6
$\begingroup$

Actually, as the corresponding integral $\frac{\ln(\cos x)}{1-x}$ or something like that) cannot be expressed in closed form by Liouville theorem, this shouldn't exist, but I believe I have seen it shown somewhere using Fourier series. By the way, the result : $$\sum_{n=0}^{\infty}\frac{1}{n^2+1}= \frac{1+ \pi\coth\pi}{2}$$ is known by WolframAlpha. But my question is : does there exist a general theory of similar results (something like exact formulas for $\sum_{n=0}^{+\infty}\frac{1}{an^2+bn+c}$, at least when $a$, $b$ and $c$ are integers)?

$\endgroup$
3
  • 3
    $\begingroup$ Yes. We may interpret $f(n)$ as a residue of function $f(x)\cot (\pi x)$ in a point $n$. Then note that sum of residues of this function equals to 0 as integral over appropriate large circuit tends to 0. $\endgroup$ Nov 30, 2015 at 15:13
  • 2
    $\begingroup$ There is no relation between existence of closed form for $\int f(x)\;dx$ and closed form for $\sum f(n)$. $\endgroup$ Nov 30, 2015 at 15:15
  • 3
    $\begingroup$ Integral not of f(x), but of $f(x)\cot \pi x$ $\endgroup$ Nov 30, 2015 at 15:30

3 Answers 3

16
$\begingroup$

To do $\sum\frac{1}{an^2+bn+c}$, factor the denominator and get a digamma answer. $$ \sum_{n=0}^\infty \frac{1}{(x+p)(x+q)} = \frac{\psi(p)-\psi(q)}{p-q} $$ And note that digamma of a rational can be evaluated in terms of logarithms and trig functions (Gauss's digamma theorem). $$ \sum_{n=0}^\infty\frac{1}{(n+\frac{1}{4})(n+\frac{1}{3})} = 36\log 2+6\pi-2\pi\sqrt{3}-18\log 3 . $$

$\endgroup$
15
$\begingroup$

This is a standard exercise on residue theory. If $f$ is a rational function with zero of order $\geq 2$ at infinity and no poles at integers, then $$\sum_{-\infty}^\infty f(n)=-\sum{\mathrm{res}}_af(z)\pi\cot\pi z,$$ where the summation is over all poles of $f$. For $f(z)=1/(1+z^2)$ we obtain $$\sum_0^\infty\frac{1}{1+n^2}=1+\sum_{1}^\infty\frac{1}{1+n^2}=(\pi\coth\pi+1)/2,$$ so you copied the result from Wolfram incorrectly.

$\endgroup$
2
  • $\begingroup$ Mmm ; sorry, but Wolfram (and Maple) gives the sum equals to 2.07417717505..., while (π coth π+1)/2 is 2.07667404... $\endgroup$ Nov 30, 2015 at 21:05
  • 3
    $\begingroup$ @FeldmannDenis this anser 2.0766704 is correct, and when I ask Wolframalpha it confirms this wolframalpha.com/input/… $\endgroup$ Nov 30, 2015 at 22:02
8
$\begingroup$

Alternatively, you may use expansion of cotangent $$ \pi\cot \pi x=\frac1x+\sum_{n=1}^{\infty}\frac{2x}{x^2-n^2} $$ as a blackbox (after all, complex residues are not the only way to obtain this formula: say, there is Herglotz trick, explained, for instance, in Proofs from the Book).

Just substitute $x=i$, for the sum $S=\sum_{n=0}^{\infty} \frac1{n^2+1}$ we get $\pi \cot \pi i=-i-2i(S-1)$, $2iS=i-\pi \cot \pi i=i(1+\pi\coth \pi)$, $S=(1+\pi \coth \pi)/2$.

For arbitrary rational function $f(z)$, we may find on this way the principal value of $\sum_{n\in \mathbb{Z}} f(n)$. If $f(z)=\sum_k c_k/(z_k-z)$, for each summand we have $$ \sum_{j=-n}^n\frac1{z_k-j}=\frac1{z_k}+\sum_{k=1}^n\frac{2z_k}{z_k^2-j^2}\rightarrow \frac{\pi\cot \pi z_k}2, $$ thus principal value of $\sum_{n\in \mathbb{Z}} f(n)$ equals $$ \sum_k c_k \frac{\pi\cot \pi z_k}2. $$ This works whenever $f$ has only simple poles and $f(z)=O(1/z)$ for large $z$. If additionally $f(z)=O(1/z^2)$, principal value is the same as ordinary sum. If $f$ has multiple poles, we may perturbate $f$ a bit and take a limit. Yes, this is one of approaches to residues.

If you need a sum not over integers, but over positive integers (it looks like it is the case), you instead of cotangent use the digamma $\psi$-function $$ \psi(z)=-\gamma+\sum_{n=0}^{\infty} \left(\frac1{n+1}-\frac1{n+z}\right). $$ This is the essence of Gerald Edgard's answer.

$\endgroup$
5
  • $\begingroup$ Yes, but he asked to sum $f(n)$ with a more general class of functions $f$. $\endgroup$ Dec 1, 2015 at 0:14
  • $\begingroup$ Indeed, but this works for all rational functions. $\endgroup$ Dec 1, 2015 at 5:59
  • 2
    $\begingroup$ Just to complete the picture, the expansion of the cotangent can be obtained by logarithmic differentiation from the infinite product for $\sin(\pi x)$, and the infinite product for $\sin(\pi x)$ can be obtained as dominated limit on the (easily computed) complete factorization of the polynomials $(1+iz/n)^n-(1-iz/n)^n$, that converge to $e^{iz}-e^{-iz}=2i\sin(z)$ (Is this Herglotz' proof?) $\endgroup$ Dec 1, 2015 at 11:24
  • $\begingroup$ (no: I've just found Herglotz trick on line) $\endgroup$ Dec 1, 2015 at 11:32
  • $\begingroup$ oh, I did not know this trick with polynomial $\endgroup$ Dec 1, 2015 at 12:07

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.