Solution of the Monge-Ampère equation on Wiener space for general log-concave measures (Q2368784)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Solution of the Monge-Ampère equation on Wiener space for general log-concave measures |
scientific article |
Statements
Solution of the Monge-Ampère equation on Wiener space for general log-concave measures (English)
0 references
28 April 2006
0 references
Let \(\rho\) and \(\nu\) be two probability measures on a separable Fréchet space \(W\) and let \(\Sigma(\rho,\nu)\) be the convex subset of the probability measures on the product space \(W\times W\) whose first marginal is \(\rho\) and the second one is \(\nu\). The corresponding Cameron-Martin space is denoted by \(H\). The Monge-Kantorovitch problem (MKP) for the couple \((\rho,\nu)\) consists of finding a measure \(\gamma\in\Sigma(\rho,\nu)\) which re\-a\-li\-zes the following infimum: \(\inf_{\beta\in\Sigma(\rho,\nu)}\{\int_{W\times W} | x-y| _H^2\,d\beta(x,y)\}.\) The function \(c(x, y) =| x-y| _H^2\) is called the cost function. The authors give a generalization of the polar factorization of vector fields which map a probability measure on \(W\) to another one such that one of them is spread and the two measures are at finite Wasserstein distance from each other. This is done without any absolute continuity hypothesis. As an example the authors treat in detail the case of the infinite-dimensional Gaussian measures: there are constructed two Gaussian measures which are at finite Wasserstein distance from each other; hence, one can be transported onto the other one with an explicitly constructed transport map although they are mutually singular. The main difficulty in this infinite-dimensional case stems from the lack of regularity of the transport potentials; in fact, we only know that these functions are in the Gaussian Sobolev space \(\mathbb D_{2,}\), i.e., they have only first-order Sobolev derivatives. Combining the finite-dimensional results of Caffarelli with Wiener space analysis, the authors solve MKP completely when the target measure is \(H\)-log-concave. More precisely, the authors show that the transport potential has a second-order derivative as a Hilbert-Schmidt operator valued map; hence, write the corresponding Jacobian which includes the modified Carleman-Fredholm determinant, and finally prove that the transport potential is the unique 1-convex strong solution of the Monge-Ampère equation. It is shown that all difficulties disappear if we use the natural Itô calculus and calculate the Itô Jacobian using the natural Brownian motion which is associated with the solution of the Monge problem. In fact, with Itô parametrization, the complications are absorbed by the filtrations of forward and backward transport processes (i.e., maps). The authors also give the delicate relations between the polar factorization of the absolutely continuous transformations of the Wiener measure and the Brownian motions which appear in the semimartingale decomposition of the transport process with respect to its natural filtration.
0 references
optimal mass transportation
0 references
Monge-Kan\-to\-ro\-vitch problem
0 references
Wiener space
0 references
Gaussian measures
0 references
\(H\)-log-concave densities
0 references
Îto calculus
0 references
Itô-solution
0 references
0 references
0 references