8
$\begingroup$

This is a variant of Exercise 64 in Chapter 9 of concrete mathematics.

Prove the following identity \begin{equation} \sum_{n = -\infty}^{\infty}' \frac{1 - \cos( 2\pi n k )}{n^2 } = 2 \pi^2 ( k - k^2 ) \qquad k \in [0,1] \end{equation} I came across it in a different context and was surprised that it is exact.

Let's use Euler-Maclaurin expansion to convert the sum into an integral.

Take $f(x) = \frac{1 - \cos( 2\pi x k )}{x^2 } $ \begin{equation} \sum_{n = -N+1}^{N}' f(n) = \int_{-N}^{N} f(x) dx - f(0) + \sum_{k=1}^{p} \frac{B_k}{k!}f^{(k-1)}(x)\Big|^N_{-N} + R_p \end{equation} The first two terms are already the exact result, \begin{equation} \int_{-\infty}^{\infty} f(x) dx = 2\pi k \int_{-\infty}^{\infty} \frac{1- \cos x}{x^2}dx = 2\pi^2 k \qquad f(0) = 2\pi^2 k^2 \end{equation} so one needs to prove the end point corrections(terms with $B_k$ in it) as well as the reminder terms are zero.

At any order $p$, there are only finite number of end point correction terms which goes at most like $\mathcal{O}(\frac{1}{N^2})$, so they vanish when taking the $N\rightarrow \infty$ limit. But I don't know how to show the reminder term is zero(as $N\rightarrow \infty$), \begin{equation} R_p(N) = (-1)^p \int_{-N}^{N} \frac{1}{p!} B_p( x - \lfloor x \rfloor ) f^{(p)}(x) dx \end{equation} Maybe there are other methods to prove the identity, but I personally would appreciate the proof using Euler Maclaurin, since I'm going to use it for another series, and the reminder term has \begin{equation} f(x) = \frac{g(x) - g(0) }{ x^2} - \frac{g'(0)}{x} \end{equation} where $g(x ) = g(x + N)$ is a periodic function. So you can also go ahead and prove the general $g(x)$ case.

Thanks.

Edit: The general case can also be worked out by robjohn 's method by summing over each individual the Fourier components. Though a direct proof of the vanishing Euler-Maclaurin reminder term is still absent, robjohn's method solved my problem.

$\endgroup$
1
  • $\begingroup$ What exactly does $\sum'$ mean? Is it $\lim\sum_{-N}^N$? $\endgroup$ Commented May 20 at 19:17

3 Answers 3

6
$\begingroup$

The identity is equivalent to $$\sum_{n\ge 1} \frac{1-\cos(2\pi nx)}{n^2} = \pi^2 x(1-x).$$

The sum term $$S(x) = \sum_{n\ge 1} \frac{1-\cos(2\pi nx)}{n^2}$$ is harmonic and may be evaluated by inverting its Mellin transform.

Recall the harmonic sum identity $$\mathfrak{M}\left(\sum_{k\ge 1} \lambda_k g(\mu_k x);s\right) = \left(\sum_{k\ge 1} \frac{\lambda_k}{\mu_k^s} \right) g^*(s)$$ where $g^*(s)$ is the Mellin transform of $g(x).$

In the present case we have $$\lambda_k = \frac{1}{k^2}, \quad \mu_k = k \quad \text{and} \quad g(x) = 1-\cos(2\pi x).$$

We need the Mellin transform $g^*(s)$ of $g(x).$

Now the Mellin transform of $\cos(x)$ was computed at this MSE link and found to be $$\Gamma(s) \cos(\pi s/2)$$ and therefore $$g^*(s) = -\frac{1}{(2\pi)^s} \Gamma(s) \cos(\pi s/2)$$ with fundamental strip $\langle -2, 0 \rangle$ (the constant one in $g(x)$ shifts the fundamental strip).

Hence the Mellin transform $Q(s)$ of $S(x)$ is given by $$ Q(s) = -\frac{1}{(2\pi)^s} \Gamma(s) \cos(\pi s/2) \zeta(s+2) \quad\text{because}\quad \sum_{k\ge 1} \frac{\lambda_k}{\mu_k^s} = \zeta(s+2)$$ where $\Re(s+2) > 1$ or $\Re(s) > -1$.

Intersecting the fundamental strip and the half-plane from the zeta function term we find that the Mellin inversion integral for an expansion about zero is $$\frac{1}{2\pi i} \int_{-1/2-i\infty}^{-1/2+i\infty} Q(s)/x^s ds$$ which we evaluate in the left half-plane $\Re(s)<-1/2.$

The cosine term cancels the poles of the gamma function term at odd negative integers and the zeta function term the poles at even negative integers. We are left with just three poles.

$$\begin{align} \mathrm{Res}(Q(s)/x^s; s=0) & = -\frac{\pi^2}{6} \\ \mathrm{Res}(Q(s)/x^s; s=-1) & = \pi^2 x \quad\text{and}\\ \mathrm{Res}(Q(s)/x^s; s=-2) & = -\pi^2 x^2. \end{align}$$

The pole at $s=0$ is outside the left half-plane of evaluation. Hence in a neighborhood of zero, $$S(x) = \pi^2 x(1-x)$$ as claimed.

To see that this is exact put $s= \sigma + it$ with $\sigma \le -5/2$ where we seek to evaluate $$\frac{1}{2\pi i} \int_{-5/2-i\infty}^{-5/2+i\infty} Q(s)/x^s ds.$$

Recall that with $\sigma > 1$ and for $|t|\to\infty$ we have $$|\zeta(\sigma+it)| \in \mathcal{O}(1).$$

Furthermore recall the functional equation of the Riemann Zeta function $$\zeta(1-s) = \frac{2}{2^s\pi^s} \cos\left(\frac{\pi s}{2}\right) \Gamma(s) \zeta(s)$$ which we re-parameterize like so $$\zeta(s+2) = 2\times (2\pi)^{s+1} \cos\left(-\frac{\pi (s+1)}{2}\right) \Gamma(-s-1) \zeta(-s-1)$$ which is $$\zeta(s+2) = -2\times (2\pi)^{s+1} \sin(\pi s/2) \frac{\Gamma(1-s)}{s(s+1)} \zeta(-s-1).$$

Substitute this into $Q(s)$ to obtain $$\frac{1}{(2\pi)^s} \Gamma(s) \cos(\pi s/2) \times 2 \times (2\pi)^{s+1} \sin(\pi s/2) \frac{\Gamma(1-s)}{s(s+1)} \zeta(-s-1).$$

Use the reflection formula for the Gamma function to obtain $$\frac{1}{(2\pi)^s} \cos(\pi s/2) \times 2 \times (2\pi)^{s+1} \sin(\pi s/2) \times \frac{\pi}{\sin(\pi s)} \frac{1}{s(s+1)} \zeta(-s-1),$$ in other words we have $$Q(s) = 2\pi^2 \frac{\zeta(-s-1)}{s(s+1)}.$$

This finally implies (with $\sigma< -5/2$ we have $\Re(-s-1) > 3/2$) $$|Q(s)/x^s|\sim x^{-\sigma} |t|^{-2}.$$

We see from the term in $|t|$ that the integral obviously converges. (This much we knew already.) Moreover, when $x\in(0,1)$ we have $x^{-\sigma}\to 0$ as $\sigma\to -\infty.$ The term in $x$ does not depend on the variable $t$ of the integral and may be brought to the front. This means that the contribution from the left side of the rectangular contour that we employ as we shift to the left vanishes in the limit, proving the exactness of the formula for $S(x)$ obtained earlier.

There is a theorem hiding here, namely that certain Fourier series can be evaluated by inverting their Mellin transforms which is not terribly surprising and which the reader is invited to state and prove.

$\endgroup$
3
  • $\begingroup$ Thanks for your fantastic answer. I think your main idea is to use Merlin/inverse Merlin transform of $S(x)$ to convert the sum to an contour integral. The "harmonic sum identity" and the ways to control the integrand in inverse Merlin transform is rather technical for me. I personally would like to see why the reminder term is zero in Euler-Maclaurin, because that seems to be more "elementary" to me. $\endgroup$
    – anecdote
    Commented Oct 17, 2014 at 13:52
  • $\begingroup$ These are kind words that might perhaps translate into an upvote? (This does not prevent you from accepting or upvoting an additional answer that is better suited to the intent of your question.) $\endgroup$ Commented Oct 17, 2014 at 20:32
  • $\begingroup$ That indeed deserves an upvote. I'm hoping to have an Euler-Maclaurin proof, because I think the exercise in Concrete Mathematics should be solvable by undergraduate level calculus, unfortunately I can't. $\endgroup$
    – anecdote
    Commented Oct 17, 2014 at 22:18
5
$\begingroup$

Let's start with $$ \begin{align} \sum_{n=1}^\infty\frac{\sin(2\pi nx)}{n} &=\mathrm{Im}\left(\sum_{n=1}^\infty\frac{e^{2\pi inx}}n\right)\\ &=\mathrm{Im}\left(-\log\left(1-e^{2\pi ix}\right)\right)\\ &=-\arctan\left(\frac{-\sin(2\pi x)}{1-\cos(2\pi x)}\right)\\ &=\arctan\left(\frac{2\sin(\pi x)\cos(\pi x)}{2\sin^2(\pi x)}\right)\\[9pt] &=\arctan(\cot(\pi x))\\[12pt] &=\frac\pi2\mathrm{sgn}(x)-\pi x\tag{1} \end{align} $$ for $x\in(-1,1)$.

Integrate $(1)$ and multiply by $2\pi$ to get $$ \sum_{n=1}^\infty\frac{1-\cos(2\pi nx)}{n^2}=\pi^2|x|-\pi^2x^2\tag{2} $$ Since the summand is even in $n$, we have $$ \sum_{n=-\infty}^\infty'\frac{1-\cos(2\pi nx)}{n^2}=2\pi^2|x|-2\pi^2x^2\tag{3} $$

$\endgroup$
7
  • $\begingroup$ Impressive! But there is a flaw in equation (1). When I set $x =0$ for equation (1), I got 0 on the LHS, but the RHS is $\frac{\pi}{2}$. $\endgroup$
    – anecdote
    Commented Oct 23, 2014 at 3:29
  • $\begingroup$ OK you have restrict the range of $x$ to be $(0,1)$, apparently that's because $x = 0$ will result in a divergence in the Tyler expansion of $\ln(1+x)$. The radius of convergence is $1$ and $\exp( 2\pi i x )$ is actually on the verge of the convergent ring. $\endgroup$
    – anecdote
    Commented Oct 23, 2014 at 3:39
  • $\begingroup$ @anecdote: Yes. You can set the domain to be $(-1/2,1/2)$, but then $(1)$ becomes $\frac\pi2\mathrm{sgn}(x)-\pi x$. Then $x=0$ gives $0$ on both sides. Actually, this works for $x\in(-1,1)$. Perhaps I will update my answer. $\endgroup$
    – robjohn
    Commented Oct 23, 2014 at 6:34
  • 1
    $\begingroup$ There are several subtleties. One is the convergence of the series $\sum_{n=1}^{\infty}\frac{e^{2\pi i n x}}{n}$. The second is the sector you choose for $\ln z$. The third is why can you do a integration on each term when the sum is only conditionally convergent. $\endgroup$
    – anecdote
    Commented Oct 23, 2014 at 14:26
  • 1
    $\begingroup$ Thanks, that makes sense for the first two questions. How do you justify the term by term integration for equation (1)? It's not uniformly convergent, since the RHS of (1) is not continuous. $\endgroup$
    – anecdote
    Commented Oct 23, 2014 at 16:44
2
$\begingroup$

$\newcommand{\angles}[1]{\left\langle\, #1 \,\right\rangle} \newcommand{\braces}[1]{\left\lbrace\, #1 \,\right\rbrace} \newcommand{\bracks}[1]{\left\lbrack\, #1 \,\right\rbrack} \newcommand{\dd}{{\rm d}} \newcommand{\ds}[1]{\displaystyle{#1}} \newcommand{\dsc}[1]{\displaystyle{\color{red}{#1}}} \newcommand{\expo}[1]{\,{\rm e}^{#1}\,} \newcommand{\half}{{1 \over 2}} \newcommand{\ic}{{\rm i}} \newcommand{\imp}{\Longrightarrow} \newcommand{\Li}[1]{\,{\rm Li}_{#1}} \newcommand{\pars}[1]{\left(\, #1 \,\right)} \newcommand{\partiald}[3][]{\frac{\partial^{#1} #2}{\partial #3^{#1}}} \newcommand{\root}[2][]{\,\sqrt[#1]{\vphantom{\large A}\,#2\,}\,} \newcommand{\totald}[3][]{\frac{{\rm d}^{#1} #2}{{\rm d} #3^{#1}}} \newcommand{\verts}[1]{\left\vert\, #1 \,\right\vert}$ $\ds{\left.\sum_{n=-\infty}^{\infty}\frac{1 - \cos\pars{2\pi nk}}{n^{2}} \right\vert_{\ n\ \not=\ 0}\ =\ 2\pi^{2}\pars{k - k^2}:\ {\large ?}.\qquad k \in \bracks{0,1}}$


\begin{align}&\color{#66f}{\large% \left.\sum_{n=-\infty}^{\infty}\frac{1 - \cos\pars{2\pi nk}}{n^{2}} \right\vert_{\ n\ \not=\ 0}} =\sum_{n=0}^{\infty}\,{\rm f}\pars{n,k} - 4\pi^{2}k^{2} \\[5mm]&\mbox{where}\quad \,{\rm f}\pars{x,k}\equiv \left\{\begin{array}{lcl} 2\,\dfrac{1 - \cos\pars{2\pi xk}}{x^{2}} & \mbox{if} & x \not= 0 \\[2mm] 4\pi^{2}k^{2} & \mbox{if} & x = 0 \end{array}\right. \end{align} A straightforward application of the Abel-Plana Formula yields: \begin{align}&\color{#66f}{\large% \left.\sum_{n=-\infty}^{\infty}\frac{1 - \cos\pars{2\pi nk}}{n^{2}} \right\vert_{\ n\ \not=\ 0}} =\overbrace{\int_{0}^{\infty}\,{\rm f}\pars{x,k}\,\dd x} ^{\dsc{2\pi^{2}k}}\ +\ \overbrace{\half\,{\rm f}\pars{0,k}}^{\dsc{2\pi^{2}k^{2}}}\ -\ 4\pi^{2}k^{2} \\[5mm]&=\color{#66f}{\large 2\pi^{2}\pars{k - k^2}} \end{align}

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .