52

Does $S_k= \sum \limits_{n=1}^{\infty}\sin(n^k)/n$ converge for all $k>0$?

Motivation: I recently learned that $S_1$ converges. I think $S_2$ converges by the integral test. Was the question known in general?

Quixotic
  • 22,431
curious
  • 2,599
  • 4
    I am not sure you can apply the integral test here for $k=2$. Isn't the function being non-negative a necessary condition? – Aryabhata Aug 12 '10 at 19:03
  • @Moron: you might be right. I thought I could use this one: http://www.jstor.org/stable/2305031?cookieSet=1 , but may be I made a mistake. – curious Aug 12 '10 at 19:18
  • @curious I don't think so. Suppose $f(x)=\sin(x^2)/x$, we have $f'(x)=2\cos(x^2)-\sin(x^2)/x^2$. If $\int f'(x)$ converges absolutely, then $\int \cos(x^2)$ converges absolutely, contradiction! I think integral test wouldn't work if the integrand is oscillating near $\infty$. – Yai0Phah May 16 '13 at 16:43
  • If we can bound the finite sum $\sum_{m=1}^{n}\sin(m^k), \forall n\in \mathbb{N}$, then we can use Dirichlet test to prove the convergence. – Mhenni Benghorbal Jun 07 '13 at 05:28

1 Answers1

61

This is a replacement for my previous answer. The sum converges, and this fact needs even more math than I believed before.

Begin by using summation by parts. This gives $$\sum_{n=1}^N \left(\sum_{m=1}^N \sin(m^k) \right) \left( \frac{1}{n}-\frac{1}{n+1}\right) + \frac{1}{N+1} \left(\sum_{m=1}^N \sin(m^k) \right).$$ Write $S_n:= \left(\sum_{m=1}^n \sin(m^k) \right)$. So this is $$\sum_{n=1}^N S_n/(n(n+1)) + S_N/(N+1).$$ The second term goes to zero by Weyl's polynomial equidistribution theorem. So your question is equivalent to the question of whether $\sum s_n/(n(n+1))$ converges. We may as well clean this up a little: Since $|S_n| \leq n$, we know that $\sum S_n \left( 1/n(n+1) - 1/n^2 \right)$ converges. So the question is whether $$\sum \frac{S_n}{n^2}$$ converges.

I will show that $S_n$ is small enough that $\sum S_n/n^2$ converges absolutely.

The way I want to prove this is to use Weyl's inequality. Let $p_i/q_i$ be an infinite sequence of rational numbers such that $|1/(2 \pi) - p_i/q_i| < 1/q_i^2$. Such a sequence exists by a standard lemma. Weyl inequality gives that $$S_N = O\left(N^{1+\epsilon} (q_i^{-1} + N^{-1} + q_i N^{-k})^{1/2^{k-1}} \right)$$ for any $\epsilon>0$.


Thanks to George Lowther for pointing out the next step: According to Salikhov, for $q$ sufficiently large, we have $$|\pi - p/q| > 1/q^{7.60631+\epsilon}.$$ Since $x \mapsto 1/(2x)$ is Lipschitz near $\pi$, and since $p/q$ near $\pi$ implies that $p$ and $q$ are nearly proportional, we also have the lower bound $|1/(2 \pi) - p/q|> 1/q^{7.60631+\epsilon}$.

Let $p_i/q_i$ be the convergents of the continued fraction of $1/(2 \pi)$. By a standard result, $|1/(2 \pi) - p_i/q_i| \leq 1/(q_i q_{i+1})$. Thus, $q_{i+1} \leq q_i^{6.60631 + \epsilon}$ for $i$ sufficiently large. Thus, the intervals $[q_i, q_i^{7}]$ contain all sufficiently large integers.

For any large enough $N$, choose $q_i$ such that $N^{k-1} \in [q_i, q_i^7]$. Then Weyl's inequality gives the bound $$S_N = O \left( N^{1+\epsilon} \left(N^{-(k-1)/7} + N^{-1} + N^{-1} \right)^{1/2^{k-1}}\right)$$

So $$S_N = \begin{cases} O(N^{1-(k-1)/(7\cdot 2^{k-1}) + \epsilon}) &\mbox{ if } \ k\leq 7, \\ O(N^{1-1/(2^{k-1})+\epsilon}) &\mbox{ if } \ k\geq 8, \end{cases}$$ which is enough to make sure the sum converges. ${ }{}{}{}{}$

Sungjin Kim
  • 20,102
  • 2
    I hate to say this, but I am getting worried about whether I can push Weyl's inequality to do what I want. Weyl's inequality says, if $|\chi - p/q| < 1/q^2$, with $GCD(p,q)=1$, then $S_N = O(N^{1+\epsilon} (q^{-1} + N^{-1} + q N^{-d})^{1/2^d})$. What if the convergents to $\chi$ grow so fast that either $1/q$ or $q N^{-d}$ is as large as $1/ \log N$, for any choice of $q$? I was working from my notes from a course which claimed the $O(N^{1-1/2^d+\epsilon})$ bound, but I no longer follow the proof which was written there. So I have some concerns. – David E Speyer Aug 17 '10 at 15:38
  • David, your proof can be fixed up to work perfectly. Just needs one extra ingredient - pi has finite irrationality measure – George Lowther Aug 19 '10 at 18:19
  • 4
    In fact, the mathworld website has a table of bounds for the irrationality measures of some constants, including pi. (http://mathworld.wolfram.com/IrrationalityMeasure.html) – George Lowther Aug 19 '10 at 18:39
  • Sweet! I'll edit this in, feel free to post your own answer if you'd like. – David E Speyer Aug 19 '10 at 19:09
  • I'll leave you to complete this so that there's a single complete answer. As $1/2\pi$ will have the same irrationality measure as pi, $\mu=7.6304$ it follows that for all $r>\mu$ and large enough M, you can choose $M^{1/(r-1)}\le q\le M$ in Weyl's inequality. Using $N^{k/r}\le q\le N^{k(1-1/r)}$ seems to give a nice bound and, otherwise, your original answer does not have to be changed so much (which I thought was a very nicely written answer, btw). – George Lowther Aug 19 '10 at 19:26
  • @David: will you consider writing this up somewhere (else)? I think a lot of people would be interested to see it. – Pete L. Clark Aug 19 '10 at 23:12
  • @Pete: What level of writing up are you thinking of? I'm sure this isn't interesting enough to publish. I am still curious about what we can say about $\sum \sin (\theta n^k)$ for general irrational $\theta$, without any diophantine conditions. – David E Speyer Aug 19 '10 at 23:32
  • David - for general irrational theta, I've thought of counterexamples showing can diverge (both the sin and cosine series, for every k>1, all simultaneously for the same theta). Using a similar argument as in my answer to your related MO question, which I can tidy up a bit ( might go back and edit that answer). – George Lowther Aug 20 '10 at 00:04
  • @David: it might be publishable somewhere, but I respect your high standards of what to try to publish. Anyway, how about a nice texed up writeup including references? – Pete L. Clark Aug 20 '10 at 01:56
  • 3
    Perhaps I am just addled, but it seems to me that $\sum_{n=1}^N \left(\sum_{m=1}^N \sin(m^k) \right) \left( \frac{1}{n}-\frac{1}{n+1}\right) + \frac{1}{N+1} \left(\sum_{m=1}^N \sin(m^k) \right)=\left(1-\frac{1}{N+1} \right )S_N + \frac{1}{N+1}S_N = S_N$, not $\sum_{n=1}^N S_n/(n(n+1)) + S_N/(N+1)$ as claimed (I "fixed" $s_N$ to $S_N$). – robjohn Aug 01 '11 at 22:26
  • 1
    @robjohn The expression should be $ \sum_{n=1}^N\left(\sum_{m=1}^{\fbox{$n$}}\sin(m^k)\right)\cdots $, which gives the result David announced. – jathd Jan 25 '14 at 22:59
  • @DavidESpeyer The bound for $S_N$ depends on $k$. If $k\geq 8$, then the inside paranthesis must be $O(N^{-1})$. Thus, it gives $$S_N=O(N^{1-\frac1{2^{k-1}}+\epsilon})$$ for $k\geq 8$. – Sungjin Kim Nov 12 '18 at 23:39
  • @DavidESpeyer I realized that was actually my edit 5 years ago after carefully following your argument. It seemed to me at that time the term $(k-1)$ should be on the numerator. However, I did not realize at that time that the inner error term need to be more carefully dealt with due to the extra $N^{-1}$ terms. Now, after reading your solution again, it is possible to optimize more to have a stronger result when $k<8$, and the above comment when $k\geq 8$. – Sungjin Kim Nov 12 '18 at 23:46