Variance asymptotics and scaling limits for Gaussian polytopes (Q748444)

For \(\lambda\in [1,\infty)\), let \({\mathcal P}_{\lambda}\) denote a Poisson point process of intensity \(\lambda\,\phi(x)\, dx\), where \(\phi(x) = (2\pi)^{-d/2} \exp{\left(-|x|^2/2\right)}\) is the standard normal density in \({\mathbb R}^d\), \(d\geq 2\). Here and elsewhere \(|\cdot|\) denotes the Euclidean norm. Let \({\mathcal X}_n:=(X_1,\dots, X_n)\), where \(X_i\) are i.i.d. with density \(\phi(x)\). Let \(K_{\lambda}\) and \(K_n\) be the Gaussian polytopes defined by the convex hull of \({\mathcal P}_{\lambda}\) and \({\mathcal X}_n\), respectively. The number of \(k\)-faces of \(K_{\lambda}\) and \(K_n\) are denoted by \(f_k(K_{\lambda})\) and \(f_k(K_n)\), respectively. In the remarkable paper [\textit{I. Bárány} and \textit{V. Vu}, Ann. Probab. 35, No. 4, 1593--1621 (2007; Zbl 1124.60014); correction ibid. 36, No. 5, 1998 (2008; Zbl 1148.60301)], dependency graph methods were used to establish rates of normal convergence for \(f_k(K_n)\) and \(\operatorname{Vol}(K_n)\), \(k\in\{0, 1, \ldots, d-1\}\). A key part of their work involved obtaining sharp lower bounds for \(\operatorname{Var}(f_k(K_n))\) and \(\operatorname{Var}(\operatorname{Vol}(K_n))\). Their results stop short of determining precise variance asymptotics for \(f_k(K_n)\) and \(\operatorname{Vol}(K_n)\) as \(n\to \infty\). Let \({\mathcal P}\) be the Poisson point process in \({\mathbb R}^{d-1} \times {\mathbb R}\) with intensity \(d{\mathcal P}((v,h)) = e^h dh dv\), where \((v, h)\in {\mathbb R}^{d-1} \times {\mathbb R}\). Let \(\Pi^{\downarrow}=\{(v, h)\in {\mathbb R}^{d-1}\times {\mathbb R}: h\leq -|v|^2/2\}\) and \(\omega= (v, h)\in {\mathbb R}^{d-1}\times {\mathbb R}\). Denote by \(\Pi^{\downarrow}(\omega):=\omega \oplus\Pi^{\downarrow}\), where \(\oplus\) denotes the Minkowski addition. The maximal union of parabolic grains \(\Pi^{\downarrow}(\omega)\) whose interior contains no points of \({\mathcal P}\) is \(\Phi({\mathcal P}) = \cup_A \Pi^{\downarrow}(\omega)\), where \(A=\{\omega\in {\mathbb R}^{d-1}\times {\mathbb R}: {\mathcal P}\cap \mathrm{int}(\Pi^{\downarrow}(\omega))=\emptyset\}\). Remove points of \({\mathcal P}\) not belonging to \(\partial \Phi({\mathcal P})\) and call the resulting thinned point set \(\operatorname{Ext}({\mathcal P})\). The authors prove that the re-scaled configuration of extreme points in \({\mathcal P}_{\lambda}\) (and in \({\mathcal X}_n\)) converges to \(\operatorname{Ext}({\mathcal P})\) and that the scaling limit \(\partial K_{\lambda}\) as \(\lambda\to \infty\) (and of \(\partial K_n\) as \(n\to \infty\)) coincides with \(\partial \Phi({\mathcal P})\). Fix \(u_0= (0, 0,\ldots,1)\in {\mathbb R}^d\) and let \(T_{u_0}\) denote the tangent space to the unit sphere \({\mathbb S}^{d-1}\) at \(u_0\). The exponential map \(\exp = \exp_{d-1}: T_{u_0} \to {\mathbb S}^{d-1}\) maps a vector \(v\) of \(T_{u_0}\) to the point \(u\in {\mathbb S}^{d-1}\) such that \(u\) lies at the end of the geodesic of length \(|v|\) starting at \(u_0\) and having initial direction \(v\). For \(\lambda\in [1, \infty)\) put \[ R_{\lambda} =\sqrt{2\log\lambda - \log(2-(2\pi)^d\cdot \log \lambda)}. \] Choose \(\lambda_0\) so that for \(\lambda\in [\lambda_0, \infty)\) we have \(R_{\lambda}\in [1, \infty)\). Let \({\mathbb B}_{d-1}(\pi)\) be the closed Euclidean ball of radius \(\pi\) and centered at the origin in the tangent space of \({\mathbb S}^{d-1}\) at the point \(u_0\). It is also the closure of the injectivity region of \(\exp_{d-1}\), that is \(\exp({\mathbb B}_{d-1}(\pi))= {\mathbb S}^{d-1}\). For \(\lambda\in [\lambda_0, \infty)\) define the scaling transform \(\displaystyle T^{(\lambda)}: {\mathbb R}^d \to {\mathbb R}^{d-1}\times {\mathbb R}\) by \(\displaystyle T^{(\lambda)}(x)= \left(R_{\lambda}\exp_{d-1}^{-1} {x\over {|x|}}, R^2_{\lambda}(1-{{|x|}\over{R_{\lambda}}})\right)\), \(x\in {\mathbb R}^d\), where \(\exp_{d-1}^{-1}(\cdot)\) is the inverse exponential map. There are several ways in which this paper differs from [the first author et al., Ann. Probab. 41, No. 1, 50--108 (2013; Zbl 1278.60020)], which considers functionals of convex hulls on i.i.d. uniform points in \(B_d({\mathbf 0}, 1)\) (\(B_d({\mathbf 0}, r)\) is the closed \(d\)-dimensional Euclidean ball centered at \({\mathbf 0}\) and radius \(r\)). First, as the extreme points of a Gaussian sample are concentrated in the vicinity of the critical sphere \(\partial B_d({\mathbf 0}, R_{\lambda})\). Second, the Gaussian sample \({\mathcal P}_{\lambda}\), when transformed by \(T^{(\lambda)}\), converges to a non-homogeneous limit point process \({\mathcal P}\), which is carried by the whole of \({\mathbb R}^{d-1}\times {\mathbb R}\). This contrasts with [the first author et al., loc. cit.], where the limit point process is simpler since it is homogeneous and confined to the upper half-space. This paper is organized as follows. Section 2 introduces scaling limit functionals of germ-grain models having parabolic grains. Section 3 shows that, for each \(\lambda\in [1, \infty)\), the scaling transform \(T^{(\lambda)}\) maps the Euclidean convex hull geometry into nearly parabolic convex geometry, which in the limit \(\lambda\to \infty\) becomes parabolic convex geometry. Section 4 establishes that the re-scaled \(k\)-face and volume functionals localize in space, which is crucial to show the convergence of their means and covariances to the respective means and covariance of their scaling limits. Finally, Section 5 provides the proofs of the main results.