7
$\begingroup$

We consider the function

$$f(x) = 1- \frac{1}{N} \sum_{i=1}^N \frac{\sin\left(\tfrac{\pi i}{N}\right)^2}{1+\sin\left(\tfrac{\pi i}{2N}\right)^2-x}.$$

The arguments of the two sines differ by a factor of $2$, and the function as a whole has a bit of the flavour of a Riemann sum.

By monotonicity, there is precisely one root of $f$ in each interval $\Big(1+\sin\big(\tfrac{\pi (i-1)}{2N}\big)^2,1+\sin\big(\tfrac{\pi i}{2N}\big)^2\Big)$ and precisely one root to the left of $ 1+\sin\big(\tfrac{\pi}{2N}\big)^2$. Now I wonder about more refined estimates for the roots of $f$.

Which roots are on the left and which ones are on the right half of the intervals $\Big(1+\sin\big(\tfrac{\pi (i-1)}{2N}\big)^2,1+\sin\big(\tfrac{\pi i}{2N}\big)^2\Big)?$

Regarding answers as of Dec 15, 2022: I am looking for rigorous estimates with error control.

$\endgroup$

3 Answers 3

5
+250
$\begingroup$

If my computations are correct, there is a root of the form $x=1+\sin^2(\frac\theta2)$ with:

$$\theta \in \left(\frac{k\pi}N,\frac{k\pi}N+\frac\pi{2N}\right) \text{ if }k < \frac{N}3-\frac12$$ $$\theta \in \left(\frac{k\pi}N-\frac{\pi}{2N},\frac{k\pi}N\right)\text{ if }k> \frac{N}3+\frac12.$$

I localize the zeros using another expression of $f(1+x)$.

For every $n \ge 1$, call $T_N$ the $n^{\rm th}$ Tchebycheff polynomial: for all $\theta \in \mathbf{R}$, $$T_n(\cos\theta) = \cos(n\theta).$$

Set $c_k = \cos(k\pi/(2N))$ and $s_k = \sin(k\pi/(2N))$ for $0 \le k \le 2N$. Then $1=c_0>\ldots>c_{2N}=-1$. The $c_k$ for $1 \le k \le 2N-1$ are double roots of $T_{4N}-1$ whereas $-1$ and $1$ are simple roots. The leading coefficient of $T_{4N}-1$ is $2^{4N-1}$. Thus $$T_{4N}(X)-1 = 2^{4N-1}(X-1)(X+1) \prod_{k=1}^{2N-1}(X-c_k)^2.$$ Since $T_{4N}(X) = T_{2n}(T_2(X))$ and $c_{2N-k} = -c_k$ for $0 \le k \le 2N$, we derive $$T_{2N}(2X^2-1)-1 = 2^{4N-1}(X^2-1)X^2 \prod_{k=1}^{N-1}(X^2-c_k^2)^2.$$ Therefore, $$T_{2N}(2X-1)-1 = 2^{4N-1}(X-1)X \prod_{k=1}^{N-1}(X-c_k^2)^2.$$ $$T_{2N}(2X-1)-1 = 2^{4N-1}\frac{X-1}{X} \prod_{k=1}^{N}(X-c_k^2)^2.$$ Taking logarithmic derivatives, we get $$\frac{2T'_{2N}(2X-1)}{T_{2N}(2X-1)-1} = \frac{1}{X-1} - \frac{1}{X} + 2\sum_{k=1}^{N} \frac{1}{X-c_k^2}.$$ Replacing $X$ with $1-X$, we get $$\frac{2T'_{2N}(1-2X)}{T_{2N}(1-2X)-1} = -\frac{1}{X} - \frac{1}{1-X} + 2\sum_{k=1}^{N} \frac{1}{s_k^2-X}.$$ $$\sum_{k=1}^{N} \frac{1}{s_k^2-X} = \frac{T'_{2N}(1-2X)}{T_{2N}(1-2X)-1} + \frac{1}{2X(1-X)}.$$ Observe now that for every $k$, $\sin^2(k\pi/N) = (2s_kc_k)^2 = 4s_k^2(1-s_k^2)$, so $$\frac{1}{4}\frac{\sin^2(k\pi/N)}{s_k^2-X} = \frac{s_k^2}{s_k^2-X} - \frac{s_k^4}{s_k^2-X}.$$ $$\frac{1}{4}\frac{\sin^2(k\pi/N)}{s_k^2-X} = \Big(1 + \frac{X}{s_k^2-X}\Big) - \Big(s_k^2 + \frac{s_k^2X}{s_k^2-X}\Big).$$ $$\frac{1}{4}\frac{\sin^2(k\pi/N)}{s_k^2-X} = \Big(1 + \frac{X}{s_k^2-X}\Big) - \Big(s_k^2 + X\Big(1 + \frac{X}{s_k^2-X}\Big)\Big).$$ $$\frac{1}{4}\frac{\sin^2(k\pi/N)}{s_k^2-X} = c_k^2 - X + \frac{X-X^2}{s_k^2-X}.$$ By summation over all $k \in \{1,\ldots,N\}$, since $c_N=0$ and by the equalities $c_{N-k}=s_k$ and $c_k^2+s_k^2=1$, $$\sum_{k=1}^{N-1} c_k^2 = \sum_{k=1}^{N-1} s_k^2 = \frac{N-1}{2},$$ we get $$\frac{1}{4} \sum_{k=1}^{N}\frac{\sin^2(k\pi/N)}{s_k^2-X} = \frac{N-1}{2} - NX + X(1-X) \sum_{k=1}^{N} \frac{1}{s_k^2-X}.$$ Putting thing together, we get $$\frac{1}{4} \sum_{k=1}^{N}\frac{\sin^2(k\pi/N)}{s_k^2-X} = \frac{N-1}{2} - NX + X(1-X) \frac{T'_{2N}(1-2X)}{T_{2N}(1-2X)-1} + \frac{1}{2}.$$ $$1 - f(X+1) = \frac{1}{N} \sum_{k=1}^{N}\frac{\sin^2(k\pi/N)}{s_k^2-X} = 2 - 4X + \frac{4}{N}X(1-X) \frac{T'_{2N}(1-2X)}{T_{2N}(1-2X)-1}.$$ $$f(X+1) = -1 +4X - \frac{4}{N}X(1-X) \frac{T'_{2N}(1-2X)}{T_{2N}(1-2X)-1}.$$ Let us evaluate at $X = (1-\cos\theta)/2 = \sin^2(\theta/2)$. $$f(1+\sin^2(\theta/2)) = -1 + 2(1-\cos\theta) - \frac{1}{N}\sin^2\theta \frac{T'_{2N}(\cos\theta)}{T_{2N}(\cos\theta)-1}.$$ By derivating $T_{2N}(\cos\theta) = \cos(2N\theta)$, we get $\sin\theta \times T'_{2N}(\cos\theta) = 2N\sin(2N\theta)$. Hence $$f(1+\sin^2(\theta/2)) = 1 - 2\cos\theta - 2\sin\theta \frac{\sin(2N\theta)}{\cos(2N\theta)-1}.$$ $$f(1+\sin^2(\theta/2)) = 1 - 2\cos\theta + 2\sin\theta \frac{\cos(N\theta)}{\sin(N\theta)}.$$ Now assume that $\theta = (k\pi+\epsilon)/N$ with $k \in \{1,\ldots,N-1\}$ and $|\epsilon|<\pi/2$. Then $$f(1+\sin^2(\theta/2)) = 1 - 2\cos\Big(\frac{k\pi+\epsilon}{N}\Big) + 2\sin\Big(\frac{k\pi+\epsilon}{N}\Big) \frac{\cos\epsilon}{\sin\epsilon}.$$ Let us apply the intermediate value theorem. $$\lim_{\epsilon \to 0+}f(1+\sin^2(\theta/2)) = +\infty.$$ $$\lim_{\epsilon \to 0-}f(1+\sin^2(\theta/2)) = -\infty.$$ $$\lim_{\epsilon \to \pi/2}f(1+\sin^2(\theta/2)) = 1 - 2\cos\Big(\frac{k\pi+\pi/2}{N}\Big).$$ $$\lim_{\epsilon \to -\pi/2}f(1+\sin^2(\theta/2)) = 1 - 2\cos\Big(\frac{k\pi-\pi/2}{N}\Big).$$

Since $k+\frac12 < \frac{N}3$ implies $2\cos((k\pi+\pi/2)/N)>1$, and $k-\frac12 > \frac{N}3$ implies $2\cos((k\pi-\pi/2)/N)<1$, this gives the answer stated above.

One could refine the method and look at the sign of $f(1+\sin^2(\theta/2))$ when $\epsilon = \pm \pi/4$ for example.

$\endgroup$
4
  • 3
    $\begingroup$ You copied the expression for $f(x)$ incorrectly. $\endgroup$ Commented Dec 15, 2022 at 16:56
  • $\begingroup$ I modified my post to answer the right question. $\endgroup$ Commented Dec 16, 2022 at 13:32
  • $\begingroup$ Can you provide details on your last two sentences? The last statement does not seem to hold for $(N,k)=(10,4)$ and $(N,k)=(20,7)$. $\endgroup$ Commented Dec 18, 2022 at 16:21
  • $\begingroup$ @Iosif Pinellis I had been to quick. This explains why the last statement did not hold for $(N,k) = (10,3)$ and $(N,k) = (20,7)$. $\endgroup$ Commented Dec 18, 2022 at 17:06
3
$\begingroup$

Let $$a_i=\sin ^2\left(\frac{\pi i}{N}\right)\qquad \text{and} \qquad b_i=1+\sin ^2\left(\frac{\pi i}{2 N}\right)$$ and consider $$f(x)=1-\frac{1}{N} \sum_{i=1}^N \frac{a_i}{b_i-x}$$ For the root between $b_k$ and $b_{k+1}$, consider instead the function $$\color{red}{g(x)=(b_k-x)(b_{k+1}-x)f(x)}$$ which, detailed to avoid the problem of limits, write $$\color{blue}{g(x)=(b_k-x)(b_{k+1}-x)-}$$ $$\color{blue}{\frac{1}{N} \Big[\sum_{i=1}^{k-1} \frac{a_i(b_k-x)(b_{k+1}-x)}{b_i-x}+a_k(b_{k+1}-x) +a_{k+1}(b_{k}-x)+\sum_{i=k+2}^{N} \frac{a_i(b_k-x)(b_{k+1}-x)}{b_i-x}\Big]}$$ No more vertical asymptotes and a smooth function.

Computing at the bounds

$$g(b_k)=-\frac{a_k(b_{k+1}-b_k)} N \qquad \text{and} \qquad g(b_{k+1})=-\frac{a_{k+1}(b_k-b_{k+1})}N$$

Just using the secant gives the estimate $$\color{red}{x_0=\frac{a_{k+1}\,b_k+a_k\,b_{k+1}}{a_{k}+a_{k+1} }}$$

Trying for $N=10$ and $k=4$, the above will give as a first estimate $$x_0=\frac{132-7 \sqrt{5}}{82}=1.41887$$ while the solution is $x=1.43987$ and the bounds $b_4=\frac{13-\sqrt{5}}{8} =1.34549$, $b_5=\frac 32$.

The first iterates of Newton method are $x_1=1.44146$, $x_2=1.43988$.

Now, may I confess that I spent (wasted ?) more than thirty years with this kind of equations. In chemical engineering, this is the so-called Leibovici & Neoschil method.

$\endgroup$
0
$\begingroup$

Put $a_k=1+\sin(k\pi/(2N))^2$, so $f(x)=1-N^{-1}\sum_{k=1}^N(a_k-1)/(a_k-x)$. Let $b_k$ be the unique root of $f(x)$ lying in $[a_{k-1},a_k]$, and put $t_k=(b_k-a_{k-1})/(a_k-a_{k-1})$, so $t_k$ measures the position of $b_k$ in the interval $[a_{k-1},a_k]$. Experimental calculation makes it clear that $t_k$ is very well approximated by a linear function $t_k\approx m_Nk+c_N$ with $1/(2N-3)<m_N<1/(2N-2)$ and so $2Nm_N\to 1$ as $N\to\infty$. In particular, it seems that $t_k<1/2$ unless $k$ is very close to $N$. When $N=100$ we have $t_k<1/2$ iff $k\leq 97$.

$\endgroup$
3
  • $\begingroup$ Thank you. What do you mean by "experimental calculations"? $\endgroup$ Commented Dec 12, 2022 at 16:20
  • $\begingroup$ I just mean that I calculated everything in Maple for $10\leq N\leq 100$ $\endgroup$ Commented Dec 12, 2022 at 16:40
  • 1
    $\begingroup$ You copied the expression for $f(x)$ incorrectly. $\endgroup$ Commented Dec 15, 2022 at 16:56

You must log in to answer this question.

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