Recursive least squares with stabilized inverse factorization

Recently developed recursive least squares schemes, where the square root of both the covariance and the information matrix are stored and updated, are known to be particularly suited for parallel implementation. However, when finite precision arithmetic is used, round-off errors apparently accumulate unboundedly, so that after a number of updates the computed least squares solutions turn out to be useless. In this paper, a Jacobi-type correction scheme is described, that continuously annihilates accumulated errors and thus stabilizes the overall scheme. Furthermore, it is shown how the resulting RLS-algorithm can be implemented on a systolic array.