Oja's algorithm for graph clustering, Markov spectral decomposition, and risk sensitive control

Given a positive definite matrix M and an integer Nm≥1, Oja's subspace algorithm will provide convergent estimates of the first Nm eigenvalues of M along with the corresponding eigenvectors. It is a common approach to principal component analysis. This paper introduces a normalized stochastic-approximation implementation of Oja's subspace algorithm, as well as new applications to the spectral decomposition of a reversible Markov chain. Recall that this means that the stationary distribution satisfies the detailed balance equations (Meyn & Tweedie, 2009). Equivalently, the statistics of the process in steady state do not change when time is reversed. Stability and convergence of Oja's algorithm are established under conditions far milder than that assumed in previous work. Applications to graph clustering, Markov spectral decomposition, and multiplicative ergodic theory are surveyed, along with numerical results.