3
$\begingroup$

Let $\omega_0, \ldots, \omega_{n-1}$ be the $n$-th roots of unity, and $a_0, \ldots, a_{n-1}$ be real numbers in the $(0, 1]$ interval. Define $z_i = a_i \omega_i$ as the vertices of a polygon on the complex plane, and $s_i$ as the area of a triangle whose vertices are $(0, z_i, z_{i+1})$, indices taken modulo $n$.

For the case illustrated below, with $n=4$, is it possible to determine the values of $(a_0, \ldots, a_3)$ given the values of $(s_0, \ldots, s_3)$?

What I suspect: for odd $n$, this problem can be reduced to a circulant linear system with a unique answer. For even $n$, the same approach is not adequate because it leads to an underdetermined system.

Find vertices $z_i$ given areas $s_i$

$\endgroup$

2 Answers 2

3
$\begingroup$

The area $s_k$ is $a_k a_{k+1}$ times the area of the triangle with the vertices $0, 1, \omega$, that is $$ \tag{*} s_k = \frac 12a_k a_{k+1} \sin\frac{2 \pi}{n} \, . $$ For odd $n$ is $$ \frac{s_0 s_2 \cdots s_{n-3}s_{n-1}}{s_1 s_3 \cdots s_{n-2}} = \frac 12a_0^2 \sin\frac{2 \pi}{n} $$ so that $a_0$, and by circular permutation, all $a_k$, are uniquely determined.

However, given positive numbers $s_0, \ldots, s_{n-1}$, the so computed lengths $a_0, \ldots, a_{n-1}$ are positive, but not necessarily $\le 1$.

If $n$ is even then both $s_0 s_2 \cdots s_{n-2}$ and $s_1 s_3 \cdots s_{n-1}$ are equal to $$ a_0 a_1 \cdots a_{n+1} \left( \frac 12 \sin\frac{2 \pi}{n}\right)^{n/2} \, , $$ so that $$ s_0 s_2 \cdots s_{n-2} = s_1 s_3 \cdots s_{n-1} $$ is a necessary condition for the existence of a solution. If this condition is satisfied then there is one degree of freedom. If one of the $a_k$ is given then all other $a_j$ are uniquely determined by equation $(*)$.

Again, the so computed lengths $a_0, \ldots, a_{n-1}$ are positive, but not necessarily $\le 1$.

$\endgroup$
2
$\begingroup$

Equations are ($n \ge 3$):
$\forall i=0 \dots n-1, a_i a_{i+1} = 2s_i / \sin(\frac {2\pi} n)$, with $a_n = a_0$.
Taking the $\log$ transforms these into linear equations: if $l_i = \log a_i$, the system is equivalent to
$\forall i=0 \dots n-1, l_i + l_{i+1} = \log(2s_i / \sin(\frac {2\pi} n))$.

The matrix has its diagonal made of $1$, plus a second diagonal above of $1$, and a $1$ at the lower left corner, and zeroes everywhere else.

Exemple for $n=4$: $\begin{pmatrix} 1 & 1 & 0 & 0\\ 0 & 1 & 1 & 0\\ 0 & 0 & 1 & 1\\ 1 & 0 & 0 & 1 \end{pmatrix}$

When developing the determinant by the first line:

  • If we take the $1$ on $(0,0)$, then on column $1$ we cannot take the $1$ on $(0,1)$ so we take the one on $(1,1)$, and then we are forced to take the $1$ on $(2,2)$, etc.: this is the principal diagonal.
  • If we take the $1$ on $(0,1)$, then on line $1$ we cannot take the $1$ on $(1,1)$ so we take the $1$ on $(1,2)$, and then, etc.: we take the second diagonal of $1$s, including the $1$ on the lower left corner as last element.

The first diagonal has the identity permutation, so it has even parity, which makes a $+1$.
The second diagonal above has a cycle permutation, so its parity is even if $n$ is odd, and odd if $n$ is even - this is because an $n$ cycle decomposes into $n-1$ transpositions. Which makes $(-1)^{n-1}$

So the determinant is $1 + (-1)^{n-1}$, i.e. $0$ when $n$ is even and $2$ when $n$ is odd: the system has a unique solution for $n$ odd, and either no solution or an infinity, for $n$ even.

For $n$ even, another way to see it is to use an alternate sum of equations $i=1 \dots n-1$. This gives the left handside of equation $i=0$, so the system is underdetermined: either no solution or an infinity.

The case $n=4$ allows still another proof: one can apply a scaling by $\alpha > 0$ on the $x$ axis and a scaling by $1/ \alpha$ on the $y$ axis, this changes the $(a_i)$ but not the $(s_i)$.

$\endgroup$
4
  • $\begingroup$ It does not matter for the solvability of the system, but I think the triangle area is something like $s_i = \frac 12a_i a_{i+1} \sin\frac{2 \pi}{n}$. $\endgroup$ Commented Dec 9, 2023 at 18:56
  • 1
    $\begingroup$ @MartinR Yes, you are right, I was absent-minded. $\endgroup$ Commented Dec 9, 2023 at 19:09
  • $\begingroup$ Laplace expansion along the first column also shows that the determinant is $1 + (-1)^{n-1}$. $\endgroup$ Commented Dec 9, 2023 at 19:39
  • $\begingroup$ Thanks everyone for discussing those points! $\endgroup$ Commented Dec 18, 2023 at 15:24

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.