2
$\begingroup$

Can the following quantity be upper bounded by the (standardized third moment) Skewness of $X$ ($\mu_3/\sigma^3$)? $$\left|\mathbb{P}(X \geq \mathbb{E}X) - \mathbb{P}(X \leq \mathbb{E}X)\right|$$

I am particularly interested in the case where $X \sim $Bin$(k,p)$, where the skewness is given by $(1-2p)/\sqrt{kp(1-p)}$.

$\endgroup$
5
  • $\begingroup$ Do you mean the population standardized third-moment skewness $\mu_3/\sigma^3$? Or some sample quantity? Or some other skewness measure? $\endgroup$
    – Glen_b
    Commented Sep 28, 2014 at 6:50
  • $\begingroup$ Is this for some course/subject? $\endgroup$
    – Glen_b
    Commented Sep 28, 2014 at 7:02
  • $\begingroup$ I do mean the population standardized third-moment skewness. For the binomial, this quantity, as you observed in your answer below, is $(1-2p)/\sqrt{kp(1-p)}$. And, this is not for a course. $\endgroup$
    – Skoro1
    Commented Sep 28, 2014 at 15:48
  • $\begingroup$ You have me curious how such a question would arise, outside some form of study. Is this self-learning? $\endgroup$
    – Glen_b
    Commented Sep 28, 2014 at 23:09
  • 1
    $\begingroup$ I am a researcher, and this is a toy version of a toy version of a toy problem :) Thanks for your help! $\endgroup$
    – Skoro1
    Commented Sep 29, 2014 at 19:19

1 Answer 1

1
$\begingroup$

This looks like it might be related to bookwork, so for the moment I'll just explore some possibilities.

In general I wouldn't hold much hope for a very useful bound, but by restricting it to binomial $(n,p)$ (sorry, I was working with $n$ in place of your $k$). I think it might be reasonably likely you can find something that will work pretty well.

In the case of the binomial, the population (standardized-third-moment) skewness is

$$\frac{1-2p}{\sqrt{np(1-p)}}$$

You could try rewriting your probability in terms of the pdf and the cdf. You might like to consider the case where $E(X)$ is not an integer separately from the case where it is an integer (however, because of the way that the equality will be nonzero at the same time, the terms cancel, so things might simplify enough not to keep them separate).

Note that the cdf of the binomial is $\mathbb{P}(X \leq \mathbb{E}X)=\text{I}_{1-p}(n - k, 1 + k)$, (the regularized incomplete beta function) for which some approximations and bounds can be found, but there are also a variety of other bounds for the cdf.

One of these might allow you to relate such bounds to the skewness.

$\endgroup$
1
  • 1
    $\begingroup$ Thanks, I will play around with this. One approach I tried is to replace the binomials by Gaussians. Then, Berry-Esseen seems to give a pretty closely related upper bound: $(2p^2 + (1-2p))/\sqrt{kp(1-p)}$. I am not sure if this is tight. $\endgroup$
    – Skoro1
    Commented Sep 28, 2014 at 15:54

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.