The multistochastic Monge-Kantorovich problem (Q2235808)

The authors study a multistochastic Monge-Kantorovich problem on the product space $X = \prod_{i=1}^n X_i$, which is a generalization of the multimarginal Monge-Kantorovich problem. More precisely, given $1\leq k < n$, the authors consider the minimization problem \[ \inf_\pi\int c\ \mathrm{d}\pi,\tag{1} \] where $c$ is a given cost function, and $\pi$ has given marginals on any product space of the form $X_{i_1}\times\dots\times X_{i_k}$, $\{i_1,\dots, i_k\}\subseteq \{1,\dots, n\}$. That is, the marginal of $\pi$ is prescribed on the product of $k$ spaces, rather than on single spaces $X_i$, as it was the case for the classic multimarginal Monge-Kantorovich problem. The authors study basic properties of the multistochastic problem, including well-posedness, existence of a dual solution, boundedness and continuity of a dual solution. Under the assumption that the cost function $c$ is lower semicontinuous, the authors show that, as long as the set of admissible uniting measures $\pi$ is non empty, then (1) admits a solution. However, as the multistochastic problem is overdetermined, such $\pi$ does not always exist, and finding conditions on $k, n$ under which uniting measures exist is tricky. The authors prove that, for given $1\leq k < n$, there exists $\lambda_{kn}$ such that the following property holds: given a family of consistent measures $\mu_\alpha$, and another family $\nu_i$, such that each $\mu_\alpha\ll \nu_\alpha =\prod_{i\in\alpha}\nu_i$, and the Radon-Nikodym densities $\frac{\mathrm{d}\mu_\alpha}{\mathrm{d}\nu_\alpha}$ are all bounded from above and away from zero, then the set of uniting measures is non empty. As an example, the authors then show that, when considering the Lebesgue measures on $[0, 1]^2$, with cost function $c(x, y, z) = xyz$, then there is a unique solution to (1), and such uniting measure is concentrated on the Sierpinski tetrahedron.