A conservative method for selecting the best simulated system (Q2564628)

This paper discusses a lower bound for the simultaneous confidence coefficient for multinomial proportions. Let \({\mathbf p}= (p_1,p_2, \dots, p_k)\), \({\mathbf N}= (N_1,N_2, \dots, N_k)\) and \(\widehat {\mathbf p}= (\widehat p_1, \widehat p_2, \dots, \widehat p_k)\) with \(\widehat p_i= N_i/n\) be the vector of the actual cell probabilities in a multinomial experiment, the vector of the observed cell counts in a sample size \(n\) and the point estimator for \({\mathbf p}\). The authors develop a lower bound for the simultaneous confidence coefficient \(\Pi(k,{\mathbf p}; n,s/ \sqrt n)= P(\cap^k_{i=1} \{|\widehat p_i- p_i|<s/ \sqrt n\})\) using an inequality of Bennett, and then describe an algorithm for determining an \(s\) such that \(\Pi(k,{\mathbf p}; n,s/ \sqrt n) \geq 1- \alpha\) for fixed \(\alpha\). This lower bound is called conservative in the sense that it does not depend on the asymptotic distribution of \(\widehat p_i\).