20

EDIT

After a long search, it seems that the solution is related to Jacobi elliptical functions.

For instance, if we try using this itegral representation

$$\boxed{\frac{1}{m^2+ n^2}=\int_{0}^{\infty}e^{-(m^2+n^2)t}dt}$$

We end up with sums of the following form, which as far as I found, are related to Jacobi elliptical functions.

$$\sum_{m=1}^{\infty}e^{-m^2t}$$

Another strategy would be starting with the following identity

$$\boxed{\sum_{n=1}^{\infty}\frac{(-1)^n}{m^2+n^2}=\frac{\pi \mathrm{csch}(\pi m)}{2m}-\frac{1}{2m^2}}$$

Also leads to evaluate a sum of the form below, which also belongs to the elliptical functions family

$$\sum_{m=1}^{\infty}\frac{1}{m \sinh(\pi m)}$$

Even the infinite product $$\prod_{k=1}^{\infty} \Big(1+e^{-\pi k2} \Big)$$

seems to be related to these special functions!


First attempt

$$\boxed{\sum_{n=1}^{\infty}\sum_{m=1}^{\infty}\frac{(-1)^{n-1}}{n^2+m^2}=\frac{\pi^2}{24}+\frac{\pi \ln(2)}{8}}$$

Consider

$$S=\sum_{n,m=1}^{\infty}\frac{(-1)^{n-1}}{n^2+m^2}$$

$$S=-\sum_{n=1}^{\infty}(-1)^n\sum_{m=1}^{\infty}\frac{1}{n^2+m^2}$$

Recall

$$\boxed{\sum_{m=1}^{\infty}\frac{1}{n^2+m^2}=\frac{\pi \coth(\pi n)}{2n}-\frac{1}{2n^2}}$$

$$S=-\sum_{n=1}^{\infty}(-1)^n\frac{\pi \coth(\pi n)}{2n}-\frac{1}{2}\sum_{n=1}^{\infty}\frac{(-1)^{n-1}}{n^2}$$

$$S=-\frac{\pi}{2}\sum_{n=1}^{\infty}(-1)^n\frac{ \coth(\pi n)}{n}-\frac{\pi^2}{24}$$

Also

$$\boxed{\coth(\pi n)-1=\frac{2}{e^{2 \pi n}-1}}$$

$$S=-\frac{\pi}{2}\sum_{n=1}^{\infty}\frac{(-1)^n}{n} \bigg\{\frac{2}{e^{2 \pi n}-1}+1 \bigg\}-\frac{\pi^2}{24}$$

$$S=-{\pi}\sum_{n=1}^{\infty}\frac{(-1)^n}{n} \bigg\{\frac{1}{e^{2 \pi n}-1}\bigg\}-\frac{\pi}{2}\sum_{n=1}^{\infty}\frac{(-1)^n}{n}-\frac{\pi^2}{24}$$

$$S=-{\pi}\sum_{n=1}^{\infty}\frac{(-1)^n}{n} \bigg\{\frac{1}{e^{2 \pi n}-1}\bigg\}+\frac{\pi \ln(2)}{2}-\frac{\pi^2}{24}$$

$$S=-{\pi}\sum_{n=1}^{\infty}\frac{(-1)^ne^{-2 \pi n}}{n}\sum_{k=0}^{\infty}e^{-2 \pi n k} +\frac{\pi \ln(2)}{2}-\frac{\pi^2}{24}$$

$$S=-\pi\sum_{k=0}^{\infty}\sum_{n=1}^{\infty}\frac{(-1)^ne^{-2 \pi n(k+1)}}{n}+\frac{\pi \ln(2)}{2}-\frac{\pi^2}{24}$$

$$S=\pi\sum_{k=1}^{\infty}\ln(1+e^{-2\pi k})+\frac{\pi \ln(2)}{2}-\frac{\pi^2}{24}$$

$$S=\pi\ln(\prod_{k=1}^{\infty}1+\big(e^{-\pi k}\big)^2)+\frac{\pi \ln(2)}{2}-\frac{\pi^2}{24}$$ From this point on, I can´t finish. Any help is welcome.


Second attempt

I tried the method suggest bellow by @NoName, but I still missing one factor $\frac{1}{4}$ before the $\log(2)$. I suspect that this method fails because $\sum_{m=1}^{\infty}\frac{1}{m}\sin(mt)=\frac{\pi}{2}-\frac{t}{2}$ is only valid betweem $0$ and $2 \pi$.

$$S=\sum_{n=1}^{\infty}(-1)^{n+1}\sum_{m=1}^{\infty}\frac{1}{n^2+m^2}$$

rewrite

$$\frac{1}{n^2+m^2}=\frac{1}{m}\int_{0}^{\infty}\sin(mt)e^{-nt}dt$$

$$S=\sum_{n=1}^{\infty}(-1)^{n+1}\sum_{m=1}^{\infty}\frac{1}{m}\int_{0}^{\infty}\sin(mt)e^{-nt}dt$$

$$S=\int_{0}^{\infty}\sum_{n=1}^{\infty}(-1)^{n+1}e^{-nt} \Big\{\sum_{m=1}^{\infty}\frac{1}{m}\sin(mt) \Big \}dt$$

$$S=\int_{0}^{\infty}\sum_{n=1}^{\infty}(-1)^{n+1}e^{-nt} \Big\{ \frac{\pi}{2}-\frac{t}{2} \Big \}dt$$

$$S=\sum_{n=1}^{\infty}(-1)^{n+1}\int_{0}^{\infty}e^{-nt} \Big\{ \frac{\pi}{2}-\frac{t}{2} \Big \}dt$$

$$S=\sum_{n=1}^{\infty}(-1)^{n+1} \Big\{ \frac{\pi}{2n}-\frac{1}{2n^2}\Big \}$$

$$S=\frac{\pi}{2}\sum_{n=1}^{\infty} \frac{(-1)^{n+1}}{n}-\frac{1}{2}\sum_{n=1}^{\infty} \frac{(-1)^{n+1}}{n^2}$$

$$S=\frac{\pi \log(2)}{2}+ \frac{\pi^2}{24}$$

Nero
  • 3,769
Ricardo770
  • 2,761
  • 1
    After writing $\frac{1}{e^{2\pi n} -1}$ as an infinite geometric series and changing the order of summation, it only remains to calculate $\prod_{k=1}^{\infty} (1+e^{-2\pi k} )$. – Vishu Apr 21 '21 at 17:59
  • One can show that $$\sum_{n\ge1}\frac{(-1)^n}{n(e^{2\pi n}-1)}=\frac1{2\pi i}\int_{c-i\infty}^{c+i\infty}\frac{2^{-s}-1}{(2\pi)^s}\Gamma(s)\zeta(s)\zeta(s+1)ds,$$ but I don't know how to evaluate the integral – clathratus Apr 21 '21 at 19:59
  • $\sum (-1)^n /n = - \ln 2$. The sign of $(\ln 2)/2$ has to be reversed. – Sungjin Kim Apr 21 '21 at 21:00
  • fixed, thank you – Ricardo770 Apr 21 '21 at 21:03
  • $\sum \frac{\sin nt}n = \frac{\pi - t}2$ is for $t\in (0,2\pi)$. Then the sum becomes $2\pi$-periodic. – Sungjin Kim Apr 22 '21 at 13:12
  • I tried $$\frac{1}{n}\int_{0}^{\infty}e^{-nt}\cos(mt)dt=\frac{1}{m^2+n^2}$$ instead. But it leads to the integral of $$\int_{0}^{\infty}\cos(mt)\log(1+e^{-t})dt$$ ultimeately which gives hyperbolic cosecant and takes me back to the same place as the first method. Leads to a loop. – Ricardo770 Apr 22 '21 at 13:16
  • 1
    Even if you use that $\sum \frac{\sin(mt)}{m}=\pi\left\lfloor \frac{t}{2\pi}\right\rfloor+\frac{\pi-x}{2}$, you end up needing to find either $\sum_{m=1}^{\infty}\ln\left(1+e^{-2m\pi}\right)$ or equivalently, $\sum_{k=1}^{\infty}\frac{\left(-1\right)^{k+1}}{k}\left(\frac{e^{-k2\pi}}{1-e^{-k2\pi}}\right)$. – Varun Vejalla Apr 22 '21 at 15:10
  • That´s exactly where the first method led to. – Ricardo770 Apr 22 '21 at 15:28
  • $\int_{0}^{\infty}e^{-nt} \Big{ \frac{\pi}{2}-\frac{t}{2} \Big }dt=\Big{ \frac{\pi}{2n}-\frac{1}{2n^2}\Big }$ doesn't make much sense, ${.}$ is the fractional part, not smooth in $n$ small in contrary to the integral. – reuns Apr 23 '21 at 01:21
  • 1
    @reuns I also pointed this out in the earlier comment, and OP seems to realize the mistake too. – Sungjin Kim Apr 23 '21 at 01:22
  • 2
    See a slightly more complicated sum of same kind : https://math.stackexchange.com/q/2469841/72031 and my answer there:https://math.stackexchange.com/a/2482542/72031 – Paramanand Singh Apr 23 '21 at 01:42
  • 1
    +1 for adding context!! People need to take a lesson from you. Compare this with the question I linked above which has almost no context. – Paramanand Singh Apr 23 '21 at 02:49

2 Answers2

7

Show that $$\lim_{s\to 1^+} (s-1)\sum_{n,m\ne 0,0} |n+im|^{-2s}=\lim_{s\to 1^+} (s-1)\sum_{n,m\ne 0,0} (n^2+m^2)^{-s}$$ $$= \lim_{s\to 1^+} (s-1) \int_{|x|>1,|y|>1} (x^2+y^2)^{-s}dxdy=\pi$$ From $(1-|1+i|^{-2s}) \sum_{n,m\ne 0,0} |n+im|^{-2s}$ $=\sum_{n,m, 2\ \nmid \ |n+im|^2} |n+im|^{-2s}$ $= 2\sum_{n\ne 0,m} |2n+i(2m+1)|^{-2s}$ you'll get for $s >1$

$$ \begin{eqnarray}F(s)&=&\sum_{n\ge 1,m\ge 1} (-1)^{n-1} (n^2+m^2)^{-s}\\ &=&\frac14\left(\sum_{n,m\ne 0,0} (-1)^{n-1} |n+im|^{-2s} -\sum_{n\ne 0}(-1)^{n-1} |n|^{-2s}+\sum_{m\ne 0}|m|^{-2s}\right) \\&=& \frac14(1- (1-2^{-s}) -2^{1-2s}) \sum_{n,m\ne 0,0} |n+im|^{-2s}+ 2^{-2s} \zeta(2s) \end{eqnarray}$$

$\zeta(2)=\pi^2/6$ will give $$\sum_{n\ge 1}\sum_{m\ge 1} (-1)^{n-1}(n^2+m^2)^{-1} =\lim_{s\to 1^+}F(s)=\frac{\pi}8\log 2 + \frac{\pi^2}{24}$$ where I'm using that $\sum_{n\ge 1,m\ge 1} ( (2n-1)^2+m^2)^{-1}- ((2n)^2+m^2)^{-1})$ is absolutely convergent to write the double series as $\lim_{s\to 1^+}F(s)$.

reuns
  • 77,999
  • Since the original sum is not absolutely convergent how can we be sure that the limit gives us the correct value? – user Apr 21 '21 at 19:40
  • 2
    Use that $\sum_{n,m} ( (2n-1)^2+m^2)^{-1}- ((2n)^2+m^2)^{-1})$ is absolutely convergent – reuns Apr 21 '21 at 19:46
  • It is quite interesting to see that OP's method will give the following identity: $$\sum_{n=1}^{\infty}\left(\frac1{n(e^{2\pi n }-1)}-\frac1{n(e^{4\pi n}-1)}\right) = -\frac{3\log 2}{8} + \frac{\pi}{12}.$$ – Sungjin Kim Apr 22 '21 at 00:47
  • @SungjinKim Yes but it is quite different to $\sum_{n=1}^{\infty}\frac1{n(e^{2\pi n }-1)}$, one is $\log \Delta(i)$ the other is $\log \Delta(i)/\Delta(2i)$, where $\Delta(z)/\Delta(2z)$ is algebraic (and radical) over $\Bbb{Q}[j(z)]$ – reuns Apr 22 '21 at 01:16
  • Just for curiosity, is there a closed form to $\sum_{n=1}^{\infty} \frac1{n(e^{2\pi n}-1)}$? – Sungjin Kim Apr 22 '21 at 01:27
  • @SungjinKim Yes, in term of $\Gamma(1/4)$, it follows from expressing $B(1/4,1/2)=\int_0^1 t^{-3/4}(1-t)^{-1/2}dt$ in term of the elliptic integral $\int_0^\infty \frac{dx}{\sqrt{4x^3-g_2(i)x}}dx$ which translates to $\int_0^{1/2} dz$ on the complex torus $\Bbb{C/(Z+iZ)}$, from which you know $g_2(i)$ and hence $\Delta(i)$. – reuns Apr 22 '21 at 01:37
  • 1
    Nice solution. I wonder whether we can find another solution using $$\displaystyle {\frac{1}{m^2+n^2}} = \frac{1}{m}{\int_{0}^{\infty} \sin(mt)e^{-nt};{\mathrm dt} }.$$

    i.e. $$\displaystyle S = \sum_{m=1}^{\infty}\sum_{n=1}^{\infty}\frac{(-1)^{n-1}}{m}\int_0^\infty \sin(mt)e^{-nt}, \mathrm dt $$

    Then maybe arguing for order switch and appealing to Fourier series? Do you think this is feasible?

    – NoName Apr 22 '21 at 02:10
  • 2
    @NoName I didn't check full details, but I think it will lead to the sum $\sum (1/n(e^{2\pi n} -1) - 1/n(e^{4\pi n}-1)$. – Sungjin Kim Apr 22 '21 at 02:26
  • 2
    @NoName: such sums were handled by Plouffe. See this thread. – Paramanand Singh Apr 23 '21 at 02:44
  • Just to confirm, this looks similar to the use of Kronecker limit formula? – Paramanand Singh Apr 23 '21 at 02:59
  • @ParamanandSingh I'm using only the first term $E(i,s) =\pi/(s-1)+O(1)$ and that $2^{1-s} E(2i,s)=(1-2^{-s})E(i,s) $ which is specific to $\tau=i$. No need of any theorem and special function. OP's Fourier expansion somewhat gives the second term of Kronecker limit formula for $E(i,s)-2^{1-s} E(2i,s)$. – reuns Apr 23 '21 at 12:23
4

I liked your first approach and the following answer completes it.

Let $q=e^{-\pi} $ and you wish to evaluate the product $$F(q)=\prod_{n=1}^{\infty} (1+q^{2n})\tag{1}$$ We can rewrite the above product as $$F(q) =\prod_{n\geq 1}\frac{1+q^n}{1+q^{2n-1}}=\prod_{n\geq 1}\frac{1-q^{2n}}{(1-q^n)(1+q^{2n-1})}=\prod_{n\geq 1}\frac{1}{(1-q^{2n-1})(1+q^{2n-1})} \tag{2}$$ Now we bring some elliptic integrals and a bit of Ramanujan into picture. Let $k$ be the elliptic modulus corresponding to nome $q$ and $K$ be the corresponding complete elliptic integral of first kind and $k'=\sqrt{1-k^2}$.

We have by definition of Ramanujan class invariant $$G_m=G(q) =2^{-1/4}q^{-1/24}\prod_{n\geq 1}(1+q^{2n-1})\tag{3}$$ and $$g_m=g(q) =2^{-1/4}q^{-1/24}\prod_{n\geq 1}(1-q^{2n-1}) \tag{4}$$ where $q=e^{-\pi\sqrt{m}} $. Here we have $q=e^{-\pi} $ so that $m=1$. And the class invariants defined above are also linked to the elliptic moduli $k, k'$ via $$G(q) =(2kk') ^{-1/12}, g_m=g(q)=(2k/k'^2)^{-1/12} \tag{5}$$ For $q=e^{-\pi} $ we have $k=k'=1/\sqrt{2}$ and thus $$G_1=G(q)=1, g_1=g(q)=2^{-1/8}\tag{6}$$ Now $$G(q) g(q) =2^{-1/2}q^{-1/12}/F(q)$$ and putting $q=e^{-\pi} $ and using $(6)$ we get $$F(q) = 2^{-3/8}e^{\pi/12}\tag{7}$$ Based on this your sum in question is $$S=\pi\log F(q) +\frac{\pi\log 2}{2}-\frac{\pi^2}{24}$$ which via $(7)$ equals $$S=\frac{\pi^2}{24}+\frac{\pi\log 2}{8}$$

  • Thank you very much for taking your time and provide an answer, $+1$! I confess you that the knowledge that lies in your answer is far beyond my knowledge so I can´t fully appreciate it. Nevertheless it completes the proof I iniated, and this is what I was looking for. For no reason, I was expecting a more elementary answer, but, maybe there is none... That´s the reason more powerfull tools are developed. I´ll wait a little bit more to see if someone comes up with an answer that utilizes a more elementary approach before I give it the "accept". Thank you again! – Ricardo770 Apr 23 '21 at 12:41
  • 2
    @Ricardo770: the topic of theta functions, elliptic integrals, elliptic functions is the most fascinating one and you should take time to study it. I instantly fell in love with it and wrote a whole blog mostly containing these ideas. You may have a look at the archives page:https://paramanands.blogspot.com/p/archives.html?m=0 I have tried to use elementary methods to present the theory because I am myself not familiar with the advanced theories. – Paramanand Singh Apr 23 '21 at 12:45
  • @ Paramanand Singh I will certainly take time to learn it. BTW, beautiful blog. I may start studying from there. – Ricardo770 Apr 23 '21 at 13:43