Transition kernels and the conditional extreme value model (Q488095)

Multivariate modelling of extremes proposed in [\textit{J. E. Heffernan} and \textit{J. A. Tawn}, J. R. Stat. Soc., Ser. B, Stat. Methodol. 66, No. 3, 497--546 (2004; Zbl 1046.62051)] was reformulated as the conditional extreme value model (CEVM) in [\textit{J. E. Heffernan} and \textit{S. I. Resnick}, Ann. Appl. Probab. 17, No. 2, 537--571 (2007; Zbl 1125.60049)]. Restricting to two dimensions, the article under review looks at how a formulation based on asymptotic theory of transition kernels fits with the general CEVM theory expressed in terms of vague convergence of measures and how constructing the CEVM from transition kernels restricts the class of measures. For this, the kernel domain of attraction condition used in [\textit{S. I. Resnick} and \textit{D. Zeber}, Adv. Appl. Probab. 45, No. 1, 186--213 (2013; Zbl 1270.60061)] is extended to accommodate general linear normalization in both the initial state and the distribution of the next state. It is shown that in the standard case when both components are scaled by the same function, kernel convergence always implies the CEVM. A counterexample is given to show that the converse fails. The authors also study the case when both components are scaled by different normalizing functions.