Biased Regression Algorithms in the Quaternion Domain

Abstract

The ill-conditioned matrix problem in quaternion linear regression models is addressed in this paper and several dimension-reduction based regression methods for circumventing this problem are suggested. The algorithms are formulated in a general way and can be easily adapted to different scenarios: widely linear, semi-widely linear and strictly linear processing, in accordance with the properness properties presented by quaternion random vectors. A comparison with existing solutions is carried out by using both laboratory data and a color face database.