An improved dqds algorithm
For numerical linear algebra practitioners, this is an incremental improvement to a well-known algorithm, offering faster convergence for badly scaled matrices.
The paper presents an improved dqds algorithm (V5) for computing all singular values of a bidiagonal matrix with high relative accuracy, achieving 1.2x–4x speedup over DLASQ and up to 3x–10x on matrices with slow convergence, without accuracy loss.
In this paper we present an improved dqds algorithm for computing all the singular values of a bidiagonal matrix to high relative accuracy. There are two key contributions: a novel deflation strategy that improves the convergence for badly scaled matrices, and some modifications to certain shift strategies that accelerate the convergence for most bidiagonal matrices. These techniques together ensure linear worst case complexity of the improved algorithm (denoted by V5). Our extensive numerical experiments indicate that V5 is typically 1.2x--4x faster than DLASQ (the LAPACK-3.4.0 implementation of dqds) without any degradation in accuracy. On matrices for which DLASQ shows very slow convergence, V5 can be 3x--10x faster. At the end of this paper, a hybrid algorithm (HDLASQ) is developed by combining our improvements with the aggressive early deflation strategy (AggDef2 in [SIAM J. Matrix Anal. Appl., 33(2012), 22-51]). Numerical results show that HDLASQ is the fastest among these different versions.