4
$\begingroup$

Suppose that $x_i \in [-1,1], i=0,n-1$ and consider the root of unity $\omega = \cos(2\pi/n)+i\sin(2\pi/n)$ for some $n \geq 2$. Consider complex numbers of the form $$ z = \sum_{i=0}^{n-1} x_i \omega^{i}$$

Are there any known lowest upper bounds for $|z|, |\operatorname{Re} z|, |\operatorname{Im} z|$?

I came across this type of questions when evaluating some numerical errors. Some trivial estimations show that for $n=3$ one should have $\operatorname{Re} z \leq 2$, for example. I was wondering if the bounds are known for general $n$.

$\endgroup$
10
  • $\begingroup$ What is wrong with $\lvert z \rvert \leq n$? $\endgroup$
    – Leo Moos
    Commented Jun 4, 2021 at 15:05
  • $\begingroup$ It’s clear that the worst case for $\def\Im{\operatorname{Im}}\Im z$ is when $x_i=\operatorname{sgn}(\Im \omega^i)$, thus $\Im z=\sum_{i<n}|\Im\omega^i|$. By a back of the envelope calculation, this is about $4/\sin(2\pi/n)\sim\frac2\pi n$. Just about the same holds for $\operatorname{Re}z$. Thus, you can’t improve much over the trivial bound given by Leo Moos. $\endgroup$ Commented Jun 4, 2021 at 15:11
  • 2
    $\begingroup$ Let me put it this way. The arguments above serve to show that in all cases, the optimal bound is attained when all the $x_i$ are $\pm1$, with $+1$ in one half-plane and $-1$ in the other half. For Re z, take the right half-plane; for Im z, take the upper half-plane; for |z|, it makes no difference what half-plane you take, as the relevant sums will have the same absolute value. Then it is a matter of summing a geometric series and computing its Re/Im/|z|. The latter is kind of tedious as there will be annoying minor differences depending on $n\mod4$, hence I’m not volunteering to do it. $\endgroup$ Commented Jun 4, 2021 at 16:40
  • 1
    $\begingroup$ Another way of putting Emil Jeřábek's argument is that by rotational symmetry, the problems with $\mathrm{Re}(z)$, $\mathrm{Im}(z)$ and $|z|$ are all (roughly) equivalent. $\endgroup$ Commented Jun 4, 2021 at 20:13
  • 3
    $\begingroup$ I did the calculation after all. It turns out to reduce to just 4 cases. First, by symmtery considerations, max Re z equals either max |z| if $n\not\equiv0\pmod4$, or max Im z if $n\equiv0\pmod4$. The rest only depend on the parity of $n$. If $n$ is even, then max |z| is $2\csc\frac\pi n$ and max Im z is $2\cot\frac\pi n$. If $n$ is odd, then max |z| is $\csc\frac\pi{2n}$ and max Im z is $\csc\frac\pi n+\cot\frac\pi n$. This is all valid for $n\ge2$. I don’t have the time to write it up now, but may do so some time next week if needed. $\endgroup$ Commented Jun 5, 2021 at 11:36

1 Answer 1

7
$\begingroup$

$\DeclareMathOperator\re{Re}\DeclareMathOperator\im{Im}\DeclareMathOperator\sgn{sgn}$Let me write $$\begin{align*} A(n)&=\max_{x\in[-1,1]^n}\biggl|\sum_{j<n}x_j\omega^j\biggr|,\\ R(n)&=\max_{x\in[-1,1]^n}\re\sum_{j<n}x_j\omega^j,\\ I(n)&=\max_{x\in[-1,1]^n}\im\sum_{j<n}x_j\omega^j. \end{align*}$$ First, it’s clear that $\re\sum_{j<n}$ is maximized if we maximize the $\re$ of each individual term of the sum, which is done by putting $x_j=\sgn\re\omega^j$. Thus, using also $\sum_{j<n}\omega^j=0$, $$\begin{align*} R(n)&=2\re\sum_{-\frac n4<j<\frac n4}\omega^j,\tag1\\ I(n)&=2\im\sum_{j<n/2}\omega^j.\tag2 \end{align*}$$ That is, the maximum is attained by setting $x_j\in\{-1,+1\}$, where we consider a line going through the origin, and set $x_j=+1$ in one half-plane determined by the line, and $x_j=-1$ in the other half-plane; for $R(n)$, we take the vertical line, and for $I(n)$, the horizontal line.

I claim that the same holds for $A(n)$, for an arbitrary choice of the dividing line. (Edge cases: if the line meets two $\omega^j$’s, one of these $x_j$ gets $+1$ and the other one $-1$; if the line meets one $\omega^j$, the choice of $x_j$ does not matter.) That is, for any $k<n$, $$\tag3A(n)=\biggl|\sum_{j<\lfloor n/2\rfloor}\omega^{k+j}-\sum_{j<\lceil n/2\rceil}\omega^{k+\lfloor n/2\rfloor+j}\biggr|=2\biggl|\sum_{j<\lfloor n/2\rfloor}\omega^{k+j}\biggr|=2\biggl|\sum_{j<\lceil n/2\rceil}\omega^{k+j}\biggr|.$$ That all these expressions are equal follows from $\bigl|\sum_j\omega^{k+j}\bigr|=\bigl|\omega^k\sum_j\omega^j\bigr|=\bigl|\sum_j\omega^j\bigr|$ and from $\sum_{j<n}\omega^j=0$. It’s also obvious that the expression in (3) is a lower bound on $A(n)$. To prove that it is an upper bound, consider $$z=\sum_{j<n}x_j\omega^j$$ for some $x\in[-1,1]^n$, and put $\alpha=\overline z/|z|$ so that $|\alpha|=1$ and $|z|=z\alpha$. Then $$|z|=\sum_{j<n}\alpha x_j\omega^j=\re\sum_{j<n}\alpha x_j\omega^j=\sum_{j<n}x_j\re(\alpha\omega^j),$$ thus putting $x'_j=\sgn\re(\alpha\omega^j)$, $$|z|\le\sum_{j<n}x'_j\re(\alpha\omega^j)=\re\sum_{j<n}x'_j\alpha\omega^j\le\biggl|\sum_{j<n}x'_j\alpha\omega^j\biggr|=\biggl|\sum_{j<n}x'_j\omega^j\biggr|,$$ which is a sum as in (3) as long as $x'_j\in\{-1,+1\}$. (I leave it as an exercise to fix the issues with $x'_j=0$.)

Now we need to evaluate the given expressions. To simplify matters, first observe $$R(n)=\begin{cases}A(n),&n\not\equiv0\pmod4,\\I(n),&n\equiv0\pmod4.\end{cases}$$ Indeed, the sum in (1) is real positive, hence its real part is also its absolute value, and it can serve as (3) as long as the vertical axis does not hit any $\omega^j$; in the latter case (i.e., $n\equiv0\pmod4$), the sums in (1) and (2) differ by rotation by $\pi/2$ (ignoring the points that hit the line, which do not contribute to the final expression).

For $n$ even, the geometric series formula gives $$A(n)=2\biggl|\sum_{j<n/2}\omega^j\biggr|=2\frac2{|1-\omega|}=\frac2{\sin\frac\pi n}=2\csc\frac\pi n$$ and $$I(n)=2\im\frac2{1-\omega}=2\im\frac{2(1-\overline\omega)}{|1-\omega|^2}=\frac{\sin\frac{2\pi}n}{\sin^2\frac\pi n}=\frac{2\sin\frac\pi n\cos\frac\pi n}{\sin^2\frac\pi n}=2\cot\frac\pi n.$$ For $n$ odd, symmetry along the vertical axis yields $$I(n)=\frac12I(2n)=\cot\frac\pi{2n},$$ and I omit the details of $$A(n)=2\biggl|\sum_{j<\lfloor n/2\rfloor}\omega^j\biggr|=2\frac{|1+\omega^{1/2}|}{|1-\omega|}=\frac{2\cos\frac\pi{2n}}{\sin\frac\pi n}=\csc\frac\pi{2n}.$$

All in all, we have (for all $n\ge1$):

$n\bmod 4$ $A(n)$ $R(n)$ $I(n)$
$0$ $2\csc\frac\pi n$ $2\cot\frac\pi n$ $2\cot\frac\pi n$
$2$ $2\csc\frac\pi n$ $2\csc\frac\pi n$ $2\cot\frac\pi n$
$1,3$ $\csc\frac\pi{2n}$ $\csc\frac\pi{2n}$ $\cot\frac\pi{2n}$

Note that asymptotically, all the expressions in this table are $\frac2\pi n+O(n^{-1})$ as $n\to\infty$.

$\endgroup$
4
  • $\begingroup$ Thank you very much for performing the computations! Nice answer. $\endgroup$ Commented Jun 8, 2021 at 15:15
  • $\begingroup$ You’re welcome. $\endgroup$ Commented Jun 8, 2021 at 17:50
  • $\begingroup$ $\sqrt{|1-\omega|}$ is not equal to $\sin(\frac{\pi}{n})$, but $\sqrt{1-\cos(\frac{\pi}{n})}$. $\endgroup$ Commented Oct 10 at 6:09
  • 1
    $\begingroup$ @JingweiChen I don’t know what exactly you are referring to, as “$\sqrt{|1-\omega|}$” does not appear anywhere in my post. Anyway, you are wrong. As I am using repeatedly, $|1-\omega|=2\sin\frac\pi n$; you can readily show this by doing the algebra, or alternatively, just look at a picture ($A=0$, $D=1$, $B=\omega$, $\theta=\pi/n$, $|1-\omega|=|BD|$). It is certainly larger than $1-\cos\frac\pi n$ (which is the distance from $C$ to the intersection point of the circle with line $AC$). $\endgroup$ Commented Oct 10 at 9:16

You must log in to answer this question.

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