Frequency-weighted optimal Hankel-norm approximation of stable transfer functions (Q2265974)

This paper shows how to modify the optimal Hankel-norm approximation problem for scalar, finite-dimensional, linear, time-invariant systems to allow for frequency weighting G. Here G(z) is assumed to be a strictly minimum phase, strictly stable, real rational scalar transfer function. Let r be the smallest integer such that \(\lim_{z\to \infty}z^ rG(z)\) is nonzero. Then, define the tilde operation by \(\tilde G(z)=z^{- r}G(z^{-1}).\) We first restate the optimal Hankel-norm approximation method. Let \(\Gamma\) (F) denote the Hankel matrix of a proper stable rational transfer function F(z) which has McMillan degree n. Then, for any integer \(k<n\) there exists a unique bounded Hankel matrix \({\bar \Gamma}\) of rank k such that \(\| \Gamma (F)-{\bar \Gamma}\| =\sigma_{k+1}(F),\) where \(\sigma_{k+1}(F)\) is the \((k+1)\)-st singular value of \(\Gamma\) (F) (the singular values being ordered in descending magnitude). Further the matrix \({\bar \Gamma}\) is the Hankel matrix of some rational \(\chi\) (z) with \(\chi (z)=F(z)-\sigma_{k+1}(F)\phi (z),\) where \(\phi\) is an all-pass function with exactly k strictly stable poles. \(\chi\) (z) is a unique best \(L^{\infty}\)-approximation to F(z). When setting \(F=H\tilde G\), we identify \(\chi\) \(\tilde G^{-1}\) as an \(L^{\infty}\)-approximation to H(z), and \(\bar W=[W]_+=[\chi \tilde G^{-1}]_+\) as a stable reduced order approximation to H(z) where \([W]_+\) denotes the strictly stable part of W. \(\bar W\) can be expressed as \(\bar W=H-\sigma_{k+1}(H\tilde G)[\phi \tilde G^{- 1}]_+.\) We can take \(\bar W\) as a stable k-th order frequency-weighted approximation to H(z). A lower bound of \(L^{\infty}\)-errors of the approximation is given. For the continuous-time case, the tilde operation is defined by \(\tilde G(s)=G(-s)\) where G(s) is proper, strictly stable, and minimum phase and \(\lim_{s\to \infty}G(s)\) is finite and nonzero. An optimal Hankel-norm reduction can be performed on \(F(s)=[H(s)\tilde G(s)]_+,\) where H(s) is strictly stable and proper of order n. We can take \(\tilde W=[\chi \tilde G^{-1}]_+\) as the k-th order frequency- weighted approximation to H(s). A numerical example is also given.