Widening the scope of an eigenvector stochastic approximation process and application to streaming PCA and related methods (Q2222226)

The authors propose a general framework for stochastic approximation and apply it to the case of streaming principal component analysis (PCA). They prove the almost sure convergence of Oja-type processes to eigenvectors of the expectation \(B\) of a random matrix while relaxing the i.i.d. assumption on the observed random matrices \(B_n\) and assuming either \(B_n\) converges to \(B\) or \(E[B_n|T_n]\) converges to \(B\) where \(T_n\) is the sigma-field generated by the events before time \(n.\) As an application of this generalization, the online PCA of a random vector \(Z\) can be performed when there is a data stream of i.i.d. observations of \(Z,\) even when both the metric \(M\) used and the expectation of \(Z\) are unknown and estimated online. All previous observations up to the current step can be used without having to store them. The general framework of this process also covers the case of factorial methods related to PCA.