Stabilized least squares estimators for time-variant processes

To use recursive-least-square (RLS) parameter estimators in an adaptive setting, the estimation matrix must not be asymptotically singular. The two classical modifications of the standard RLS algorithms, the linear and the exponential forgetting, lead to a blowup of the estimator for a nonpersistent excitation. For a well-behaved LS algorithm the estimation matrix and its inverse must be bounded. A family of simple stabilized LS algorithms is proposed. The increase in computational complexity with respect to the standard RLS is negligible