Accelerating the SVD block-Jacobi method (Q2487206)

A block version of the one-sided Jacobi method for the singular value decomposition of a rectangular matrix is formulated and a rough operation count is performed. Subsequently, the algorithm is re-analysed in detail to save on the number of flops and increase its efficiency. This is obtained by three improvements. First a preprocessing step is proposed which consists of a QR factorization, followed by an LQ factorization. This should reduce the number of iteration steps till convergence. Secondly, a (block) cosine-sine (CS) decomposition is used inside every iteration step, which reduces the number of flops and hence the computation time. Finally, some implementation details and tricks are given which boost the overall efficiency of the algorithm. For example an efficient stopping criterion is proposed based on the magnitude of the off-diagonal elements. An operation count of the modified algorithm shows the gain in flops and some numerical experiments illustrate the algorithm.