6
$\begingroup$

I need help to prove the following lower bound for tail probability. I have tried using well-known inequalities like Chebyshev and Paley-Zygmund, but cannot get the required bound.

Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space, and $X$ be a random variable with $\mathbb{P}(X\in [-M,M])=1$ for some positive constant $M$. Show that $\mathbb{P}\left(|X-\mu|\geq \frac{\sigma^2}{4M} \right)\geq \frac{\sigma^2}{8M^2}$, where $\mu= \mathbb{E}(X)$ and $\sigma^2=\text{Var}(X)$.

Thank you.

Additional question: How can I also prove that $\mathbb{P}\left(X\geq \mu+ \frac{\sigma^2}{4M} \right)\geq \frac{\sigma^2}{8M^2}$ ?

$\endgroup$
1
5
$\begingroup$

The function $f(t):= \mathbb P(|X-\mu|\ge t)-\frac t{2M}$ is strictly decreasing on $[0,\infty)$ and we have $f(0)=1$, $f(2M)=-1$. Let $s=\inf\{t\ge0\mid f(t)\le 0\}$. Then $$ \sigma^2 = \int (x-\mu)^2 d\mu =\int_0^\infty 2r\mathbb P(|X-\mu|\ge r) dr\\ \le\int_0^s2r\cdot1\,dr+\int_s^{2M}2r\cdot\frac s{2M}dr\\ =s^2+2Ms-\frac{s^3}{2M}. $$ If $s=0$, then this shows $\sigma^2=0$ and thus the claim we want to show is the trivial statement $\mathbb P(|X-\mu|\ge0)\ge0$. Therefore, we may assume $0< s\le 2M$ and conclude $$ \sigma^2\le s^2+2Ms-\frac{s^3}{2M}< 2Ms+2Ms-0=4Ms.$$ But then $\frac{\sigma^2}{4M}<s$ implies $f\left(\frac{\sigma^2}{4M}\right)>0$, i.e. $$ \mathbb P\left(|X-\mu|\ge \frac{\sigma^2}{4M}\right)>\frac{\sigma^2}{8M^2}.$$ Apparently, the only case where $\le$ canot be replaced with $>$ is the case of an (almost surely) constant variable.

$\endgroup$
1
  • $\begingroup$ Thanks a lot - clever argument! I'm curious... Are there more inequalities of this type? Could you point me towards a reference? $\endgroup$ Nov 2 '13 at 23:07
0
$\begingroup$

The fourth moment bound says: $$ E[|X|] \ge \frac{E[X^2]^{3/2}}{E[X^4]^{1/2}}.\\ $$ Let $s = \max\{|\mu+M|, |\mu-M|\} \le 2M$, that is the furthest the mean can be from $X$. Then we have $E[(X-\mu)^4] \le s^2 E[(X-\mu)^2]$.

Thus, by Paley–Zygmund's inequality: $$ \Pr[|X-\mu| \ge \alpha\sigma^2] \ge (1-\alpha)^2\frac{E[|X-\mu|]^2}{E[|X-\mu|^2]} \ge (1-\alpha)^2\frac{E[|X-\mu|^2]^2}{E[|X-\mu|^4]} \ge (1-\alpha)^2\frac{E[|X-\mu|^2]}{s^2}. $$

Setting $\alpha=1/(4M)$ we have $$ \Pr[|X-\mu| \ge \frac{\sigma^2}{4M}] \ge (1-1/(4M))^2\frac{\sigma^2}{s^2} \ge \frac{\sigma^2}{8M^2} , $$ where the last bound requires $M\ge 1$. However, that requirement could probably be removed by using a slightly stronger version of Paley–Zygmund.

$\endgroup$

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.