Title says it all.

Well, maybe some backstory.

Flipping through my past notebooks, I found this:

$$ \vdots $$ $$= \sum^\infty_{k=0} \frac{2(-1)^k}{(2k+1)\pi\cosh[(2k+1)\pi/2]}\\=\frac14\quad(?)$$

[end of page]

Ah, yes. My engineering numerical analysis professor gave me one homework problem:

Make a contour plot of the steady-state temperature distribution of a square plate, with the temperature of one of its side maintained at $1$ unit, the other three sides maintained at $0$ units.

Steady-state temperature distribution $T$ follows $\nabla^2T=0$.

I was calculating the temperature of the center of the plate analytically to sanity-check my program output, when I encountered this weird-looking sum. I remember I made a spreadsheet to numerically evaluate the sum, and Excel said 0.2500000... So I jot down $\frac14$ with a question mark beside it. The next page is filled with my fruitless attempts to find out the exact sum, to convince myself that the sum is exactly $1/4$.

I posted a bounty on my social media accounts, whoever can show the proof (or disproof) gets free beer.

After a few days, the bounty was still unclaimed, but I realized, "duuuuuhhhh! I could just exploit the symmetry. If I rotate the plate $90^\circ$ three times and superpose all the temperature distribution with the original one, temperature is $1$ unit everywhere. By linearity and symmetry, the temperature of the center point is therefore $1/4$."

I treated myself a few rounds of beer, and moved on. Case closed.

Looking again at this, I now realize that I might be missing out on some summation tactics I never knew.

Anybody can show me how is this equal to $1/4$ just by algebraic manipulation?

  • 5
    $\begingroup$ +1 for the story, +1 more for the solution, and +1 even more for trying to learn greater things. $\endgroup$ – Simply Beautiful Art Jan 12 '17 at 23:22
  • $\begingroup$ The residue theorem might come in useful. $\endgroup$ – Michael M Jan 13 '17 at 0:23
  • $\begingroup$ Thanks for the suggestion and the upvote. Probably will take me a while to work that out, as I never took a course in analysis. But now I realize that math is wonderful, there are always more beautiful things to learn. $\endgroup$ – f1garo Jan 13 '17 at 1:25
  • $\begingroup$ related math.stackexchange.com/questions/898955/… $\endgroup$ – Math-fun Jan 13 '17 at 9:49

I am busy as hell today but this is such a nice problem (and amusingly posed question) i can't resist... :)

The key idea here is simple: Look for a function in the complex plane $f(z)$ which has poles at the correct values $z_n$ and integrate it over an appropriate contour $C$.

It turns out that the function (which is also the simplest guess i can imagine)

$$ f(z)=\frac{1}{z \cos(z)\cosh(z)} $$

is the correct choice. Its poles are given by $z_0=0$, $z_k=\frac{(2 k-1)\pi}{2} $ and $\tilde{z}_k=i \frac{(2 k-1)\pi}{2} $ with $k \in \mathbb{Z/0}$. We furthermore have

$$ \text{res}(z_0)=1 \\ \text{res}(z_k)=\text{res}(\tilde{z}_k)=\frac{2}{\pi}\frac{(-1)^{k}}{(2k-1)\cosh\left(\pi\frac{2k-1}{2}\right)}\\ \text{res}(z_{-k})=\text{res}(z_k) $$

since $z\cos(z)\cosh(z)\sim |z| e^{a |z|}$ with $\Re(a)>0$ as $|z|\rightarrow\infty$ we can choose a big circle traversed anticlockwise with radius choosen such that we hit no pole of $f(z)$ as the integration contour $C$ (Thanks to @Dr. MV for rigorizing this point).

Then, we get by applying the residue theorem, we get in the limit of infinte radius

$$ \oint_Cf(z)dz=2\pi i\text{res}(z_0)+4\pi i \sum_{k\geq1}\text{res}(z_k)+4\pi i\sum_{k\geq1}\text{res}(\tilde{z}_k)=0 $$


$$ 8\pi i \sum_{k\geq1}\text{res}(z_k)=-2\pi i $$

which can be rewritten as (after shifting $k\rightarrow k+1$)

$$ \sum_{k\geq0}\frac{2}{\pi}\frac{(-1)^{k}}{(2k+1)\cosh\left(\pi\frac{2k+1}{2}\right)}=\frac{1}{4}\\ \textbf{Q.E.D.} $$

To make things clearer, here is a sketch of the integration contour $\color{blue}{C}$ and the singularities $\color{red}{z_n}$

enter image description here

  • 1
    $\begingroup$ (+1) This is extremely close to Ramanujan original argument, I'd say it is perfectly fine! $\endgroup$ – Jack D'Aurizio Jan 13 '17 at 12:51
  • 1
    $\begingroup$ @JackD'Aurizio thanks! This is a relief...still wondering how this complicated sum can be derived with nearly zero effort $\endgroup$ – tired Jan 13 '17 at 12:53
  • 1
    $\begingroup$ A big (+1). I was working on this last night, had identified $f(z)$ and then floundered to find a suitable contour. Now, I feel stupid. Well done my friend! -Mark $\endgroup$ – Mark Viola Jan 13 '17 at 14:42
  • 1
    $\begingroup$ I was trying to isolate the poles on the real axis by taking a rectangular contour in hopes that the top and bottom halves would cancel. They didn't. I even recognized the symmetry between the residues on real and imaginary axes. For some reason (it was getting late and I was tiring) I just couldn't see the obvious "big circle" contour was the way forward. One note; we must take the limit with discrete radii to ensure the contour doesn't hit a pole. $\endgroup$ – Mark Viola Jan 13 '17 at 14:55
  • 1
    $\begingroup$ In general, the function $\pi \sec(\pi z) f(z)$ can be used to evaluate infinite sums of the form $\sum_{n \in \mathbb{Z}} (-1)^n f(n+1/2)$, while the function $\pi \tan(\pi z) f(z)$ can be used to evaluate infinite sums of the form $\sum_{n \in \mathbb{Z}} f(n+1/2)$. In some cases, of course, the contour integral won't vanish in the limit. $\endgroup$ – Random Variable Jan 13 '17 at 19:56

As mentioned by Zucker in The summation of series of hyperbolic functions, in Question 358, J. Indian Math. Soc., 4 (1912), p. 78. Ramanujan proves (through the residue theorem) that $$ \sum_{n\geq 1}(-1)^{n+1}(2n-1)^{4m-1}\,\text{sech}\left[(2n-1)\frac{\pi}{2}\right] = 0 $$ holds for every $m\geq 0$. Your identity then follows from the $m=0$ case.

It is truly remarkable to know that such identity can be proved through Dirichlet's problem about the eigenvalues of the Laplacian operator. Can you be more specific about the relation between such a series and the physical problem?$^{(0)}$ A brilliant piece of math might arise from there.

$^{(0)}$Update: I found the connection - such series are related with the Green function of a square.

I will show an interesting technique for deriving a similar identity.

From the Weierstrass product $$ \cosh(\pi x/2) = \prod_{m\geq 0}\left(1+\frac{z^2}{(2m+1)^2}\right)\tag{1} $$ by applying $\frac{d^2}{dz^2}\log(\cdot)$ to both sides we get: $$ \frac{\pi^2}{8\cosh^2(\pi x/2)}=\sum_{m\geq 0}\frac{(2m+1)^2-z^2}{((2m+1)^2+z^2)}\tag{2} $$ If we replace $z$ with $(2n+1)$ and sum over $n\geq 0$, $$ \sum_{n\geq 0}\frac{\pi^2}{8\cosh^2(\pi(2n+1)/2)} = \sum_{n\geq 0}\sum_{m\geq 0}\frac{(2m+1)^2-(2n+1)^2}{((2m+1)^2+(2n+1)^2)^2} \tag{3}$$ where the RHS of $(3)$ can also be written as $$ \sum_{n\geq 0}\sum_{m\geq 0}\int_{0}^{+\infty}\cos((2n+1)x)x e^{-(2m+1)x}\,dx = \sum_{n\geq 0}\int_{0}^{+\infty}\frac{x\cos((2n+1)x)}{2\sinh(x)}\,dx\tag{4}$$ or, by exploiting integration by parts, $$ \sum_{n\geq 0}\int_{0}^{+\infty}\frac{x\cosh(x)-\sinh(x)}{2\sinh(x)}\cdot\frac{\sin((2n+1)x)}{2n+1}\,dx\tag{5}$$ On the other hand, $\sum_{n\geq 0}\frac{\sin((2n+1)x)}{2n+1}$ is the Fourier series of a $2\pi$-periodic rectangle wave that equals $\frac{\pi}{4}$ over $(0,\pi)$ and $-\frac{\pi}{4}$ over $(\pi,2\pi)$. That implies, by massive cancellation: $$ \sum_{n\geq 0}\frac{1}{\cosh^2(\pi(2n+1)/2)}=\frac{1}{2\pi}.\tag{6} $$ On the other hand, the Fourier transform of $\frac{1}{\cosh^2(\pi x)}$ is given by by $\frac{s\sqrt{8\pi}}{\sinh(\pi s)}$.
By Poisson's summation formula, $$ \sum_{n\geq 1}\frac{(-1)^{n+1}n}{\sinh(\pi n)}=\frac{1}{4\pi}.\tag{7}$$

  • $\begingroup$ Hey Jack, since i can't read the source you mentioned, could you please check if their derivation fits my approach? i'm quiet unsure if i made no big mistake here (it seems to easy to be true). thanks! $\endgroup$ – tired Jan 13 '17 at 12:38
  • $\begingroup$ For the pyhsical background you might be interested in this presentation: cedricthieulot.net/diva/05.pdf. The symmetry considerations mentioned by op are essentially saying that: By superposition, adding four plates with one side on a different temperature then the other three (and center temperature given by the complicated sum $S$) gives one plate with constant temperature with core temerature is trivially one. From this $S=1/4$ follows directly $\endgroup$ – tired Jan 13 '17 at 13:03
  • 1
    $\begingroup$ It looks like my predictions were true: sciencedirect.com/science/article/pii/S0955799706000683 $\endgroup$ – Jack D'Aurizio Jan 13 '17 at 13:12
  • 1
    $\begingroup$ @tired: thank you. I always find very pleasing to work on some problem with you. $\endgroup$ – Jack D'Aurizio Jan 13 '17 at 13:13
  • 1
    $\begingroup$ D'Aurizo Yeah, this is a nice proof of concept that physicists and mathematicans CAN work together ;-) $\endgroup$ – tired Jan 13 '17 at 13:17

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.