# Limit of sum of $\sqrt{\frac{\delta^2}4+N\left(N+\delta\right)\sin^2\left(\frac{\pi n}N\right)}$

How can we prove that this limit exists? It would be even better if it could be computed ( even in terms of an infinite series would be good). ($$N$$ is obviously an integer)

$$f\left(\delta\right)=\lim\limits_{N\rightarrow\infty}\left(N-2\sum\limits_{n=1}^{\left\lfloor\frac N2\right\rfloor}\left(\sqrt{\frac{\delta^2}4+N\left(N+\delta\right)\sin^2\left(\frac{\pi n}N\right)}-\sqrt{\frac{\delta^2}4+N\left(N+\delta\right)\sin^2\left(\frac{\pi n}N-\frac\pi{2N}\right)}\right)\right)$$

I found the limit at $$\delta=0$$ to be $$\frac\pi4$$ as then the square roots disappear and the sum of $$\sin$$ terms can be evaluated directly. or the other cases I tried proving that the sequence was Cauchy, but I couldn't do that.

The expression under the limit sign equals (exactly) $$f_N(\delta):=\sum_{n=1}^{2N}(-1)^{n-1}\sqrt{\frac{\delta^2}{4}+N(N+\delta)\sin^2\frac{n\pi}{2N}}.$$ As $$\sum_{n=1}^{2N}(-1)^{n-1}\sin\frac{n\pi}{2N}=\tan\frac{\pi}{4N}$$ (yes, the case $$\delta=0$$ is easy), we can write $$f_N(\delta)=\Big(N+\frac{\delta}{2}\Big)\tan\frac{\pi}{4N}-\frac{|\delta|}{2}+2\sum_{n=1}^{N-1}(-1)^{n-1}\times{}\\{}\times\left(\sqrt{\frac{\delta^2}{4}+N(N+\delta)\sin^2\frac{n\pi}{2N}}-\Big(N+\frac{\delta}{2}\Big)\sin\frac{n\pi}{2N}\right).$$ Here, the limit can be taken termwise — after pairing "odds and evens", this is justified by (a discrete version of) the dominated convergence theorem. This immediately gives $$\color{blue}{f(\delta)=\frac{\pi}{4}-\frac{|\delta|}{2}+\sum_{n=1}^{\infty}(-1)^{n-1}\big(\sqrt{\delta^2+n^2\pi^2}-n\pi\big)}$$ which also has an integral representation: $$\color{blue}{f(\delta)=\displaystyle\frac{1}{\pi}\int_{|\delta|}^{\infty}\frac{\sqrt{x^2-\delta^2}}{\sinh x}\,dx}$$ (see the linked question). Below is my older answer obtaining this integral another (somewhat more complicated) way.
$$f_N(\delta)=\frac{N+\delta}{2}\left(g_{2N}\Big(\frac{N}{N+\delta}\Big)-2g_N\Big(\frac{N}{N+\delta}\Big)\right),\qquad(\delta>0)\\ f_N(\delta)=\frac{N}{2}\left(g_{2N}\Big(\frac{N+\delta}{N}\Big)-2g_{N}\Big(\frac{N+\delta}{N}\Big)\right),\qquad(\delta<0)\\g_N(k):=\sum_{n=0}^{N-1}\sqrt{1-2k\cos\frac{2n\pi}{N}+k^2}\qquad(|k|<1)$$ Now the idea is to use Fourier expansion of $$\sqrt{1-2k\cos x+k^2}$$, together with $$\sum_{n=0}^{N-1}\cos\frac{2mn\pi}{N}=\begin{cases}N,& N\mid n\\0,& N\not\mid n\end{cases}.$$ The expansion can be obtained indirectly from the binomial series $$(1-z)^{1/2}=\sum_{m=0}^{\infty}b_m z^m,\qquad b_m=(-1)^m\binom{1/2}{m}=-\frac{1}{\pi}\mathrm{B}\Big(m-\frac{1}{2},\frac{3}{2}\Big)$$ (where $$\mathrm{B}$$ denotes the beta function; we'll need this later): \begin{align} \sqrt{1-2k\cos x+k^2}&=(1-ke^{ix})^{1/2}(1-ke^{-ix})^{1/2} \\&=\left(\sum_{m=0}^{\infty}b_m k^m e^{imx}\right)\left(\sum_{m=0}^{\infty}b_m k^m e^{-imx}\right)\\&=a_0(k)+2\sum_{m=1}^{\infty}a_m(k)\cos mx, \\a_m(k)&:=\sum_{n=0}^{\infty}b_n b_{m+n}k^{m+2n}. \end{align} So, for $$m>0$$, using the integral representation of $$\mathrm{B}$$, we have \begin{align} a_m(k)&=-\frac{1}{\pi}\sum_{n=0}^{\infty}b_n k^{m+2n}\int_0^1 t^{m+n-3/2}(1-t)^{1/2}\,dt\\&=-\frac{k^m}{\pi}\int_0^1 t^{m-3/2}\sqrt{(1-t)(1-k^2 t)}\,dt, \end{align} which gives \begin{align} g_N(k)&=\sum_{n=0}^{N-1}\left(a_0(k)+2\sum_{m=1}^{\infty}a_m(k)\cos\frac{2mn\pi}{N}\right)=Na_0(k)+2N\sum_{m=1}^{\infty}a_{mN}(k)\\&=Na_0(k)-\frac{2N}{\pi}k^N\int_0^1\frac{t^{N-3/2}}{1-(kt)^N}\sqrt{(1-t)(1-k^2 t)}\,dt. \end{align} Note that $$a_0(k)$$ can also be expressed in terms of elliptic integrals: $$a_0(k)=\frac{2}{\pi}\big(2\mathrm{E}(k)-(1-k^2)\mathrm{K}(k)\big),$$ but we won't need it, since the preceding equality implies $$g_{2N}(k)-2g_N(k)=\frac{4N}{\pi}k^N\int_0^1\frac{t^{N-3/2}}{1-(kt)^{2N}}\sqrt{(1-t)(1-k^2 t)}\,dt.$$ This can be plugged back into $$f_N(\delta)$$. For $$\delta>0$$, substituting $$t=1-x/N$$, we get $$f_N(\delta)=\frac{2}{\pi}\Big(\frac{N}{N+\delta}\Big)^N\int_0^N\frac{(1-x/N)^{N-3/2}}{1-\left(\frac{N-x}{N+\delta}\right)^{2N}}\sqrt{x\left(x+2\delta+\frac{\delta^2}{N}\right)}\,dx,$$ and taking $$N\to\infty$$ is easy now. Doing the same for $$\delta<0$$, we get the integral above.
• @metamorphy, this answer also works for $\delta=0$, as well as $\delta>0$ and $\delta<0$. This REALLY REALLY helped my work. Thanks a lot. If you are interested, $f\left(\delta\right)+\mathrm{min}\left(0,\delta\right)$ is the dominant period of the structure form factor of a periodic Ising model (the Fourier transform of the two-point correlator of the spectral density). Anyway, thanks a lot again for this. – Chetan Vuppulury Jul 9 '19 at 22:53