Adaptive rank estimation for spherical subspace trackers

We develop a rank tracking method for spherical subspace trackers. The method adaptively sets a threshold based on the averaged noise eigenvalue. The signal eigenvalues are then compared with the threshold to reach a decision on the subspace rank. The threshold itself is chosen to balance the error probabilities due to rank underfitting and overfitting. Simulation results show that our method performs as well as (and, for very low SNRs, even better than) information theoretic criteria.