scientific article; zbMATH DE number 3919594

From MaRDI portal
Publication:3694487





A special modification of the Robbins-Monro stochastic approximation procedure is investigated in the situation when the regression function varies with time and a process converging to the root of the regression function from below (in some sense) is sought. The modified Robbins-Monro procedure defined by (*) \textit{D. Anbar} [Ann. Stat. 5, 229-234 (1977; Zbl 0359.62066)] \(X_{n+1}=X_ n-a_ n(y_ n+b_ n),\) where \(\{a_ n\}\) is a sequence of positive numbers, \(\{b_ n\}\) is a sequence of real numbers, \(E\{y_ n| X_ 1,...,X_ n\}=M_ n(X_ n)\quad a.s.,\) \(M_ n(X)\) is the regression function at time n and there exists a unique root \(\theta_ n\) of the equation \(M_ n(X)=0\), is proposed. The author proves that under certain assumptions \(X_ n-\theta_ n\to 0\) a.s. and, with probability one, \(X_ n-\theta_ n>0\) only finitely many times. The special cases treated in Anbar (*) and \textit{V. Dupač} [Ann. Math. Stat. 36, 1695-1702 (1965; Zbl 0137.365)] are mentioned.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3694487)