4
$\begingroup$

This is a follow up question on from How to solve a non-homogeneous quadratic matrix equation?.

Given the matrix $G = A(A^{-1}M)^{1/2}=A^{1/2}(A^{-1/2}MA^{-1/2})^{1/2}A^{1/2}$, where $A=-H^{-1}$, with $H$ symmetric negative definite and $M$ symmetric positive definite.

Is it possible to establish a clear relationship between the eigenvectors and eigenvalues of $G$, and those of $H$ and $M$?

Also, I have read that $G$ is on the midpoint of the geodesic linking $A$ and $M$, in the space of symmetric positive definite matrices, with distance between two matrices $A$ and $M$ defined as $$ d(A,M) = \| \log(A^{-1/2}MA^{-1/2}) \|_{\mathrm{HS}},$$ where $ \| X \|_{\mathrm{HS}} = \mathrm{Trace} (X^TX)^{1/2}$ . Do you know if it would be possible to have a 3D graphic representation of this relationship for $2\times 2$ matrices, and if so, how one would go about it?

Once again, many thanks.

$\endgroup$

2 Answers 2

2
$\begingroup$

As for visualization, I have seen SPD 2x2 matrices represented as ellipsoids (such as this one, which would be a 3D case), and the segment joining $A$ and $B$ in that Riemannian distance as a sequence of small ellipsoid slowly morphing one into the other (moving either in space or in time). It is a nice geometric visualization, but from a mathematical point of view it does not add too much insight.

I do not know of simple relations between the eigenvalues of $A$, $B$ and $A \#B$, apart from trivial cases where $A$ and $B$ have a common eigenvector.

This might be because in some sense the "natural" operation on that matrix space is not orthogonal conjugation $A\mapsto QAQ^T$ but generic conjugation $A\mapsto MAM^T$, with $M$ not necessarily orthogonal. As you can read in Robert Bryant's answer, a suitable transformation of that kind can diagonalize two generic matrices $A$ and $B$ at the same time. So "what is preserved under orthogonal conjugation" might not be the best question to ask, in an abstract setting. In a concrete setting where you have the matrices and you want to do computations with them, well, I understand perfectly the logic of your question but I do not have a ready answer for it.

(Incidentally, if you are doing computations with matrix means, you might be interested in Bruno Iannazzo's Matlab toolbox

$\endgroup$
2
$\begingroup$

There are several known relations between eigenvalues of $G$ and $A, M$.

However, as the following equality for $2\times 2$ matrices shows, these relations are as rich as the standard inequalities between eigenvalues of sums / products of positive matrices.

More specifically, let $G$ denote the matrix geometric mean, then for $2\times 2$ matrices with unit determinant: \begin{equation*} G = \frac{A+M}{\sqrt{\det(A+M)}}, \end{equation*} so that $\lambda(G) \propto \lambda(A+M)$. For not unit determinants, one can suitably scale the matrices above to obtain a corresponding expression.


$\newcommand{\reals}{\mathbb{R}}\newcommand{\gm}{\sharp}$ For the general case, the following may be of interest. I'll use $A\gm_t M$ to denote the point on the geodesic between $A$ and $M$ given by \begin{equation*} A\gm_t M := A^{1/2}(A^{-1/2}MA^{-1/2})^tA^{1/2}, \end{equation*} so in particular the midpoint $G=A\gm_{1/2}M$ is the matrix geometric mean.

Theorem. Let $\prec_{\log}$ denote the log-majorisation order, i.e., for $x, y \in \reals_{++}^n$ ordered nonincreasingly, we say $x \prec_{\log} y$ if $\prod_{i=1}^{n-1}x_i \le \prod_{i=1}^{n-1}y_i$ and $\prod_{i=1}^nx_i=\prod_{i=1}^ny_i$. Then, for $A, M > 0$ and $t \in [0,1]$, we have the log-majorisation between the eigenvalues: \begin{equation*} \lambda(A\gm_t M) \prec_{\log} \lambda(A^{1-t}M^t) \prec_{\log} \lambda(A^{1-t})\lambda(M^t). \end{equation*}

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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