Some bounds on the deviation probability for sums of nonnegative random variables using upper polynomials, moment and probability generating functions (Q987944)

The upper bounds for deviation probability of the type \[ P\equiv Pr\{E(S)-S\geq t\}\leq e^{-t^2/A}\eqno(1) \] are considered; \(X_i\) are i.r.v. with \(P(X_i\geq 0)=1,\;E(X_i^2)<\infty,\) \(i=1,2,\cdots\) \(,m,\) and \(S=\sum_1^mX_i,\;t\geq 0.\) In section 1, the author proves inequality \[ P\leq \exp\,\left\{\frac{-(t-(1-a)E(S))^2}{4b\sum_1^mE(X_i)^2}\right\} \] with \(\frac{E(S)-t}{E(S)}\leq a\leq 1,\;b=(1-\sqrt{1-a^2})/2,\) which always improuves on the bounds derived earlier by \textit{I. S. Pinelis} and \textit{S. A. Utev} [Theory Probab. Appl. 34, No. 2, 340--346 (1989; Zbl 0693.60036); translation from Teor. Veroyatn. Primen. 34, No. 2, 384--390 (1989); correction Ann. Probab. 27, No. 4, 2119 (1999)] and rediscovered by \textit{A. Maurer} [JIPAM, J. Inequal. Pure Appl. Math. 4, No. 1, Paper No. 15, 6 p., electronic only (2003; Zbl 1021.60036)] (bound in (1) with \(A=2\sum_1^mE(X_i)^2\)) and in some cases is better than the Bentkus's bound ((1) with \(A=2\sum_1^m W_i,\) \(W_i=\max((E(X_i))^2,V(X_i)),\) \(V(X_i)=E(X_i^2)-(E(X_i))^2\)). It was also shown that in the latter case neither bound is uniformly better than the other bound. The other two sections are dedicated to numerical study and infinitely divisible cases. In the first one the author computes the actual deviation probability, his bound and abovementioned upper bounds when \(X_i\) have Weibull distributions for comparison. The approximated value of the left-hand side \(P\) in (1) was found using one million runs of Monte-Carlo simulation. In the second one he points out some connections between all bounds and also discusses how some of them can be improved if \(X_i\) have moment or probability generating functions that are of the infinite divisibility type.