0
$\begingroup$

In the appendix of this paper appendix, Lemma I.4 the following Gaussian concentration lemma is given.

Consider $d$- dimensional multivariate Gaussian distribution $\overline{\xi}_t^k \sim \mathcal{N} (0, H \nu_k(\delta)(\Sigma^t_k)^{-1})$. For any $\delta > 0$, with probability $1-\delta$

\begin{equation} \|\bar{\xi}_t^k\|_{\Sigma_t^k} \leq c\sqrt{H d \nu_k(\delta)\log(d/\delta)} \end{equation} where c is some absolute constant.

Here $\Sigma^t_k$ is a positive definite matrix and $\nu_k$ is a function of $\delta$. In the paper they use a specific definition for $\nu_k$ but from the lemma description, it seems this lemma comes from more general statement. Could someone point me to a reference where I can find a more general statement of this? My guess there is something of the following form:

Consider random variable $X \sim \mathcal{N}(0,\alpha(\delta) \Sigma)$ where $\alpha$ is a scalar function of $\delta$ and $\Sigma \in \mathbb{R}^{d\times d}$. Then for any $\delta > 0$, with probability $1-\delta$

\begin{equation} \|X\|_{\Sigma^{-1}} \leq c\sqrt{d \alpha(\delta)\log(d/\delta)} \end{equation} where c is some absolute constant.

$\endgroup$
2
$\begingroup$

If $G$ is a random vector with $d$ iid $N(0,1)$ entries then $P(|\|G\|\le E[\|G\|] + \sqrt{2\log(1/\delta)} ) \ge 1-\delta $; this follows by Gaussian concentration of $\|G\|$ which is a 1-Lipschitz of the entries of $G$, see for instance Theorems 5.5, 5.6 in Concentration Inequalities: A Nonasymptotic Theory of Independence by Boucheron Gabor Lugosi and Massart. Furthermore $E[\|G\|]\le \sqrt d$ by Jensen's inequality.

Now in your problem you may apply this result to $G=E[XX^T]^{-1/2} X$ where $E[XX^T]=\Sigma \alpha(\delta)$. It looks that the above concentration bound gives $P\Big(|\|\Sigma^{-1/2}X\|\le \sqrt{\alpha(\delta)}\big( \sqrt{d}+ \sqrt{2\log(1/\delta) }\big) \Big) \ge 1-\delta $.

Section 3.1 of the book https://www.math.uci.edu/~rvershyn/papers/HDP-book/HDP-book.html by Vershynin is also relevant.

| 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.