Stabilization of fast recursive least-squares transversal filters for adaptive filtering

This paper investigates the numerical stability of the well-known Fast Transversal Filters algorithms. Two different modes of divergence are indicated with an effort to mathematically explain the reason of the appearance of each mode. An original stabilization method is thus introduced. It basically consists in introducing redundancy in the algorithm by computing some quantity in two different ways. The difference of the two values (nonzero because of finite precision) is fed back to the algorithm in order to correct it. The stabilized algorithm requires some more operations but it is always linear in the order of the filter per iteration.