2

Sometimes I come across integrals that seem like they should have a nice asymptotic expansion, but I get nervous when the terms of the expansion seem to fail to converge. The following is an example integral that captures the features I want to better understand (variants of this integral appear in the radiation problem for 2d electrodynamics).

Consider the integral $$I(r) = \int_{0}^{\infty} \frac{\cos(x)}{\sqrt{x+r}} dx$$ as $r \to \infty$. I want to consider an asymptotic expansion in powers of $\sqrt{\frac{1}{r}}$.

What is the asymptotic expansion? What is the "leading" non-analytic piece?


Here are my thoughts. I'll play fast and loose: I'm Taylor expanding past radii of convergence, exchanging integrals and sums, and renormalizing non-convergent integrals.

First, if I attempt to naively expand the denominator, there is trouble already at the lowest nontrivial order:

$$I(r) \sim \sum_{n=0}^{\infty} \frac{1}{r^{n+\frac{1}{2}}} (-1)^n \frac{\binom{2n}{n}}{2^{2n}} \int_{0}^{\infty} \cos(x) x^{n} dx$$

This makes me want to regularize the integrals, such as considering $\lim_{\epsilon \to 0^+} \int_{0}^{\infty} \cos(x) x^{n} e^{-\epsilon x} dx $. This limit results in $(-1)^{m+1} (2m+1)!$ for $n=2m+1$ is and $0$ for even $n$.

This results in

$$I(r) \sim \sum_{m=0}^{\infty} \frac{1}{r^{2m+\frac{3}{2}}} (-1)^m \frac{\binom{4m+2}{2m+1}}{2^{4m+2}} (2m+1)!$$

If I naively go by "the smallest term in the sum sets the scale of the hyperasymptotics," then the leading non-analytic piece would appear to be $e^{-r}=e^{-1/(1/r)}$, dropping many factors. However, this almost seems too good to be true, since this is quite a small piece.

It's possible there are typos above; please let me know if there are any algebra mistakes.

user196574
  • 2,294
  • 3
    Try to integrate by parts several times - integrate cos(x) and differentiate the root function. – Pavel R. May 12 '24 at 22:34
  • What do you mean by "non-analytic piece"? The Stokes lines closest to $\arg r=0$ are $\arg r=-\frac{\pi}{2}$ and $\arg r=\frac{3\pi}{2}$. – Gary May 13 '24 at 01:22
  • @Gary Very roughly: $f(t) = \int_{0}^\infty \frac{e^{-x}}{x+t}$ has an asymptotic expansion at large $t$. The part of $f(t)$ that is not captured by the expansion in powers of $t$ is approximately bounded by $e^{-t}$. This term $e^{-t}$ is $e^{-1/(1/t)}$, which can't be expanded in powers of $1/t$, and so it's invisible to the asymptotic expansion. I view it as setting the error in the asymptotic series. Some of this is a bit handwavy, and I might want to try to make it more rigorous. – user196574 May 13 '24 at 01:27
  • @PavelR. Very nice! I'm embarrassed to have forgotten to try that. That's exactly what I was missing, and it makes everything easier to think about rather than my uncontrolled series and limits above. I'll give it a shot and see if I get the same thing or something new. – user196574 May 13 '24 at 01:31
  • I started from scratch and hav the same result as you (replace the bnomial coefficient by the gamma function). – Claude Leibovici May 13 '24 at 08:17

4 Answers4

1

Performing the change of variables from $x$ to $t$ via $t=rx$ in the integral representation provided by MathArt, we obtain: $$\tag{1}\label{1} I(r) = \frac{1}{{\sqrt \pi }}\int_0^{ + \infty } {\frac{{{\rm e}^{ - rx} x^{1/2} }}{{1 + x^2 }}{\rm d}x} = \frac{1}{{\sqrt \pi }}\frac{1}{{r^{3/2} }}\int_0^{ + \infty } {\frac{{{\rm e}^{ - t} t^{1/2} }}{{1 + (t/r)^2 }}{\rm d}t} $$ provided $|\arg r|<\frac{\pi}{2}$. For any $N\ge 0$, $t>0$ and $r$ with $|\arg r|<\frac{\pi}{2}$, we substitute the identity $$ \frac{1}{{1 + (t/r)^2 }} = \sum\limits_{n = 0}^{N - 1} {t^{2n} \frac{{( - 1)^n }}{{r^{2n} }}} + \frac{{( - 1)^N }}{{r^{2N} }}\frac{{t^{2N} }}{{1 + (t/r)^2 }} $$ into \eqref{1}. Integrating term-by-term, we get $$ I(r) = \frac{1}{{r^{3/2} }}\left( {\sum\limits_{n = 0}^{N - 1} {( - 1)^n \frac{{\left( {\frac{1}{2}} \right)_{2n + 1} }}{{r^{2n} }}} +R_N (r) } \right) $$ for any $N\ge0$ and $r$ with $|\arg r|<\frac{\pi}{2}$. Here $(w)_p=\Gamma(w+p)/\Gamma(w)$ is the Pochhammer symbol. The remainder term $R_N(r)$ is given by $$\tag{2}\label{2} R_N (r) = \frac{{( - 1)^N }}{{r^{2N} }}\frac{1}{{\sqrt \pi }}\int_0^{ + \infty } {\frac{{{\rm e}^{ - t} t^{2N + 1/2} }}{{1 + (t/r)^2 }}{\rm d}t} = ( - 1)^N \frac{{\left( {\frac{1}{2}} \right)_{2N + 1} }}{{r^{2N} }}\Pi _{2N + 3/2} (r), $$ where $\Pi _p (w)$ is one of Dingle's basic terminants: $$ \Pi _p (w) = \frac{1}{{\Gamma (p)}}\int_0^{ + \infty } {\frac{{{\rm e}^{ - t} t^{p - 1} }}{{1 + (t/w)^2 }}{\rm d}t} ,\quad |\arg w|<\frac{\pi}{2}, \quad p>0. $$ Through analytic continuation, the function $\Pi _p (w)$ and the final expression in \eqref{2} can be extended across the entire Riemann surface associated with the logarithm. The Stokes phenomenon associated with $I(r)$ is fully determined by the basic terminant. Specifically, since $$ \Pi _p (w) = \frac{1}{{\Gamma (p)}}\int_0^{ + \infty } {\frac{{{\rm e}^{ - t} t^{p - 1} }}{{1 + (t/w)^2 }}{\rm d}t} \pm \pi {\rm i}\frac{{{\rm e}^{ \mp \frac{\pi }{2}{\rm i}p} }}{{\Gamma (p)}}w^p {\rm e}^{ \pm {\rm i}w} $$ for $\frac{\pi }{2} < \pm \arg w < \frac{{3\pi }}{2}$, we deduce $$ r^{3/2} I(r) \sim \sum\limits_{n = 0}^\infty {( - 1)^n \frac{{\left( {\frac{1}{2}} \right)_{2n + 1} }}{{r^{2n} }}} +\begin{cases}0 & \text{if }\; |\arg r| < \frac{\pi}{2},\\[1ex] \frac{{\sqrt \pi }}{2}{\rm e}^{ \mp \frac{\pi }{4}{\rm i}} {\rm e}^{ - \left| r \right|} & \text{if }\; \arg r = \pm \frac{\pi }{2},\\[1ex] \sqrt \pi {\rm e}^{ \mp \frac{\pi }{4}{\rm i}} {\rm e}^{ \pm {\rm i}r} & \text{if }\; \frac{{\pi }}{2} < \pm\arg r < \frac{3\pi }{2}, \end{cases} $$ as $|r|\to+\infty$. From the bounds for $\Pi _p (w)$ given in this paper, we immediately obtain $$ \left| {R_N (r)} \right| \le \frac{{\left( {\frac{1}{2}} \right)_{2N + 1} }}{{\left| r \right|^{2N} }} \times \begin{cases} 1 & \text{ if } \; |\arg r| \le \frac{\pi}{4}, \\[1ex] \min\! \left(|\csc ( 2\arg r)|,\frac{1}{2}\chi\!\left(2N+\frac{3}{2}\right)+1\right) & \text{ if } \; \frac{\pi}{4} < |\arg r| \le \frac{{\pi }}{2}, \\[1ex] \cfrac{\sqrt {2\pi \left( 2N + \frac{3}{2} \right)} }{2\left| \sin (\arg r)\right|^{2N + 3/2} } + \frac{1}{2}\chi\!\left(2N+\frac{3}{2}\right)+1 & \text{ if } \; \frac{\pi}{2} < |\arg r| < \pi . \end{cases} $$ Here $\chi (w) = \sqrt \pi \Gamma \left( {\frac{w}{2} + 1} \right)/\Gamma \left( {\frac{w}{2} + \frac{1}{2}} \right)$ for $w>0$.

Gary
  • 36,640
  • 1
    Thanks for the answer. This is really neat with the focus on understanding the asymptotic expansion throughout the complex plane. I actually hadn't fully appreciated Stokes phenomena before your comment and post (and I think my understanding is still weak, but now I know what I don't know!). – user196574 May 13 '24 at 06:21
  • @user196574 I revised and corrected my answer. – Gary May 14 '24 at 00:14
1

This will not be an answer but might be useful.

Your original integral can be transformed into one form with decaying factor of $e^{-rx}$ by Laplace transform

\begin{align} I(r)&=\int_{0}^{\infty} \frac{\cos(x)}{\sqrt{x+r}}\mathrm dx\\ &=\int_{0}^{\infty}\mathcal{L}[\cos(t)](x)\mathcal{L}^{-1}\left[\frac{1}{\sqrt{s+r}}\right](x)\mathrm dx\\ &=\int_{0}^{\infty}\frac{x}{x^2+1}\frac{e^{-rx}}{\sqrt{\pi x}}\mathrm dx\\ &=\frac{1}{\sqrt{\pi}}\int_{0}^{\infty}e^{-rx}\frac{\sqrt{x}}{x^2+1}\mathrm dx. \end{align} The last integral is evaluated here by Wolfram.

MathArt
  • 1,568
  • This is interesting! I think I'm not familiar with the method of replacing one function in the product with a Laplace transform and the other with an inverse Laplace transform. What is this technique called? – user196574 May 13 '24 at 15:59
  • 1
    This post may help you :-) https://math.stackexchange.com/a/4888909/319307 – MathArt May 13 '24 at 17:14
0

Here's a partial answer to my question.

As Pavel R. pointed out, integration by parts gives a nice way to generate the asymptotic series.

$$\int_0^\infty \frac{\cos(x)}{\sqrt{x+r}} dx \sim \frac{\sin(x)}{(x+r)^{1/2}}\bigg|_{0}^\infty - \frac{1}{2}\frac{\cos(x)}{(x+r)^{3/2}}\bigg|_{0}^\infty - \frac{1}{2} \frac{3}{2} \frac{\sin(x)}{(x+r)^{3/2}}\bigg|_{0}^\infty + \frac{1}{2} \frac{3}{2} \frac{5}{2} \frac{\cos(x)}{(x+r)^{5/2}}\bigg|_{0}^\infty + \, ...$$

We can quickly see the result by noting that every other contribution to the series vanishes by virtue of $\sin(0)=0$, and alternates by virtue of the double integral of $\cos(x)$ being $-\cos(x)$.

That is, we have

$$\int_0^\infty \frac{\cos(x)}{\sqrt{x+r}} \sim \sum_{m=0}^{\infty} \frac{1}{r^{2m+\frac{3}{2}}}(-1)^m \frac{(4m+1)!!}{2^{2m+1}}$$

The first couple terms are $\frac{1}{2}\frac{1}{r^{3/2}}-\frac{15}{8} \frac{1}{r^{7/2}} + \, ...$.

It's straightforward to verify that this matches the series found in the original post via unorthodox means! The integration by parts is more straightforward and much more sensible, so thanks Pavel!


Integration by parts also allows us to find bounds on the error term. The error term after truncating the sum at $\frac{1}{r^{2m+\frac{3}{2}}}$

$$E_{2m+\frac{3}{2}} = (-1)^m \frac{(4m+3)!!}{2^{2m+2}} \int_0^\infty \frac{\cos(x)}{(x+r)^{2m+5/2}}$$

I'll stop here, but I'll note that the error term where $m$ is order $r$ appear to be exponentially bounded above by $r$, pointing to the $e^{-r}$-style leading "non-analytic-in-$\frac{1}{r}$" piece.

user196574
  • 2,294
0

Since $$\int \frac{e^{i x}}{\sqrt{x+r}}\,dx=\sqrt[4]{-1} \,e^{-i r} \,\,\Gamma\left(\frac{1}{2},-i (r+x)\right)$$ $$\sqrt{\frac{2}{\pi }}\int_0^\infty \frac{e^{i x}}{\sqrt{x+r}}\,dx=\left(1-2 C\left(\sqrt{\frac{2r}{\pi }} \right)\right) \cos (r)+$$ $$\left(1-2 S\left(\sqrt{\frac{2}{\pi }} \sqrt{r}\right)\right) \sin(r)$$ where appear Fresnel integrals.

Asymptotically $$\color{blue}{I=\int_0^\infty \frac{\cos(x)}{\sqrt{x+r}}\,dx= \frac{1}{\sqrt{\pi }}\sum_{n=0}^\infty (-1)^n\,\frac{\Gamma \left(2 n+\frac{3}{2}\right) } {r^{2 n+\frac{3}{2}} }}$$