2
$\begingroup$

Let $0<a<b$, and let $A, B \in \mathbb{R}^{n \times n}$ be two positive definite symmetric matrices (that do not commute).

My question is: are the matrices $M, P \in \mathbb{R}^{2n \times 2n}$ defined by \begin{align*} M = \begin{bmatrix} bABA & aAB\\ aBA & bBAB \end{bmatrix},\qquad P = \begin{bmatrix} bA & aAB\\ aBA & bB \end{bmatrix} \end{align*} positive definite or not? (In the sense $x^\intercal M x >0$ for any nonzero $x \in \mathbb{R}^{2n}$.)

Why this question: I am trying to find a positive definite symmetric matrix $Q \in \mathbb{R}^{2n \times 2n}$ such that the product \begin{align*} Q \begin{bmatrix} 0 & A^{-1}\\ B^{-1} & 0 \end{bmatrix} \end{align*} is a symmetric matrix. My candidates at the moment are M and P, which are such that \begin{align*} M \begin{bmatrix} 0 & A^{-1}\\ B^{-1} & 0 \end{bmatrix} = \begin{bmatrix} aA & bAB\\ bBA & aB \end{bmatrix}, \qquad P \begin{bmatrix} 0 & A^{-1}\\ B^{-1} & 0 \end{bmatrix} = \begin{bmatrix} aA & bI_n\\ bI_n & aB \end{bmatrix} \end{align*} but I don't know if they are positive definite.

What I tried: I tried to find an invertible matrix $K$ such that the product $K^\intercal P K$ (or $K^\intercal M K$ ) is easier to study, since $P$ is positive definite if and only if $K^\intercal P K$ is. For example, for \begin{align*} K = \begin{bmatrix} A^{-1} & A^{-1}\\ B^{-1} & B^{-1} \end{bmatrix} \end{align*} we have \begin{align*} K^\intercal P K = \begin{bmatrix} bA^{-1} + 2a I_n + bB^{-1} & bA^{-1} - bB^{-1} \\ bA^{-1} - bB^{-1} & bA^{-1} - 2a I_n + bB^{-1} \end{bmatrix}. \end{align*} But I don't know how to proceed for this matrix either.

$\endgroup$
1
$\begingroup$

First of all, we must assume that $b>0$ for either matrix to be positive definite.

For $M$, we compute the Schur complement $$ bABA - \frac{a^2}{b}(AB)(BAB)^{-1}(BA)= bABA - \frac{a^2}{b}A. $$ Since $bABA$ is positive definite, $M$ will be positive definite if and only if $bABA - \frac{a^2}{b}A$ is positive definite. With the Loewner order, we can write this condition as $$ bABA - \frac{a^2}{b}A > 0 \iff bABA > \frac{a^2}{b}A \iff \\ ABA > \frac{a^2}{b^2} A \iff A^{-1/2}(ABA)A^{-1/2} > A^{-1/2}[\frac{a^2}{b^2} A]A^{-1/2} \iff\\ A^{1/2}BA^{1/2} > \frac{a^2}{b^2}I. $$ That is, $M$ will be positive definite if and only if the eigenvalues of $A^{1/2}BA^{1/2}$ are greater than $a^2/b^2$. We see that the matrix $AB$ is similar since $$ AB = A^{1/2}(A^{1/2}BA^{1/2})A^{-1/2}, $$ so $M$ will be positive definite if and only if the eigenvalues of $AB$ are greater than $a^2/b^2$.

For $P$, we compute the Schur complement to be $$ bA - \frac{a^2}{b} ABA $$ and a similar analysis can be applied. We find that $P$ will be positive definite if and only if either $a=0$ or the eigenvalues of $AB$ are less than $b^2/a^2$.

In summary, $M$ and $P$ will be positive definite if and only if $a=0$ or the eigenvalues of $AB$ lie inside the interval $(a^2/b^2,b^2/a^2)$.


A potentially helpful insight: we can reduce your original problem to a possibly simpler case by considering a congruent matrix. For instance: $$ \pmatrix{A^{1/2}\\&B^{1/2}} \pmatrix{0&A^{-1}\\B^{-1}&0} \pmatrix{A^{1/2}\\&B^{1/2}} = \pmatrix{0&A^{-1/2}B^{1/2}\\B^{-1/2}A^{1/2} & 0}\\ % \pmatrix{A^{1/2}\\&A^{1/2}} \pmatrix{0&A^{-1}\\B^{-1}&0} \pmatrix{A^{1/2}\\&A^{1/2}} = \pmatrix{0&I\\A^{1/2}B^{-1}A^{1/2} & 0} $$

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you for your answer and the the insight. I will now try to find a $Q$ that doesn't put such constraints on $A$ and $B$. $\endgroup$ – user344045 Nov 19 '19 at 19:59

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.