I'm interested in the asymptotics of $$\int_0^\infty \frac{x^{2z}}{\Gamma(1+z)}\,dz$$ as $x\to\infty$. I expect the results to behave similarly to $e^{x^2}=\sum_{k\ge 0}\frac{x^{2k}}{k!}$. However, I'm not quite sure how to develop the leading asymptotics of the integrals. I first thought that for large $x$, the integral should be dominated by its integral over a small ball around the maximum of the integrand $\frac{e^{2z\log x}}{\Gamma(1+z)}$.

To find this maximum, I computed $$f'(z)=\left(2\log(x)-\psi^{(0)}(1+z)\right)f(z)$$ with $\psi^{(0)}(z)$ the logarithmic derivative of $\Gamma(z)$. Since $f$ never vanishes, the maximum must occur at $z\in(0,\infty)$ such that $\psi^{(0)}(1+z)=2\log x$. Assuming that we will take $x\to\infty$ as well as the asymptotics $\psi(z)=\log z +O(1/z)$ for large and positive $z$, we seek to solve $2\log x= \log(1+z) +O(1/z)$. Exponentiating, we find that $x^2=1+z+O(1)$. Thus, we have that $\operatorname{arg max} f(z)\sim x^2$ is asymptotically correct for large $x$. Let $z_0=x^2$. We can now rewrite the integral as dominated by $$\frac{e^{2x^2\log x}}{\Gamma(1+x^2)}\int_{z_0-\epsilon}^{z_0+\epsilon} e^{2(z-z_0)\log x} \frac{\Gamma(1+z_0)}{\Gamma(1+z)}\,dz.$$ However, I'm not sure what size to take $\epsilon$ as a function of $x$. I do know that the expression outside of the integral is asymptotic to $$(2\pi)^{-1/2} \frac{e^{x^2}}{x}.$$ I'm not sure how to deal with the actual integral though. Input is much appreciated.

Edit: Math Stack Exchange cross-post

  • $\begingroup$ Consider $g(x)=\int_0^\infty\frac{x^t}{\Gamma(t+1)}dt$. You differentiate and then split the integral at 1, you find that $g(x)/e^x$ has bounded derivative. $\endgroup$
    – user473423
    Dec 31, 2022 at 8:06
  • $\begingroup$ Ramanujan observed that $$ \int_0^{ + \infty } {\frac{{x^{2z} }}{{\Gamma (1 + z)}}{\rm d}z} = {\rm e}^{x^2 } - \int_{ - \infty }^{ + \infty } {\frac{{\exp ( - x^2 {\rm e}^t )}}{{t^2 + \pi ^2 }}{\rm d}t} . $$ $\endgroup$
    – Gary
    Feb 8 at 6:13

2 Answers 2


$\newcommand\Ga\Gamma$Let $g(x)$ denote your integral. Then $$g(x)\sim e^{x^2} \tag{1}\label{1} $$ as $x\to\infty$.

$$g(x)=\int_0^\infty dz\,\frac{x^{2z}}{\Ga(1+z)}. \tag{1.5}\label{1.5} $$ So, for $x>0$, $$g'(x)=\int_0^\infty dz\,\frac{2zx^{2z-1}}{\Ga(1+z)} =\int_0^\infty dz\,\frac{2x^{2z-1}}{\Ga(z)} \\ =2\int_{-1}^\infty dt\,\frac{x^{2t+1}}{\Ga(1+t)} =2x\int_{-1}^\infty dz\,\frac{x^{2z}}{\Ga(1+z)}.$$ So (and this is the key), $g$ is a solution of the ODE $$g'(x)=2xg(x)+h(x),$$ where $$h(x):=2x\int_{-1}^0 dz\,\frac{x^{2z}}{\Ga(1+z)}.$$ Also, $g(0)=0$. So, $$g(x)=e^{x^2}G(x), \tag{2}\label{2} $$ where $$G(x):=\int_0^x du\, e^{-u^2}h(u). \tag{3}\label{3} $$ As $x\to\infty$, $$G(x)\to\int_0^\infty du\, e^{-u^2}h(u) =\int_0^\infty du\, e^{-u^2}2u\int_{-1}^0 dz\,\frac{u^{2z}}{\Ga(1+z)} \\ =\int_0^\infty dv\, e^{-v}\int_{-1}^0 dz\,\frac{v^z}{\Ga(1+z)} =\int_{-1}^0 \frac{dz}{\Ga(1+z)}\int_0^\infty dv\, v^z e^{-v} \\ =\int_{-1}^0 \frac{dz}{\Ga(1+z)}\,\Ga(1+z)=1. $$ Now \eqref{1} follows from \eqref{2}. $\quad\Box$

Remark: Note that $-\Ga'(1)$ is Euler's $\gamma>0$. Therefore and because $\Ga$ is log convex on $(0,\infty)$, we see that $\Ga$ is decreasing on $(0,1]$. So, for $x\ge1$ we have $$h(x)<\frac{2x}{\Ga(1)}\int_{-1}^0 dz\,x^{2z}<2x.$$ So, by \eqref{2}--\eqref{3}, for $x\ge1$ the relative error of the asymptotic approximation \eqref{1} is $$0<R(x):=\int_x^\infty du\, e^{-u^2}h(u) <\int_x^\infty du\, e^{-u^2}2u=e^{-x^2},$$ which goes to $0$ very fast as $x\to\infty$.

This bound on the relative error seems hard (if at all possible) to get by the Laplace method/Watson lemma applied to the original integral expression \eqref{1.5} of $g(x)$, which gives an asymptotic expansion in integral powers of $x$.

However, one can apply the Watson lemma to the expression \eqref{2} of $g(x)$ to get the following asymptotic expansion of the absolute error of approximation \eqref{1}, in integral powers of $\ln x$: $$E(x):=e^{x^2}-g(x) =e^{x^2}\int_x^\infty du\, e^{-u^2}h(u) \sim\sum_{k\ge0}\frac{(-1)^k c_k}{(2\ln x)^{k+1}} \tag{4}\label{4}$$ as $x\to\infty$, where $c_k:=\frac{d^k}{dz^k}\frac1{\Ga(1+z)}\big|_{z=0}$, so that $c_0=1$. $\quad\Box$

For an illustration, below are the graphs $\{(x,g(x)/e^{x^2})\colon0\le x\le 4\}$ (solid black), $\{(x,E(x)/E_0(x))\colon5\le x\le100\}$ (solid red), $\{(x,E(x)/E_1(x))\colon5\le x\le100\}$ (solid green), and $\{(x,E(x)/E_2(x))\colon5\le x\le100\}$ (solid blue), where (cf. \eqref{4}) $E_m(x):=\sum_{k=0}^m\frac{(-1)^k c_k}{(2\ln x)^{k+1}}$:

enter image description here

enter image description here


The saddle point of the integrand is at $z=x^2$ for large $x$, this gives the saddle point approximation $$\int_0^\infty \frac{x^{2z}}{\Gamma(1+z)}\,dz\rightarrow \frac{x^{2x^2}}{\Gamma(1+x^2)}.$$ You may then obtain higher order terms by performing the gaussian integral around the saddle point,$^\ast$ which gives a pre-exponential factor $\sqrt{2\pi}\,x$. Substituting also the large-$x$ asymptotics of the $\Gamma$ function, I arrive at $$\int_0^\infty \frac{x^{2z}}{\Gamma(1+z)}\,dz\rightarrow e^{x^2}.$$

The $e^{x^2}$ approximation is already quite good for relatively small values of $x$; see the plot (blue curve = exact, orange curve = approximation):

$^\ast$ The integrand $e^{F(z)}$ with $F((z)=2z\ln x-\ln\Gamma(1+z)$ is expanded to second order about the saddle point $z_0$, where the first derivative vanishes, $F'(z_0)=0$. For $x\gg 1$ one has $z_0=x^2$ and $$F(z)\approx2z_0\ln x-\ln\Gamma(1+z_0)-\tfrac{1}{2}x^{-2}(z-z_0)^2.$$ The integral $\int_{-\infty}^\infty e^{F(z)}\,dz$ then gives the asymptotic approximation $e^{x^2}$.

  • 1
    $\begingroup$ Re: The discriminatory power of graphical comparisons, have you tried plotting $e^{x^2} $ or $x^{-1} e^{x^2} $ just for fun as well? $\endgroup$ Dec 31, 2022 at 14:54
  • 1
    $\begingroup$ Incidentally, I cross posted the question to math.stack.exchange (with the link to the post now in my mathoverflow post), and an answer there finds the asymptotics to be purely $e^{x^2}$. $\endgroup$
    – Dispersion
    Dec 31, 2022 at 22:19
  • $\begingroup$ What tool do you use for plotting? $\endgroup$ Dec 31, 2022 at 22:29
  • $\begingroup$ the plots are Mathematica output... $\endgroup$ Dec 31, 2022 at 22:53
  • 3
    $\begingroup$ Based on your Taylor expansion I would expect a prefactor of $\sqrt{2\pi} x$ rather than $\sqrt{2\pi x}$, which should explain the discrepancy with the other answers. $\endgroup$
    – Terry Tao
    Jan 1, 2023 at 2:29

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.