8
$\begingroup$

$$\forall i,\forall j\neq i,\quad\sum_{k=1}^n\frac{a_i-a_k}{a_i+a_k}\cdot\frac{a_j-a_k}{a_j+a_k}=0.$$

We can't have two different $a_i=0$ because of the denominators; but we can allow one $a_i=0$, if the terms $k=i$ and $k=j$ are excluded from the sum.

For $n=3$, these equations are easy to solve: \begin{align*} (1,2):\quad&\frac{a_1-a_3}{a_1+a_3}\cdot\frac{a_2-a_3}{a_2+a_3}=0\\ (1,3):\quad&\frac{a_1-a_2}{a_1+a_2}\cdot\frac{a_3-a_2}{a_3+a_2}=0\\ (2,3):\quad&\frac{a_2-a_1}{a_2+a_1}\cdot\frac{a_3-a_1}{a_3+a_1}=0. \end{align*} Indeed we just get $a_1=a_2=a_3$.

For $n=4$, the first of $6$ equations is

$$(1,2):\quad\frac{a_1-a_3}{a_1+a_3}\cdot\frac{a_2-a_3}{a_2+a_3}+\frac{a_1-a_4}{a_1+a_4}\cdot\frac{a_2-a_4}{a_2+a_4}=0.$$

(For the other $5$, just permute the indices.) I multiplied to clear the denominators, then added equations $(1,2)$ and $(3,4)$ to get

$$4(a_1a_2-a_3a_4)^2=0$$

and thus

$$a_1a_2=a_3a_4,\quad a_1a_3=a_2a_4,\quad a_1a_4=a_2a_3.$$

These imply that $a_1^2=a_2^2=a_3^2=a_4^2$; and we can't have $a_i=-a_j$, again because of the denominators. So $a_1=a_2=a_3=a_4$.

Does this continue for $n\geq5$?

If the variables are non-negative real numbers, then we can arrange them in order, $a_1\geq a_2\geq a_3\geq\cdots\geq a_n\geq0$; equation $(1,2)$ is then a sum of non-negative terms, so each term must vanish, which gives $a_2=a_3=\cdots=a_n$. Then equation $(2,3)$ has only its first term remaining, which gives $a_1=a_2$.

What if some of the variables are negative, or complex numbers?


We might define $b_{ij}=\dfrac{a_i-a_j}{a_i+a_j}$ to simplify the equations to $\sum_kb_{ik}b_{jk}=0$. Collecting these into an antisymmetric matrix $B$, we see that the system of equations is just saying that

$$BB^T=-B^2=B^TB=D$$

is some diagonal matrix. But I don't think this tells us enough about $B$ itself.

The defining equation for $b_{ij}$ can be rearranged to

$$a_j=\frac{1-b_{ij}}{1+b_{ij}}a_i$$

so in particular

$$a_3=\frac{1-b_{13}}{1+b_{13}}a_1=\frac{1-b_{23}}{1+b_{23}}a_2=\frac{1-b_{23}}{1+b_{23}}\cdot\frac{1-b_{12}}{1+b_{12}}a_1;$$

cancelling $a_1$,

$$(1+b_{31})(1+b_{23})(1+b_{12})=(1-b_{31})(1-b_{23})(1-b_{12});$$

expanding,

$$2b_{12}b_{23}b_{31}+2b_{12}+2b_{23}+2b_{31}=0.$$

In this process I divided by some things that might be $0$, but this resulting cubic equation is valid nonetheless.

I think we can dispense with $a_i$ now. In summary, we need to solve the system of equations \begin{align*} \forall i,\forall j,\quad&b_{ij}+b_{ji}=0\\ \forall i,\forall j,\forall k,\quad&b_{ij}b_{jk}b_{ki}+b_{ij}+b_{jk}+b_{ki}=0\\ \forall i,\forall j\neq i,\quad&\sum_kb_{ik}b_{jk}=0. \end{align*} Is the only solution $b_{ij}=0$?

$\endgroup$
5
+150
$\begingroup$

$\def\C{\mathbb{C}}$This answer solves the system of equations\begin{gather*} \sum_{k = 1}^n \frac{a_i - a_k}{a_i + a_k} · \frac{a_j - a_k}{a_j + a_k} = 0 \quad (\forall i ≠ j) \tag{$*$} \end{gather*} in $\C$ and the italic letter $i$ is not the imaginary unit $\mathrm{i}$.

On the one hand, suppose $(a_1, \cdots, a_n) \in \C^n$ is a solution to ($*$). For any $i, j, k$,\begin{gather*} \frac{a_i - a_k}{a_i + a_k} - \frac{a_j - a_k}{a_j + a_k} = \frac{2(a_i - a_j) a_k}{(a_i + a_k) (a_j + a_k)}\\ = \frac{a_i - a_j}{a_i + a_j} · \frac{2(a_i + a_j) a_k}{(a_i + a_k) (a_j + a_k)} = \frac{a_i - a_j}{a_i + a_j} \left( 1 - \frac{a_i - a_k}{a_i + a_k} · \frac{a_j - a_k}{a_j + a_k} \right), \end{gather*} thus\begin{gather*} \sum_{k = 1}^n \left( \frac{a_i - a_k}{a_i + a_k} - \frac{a_j - a_k}{a_j + a_k} \right) = \sum_{k = 1}^n \frac{a_i - a_j}{a_i + a_j} \left( 1 - \frac{a_i - a_k}{a_i + a_k} · \frac{a_j - a_k}{a_j + a_k} \right)\\ = n · \frac{a_i - a_j}{a_i + a_j} - \frac{a_i - a_j}{a_i + a_j} \sum_{k = 1}^n \frac{a_i - a_k}{a_i + a_k} · \frac{a_j - a_k}{a_j + a_k} \stackrel{(*)}{=} n · \frac{a_i - a_j}{a_i + a_j}. \tag{1} \end{gather*} Define $c_i = \dfrac{1}{n} \sum\limits_{k = 1}^n \dfrac{a_i - a_k}{a_i + a_k}$ for all $i$, then (1) implies that $\dfrac{a_i - a_j}{a_i + a_j} = c_i - c_j$, i.e.\begin{gather*} (1 - c_i + c_j) a_i = (1 - c_j + c_i) a_j \quad (\forall 1\leqslant i, j \leqslant n). \tag{2} \end{gather*} Note that for any $1 \leqslant i < j < k \leqslant n$ with $a_i, a_j, a_k ≠ 0$,$$ \begin{cases} (1 - c_i + c_j) a_i = (1 - c_j + c_i) a_j\\ (1 - c_j + c_k) a_j = (1 - c_k + c_j) a_k\\ (1 - c_k + c_i) a_k = (1 - c_i + c_k) a_i \end{cases} $$ imply that$$ (1 - c_i + c_j)(1 - c_j + c_k)(1 - c_k + c_i) = (1 - c_j + c_i)(1 - c_k + c_j)(1 - c_i + c_k), $$ which is simplified to be\begin{gather*} (c_i - c_j)(c_j - c_k)(c_k - c_i) = 0. \tag{3} \end{gather*}

Case 1: If $a_{i_0} = 0$ for some $i_0$, then $a_i ≠ 0$ for all $i ≠ i_0$ because of non-zero denominators in ($*$), and (2) implies that $c_i = c_{i_0} + 1$ for all $i ≠ i_0$. Thus for any $i, j ≠ i_0$, (2) implies that $a_i = a_j$.

Case 2: If $a_i ≠ 0$ for any $i$, then (3) implies that among any $c_i, c_j, c_k$, there are at least two equal to each other. Thus all $c_i$'s assume at most two values, and whenever $c_i = c_j$ for some $i$ and $j$, (2) implies that $a_i = a_j$.

To summarize, all possible $(a_1, \cdots, a_n)$'s (up to permutation) are of the form$$ (\underbrace{a, \cdots, a}_{m \text{ copies of } a}, \underbrace{b, \cdots, b}_{n - m \text{ copies of } b}) $$ where $2 \leqslant m \leqslant n$ (since $n \geqslant 3$), $a, b \in \C$ and $a ≠ b$. Now without loss of generality assume that $a_1 = a_2 = a$, then$$ 0 \stackrel{(*)}{=} \sum_{k = 1}^n \frac{a_1 - a_k}{a_1 + a_k} · \frac{a_2 - a_k}{a_2 + a_k} = \sum_{k = 1}^n \left( \frac{a - a_k}{a + a_k} \right)^2 = (n - m) \left( \frac{a - b}{a + b} \right)^2, $$ which implies that $m = n$. Therefore all $a_i$'s are equal.

On the other hand, it is easy to verify that $(a_1, \cdots, a_n) = (a, \cdots, a)$ $(a \in \C^*)$ are indeed solutions to ($*$). Therefore they are all the solutions.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ This could be simplified by using $b_{ij}$ instead of $a_i$. $\endgroup$ – mr_e_man Jul 8 at 15:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.