Jacobi-type SVD and its Floating-point Realization Based on Fast Rotations

摘要

In this paper a Jacobi-type algorithm for computing SVD of a square matrix is presented. Fast rotations are used as approximate rotations. These fast rotations are closely related to CORDIC arithmetic and perform orthogonal rotation over a certain angle at a very low cost in realization. Both the computation of the approximate rotation angles and the rotation mode are performed by execuition of fast rotations. A floating-point CORDIC-like architecture for realization of fast rotation operations with full arithmetic accuracy is also presented.