2
$\begingroup$

Let: $F_{j+1,y}(s)$ be the cumulative distribution function of a binomial distribution with mean $y$, $j+1$ independent trials considered for $s$ successes. Is it possible to show in any way that:

$\frac{1}{F_{j+1,y}(s+2)}-\frac{2}{F_{j+1,y}(s+1)}+\frac{1}{F_{j+1,y}(s)}>0, 0\leq s\leq j-1$

?

This can be helpful when comparing the expected value of random processes governed by poisson-binomial distributions against ones governed by binomial distribution. And in infer statistical dominance of a process over the other, indeed is a frequent term in bandit algorithms problems Thank you for your kind attention

$\endgroup$

1 Answer 1

1
$\begingroup$

The answer is yes. Indeed, the cdf $F$ of the binomial distribution is log concave -- see e.g. Theorem 2 on p. 152, used with $\alpha=1$, $r=\infty$, and $q$ being the probability mass function of the binomial distribution. So, $1/F$ is log convex on the set $S:=\{0,\dots,j+1\}$ of all atoms of the distribution. So, $1/F$ is strictly convex on $S$.


Detail on the last statement: Let $g:=1/F$, $i\in\{1,\dots,j\}$, $x:=g(i-1)$, $y:=g(i)$, $z:=g(i+1)$. We have $x>z$ and $(\ln x+\ln z)/2\ge\ln y$. We want to show that then $(x+z)/2>y$. But $(\ln x+\ln z)/2\ge\ln y$ means that $y\le\sqrt{xz}$ and hence, by the inequality of arithmetic and geometric means, $(x+z)/2\ge\sqrt{xz}\ge y$, and the first of the latter two inequalities is strict, because $x>z$ and hence $x\ne z$. Thus, $(x+z)/2>y$.

$\endgroup$
5
  • $\begingroup$ Thank you @IosifPinelis ! Where do I find the reference that the reciprocal of a log concave function is log convex? It appears to me that the remark states for the pdf to be logconcave not for the cdf, am I missing someething? $\endgroup$ Commented Jan 29 at 15:05
  • $\begingroup$ @MarcoMaxFiandri : (i) That the reciprocal of a log concave function is log convex follows immediately from the definitions of such functions -- similarly to the fact that the function $(-f)$ is convex if $f$ is concave. (ii) A better reference than that remark is Theorem 2 of the same linked paper. I have also added details on this. $\endgroup$ Commented Jan 29 at 17:21
  • $\begingroup$ I see! I understood thank you for your patience! However I don't get how we retrieve the last inequality. Sure we've proved that (calling B_j the reciprocal of the cdf) $\ln(B_{j+2})-2\ln(B_{j+1})+ln{B_{j}}>0$ but this implies that the inequality also hold true for $B_{j}$? $\endgroup$ Commented Jan 29 at 17:41
  • $\begingroup$ @MarcoMaxFiandri : I have added a detail on this as well -- this is just the AM-GM inequality. $\endgroup$ Commented Jan 29 at 18:21
  • $\begingroup$ thank you so much I will acknowledge your priceless contribution in the article $\endgroup$ Commented Jan 29 at 18:41

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.