Almost everywhere convergence of convolution powers without finite second moment (Q646391)

Let \(\mu(k)\), \(k \in \mathbb{Z}\), be a probability measure on \(\mathbb{Z}\). Its convolution \(\mu * \mu\) is defined by \[ \mu * \mu(k) = \sum_{j \in Z} \mu(j) \mu(k - j), \] and the \(n\)-fold convolution \(\mu^n(k)\) is defined inductively by \(\mu^n(k) = \mu * \mu^{n - 1}(k)\). The Fourier transform of \(\mu\) is denoted \(\theta(t)\), \(t \in [-1/2, 1/2)\), and is equal to \[ \theta(t) = \sum_{k \in Z} \mu(k)e^{2 \pi i t}. \] The \(p^{\text{th}}\) moment of \(\mu \) is the sum \(m_p(\mu) = \sum |k|^p \mu(k)\), and \(\mu \) has a \(p^{\text{th}}\) moment if the sum is finite. Here, \(p\) is allowed to take non-integral values, the expectation of \(\mu\) is \(E(\mu) = \sum_{k \in Z} k \mu(k)\), and \(\mu\) is called strictly aperiodic if the support of \(\mu\) is not contained in a proper coset of the integers, which is known to be equivalent to \(|\theta(t)| \neq 1\) for all \(t \neq 0\). The author considers the convergence of the sequence \(\mu^n f(x)\), where \(T:X \to X\) is an invertible measure preserving transformation of a probability space \((X, \mathcal{B}, \lambda)\) and \[ \mu^n f(x) = \sum_{k \in Z} \mu^n(k) f(T^k x),\;f \in L^1(X, \lambda). \] There is a procedure for considering this question which uses the Calderón transfer principle to reduce the question to a weak type estimate for a maximal operator, associated with the \(\mu^n\). Using these methods, \textit{A. Bellow} and \textit{A. P. Calderón} proved the almost everywhere convergence of \(\mu^nf(x)\) when \(\mu\) is strictly aperiodic, \(E(\mu) = 0\) and \(m_2(\mu) < \infty\) [``A weak-type inequality for convolution products'', in: M. Christ (ed.) et al., Harmonic analysis and partial differential equations. Proceedings of a conference, University of Chicago, IL, USA, February 1996, 41--48 (1999; Zbl 0960.28011)]. \textit{V. Losert} proved that the result cannot be extended to measures with \(m_p(\mu) < \infty\) for \(p < 2\) [Ill. J. Math. 43, No.~3, 465--479 (1999; Zbl 0963.28014); Ergodic Theory Dyn. Syst. 21, No.~1, 115--119 (2001; Zbl 0972.37002)]. However, it was shown by \textit{A. Bellow}, \textit{R. L. Jones} and \textit{J. Rosenblatt} that, if \(\mu\) is symmetric and \(\mu(k) \geq \mu(k + 1)\) for all \(k>0\), then \(\mu^nf(x)\) converges a.e.\ for \(f \in L^1(X)\) [Math. Ann. 293, No.~3, 399--426 (1992; Zbl 0736.28008); Ergodic Theory Dyn. Syst. 14, No.~3, 415--432 (1994; Zbl 0818.28005)]. The author gives a result which weakens the requirement that the second moment be finite in the Calderón-Bellow theorem. A~corollary of his main result is the following. Theorem. Suppose that \(\mu\) is a strictly aperiodic, symmetric measure on \(\mathbb{Z}\) and, for some \(0 < \alpha \leq 1\), \[ \sum_{|k| \leq n } k^n \mu(k) = O(n^{1 - \alpha}). \] \noindent Suppose that \(\theta^{\prime \prime}(t) \) exists in some set \(0 < |t| < \delta\) and that \(\operatorname{Re}(\theta^{\prime \prime}(t)) = p(t) + O(1) < 0\), where \(p(t)\) is non-decreasing in this set. Then \(\{ \mu^nf(x) \}\) converges a.e.\ for all \(f \in L^1(X)\).