Convergence and convergence rate of stochastic gradient search in the case of multiple and non-isolated extrema (Q2018557)

Given a continuously differentiable function \(f=f (\theta)\), \(\theta \in \mathbb{R}^n\), the problem is to determine a solution \(\hat{\theta}\) of the equation \(\nabla f (\theta)=0\). Having an estimate of the gradient \(\nabla f (\theta)\), for the iterative computation of a zero of the gradient, a standard stochastic gradient procedure is considered. Under certain assumptions on the additive noise term and the function \(f\), results on the almost sure convergence and convergence rate are provided. The results are illustrated by means of examples from principle component analysis and maximum likelihood estimation.