State-space formulae for all stabilizing controllers that satisfy an \(H_{\infty}\)-norm bound and relations to risk sensitivity (Q1119528): Difference between revisions

Let a linear system be described by the state equation \(\dot x(T)=Ax(t)+B_ 1w(t)+B_ 2u(t)\) \(z(t)=C_ 1x(t)+D_{11}w(t)+D_{12}u(t)\) \(y(t)=C_ 2x(t)+D_{21}w(t)+D_{22}u(t),\) with the transfer function \[ P(s)=\left[ \begin{matrix} P_{11}\\ P_{21}\end{matrix} \begin{matrix} P_{12}\\ P_{22}\end{matrix} \right]=\left[ \begin{matrix} D\quad_{11}\\ D_{21}\end{matrix} \begin{matrix} D_{12}\\ D_{22}\end{matrix} \right]+\left[ \begin{matrix} C_ 1\\ C_ 2\end{matrix} \right](\quad sI-A)^{-1}[B_ 1B_ 2] \] (the matrices, A, \(B_ j\), \(C_ j\), \(D_{ij}\) are constant). For the linear controller with transfer function K(s) connected from y to u, the closed-loop transfer function w to z will be denoted \[ F(P,K)=P_{11}+P_{12}K(I-P_{22}K)^{-1}P_{21}. \] The problem of finding all internally stabilizing K such that \(\| F(P,K)\|_{\infty}<\gamma\) for some prespecified \(\gamma\) is considered. A state-space parametrization of all such controllers is given, which involves two algebraic Riccati equations each of degree equal to the size of A. Interesting relations between this \(H_{\infty}\)-bounded controller problem and the risk sensitive LQG control are pointed out. The latter problem has to do with the discrete time system \(x_ t=Ax_{t-1}+B_ 1w_ t+B_ 2u_{t-1}\) \(z_ t=C_ 1x_{t- 1}+D_{11}w_ t+D_{12}u_{t-1}\) \(y_ t=C_ 2x_{t-1}+D_{21},\) where \(w_ t\) is Gaussian with unit variance. Then the risk sensitive optimal controller minimizes \[ -\frac{2}{\theta T}\log E[\exp (- \frac{\theta}{2}(x^*_ T\pi x_ T+\sum^{T-1}_{t=0}z^*_ tz_ t))], \] where E denotes expectation and \(\theta\) is a scalar parameter (risk sensitivity).