Eigensubspace Estimation Using Discrete Recurrent Neural Networks

This paper proposes two models of discrete recurrent neural networks to study the problem of eigensubspace estimation for positive definite symmetric matrix. The first model is a class of nonlinear neural networks. It is used for estimating the largest eigenvalue and one of its corresponding eigenvectors. The second model is a class of linear neural networks which estimates the eigensubspace corresponding to the largest eigenvalue. Dynamic properties of these two classes of discrete recurrent neural network models are studied and used for eigensubspace estimation.