6
$\begingroup$

Let $\binom{n}{k}:=\frac{\Gamma(n+1)}{\Gamma(k+1)\Gamma(1-k+n)}$ be the generalized binomial coefficient then I noticed by playing around with Mathematica that the function $f:[0,n/2] \rightarrow \mathbb R$

$$f(x) = \log\left(\binom{n}{n/2+x} \right)-n \alpha(1-(2x/n)^2)$$

has very interesting properties.

For $\alpha\le \tfrac{1}{2}$ the global maximum of this function is attained at $x_n=0$ for sufficiently large $n.$

Yet, for $\alpha>\tfrac{1}{2}$ the global maximum seems to be attained at some $x_n>0.$

My question is: Is it possible to analytically verify this property $(x_n>0)$ and in particular can anybody shed some light on what is so special about $\alpha =1/2$? (The first comment below this thread seems to indicate already that $\alpha=1/2$ is indeed the right threshold)

In particular, Mathematica gave me the following equation for the derivative of $f$

$$f'(x)= \frac{8 \alpha x}{n} + \operatorname{HarmonicNumber}(n/2 - x) - \operatorname{HarmonicNumber}(n/2 + x).$$

$\endgroup$
  • 1
    $\begingroup$ Expanding $f'(x)$ to $O(n^{-3})$ gives $ f'(x)=8 x (\alpha-\frac{1}{2})\ n^{-1}+ 4\ x\ n^{-2} -\frac{8}{3}\ x \ (1+2 x^2)\ n^{-3}+O(n^{-4}) $. We get three solutions of $f'(x)=0$: $x=0$ and two other solutions, which are real for $\alpha\ge \alpha_0$. The critical alpha is $\alpha_0= \frac{1}{2}-\frac{1}{2n}+\frac{1}{3 n^2}+O(n^{-3})$. (all calculations done with Mathematica 11) $\endgroup$ – Johannes Trost Nov 11 '18 at 9:10
  • 6
    $\begingroup$ I think that you observed the de Moivre's local limit theorem that states $$ 2^{-n/2-x}\binom{n}{n/2+x}\sim\frac{1}{\sqrt{n\pi/2}} e^{-2x^2/n} $$ as $n\to\infty$ $\endgroup$ – Liviu Nicolaescu Nov 11 '18 at 9:52
6
$\begingroup$

We have \begin{equation} f'(x)=\frac{8 \alpha x}{n}+\psi\left(\frac{n}{2}-x+1\right)-\psi\left(\frac{n}{2}+x+1\right), \end{equation} where $\psi:=(\ln\Gamma)'$. By the Gauss formula, Theorem 1.6.1, page 26 \begin{equation} \psi(x)=\int_0^\infty\Big(\frac{e^{-z}}{z}-\frac{e^{-xz}}{1-e^{-z}}\Big)dz, \end{equation} we have \begin{equation} f'(x)=\frac{8 \alpha x}{n}-2\int_0^\infty\frac{dz}{1-e^{-z}}\,e^{-(1+n/2)z}\sinh xz. \end{equation} Since $\sinh$ is strictly convex on $(0,\infty)$, $f'$ is strictly concave on $(0,n/2)$. Also, $f'(0)=0$ and $f'(\frac n2)\to-\infty$ as $n\to\infty$. So, eventually (i.e., for all large enough $n$), either $f'<0$ on $(0,n/2)$ (which will be the case if $f''(0)\le0$) or $f'$ changes sign only once on $(0,n/2)$, from $+$ to $-$, at some $x_n\in(0,n/2)$. So, either $f$ is decreasing on $(0,n/2)$ (which will be the case if $f''(0)\le0$) or increasing-decreasing, attaining its only maximum at some $x_n\in(0,n/2)$.

As $x\downarrow0$, \begin{equation} f'(x)=\Big(\frac{8\alpha}{n}-2\psi'(n/2+1)\Big)x+O(x^3), \end{equation} and \begin{equation} \frac{8\alpha}{n}-2\psi'(n/2+1)=\frac{4\alpha-2}{n}+\frac2{n^2}+O(n^{-3}) \end{equation} as $n\to\infty$.

So, if $\alpha<1/2$, then $f''(0)<0$ eventually, and so, $f$ is decreasing on $(0,n/2)$, with the maximum at $0$.

If $\alpha=1/2$, then $f'(c\sqrt n)=\left(4 c-\frac{16 c^3}{3}\right) \left(\frac{1}{n}\right)^{3/2}+O\left(\left(\frac{1}{n}\right)^{5/2}\right)$ for each real $c>0$ as $n\to\infty$, and so, eventually $f$ attains its only maximum at some $x_n\sim c\sqrt n$, where $c=\sqrt3/2$.

Finally, if $\alpha>1/2$ and $c\in(0,1/2)$, then $f'(cn)=g(c)+O(1/n)$ as $n\to\infty$, where $g(c):=8 \alpha c+\ln \frac{1-2 c}{2 c+1}$ for each real $c>0$. We have $g(c)=0$ iff $\alpha=a(c):=-\frac1{8c}\,\ln\frac{1-2 c}{2 c+1}$. Note that $a(c)$ increases from $1/2$ to $\infty$ as $c$ increases from $0$ to $1/2$. So, eventually $f$ attains its only maximum at some $x_n\sim cn$, where $c=a^{-1}(\alpha)$.

$\endgroup$

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.