A modified Bernstein-technique for estimating noise-perturbed function values (Q1080261)

Bounds for the number of observations to be made in order to obtain fixed confidence intervals for the estimation of function values are deduced. Bienaymé-Chebyshev inequality fixes a property for random variables \(\xi\) and a bound for the sample size n may be obtained using it. Markov inequality fixes properties of exp(v(\(\xi\)-E(\(\xi)\))) for \(v>0\) and the improvement due to Bernstein permits to sharpen the Bienaymé-Chebyshev bound of n. This paper is devoted to the ''sharpenings'' of n for noise- perturbed values. The estimation of a probability is studied as a particular case. Taking \(h: R^ n\times R^ q\to R\) as a Borel-measurable function and observations of \(R^ q\), a sufficiently small n is desired when the expectation of h is to be estimated. Lemma 2.1 fixes a bound for the probability of observing deviations larger than a certain \(\epsilon\) when \(\xi\) has a symmetrical distribution. In Lemma 2.2 a bound for the distribution is obtained under the hypothesis \(v\in [0,V]\). If the assumptions of these lemmas hold a bound for n may be deduced - Theorem 2.1. Bounds are obtained, by applying the technique developed, in different cases. Corollary 2.1 is dedicated to the obtaining of a bound for n when a probability P(A) is estimated. The results of this paper are of interest in stochastic programming and other disciplines using simulation techniques in the analysis of certain particular problems.