Convergence of marked point processes of excesses for dynamical systems (Q1783998)

Let \((\mathcal{M},\mathcal{B}, \mathbb{P},f)\) be a dynamical system on a Riemannian manifold \(\mathcal{M}\). Any observable \(\varphi : \mathcal{M}\to \overline{\mathbb{R}}\) gives rise to a stationary process \((X_n:=\varphi\circ f^n)\). Fix thresholds \((u_n)\) such that \(\lim_{n\to\infty} n\mathbb{P}(X_0>u_n) = \tau >0\). For any Borel set \(B\in\mathcal{B}\) and \(x\in \mathcal{M}\), denote by \(r_B(x)\) the return time of \(f^n(x)\) into \(B\), and define the \(j\)-th inter-hitting time by \(w^1_B=r_B\) and \(w_B^j(x) := r_B\Big(f^{\sum_{i=1}^{j-1}w_B^i(x)}(x)\Big)\); so that the \(j\)-th hitting time is \(r^j_B = \sum_{i=1}^{j}w_B^i\). The authors are interested in counting exceeding events, such as the following examples: for any time interval \(I\subset \mathbb{N}^*\) and threshold \(u\), \[ m_u\big\{X_j, j\in I \big\} := \text{ either }\sum_{j\in I}(X_j-u)^+,\text{ or }\max_{j\in I}(X_j-u)^+,\text{ or }\sum_{j\in I}1_{\{X_j>u\}}. \] But actually, they intend to consider such events as aggregated in clusters. These clusters are defined as subsequences of successive exceeding events which are separated by inter-hitting times less than some given \(p\). Thus, if \(I_1,\ldots,I_N\) denote the successive \(p\)-clusters within \(I\), the authors preferably consider \[ \mathcal{A}_u(I) := \sum_{j=1}^N m_u\big\{X_j, j\in I_j \big\}. \] Among the three examples above, aggregating by clusters only modifies the second one. In order to get a limit law, they dilate the observation intervals by \(v_n:=\mathbb{P}(X_0>u_n)^{-1}\), and then actually consider \(A_n\Big(\sqcup_{i=1}^k J_i\Big) := \sum_{i=1}^k \mathcal{A}_{u_n}(v_nJ_i)\), for any disjoint union of intervals \(J_1,\ldots,J_k\). Then the authors make some hypotheses, more or less technical, in order to force the clusters, of convenient maximal inter-times \(p\), to have some mean size \(\theta^{-1}\) and to have a small enough probability to be too long, and to be separated by \(p\)-clusters-inter-times going (after renormalization by some deterministic sequence \(a_n\)) to some law \(\mu\). Furthermore, they suppose that the dynamical system under study satisfies some mixing properties. Then their main result is the convergence in law of the ``marked point process of rare events'' \(a_nA_n\), towards a compound Poisson point process \(\sum_{n\in\mathbb{N}^*}\kappa_n1_{\{S_n\leq t\}}\) with intensity \(\theta\) and multiplicity law (law of the i.i.d. sequence \(\{\kappa_n\}\)) \(\mu\). Moreover the authors show that their mixing (rather technical) assumptions follow from some convenient decay-of-correlations condition.