I would like to find a matrix (or ideally all matrices) $\Sigma$ such that $$ (1) \quad P + (P \Sigma) + (P \Sigma)^T + (P \Sigma H) P^{-1} (P \Sigma H)^T \leq 0, $$ where '$\leq$' denotes negative semi-definiteness and $P$ and $H$ are prescribed symmetric positive definite square matrices. This is a similar to the Riccati equation $$ (2) \quad Q + (XA) + (XA)^T - (XB) R^{-1} (XB)^T \leq 0. $$ In fact (1) is obtained from (2) by letting \begin{align} P &= Q = X = -R \\ \Sigma &= A \\ \Sigma H &= B. \end{align} Most literature about the Riccati equation seems to be focused on finding $X$ given $A$, $B$ and $R$, or in my case, finding $P$ given $\Sigma$ and $H$. But what can be said about the converse? Given $P$ and $H$, does a $\Sigma$ exist such that (1) is satisfied? If so, how can it be found?

  • $\begingroup$ The solution part seems easier to me. You can directly transform the equation into an LMI with $\Sigma$ the unknown matrix. $\endgroup$ – RTJ Apr 26 '17 at 10:53
  • $\begingroup$ @CTNT Thank's for your input. Could you elaborate a little on your idea? I'm not sure I follow. $\endgroup$ – ekkilop Apr 26 '17 at 11:56
  • $\begingroup$ I just noticed that you need negative semi-definiteness and not necessarily negative definiteness. What I meant is that in the latter case, you can use Schur's complement idea to transform the nonlinear Riccati ineq. into a larger linear inequality which can be easily solved with the LMI toolbox of MATLAB. $\endgroup$ – RTJ Apr 26 '17 at 12:33
  • 2
    $\begingroup$ Specifically, if that was a strict inequality then the respective LMI is $$\left[\array{P+P\Sigma+\Sigma^TP & P\Sigma H\\ H^T\Sigma^TP & -P}\right]<0$$ $\endgroup$ – RTJ Apr 26 '17 at 13:43

Let $Q=HP^{-1}H$ and $X=Q^{1/2}P\Sigma Q^{1/2}$. Left- and right-multiply your equation $(1)$ by $Q^{1/2}$, we get $Q^{1/2}PQ^{1/2} + X + X^T + XX^T\le0$, i.e. $$ I - Q^{1/2}PQ^{1/2} \ge (I+X)(I+X)^T.\tag{3} $$ Since the RHS of $(3)$ is positive semidefinite, the equation is solvable only when the spectral norm/radius of $Q^{1/2}PQ^{1/2}$ is at most $1$. If this is the case and we orthogonally diagonalise $I-Q^{1/2}PQ^{1/2}$ as $USU^T$, then any matrix of the form $X=UDV^T-I$, where $V$ is any real orthogonal matrix and $D$ is any diagonal matrix that satisfies $0\le D\le S^{1/2}$, is a solution to $(3)$, so that $\Sigma=P^{-1}Q^{-1/2}XQ^{-1/2}$ solves your equation.


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.