How do I find for which values of $\alpha \in \mathbb{R}$ the sum converges?

$$\sum_{n=1}^\infty (1 - \cos \frac1n)^\alpha\log n$$

I have tried using the following techniques:

  • Comparison test. Hard to compare with $\sum_{n=1}^\infty\frac1{n^\alpha}$ as there is no similarity.
  • Limit comparison test. Comparing with $\sum_{n=1}^\infty \frac1{n^\alpha}$, I get no information, as the limit is zero and $\sum_{n=1}^\infty \frac1n$ diverges. (Also tried for small cases of $n$.)

The other convergence criteria do not seem to be handy in this particular exercise at all.

This should be very easy, what am I doing wrong?


  • $\begingroup$ Do you see how to bound $\cos(1/n)$ in terms of $n^{-2}$? $\endgroup$ – Gerry Myerson Aug 9 '13 at 11:46
  • $\begingroup$ @GerryMyerson, I do not. Maybe the same way Harald suggested in his post? $\endgroup$ – David Aug 9 '13 at 11:59

Hint: $$1-\cos\frac1n \sim \frac1{2n^2}$$ as $n\to\infty$. So you can compare your series with $$\sum_{n=1}^\infty\frac{\log n}{n^{2\alpha}}.$$ And you can compare that from one side with $$\sum_{n=1}^\infty\frac{1}{n^{2\alpha}},$$ and from the other with $$\sum_{n=1}^\infty\frac{1}{n^{2\alpha-\varepsilon}}$$ for any $\varepsilon>0$.

(Edited to include more detail after the discussion in the comments.)

Yet another edit, adding more detail: For the last bit, you need to know that $\log n<n^\varepsilon$ for large $n$, if $\varepsilon>0$. If you put $x=n^\varepsilon$, this can be rewritten as $\log x^{1/\varepsilon}<x$ for large $x$, which in turn simplifies to $\log x<\varepsilon x$ when $x$ is large enough. But that follows from $$\lim_{x\to\infty}\frac{\log x}{x}=0.$$ This very fundamental equality can easily be shown using l'Hôpital's rule, but it is so important you should know it by heart.

  • $\begingroup$ Missing a $1/2$? $\endgroup$ – Gerry Myerson Aug 9 '13 at 11:52
  • $\begingroup$ I got this; this comes from the fact that $\lim_{n \to\infty} \frac{1 - \cos\frac1n}{\frac1{n^2}} = \frac12$. I have noticed this, but comparing the sum with both $\sum_{n=1}^\infty \frac1{n^\alpha}$ and $\sum_{n=1}^\infty \frac1{n^{2\alpha}}$, I got nowhere. $\endgroup$ – David Aug 9 '13 at 11:58
  • $\begingroup$ The point is, you can use this to compare your sum to $\sum n^{-2a}\log n$. Then the question is, can you tell which values of $a$ make $\sum n^{-2a}\log n$ converge? $\endgroup$ – Gerry Myerson Aug 9 '13 at 12:03
  • $\begingroup$ That seems to be the problem. When I reduced the problem to the convergence of $\sum n^{-2\alpha}\log n$, it seemed to be improper form for appropriate discussion. I know that $\sum \frac1{n^\alpha}$ will converge $\forall \alpha > 1$. But I have to compensate for $\log n$ in the numerator. I could divide both the numerator and the denominator by $n$ getting $\sum \frac{\frac{\log n}n}{n^{2\alpha - 1}}$, but I could also divide by $\sqrt{n}$ getting different constraints for the values of $\alpha$ for which the series converges. $\endgroup$ – David Aug 9 '13 at 12:12
  • $\begingroup$ Try thinking about it this way: $\log(n)$ is eventually smaller than any positive power of $n$. So, if $2\alpha>1$, write $2\alpha=1+\epsilon$, $\epsilon>0$, and note that $\log(n)$ is much smaller than $n^{\epsilon/2}$. Then, you have to consider the case when $2\alpha\leq 1$; what can you do comparison-wise with this? $\endgroup$ – Nick Peterson Aug 9 '13 at 12:42

In this answer, all the bounds will be explicit. There are no "for large enough $x$" type statements.

Since $\cos(2x) = \cos^2(x)-\sin^2(x) =1-2\sin^2(x) $, $1-\cos(x) =2\sin^2(x/2) $.

So $S(\alpha) =\sum_{n=1}^\infty (1 - \cos \frac1n)^\alpha\log n =\sum_{n=1}^\infty (2\sin^2\frac1{2n})^\alpha\log n =2^\alpha\sum_{n=1}^\infty (\sin\frac1{2n})^{2\alpha}\log n $.

There are two simple ways to bound $\sin(\frac1{2n})$.

First: For $0 \le x \le \pi/2$, $2x/pi \le \sin(x) \le x$, so $1/(\pi n) \le\sin(1/2n) \le 1/(2n) $.

Second: Since $x \ge \sin(x) \ge x-x^3/6$, for $n \ge 1$, $\sin(\frac1{2n}) \ge \frac1{2n}-(\frac1{2n})^3/6 = \frac1{2n}-\frac1{48n^3} \ge \frac1{n}(\frac1{2}-\frac1{48}) = \frac{23}{48n} $.

The first bound shows that $S(\alpha)$ is between $2^\alpha\sum_{n=1}^\infty (\frac1{\pi n})^{2\alpha}\log n =\left(\frac{2}{\pi^2}\right)^\alpha\sum_{n=1}^\infty \frac1{ n^{2\alpha}}\log n $ and $2^\alpha\sum_{n=1}^\infty (\frac1{2 n})^{2\alpha}\log n =\left(\frac{1}{2}\right)^\alpha\sum_{n=1}^\infty \frac1{ n^{2\alpha}}\log n $, so that $S(\alpha)$ converges exactly when $\sum_{n=1}^\infty \frac1{ n^{2\alpha}}\log n $ converges.

I will now show that, if $x > e^{2/\epsilon}(1/(\epsilon\ d))^{2/(\epsilon\ln 2)} $ then $\ln x/x^\epsilon < d$.

This is, of course, far from the best (or even a good) bound, but it will be good enough to show what we want and derive the convergence result for $S(\alpha)$.

Since $e^x > x^m/m! $ for any positive integer $m$, $x/e^x < m!/x^{m-1} \le m^{m-1}/x^{m-1} = (m/x)^{m-1} $.

Therefore, if $x > 2m$, $x/e^x < (1/2)^{m-1} $. Note that the "$2$" here is moderately arbitrary: it can be replaced by any value $> 1$. This is made explicit later on, but I will use $2$ for now.

Replacing $x$ by $\ln x$, if $x > e^{2m}$, $\ln x/x < (1/2)^{m-1} $.

Putting $x^c$ for $x$ (I write $c$ for $\epsilon$ since I'm lazy), for $c > 0$, if $x^c > e^{2m}$, $\ln x^c/x^c = c\ln x/x^c < (1/2)^{m-1} $.

Therefore, for any $c > 0$, if $x > e^{2m/c}$, $\ln x/x^c < (1/2)^{m-1}/c $.

To make $(1/2)^{m-1}/c < d$, for any $d > 0$, we want $2^{m-1} > \frac1{c\ d}$ or $m>1+\ln(\frac1{c\ d})/\ln 2$.

This gives $e^m > e^{1+\ln(1/(c\ d))/\ln 2} = e (1/(c\ d))^{1/\ln 2} $ so $e^{2m/c} > e^{2/c}(1/(c\ d))^{2/(c\ln 2)} $.

Replacing $c$ by $\epsilon$, we get this:

If $x > e^{2/\epsilon}(1/(\epsilon\ d))^{2/(\epsilon\ln 2)} $, then $\ln x/x^\epsilon < d$.

The following makes explicit what I said above about the "$2$" being arbitrary. It also simplifies things.

Choose any $r > 1$.

If $x > rm$, replacing $x$ by $\ln x$, if $x > e^{rm}$, $\ln x/x < (1/r)^{m-1} $.

To make $(1/r)^{m-1} < d$, where $1 > d > 0$, we want $r^{m-1} > 1/d$ or $m > 1+\ln(1/d)/\ln r$. Putting this in the bound for $x$, above, if $x > e^{r(1+\ln(1/d)/\ln r)}$, $\ln x/x < d $. Note that $e^{r(1+\ln(1/d)/\ln r)} =e^r (1/d)^{r/\ln r} $, so the bound for $x$ becomes $x > e^r (1/d)^{r/\ln r}$.

Letting $r = e$, this becomes if $x > (e/d)^e$, $\ln x/x < d $.

To get a bound on $x$ when $\ln x/x^c <d $, where $c > 0$, this becomes $\ln(x^c)/x^c =c\ln x/x^c < cd $. Putting $x^c$ for $x$, the bound is $x^c > e^r (1/d)^{r/\ln r}$ or $x > e^{r/c} (1/d)^{r/(c\ln r)}$. Letting $r = e$, the bound is $x > e^{e/c} (1/d)^{e/c} = (e/d)^{e/c} $.

  • $\begingroup$ Good. Maybe you'll edit out the formatting typos --- I'd do it myself, but I'm worried I'd get the math wrong. $\endgroup$ – Gerry Myerson Aug 10 '13 at 0:49
  • $\begingroup$ I'll do it. Thanks for the approval. OK, I fixed a missing $\$$. I didn't see any others. Feel free to edit it. $\endgroup$ – marty cohen Aug 10 '13 at 1:37

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.