Semiparametric estimation for isotropic max-stable space-time processes (Q2325332)

The authors of the paper consider problems related with tail dependence coefficient of strictly stationary isotropic space-time process. A strictly stationary stochastic space-time process \(\{\eta(\mathbf{s},t), (\mathbf{s},t)\in\mathbb{R}^{d-1}\times [0,\infty)\}\) is called regularly varying if there exists an unboundedly increasing sequence \(a_n\) such that: \[ \mathrm{(i)}\ \mathbb{P}(|\eta(\mathbf{0},0)|>a_n)\mathop{\sim}\limits_{n\rightarrow\infty} n^{-d}, \] \[\mathrm{(ii)}\ n^d\mathbb{P}\Big(\frac{(\eta(\mathbf{s},t):(\mathbf{s},t)\in I)^T}{a_n}\in B\Big)\mathop{\rightarrow}\limits_{n\rightarrow\infty}^{v}\mu_I(B) \] for each finite set \( I\subset\mathbb{R}^{d-1}\times [0,\infty)\), for each Borel set \(B\subset \overline{\mathbb{R}}^{\,|I|}\backslash {\{\mathbf{0}\}}\) and for some non-null Radon measure \(\mu_I\) under condition \[ \mu_I(xC)=x^{-\beta}\mu_I(C), x>0, C\in\mathcal{B}\big(\overline{\mathbb{R}}^{\,|I|}\backslash {\{\mathbf{0}\}}\big ), \beta>0. \] It is supposed that the space-time processes considered in the paper are spatially isotropic. This together with the assumption of strict stationarity means that extremal dependence between two space-time points is only driven by the spatial and temporal lags. For a regularly varying strictly stationary isotropic space-time process \(\eta\) the tail dependence coefficient \(\chi\) is defined by the equality \[ \chi(v,u)=\lim\limits_{n\rightarrow\infty}\frac{\mathbb{P}(\eta(\mathbf{s}_1,t_1)/a_n>1,\eta(\mathbf{s}_2,t_2)/a_n>1)} {\mathbb{P}(\eta(\mathbf{s}_1,t_1)/a_n>1)}, \] where \(v=\|\mathbf{s}_1-\mathbf{s}_2\|\) and \(u=|t_1-t_2|\). The authors of the paper introduce a new semi-parametric estimation procedure to estimate \(\chi(v,u)\) by supposing, in addition, that this tail dependence coefficient can be linearly parametrised with respect to the temporal lag \(u\) and the spatial lag \(v\) separately. The obtained results are applied to the Brown-Resnick process together with verification of all required conditions. The presented estimation procedure is tested using a simulation study. The authors show that the new method works remarkably well and produces reliable estimates that are much faster to compute than composite likelihood estimates.