1
$\begingroup$

Let $f$ be a smooth probability distribution on the unit square $S$ such that $f(x)>0$ on $S$. Let $\{g_i\}$ be a sequence of smooth probability distributions such that $g_i(x)>0$ on $S$ as well and such that the Wasserstein distance $W_1(f,g_i)$ approaches zero. Is it true that $$\iint_S \sqrt{f(x)} dx \geq \limsup_{i\to\infty} \iint_S \sqrt{g_i(x)}dx$$ assuming both quantities exist? The other direction of the inequality is definitely not true because we could construct $\{g_i\}$ that is almost discrete, which would have very tiny values of $\iint_S \sqrt{g_i(x)}dx$.

$\endgroup$

1 Answer 1

1
$\begingroup$

Yes. Also, both quantities are guaranteed to exist: the left hand side because $f$ is smooth and bounded, and the right hand side because a lim sup always exists, and it's finite because the LHS is finite and the inequality is true.

Here is a sketch for the proof of (a slightly generalized version of) the inequality: suppose $f: \mathbb{R}^d \to \mathbb{R}$ is a uniformly continuous density and let $\delta$ be its modulus of continuity. For $\epsilon > 0$, let $$ \eta_\epsilon(x) = \frac{1}{|B_\epsilon|} \chi_{B_\epsilon}(x), $$ where $B_\epsilon$ is the Euclidean ball of radius $\epsilon$ in $\mathbb{R}^d$.

Lemma If $\|f - \eta_\epsilon * g\|_\infty \ge 4 \delta(\epsilon)$ then $W_1(f, g) \ge C_d \delta(\epsilon)^2 \epsilon^{d+1}$, where $C_d$ is a constant depending only on $d$.

Proof (sketch): Choose a point $x$ such that $|f(x) - (\eta_\epsilon * g)(x)| \ge 3 \delta(\epsilon)$. Suppose that $f(x) \ge (\eta_\epsilon * g)(x) + 3 \delta$ (the other direction is similar). Consider a ball $B$ of radius $\epsilon$ around $x$: by the definition of $\delta$, $f(y) \ge f(x) - \delta \ge (\eta_\epsilon * g)(x) + 2 \delta$ for all $y \in B$. On the other hand, $(\eta_\epsilon * g)(x)$ is the average value of $g$ on $B$. Using this, you can find $\epsilon' < \epsilon$ so that $f$ has $\delta$ more mass on $B(x, \epsilon')$ than $g$ has on $B(x, \epsilon)$. To move $g$ onto $f$, you will have to move at least $\delta$ mass from outside of $B(x, \epsilon)$ into $B(x, \epsilon')$, which costs at least $\delta (\epsilon - \epsilon')$ in $W_1$ distance. Then you need to compute how small $\epsilon'$ can be: you'll get $\epsilon - \epsilon' \asymp \delta \epsilon^{d+1}$.

Since $W_1(f, g_i) \to 0$, the lemma implies that $$ \lim_{\epsilon \to 0} \lim_{i \to \infty} \|f - \eta_\epsilon * g_i\|_\infty \to 0. $$ Since $f$ and $\eta_\epsilon * g_i$ are both bounded, for every continuous $\phi: \mathbb{R} \to \mathbb{R}$ we have $$ \lim_{\epsilon \to 0} \lim_{i \to \infty} \int \phi(\eta_\epsilon * g) = \int \phi(f). $$ Finally, if $\phi$ is a convex function (like the square root) then convolution with $\eta_\epsilon$ is an averaging operation and so Jensen's inequality implies $\phi(\eta_\epsilon * g) \ge \eta_\epsilon * \phi(g)$. So $$ \int \phi(\eta_\epsilon * g_i) \ge \int \eta_\epsilon * \phi(g_i) = \int \phi(g_i) $$ for every $\epsilon$ and every $i$.

$\endgroup$

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.