Search

In recent years, several algorithms for the iterative calculation of a polynomial matrix eigenvalue decomposition (PEVD) have been introduced. The PEVD is a generalisation of the ordinary EVD and uses paraunitary operations to diagonalise a parahermitian matrix. This paper addresses potential computational savings that can be applied to existing cyclic-by-row approaches for the PEVD. These savings are found during the search and rotation stages, and do not significantly impact on algorithm accuracy. We demonstrate that with the proposed techniques, computations can be significantly reduced. The benefits of this are important for a number of broadband multichannel problems.

In this work we present a new method of controlling the order growth of polynomial matrices in the multiple shift second order sequential best rotation (MS-SBR2) algorithm which has been recently proposed by the authors for calculating the polynomial matrix eigenvalue decomposition (PEVD) for para-Hermitian matrices. In effect, the proposed method introduces a new elementary delay strategy which keeps all the row (column) shifts in the same direction throughout each iteration, which therefore gives us the flexibility to control the polynomial order growth by selecting shifts that ensure non-zero coefficients are kept closer to the zero-lag plane. Simulation results confirm that further order reductions of polynomial matrices can be achieved by using this direction-fixed delay strategy for the MS-SBR2 algorithm.

In this paper, we show that the paraunitary (PU) matrices that arise from the polynomial eigenvalue decomposition (PEVD) of a parahermitian matrix are not unique. In particular, arbitrary shifts (delays) of polynomials in one row of a PU matrix yield another PU matrix that admits the same PEVD. To keep the order of such a PU matrix as low as possible, we propose a row-shift correction. Using the example of an iterative PEVD algorithm with previously proposed truncation of the PU matrix, we demonstrate that a considerable shortening of the PU order can be accomplished when using row-corrected truncation.

In this paper, we present an improved version of the second order sequential best rotation algorithm (SBR2) for polynomial matrix eigenvalue decomposition of para-Hermitian matrices. The improved algorithmis entitledmultiple shift SBR2 (MS-SBR2) which is developed based on the original SBR2 algorithm. It can achieve faster convergence than the original SBR2 algorithm by means of transferring more off-diagonal energy onto the diagonal at each iteration. Its convergence is proved and also demonstrated by means of a numerical example. Furthermore, simulation results are included to compare its convergence characteristics and computational complexity with the original SBR2, sequential matrix diagonalization (SMD) and multiple shift maximum element SMD algorithms.

A number of algorithms for the iterative calculation of a polynomial matrix eigenvalue decomposition (PEVD) have been introduced. The PEVD is a generalisation of the ordinary EVD and will diagonalise a parahermitian matrix via paraunitary operations. This paper addresses savings - both computationally and in terms of memory use - that exploit the parahermitian structure of the matrix being decomposed, and also suggests an implicit trimming approach to efficiently curb the polynomial order growth usually observed during iterations of the PEVD algorithms. We demonstrate that with the proposed techniques, both storage and computations can be significantly reduced, impacting on a number of broadband multichannel problems.

A polynomial eigenvalue decomposition of paraher- mitian matrices can be calculated approximately using iterative approaches such as the sequential matrix diagonalisation (SMD) algorithm. In this paper, we present an improved SMD algorithm which, compared to existing SMD approaches, eliminates more off-diagonal energy per step. This leads to faster convergence while incurring only a marginal increase in complexity. We motivate the approach, prove its convergence, and demonstrate some results that underline the algorithm’s performance.