Bidiagonal SVD Computation via an Associated Tridiagonal Eigenproblem

In this paper, we present an algorithm for the singular value decomposition (SVD) of a bidiagonal matrix by means of the eigenpairs of an associated symmetric tridiagonal matrix. The algorithm is particularly suited for the computation of a subset of singular values and corresponding vectors. We focus on a sequential version of the algorithm, and discuss special cases and implementation details. We use a large set of bidiagonal matrices to assess the accuracy of the implementation in single and double precision, as well as to identify potential shortcomings. We show that the algorithm can be up to three orders of magnitude faster than existing algorithms, which are limited to the computation of a full SVD. We also show time comparisons of an implementation that uses the strategy discussed in the paper as a building block for the computation of the SVD of general matrices.