Eigenvalue distribution of optimal transportation (Q2344060)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Eigenvalue distribution of optimal transportation |
scientific article |
Statements
Eigenvalue distribution of optimal transportation (English)
0 references
12 May 2015
0 references
Let \(\mu\) and \(\nu\) be two absolutely continuous probability measures in \({\mathbb R}^n\). It was discovered by \textit{Y. Brenier} [Commun. Pure Appl. Math. 44, No. 4, 375--417 (1991; Zbl 0738.46011)] and \textit{R. J. McCann} [Duke Math. J. 80, No. 2, 309--323 (1995; Zbl 0873.28009)] that there exists a convex function \(\Phi\) in \({\mathbb R}^n\) such that \(\int_{{\mathbb R}^n}b(\nabla\Phi(x))\, d\mu(x) = \int_{{\mathbb R}^n}b(x)\, d\nu(x)\) for any \(\nu\)-integrable function \(b: {\mathbb R}^n\to{\mathbb R}\). Moreover, the Brenier map \(x\mapsto \nabla\Phi(x)\) is uniquely determined \(\mu\)-almost everywhere. In this paper, the authors consider the case where \(\mu\) and \(\nu\) are log-concave probability measures. An absolutely continuous probability measure in \({\mathbb R}^n\) is called log-concave if it has a density \(\rho\) which satisfies \(\rho(\lambda x + (1-\lambda) y) \geq \rho(x)^\lambda \rho(y)^{1-\lambda}\) for any \(x, y\in {\mathbb R}^n\) and \(\lambda\in(0,1)\). The uniform measure on any convex domain is log-concave, because it is the Gaussian measure. Let \(\mathrm{Supp}(\mu)\) be the interior of the support of \(\mu\), which is an open, convex set in \({\mathbb R}^n\). It follows from work by \textit{L. A. Caffarelli} [Ann. Math. (2) 131, No. 1, 129--134 (1990; Zbl 0704.35045); J. Am. Math. Soc. 5, No. 1, 99--104 (1992; Zbl 0753.35031)] that the function \(\Phi\) is \(C^2\)-smooth in \(\mathrm{Supp}(\mu)\) when each of the measures \(\mu\) and \(\nu\) satisfies the following regularity condition: either the support of the measure is the entire \({\mathbb R}^n\) or else the support is a bounded, convex domain and the density of the measure is bounded away from zero and from infinity in this convex domain. As it turns out, the positive-definite Hessian matrix \(D^2\Phi(x)\) exhibits remarkable regularity in the behavior of its eigenvalues. The results of the paper are the following two assertions. Theorem 1. Let \(\mu\) and \(\nu\) be absolutely continuous, log-concave probability measures in \({\mathbb R}^n\). Let \(\nabla\Phi\) be the Brenier map between \(\mu\) and \(\nu\) and \(\Phi\) satisfies the regularity condition. Write \(0< \lambda_1(x)\leq\cdots \leq \lambda_n(x)\) for the eigenvalues of the matrix \(D^2\Phi(x)\), repeated according to their multiplicity. Let \(X\) be a random variable in \(\mathbb R^n\) that is distributed according to \(\mu\). Then, for \(i=1,\dots, n\), we have \(\mathrm{Var}[\log\lambda_i(X)]\leq 4\), where \(\mathrm{Var}\) stands for variance. Theorem 2. Under the notation and assumptions of Theorem 1, fix \(v\in\mathbb R^n\) and let \(Y=\log(D^2\Phi(X) v\cdot v)\), where \(\log\) stands for the natural logarithm and \(x\cdot y\) stands for the standard scalar product of \(x, y\in\mathbb R^n\). Then, the following hold true. {\parindent=0.5cm\begin{itemize}\item[1)] \(\mathrm{Var}[Y]\leq 4\);\item[2)] for any locally Lipschitz function \(f:\mathbb R \to\mathbb R\) with \(E| f(Y)|<\infty\), we have \(\mathrm{Var} [f(Y)]\leq 4 E| f'|^2(Y)\), where \(|\cdot|\) stands for the standard Euclidean norm in \(\mathbb R^n\); \item[3)] for any 1-Lipschitz function \(f: \mathbb R \to \mathbb R\), denoting \(A=Ef(Y)\), we have that \(A\in\mathbb R\) and \(E\exp\{c| f(Y) - A|\}\leq 2\), where \(c\) is a constant. \end{itemize}} The paper is organized as follows. Section 2 recalls some well-known constructions related to positive-definite matrices. In Sections 3 and 4, the authors prove the main results under regularity assumptions by employing the Bakry-Émery \(\Gamma_2\)-calculus. Section 5 is devoted to the elimination of these regularity assumptions. In Section 6, the authors complete the proofs of Theorems 1 and 2.
0 references
log-concave measure
0 references
optimal transportation
0 references
eigenvalue distribution
0 references