Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Let $X_1,\dots,X_n\in {\bf R}^d$ be $d$-dimensional iid zero mean random vectors with covariance matrix $\Sigma$. I am interested in tail bounds for the Euclidean norm $$N_n\equiv \frac{1}{\sqrt{n}}\|X_1 + X_2 + \dots X_n\|.$$ The main assumption should be that the one-dimensional marginals of $X_1$ have finite $L^p$ norm for some $p>4$, ie that there exists $c>0$ such that: $${\bf (A)}_p:\forall v\in {\bf R}^d, {\bf Ex}\langle X_1,v\rangle^p\leq c \langle v,\Sigma v\rangle^{p/2}$$ where $\langle,\rangle$ is the usual inner product (edit: fixed typo in above display). More specifically, I am interested in Central-Limit-Theorem like bounds of the form: $$(\star)\Pr(N_n>\sqrt{a_1 tr(\Sigma) + a_2 \|\Sigma\|_{{\rm op}}t})\leq e^{-t} + C n^{-\beta}$$ where $a_1,a_2$ are universal and $\beta,C$ depend only on $c$ and $p$ (and not on the dimension $d$); $tr(\Sigma)$ is the trace; and $\|\Sigma\|_{{\rm op}}$ is the operator norm.

If the $X_i$ are Gaussian, this result follows from an explicit calculation without the $n^{-\beta}$ term. A similar result holds if the $X_i$ are subgaussian [1].

Much less seems to be known for more general distributions. All results I found (see eg. [2]) assume less about the $X_i$'s but also give poorer bounds with respect to dependence on d. On the other hand, these results also go for an actual CLT whereas I just want the looser inequality $(\star)$.

Is there a (perhaps obvious) obstruction to this? Do you know any references that prove/disprove $(\star)$ or something close?

[1] A tail inequality for quadratic forms of subgaussian random vectors. Daniel Hsu, Sham M. Kakade, and Tong Zhang. Electronic Communications in Probability, 17(52):1-6, 2012.

[2] On the rate of convergence of the multivariate CLT. F. Götze Annals of Probability, 19 (2): 724-739, 1991.

share|improve this question
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Browse other questions tagged or ask your own question.