Famous Gaussian concentration inequality states that:

If $\mathrm{F}$ is 1 -Lip, and $\mathbb{E} F(X)=0,$ and $X=(X_1,...,X_n) \sim N\left(0, I_{n}\right),$ then we have for some absolute constant $C>0$ $$ \|F(X)\|_{\psi_{2}} \leq C $$ uniformly for all $\mathrm{n}$.

I always thought the independent condition for $X_1,...,X_n$ here is indispensable. But today I came across a special case: for finite measure $G(\cdot)$ on $\mathbb{N}$, and $F(X):=\log \sum_i \exp(X_i)G(i)$. If we have $\sup_i \mathbb{E}X_i^2 <C^2$, then

$$ \|F(X)-\mathbb{E}F(X)\|_{\psi_{2}} \leq C $$ even when $X_1,X_2,...$ are correlated. I wonder if there is a generalization of such concentration inequality for correlated Gaussians for a class of function $F$.


1 Answer 1


It seems that a more general version of Gaussian isoperimetric inequality (i.e. Bakry Emery) applies to distribution with density function of the form: $\propto e^{-H(t)}dx$ where $\nabla^2H\ge K,K\in\mathbb{R}^+$ (this means $\ge KI_n$). Bakry-Emery then states: the log-Sobolev inequality holds with $\gamma \leq \frac{2}{K}$ : $$ S(f) \leq \frac{2}{K} D(\sqrt{f}) $$ where $S$ is entropy, $D$ is Dirichlet form. Concentration result like the one in my problem statement will follow from Herbst argument with Lipschitz test functions.

In the particular log-sum-expo case I gave, i.e. for correlated Gaussian, $\nabla^2H=\Sigma^{-1}>\lambda_n(\Sigma^{-1})=(\lambda_1(\Sigma))^{-1}\ge n/C^2$ by Schur-Horn inequality. So concentration holds with subgaussian norm O($C$)


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.