6
$\begingroup$

While testing implementations of Wynn's $\epsilon$-algorithm and Levin's u-transformation I need the value of $$\sum_{n=0}^{\infty} \frac{(-1)^n}{\ln(n+2)} \cdot$$ The results of my algorithms are in agreement with the Pari/GP sumalt value of $0.92429989722293885595957$. But Wolfram Alpha gives the following approximated sum when entering

sum (-1)^n/(ln(n+2))

(a direct link from Math.SE will be mangled and does not work, here is the eq.):

$$\sum_{n=0}^{\infty}\dfrac{(-1)^n}{\log(2+n)}\approx1.00766524110155\ldots$$

Questions:

  • Are the values from Pari and my algorithms correct?
  • Is there a closed form analytical result?
$\endgroup$
  • $\begingroup$ Concerning pari/gp's function sumalt (which will often return finite result even for divergent series) a trick is to gradually increase precision ( \p 100 and more) : the correct digits should remain as precision increases. $\endgroup$ – Raymond Manzoni Dec 15 '13 at 9:01
  • $\begingroup$ Maybe this helps: mathematica.stackexchange.com/questions/149090/… In particular, for n_max = 8 I get 0.924119 with a maximum error 0.000436263 $\endgroup$ – TeM Jun 26 '17 at 15:44
  • $\begingroup$ @Manu: Thank you for the info. I tested it on Raspberry Pi with a[n_] := 1/Log[n + 2]; EulerTransformation[0, 20] and can reproduce your result. But the result is always displayed with 4 digits even for error $\approx 10^{-8}$ and for nmax_ >= 22 I get obscure errors. $\endgroup$ – gammatester Jun 26 '17 at 16:22
  • $\begingroup$ @gammatester: The displayed digits do not match those stored. For example, by writing NumberForm[EulerTransformation[0, 20], 15] you can display 15 decimal digits for both numbers. As for your example, with my Dell I can run the routine up to n = 23, then it does not give any answer. But this is in line with any other numeric algorithm: only effective up to a certain n_max. $\endgroup$ – TeM Jun 26 '17 at 16:37
  • $\begingroup$ @gammatester: For a much more efficient algorithm than Euler's, see the last change that I made to the linked answer; the latter I believe is the best that it is nowadays. $\endgroup$ – TeM Jun 27 '17 at 10:08
7
$\begingroup$

For $x \in (0,1)$ consider

$$f(x) = \sum_{n=2}^{\infty} \frac{(-1)^n}{\log{(n+2)}} x^{\log{(n+2)}}$$

Then

$$f'(x) = \frac{1}{x} \left [ 1-\left(1-2^{1+\log{x}}\right) \zeta(-\log{x})\right]$$

where $\zeta$ is the Riemann zeta function. Using $f(0)=0$, I get that the sum may be expressed in terms of the following integral:

$$\int_0^{\infty} du \left [ 1-\left ( 1 - \frac{1}{2^{u-1}}\right ) \zeta(u)\right]$$

$\endgroup$
  • $\begingroup$ Nice and tricky! But I guess your lower summation limit is 0? $\endgroup$ – gammatester Aug 26 '13 at 13:23
  • $\begingroup$ @gammatester: I added and subtracted the $n=1$ term in the sum to get the zeta function; I hope I answered your question. $\endgroup$ – Ron Gordon Aug 26 '13 at 13:24
  • $\begingroup$ Thank you for the final integral. I evaluated it with Pari and got the 0.92429989722293885595957 again. I will accept your answer because of the analytical result (and up-vote Artin's, and I hope I will avoid multipart questions in the future). $\endgroup$ – gammatester Aug 26 '13 at 14:01
  • $\begingroup$ @gammatester: thanks. I like the integral representation because the series converges incredibly slowly. $\endgroup$ – Ron Gordon Aug 26 '13 at 14:37
1
$\begingroup$

Since $\frac{1}{\log 2}-\frac 1{2\log 3}<1$ already, there is no way $1.07\dots$ can be anywhere close. To get high precision by hand, use Euler-Maclaurin. With Wolfram Alpha, I got $1.01845-2.45012\cdot 10^{-13}i$ . Remember, however, that WA was designed to impress calculus students and educators, not for any minimally non-trivial task...

$\endgroup$

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.