Computation of signature symmetric balanced realizations (Q1407244)

The aim of this paper is to present a numerical algorithm for computing balanced signature symmetric realizations. More exactly, a Jacobi-type algorithm is developed to minimize the cost function \[ f_N: M(\Sigma)\to \mathbb{R},\qquad f_N(W)= \text{tr}(NW), \] where \[ M(\Sigma):= \{T\Sigma T'\mid T\in O^+_{pq}(\mathbb{R})\}, \] \(N= \text{diag}(\mu_1,\dots, \mu_p,\nu_1,\dots,\nu_q)\) with \(0< \mu_1<\cdots< \mu_p\) and \(0< \nu_1< \nu_2<\cdots< \nu_q\), the set \[ O_{pq}(\mathbb{R}):= \{T\in\mathbb{R}^{n\times n}\mid T I_{pq}T'= I_{pq}\} \] is the real Lie group of pseudo-orthogonal \((n\times n)\)-matrices stabilizing \(I_{pq}\) by congruence, the set \(O^+_{pq}(\mathbb{R})\) denotes the identity component of \(O_{pq}(\mathbb{R})\), and \(\Sigma= \text{diag}(\sigma_1,\dots,\sigma_n)\) where the \(\sigma_1,\dots,\sigma_n\) are the Hankel singular values. Given an asymptotically stable linear system \((A,B,C)\), the controllability Gramian \(W_c\) and the observability Gramian \(W_0\) are defined, respectively, by \[ W_c= \int^\infty_0 e^{tA} BB' e^{tA'}dt\quad\text{and}\quad W_0= \int^\infty_0 e^{tA'} C'Ce^{tA} \,dt. \] A realization \((A,B,C)\) is called signature symmetric if \[ (AI_{pq})'= AI_{pq},\quad (CI_{pq})'= B, \] and balanced if \(W_c= W_0= \Sigma\). Local quadratic convergence of the Jacobi-type algorithm for balancing is shown. One numerical example is presented verifying the theoretical results.