Let $T \subset \mathbb R^n$, and assume it's a finite set if that helps. Consider the symmetric Gaussian process $(X_t)_{t\in T}$ defined by $X_t = \langle G, t\rangle$, where $G$ is a standard Gaussian in $\mathbb{R}^n$ and $\langle\cdot, \cdot \rangle$ is the standard inner product. In other word, $\mathbb{E} X_t = 0$, and $\mathbb{E} (X_s - X_t)^2 = \|s-t\|_2^2$ for all $s,t\in T$

I want to know what happens to $\mathbb E \sup_{t\in T} X_t$ when we apply a contraction map to the index set $T$?

I.e. take a contraction map $f:\mathbb R^n \to \mathbb R^n$, $\|f(x)-f(y)\|_2 \leq \|x-y\|_2$ for all $x,y\in\mathbb R^n$, and define the process $(Y_t)_{t\in T}$ as $Y_t = \langle G, f(t)\rangle$. (Equivalently, $\mathbb{E} Y_t = 0$, and $\mathbb{E} (Y_s - Y_t)^2 = \|f(s)-f(t)\|_2^2$.) An application of Talagrand's majorizing measures theorem shows that for some absolute constant $L$,

$$ \mathbb E \sup_t Y_t \leq L\ \mathbb E \sup_t X_t. $$

Here are two questions:

  1. Is the bound true with $L=1$?

  2. Is there a direct proof that does not use Talagrand's theorem?

I am even interested in the simple map $f(t_1, \ldots, t_n) = (|t_1|, \ldots, |t_n|)$. Is there a simple direct proof in this case? Is the inequality true with constant 1 in this case?


It is true, and it follows from the following fact, sometimes referred to as Sudakov-Fernique inequality, sometimes as Slepian-Fernique lemma:

Assume that $(X_t)$ and $(Y_t)$ are two centered Gaussian processes. If $\|Y_s - Y_t\|_2 \leqslant \|X_s - X_t\|_2$ then $\mathbb{E} \sup Y_t \leqslant \mathbb{E} \sup X_t$.

In this inequality the index set $T$ is assumed to be countable so that we do not have any problems with measurability.

  • $\begingroup$ Thank you Mateusz! I now realize this is something I should know, for example as a step in proving Sudakov's minoration. $\endgroup$ Apr 24 '15 at 17:00

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.