Suppose $s$ is a complex number with $\Re(s) \in (0,1]$ and $\{a_n\}$ is a complex sequence converging to $a \neq 0$. Must the Dirichlet series $$\sum_{n=1}^\infty\frac{a_n}{n^s}$$ diverge?

I asked this question on math.stackexchange on 2011-12-13. Here is the link: https://math.stackexchange.com/questions/91218/divergence-of-dirichlet-series.

For reference, put $s = r + it$. It is easy to see that the above series must converge if $r > 1$ and must diverge if $r \leq 0$ or if $s = 1$. On math.stackexchange I was able to prove divergence for $0 < r < 1$, using a result from Knopp. This leaves the question open for $s = 1 + it$, where $t \neq 0$.

  • $\begingroup$ The results you mention seem to follow from math.nus.edu.sg/~chanhh/MA4263/Chapter6.pdf theorem 6.3.1 $\endgroup$ – Igor Rivin Dec 22 '11 at 15:38
  • $\begingroup$ I added a small correction to your proof for $s=1$ on SE (it awaits "peer review" as I am not registered there). Your explanation that either the real or the imaginary part of $(a_n)$ is bounded away from zero was not sufficient, e.g. $\sum_n (-1)^n/n$ has this feature but it converges. You need that either the real or the imaginary part of $(a_n)$ tends to a nonzero real number. $\endgroup$ – GH from MO Dec 23 '11 at 13:33
  • $\begingroup$ Yes, the theorem at math.nus.edu is much like the one I used from Knopp. Also, thanks, GH--I have clarified my argument on SE. $\endgroup$ – Richard Hevener Dec 29 '11 at 23:39

The answer to your question is yes, the series must always diverge when $\Re (s)\in (0,1]$. The proof is a little delicate but here is the argument:

First of all to reiterate what you already mentioned, since the $a_n$'s have a non-zero limit, the abscissa of absolute convergence and the abscissa of convergence are both equal to $1$. Also the series clearly does not converge absolutely for $\Re (s)=1$, and it doesn't converge at all for $s=1$, so the only question is whether or not it converges conditionally for $s=1+it,~t\not= 0$.

Now for $t\not= 0$ let $$S_N=\sum_{n=1}^N\frac{a_n}{n^{1+it}}.$$ By a simple power series argument, for all $K,N\in\mathbb{N}$ with $K\le N/2$ we have that $$t(\log (N+K)-\log N)<\frac{ctK}{N},$$ for some universal constant $c>0$. Thus for $\epsilon >0$ if $N$ is large enough (depending on $t$) then choosing $$K=\left[\frac{\epsilon N}{ct}\right],$$ gives that $t(\log (N+K)-\log N)<\epsilon$.

Finally assume that $\epsilon$ is small and that $N$ is chosen as large as you need so that the quantities $a_n\exp(-it\log n)$ are all close to some constant $c'$, with $|c'|\approx |a|$, for all $N< n\le N+K.$ Then since the argument of the summand can not change by more than $\epsilon$ over this range we have that $$|S_{N+K}-S_N|=\left|\sum_{n=N+1}^{N+K}\frac{a_n\exp(-it\log n)}{n}\right|\gg |a|\sum_{n=1}^K\frac{1}{N+n}.$$ Pulling out the $N$ from the denominator gives $$|S_{N+K}-S_N|\gg \frac{|a|K}{N}\gg 1.$$ The final implied constant depends on $\epsilon$ and $t$, but both of these parameters are fixed (note that we don't have to let $\epsilon$ tend to $0$ to make this work, we just need it to be small). Therefore the sequence $S_N$ is not a Cauchy sequence and the Dirichlet series is not conditionally convergent for any value of $t$.

The underlying reason why the series don't converge is that the phase of the summand is not increasing fast enough to cause the necessary amount of cancellation.


I had a devil of a time figuring out the previous answer, but I now believe it to be fundamentally sound. Based on that answer, I am providing another one with a number of gaps filled in.

First note that we can assume WLOG that $a = 1$. Under this assumption we will show that either the real or the imaginary part of $\sum_{n=1}^\infty \frac {a_n} {n^s}$ diverges, where $s = 1 + it$ and $t \neq 0$. We also assume WLOG that $t > 0$. (Otherwise, replace $t$ by $-t$ in appropriate places below.) Put $a_n = u_n + iv_n$, where $u_n \rightarrow 1$ and $v_n \rightarrow 0$. Note that $$\Re(\frac {a_n} {n^s}) = (u_n \cos (t \log (n)) + v_n \sin (t \log (n))) / n$$ and $$\Im(\frac {a_n} {n^s})= (-u_n \sin (t \log (n)) + v_n \cos (t \log (n))) / n.$$

There must either be infinitely many values of n for which $|\cos (t \log (n))| > 1/2$ or infinitely many values of n for which $|\sin (t \log (n))| > 1/2$ . We assume the former. (Otherwise, switch the roles of sin and cos, and work with the imaginary part of the series instead of the real part; the necessary changes are fairly straightforward.) Likewise, WLOG we assume that there are inf. many values of n for which $\cos (t \log (n)) > 1/2$, as opposed to $< -1/2$. (Otherwise, replace cos by -cos and $\Re$ by $-\Re$ in appropriate places below.)

Now $\cos(x)$ is uniformly continuous on the real line. Select $\delta > 0$ such that if $|x - y| < \delta$, then $|\cos(x) - \cos(y)| < 1/4$ . We can also assume that $\delta < t$.

To prove divergence under the above assumptions, we will show that the sequence of partial sums of the real part of our series is not Cauchy. Let $M$ be a positive integer.

First find $M_1 > M$ such that for all $n > M_1$, $u_n > 1/2$ and $|v_n| < 1/16$. Choose $N$ an integer such that $N > \max(M_1, 2t / \delta)$ and $\cos (t \log (N)) > 1/2$ . Now $N \delta / 2t > 1$, whence $N \delta / t - N \delta / 2t > 1$. Thus, we can select $K$ an integer with $1 < N \delta / 2t < K < N \delta / t$ . Let $n$ be any integer with $n > N$. From calculus we know that $$t(\log (n) - \log (N)) < t(n - N) / N.$$
Observe that if $N < n \leq N + K$ , then $$t(\log (n) - \log (N)) < tK / N < \delta,$$ whence $$|\cos(t \log (n)) - \cos(t \log (N))| < 1/4.$$ Therefore, $$ \eqalign{&u_n \cos (t \log (n)) + v_n \sin (t \log (n))\geq u_n \cos (t \log (n)) - |v_n \sin (t \log (n))|\cr&\qquad\gt \cos (t \log (n)) / 2 - 1/16\cr&\qquad\geq (\cos (t \log (N)) - |\cos (t \log (n)) - \cos (t \log (N))|) / 2 - 1/16\cr&\qquad\gt (1/2 - 1/4) / 2 - 1/16 = 1/16.\cr}$$

It follows that $$\eqalign{\sum_{n=N+1}^{N+K} \Re(\frac {a_n} {n^s}) &\gt \sum_{n=N+1}^{N+K} 1 / 16n \gt K / (16(N + K))= 1 / (16(N / K + 1))\cr& \gt 1 / (16(2t / \delta + 1))= \delta / (16(2t + \delta)) \gt \delta / 48t.\cr}$$
Thus, the sequence of partial sums of the real part of our series is not Cauchy.

  • $\begingroup$ Richard, if you're still around, I've done rather a lot of reformatting of your answer. You might want to check to see whether I've messed anything up. $\endgroup$ – Gerry Myerson Dec 6 '12 at 0:13

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.