7
$\begingroup$

Let $\mathbb{R}^n$ be the Euclidean space and $A \subseteq \mathbb{R}^n$ be a sufficiently regular set, e.g., one that has smooth boundary or is convex. We define the $\epsilon$-neighbor of $A$ in the $\ell_2$ sense as \begin{align} A^{\epsilon} = \{ y \in \mathbb{R}^n \colon \text{there exists}~ x \in A ~\text{such that}~ \| x - y \|_{2} \leq \epsilon \}. \end{align} We denote $\gamma$ as the standard $n$-dimensional Gaussian measure $N(0, I_n)$, then the Gaussian surface measure is defined as \begin{align} \tau (A) = \lim_{\epsilon \rightarrow 0} \frac{ \gamma( A^{\epsilon} \setminus A) }{ \epsilon}. \end{align} By results of Keith Ball (the reverse isoperimetric problem for Gaussian measure https://link.springer.com/article/10.1007/BF02573986), there is a universal upper bound of the Gaussian surface area of any convex set.

My questions are: 1) when is Keith Ball's upper bound tight? 2) what is a tight upper bound of the Gaussian surface area of the cone of positive semidefinite matrices?

$\endgroup$
  • 3
    $\begingroup$ What do you mean by the "positive semidefinite cone"? The cone of positive semidefinite matrices? An orthant? For any cone $A$ with apex at the origin, the Gaussian perimeter is proportional to the $n-2$-dimensional measure of $A \cap \{x : |x|=1\}$, the constant of proportionality being $(2\pi)^{-n/2} \int_0^\infty r^{n-2} \exp(-r^2/2) dr = 2^{-3/2} \pi^{-n/2}\Gamma((n-1)/2)$. $\endgroup$ – Mateusz Kwaśnicki Aug 24 '17 at 8:38
  • $\begingroup$ @MateuszKwaśnicki Sorry for the confusion. I meant the cone of positive semidefinite matrices. I will revise my statement in the question. Could you detail your claim in an answer? Thanks a lot! $\endgroup$ – Minkov Aug 24 '17 at 11:28
  • $\begingroup$ If $A$ is a "cone" (a positively homogeneous set) with Lipschitz boundary, then $\partial A$ is also a "cone". Denote $F=\partial A\cap \partial B(0,1)$. The $n-1$-dimensional Hausdorff measure $\sigma_{n-1}$ on $\partial A$ can be written as $\sigma_{n-1}(dx)=r^{n-2}\sigma_{n-2}(dz)dr$, where $x=rz$, $r>0$, $z\in F$ and $\sigma_{n-2}$ is the $n-2$-dimensional Hausdorff measure on $F$. Now apply this to the equivalent definition of the Gaussian perimeter: $\tau(A)=\int_{\partial A}\gamma(x)\sigma_{n-1}(dx)$, where $\gamma(x)$ is the density of the Gaussian measure. $\endgroup$ – Mateusz Kwaśnicki Aug 24 '17 at 11:49
  • $\begingroup$ @MateuszKwaśnicki For the cone of positive semidefinite matrices, is it possible to get a closed form or order estimation for the $n-2$-dimensional measure of $A\cap \{x: |x| = 1\}$? $\endgroup$ – Minkov Aug 24 '17 at 23:04
  • $\begingroup$ I do not know the answer, although I guess it follows from Weyl's integration formula, see mathoverflow.net/a/95256/108637. $\endgroup$ – Mateusz Kwaśnicki Aug 25 '17 at 8:40
3
$\begingroup$

Regarding the tightness of Ball's bound, which is $O(n^{1/4})$ for a convex set in dimension $n$, this was proven to be sharp (up to a constant factor) by Nazarov, in

Nazarov, Fedor, On the maximal perimeter of a convex set in $\mathbb{R}^n$ with respect to a Gaussian measure, Milman, V. D. (ed.) et al., Geometric aspects of functional analysis. Proceedings of the Israel seminar (GAFA) 2001--2002. Berlin: Springer (ISBN 3-540-00485-8/pbk). Lect. Notes Math. 1807, 169-187 (2003). ZBL1036.52014.

| cite | improve this answer | |
$\endgroup$
3
$\begingroup$

The Gaussian surface area of the positive semidefinite cone $\mathbb{S}_d^{+} = \{ M \in \mathbb{R}^{d\times d} \colon M \succeq 0 \}$ is bounded by a constant in $[0,1]$ uniformly, according to ''Learning Geometric Concepts via Gaussian Surface Area'' by Adam R. Klivans, Ryan O'Donnell and Rocco A. Servedio.

More specifically, as shown in Definition 2 of this paper, the Gaussian surface area of any convex set $A\in \mathbb{R} ^{d\times d} $ can be written as \begin{align} \Gamma(A) = \int_{\partial A} \varphi(x) d \sigma (x),~~~~[1] \end{align} where $\varphi$ is the Gaussian measure on $\mathbb{R}^{d\times d}$ and $\sigma$ is the Euclidean surface area. By Nazarov's inequality, for any convex set $A$ containing the origin, it holds that \begin{align} \int_{\partial A} \frac{\phi (x)}{1 + h(x) } d \sigma (x) \leq 1 - \gamma (A) \leq 1, ~~~~[2] \end{align} where $\gamma $ is the Gaussian measure and function $h$ is defined as the distance from the origin to the tangent hyperplane of $A$ which contains $h\in \partial A$. In more details, for any $h \in \partial A$, let $n_y$ be the unit normal vector to $\partial A$ at $y$. Then we define $h$ by $h(y)= \| y\|_2 \cdot \cos ( \langle y, n_y \rangle )$.

Now we go back to the PSD cone. The boundary of $\mathbb{S}_d^{+}$ is defined as \begin{align} \partial \mathbb{S}_d^{+} = \{ M\colon M \succeq 0, \textrm{rank}(M) < d\}. \end{align} For any $M \in \partial \mathbb{S}_d^{+}$, the tangent plane at $M$ is given by \begin{align} \mathcal{T}(M) = \left \{ X M + MX^{\top} \colon X \in \mathbb{R}^{d\times d} \right \}, \end{align} which implies that $M\in \mathcal{T} (M)$. Thus we have $h(M) = 0$. Combining $[1]$ and $[2]$, we conclude that the Gaussian surface area of the PSD cone is no larger than $1 - \gamma (S_d^{+})$.

| cite | improve this answer | |
$\endgroup$

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.