# Is there any numerical technique to sum x^(n^alpha), n=0,1,…?

I've come across this infinite series: $\sum_{n=0}^\infty x^{n^\alpha}$, with $0<x<1$ and $\alpha > 0$.

Does this series have a name and/or is there a method for computing it (besides brute force, obviously)? Thanks!

-

## 2 Answers

For $\alpha > 1$, this sum will converge extremely rapidly, even if $x$ is fairly close to $1$ (you can't express in floating point numbers those numbers close enough to $1$ where this would converge slowly). The only case that is difficult, convergence wise, is when $\alpha$ is very close to 0.

In that case, your best bet are Levin's U-transform and related algorithms. Even better, just link in to GSL, which already implements that.

-
What do you mean by "For $n\geq2$"? – Jonas Meyer Mar 17 '10 at 18:17
And note the reference to the extensive review paper arxiv.org/abs/math/0005209 found in the second link above. – Harald Hanche-Olsen Mar 17 '10 at 18:19
@Jonas: I bet that's just a slip o' ye olde keyboard and shouldn't have been there. The answer seems to make good sense if you just delete it. – Harald Hanche-Olsen Mar 17 '10 at 18:24
Yep, that $n\geq 2$ was a simple mistake [I had too many variants of the equation in my Maple session where I 'played around' with it first before posting my answer. Error edited away, thanks] – Jacques Carette Mar 17 '10 at 18:53
I thought so. I wouldn't mind all comments referring to it being deleted for posterity, but if I only delete mine it may be confusing. – Jonas Meyer Mar 17 '10 at 19:12

The $k$th power is the generating function for the ways of expressing $n$ as a sum of $k$ $\alpha$th powers of positive integers. However, I don't think this helps much even for most integer values of $\alpha$.

When $\alpha = 1$, this is a geometric series.

When $\alpha = 2$, this is a theta function related to the Jacobi triple product formula

$$\prod_{m=1}^\infty (1-x^{2m})(1+x^{2m-1}y^2)(1+x^{2m-1}y^{-2}) = \sum_{n=-\infty}^\infty x^{n^2}y^{2n}$$ since $$(\frac 12 \text{RHS}+\frac12) \bigg|\_{y=1} = \sum\_{n=0}^\infty x^{n^2} .$$

You may be able to compute the sum when $\alpha=2$ more efficiently using one of the integral formulas or other properties for elliptic theta functions. Other than that, I don't know of special cases.

This doesn't say anything about whether some rapid series acceleration technique exists.

-
For $\alpha=2$ you can use the functional equation for the theta function to get a massive speed-up. Set $\theta(t)=\sum_ne^{-\pi n^2 t^2}$, where the sum is over all integers. This converges for $t>0$ real. Much less clear, but completely standard, is that $\theta(1/t)=t\theta(t)$. Using this trick you can reduce to the case $0<t\leq 1$ which corresponds to $0<x<0.0432...$ where the series is converging super-fast. – Kevin Buzzard Mar 17 '10 at 19:47
As an explicit example, if $x=0.9$ and $\alpha=2$ then summing 1000 terms gives an answer of 3.230272... . But solving $e^{-t^2}=x$ gives t=0.183... so $u:=1/t=5.4605...$ and summing just one term for the series with $x=e^{-\Pi u^2}$ gives 1, correct to 40 or so dec places, so $\theta(t)$ is approximately $u$ (to about 40 dec places!) so the sum we're interested in is hence (1+u)/2. In summary, the answer with $x=0.9$ and $\alpha=2$ is very very close to $(1+u)/2$ with $u=(-\log(0.9)/\pi)^{-1/2}$. As $x$ gets closer to 1 this formula gets better and better. – Kevin Buzzard Mar 17 '10 at 19:55
[...and if you want more accuracy then use more than one term when expanding $\theta(u)$!] – Kevin Buzzard Mar 17 '10 at 19:57
[aargh slip above "reduce to the case $t\geq1$". When can we edit comments?? – Kevin Buzzard Mar 17 '10 at 19:59
For $\theta(1/t)=t\theta(t)$ read $\theta(1/t)=\sqrt{t}\theta(t)$. (Not that this helps much with the original question). So again, when can we edit comments? – Robin Chapman Mar 17 '10 at 21:00