20

I saw this amazing identity elsewhere : $$ \bbox[5px,border:2px solid #C0A000]{\sum_{n\in\mathbb Z} \tan^{-1} (an+b) =\lim_{N\to \infty} \sum_{-N}^N \tan^{-1} (an+b) = \tan^{-1} \left( \tan\frac{b\pi}{a} \cdot \coth \frac{\pi}{a} \right) +\pi\text{sgn}(a)\bigg( \text{ceil}\left(\frac ba+\frac 12\right) -1 \bigg)}$$

for $a\ne 0$. Here, the principal branch of $\tan^{-1} x$ is taken, i.e. $\left(-\frac{\pi}{2}, \frac{\pi}{2}\right)$.

I feel like proving this involves the use of Euler’s sin product in some way. I tried integrating w.r.t $r$, the identity $$\sum_{\mathbb Z} \frac{1}{(n+r)^2 +x^2}=\frac{\pi\sinh(2\pi x)}{x(\cosh(2\pi x) -\cos(2\pi r))} $$ using the Weierstrass substitution to get $$\sum_{\mathbb Z}\tan^{-1} \left( \frac{n+r}{x} \right) =\tan^{-1} \left( \tan\pi r \cdot \coth \pi x \right) +C $$ I could then change variables, $\frac 1x \mapsto a$ and $\frac rx \mapsto b $. The problem is with the $+C$: it should equal the second term of the RHS, but I don’t know how I could show that. Any ideas for a proof?

Vishu
  • 14,469
  • 2
    Replace $\tan^{-1}$ by a sum of two $\log$ then use the sine product. The second $sign$ term comes from the branches of $\log$. @JeanMarie – reuns Apr 29 '21 at 22:14
  • If it helps: The identity "elsewhere" for a fixed parameter "a" can be found here: Dr. Andreas Dieckmann – stocha May 05 '21 at 08:49
  • $$\sum _{k=-\infty }^{\infty } \text{ArcTan}[x+k]=\text{ArcTan}[\text{Coth}[\pi ] \text{Tan}[\pi x]]+\pi \text{Floor}\left[x+\frac{1}{2}\right]$$ can be verified by differentiating on both sides, summation and integration – stocha May 05 '21 at 10:00
  • @stocha Interesting, I’m seeing that page for the first time. Also, wouldn’t that second term vanish under differentiation? – Vishu May 05 '21 at 10:02
  • @Tavish: I did the same method and proved the identity. My problem, I don't have the time to write it down. Therefore I will give you the instruction for the proof. The trick is to differentiate with respect to "b" or just replace b by x. Mathematica give you two expressions, the first expression for two cases for choosing the parameter x. Both give no contribution. If you integrate the last expression, you can proof the equivalence between the expression on the left hand and right hand side. You only have to differentiate the summand to "b" or "x" and sum it up, then integrate it. – stocha May 05 '21 at 11:49
  • Mathematica can't evaluate it, it is a bit tricky, but I guess you only have to do some transformation. The same method can be used for the simpler expression. Here the integral after summation can be done easily. – stocha May 05 '21 at 11:50
  • @stocha Let me see if I understand correctly. I differentiate the equality in my post with respect to $b$, then evaluate the sum on the left side, then integrate both sides and I should end up with exactly a tautology, i.e. $S=S$ where $S$ is the current RHS? – Vishu May 05 '21 at 11:56
  • Yes this works directly with the simpler expression a=1, for any a I did not manage to completely solve the integral on the left hand side after summation, but I guess only some transformation have to be done. otherwise you just need to evaluate the left hand side numerical – stocha May 05 '21 at 12:00
  • Just try this procedure with the simpler expression first to get an idea – stocha May 05 '21 at 12:01
  • But the problem is that we lose the second term of the RHS on differentiation, so how do I retrieve it? – Vishu May 05 '21 at 12:15
  • 1
    @Tavish: As soon I have time, I will post my solution! – stocha May 05 '21 at 12:21
  • The second term is not lost, one has to study besides of the indefinite Integral also the definite integral e.g. from 0 to 1/2 +$\epsilon$ on both sides in order to recover the second term. It is the same problem like Dr. Wolfgang Hintze – stocha May 05 '21 at 17:33

1 Answers1

6

Disclaimer: This is far from a completed answer. It contains

A) A proof of convergence that I felt necessary to establish at first (I didn't considered it as evident) providing as a side result, an understanding of the second term of the RHS part:

$$\pi \bigg( \text{ceil}\left(\frac ba+\frac 12\right) -1 \bigg) \tag{1}$$

(in which we have assumed $a>0$ WLOG).

B) The use of integral formula

$$\tan^{-1}\left(\tfrac{u}{t}\right)=\int_0^{\infty}\dfrac{e^{-xt}}{x}\sin(ux)dx\tag{2}$$

that looks promising but is still with the status "work in progress".

Part A:

Let

$$t_n:=\tan^{-1} (an+b)$$

Grouping symmetrical terms, one gets

$$t_{-n}+t_n= \tan^{-1}\dfrac{2b}{1-b^2+a^2n^2}+R_n\tag{3} $$

due to relationship (that can be understood with the following figure)

$$\tan^{-1}A+\tan^{-1}B=\tan^{-1}\dfrac{A+B}{1-AB}+R\tag{4}$$

where $R=-\pi,0,\pi$ according to the region in which $(A,B)$ is situated.

enter image description here

Fig. 1: Graphical representation of $R=\tan^{-1}A+\tan^{-1}B-\tan^{-1}\frac{A+B}{1-AB}$, with "plateaus" for $R=-\pi,0,\pi$. This is established in different references such as this one.

This "residual" $R_n$ is $0$ if

  • if $A$ and $B$ have the same sign with $AB<1$ or

  • if $A$ and $B$ haven't the same sign ; we are in this case here, when is $n$ large enough (as it is a proof of convergence, we can drop some terms without harm).

Otherwise said, $R_n$ becomes zero in (3) when $n$ is such that:

$$(an+b)(-an+b)<0 \ \iff \ b^2-n^2a^2<0$$

which amounts to say

$$n\ge n_0 \ \text{where} \ n_0:=\operatorname{ceil} \big|\tfrac{b}{a}\big|\tag{5}$$

In this way, it remains to sum up a series equivalent to the Riemann series $\sum 1/n^2$, therefore convergent.

Remark: Please note the close relationship between (5) and (1): (1) accounts for the "exceptional" terms $t_1+t_{-1}$, $t_2+t_{-2},...$ each one contributing with a $R_1=R_2=... = \pi$.

Remark: There is an exceptional term $t_0=\tan^{-1}(b)$ that will have to be taken into account later...

Part B: In order to "approach" expression

$$\tan^{-1} \left( \tan\frac{b\pi}{a} \cdot \coth \frac{\pi}{a} \right),\tag{6}$$

let us use formula (2) with $t=1$ and $u=an+b$: giving:

$$t_n:=\tan^{-1}(an+b)=\int_0^{\infty}\dfrac{e^{-x}}{x}\sin((an+b)x)dx$$

from which

$$I_n:=t_n+t_{-n}=2 \int_0^{\infty}\dfrac{e^{-x}}{x}\sin(bx)\cos(2anx)dx$$

otherwise expressed with the real part function $\Re$:

$$I_n=2 \int_0^{\infty}\dfrac{e^{-x}}{x}\sin(bx)\Re (e^{2iax})^n dx\tag{7}$$

Set apart the exceptional terms that we have considered just before, we need to compute the following sum (where $n_0$ has been defined in (5)):

$$S:=\sum_{n=n_0}^{\infty}I_n=2 \int_0^{\infty}\dfrac{e^{-x}}{x}\sin(bx)\Re \left(e^{2ian_0x}\dfrac{1}{1-e^{2iax}}\right)dx\tag{8}$$

$$S= \int_0^{\infty}\dfrac{e^{-x}}{x}\sin(bx) \left(\dfrac{\sin((2n_0-1)ax)}{\sin(ax)}\right)dx\tag{9}$$

It remains to match expressions (9) and (6)...

Please note that the parenthesized fraction in (9) can be expressed as $U_{2n_0-2}(\cos(ax))$ (Chebyshev polynomial of second kind), therefore doesn't need special care at the (removable) poles $x=\tfrac{k\pi}{a}$.

Jean Marie
  • 81,803
  • 2
    This shows the series converges when viewed as the symmetric limit but how does it address OPs request for proof of the closed-form? – Integrand Apr 29 '21 at 17:29
  • 1
    @FearfulSymmetry It is why you find on the first line "proof of convergence" (a kind of disclaimer) which, IMHO, must be first adressed before going further. Besides, see the Edit I just wrote indicating a classical mean to cope with series having their general term in $atan(...)$. – Jean Marie Apr 29 '21 at 17:36
  • 1
    With respect, I don't feel like your post answers the question. OP is not asking why the series converges, OP is asking why it converges to that expression. – Integrand Apr 29 '21 at 17:38
  • @JeanMarie: a useful complement to the full answer. – A rural reader Apr 29 '21 at 22:54
  • 1
    I fully agree. While you were sending this comment, I had already cancelled this proposition in the new "revamped" version, still "work in progress" as you will see. – Jean Marie Apr 30 '21 at 09:56
  • 1
    My idea was to calculate the InverseLaplacetransform of the expression tan1^(-1)(a*n+b), where Mathematica finds a closed form solution of the sum and then show that the Laplacetransform results in the expression of the right hand side. – stocha May 04 '21 at 09:48
  • @stocha But variable $n$ in term $\tan^{-1}(an+b)$ isn't a continuous variable... – Jean Marie May 04 '21 at 10:17
  • 1
    Now the idea is to use the Poisson Summation Formula: see e.g. Dr. Wolfgang Hintze and my suggestions. I'm working on proof to show an alterantive form of the Poisson Summtion Formula by the Laplacetransformation, but I did not post it yet. In this case the solution was the right hand side of the functional equation, similar to the famous theta functional equation. – stocha May 04 '21 at 10:33
  • 1
    @stocha Very interesting – Jean Marie May 04 '21 at 11:46