Householder transformation for the regularized least square problem on iPSC/860

Abstract

Discusses a householder factorization algorithm for a special type of matrix arising from the application of the Tikhnov regularization method to an ill-conditioned least square problem. The matrix involved is half dense and half sparse. The algorithm has been implemented on iPSC/860 hypercubes. By overlapping communications with computations, the code has been optimized to take advantage of the special structure of the matrix and minimize inter-node communications. Super-linear speed-up was observed in the numerical experiment for large problems. The algorithm has been used as a core routine in the program solving parameter identification problems in reservoir simulations.<>