1
$\begingroup$

Question 1

Given probability measures $\mu$ and $\nu$ on the same metric space $X=(X,d)$, and $\alpha \in [0, 1]$, is it always possible to find another probability measure $\lambda_\alpha$ on $X$ such that $W_1(\mu,\lambda_\alpha) \le \alpha W_1(\mu,\nu)$ and $W_1(\nu,\lambda_\alpha) \le (1 - \alpha) W_1(\mu,\nu)$ ?

Observations. Thus $\lambda_\alpha$ is some sort of interpolation between $\mu$ and $\nu$. Maybe the Wasserstein barycenter of $\mu$ and $\nu$ (with weights $\alpha$ and $1-\alpha$) or the Mccann interpolation can do the job ?

Question 2

In case the answer is negative, are there any other metrics (or just divergences like KL, etc.) between probability measures with such a property ?

Question 3

Same questions for the weaker condition: $\alpha W_1(\mu,\lambda_\alpha) + (1-\alpha)W_1(\nu,\lambda_\alpha) \le W_1(\mu,\nu)$.

$\endgroup$
  • 1
    $\begingroup$ The McCann interpolation satisfies the inequalities as equalities for $p$-Wasserstein distance with $p>1$. See Section 7.2 of the book Gradient flows by Ambrosio, Gigli and Savare. $\endgroup$ – O. Richard Aug 17 '18 at 20:34
  • $\begingroup$ Indeed. Thanks! If you post your comment as an answer, I'll be happy to up-vote and accept it. $\endgroup$ – dohmatob Aug 17 '18 at 20:58
5
$\begingroup$

The following discussion is based on the book Gradient Flows by Ambrosio, Gigli, and Savare (2008).

Consider $p$-Wasserstein distance with $p>1$ on a Hilbert space (for the sake of uniqueness). Let $\gamma$ be the optimal transport plan between $\mu$ and $\nu$ under the $p$-Wasserstein distance. Denote by $\pi^i$ be the $i$-th projection ($i=1,2$) and by $\#$ the pushforward of measures (see Section 5.2). Define $\lambda_\alpha = ((1-\alpha) \pi^1 + \alpha \pi^2)_\# \gamma$. Then according to Section 7.2, we have $W_p(\mu,\lambda_\alpha)=\alpha W_p(\mu,\nu)$, $W_p(\lambda_\alpha,\nu)=(1-\alpha)W_p(\mu,\nu)$ and $W_p(\mu,\lambda_\alpha) + W_p(\lambda_\alpha,\nu)=W_p(\mu,\nu)$.

| cite | improve this answer | |
$\endgroup$

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.