Proceedings of the Edinburgh Mathematical Society (1990) 33, 337-366 ©
A SCHUR-COHN THEOREM FOR MATRIX POLYNOMIALS
by HARRY DYM and NICHOLAS YOUNG
(Received 14th July 1988)
Let N(X) be a square matrix polynomial, and suppose det N is a polynomial of degree d. Subject to a certain
non-singularity condition we construct a d by d Hermitian matrix whose signature determines the numbers of
zeros of N inside and outside the unit circle. The result generalises a well known theorem of Schur and Cohn
for scalar polynomials. The Hermitian "test matrix" is obtained as the inverse of the Gram matrix of a natural
basis in a certain Krein space of rational vector functions associated with N. More complete results in a
somewhat different formulation have been obtained by Lerer and Tismenetsky by other methods.
1980 Mathematics subject classification (1985 Revision): 30C15, 46D05, 93D20.
1. Introduction
A celebrated paper of I. Schur [11] contains, among much else, a criterion for a
polynomial p(X) to have all its zeros in the open unit disc D: one forms a certain
Hermitian matrix H from the coefficients of p, and the assertion is that the zeros of p lie
in D if and only if H is positive definite. The theorem was generalized by A. Cohn [4]: if
the zeros of p are all non-conjugate with respect to the unit circle T then the numbers
of zeros inside and outside the unit circle are equal to the numbers of positive and
negative eigenvalues of H respectively. The corresponding problem for polynomial
matrices holds considerable interest, both for its own sake and for application to
multivariable systems theory: the stability of a linear system with many inputs and
outputs depends on the zeros of an associated matrix polynomial lying in D. Now the
zeros of a square matrix polynomial N(X) (also known as the eigenvalues of N) are
defined to be the zeros of the scalar polynomial det N(X), and so one way of testing N is
to apply Schur's criterion to det N. However, since the calculation of the determinant of
a matrix polynomial is lengthy and numerically somewhat ill behaved, it is natural to
look for a more direct way of associating with N(X) a Hermitian matrix whose signature
indicates the location of the zeros of N with respect to T.
One way to approach this question is through Bezoutian matrices. In the scalar case
the Schur matrix can be regarded as the Bezoutian of two polynomials, and there are
several extensions of the notion of Bezoutian to matrix polynomials. The most
successful one appears to be that due to Anderson and Jury [2], who asked whether
their form of Bezoutian could be used to prove an analogue of the Schur-Cohn
337