7
$\begingroup$

Let $(X,d)$ be a complete separable metric space, and let $(\mathcal{P}_2 (X), W_2)$ be the space of probability measures on $X$ with finite second moments, equipped with the 2-Wasserstein distance. It is known that discrete measures are dense inside $(\mathcal{P}_2 (X), W_2)$ - namely, given any $\mu \in \mathcal{P}_2 (X)$, and $\delta>0$, one can find a discrete measure $\mu_\delta$ with $W_2 (\mu, \mu_\delta)<\delta$.

Now, let $\mu_0, \mu_1 \in \mathcal{P}_2 (X)$, and let $\mu_t$ be a $W_2$ geodesic connecting $\mu_0$ and $\mu_1$ (a.k.a. $\mu_t$ is a [not necessarily unique] displacement interpolation between $\mu_0$ and $\mu_1$). Is the displacement interpolation stable under discrete approximation? That is, can one pick discrete $\mu_{0,n}, \mu_{1,n}$ such that $\mu_{t,n}$ is close to $\mu_t$ for all $t\in[0,1]$?

$\endgroup$

1 Answer 1

3
$\begingroup$

The displacement interpolation $\mu_t$ should not be fixed a priori, due to nonuniqueness of Wasserstein Geodesics. Thus, the correct question should be: fix the approximating sequences $(\mu_{0,n}),(\mu_{1,n})$ and $W_2$ geodesics $\mu_{t,n}$, and ask if there exists one $\mu_t$ close to $\mu_{t,n}$ for $t \in [0,1]$.

$\endgroup$
6
  • 1
    $\begingroup$ This is a comment rather than an answer, but I could not post it as a comment. Anyway, something useful in this direction can be found in Lemma 4.4 arxiv.org/pdf/1609.00782.pdf which, combined with Proposition 4.8 of arxiv.org/pdf/1311.4907.pdf gives you $W_2$ close $\mu_{t,n}$'s. $\endgroup$ Aug 19, 2020 at 8:10
  • 1
    $\begingroup$ Certainly there are non-uniqueness issues. I actually meant something like: given a Wasserstein geodesic $\mu_t$, can we produce sequences $(\mu_{0,n})$ and $(\mu_{1,n})$ such that $(\mu_{t,n})$ converges to $\mu_t$ in some suitable sense. $\endgroup$ Aug 19, 2020 at 8:35
  • $\begingroup$ These links do look quite helpful, thank you! $\endgroup$ Aug 19, 2020 at 8:36
  • $\begingroup$ I think is very hard to fix $\mu_t$ and produce, afterwards, approximating marginals $(\mu_{0,n}),(\mu_{1,n})$. If you stick to the other way, as in my answer, you can try to argue essentially by tightness to get $\mu_t$ from $\mu_{t,n}$. Finally, you just need to show that the limit $\mu_t$ is a $W_2$ geodesic. The first link I gave you follows this path. With this approach, is very hard to control whether you are converging to your fixed a priori Wasserstein geodesic, or to another. $\endgroup$ Aug 19, 2020 at 8:43
  • 1
    $\begingroup$ There is not only the problem of uniqueness, but also of existence. Under the present assumptions, $\mu_{0,n}$ and $\mu_{1,n}$ may not be connected by a $W_2$-geodesic. $\endgroup$
    – MaoWao
    Aug 19, 2020 at 9:21

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.