Limit distribution of the sample volume fraction of Boolean set

Hermine Biermé, Olivier Durieu, Donatas Surgailis · 2025

Read Limit distribution of the sample volume fraction of Boolean set by Hermine Biermé, Olivier Durieu, Donatas Surgailis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-13340.txt.

Introduction

Hermine Biermé and co-authors wrote a 2025 probability paper on measuring how much of a window a random set fills. A Boolean set scatters random grains across space. The sample volume fraction estimates its true coverage. As the window grows huge, that estimate settles into a limit law.

The authors give conditions under which the limit is an alpha-stable law. Those laws allow heavier tails than the normal curve. A broad class of models with randomly scaled grains fits the conditions. They also treat windows sliced by lower-dimensional planes. The results serve materials and image analysis where coverage matters.

Catalog record

CREATORHermine Biermé, Olivier Durieu, Donatas Surgailis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-13340

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Limit distribution of the sample volume fraction of Boolean set Hermine Biermé, Olivier Durieu, Donatas Surgailis Abstract We study the limit distribution of the volume fraction estimator p ^ λ , A \widehat{p}_{\lambda,A} (= the Lebesgue measure of the intersection 𝒳 ∩ ( λ ​ A ) \mathcal{X}\cap(\lambda A) of a random set 𝒳 \mathcal{X} with a large observation set λ ​ A \lambda A , divided by the Lebesgue measure of λ ​ A \lambda A ), as λ → ∞ \lambda\to\infty , for a Boolean set 𝒳 \mathcal{X} formed by uniformly scattered random grains Ξ ⊂ ℝ ν \Xi\subset\mathbb{R}^{\nu} . We obtain general conditions on generic grain set Ξ \Xi under which p ^ λ , A \widehat{p}_{\lambda,A} has an α \alpha -stable limit distribution with index 1 < α ≤ 2 1<\alpha\leq 2 . A large class of Boolean models with randomly homothetic grains satisfying these conditions is introduced. We also discuss the limit distribution of the sample volume fraction of a Boolean set observed on a large subset of a ν 0 \nu_{0} -dimensional ( 1 ≤ ν 0 ≤ ν − 1 CLOSE (1\leq\nu_{0}\leq\nu-1 ) hyperplane of ℝ ν \mathbb{R}^{\nu} . Keywords: Poisson process, Boolean set, random grain model, long-range dependence, volume fraction estimator on hyperplane, stable limit distribution, randomly homothetic grain 1 Introduction Let (\Omega,\mathcal{A},\mathrm{P}) be a complete probability space. Let \{\boldsymbol{u}_{j};j\geq 1\} be a stationary Poisson process on \mathbb{R}^{\nu} with unit intensity and \{\Xi,\Xi_{j};j\geq 1\} be an i.i.d. sequence of random sets (‘grains’) in \mathbb{R}^{\nu} , independent of \{\boldsymbol{u}_{j}\}. It is usual in stochastic geometry to consider grains as closed random sets and we denote (\mathbb{F}(\mathbb{R}^{\nu}),\mathcal{B}(\mathbb{F}(\mathbb{R}^{\nu}))) the measurable space of closed sets of \mathbb{R}^{\nu} , endowed with the \sigma algebra \mathcal{B}(\mathbb{F}(\mathbb{R}^{\nu})) induced by Fell topology (see [ 3 ] Chapter 9 for instance). Assuming that \Xi is a random closed set means that \Xi:(\Omega,\mathcal{A})\rightarrow(\mathbb{F}(\mathbb{R}^{\nu}),\mathcal{B}(\mathbb{F}(\mathbb{R}^{\nu}))) is measurable and we denote by \mathrm{P}_{{}_{\Xi}} its probability distribution. A Boolean set is defined as the union of all grains: \displaystyle{\mathcal{X}}:=\bigcup_{j=1}^{\infty}(\boldsymbol{u}_{j}+\Xi_{j})\ \subset\ \mathbb{R}^{\nu}. (1.1) The Boolean model is a basic model in stereology and stochastic geometry, see [ 22 ] . A rigorous definition of ( 1.1 ) and random set are given later. A closely related random grain (RG) model is defined as superposition of the indicator functions of grains \boldsymbol{u}_{j}+\Xi_{j} , viz., \displaystyle X(\boldsymbol{t})=\sum_{j=1}^{\infty}\mathbb{I}(\boldsymbol{t}\in(\boldsymbol{u}_{j}+\Xi_{j})),\quad\boldsymbol{t}\in\mathbb{R}^{\nu}. (1.2) The Boolean set in ( 1.1 ) can be identified with its indicator function \widehat{X}(\boldsymbol{t}):=\mathbb{I}(\boldsymbol{t}\in{\mathcal{X}}),\boldsymbol{t}\in\mathbb{R}^{\nu} or a simple nonlinear transformation of the linear RF in ( 1.2 ) \displaystyle\widehat{X}(\boldsymbol{t})=X(\boldsymbol{t})\wedge 1, (1.3) where a\wedge b=\min(a,b) for real values a,b . The basic assumption guaranteeing the convergence of ( 1.2 ) is \displaystyle\mu:=\mathrm{E}\,{\rm Leb}_{\nu}(\Xi)<\infty. (1.4) In this paper, a random closed set satisfying ( 1.4 ) is called a random grain and the RG model in ( 1.2 ) is well-defined. It has marginal Poisson distribution with mean \mu and a nonnegative covariance function \displaystyle{\rm Cov}(X(\boldsymbol{0}),X(\boldsymbol{t})) \displaystyle= \displaystyle\mathrm{E}{\rm Leb}_{\nu}(\Xi\cap(\Xi-\boldsymbol{t}))\geq 0,\quad\boldsymbol{t}\in\mathbb{R}^{\nu}. (1.5) The volume fraction of the stationary Boolean set \mathcal{X} in ( 1.1 ) is the mean of the ‘volume’ of \mathcal{X} in the unit ‘cube’ ]\boldsymbol{0},{\bf 1}] : \displaystyle p:=\mathrm{E}{\rm Leb}_{\nu}(\mathcal{X}\cap]\boldsymbol{0},{\bf 1}])=\int_{]\boldsymbol{0},{\bf 1}]}\mathrm{E}\widehat{X}(\boldsymbol{t})\mathrm{d}\boldsymbol{t}=\mathrm{E}\widehat{X}(\boldsymbol{0})=1-\mathrm{P}(X(\boldsymbol{0})=0), (1.6) leading to \displaystyle p=1-\mathrm{e}^{-\mu}. (1.7) The volume fraction is the most important parameter of a Boolean set, the analog of the mean (expectation) of a stationary process on \mathbb{R}^{\nu} . By stationarity, p=\mathrm{E}{\rm Leb}_{\nu}(\mathcal{X}\cap A)/{\rm Leb}_{\nu}(A) for any Borel set A,0<{\rm Leb}_{\nu}(A)<\infty . The natural estimator of p from observations of {\mathcal{X}} on a (large) ‘inflated’ set \lambda A\subset\mathbb{R}^{\nu} is the ratio \displaystyle\widehat{p}_{\lambda,A} \displaystyle:= \displaystyle\frac{\widehat{X}_{\lambda}(A)}{{\rm Leb}_{\nu}(\lambda A)}, (1.8) called the sample volume fraction , where \widehat{X}_{\lambda}(A):={\rm Leb}_{\nu}({\mathcal{X}}\cap\lambda A) is the ‘volume’ of the intersection of the Boolean set with \lambda A . Then \displaystyle\lambda^{\nu}(\widehat{p}_{\lambda,A}-p) \displaystyle= \displaystyle\frac{\widehat{X}_{\lambda}(A)-\mathrm{E}\widehat{X}_{\lambda}(A)}{{\rm Leb}_{\nu}(A)} (1.9) and finding the limit distribution of \widehat{p}_{\lambda,A} reduces to that of the numerator of the last fraction. A stationary RF Y=\{Y(\boldsymbol{t});\boldsymbol{t}\in\mathbb{R}^{\nu}\} with finite variance and covariance r_{Y}(\boldsymbol{t})={\rm Cov}(Y(\boldsymbol{0}), Y(\boldsymbol{t})) is said long-range dependent (LRD) if \int_{\mathbb{R}^{\nu}}|r_{Y}(\boldsymbol{t})|\mathrm{d}\boldsymbol{t}=\infty and short-range dependent (SRD) if \int_{\mathbb{R}^{\nu}}|r_{Y}(\boldsymbol{t})|\mathrm{d}\boldsymbol{t}<\infty,\int_{\mathbb{R}^{\nu}}r_{Y}(\boldsymbol{t})\mathrm{d}\boldsymbol{t}\neq 0 . For RG RF with covariance in ( 1.5 ) we see that \displaystyle\int_{\mathbb{R}^{\nu}}{\rm Cov}(X(\boldsymbol{0}),X(\boldsymbol{t}))\mathrm{d}\boldsymbol{t} \displaystyle= \displaystyle\mathrm{E}\int_{\mathbb{R}^{\nu}}\mathbb{I}(\boldsymbol{s}\in\Xi)\mathrm{d}\boldsymbol{s}\int_{\mathbb{R}^{\nu}}\mathbb{I}(\boldsymbol{s}+\boldsymbol{t}\in\Xi)\mathrm{d}\boldsymbol{t} \displaystyle= \displaystyle\mathrm{E}{\rm Leb}_{\nu}(\Xi)^{2}. Therefore, a RG model in ( 1.2 ),( 1.4 ) is LRD if \mathrm{E}{\rm Leb}_{\nu}(\Xi)^{2}=\infty and SRD if \mathrm{E}{\rm Leb}_{\nu}(\Xi)^{2}<\infty . Clearly, a RG model is LRD if {\rm Leb}_{\nu}(\Xi) has a regularly decaying tail as \displaystyle\mathrm{P}({\rm Leb}_{\nu}(\Xi)>x)\ \sim\ c_{\Xi}\,x^{-\alpha},\quad x\to\infty\qquad(\exists\ c_{\Xi}>0,1<\alpha<2) (1.10) In most of the literature on LRD RG models [ 11 , 6 , 7 , 23 ] , it is assumed that randomness of \Xi is due to dilation of a deterministic set \Xi^{0} by a random factor R^{1/\nu} , viz., \displaystyle\Xi=R^{1/\nu}\,\Xi^{0}, (1.11) where R>0 is a r.v. with regularly decaying \alpha -tail, \alpha\in(1,2) . For \Xi in ( 1.11 ), the results in [ 23 ] imply that the sample volume fraction in ( 1.8 ) has an asymmetric \alpha -stable limit distribution, for arbitrary bounded Borel set A . ( 1.11 ) comprise a very special class of random set with all grains homothetic to each other. The present paper extends the result in [ 23 ] to much more general grain class. One of the major results of this work is Theorem 1 which says that condition ( 1.10 ) together with \displaystyle\mathrm{E}{\rm Leb}_{\nu}\big(\Xi\cap\{|\boldsymbol{t}|>\lambda\}\big) \displaystyle= \displaystyle o(\lambda^{(1-\alpha)\nu/\alpha}),\qquad\lambda\to\infty (1.12) imply that the sample volume fraction \widehat{p}_{\lambda,A} has an \alpha -stable limit distribution. We note that conditions ( 1.10 ) and ( 1.12 ) involve the Lebesgue measure of \Xi and \Xi\cap\{|\boldsymbol{t}|>\lambda\} alone and do not impose any structural assumptions on the Boolean set in contrast to ( 1.11 ); moreover, the sufficient condition ( 1.12 ) is sharp in the sense that the exponent (1-\alpha)\nu/\alpha cannot be improved in general. Theorem 1 refers to LRD RG model; in the SRD case \mathrm{E}{\rm Leb}_{\nu}(\Xi)^{2}<\infty we prove the CLT for the sample volume fraction (Theorem 2 ) without any additional conditions. We note that Gaussian limits for estimators of p were obtained in [ 4 , 17 , 18 ] and other works under more stringent assumptions on \Xi and the observation set. Note also that excursion sets of RG RF are given for any k=1,2,\ldots and u\in[k-1,k) by \{\boldsymbol{t};X(\boldsymbol{t})>u\}=\{\boldsymbol{t};X(\boldsymbol{t})\geq k\}, and the Boolean set {\mathcal{X}} corresponds to the excursion set of level u\in[0,1) . There is a growing interest in the study of the mean geometry of excursion sets of random fields in view of their links with extremal properties [ 1 , 2 ] . Lots of work concern Gaussian stationary RF but there are also results concerning shot noise random fields [ 5 ] or more generally, infinitely divisible random fields. Especially, CLT for the excursion set volumes have been investigated in [ 9 ] . Moreover a new notion of SRD, based on excursion sets, has been introduced in [ 13 ] , which is the only invariant with respect to monotone transformations of the marginal. More precisely a measurable stationary field Y is said SRD if \int_{\mathbb{R}^{\nu}}\int_{\mathbb{R}}\int_{\mathbb{R}}{\rm Cov}(\mathbb{I}(Y(\boldsymbol{0})>u),\mathbb{I}(Y(\boldsymbol{t})>v))\mathrm{d}\mu(u)\mathrm{d}\mu(v)\mathrm{d}\boldsymbol{t}<+\infty, (1.13) for all probability measures \mu on \mathbb{R} . A sufficient condition for infinitely divisible fields to satisfy ( 1.13 ) is also given in [ 16 ] . That motivates the more general study of the boolean fields given

Read the complete text