On the resultant property of the Fisher information matrix of a vector ARMA process (Q2484496)

A new multiple resultant matrix is introduced (by generalizing the resultant matrix of two scalar polynomials), as well as some of its useful representations. A matrix is called a multiple resultant matrix associated to two matrix polynomials when it is singular if and only if the two matrix polynomials have at least one common eigenvalue. In this paper it is shown that the Fisher information matrix (FIM) of a stationary vector autoregressive and moving average (ARMA) time series process is a multiple resultant matrix. By using a set of matrix differential rules introduced by the authors in some previous papers, two compact representations for the FIM are derived in terms of two structured matrices, the multiple Sylvester matrix and the tensor Sylvester matrix. It is shown that the FIM is singular if and only if the corresponding tensor Sylvester matrix is singular, a property which is equivalent to the existence of at least one common eigenvalue of the two matrix polynomials. Two numerical examples emphasize the fact that while the tensor Sylvester matrix has this property, the multiple Sylvester matrix can be non-singular for matrix polynomials with common eigenvalues and it can be singular although the matrix polynomials do not have common eigenvalues. The usefulness of the tensor Sylvester matrix approach is emphasized by a numerical example which shows that the FIM fails to detect common eigenvalues, due to a numerical instability, contrary to the tensor Sylvester matrix.