A new efficient algorithm for singular value decomposition

Abstract

The popular CORDIC-based SVDs suffer from the disadvantages of inherently sequential operations and requiring area-consuming barrel shifters. On the other hand, the basic-arithmetic SVD needs nasty division and square-root operations. To alleviate these problems, a new algorithm for SVD computation is proposed. Instead of using CORDIC algorithm, the new algorithm is based on multiplication-and-addition operations (without nasty division and square-root operations), combined with an approximate rotation and table lookup scheme, As a result, the new algorithm has a smaller computational complexity than those of the conventional SVD algorithms, in terms of number of addition operations. In realization, it also has a higher parallelism and smaller area than the conventional CORDIC-based SVD processors.