A Robust RLS Algorithm for Adaptive Canonical Correlation Analysis
Javier Vía, Ignacio Santamarı́a, Jesus Perez · 2006
Canonical correlation analysis (CCA) is a classical tool in statistical analysis that measures the linear relationship between two data sets. In this paper we show that CCA can be reformulated as a pair of coupled least squares (LS) problems. By exploiting this idea, we first present an iterative batch procedure to extract all the canonical vectors through a regression procedure. Then, we derive a recursive least squares (RLS) algorithm for on-line CCA. This algorithm can be further improved to increase its robustness against outliers and impulsive noise. The proposed algorithm is applied to blind identification of multichannel FIR systems, and its performance is illustrated through simulations.