2
$\begingroup$

Suppose we have nonindependent random variables $X \sim P$ and $Y \sim Q$, where $P$ and $Q$ denote their marginal distributions. We are interested in upper bounding $$ |\mathbf{E}_{X, Y\sim P \otimes Q } f(X, Y) - \mathbf{E}_{X, Y\sim P, Q } f(X, Y)| $$ for a given function $f$. Here $P \otimes Q$ denotes the product distribution of $P$ and $Q$, while $P, Q$ denotes their joint distribution (or equivalently a coupling).

Question: Can we upper bound such a mean difference using the Wasserstein distance between $P$ and $Q$, or the Wasserstein distance between $P \otimes Q$ and $P, Q$?

If needed, we are allowed to impose regularity conditions, such as Lipschitz continuity, on $f$.

$\endgroup$
1
$\begingroup$

By the dual representation of the Wasserstein metric $W_1$, $$ |\mathbf{E}_{X, Y\sim P \otimes Q } f(X, Y) - \mathbf{E}_{X, Y\sim P, Q } f(X, Y)|\le\text{Lip}(f)\,W_1(P \otimes Q,(P, Q)), $$ where $\text{Lip}(f)$ is the Lipschitz constant of $f$.

On the other hand, the closeness of $P$ to $Q$ in any metric will not by itself provide any interesting upper bound on $|\mathbf{E}_{X, Y\sim P \otimes Q } f(X, Y) - \mathbf{E}_{X, Y\sim P, Q } f(X, Y)|$; e.g., take $P=Q$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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