An invariance principle for a finite dimensional stochastic approximation method in a Hilbert space (Q799302)

For the application of the classical Robbins-Monro stochastic approximation procedure in an infinite dimensional Hilbert space H a modified procedure which works in finite dimensional subspaces of growing dimension is proposed, viz. \(X_{n+1}=X_ n-(1/n)\phi_ n(f(X_ n)- V_ n)\) where f:\(H\to H\) is a regression function, \(V_ n\), \(n\in {\mathbb{N}}\), are H-valued random elements, \(\phi_ n\), \(n\in {\mathbb{N}}\), are projections from H onto \(L_ n\), \(L_ 1\subset L_ 2..\). are finite dimensional subspaces of H. The invariance principle for this procedure is given. Two applications are considered - estimation of the expectation of a stochastic process and estimation of a conditional expectation.