Subspace tracking with a correlation-based decomposition

In signal processing applications where dominant/subdominant subspace is computed, one may justify using the eigenvalue decomposition (EVD) over the singular value decomposition (SVD) as it is computationally cheaper to compute and its round-off errors are often overshadowed by the effects of noise. Stewart (1992) has further proposed the URV algorithm as a computationally cheaper alternative to the SVD. By forming the cross-product of the URV decomposition with its transpose, a correlation domain decomposition can be produced. We show how to update this cross-product RV decomposition (CRV) and justify its effectiveness as a subspace tracking technique.