A spherical subspace based adaptive filter

Abstract

The authors use the adaptation mechanism of the spherical subspace tracker together with the weighting scheme of total least squares (TLS) to construct an adaptive filter that tracks solutions to time-varying ordinary least squares. TLS, data least squares, and reduced rank problems. To study convergence properties, they relate this filter to Thompson's constrained stochastic gradient eigenfilter. They present a convergence rate acceleration scheme that keeps the filter from being slowed down by saddle points in the performance surface. Simulation results verify the theoretical development. The filter behaves well in the full rank case and is more sensitive and slow to converge in certain reduced rank problems.<>