When does $\mbox{tr} \left( A^2 \right) = \mbox{tr} (A)^2$ hold?

Suppose $A$ is a square matrix whose eigenvalues are $\lambda_1, \lambda_2, \dots, \lambda_n$. The equation

$$\mbox{tr} \left( A^2 \right) = \mbox{tr} (A)^2$$

implies that $$\sum_{i\not=j}\lambda_i\lambda_j=0$$

What can we learn about the matrix $A$ from the equation above? What kind of matrix has this property?


Solution 1:

The particular case $n=2$ has a nice specific proof.

As the characteristic equation of $A$ can be written under the form

$$\lambda^2- \operatorname{tr}(A)\lambda +\det(A)=0,$$

Cayley–Hamilton theorem gives

$$A^2-\operatorname{tr}(A) A +\det(A)I=0.$$

Taking the trace (linear operator) :

$$\operatorname{tr}(A^2)-\operatorname{tr}(A)^2=-2\det(A)$$

Therefore the issue is simply to look for matrices with zero determinant, i.e., matrices such that $\lambda_1\lambda_2=0$.

See this excellent document by Darij Grinberg.

See as well : (https://math.stackexchange.com/q/1267700.)

Solution 2:

This is really more a question about collections of numbers than about linear algebra.

As you know, $\operatorname{tr}A=\sum_i\lambda_i$ and $\operatorname{tr}A^2=\sum_i\lambda_i^2$. Let $\mu=\frac1n\sum_i\lambda_i$ and $\sigma^2=\left(\frac1n\sum_i\lambda_i^2\right)-\mu^2$ be the mean and variance of the eigenvalues. (The latter is actually the pseudo-variance if the eigenvalues are complex.) Then, $\operatorname{tr}A=n\mu$ and $\operatorname{tr}A^2=n(\mu^2+\sigma^2)$. So $(\operatorname{tr}A)^2=\operatorname{tr}A^2$ if and only if $n\mu^2=\mu^2+\sigma^2$, i.e. $$\mu=\pm\frac\sigma{\sqrt{n-1}}.$$

To me there does not seem to be anything very special about this property. For any matrix $A$ you can add a scalar multiple of the identity to obtain a matrix which satisfies it, i.e. $\operatorname{tr}^2(A+kI)=\operatorname{tr}(A+kI)^2$ for $k=\pm\sigma/\sqrt{n-1}-\mu$.

Solution 3:

$\textbf{Part 1.}$ Let $Z=\{X\in M_n(\mathbb{C});\;(*)\;tr(X^2)=(tr(X))^2\}$; since the studied relation $(*)$ is homogeneous, $Z$ is a complex algebraic cone of dimension $n^2-1$ (that is, $n^2-1$ independent complex parameters).

$\textbf{Part 2.}$ The relation $(*)$ can be written $tr(X^2-tr(X)X)=0$, that is equivalent to

there are $U,V\in M_n(\mathbb{C})$ s.t. $(**)$ $X^2-tr(X)X=UV-VU$.

Conversely, if you want to construct such matrices $X$ without using its eigenvalues, then you can do as follows -when $n\geq 3$ (the case $n=2$ has been solved by Jean Marie)-

i) Randomly choose $U,V$

ii) Solve the equation $(**)$. In general (generic choice of $U,V$) , this equation has $2,8,22,52$ solutions (pairwise opposite) when $n=3,4,5,6$.

Solution 4:

Essentially, we have the following equation in $\mathrm x \in \mathbb C^n$

$$\mathrm x^\top \left( \mathrm I_n - 1_n 1_n^\top \right) \mathrm x = 0$$

Note that matrix $\mathrm I_n - 1_n 1_n^\top$ has eigenvalue $1$ with multiplicity $n-1$ and eigenvalue $1-n$ with multiplicity $1$. Hence, the equation above can be rewritten in the eigenvector basis as follows

$$y_1^2 + y_2^2 + \cdots + y_{n-1}^2 - (n-1) y_n^2 = 0$$

Note that $\mathrm y = 1_n$ is a solution.