Concentration inequality to bound expectation

308 Views Asked by At

Let $X$ be a non-negative r.v. so that $$ P(X \geq t) \leq C \exp\bigg\{\frac{-t^2/2}{\sigma^2 + bt}\bigg\}$$ for positive $\sigma, b$ and $C\geq 1$. Show that

$$ E[X] \leq 2\sigma (\sqrt{\pi} + \sqrt{\log C}) + 4b(1 + \log C) $$

I know how to usually get concentration bound or bound the deviation from the mean, but no idea how to bound this expectation here. Any help would be really appreciated.

1

There are 1 best solutions below

4
On

Hold on, this inequality doesn't appear true. Let's take $C = 100, \sigma^2 = 1, b = 10^{-6}$ and suppose the tail-bound were exact. Then \begin{align*} E[X] = 100\int_0^\infty \exp\left(\frac{-t^2}{2(1 + 10^{-6}t)}\right)dt \approx 125.33 \end{align*} Yet \begin{align*} 2\cdot 1 \cdot (\sqrt{\pi}+\sqrt{\log 100}) + 4\cdot 10^{-6}\cdot(1+\log 100) \approx 7.84 \end{align*}