Binomial approximation to the Poisson binomial distribution (Q757966): Difference between revisions

Let \(X_ 1,...,X_ n\) be independent Bernoulli random variables with \(P(X_ i=1)=p_ i=1-q_ i.\) Let \(S=X_ 1+...+X_ n,\) \(P(k)=P(S=k)\) and \(Q(k)=\left( \begin{matrix} n\\ k\end{matrix} \right)p^ kq^{n-k}=P(Z=k)\) with \(p=n^{-1}(p_ 1+...+p_ n),\) \(q=1-p.\) The authors derive a lower and an upper bound for the variation distance \(\sum | P(k)- Q(k)|,\) both of which are, roughly, proportional to the nonnegative quantity 1-(Var S/Var Z). Proofs follow the method of \textit{C. Stein} [Approximate computation of expectations (Hayward, CA, 1986)]. Logconcavity of P(k)/Q(k) leads to similar bounds for the sup-distance between distribution functions. The hypergeometric distribution has probabilities P(k) for suitable \(p_ 1,...,p_ n\) [see \textit{V. A. Vatutin} and \textit{V. G. Mikhailov}, Theory Probab. Appl. 27, 734-743 (1982); translation from Teor. Veroyatn. Primen. 27, No.4, 684-692 (1982; Zbl 0517.60008)]. From this result bounds for the variation distance between the multivariate hypergeometric and the multinomial distribution are derived.