On the almost sure convergence of series of stationary and related nonstationary variables (Q1902948)

This paper addresses the classical problem of the almost sure convergence of partial sums of weighted random variables, but in the case of a broad class or dependent random variables. The author briefly reviews existing results on certain classes of dependent random variables (orthonormal sequences, lacunary systems, quasistationary systems and mixing random variables) and proceeds to develop a single framework for such examples. This framework mainly assumes only finite \(p\)th moments, \(1 \leq p \leq 2\), rather than a finite \(2 + \delta\) moment. The conditions are largely complementary to existing conditions. The class of dependency addressed here is one where there exists a Fourier representation for the sequence and an associated ``spectrum''. Precisely, \(\{X_i : i \in Z\}\) is assumed to be \((p,q)\)-bounded, i.e. for every finite selection \(\{\alpha_i : i \leq N\}\) \[ \left( E \left |\sum^N_{n = -N} \alpha_n X_n \right |^p \right)^{1/p} \leq C \left( \int^{\pi}_{- \pi} \left |\sum^N_{n = - N} \alpha_n e^{in \theta} \right |^q d \theta \right)^{1/q}. \] The main theorem (Theorem 3.1) then states that if \(\{X_i : i \in Z\}\) is \((p,q)\)-bounded, \(1 \leq p \leq 2 \leq q \leq + \infty\), and \(\sum |a_n |^22 |n |^{(q - 2)/q} \log^2 (1 + |n |) < + \infty\), then \(\sum^N_{n = - N} a_n X_n\) converges almost surely. The conditions are sharp, and the \(\{X_i : i \in Z\}\) need not be stationary. A key lemma (Lemma 2.1) reduces the problem to \((2,q)\)-bounded sequences. The proof is then an adaptation of that for Rademacher-Menchov theorem. The \((p,q)\)-boundedness condition enters in several places to provide a bound on the second moment of sums of the sequence over dyadic blocks of indices. The integral on the right-hand side of the \((2,q)\) inequality is handled by the assumed convergence of the series in \(a_n\) given in the statement of the theorem. Corollaries 3.3, 3.5, and 3.6 specialize Theorem 3.1 to weakly stationary sequences and lacunary \(S_p\) systems, the case of iterated contractions, and harmonizable symmetric \(\alpha\)-stable sequences. Theorems 3.8, 3.10, and 3.12 show how to extend results on orthonormal sequences to \((p,q)\)-bounded sequences. Theorem 3.13 provides partial results when \(p \leq 2 \leq q\) fails. The last result, Theorem 3.14, provides a rate at which the estimators of regression coefficients converge to the true parameter values when error variables are \((p,2)\)-bounded.