Refinement of convergence rates for tail probabilities (Q2576807)

There are many results aimed at providing necessary and sufficient conditions for the convergence of the sum \[ f(x)=\sum_n a_nP(|S_n|>xb_n),\;\;\;x>0, \] where \(S_n=X_1+\cdots+X_n,\) \(\{X,\;X_n;\;n\geq 1\}\) is a sequence of i.i.d. r.v., such that \(P(X\neq 0)>0,\;EX=0\;(E|X|<\infty),\) \(a_n>0,\;\sum_n a_n=\infty\) and \(b_n\) is either \(n^{1/p},\;0<p<2,\) \(\sqrt{n\log n}\) or \(\sqrt{n\log\log n}.\) Thus these results offer rates of convergence for the tail probabilities \(P(|S_n|>xb_n)\) concerning either the law of large numbers, moderate deviations or the law of the iterated logarithm. From the theory of branching processe the impetus comes to strengthen theorems in the area, and many authors relying on some inequalities, stopping time techniques and so forth have tried to get conditions under which \(\int_{\varepsilon}^{\infty}f(x)dx<\infty,\) \(\varepsilon>\) some \(\delta.\) But the authors of this paper use a different approach and derive necessary and sufficient conditions for \(\int_{\varepsilon}^{\infty}f(x^q)dx<\infty,\) \(\varepsilon>\delta,\) where \(q>0,\) and so they further refine the rate of convergence for \(P(|S_n|>xb_n).\) The main idea is based, via symmetrization and the Hoffman-Jørgensen inequality, on a device allowing to get information about the tail probabilities \(P(|S_n|>xb_n)\) from the probabilities \(P(\max_{1\leq k\leq n}|S_k|>xb_n),\) and further, to obtain information about \(P(\max_{1\leq k\leq n}|S_k|>xb_n)\) from \(P(|S_n|>xb_n),\) by using the fact that \(S_n/b_n {\overset{P}{\to}} 0.\)