11

How to derive an upper bound of

$$\sum_{n=1}^N |1-z^n|$$ where $z\in\mathbb{C}$ and $|z| \leq 1$?

A trivial upper bound would be $2N$ since each $|1-z^n| \leq 2$. But I am hoping for tighter bounds. I ran some numerical experiments and believe the bound should be $\frac{3}{2}N$ but don't know how to prove it.

  • 1
    For some reason, this problem gives me the same sort of vibe as https://math.stackexchange.com/questions/3906952/if-z-in-mathbb-c-with-z-leqslant-frac45-then-sum-n-in-szn-neq-f . Not sure if theres a connection or not and just wanted to put that out there – QC_QAOA May 10 '22 at 06:47
  • 1
    Could maybe conjure something out of $|1-z^n| \leqslant |1-z| \sum _{k=0}^{n-1}|z|^k = |1-z| \frac{(1-|z|^{n-1})}{(1-|z|)}$ – AlvinL May 10 '22 at 07:23
  • 2
    Maybe the following could help:The maximum of this sum is attained on the circle $|z|=1$, see https://math.stackexchange.com/questions/96257/maximum-of-sum-of-finite-modulus-of-analytic-function – Gerd May 10 '22 at 07:36
  • 2
    @AlvinL Does not help. The intermediate sum is maximal fo $z=-1$ and gives the obvious bound $2n.$ – Ryszard Szwarc May 10 '22 at 08:10
  • for large enough $N$ (probably $N \ge 6$ would do) one can prove a $N\sqrt {5/2}$ bound which is a bit bigger than $3N/2$, but for say $N=2$ one can easily see that we can get $2\sqrt 3$ for $z^3=1, z\ne 1$ because $1,z,z^2$ form an equilateral triangle with side $\sqrt 3$, so for small enough $N$ a $3N/2$ bound is not correct – Conrad May 10 '22 at 16:22

4 Answers4

6

Here is an upper bound:

According to the link given by @Gerd in comment (Maximum of sum of finite modulus of analytic function.), the maximum of $\sum_{n=1}^N |1 - z^n|$ is attained on $|z| = 1$.

Let $z = \mathrm{e}^{\mathrm{i}\theta}$ with $\theta \in [0, 2\pi]$. Using AM-QM, we have \begin{align*} \sum_{n=1}^N |1 - z^n| &\le \sqrt{N\sum_{n=1}^N |1 - z^n|^2}\\ &= \sqrt{N\sum_{n=1}^N (2 - 2\cos n\theta)}\\ &= \sqrt{2N^2 + N \frac{\sin \frac{\theta}{2} - \sin\frac{(2N + 1)\theta}{2}}{\sin \frac{\theta}{2}}}. \tag{1} \end{align*}

Fact 1: Let $x\in [0, \pi/2]$ and $N \ge 20$. Then $$\frac{\sin x - \sin (2N + 1)x}{\sin x} \le \frac{N}{2}.$$ (The proof is not difficult.)

By Fact 1 and (1), we have, for all $N \ge 20$, $$\sum_{n=1}^N |1 - z^n| \le \sqrt{5/2}\, N.$$


Some thoughts:

We have \begin{align} \sum_{n=1}^N |1 - z^n| &= \sum_{n=1}^N \sqrt{2 - 2\cos n\theta }\\ &= \sum_{n=1}^N 2\left|\sin \frac{n\theta}{2}\right|\\ &= \frac{4}{\pi}N - \sum_{n=1}^N \sum_{k=1}^\infty \frac{8}{\pi(4k^2 - 1)}\cos kn\theta\\ &= \frac{4}{\pi}N - \sum_{k=1}^\infty \sum_{n=1}^N \frac{8}{\pi(4k^2 - 1)}\cos kn\theta\\ &= \frac{4}{\pi}N + \sum_{k=1}^\infty \frac{8}{\pi(4k^2 - 1)} \frac{\sin\frac{k\theta}{2} - \sin \frac{(2N + 1)k \theta}{2}}{2\sin \frac{k\theta}{2}}\\ &= \frac{4}{\pi}N + \sum_{k=1}^\infty \frac{4}{\pi(4k^2 - 1)} + \sum_{k=1}^\infty \frac{4}{\pi(4k^2 - 1)} \frac{ - \sin \frac{(2N + 1)k \theta}{2}}{\sin \frac{k\theta}{2}}\\ &= \frac{4}{\pi}N + \frac{2}{\pi} + \sum_{k=1}^\infty \frac{4}{\pi(4k^2 - 1)} \frac{ - \sin \frac{(2N + 1)k \theta}{2}}{\sin \frac{k\theta}{2}} \end{align} where we have used the identity $$\left|\sin y\right| = \frac{2}{\pi}-\sum_{k = 1}^\infty \frac{4}{\pi(4k^2-1)}\cos(2k y).$$ (Note: The RHS is the Fourier expansion of the LHS.)

We need to find bounds for $$\sum_{k=1}^\infty \frac{4}{\pi(4k^2 - 1)} \frac{ - \sin \frac{(2N + 1)k \theta}{2}}{\sin \frac{k\theta}{2}}.$$

River Li
  • 37,323
  • I think it may work for $N \ge 6$ per wolfram alpha (see my comment under the OP); not sure one can do better in a simple way – Conrad May 11 '22 at 03:49
  • @Conrad Yes. For small $N$, we can verify it directly. Also, if $N> 100$, we may improve the bound to $\sqrt{2.45}, N$. – River Li May 11 '22 at 03:54
  • Yes - one can drive down a little the constant but still not sure one can get $1.5$ easily even for large $N$ – Conrad May 11 '22 at 03:58
  • @Conrad My approach can not get 1.5N. Numerically, 1.5N is a bound for $N > 15$. – River Li May 11 '22 at 04:00
  • Cauchy-Schwarz to transform the sum of $|\sin k\theta/2|$ into the corresponding sum of squares as above loses something; one can try using Fourier series instead and summing inside and using the bound above for the sum of cosines that way, but still didn't get anything better though there is extra cancelation so who knows... gets complicated though – Conrad May 11 '22 at 04:04
  • @Conrad We may deal with $\sum |\sin k\theta/2|$ directly. Let me try. – River Li May 11 '22 at 04:18
3

The behavior in the limit (as $N\to\infty$) looks interesting: the value $$L:=\lim_{N\to\infty}\frac1N\sup_{|z|\leqslant 1}\sum_{n=1}^N|1-z^n|\approx1.4492227\ldots\color{red}{<1.5}$$ equals $2\sin\lambda$, where $\lambda$ is the smallest positive root of $\cos\lambda+\lambda\sin\lambda=1$.

Here is a sketch towards a proof. As already noted, the $\sup_{|z|\leqslant 1}$ is attained at $|z|=1$: $$\sup_{|z|\leqslant 1}\sum_{n=1}^N|1-z^n|\underset{\big[z=e^{2it}\big]}{=}2\sup_{0\leqslant t\leqslant\pi/2}\sum_{n=1}^N|\sin nt|.$$ Suppose the maximum is attained at $t=t_N$, and let $\lambda_N=Nt_N$. Then it turns out that $\lambda_N<\pi$ for all $N$, and that the limit $\lambda=\lim_{N\to\infty}\lambda_N$ exists. Assuming this is proven, $t=t_N$ is the smallest positive root of $\sum_{n=1}^N n\cos nt=0$, that is of $(N+1)\cos Nt-N\cos(N+1)t=1$, so that $$\left(N+1-N\cos\frac{\lambda_N}N\right)\cos\lambda_N+N\sin\frac{\lambda_N}N\sin\lambda_N=1,$$ which gives $\cos\lambda+\lambda\sin\lambda=1$ after $N\to\infty$. And then $$L=\lim_{N\to\infty}\frac{\cos\frac{t_N}2-\cos\left(N+\frac12\right)t_N}{N\sin\frac{t_N}2}=\frac{1-\cos\lambda}{\lambda/2}=2\sin\lambda.$$

metamorphy
  • 39,111
  • I am not doubting that your result is correct, but how do you argue that the maximum is attained at a root of $\sum_{n=1}^N n\cos (nt)=0$? Do you assume that all $\sin(nt)$ terms are positive at that maximum? – Martin R May 11 '22 at 12:37
  • @MartinR: Yes, I'm assuming that (as said in $\lambda_N<\pi$). This is seen in the plots of $t\mapsto\sum_{n=1}^N|\sin nt|$, but is of course yet to be shown (not too hard I believe). – metamorphy May 11 '22 at 13:33
  • Yes, you are right. I had overlooked the scaling by the factor $N$. – Martin R May 11 '22 at 14:17
  • @metamorphy I also observed that $\lambda_N < \pi$. I have not yet proved it. (Your answer is nice. Already +1 some hours ago). – River Li May 11 '22 at 15:40
2

Some thoughts:

As already noted, the maximum of $\sum_{n=1}^N |1 - z^n|$ is attained on $|z| = 1$.

Let $z = \mathrm{e}^{\mathrm{i}\theta}$ with $\theta \in [0, 2\pi]$. We have $$\sum_{n=1}^N |1 - z^n| = \sum_{n=1}^N 2\left|\sin \frac{n\theta}{2}\right|.$$

Conjecture 1: The maximum of $\sum_{n=1}^N 2\left|\sin \frac{n\theta}{2}\right|$ on $[0, 2\pi]$ is attained at some $x_0 \in [0, 2\pi/N]$. (Note: Numerical experiment supports the claim. )

If Conjecture 1 is true, when $\theta \in [0, 2\pi/N]$, we have $$ \sum_{n=1}^N 2\left|\sin \frac{n\theta}{2}\right| = \sum_{n=1}^N 2\sin \frac{n\theta}{2} = \frac{2\sin^2 \frac{N\theta }{4} \cos \frac{\theta}{4}}{\sin \frac{\theta}{4}} + \sin \frac{N\theta}{2} = \frac{2\sin^2 y \cos \frac{y}{N}}{\sin \frac{y}{N}} + \sin 2y $$ where $y = N\theta/4 \in [0, \pi/2]$. Then we have $$\frac{2\sin^2 y \cos \frac{y}{N}}{\sin \frac{y}{N}} + \sin 2y \le \frac{2\sin^2 y}{\frac{2y}{\pi}\sin \frac{\pi}{2N}} + 1 = \frac{\pi}{\sin \frac{\pi}{2N}}\frac{\sin^2 y}{y} + 1 \le \frac{\pi}{\sin \frac{\pi}{2N}}\, \frac{29}{40} + 1$$ where we have used $\sin \frac{y}{N} \ge \frac{2y}{\pi}\sin \frac{\pi}{2N} $ and $\frac{\sin^2 y}{y} \le \frac{29}{40}$ for all $y\in [0, \pi/2]$. It is not difficult to prove that, for all $N > 21$, $$\frac{\pi}{\sin \frac{\pi}{2N}}\, \frac{29}{40} + 1 < \frac{3}{2}N.$$

River Li
  • 37,323
1

This is not an an exact answer: I just want to present a physical-geometrical approach to the problem which can help to "visualize" and guide to a rigorous mathematical solution.

Please allow me to slightly change the notation and use a more practical small $n$ instead of $N$.
So let's consider $$ \left\{ \matrix{ S_n = \sum\limits_{1 \le k \le n} {\left| {1 - z^k } \right|} = \sum\limits_{1 \le k \le n} {\left| {z^k - 1} \right|} = \sum\limits_{0 \le k \le n} {\left| {z^k - 1} \right|} \hfill \cr z = re^{i\alpha } \quad \left| {\;r \le 1} \right. \hfill \cr} \right. $$

Now, we can look at $\left| {w - 1} \right|$ as being the distance from $U=(1,0) = 1+i0$, and thus the first moment of a unit mass placed at $w$ with respect to $U$.
Then $$ \sum\limits_{0 \le k \le n} {\left| {e^{ik\alpha } - 1} \right|} $$ is the sum of the moments of $n+1$ unit masses, placed on the unit circle at constant angular separation $\alpha$ with the first being at $U$.
If $r < 1$ the $n+1$ unit masses will be placed instead on the log-spiral an starting from $U$. $$ \rho = \left| {r^k e^{ik\alpha } } \right| = \left| {r^{\left( {k\alpha } \right)/\alpha } e^{i\left( {k\alpha } \right)} } \right| = r^{\theta /\alpha } = e^{{{\ln r} \over \alpha }\theta } = e^{\ln r\;k} $$ again with constant angular separation, $\alpha$ or better $k$ if we eliminate $\alpha$ from the constant at the exponent, and again starting from $U$ and proceeding towards the origin.

Therefore $S_n$ is the moment of a mass $n+1$ placed at the barycenter of the configuration of the single masses.

It becomes then geometrically clear that

  • For large $n$ and $r=1$ the largest distance of the baricenter from $U$ is attained when most of the points are lying in the 2nd and 3rd quadrant; this means to correspondingly decrease $\alpha$ so that the almost continuous circular arc is about $270^{circ}$.
  • In the limit, for $n \to \infty$, $S_n /(n+1)$ will tend to the barycenter of an arc of unitary mass of which to maximize the barycenter distance.
  • For a spiral configuration, also the barycenter distance will be maximized when most of the points are in the 2nd and 3rd quadrant
  • It is also geometrically "visible" that given a maximized spiral configuration, we can radially expand it to the circle, thereby increasing the distance from $U$ except for the points close to that.
G Cab
  • 35,272