Efficient algorithm for matrix spectral factorization (Q1071712)

Spectral factorization represents an essential mathematical tool for solving optimal filtering problems or performing synthesis of linear electrical n-ports in the frequency domain among other important areas of applications. Since continuous-time and discrete-time systems have to be distinguished from each other problems are as follows: Find matrices A so that known positive real polynomial and symmetric matrices B are factored according to \(B(p)=A(p)A^ T(-p)\) in the continuous-time case or \(B(z)=A(z)A^ T(\frac{1}{z})\) in the discrete-time case. The spectral factors A exist and are unique up to a constant orthogonal matrix. In the past various approaches were proposed, among them a suitable partition of all roots of det B or iterative solutions which make appeal of matrix Riccati equations, perform triangular factorization of Toeplitz matrices or are based on Newton's method. All those methods suffer from certain deficiencies. The paper suggests a new algorithm for the spectral factorization of polynomial matrices. Though Newton's method is applied again, the algorithm is more efficient. This improvement in performance is obtained by using fully the structure of the equations to be solved in each iteration. The theory of symmetric polynomial equations represents the mathematical basis of the investigations. The resulting algorithm was implemented in PL/I language and tests are performed using 8 bytes floating-point format. A precision of 8 decimal digits was usually met after 6-8 iterations. Further, it turns out, the farther from the stability boundary the roots of det B lie, the better the computations proceed. The algorithm works even with roots just on the boundary: Before approaching the singularity and then collapsing, the result is calculated to several decimal digits. In such cases the rate of convergence is degraded from a quadratic to a geometric one.