7

It is known that for independent sub-exponential random variables, the following Bernstein-type inequality holds:

\begin{align} \mathbb{P}\biggl(\biggl| \sum_{i=1}^N a_i X_i\biggr| >t \biggr) \leq 2 \exp\left[ -c\min \left(\frac{t^2}{K^2 \| \vec{a}\|_2^2}, \frac{t}{ K \| \vec{a}\|_{\infty}} \right)\right], \end{align} where $K = \max \| X_1\|_{\psi_1}$ and $\vec{a}\in\mathbb{R}^N$.

I wonder if similar concentration inequality holds for heavy-tailed random variables where $X_i$ satisfies $\mathbb{P}(X_i > t) \leq C\exp(c t^{-\alpha}) $ for $\alpha\in(0,1)$.

amoeba
  • 93,463
  • 28
  • 275
  • 317
Steve
  • 287
  • 1
  • 6

1 Answers1

9

Yes, see Theorem 6.21 of [LT13], Michel Ledoux and Michel Talagrand. Probability in Banach Spaces: isoperimetry and processes, volume 23. Springer Science & Business Media, 2013.

For simplicity you may also look at section 8 of my paper.

http://arxiv.org/pdf/1507.06370v2.pdf

(I just summarized those theorems -- the purpose of the paper is completely different)

Alex Wenxin Xu
  • 191
  • 1
  • 2
  • It seems like you're defining $\psi_\alpha$ somewhat differently from talagrand when $alpha<1$? He says $e^{x^\alpha}-1$ (linear near the origin). You say $x^\alpha-1$. I don't see how your definition can give tail bounds, since we're only considering very tiny moments, but probably I misunderstood something.. – Thomas Ahle Dec 27 '18 at 09:30
  • I guess you just forgot an exp() in the second part of definition 8.1? Then you also match the definition in [vdVW00]. Btw, do you a more precise reference for your lemma 8.3? The vdVW book is long – Thomas Ahle Jan 02 '19 at 12:25