Two hypotheses on the exponential class in the class of \(O\)-subexponential infinitely divisible distributions (Q2031014)

For a probability distribution \(\rho\) on \(\mathbb{R}\) denote its tail by \(\overline{\rho}(x) := \rho((x,\infty))\). Let \(\gamma \geq 0\) and denote by \(\mathcal{L}(\gamma)\) the class of all probability distributions on \(\mathbb{R}\) such that \(\overline{\rho}(x) > 0\) for all \(x>0\) and such that \(\lim_{x\to\infty} \rho(x+a) / \rho(x) = e^{-\gamma a}\) for every \(a\in \mathbb{R}\). A probability distribution \(\rho\) is said to belong to the class \(\mathcal{OS}\) if \(\overline{\rho}(x) > 0\) for all \(x> 0\) and \[0 < \liminf_{n\to\infty} \frac{\overline{\rho * \rho} (x)}{\overline{\rho}(x)} \leq \limsup_{n\to\infty} \frac{\overline{\rho * \rho} (x)}{\overline{\rho}(x)} < \infty.\] Finally, a distribution \(\rho\) is infinitely divisible if for every \(n\in \mathbb{N}\) there exists a probability distribution \(\rho_n\) such that \(\rho_n^{n\ast} = \rho\), where \(\rho_n^{n\ast}\) denotes the \(n\)'fold convolution of \(\rho_n\) with itself, and the class of infinitely divisible probability distributions is denoted by \(\mathcal{ID}\). The author shows that for \(\mu \in \mathcal{OS}\cap \mathcal{ID}\) for which there exists an \(n\in \mathbb{N}\) such that both \(\mu^{n\ast}\) and \(\mu^{(n+1)\ast}\) are in \(\mathcal{L}(\gamma)\), then both \(\mu^{(n+2)\ast}\) and \(\mu\) are in \(\mathcal{L}(\gamma)\). The author then investigates the question whether for every \(\mu \in \mathcal{OS} \cap \mathcal{ID}\) such that \(\mu^{n\ast} \in \mathcal{L}(\gamma)\) for some \(n\in \mathbb{N}\), is it implied that \(\mu^{(n+1)\ast} \in \mathcal{L}(\gamma)\)? He also considers the related question whether \(\mu \in \mathcal{OS} \cap \mathcal{ID}\) such that \(\mu^{n\ast} \in \mathcal{L}(\gamma)\) for some \(n\in \mathbb{N}\) always implies that \(\mu \in \mathcal{L}(\gamma)\). While the answers to the questions are not completely settled, the author shows that they are equivalent in the sense that they are either both true or both wrong, and that they are further equivalent (in the sense that they are true) to the fact that there does not exist an \(a>0\) such that the \(t\)-th convolution power \(\mu^{t\ast}\) of \(\mu\) (where \(t>0\)) is in \(\mathcal{L}(\gamma)\) if and only if \(t\in a \mathbb{N}\). Various examples of distributions where no such \(a>0\), as appearing in the last condition, exists are discussed.