ROCKET: a reduced order correlation kernel estimation technique

Abstract

The ROCKET (reduced order correlation kernel estimation technique) algorithm is a new reduced rank autoregressive (AR) spectrum estimation technique which is substantially more robust to signal rank underestimation and significantly more computationally efficient then conventional reduced rank techniques based on principal component analysis. Perhaps more importantly, ROCKET's reduce rank performance has the potential to surpass the performance of full rank AR spectrum estimation techniques. ROCKET is based on the observation that the reduced rank subspace of importance is the one that best predicts the desired signal from the data. ROCKET's subspace is formed in an iterative manner from the cross-correlation vectors defined by a specified desired signal and data. Projecting the desired signal onto this new subspace allows for a significantly reduced dimensional weight vector with the aforementioned properties and benefits.