3
$\begingroup$

Let $\gamma_d = \gamma_1 \otimes \ldots \otimes \gamma_1$ be the standard Gaussian distribution on $\mathbb R^d$, where $d$ is a large positive integer. Given $\epsilon \ge 0$ and a measurable $A \subseteq \mathbb R^d$, let $A^\epsilon := \{x \in \mathbb R^d \mid \mbox{dist}(x,A) \le \epsilon\}$ be its epsilon-neighborhood, where $\mbox{dist}(x,A) := \inf_{a \in A} \|x-a\|$ is the distance of between $x$ and the closest point in $A$. Finally, let $0 < p \le q \le 1$.

Question. What is a good lower-bound for $r_d(\epsilon,p,q) := \underset{p \le \gamma_d(A) \le q}{\inf} \gamma(A^\epsilon)$ ?

Note. Understanding the special case $p \to 0^+$ would already be interesting. And of course, I'm perfectly fine with a bound which looks at different regimes of $\epsilon$, $p$, and $q$.

$\endgroup$

1 Answer 1

4
$\begingroup$

The answer is $\inf_{p \leq \gamma_{d}(A) \leq q} \gamma(A^{\varepsilon}) = \Phi(\Phi^{-1}(p)+\varepsilon)$ where $\Phi(x) = \int_{-\infty}^{x} \frac{e^{-s^{2}/2}}{\sqrt{2\pi}}ds$.

Indeed, all you need is the following claim:

For any measurable $A \subset \mathbb{R}^{d}$ and any $\varepsilon>0$ we have $\gamma_{d}(A^{\varepsilon})\geq \Phi(\Phi^{-1}(\gamma_{d}(A))+\varepsilon)$. The equality is attained if $A$ is any hafspace $H$ with $\gamma_{d}(A)=\gamma_{d}(H)$.

Since both $\Phi$ and $\Phi^{-1}$ are increasing the answer to your questions follows from the claim.

The claim follows from the Ehrhard inequality: for any measurable sets $U, V \subset \mathbb{R}^{d}$ and any $\alpha, \beta \geq 0$ such that $\alpha+\beta \geq 0$, $|\alpha-\beta|\leq 1$, $\alpha U+\beta V$ is measurable (here "+" is Minkowski sum) we have $$ \gamma_{d}(\alpha U +\beta V) \geq \Phi(\alpha \Phi^{-1}(\gamma_{d}(U))+\beta\Phi^{-1}(\gamma_{d}(V))). $$

So, let $B$ be the unit ball, and $\lambda \in (0,1)$. Then by Ehrhard we have $$ \gamma_{d}(A^{\varepsilon}) = \gamma_{d}(\lambda [\lambda^{-1}A]+(1-\lambda)[(1-\lambda)^{-1}\varepsilon B]) \geq \Phi(\lambda \Phi^{-1}(\gamma_{d}(\lambda^{-1}A))+(1-\lambda)\Phi^{-1}(\gamma_{d}((1-\lambda)^{-1} \varepsilon B))) $$ Next, let $\lambda \to 1$, $\lambda<1$. Then $\lambda \Phi^{-1}(\gamma_{d}(\lambda^{-1}A)) \to \Phi^{-1}(\gamma_{d}(A))$, and also $(1-\lambda)\Phi^{-1}(\gamma_{d}((1-\lambda)^{-1} \varepsilon B)) \to \varepsilon$. The first limit is simple. To verify the second limit we need to show that $\lim_{r \to \infty}\frac{1}{r} \Phi^{-1}(\gamma_{d}(r B))=1$. Perhaps there are many ways to show this. The most straightforward one will be to honestly compute all asymptotic:

$$ \gamma_{d}(r B) = 1- \frac{\sigma_{d}}{(2\pi)^{d/2}}\int_{r}^{\infty}e^{-r^{2}/2}r^{d-1}dr \stackrel{r \to \infty}{=}1-\frac{\sigma_{d}}{(2\pi)^{d/2}} e^{-r^{2}/2} r^{d-2}+o(e^{-r^{2}/2} r^{d-2}), $$

where $\sigma_{d}$ is the surface area measure of the unit sphere. Thus

$$ \Phi^{-1}(s) \stackrel{s \to 1^{-}}{=} (-2\ln(1-s))^{1/2}+o((-\ln(1-s))^{1/2}), $$ therefore $$ \lim_{r \to \infty} \frac{1}{r} \Phi^{-1}(\gamma_{d}(r B)) = \lim_{r \to \infty} \frac{1}{r} [-2\ln(r^{d-2}e^{-r^{2}/2})]^{1/2}=1. $$

$\endgroup$
1
  • $\begingroup$ Indeed, it follows directly from Gaussian isoperimetric inequality. Thanks! $\endgroup$
    – dohmatob
    Dec 2, 2020 at 15:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.