Optimality of Kantorovich potentials for the squared distance

102 Views Asked by At

This question comes from Villani's book, Optimal Transport: Old and New. Consider the cost function $c(x, y) = |x - y|^2$ on $X \times Y$, where $X$ is the right half of the unit ball, and $Y$ is the right half of the unit ball, shifted one unit to the right. Let $\mu$ be the uniform distribution on $X$ and $\nu$ the uniform distribution on $Y$.

The optimal transport map $T$ is given by $(x, y) \mapsto (x + 1, y)$, and by a theorem from chapter 10 we know that $T(x, y) = (x,y) + \nabla \psi(x, y)$ for some $c$-convex function $\psi$ on $X$. Since this means that $\psi(x, y) = (1, 0)$ we have $\psi(x, y) = x$.

My question: Is this $\psi$ equal to the optimal $\psi$ in the Kantorovich dual problem? That is, is the following inequality true: $$ \int_X c(x, Tx)\, d\mu(x) = \int_Y \psi^c \, d\nu - \int_X \psi\, d\mu, $$ where $\psi^c(y) := \inf_x \psi(x) + c(x, y)$. I've tried computing the $c$-transform and as far as I can tell the answer to my question is negative, since I am getting that $\partial_c \psi(x)$ is not the translate by 1, which it should be to guarantee optimality by Kantorovich duality.

Any help is much appreciated!

1

There are 1 best solutions below

0
On

I've since figured out the answer, which is somewhat annoyingly trivial. The transport map is given by $x + \frac{1}{2}\nabla \psi(x)$, not by $x + \nabla \psi (x)$; the latter occurs for $c = d^2/2$. Then we get in fact that $\psi(x) = 2|x|$, which can also be checked is optimal for the dual problem.