7
$\begingroup$

I have to solve find the value of $$\sum_{k=1}^{n/2} k\log k$$ as a part of question.

How should I proceed on this ?

$\endgroup$
8
  • $\begingroup$ @user8250: It's submission or summation. $\endgroup$
    – anonymous
    Mar 14, 2011 at 16:12
  • $\begingroup$ I may be asking a silly question, but what is a "submission"? $\endgroup$
    – JavaMan
    Mar 14, 2011 at 16:13
  • $\begingroup$ @DJC srry :) it was a typo .... i meant sigma and i am reading more on how to use tags to write equations here.... $\endgroup$
    – user8250
    Mar 14, 2011 at 16:14
  • 3
    $\begingroup$ I doubt you will find a "closed form" formula. Maybe an asymptotic approximation? $\endgroup$
    – Aryabhata
    Mar 14, 2011 at 16:18
  • 5
    $\begingroup$ $log(1^1) + log(2^2) + log(3^3) + ... + log((n/2)^{n/2}) = log(1*4*27*...)$? $\endgroup$ Mar 14, 2011 at 16:22

2 Answers 2

7
$\begingroup$

Got it. The constant in Moron's answer is $C = \log A$, where $A$ is the Glaisher-Kinkelin constant. Thus $C = \frac{1}{12} - \zeta'(-1)$.

The expression $H(n) = \prod_{k=1}^n k^k$ is called the hyperfactorial, and it has the known asymptotic expansion

$$H(n) = A e^{-n^2/4} n^{n(n+1)/2+1/12} \left(1 + \frac{1}{720n^2} - \frac{1433}{7257600n^4} + \cdots \right).$$ Taking logs and using the fact that $\log (1 + x) = O(x)$ yields an asymptotic expression for the OP's sum $$\sum_{k=1}^n k \log k = C - \frac{n^2}{4} + \frac{n(n+1)}{2} \log n + \frac{\log n}{12} + O \left(\frac{1}{n^2}\right),$$ the same as the one Aryabhata obtained with Euler-Maclaurin summation.


Added: Finding an asymptotic formula for the hyperfactorial is Problem 9.28 in Concrete Mathematics (2nd ed.). The answer they give uses Euler-Maclaurin, just as Aryabhata's answer does. They also mention that a derivation of the value of $C$ is in N. G. de Bruijn's Asymptotic Methods in Analysis, $\S$3.7.

$\endgroup$
6
  • $\begingroup$ +1: Very nice ! :-) I am surprised the Plouffe inverter does not have this, though. $\endgroup$
    – Aryabhata
    Mar 14, 2011 at 20:53
  • 1
    $\begingroup$ @Moron: Thanks! It's satisfying to have finally figured out what that constant is. :) $\endgroup$ Mar 14, 2011 at 20:55
  • $\begingroup$ @Moron: It looks like Plouffe does have it (or a scaled version of it) after all. The numerical estimate I was using before wasn't precise enough. $\endgroup$ Mar 14, 2011 at 20:59
  • $\begingroup$ I see. I was about to suggest maybe you should add it there :-) $\endgroup$
    – Aryabhata
    Mar 14, 2011 at 21:25
  • 1
    $\begingroup$ I suppose one proof of this would be to start from $\zeta'(s) = -\sum \log k/k^s$ for $Re(s) \gt 1$ and extend it for all $s$ (which can be done using Euler McLaurin I believe), similar to the Riemann Zeta function. $\endgroup$
    – Aryabhata
    Mar 14, 2011 at 21:34
6
$\begingroup$

Here is an asymptotic expression using EulerMcLaurin Summation.

$$ \sum _{k=1}^{n} k \log k = \int_{1}^{n} x \log x\ \text{d}x + (n\log n)/2 + C' + (\log n + 1)/12+ \mathcal{O}(1/n^2)$$

$$ = n^2(2 \log n - 1)/4 + (n\log n)/2 + (\log n)/12 + C + \mathcal{O}(1/n^2)$$

for some constant $C$.

$\endgroup$
9
  • $\begingroup$ +1, although I get $C = \frac{1}{4}$, $(\log n)/12$ rather than $(\log n)/18$, and $O(\frac{1}{n^2})$. (And I have verified this numerically.) $\endgroup$ Mar 14, 2011 at 17:03
  • $\begingroup$ @MIke: You are correct. I have edited the answer. Thanks. That C=1/4 would need a proof, but that is a neat value for that constant. $\endgroup$
    – Aryabhata
    Mar 14, 2011 at 18:24
  • $\begingroup$ @Moron: You're right to be suspicious of $C = 1/4$. It's not correct. The value of $C$ to six decimal places appears to be $0.248755$ - close to $1/4$ but not quite $1/4$. I'm not sure how to prove that or get an explicit expression, though. $\endgroup$ Mar 14, 2011 at 19:23
  • $\begingroup$ @Mike: Perhaps the Plouffe inverter has something. We would probably need more than 6 digits though. $\endgroup$
    – Aryabhata
    Mar 14, 2011 at 19:45
  • 1
    $\begingroup$ But you're right that the same procedure for finding Stirling's constant ought to work here. There's got to be a way to get this! :) $\endgroup$ Mar 14, 2011 at 20:14

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .