Extremes of the standardized Gaussian noise (Q550150)

The author derives some extreme value asymptotics for ``increments'' of partial sums of independent standard normal random variables. More precisely, let \(\{\xi_{\mathbf k}\), \({\mathbf k}\in\mathbb Z^d\}\) be a \(d\)-dimensional array of i.i.d. \(N(0,1)\)-random variables, set \(S(A)=\sum_{{\mathbf k}\in A}\xi_{\mathbf k}\), and let \(|A|\) be the number of points in a finite set \(A\subset \mathbb Z^d\). Define \[ u_n(\tau)= \sqrt{2d\log n}+ \frac{\frac 12\log(d\log n) +\log\frac{(2d)^d J_d}{\sqrt{\pi}}+\tau}{\sqrt{2d\log n}},\qquad\tau\in\mathbb R, \] where \(J_d\in (0,\infty)\) is an explicitly given constant. The author's first main result proves that, for every \(\tau\in\mathbb R\), \[ \lim_{n\to\infty}\text{P}\left[\max_{A\in{\mathfrak C}_n^d} \frac{S(A)}{\sqrt{|A|}} \leq u_n(\tau)\right]=e^{-e^{-\tau}}, \] where \({\mathfrak C}_n^d\) denotes the set of all lattice-point cubes contained in \(\{1,\dots,n\}^d\). The second result provides a corresponding Gumbel-type extreme value asymptotic, in which \(A\) now varies over the set \(\mathfrak R_n^d\) of all discrete rectangles contained in \(\{1,\dots,n\}^d\). Moreover, corresponding continuous-time analogs of the above asymptotics are derived for suprema of standardized ``increments'' \(W(A)/\sqrt{\lambda(A)}\), where \(W\) is a zero-mean Gaussian process indexed by the collection \({\mathcal B}_b(\mathbb R^d)\) of all Borel sets \(A\) in \(\mathbb R^d\) with finite Lebesgue measure \(\lambda(A)\) such that, for every \(A_1,A_2\in{\mathcal B}_b(\mathbb R^d)\), \[ \text{Cov}(W(A_1),W(A_2))=\lambda(A_1 \cap A_2). \] Again, two results are presented covering the cases of subsets \(A\) varying in the collection of \(d\)-dimensional cubes \({\mathcal C}_n^d\) or rectangles \({\mathcal R}_n^d\), respectively, which are contained in \([0,n]^d\). Finally, the notion of asymptotic extreme-value rate for a sequence \(X_n\) of zero-mean unit-variance Gaussian random fields defined on some parameter space \(T_n\) \((n\in\mathbb{N})\) is introduced. The latter is then applied to evaluate the four extreme value asymptotics above with respect to their extreme-value rates.