3

It is well-known that the harmonic sum $H_{n}= \sum_{k=1}^{n}\frac{ 1}{k}$ has the following asymptotic expansion for $n\to\infty$

$$H_n = \sum_{k=1}^{n}\frac{1}{k}\sim \gamma+\log \left(n\right)+\frac{1}{2 n}-\frac{1}{12 n^2}+\frac{1}{120 n^4}-\frac{1}{252 n^6}\pm \ldots\tag{1}$$

The alternating harmonic sum is defined as

$$\overline {H}_{n} = \sum_{k=1}^{n}\frac{(-1)^{k+1}}{k}\tag{2}$$

and we ask for its asymptotic expansion.

At first I tried to use the representation

$$\overline{H}_{n} =\log (2)+ (-1)^{n+1} \Phi (-1,1,n+1)\tag{3}$$

where $ \Phi (z,s,a)=\sum_{k=0}^{\infty} \frac{z^k}{(k+a)^s}$ is a special function called Lerch transcendent (https://en.wikipedia.org/wiki/Lerch_zeta_function) which is just the tail of the expansion of $\log(2)$ starting at the $(n+1)$st term. But I couldn't find the asymptotics of $\Phi$. Also Mathematica wouldn't do it.

So I came up with another idea and found

$$\overline{H}_{n} \sim \log(2) -(-1)^n \left(\frac{1}{2n}-\frac{1}{4 n^2} +\frac{1}{8n^4} - \frac{1}{4n^6}+\ldots\right)\tag{4}$$

I have looked up possibly related proofs. This reference contains two of them.

Asymptotic expansion at order 2 of $\int_0^1 \frac{x^n}{1+x} \, dx$

But mine was still different.

What would be your proof?

2 Answers2

3

Repetitive integration by parts: \begin{align} \Phi(-1,1,n+1) &= \int_0^1 \frac{x^n}{x+1} \, {\rm d}x \\ &= \int_0^1 x^{n-1} \frac{x}{x+1} \, {\rm d}x \\ &= \frac{x^n}{n} \, \frac{x}{x+1} \Bigg|_0^1 - \frac{1}{n} \int_0^1 x^{n-1} \left( x \frac{{\rm d}}{{\rm d}x} \right) \frac{x}{x+1} \, {\rm d}x \\ &= \frac{x^n}{n} \, \frac{x}{x+1} \Bigg|_0^1 - \frac{x^{n}}{n^2} \left( x \frac{{\rm d}}{{\rm d}x} \right) \frac{x}{x+1} \Bigg|_0^1 + \frac{1}{n^2} \int_0^1 x^{n-1} \left( x \frac{{\rm d}}{{\rm d}x} \right)^2 \frac{x}{x+1} \, {\rm d}x \\ &=\cdots \\ &=\sum_{k=0}^{N-1} (-1)^k \frac{x^n}{n^{k+1}} \left( x \frac{{\rm d}}{{\rm d}x} \right)^k \frac{x}{x+1} \Bigg|_0^1 + \frac{(-1)^N}{n^N} \int_0^1 x^{n-1} \left( x \frac{{\rm d}}{{\rm d}x} \right)^N \frac{x}{x+1} \, {\rm d}x \, . \end{align}

As far as I'm aware, the last term is problematic. By the identity $$\left( x \frac{{\rm d}}{{\rm d}x} \right)^N \frac{x}{x+1} = \sum_{k=1}^N {N\brace k}k! \, x^k \left(\frac{-1}{x+1}\right)^{k+1}$$ with Stirling numbers of the second kind $\left\{\cdot\right\}$, the last integral can be bounded $$\left|\int_0^1 x^{n-1} \left( x \frac{{\rm d}}{{\rm d}x} \right)^N \frac{x}{x+1} \, {\rm d}x\right| \leq \sum_{k=1}^N {N \brace k} (k-1)! \left(1-2^{-k}\right) \leq \sum_{k=0}^N {N\brace k}k! = a(N)$$ where $a(N)$ is the ordered Bell number. Its asymptotics $$a(N) \sim \frac{N!}{2(\log 2)^{N+1}}$$ show that the series is only asymptotic.

Hence, $$\bar{H}_n = \log 2 + (-1)^{n+1} \int_0^1 \frac{x^n}{x+1} \, {\rm d}x \\ =\log 2 + \frac{(-1)^{n+1}}{2n} + (-1)^n \sum_{k=1}^{N-1} \frac{(-1)^k}{n^{k+1}} \sum_{l=1}^k {k\brace l} l! (-1)^{l} 2^{-l-1} + {\cal O}(n^{-N-1})$$ for fixed $N$ and $n\rightarrow \infty$.

E.g. for $N=16$ this becomes $$\bar{H}_n = \log 2 + (-1)^{n+1} \left( \frac{1}{2n} - \frac{1}{4n^2} + \frac{1}{8n^4} - \frac{1}{4n^6} + {\frac {17}{16\,{n}^{8}}}-{\frac {31}{4\,{n}^{10}}}+{\frac {691}{8\,{n}^{12}}}-{ \frac {5461}{4\,{n}^{14}}}+{\frac {929569}{32\,{n}^{16}}} \right) \, .$$


Another way (which is however similar to the link you gave) I came up with is $(x=e^{-u/n})$: \begin{align} \int_0^1 \frac{x^n}{x+1} \, {\rm d}x &= \frac{1}{n} \int_0^\infty \frac{e^{-u}}{e^{u/n}+1} \, {\rm d}u \\ &=\frac{1}{n} \int_0^\infty {\rm d}u \, e^{-u} \sum_{k=0}^\infty (-1)^k e^{-u(k+1)/n} \\ &=\frac{1}{n} \int_0^\infty {\rm d}u \, e^{-u} \sum_{k=0}^\infty (-1)^k \sum_{m=0}^\infty \frac{\left(-u(k+1)/n\right)^m}{m!} \\ &=\frac{1}{n} \int_0^\infty {\rm d}u \, e^{-u} \sum_{m=0}^\infty \frac{\left(-u/n\right)^m}{m!} \, \eta(-m) \\ &=\sum_{m=0}^\infty \frac{\left(-1\right)^m \, \eta(-m)}{n^{m+1}} \end{align} where $\eta$ is the Dirichlet $\eta$-function, evaluated by analytic continuation which destroys convergence and makes it only an asymptotic series.


A third method I managed was by contour integration. The basic principle is to express the denominator of the above integrand as $$\frac{1}{n} \, \frac{1}{e^{u/n}+1} = \frac{1}{2\pi i} \oint_C \frac{1}{e^{s}+1} \, \frac{{\rm d}s}{ns-u}$$ where $C$ is the contour encircling the positive $x$-axis including $0$ counter-clockwise, while leaving out the poles of $\frac{1}{e^s+1}$. The evaluated $u$-integral gives $-e^{-ns} {\rm Ei_1}(-ns)$ and it remains to calculate $$\frac{-1}{2\pi i} \oint_C \frac{{\rm Ei_1}(-ns)}{e^s+1} \, e^{-ns} \, {\rm d}s \, .$$ The trick now is to use ${\rm Ei_1}(z) = {\rm Ein}(z) - \ln(z) - \gamma$, that is since ${\rm Ein}$ is holomorphic the only contribution comes from the logarithm at the cut. Therefore, repeatedly integrating by parts, the last expression becomes \begin{align} \frac{1}{2\pi i}\oint_C \frac{\ln(-s)}{e^s+1} \, e^{-ns} \, {\rm d}s &= \frac{1}{2\pi i \, n}\oint_C e^{-ns} \left(\frac{1}{s} + \ln(-s) \frac{{\rm d}}{{\rm d}s} \right) \frac{1}{e^s+1} {\rm d}s \\ &=\frac{1}{2\pi i \, n}\oint_C e^{-ns} \left(\frac{1}{s} + \frac{1}{ns} \frac{{\rm d}}{{\rm d}s} + \frac{\ln(-s)}{n} \frac{{\rm d}^2}{{\rm d}s^2} \right) \frac{1}{e^s+1} {\rm d}s \\ &= \dots \\ &= \frac{1}{2\pi i \, n}\oint_C e^{-ns} \frac{{\rm d}s}{s} \sum_{k=0}^\infty \frac{1}{n^k} \frac{{\rm d}^k}{{\rm d}s^k} \frac{1}{e^s+1} \\ &= \sum_{k=0}^\infty \frac{1}{n^{k+1}} \frac{{\rm d}^k}{{\rm d}s^k} \frac{1}{e^s+1} \Bigg|_{s=0} \, . \end{align}

Of course the same result would have been immediately obtained by the Borel sum of the Borel transform $\frac{1}{e^s+1}$; $$\int_0^\infty \frac{e^{-u}}{e^{u/n}+1} \, {\rm d}u = \int_0^\infty {\rm d}u \, e^{-u} \sum_{k=0}^\infty \frac{u^k}{n^k} \, \frac{\frac{{\rm d}^k}{{\rm d}s^k} \frac{1}{e^s+1} \Big|_{s=0}}{k!} = \sum_{k=0}^\infty \frac{1}{n^k} \frac{{\rm d}^k}{{\rm d}s^k} \frac{1}{e^s+1} \Bigg|_{s=0} \, .$$

Diger
  • 6,197
  • @ Diger Thank you. But you didn't tell the beginning of the story. Also this proof is a repetition of that by Olivier Oloa and Mark Viola who also include a strict evaluation of the remaining term which can be found in the quoted reference of related proofs. – Dr. Wolfgang Hintze Feb 29 '20 at 16:13
  • What beginning? – Diger Feb 29 '20 at 16:21
  • Can you link that reference? I can't find it. The only thing I find is a similar - but not the same - approach by Olivier Oloa here: https://math.stackexchange.com/questions/1589789/asymptotic-expansion-at-order-2-of-int-01-fracxn1x-dx – Diger Feb 29 '20 at 16:23
  • In fact - correct me if I'm wrong - that link only asks for order 2, this is order arbitrary. – Diger Feb 29 '20 at 16:27
  • I mean asymptotic still means, that for any fixed $N$ and $n\rightarrow \infty$ the error term vanishes as $n^{-N-1}$, which is trivially seen. It just doesn't converge for fixed $n$ and $N\rightarrow \infty$. – Diger Feb 29 '20 at 16:46
  • @ Diger I appreciate your effort. My answers 1) The "beginning" is the problem I have formulated. 2) I had assumed that you have read the post of the problem. There you find the link to related proofs, and 3) I had hoped to see other proofs. – Dr. Wolfgang Hintze Feb 29 '20 at 17:40
  • What do you mean by other proofs? Can you show me the link for which my solution is a repetition? – Diger Feb 29 '20 at 17:49
  • @ Dinger Ok, today is my tolerance day, hence for the third (!) time, now step by step: go to the third but last line of my OP, make click on the blue link. Open it, read it, and ... voliá you see two different proofs. One of these was repeated in your proof, surely unintended, but still ... 5 minutes ago I saw you extended proof, and I have upvoted it. – Dr. Wolfgang Hintze Feb 29 '20 at 18:39
  • You must be clearly blind. The proof by PI from Olivier Oloa is done in a completely different way. Go check it out. You clearly state you couldn't find the asymptotics of $\Phi$, which I'm merely doing above; namely purely in powers of $1/n$. The proof from Olivier does not yield any power expansion; it is by inverse pochhammer i.e. $\frac{1}{(n+1)(n+2)\cdots}$. – Diger Feb 29 '20 at 18:39
  • @ Diger I suggest you return to decent behaviour. The series in terms of Pochhammer powers is trivially converted into inverse powers on $n$. The main idea is partial integration, as was done by Oloa et. al. That's it. – Dr. Wolfgang Hintze Feb 29 '20 at 18:47
  • Whatever... I guess you should clearly specify what you want! "But mine was still different." Different??? How did you proceed/How did you arrive at your result? "What would be your proof" What do you expect??? Saying you don't know how to obtain the asymptotics, to me, implies that you'll be looking for something like that... – Diger Feb 29 '20 at 19:24
  • @ Diger There is an old saying "He who can read is better off". I have stated clearly in the OP that I have a proof and I gave a link to two other different proofs, and that I'm waiting for new ideas. Yours was not new but was contained in the link I gave. And now, if you don't mind, I would like to return to interesting topics. – Dr. Wolfgang Hintze Feb 29 '20 at 19:40
  • Not sure if you genius care, but I added 2 further methods. – Diger Feb 29 '20 at 23:53
  • @ Diger IMHO only the contour integration approach which reformuates the complete solution of @Start wearing purple is novel. I repeat: I appreciate your effort, and that's why I upvoted your contribution. Have a nice sunday. – Dr. Wolfgang Hintze Mar 01 '20 at 09:10
1

My idea was to express $\overline{H}_k$ by $H_k$ and then use the asyptotic expansion of $H_k$.

Indeed, $\overline{H}_n$ can be expressed as follows ($m=1,2,3,\ldots$}

$$\overline{H}_{2m} = H_{2m} -H_{m}\tag{5a}$$

$$\overline{H}_{2m+1} = H_{2m+1} -H_{m}\tag{5b}$$

The (simple) proof is left as an exercise to the reader.

For the asymptotic expressions of the even version we find from $(1)$

$$\overline{H}_{2m}\overset{m\to\infty,m->\frac{n}{2}} = \log (2) \\-\frac{1}{2 n}+\frac{1}{4 n^2}-\frac{1}{8 n^4}+\frac{1}{4 n^6} -\frac{17}{16 n^8}\pm\ldots\tag{6a}$$

For the odd version we have, to begin with,

$$\overline{H}_{2m+1}\overset{m\to\infty, m->\frac{n-1}{2}}=\log (2) \\ +\frac{1}{2 (n-1)}-\frac{3}{4 (n-1)^2}+\frac{1}{(n-1)^3}-\frac{9}{8 (n-1)^4}+\frac{1}{(n-1)^5}-\frac{3}{4 (n-1)^6} \\ +\frac{1}{(n-1)^7}-\frac{33}{16 (n-1)^8}+\frac{1}{(n-1)^9}\mp\ldots$$

Taking the asymptotics of this in turn we get

$$\overline{H}_{2m+1}\overset{m\to\infty, m->\frac{n-1}{2}}=\log (2)\\+ \frac{1}{2 n}-\frac{1}{4 n^2}+\frac{1}{8 n^4}-\frac{1}{4 n^6}+\frac{17}{16 n^8}\mp\ldots\tag{6b}$$

Finally, combining $(6a)$ and $(6b)$ gives the expression $(4)$ of the OP.

Combining this with $(3)$ we have also derived the asymptotics of the Lerch $\Phi$ function from that of the harmonic number.