Convergence and robustness of the Robbins-Monro algorithm truncated at randomly varying bounds (Q1095544)

In this paper the Robbins-Monro algorithm with step-size \(a_ n=1/n\) and truncated at randomly varying bounds is considered under mild conditions imposed on the regression function. It is proved that for its a.s. convergence to the zero of a regression function the necessary and sufficient condition is \[ (1/n)\sum^{n}_{i=1}\xi_ i\to_{n\to \infty}0\quad a.s. \] where \(\xi_ i\) denotes the measurement error. It is also shown that the algorithm is robust with respect to the measurement error in the sense that the estimation error for the sought- for zero is bounded by a function g(\(\epsilon)\) such that \[ g(\epsilon)\to_{\epsilon \to 0}0\text{ if } \limsup_{n\to \infty}(1/n)\| \sum^{n}_{i=1}\xi_ i\| \equiv \epsilon >0. \]