Almost sure convergence for the maxima and minima of strongly dependent nonstationary multivariate Gaussian sequences (Q6112868)

The authors prove a multivariate version of an almost sure max-limit theorem in [\textit{F. Lin}, Electron. Commun. Probab. 14, 224--231 (2009; Zbl 1189.60065); Comput. Math. Appl. 62, No. 2, 635--640 (2011; Zbl 1228.60042)] for the joint limit of the maximum and minimum of certain strongly dependent nonstationary sequences \((X_n=(X_{n1},\ldots,X_{nd}))_{n\in\mathbb N}\) of Gaussian random vectors in \(\mathbb R^d\). Let \[ A_n=\{-u_n(y)<\textstyle\min_{1\leq i\leq n}X_i\leq\textstyle\max_{1\leq i\leq n}X_i\leq u_n(x)\}, \] where \(u_n(x)=a_nx+b_n\mathbf1\) for \(x\in\mathbb R^d\) with normalizing sequences \((a_n)\), \((b_n)\) coming from the MDA of the univariate Gumbel distribution. Under certain assumptions on strong dependency implying that \[ |\mathrm{Cov}(X_{ki},X_{(k+n)j})\log n-\varrho_{ij}|=o(1) \] for some \(\varrho_{ij}\geq 0\), the authors first show that \(P(A_n)\) converges to a Gaussian location mixture of a multivariate Gumbel distribution. Under an additional assumption on the rate of convergence for strong dependency \[ |\mathrm{Cov}(X_{ki},X_{(k+n)j})\log n-\varrho_{ij}|(\log\log n)^{1+\varepsilon}=O(1) \] the authors then derive almost sure convergence of \((\log n)^{-1}\sum_{k=1}^nk^{-1}1_{A_n}\) to the same limit.