Series of moderate deviation probabilities for martingales (Q2484190)

As a complement to (extensions of) the law of large numbers there has been much work concerning Borel-Cantelli type sums related to the corresponding law, starting with legendary results by \textit{P. L. Hsu} and \textit{H. Robbins} [Proc. Natl. Acad. Sci. USA 33, 25--31 (1947; Zbl 0030.20101)] and \textit{F. Spitzer} [Proc. Am. Math. Soc. 7, 164--171 (1956; Zbl 0070.13802)]. Let \(X_1,X_2,\dots\) be i.i.d. random variables with partial sums \(S_n\), \(n\geq 1\). A theorem due to \textit{J. A. Davis} [Ann. Math. Stat. 39, 1479--1485 (1968; Zbl 0174.49902)] asserts that \[ \sum^\infty_{n=2} {\log n\over n} P(|S_n|> \varepsilon\sqrt{n\log n})< \infty \] for all \(\varepsilon> 0\) iff \(EX^2<\infty\) and \(EX = 0\). The first result of the present paper tells us that the sum remains convergent if \(X_1,X_2,\dots\) is an \(L^p\)-bounded martingale difference sequence for some \(p> 4\), and that the result fails if \(p\in[1, 4)\). The second result is an analog pertaining to \[ \sum^\infty_{n=3} {1\over n\log n}P(|S_n|> \varepsilon\sqrt{n\log\log n}). \] The boundary between convergence and failure is \(p= 2\) in this case. In both cases it is not known whether \(p= 4\) and \(p= 2\), respectively, belong to the convergence case or not.