7
$\begingroup$

The integral is:

$$f(a) = \int\limits_{-\infty}^\infty \frac{x e^{-a^2 x^2}}{\tanh(x)}dx$$

which seems to converge for all $a>0$. But I don't know how to get a sense of the function $f(a)$ such as writing it as a convergent series. The usual Taylor series has infinity for each term. Any ideas?

Edit:

I believe (using answers below) that when $a$ is close to zero we have (doing a substitution):

$$f(a) = \frac{1}{a^2}\int\limits_{-\infty}^\infty \frac{x e^{-x^2}}{\tanh (x/a)}dx$$

But using $\tanh(x/a)\rightarrow\operatorname{sign}(x)$ as $a\rightarrow 0^+$. So the above should become: $$f(a) \approx \frac{2}{a^2}\int\limits_{0}^\infty x e^{-x^2}dx = \frac{1}{a^2}$$

So that gives the behaviour of $f(a)$ when $a$ is small. But I don't know how to give extra terms. (Also not sure if this is mathematically correct). From numerical calculations I find that near $0$, $$ f(a)\approx \frac{1}{a^2} + \frac{\pi^6}{6} - \frac{\pi^4 }{60}a^2+\frac{\pi^6 }{252}a^4+... = \frac{1}{a^2}\sum\limits_{n=0}^\infty \frac{B_{2n} a^{2n}\pi^{2n}}{n!} $$ although apparently this doesn't converge?

Comment

The answers below are two asymptotic series depending on whether $a$ is small or large. These give good approximations if we truncate the summation before begins to diverge. In the mid-range, when $a^2=1/\pi$, these two sums become term-by-term equal and the closest the truncated sum get to the true answer of $f(1/\sqrt{\pi})$ is to about 1% error. Using both these sums, we can know any value to within about 1%-2% error, and if $a$ is small or large then much more accurately.

$\endgroup$
2
  • 1
    $\begingroup$ Just from calculations it seems like a good approximation when $a$ is close to zero is $f(a)\approx 1.6 + 1/a^2$. (Just putting values like $a=1/1000$ into Wolfram alpha. $\endgroup$
    – zooby
    May 24, 2022 at 22:34
  • 3
    $\begingroup$ Your conjectured result is correct, up to a missing $\pi^{2n} $ - I'll write an answer. $\endgroup$ May 25, 2022 at 3:31

2 Answers 2

7
$\begingroup$

A Taylor series exists in powers of $1/a$: $$f(a) = \int\limits_{-\infty}^\infty \frac{x e^{-a^2 x^2}}{\tanh x}\,dx=a^{-2}\int\limits_{-\infty}^\infty x e^{-x^2}\,\text{cotanh}\, (x/a)\,dx$$ $$=\sum_{n=0}^{\infty}\frac{2^{2 n} B_{2 n} \,\Gamma \left(n+\frac{1}{2}\right)}{(2 n)!a^{2 n+1}},$$ with $B_{2n}$ the Bernoulli number.


To assess whether this asymptotic series is useful for $a\gtrsim 1$, below I plot the sum $\sum_{n=0}^{10}$, so the first eleven terms, for $a=1,2,3\ldots 10$ (blue data points), and compare with a numerical evaluation of the integral (blue curve).

$\endgroup$
7
  • 2
    $\begingroup$ Good stuff. Is it convergent? If range of convergence is $1/|a|<R$ for some $R$ it could mean that is doesn't converge near $a=0$. $\endgroup$
    – zooby
    May 24, 2022 at 21:25
  • 4
    $\begingroup$ Also in terms of series of $_1F_1$ hypergeometric functions $$f(a)=\sum_{n=0}^\infty\frac{2(1-2^{2n-1})\,B_{2n}}{(2n)!\,|a|^{2n+1}}\cdot\,_1F_1(n+1/2;1/2;1/(4a^2))$$ $\endgroup$ May 24, 2022 at 21:38
  • 7
    $\begingroup$ Alas the power series in powers of $1/a^2$ is only an asymptotic expansion at infinity; it does not converge for any $a$. [That is to be expected, because the series was obtained by integrating termwise the Taylor expansion of $\tanh x$ about $x=0$, and this Taylor expansion converges only for $|x| < \pi/2$ (the distance from $x=0$ to the nearest complex pole), whereas the integral is over all real $x$.] $\endgroup$ May 24, 2022 at 22:20
  • $\begingroup$ @JorgeZuniga Is that a convergent sum for all $a$? $\endgroup$
    – zooby
    May 26, 2022 at 17:09
  • $\begingroup$ I don't see how @JorgeZuniga 's series agrees with the integral; for example, for $a=5$ that sum gives 0.2006, while the correct value is 0.3568. $\endgroup$ May 26, 2022 at 17:24
12
$\begingroup$

To obtain the series in $a$, separate off the leading term proportional to $1/a^2 $ and expand the Gaussian instead of the hyperbolic cotangent: \begin{eqnarray} f(a)&=& 2\int_{0}^{\infty } dx\, x \ e^{-a^2 x^2 } \left[ \coth x -1 +1 \right] \\ &=& 2\int_{0}^{\infty } dx\, x \ e^{-a^2 x^2 }+ 2\sum_{n=0}^{\infty } \frac{(-a^2)^{n}}{n!} \int_{0}^{\infty} dx\, x^{2n+1} [\coth x -1] \\ &=& \frac{1}{a^2 } + 2\sum_{n=0}^{\infty } \frac{(-a^2)^{n}}{n!} \frac{ (-1)^n \pi^{2n+2} B_{2n+2} }{2n+2} \\ &=& \frac{1}{a^2} + \frac{1}{a^2} \sum_{n=1}^{\infty } \frac{(\pi a)^{2n} B_{2n} }{n!} \\ &=& \frac{1}{a^2} \sum_{n=0}^{\infty } \frac{(\pi a)^{2n} B_{2n} }{n!} \end{eqnarray} Also this is only an asymptotic expansion - the Bernoulli numbers diverge more rapidly than $n!$ at large $n$.

$\endgroup$
4
  • 1
    $\begingroup$ Wow. I take it the coth x integral is just one you have to know. I take it also it's an asymptotic sum as someone else mentioned. ( I tried it out numerically and it seems to diverge for all x after long enough - although that could just be machine errors). $\endgroup$
    – zooby
    May 25, 2022 at 3:56
  • 1
    $\begingroup$ Well, I suppose there's no reason an integral should have a representation as a convergent series, I guess. $\endgroup$
    – zooby
    May 25, 2022 at 4:05
  • 3
    $\begingroup$ @zooby - Since $\coth x -1 = 1/(e^{2x} -1)$, the integral is one of the pretty standard ones that lead to Bernoulli polynomials ... $\endgroup$ May 25, 2022 at 4:06
  • 2
    $\begingroup$ By truncating the expansion of $e^{-a^2x^2}$ with a remainder term having known bound, you should be able to do that for the final expansion too. Then you will have rigorous bounds on $f(a)$. $\endgroup$ May 25, 2022 at 4:10

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.