An adaptive quasi-Newton algorithm for eigensubspace estimation. (Q2734429)

The aim of the paper is to study the landscape of the cost function and to derive a new quasi-Newton (QN) adaptive algorithm for the principal component analysis. The proposed QN adaptive eigensubspace algorithm estimates first the principal eigenvector and then estimates the minor eigenvectors sequentially. The new QN algorithm is compared with recursive least-squares-type algorithms and shown to have faster and better tracking abilities. Compared with another quasi-Newton algorithm [G. Mathew, V. U. Reddy and S. Dasgupta (1995)], the new QN algorithm does not need any a priori information of the data covariance matrix for the choice of the penalty coefficient, and provides superior tracking performance. Extensive experiments of the investigated algorithms with stationary and non-stationary data are exposed, and interesting considerations on their computational complexity are given.