Multi-species optimal transportation (Q2302745)

The article focuses on optimal transport problem for vector-valued positive measures in Kantorovich's formulation. First, the problem's scalar formulation is extended to the vector case, yielding a linear primal problem. In this formulation, two probability distributions $\mu$ and $\nu$ of two measurable spaces $X$ resp. $Y$, are each decomposed as a sum of $n$ positive measures $\mu_1, \dotsc, \mu_n$ and $\nu_1, \dotsc, \nu_n$. Given a cost $\mathbf{c}$, the main question is which is the a transport plan $\gamma$ such that the transport cost from $\mu$ to $\nu$ is minimal. In the formulation it is possible to transfer any part of a component of $\mu$ to any component of $\nu$, with the constraints that each component of $\mu$ is fully transferred to $\nu$ and each component of $\nu$ is ``made up'' exclusively from contributions from $\mu$. The article shows several properties of the set of admissible transport plans and that, under certain conditions for the cost function, an optimal transport plan exists that satisfies the constraints. In a second step, a dual formulation of the problem is given. In this formulation, to each component of $\mu$ and $\nu$, a function is associated which can be interpreted as the benefit of having ``cleaned'' each component of $\mu$ and ``filled'' each component of $\nu$. These benefits are constrained so that the total benefit of moving a part from a component of $\mu$ to a component of $\nu$ cannot be greater than the associated component of $\mathbf{c}$. The article shows a weak duality result relating the dual and the primal problems. Additionally, it is shown that an optimum of the dual problem exists, and that the total optimal cost equals the total optimal benefit. Finally, necessary and sufficient optimality conditions are shown that link both problems. Finally, the article proposes an extension to the Wasserstein metric. To this end, it considers the case in which the $X$ is a Polish space and $Y=X$. The article shows that, if all the cost function's components are finite, symmetric and non-negative functions on $X\times X$; they satisfy a set of mixed triangular inequalities; and those located on the diagonal of $\mathbf{c}$ are metrics over $X$, then, for each $p \in [1, \infty]$, the $p$-th root of the solution of the associated transport problem is a metric over the space of vector-valued measures.