29

I have found a closed form of the following new series involving non-linear harmonic numbers.

Proposition. $$\sum_{n=1}^{\infty} \dfrac{H_n^2-(\gamma + \ln n)^2}{n} = \dfrac{5}{3}\zeta(3)-\dfrac{2}{3}\gamma^3-2\gamma \gamma_{1}-\gamma_{2} $$

where \begin{align} & H_{n}: =\sum_{k=1}^{n}\frac{1}{k} \\ &\gamma: =\lim_{n\to\infty} \left(\sum_{k=1}^n \frac{1}{k}-\ln n\right) \\ & \gamma_{1}:=\lim_{n\to\infty} \left(\sum_{k=1}^n \frac{\ln k}{k}-\frac{1}{2}\ln^2 n\right)\\& \gamma_{2}: =\lim_{n\to\infty} \left(\sum_{k=1}^n \frac{\ln^2 k}{k}-\frac{1}{3}\ln^3 n\right), \end{align} $\gamma_1, \gamma_2$ being Stieltjes constants.

What tool would use to prove it?

Olivier Oloa
  • 120,989
  • I'm slightly confused as to how you were able to find it without proving it, or at least getting part of the way... is this a numerical guess, or do you have some heuristics for it? – Andrew D Jul 13 '14 at 22:39
  • @Andrew D, for the moment neither Mathematica nor Maple is able to give a closed form for the series. In addition, all 'inverse symbolic calculators' fail to give something interesting for this series. I set this result as a challenge problem for those who are interested in it. – Olivier Oloa Jul 13 '14 at 23:03
  • 4
    Could you* prove it?* - No! $\big($ OK, now I've answered the only question that's actually being asked in this entire post $)$. – Lucian Jul 14 '14 at 00:58
  • I meant (typo corrected) $$\sum_{n=1}^{\infty} \dfrac{H_n-(\gamma + \ln n)}{n} = \dfrac{1}{2}\zeta(2)-\dfrac{1}{2}\gamma^2-\gamma_{1}.$$ – Olivier Oloa Jul 14 '14 at 07:00
  • @OlivierOloa Have you been able to find or conjecture values of $\sum_{n=1}^{\infty} \dfrac{H_n^m-(\gamma + \ln n)^m}{n}$ for $m>2$? – Vladimir Reshetnikov Dec 24 '14 at 20:06
  • @VladimirReshetnikov I did not write down a general formula, but I think I can obtain one, I mean some recurrent relation between series of different 'orders'. Thank you for your interest. – Olivier Oloa Dec 25 '14 at 00:30
  • @OlivierOloa Could you please post an explicit result for $m=3$? – Vladimir Reshetnikov Dec 25 '14 at 18:34
  • I assume, the $\gamma$ are the Stieltjes-constants? (I just tried to confirm it numerically using $n=100$,$n=1000$, $n=10000$ and $n=200000$ and seems to approximate Stieltjes $\gamma_1$ and $\gamma_2$ – Gottfried Helms Jul 08 '15 at 07:10
  • @GottfriedHelms Yes, the $\gamma$ are Stieltjes constants! By the way, I'm interested in references to a simple Mathematica/Maple program to compute Stieltjes constants. Thanks! – Olivier Oloa Jul 08 '15 at 09:34
  • I have a Pari/GP-routine which makes it really simple. It uses sumalt , a procedure for divergent summation (based on Zagier's ideas) on formal powerseries. Possibly in Math'ica this is possible as well. I can post it here if you're interested. (Or search for "stieltjes" and my user name, possible I've already posted that routine here or in MO) – Gottfried Helms Jul 08 '15 at 10:45
  • @GottfriedHelms Thank you very much. – Olivier Oloa Jul 08 '15 at 11:47

1 Answers1

19

Here is a nice solution from R. Tauraso, for those interested in this problem:

Solution proposed by Roberto Tauraso, Roma, Italy.

The Multiple Harmonic Sum is defined by $$H_{n}(s_{1}, \dots, s_{l}) == \sum_{0 < k_{1} < k_{2} < \dots < k_{l} \leq n}\prod_{i = 1}^{l}\frac{1}{k_{i}^{s_{i}}}$$ Then, by the known properties of the stuffle product, \begin{align} H_{n}^{2}(1) &= 2H_{n}(1, 1) + H_{n}(2)\notag\\ H_{n}^{3}(1) &= 6H_{n}(1, 1, 1) + 3H_{n}(1, 2) + 3H_{n}(2, 1) + H_{n}(3)\notag \end{align} Therefore, for $N > 0$, \begin{align} \sum_{n = 1}^{N}\frac{H_{n}^{2}}{n} &= 2\sum_{n = 1}^{N}\frac{H_{n}(1, 1)}{n} + \sum_{n = 1}^{N}\frac{H_{n}(2)}{n}\notag\\ &= 2H_{N}(1, 1, 1) + 2H_{N}(1, 2) + H_{N}(3) + H_{N}(2, 1)\notag\\ &= \frac{1}{3}H_{N}^{3}(1) + \frac{2}{3}H_{N}(3) + H_{N}(1, 2)\notag\\ &= \frac{1}{3}H_{N}^{3}(1) + \frac{2}{3}\zeta(3) + \zeta(2, 1) + o(1)\notag\\ &= \frac{1}{3}H_{N}^{3}(1) + \frac{5}{3}\zeta(3) + o(1)\tag{1} \end{align} because $\zeta(2, 1) = \zeta(3)$ (see for example Thirty-two Goldbach Variations by J. M. Borwein and D. M. Bradley).

Moreover, $H_{N}(1) = \ln N + \gamma + O(1/N)$ implies that \begin{align} \frac{1}{3}H_{N}^{3}(1) - \sum_{n = 1}^{N}\frac{(\gamma + \ln n)^{2}}{n} &= \frac{1}{3}H_{N}^{3}(1) - \gamma^{2}H_{N}(1) - 2\gamma\sum_{n = 1}^{N}\frac{\ln n}{n} - \sum_{n = 1}^{N}\frac{\ln^{2}n}{n}\notag\\ &= \frac{\ln^{3}N}{3} + \frac{\gamma^{3}}{3} + \gamma\ln^{2}N + \gamma^{2}\ln N - \gamma^{2}(\ln N + \gamma)\notag\\ &\,\,\,\,\,\,\,\, -2\gamma\left(\gamma_{1} + \frac{\ln^{2}N}{2}\right) - \left(\gamma_{2} + \frac{\ln^{3}N}{3}\right) + o(1)\notag\\ &= -\frac{2}{3}\gamma^{3} - 2\gamma\gamma_{1} - \gamma_{2} + o(1)\tag{2} \end{align} By adding together $(1)$ and $(2)$, and by taking the limit as $N \to \infty$, we get the result.

Paramanand Singh
  • 87,309
Olivier Oloa
  • 120,989