Asymptotic confidence regions of stochastic approximation procedures in Hilbert spaces (Q2641050)

The author presents two main theorems about random variables on a probability space (\(\Omega\),\({\mathcal A},P)\) with values in a separable Hilbert space \({\mathcal H}\). The first of them states the following: Let denote \(K_ n\), \(n\in {\mathbb{N}}\), a sequence of random covariance operators converging to a non-random covariance operator K in an appropriate way and let \(G_{K_ n(\omega)}(\omega ')\) and \(G_ K(\omega ')\) be centered Gaussian random variables in \({\mathcal H}\) with covariance operators \(K_ n(\omega)\) and K, resp. Then for each p with \(0<p<1\) and each \(\omega\in \Omega\) there exist \(r_ n(\omega)\) such that \[ P\{\omega '\in \Omega:\;\| G_{K_ n(\omega)}(\omega ')\| \leq r_ n(\omega)\}=1-p,\quad n\in {\mathbb{N}}, \] where the \(r_ n\) converge P-a.e. to a suitable non-random \(r>0\) as \(n\to \infty\) with \[ P\{\omega '\in \Omega:\;\| G_ K(\omega ')\| \leq r\}=1-p. \] The second theorem states that for a sequence of arbitrary random variables \(Z_ n\) in \({\mathcal H}\) obeying a Donsker type invariance principle and a sequence of stopping rules \(N_ n\) fulfilling a certain growth condition, the probability \(P\{\| Z_{N_ n}\| \leq r_{N_ n}\}\) tends to 1-p as \(n\to \infty.\) In the major part of the paper these results are applied in order to deal with a Robbins-Monro method in a Hilbert space. In connection with a sequence of empirical covariance operators an estimator of the unknown radius of a ball is described, for which the Gaussian limit distribution takes a given value. Moreover, a stopping rule is proposed leading to asymptotic confidence balls with a fixed radius.