1
$\begingroup$

In a previous question here, I asked the question below for block matrices and received an answer showing the question is true if $\mathcal B$ is hermitian and false, in general if $\mathcal B$ is non-hermitian. However, numerical experiments suggest it is still true if we are talking only about matrices rather than block matrices and this is the content of this question. We consider matrices $$\mathcal A = \begin{pmatrix} 0 & a\\\bar a& 0 \end{pmatrix}$$ and $$\mathcal B = \begin{pmatrix} 0 & b\\c & 0 \end{pmatrix}$$ with $a,b,c \in \mathbb C.$

Then we define the new matrix $$T(t) = \begin{pmatrix} \mathcal A+t & \mathcal B \\ \mathcal B^* & \mathcal A-t\end{pmatrix}.$$

Numerical experiments seem to show that the eigenvalues of $[0,\infty) \ni t\mapsto T(t)$ have the property that their absolute values are monotonically increasing in $t \ge 0.$ However, I do not have a proof of this, does anybody know how this follows? (The eigenvalues of $T(t)$ seem to come in pairs $\pm \lambda$ with $\lambda = \lambda(t) \ge 0$, i.e. $+\lambda(t)$ is increasing, while $-\lambda(t)$ is decreasing.

To illustrate the effect, consider

$$T(t)=\begin{pmatrix} t & 1& 0& 2\\ 1 & t & 0& 0\\ 0 & 0& -t & 1\\ 2& 0 & 1 & -t \end{pmatrix},$$ then the eigenvalues of $T(t)$ are $$ \pm 1 \mp \sqrt{2+t^2}.$$

Please let me know if you have any questions.

$\endgroup$
4
  • $\begingroup$ How does this new question differ from [1] ? $\endgroup$
    – Kurt G.
    Commented Mar 17, 2022 at 12:05
  • $\begingroup$ @KurtG. [1] is about block-matrices, the question has been answered when $B$ is self-adjoint and shown to be false when $B$ is not self-adjoint (in general). Here, I consider now the case that $A$ and $B$ are not block matrices but matrices, in which case the statement still seems to hold, even when $B$ is not hermitian. $\endgroup$
    – Sascha
    Commented Mar 17, 2022 at 12:20
  • $\begingroup$ The previous question was too general so that it was false in general and true in the special case of $B$ self-adjoint. The new question is a special case of the previous question, but hopefully true even if $B$ is not self-adjoint. The self-adjoint case is already proven. $\endgroup$
    – Sascha
    Commented Mar 17, 2022 at 12:34
  • $\begingroup$ @KurtG. sure, I think I did this 5 minutes ago, right? :) $\endgroup$
    – Sascha
    Commented Mar 17, 2022 at 12:39

1 Answer 1

4
$\begingroup$

The characteristic polynomial is even in both $X$ and $t$ : $P_t(X)=Q(X^2,t^2)$ where $$Q(Y,s)=(Y-s)^2-(2|a|^2+|b|^2+|c|^2)(Y-s)-4|a|^2s+|a^2-b\bar c|^2.$$ The variation of $s\mapsto Y(s)$ is given by the derivative $$2(Y-s)(Y'-1)-(2|a|^2+|b|^2+|c|^2)(Y'-1)-4|a|^2=0.$$ The sign of $Y'$ changes when $s$ crosses the value $$s_0=\frac1{4|a|^2}(2|a|^2+|b|^2+|c|^2)(2|a|^2-|b|^2-|c|^2).$$ This value is admissible for $t_0^2$ if it is positive, that is is $2|a|^2>|b|^2+|c|^2$.

In conclusion, the absolute values of the eigenvalues are monotonous functions of $s$ whenever $2|a|^2\le|b|^2+|c|^2$. If on the contrary $2|a|^2>|b|^2+|c|^2$, then one of the eigenvalues is not monotone, and the change happens when $t$ crosses $$t_0=\frac1{2|a|}\sqrt{(2|a|^2+|b|^2+|c|^2)(2|a|^2-|b|^2-|c|^2)}.$$

$\endgroup$
1
  • $\begingroup$ Nice work. Deleted my answer because I only ended up calculating the same characteristic polynomial. $\endgroup$
    – Kurt G.
    Commented Mar 18, 2022 at 12:54

You must log in to answer this question.

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