5
$\begingroup$

This has received no full solution at StackExchange.

As per https://dlmf.nist.gov/8.10#E13 we have

$$\frac{\Gamma\left(n,n\right)}{\Gamma\left(n\right)}<\frac{1}{2}<\frac{\Gamma% \left(n,n-1\right)}{\Gamma\left(n\right)}$$ for $n=1,2, \ldots$.

My question is: 1) show that both of these are monotone in $n$ AND 2) replacing $n$ by real $s>1$, show that these are monotone in $s$ (of course 1 follows from 2, but perhaps 1 is easier).

Note: For large $n$ both of these converge to $\frac{1}{2}$: rewriting via repeated integration by parts we get $\frac{\Gamma% \left(n,n-1\right)}{\Gamma\left(n\right)}=\exp\{-(n-1)\}\sum_{i=0}^{n-1} \frac{(n-1)^i}{i!}$ which is the CDF of Poisson distribution with $\lambda=n-1$ evaluated at $x=\lambda=n-1$; as $n$ increases this approaches CDF of normal with mean and variance $\lambda$, which at $x=\lambda$ is $\frac{1}{2}$. (I would be interested in alternative proofs of this fact as well.)

$\endgroup$
2
  • $\begingroup$ Now your second question is completely answered as well. $\endgroup$ Commented Oct 14, 2020 at 0:36
  • $\begingroup$ Beautiful, thank you. I'll keep the "accept" on the first answer, since it is a bit simpler and does solve the integer case. They are both great! $\endgroup$
    – Max M
    Commented Oct 15, 2020 at 13:28

2 Answers 2

2
$\begingroup$

$\newcommand\Ga\Gamma$Using integration by parts, we have $$\Ga(n,t)=t^{n-1}e^{-t}+(n-1)\Ga(n-1,t)$$ for real $t>0$.

So, for $n\ge2$ the sign of $$\frac{\Ga(n,t)}{\Ga(n)}-\frac{\Ga(n-1,t-1)}{\Ga(n-1)}$$ is the same as the sign of $$t^{n-1}e^{-t}-(n-1)\int_{t-1}^{t} x^{n-2}e^{-x}\,dx.\tag{1}$$


Letting here $t=n-1$, we see that for $n\ge2$ the inequality $$\frac{\Ga(n,n-1)}{\Ga(n)}<\frac{\Ga(n-1,n-2)}{\Ga(n-1)}$$ can be rewritten as $$(n-1)^{n-2}e^{-(n-1)}<\int_{n-2}^{n-1} x^{n-2}e^{-x}\,dx,$$ which is true, because $x^{n-2}e^{-x}$ is decreasing in $x\in(n-2,\infty)$ and hence $x^{n-2}e^{-x}>(n-1)^{n-2}e^{-(n-1)}$ for $x\in(n-2,n-1)$. Thus, $\frac{\Ga(n,n-1)}{\Ga(n)}$ is decreasing in natural $n$ (and also in $n\in\{a,a+1,\dots\}$ for any real $a\ge1$).


Somewhat similarly, one can show that the sign of (1) is $+$ for $t=n$, so that $\frac{\Ga(n,n)}{\Ga(n)}$ is increasing in natural $n$. Indeed, the statement that the sign of (1) is $+$ for $t=n$ can be rewritten as $$1>\int_{n-1}^n e^{h(x)}\,dx,\tag{2}$$ where $h(x):=(n-2) \ln x-(n-1) \ln n+n-x$. Note that $h'(x)x=n-2-x$, so that $h$ is decreasing on $[n-1,n]$. Also, $h(n-1)<1-\ln n<0$ for $n\ge3$; so, $h<0$ on $[n-1,n]$, which does yield (2) for $n\ge3$. So, $\frac{\Ga(n,n)}{\Ga(n)}$ is increasing in natural $n\ge2$. The inequality $\frac{\Ga(n,n)}{\Ga(n)}>\frac{\Ga(n-1,n-1)}{\Ga(n-1)}$ in the case $n=2$ is immediate. So, $\frac{\Ga(n,n)}{\Ga(n)}$ is increasing in all natural $n$.


That $\frac{\Ga(n,t)}{\Ga(n)}\to1/2$ for $n\to\infty$ and $t=n+o(\sqrt n)$ follows immediately by the central limit theorem, because $\frac{\Ga(n,t)}{\Ga(n)}=P(X_1+\cdots+X_n>t)$, where the $X_i$'s are iid standard exponential random variables.

$\endgroup$
3
$\begingroup$

$\newcommand\Ga\Gamma \newcommand\tp{\tilde p} \newcommand\tq{\tilde q} \newcommand\tr{\tilde r}$First, let us show that $\dfrac{\Ga(s,s-1)}{\Ga(s)}$ decreases in real $s>1$. This is equivalent to $\dfrac{\Ga(s)-\Ga(s,s-1)}{\Ga(s,s-1)}$ being increasing in real $s>1$. Note that \begin{align*} \Ga(s)-\Ga(s,s-1)&=\int_0^{s-1} t^{s-1}e^{-t}\,dt=(s-1)^s\int_0^1 x^{s-1}e^{-(s-1)x}\,dx, \\ \Ga(s,s-1)&=\int_{s-1}^\infty t^{s-1}e^{-t}\,dt=(s-1)^s\int_1^\infty x^{s-1}e^{-(s-1)x}\,dx, \\ \end{align*} So, what we need to show is that \begin{equation} R(u):=\frac{I(u)}{J(u)} \end{equation} is increasing in real $u>0$, where \begin{align*} I(u)&:=\int_0^1 f(x)^u\,dx=\int_0^1 z^u\,p(z)\,dz, \\ J(u)&:=\int_1^\infty f(x)^u\,dx=\int_0^1 z^u\,q(z)\,dz, \end{align*} where $f(x):=xe^{1-x}$, $p(z):=x_1'(z)>0$, $q(z):=-x_2'(z)>0$, $x_1(z)$ is the only root $x\in(0,1)$ of the equation $f(x)=z$ for $z\in(0,1)$, and $x_2(z)$ is the only root $x\in(1,\infty)$ of the equation $f(x)=z$ for $z\in(0,1)$.

To show this, note that \begin{align*} 2J(u)^2R'(u)&=2\int_0^1\int_0^1 dx\,dy\,(xy)^u p(x)q(y)(\ln x-\ln y) \\ &=2\int_0^1\int_0^1 dy\,dx\,(yx)^u p(y)q(x)(\ln y-\ln x) \\ &=\int_0^1\int_0^1 dy\,dx\,(xy)^u [p(x)q(y)-p(y)q(x)] (\ln x-\ln y)\\ &=\int_0^1\int_0^1 dy\,dx\,(xy)^u\,p(y)q(y) [r(x)-r(y)](\ln x-\ln y), \end{align*} where \begin{equation} r:=p/q. \end{equation} So, it remains to show that $r$ is increasing on $(0,1)$.

We have \begin{equation} p=\frac{x_1}{(1-x_1) z},\quad q=\frac{x_2}{(x_2-1) z}, \end{equation} \begin{equation} r'=\frac{x_1(x_2-x_1)(x_1+x_2-2)}{(1-x_1)^3 (x_2-1) x_2 z}. \end{equation}

So, it remains to show that $x_1+x_2-2>0$ or, equivalently, $x_2>2-x_1$ or, equivalently, $f(x_2)<f(2-x_1)$ or, equivalently, $f(x_1)<f(2-x_1)$.

So, it remains to show that $f(x)<f(2-x)$ for $x\in(0,1)$ or, equivalently, $1<h(x):=(2/x-1)e^{2x-2}$ for $x\in(0,1)$, which follows because $h(1)=1$ and $h'(x)=-\dfrac{2 e^{2 x-2} (1-x)^2}{x^2}<0$, so that $h$ is decreasing on $(0,1)$. $\Box$


Let us also prove that $\dfrac{\Ga(s,s)}{\Ga(s)}$ increases in real $s>0$. This proof is similar to the one above. Here in place of $p,q,r=p/q$ we get \begin{equation} \tp:=pe^{-x_1},\quad \tq:=qe^{-x_2},\quad \tr:=\tp/\tq, \end{equation} with \begin{equation} \tr'=\frac{(x_2-x_1) (x_1 x_2-1)}{(1-x_1)^3 (x_2-1) z} \end{equation}

So, it remains to show that $x_1x_2-1<0$ or, equivalently, $x_2<1/x_1$ or, equivalently, $f(x_2)>f(1/x_1)$ or, equivalently, $f(x_1)>f(1/x_1)$.

So, it remains to show that $f(x)>f(1/x)$ for $x\in(0,1)$ or, equivalently, $1<g(x):=x^2e^{1/x-x}$ for $x\in(0,1)$, which follows because $g(1)=1$ and $g'(x)=-(1-x)^2e^{1/x-x}<0$, so that $g$ is decreasing on $(0,1)$. $\Box$


Another way to prove the inequalities $x_1+x_2-2>0$ and $x_1x_2-1<0$, used in the proofs above, is to note that the condition $f(x_1)=f(x_2)$ means that the logarithmic mean of $x_1,x_2$ is $1$, and then use the arithmetic-logarithmic-geometric mean inequality.


The same method can be used to establish monotonicity of the ratio $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ for real $a\notin\{-1,0\}$. For any given real $a$, it will follow that (i) $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ is increasing in real $s>\max(0,-a)$ if the function $$m_a:=(a+1) (x_1 x_2-1)-a (x_1+x_2-2)$$ is $<0$ on $(0,1)$ and (ii) $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ is decreasing in real $s>\max(0,-a)$ if $m_a>0$ on $(0,1)$. So, it follows from the inequalities $x_1 x_2-1<0<x_1+x_2-2$, proved above, that (i) $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ is increasing in real $s>0$ for each real $a\ge0$ and (ii) $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ is decreasing in real $s>-a$ for each real $a\le-1$.

Moreover, one can show that $m_a>0$ on $(0,1)$ iff $a\le-1/3$. So, $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ is decreasing in real $s>-a$ for each real $a\le-1/3$. Furthermore, one can show that $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ is neither decreasing nor increasing in real $s>-a$ for each real $a\in(-1/3,0)$.

We conclude that $\dfrac{\Ga(s,s+a)}{\Ga(s)}$ is

(i) increasing in real $s>0$ for each real $a\ge0$;

(ii) decreasing in real $s>-a$ for each real $a\le-1/3$;

(iii) non-monotonic in real $s>-a$ for each $a\in(-1/3,0)$.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .