4
$\begingroup$

I am looking to solve the following matrix equation for $G$

$$GHG + M = 0$$

where $G$, $H$, and $M$ are square, symmetric, real matrices. $H$ is negative-definite and $M$ is positive-definite. $G$ should also be positive-definite.

Is it possible?

Many thanks!

$\endgroup$
4
  • $\begingroup$ Are you sure about those signs? If $G$ has to be symmetric, then $GHG$ has the same signature as $H$, that is, negative definite, but $GHG=M$ tells you that it needs to be positive definite. $\endgroup$ Commented Oct 13, 2014 at 9:43
  • $\begingroup$ @RobertBryant I saw your answer only briefly -- why did you delete it? It looked correct to me, and it added some insight on the origin of the solution which my answer currently lacks. $\endgroup$ Commented Oct 13, 2014 at 12:28
  • $\begingroup$ @FedericoPoloni: I deleted it when I saw that you had already posted an answer. (You must have posted after I started typing, because I didn't see your answer until I hit 'post'.) I didn't feel that my answer really added anything after yours, so I deleted it to keep people from voting for it and thus drawing votes from your prior answer. $\endgroup$ Commented Oct 13, 2014 at 12:42
  • 3
    $\begingroup$ @RobertBryant Thanks for your consideration, but I encourage you to undelete it, I think it was useful. There is nothing wrong with two answers that give different views and information. People can upvote us both. :p $\endgroup$ Commented Oct 13, 2014 at 12:48

2 Answers 2

8
$\begingroup$

(commenting about the equation with the plus sign, I hope that the correction was right).

This is one of the few quadratic matrix equations that have a closed form solution. Set $A=-H^{-1}$; then $G = A(A^{-1}M)^{1/2}=A^{1/2}(A^{-1/2}MA^{-1/2})^{1/2}A^{1/2}$. Here the operation $X^{1/2}$, for a matrix $X$, denotes the unique square root with spectrum in the right half-plane. The solution $G$ is known as the matrix geometric mean of the positive definite matrices $M$ and $A$, and denoted by $A\operatorname{\#}M$. You can see Section 4.1 of Bhatia's book Positive definite matrices for more information and proofs of these facts.

$\endgroup$
5
$\begingroup$

Yes, this is always possible, and $G$ is unique. Here is how you can see this:

Consider the pair of positive definite symmetric matrices $(-H^{-1}, M)$. By a well-known theorem (simultaneous diagonalization of quadratic forms), there exists an invertible matrix $A$ such that $$ -H^{-1} = AA^T\qquad\text{and}\qquad M = A D A^T $$ where $D$ is positive definite and diagonal. Then the equation you want to solve becomes $$ G (A^T)^{-1} A^{-1} G = ADA^T. $$ Setting $\bar G = A^{-1}G(A^T)^{-1}$, this becomes $\bar G^2 = D$, which has a unique positive definite symmetric (in fact, diagonal) solution, $\bar G = \sqrt{D}$. Now just set $$ G = A\bar G A^T = A\sqrt{D}\,A^T. $$

Note: Explicitly finding $A$ is a matter of first finding the diagonal elements of $D$ as the roots (which are all positive) of the equation $\det(M+\lambda H^{-1}) = 0$, or, equivalently $\det(MH+\lambda I) = 0$. Then, using these, you get the columns of $A$ as the corresponding eigenvectors.

$\endgroup$

You must log in to answer this question.

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