Polynomial parallel volume, convexity and contact distributions of random sets (Q2492621)

Modern data frequently arise as images of random structures in the \(d\)-dimensional Euclidean space \(\mathbb R^d\), \(d\geq 1\). It is of the main purposes of stochastic geometry to provide models for such random data. The basic, most flexible and frequently used model is still the Boolean model. A stationary Boolean model \(Z\) in \(\mathbb R^d\) is a random closed set \(Z=\bigcup_{n\in \mathbb N}(Z_n+\xi_n)\) where \(\xi_n\) form a stationary Poisson process \(\Xi\) in \(\mathbb R^d\) with intensity \(\gamma>0\); and the grains \(Z_1\), \(Z_2\),... are independent, identically distributed nonempty random compact sets, which are also independent of \(\Xi\). A random closed set in the sense of Matheron is a measurable map into the space of closed subsets of \(\mathbb R^d\). In particular, a random compact set is a random closed set which is almost surely compact. The distribution of the Boolean model \(Z\) is determined by \(\gamma\) and the distribution of the typical grain \(X\), a random compact set having distribution of the \(Z_i\). In order to fit a Boolean model to given data, the statistical problem consists in finding appropriate estimates for these two parameters. A simple yet powerful set of tools which is available with most image analyzing equipment is given by the contact distribution functions, see the authors [in: Statistical physics and spatial statistics. Lect. Notes Physics 600, 317--357 (2002)] for a recent survey. For a compact convex set \(B\subset \mathbb R^d\) containing the origin 0, the contact distribution function \(H_B\) of \(Z\) (with structuring element \(B\)) is defined as the distribution function of the ``\(B\)-distance'' \(d_B(0, Z)\) from 0 to \(Z\), given that 0 is not covered by \(Z\): \(H_B(r) = {\mathbb P}(d_B(0, Z)\leq r| 0\notin Z)\), \(r\geq 0\), with \(d_B(x, Z) =\inf\{t\geq 0: (x+tB)\cap Z\neq\emptyset\}\), \(x\in \mathbb R^d\). As a consequence of the Poisson properties of the Boolean model \(Z\) one easily gets \(H_B(r)= 1-\exp\{- \gamma E[V_d(X+rB^*) - V_d(X)]\}\), where \(E\) denotes mathematical expectation, \(V_d\) is the volume in \(\mathbb R^d\), and \(X+ r B^*\) is the Minkowski sum of the random compact set \(X\) and the reflection of \(rB\) in the origin. At this stage, a commonly made assumption is that the grains are (almost surely) convex. For a Boolean model with convex grains the contact distribution function has the form \[ H_B(r) =1 -\exp\left\{-\sum_{k=1}^d r^k \binom{d}{k}\gamma \overline{V}_{d-k, B}\right\} \] with mean values (densities) \(\overline{V}_{d-k, B} = EV(X[d-k], B^*[k])\) of the mixed volumes of the grains. Popular choices for \(B\) are the unit ball \(B^d\) (spherical contact distribution) or a unit segment \([0,u]\) with fixed or varying direction \(u\) (a linear contact distribution). While \(-\ln(1-H_{B^d})\) is a polynomial of order \(d\) (if \(X\) is almost surely convex) with \(\gamma V_d(B^d)\) as the leading coefficient and with quermass densities of the grains as the other coefficients, \(-\ln(1-H_{[0,u]})\) is a linear function and the slope is given by \(\gamma\) times the mean grain projection orthogonal to \(u\). The aim of the paper is to explore and clarify the connection between the polynomial behavior of logarithmic contact distribution functions of a Boolean model and the convexity of the grains. The authors prove that the polynomial behavior (or, more precisely, the linearity) of \(-\ln(1-H_{[0,u]})\) does in fact imply that the grains are convex, if either \(Z\) is isotropic or if we average over all directions \(u\). A similar result holds for the disc contact distributions \(H_B\) with certain two-dimensional convex bodies \(B\). But a corresponding result is not valid, for example, if \(B\) is a ball of dimension at least three. The average logarithmic linear contact distribution (ALLC) function \(L\) of a stationary Boolean model \(Z\) is given by \[ L(r) = - \int_{S^{d-1}} \ln(1-H_{[0,u]}(r)) \sigma(du), \quad r\geq0, \] where \(\sigma\) is the invariant probability measure on the unit sphere \(S^{d-1}\). The average logarithmic disc contact distribution (ALDC) function \(D_B\) of \(Z\) (with respect to \(B\)) is defined as \[ D_B(r) = -\int_{\text{SO}_d} \ln(1- H_{\theta B}(r))\, \nu(d\theta), \quad r\geq0, \] where \(\nu\) is the Haar probability measure on the rotation group SO\(_d\). The paper is organized as follows: Section 2 recalls some facts and tools from geometry. Section 3 is devoted to the study of dilatations by random segments. Section 4 is devoted to the investigation of dilatations by random disc bodies. In Section 5 using definitions of ALLC- and ALDC-functions the authors derive conditions for almost sure convexity of a typical grain.