15
$\begingroup$

Mathematica obtains

$$\int_0^{\infty}\frac{x}{\sinh(\pi x)}\ln(1+x^2) \ \mathrm{d}x=-\frac{3}{4}+\frac{\ln 2}{6}+9\ln(A)-\frac{\ln \pi}{2}$$ where $A$ is the Glaisher-Kinkelin constant.

A numerical approximation of the integral strongly suggests that this is incorrect. What is the correct value?

$\endgroup$
3
  • $\begingroup$ I'm curious, how do you know it's incorrect? $\endgroup$ – MasterOfBinary Nov 10 '13 at 23:33
  • $\begingroup$ Try to evaluate the integral numerically using Maple or Mathematics. $\endgroup$ – Mhenni Benghorbal Nov 10 '13 at 23:38
  • $\begingroup$ This expression seems to be around 1.03 and the value of the integral around 0.08, not even close! $\endgroup$ – Bennett Gardiner Nov 11 '13 at 0:02
14
$\begingroup$

Cody's answer gave me the idea to look at $\displaystyle \int_{0}^{\infty} \frac{\sin ( s \arctan t)}{(1+t^{2})^{s/2} \sinh \pi t} \, dt $.

First add the restriction $ \text{Re}(s) >1$.

Then

$$ \begin{align} &\int_{0}^{\infty} \frac{\sin ( s \arctan t)}{(1+t^{2})^{s/2} \sinh \pi t} \, dt \\ &= \frac{1}{2}\int_{-\infty}^{\infty} \frac{\sin(s \arctan t)}{(1+t^{2})^{s/2} \sinh \pi t} \, dt \\&= \frac{1}{2} \int_{-\infty}^{\infty} \text{Im} \ \frac{1}{(1-it)^{s} \sinh \pi t} \, dt \\ &= \frac{1}{2} \, \text{Im} \, \text{PV} \int_{-\infty}^{\infty} \frac{1}{(1-it)^{s} \sinh \pi t} \, dt \\ &= \frac{1}{2} \, \text{Im} \left(\pi i \ \text{Res}\left[\frac{1}{(1-iz)^{s} \sinh \pi z},0\right] + 2 \pi i \sum_{n=1}^{\infty} \text{Res} \left[\frac{1}{(1-iz)^{s} \sinh \pi z},in \right] \right) \\ &= \frac{1}{2} \text{Im} \left(\pi i \left(\frac{1}{\pi} \right) + 2 \pi i \sum_{n=1}^{\infty} \frac{(-1)^{n}}{\pi (1+n)^{s}} \right) \\ &= \frac{1}{2} \, \text{Im} \Big(i + 2i \big( (1-2^{1-s}) \zeta(s)-1\big) \Big) \\& = (1-2^{1-s})\zeta(s) - \frac{1}{2}. \end{align}$$

By analytic continuation, the result is valid for all complex values of $s$.

Differentiating under the integral sign and letting $s=-1$ we get

$$ \begin{align} \frac{1}{2} \int_{0}^{\infty} \frac{t \log(1+t^{2})}{\sinh \pi t} \, dt + \int_{0}^{\infty} \frac{\arctan t}{\sinh \pi t} \, dt &= 2^{1-s} \log 2 \ \zeta(s) + (1-2^{1-s}) \zeta'(s) \Bigg|_{s=-1} \\ &= 4 \log (2) \left(-\frac{1}{12} \right) - 3 \zeta'(-1) \\ &= - \frac{\log 2}{3} - 3 \zeta'(-1) . \end{align}$$

So we need to evaluate $ \displaystyle \int_{0}^{\infty} \frac{\arctan t}{\sinh \pi t} \, dt $.

Let $ \displaystyle I(z) = \int_{0}^{\infty} \frac{\arctan \frac{x}{z}}{\sinh \pi x} \, dx $.

Then

$$ \begin{align} I(z) &= \int_{0}^{\infty} \frac{1}{\sinh \pi x} \int_{0}^{\infty} \frac{\sin tx}{t} e^{-zt} \, dt \, dx \\ &= \int_{0}^{\infty} \frac{e^{-zt}}{t} \int_{0}^{\infty} \frac{\sin t x}{\sinh \pi x} \, dx \, dt \\ &= \frac{1}{2} \int_{0}^{\infty} \frac{e^{-zt}}{t} \tanh \left( \frac{t}{2} \right) \, dt . \end{align} $$

Now differentiate under the integral sign.

$$ \begin{align} I'(z) &= - \frac{1}{2} \int_{0}^{\infty} \tanh \left(\frac{t}{2} \right) e^{-zt} \, dt \\ &= - \frac{1}{2} \int_{0}^{\infty} \frac{1-e^{-t}}{1+e^{-t}} e^{-zt} \, dt \\ &= - \frac{1}{2} \int_{0}^{\infty} \left( -1 + 2 \sum_{n=0}^{\infty} (-1)^{n} e^{-tn}\right) e^{-zt} \, dt \\ &= \frac{1}{2z} - \sum_{n=0}^{\infty} (-1)^{n} \frac{1}{z+n} \\ &= \frac{1}{2z} - \frac{1}{2} \psi \left(\frac{z+1}{2} \right) + \frac{1}{2} \psi \left( \frac{z}{2} \right) \tag{1} \end{align}$$

Integrating back and using Stirling's formula to determine the value of the constant of integration, we find

$$ I(z) = \frac{\ln z}{2} - \log \Gamma \left(\frac{z+1}{2} \right) + \log \Gamma \left(\frac{z}{2} \right) - \frac{\log 2}{2} .$$

So

$$ \int_{0}^{\infty} \frac{\arctan x}{\sinh \pi x} \, dx = I(1) = \frac{\log \pi}{2} - \frac{\log 2}{2} .$$

Combining this result with the first result we have

$$ \begin{align} \int_{0}^{\infty} \frac{x \log(1+x^{2})}{\sinh \pi x} \, dx &= 2 \left(-\frac{\log 2}{3} - 3 \zeta'(-1) - \frac{\log \pi}{2} + \frac{\log 2}{2} \right) \\&=\frac{\log 2}{3} - \log \pi - 6 \zeta'(-1) \\ &= \frac{\log 2}{3} - \log \pi - \frac{1}{2} + 6 \log A \\ &\approx 0.0788460364 . \end{align}$$

$ $

$(1)$ http://mathworld.wolfram.com/DigammaFunction.html (5)

$\endgroup$
1
  • $\begingroup$ There ya' go, buddy. :):) Very nice indeed. $\endgroup$ – Cody Nov 13 '13 at 20:45
10
$\begingroup$

I'll conjecture that the correct value is : $$I:=\frac{\ln\;2}3-\ln\, \pi-6\,\zeta'(-1)$$ ($\zeta'(-1)=\frac 1{12}-\ln\,A\;$ with $A$ is the Glaisher-Kinkelin constant)

Let's obtain this solution using the Abel-Plana formula (a little as proposed by Cody) :

$$\sum_{n=0}^\infty (-1)^n\,f(n)=\frac{f(0)}2+i\int_0^\infty\frac{f(it)-f(-it)}{2\,\sinh(\pi\,t)}dt$$

Setting $\,f(z):=z\,\ln(1+z)\,$ returns quickly the equivalent sum $\,\displaystyle 2\sum_{k=1}^\infty (-1)^k\,(k-1)\,\ln(k)$. Unfortunately this sum is divergent ('regularization' returns the correct value) so let's try :

$\,\displaystyle f(z):=z\;\ln\left(1+\frac 1z\right)-1\,$ and get (observing that $\,\lim_{z\to 0^+} f(z)=-1$) :

\begin{align} S&:=-1+\sum_{n=1}^\infty (-1)^n\,\left(n\,\ln\left(1+\frac 1n\right)-1\right)\\ &=-\frac 12+i\int_0^\infty\frac{it\ln\left(1+\frac 1{it}\right)+it\ln\left(1+\frac 1{-it}\right)}{2\,\sinh(\pi\,t)}dt\\ &=-\frac 12-\frac 12\int_0^\infty\frac{t\ln\left(1+\frac 1{t^2}\right)}{\sinh(\pi\,t)}dt\\ &=-\frac 12-\frac 12\int_0^\infty\frac{t\ln(1+t^2)}{\sinh(\pi\,t)}dt+\int_0^\infty\frac{t\ln(t)}{\sinh(\pi\,t)}dt\\ &=-\frac 12(1+I)+J \end{align}

The integral $\,\displaystyle J:=\int_0^\infty\frac{t\ln(t)}{\sinh(\pi\,t)}dt\;$ may be evaluated using the integral representation of the $\eta$ function (see $(*)$) $\quad\displaystyle \frac{\eta(s)}{\sin(s\pi/2)}=-\int_0^\infty \frac{t^{-s}}{\sinh(\pi\,t)}dt\,$
by computing the derivative under the integral sign at $s=-1$ using $t^{-s}=e^{-s\ln(t)}\,$ to get $$J=\dfrac{\ln(2)}3+3\,\zeta'(-1)$$

while $I$ will be given by : $$I=-1-2(S-J)$$ with \begin{align} S&=-1+\sum_{n=1}^\infty (-1)^n\,\left(n\,\ln\left(1+\frac 1n\right)-1\right)\\ &=-1+\sum_{n=1}^\infty (-1)^n\,\left(-1+n\sum_{k=1}^\infty\frac{(-1)^{k-1}}{k\;n^k}\right)\\ &=-1+\sum_{n=1}^\infty (-1)^n\,\sum_{k=2}^\infty\frac{(-1)^{k-1}}{k\;n^{k-1}}\\ &=-1+\sum_{m=1}^\infty\frac{(-1)^m}{m+1}\sum_{n=1}^\infty \frac{(-1)^n}{n^m}\quad\text{for}\;m:=k-1\\ &=\frac{\ln(2)}2-1-\sum_{m=2}^\infty\frac{(-1)^m}{m+1}\left(1-2^{1-m}\right)\zeta(m)\quad(*)\\ &=\frac{\ln(2)}2-1+\sum_{m=2}^\infty\frac{\zeta(m)}{m+1}\left((-1)^{m+1}-4\left(\frac {-1}2\right)^{m+1}\right)\\ \end{align} $(*)$ : from the Dirichlet eta definition and property $\,\displaystyle \eta(m)=-\sum_{n=1}^\infty \frac{(-1)^n}{n^m}=\left(1-2^{1-m}\right)\zeta(m)$

But for $\psi$ the digamma function we have $\,\displaystyle\displaystyle \psi(1+x)=-\gamma-\sum_{m=2}^\infty \zeta(m)\;(-x)^{m-1}$
and $\,\displaystyle\int (\psi(1+x)+\gamma)x\,dx=\int \sum_{m=2}^\infty \zeta(m)\;(-x)^{m} dx=-\sum_{m=2}^\infty \frac{\zeta(m)}{m+1}\;(-x)^{m+1}$

If we observe that $\,\displaystyle\int \psi(x)\,x\,dx=[\ln(\Gamma(x))\,x]-\int\ln(\Gamma(x))\,dx$ we may use the formulae obtained by Glaisher, Gosper and Adamchik for the 'Negapolygammas' $(11)$ with $\zeta(s,q)$ the Hurwitz zeta function : $$K(q):=\int_0^q\ln(\Gamma(x))\,dx=\frac{(1-q)q}2+\frac q2\ln(2\pi)-\zeta'(-1)+\zeta'(-1,q)$$ with the special values $\,\displaystyle K(1)=\frac 12\ln(2\pi)\,$ and $\,\displaystyle K\left(\frac 12\right)=\frac 18+\frac 5{24}\ln(2)+\frac 14\ln(\pi)-\frac 32\zeta'(-1)$

obtaining (since $\,\psi(1+x)=\psi(x)+\frac 1x\,$) : $$\sum_{m=2}^\infty \frac{\zeta(m)}{m+1}\;(-x)^{m+1}=-x-\frac{\gamma}2x^2-\int \psi(x)\,dx=x-\frac{\gamma}2x^2+K(x)-\ln(\Gamma(x))\,x$$ getting (very laboriously I'll admit...) : \begin{align} S&=\frac{\ln(2)}2-1+\sum_{m=2}^\infty\frac{\zeta(m)}{m+1}\left((-1)^{m+1}-4\left(\frac {-1}2\right)^{m+1}\right)\\ &=\frac{\ln(2)}2-1+1+\frac 12\ln(2\pi)-4\left(\frac 18+\frac 5{24}\ln(2)+\frac 14\ln(\pi)-\frac 32\zeta'(-1)-\frac 12\ln\sqrt{\pi}\right)\\ &=-\frac 12+\frac 16\ln(2)+\frac {\ln(\pi)}2+6\zeta'(-1) \end{align} And finally $$I=-1-2\left(-\frac 12+\frac 16\ln(2)+\frac {\ln(\pi)}2+6\zeta'(-1)-\dfrac{\ln(2)}3-3\,\zeta'(-1)\right)\\ \boxed{\displaystyle I=\frac {\ln(2)}3-\ln(\pi)-6\,\zeta'(-1)}$$

Of course much more direct derivations exist but an interesting journey anyway!

$\endgroup$
8
  • 1
    $\begingroup$ could maybe be a comment? $\endgroup$ – Bennett Gardiner Nov 11 '13 at 0:08
  • 1
    $\begingroup$ @Bennett: Isn't the correct value the question? $\endgroup$ – Raymond Manzoni Nov 11 '13 at 0:09
  • $\begingroup$ I suppose. My apologies, I'm looking for a proof as well for my curiosity. $\endgroup$ – Bennett Gardiner Nov 11 '13 at 0:10
  • $\begingroup$ @Bennett: Of course a proof would be a much better answer ! $\endgroup$ – Raymond Manzoni Nov 11 '13 at 0:12
  • 1
    $\begingroup$ (+1) nice answer which answers the question in consideration. I believe that you put some effort to come up with the right answer. So, you deserve a credit for doing that. $\endgroup$ – Mhenni Benghorbal Nov 11 '13 at 4:01
6
$\begingroup$

If I may add something that may help.

It would appear this integral can be tied to the integrals of log-Gamma or Barnes G in some form or another.

If we write the $\displaystyle csch(\pi x)$ in terms of its exponential, it is then very similar to the known result that can be found by using Hermite's formula for the Hurwitz zeta.

Hermite's formula for $\zeta(s,1)$ is:

$\displaystyle \zeta(s,1)=1/2+\frac{1}{s-1}+2\int_{0}^{\infty}\frac{\sin(s\cdot \tan^{-1}(x))}{(x^{2}+1)^{s/2}(e^{2\pi x}-1)}dx$.

Which can be differentiated w.r.t s, leading to:

$\displaystyle \int_{0}^{\infty}\frac{x\log(x^2+1)}{e^{2\pi x}-1}dx=1/2\log(2\pi)-2/3-\log(A)$.

The integral in question is also equal to the series:

$\displaystyle \sum_{n=2}^{\infty}(-1)^{n+1}\log\left(1-\frac{1}{n}\right)+\sum_{n=2}^{\infty}(-1)^{n+1}n\cdot \log\left(1+\frac{1}{n-1}\right)$

I will look into it later when I find more time. But, if anyone can use these to derive the solution, please feel free.

$\endgroup$
1
  • $\begingroup$ Thanks @Cody ! I had found this one too as a consequence of the Abel-Plana formula but a $e^{\pi\,x}\;$ factor is still missing in the integral. Perhaps that for another $f$ function... (the derivation doesn't look easy). $\endgroup$ – Raymond Manzoni Nov 13 '13 at 16:58
1
$\begingroup$

Noting the tabulated cosine transforms of $\ln(1+1/x^2)$ and $x/\sinh(\pi x)$, by the Parseval relation for this transform $$\int_0^{\infty}\frac{x \ln(1+1/x^2)}{\sinh(\pi x)}dx=\int_0^{\infty}\frac{dy}{y}\frac{1+\sinh y -\cosh y}{1+\cosh y}$$ which has the known value $12\ln A-\ln \pi -1-\frac{1}{3}\ln 2$. It is also known that $$\int_0^{\infty}\frac{x \ln x^2}{\sinh(\pi x)}dx=\frac{1}{2}+\frac{2}{3}\ln 2-6\ln A =.$$ By combining these results one gets the desired value. This points to a disturbing bug in Mathematica.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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