Would anyone know how to prove the following, coming from the proof of theorem 2 in this paper (https://arxiv.org/pdf/1605.08671.pdf)?

Consider i.i.d. Sub Gaussian random variables $(X_t)_{t\geq 1}$ with parameter $\sigma$ and mean $\mu$. Let $u \geq 1$. Then $P[\exists v \in \{2^u,...,2^{u+1} \} | \frac{1}{v} \sum_{t=0}^v X_t - \mu| > \frac{\delta}{\sqrt{v}}] \leq \exp(-\frac{\delta^2}{2 \sigma^2})$.


1 Answer 1


In the mentioned paper, they have $\sum_{t=1}^v$ rather than $\sum_{t=0}^v$. Anyhow, the inequality is incorrect in general. Indeed, without loss of generality, $\mu=0$ and $\sigma=1$. Suppose that (say) the $X_i$'s are just standard normal. Then, letting $n:=2^{u+1}[\ge4]$ and $S_v:=\sum_{t=1}^v X_t$, we see that $S_{n/2},\dots,S_n$ are jointly normal, and so, they have a bounded joint density. So, the lhs of the inequality is $1-O(\delta^{n/2+1})=1-o(\delta^2)$ as $\delta\downarrow0$, whereas the rhs is $1-\delta^2/(2+o(1))$.

However, the inequality can be easily fixed as follows, likely without any serious damage to the proof of that theorem in that paper (again assuming $\mu=0$ and $\sigma=1$): In view of Doob's maximal inequality for submartingales and the sub-Gaussian property, the lhs of your inequality (with $\sum_{t=1}^v$ rather than $\sum_{t=0}^v$) is no greater than \begin{align*} P(\max_{0\le v\le n}|S_v|>\delta\sqrt{n/2}) &\le e^{-h\delta\sqrt{n/2}}Ee^{h|S_n|} \\ & \le e^{-h\delta\sqrt{n/2}}(Ee^{hS_n}+Ee^{-hS_n}) \\ & \le 2\exp\{-h\delta\sqrt{n/2}+nh^2/2\} =2\exp\{-\delta^2/4\}, \end{align*} where $h:=\delta/\sqrt{2n}$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.