Stability of Oja's PCA Subspace Rule

Principal eigenvectors of the data covariance matrix or the subspace spanned by them, called PCA subspace, provide optimal solutions to several information representation tasks. Recently, many neural approaches have been proposed for learning them (see, e.g., Hertz r t a/. 1991; Oja 1992). A well-known algorithm for learning the PCA subspace of the input vectors is so-called Op’s subspace rule (Oja 1989; Hertz et a / . 1991):