Implicit renewal theorem for trees with general weights (Q444354): Difference between revisions

The main result of the paper is an extension of \textit{C. M. Goldie}'s implicit renewal theorem [Ann. Appl. Probab. 1, No. 1, 126--166 (1991; Zbl 0724.60076)]. The latter theorem allows to determine the tails of solutions of certain distributional fixed-point equations, for instance, the tails of perpetuities. In an earlier paper by the authors [Adv. Appl. Probab. 44, No. 2, 528--561 (2012; Zbl 1253.60076)], the implicit renewal theorem has been extended to a version which can be applied to determine the solutions to distributional fixed-point equations involving branching such as the fixed-point equation of the smoothing transform. However, the result in [the authors, loc. cit.] is restricted to situations in which all random variables appearing are non-negative. In the paper under review, the authors give a new proof, based on a matrix renewal theorem due to \textit{M. S. Sgibnev} [Sb. Math. 197, No. 3, 369--386 (2006); translation from Mat. Sb. 197, No. 3, 69--86 (2006; Zbl 1144.60052)], of their implicit renewal theorem that covers also the case of random variables taking negative values with positive probability. To be more precise, what is new is the following. Let \((N,C_1, C_2, \dots)\) be a sequence of real-valued random variables with \(N\) taking values in \(\mathbb{N}_0 \cup \{\infty\}\), assume that \(0 < \operatorname{E} \big[ \sum_{j=1}^N |C_j|^{\alpha} \log |C_j| \big] < \infty\), \(\operatorname{E} \big[\sum_{j=1}^N |C_j|^{\alpha}\big] = 1\) and \(\operatorname{E} \big[\sum_{j=1}^N |C_j|^{\gamma}\big] < \infty\) for some \(0 \leq \gamma < \alpha\). Further, assume that \(\operatorname{P}(N \geq j, C_j < 0) > 0\) for some \(j \in \mathbb{N}\) and that \((N,C_1, C_2, \dots)\) satisfies an appropriate non-lattice condition. Then, if \(R\) is a real-valued random variable independent of \((N,C_1, C_2, \dots)\) with \(\operatorname{E} |R|^{\beta} < \infty\) for all \(0 < \beta < \alpha\) and \[ \int_0^{\infty} \bigg|\operatorname{P}(\pm R > t) - \operatorname{E} \bigg[\sum_{j=1}^N 1_{\{\pm C_j R > t\}} \bigg] \bigg| t^{\alpha-1} \, dt ~<~\infty, \] one has \(\operatorname{P}(\pm R > t) \sim H_{\pm} t^{-\alpha}\) as \(t \to \infty\) for some \(H_{\pm} \in [0,\infty)\). An implicit representation for \(H_{\pm}\) is given but it cannot be ruled out that \(H_{\pm}=0\) in some applications. The paper further contains extensions to the lattice case and applications of the main results to the fixed-point equation of the smoothing transform.