5

I am writing perturbation theory for some linear operator, finding the first-order corrections to the components of eigenvectors boils down to the following sum: $$\sum\limits_{{}^{m=1}_{m\ne n}}^{N} (1 + (-1)^{m+n}) \frac{\sin \left( \frac{\pi m k}{N+1} \right) \sin \left( \frac{\pi m}{N+1} \right)}{\cos\left(\frac{\pi n}{N+1}\right) - \cos\left(\frac{\pi m}{N+1}\right)},$$ where $k,n=1,2,\ldots,N$, and $N$ are integers. It is really desirable that this answer has a simple analytical expression, as otherwise it has no use, but at the moment it does not look like exactly summable.

Does anyone have ideas on how to approach this? Even the answer in the limit of large $N \gg 1$ would be nice.

UPDATE: I think this is impossible. Here is why. Let us just ignore the factor in front for simplicity (it will erase half of the terms simply), and our goal is to get rid from this difference in the denominator. Assume the part of the sum for which $m>n$. In this case $\cos(\pi n/(N+1))>\cos(\pi m/(N+1))$, and we can write:

$$ \frac{1}{\cos\left(\frac{\pi n}{N+1}\right)} \sum\limits_{m=n+1}^{N} \sin \left( \frac{\pi m k}{N+1} \right) \sin \left( \frac{\pi m}{N+1} \right) \frac{1}{1 - \frac{\cos\left(\frac{\pi m}{N+1}\right)}{\cos\left(\frac{\pi n}{N+1}\right)}} = \\\frac{1}{\cos\left(\frac{\pi n}{N+1}\right)} \sum\limits_{l=0}^{\infty} \cos^{-l}\left(\frac{\pi n}{N+1}\right) \sum\limits_{m=n+1}^{N} \sin \left( \frac{\pi m k}{N+1} \right) \sin \left( \frac{\pi m}{N+1} \right) \cos^l\left(\frac{\pi m}{N+1}\right).$$ Obviously, for $n>m$ one has to do the inverse, and perform a similar expansion of the $\frac{1}{1-x}$ function.

The only option one has is to convert all trigonometric functions into exponents. $\cos^l\left(\frac{\pi m}{N+1}\right) = \frac{1}{2^l} \sum\limits_{p=0}^{l} C(l,p) e^{i m (l-p) \phi } e^{-i p m \phi },$ where $C(l,p)$ are the binomial coefficients. So, one can get rid of the summation over $m$, but will be left with other two (over $p$, and $l$), and it does not look like any special function or smth like this.

UPDATE OF 27.07.2021: As metamorphy has proven, I was wrong, and it is possible to find a closed-form solution :). The final answer is:

$$ - \sin\left( \dfrac{\pi n k}{N+1} \right) \cot\left( \dfrac{\pi n}{N+1}\right) - (2k-N-1) \cos\left( \dfrac{\pi n k}{N+1} \right) $$

Sl0wp0k3
  • 516

1 Answers1

4

Using $2\sin A\sin B=\cos(A-B)-\cos(A+B)$, we see that the sum is $$\frac12\big(S_{N+1,n,k-1}-S_{N+1,n,k+1}+(-1)^n(S_{N+1,n,N+k}-S_{N+1,n,N+k+2})\big)$$ where, for $0<n<N$ and $0\leqslant k\leqslant 2N$, $$S_{N,n,k}=\sum_{\substack{0<m<N\\m\neq n}}\frac{\cos(mk\pi/N)}{\cos(n\pi/N)-\cos(m\pi/N)}.$$

Here are steps to get a closed form of the latter. In this answer I show that $$\sum_{m=0}^{N-1}\frac{\exp(2mk\pi i/N)}{1-2z\cos(2m\pi/N)+z^2}=\frac{N(z^k+z^{N-k})}{(1-z^2)(1-z^N)}$$ for $0\leqslant k\leqslant N$ and $z\in\mathbb{C}$ not a singularity. Now we replace $N$ by $2N$, put $z=e^{it}$, and use $\sum_{m=0}^{2N-1}a_m=a_0+a_N+\sum_{m=1}^{N-1}(a_m+a_{2N-m})$. This gives (for $0\leqslant k\leqslant 2N$ now) $$\sum_{m=1}^{N-1}\frac{\cos(mk\pi/N)}{\cos t-\cos(m\pi/N)}=\frac12\left(\frac1{1-\cos t}-\frac{(-1)^k}{1+\cos t}\right)-N\frac{\cos(N-k)t}{\sin t\cdot\sin Nt}.$$ To get $S_{N,n,k}$, subtract the term with $m=n$ and take the limit as $t\to n\pi/N$.

metamorphy
  • 39,111
  • Hi, @metamorphy! You have solved one more puzzle for me! Thanks! :) – Sl0wp0k3 Jul 23 '21 at 00:12
  • Dear @metamorphy, from your answer above I have tried to obtain the closed form for the sum $$ \sum\limits_{m=1,m \ne n}^{N} \dfrac{ \sin^2\left(\frac{\pi m}{N+1}\right) }{\left( \cos\left(\frac{\pi n}{N+1}\right) - \cos\left(\frac{\pi m}{N+1}\right) \right)^2} $$, which can be related to the sum above by differentiating by $t$, but for some strange reason I failed, mb there is a problem with interchanging the limit and the differentiation. Do you have a simple idea how to do that? – Sl0wp0k3 Jul 29 '21 at 10:18
  • @Sl0wp0k3: Differentiation by $t$ should have worked (together with $\sin^2\alpha=(1-\cos2\alpha)/2$ in the numerator, reducing to $k\in{0,2}$). – metamorphy Jul 29 '21 at 11:39
  • Dear @metamorphy, sorry, I was mistaken, the actual series are $\sum\limits_{m=1,m\ne n}^{N} \dfrac{ \sin^2\left( \frac{\pi m}{N+1} \right) \left( 1 + (-1)^{n+m} \right) }{ \left( \cos \left( \frac{\pi n}{N+1} \right) - \cos \left( \frac{\pi m}{N+1} \right) \right)^2 }$. Could you give me an advice on how to get rid of that sign-altering part? The same trick from the post of yours does not seem to work now as there is no $k$. – Sl0wp0k3 Jul 30 '21 at 15:22
  • @Sl0wp0k3: $(-1)^m\cos m\alpha=\cos m(\pi+\alpha)$ at $\alpha=0$ and $\alpha=2\pi/(N+1)$, much like it's done at the beginning of the answer. – metamorphy Jul 31 '21 at 04:04
  • Dear @metamorphy, yes, that was a dumb question, I realized the answer right after posting it :). I was able to obtain the final answer for the sum, even though it is more involving than the original one with things like applying 4 times the l'Hotipal's rule to find the limit :). You helped me a lot with all that, and, at the very least, I wanted to mention you in the acknowledgement section of the manuscript I will be preparing somewhen during the autumn. So, how should I mention you there? – Sl0wp0k3 Aug 03 '21 at 12:16
  • 1
    thanks! Nice, I am also from RF :) – Sl0wp0k3 Aug 11 '21 at 16:29
  • By the way, is that possible to find $\sum\limits_{m=1}^{N−1}\dfrac{\exp(imk\pi/N)}{\cos(t)−cos(m\pi/N)}$ What I see is that bcoz of symmetry it is always cos() appearing in the numerator – Sl0wp0k3 Mar 23 '22 at 11:17
  • excuse me, did not get it. What did you suggest as a starting point? – Sl0wp0k3 Mar 23 '22 at 13:18
  • @Sl0wp0k3: (oops) the sum with $\cos(\Box)$ instead of $\exp(i\Box)$ can be evaluated, since there is a simple relation between $\sum_{m=1}^{N-1}$ and $\sum_{m=0}^{2N-1}$ (the latter is over the full system of $2N$-th roots of unity, so that "everything works"), but the sum with $\sin(\Box)$ is problematic. – metamorphy Mar 23 '22 at 13:57
  • Exactly, the roots are always symmetric with respect to real axis, so if summed over a full cycle - you get $\cos()$ anyways. I keep that in mind, and will write a solution if it comes to my mind. – Sl0wp0k3 Mar 24 '22 at 13:22
  • What I naively thought about as that the standard roots of unity are symmetric with respect to the real axis, mb I need to rotate it by $\cdot \exp(i \frac{\pi}{2})$, but seems like did not work either :). Really tricky with this symmetry bcoz of $N \to 2 N$ substitution. – Sl0wp0k3 Mar 25 '22 at 13:36