4
$\begingroup$

Suppose $c>2$ is a real number, and $x$ solves equation $f(\frac{1}{2}+x,m,2m)=(2m+1)cx$ for some $m>0$ integer number, where $f$ is probability mass function for binomial distribution, with probability $\frac{1}{2}+x$ and population size $2m$.

Simulations show that it must hold: $Pr(X\geq m+1)\leq \frac{1}{2}+\frac{1}{c}$, where $X$ random variable is distributed binomially with probability $\frac{1}{2}+x$ and population size $2m+1$. Hints how to prove the inequality?

$\endgroup$
4
$\begingroup$

Let us prove a result stronger than what was requested. As in the OP, let $X=X_p$ be a random variable (r.v.) distributed binomially with parameters $2m+1$ and $p:=\frac12+x$ (so that $x=p-\frac12$) and let $Y=Y_p$ be a r.v. distributed binomially with parameters $2m$ and $p=\frac12+x$.

Then the condition $f(\frac12+x,m,2m)=(2m+1)cx$ in the OP can be rewritten as \begin{equation*} c=c_p:=\frac{P(Y_p=m)}{(2m+1)x}, \end{equation*} so that $0<x\le\frac12$. The conjecture in the OP can now be rewritten as \begin{equation*} g(p):=P(X_p\ge m+1)\overset{\text{(?)}}\le\frac12+\frac1c=\frac12+\frac{(2m+1)x}{P(Y_p=m)} \end{equation*} for $p\in(\frac12,1]$.

Let us prove the following much stronger result: \begin{equation*} P(X_p\ge m+1)=g(p)\overset{\text{(?)}}\le h(p):=\tfrac12+(2m+1)x\,P(Y_{1/2}=m) \tag{1} \end{equation*} for $p\in[\frac12,1]$.

We have \begin{equation*} g(p)=\sum_{j=m+1}^{2m+1}\binom{2m+1}j p^j q^{2m+1-j}, \end{equation*} where $q:=1-p$, whence \begin{align*} g'(p)&=\sum_{j=m+1}^{2m+1}\frac{(2m+1)!}{j!(2m+1-j)!} jp^{j-1} q^{2m+1-j} \\ &-\sum_{j=m+1}^{2m+1}\frac{(2m+1)!}{j!(2m+1-j)!}(2m+1-j)p^j q^{2m-j} \\ &=(2m+1)\sum_{j=m+1}^{2m+1}\binom{2m}{j-1} p^{j-1} q^{2m+1-j} \\ &-(2m+1)\sum_{j=m+1}^{2m}\binom{2m}j p^j q^{2m-j} \\ &=(2m+1)\binom{2m}m p^m q^m \\ &\le(2m+1)\binom{2m}m (\tfrac12)^{2m} =(2m+1)P(Y_{1/2}=m). \end{align*} So, for $p\in[\frac12,1]$, \begin{equation*} g(p)\le g(\tfrac12)+(2m+1)P(Y_{1/2}=m)(p-\tfrac12)= \tfrac12+(2m+1)P(Y_{1/2}=m)x=h(p), \end{equation*} so that the question mark in (1) can be removed.

Remark. Using the first 5 lines of the above 6-line display, we can also get the following lower bound on $P(X_p\ge m+1)$: \begin{equation*} P(X_p\ge m+1)\ge \tfrac12+(2m+1)x\,P(Y_p=m); \end{equation*} cf. the upper bound on $P(X_p\ge m+1)$ in (1).

Graphs of the probability $P(X_p\ge m+1)$ to bound, of the upper and lower bounds on it given in this answer, and of the upper bound conjectured in the OP (as functions of $p$) are shown in the picture below for $m=5$:

enter image description here

$\endgroup$
2
  • 1
    $\begingroup$ Maybe I am confused by a notation, but isn't $P(X\geq m+1)$ exactly equal to $\frac{1}{2}$ with parameters $2m+1$ and $\frac{1}{2}$, because of symmetry? I think the answer is wrong. $\endgroup$ – anagram Feb 26 '19 at 23:55
  • $\begingroup$ Oops! You are right. I have now given a completely different answer, not only confirming your conjecture, but also improving it quite a bit. A similar lower bound on $P(X_p\ge m+1)$ is now also given. $\endgroup$ – Iosif Pinelis Feb 27 '19 at 2:15

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.