I am interested in the double series


I believe that this series is not absolutely convergent but converges by rows or columns,


and would like to find both a proof of convergence and the value of $S$ in closed form.

As a first step, I considered the double power series:

$$F(x,y) = \sum_{m=1}^{\infty}\sum_{n=1}^{\infty}\frac{(-1)^{m+n}mn}{(m+n)^2}x^my^n$$

which I believe converges absolutely for $|x|,|y|<1$. For $x=y$, I was able to sum diagonally

$$F(x,x) = \sum_{m=2}^{\infty}\sum_{k=1}^{m-1}\frac{(-1)^{m}k(m-k)}{m^2}x^m =\sum_{m=2}^{\infty}\frac{(-1)^{m}x^m}{m^2}\sum_{k=1}^{m-1}k(m-k)=\sum_{m=2}^{\infty}\frac{(-1)^{m}x^m}{6}\Big(m-\frac{1}{m}\Big)=\frac{1}{6}\Big[\log(1+x)-\frac{x}{(1+x)^2}\Big].$$

and find the limit

$$\lim_{x\rightarrow1-}F(x,x) = \frac{1}{6}\Big(\log2-\frac{1}{4}\Big).$$

Note that the diagonal sum of $F(1,1)$ does not converge, but oscillates between $\pm \infty.$

My conjecture is that $S = \frac{1}{6}\Big(\log2-\frac{1}{4}\Big)$ but I have not found a way to directly sum by rows or columns. I also wonder if some extension of Abel's limit theorem can be applied here.


2 Answers 2


To compute the sum $S$, one can start with the identity $$ \frac1{a^2}=\int_0^\infty \mathrm e^{-at}t\mathrm dt, $$ valid for every positive $a$. Using it for each $a=m+n$ allows to get rid of the denominators $(m+n)^2$ and yields $$ S=\int_0^\infty\sum_{m,n}(-1)^{m+n}mnt\mathrm e^{-(m+n)t}t\mathrm dt=\int_0^\infty\left(\sum_n(-1)^nn\mathrm e^{-nt}\right)^2t\mathrm dt. $$ Identifying the inner sum yields $$ S=\int_0^\infty\left(\frac{-\mathrm e^{-t}}{(1+\mathrm e^{-t})^2}\right)^2t\mathrm dt=\int_0^\infty\frac{\mathrm e^{2t}}{(1+\mathrm e^t)^4}t\mathrm dt. $$ Integrating by parts, one gets $$ S=\int_0^\infty\frac{1+3\mathrm e^t}{6(1+\mathrm e^t)^3}\mathrm dt. $$ The change of variable $x=\mathrm e^t$ reduces this to $$ S=\int_1^\infty\frac{1+3x}{6x(1+x)^3}\mathrm dx. $$ Note that $$ \frac{1+3x}{x(1+x)^3}=\frac1x-\frac1{1+x}-\frac1{(1+x)^2}+\frac2{(1+x)^3}, $$ hence $$ 6S=\left.\log\left(\frac{x}{1+x}\right)+\frac1{1+x}-\frac1{(1+x)^2}\right|_1^\infty=\log2-\frac12+\frac14, $$ hence $$ S=\frac16\log2-\frac1{24}. $$

  • 2
    $\begingroup$ Using the first identity requires a justification in changing the order of $\int_0^{\infty}$ and $\sum$. $\endgroup$ May 30, 2014 at 9:12
  • $\begingroup$ @i707107 Yes it does. $\endgroup$
    – Did
    May 30, 2014 at 9:27
  • 1
    $\begingroup$ @Did: How can it be justified? Actually you need to change $\sum_m \sum_n \int$ to $\int \sum_m \sum_n$. $\endgroup$ May 31, 2014 at 1:18
  • $\begingroup$ @i707107 You already said that and I perfectly understood the first time, thank you. $\endgroup$
    – Did
    May 31, 2014 at 7:10
  • $\begingroup$ @Did: (+1) Never mind. I know how to justify it now. Thank you anyway. $\endgroup$ Jun 1, 2014 at 9:48

It's certainly not absolutely convergent, because e.g. the $m=n$ terms are all $1/4$.

FWIW, Maple's numerical summation (which uses Levin's U transform) agrees with your conjecture:




$$ - \dfrac{1}{24} + \dfrac{1}{6} \ln(2) $$


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.