.
We say that $(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Sign}},\operatorname{\mathsf{Vrfy}})$ is *reliable*, if $\operatorname{\mathsf{Vrfy}}({\textit{pk}},m,\operatorname{\mathsf{Sign}}({\textit{sk}},m))=1$ for all ${\textit{pk}}$, ${\textit{sk}}$ and $m$.
A *forger* $\operatorname{\mathsf{Fo}}$ is a probabilistic polynomial time algorithm that upon input ${\textit{pk}}$ and oracle access to $\operatorname{\mathsf{Sign}}_{{\textit{sk}}}$ tries to produce a pair $(m,\sigma)$ such that $\operatorname{\mathsf{Vrfy}}_{{\textit{pk}}}(m,\sigma)=1$. Formally, this is defined via the following experiment $\operatorname{\mathsf{Sig-Forge}}$:
Forger $\operatorname{\mathsf{Fo}}$, Signature Scheme $\operatorname{\mathsf{SIG}}=(\operatorname{\mathsf{Gen}},\operatorname{\mathsf{Sign}},\operatorname{\mathsf{Vrfy}})$ length $\kappa$
$({\textit{pk}},{\textit{sk}}) \gets \operatorname{\mathsf{Gen}}(1^{\kappa})$ $(m,\sigma) \gets \operatorname{\mathsf{Fo}}^{\operatorname{\mathsf{Sign}}_{{\textit{sk}}}}({\textit{pk}})$ Let $Q$ be the set of messages given to $\operatorname{\mathsf{Sign}}_{{\textit{sk}}}$ by $\operatorname{\mathsf{Fo}}$
A signature scheme $\operatorname{\mathsf{SIG}}$ is called *existentially unforgeable*, if for every forger $\operatorname{\mathsf{Fo}}$, there is a negligible function $\operatorname{\mathsf{negl}}$ such that $$\begin{aligned}
\operatorname{\mathbf{Adv}}^{\operatorname{sig}}_{\operatorname{\mathsf{Fo}},\operatorname{\mathsf{SIG}}}(\kappa) :=
\Pr[\operatorname{\mathsf{Sig-Forge}}_{\operatorname{\mathsf{Fo}},\operatorname{\mathsf{SIG}}}(\kappa)=1]\leq \operatorname{\mathsf{negl}}(\kappa).\end{aligned}$$
The maximal advantage of any forger against $\operatorname{\mathsf{SIG}}$ is called the *insecurity* of $\operatorname{\mathsf{SIG}}$ and is defined as $$\begin{aligned}
\operatorname{\mathbf{InSec}}_{\operatorname{\mathsf{SIG}}}^{\operatorname{sig}}(\kappa) := \max_{\operatorname{\mathsf{Fo}}}\{\operatorname{\mathbf{Adv}}^{\operatorname{sig}}_{\operatorname{\mathsf{Fo}},\operatorname{\mathsf{SIG}}}(\kappa)\}.\end{aligned}$$
For $({\textit{pk}},{\textit{sk}})\in
\operatorname{supp}({\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa}))$, let $\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}$ be the encryption scheme with
- ${\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\!.\!\operatorname{\mathsf{Gen}}}={\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}}$, [i.e.]{}the key generation algorithm remains the same.
- The encryption algorithm ${\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\!.\!\operatorname{\mathsf{Enc}}}$ is given as:
key $k$, message $m$
$c \gets {\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(k,m)$ $\gets {\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Sign}}}({\textit{sk}},c)$
- Similarly, the decryption algorithm ${\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\!.\!\operatorname{\mathsf{Dec}}}$ is given as:
key $k$, ciphertext $(c,\sigma)$
By using this family $$\begin{aligned}
\mathcal{F}(\operatorname{\mathsf{SES}},\operatorname{\mathsf{SIG}})=\{\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\}_{({\textit{pk}},{\textit{sk}})\in \operatorname{supp}({\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa}))},\end{aligned}$$
we can derive the following upper bound on the rate of each universal ASA:
Let $\operatorname{\mathsf{SES}}$ be a symmetric encryption scheme, $\operatorname{\mathsf{SIG}}$ be a signature scheme and $\mathcal{F}=\mathcal{F}(\operatorname{\mathsf{SES}},\operatorname{\mathsf{SIG}})$ be defined as above. For every universal algorithm substitution attack $\operatorname{\mathsf{ASA}}$ against $\operatorname{\mathsf{SES}}$, there exist a forger $\operatorname{\mathsf{Fo}}$ on $\operatorname{\mathsf{SIG}}$ with advantage at least $$\begin{aligned}
1-\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\mathcal{F}}(\kappa)-\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\mathcal{F}}(\kappa)-\varphi(\operatorname{\mathsf{ASA}},\kappa)
\end{aligned}$$ for every $\kappa$, where $$\begin{aligned}
\varphi(\operatorname{\mathsf{ASA}},\kappa)=\frac{({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)\cdot {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\mathcal{F},\kappa))^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)}}{2^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}}.
\end{aligned}$$
The proof is analogue to the proof of [@Berndt2016 Theorem 13].
Fix $\kappa\in {\mathbb{N}}$ and $({\textit{pk}},{\textit{sk}})\in \operatorname{supp}({\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa}))$. We will now construct an forger on $\operatorname{\mathsf{SIG}}$ with the help of the algorithm substitution attacker $\operatorname{\mathsf{ASA}}$. Choose a random attacker message ${\textit{am}}^{*}\gets
\{0,1\}^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$, a random attacker key ${\textit{ak}}^{*}\gets {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$, a random message $m^{*}\gets \{0,1\}^{{\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}$ and a random key $k^{*}\gets {\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Gen}}}(1^{\kappa})$.
The forger now simulates the run of the algorithm substitution attack ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}^{{\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\!.\!\operatorname{\mathsf{Enc}}}(\cdot,\cdot;\cdot)}({\textit{ak}}^{*},{\textit{am}}^{*},k^{*},m^{*})$ against the symmetric encryption scheme $\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}$. Whenever ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}$ makes an access $(k,m;r)$ to its encryption oracle, the forger computes $c={\operatorname{\mathsf{SES}}\!.\!\operatorname{\mathsf{Enc}}}(k,m;r)$ and uses its signing oracle ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Sign}}}_{{\textit{sk}}}$ upon $c$. This returns a valid signature $\sigma$ for $c$ and the forger returns $(c,\sigma)$ to ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}$. This simulation hence yields the same result as ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}^{{\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\!.\!\operatorname{\mathsf{Enc}}}(\cdot,\cdot;\cdot)}({\textit{ak}}^{*},{\textit{am}}^{*},k^{*},m^{*})$. Denote the first document produced by the run of the algorithm substitution attack ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}^{{\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\!.\!\operatorname{\mathsf{Enc}}}(\cdot,\cdot;\cdot)}({\textit{ak}}^{*},{\textit{am}}^{*},k^{*},m^{*})$ as $(\widehat{c},\widehat{\sigma})$. By , the probability that the pair $(\widehat{c},\widehat{\sigma})$ does not belong to to the support $\operatorname{supp}({\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}\!.\!\operatorname{\mathsf{Enc}}}(k,m))$ ([i.e.]{}it is no valid ciphertext-signature pair) is bounded by $\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}}(\kappa)$. Furthermore, implies that the probability that $(\widehat{c},\widehat{\sigma})$ is equal to any $(c,\sigma)$ which was given to the ASA is at most $\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}}(\kappa)+\varphi(\operatorname{\mathsf{ASA}},\kappa)$. We can thus conclude that with probability $$\begin{aligned}
1-&\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}}(\kappa)-\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}}(\kappa)-\\
&\frac{({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)\cdot {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}},\kappa))^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)}}{2^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}},
\end{aligned}$$ the ciphertext-signature pair $(\widehat{c},\widehat{\sigma})$ is a valid ciphertext-signature pair and was not produced by the oracle ${\operatorname{\mathsf{SIG}}\!.\!\operatorname{\mathsf{Sign}}}_{{\textit{sk}}}$ The advantage of the forger against the signature scheme $\operatorname{\mathsf{SIG}}$ is thus at least $$\begin{aligned}
1-&\operatorname{\mathbf{InSec}}^{\operatorname{enc-watch}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}}(\kappa)-\operatorname{\mathbf{UnRel}}_{\operatorname{\mathsf{ASA}},\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}}}(\kappa)-\\
&\frac{({\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)\cdot {\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{query}}}(\operatorname{\mathsf{SES}}_{{\textit{pk}},{\textit{sk}}},\kappa))^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ol}}}(\kappa)}}{2^{{\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{ml}}}(\kappa)}},
\end{aligned}$$ The running time of the forger is polynomial in $\kappa$ due to the polynomial running time of ${\operatorname{\mathsf{ASA}}\!.\!\operatorname{\mathsf{Enc}}}$.
This allows us to conclude the following corollary bounding the number of bits embeddable into a single ciphertext by a universal algorithm substitution attack.
\[cor:bound\] There is no universal algorithm substitution attack that embeds more than $\mathcal{O}(1)\cdot \log(\kappa)$ bits per ciphertext (unless one-way functions do not exist).
Conclusions
===========
In this work, we proved that ASAs in the strong undetectability model of Bellare, Jaeger and Kane [@bellare2015asa] are a special case of stegosystems on a certain kind of channels described by symmetric encryption schemes. This gives a rigorous proof of the well-known connection between steganography and algorithm substitution attacks. We make use of this relationship to show that a wide range of results on ASAs are already present in the steganographic literature. Inspired by this connection, we define *universal ASAs* that work with no knowledge on the internal implementation of the symmetric encryption schemes and thus work for *all* such encryption schemes with sufficiently large min-entropy. As almost all known ASAs are universal, we investigate their rate – the number of embedded bits per ciphertext – and prove a logarithmic upper bound of this rate.
[00]{}
\#1 \#1[[DOI:]{}0[\#1]{} ]{} \#1 \#1 \#1 \#1 \#1[\#1]{} \#1[\#1]{} \#1[\#1]{}
[anderson1996limits]{} . . In , Vol. . , .
[ateniese2015sig]{} . . In . ACM, .
[backes2005public]{} . . In , Vol. . , .
[ball2013revealed]{} . . ().
[bellare1997concrete]{} . . In . , .
[bellare2015asa]{} . . In . ACM, .
[bellare2014asa]{} . . In , Vol. . .
[bellare1993randomoracle]{} . . In . , .
[Berndt2016]{} . . In , Vol. . , .
[Cachin2004]{} . . , (), .
[checkoway2014dual]{} . . In . , .
[dedic2009upper]{} . . , (), .
[degabriele2015asa]{} . . In , Vol. . , .
[greenwald2014no]{} . . .
[hopper2002provably]{} . . In . , Vol. . , .
[hopper2009provably]{} . . , (), .
[lindell2007introduction]{} . . .
[katzenbeisser2002defining]{} . . In . SPIE, .
[mitzenmacher2005probability]{} . . .
[pasquini2017decoy]{} . . In , Vol. . Springer, .
[perlroth2013nsa]{} . . ().
[russell2016cliptography]{} . . In , Vol. . Springer, .
[russell2016destroying]{} . (), .
[schneier2007did]{} . . ().
[schneier2015survey]{} . . (), .
[shumow2007back]{} . . . ().
[simmons1984prisoners]{} . . In . Springer, .
[simmons1998history]{} . . , (), .
[young1996dark]{} . . In , Vol. . Springer, .
[young1997kleptography]{} . . In , Vol. . Springer, .
[^1]: <https://www.openssl.org/>
[^2]: In [@bellare2015asa], this is called the *key recovery security*.
[^3]: To be more precise, their attacks only replace the signing algorithm $\operatorname{\mathsf{Sign}}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $f$ be a Hecke-Maass or holomorphic primitive cusp form of arbitrary level and nebentypus, and let $\chi$ be a primitive character of conductor $M$. For the twisted $L$-function $L(s,f\otimes \chi)$ we establish the hybrid subconvex bound $$L\left(\frac{1}{2}+it,f\otimes\chi\right)\ll (M(3+|t|))^{\frac{1}{2}-\frac{1}{18}+\varepsilon},$$ for $t\in \mathbb R$. The implied constant depends only on the form $f$ and $\varepsilon$.'
address: 'School of Mathematics, Tata Institute of Fundamental Research, 1 Homi Bhabha Road, Colaba, Mumbai 400005, India.'
author:
- Ritabrata Munshi
title: 'The circle method and bounds for $L$-functions - I'
---
Introduction {#intro}
============
Statement of result
-------------------
In this paper we prove the following hybrid subconvex bound:
Let $f$ be a Hecke-Maass or holomorphic primitive cusp form for $\Gamma_0(P)\subset SL(2,\mathbb Z)$ with nebentypus $\psi$. Let $\chi$ be a primitive Dirichlet character of modulus $M$. Then we have $$L\left(\tfrac{1}{2}+it,f\otimes\chi\right)\ll_{f,\varepsilon} (MT)^{\frac{1}{2}-\frac{1}{18}+\varepsilon},$$ where $T=3+|t|$.
Our result beats the convexity bound $L\left(\tfrac{1}{2}+it,f\otimes\chi\right)\ll_{f,\varepsilon} (MT)^{\frac{1}{2}+\varepsilon}$, simultaneously in the $t$ and $M$ aspect. This is one of the most extensively studied subconvexity problems in the literature. In the $t$-aspect, subconvexity was accomplished by Good [@G] (in the holomorphic case of full level) and Meurman [@Me-2] (in the case of Maass forms of full level). For number fields, $t$-aspect subconvexity was achieved by Cogdell, Piatetski-Shapiro and Sarnak [@CPS], Petridis and Sarnak [@PS] and Diaconu and Garrett [@DG]. The first subconvex bound in the $M$ aspect (for $t=0$) was obtained by Duke, Friedlander and Iwaniec [@DFI-1] (for holomorphic forms of full level), Bykovskii [@By] (for general holomorphic forms) and later by Harcos [@H], Michel [@Mi] and Blomer, Harcos and Michel [@BHM] (for Maass forms). Over number fields this was established by Venkatesh [@V] and Blomer and Harcos [@BH-2].\
The first hybrid subconvex bound was given by Blomer and Harcos [@BH]. They obtained $$L\left(\tfrac{1}{2}+it,f\otimes\chi\right)\ll (MT)^{\frac{1}{2}-\frac{1}{40}+\varepsilon},$$ for $f$ Maass or holomorphic form of general level. They however only tackle the case of trivial nebentypus. In [@M] we established the weaker exponent $\frac{1}{2}-\frac{1}{118}$, for holomorphic $f$ of general level and general nebentypus. For number fields, Michel and Venkatesh [@MV] have established such a hybrid subconvex bound with an unspecified exponent.\
At present we have the Weyl-type bound in the $t$-aspect, and the Burgess-type bound in the $M$-aspect [@BH-2] over general number fields, under the Ramanujan conjecture. (This is expected to be the natural boundary of the current methods.) To complete the story one seeks to prove a Burgess-type hybrid bound over general number fields (assuming Ramanujan conjecture). But this has not been achieved yet, not even over $\mathbb Q$. In the context of subconvexity, the only Burgess-type hybrid bound known is due to Heath-Brown [@HB] in the case of Dirichlet $L$-function $L(s,\chi)$.
Sketch of proof
---------------
We will briefly describe our method, which we believe to be new. The experts will have no problem in filling the gaps. We start with the approximate functional equation which reduces the problem to getting cancellation in the sum $$\sum_{n\sim N}\lambda_f(n)\chi(n)n^{-it},$$ for $N\ll (MT)^{1+\varepsilon}$, where $\lambda_f(n)$ are the normalized Fourier coefficients of $f$. Let us consider the worst case scenario namely $N=MT$. We write this sum as $$\mathop{\sum\sum}_{n,m\sim N}\lambda_f(n)\chi(m)m^{-it}\delta_{n,m}$$ where $\delta_{n,m}$ is the Kronecker symbol. Then we apply additive harmonics - the circle method, to detect the equation $n=m$ (see Section \[setup\]). Suppose we use the $\delta$-method of Duke, Friedlander and Iwaniec. Roughly speaking this will yield $$\frac{1}{Q^2}\sum_{q\leq Q}\;\sideset{}{^\star}\sum_{a\bmod{q}}\mathop{\sum\sum}_{n,m\sim N}\lambda_f(n)\chi(m)m^{-it}e_q(an-am)$$ where $e_q(x)=e^{2\pi ix/q}$ and $Q=\sqrt{N}=\sqrt{MT}$.\
Next we apply the Poisson summation on the sum over $m$ and the Voronoi summation on the sum over $n$ (see Section \[gl1\_twists\]). This will reduce the above sum roughly to (see Lemma \[sum-form\]) $$\sum_{q\leq \sqrt{MT}}\mathop{\sum\sum}_{\substack{|n|\ll 1\\|m|\ll \sqrt{MT}}}\lambda_f(n)\bar\chi(m)\chi(q)m^{it}q^{-it}e_q(-M\bar m n).$$ (Here one needs to use stationary phase to deal with some exponential integrals.) Trivial estimation of this sum yields the convexity bound. Any further saving will give us subconvexity. However it is not useful to execute the sum over $q$, as the modulus of the sum (after reciprocity) is of size $(MT)^{3/2}$, whereas the length is just $\sqrt{MT}$. As in [@M] we can now use the bound for bilinear forms with Kloosterman fractions, and get a subconvexity result with exponent $\frac{1}{2}-\frac{1}{118}$. However we will avoid this result, and as an alternative use a very simple idea to get a better exponent.\
Suppose the collection of moduli $q$ we are using in the circle method, has a multiplicative structure - namely each $q$ factorizes uniquely as $q=q_1q_2$, with $q_1\leq Q_1$ and $q_2\leq Q_2$ (with $Q_1Q_2=\sqrt{MT}$). Then applying Cauchy to the above sum one arrives at $$\sqrt{Q_2}(MT)^{\frac{1}{4}}\left[\sum_{q_2\leq Q_2}\mathop{\sum\sum}_{\substack{|n|\ll 1\\|m|\ll \sqrt{MT}}}\left|\sum_{q_1\leq Q_1}\chi(q_1)q_1^{-it}e_{q_1q_2}(-M\bar m n)\right|^2\right]^{\frac{1}{2}}.$$ Again for subconvexity we just need some cancellation in the remaining sum. We see that we need $Q_1$ to have some size, because that is exactly the amount we save in the diagonal. For the off-diagonal we will again apply Poisson on the sum over $m$. We save in the off-diagonal as long as the modulus which is of the size $Q_1^2Q_2=(MT)/Q_2$ is smaller than the square of the length of the $m$-sum, which is $\sqrt{MT}$. So the off-diagonal will be satisfactory if $Q_2$ has some size. (We explain this in Section \[conclusion\].)\
Of course to get an inbuilt bilinear structure in the circle method itself, we need to use a more flexible version of the circle method - the one investigated by Jutila. This version comes with an error term which is satisfactory, as we shall find out, as long as we allow the moduli to be slightly larger than $\sqrt{MT}$ (see Section \[setup\]). For another application of this idea see [@M2]. Further applications of this idea in the context of subconvexity will be given in a follow up paper.
Preliminaries {#prelim}
=============
Preliminaries on Maass forms
----------------------------
For the sake of exposition we shall only present the case of Maass forms of weight $0$, level $P$ and nebentypus $\psi$. The case of holomorphic forms is just similar (or even simpler). Let $f:\mathbb H\rightarrow \mathbb C$ be a Hecke-Maass cusp form with Laplace eigenvalue $\frac{1}{4}+\mu^2\geq 0$, and with Fourier expansion $$\sqrt{y}\sum_{n\neq 0}\lambda_f(n)K_{i\mu}(2\pi|n|y)e(nx).$$ Let $\chi$ be a primitive Dirichlet character of modulus $M$. For simplicity we will assume that $(P,M)=~1$. The twisted $L$-series $L(s,f\otimes\chi)$, which in the right half plane $\sigma>1$ is defined by the absolutely convergent Dirichlet series $L(s,f\otimes\chi)=\sum_{n=1}^\infty \lambda_f(n)\chi(n)n^{-s}$, extends to an entire function and satisfies the functional equation $\Lambda(s,f\otimes \chi)=\varepsilon(f\otimes\chi)\Lambda(1-s,f\otimes\chi)$. The completed $L$-function is given by $$\Lambda(s,f\otimes\chi)=\left(\frac{\sqrt{P}M}{\pi}\right)^s\Gamma\left(\frac{s+\delta+i\mu}{2}\right)\Gamma\left(\frac{s+\delta-i\mu}{2}\right)L(s,f\otimes\chi)$$ where $\delta=0, 1$ depending on the parity of $f\otimes\chi$. The root number satisfies $|\varepsilon(f\otimes\chi)|=1$.\
The functional equation, together with the Stirling approximation and Phragmen-Lindelöf principle, implies the convexity bound $L\left(\tfrac{1}{2}+it,f\otimes\chi\right)\ll_{f,\varepsilon} (MT)^{\frac{1}{2}+\varepsilon}$. Alternatively the functional equation yields an expression for the $L$-values $L(\frac{1}{2}+it,f\otimes\chi)$ as a rapidly convergent series, called the approximate functional equation. Taking a dyadic subdivision of the approximate functional equation, we get the bound $$\begin{aligned}
\label{afe}
L\left(\tfrac{1}{2}+it,f\otimes\chi\right)\ll_{f,A} \sum_{N \;\text{dyadic}}\frac{1}{\sqrt{N}}\left|\sum_{n\in\mathbb Z}\lambda_f(n)\chi(n)n^{-it}h\left(\frac{n}{N}\right)\right|\left(1+\frac{N}{MT}\right)^{-A},\end{aligned}$$ where $h$ is a smooth function supported in $[1,2]$, $A>0$, $t\in\mathbb R$ and $T=(3+|t|)$. Estimating the inner sums using Cauchy and the Rankin-Selberg bound $\sum_{1\leq n\leq x}|\lambda_f(n)|^2\ll_{f,\varepsilon}x^{1+\varepsilon}$, one recovers the convexity bound $L\left(\tfrac{1}{2}+it,f\otimes\chi\right)\ll_{f,\varepsilon} (MT)^{\frac{1}{2}+\varepsilon}$.
Voronoi summation formula {#vsf}
-------------------------
We will use the following Voronoi type summation formula. This was first established by Meurman [@Me] in the case of full level.
Let $f$ be as above, let $v$ be compactly supported smooth function on $(0,\infty)$, and suppose $P|q$ and $(a,q)=1$. We have $$\begin{aligned}
\label{voronoi2}
\sum_{n=1}^\infty \lambda_f(n)e_q\left(an\right)v(n)=\frac{\bar\psi(a)}{q}\sum_{\pm}\sum_{n=1}^\infty \lambda_f(\mp n)e_q\left(\pm\bar{a}n\right)V^{\pm}\left(\frac{n}{q^2}\right)\end{aligned}$$ where $\bar{a}$ is the multiplicative inverse of $a\bmod{q}$, and $$\begin{aligned}
V^-(y)=&-\frac{\pi}{\cosh \pi\mu}\int_0^\infty v(x)\{Y_{2i\mu}+Y_{-2i\mu}\}\left(4\pi\sqrt{xy}\right)dx\\
V^+(y)=&4\cosh \pi\mu\int_0^\infty v(x)K_{2i\mu}\left(4\pi\sqrt{xy}\right)dx.\end{aligned}$$
Note that if $v$ is supported in $[Y,2Y]$, satisfying $y^jv^{(j)}(y)\ll_j 1$, then the sums on the right hand side of are essentially supported on $n\ll q^2(qY)^{\varepsilon}/Y$ (where the implied constant depends on the form $f$ and $\varepsilon$). The contribution from the terms with $n\gg q^2(qY)^{\varepsilon}/Y$ is negligibly small. For smaller values of $n$ we will use the trivial bound $V^{\pm}(n/q^2)\ll Y$.
Circle method
-------------
We will be using a variant of the circle method, with overlapping intervals, which has been investigated by Jutila ([@J-1], [@J-2]). For any set $S \subset \mathbb R$, let $\mathbb I_S$ denote the associated characteristic function, i.e. $\mathbb I_S(x)=1$ for $x\in S$ and $0$ otherwise. For any collection of positive integers $\mathcal Q \subset [1,Q]$ (which we call the set of moduli), and a positive real number $\delta$ in the range $Q^{-2}\ll \delta \ll Q^{-1} $, we define the function $$\tilde I_{\mathcal Q,\delta} (x)=\frac{1}{2\delta L}\sum_{q\in\mathcal Q}\;\sideset{}{^\star}\sum_{a\bmod{q}}\mathbb{I}_{[\frac{a}{q}-\delta,\frac{a}{q}+\delta]}(x),$$ where $L=\sum_{q\in\mathcal Q}\phi(q)$. This is an approximation for $\mathbb I_{[0,1]}$ in the following sense:
\[jutila-lemma\] We have $$\begin{aligned}
\label{jutila}
\int_0^1\left|1-\tilde I_{\mathcal Q,\delta}(x)\right|^2dx\ll \frac{Q^{2+\varepsilon}}{\delta L^2}.\end{aligned}$$
This is a simple consequence of the Parseval theorem from Fourier analysis.
Application of circle method {#setup}
============================
We will apply the circle method directly to the smooth sum $$S(N)=\sum_{n\in \mathbb Z}\lambda_f(n)\chi(n)n^{-it}h\left(\frac{n}{N}\right),$$ which appears in . Let’s recall that the function $h$ is smooth, supported in $[1,2]$, and satisfies the bound $h^{(j)}(x)\ll 1$, where the implied constant depends only on $j$. We shall approximate $S(N)$ by $$\tilde S(N)=\frac{1}{L}\sum_{q\in\mathcal Q}\;\sideset{}{^\star}\sum_{a\bmod{q}}\mathop{\sum\sum}_{n,m\in \mathbb Z}\lambda_f(n)\chi(m)m^{-it}e_q(a(n-m))F(n,m)$$ where $e_q(x)=e^{2\pi ix/q}$, and $
F(x,y)=h\left(\frac{x}{N}\right)h^\star\left(\frac{y}{N}\right)\frac{1}{2\delta}\int_{-\delta}^{\delta}e(\alpha(x-y))d\alpha.
$ Here $h^\star$ is another smooth function with compact support in $(0,\infty)$, and $h^\star(x)=1$ for $x$ in the support of $h$. Also we choose $\delta=N^{-1}$ so that $
\frac{\partial^{i+j}}{\partial^i x\partial^j y}F(x,y)\ll_{i,j} \frac{1}{N^{i+j}}.
$\
\[circ\] Let $\mathcal Q\subset [1,Q]$, with $L=\sum_{q\in\mathcal Q}\phi(q)\gg Q^{2-\varepsilon}$ and $\delta=N^{-1}$. Then we have $$S(N)=\tilde S(N)+O_{f,\varepsilon}\left(N\frac{\sqrt{N}(QN)^{\varepsilon}}{Q}\right).$$
Set $$G(x)=\mathop{\sum\sum}_{n,m\in \mathbb Z}\lambda_f(n)\chi(m)m^{-it}h\left(\frac{n}{N}\right)h^\star\left(\frac{m}{N}\right)e(x(n-m)).$$ Observe that $S(N)=\int_0^1G(x)dx$ and $\tilde S(N)=\int_0^1\tilde I_{\mathcal Q,\delta} (x)G(x)dx$. Hence $$\left|S(N)-\tilde S(N)\right|\leq \int_0^1\left|1-\tilde I_{\mathcal Q,\delta} (x)\right|\left|\mathop{\sum}_{n\in \mathbb Z}\lambda_f(n)e(xn)h\left(\frac{n}{N}\right)\right|\left|\sum_{m\in\mathbb Z}\chi(m)m^{-it}e(-xm)h^\star\left(\frac{m}{N}\right)\right|dx.$$ For the middle sum we have the point-wise bound $\mathop{\sum}_{n\in \mathbb Z}\lambda_f(n)e(xn)h\left(\frac{n}{N}\right)\ll_{f,\varepsilon} N^{\frac{1}{2}+\varepsilon}$. Using Cauchy we now arrive at $$\left|S(N)-\tilde S(N)\right|\ll_{f,\varepsilon} N^{\frac{1}{2}+\varepsilon}\left[\int_0^1\left|1-\tilde I_{\mathcal Q,\delta} (x)\right|^2dx\right]^{\frac{1}{2}}\left[\int_0^1\left|\sum_{m\in\mathbb Z}\chi(m)m^{-it}e(-xm)h^\star\left(\frac{m}{N}\right)\right|^2dx\right]^{\frac{1}{2}}.$$ For the last sum we open the absolute value square and execute the integral. We are left with only the diagonal, which has size $N$. For the other sum we use Lemma \[jutila-lemma\]. It follows that $$\left|S(N)-\tilde S(N)\right|\ll_{f,\varepsilon} \frac{(QN)^{1+\varepsilon}}{\sqrt{\delta}L}\ll_{f,\varepsilon} N\frac{\sqrt{N}(QN)^{\varepsilon}}{Q}.$$
We will choose the set of moduli in Section \[conclusion\]. We pick the size of the moduli to be $Q=N(MT)^{\eta-\frac{1}{2}}$, so that the contribution of the error term in Lemma \[circ\] to is bounded by $(MT)^{\frac{1}{2}-\eta+\varepsilon}$. For $N\leq (MT)^{1-2\eta}$, the trivial bound for $S(N)$ is good enough for our purpose. Now we proceed towards the estimation of $\tilde S(N)$.
Estimation of $\tilde S(N)$ {#gl1_twists}
===========================
Applying Poisson and Voronoi summation
--------------------------------------
We will now assume that each member of $\mathcal Q$ is a multiple of $P$, the level of the Maass form $f$, and is coprime to $M$, the modulus of the character $\chi$. Set $$\begin{aligned}
\label{approx-sn}
\tilde S_x(N)=\frac{1}{L}\sum_{q\in\mathcal Q}\;\sideset{}{^\star}\sum_{a\bmod{q}}\mathop{\sum\sum}_{n,m\in \mathbb Z}\lambda_f(n)\chi(m)m^{-it}e_q(a(n-m))h\left(\frac{n}{N}\right)h^\star\left(\frac{m}{N}\right)e(x(n-m))\end{aligned}$$ so that $\tilde S(N)=(2\delta)^{-1}\int_{-\delta}^{\delta}\tilde S_x(N)dx$.
\[sum-form\] We have $$\begin{aligned}
\label{sum-form-eq}
\tilde S_x(N)=\tfrac{N^{1-it}\psi(M)\varepsilon_{\chi}}{\sqrt{M}L}\sum_{q\in\mathcal Q}\tfrac{\chi(q)}{q}\;\sum_{\pm}\sum_{n=1}^\infty\mathop{\sum}_{\substack{m\in \mathbb Z\\(m,q)=1}}\lambda_f(\mp n)\bar\chi(m)\bar\psi(m)e_q(\pm M\bar mn)H_x^\star(m;q)V_x^{\pm}\left(\tfrac{n}{q^2}\right),\end{aligned}$$ where $\varepsilon_{\chi}$ is the sign of the Gauss sum associated to $\chi$, $H_x^\star$ is defined in , and $V_x^{\pm}$ are as in corresponding to $v_x(n)=h\left(\frac{n}{N}\right)e(xn)$.
First we apply the Poisson summation formula to the sum over $m$ in , after breaking it up modulo $Mq$. This gives $$\mathop{\sum}_{m\in \mathbb Z}\chi(m)m^{-it}e_q(-am)h^\star\left(\frac{m}{N}\right)e(-xm)=\frac{N^{1-it}\varepsilon_M}{\sqrt{M}}\mathop{\sum}_{\substack{m\in \mathbb Z\\m\equiv Ma\bmod{q}}}\bar\chi(m)\chi(q)H_x^\star(m;q),$$ where $$\begin{aligned}
\label{h-trans}
H_x^\star(m;q)=\int_{\mathbb R}h^\star(y)y^{-it}e(-xyN)e_{Mq}(-mNy)dy.\end{aligned}$$ To the sum over $n$ we apply Voronoi summation formula to get $$\sum_{n=1}^\infty \lambda_f(n)e_q(an)h\left(\frac{n}{N}\right)e(xn)=\frac{\bar\psi(a)}{q}\sum_{\pm}\sum_{n=1}^\infty \lambda_f(\mp n)e_q\left(\pm\bar{a}n\right)V_x^{\pm}\left(\frac{n}{q^2}\right),$$ where $V_x^{\pm}$ are as in corresponding to $v_x(n)=h\left(\frac{n}{N}\right)e(xn)$.
Estimates for the integrals
---------------------------
By repeated integration by parts it follows that $H_x^\star(m;q)$, as given in , is negligibly small if $|m|\gg MTQ^{1+\varepsilon}N^{-1}$, where $T=3+|t|$. For smaller values of $m$, we change variables to get $$\begin{aligned}
\label{h-trans-est}
H_x^\star(m;q)=&\frac{|m|^{it}}{m}\left(\frac{Mq}{N}\right)^{1-it}\int_{\mathbb R}h^\star\left(\frac{Mqy}{mN}\right)e\left(-\frac{Mqxy}{m}\right)|y|^{-it} e(-y)dy\\
\nonumber =&\frac{|m|^{it}}{m}\left(\frac{Mq}{N}\right)^{1-it}H_x^\sharp(m;q).\end{aligned}$$ Differentiating under the integral sign and using the second derivative bound for the exponential integral we get $$\begin{aligned}
\label{h-trans-est-2}
u^j\frac{\partial^j}{\partial u^j}H_x^\sharp(u;q)\ll_j \frac{|u|N}{Mq\sqrt{T}}\end{aligned}$$ (as $|x|\ll N^{-1}$). For the $n$-sum in Lemma \[sum-form\], we use the properties of the integral $V_x^{\pm}$ which we have noted in Section \[vsf\]. The effective support of the $n$-sum is given by $1\leq |n|\ll Q^{2+\varepsilon}/N\ll (MT)^{2\eta}Q^{\varepsilon}$. (So the $n$-sum is short and the $m$-sum is relatively long.) For small $n$ we will use the trivial bound $$V_x^{\pm}\left(\frac{n}{q^2}\right)\ll N.$$\
We will now use these bounds to show that the contribution from small $m$ in is good enough for our purpose. Let $$\begin{aligned}
\tilde S_x(N;X)=\frac{N^{1-it}\psi(M)\varepsilon_{\chi}}{\sqrt{M}L}\sum_{q\in\mathcal Q}\frac{\chi(q)}{q}\;\sum_{\pm}\sum_{n=1}^\infty\mathop{\sum}_{\substack{|m|\sim X\\(m,q)=1}}\lambda_f(\mp n)\bar\chi(m)\bar\psi(m)e_q(\pm M\bar mn)H_x^\star(m;q)V_x^{\pm}\left(\frac{n}{q^2}\right),\end{aligned}$$ where $|m|\sim X$ means that $X\leq |m|<2X$.
We have $$\begin{aligned}
\tilde S_x(N;X)\ll \frac{NX}{\sqrt{MT}}Q^{\varepsilon}.\end{aligned}$$
This lemma yields the bound $\tilde S(N)\ll N(MT)^{\eta}$. Hence for larger value of $X$ any further saving will yield subconvexity. One way will be to appeal to the large sieve inequality of Duke, Friedlander and Iwaniec for Kloosterman fractions. But a more interesting and fruitful way will be to use the flexibility of the set $\mathcal Q$ to build a bilinear structure in the circle method itself.
Estimation of $\tilde S(N)$ : conclusion {#conclusion}
========================================
Applying Cauchy and Poisson
---------------------------
We choose the set of moduli $\mathcal Q$ to be the product set $P\mathcal Q_1 \mathcal Q_2$, where $\mathcal Q_i$ consists of primes in the dyadic segment $[Q_i,2Q_i]$ (and not dividing $PM$) for $i=1,2$, and $Q_1Q_2=Q=N(MT)^{\eta-\frac{1}{2}}$. Also we pick $Q_1$ and $Q_2$ (whose optimal sizes will be determined later) so that the collections $\mathcal Q_1$ and $\mathcal Q_2$ are disjoint. Using Cauchy we have $$\tilde S_{x}(N;X)\ll Q^{\varepsilon}\frac{\sqrt{MQ_2}}{Q^2}\left\{\sum_{q_2\in\mathcal Q_2}\;\sum_{\pm}\tilde S_{x}(N;X,q_2,\pm)\right\}^{\frac{1}{2}}+N^{-A}.$$ where $A>0$, and $\tilde S_{x}(N;X,q_2,\pm)$ is given by $$\mathop{\sum}_{\substack{m\in \mathbb Z\\(m,Pq_2)=1}}\frac{W\left(\frac{m}{X}\right)}{m}\left|\sum_{\substack{q_1\in\mathcal Q_1\\(m,q_1)=1}}\chi(q_1)\sum_{1\leq n\ll \frac{Q^2}{N}Q^{\varepsilon}}\lambda_f(\mp n)e_{Pq_1q_2}(\pm M\bar mn)H_x^\sharp(m;Pq_1q_2)V_x^{\pm}\left(\frac{n}{Pq_1^2q_2^2}\right)\right|^2$$ with $(MT)^{\frac{1}{2}-\eta}Q^{-\varepsilon}\ll X\ll (MT)^{\frac{1}{2}+\eta}Q^{\varepsilon}$ and $W$ is non-negative smooth function supported in $[-2,2]$ such that $W(x)=1$ for $x\in [-1,1]$. Opening the absolute square and interchanging the order of summations we get $$\begin{aligned}
\label{final}
\tilde S_{x}(N;X,q_2,\pm)=\mathop{\sum\sum}_{q_1, q_1'\in\mathcal Q_1}\chi(q_1\overline{q_1'})\mathop{\sum\sum}_{1\leq n, n'\ll \frac{Q^2}{N}Q^{\varepsilon}}\lambda_f(\mp n)\bar \lambda_f(\mp n')V_x^{\pm}\left(\frac{n}{Pq_1^2q_2^2}\right)V_x^{\pm}\left(\frac{n'}{Pq_1'^2q_2^2}\right)\mathcal T\end{aligned}$$ where $$\mathcal T=\sum_{\substack{m\in\mathbb Z\\(m,Pq_1q_1'q_2)=1}}e_{Pq_1q_2}(\pm M\bar mn)e_{Pq_1'q_2}(\mp M\bar mn')H_x^\sharp(m;Pq_1q_2)\bar H_x^\sharp(m;Pq_1'q_2)\frac{1}{m}W\left(\frac{m}{X}\right).$$ Now again we apply Poisson to the sum over $m$. With this we arrive at
We have $$\begin{aligned}
\mathcal T=\frac{1}{Pq_1q_1'q_2}\sum_{m\in\mathbb Z}S\left(\mp M(q_1n'-q_1'n),m;Pq_1q_1'q_2\right)\mathcal I\end{aligned}$$ where $S\left(\mp M(q_1n'-q_1'n),m;Pq_1q_1'q_2\right)$ is the Kloosterman sum, and $$\begin{aligned}
\mathcal I=\int_{\mathbb R}y^{-1}W\left(y\right) H_x^\sharp(yX;Pq_1q_2)\bar H_x^\sharp(yX;Pq_1'q_2)e\left(-\frac{Xm}{Pq_1q_1'q_2}y\right)dy.\end{aligned}$$
Using and repeated integration by parts we find that $\mathcal I$ is negligibly small unless $|m|\ll Q_1Q^{1+\varepsilon}X^{-1}$. For smaller values of $m$ we use the bound $$\mathcal I\ll \frac{X^2N^2}{M^2Q^2T}\asymp \frac{X^2}{M(MT)^{2\eta}},$$ which again follows from .
Final estimates
---------------
Using the Weil bound for the Kloosterman sums we get $$\mathcal T\ll \frac{X^2}{M(MT)^{2\eta}Q_1\sqrt{Q_2}}\sum_{|m|\ll \frac{Q_1Q}{X}Q^{\varepsilon}}\left(M(q_1n'-q_1'n),m,q_1q_1'q_2\right)^{\frac{1}{2}}.$$ Suppose $Q_1\gg (MT)^{2\eta}Q^{\varepsilon}$. Since $\mathcal Q_1$ consists of primes (not dividing $M$), it follows that $q_1|(q_1n'-q_1'n)$ if and only if $q_1=q_1'$, as $1\leq |n|\ll (MT)^{2\eta}$ and $q_1$ is a prime of larger size. Also taking $Q_1\ll (MT)^{\frac{1}{4}-\frac{\eta}{2}}Q^{-\varepsilon}$, we can guarantee that each $q_2\in\mathcal Q_2$ is a large enough prime, so that $q_2|m$ if and only if $m=0$. Also as $N>(MT)^{1-2\eta}$ and $Q_1Q_2=N(MT)^{\eta-\frac{1}{2}}$, it follows from the bound $Q_1\ll (MT)^{\frac{1}{4}-\frac{\eta}{2}}Q^{-\varepsilon}$ that $Q_2>Q_1(MT)^{2\eta}Q^{\varepsilon}$. Hence $q_2|(q_1n'-q_1'n)$ if and only if $q_1=q_1'$ and $n=n'$. Using these observations we first conclude that $$\begin{aligned}
\label{sum-t}
\mathcal T\ll \frac{XQ_1\sqrt{Q_2}}{M(MT)^{2\eta}}+\frac{X^2}{M(MT)^{2\eta}Q_1\sqrt{Q_2}}\left((q_1n'-q_1'n),q_1q_1'q_2\right)^{\frac{1}{2}},\end{aligned}$$ where the first term is the contribution of non-zero $m$, and the last term accounts for $m=0$. Also it follows that the gcd $\left((q_1n'-q_1'n),q_1q_1'q_2\right)=1$ if $q_1\neq q_1'$, $\left((q_1n'-q_1'n),q_1q_1'q_2\right)=q_1$ if $q_1=q_1'$ but $n\neq n'$, and $\left((q_1n'-q_1'n),q_1q_1'q_2\right)=q_1^2q_2$ if $q_1= q_1'$ and $n=n'$.\
Next we substitute the bound for $\mathcal T$ in . We use Cauchy and the bound $\sum_{1\leq n\leq x}|\lambda_f(n)|^2\ll x^{1+\varepsilon}$ for the Fourier coefficients to conclude - $$\tilde S_{x}(N;X)\ll \left(\frac{N^{\frac{3}{2}}(MT)^{\eta}}{\sqrt{MT}Q_2^{\frac{1}{4}}}+\frac{\sqrt{MNT}}{\sqrt{Q_1}}\right)Q^{\varepsilon}+N^{-A}.$$ The first term on the right hand side accounts for the contribution of the first term in , and the second term comes from the second term in . For any given $\eta$, the optimum choice of $Q_1$ is obtained by equating the two terms and using the relation $Q_1Q_2=N(MT)^{\eta-\frac{1}{2}}$. It follows that $$Q_1=\frac{(MT)^{\frac{7}{6}}}{N(MT)^{\eta}}.$$ This satisfies our requirement that $(MT)^{\frac{1}{4}-\frac{\eta}{2}}Q^{-\varepsilon}\gg Q_1\gg (MT)^{2\eta}Q^{\varepsilon}$ if $\eta<\frac{1}{18}$. To get the optimal value of $\eta$ we compare this bound with the bound for the error term in Lemma \[circ\], i.e. $$\frac{\sqrt{MT}N}{(MT)^{\frac{7}{12}-\frac{\eta}{2}}}=\sqrt{N}(MT)^{\frac{1}{2}-\eta}.$$ It follows that $\eta=\frac{1}{18}-\varepsilon$ is the optimal choice. This completes the proof of the theorem.
[99]{}
V. Blomer; G. Harcos: Hybrid bounds for twisted $L$-functions. J. reine angew. Math. **621** (2008), 53–79.
V. Blomer; G. Harcos: Twisted $L$-functions over number fields and Hilbert’s eleventh problem. GAFA **20** (2010), 1–52.
V. Blomer; G. Harcos; P. Michel: A Burgess-like subconvex bound for twisted L-functions. Forum Math. **19** (2007), 61–106.
V. A. Bykovskii: A trace formula for the scalar product of Hecke series and its applications. Translated in J. Math. Sci (New York) **89** (1998), 915–932.
J. Cogdell; I. Piatetski-Shapiro; P. Sarnak: Estimates for Hilbert modular $L$-functions and applications. (Unpublished)
A. Diaconu; P. Garrett: Subconvexity bounds for automorphic $L$-functions. Journal of the Inst. of Math. Jussieu **9** (2010), 95–124.
W. Duke; J.B. Friedlander; H. Iwaniec: Bounds for automorphic L-functions. Invent. Math. **112** (1993), 1–8.
A. Good: The square mean of Dirichlet series associated with cusp forms. Mathematika **29** (1982), 278–295.
G. Harcos: An additive problem in the Fourier coefficients of cusp forms. Math. Ann. **326** (2003), 347–365.
D. R. Heath-Brown: Hybrid bounds for Dirichlet $L$-functions II. Quart. J. Math. Oxford Ser. (2) **31** (1980), 157–167.
M. Jutila: Transformations of exponential sums. Proceedings of the Amalfi Conference on Analytic Number Theory (Maiori 1989), Univ. Salerno, Salerno, (1992) 263–270.
M. Jutila: The additive divisor problem and its analogs for Fourier coefficients of cusp forms. I., Math. Z. **223** (1996), 435–461; II., ibid **225** (1997), 625–637.
T. Meurman: On exponential sums involving the Fourier coefficients of Maass wave forms. J. Reine Angew. Math. **384** (1988), 192–207.
T. Meurman: On the order of the Maass $L$-function on the critical line. in Number Theory, Vol. I (Budapest, 1987), Colloq. Math. Soc. János Bolyai, vol. 51, pp. 325–354, North-Holland, Amsterdam, 1990.
P. Michel: The subconvexity problem for Rankin–Selberg L-functions and equidistribution of Heegner points. Ann. of Math. **160** (2004), 185–236.
P. Michel; A. Venkatesh: The subconvexity problem for $\rm{GL}_2$. Publ. Math. Inst. Hautes ' Etudes Sci. **111** (2010), 171–271.
R. Munshi: On a hybrid bound for twisted $L$-values. Archiv der Math. **96** (2011), 235–245.
R. Munshi: Shifted convolution sums for $GL(3)\times GL(2)$. arXiv:1202.1157 (2012).
Y. Petridis; P. Sarnak: Quantum unique ergodicity for $SL_2(\mathcal O)\backslash \mathbb H^3$ and estimates for $L$-functions. J. Evol. Eqns. **1** (2001), 277–290.
P. Sarnak: Estimates for Rankin-Selberg L-functions and Quantum Unique Ergodicity. J. Funct. Analysis **184** (2001), 419–453.
A. Venkatesh: Sparse equidistribution problems, period bounds and subconvexity. Annals of Math. **172** (2010), 989–1094.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the geometric structure of the statistical models for two-by-two contingency tables. One or two odds ratios are fixed and the corresponding models are shown to be a portion of a ruled quadratic surface or a segment. Some pointers to the general case of two-way contingency tables are also given and an application to case-control studies is presented.'
author:
- 'E. Carlini and F. Rapallo'
bibliography:
- 'carlinirapallo.bib'
title: 'The Geometry of Statistical Models for Two-Way Contingency Tables with Fixed Odds Ratios'
---
Introduction {#intro}
============
A two-way contingency table gives the joint distribution of two random variables with a finite number of outcomes. If we denote by $\{0, \ldots, I-1\}$ and $\{0, \ldots, J-1\}$ the outcomes of $X_1$ and $X_2$ respectively, the contingency table is represented by a matrix $P=(p_{ij})$, where $p_{ij}$ is the probability that $X_1=i$ and $X_2=j$. The table $P$ is also called an $I \times J$ contingency table, in order to emphasize that the variable $X_1$ has $I$ outcomes and the variable $X_2$ has $J$ outcomes.
In the analysis of contingency tables odds ratios, or cross-product ratios, are major parameters, and their use in the study of $2 \times 2$ tables goes back to the 1970’s. For an explicit discussion on this approach see, e.g., [@fienberg:80].
For a $2 \times 2$ table of the form: $$\label{tabella2per2}
\begin{pmatrix}
p_{00} & p_{01} \\
p_{10} & p_{11} \\
\end{pmatrix}$$ there is only one cross-product ratio, namely: $$r = \frac {p_{00}p_{11}} {p_{01}p_{10}} \, .$$ In the general $I \times J$ case, there is one cross-product ratio for each $2 \times 2$ submatrix of the table. Thus, they have the form $$\frac {p_{ij}p_{kh}} {p_{ih}p_{kj}}$$ for $0 \leq i < k \leq I-1$ and $0 \leq j < h \leq J-1$, see [@agresti:02 Chapter 2]. In this paper we will consider the cross-product ratio and other ratios naturally defined.
Odds ratios are used in a wide range of applications, and in particular in case-control studies in pharmaceutical and medical research. Following the theory of log-linear models, the statistical inference for the odds ratios is made under asymptotic normality, see for example . More recently, some methods for exact inference have been introduced, see [@agresti:02] and [@agresti:01] for details and further references. For the theory about the Bayesian approach, see [@lindley:64].
From the point of view of Probability and Mathematical Statistics, different descriptions of the geometry of the statistical models for contingency tables are presented in [@collombier:80 Chapter 2], and in , using vector space theory. An earlier approach to the geometry of contingency tables with fixed cross-product ratio can be found in . In the last few years, the introduction of techniques from Commutative Algebra gave a new flavor to the geometrical representation of statistical models, as shown in, e.g., , , and [@slavkovic:04].
In this paper we use Algebraic and Geometric techniques in order to describe the structure of some models for two-way contingency tables described through odds ratios.
We first consider the case of $2 \times 2$ contingency tables of the form (\[tabella2per2\]) with the constraints $p_{ij} > 0$ for all $i,j=0,1$ and $p_{00}+p_{01}+p_{10}+p_{11}=1$. If we allow some probabilities to be zero, notice that the ratios are either zero or undefined. Thus we restrict the analysis to the strictly positive case.
In a $2 \times 2$ table we consider the three odds ratios: $$r_\times = \frac {p_{00}p_{11}} {p_{01}p_{10}} \, ,$$ $$r_{||} = \frac {p_{00}p_{10}} {p_{01}p_{11}} \, ,$$ $$r_= = \frac {p_{00}p_{01}} {p_{10}p_{11}} \, .$$ The meaning of the three odds ratios above will be fully explained in Section \[application\].
Let $r_\times = \alpha^2$, $r_{||} = \beta^2$ and $r_==\gamma^2$. For further use, it is useful to make explicit the following identities. Considering $r_=$ and $r_{||}$, it is easy to check that: $$\label{rel1}
\beta \gamma = \frac{p_{00}} {p_{11}} \, ,$$ and $$\label{rel2}
\frac \beta \gamma = \frac{p_{01}} {p_{10}} \, .$$
In Section \[oddsSEC\], we study the geometric properties of some statistical models for $2 \times 2$ contingency tables. We consider models obtained by fixing two odds ratios, showing that the model is represented by a segment in the probability simplex and studying the behavior of the third ratio. In particular, an expression for tables with three fixed ratios is derived. We also recover classical results about models with a fixed odds ratio. In Section \[2x3SEC\], we give a glimpse of the general situation of $I \times J$ contingency tables. We focus our attention on $2\times 3$ tables and we present some of the difficulties arising in the general case. An application to case-control studies is presented in Section \[application\].
Odds Ratios {#oddsSEC}
===========
In this section, we use basic geometric techniques to study the $2\times 2$ tables having two out of the three ratios $r_\times,
r_=$ and $r_{||}$ fixed.
We consider a $2\times 2$ matrix as a point in the real affine 4-space ${\mathbb{A}}^4$. In particular, with the notation of Equation (\[tabella2per2\]), the $p_{ij}$’s are coordinates in ${\mathbb{A}}^4$. A $2\times 2$ [*table*]{} is a matrix in the open [*probability simplex*]{} $$\Delta=\left\lbrace P=(p_{ij})\in{\mathbb{A}}^4 : \sum p_{ij}=1, p_{ij}> 0 ,i,j=0,1 \right\rbrace \, .$$
As our goal is to describe odds ratios for tables, we may assume the ratios to be non-zero positive numbers.
Fixed the first two ratios $$r_\times=\alpha^2\mbox{\ \ and \ \ } r_{||}=\beta^2\, ,$$ we want to answer the following question:
> Q1: How can we describe the locus of tables having the assigned two ratios?
and also
> Q2: What are the possible values of the third ratio?
These questions were posed in the AIM computational algebraic statistics plenary lecture by Stephen Fienberg in 2003. In this situation, some interesting comments about treating questions Q1 and Q2 were also made.
Consider the quadratic hypersurfaces of ${\mathbb{A}}^4$:
$$Q_\alpha: \alpha^2p_{01}p_{10}=p_{00}p_{11}$$ and $$Q_\beta: \beta^2p_{01}p_{11}=p_{00}p_{10}\, .$$
Notice that a matrix in $Q_\alpha\cap Q_\beta$ is such that $r_\times=\alpha^2$ and $r_{||}=\beta^2$ as soon as the ratios are defined. Hence, to answer the first question, it is enough to study $$Q_\alpha\cap Q_\beta\setminus Z\, ,$$ where $Z=\left\lbrace P=(p_{ij})\in{\mathbb{A}}^4 :
p_{00}p_{01}p_{10}p_{11}=0 \right\rbrace$.
We readily see that $Q_\alpha\cap Q_\beta$ contains the 2-dimensional skew linear spaces $$p_{00}=p_{01}=0 \mbox{ \ \ and \ \ } p_{10}=p_{11}=0$$ and by general facts on quadrics (see [@Harris page 301]) we know that there exist two more 2-dimensional skew linear spaces, $R$ and $S$, such that $$Q_\alpha\cap Q_\beta=\lbrace p_{00}=p_{01}=0\rbrace\cup\lbrace p_{10}=p_{11}=0\rbrace\cup R\cup S \, .$$
Manipulating equations we notice that a point in $Q_\alpha\cap
Q_\beta\setminus Z$ is such that $${p_{00}\over p_{01}}=\alpha^2{p_{10}\over p_{11}}=\beta^2{p_{11}\over p_{10}}$$ and $${p_{10}\over p_{11}}=\beta^2{p_{01}\over p_{00}}={1\over\alpha^2}{p_{00}\over p_{01}} \, .$$ Hence, $R$ and $S$ lie in the intersection of the two 3-dimensional spaces $$(\alpha p_{10}-\beta p_{11})(\alpha p_{10}+\beta p_{11})=0$$ and $$(\beta p_{01}-{1\over\alpha}p_{00})(\beta p_{01}+{1\over\alpha}p_{00})=0 \, ,$$ where $\alpha$ and $\beta$ are chosen to be positive. Only two out of the four resulting 2-dimensional linear spaces lie in both $Q_\alpha$ and $Q_\beta$ and these are $R$ and $S$: $$R: \alpha p_{10}-\beta p_{11}=\beta p_{01}-{1\over\alpha}p_{00}=0 \, ,$$ $$S: \alpha p_{10}+\beta p_{11}=\beta p_{01}+{1\over\alpha}p_{00}=0 \, ,$$ which have parametric presentations $$R=\{(\beta u,{1\over\alpha}u,\beta v,\alpha v): u,v\in{\mathbb{R}}\} \, ,$$ $$S=\{(\beta u,-{1\over\alpha}u,\beta v,-\alpha v): u,v\in{\mathbb{R}}\} \, .$$ Summing all these facts up, we get
Fix the ratios $r_\times=\alpha^2$ and $r_{||}=\beta^2$. Then, a matrix has the given ratios if and only if it has the form $$\left(\begin{array}{cc}
\beta u & {1\over\alpha}u \\
\beta v & \alpha v
\end{array}\right)
\mbox{ or } \left(\begin{array}{cc}
\beta u & -{1\over\alpha}u \\
\beta v & -\alpha v
\end{array}\right)$$ with $u,v$ non-zero real parameters.
Finally, we have to intersect $R$ and $S$ with the probability simplex. As we can choose $\alpha$ and $\beta$ to be positive, we immediately see that $S\cap\Delta=\emptyset$ (there is always a non-positive coordinate).
To determine $R\cap\Delta$, notice that $R\cap\{\sum p_{ij}=1\}$ is obtained by taking $$u={1-(\beta+\alpha)v\over{\beta+{1\over\alpha}}}$$ in the parametric presentation of $R$. Hence, we get
\[x=PROP\] Fix the ratios $r_\times=\alpha^2$ and $r_{||}=\beta^2$. Then, a [*table*]{} has the given ratios if and only if it has the form $$\left(\begin{array}{cc}
{\beta\over{\beta+{1\over\alpha}}}[1-(\beta+\alpha)v] & {1\over\alpha\beta+1}[1-(\beta+\alpha)v] \\
\beta v & \alpha v
\end{array}\right)$$ where $0< v<{1\over\alpha+\beta}$.
This answers question Q1: fixed the two ratios, the tables with those ratios describe a segment in the probability simplex.
In , a parametric description of the tables with $r_\times=1$ is written in the form $$\label{par1}
\left(
\begin{array}{cc}
st & s(1-t) \\
(1-s)t & (1-s)(1-t)
\end{array}
\right) \, .$$ Let us check that our parametrization contains this as a special case. In order to do this, we will compute the marginal sums $$\label{par2}
\left(\begin{array}{ccc|c}
\beta[\frac{1}{\beta+1}-v] & \ \ & [\frac{1}{\beta+1}-v] & 1-(\beta+1)v \\
& & & \\
\beta v & & v & (\beta+1)v \\ \hline
\frac{\beta}{\beta+1} & & \frac{1}{\beta+1} & 1 \end{array}\right)
\, .$$ Hence, the parametrizations in Equations (\[par1\]) and (\[par2\]) are just the same, simply let $t=\frac{\beta}{\beta+1}$ and $s=1-(\beta+1)v$.
Suppose to fix $r_\times$ and to ask for a geometric description of the locus of tables with this ratio. Using Proposition \[x=PROP\] we can easily get an answer. For each value of $r_{||}$ we get a segment of tables, and making $r_{||}$ to vary this segment describes a portion of a ruled quadratic surface. Notice that, for $r_\times=1$, this is the result contained in . In particular, we recall that matrices such that $r_\times$ is fixed form a so called Segre variety (i.e., in this case, a smooth quadric surface in the projective three space). For more on this see, e.g., .
Answering question Q2 is just a computation, and we see that $$r_=={1\over \alpha\beta+1}{[1-(\beta+\alpha)v]^2\over v}\, ,$$ where $r_\times=\alpha^2$ and $r_{||}=\beta^2$. Notice that, fixed $r_\times$ and $r_{||}$, the third ratio can freely vary in $(0,+\infty)$.
\[INVrem\] We expressed $r_=$ as a function $r_=(\alpha,\beta,v)$, and standard computations show that this is an invertible function of $v$. In particular, we get $$v=\frac{1}{\alpha+\beta+\sqrt{(\alpha\beta+1)r_=}} \, .$$ Thus, given $r_\times=\alpha^2,r_{||}=\beta^2$ and $r_=$, we have an explicit description of the [*unique*]{} table with these ratios (use Proposition \[x=PROP\]).
Clearly, completely analogous results hold if we fix the ratios $r_\times$ and $r_=$.
If we fix the ratios $r_{||}=\beta^2, r_==\gamma^2$ and we argue as above, we get the following:
\[==PROP\] Fix the ratios $r_{||}=\beta^2$ and $r_==\gamma^2$. Then, a [*table*]{} has the given ratios if and only if it has the form $$\left(\begin{array}{cc}
{\beta\over{\beta+{1\over\gamma}}}[1-(\beta+\gamma)v] & \gamma v \\
\beta v & {1\over\beta\gamma+1}[1-(\beta+\gamma)v]
\end{array}\right)$$ where $0< v<{1\over\beta+\gamma}$.
Again, a trivial computation yields: $$r_\times=\left({\beta\over \beta\gamma+1}\right)^2{[1-(\beta+\gamma)v]^2\over v^2} \, ,$$ and hence, fixed $r_=$ and $r_{||}$, the third ratio can freely vary in $(0,+\infty)$, see Remark \[INVrem\].
In recent literature, there is an increasing attention to the geometrical structure of probability models for contingency tables. In particular, in [@slavkovic:04 Chapter 3] the author presents some results about the geometrical characterization of probability models for $2 \times 2$ contingency tables in terms of the cross-product ratio and the conditional distributions. In the same work the connections between the odds ratios and the classical log-linear and ANOVA-type representations of the probability models are clearly stated. We remark that our notation slightly differs from the one used by A. Slavkovic in her Ph.D. dissertation.
In the same direction, in the graphical visualization of joint, marginal and conditional distributions on the probability simplex for $2 \times 2$ contingency tables is presented.
The $2\times 3$ case {#2x3SEC}
====================
The study of tables with more than two rows and columns would be of great interest, but the complexity of the problem readily increases as we show in the $2 \times 3$ case.
Consider the $2\times 3$ matrix $$\left(\begin{array}{ccc}p_{00} & p_{01} & p_{02}\\
p_{10} & p_{11} & p_{12}\end{array}\right)$$ and define odds ratios as above for each $2\times 2$ submatrix. We complete our previous notation by adding a superscript to denote the deleted column, e.g. $$r_=^{(1)}={p_{00}p_{02}\over
p_{10}p_{12}} \, .$$ Again, we consider a matrix as a point in a real affine space, in this case ${\mathbb{A}}^6$. Notice that the ratios are well defined for matrices in ${\mathbb{A}}^6\setminus Z$, where $Z$ denotes the set of matrices with at least a zero entry.
Relations among the ratios are the cause of the increased complexity of the higher dimensional cases. For example, as we will see, two of the ratios can always be freely fixed. But, as soon as three ratios are considered, constraints come in the picture.
Easy calculations show that the following relations hold: $$r_{||}^{(0)}r_{||}^{(2)}=r_{||}^{(1)} \, ,$$ $$r_{\times}^{(0)}r_{\times}^{(2)}=r_{\times}^{(1)}$$ and also $$r_\times^{(0)}=r_=^{(2)}(r_=^{(1)})^{-1}\, ,$$ $$r_\times^{(1)}=r_=^{(2)}(r_=^{(0)})^{-1}\, ,$$ $$r_\times^{(2)}=r_=^{(1)}(r_=^{(0)})^{-1}\, .$$
These relations, beside producing constraints on the numerical choice of the ratios, lead to a much more complex geometric situation. We illustrate this by exhibiting some explicit examples (worked out with the Computer Algebra systems [**Singular**]{} and [**CoCoA**]{}). As references for the software, see [@cocoa] and [@singular].
More precisely, we fix some of the ratios and we describe the locus of matrices satisfying these relations in $$\Sigma^\circ=\{P=(p_{ij})\in{\mathbb{A}}^6 : \sum p_{ij}=1\}\setminus Z \, ,$$ i.e. the space of matrices with non-null entries of sum one. For the sake of simplicity, we do not consider the positivity conditions defining the simplex.
In our geometric descriptions, we will slightly abuse terminology, e.g. we will call a line in $\Sigma^\circ$ a line in ${\mathbb{A}}^6$ not contained in $Z$; notice that our lines may have some holes (i.e. the points of intersection with $Z$).
We start by considering the easiest case where two of the ratios are fixed. Already at this stage, a dichotomy arises and we have two different situations, as shown in the following examples: $$\label{2eq1}
{r_{\times}^{(1)}}={r_{\times}^{(2)}}=1 \, ,$$ $$\label{2eq2}
{r_{\times}^{(1)}}={r_{=}^{(2)}}=1\mbox{ or
}{r_{=}^{(1)}}={r_{||}^{(2)}}=1\mbox{ or }
{r_{||}^{(1)}}={r_{||}^{(2)}}=1\mbox{ or }
{r_{=}^{(1)}}={r_{=}^{(2)}}=1 \, .$$ The locus of matrices in $\Sigma^\circ$ satisfying one of conditions (\[2eq2\]) is a 3-dimensional variety of degree 4, while condition (\[2eq1\]) describes a 3-dimensional variety of degree 3. Roughly speaking, the degree (see [@Harris page 16] and [@Shaf page 41]) is a measure of the complexity of the variety. For a surface in 3-space, for example, the degree bounds the number of intersections with a line and, in a certain sense, measures how the surface is folded.
Next, we try to fix three of the ratios, for example: $$\label{3eq1}
{r_{\times}^{(0)}}={r_{=}^{(1)}}={r_{||}^{(2)}}=1\mbox{ or }
{r_{\times}^{(0)}}={r_{\times}^{(1)}}={r_{||}^{(2)}}=1 \, ,$$ $$\label{3eq2}
{r_{\times}^{(0)}}={r_{\times}^{(1)}}={r_{=}^{(2)}}=1 \, ,$$ $$\label{3eq3}
{r_{\times}^{(0)}}=4,{r_{\times}^{(1)}}=3,{r_{=}^{(2)}}=2 \, .$$ The locus of matrices in $\Sigma^\circ$ satisfying one of conditions (\[3eq1\]) is the union of two quadratic surfaces, while condition (\[3eq2\]) gives a plane. Moreover, if we consider the same ratios but we vary their values, as in (\[3eq3\]), the locus of matrices is now described by a single quadratic surface.
Finally, a glimpse of the case of four fixed ratios: $$\label{4eq1}
{r_{\times}^{(0)}}={r_{\times}^{(1)}}={r_{||}^{(1)}}={r_{||}^{(2)}}=1
\, ,$$ $$\label{4eq2}
{r_{\times}^{(0)}}={r_{\times}^{(1)}}={r_{=}^{(1)}}={r_{=}^{(2)}}=1 \,
,$$ In both cases, the locus is described by a curve as expected. But, condition (\[4eq1\]) produces the union of four lines, while condition (\[4eq2\]) is satisfied by a single line in $\Sigma^\circ$.
The Computer Algebra systems [**Singular**]{} and [**CoCoA**]{} were used to compute primary decompositions (giving the irreducible components of the loci) and Hilbert functions (giving the dimension and the degree of the loci).
An application. The case-control studies {#application}
========================================
Two-by-two contingency tables are natural models for a large class of problems known, in medical literature, as case-control studies. Let us consider a table coming, e.g., from the study of a new pharmaceutical product, or clinical test, designed for the detection of a disease. This is an example of a case-control study.
In a case-control study there are two random variables. The first variable $X_1$ encodes the presence (level 1) or absence (level 0) of the disease. The second variable $X_2$ encodes the result of the clinical test (level 1 if positive, level 0 if negative).
The joint variable $(X_1,X_2)$ has $4$ outcomes, namely: $$(0,0),(0,1),(1,0),(1,1)\, .$$ Its probabilities form a $2 \times 2$ contingency table: $$\begin{pmatrix}
p_{00} & p_{01} \\
p_{10} & p_{11} \\
\end{pmatrix} \, .$$
The probabilities $p_{00}$ and $p_{11}$ are called the probability of true negative and of true positive, respectively. They correspond to the cases of correct answer of the clinical test. The probabilities $p_{10}$ and $p_{01}$ are called the probability of false positive and of false negative, respectively. They correspond to the two types of error which can show in a case-control study. For example, the probability of false negative is the probability that a diseased subject is incorrectly classified as not diseased.
A perfect clinical test which correctly classifies all the subjects would have $p_{01}$ and $p_{10}$ as low as possible, implying a large value of the odds ratio $r_{\times}$. Therefore, the odds ratio $r_\times$ measures the validity of the clinical test. In particular, when $r_{\times}=1$, the random variables are statistically independent. In our framework this means that, when $r_\times=1$, the result of the clinical test is independent from the presence or absence of the disease. Unless one obtains a large value of $r_{\times}$, the clinical test is judged as non efficient. The odds ratio $r_{\times}$ is also called Diagnostic Odds Ratio (DOR) in medical literature.
In such a case-control study, two essential indices are the specificity and the sensitivity, defined as: $${\rm specificity} = \frac {p_{00}} {p_{00}+p_{01}}$$ and $${\rm sensitivity} = \frac {p_{11}} {p_{10}+p_{11}} \, .$$ Specificity is the proportion of true negative among the diseased subjects, while sensitivity is the proportion of true positive among the non-diseased subjects.
Straightforward computations show that $$r_{\times} = \frac {{\rm specificity} / (1 - {\rm specificity})}
{(1-{\rm sensitivity}) / {\rm sensitivity} } \, .$$
In view of the definition above, it is easy to show that the relative magnitude of the sensitivity and specificity is measured by the odds ratio $r_{||}$. In fact one can show that $$\frac {{\rm sensitivity} / (1- {\rm sensitivity})} {{\rm
specificity}/(1 - {\rm specificity})} = \frac 1
{r_{||}}\, .$$ The ratio above is called Error Odds Ratio (EOR).
In recent literature, the DOR and the EOR are relevant parameters for the assessment of the validity of a clinical test. They have received increasing attention in the last few years and a huge amount of literature has been produced. Hence, we refrain from any tentative description and refer the interested reader to, for example, [@knottnerus:01].
The meaning of the third ratio $r_{=}$ is not straightforward as explained in . However its statistical meaning can be derived using Equations (\[rel1\]) and (\[rel2\]) shown in Section \[intro\].
Finally, we remark that the geometrical structure of the statistical models for case-control studies is very simple. From the results in Section \[oddsSEC\], one readily sees that the models are segments or portions of ruled quadratic surfaces. Moreover, from a Bayesian point of view, Propositions \[x=PROP\] and \[==PROP\] allow to compute the exact range of the free odds ratio.
[**Acknowledgement.**]{} We wish to thank an anonymous referee for his/her valuable suggestions and comments for the improvement of the paper.
[**AMS Subject Classification: 14Q99, 62H17.**]{}\
Enrico CARLINI\
Department of Mathematics\
Politecnico di Torino\
corso Duca degli Abruzzi, 24\
10129 Torino, ITALY\
email: `[email protected]`\
Fabio RAPALLO\
Department of Mathematics\
University of Genova\
via Dodecaneso, 35\
16146 Genova, ITALY\
email: `[email protected]`\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we prove some results concerning stability of hypersurfaces in the four dimensional Euclidean space with zero scalar curvature. First we prove there is no complete stable hypersurface with zero scalar curvature, polynomial growth of integral of the mean curvature, and with the Gauss-Kronecker curvature bounded away from zero. We conclude this paper giving a sufficient condition for a regular domain to be stable in terms of the mean and the Gauss-Kronecker curvatures of the hypersurface and the radius of the smallest extrinsic ball which contains the domain.'
address:
- 'Instituto de Matem'' atica, Universidade Federal de Alagoas, Maceió, AL, 57072-900, Brasil'
- 'Instituto Nacional de Matem'' atica Pura e Aplicada, Rio de Janeiro, RJ, 22460-320, Brasil'
- 'Instituto de Matem'' atica, Universidade Federal de Alagoas, Maceió, AL, 57072-900, Brasil'
author:
- Hilário Alencar
- Manfredo do Carmo
- Gregório Silva Neto
date: 'January 23, 2015'
title: Stable hypersurfaces with zero scalar curvature in Euclidean space
---
[^1]
[^2]
Introduction {#intro}
============
Let $M^3$ be a hypersurface of $\mathbb{R}^4$ with scalar curvature $R=0$ and whose mean curvature $H$ is nowhere zero. Let $\Omega\subset M$ be a regular domain, i.e., a domain with compact closure and piecewise smooth boundary. We recall that hypersurfaces of ${\mathbb{R}}^4$ with zero scalar curvature are critical points of the functional $$\mathcal{A}_1(\Omega) = \int_{\Omega} H dM$$ under all compactly supported variations in $\Omega,$ see [@AdCE]. Thus, the notion of stability makes sense and we can ask for a condition to ensure that a regular domain $\Omega\subset M$ be stable.
Since $H$ is nowhere zero, depending on choice of orientation we have $H>0$ or $H<0$ everywhere. Let $\Omega\subset M$ be a regular domain. If we choose an orientation such that $H>0$ everywhere, then the domain $\Omega$ will be stable if $\left.\dfrac{d^2\mathcal{A}_1}{dt^2}\right|_{t=0}>0$ under all compactly supported variations in $\Omega.$ Otherwise, i.e., if we choose an orientation such that $H<0,$ then the domain $\Omega$ is stable if $\left.\dfrac{d^2\mathcal{A}_1}{dt^2}\right|_{t=0}<0$ under all such variations. We say that $M$ is stable if all regular domains of $M$ are stable. For more information about the concept of stability, we refer to [@Smale; @reilly; @HL2; @AdCE].
We say that $M^3$ has polynomial growth of the $1-$volume if there exist constants $C>0$ and $\alpha>0$ such that $$\int_{\mathcal{B}_r(q)}HdM\leq C r^\alpha$$ for all $r>0$ and $q\in M,$ where $\mathcal{B}_r(q)$ denotes the geodesic ball of $M$ with center $q$ and radius $r.$
If $M^3$ is a hypersurface of ${\mathbb{R}}^4$ with zero scalar curvature and Gauss-Kronecker curvature nowhere zero then, by the Gauss equation $(3H)^2 = \|A\|^2 + 6R$ (where $\|A\|$ is the matrix norm of the second fundamental form and $R$ is the scalar curvature of $M^3$), the mean curvature is nowhere zero. Thus, the notion of stability makes sense in this case.
Alencar, do Carmo and Elbert, see [@AdCE p.215], posed a conjecture, which for the case $n=3$ can be written as follows:
*There is no complete, stable hypersurface $M^3$ of ${\mathbb{R}}^4$ with zero scalar curvature and everywhere non-zero Gauss Kronecker curvature.* Our first result is a partial answer to this conjecture.
\[T-1\] There is no stable, complete, non-compact, hypersurface $M^3$ of ${\mathbb{R}}^4$ with zero scalar curvature, Gauss-Kronecker curvature bounded away from zero, and polynomial growth of the $1-$volume.
Since the image of the Gauss map for graphs lies in a open hemisphere, they are stable, see [@AdCE p.201, Theorem 1.1]. Moreover, by [@A-S-Z p.3310, Proposition 4.2], graphs have polynomial growth of the $1-$volume. Thus an immediate corollary of Theorem \[T-1\] is the following Bernstein type result.
There is no complete graph in ${\mathbb{R}}^4$ with zero scalar curvature and Gauss-Kronecker curvature bounded away from zero.
[We point out that some condition on the Gauss-Kronecker curvature is needed. In fact, cylinders over positively curved curves are examples of stable hypersurfaces with zero scalar curvature and everywhere zero Gauss-Kronecker curvature. If we choose, for example, the curve as the graph of the polynomial function $y(x)=x^2,$ see [@SN p.492, Example 4.2], we obtain a cylinder which is a graph with polynomial growth of the $1-$volume. ]{}
In the direction of Theorem \[T-1\], by using a technique which holds only in dimension 3, the third author, see [@SN p.483, Theorem A], proved the following result:
*There is no stable complete hypersurface $M^3$ of ${\mathbb{R}}^4$ with zero scalar curvature, polynomial volume growth and such that $$-\frac{K}{H^3}\geq c>0$$ everywhere, for some constant $c>0$. Here $H$ denotes the mean curvature and $K$ denotes the Gauss-Kronecker curvature of the immersion.*
Let $r_\Omega$ be the radius of the smallest extrinsic ball which contains the domain $\Omega\subset M.$ Our second result gives a sufficient condition for a regular domain to be stable in terms of the mean and the Gauss-Kronecker curvatures of the hypersurface and the radius of the smallest extrinsic ball which contains the domain. We have the following result.
\[stability\] Let $M^3$ be a hypersurface of ${\mathbb{R}}^4$ with zero scalar curvature and such that $H\neq0.$ Let $\Omega\subset M$ be a regular domain. If $$\sup_\Omega\left(\frac{-3K}{H}\right)\leq \dfrac{3}{2r_\Omega^2},$$ then $\Omega$ is stable.
[By using essentially the same proofs, the Theorems \[T-1\] and \[stability\] extend to the case of a hypersurface $M^n$ in ${\mathbb{R}}^{n+1},$ $n$ arbitrary, with zero scalar curvature and non-vanishing of the third symmetric function of the principal curvatures, rather than non-vanishing of Gauss-Kronecker curvature. ]{}
[*Acknowledgements.*]{} The authors would like to thank to the referee for the comments.
Proofs of the Theorems {#sec:1}
======================
In what follows we introduce a second order differential operator which will play a role for hypersurfaces with zero scalar curvature similar to that of Laplacian for minimal hypersurfaces. For that, consider the linear operator $P_1:TM{\rightarrow}TM$ given by $$P_1 = 3HI-A,$$ where $A:TM{\rightarrow}TM$ is the linear operator associated with the second fundamental form of the immersion of $M^3$ into ${\mathbb{R}}^4$ and $I:TM{\rightarrow}TM$ is the identity operator. We define $$\label{def.L1}
L_1(f) = \operatorname{div}(P_1({\nabla}f)),$$ where $\operatorname{div}X$ denotes the divergence of the vector field $X$ and ${\nabla}f$ denotes the gradient of the function $f$ in the induced metric. By a result of Hounie and Leite, see [@HL p.873, Proposition 1.5], when the scalar curvature $R=0,$ the differential operator $L_1$ is elliptic if and only if $K\neq0$ everywhere. Since $L_1$ is an elliptic and self-adjoint operator, it has a discrete spectrum and thus we can consider the eigenvalues of $L_1$ for regular domains $\Omega\subset M.$ The first eigenvalue $\lambda_1^{L_1}(\Omega)$ has an associated positive eigenfunction $g$, i.e., a function such that $L_1g + \lambda_1^{L_1}(\Omega)g=0$ in $\Omega.$ Set $$\|g\|_{H_0^1}=\left(\int_\Omega (|g|^2 + |{\nabla}g|^2) dM\right)^{1/2}$$ and let $H_0^1(\Omega)$ be the completion of $\mathcal{C}^\infty_0(\Omega)$ with respect to the norm $\|\cdot\|_{H_0^1}.$ It is well known that $H_0^1(\Omega)$ is called a Sobolev space over $\Omega.$ It can be proven, see [@Smale p.1052, Lemma 4(a)], that $$\begin{split}
\lambda_1^{L_1}(\Omega) &=\inf\left\{\frac{\int_\Omega -gL_1g dM}{\int_\Omega g^2 dM}: g\in H_0^1(\Omega) ,\ g\not\equiv0\right\}\\
& = \inf\left\{\frac{\int_\Omega {\langle}P_1({\nabla}g),{\nabla}g{\rangle}dM}{\int_\Omega g^2 dM}:g\in H_0^1(\Omega),\ g\not\equiv0 \right\}.\\
\end{split}$$
Let $M^3$ be a complete and non-compact hypersurface of ${\mathbb{R}}^4$ with zero scalar curvature and Gauss-Kronecker curvature $K\neq0$ everywhere. It is known that $$\label{desig}
HK\leq \frac{3}{2}R^2.$$ In fact, if $k_1, \ k_2,$ and $k_3$ denotes the principal curvatures of the hypersurface, then $$\begin{split}
(3R)^2 - 6HK& = (k_1k_2 + k_1k_3 + k_2k_3)^2 - 2(k_1 + k_2 + k_2)(k_1k_2k_3)\\
& = [k_1^2k_2^2 + k_1^2k_3^2 + k_2^2k_3^2 + 2(k_1^2k_2k_3 + k_1k_2^2k_3 + k_1k_2k_3^2)]\\
&\qquad - 2(k_1^2k_2k_3 + k_1k_2^2k_3 + k_1k_2k_3^2)\\
& = k_1^2k_2^2 + k_1^2k_3^2 + k_2^2k_3^2 \geq0,
\end{split}$$ and the equality holds if, and only if, two of the principal curvatures are zero. If, without loss of generality, we choose an orientation of $M^3$ such that $H>0,$ then the hypothesis $R=0,$ $K\neq0,$ and the (\[desig\]) imply $K<0.$
The proof of the Theorem will be made by showing the existence of unstable domains in $M.$ Let $\{\Omega_i\}_{i=1}^\infty$ be a family of regular domains in $M$ such that $\Omega_i\subset \Omega_{i+1}$ and $\displaystyle{\bigcup_{i=1}^\infty \Omega_i=M.}$ The second variation of the functional $\displaystyle{\mathcal{A}_1(\Omega_i)=\int_{\Omega_i} H dM}$ is $$\begin{split}
\left.\frac{d^2\mathcal{A}_1}{dt^2}(g_i)\right|_{t=0} &=-\int_{\Omega_i}(g_iL_1g_i - 3Kg_i^2)dM\\
\end{split}$$ where $g_i:M{\rightarrow}{\mathbb{R}}$ is any piecewise smooth function defined on $\Omega_i$ with $g_i|_{\partial \Omega_i}=0,$ see [@AdCE p.207]. Let $g_i$ be the first eigenfunction of $L_1$ over $\Omega_i\subset M.$ Thus we have $$\left.\frac{d^2\mathcal{A}_1}{dt^2}(g_i)\right|_{t=0} = \lambda_1^{L_1}(\Omega_i)\int_{\Omega_i} g_i^2 dM + \int_{\Omega_i} 3K g_i^2 dM,$$ i.e., $$\label{eq.1}
\displaystyle{\frac{\left.\dfrac{d^2\mathcal{A}_1}{dt^2}(g_i)\right|_{t=0}}{\int_{\Omega_i} g_i^2 dM} = \lambda_1^{L_1}(\Omega_i) - \frac{\int_{\Omega_i}(-3K) g_i^2 dM}{\int_{\Omega_i} g_i^2 dM}}.$$ Since $M^n$ has polynomial growth of the 1-volume, see [@Elbert p.259, Lemma 3.12], gives $\lambda_1^{L_1}(M)=\inf\{\lambda_1^{L_1}(\Omega)| \Omega\subset M\}=0.$ By using that $\Omega_i\subset\Omega_{i+1},$ see [@Smale p.1051, Lemma 2], we have $\lambda^{L_1}_1(\Omega_i)\geq \lambda_1^{L_1}(\Omega_{i+1}).$ This implies $$\label{eq.2}
\lim_{i{\rightarrow}\infty} \lambda_1^{L_1}(\Omega_i)=0.$$ The second member of the expression $$\label{eq.3}
\lim_{i{\rightarrow}\infty}\left\{ \lambda_1^{L_1}(\Omega_i) - \frac{\int_{\Omega_i}(-3K) g_i^2 dM}{\int_{\Omega_i} g_i^2 dM}\right\}$$ is the limit of the *mean value* of $3K$ in $\Omega_i$ with respect to the volume element $g_i^2dM.$ There are three possibilities for the limit of the quotient of the integrals in $(\ref{eq.3}):$
- It may be infinite, in which case, because $\lambda_1^{L_1}(\Omega_i){\rightarrow}0,$ the expression (\[eq.3\]) is negative after some $i_0;$
- It may be finite but non-zero, in which case, by the same reason, the expression is negative after some $i_0;$
- It might be zero. Then we use for the first time the hypothesis that $K$ is bounded away from zero to conclude this case cannot happen.
Therefore, $M$ is unstable, thus proving Theorem \[T-1\].
In order to prove Theorem \[stability\] we need the following Poincaré type inequality.
\[mod.poincare\] Let $M^3$ be a hypersurface of ${\mathbb{R}}^4$ such that $H>0$ and $R=0.$ Let $\Omega\subset M$ be a regular domain. If $u\in H_0^1(\Omega)$ is a non-negative function, then $$\label{new.sobolev.2}
\int_\Omega uHdM \leq \frac{r_\Omega}{\sqrt{6}}\int_\Omega H^{1/2}\left{\langle}P_1({\nabla}u),{\nabla}u\right{\rangle}^{1/2}dM,$$ where $r_\Omega$ denotes the radius of the smallest extrinsic ball which contains $\Omega.$
Initially, let $B_{r_\Omega}(x_0), x_0\in{\mathbb{R}}^4$ be the smallest ball of ${\mathbb{R}}^4$ containing $\Omega$ and $\rho(x)=\rho(x_0,x)$ be the extrinsic distance from $x_0$ to $x\in M.$ Since $\Omega\subset B_{r_\Omega}(x_0),$ then, for all $x\in\Omega,$ $$\label{diam}
\rho(x)\leq r_\Omega.$$ We claim that $$\operatorname{div}_M(P_1(\rho{\nabla}\rho)) = 6H.$$ In fact,$$\begin{split}
\operatorname{div}(P_1(\rho{\nabla}\rho))&=\operatorname{div}(3H\rho{\nabla}\rho-A(\rho{\nabla}\rho))\\
&=3{\langle}{\nabla}H,\rho{\nabla}\rho{\rangle}+ 3H\operatorname{div}(\rho{\nabla}\rho) - \operatorname{div}(A(\rho{\nabla}\rho)).
\end{split}$$ Since $\operatorname{div}(A(\rho{\nabla}\rho))=3H + 3{\langle}\rho{\nabla}\rho,{\nabla}H{\rangle},$ we obtain our claim. This implies $$\begin{split}
\operatorname{div}(uP_1(\rho{\nabla}\rho))&=u\operatorname{div}(P_1(\rho{\nabla}\rho)) + {\langle}{\nabla}u,P_1(\rho{\nabla}\rho){\rangle}\\
&=6uH + {\langle}{\nabla}u,P_1(\rho{\nabla}\rho){\rangle}.\\
\end{split}$$ Integrating the expression above over $\Omega$ and by using the divergence theorem, we have $$0=6\int_\Omega uH dM + \int_\Omega{\langle}{\nabla}u,P_1(\rho{\nabla}\rho){\rangle}dM,$$ i.e., $$\label{p-1}
\int_\Omega uH dM = \frac{1}{6}\int_\Omega{\langle}{\nabla}u,P_1(-\rho{\nabla}\rho){\rangle}dM.\\$$ Since $R=0$ and $H>0,$ then $P_1$ is positive semi-definite. In fact, if $R=0$ then $(3H)^2 = |A|^2 \geq k_i^2,$ for all $i=1,2,3,$ where $k_i$ are the principal curvatures of $M^3.$ Thus $0\leq (3H)^2 - k_i^2 = (3H - k_i)(3H + k_i)$ which implies that all eigenvalues $3H - k_i$ of $P_1$ are non-negative, provided $H>0,$ i.e., $P_1$ is positive semi-definite. Thus, by using Cauchy-Schwarz inequality, we obtain $$\begin{split}
{\langle}{\nabla}u, P_1(-\rho{\nabla}\rho){\rangle}& = {\langle}\sqrt{P_1}({\nabla}u),\sqrt{P_1}(-\rho{\nabla}\rho){\rangle}\\
& \leq |\sqrt{P_1}({\nabla}u)||\sqrt{P_1}(-\rho{\nabla}\rho)|\\
& = {\langle}P_1({\nabla}u),{\nabla}u{\rangle}^{1/2}{\langle}P_1(\rho{\nabla}\rho),\rho{\nabla}\rho{\rangle}^{1/2}\\
& \leq (\operatorname{tr}_M P_1)^{1/2}\rho {\langle}P_1({\nabla}u),{\nabla}u{\rangle}^{1/2} |{\nabla}\rho|\\
& \leq \sqrt{6}H^{1/2}\rho {\langle}P_1({\nabla}u),{\nabla}u{\rangle}^{1/2}.
\end{split}$$ Introducing inequality above into (\[p-1\]), we have $$\int_\Omega uH dM \leq \frac{1}{\sqrt{6}}\int_\Omega \rho H^{1/2}{\langle}P_1({\nabla}u),{\nabla}u{\rangle}^{1/2} dM.$$ Therefore, by using (\[diam\]), $$\int_\Omega uH dM \leq \frac{r_\Omega}{\sqrt{6}}\int_\Omega H^{1/2}{\langle}P_1({\nabla}u),{\nabla}u{\rangle}^{1/2} dM.$$ This proves the Proposition \[mod.poincare\].
We now prove Theorem \[stability\].
Without loss of generality, choose an orientation of $M^3$ such that the mean curvature $H>0.$ The hypothesis $R=0$ and the inequality (\[desig\]), p. , imply $K\leq0.$ Since $L_1(\frac{1}{2}g^2)=gL_1g + {\langle}P_1({\nabla}g),{\nabla}g{\rangle}$ for any $g$ with compact support, after integrating and using the divergence theorem, stability becomes equivalent to $$\label{ineq.stab}
-3\int_{\Omega} K g^2 dM \leq \int_{\Omega}{\langle}P_1({\nabla}g),{\nabla}g{\rangle}dM,$$ for any smooth function $g:\Omega\subset M{\rightarrow}{\mathbb{R}}$ with compact support, satisfying $g|_{\partial\Omega}=0,$ where $\Omega$ is a regular domain. The proof will be made by contradiction. Suppose $\Omega$ is unstable. Then there exists a smooth function $g:\Omega{\rightarrow}{\mathbb{R}},$ with compact support, satisfying $g|_{\partial\Omega}=0,$ such that $$\label{eq.est.1}
-3\int_{\Omega}Kg^2 dM > \int_{\Omega} {\langle}P_1 ({\nabla}g), {\nabla}g{\rangle}dM.$$ Choosing $u=g^2$ in the inequality (\[new.sobolev.2\]) of Proposition \[mod.poincare\], we have $$\int_{\Omega} g^2 HdM\leq \frac{2r_\Omega}{\sqrt{6}}\int_{\Omega} H^{1/2}g{\langle}P_1({\nabla}g),{\nabla}g{\rangle}^{1/2}dM.$$ By using the Cauchy-Schwarz inequality in the right hand side of the inequality above, we have $$\int_{\Omega} H^{1/2}g{\langle}P_1({\nabla}g),{\nabla}g{\rangle}^{1/2}dM\leq \left(\int_{\Omega} g^2 HdM\right)^{1/2}\left(\int_{\Omega} {\langle}P_1({\nabla}g),{\nabla}g{\rangle}dM\right)^{1/2},$$ and therefore, $$\left(\int_{\Omega} g^2 HdM\right)^{1/2}\leq\frac{2r_\Omega}{\sqrt{6}}\left(\int_{\Omega} {\langle}P_1({\nabla}g),{\nabla}g{\rangle}dM\right)^{1/2}.$$ By using the hypothesis (\[eq.est.1\]), we have $$\begin{split}
\int_{\Omega} g^2 H dM &\leq \frac{2r_\Omega^2}{3} \int_{\Omega} {\langle}P_1({\nabla}g),{\nabla}g{\rangle}dM\\
&<\frac{2r_\Omega^2}{3}\int_\Omega (-3K)g^2dM\\
&\leq \frac{2r_\Omega^2}{3}\sup_\Omega\left(\dfrac{-3K}{H}\right)\int_\Omega g^2 H dM,\\
\end{split}$$ i.e., $$1<\frac{2r_\Omega^2}{3}\sup_\Omega\left(\dfrac{-3K}{H}\right)$$ which is a contradiction.
[^1]: Hilário Alencar and Manfredo do Carmo were partially supported by CNPq of Brazil
[^2]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using density-functional molecular dynamics simulations we analyzed the cooling-rate effects on the physical properties of GeS$_2$ chalcogenide glasses. Liquid samples were cooled linearly in time according to $T(t) = T_0 - \gamma t$ where $\gamma$ is the cooling rate. We found that our model leads to a promising description of the glass transition temperature $T_g$ as a function of $\gamma$ and gives a correct $T_g$ for experimental cooling rates. We also investigated the dependence of the structural properties on the cooling rate. We show that, globally, the properties determined from our simulations are in good agreement with experimental values and this even for the highest cooling rates. In particular, our results confirm that, in the range of cooling rates studied here, homopolar bonds and extended charged regions are always present in the glassy phase. Nevertheless in order to reproduce the experimental intermediate range order of the glass, a maximum cooling rate should not be exceeded in numerical simulations.'
address: |
Laboratoire de Physicochimie de la Matière Condensée - Institut Charles Gerhardt\
Université Montpellier 2, Place E. Bataillon, Case 03,\
34095 Montpellier, France\
author:
- Sébastien Le Roux and Philippe Jund
title: 'Influence of the cooling-rate on the glass transition temperature and the structural properties of glassy GeS$_2$: an [*ab initio*]{} molecular dynamics study'
---
Introduction
============
During the last fifteen years the unceasingly growing interest for chalcogenide glasses has lead to numerous works. This is in particular true for germanium disulfide glasses for which a large amount of experimental [@1999JPCS...60.1473H; @1986PhRvB..33.5421B; @2001JNCS...279.186; @1996JNCS...202.248] and theoretical [@2003PhRvB..67i4204B; @2004PhRvB..70r4210B; @2004PhRvB..69f4201B; @PhysRevB.60.R14985] studies have been carried out. Indeed its known properties used for example in optical amplifiers, memory switching devices or anti-reflection coatings [@1999JNS...243.116] make him a good candidate for intensive research.\
Among the different research means, molecular dynamics (MD) simulations can be a very interesting tool to provide detailed information on the physical properties of such glassy systems. Firstly because they allow to investigate the structure in full microscopic detail giving access to the position of the atoms and, secondly because they are useful to study dynamical phenomena accessible to such simulations [*e.g.*]{}, for time scales between 10$^{-13}$ and 10$^{-8}$ s.\
If the temperature of a liquid is decreased so much that the relaxation time of the system becomes longer than the time scale of the computer simulation or of the experiment, the system undergoes a kinetic arrest and, provided that it does not crystallize, will undergo a glass transition and remain trapped in a disordered configuration. However it has been demonstrated in both, experiment [@1992PhRvB..4611318B; @1994PhRvB..49.3124B] and computer simulations [@1989PhRvA..40.6007M; @1996PhRvB..5415808V], that the properties of the resulting glass, like the density or the glass transition temperature, will depend on its thermal history and in particular on the rate at which the sample is cooled down.\
Previous studies [@2003PhRvB..67i4204B; @2004PhRvB..70r4210B; @2004PhRvB..69f4201B] have already validated our “Cook and Quench” model to produce and study GeS$_2$ glasses using approximate [*ab initio*]{} molecular dynamics simulations. Nevertheless, because of the time scale of our computer simulations which is many orders of magnitude shorter than the typical experimental one, the glass transition temperature, for example, appears to be significantly higher than the one observed in the laboratory ($\sim$750K [@1986PhRvB..33.5421B; @2001JNS...293.169; @PhysRevLett.78.4422]) and thus it is necessary to see how the properties of the so-obtained glass depend on the way it was produced.\
Thus, in the present paper, we focus on the cooling rate effects on some physical properties of glassy GeS$_2$. Firstly we investigate how it does affect the glass transition temperature $T_g$ and secondly we study how it does affect the structural properties of the glassy samples.\
The paper is consequently organized as follows. In section \[theo\] we briefly present the theoretical model used in our calculations, results and discussions are presented in section \[res\], and finally in section \[conclu\] we summarize the major conclusions of our work.
\[theo\]Theoretical framework
=============================
Computations were performed using Fireball96, an approximate [*ab initio*]{} molecular dynamics code based on the local-orbital electronic structure method developed by Sankey and Niklewski [@1989PhRvB..40.3979S]. The electronic structure is described using density functional theory (DFT) [@1964PhRv..136..864H] within the local density approximation (LDA) [@1965PhRv..140.1133K] and the non-local pseudo-potential scheme of Bachelet, Hamann and Schlüter[@1982PhRvB..26.4199B]. To reduce the CPU time we used the non-selfconsistent Harris functional [@1985PhRvB..31.1770H] with a set of four atomic orbitals (1 “s” and 3 “p”) per atom that vanish outside a cutoff radius of 5a$_0$ (2.645Å). This model has been successfully used the last ten years for several different chalcogenide systems [@2003PhRvB..67i4204B; @2004PhRvB..70r4210B; @2004PhRvB..69f4201B; @2001PhRvB..64j4206L; @1997PhRvB..56.3054C; @2000JPCM...12L..21O].\
All the calculations of the present simulations were performed in the microcanonical ensemble, with a time step $\Delta$t=2.5fs and using only the $\Gamma$ point to sample the Brillouin zone. The initial configuration of our system was a crystal of $\alpha$-GeS$_2$ in a cubic cell of 19.21Å containing 258 atoms with standard periodic boundary conditions, melted at 2000 K for 60ps in order to obtain an equilibrium liquid. Then, we relaxed this system further at 2000K for 50ps and we choose five different liquid samples (approximately every 10ps) during this process. Each sample was then quenched down to 300K through the glass transition temperature $T_g$. Quenches were carried out by a linear velocity rescaling according to $T(t) = T_0 - \gamma t$ where $\gamma$ is the cooling rate. Six different cooling rates were used: 3, 0.6, 0.3, 0.12, 0.09 and 0.06$\times$10$^{14}$ K/s. At 300K each sample was relaxed over 50ps, [*i.e.*]{}, 20000 time steps. Configurations were saved every 20 steps and the results were averaged for each sample over these 1000 configurations. Furthermore we averaged the results for the 5 samples of each cooling rate thus all the data presented below have been averaged over 5$\times$1000 configurations, and the error bars (when given) represent the usual standard deviation.
\[res\]Results
==============
Glass transition
----------------
As usually done in simulations we tried to identify the glass transition of our samples by representing the evolution of the average potential energy as a function of the temperature during the quench (an example is proposed for one of the samples cooled at 0.09$\times$10$^{14}$ K/s (Fig. \[ftg\])), the glass transition being localized by the change in the evolution of the energy. The glass transition temperature $T_g$ is determined by the intersection of the two linear regressions at high and low temperature (Fig. \[ftg\]). For each cooling rate we averaged the values of the $T_g$ obtained for each of the five GeS$_2$ samples, thus we obtained an evolution of the average $T_g$ with the cooling rate. As shown in Fig. \[Tg\] the error bars remain huge indicating an insufficient sampling (it is worth noting that the computer time needed to perform the simulations for the lowest values of $\gamma$ is of the order of 12 weeks for one sample) and it is therefore difficult to extract an accurate description of the evolution of $T_g$ with the cooling rate. Nevertheless we attempted to fit these average simulation values assuming a power-law dependence of $T_g$ with the cooling rate as suggested by the mode-coupling theory [@1992JRPP...55.241] : $$T_g(\gamma)= T_c + (A\gamma)^{1/\delta}$$ with $T_c=676\pm75 K$, $A=3.0\pm2.3\times 10^{31}$ and $\delta =17.1\pm0.1$ (the errors given for the parameters of the fit have been evaluated by fitting the extreme (lowest and highest) Tg values enclosed in the error bars). We observed that a variation of $\gamma$ by about 1 decade gives rise to a variation of $T_g$ of about 75K which is not that larger, compared to the difference of magnitude of the cooling rates, than the variation of 10K measured in real experiments for different materials [@1992PhRvB..4611318B; @1994PhRvB..49.3124B; @1992JRPP...55.241]. Then by extrapolating the results of our fit down to usual experimental cooling rates i.e. 10$^0$-10$^5$ K/s, we observed at this scale that a variation of $\gamma$ by about 1 decade gives rise to a variation of $T_g$ of about 13K, which is in agreement with the experimental variation of 10K previously mentioned. We have represented in Fig. \[Tg\] the result of this work together with a few experimental data [@1986PhRvB..33.5421B; @2001JNS...293.169; @PhysRevLett.78.4422]. The agreement between the extrapolation of the fit and the experimental values of $T_g$ shows that our model is able to give a correct [*tendency*]{} of the variation of $T_g$ with $\gamma$, even though the poor statistics prevents us from having accurate estimates of $T_g$ at a specific (high) cooling rate. This can be improved with more simulations in order to reduce the error bars.
Structural properties
---------------------
### Radial pair correlation functions and bonding properties
\
\
In glassy GeS$_2$ the basic building blocks are GeS$_4$ tetrahedra, connected together forming a random network. The structural disorder is reflected by the absence of long range order and by the wide distribution of bond lengths and bond angles. Structural information may be extracted from the radial pair correlation function $g(r)$ which can be defined for a given $\alpha,\beta$ pair by: $$\label{grij}
{\tt{g}}_{\alpha \beta}(r)= \frac{V}{4\pi r^{2} \rho N c_{\alpha} c_{\beta} } \sum_{i \ne j} \delta (r-r_{ij})$$ where $\rho$ is the number density of the system, $c_{\alpha}$ the fraction of species $\alpha$ in the system, $i$ the atoms of species $\alpha$ and $j$ the atoms of species $\beta$. For each cooling rate we averaged the radial pair correlation functions g$_{\alpha \beta}$(r) of the five samples, so we are able to compare the evolution of the average g$_{\alpha \beta}$(r) according to the cooling rate (Fig. \[GrGeGe\] and \[GrSS\]).\
The bond lengths appear to be in good agreement with experimental data[@1986PhRvB..33.5421B] since we find 2.23Å for the Ge-S bond (expe.: 2.21Å), 2.91Å and 3.49Å for respectively the edge and corner sharing Ge-Ge connections (expe: 2.91Å and 3.42Å). This good agreement is true even for the highest cooling rates. The main influence of the cooling rate is reflected in the small peak corresponding to homopolar bonds between 2.2-2.6Å for Ge (Fig. \[GrGeGe\]) and 2.1-2.45Å for S (Fig. \[GrSS\]). Our results indicate that the number of homopolar bonds decreases with the cooling rate and it is therefore justified to address the question of the existence of homopolar bonds at experimental cooling rates which is still an open question (Cai and Boolchand using Raman scattering experiment, found the existence of homopolar bonds in glassy GeS$_2$ [@PhilosMagB.82.1649] while Petri and Salmon found no evidence of such bonds in $g$GeS$_2$ using neutron diffraction [@2001JNS...293.169] studies). According to Fig. \[GrGeGe\] and Fig. \[GrSS\] it seems that the decrease of the proportion of homopolar bonds slows down for the lowest cooling rates and tends towards a limit of respectively 1.9$\%$ for the Ge atoms and 1.2$\%$ for the S atoms. These limiting values are small but nonzero and therefore our simulation results seem to confirm the existence of homopolar bonds in experimental glassy GeS$_2$.\
The simulation gives access to the positions of the atoms, therefore we can also obtain information on the connectivity of the network. In our approach we focused on the ratio between edge and corner sharing tetrahedra and the evolution of the different local environments of the Ge (Tab. \[EnvGe\]) and S atoms (Tab. \[EnvS\]).\
------ ------------------- ------------------- --------------------- --
[**Ge**]{}(S$_4$) [**Ge**]{}(S$_3$) [**Ge**]{}(GeS$_3$)
3 93.0 $\pm$2.0 2.3 $\pm$1.8 3.7 $\pm$2.1
0.6 95.0 $\pm$1.6 1.7 $\pm$1.1 1.9 $\pm$2.0
0.3 96.2 $\pm$1.8 0.3 $\pm$0.3 2.8 $\pm$2.0
0.12 96.4 $\pm$1.8 1.7 $\pm$1.8 1.9 $\pm$2.0
0.09 97.6 $\pm$1.7 0.1 $\pm$0.1 1.8 $\pm$1.0
0.06 98.1 $\pm$1.0 0.0 $\pm$0.0 1.9 $\pm$1.1
------ ------------------- ------------------- --------------------- --
: \[EnvGe\]Evolution of the local structural environment of Ge atoms as a function of the cooling rate $\gamma$
------ ------------------- --------------- ------------------- ---------------- --------------------
[**S**]{}(Ge$_2$) [**S**]{}(Ge) [**S**]{}(Ge$_3$) [**S**]{}(GeS) [**S**]{}(Ge$_2$S)
3 67.9 $\pm$2.2 14.3 $\pm$1.5 12.9 $\pm$1.8 3.1 $\pm$1.3 1.8 $\pm$0.9
0.6 71.0 $\pm$4.1 12.9 $\pm$2.0 12.6 $\pm$1.8 2.1 $\pm$0.8 0.4 $\pm$0.3
0.3 75.5 $\pm$2.6 11.4 $\pm$1.0 10.7 $\pm$1.2 2.0 $\pm$1.3 0.4 $\pm$0.6
0.12 76.1 $\pm$2.7 10.9 $\pm$1.2 10.5 $\pm$1.3 1.7 $\pm$0.8 0.6 $\pm$0.6
0.09 78.2 $\pm$3.3 10.4 $\pm$1.6 9.9 $\pm$1.4 1.1 $\pm$1.1 0.4 $\pm$0.3
0.06 78.5 $\pm$1.0 10.4 $\pm$0.4 9.8 $\pm$1.0 0.8 $\pm$0.3 0.4 $\pm$0.3
------ ------------------- --------------- ------------------- ---------------- --------------------
: \[EnvS\]Evolution of the local structural environment of S atoms as a function of the cooling rate $\gamma$
\
First it is worth noting that the proportion of edge and corner-sharing links is almost a constant independent of the cooling rate: 84$\%\pm$1.8 of corner-sharing and 16$\%\pm$1.8 of edge-sharing bounds, values in good agreement with experimental data [@1986PhRvB..33.5421B]. This is in contrast with the evolution of the proportions of Germanium (Tab.\[EnvGe\]) and Sulfur (Tab.\[EnvS\]) in their standard environment (respectively a four fold S coordination for Ge and a two fold Ge coordination for S) which appreciably increase with decreasing cooling rate. The second point concerns, as expected, the decrease of the chemical disorder with decreasing cooling rate. Indeed the proportion of under-coordinated Ge atoms (2.3$\%$ for the fastest cooling rate) disappears for the slowest cooling rate. And the proportions of non-bridging S atoms and over-coordinated S atoms, respectively 14.25$\%$ and 12.85$\%$ decrease to 10.4$\%$ and 9.8$\%$.\
These results indicate that the cooling rate has an impact on the structure of the glass. Nevertheless while certain types of structural “defects” disappear at low rate, others survive and can therefore be considered as inherent of the glassy structure.
### Neutron static structure factor
\
\
An alternative way to analyze the structure is to compute the static neutron structure factor $S(q)$ which can be directly compared to neutron scattering experiments:\
$$\label{sqeq}
S(q)=\frac{1}{N} \sum_{j,k} b_j\,b_k \left< e^{iq[r_j-r_k]} \right>$$ where N is the number of atoms, and $b_j$ is the neutron scattering factor for atom $j$.\
As for the radial pair distribution functions, we averaged for each cooling rate the total neutron structure factors of the five samples, permitting thus the comparison of the evolution of the average structure factors as a function of the cooling rate (Fig. \[Sq\]).\
First we note the accurate description of glassy GeS$_2$ reflected in the good agreement between the simulated curves and the experimental one, and this even for the highest cooling rates. Although the simulated and experimental curves present differences in the range 0-2.5Å$^{-1}$ it has already been shown [@2004PhRvB..70r4210B; @2004PhRvB..69f4201B; @blaineau-2006-] that the physical properties of the so-simulated glassy samples are in good agreement with experiment. The First Sharp Diffraction Peak (FSDP), signature of the intermediate range order (IRO) in amorphous samples, appears at $\approx$1Å$^{-1}$ and is, as expected in such simulations, slightly underestimated [@2003PhRvB..67i4204B]. Size effects can be considered as an explanation, nevertheless we note that a decrease of the cooling rate globally improves the calculated structure factor and in particular the FSDP (Fig. \[Sq\]), which highlights that the IRO is also cooling rate dependent. However there is no linear/regular evolution of S(q) with the cooling rate. To illustrate the improvement of S(q) in the FSDP region with the decrease of the cooling rate, we represent (Fig. \[dsq\]) the difference, between 0 and 2.5Å$^{-1}$, of the experimental total neutron structure factor and:\
1. the total neutron structure factor of the liquid state at 2000K:\
$\qquad \qquad S(q)_{exp} - S(q)_{liq} = \Delta S(q)_{exp-liq}$
2. the average total neutron structure factor of the simulated glass quenched down to 300K at the rate $\gamma$:\
$\qquad \qquad S(q)_{exp} - S(q)_{\gamma} = \Delta S(q)_{exp - \gamma}$
The $\Delta S(q)_{exp-liq}$ is a reference representing the biggest variation between the experimental and simulated S(q) and should be compared to $\Delta S(q)_{exp - \gamma}$. We see in Fig. \[dsq\] that by decreasing the cooling rate, the difference $\Delta S(q)_{exp - \gamma}$ decreases. In particular it appears that for cooling rates higher than 0.3x10$^{14}$ K/s the differences $\Delta S(q)_{exp - \gamma}$ are very close to the difference $\Delta S(q)_{exp-liq}$ . For the highest cooling rate (3x10$^{14}$ K/s) this difference is even quasi identical to $\Delta S(q)_{exp-liq}$ which indicates that there is no real change in the IRO between the fastest quenched glass and the liquid phase. One can thus argue that the fastest cooled samples are too similar to the liquid and can therefore not be considered as glassy GeS$_2$ samples. This defines a limit for the maximum cooling rate usable in MD simulations in order to avoid interferences between the liquid and the glassy state. In our simulations this limit appears to be between 0.6x10$^{14}$ K/s and 0.3x10$^{14}$ K/s. Nevertheless this limit is directly related to the method [*i.e. ab-initio*]{} MD simulations and indirectly related to our model [*i.e.*]{} the nature of the glass and the characteristics of the atomic pseudopotentials and therefore the numerical value of this limit can not be straightforwardly extended to other simulated glassy systems.
### Atomic charges
\
\
Even if atomic charges cannot actually be determined experimentally, relevant tools such as Löwdin [@1950JChemPhys..18.365] or Mulliken [@1955JChemPhys..23.1833] population analysis can be used to compare different configurations with the [*same*]{} description. In the present work the Löwdin description has been chosen in order to compare the dependence of the atomic charges on the cooling rate. It should however be mentioned that the non-self-consistent Harris functional is known to overestimate the charge transfers between the atoms.\
The atomic charge [*q*]{} is calculated by the difference between the number of electrons of the neutral atom and the “real” number of electrons of the atom in the glass. We found no dependence of the Löwdin charges on the cooling rate. Thus it is necessary to correlate the charges with the evolution of the proportion of each atomic type in its local structural environment with the cooling rate (Tab. \[EnvGe\] and \[EnvS\]).\
As expected the general polarity of the Ge-S bond is found with a charge transfer in an ordered Ge(S$_4$)$_{1/2}$ configuration of +0.94 for the Ge atoms and -0.46 for the S atoms. As it has already been shown in a previous work [@2004PhRvB..70r4210B] Ge charges are always positive and decrease with the number of neighbors whereas S charges are more variable with respect to the local environment: from strongly negative charges for non-bridging S atoms (-1.07) to almost neutral charges for 3 fold Ge-coordinated S atoms. And even if the existence of positively charged S atoms in the environment [**[S]{}**]{}-S is confirmed in our present work, Tab. \[EnvS\] shows that this kind of local structural environment disappears rapidly with decreasing cooling rate. These structures obtained at high cooling rate may be explained by the results shown in Tab. \[ttble\_q3\], representing the charge of an atom in a given local structural environment and the proportion of atoms in this environment in an equilibrium GeS$_2$ liquid at 2000K.
[ccc]{} Environment & [*q*]{} & Proportion\
(S$_4$) & +0.96$\pm$0.04 & 64.83$\%$\
[**Ge**]{}(S$_3$) & +1.02$\pm$0.06 & 27.41$\%$\
[**Ge**]{}(GeS$_3$) & +0.75$\pm$0.1 & 1.14$\%$\
\
Ge-[**S**]{}-Ge & -0.43$\pm$0.15 & 60.22$\%$\
Ge-[**S**]{} & -0.92$\pm$0.21 & 26.03$\%$\
[**[S]{}**]{}-Ge & -0.03$\pm$0.11 & 9.74$\%$\
Ge-[**[S]{}**]{}-S & -0.1$\pm$0.18 & 2.07$\%$\
[**[S]{}**]{}-S & 0.32$\pm$0.2 & 1.60$\%$\
Indeed it appears that the configuration of the liquid is quite similar, at least for the S atoms, to the fastest cooled glass configurations (Tab. \[EnvS\]). This result indicates that the fastest cooled glasses are, in the literal sense, frozen liquids. This confirms what we have already detected in the total neutron structure factor.\
Positively and negatively charged zones inside the glass have been reported in our previous study [@2004PhRvB..70r4210B] for the highest cooling rate. In order to see if the atomic charges measured for the lowest cooling rate confirm or reject the existence of such zones we looked at the short-range charge deviation $\Delta Q_{SR}$ of a particle $i$ defined by: $$\label{eqdqsr}
\Delta Q_{SR}(i) = q(i) + \sum_{j=1}^{n(i)} \frac{q(j)}{n(j)}$$ This allows to take into account the atomic charge $q(i)$ of a given particle $i$ as well as the charges on its $n(i)$ nearest neighbors (determined from the radial pair distribution function). Whereas for a crystalline structure, in which no bond defects are present, this value is almost zero for all the particles, positive and negative values appear for glassy samples.\
In Tab. \[dqsr\] we have reported the evolution of the number of charged zones (a charged zone contains at least 2 nearest neighbors having the same sign for $\Delta Q_{SR}$) as a function of the cooling rate.\
------ -------------------------- -------------------------- --
$\Delta Q_{SR}$ $>$ +0.3 $\Delta Q_{SR}$ $<$ -0.3
3 9.0 $\pm$1.0 20.8 $\pm$1.2
0.6 9.0 $\pm$2.0 19.2 $\pm$0.8
0.3 8.4 $\pm$1.6 17.2 $\pm$0.8
0.12 8.6 $\pm$1.4 16.0 $\pm$1.0
0.09 7.8 $\pm$1.8 15.8 $\pm$0.2
0.06 9.2 $\pm$0.8 16.4 $\pm$0.6
------ -------------------------- -------------------------- --
: \[dqsr\]Evolution of the average number of charged zones as a function of the cooling rate $\gamma$
------ -------------------------- -------------------------- --
$\Delta Q_{SR}$ $>$ +0.3 $\Delta Q_{SR}$ $<$ -0.3
3 5.7 $\pm$0.1 2.1 $\pm$0.0
0.6 5.2 $\pm$0.4 2.1 $\pm$0.1
0.3 5.2 $\pm$1.1 2.0 $\pm$0.0
0.12 5.2 $\pm$0.2 2.1 $\pm$0.0
0.09 5.2 $\pm$1.7 2.1 $\pm$0.0
0.06 4.2 $\pm$0.2 2.1 $\pm$0.0
------ -------------------------- -------------------------- --
: \[sdqsr\]Evolution of the average number of atoms per charged zones as a function of the cooling rate $\gamma$
The number of positively charged zones (particles with $\Delta Q_{SR}$ $>$ +0.3) is almost constant and equal to 9, with no dependence on the cooling rate, whereas the number of negatively charged zones (particles with $\Delta Q_{SR}$ $<$ -0.3) decreases slightly from $\sim$20 for the samples cooled at 3$\times$10$^{14}$ K/s to $\sim$16 for the samples cooled at a rate smaller than 0.3$\times$10$^{14}$ K/s (this is again coherent with the idea of a maximum cooling rate usable in MD simulations). The negatively charged zones are principally made of Ge atoms coordinated to one or more non-bridging sulfur atoms, and for the fastest cooled glasses, of a few S-[**S**]{}-Ge structures. As already shown (Tab. \[EnvS\]) these structures disappear with decreasing cooling rate. This observation correlated to the diminution of the proportion of non-bridging sulfur atoms gives an explanation to the decrease of the number of negatively charged zones with the cooling rate for rates higher than 0.3$\times$10$^{14}$K/s. The positively charged zones are exclusively made of Ge atoms linked to over-coordinated S atoms.\
In addition we have reported the average number of atoms per charged zone as a function of the cooling rate (Tab. \[sdqsr\]). This shows that the size of the positively and the negatively charged zones (respectively with 5 and 2 atoms per zone) is independent of the cooling rate and therefore remains constant. It is worth noting that the global neutrality of the glass is always respected. Our results show that the existence of charged zones in glassy GeS$_2$ is not influenced by the variation of the cooling rate. They confirm thus those of our previous study, and show that extended charged zones (whose manifestation has also been detected recently for other chalcogenide systems by Taraskin [*et al.*]{} [@PhysRevLett.97.055504]) reflect the broken chemical order of the glass and are therefore inherent to the amorphous state.
\[conclu\]Conclusion
====================
Through DFT based molecular dynamics simulations we have analyzed the effect of the cooling rate on some properties of glassy GeS$_2$. Influence of the cooling rate on the glass transition temperature as well as on the structural properties has been studied.\
Due to a lack of statistics mainly due to computer time limitations (especially for the lowest cooling rates), the detailed variation of $T_g$ with the cooling rate could not be obtained. Nevertheless the extrapolation of our results to “realistic” cooling rates is in good agreement with the experimental glass transition temperature.\
Analyzing the radial pair distribution functions and the local structural environments for each cooling rate, we find that the number of S and Ge homopolar bonds as well as the number of coordination defects decrease with the cooling rate. Nevertheless the decrease of the homopolar bonds seems limited and therefore it is reasonable to think that this type of defect is present in real glasses. However calculations at lower cooling rates should be done to confirm this observation.\
The study of the simulated total neutron structure factor has confirmed the reliability of our model in the description of glassy GeS$_2$. We have analyzed the effect of the cooling rate on the intermediate range order whose signature is the first sharp diffraction peak. The simulated FSDP is closer to the experimental one for the slowest cooled glasses. By comparing with the properties of the liquid state, we have shown that a maximum cooling rate should not be exceeded in the simulation in order to reproduce the IRO characteristic of the glassy phase. The value of this maximum cooling rate will depend on the details of the model used to describe a given system. The existence of a maximum cooling rate has also been supported by the analysis of the charges which has revealed that the electronic configuration of the fastest cooled glasses is close to the one obtained in liquid GeS$_2$. In addition the existence of positively and negatively charged regions in the amorphous state has been clearly confirmed even for the lowest cooled samples and seems therefore inherent to the glassy state. These regions will have an important impact on the properties of samples containing metallic ions as shown recently [@blaineau-2006-].
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors wish to thank Sébastien Blaineau for his help at the beginning of this work and Annie Pradel and Benoit Coasne for profitable discussions. Parts of the calculations have been performed at the “Centre Informatique National de l’Enseignement Supérieur”(CINES) in Montpellier.
References {#references .unnumbered}
==========
[10]{}
A. C. [Hannon]{} and B. G. [Aitken]{}. 1999 , [**60**]{}:1473–1477.
P. [Boolchand]{}, J. [Grothaus]{}, M. [Tenhover]{}, M. A. [Hazle]{}, and R. K. [Grasselli]{}. 1986 , [**33**]{}:5421–5434.
J. [Shojiya]{} M. [Kawamoto]{} Y. [Miyauchi]{}, K. [Qiu]{} and N. [Kitamura]{}. 2001 , [**279**]{}:186–195.
M. [Philippot]{} E. [Descôtes]{} L. [Ibanez]{}, A. [Bionducci]{} and R. [Bellissent]{}. 1996 , [**202**]{}:248–252.
S. [Blaineau]{}, P. [Jund]{}, and D. A. [Drabold]{}. 2003 , [**67**]{}(9):094204.
S. [Blaineau]{} and P. [Jund]{}. 2004 , [**70**]{}(18):184210.
S. [Blaineau]{} and P. [Jund]{}. 2004 , [**69**]{}(6):064201.
, [Arlin]{} [Briley]{}, [Shau]{} [Grossman]{}, [Dirk]{} V. [Porezag]{}, and [Mark]{} R. [Pederson]{}. 1999 , [**60**]{}(22):R14985–R14989.
J. [Malek]{} and J. [Shanelova]{}. 1999 , [**243**]{}:116–122.
R. [Br[ü]{}ning]{} and K. [Samwer]{}. 1992 , [**46**]{}:11318–11322.
R. [Br[ü]{}ning]{} and M. [Sutton]{}. 1994 , [**49**]{}:3124–3130.
H. [Miyagawa]{} and Y. [Hiwatari]{}. 1989 , [**40**]{}:6007–6013.
K. [Vollmayr]{}, W. [Kob]{}, and K. [Binder]{}. 1996 , [**54**]{}:15808–15827.
L. [Petri]{} and P.S. [Salmon]{}. 2001 , [**202**]{}:169.
, W. J. [Bresser]{}, and P. [Boolchand]{}. 1997 , [**78**]{}(23):4422–4425.
O. F. [Sankey]{} and D. J. [Niklewski]{}. 1989. , [**40**]{}:3979–3995.
P. [Hohenberg]{} and W. [Kohn]{}. 1964 , [**136**]{}:864–871.
W. [Kohn]{} and L. J. [Sham]{}. 1965 , [**140**]{}:1133–1138.
G. B. [Bachelet]{}, D. R. [Hamann]{}, and M. [Schl[ü]{}ter]{}. 1982 , [**26**]{}:4199–4228.
J. [Harris]{}. 1985 , [**31**]{}:1770–1779.
J. [Li]{} and D. A. [Drabold]{}. 2001 , [**64**]{}(10):104206.
M. [Cobb]{} and D. A. [Drabold]{}. 1997 , [**56**]{}:3054–3065.
J. [Ortega]{}, R. [P[é]{}rez]{}, and F. [Flores]{}. 2000 :L21–L27.
W. [Gotze]{} and L. [Sjogren]{}. 1992 , [**55**]{}:241–376.
L. [Cai]{} and P. [Boolchand]{}. 2002 , [**82**]{}:1649.
S. [Blaineau]{} and P. [Jund]{}. 2006 , [**74**]{}:054203.
P.-O. [Lödin]{}. 1955 , [**18**]{}(3):365.
R.S. [Mulliken]{}. 1955 , [**23**]{}(10):1833.
S.N. [Taraskin]{}, S.I. [Simdyankin]{}, S.R. [Elliott]{}. J.R. [Neilson]{}. and. T.[Lo]{} 2006 , [**97**]{}(5):055504.
Figures {#figures .unnumbered}
=======
![\[ftg\] Average potential energy as a function of the temperature and linear regressions above and below the glass transition temperature for a sample cooled at 0.09$\times$10$^{14}$ K/s.](img/Tg/Figure1-fittg.eps){width="11cm"}
![\[Tg\] Glass transition temperature as a function of the cooling rate, fit and extrapolation to very low cooling rates - comparison with experiment [@1986PhRvB..33.5421B; @2001JNS...293.169; @PhysRevLett.78.4422].](img/Tg/Figure2-tg.eps){width="11cm"}
![\[GrGeGe\] *Partial radial pair distribution function g(r)\[Ge-Ge\] as a function of the cooling rate. Enlargement corresponds to details of the peak due to homopolar bonds.*](img/Gr/Figure3-grGeGe.eps){width="17cm"}
![\[GrSS\] *Partial radial pair distribution function g(r)\[S-S\] as a function of the cooling rate. Enlargement corresponds to details of the peak due to homopolar bonds.*](img/Gr/Figure4-grSS.eps){width="17cm"}
![\[Sq\]Total simulated static neutron structure factors. Enlargement corresponds to the FSDP part of the S(q).](img/Sq/Figure5-Sq.eps){width="17cm"}
![\[dsq\] $\Delta$S(q) for the IRO, between experiment and liquid GeS$_2$, and between experiment and glassy GeS$_2$ for the different cooling rates.](img/Sq/Figure6-DSq.eps){width="11cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Boundary layers in space and astrophysical plasmas are the location of complex dynamics where different mechanisms coexist and compete eventually leading to plasma mixing. In this work, we present fully kinetic Particle-In-Cell simulations of different boundary layers characterized by the following main ingredients: a velocity shear, a density gradient and a magnetic gradient localized at the same position. In particular, the presence of a density gradient drives the development of the lower hybrid drift instability (LHDI), which competes with the Kelvin-Helmholtz instability (KHI) in the development of the boundary layer. Depending on the density gradient, the LHDI can even dominate the dynamics of the layer. Because these two instabilities grow on different spatial and temporal scales, when the LHDI develops faster than the KHI an inverse cascade is generated, at least in 2D. This inverse cascade, starting at the LHDI kinetic scales, generates structures at scale lengths at which the KHI would typically develop. When that is the case, those structures can suppress the KHI itself because they significantly affect the underlying velocity shear gradient. We conclude that depending on the density gradient, the velocity jump and the width of the boundary layer, the LHDI in its nonlinear phase can become the primary instability for plasma mixing. These numerical simulations show that the LHDI is likely to be a dominant process at the magnetopause of Mercury. These results are expected to be of direct impact to the interpretation of the forthcoming BepiColombo observations.'
author:
- 'Jérémy Dargent , Federico Lavorenti, Francesco Califano, Pierre Henri, Francesco Pucci and Silvio S. Cerri'
bibliography:
- 'biblio.bib'
title: 'Interplay between Kelvin-Helmholtz and Lower-Hybrid Drift instabilities'
---
Introduction
============
Boundary layers spontaneously emerge in many space and astrophysical plasmas, typically in the presence of two interacting different plasma environments. That is the case of the solar wind interacting with self-generated or induced magnetospheres of various Solar System’s objects, of the interaction between the local interstellar medium and the heliopause or of astrophysical jets interacting with the surrounding environment, just to mention a few. Such layers typically exhibit large-scale variations of magnetic, density and/or velocity fields. Because those gradients represent a source of free energy for a variety of plasma instabilities, their evolution feeds back into the global evolution of such systems. A typical situation of interest for the above-mentioned cases is provided by the boundary layer forming between two different magnetized plasma flows, as it occurs at planetary magnetopauses [e.g., @Fujimoto1998; @Hasegawa2003; @MastersPSS2012; @Cerri2013; @Haaland2014; @LiljebladJGRA2015; @Cerri2018; @Malara2018]. In particular, such shear flows can be unstable to Kelvin-Helmholtz instability (KHI), developing the characteristic fully rolled-up vortex structures [e.g., @Nakamura2005; @Henri2013; @Faganello2017]. Signatures of possible KHI structures have been indeed observed at several planetary magnetpauses [see e.g., @Hasegawa2004; @SundbergJGRA2012; @DelamereJGRA2013; @Paral2013; @LiljebladJGRA2014; @GershmanJGRA2015]. These vortices may in turn feed secondary instabilities, developing on the shoulder of the primary KHI. A typical example is provided by the vortex-induced magnetic reconnection in various forms [e.g., @Nakamura2008; @Faganello2009; @Nakamura2013; @Nakamura2014; @Fadanelli2018], or by the development of pressure anisotropies able to trigger kinetic instabilities [e.g., @Decamillis2016]. Moreover, when the two plasma flows have different densities, the Rayleigh-Taylor instability (RTI) may be triggered by the large-scale vortex motion [@Matsumoto2004; @Faganello2008; @Faganello2008b]. At fluid scales, the RTI can also emerge as a primary instability because of the gravity acceleration since the magnetosheath is denser than the magnetosphere. However, since the gravity acceleration is small and because the magnetic field tends to stabilize, the RTI is usually much less considered with respect the KHI, except than in some special cases [@Guglielmi2010].
At kinetic scales, a Lower-Hybrid Drift instability (LHDI) may also emerge from the density gradient itself [@Gary1993; @Daughton2003; @Daughton2004]. The LHDI is often observed in both spacecraft data [@Mozer2011; @Norgren2012; @Graham2014; @Graham2017b; @Yoo2018; @Yoo2019] and laboratory experiments [@Carter2001; @Carter2002; @Yoo2014; @Yoo2017] and has been especially studied in the context of magnetic reconnection [@Lapenta2002; @Pritchett2012; @Roytershteyn2012; @Roytershteyn2013; @Price2016; @Le2017; @Le2018]. As a consequence, the KHI and the LHDI are expected to compete and interact in a quite complex way since their fastest growing mode (FGM) grows at different scale lengths and with different growth rates. The dynamics arising from such a competition and its nonlinear development is the main focus of this paper.
The KHI is a fluid-scale instability growing in a sheared flow. Its growth rate is controlled by the velocity shear and the corresponding gradient scale length [@Chandrasekhar1961; @Miura1982; @Faganello2017]. On the other hand, the LHDI is a kinetic scale instability. This instability is driven by the coupling of ion thermal gyration with the free energy provided by the ion density gradient drift velocity. At scales where ions are demagnetized but electrons still frozen-in, this free energy efficiently feeds Lower-Hybrid waves through an inverse ion damping that makes these waves unstable [@Gary1993]. The growth rate is controlled by the frequency ratio between the electron plasma frequency and the cyclotron frequency, the plasma beta and the density gradient. In this work, we will mainly focus on the latter.
Due to the different typical scale length at which the KHI and the LHDI develop, the interplay between these instabilities has not yet been observed. The KHI including a density gradient has been investigated either by adopting a MagnetoHydroDynamic model [@Takagi2006; @Faganello2008; @Matsumoto2010; @Leroy2017] or by adopting a hybrid kinetic model but neglecting electron inertia [@Gingell2015]. In this case, however, the LHDI can not grow since the FGM scales as $k_{FGM} \propto (mi/me)^{1/2}$ [@Gary1990] and stabilizes more and more when the density gradient scale length becomes larger than a few ion inertial lengths [@Gary1993]. There also exists some fully kinetic simulations theoretically able to develop LHDI but none of them has actually shown evidence of it, either because the layer is too large [@Matsumoto2006; @Umeda2010], either because the KHI have been excited to grow quicker [@Matsumoto2010; @Umeda2014]. As a consequence, no simulation of KHI ever reported the growth of LHDI in space plasma. On the other hand, kinetic simulations of KHI revealed another kinetic effect: the dawn-dusk asymmetry, which impact the growth of the KHI depending on the sign of $\mathbf{B} \cdot \mathbf{\Omega}$, where $ \mathbf{\Omega}$ is the vorticity at the layer [@Nakamura2010; @Henri2013; @Paral2013].
In this work, in order to investigate the interplay between the KHI and LHDI, we make use of a kinetic model of a boundary layer characterized by the presence of (i) a velocity shear that would be unstable to the KHI and (ii) a density gradient that would be unstable to the LHDI. We take the scale lengths of variation of the main fields (densities, magnetic field and velocities) of the order of the ion inertial length and vary their respective values. These configurations allow us to study the relative impact of each instability on the layer and on the plasma mixing.
The present paper is organized as follows. [Sec.]{}[sec:param]{} describes the numerical model used in this paper. [Sec.]{}[sec:res]{} presents the results, and is split into the different phases of the simulations. [Sec.]{}[ssec:LHDI]{} shows the linear growth of the LHDI, when applicable, and the comparison of our numerical results with linear theory. [Sec.]{}[ssec:nonlinear]{} contains the nonlinear phase of the LHDI and describes the presence of an inverse cascade of energy responsible for the formation of large scale structures. In some regime of parameters, those structures are shown to interfere with the growth of the KHI that develops on larger time scales. This feature, together with the factors playing a role in the competition between the nonlinear phase of the LHDI and the linear growth of the KHI are studied in [Sec.]{}[ssec:KHI]{}. In [Sec.]{}[sec:dis]{}, we discuss the results of this paper and the inherent limitations of the used model. Finally, in [Sec.]{}[sec:con]{}, we present a summary of this study, together with a description of possible consequences for the dynamics of planetary magnetospheres, such as the hermean magnetopause.
Numerical setup {#sec:param}
===============
We present four fully kinetic simulations in a two-dimensional (2D) cartesian geometry using the Particle-In-Cell (PIC) code SMILEI [@Derouillat2017]. The simulations model the boundary between two plasmas characterized by different densities and separated by a velocity shear. Simulations are performed in the reference frame where the low-density plasma medium is at rest.
The data presented are normalized using ion scale quantities. The magnetic field and density are normalized to arbitrary value $B_0$ and $n_0$, respectively. We choose $B_0$ and $n_0$ such that the density and magnetic field are equal to one on the flowing side of the layer (in our simulations: the right side). The masses and charges are normalized to the proton mass $m_p$ and charge $e$, time is normalized to the inverse of the proton gyrofrequency $\omega_{ci}^{-1}=m_p/eB_0$ and length to the proton inertial length $\delta_i=c/\omega_{pi}$, where $c$ is the speed of light and $ \omega_{pi}=\sqrt{n_0e^2/m_p\epsilon_0}$ is the proton plasma frequency. Velocities are normalized to the ions’ Alfvén velocity $v_{Al} = \delta_i \omega_{ci}$.
All simulations are initialized with a single layer where density, velocity (directed along the $y$-direction) and magnetic field (directed along the $z$-direction) vary along the $x$ direction. This layer is contained in the $(x,y)$ plane in a 2D domain of size $(x_{max},y_{max})=(68,136)~\delta_i$. There are $n_x=n_y=2720$ cells in the $x$ and $y$ directions, corresponding to a grid resolution of $\Delta_x = 0.025~\delta_i$ and $\Delta_y = 0.05~\delta_i$.
The ion and electron distribution functions are initially composed by $50$ macro-particles per cell loaded using Maxwellian distributions. Plasma moments and electromagnetic forces are calculated using second order interpolation. The time step is calculated using a Courant-Friedrichs-Lewy (CFL) condition which in our simulations turns out to be $\Delta_t=8.4\cdot 10^{-4} ~\omega_{ci}^{-1}$, and the total simulation time is $400~\omega_{ci}^{-1}$. The boundary conditions are periodic in the $y$ direction and reflective in the $x$ direction for both particles and fields.
The initial density profile is given by: $$\begin{aligned}
n_i(x,y) &=& \frac{1}{n_r} \left[ 1 + \frac{n_r-1}{2} \left( 1+ \tanh \left( \frac{x-x_0}{L} \right) \right) \right] + n_{curr}
\label{eq:ni}\end{aligned}$$ where $n_r=n_{a}/n_{b}$ is the density ratio across the current sheet, $L$ the initial characteristic width of the layer, fixed at $L=1$ in this work, and $n_{curr}=\mathbf{J}^2/(e^2n_0v_{Al}^2)$ a small correction to allow ions to carry part of the current in order to avoid inconsistencies (negative electron density in case of very strong magnetic field gradient). The subscript $a$ and $b$ stands for the asymptotic values on both sides of the layer. The values of $n_r$ are listed in [Tab.]{}[tab:asym]{}. The ions velocity is initialized as: $$\mathbf{v}_i(x,y) = \left[\frac{\Delta v_{shear}}{2} \left( 1+ \tanh \left( \frac{x-x_0}{L} \right) \right)
\right] \mathbf{e}_y
~+~ \mathbf{v}_{curr}
\label{eq:vi}$$ where $\Delta v_{shear} = |v_{a} - v_{b}|$ is the velocity difference across the shear layer and $\mathbf{v}_{curr}=\mathbf{J}/(en_i)$ a small correction to allow ions to carry part of the current, in order to avoid inconsistencies, as consistent with the density correction. The values of $\Delta v_{shear}$ are listed in [Tab.]{}[tab:asym]{}. The initial magnetic field profile is given by: $$\begin{aligned}
\mathbf{B}(x,y) &=& \frac{1}{B_r} \left[ 1 + \frac{B_r-1}{2} \left(1+ \tanh \left( \frac{x-x_0}{L} \right) \right) \right] \mathbf{e}_z
\label{eq:B}\end{aligned}$$ where $B_r=|B_{a}/B_{b}|$ is the magnetic asymptotic field ratio across the current sheet. The values of $B_r$ are listed in [Tab.]{}[tab:asym]{}. From the curl of $\mathbf{B}$ we calculate the current $\mathbf{J}$ since in our model the displacement current is neglected. Finally, the electric field is initially set to: $$\mathbf{E} =-\mathbf{v}_i \times \mathbf{B} + \frac{1}{en_i} \mathbf{J} \times \mathbf{B}
\label{eq:E}$$ which correspond to an Ohm’s law where the initial electron pressure gradient term has been neglected. The $\nabla(P_e)/en_e$ term in the generalized Ohm’s law is by far the smallest contribution to the electric field. To ease the implementation of the initial conditions, we have neglected the electron pressure gradient term, and later checked and confirmed that this term can be neglected. However, note that this simplified equation does not perfectly match the Ohm’s law within the boundary, as we consider a charge separation ($n_i \ne n_e$) associated to the electric field gradient [@Pritchett1984]. Finally, electron density $n_e = n_i - \nabla \cdot \mathbf{E}$ and mean velocity $\mathbf{v}_e = (\mathbf{J}-n_i\mathbf{v}_i)/n_e$ consistently follow from $\mathbf{J}$ and Maxwell equations.
The total scalar pressure $P=P_i+P_e$ is determined in order to preserve pressure balance. The electron to ion temperature ratio is taken constant and equal to $\theta=T_e/T_i=0.2$. The plasma $\beta= 2P/B^2$ is set equal to 1 in plasma “a” (left side). A reduced mass ratio $m_i/m_e = 25$ is used for computational reasons. The consequences of using a reduced mass ratio is explained in details in [Sec.]{}[sec:dis]{}. We fix $\omega_{pe}/\omega_{ce}=4$. To optimize the layer stability, we make use of finite Larmor radius correction in the simulation setup, as described in @Cerri2014.
Simulation 1 2 3 4
-------------------- ----- ----- ----- -----
$B_r$ 0.5 0.5 0.5 0.5
$n_r$ 10 10 5 1
$\Delta v_{shear}$ 0 0.5 0.5 0.5
: Magnetic field ratio ($B_r$), density ratio ($n_r$) and velocity difference ($\Delta v_{shear}$) values characterizing the discontinuity between the two adjacent plasmas, for the 4 simulations. Simulation 2 parameters correspond to parameters consistent with Mercury’s magnetopause [@Slavin2008]. []{data-label="tab:asym"}
Numerical results {#sec:res}
=================
The simulations are characterized by three main phases. In the first one, a Lower Hybrid Drift instability (LHDI) develops very fast in all simulations except in simulation 4 where no density gradient is present. In the second phase, the LHDI saturates and enters the nonlinear stage, which is characterized in simulations 1 and 2 by the growth of large scale finger-like structures. Finally, in the last phase, the Kelvin-Helmholtz instability (KHI) may develop depending whether or not the nonlinear LHDI structures grew fast enough (i.e., to a sufficiently large amplitude) to destroy the coherence of the shear flow. In order to better illustrate the different phases, movies of the out-of-plane magnetic field time evolution are available for simulations 2, 3 and 4 as supplementary material. In the following, we discuss how the system evolves in each of these phases for each simulation.
The linear LHDI phase {#ssec:LHDI}
---------------------
![From left to right panel: in-plane electron velocity component $v_x$ at $t=7.5$ for simulations 1, 2 and 3, respectively. Simulation 4 is not plotted here because in the absence of a density gradient, the LHDI does not develop, and consequently there are no visible fluctuation at the layer this early in the simulation. []{data-label="fig:Vex7"}](fig1.jpeg){width="\linewidth"}
At the initial times of all simulations (except for simulation 4), we observe growing fluctuations along the layer at $x \simeq 34$. In particular, in [Fig.]{}[fig:Vex7]{} we show the in-plane electron velocity component $v_x$ for simulations 1, 2 and 3, where the fluctuations length scale is observed at kinetic scale. Furthermore, the amplitude of fluctuations is bigger in simulations 1 and 2 than in simulation 3 since the instability depends on the density gradient (see [Tab.]{}[tab:asym]{}). Furthermore, a wave front propagating in the $x$ direction is clearly visible as a consequence of the absence of a kinetic equilibrium in the layer’s initialization and our choice to neglect the pressure gradient term in [Eq.]{}[eq:E]{}. However, this artificial by-product of the initialization has not a significant impact on the evolution of the simulation, as typically the case for kinetic simulation. Those waves are especially clear in simulation 3 since the color scale amplitude has a reduced range with respect to simulations 1 and 2.
To identify the instability as a LHDI, we compare our results with the corresponding linear theory. We solve the linear Vlasov equation under some simplified assumptions to compute the growth rate of the LHDI [@Gary1978; @Gary1983; @Gary1993]. The simplifications are made in order to carry out the analytical model and are the following: density and temperature gradients much larger than ion gyro-radius, uniform out-of-plane magnetic field and a low plasma beta, $\beta= 2P/B^2<<1$. Our method consists in integrating the linearized Vlasov equation along unperturbed orbits, thus finding an expression for the dielectric function, and finally finding the zeros of this function numerically.
In our case, we have used a dieletric function $\varepsilon(k, \omega)$ of the form [@Gary1979; @Sgro1989; @Gary1993] : $$\begin{aligned}
\label{eq:Vlasov_dispersion}
\begin{aligned}
\varepsilon(k, \omega) &= 1 + \sum_{j} K_j(k, \omega) \\
K_j(k, \omega) &=\frac{A_j^2}{k^2}\bigg(1-[\omega-kv_{nj}] e^{-k^2 \rho^2_{Lj}}\sum_{m=-\infty}^{+\infty}\frac{I_m(k^2 \rho^2_{Lj})}{\omega + m\omega_{cj}}+ \\
&+\frac{\omega v_{nj}}{k v^2_{th,j} }e^{-k^2 \rho^2_{Lj}}\sum_{m=-\infty}^{+\infty} \frac{m \omega_{cj} I_m(k^2 \rho^2_{Lj})}{\omega + m\omega_{cj}}\bigg)-\\
&-k^3\rho^2_{Lj} v_{Tj} e^{-k^2 \rho^2_{Lj}}\sum_{m=-\infty}^{+\infty} \frac{I_m(k^2 \rho^2_{Lj})- I'_m(k^2 \rho^2_{Lj})}{\omega + m\omega_{cj}}\\
\end{aligned}\end{aligned}$$ where $k$ and $\omega$ are the wave vector and frequency, respectively, and $K_j$ is the dielectric susceptibility of the species $j$ (the sum over $j$ running on all the species, $j=(i,e)$ in our case). $I_m$ are the modified Bessel functions of order m. Moreover, $v_{th,j}= \sqrt{T_j/m_j}$ and $\rho_{Lj}= \sqrt{m_j T_j}/q_j B$ are the thermal velocity and the Larmor radius of the species $j$, respectively, and $B$ is the magnetic field modulus. Finally $m_j$ and $q_j$ are the mass and charge of species $j$, respectively, and $\omega_{cj}=q_jB/m_j$ is the gyrofrequency of the population $j$. The drift velocities associated with the density ($n_j$) and temperature ($T_j$) gradients are defined as $v_{Fj}=\epsilon_{_F} \rho_{Lj} v_{th,j}$, where $F_j=(n_j,T_j)$ and $\epsilon_{_F} = ({1}/{F_j(x)}) d F_j(x) / dx$. We also define $A_j= \omega_{pj}/v_{th,j}$. All the above formulas are in ion units ($\delta_i$ and $\omega_{ci}^{-1}$).
As already said, this method assumes a gradient configuration with a density and temperature gradient length scale ($L_F\sim\epsilon_F^{-1}$) much larger than ion gyro-radius ($\rho_{Li}/L_F\ll1$), a uniform out-of-plane magnetic field and a low plasma beta, $\beta = 2P/B^2<<1$. Although those assumptions do not strictly hold in our simulations, the neglected effects would tend to slow down the process, so that we can consider the analytical theory as an upper limit for the growth rates and a useful reference for a comparison with simulation results in the linear regime [@Davidson1977]. Indeed, since $v_{nj}$ and $v_{_{Tj}}$ are proportional to the density and temperature gradients, the LHDI growth rate is expected to be larger for steeper gradients of $n$ and $T$. In the configuration adopted for the linear theory, we shall define the inverse gradient length $\epsilon_{_{T,n}}$ as the one computed at $x \simeq x_0$, i.e. at the initial location of the shear layer (see [Sec.]{}[sec:param]{}).
The linear theory assumes a uniform magnetic field $B$ (and thus a uniform Larmor radius $\rho_{_{Li}}$), while in the simulation the magnetic field is sheared. Therefore, we must fix a characteristic (mean) value of $\rho_{Li} = \sqrt{\beta_i m_i/2 e n_i} \equiv \sqrt{\beta_i /2 n_i}$ in normalized units. We take it as the local value at the center of the magnetic gradient, i.e. at $x=x_0$, and report the results in [Tab.]{}[tab:grad]{}.
Simulation 1-2 3 4
--------------- ------ ------- -------
$\epsilon_n $ 0.82 0.67 0
$\epsilon_T $ -0.7 -0.48 -0.27
$\rho_{Li} $ 2.18 1.73 1.42
: Gradients and Larmor radius in the different simulations. []{data-label="tab:grad"}
We have computed the dispersion relation using the parameters listed in [Tab.]{}[tab:grad]{}. The corresponding results, associated to the four kinetic simulations, are reported in [Fig.]{}[fig:LHDIrate]{}.
![a) Theoretical dispersion relation of the LHDI branch obtained from [Eq.]{}[eq:Vlasov\_dispersion]{} with $m_i/m_e=25$, and for the set of parameters given in [Tab.]{}[tab:grad]{}. The dotted lines represent the growth rate and the solid lines represent the real frequencies. b) Fourier transform along y of the density field at x=34 and t=5 for simulation 1, 2 and 3. []{data-label="fig:LHDIrate"}](fig2.png){width="\linewidth"}
The main result emerging from [Fig.]{}[fig:LHDIrate]{}$a$ is that the LHDI is expected to grow on ion kinetic scales in simulations 1-2 and 3, while in simulation 4 the system is stable with respect to the LHDI. The LHDI fastest growing mode is $k_{FGM} = 3.9$ for simulations 1-2 and 3, while the growth rate is $\gamma_{LHDI}=1.3$ in simulation 1-2 and $\gamma_{LHDI}=0.8$ in simulation 3 because of the different density gradients. In [Fig.]{}[fig:LHDIrate]{}$b$ we show the density spectrum observed in our simulations at $t=5$. The peak around the FGM is in agreement with linear theory. On the other hand, simulation 4 does not develop any instability at such early times. As expected, the instability grows faster in simulation 1 and 2 than in simulation 3. We therefore conclude that the instability observed in the early stage of simulations 1, 2 and 3 corresponds to a LHDI.
The growth rate in simulations is estimated by fitting the temporal evolution of the Fourier transform of $n_i$ (see [Fig.]{}[fig:LHDIrate]{}.$b$) for $k=4$ (the fastest growing mode) during the linear phase of LHDI with an exponential function $Ae^{t \gamma_{LHDI}}$. Quantitatively, the measured growth rates in simulations 1, 2 an 3 turn out to be smaller, by a factor of three, than the ones estimated by linear theory (see [Eq.]{}[eq:Vlasov\_dispersion]{}). Such a discrepancy is explained by the inherent limitations of the theoretical model [@Davidson1977], which considers a smooth density gradient, a uniform magnetic field and a low plasma $\beta$. Finally, it is worth noticing that the presence of a velocity shear does not affect too much the instability, as shown in [Fig.]{}[fig:LHDIrate]{}$b$, where simulations 1, with no velocity shear, and 2, with a velocity shear, behave similarly.
The nonlinear LHDI phase {#ssec:nonlinear}
------------------------
![Ion density at $t=150$ for simulations 1, 2 and 3.[]{data-label="fig:fingers"}](fig3.jpeg){width="\linewidth"}
The LHDI ends its linear growth phase long before the KHI starts to develop, thus entering its nonlinear stage. In simulations 1 and 2 such transition to the non linear regime occurs at around $t_{NL}\sim8$, eventually leading to an effective inverse cascade with the formation of many fluid scale structures. This is shown in [Fig.]{}[fig:fingers]{} where we draw the density contours for simulations 1, 2 and 3 in the nonlinear stage of the LHDI. Furthermore, in simulations 1 and 2 we observe the development of a new process associated to the inverse cascade that produces elongated finger-like structures entering one into the other plasma. This is not observed in simulation 3 at similar times since the inverse cascade is less developed due to a slower growth of the LHDI. Such phenomenon of finger-like structures generation is similar with what has been previously observed in kinetic simulations [@Brackbill1984; @Gary1990; @Singh1998].
As stated above, since in simulation 3 the LHDI is less efficient, the inverse cascade occurs later and is much less intense. Even though some finger-like features start to develop at later times, their growth is slower in simulation 3, on a time scale comparable with the growth of the KHI. Indeed, in this case the competition between the growth of nonlinear LHDI structures and linear KHI actually prevents the finger-like structures to grow, eventually suppressing them completely once the KH vortices form. This last stage where KHI can develop will be further discussed in [Sec.]{}[ssec:KHI]{}.
![ Time evolution of Fourier coefficients modulus of ion density fluctuations , $|\widehat{\delta n}_{i,k}(x)|=|\mathrm{FFT_y}(n_i)|$, for some given $k$ at $x=33.25$ from simulation 1. Panel $a$ and $b$ share the same k, while panel $c$ plot other k. The k shared by all panel are plotted with thick lines in panel $c$. []{data-label="fig:evolmodes"}](fig4.jpeg){width="1\linewidth"}
In [Fig.]{}[fig:evolmodes]{}$a$ we show the time evolution of the modules of different Fourier coefficients for density of simulation 1. [Fig.]{}[fig:evolmodes]{}$b$ shows a zoom of [Fig.]{}[fig:evolmodes]{}$a$ at early times, when the energy growth is driven by the linear development of the LHDI. [Fig.]{}[fig:evolmodes]{}$c$ shows a zoom at late times of [Fig.]{}[fig:evolmodes]{}$a$, although plotting other $k$. [Fig.]{}[fig:evolmodes]{}$c$ shows small k ($k \leqslant 1$) in order to show the evolution of the inverse cascade at large scales.
As discussed in [Sec.]{}[ssec:LHDI]{} the FGM of the LHDI is at about $k=4$ (yellow curve). However, in the nonlinear phase after $t_{NL} \sim 8$, we observe that this mode stops growing and its associated energy starts to decrease. This transition marks the saturation phase of the LHDI and the beginning of the inverse cascade phase which feeds the growth of lower $k$’s modes. Indeed, after $t_{NL}$ all wave numbers larger than $k=4$ (yellow and black curves) decrease while smaller wave numbers continue to grow. After some time, the low $k$ modes begin to decrease one after the other, except for $k<1$ which continues to grow even faster. On longer times, we observe in [Fig.]{}[fig:evolmodes]{}$c$ that the inverse cascade continues, with the energy evolving towards ever larger scales (i.e. smaller $k$).
To calculate a characteristic time-scale associated with the inverse cascade process and the corresponding growth of the fluid-scale structures, we measure the position $x_0(y)$ of the center of the layer at any value of $y$. To get $x_0(y)$, we fit each cut along $x$ of the density with an hyperbolic tangent and take the inflection point of the fit as $x_0(y)$. The time evolution of the standard deviation $\sigma_0=\sqrt{<x_0^2>_y}$ gives an estimation of the growth of the finger-like structures.
![Time evolution of $\sigma_0=\sqrt{<x_0^2>_y}$ from simulations (dots) and corresponding exponential fit (dashed lines). Different colors represent different simulations (see legend). The characteristic time $\tau_{NL}$ obtained from those fits is given in the legend. The horizontal black dashed line gives the value of $\sigma_0\sim 3$ that the KHI reaches in its nonlinear stage at $t \sim 300$, in simulation 4. []{data-label="fig:stdev"}](fig5.png){width="1\linewidth"}
The time evolution of $\sigma_0$ is shown in [Fig.]{}[fig:stdev]{} for simulations 1, 2 and 3. The growth of $\sigma_0$ turns out to be fitted quite well by an exponential function of the form $Ae^{t/\tau_{NL}}$, especially for simulations 1 and 2. We therefore determine a characteristic growth time $\tau_{NL}$ of the nonlinear LHDI structures from the exponential fit of the curves in [Fig.]{}[fig:stdev]{}. Note that the exponential fit match very well an exponential growth for simulation 1 where no velocity shear flow is present. Despite the similarities between simulations 1 and 2, we observe that in the early phase simulation 2 does not match very well an exponential behavior and that the observed characteristic time of the nonlinear LHDI structures is a bit longer than the case with no velocity shear. Thus we conclude that the velocity shear has a small impact on the early growth of the large-scale fluid structures. The exponential fit match even less well for simulation 3, which we attribute to the competition between the nonlinear LHDI, weakened by a smaller density gradient, and the KHI.
The KHI phase {#ssec:KHI}
-------------
![Magnetic field along $z$ at $t=350~\omega_{ci}^{-1}$ for simulations 2, 3 and 4.[]{data-label="fig:KH"}](Bz350.jpeg){width="\linewidth"}
Since a velocity shear is present in simulations 2, 3, and 4 the onset of a KHI could be expected. [Fig.]{}[fig:KH]{} shows the out-of-plane magnetic field at $t=350~\omega_{ci}^{-1}$ for the three simulations with a velocity shear.
In simulation 4, where the LHDI does not develop, a series of KHI vortices forms. In simulation 3, despite the patchy layer produced by LHDI, the KHI dominate the large scale nonlinear dynamics to form KHI vortices. Note that the fastest growing mode is different between simulations 3 and 4, despite having the same initial velocity shear and layer width. This is likely due to the widening of the initial shear layer induced by the LHDI in simulation 3. Finally, simulation 2 does not develop KH vortices. Nevertheless, the dynamics observed in simulation 2 is very similar to that of simulation 1 (which does not have a velocity shear–not plotted here), with only one difference that finger-like structures are drifting along y in the former, while they do not in the latter. The nonlinear finger-like structures developed by the LHDI grow too quickly and reach the size of the expected KHI (black dashed line in [Fig.]{}[fig:stdev]{}) vortices long before the KHI could actually emerge. Thus, the layers coherence is long gone at the time were we would expect KHI. In simulation 2, the KHI has been killed by the nonlinear LHDI before it can develop.
Now we try to estimate the growth rate of the fastest growing mode of the KHI, i.e. $\gamma_{KH}$, in all simulations with a velocity shear. In order to do that, we use the result of @Michalke1964 (table 1 in the article): the fastest growing mode of the KHI over a layer with a hyperbolic tangent shape, $\tanh(x/L)$, is given by $k_{KH}L=0.44$, and $\gamma_{KH} = 0.095 \Delta u /L$. Such result is obtained considering an incompressible fluid, in the absence of a magnetic field and with no density asymmetry. The incompressibility approximation roughly holds because the fast magnetosonic Mach number $M_f= u/\sqrt{c_A^2 + c_s^2}$ ($c_A$ the Alfvén velocity and $c_s$ the sound velocity) is always smaller than 0.05 on both sides. We can however expect that minor compressibiliy effects would tend to just slightly reduce the growth rate of the KHI in the above incompressible limit. Then the presence of a magnetic field here also plays no relevant role because it is perpendicular to the propagation plane. Note however that the presence of a magnetic field gradient can also generate instabilities [@Huba1980b], but we neglect this term as we don’t see any $\nabla B$-drift induced instability in our simulations. For what concerns the homogeneous density approximation, in general we can expect that a density asymmetry $n_r$ would indeed alter the results of @Michalke1964. However, following @Chandrasekhar1961, we do expect that the growth rate would be modified as follows:
$$\gamma_{KH} \propto \frac{\sqrt{n_r}}{1 + n_{r}}
\label{eq:gKH}$$
Simulation 2 3 4
--------------- ------- ------- -------
$\gamma_{KH}$ 0.013 0.018 0.024
$\tau_{KH} $ 77 57 42
: Theoretical growth rates $\gamma_{KH}$ of the KHI for a layer width $L$ picked at $t=10$. $\tau_{KH}=1/\gamma_{KH}$ is the characteristic time of the KHI growth.[]{data-label="tab:KH"}
In [Tab.]{}[tab:KH]{}, we have computed the growth rate for the KHI in the three simulations with the parameter $L$ computed at $t=10.0$, when the linear growth of the LHDI is over. The width $L$ of the layer is calculated by fitting the profile of $v_y$ averaged along $y$ with [Eq.]{}[eq:vi]{}, where $x_0$ and $L$ are left as free parameters. At this time the layer width is more or less the same for all the simulations, corresponding to $L\sim2$. However, the layer width $L$ keeps growing slowly but continuously in simulations 2 and 3 because of the nonlinear LHDI. Thus, in practice, the KHI growth rate is decreasing with time in those simulations.
$\tau_{KH}$ is calculated for simulation 4 with the method used in [Fig.]{}[fig:stdev]{} (i.e. by fitting the curve of $\sigma_0(t)$ during its exponential growh, given that $\sigma_0$ is proportional to the KH wave amplitude) gives us a result of $\tau_{KH} \approx 49$. Our theoretical $\tau_{KH}$ is actually in good agreement with the observed one (42 vs 49), the slight discrepancy can be due to the error associated with $L$, and the wave vector of maximum growth matches with the one observed in the simulation $k=0.23$. These results help us to clarify what’s happening in the simulations.
In simulation 2 and 3 the LHDI grows much faster than the KHI, so we expect the LHDI to start the inverse-cascade process well before the KHI begins. Indeed we see that the cascade begins at $t\sim10$ and reaches fluid scales ($k\sim1$) at $t\sim20$. After that time the competition between the nonlinear LHDI and the KHI begins. In simulation 2, $\tau_{NL}<<\tau_{KH}$, so the nonlinear LHDI totally dominate the evolution and the KHI has no opportunity to grow. In simulation 3, $\tau_{NL} \sim 2 \tau_{KH}$ so even if the nonlinear LHDI grows significantly, the KHI can still dominate. This case is interesting as both instabilities manage to develops enough to both impact the layer structure. In simulation 4 the LHDI is suppressed because there isn’t any density gradient. So we see the formation of KHI vortices with a growth rate comparable to the one predicted by the theory.
Discussions {#sec:dis}
===========
In this work we have shown that, despite the difference in characteristic scale length and growth rate values between the KHI and the LHDI, in the presence of a relatively strong density gradient the LHDI not only contributes to the dynamics but even dominates at large scales. The validity of these results and their application to space plasmas are discussed in this section.
In conditions of a boundary layer with both a velocity shear and a density gradient, the KHI and LHDI might compete. This competition relies on the linear time scales at which the instabilities develop and on the nonlinear time scale at which the layer diffuses and relaxes. All that can be summarized by three basic factors: the velocity shear amplitude, the density gradient amplitude and the layer’s width. When the importance of the density gradient dominates with respect to the velocity shear, then the structures generated during the nonlinear phase of the LHDI broaden, eventually smoothing the average velocity shear layer, thus slowing down (e.g. simulation 3) or even preventing (e.g. simulation 2) the KHI growth. A broadening of the layer then tends to decrease the growth rates of both KHI and LHDI. Moreover, for layers larger than a few ion inertial lengths, the LHDI growth rate rapidly decreases and eventually becomes stable. For example, with simulation 1 parameters, the LHDI becomes stable for $L \gtrsim 8$. Therefore, when the velocity shear and the density gradient are characterized by nearly the same scale length, the layer width controls whether or not the LHDI develops. This is the reason why most of the planetary magnetopauses might be stable to the LHDI, as they are usually larger than a few ion inertial lengths, and why previous fully kinetic simulations of KHI at the Earth’s magnetopause did not see it [@Matsumoto2006; @Umeda2010; @Nakamura2014; @Nakamura2017]. On the contrary, the hermean magnetopause, characterized by a much smaller length scale of the boundary layer width, is potentially thin enough for the LHDI to develop and drive the dynamics eventually up to inhibiting the KHI development. Such a study of KHI in the fully kinetic regime is therefore especially relevant for Mercury. As a matter of fact, simulation 2 which was designed with parameters expected in the magnetosphere of Mercury, shows that LHDI dominates the dynamics of such a strongly inhomogenous shear layer.
The main topic of this study is the competition between the LHDI and the KHI in an inhomogeneous boundary layer and its nonlinear saturation. In this last phase, the mechanism behind the inverse cascade of the LHDI is out of the scope of this paper. However, it is worth reminding some previous results regarding the nonlinear phase of the LHDI concerning the inverse cascade [@Davidson1978; @Huba1978; @Gary1979; @Drake1984; @Brackbill1984; @Shapiro1994]. On the one hand, the theory of @Drake1984, which relies on mode coupling to short-wavelength damped modes, works quite well to explain the inverse cascade observed in this work. In particular, the temporal evolution of the spectra associated with the inverse cascade observed in our simulations is thoroughly described by @Drake1984. On the other hand, the modulational instability theory developed by @Shapiro1994 might explain the exponential growth observed in the nonlinear phase of the LHDI (see [Fig.]{}[fig:stdev]{}). Moreover, @Shapiro1994 predicted the formation of large-scale structures similar to those observed in our simulation at later times (see [Fig.]{}[fig:KH]{}.$1$). These two explanations might be not exclusive.
For practical reasons linked to the full PIC modeling, our simulations use (i) a reduced mass ratio (ii) and a reduced geometry (2D). Thus, the scales separation is much smaller than for a realistic mass ratio and it affects the LHDI development. Typically, the fastest growing mode of the LHDI in our simulations is for $k_{FGM}\sim 4$, while for a realistic mass ratio it would been $k_{FGM}\sim 40$ [@Gary1993]. @Drake1984 argued that a realistic mass ratio will support the nonlinear evolution of the LHDI due to two reasons. First, the ratio of the rate of change of the magnetic field energy to particle drift energy scales as $m_i/m_e$ in a finite $\beta$ plasma [@Drake1981], so much more energy will be available to supply the inverse cascade. Second, the number of unstable modes scales as $\sqrt{m_i/m_e}$ [@Huba1980]. So, for realistic values, we expect a much broader spectrum of unstable modes to be excited. For these reasons, we expect the LHDI nonlinear phase to develop faster for a realistic mass ratio as compared to our simulations. To support our claim, we observe that the growth of the nonlinear LHDI structures is faster than ours in the paper of @Gary1990, where the mass ratio is $m_i/m_e=100$, despite a slightly smoother density gradient. Future studies will look at the impact of the mass ratio for realistic magnetopause parameters.
As this work has been performed in a reduced 2D configuration, understanding the changes that would arise for a 3D configuration to the linear and nonlinear development of the LHDI is a future necessary step. First of all, in 3D the instability is no more confined to the plane perpendicular to B, so we expect growing modes with $\mathbf{k}\cdot \mathbf{B}\neq0$. The nonlinear relaxation in 3D could lead to the formation of elongated cigar-shaped structures [@Shapiro1994], to the acceleration of electrons parallel to the magnetic field [@Singh1998; @Bingham2002], and to mode conversion towards different kinds of waves such as, e.g., whistlers mode [@Camporeale2012]. How such effects would affect the competition and interaction between the LHDI and the KH is unclear and will be investigated in the future.
The work described in this paper highlights the importance of processes other than KHI to generate large scale structures responsible for plasma mixing along the magnetopause. In regions characterized by a strong density inhomogeneities, such as the magnetopause of Mercury, the LHDI provides another efficient mechanism for plasma mixing, together with the KHI reported by MESSENGER observations [@Slavin2008]. Some differences between the structures generated by KHI and LHDI should help us to identify them. Typically, during its late nonlinear phase, the KHI generates a diffuse layer [@Matsumoto2010], while the LHDI generates large scale finger-like structures that do not evolve into a diffusive layer [@Brackbill1984; @Gary1990; @Singh1998; @Bingham2002]. Future studies should be dedicated to the interplay between KHI and LHDI structures in observations in order to understand the actual role of the LHDI in an hermean-like magnetosphere. This requires observations at the lower hybrid frequency, not available from MESSENGER data, but planned with the BepiColombo mission, especially with the PWI consortium onboard the Mio (MMO) spacecraft [@Kasaba2010].
Conclusion {#sec:con}
==========
In this paper, we have performed several 2D-3V kinetic simulations to study the linear and nonlinear evolution of a magnetized shear flow separating plasmas of different densities and magnetic field intensity. This study shows that the large scale structures that emerge from the nonlinear phase of the LHDI can interfere with the KHI development if the nonlinear LHDI driven dynamics modifies the velocity shear layer. In a layer subjected to both a velocity shear and a density gradient, we have shown that, at early time and within the parameters range used in this study, a LHDI develops first. Then, the LHDI enters a nonlinear phase and energy starts to cascade from small to large scales. Depending on the growth rate of those structures, the system can become totally dominated by the nonlinear LHDI and the KHI won’t be able to develop. On the contrary, if the nonlinear LHDI growth rate is smaller than the KHI growth rate the KHI dominates and leads to the formation of vortices. However, those vortices are already partially mixed due to the simultaneous, even if less efficient, development of the LHDI at smaller scales.
The range of parameters used in this study encompasses those expected in the magnetosphere of Mercury (e.g. simulation 2). We therefore expect that these kind of structures could be detected by future observations provided by the BepiColombo space mission.
Acknowledgments {#acknowledgments .unnumbered}
===============
This project (JD, FC) has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 776262 (AIDA). We acknowledge ISCRA for awarding us access to the supercomputer Marconi at CINECA, Italy, where the calculations were performed. We thank M. Guarrasi (CINECA) for useful discussion about code implementation on Marconi. The work at LPC2E/CNRS was supported by CNES and by ANR under the financial agreement ANR-15-CE31-0009-01. The work by F. Pucci has been supported by Fonds Wetenschappelijk Onderzoek - Vlaanderen (FWO) through the postdoctoral fellowship 12X0319N. S.S.C. is supported by the National Aeronautics and Space Administration under Grant No. NNX16AK09G issued through the Heliophysics Supporting Research Program.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a hidden variable analysis of collapse dynamics in which the state’s reduction process may take a finite time $\delta t$. A full characterization of the model is given for the case of black boxes. By introducing nonlocal perfect correlations to a two black-boxes scenario, it is shown that in order to avoid faster than light communication, the reduction time associated to the system must be strictly null. Furthermore we prove that the result above holds even when there is a time window between the choice of both part’s inputs. Our results represent a new evidence of the instantaneous nature of the wave function collapse process which could have implications in foundations of quantum mechanics and information science.'
author:
- 'M. G. M. Moreno'
- Alejandro Fonseca
- 'Márcio M. Cunha'
bibliography:
- 'Bib.bib'
title: 'Finite-time collapse process and non-local correlations are incompatible with non-signaling theories'
---
Introduction
============
After almost a century of Quantum Theory, one of its fundamental postulates represents yet an issue: the measurement problem. This widely discussed phenomenon, in principle forbidden by the linearity of Schrödinger’s equation, motivated many works aiming to present an explanation on the mechanism leading the system, in a probabilistic way to the collapsed state, in perfect harmony with Born’s rule and Lüders postulate. From the perspective of the Copenhagen interpretation those problems are solved by taking them as fundamental postulates [@bohr1928quantum]. Following the *Many World* interpretation, all possible outcomes from a measurement coexist in different universes [@everett1957relative], avoiding the necessity of Lüders postulate nevertheless it doesn’t explain the mechanism behind Born’s rule. Decoherence [@zurek2003decoherence] approaches the problem by showing that a quantum system in contact with a environment should swiftly reach a classic statistical distribution, nonetheless closed systems remain a problem. Among the attempts to solve the measurement problem, the most plausible explanation is that Quantum Mechanics is an approximated theory, in the sense that there must a nonlinear equation that rules the dynamics of all systems which in the microscopic limit becomes approximately linear in accordance with Schrödinger equation, after all no microscopic experiment ever indicated Quantum Theory to be wrong. Such a theory should also be stochastic, since any nonlinear deterministic extra term in Schrödinger’s equation leads to signaling [@gisin1990weinberg]. That approach motivated many collapse models introduced in the last few decades, as well as several experiments were proposed to investigate the features of the wave function collapse [@bahrami2014proposal; @diosi2015testing; @goldwater2016testing; @nimmrichter2014optomechanical; @genoni2016unravelling; @bilardello2017collapse; @vinante2016upper; @vinante2017improved]. For a complete review on the current state of the area see [@bassi2013models].
If collapse comes from a dynamical process, a fundamental question should be addressed: *once initiated, how long, in average, does the process of collapse take to be accomplished?* The mean time of the collapse may bring a deep insight on the dynamics that rules it. This problem, which has already been considered and experimental tests to verify the duration of the collapse have been proposed [@parisio2011estimating; @moreno2013investigation; @moreira2018toward], will be the central issue of this work.
In order to investigate the mean time associated with the collapse process, the present work brings into play a quite general approach: a hidden variables model, in a Device Independent (DI) scenario. A hidden variables model allows for the contemplation of possible consequences and effects of unknown parameters, and its dynamics, that may be playing some rule on a given problem. Under this approach, we are particularly interested into hypotheses on the set of hidden variable which can generate appreciable differences in the outputs of some experiment. This method is well know by the seminal work of J. Bell [@Bell1964]. Recently Bedingham employed this concept to perform a link between the collapse dynamics and the Bohmian mechanics [@1751-8121-44-27-275303] (see also [@1751-8121-44-47-478001]). On the other hand, the DI certification program provides robust results, for it only relies on the statistics of a given experiment and it is not necessary to make any extra assumption on the system to be tested. A system under the DI procedure is treated as an assortment of boxes equipped with buttons which after being pushed produce one out of an array of outcomes, in general different in each run of the experiment. Associated with hidden variable models, the DI idea has been of remarkable importance on protocols of certification. For an introduction to the subject, we refer the reader to [@scarani2012device]. At the end of the day, from the set of frequencies it is possible to infer underlying properties of the whole system. In this paper, by using nonlocally correlated systems and applying the ideas above we obtain general results that are independent of Quantum Mechanics, however, they lead to very remarkable consequences in its context.
The paper is organized as follows: In section II we show the usual hidden variable (HV) model in the context of non-local correlations. After in section III we introduce a HV model to treat the problem of a single system subject to an arbitrary collapse dynamics. Section IV is devoted to make a connection between nonlocal correlations and collapse dynamics under the HV models overlook, and we present some implications of the collapse dynamics. In section V we expose our main conclusions.
Hidden variables and non-local correlations
===========================================
The most known application of hidden variables models is perhaps the definition of local correlations, associated with the derivation of Bell’s inequalities. In this problem it is investigated the statistical behavior $P(a,b|x,y)$ of two separated parts in which inputs are performed, $x$ in one and $y$ in the other side, generating outputs $a$ and $b$ respectively. We can always write: $$\label{HV1}
P(a,b|x,y)=\int_{\Lambda}d\lambda\zeta(\lambda|x,y)P(a,b|x,y,\lambda).$$ In the above expression, $\lambda$ represents variable(s) which are sampled from a set $\Lambda$, following a distribution $\zeta (\lambda|x,y)$, responsible for the probability of the system, and yet out of the reach for the experimenter.
Equation (\[HV1\]) represents the basic assumption behind hidden variable models, however one can always add extra hypothesis on them. For instance, one can assume that superluminal communication between parts is forbidden, i. e. the non-signaling assumption, in this case the marginal probabilities of each part should be independent of what happens in the other: $$\begin{aligned}
\nonumber
P(a|x,y,\lambda)=P(a|x,\lambda)\\
\nonumber
P(b|x,y,\lambda)=P(b|y,\lambda),\end{aligned}$$ then: $$P(a,b|x,y,\lambda)=P(a|x,\lambda)\cdot P(b|y,\lambda).$$ We can go further and assume that all correlations come from the $\lambda$ variables and that their distribution is well defined despite the inputs $x$ and $y$, i. e., $\zeta(\lambda|x,y)=\zeta(\lambda)$, which is known as the measurement independence assumption. Moreover note that the inputs $x$ and $y$ do not depend on the set of hidden variables (free will assumption). Considering this, correlations described by a local hidden variables model may be written as: $$\label{LV}
P(a,b|x,y)=\int_{\Lambda}d\lambda~\zeta(\lambda) P(a|x,\lambda)P(b|y,\lambda),$$ attainable by any classically correlated composed system.
This example illustrates the power of hidden variables assumptions and how to handle them in order to get valuable information on the system under consideration.
Hidden Variables and Collapse
=============================
Imagine Alice receives a closed box containing a flipped coin. Right before Alice looks inside the box, she would say that the probability of getting either heads or tails is one half (assuming a faithful coin). If Alice finds that the output was tails (heads) any further observation of the same coin will yield the output tails (heads) with probability one. We may say that Alice’s system collapses after it is measured. So far the problem seems very trivial: the position of the coin is well defined from the moment the system was created, and the act of looking to the coin just means to learn the value of some unknown well defined variable. This is not always the case, for instance one may consider that instead of a coin, there is an electron inside the box prepared in a spin state $(\ket{0}+\ket{1})/\sqrt{2}$, for $\ket{0}$ and $\ket{1}$ representing the eigenstates of $\hat{\sigma}_z$. When Alice carries out a spin measurement in $z$ direction, quantum theory states that there are not *a priori* established variables hidden from Alice defining the corresponding outcomes. In contrast with the former case, now a physical process is expected to take place, leading to the final outputs. This is what we mean by *collapse*.
Following the above scenario, now we consider Alice receiving a box, on which she can provide an input $x$ from a set of inputs $\mathcal{X}$ that generates some output $a\in\mathcal{A}$, with a well defined probability. Before making any further assumption, it is useful to introduce a formal distinction between the two classes of inputs that Alice may provide to her box. On one hand, we have inputs leading to outcomes in a non-deterministic way -hereafter *collapse triggering* operations (CT). On the other hand, operations conducting to deterministic outputs, defined here as *non collapse triggering* operations (NCT). Hence it is possible to divide the set of inputs $\mathcal{X}$ in two parts: $\mathcal{X}=\mathcal{X}_{CT}\cup\mathcal{X}_{NCT}$.
In order to contemplate any possible collapse dynamics (consequence of a CT input) and its effect on the description of an arbitrary system, we propose a hidden variable approach similar to that introduced in previous section. Let state some assumptions: (i) the collapse is triggered by an input $x\in\mathcal{X}_{CT}$ on the box at an instant $\tau$ which returns some output $a\in \mathcal{A}$, (ii) the system takes a time $\delta t_a$ to collapse (i.e. to generate an output), (iii) the probability of obtaining the output “$a$" as a consequence of the first input $x$ is known to be $P_0(a|x)$, and (iv) the collapse time $\delta t_a$ depends in a non trivial way on the probability of its output $P_0(a|x)$. Considering that, the most general expression describing the probability of an output $a'$ given a second input $x$ in a posterior time $\tau\leq t\leq \tau+\delta t$, is: $$\label{HV2}
P(a'|x;t>\tau)=\int_{\Gamma }d\gamma(t)\chi(\gamma(t)|x)P(a'|x;\gamma(t)),$$ here $\Gamma$, $\chi$ and $\gamma$ play the same role as $\Lambda$, $\zeta$ and $\lambda$ in equation \[HV1\], respectively. This model encompasses any possible description behind the phenomenon of collapse. In fact we could make $\chi(\gamma(t)|x)=\delta[\gamma(t)-\hat{\rho}(t)]$, where $\delta[.]$ is the Dirac’s delta and $\hat{\rho}(t)$ is a density operator, and considering the POVM $\hat{M}_x=\lbrace \hat{E}_a^x|a\in \mathcal{A}\rbrace$, and $P(a'|x;\gamma(t))=\tr(\gamma(t)\cdot\hat{E}_{a'}^x)$, then we have:$$P(a'|x;t>\tau)=\tr(\hat{\rho}(t)\cdot\hat{E}_{a'}^x),$$ which corresponds to the standard formulation in quantum mechanics [@scarani2012device].
After the largest among the collapse times ($t\geq \tau+\delta t^*$), where $\delta t^*=\max\lbrace\delta t_a\rbrace_{a\in\mathcal{A}}$, we expect the box to evolve in such a way that $P(a'|x;t\geq\tau+\delta t)=\delta_{a,a'}$, where $\delta_{a,a'}$ is the Kronecker’s delta, and $a$ represents the first output. Without loss of generality we can divide the set of variables $\Gamma$ into subsets $\Gamma_a$, each containing all possible $\gamma(t)$ leading to every output “$a$". As we know *a priori* that the result “$a$" should be obtained with probability $P_0(a|x)$, then we can write: $$P(a'|x;t)=\sum_{a\in\mathcal{A}}P_0(a|x)\int_{\Gamma_a }d\gamma(t)\chi_a(\gamma(t)|x)P(a'|x;\gamma(t)),$$ for $\tau\leq t\leq \tau+\delta t^*$. To gain some insight on this particular step, it is possible to consider the one-dimensional random walker, which after $n$ steps has a probability $P(j|n)$ of being found in the position $j$. There may exist several possible paths leading to this configuration, thus one can assemble all these paths together in the set $\Gamma_j$ and argue that with probability $P(j|n)$ a path from this set is sorted out. Also notice that no knowledge from the outputs to be obtained is required to conceive the existence of this partition, only the assumption that one of the possible results will happen.
We can simplify the above equation, by defining the functions $f_{aa'}(t)$: $$f_{aa'}(t)=\int_{\Gamma_a }d\gamma\chi(\gamma(t)|x)P(a'|x;\gamma(t)),$$ which should respect the following bounds: $$\begin{aligned}
\label{bc1}
\nonumber
&f_{aa'}(\tau) =P_0(a'|x),&\\
&f_{aa'}(t'\geq\tau +\delta t_a) = \delta_{a,a'},&\\
\nonumber
&\sum_{a'} f_{aa'}(t'\geq\tau) = 1.&\end{aligned}$$ The first condition is related to the initial probability distribution of the box, the second sets the final configuration after the collapse and the last one ensures that normalization is satisfied.
Thus using these new functions, we have that: $$\label{col0}
P(a'|x;t\geq \tau)=\sum_{a} f_{aa'}(t)P_0(a|x).$$ Note that we have not considered any specific dynamics so that this result remains as general as possible. Furthermore, the collapse time intervals $\delta t_a$ can be taken as zero or finite without loss of generality.
Equation (\[col0\]) and relations (\[bc1\]) lead to the following result: if the second input $x$ is performed at an instant $t\geq\delta t^*$, the output $a$ will occur with probability $P_0(a|x)$, however if $t\leq\delta \tilde{t}$, for $\delta\tilde{t}=\min\lbrace\delta t_a \rbrace_{a\in\mathcal A}$, then the probability distributions will depend on the functions $f_{aa'}(t)$. In particular, we can consider the quantity: $$P(a|x;\tau\leq t\leq \delta\tilde{t})-P_0(a|x),$$ which can be experimentally assessed. For a dichotomic system, the only condition that allows it to be zero would be $P(a|x)=\frac{1}{2}$, otherwise this quantity is non-vanishing.
Hidden Variables, Non-locality and collapse {#3}
===========================================
Following the previous reasoning, an extension to the case of two separated boxes sharing some correlation is presented. Two balls, one black and the other white, are randomly placed into Alice and Bob’s boxes respectively. If Alice looks inside her box and learns the colors of her ball, then due to the correlation they shared, she also learns that of Bob. Setting $a=\{0,1\}$ as the color of the ball and $x=1$ to the act of measuring it, the first time the input $x=1$ is provided, the output $a$ is obtained with probability $P_0(a|x=1)$, however any further “color measurement" will yield the output $a'$ with probability $P(a'|x=1)=\delta_{a,a'}$, given that we are leading with a locally-correlated system. The same behavior is observed in Bob’s box. Like the first example, in this case there are variables that could give a complete description of the box at first, but are not revealed to the parts. Thus the collapse represents only the knowledge of some hidden variable. No dynamical process is expected here, for there is no evidence of physical changes. To observe some collapse dynamics as discussed above, one must look for correlations that cannot be represented by a local hidden variables model (eq. \[LV\]), where the collapse represents a physical transformation in both parts.
With this in mind, assume we have two arbitrarily separated parts, Alice and Bob as usual, both in inertial reference frames, possessing nonlocally correlated boxes. Alice can provide either an input or $x=0\in\mathcal{X}_{NCT}$ or $x=1\in\mathcal{X}_{CT}$, such that $P(a|0)=\delta_{a,0}$, obtaining some output $a\in\mathcal{A}$, and Bob also gives either an input $y=0\in\mathcal{Y}_{NCT}$ or $y=1\in\mathcal{Y}_{CT}$ where $P(b|0)=\delta_{b,0}$ returning some output $b\in\mathcal{B}$. Furthermore, assume that the correlation is such that given $x=1$ and $y=1$, then $a=b$, and that the probability of the first measurement in the system $P_0(a,b|x=1,y=1)=P_0(a|x=1)=P_0(b|y=1)$ is known, where the equalities hold due to the correlation which forbids results where $a\neq b$. Notice that the input pairs $(x=0,y=1)$, $(x=1,y=0)$, $(x=1,y=1)$ are collapse triggering.
Suppose Alice and Bob agree that at an instant $\tau$ in Bob’s watch, she decides the value for $x$, while Bob sets $y=0$. And at an instant $t'\geq\tau$, Alice provides the input $x=0$, and Bob $y=1$. In this point the natural step is to compare both scenarios $x=0$ and $x=1$, given by Alice’s initial choice. This is a crucial aspect in our approach, for usually when the subject of collapse is tested with respect to whether it is signaling or not, only collapse triggering inputs are considered [@bedingham2009dynamical].
Whenever Alice chooses $x=0$, Bob providing $y=1$ in $t'$ will be able to observe that the outputs $b$ follow the known distribution: $$\begin{aligned}
\label{X1}
P(b|0,1;t')=P_0(b|1).\end{aligned}$$ Nevertheless, when she supplies $x=1$, the system as a whole starts to collapse, and according to equation (\[HV2\]) the probability in $t'$ can be described by: $$\nonumber
P(a,b|0,1;t)=\int_{\Gamma}d\gamma(t)\chi(\gamma(t)|0,1)P(a,b|0,1;\gamma(t)).$$ Following the treatment for the single box, it is a fact that the system collapses to some output $(a,b)$ with probability $P_0(a,b|1,1)=\delta_{a,b}P_0(b|1,1)$, assuming perfect correlation:$$\nonumber
P(b'|1,1;t)=\sum_{b\in\mathcal B}P_0(b|1)\int_{\Gamma_b}d\gamma(t)\chi_b(\gamma(t)|1)P(b|1,1;\gamma(t)).$$ Here we define: $$f_{bb'}(t)=\int_{\Gamma_b}d\gamma(t)\chi(\gamma(t)|1)P(b|1,1;\gamma(t)),$$ such that: $$\begin{aligned}
\nonumber
&f_{bb'}(\tau)=P_0(b'|y),&\\
\nonumber
&f_{bb'}(t'\geq\tau +\delta t) = \delta_{b,b'},&\\
\nonumber
&\sum_{b'} f_{bb'}(t'\geq\tau) = 1,&\end{aligned}$$ so we can write: $$\label{X2}
P(b|1,1;t') = \sum_a f_{bb'}(t')P_0(b|1).$$ It is always possible to find distributions for which equations (\[X1\]) and (\[X2\]) are in agreement with non-signaling conditions if and only if $f_{00}(t')=f_{11}(t')=1$ and $f_{01}(t')=f_{10}(t')=0$ for $t'>\tau$, representing an instantaneous collapse. Otherwise Alice’s choice affects Bob’s statistics. This analysis suggests that only instantaneous collapses are compatible with non-signaling theories.
The results shown above are quite general in the sense that no assumptions are made on the dynamics behind the process. However this scheme does not contemplate all possible scenarios yet. Particularly, we have assumed that Alice and Bob can choose the specific time in which the inputs are delivered, which is not necessarily feasible. For instance, following Quantum Theory one cannot choose exactly the instant in which a photon is emitted nor when it will hit the detector. Now, we repeat our analysis by taking into account that Alice and Bob can only decide the time window $\Delta t>\delta \tilde{t}$ in which the input acts happen. In addition, we consider that each of them can only perform one input: Alice deciding $x$ and Bob applying $y=1$.
When the time window starts we assume that an input takes place according to some probability distribution $g(t)$. Thus, the probability of it to happen at an instant $t'$, $\tau\leq t'\leq \tau + \Delta t$, is: $$\begin{aligned}
\mathcal{P}(t')=\int_{\tau}^{t'}g(t)dt,\end{aligned}$$ where $\int_{\tau}^{\tau+ \Delta t}g(t)dt=1$.
Now we are interested in what happens if the time interval between the instant in which Alice’s input happens, $t_A$, and the moment Bob’s input occur, (hereafter $t_B$), are smaller than $\delta \tilde{t}$. This is because if the time interval is larger, then the collapse process will be completed and the observed effect in the previous case will not play any role here. In general, we have:
$$\begin{aligned}
\nonumber
\mathcal{P}(|t_B-t_A|<\delta \tilde{t})=\int_{\tau}^{\tau+\Delta t}g(t_A)\int_{t_A-\delta \tilde{t}}^{t_A+\delta \tilde{t}}g(t_B)dt_B dt_A,\end{aligned}$$
hereafter $\Theta=\mathcal{P}(t_B-t_A\leq\delta \tilde{t})$.
Let us investigate the partial probability distribution for Bob. Once again, when Alice decides $x=0$, Bob observes that the statistics associated to his outputs is equal to the already know distribution $P_0(b|y)$. Alternatively, when she chooses $x=1$, triggering the collapse, then Bob’s probability can be described in the following way: where $\Omega=\int_{0}^{\delta \tilde{t}}p(t)dt$.
The positivity of the probability distributions guarantees that equation (\[py1\]) may be different from the a priori known distribution $P_0(b|y)$, unless $f_{ab}(t')=\delta_{a,b}$ for $t'\geq\tau$. Hence the only way to avoid signaling for any distribution is an instantaneous collapse dynamics.
Conclusion
==========
We have addressed the question of the finiteness of collapse time for two different scenarios, on one hand a single system and on the other, a bipartite correlated one, treated as boxes in analogy to Bell scenarios. Based on our idea of collapse, the inputs are divided in two sets: collapse and non-collapse triggering. A hidden variables approach is employed to model an arbitrary collapse dynamics. For the first case we demonstrate that in principle it is possible to distinguish finite time from instantaneous collapse dynamics.
For the case of two nonlocally correlated parts, we derive some conditions necessary to such correlations do not violate non-signaling constraints during the collapse process. The obtained result is quite general, and suggests that any collapse dynamics with finite time is incompatible with non-signaling constraints. Our results are particularly relevant in the context of Quantum Foundations, because it can can bring some insights on the measurement problem, still an open question nowadays.
Acknowledgements
================
We thank Rafael Chaves and Barbara Amaral for the comments and suggestions. Márcio M. Cunha is supported by FACEPE-FULBRIGHT BCT 0060-1.05/18 grant. Financial support from Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) through its program INCT-IQ, Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), and Fundação de Amparo à Ciência e Tecnologia do Estado de Pernambuco (FACEPE) is acknowledged.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate antiferromagnetic spin chains, which are coupled by a weak antiferromagnetic exchange interaction. The spins are located on a hexagonal lattice, i.e. frustration is present when three-dimensional order sets in. Typical realizations of such systems are the halides ABX$_3$. In this work we particularly study the role of the long-range dipolar interaction within the framework of a Heisenberg model with nearest-neighbor exchange and additional dipolar interaction. We perform a classical ground-state analysis and show that the spin configuration is sensitively dependent on $\kappa''$, the ratio of the dipolar interaction to the interchain interaction, as a consequence of their competing character. Several commensurate and incommensurate phases arise in the different regions of the parameter space. The ground-state investigations are supplemented by a stability analysis by means of a linear spin-wave calculation. From the magnon spectra we can show that all commensurate phases are stable against fluctuations. In comparison with experiments (CsMnBr$_3$, RbMnBr$_3$) we obtain good agreement for the energy gaps. From this we conclude that the dipolar interaction is the most important source of anisotropy in these Mn-compounds.'
address:
- 'Physik-Department, Technische Universität München, D-85747 Garching, Germany'
- 'Physics Department, University of California, Santa Cruz, CA 95064'
author:
- 'M. Hummel and F. Schwabl'
- 'C. Pich'
title: 'Phase diagram and magnons in quasi-one-dimensional dipolar antiferromagnets'
---
Introduction {#sec:intro}
============
Unconventional magnetic systems have attracted the interest of experimental and theoretical physicists in the last few years[@Sherrington]. In these systems competing interactions and/or geometric frustration due to the underlying lattice can lead to unconventional ground states, magnon spectra and magnetic phase diagrams[@diep:94]. Furthermore, fluctuations are enhanced in systems with frustrated ground states as well as in low dimensions.
An interaction, which is often competing with respect to the exchange interaction is the dipole-dipole interaction[@pich97]. In real systems the dipole-dipole interaction (DDI) is always present in addition to the short-range exchange interaction. Although its energy is lower than the exchange energy it plays an important role in low-dimensional systems due to its anisotropic and long-range character.
The most famous quasi-one-dimensional systems are the ternary compounds ABX$_3$ (A alkaline, B transition metal and X halogen), which have been studied intensively theoretically and experimentally in the context of Haldane’s phase [@Haldane] and solitonic excitations [@Steiner91]. In these systems the carrier of the magnetic moment, the B-ions, are located on a hexagonal lattice[@Collins:98]. In this work, the Mn compounds are of particular interest. Because the angular momentum $L$ is zero, no crystal-field splitting occurs in these systems and the dipole-dipole interaction should be the most important anisotropy.
The influence of the DDI in quasi-one-dimensional, antiferromagnetic spin chain systems has not yet been studied very thoroughly. Instead, the DDI is often replaced by a single-ion anisotropy[@kadowaki] or the coupling between spin chains is neglected[@dietz].
Model {#sec:model}
=====
The Hamiltonian of the dipolar antiferromagnet reads $$H=-\sum_{l \neq l'} \sum_{\alpha \beta} \bigl( J_{ll'} \delta_{\alpha
\beta} + A_{ll'}^{\alpha\beta} \bigr) S_l^{\alpha} S_{l'}^{\beta} \; ,
\label{Hamiltonian}$$ with spins ${\bf{S}}_l$ at hexagonal lattice sites ${\bf{x}}_l$. The first term describes the exchange interaction $J_{ll'}$ which includes the intrachain as well as the interchain interaction. In the following we consider only nearest-neighbor exchange, i.e. $$J_{l l'} = \left\{
\begin{array}{lll}
-J & l,l' & \mbox{along the chains}\\
-J' & l,l' & \mbox{within the basal plane}
\end{array}
\right. \quad .$$ For a hexagonal lattice the Fourier transform of the exchange energy is given by $$J_{\bf{q}}=-2J \cos q_z - 2J' \Bigl(
\cos q_x + 2 \cos \bigl( \frac{q_x}{2} \bigr) \cos \bigl(
\frac{\sqrt{3}}{2} q_y \bigr) \Bigr) \; .$$ Here and in the following we measure wave vectors in chain direction in units of $1/c$, and wave vectors within the basal planes in units of $1/a$. The second term in Eq. (\[Hamiltonian\]) is the classical dipole-dipole interaction $$\begin{aligned}
A_{ll'}^{\alpha\beta}= -\frac{1}{2} {(g \mu_B)^2} && \Biggl(
\frac{\delta_{\alpha\beta}}{{|{\bf{x}}_l-{\bf{x}}_{l'}|}^3}\nonumber\\
&&- \frac{3 ({{\bf{x}}_l-{\bf{x}}_{l'})}_{\alpha}
{({\bf{x}}_l-{\bf{x}}_{l'})}_{\beta}}{{|{\bf{x}}_l-{\bf{x}}_{l'}|}^5}
\Biggr) \, .
\label{ddw}
\end{aligned}$$ This term is evaluated by means of the Ewald summation technique[@Ewald; @Maradudin], which allows the consideration of the long-range nature of the three-dimensional DDI in terms of fast convergent sums.
Ground states {#sec:gs}
=============
In this section we calculate the classical ground states of the system as a function of the ratio of the dipolar energy to the interchain exchange interaction $$\kappa' = \frac{(g\mu_B)^2}{V_z J'} \; ,$$ where $V_z=\frac{\sqrt{3}}{2}a^2 c$ is the volume of the primitive cell. Due to the quasi-one-dimensionality of the systems (the ratio $J'/J$ is $10^{-2}
\ldots 10^{-3}$) we consider only antiferromagnetic spin configurations along the chain axis, i.e., we restrict $q_z$ to $\pi$. The ground-state energy reads $$E_g = - \sum_{l \neq l'} S_l M_{l l'} S_{l'}
= - \sum_{\bf{q}} S_{\bf{q}} M_{\bf{q}} S_{-\bf{q}}$$ with $$M_{\bf{q}}=\begin{pmatrix}
J_{\bf q} + A_{\bf q}^{11} & A_{\bf q}^{12} & A_{\bf q}^{13}\\
A_{\bf q}^{12} & J_{\bf q} + A_{\bf q}^{22} & A_{\bf q}^{23}\\
A_{\bf q}^{13} & A_{\bf q}^{23} & J_{\bf q} + A_{\bf q}^{33} \\
\end{pmatrix}
\; ,
\label{matrix}$$ where $A_{\bf q}^{13} = A_{\bf q}^{23} = 0$ for $q_z=\pi$. The ground states are specified by those wave vectors which belong to the largest eigenvalues of the matrix $M_{\bf{q}}$ in Eq. (\[matrix\]). We obtain the following phases for decreasing $\kappa'$ (for the lattice constants we took the values of CsMnBr$_3$ [@goodyear]):
1. Ferromagnetic phase: $\kappa' > \kappa_1' = 200.50$
In the region where the dipolar energy is large compared to the interchain exchange, the minimum of the ground-state energy is reached at ${\bf q}_1 = (0,0,\pi)$. This means that spins within basal planes are ordered ferromagnetically, but still antiferromagnetically along the chains.
2. Incommensurate phase I: $200.12 = \kappa_2' < \kappa' < \kappa_1'$
The ground state is an incommensurate phase in this parameter region. The wave vector moves continuously from ${\bf q}_1$ to the wave vector ${\bf q}_2 = (0,\frac{2\pi}{\sqrt{3}},\pi)$ (thick solid line in Fig. \[figbz\]) or any other path rotated by 60$^\circ$ (dashed lines in Fig. \[figbz\]).
3. Collinear phase: $17.25 = \kappa_3' < \kappa' < \kappa_2'$
Spins within basal planes are oriented ferromagnetically in chains, that are aligned antiferromagnetically to one another (see Fig. \[fig2\]). Because of the six-fold symmetry, there are six such ground states resulting from rotation of the ferromagnetic chains. Note that the continuous degeneracy is lifted.
4. Incommensurate phase II: $0 < \kappa' < \kappa_3'$
In this incommensurate phase the wave vector moves from ${\bf q}_2$ to ${\bf q}_3 = (\frac{2 \pi}{3},\frac{2\pi}{\sqrt{3}},\pi)$. The incommensurability appears because of the finite slope of the dipolar tensor at ${\bf q}_3$ [@Shiba] and the parabolic behavior of the exchange energy.
5. 120$^\circ$-structure: $\kappa' = 0$
This ground state is characterized by a three-sublattice spin configuration in each basal plane.
The coupling of the spin space to the real space induced by the DDI forces the spins to align within the lattice basal planes for all four phases, in which the DDI is non-zero. Thus, the DDI leads to an in-plane anisotropy.
![\[figbz\]Paths of the wave vector that minimize the ground-state energy in the Brillouin zone of the hexagonal lattice ($q_z=\pi$).](figure1.ps){width="30.00000%"}
The phase diagram for the whole parameter region of $\kappa'$ is shown in Fig. \[fig2\].
![\[fig2\]Spin configurations within the basal plane for different ratios of dipolar to interchain interaction $\kappa'$. Only for $\kappa'=0$ the 120$^\circ$-structure is established. For infinitesimal $\kappa'$ the phase IC II is favored. The spin configurations of incommensurate phases are not sketched.](figure2.ps){width="47.00000%"}
In summary, we find three commensurate and two incommensurate phases for arbitrary value of $\kappa'$.
Magnon spectra {#sec:ms}
==============
The spin-wave calculation is of interest in its own right and also serves to scrutinize the stability of the phases found against fluctuations. To that end we write the Hamiltonian (\[Hamiltonian\]) in terms of creation and annihilation operators employing the Holstein-Primakoff transformation [@HP]. For low temperatures an expansion up to bilinear terms can be used, leading to linear spin-wave theory. The Hamiltonian then is diagonalized with a Bogoliubov transformation from which we obtain the spin-wave frequencies. This investigation is restricted to the commensurate phases, because the infinite primitive cell of the incommensurate phases resists such an analysis.
[**Ferromagnetic phase**]{}
The magnon spectrum of the ferromagnetic phase is stable for the parameter region given in Sec. \[sec:gs\]. The rotational invariance of the spins around the chain axis leads to a Goldstone mode in the spectrum[@hummel]. [**Collinear phase**]{}
The collinear phase of Sec. \[sec:gs\] is also stable against fluctuations. There is no Goldstone mode due to the discrete degeneracy of the ground state.
[**120$^\circ$-structure**]{}
We argued in Sec. \[sec:gs\] that the 120$^\circ$-structure is unstable for infinitesimal dipolar energy due to a linear slope of the dipolar tensor at the ordering wave vector. However, we performed a spin-wave calculation based on a commensurate 120$^\circ$-structure, where the spins are located within the basal planes of the lattice[@oyedele]. It turns out that the spin-wave spectrum is stable for weak dipolar energies, from which we conclude that this commensurate ground state is a good approximation.
![\[figspectrum\]Magnon spectrum of the 120$^\circ$-structure perpendicular to the chain direction.](figure3.ps){width="39.00000%"}
The spin-wave frequencies resulting for the dipolar 120$^\circ$-structure for CsMnBr$_3$ are shown in Fig. \[figspectrum\], where we used $J=215$ GHz and $J'=0.41$ GHz respectively[@falk]. This leads to $\kappa' = 0.774$, i.e., this substance is in the IC II region of the phase diagram in Fig. \[fig2\]. Including the DDI, from the three Goldstone modes only one survives reflecting the unchanged rotational symmetry around the chain axis.
The spin-wave gaps at ${\bf q} = 0$ amount to $E_{0,1} = 198$ GHz, $E_{0,2} =
295$ GHz and $E_{0,3} = 410$ GHz for CsMnBr$_3$, which compares favorably with the experimental values [@falk; @gaulin]. Note that this calculation has no free parameter to fit, since the dipolar energy is determined by the lattice constants. We also calculated the spin-wave gaps for RbMnBr$_3$; neglecting crystal distortions we also obtain good agreement with the experiment[@heller].
Thus, we do not need any single-ion anisotropy to explain these results.
Summary
=======
We found three commensurate and two incommensurate phases for different values of the ratio of dipolar to interchain interaction due to the competing character of those two interactions. We showed via linear spin-wave theory that all commensurate phases are stable against fluctuations and that the incommensurate phase IC II can be approximated by a 120$^\circ$-structure for weak dipolar energies.
The spin-wave gaps of CsMnBr$_3$ and RbMnBr$_3$ are in good agreement with the experiment, which shows that the dipolar energy is the most important source of anisotropy in these Mn-compounds.
This work has been supported by the BMBF under contract number 03-SC5-TUM 0 and the DFG under contract number PI 337/1-2.
A.T. Skjeltorp and D. Sherrington, editors, [ *Dynamical Properties of Unconventional Magnetic Systems*]{}, (Kluwer Academic Publishers, Dordrecht, Boston and London, 1998).
H.T. Diep, editor, [*Magnetic Systems with Competing Interactions*]{}, World Scientific, Singapore (1994).
C. Pich and F. Schwabl, Z. Phys. B [**104**]{}, 165 (1997).
F.D.M. Haldane, Phys. Rev. Lett. [**50**]{}, 1153 (1983).
M. Steiner, M.J. Mikeska, Adv. Phys. [**40**]{}, 191 (1991).
M.F. Collins and O.A. Petrenko, Can. J. Phys. [**9**]{}, 75 (1997).
H. Kadowaki, K. Hirakawa and K. Ubukoshi, J. Phys. Soc. Jap. [**52**]{}, 1799 (1983).
R.E. Dietz, L.R. Walker, F.S.L. Hsu and W.H. Haemmerle, Sol. St. Comm. [**15**]{}, 1185 (1974).
P.P. Ewald, Ann. Phys. (Leipz.) [**[64]{}**]{}, 253 (1921).
L. Bonsall and A.A. Maradudin, Phys. Rev. B [**15**]{}, 1959 (1977).
J. Goodyear and D.J. Kennedy, Acta Cryst. B [**28**]{}, 1640 (1972).
H. Shiba, N. Suzuki, J. Phys. Soc. Jpn. [**51**]{}, 3488 (1982).
T. Holstein and H. Primakoff, Phys. Rev. [**[58]{}**]{}, 1098 (1940).
M. Hummel, C. Pich and F. Schwabl, to be published.
J.A. Oyedele and M.F. Collins, Can. J. Phys. [**56**]{}, 1482 (1978).
U. Falk, A. Furrer, H.U. Güdel and J.K. Kjems, Phys. Rev. B [**35**]{}, 4888 (1987).
B.D. Gaulin, M.F. Collins and W.J.L. Buyers, J. Appl. Phys. [**61**]{}, 3409 (1987).
L. Heller, M.F. Collins, Y.S. Yang and B. Collier, Phys. Rev. B [**49**]{}, 1104 (1994).
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.