1
$\begingroup$

Let $X_1,...,X_N$ be IID, mean-zero random variables whose tail is bounded by a subgaussian-tailed variable to the fourth moment, i.e., for some $t \ge t_0 > 0$ $$ P(|X_i| > t) \le C\exp\left( -C t^{1/2} \right) $$ Then what is the best upper bound one can get for the $$ P\left( \left|\frac{1}{\sqrt{N}} \sum_{i=1}^N X_i \right| > t\right) $$ ?

$\endgroup$
2
$\begingroup$

For $\gamma\gt 0$ and a random variable $X$, define $$\lVert X\rVert_{\Phi_\gamma}:=\inf\left\{c\gt 0;\mathbb E\left[\exp\left(\left|X/c\right|^\gamma\right)\right] \leqslant 2\right\}.$$ For $\gamma\geqslant 1$, this defines a norm, and for $0\lt \gamma\lt 1$, this is equivalent to a norm. Obverse that Markov's inequality and the monotone convergence theorem entail $$\mathbb P \left\{\left|X\right|\geqslant x\right\}\leqslant 2\exp\left(-\frac{x^\gamma}{\lVert X\rVert^\gamma_{\Phi_\gamma}}\right),$$ and rewriting the expectation of the exponential in terms of norm, we derive that if $Y$ satisfies $$\mathbb P \left\{\left|Y\right|\geqslant t\right\}\leqslant K\exp\left(-\frac{t^\gamma}{\lambda^\gamma}\right)$$ for each positive $t$, then $\lVert Y\rVert_{\Phi_\gamma}\leqslant \left(1+K/2\right)^{1/\gamma}\lambda$. Therefore, the problem reduces to find a good upper bound for $\lVert \sum_{i=1}^nX_i\rVert_{\Phi_\gamma}$. This is precisely the purpose of Theorems 3 and 4 in Isoperimetry and integrability of the sum of independent Banach-space valued random variables by Michel Talagrand. In particular, this gives that if $X_1$ satisfies $$\mathbb P \left\{\left|X_1\right|\geqslant x\right\}\leqslant c_1\exp\left(-c_2t^\gamma\right)$$ then the quantity $\lVert \sum_{i=1}^nX_i\rVert_{\Phi_\gamma}/\sqrt n$ can be bounded independently of $n$.

$\endgroup$
  • $\begingroup$ Seems very helpful. Will read more into the paper. Thanks! $\endgroup$ – Koltchinskii Apr 2 '17 at 11:50
  • $\begingroup$ In our case theorem 3 should be very helpful $\endgroup$ – Koltchinskii Apr 2 '17 at 12:05

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.