Systems & Control Letters 57 (2008) 1058–1066 Contents lists available at ScienceDirect Systems & Control Letters journal homepage: www.elsevier.com/locate/sysconle Numerical J -spectral factorization of general para-hermitian matrices Jovan Stefanovski Control & Informatics Div., JP ‘‘Streževo’’, Bitola, Former Yugolav Republic of Macedonia, The article info Article history: Received 17 December 2007 Received in revised form 9 April 2008 Accepted 13 July 2008 Available online 15 August 2008 Keywords: Para-hermitian matrices J -spectral factorization Zeros at infinity abstract A numerical algorithm for J -spectral factorization of para-hermitian rational matrices is presented, based on a generalization of the notion of separated realization of a para-hermitian matrix. The transformations used preserve the para-hermitian matrix pencil structure. © 2008 Elsevier B.V. All rights reserved. 1. Introduction Consider a possibly singular, with possible zeros on the imaginary axis (including the infinity), and possibly indefinite on the imaginary axis m × m-dimensional real rational matrix Π (λ), which is para-hermitian, i.e. Π T (λ) = Π (λ). A problem of J - spectral factorization is to find a real square rational matrix Φ(λ) nonsingular for all R[λ] > 0C, and indices m + , m 0 and m satisfying m + + m 0 + m = m, such that Π (λ) = Φ T (λ)J Φ(λ), J = diag{I m + , 0 m 0 ×m 0 , I m }. (1.1) The J -spectral factorization formulated in its most general form has many applications (see [7,5] and references therein). Another variants of the problem is to find a possibly nonsquare Φ (1.1) and J = diag{I m + , I m }, such that Φ has full row rank for all R[λ] > 0, and to find spectral factor Φ such that, besides of full rank in R[λ] > 0, it is analytic in that region. Notation and some standard results. By the superscript T we denote matrix transposition. By the superscript we denote conjugate transpose. The identity matrix is denoted by I , or I n if the matrix dimension requires. The sets of complex and real numbers are denoted by C and R. By null(K ) we denote a right annihilator of matrix K . The abbreviation FGE means finite generalized eigenvalue, and the abbreviation IGE means infinite generalized eigenvalue. Tel.: +389 47207838; fax: +389 47207836. E-mail address: jovanstef@t-home.mk. Kronecker canonical form of matrix pencil λM N is: P M N )Q = diag [λI n r A c , B c ]M I , λI n f J f , λI n l A o C o  (1.2) where (A c , B c ) is a controllable pair, (A o , C o ) is observable, M is a nilpotent matrix of the form M = diag{M i , i = 1,..., s}, where M i = 0 I ν i 1 0 0 are ν i × ν i -dimensional matrices. The indices ν i , i = 1,..., s are multiplicities of the IGEs. J f is a matrix that contains the FGEs of λM N and P and Q are non-singular matrices [20], or Kronecker canonical form is P M N )Q = 0 0 0 λM r N r , (1.3) where λM r N r is a regular matrix pencil. Matrix pencil λM N is para-hermitian if M T = M and N T = N . If λM N is para-hermitian then n l = n r in (1.2), and the FGEs of λM N that are not purely imaginary appear in symmetric pairs, in respect to imaginary axis. The zeros of a rational matrix are defined through its McMillan form. A rational matrix is minimum phase if it has no zeros in R[λ] > 0. A rational matrix is stable if it is analytic in R[λ] > 0. A square matrix is stable if it has no eigenvalues in R[λ] > 0. Analogously we define stable FGEs. The following notions and results are taken from [13]. The descriptor realization (E , A, B, C , D) of transfer matrix G(λ) = D + C E A) 1 B is weakly minimal if and only if matrices [λE A, B] and [E , B] have full row ranks (finite and infinite 0167-6911/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.sysconle.2008.07.002