Well-conditioned computations for \(H^ \infty\) controller near the optimum (Q1373133)

According to well-known results, the state-space solution to the \(\gamma\)-attenuation problem \((\gamma\)-suboptimal controller) can be easily determined by using the stabilizing positive semidefinite solutions \(X(\gamma)\) and \(Y(\gamma)\) to a pair of game theoretic Riccati equations (GTRE), assuing that such solutions exist and the spectral norm of their product is less than \(\gamma^2\) (or equal). It is also a known fact that when approaching the optimum \(\gamma_0\), the implementation of explicit formulae for the suboptimal controller may determine ill-conditioned computations; such a situation appears when the spectral norm condition fails before the ones refering to the solutions of GTREs. This fact is due to the matrix \([\gamma^2 I-X(\gamma)Y(\gamma)]^{-1}\) which is contained in the realization of the solution to the suboptimal controller and which is large near the optimum, since the matrix \(\gamma^2 I-X(\gamma)Y(\gamma)\) tends to be singular when \(\gamma\) tends to \(\gamma_0\). In order to avoid such a situation, a descriptor form for the suboptimal controller was adopted which leads to a descriptor form for the optimal controller. A different approach was proposed in order to determine a well-conditioned solution near and at the optimum \(\gamma_0\) starting with a (not strictly) proper \(\gamma\)-suboptimal controller. The authors propose an alternative theory to avoid the ill-conditioning when approaching the optimum. The theory provides not only a procedure but leads to explicit formulae for the optimal controller; the construction of the optimal controller is based on using a balanced realization of the given plant with respect to the stabilizing solutions of game-theoretic Riccati equations and performing the usual reductions according to singular perturbation methods.