Weighted Hankel-norm approximation: Calculation of bounds (Q1078131)

This paper gives bounds on the frequency-domain error associated with the frequency-weighted optimal Hankel-norm approximation. Let C(s) be a real stable rational transfer function of degree n. Let \(\hat G(\)s) be the frequency-weighting function. Suppose that \(\hat G(s)\) is a table rational minimum-phase tranfer function, and further \(\hat G(\infty)=0\). Let us set \(\check G(s)=\hat G(-s)\). Let X be a degree-k Hankel-norm approximation of \(C\check G.\) Set \(\hat C=[X\check G^{-1}]\) where \([Z]_-\) denotes taking the constant part and the strictly proper stable part of Z. Then \(\hat C\) has degree k and is stable, and \(\| [C-\hat C]\check G\|_ H\) is minimized over all stable \(\hat C\) of degree k, where \(\|.\|_ H\) denotes the Hankel-norm. Accordingly, C is a frequency-weighted optimal Hankel-norm approximation of C. Let \(E=[C\check G_- -X=[(C-\hat C)\check G]_-\). The bounds on \(\| E\|_{\infty}\) are given by \(\| E\|_{\infty}\leq \sigma_{k+1}+...+\sigma_ n\) where the \(\sigma_ i's\) are the Hankel singular values, and the \(L^{\infty}\)-norm is taken on the \(i\omega\) axis. Let us assume that we are working in the discrete-time domain (z- domain) by setting \(z=(1+s)/(1-s).\) The bounds on the error of the frequency-weighted optimal Hankel-norm approximation are give by \(\| [C-\hat C]\|_{\infty}\leq \| E\|_{\infty}+\mu \| z^{-1}E\|_ H\cdot \| \hat G^{- 1}\|_ 2\cdot \| \hat G\|_ H\leq [1+\mu (\check G)\| \hat G^{-1}\|_ 2\cdot \| \hat G\|_ h]\| E=_{\infty}\) where \(\mu^ 2\) is an upper bound on the ratio \(\| \check G\|^ 2_{\infty}/\| \check G\|^ 2_ 2\), computable for \(\check G\) and depending only on the poles of \(\check G\), and the \(L^{\infty}\)- norm is taken on the unit circle. Some numerical examples are given in order to illustrate the computation.