A new inequality of Menshov-Rademacher type and the strong law of large numbers (Q1911365)

Let \(X_1, X_2, \dots, X_N\) be r.v.'s; \(u_1, u_2, \dots, u_N\) nonnegative real numbers, and \[ {\mathcal M} (1,N) : = \max_{1 \leq n \leq N} \left |\sum^N_{k = 1} X_k \right |, \quad U : = \sum^N_{k = 1} u_k. \] First main result (Theorem 1): Assume that for some \(\gamma > 1\) and for all permutations \(\ell (1)\), \(\ell (2), \dots, \ell (N)\) of the integers \(1,2, \dots, N\) we have \[ E \left |\sum^n_{k = m} X_{\ell (k)} \right |^\gamma \leq \sum^n_{k = m} u_{\ell (k)} \quad (1 \leq m \leq n \leq N). \tag{1} \] Then for every \(\varepsilon\), \(0 < \varepsilon \leq \gamma\), there exists a constant \(C\) depending only on \(\gamma\) and \(\varepsilon\) such that \[ E {\mathcal M}^\gamma (1,N) \leq C (\log 2N)^\varepsilon \sum^N_{k = 1} u_k \left( \log {2U \over u_k} \right)^{\gamma - \varepsilon}. \tag{2} \] The classical Menshov-Rademacher inequality [see e.g. \textit{J. L. Doob}, ``Stochastic processes'' (1953; Zbl 0053.26802), p. 156] is a particular case of (2). Particular cases of the inequality (2) were given by the third author [Acta Sci. Math. 26, 249-251 (1965; Zbl 0136.05001)] for orthogonal r.v.'s, \(\gamma = 2\), \(\varepsilon = 1\); and by the second author [Z. Wahrscheinlichkeitstheorie Verw. Geb. 35, 299-314 (1976; Zbl 0314.60023)]. Now, let \(\{X_k : k \geq 1\}\) be a sequence of r.v.'s, \(\{u_k : k \geq 1\}\) a sequence of nonnegative real numbers, \(I_p : = \{2^p + 1,2^p + 2, \dots, 2^{p + 1}\}\) and \(U_p : = \sum_{k \in I_p} u_k\) \((p \geq 0)\). As a consequence of the maximal inequality (2), the authors get the following strong law of large numbers (SLLN) theorem: Assume that for some \(\gamma > 1\) and for all permutations \(\ell (1)\), \(\ell (2), \dots, \ell (2^p)\) of \(I_p\) in the case of \(p\) large enough we have (1). If for some \(0 < \varepsilon \leq \gamma\) we have \[ \sum^\infty_{ p = 0} \sum_{k \in I_p} {u_k \over k^\gamma} (\log k)^\varepsilon \left( \log {2U_p \over u_k} \right)^{\gamma - \varepsilon} < \infty, \tag{3} \] then \[ \lim_{n \to \infty} {1 \over n} (X_1 + X_2 + \cdots + X_n) = 0 \text{ a.s.} \] For orthogonal square-integrable r.v.'s \(X_k\), we recognize a classical SLLN theorem, by doing \(\varepsilon = \gamma = 2\). In the last part of this paper, the second and third named authors characterize those sequences \(\{X_k : k \geq 1\}\) of square-integrable r.v.'s on the unit interval \([0,1]\) such that, if \(\sigma^2_k = E(X_k^2)\), the sequence \(\{\sigma_k/k : k \geq 1\}\) is nonincreasing, and which obey the SLLN. This final part is very rich and contains illuminating counterexamples.