I'm trying to make sense of the Wasserstein metric, it says in wikipedia

enter image description here

Say the marginal distributions $\mu$ and $\nu$ are given, what properties does a joint distribution from the two need to satisfy? For example like $\mu(x)=\int p(x,y)dx$ and $\nu(y)=\int p(x,y)dy$?


1 Answer 1


The joint distribution just needs to be a valid probability distribution, with matching marginals. That's it.

It turns out that you can characterize the set of such distributions via copulas. That is, if we assume for now that $X$ and $Y$ are each real-valued, every such distribution has cdf of the form $$ \Pr(X \le x, Y \le y) = C(\Pr(X \le x), \Pr(Y \le y)) $$ where $C$ is the joint cdf of a distribution that has uniform margins on the unit square. The converse also holds: any such $C$ gives you a valid joint distribution. I'm pretty sure something analogous holds for $X$ and $Y$ taking values in $\mathbb R^n$, but apparently it only holds in one direction for infinite-dimensional Hilbert spaces.

That doesn't necessarily help you in computing the Wasserstein distance, but it might help in thinking about what it means.

  • 1
    $\begingroup$ Ohhh, cool paper! $\endgroup$ Commented Feb 13, 2017 at 15:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.