A characterization for solutions of the Monge-Kantorovich mass transport problem (Q303608)

The Kantorovich problem is a relaxed version of the Monge optimal transport problem as a linear optimization problem on a convex domain. Consider two Polish probability spaces \((X,\mu)\) and \((Y,\nu)\) and a continuous function \(c:X\times Y \to \mathbb{R}\). Denote by \(\Pi(\mu,\nu)\) the set of Borel probability measures on \(X\times Y\) which have \(\mu\) and \(\nu\) as marginals. The transport cost associated to a transport plan \(\pi \in \Pi(\mu, \nu)\) is given by \(l_c(\pi) = \int_{X\times Y}c(x,y) d\pi\). The Kantorovich problem (MK) is to minimize \(\inf\{l_c(\pi);\pi \in \Pi(\mu, \nu)\}\). The Kantorovich problem always admits solutions as soon as the cost function \(c\) is a non-negative lower semi-continuous function. Generalized twist condition: Assume that \(x\mapsto c(x,y)\) is differentiable for all \(y\in Y\). It satisfies the generalized twist condition if for any \(x_0 \in X\) and \(y_0\in Y\) the set \(\{y: \partial c(x_0,y)/\partial x= \partial c(x_0,y_0)/\partial x\}\) is a finite subset of \(Y\). \(m\)-twist condition: Let \(m\in \mathbb{N}\). The function \(c\) satisfies the \(m\)-twist condition if for any \(x_0\in X\) and \(y_0\in Y\) the cardinality of the set \(\{y: \partial (c(x_0,y)/\partial x= \partial c(x_0,y_0)\}\) is at most \(m\). The function \(c\) satisfies the \(m\)-twist condition locally if for any \(x_0\in X\) and \(y_0\in Y\) the above cardinality condition is valid only for \(y\in U\), where \(U\) is a neighborhood of \(y_0\). The main result of the paper is the Theorem 1.3: Let \(X\) be a complete separable Riemannian manifold and \(Y\) be a Polish space equipped with Borel probability measures \(\mu\) on \(X\) and \(\nu\) on \(Y\). Let \(c:X\times Y \to \mathbb{R}\) be a bounded continuous cost function and assume that 1) the cost function \(c\) satisfies the \(m\)-twist condition, 2) \(\mu\) is non-atomic and any \(c\)-concave function on \(X\) is differentiable \(\mu\)-almost surely on its domain. Then, for each optimal plan \(\gamma\) of \((MK)\), there exist \(k\in \{1,\dots m\}\), a sequence \(\{\alpha_i\}_{i=1}^k\) of non-negative functions from \(X\) to \([0,1]\) and Borel measurable maps \(G_1,\dots ,G_k\) from \(X\) to \(Y\) such that \(\gamma = \sum_{i=1}^k\alpha_i(Id\times G_i)_{\sharp\mu}\) (i.\,e. \(\gamma(S) =\sum_{i=1}^k\int_X\alpha_i(x)\chi_S(x, T_ix)d\mu\), where \(S\) is a measurable set \(S\subset X\times Y\) and \(\chi_S\) is the indicator function of the set \(S\)). The relation \(\sum_{i=1}^k\alpha_i(x) =1\) is valid for \(\mu\)-almost every \(x\in X\). Moreover, if \(G_i(x) =G_j(x)\) for some \(x\in X \), then \(\alpha_i(x)=\alpha_j(x)\). In the final section, the author studies the uniqueness issue for \(m\)-twisted cost functions.