Is this decomposition of a measure into the sum of other measures on each cube correct?

48 Views Asked by At

I'm reading the paper Convergence of Entropic Schemes for Optimal Transport and Gradient Flows in which I suspect one argument in Lemma 2.11 is wrong. First, I start with the notation used in the paper.

  • $\mathcal X = \mathbb R^d$ and $c$ is the Euclidian distance on $\mathcal X$.

  • $\mathcal P_2(\mathcal X)$ be the set of Borel probability measures on $\mathcal X$ with finite second moment.

  • $\mu,\nu \in \mathcal P_2(\mathcal X)$.

  • $\Pi(\mu,\nu)$ is the set of all probability measures on $\mathcal X^2$ with marginals $\mu$ on the first $\mathcal X$ and $\nu$ on the second $\mathcal X$.

  • We write $(c, \gamma)$ for $\int_{\mathcal X^2} c(x,y) \, \mathrm d \gamma(x,y)$ and define $W^2 (\mu, \nu) = \inf_{\gamma \in \Pi(\mu,\nu)} (c, \gamma)$.

Then I state the Lemma 2.11.

Let $(Q_i)_{i \in I}$ be a countable partition of $\mathcal X$ into Borel sets with $\sup _{i \in I} \operatorname{diam}\left(Q_{i}\right) \leq C< +\infty$, i.e., $\|x-y\|^{2} \leq C^{2}$ for any $x, y \in Q_{i}$ for any $i \in I$. Let $\mu,\nu \in \mathcal P_2 (\mathcal X)$ such that $\mu(Q_i) = \nu(Q_i)$ for all $i \in I$. Then $W^{2}(\mu, \nu) \leq C^{2}$.

Below is the proof provided in the paper:


Denote by $\hat I$ the subset of $I$ such that $\mu\left(Q_{i}\right) = \nu\left(Q_{i}\right)>0$ for $i \in \hat {I}$. For $i \in \hat I$ and every Borel $\sigma \subset \mathcal X$ let

$$ \mu_{i}(\sigma)=\frac{\mu\left(\sigma \cap Q_{i}\right)}{\mu\left(Q_{i}\right)} $$

and analogously define $\nu_{i} .$ Clearly all $\mu_{i}, \nu_{i} \in \mathcal{P}_{2}\left(\mathcal X\right)$, with support contained in $\overline{Q_{i}}$. For every $i \in \hat{I}$ let $\gamma_{i} \in \Pi\left(\mu_{i}, \nu_{i}\right)$ and have spt $\gamma_{i} \subset \overline{Q_{i}}^{2}$ and thus

$$ \left(c, \gamma_{i}\right)=\int_{\mathcal X^2}\|x-y\|^{2} \, \mathrm d \gamma_{i}(x, y)=\int_{Q_{i} \times Q_{i}}\|x-y\|^{2} \, \mathrm d \gamma_{i}(x, y) \leq C^{2} $$

One finds $\color{blue}{\gamma=\sum_{i \in \hat{I}} \mu\left(Q_{i}\right) \gamma_{i}} \in \Pi(\mu, \nu)$ and consequently

$$ W^{2}(\mu, \nu) \leq(c, \gamma)=\sum_{i \in \hat{I}} \mu\left(Q_{i}\right)\left(c, \gamma_{i}\right) \leq \sum_{i \in \hat{I}} \mu\left(Q_{i}\right) C^{2}=C^{2}. $$


The measure $\gamma \in \Pi(\mu, \nu)$ is defined on the whole $\mathcal X^2$, but $\operatorname{supp}(\gamma_i) \subseteq \overline{Q_{i}}^{2}$ and $\bigcup_i \overline{Q_{i}}^{2} \subsetneq \mathcal X^2$. This is why I suspect $\gamma=\sum_{i \in \hat{I}} \mu\left(Q_{i}\right) \gamma_{i}$ is not correct.

Could you please elaborate on this point?

1

There are 1 best solutions below

0
On BEST ANSWER

I misunderstood the argument by the authors. Actually, we need to verify that such $\gamma$ with that particular form belongs to $\Pi(\mu, \nu)$. And for sure $W^2 (\mu, \nu)$ is not greater than the cost induced by this particular subset of $\gamma$'s.