Capturing the multivariate extremal index: bounds and interconnections (Q2271714)

Consider \(D\)-variate stationary max-stable processes \(\mathbf{Y}=\{\mathbf{Y}_i = (Y_{i,1}\dots,Y_{i,D}),i\in\mathbb Z\}\) and assume that the univariate marginal distribution functions \(F_d\) are standard Fréchet, \(F_d(x_d)=\exp\{-x_d^{-1}\}\) for \(x_d>0, d = 1,\dots,D\). Let \(\tilde\mathbf{Y}=\{\tilde\mathbf{Y}_i = (\tilde{Y}_{i,1}\dots,\tilde{Y}_{i,D}),i\in\mathbb Z\}\) be the associated \(D\)-variate sequence of i.i.d. random vectors with the same marginal distribution and let \(\mathbf{M}_n=(\max_{i=1,\dots,n}Y_{i,1},\dots,\max_{i=1,\dots,n}Y_{i,D})\), and \(\tilde\mathbf{M}_n\) similarly, denote the sequences of componentwise maxima. Then, for any \(\mathbf{x}=(x_1,\dots,x_D)\in\mathbb R^D_+\) and \([\mathbf{0},\mathbf{x}]^c=[\mathbf0,\infty]\setminus[\mathbf0,\mathbf{x}]\) we have by Theorem 3.1 in \textit{G. Samorodnitsky} [Ann. Probab. 32, No. 2, 1438--1468 (2004; Zbl 1049.60027)] and a tightness argument that \(\lim_{n\to\infty} P(n^{-1}\mathbf{M}_n\leq\mathbf{x})=\exp\{-\mu([\mathbf0,\mathbf{x}]^c)\}=G(x)\), \(\lim_{n\to\infty} P(n^{-1}\tilde\mathbf{M}_n\leq\mathbf{x})= \exp\{-\tilde\mu([\mathbf0,\mathbf{x}]^c)\}=\tilde G(x)\), where \(\mu(x)\) and \(\tilde\mu(x)\) denote the exponent measures as in \textit{S. I. Resnick} [Extreme values, regular variation, and point processes, Springer-Verlag (1987; Zbl 0633.60001)]. Then, for \(\mathbf{v}\in[\mathbf0,\infty)\setminus\{\mathbf0\}\subseteq\mathbb R^D_+\), the function \(\theta(\mathbf{v})=\mu([\mathbf0,{\mathbf{v}}^{-1}]^c)/ \tilde\mu([\mathbf0,{\mathbf{v}}^{-1}]^c)\), introduced by \textit{S. Nandagopalan} [J. Res. Natl. Inst. Stand. Technol. 99, No.4, 543--550 (1994; Zbl 0881.60050)] is called the multivariate extremal index. For \(D=1\) the quotient of the exponent measures reduces to the well-known univariate extremal index. \(\theta(\mathbf{v})\) is the univariate extremal index of the series \(\{\max_d v_dY_{i,d},i\in\mathbb Z\}\). Let \(\mathbf{\theta}=(\theta_1,\dots,\theta_D)\), and let \(\mathbf{\theta v}= (\theta_1v_1,\dots,\theta_Dv_D)\) denote the componentwise multiplication. Then, \(\theta(\mathbf{v})=l(\mathbf{\theta v})/\tilde{l}(\mathbf{v}),\) where \(l\) and \(\tilde{l}\) are the two stable tail dependence functions. For example, \(l({\mathbf z}^{-1})=\mu([\mathbf0,\mathbf{x}]^c),z_d=-(\ln{G_d(x_d)})^{-1},d=1,\dots,D\). The authors extend the set of common properties of the multivariate extremal index function and derive sharp bounds for the entire function given only marginal dependence. It is shown how another popular dependence measure, the extremal coefficient \(\tilde{\phi}=\tilde{l}(\mathbf1)\), is closely related to the multivariate extremal index. An obvious connection between the univariate extremal indices and the extremal coefficient is discussed and an improved upper bound for the dependence adjusted extremal coefficient, \({\phi}={l}(\mathbf{\theta}\mathbf1)\), a counterpart of \(\tilde{\phi}\) that applies to stationary sequences, is given. In the main, however, the authors concentrate on the fact that \({\theta}(\mathbf1)={\phi}/\tilde{\phi}\). Knowledge of \(\tilde{\phi}\) or \({\phi}\), respectively, allows a significant improvement of the unrestricted bounds for \({\theta}(\mathbf{v})\). Throughout the text various example processes are discussed.