Large deviations of the empirical volume fraction for stationary Poisson grain models (Q1774224)

The Poisson grain model (PGM; also known as the Boolean model) is the basic model in stereology and stochastic geometry. Statistical analysis of a stationary PGM is based on a single realization of the union set of clumps in some region \(W\) which assumed to expand unboundedly in all directions of a Euclidean space \(\mathbb R^d\), \(d\geq 1\) [see \textit{I. Molchanov}, ``Statistics of the Boolean model for practitioners and mathematicians'' (1997; Zbl 0878.62068)]. A stationary PGM is the union \(\Xi := \bigcup_{i\geq1} (\Xi_i + X_i)\) of independent copies \(\Xi_1\), \(\Xi_2\),\dots (grains) of a random compact set \(\Xi_0\) (typical grain), where the grains are independently shifted by the atoms \(X_1\), \(X_2\), \(\dots\) (germ points) of a stationary Poisson process \(\mathbb \Pi_\lambda =\sum_{i\geq 1} \delta_{X_i}\) with intensity \(\lambda>0\). Note that \(\Xi\) is a closed set if \(E| \Xi_0 +B_r(O)| <\infty\), for \(r>0\), where \(E\) stands for expectation, \(B_r(x)\) denotes the closed ball with radius \(r>0\) centered at \(x\in \mathbb R^d\) and \(| \cdot| \) denotes the Lebesgue measure in \(\mathbb R^d\) [see \textit{L. Heinrich}, Statistics 23, No. 3, 271--286 (1992; Zbl 0811.60034)]. The main aim of this paper is to prove the existence and analyticity of the (thermodynamic) limit \(L(z)\) (as \(n\to \infty\)) of the scaled cumulant-generating function \[ L_n(z) = | W_n| ^{-1} \log E\exp\{z\, | \Xi\cap W_n| \} \] of the empirical volume fraction \(\widehat p_n :=| \Xi\cap W_n| /| W_n| \) on \(D_{\Delta}:=\{z\in C^1: \, | z| < 1/\Delta\}\) for some \(0<\Delta<\infty\) provided that an exponential moment of the volume \(| \Xi_0| \) exists, that is \(M(a) := E\exp\{a\, | \Xi_0| \} < \infty\) for some \(a>0\) and \((W_n)\) is a convex averaging sequence of sets in \(\mathbb R^d\), that is, each \(W_n\) is a (deterministic) compact convex set, \((W_n)\) is nondecreasing and its union is \(\mathbb R^d\). The second aim, which is closely connected with the first one, is to derive inequalities and asymptotic relationships (in the sense of Cramér and Chernoff) for probabilities of large deviations of \(\widehat p_n\) from its mean \(p : = E| \Xi\cap [0, 1)^d| = P(O\in \Xi)\). In the special case of a bounded typical grain, that is, \(\Xi_0\subset B_R(O)\) for some \(0< R< \infty\), both problems were solved satisfactorily by \textit{F. Götze, L. Heinrich}, and \textit{C. Hipp} [Scand. J. Stat. 22, No. 2, 183--195 (1995; Zbl 0841.60033)] using the device of \(m\)-dependent random fields with block representation. The proving technique in the present paper is completely different from that in the last mentioned paper and does not require any strong mixing properties of the PGM \(\Xi\) as one would expect [see also \textit{L. Heinrich} and \textit{I. S. Molchanov}, Adv. Appl. Probab. 31, No. 2, 283--314 (1999; Zbl 0941.60025)]. This paper is organized as follows. Section 2 investigates \(L_n(z)\) for a quite general random set model (Lemmas 1 and 2) and puts together the required tools from point process theory presented in a rather general random setting (Lemma 3). In Section 3 the existence and analyticity of the thermodynamic limit \(L(z)=\lim L_n(z)\) are proved. Section 4 contains, among other things, Cramér's large deviations relationships for the random sequence \(| \Xi\cap W_n| \) and an optimal Berry-Esseen bound of the distance between \(F_n(x) : = P(\sqrt{| W_n| } (\widehat p_n - p)\leq x\, \sigma_n)\) and the standard normal distribution function, where \(\sigma_n^2 := (\text{Var}(| \Xi\cap W_n| ))/ | W_n|\).