For Lebesgue-absolutely continuous probability measures $\rho\ll \mathcal{L}^d$ in the whole space $\mathbb{R}^d$ with finite second moments (i-e $\rho\in \mathcal{P}^2_{ac}(\mathbb{R}^d)$), let $$ \mathcal{H}(\rho)=\int_{\mathbb{R}^d}\rho\log \rho \,dx $$ be the Boltzmann entropy.

Question: Is $\mathcal{H}$ continuous with respect to the quadratic Wasserstein distance $\mathcal{W}_2$? i-e is it true that $$ \mathcal{W}_2(\rho_n,\rho)\to 0 \quad\Rightarrow\quad \mathcal{H}(\rho_n)\to\mathcal{H}(\rho)\qquad ??? $$

I guess this should be known and written down somewhere. I looked at the classical references (Villani's books [topics...] and [old and new...], also in [Ambrosio-Gigli-Savaré]) but surprisingly enough I couldn't find the precise statement.

It is well known that $\mathcal{H}$ is $0$-displacement convex (in the sense of McCann). By analogy with the Euclidean setting, where any convex function $\phi:\mathbb{R}^d\to \mathbb{R}$ is continuous, I suspect that any displacement convex functional $\mathcal{F}:\mathcal{P}_2(\mathbb{R}^d)\to\mathbb{R}$ should be continuous with respect to the wasserstein distance $\mathcal{W}_2$. But maybe not...

This question arised in the context of my research: I am trying to construct weak solutions to some system of PDEs by means of the Jordan-Kinderlehrer-Otto/DeGiorgi's minimizing scheme, and I need the above statement in some technical step. Unfortunately the mere lower semi-continuity would not be enough for my purpose, I really need full continuity.


Let $A_n=\bigcup_{i=0}^{n-1}[2i/(2n),(2i+1)/2n)$, $\rho_n=2\chi_{A_n}$, $A=[0,1]$ and $\rho=\chi_{[0,1]}$. Then it seems pretty clear that $\mathcal W_2(\rho_n,\rho)\to 0$. On the other hand $H(\rho_n)\not\to H(\rho)$, right?

  • $\begingroup$ yes you're perfectly right, very nice counterexample. Yet another naive belief goes to the bin... $\endgroup$ – leo monsaingeon Jun 26 '15 at 9:04
  • $\begingroup$ what about lower semi-continuity then? after thinking on it for a while this should be enough for my purpose, contrarily to what I said in my post $\endgroup$ – leo monsaingeon Jun 26 '15 at 10:00
  • $\begingroup$ Yes, it is l.s.c; You have that $\rho_n\to \rho$ weakly (as probability measures). Since $\rho_n\to H(\rho_n|\rho)$ is l.s.c., you are done. $\endgroup$ – ofer zeitouni Jun 26 '15 at 15:48
  • $\begingroup$ Well, sure, if the relative H is lsc then of course the original one is lsc too. But do you have a reference? $\endgroup$ – leo monsaingeon Jun 26 '15 at 20:02
  • $\begingroup$ The relative entropy is the rate function in Sanov's theorem of large deviations, and thus is by definition l.s.c. The proof is not hard - H can be written as a Legendre transform, that is a supremum over continuous functions. Any book on large deviations (in particular, Deuschel-Stroock, Varadhan's lecture notes, Dembo-Zeitouni) contains a proof of these facts. $\endgroup$ – ofer zeitouni Jun 26 '15 at 21:05

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.