Consistent Low - Complexity Estimation of Activeparameters in Large Linear Regressions

Some important practical signals and systems can be modeled by very large linear regression models where it is reasonable that most of the parameters are zero. We give an eecient method to solve this combined estimation and structure determination problem. It is related to Akaike-like criteria, and is based on one LMS lter and thus it is of low complexity. Asymptotic analysis shows that the method is consistent for nite impulse response models. A recursive algorithm is derived, which can be applied to time-varying systems as well. An example shows the eeciency of the approach.