2
$\begingroup$

I have the following quadratic matrix equation:

$$ XAX+X = B $$

where both $A$ and $B$ are given positive definite matrices, and $X$ is a covariance matrix and, hence, positive definite.

When there is no constraint, the equation can be solved via Bernoulli iteration in the following form:

$$X_{k+1} = -A^{-1}(I-BX_k^{-1})$$

However, this does not seems to preserve positive semidefinite.

Any guidance would be appreciated. Thank you.

$\endgroup$
  • $\begingroup$ another iteration would be $X_{k+1}=B-X_k A X_k$. This would preserve the symmetry of the matrix and, if you start close enough to a solution, also the positive definiteness. If it converges depends however on $A$ and $B$. $\endgroup$ – Markus Sprecher Nov 7 '19 at 10:04
  • 1
    $\begingroup$ Let $f(X)=XAX+X$ we have $f(X+dX)=f(X)+dX\cdot A\cdot X+X\cdot A\cdot dX+dX+O(|dX|^2)$. We could define an iteration by $X_{k+1}=X_k+dX$ where $dX$ is the solution to $f(X_k)+dX\cdot A\cdot X_k+X_k\cdot A\cdot dX+dX=B$. This is a continuous Lyapunov equation. $\endgroup$ – Markus Sprecher Nov 7 '19 at 10:09
2
$\begingroup$

Turning my previous comments into an answer.

What you have is an algebraic Riccati equation: indeed, setting $F=-\frac12 I$, you get $F^TX+XF + B = XAX$. Since $A$ and $B$ are positive definite, the pairs $(F,A)$ and $(F^T,B)$ are controllable and observable, so the classical solvability criteria are satisfied and the equation has a unique stabilizing, positive definite solution. You will find solution algorithms implemented in most numerical software; among them, Matlab's care and icare, or scipy.linalg.solve_continuous_are.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you Poloni! I have just tried it and it works very well! $\endgroup$ – lisi Nov 10 '19 at 1:09

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.