Efficient implementation of SVD-updating algorithms for subspace tracking

An efficient parallel implementation of the SVD (singular value decomposition) updating algorithm is addressed. The modifications applied to the algorithm are: approximation of Kogbetliantz's SVD algorithm, approximation of the SVD-rotation evaluations, and factorization of the rotations. This results in a square root and division free SVD-updating algorithm, which combines the speedups and hardware reductions obtained by the different modifications. The proposed SVD-based subspace tracking algorithm only requires the implementation of (+, *) operations in all processor cells.<
>