Sampling of probability measures in the convex order by Wasserstein projection

Abstract: In this paper, for $\mu$ and $\nu$ two probability measures on $\mathbb{R}^d$ with finite moments of order $\rho\ge 1$, we define the respective projections for the $W_\rho$-Wasserstein distance of $\mu$ and $\nu$ on the sets of probability measures dominated by $\nu$ and of probability measures larger than $\mu$ in the convex order. The $W_2$-projection of $\mu$ can be easily computed when $\mu$ and $\nu$ have finite support by solving a quadratic optimization problem with linear constraints. In dimension $d=1$, Gozlan et al.~(2018) have shown that the projections do not depend on $\rho$. We explicit their quantile functions in terms of those of $\mu$ and $\nu$. The motivation is the design of sampling techniques preserving the convex order in order to approximate Martingale Optimal Transport problems by using linear programming solvers. We prove convergence of the Wasserstein projection based sampling methods as the sample sizes tend to infinity and illustrate them by numerical experiments.