RLS parameter convergence with overparameterized models (Q1119237)

This paper makes use of the Lyapunov type function approach to generalize in several directions the existing results on the convergence properties of the recursive least-squares (RLS) algorithm applied to (noise-free) linear time-invariant difference equations. More specifically, the paper treats the case of overparameterized models, without assuming a priori knowledge of the degree of overparameterization; in addition, the weakest possible condition of excitation is imposed. The RLS estimate is shown to converge in a parameter set H the elements of which correspond to reducible parameterizations of the (irreducible) true transfer function. The initial conditions of the RLS algorithm are shown to be qualitatively related to the convergence points of the RLS. In particular, if the condition number of the initial covariance matrix in the RLS algorithm is chosen to be equal to one, then the algorithm will converge to the unique point in H which is closest to the initial parameter estimates (such a point always exists as H is a closed convex set).