Neural networks for computing eigenvalues and eigenvectors (Q1202354)

The authors consider the problem of computing an eigendecomposition of a square matrix. They formulate the problem as a constrained optimization problem and construct a penalty function to be minimized. They solve the resulting unconstrained optimization problem by designing neural networks and applying a back-propagation learning scheme, which is similar to the steepest descent algorithm in numerical optimization parlance. The result of numerical simulations on some small test problems are presented.