Estimation of the signal subspace without estimation of the inverse covariance matrix

Abstract

Let a high-dimensional random vector $vecX$ can be represented as a sum of two components - a signal $vecS$, which belongs to some low-dimensional subspace $mathcalS$, and a noise component $vecN$. This paper presents a new approach for estimating the subspace $mathcalS$ based on the ideas of the Non-Gaussian Component Analysis. Our approach avoids the technical difficulties that usually exist in similar methods - it doesn't require neither the estimation of the inverse covariance matrix of $vecX$ nor the estimation of the covariance matrix of $vecN$.