Montononicity of the optimal cost in the discrete-time regulator problem and Schur complements (Q5950251)

A discrete-time regulator problem \[ x(t+ 1)= Fx(t)+ Gu(t),\quad x(0)= x_0,\quad y(t)= Hx(t),\quad t= 0,1,\dots,\tag{1} \] \[ \text{minimize}\quad J(x_0, u)= \sum^\infty_{t=0} (y(t)^* y(t)+ u(t)^* u(t)) \] is considered. If the system is output stabilizable, an optimal control \(u_{\text{opt}}\) exists. Let \(J_{\text{opt}}(x_0)= J(x_0, u_{\text{opt}})\), the optimal performance index. A matrix \(D= \left[\begin{smallmatrix} H^*H & F^*\\ F & -GG^*\end{smallmatrix}\right]\) is associated with the system. It is known that if the system (1) is output stabilizable, then so is any system \((\widetilde F,\widetilde G,\widetilde H)\) whose associated matrix \(\widetilde D\) is such that \(D\geq\widetilde D\), i.e., \(D-\widetilde D\) is positive semidefinite. The main result: If the system (1) is output stabilizable and \(D\geq\widetilde D\), then for the optimal performance indices we have \(J_{\text{opt}}(x_0)\geq \widetilde J_{\text{opt}}(x_0)\) for every \(x_0\). The main result is extended to the more general performance index \[ \sum^\infty_{t=0} [x(t)^* u(t)^*]\Biggl[\begin{matrix} Q & T^*\\ T & R\end{matrix}\Biggr]\Biggl[\begin{matrix} x(t)\\ u(t)\end{matrix}\Biggr], \] where \(\left[\begin{smallmatrix} Q &T^*\\ T & R\end{smallmatrix}\right]\) is positive semidefinite and \(R\) is positive definite.