Second-Order Geometric Statistics

• Second-Order Geometric Statistics is a framework that measures dependencies among geometric functionals like volume and surface area in random spatial structures.
• It utilizes add-one operators, Fock space expansions, and stabilization techniques to derive asymptotic covariances and establish central limit theorems.
• The theory underpins applications in Boolean models, tessellations, geometric graphs, and topological data analysis for robust statistical inference.

Second-order geometric statistics quantify and control the joint fluctuations and dependencies of geometric functionals—such as volume, surface area, connectivity, and topological invariants—associated with random spatial structures. These statistical properties permeate the study of stochastic geometry, random spatial tessellations, geometric graphs, and related models, where understanding covariances and limit laws is essential for both inference and theory. Rigorous asymptotic covariances, sharp central limit theorems (CLT), and the development of multivariate Berry–Esseen-type bounds underpin a modern second-order theory.

1. Formalism and Model Classes

A geometric functional is an additive, translation-invariant, locally bounded map $\psi$ on the convex ring or on random geometric structures such as Boolean models, tessellations, geometric graphs, or general functionals of Poisson processes. The prototypical setting involves a stationary Poisson process $\eta$ with intensity $\gamma$ (on convex bodies or points) and a compact observation window $W$; the vector of interest is typically

$$\Psi(W) = \left( \psi_1(Z \cap W), \dots, \psi_m(Z \cap W) \right)$$

for a random set $Z$ (e.g., Boolean models: $Z = \bigcup_{K\in\eta} K$) or a sum over Poisson points with stabilizing score functions (e.g., geometric graphs, random tessellations) (Hug et al., 2013, Hug et al., 2016, Schreiber et al., 2010, Schulte et al., 2021).

Functionals of this type include intrinsic volumes ($V_0,\dots,V_d$), Minkowski functionals, Euler characteristic, edge or vertex counts, subgraph counts, and other local geometric measurements.

2. Asymptotic Covariances: Fock Space Expansions and Explicit Formulas

Second-order statistics are centered on asymptotic covariance matrices: $$\sigma_{\psi,\varphi} = \lim_{r(W)\to\infty} \frac{\operatorname{Cov}(\psi(Z\cap W), \varphi(Z\cap W))} {V_d(W)}$$ and multivariate analogues. For Poisson–Boolean models, expansion via the Fock space (Wiener–Itô chaos) yields

$$\operatorname{Cov}(F, G) = \sum_{n=1}^{\infty} \frac{1}{n!} \int E[D^n_{K_1,\dotsc,K_n} F] \, E[D^n_{K_1,\dotsc,K_n} G] \; \Lambda^{n}(dK_1,\dotsc,dK_n)$$

where $D^n_{K_1,\dots,K_n} F$ iterates add-one-cost operators over $n$ grains $K_i$ (Hug et al., 2013). Asymptotic covariances between general geometric functionals have compact representations using “add-a-grain” transforms $\psi^*(K) = E[\psi(Z \cap K)] - \psi(K)$ and mixed moment measures (Hug et al., 2016): $$\sigma(\psi,\varphi) = \rho(\psi^*, \varphi^*), \quad \rho(\psi,\varphi) = \sum_{k=1}^\infty \frac{\gamma}{k!} \int_{\mathcal{K}^n \times \dotsb \times \mathcal{K}^n} \psi(K_1 \cap \dotsb \cap K_k)\,\varphi(K_1 \cap \dotsb \cap K_k) \cdots$$ For intrinsic volumes, explicit integral formulas in terms of curvature measures, covariograms, and overlap geometry follow (Hug et al., 2016).

In other stochastic geometric models (e.g., STIT tessellations (Schreiber et al., 2010)), the covariance, cross-covariance, and pair-correlation structure of vertex and edge measures is given by integral or series representations depending on tessellation rules, often specialized for the stationary–isotropic regime.

3. Central Limit Theorems and Quantitative Rates

Second-order statistics drive multivariate normal approximation for geometric functionals by determining limiting covariance matrices and controlling convergence rates. For Poisson–Boolean models and similar settings, the multivariate CLT holds for

$$\mathbf{X}(W) = V_d(W)^{-1/2} \left( \Psi(Z\cap W) - \mathbb{E}\Psi(Z\cap W) \right),$$

yielding Gaussian limits $N_\Sigma$ in $\mathbb{R}^m$ with covariance matrix $\Sigma$ given by the asymptotic second-order theory (Hug et al., 2013, Hug et al., 2016).

Sharp Berry–Esseen-type bounds are established via Malliavin–Stein or second-order Poincaré inequalities, with rates characterized as follows:

• Under sufficient moment conditions, the error in $d_3$ (smooth-test) distance decays as $O(r(W)^{-1})$ or $O(V_d(W)^{-1/d})$, unimprovable in general (Hug et al., 2013, Hug et al., 2016, Schulte et al., 2021, Trauthwein, 2024).
• In generalized Poisson input and geometric graph models, the sharp rate for comparing centered, scaled statistics to the limiting Gaussian is $s^{-1/d}$, for $s$ the intensity parameter (Schulte et al., 2021).
• Refined second-order $p$-Poincaré inequalities permit extension of these rates to settings with only $(2+\epsilon)$ moment conditions on add-one costs, matching optimal $t^{-1/2}$ rates for statistics outside the reach of classical fourth-moment approaches (Trauthwein, 2024).

4. Methodological Framework: Add-One Operators and Stabilization

The modern approach is grounded in difference operators (add-one and iterated add-one costs) for Poisson functionals: $$D_x F = F(\eta+\delta_x) - F(\eta) \,, \quad D^2_{x,y} F = F(\eta+\delta_x+\delta_y) - F(\eta+\delta_x) - F(\eta+\delta_y) + F(\eta)$$ These provide analytic access to covariance and higher-order dependencies.

For geometric probability and spatial statistics, “score function” representations and stabilization play a pivotal role. Exponential stabilization (i.e., bounded probability of large influence radii) ensures weak dependence, and, together with moment control, delivers tractable rates and asymptotic normality for a vast class of geometric and topological statistics (Schulte et al., 2021). These conditions enable application of second-order Poincaré inequalities and coupling techniques delivering explicit rates and limit theorems.

5. Explicit Examples and Applications

Boolean models with general grains: All formulas above apply for convex grains; explicit covariance structure is accessible for intrinsic volumes, support measures, and in anisotropic examples (e.g., planar aligned rectangles), where analytic results match simulations up to sampling fluctuations (Hug et al., 2016).

Geometric graphs: Covariances and CLTs are provided for statistics in the $k$-NN and Gilbert graphs—lengths, degree counts, subgraph counts—with precise rates ($s^{-1/d}$), relying on stabilizing score representations (Schulte et al., 2021, Trauthwein, 2024).

STIT Tessellations: Second-order analysis yields closed formulas for covariance, pair-correlation, and cross-correlation of vertex and edge measures in both nonstationary and stationary–isotropic settings. The variance scaling exhibits weak long-range dependence ($\sim R^2 \ln R$ for window size $R$), intermediate between Poisson-Voronoi and Poisson-line tessellations (Schreiber et al., 2010).

Topological Data Analysis: Counts of critical points (e.g., Morse indices) in Poisson–Boolean unions fall under the stabilized functional framework, inheriting explicit rate control and Gaussian approximation in sparse regimes (Schulte et al., 2021).

6. Extensions, Optimality, and Limitations

Second-order theories generalize to abstract metric space settings (under suitable volume growth conditions), with scaling exponents derived from Minkowski dimension (Schulte et al., 2021). Rates $s^{-1/d}$ (or $s^{-1/\gamma}$) are generally unimprovable, as established via lower-bound examples (e.g., joint vertex-edge counts in the Gilbert graph).

Limitations include requirements for convex grains, stationarity, and high-moment existence for functionals such as intrinsic volumes. In random geometric graphs, heavy-tailed edge-length effects complicate analysis—recent advances in $p$-Poincaré inequalities circumvent some moment restrictions (Trauthwein, 2024). Nonconvex or unbounded grains, weaker stabilization, and extensions to inhomogeneous or anisotropic models remain active areas of investigation.

7. Interpretive and Applied Significance

Second-order geometric statistics encode information far beyond first-order means:

• Sensitivity to anisotropy: distinct behavior for aligned vs. random orientations, as manifest in off-diagonal entries of covariance matrices (Hug et al., 2016).
• Detection of shape variability and clustering: Second-order terms distinguish rectangles from disks, or quantify loop/hole formation via sign and magnitude (e.g., $\sigma(V_0, V_1)$).
• Design of statistical inference tools: Explicit CLTs with rates underpin confidence intervals, hypothesis tests, and goodness-of-fit methods for random media and spatial structures (Schreiber et al., 2010).

These theoretical foundations are instrumental across stochastic geometry, image analysis, spatial statistics, and materials science. Second-order analysis provides not only rigorous asymptotic distributions but also explicit formulas enabling simulation validation, model comparison, and inference for physical properties.