2
$\begingroup$

In Wikipedia and many textbooks, they explain Markov's inequality for $X \geq a$ for a non-negative random variable $X$ and any $a > 0$ as follows: $$ \text{Pr}(X \geq a) \leq \frac{\mathbb{E}(X)}{a}$$ where $\mathbb{E}(X)$ is the expectation of $X$. Then, how to get the bound of $\text{Pr}(X > a)$ (i.e., without the equality in $\text{Pr}(X \geq a)$)? If the bound exists, is it meaningful to obtain the bound?

My guesses are:

  • under a p.d.f., $\text{Pr}(X>a)$ has the same bound as $\text{Pr}(X \geq a)$ according to Proof 1.
  • under a p.m.f., for a non-negative integer random variable X, the bound of $\text{Pr}(X>a)$ is represented as follows:
    • B-PMF-1) $\text{Pr}(X \geq a+1)=\text{Pr}(X>a) \leq \frac{\mathbb{E}(X)}{a+1}$, or
    • B-PMF-2) $\text{Pr}(X \geq a)= \text{Pr}(X > a) + \text{Pr}(X=a) \leq \frac{\mathbb{E}(X)}{a} \Leftrightarrow \text{Pr}(X > a) \leq \frac{\mathbb{E}(X)}{a} - \text{Pr}(X=a)$
    • B-PMF-2 will has the same bound as B-PMF-1 or tighter one than B-PMF-1.

Are those claims true?

Proof 1) Suppose that we have very small $\delta \geq 0$. Then, $\mathbb{E}(X)$ is $$ \mathbb{E}(X) = \int_{0}^{\infty}xp(x)dx = \int_{0}^{a-\delta}xp(x)dx + \int_{a-\delta}^{\infty}xp(x)dx \geq \int_{a-\delta}^{\infty}xp(x)dx \geq (a - \delta)\int_{a-\delta}^{\infty}p(x)dx = (a-\delta)\text{Pr}(X \geq a - \delta)$$ Note that $\text{Pr}(X \geq a - \delta) \geq \text{Pr}(X > a)$. Hence, $$ \mathbb{E}(X) \geq (a-\delta)\text{Pr}(X > a) \Leftrightarrow \text{Pr}(X > a) \leq \frac{\mathbb{E}(X)}{a - \delta}$$ As $\delta \rightarrow 0$, the bound converges to $\frac{\mathbb{E}(X)}{a }$.

Another way is that $\text{Pr}(X > a)=\text{Pr}(X \geq a) - \text{Pr}(X = a) \leq \frac{\mathbb{E}(X)}{a} - \text{Pr}(X = a)$, and $\text{Pr}(X = a) = 0$ when $p(\cdot)$ is a p.d.f. Thus, $\text{Pr}(X > a) \leq \frac{\mathbb{E}(X)}{a}$.

$\endgroup$
  • 2
    $\begingroup$ Huh? But wait, $$\{X>a\}\subseteq\{X\geqslant a\}$$ hence $$P(X>a)\leqslant P(X\geqslant a)$$ right? And since Markov inequality says that $$P(X\geqslant a)\leqslant E(X)/a$$ you are done, no? $\endgroup$ – Did Nov 12 '17 at 16:16
  • $\begingroup$ Thanks for your comment. You're right, and your answer solved my curiosity clearly. $\endgroup$ – 정진홍 Nov 14 '17 at 2:42
1
$\begingroup$

Do you remember the one-line proof of Markov's inequality ? It readily adapts:

$$a\mathbb 1_{X>a}\leq X$$ still holds true. Taking expectations on both sides, $aP(X>a)\leq E(X)$, hence $$P(X>a)\leq \frac{E(X)}a$$

$\endgroup$
  • 1
    $\begingroup$ Of course, this also follows directly from the original inequality, and the trivial fact that $P(X > a) \le P(X \ge a)$. $\endgroup$ – Ilmari Karonen Nov 12 '17 at 16:33
0
$\begingroup$

The bounds are correct, all comes from the equation you wrote
$P(X\geq a)-P(X=a)\leq \frac{E(X)}{a}-P(X=a)$

And clearly in the continuous case $P(X=a)=0$

$\endgroup$

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.