betaFeedback

WorksMathematics

Testing isotropy of random fields with values in abstract spaces

Abstract

We study the problem of testing isotropy, i.e., distributional invariance under orthogonal transformations, for random fields taking values in abstract spaces. We propose a novel approach based on computing local directional characteristics, which are uniformly distributed under the assumption of isotropy. Our work is motivated by an application in material science, where we are testing the isotropy of function-valued porosity profiles of paper sheets, whose morphology is resolved by μμ-CT imaging. While considerably more general, the presented methodology is competitive with existing isotropy tests for real-valued random fields.

Introduction

Motivation

A random field is stationary if its distribution is shift-invariant, and it is isotropic if its distribution is invariant under any orthogonal transformation (rotation or reflection) at the origin. If the random field is both, stationary and isotropic, we call it motion-invariant. For the analysis of spatial data, the assumption of motion-invariance allows for strong simplifications [10]. In particular, the covariance between two locations depends only on their Euclidean distance. Consequently, several statistical tests have been proposed in the literature to decide whether the assumption of isotropy for random fields is justified by the data. See Weller and Hoeting [48] for an overview of nonparametric tests. In recent years, further tests have been developed, including tests based on the excursion set of a given random field [1, 3], spectral domain tests [14] (Weller and Hoeting 2020), and also isotropy tests on the sphere [42].

A critical gap is that existing literature addresses primarily real-valued random fields. While multivariate extensions are scarce (see e.g. Jona-Lasinio [23]) there is currently a lack of methodology for high-dimensional or non-Euclidean data. Driven by an application in which we analyze local porosity profiles in paper-based materials represented as point-wise functional data, we aim to bridge this gap by introducing an isotropy test designed for random fields with values in abstract spaces. The analysis of local porosity in paper-based materials is of particular interest for industrial applications because it influences effective physical properties such as air permeance [52, 27]. In general, porosity profiles are frequently used descriptors for analyzing the morphology of many functional materials, such as batteries [19, 12], fuel cells [4], or cement [41].

Also in the context of materials science, SO(3)SO(3)-valued random fields provide another instance of non-Euclidean random fields, emerging naturally in the 3D imaging of polycrystalline materials [25, 24, 43, 40]. In this context, the values of the random field represent the crystallographic orientations of the material in space. More generally, tensor- and spinor-valued random fields provide an appropriate mathematical framework to model phenomena at the interface between materials science and continuum physics, as well as in cosmology [30].

Dataset

The methodology in this paper was developed to test for isotropy of local volume fractions in paper sheets compressed by hard-nip calendering. Local volume fractions are computed based on experimental image data. For this purpose, we consider one 3D image of the data set measured at the PETRA III storage-ring of DESY in Hamburg [18, 50] described in Neumann et al. [35]. This material has been statistically investigated to quantify interdependencies between local morphological descriptors Neumann et al., [33] and to elucidate the impact of morphology on air permeance [16].

A 3D rendering is provided in Figure 1a. The total image consists of 1920×1111×1341920 \times1111 \times134 cubic voxels with an edge length of px=1.3 μm\mathrm{px} = 1.3\,\mu\mathrm{m}, where each voxel is either labeled as pore or solid. Note that the solid phase is the union of cellulose fibers. The images are smoothed in horizontal direction with a 2D Gaussian kernel with standard deviation σG=13 μm\sigma_{\mathrm{G}} = 13\,\mu\mathrm{m}. Doing so, we obtain an image, where each voxel contains a number between 0 and 1 representing the local volume fraction of cellulose material.

3D rendering of the microstructure and definition of functional observations

Figure 1. 3D rendering of the microstructure of a compressed paper sheet, for which we test isotropy of local volume fractions (a), functional observations defined as volume fractions read in thickness direction (b), where the zz-direction plays the role of tt in (1).

As evident from the 3D rendering in Figure 1a, the fibers are concentrated in layers parallel to the surface. When assessing isotropy, we are interested in whether the distribution is invariant under plane-preserving transformations, i.e., rotations or reflections of the paper that do not involve tilting it. Hence, we interpret this 3-dimensional image as a realization of a function-valued field on a 2-dimensional domain D\mathcal{D},

ξ={ξ(x):x∈D},ξ(x)={ξ(x)(t):t∈[0,T]},(1)\xi= \{\xi(x) : x \in\mathcal{D}\}, \quad\xi(x) = \{\xi(x)(t): t \in[0,T]\}, \tag*{(1)}

where the tt corresponds to the position in through-plane direction observed at some 134 equidistant points and ξ(x)\xi(x) represents the function of local volume fraction through the paper sheet at position xx. This interpretation is illustrated in Figure 1b.

Testing for isotropy of this function-valued field corresponds to assessing the planar isotropy of the material at microscale. For the particular example of function-valued random fields, it is also possible to apply a test developed for spatio-temporal random fields by Li et al. (2007). However, the latter test assumes stationarity in time, which is not required in the framework developed in the present work and not given in the dataset of interest.

Outline of this paper

The core idea of our work is to recast the problem of testing for isotropy as a problem of testing for uniform distribution in directions. The only structural ingredient we require from the observation space is a pairwise dissimilarity function, formalized by a semi-metric. In normed spaces, this would naturally be the norm of the difference between points.

For each sampling location, we measure the dissimilarity between all points that lie at a fixed distance and record the direction associated with the greatest dissimilarity. We will show that under isotropy, these directions are uniformly distributed. For certain choices of the semi-metric, this procedure has conceptual parallels to methods comparing directional autocovariance, e.g., Guan et al. [17].

The remainder of the paper is structured as follows: In § 2, we establish the methodological framework of the proposed hypothesis test. In § 3, we address aspects concerning practical implementation. This includes the estimation of distributional parameters for the test statistic, the choice of hyperparameters, and an outline of the algorithmic procedure. Note that an accompanying R package is also introduced here. § 4 presents an empirical finite-sample power analysis in the scalar-valued case, drawing comparisons to the methods by Guan et al. [17] and Weller and Hoeting (2020). Subsequently, § 5 analyzes the test in the context of a function-valued random field setting, which arises from assessing planar isotropy of the microstructure in paper-based materials. We apply the procedure to a synthetic data set generated from a stochastic fiber model as well as on the real data described in § 1.2. Finally, § 6 contains concluding remarks and outlines possible future directions. Theoretical justifications of our method are deferred to the Appendix.

Notation and setup

The object of interest for this paper is a strictly stationary random field ξ={ξ(x):x∈D}\xi= \{\xi(x): x \in D\} taking values in some measurable space V\mathcal{V} and domain D⊂ZdD \subset\mathbb{Z}^{d}, with sample size n(D)n(D). Without loss of generality, we assume that the grid has unit spacing b=1b = 1.

Our methodology can be extended to continuous domains or data observed on irregular spatial locations, but we refrain from developing all three formulations in parallel and leave this open for future works, as our data are sampled on a grid and our algorithms operate in the discrete setting as well.

The principal target of this work is to develop a statistical test for the hypothesis

H0:ξ is isotropic.H_0: \xi\text{ is isotropic}.

Recall that isotropy means that the law of the random field is invariant under rotations and reflections, i.e., if x1,…,xk∈Dx_1, \ldots, x_k \in D and if QQ is an element of the orthogonal group O(d)O(d) such that Qx1,…,Qxk∈DQx_1,\ldots,Qx_k \in\mathcal{D}, we have equality of the joint laws, i.e.

(ξ(x1),…,ξ(xk))=L(ξ(Qx1),…,ξ(Qxk)).\left(\xi(x_1),\ldots,\xi(x_k)\right) \overset{\mathcal{L}}{=} \left(\xi(Qx_1),\ldots,\xi(Qx_k)\right).

Along with the required stationarity, the field is said to be motion invariant.

In order to define our test statistic we transform the process ξ(x)\xi(x) into a new field taking values on H(x)(r)\mathbb{H}^{(x)}(r), a set of points in D\mathcal{D} with Euclidean distance rr from xx, i.e. H(x)(r)⊆Sxd−1(r)∩D\mathbb{H}^{(x)}(r) \subseteq\mathbb{S}^{d-1}_x(r) \cap\mathcal{D} where Sxd−1(r)\mathbb{S}^{d-1}_x(r) denotes the hypersphere in Rd\mathbb{R}^d with radius rr and center xx. We suppose that (V,d)(\mathcal{V},d) is an abstract space equipped with a semi-metric d:V×V→[0,∞)d:\mathcal{V} \times\mathcal{V} \to[0,\infty) which is symmetric in its arguments and satisfies d(v,w)=0d(v,w)=0 if and only if v=wv=w [51]. Now we define

M(x,r):=arg⁡max⁡y∈H(x)(r)d(ξ(y),ξ(x))andM0(x,r):=M(x,r)−x,(2)M(x,r) := \mathop{\arg\max}_{y \in\mathbb{H}^{(x)}(r)} d(\xi(y),\xi(x)) \quad\text{and}\quad M_0(x,r) := M(x,r)-x, \tag*{(2)}

which means that M(x,r)M(x,r) identifies the point in H(x)(r)\mathbb{H}^{(x)}(r) at which the field exhibits the largest dissimilarity to ξ(x)\xi(x) within the set of candidates, and M0(x,r)M_0(x,r) identifies the lag corresponding to M(x,r)M(x,r).

If the maximizer in (2) is not unique, we will break ties with a uniform random variable on the set of maximizers, which is independent of everything else. The tiebreakers for different xx are also supposed to be mutually independent.

We fix a basis subset of lags at length rr given by H(0)(r)={h1,…,hm}⊂Zd∩S0d−1(r)\mathbb{H}^{(0)}(r)=\{h_1,\ldots,h_m\}\subset\mathbb{Z}^d\cap\mathbb{S}^{d-1}_0(r), and suppose H(x)(r)=x+H(0)(r)\mathbb{H}^{(x)}(r)=x+\mathbb{H}^{(0)}(r) whenever all required neighbors of xx are observed. Further, we denote by D0\mathcal{D}_0 the set of such x∈Dx\in\mathcal{D}, where x+H(0)(r)⊂Dx+\mathbb{H}^{(0)}(r)\subset\mathcal{D}. M0(x,r)M_0(x,r) is thus a field taking values in H(0)(r)\mathbb{H}^{(0)}(r). Examples of the sets H(0)(r)\mathbb{H}^{(0)}(r) as well as D\mathcal{D} and D0\mathcal{D}_0 are illustrated in Figure 2.

Illustration of the sets $\mathcal{D}$, $\mathcal{D}_0$, and $\mathbb{H}^{(x)}(r)$

Figure 2. Illustration of the sets D\mathcal{D}, D0\mathcal{D}_0, and H(x)(r)\mathbb{H}^{(x)}(r): The points for which the random field is observed (D\mathcal{D}) are drawn in gray and black. The points feasible for further analysis (D0\mathcal{D}_0) are shown in black. Moreover, the set of neighbors used in maximizing dissimilarity H(x)(r)\mathbb{H}^{(x)}(r) are illustrated on circles. The arrows indicate lags H(0)(r)\mathbb{H}^{(0)}(r).

From now on, we suppose rr is fixed, and write M0(x)M_0(x) and H(0)\mathbb{H}^{(0)} instead of M0(x,r)M_0(x,r) and H(0)(r)\mathbb{H}^{(0)}(r).

Methodology

Formulation of the test on isotropy

The key idea behind our method is that under motion invariance, for any fixed point xx in the domain, M0(x)M_0(x) has a uniform distribution on H(0)\mathbb{H}^{(0)} provided the following notion applies to H(0)\mathbb{H}^{(0)}:

Definition 2.1. A non-empty set H⊂Zd\mathbb{H}\subset\mathbb{Z}^d is called transitive under its orthogonal symmetry group if there is a generating element h∈Hh\in\mathbb{H} so that for every h~∈H\tilde{h}\in\mathbb{H}, there exists an orthogonal transformation Q∈O(d)Q\in O(d) with Qh=h~Qh=\tilde{h} and H\mathbb{H} invariant under QQ, i.e. QH=HQ\mathbb{H}=\mathbb{H}.

This definition is closely related to the algebraic concept of orbits. Specifically, H\mathbb{H} is a transitive orbit of a finite subgroup GHG_{\mathbb{H}} of O(d)O(d), namely that, which leaves H\mathbb{H} invariant. In the following, we write A′A' for the transpose of A∈Rn×mA\in\mathbb{R}^{n\times m} with m,n≥1m,n\geq1.

Example 2.1. Sets H⊂Z2\mathbb{H}\subset\mathbb{Z}^2 that are transitive under their respective orthogonal symmetry groups include the following:

Ha={(±2,±1),(±1,±2)},Hb={(±1,±1)},\mathbb{H}_a=\{(\pm2,\pm1),(\pm1,\pm2)\}, \qquad \mathbb{H}_b=\{(\pm1,\pm1)\},
Hc={(±1,0),(0,±1)},Hd={(1,0),(0,1)}.\mathbb{H}_c=\{(\pm1,0),(0,\pm1)\}, \qquad \mathbb{H}_d=\{(1,0),(0,1)\}.

where (±a,±b)(\pm a,\pm b) denotes the four possible sign configurations. For H⊂Zd\mathbb{H}\subset\mathbb{Z}^d a general construction with n(H)=2dd!n(\mathbb{H})=2^d d!, is the set generated by h=(1,…,d)′h=(1,\ldots,d)':

H={(±τ(1),…,±τ(d)):τ is a permutation of {1,…,d}}.\mathbb{H}=\{(\pm\tau(1),\ldots,\pm\tau(d)):\tau\text{ is a permutation of }\{1,\ldots,d\}\}.

Furthermore, all sets H={h1,h2}\mathbb{H}=\{h_1,h_2\} with ∥h1∥=∥h2∥\lVert h_1\rVert=\lVert h_2\rVert are transitive under their respective orthogonal symmetry (by reflecting along h1+h2h_1+h_2).

Example 2.2. Note that symmetry and equal distance from the origin are not sufficient. E.g., the set

H={(±3,±4),(±4,±3),±(5,0),±(0,5)},\mathbb{H} = \{(\pm3,\pm4),(\pm4,\pm3),\pm(5,0),\pm(0,5)\},

is not transitive under its orthogonal symmetry group.

Some necessary conditions for H(0)\mathbb{H}^{(0)} to be transitive under its orthogonal symmetry group are discussed in § 8.1 of the Appendix.

The following lemma establishes uniform distribution of M0(x)M_0(x) under isotropy given the above condition holds for H(0)\mathbb{H}^{(0)}. If the condition fails, then M0(x)M_0(x) will generally not be uniformly distributed. In § 8.2 of the Appendix we give a counter-example.

Lemma 2.1. Let ξ\xi be a motion-invariant random field in Zd\mathbb{Z}^d and let H(0)\mathbb{H}^{(0)} be transitive under its orthogonal symmetry group. Then, for all x∈D0x \in\mathcal{D}_0 we have that M0(x)M_0(x) is uniformly distributed on H(0)\mathbb{H}^{(0)}.

We will now develop a test which assesses the new null-hypothesis H0′⊃H0H'_0 \supset H_0 given by

H0′:M0(x) is uniformly distributed on H(0).H'_0: M_0(x) \text{ is uniformly distributed on } \mathbb{H}^{(0)}.

A more detailed discussion regarding the relationship between H0H_0 and H0′H'_0 is deferred to § 2.2 below. We test uniform distribution of M0M_0 by adapting Pearson’s χ2\chi^2-test for spatially dependent data. To this end, we transform the categorical variables M0(x)M_0(x) into a dummy vector I(x)\mathcal{I}(x), whose ii-th component is equal to 1 if M0(x)=hiM_0(x)=h_i. We denote 1m:=(1,…,1)′∈Rm\mathbf{1}_m := (1,\ldots,1)^\prime\in\mathbb{R}^m. Define

S:=1n(D0)∑x∈D0[I(x)−1m1m].(3)S := \frac{1}{\sqrt{n(\mathcal{D}_0)}}\sum_{x\in\mathcal{D}_0}\left[\mathcal{I}(x)-\frac{1}{m}\mathbf{1}_m\right]. \tag*{(3)}

Let Γ^\widehat{\Gamma} be some positive semi-definite estimator of the long-run covariance matrix given by

Γ=∑h∈ZdCov⁡(I(0),I(h))\Gamma=\sum_{h\in\mathbb{Z}^d}\operatorname{Cov}(\mathcal{I}(0),\mathcal{I}(h))

(see § 3), and suppose F^\widehat{F} denotes the weighted χ2\chi^2 distribution with weights given by the eigenvalues of Γ^\widehat{\Gamma}. Define T:=∥S∥2T:=\lVert S\rVert^2. Our test for isotropy is based on the following decision rule:

Reject H0′ if T>q1−α(F^),(4)\text{Reject } H'_0 \text{ if } T>q_{1-\alpha}(\widehat{F}), \tag*{(4)}

where qα(.)q_{\alpha}(.) is the α\alpha-quantile.

Before proceeding, a few remarks are in order:

Remark 2.1 (Asymptotic validity). Conditions which guarantee that this is a consistent and asymptotically valid level-α\alpha test will be given in the Appendix.

Remark 2.2 (Permuting labels). If we require Γ^\widehat{\Gamma} to be affine equivariant, then the test statistic TT is also affine equivariant. Particularly, this implies that permuting the labels in M0(x)M_0(x) (which is replacing I(x)\mathcal{I}(x) by PI(x)P\mathcal{I}(x) with a suitable permutation matrix PP) does not change TT.

Remark 2.3 (Multiple sets of lags). While the above methodology only derives a decision rule for rejecting H0′H'_0 based on investigating uniform distribution along a single set H(0)(r)\mathbb{H}^{(0)}(r), one may consider multiple such sets H(0)(rj)\mathbb{H}^{(0)}(r_j), j=1,…,Jj=1,\ldots,J. Denote their respective cardinalities by mjm_j and the corresponding dummy variables by Ij\mathcal{I}_j. Stacking these dummy variables to obtain I⋆:=(I1′,…,IJ′)′\mathcal{I}^{\star}:=(\mathcal{I}'_1,\ldots,\mathcal{I}'_J)' and S⋆S^{\star} through centering by (1m11m1′,…,1mJ1mJ′)′(\frac{1}{m_1}\mathbf{1}'_{m_1},\ldots,\frac{1}{m_J}\mathbf{1}'_{m_J})' gives an analogous decision rule.

Remark 2.4 (Testing for reflection symmetry). This work focuses on testing for isotropy, i.e., distributional invariance under all orthogonal transformations. However, there are also weaker formulations, like axial, diagonal, or complete symmetry, which limit to invariance under certain reflections only, see Lu and Zimmerman (2005) for a hypothesis test of second-order symmetry characteristics. Our approach can be extended to such weaker formulations, given that H(0)\mathbb{H}^{(0)} is transitive under its orthogonal symmetry group limited to the respective reflectional transformations. Similarly, one may test for rotational symmetry only.

Relationship between H0H_0 and H0′H'_0

We have seen that H0⊂H0′H_0 \subset H'_0 and for the latter hypothesis, we have developed a statistical test. In simplified settings it is possible to show that the implication can be reversed, that is H0′⊂H0H'_0 \subset H_0 holds. Evidently, this will depend on the choice of dd and H(0)\mathbb{H}^{(0)}. By the generality of our approach, we cannot expect universally applicable results. Instead, we illustrate the problem on the basis of an example:

Example 2.3. Let ξ\xi be a stationary univariate Gaussian random field on Z2\mathbb{Z}^2 with geometric anisotropy, meaning that its covariance function has the form

C(h):=Cov⁡(ξ(0),ξ(h))=ψ(h′Ah),C(h) := \operatorname{Cov}(\xi(0),\xi(h)) = \psi(h'Ah),

where ψ:[0,∞)↦[0,∞)\psi:[0,\infty)\mapsto[0,\infty) is strictly decreasing to zero and AA is a positive definite matrix. We set H(0):={h1,h2}\mathbb{H}^{(0)}:=\{h_1,h_2\} with ∥h1∥=∥h2∥\lVert h_1\rVert=\lVert h_2\rVert as described in Example 2.1. For x∈Z2x\in\mathbb{Z}^2, let M0(x)M_0(x) be the lag maximizing

d(ξ(x),ξ(x+hi)):=∣ξ(x+hi)−ξ(x)∣,i∈{1,2}.d(\xi(x),\xi(x+h_i)):=|\xi(x+h_i)-\xi(x)|,\quad i\in\{1,2\}.

A straightforward computation yields

P(M0(x)=h1)=12−1πarcsin⁡Corr⁡(ξ(h1)+ξ(h2)−2ξ(0),ξ(h1)−ξ(h2)).\mathbb{P}(M_0(x)=h_1)=\frac{1}{2}-\frac{1}{\pi}\arcsin\operatorname{Corr}(\xi(h_1)+\xi(h_2)-2\xi(0),\xi(h_1)-\xi(h_2)).

Observing that Cov⁡(U,V)=2(C(h2)−C(h1))\operatorname{Cov}(U,V)=2(C(h_2)-C(h_1)), we conclude that H0=H0′H_0=H'_0 unless C(h1)=C(h2)C(h_1)=C(h_2).

The example shows that in the Gaussian case, geometric anisotropy generally implies a non-uniform distribution of M0(x)M_0(x) even if H(0)\mathbb{H}^{(0)} contains only two lags. In the exceptional case where C(h1)=C(h2)C(h_1)=C(h_2), we can adapt the test by using a different set of lags, as evident in the following example. There, we extend the above observation to multivariate fields on higher-dimensional lattices, provided all components share the same geometric anisotropy.

Example 2.4. Let ξ\xi be a stationary Gaussian random field on Zd\mathbb{Z}^{d}, d≥2d \ge2, with values in V=Rp\mathbb{V} = \mathbb{R}^{p} so that its matrix-valued covariance function has the form

C(h)=Θdiag⁡(ψ1(h′Ah),…,ψp(h′Ah))Θ′,C(h) = \Theta\operatorname{diag}\left(\psi_{1}(h^{\prime}Ah),\ldots,\psi_{p}(h^{\prime}Ah)\right)\Theta^{\prime},

where Θ∈O(p)\Theta\in O(p), AA is a symmetric positive definite matrix and ψ1,…,ψp:[0,∞)→[0,∞)\psi_{1},\ldots,\psi_{p} : [0,\infty) \to[0,\infty) are strictly decreasing. Like in the previous example we choose H(0)={h1,h2}\mathbb{H}^{(0)} = \{h_{1},h_{2}\} and select the Euclidean distance for d\mathbf{d}. Then,

sign⁡(P(M0(x)=h1)−12)=sign⁡(h1′Ah1−h2′Ah2),\operatorname{sign}\left(\mathbb{P}\left(M_{0}(x)=h_{1}\right)-\frac{1}{2}\right)=\operatorname{sign}\left(h_{1}^{\prime}Ah_{1}-h_{2}^{\prime}Ah_{2}\right),

Furthermore, there exists a selection of lag sets H1(0),…,H(d−1)(d+2)/2(0)\mathbb{H}_{1}^{(0)},\ldots,\mathbb{H}_{(d-1)(d+2)/2}^{(0)}, each of cardinality two, so that M0(x)M_{0}(x) is uniformly distributed with respect to all lag sets if and only if AA is a positive multiple of the identity matrix, i.e. when ξ\xi is isotropic.

Remark 2.5. The construction of the field in Example 2.4 is related to independent component analysis (ICA), where multivariate observations are modeled as linear mixtures of mutually independent latent sources Hyvärinen et al. [20]. Extensions of this principle to random fields have been considered in the spatial blind source separation literature, see, e.g., Nordhausen et al. [37]. Our setting corresponds to a special case, where the mixture matrix is orthogonal.

The claims of the example as well as the construction of the lag sets H1(0),…,H(d−1)(d+2)/2(0)\mathbb{H}_{1}^{(0)},\ldots,\mathbb{H}_{(d-1)(d+2)/2}^{(0)} are further discussed in § 8.3 of the Appendix. It supports Remark 2.3: changing lags or adding lags to H(0)\mathbb{H}^{(0)} will make the test applicable against a bigger class of alternatives. In the example above, H0=H0′H_{0}=H_{0}^{\prime} if we require uniform distribution of M0(x)M_{0}(x) when finding maximal dissimilarity with respect to the lag sets H1(0),…,H(d−1)(d+2)/2(0)\mathbb{H}_{1}^{(0)},\ldots,\mathbb{H}_{(d-1)(d+2)/2}^{(0)} simultaneously.

Practical implementation

Estimation of the long-run covariance

Our test statistic is based on the mean of weakly dependent data multivariate, for which Pantle et al. [38] introduce a positive semi-definite long-run covariance estimator on a continuous domain. We apply this estimator to the lattice domain case similarly to Sherman [44] but extended to higher dimensions in a way that preserves mean-square consistency, which we prove in the Appendix.

To compute the estimator, the domain D0\mathcal{D}_{0} is divided into sub-domains, over which the empirical means and their empirical covariance operator are subsequently computed. To this end, select a basis sub-domain D⊂D0\mathcal{D} \subset\mathcal{D}_{0} (which is, e.g., rectangular) and let Dℓ:=D+sℓ\mathcal{D}_{\ell} := \mathcal{D} + s_{\ell}, ℓ=1,…,N\ell= 1,\ldots,N be pairwise disjoint translations of D\mathcal{D} that fully intersect with the domain D0\mathcal{D}_{0}.

Define

μ^ℓ=1n(D)∑x∈DℓI(x)−1m1m∈Rm,andμ‾:=1N∑ℓ=1Nμ^ℓ.\widehat{\mu}_{\ell} = \frac{1}{n(\mathcal{D})}\sum_{x \in\mathcal{D}_{\ell}} I(x)-\frac{1}{m}\mathbf{1}_{m} \in\mathbb{R}^{m}, \quad\text{and} \quad\overline{\mu}:=\frac{1}{N}\sum_{\ell=1}^{N}\widehat{\mu}_{\ell}.

The positive semi-definite long-run covariance estimator is then given by

Γ^=n(D)N−1∑ℓ=1N(μ^ℓ−μ‾)(μ^ℓ−μ‾)′.\widehat{\Gamma}=\frac{n(\mathcal{D})}{N-1}\sum_{\ell=1}^{N}\left(\widehat{\mu}_{\ell}-\overline{\mu}\right)\left(\widehat{\mu}_{\ell}-\overline{\mu}\right)^{\prime}.

For extensions to general weight functions and non-disjoint sub-domains, we refer to Pantle et al. [38].

User-selected parameters

The semi-metric: The primary requirement for dd is that it must reflect that algebraic or geometric properties of the sample space V\mathcal{V} and it should be capable of measuring the specific anisotropy of interest. For example, in the analysis of polycrystalline materials, quotient spaces SO(3)/GSO(3)/G of SO(3)SO(3) play an important role, where GG is a subgroup of SO(3)SO(3). In this case, a physically meaningful semi-metric is

d(v,w)=min⁡g∈Garccos⁡(tr⁡(v−1gw)−12),d(v,w)=\min_{g\in G}\arccos\left(\frac{\operatorname{tr}\left(v^{-1}gw\right)-1}{2}\right),

see Equation (6) in Karafiátová et al. [24]. Here we interpret vv, ww and gg as the matrix representation of the elements in SO(3)SO(3). For our data example, where V=L2([0,1])\mathcal{V}=L^2([0,1]), the Hilbert space of square integrable functions, we will use the standard metric d(v,w)=∥v−w∥d(v,w)=\lVert v-w\rVert, where ∥⋅∥\lVert\cdot\rVert is the L2L^2-norm for square integrable functions.

It is possible to combine tests for multiple semi-metrics into a single procedure, analogous to the treatment of multiple lags discussed below and in Remark 2.3.

The spatial lags: We need to choose a radius rr for the set of lags H(0)(r)\mathbb{H}^{(0)}(r). Obviously, r≥1r\geq1, and there must exist integers j1,…,jdj_1,\ldots,j_d such that r2=j12+⋯+jd2r^2=j_1^2+\cdots+j_d^2, in order that H(0)(r)≠∅\mathbb{H}^{(0)}(r)\ne\varnothing. Another practical limitation is the size of the domain. If the radius is large, the feasible domain D0D_0 becomes small.

A large radius may also cause vanishing spatial dependence between ξ(x)\xi(x) and ξ(x+h)\xi(x+h) leading to diminished statistical power. This effect can be seen in Example 2.3: if hih_i is replaced by rhirh_i then for r→∞r\to\infty, M0(x)M_0(x) becomes uniformly distributed. On the other hand, at small radii, depending on the data situation, the anisotropy may be only barely detectable.

So in practice the choice of rr is a tradeoff. In general we propose using multiple smaller lags, and potentially aggregating those into a single hypothesis test (avoiding the need for any a posteriori corrections, such as, e.g., Bonferroni or Šidák corrections), see Remark 2.3. This procedure is outlined in Algorithm 2.

For the case where ξ\xi is real-valued, as a rule of thumb, we suggest selecting rr such that Corr⁡(ξ(0),ξ(h))≈12\operatorname{Corr}(\xi(0),\xi(h))\approx\frac{1}{2} for ∥h∥=r\lVert h\rVert=r, taking r=1r=1 if the correlation is already below 12\frac{1}{2} at ∥h∥=1\lVert h\rVert=1. This can be determined using a direction-independent (isotropic) correlation estimator.

The basis sub-domain DD: The region DD must be chosen large enough to adequately cover the long-range dependence of the field. The specific choice of DD requires careful consideration, as it directly influences both the size and the power of the test. For a detailed derivation of the asymptotically optimal domain size in the two-dimensional setting, see Sherman [44].

Algorithm

In the following, we synthesize the concepts from the preceding subsections into a step-by-step computational procedure. Our algorithms are also available in form of an R package (Baumhakel et al., [5]).

Algorithm 1 outlines the testing procedure for a single set of lags and a single semi-metric. The computations in this basic algorithm are visually summarized in Figure 3, which illustrates the local computation of maximizing lags at the grid level, and global aggregation. Furthermore, Algorithm 2 details how this approach naturally extends to a combined test for multiple sets of lags and semi-metrics by appropriately stacking the test statistics. To determine the quantiles of the weighted χ2\chi^2 distribution in the final step of both algorithms, we suggest Davies’ method (Davies, [11]).

Illustration of computing dissimilarity-maximizing lags and comparing proportions to a uniform distribution

Figure 3. Illustration of Algorithm 1: Computing dissimilarity-maximizing lags at every grid point (pixel) and comparing proportions to uniform distribution. The example image is a realization of an isotropic Gaussian field.

Algorithm 1 Hypothesis test for isotropy

  1. Select distance rr, lags H(0)(r)\mathbb{H}^{(0)}(r), and semi-metric d\mathsf{d}

  2. Determine D0D_0, the feasible region for optimization

  3. For each x∈D0x \in D_0, find M0(x)M_0(x)

  4. Compute SS according to (3)

  5. Select basis sub-domain DD and non-overlapping shifted versions D1,…,DND_1,\ldots,D_N

  6. Compute Γ^\widehat{\Gamma} and its eigenvalues λ^1,…,λ^m\widehat{\lambda}_1,\ldots,\widehat{\lambda}_m

  7. Determine the (1−α)(1-\alpha)-quantile of a weighted χ2\chi^2 distribution with weights λ^1,…,λ^m\widehat{\lambda}_1,\ldots,\widehat{\lambda}_m

  8. Compute T=∥S∥2T=\lVert S\rVert^2 and compare it to the quantile

Algorithm 2 Combined hypothesis test for multiple lag sets and semi-metrics

  1. Select multiple sets of parameters (H(0)(r1),d1),…,(H(0)(rJ),dJ)(\mathbb{H}^{(0)}(r_1),\mathsf{d}_1),\ldots,(\mathbb{H}^{(0)}(r_J),\mathsf{d}_J)

  2. Determine the global feasible region D0D_0 via the intersection of all individual feasible sets

  3. for each parameter set j=1,…,Jj=1,\ldots,J do

  4. Compute SjS_j by executing Steps 3–4 of Algorithm 1

  5. end for

  6. Stack S1,…,SJS_1,\ldots,S_J to form the combined test statistic S∗S^*

  7. Execute Steps 5–8 of Algorithm 1 for S∗S^*

Finite sample behavior and power analysis

While the primary motivation of our work lies in accommodating high-dimensional and functional data objects, we will now evaluate its finite-sample performance in the scalar setting and compare to specialized tests. Several nonparametric tests for spatial isotropy have been proposed for scalar-valued fields (see e.g. Weller and Hoeting [48]), and we benchmark the empirical size and power of our approach against some of these established baselines.

To this end, we replicate the simulation experiment detailed by Weller and Hoeting (2020). Specifically, we compare our proposed test (UNI, Uniform Neighborhood Isotropy testing) to the spectral domain test developed by Weller and Hoeting (2020) (SIS) and the covariance-based test by Guan et al. [17] (GSC). Note that the GSC test also requires to select a set of spatial lags H(0)\mathbb{H}^{(0)} with equal lengths on which it tests for equality of spatial covariance as well as a subsampling window, while SIS does neither.

As data generating process we consider a centered, stationary Gaussian random field ξ={ξ(x):x∈D⊂Z2}\xi= \{\xi(x) : x \in D \subset\mathbb{Z}^2\} and draw realizations on a regular grid. The spatial dependence is governed by a Matérn covariance function, given by

C(h)=σ22κ−1Γ(κ)(∥A(θ,R)h∥ϕ)κKκ(∥A(θ,R)h∥ϕ),C(h) = \frac{\sigma^2}{2^{\kappa-1}\Gamma(\kappa)}\left(\frac{\lVert A(\theta,R)h\rVert}{\phi}\right)^\kappa K_\kappa\left(\frac{\lVert A(\theta,R)h\rVert}{\phi}\right),

where σ2>0\sigma^2 > 0 is the marginal variance, ϕ>0\phi> 0 is the range parameter, κ>0\kappa> 0 is a parameter related to smoothness, and KκK_\kappa is the modified Bessel function of the second kind of order κ\kappa. The linear transformation matrix A(θ,R)A(\theta,R) induces geometric anisotropy and is given by

A(θ,R)=(1001/R)(cos⁡θ−sin⁡θsin⁡θcos⁡θ),A(\theta,R) = \begin{pmatrix} 1 & 0 \\ 0 & 1/R \end{pmatrix} \begin{pmatrix} \cos\theta& -\sin\theta\\ \sin\theta& \cos\theta \end{pmatrix},

where θ∈[0,π/2)\theta\in[0,\pi/2) is the angle of the principal axis of anisotropy and R≥1R \ge1 is the anisotropy ratio. Starting out with an isotropic field, those parameters govern in which direction (θ\theta) and how much (RR) the field is stretched in order to obtain the anisotropic one.

Following the specifications of Weller and Hoeting (2020), we set the covariance parameters to σ2=1\sigma^2 = 1, 1/ϕ=1.521/\phi= 1.52, and κ=1\kappa= 1. The fields are generated and evaluated on different square grids to assess the effect of the sample size: 15×1515 \times15 up to 50×5050 \times50 in 5-step increments for each side length.

The null hypothesis of strict spatial isotropy corresponds to R=1R = 1, under which the angle θ\theta is unidentifiable (set to θ=0\theta= 0 for convenience). To evaluate the empirical power, we simulate fields under combinations of anisotropy ratios R∈{2,2}R \in\{\sqrt{2}, 2\} and rotation angles θ∈{π/8,π/4}\theta\in\{\pi/8,\pi/4\}.

We set GSC up to test for equality of autocovariance along the lags (1,0)(1,0) and (0,1)(0,1), as well as (1,1)(1,1) and (−1,1)(-1,1). We further choose subwindow dimensions for long-run covariance estimation as 2×22 \times2. Due to constraints in computation time, GSC is not computed for domains larger than 30×3030 \times30. Regarding UNI, we similarly test for uniform distribution along lags {(±1,0),(0,±1)}\{(\pm1,0),(0,\pm1)\}, and {(±1,±1)}\{(\pm1,\pm1)\}. For estimating the long-run covariance matrix, we set the subwindow size to 4×44 \times4 here. While SIS has no user-defined parameters, square grids lead to inflated size, which is why instead of computing the statistic on the full ℓ×ℓ\ell\times\ell grid, we compute it on ℓ×(ℓ−1)\ell\times(\ell- 1) data points only.

All tests are run in R ([39]). For GSC, the implementation of the package spTest ([47]) is used. The empirical size and power results at the nominal 0.05 significance level as well as computation times for selected sample sizes are summarized in Figure 4.

Figure 4. (a) Proposed test (UNI). (b) Weller and Hoeting (2020) (SIS). (c) Guan et al. [17] (GSC). (d) Average computation time in log⁡10(seconds)\log_{10}(\text{seconds}). Empirical size and power as well as average computation time for testing spatial isotropy at the α=0.05\alpha= 0.05 significance level. Results are based on 1000 Monte Carlo simulations for various anisotropy ratios (RR) and angles (θ\theta). Sample size is increased in steps of 5 units per direction, from a grid of size 15×1515 \times15 up to 50×5050 \times50. GSC is only computed up to side length 30 due to computation time. The corresponding sample size up to which GSC is computed is marked with a vertical gray line in each image.

In the investigated scalar setting, UNI and SIS exhibit similar statistical power and computation times. GSC, in contrast, achieves higher statistical power, but is substantially more expensive in computation: depending on sample size, its computation time is up to two orders of magnitude larger than that of UNI and SIS. This additional computational cost is mainly due to the more involved estimation of the long-run covariance. Moreover, both GSC and UNI require the selection of a specific set of lags and a subwindow size. SIS instead considers all lags simultaneously and avoids long-run covariance estimation via subsampling. However, this computational simplicity comes at a statistical cost, as SIS exhibits inflated test size for square grids. Overall, these results indicate that the proposed UNI test is competitive with existing methods even in the scalar setting, despite it being applicable much greater generality.

Application to fiber-based materials

Before applying our UNI procedure to the dataset, we investigate its size and power empirically on synthetic fiber data, via a stochastic fiber model. With the fiber model, we mimic porous structures of a similar type as the data.

Stochastic fiber model

We use a slightly modified version of the soft-core fiber model introduced in Section II.A of Altendorf and Jeulin [2]. The solid phase is modeled by a Boolean model Ξ\Xi, see Chapter 6 in Jeulin [22], with fibers as grains. For this purpose, let P1,P2,…P_1, P_2, \ldots be a homogeneous Poisson point process in R2×[0,K]\mathbb{R}^2 \times[0,K] with intensity η>0\eta> 0 [26]. Moreover, let Ξ1,Ξ2,…\Xi_1,\Xi_2,\ldots be a sequence of i.i.d. copies of the random compact set Ξ0\Xi_0, defined as follows. For a fixed nb∈Nn_b \in\mathbb{N}, we define a random sequence {p0,p1,…,pnb}\{p_0,p_1,\ldots,p_{n_b}\}, where p0=0p_0=0 is the origin in R3\mathbb{R}^3. Then

Ξ0=⋃i=0nb(pi+Brf(0)),(5)\Xi_0 = \bigcup_{i=0}^{n_b} \left(p_i + B_{r_f}(0)\right), \tag*{(5)}

where Brf(0)B_{r_f}(0) denotes the closed ball with radius rf>0r_f > 0 centered at the origin. Here, rf>0r_f > 0 is a model parameter that describes the thickness of the fibers. The random sequence p1,…,pnbp_1,\ldots,p_{n_b} models the medial axis of a fiber with radius rf>0r_f > 0. In the following, we describe the distribution of this sequence. First, the main orientation μ0∈S2\mu_0 \in S^2 is modeled by a Bingham distribution, see Section 9.4.3 in Mardia and Jupp [31], i.e. the probability density function fμ0:S2→[0,∞)f_{\mu_0}: S^2 \to[0,\infty) of μ0\mu_0 is given by

fμ0(x)=cBexp⁡(x′Bx),(6)f_{\mu_0}(x) = c_B \exp\left(x^\prime Bx\right), \tag*{(6)}

for each x∈S2x \in\mathbb{S}^{2}, where B∈R3×3B \in\mathbb{R}^{3 \times3}, positive semi-definite, is a model parameter and cB>0c_B > 0 denotes the normalizing constant. Then, we model p1=rfμ0/2p_1 = r_f\mu_0/2. The remaining p2,…,pnbp_2, \ldots, p_{n_b} are modeled by pj=pj−1+μj−1rf/2p_j = p_{j-1} + \mu_{j-1}r_f/2, where conditional on μ0\mu_0, the directions μ1,…,μnb−1\mu_1, \ldots, \mu_{n_b-1} form a Markov chain in S2\mathbb{S}^{2}. For each 2≤j≤nb2 \leq j \leq n_b, the conditional probability density function is

fμj∣μj−1(x)=∣κ1μ0+κ2μj−1∣2π(exp⁡(∣κ1μ0+κ2μj−1∣)−exp⁡(−∣κ1μ0+κ2μj−1∣))exp⁡(κ1μ0′x+κ2μj−1′x).f_{\mu_j \mid\mu_{j-1}}(x) = \frac{\lvert\kappa_1\mu_0 + \kappa_2\mu_{j-1}\rvert}{2\pi\left(\exp(\lvert\kappa_1\mu_0 + \kappa_2\mu_{j-1}\rvert) - \exp(-\lvert\kappa_1\mu_0 + \kappa_2\mu_{j-1}\rvert)\right)} \exp\left(\kappa_1\mu_0^{\prime}x + \kappa_2\mu_{j-1}^{\prime}x\right).

for x∈S2x \in\mathbb{S}^{2}. The parameters κ1,κ2>0\kappa_1, \kappa_2 > 0 are controlling the influence of the main direction μ0\mu_0 and the previous direction μj−1\mu_{j-1}. Note that μj\mu_j follows a von Mises-Fisher distribution conditioned on μj−1\mu_{j-1}, see Section 9.3.2 in Mardia and Jupp [31]. Finally, the entire sequence {0,p1,…,pnb}\{0, p_1, \ldots, p_{n_b}\} is rotated around the origin, such that the main direction of the fiber, which is pnb/∣pnb∣p_{n_b}/\lvert p_{n_b}\rvert, coincides with μ0\mu_0. The latter is described in detail in Section II.A of Altendorf and Jeulin [2]. With this definition of the medial axis, we have specified the model of individual fibers given in (5). The solid phase, which is the union of individual fibers, is then defined by

Ξ=⋃i=1∞(Pi+Ξi).(7)\Xi= \bigcup_{i=1}^{\infty} (P_i + \Xi_i). \tag*{(7)}

For the Boolean model in (7), the expected volume fraction is known, i.e.

1λ(W)E[λ(Ξ∩W)]=1−exp⁡(−ηE[λ(Ξ0)]),(8)\frac{1}{\lambda(W)}\mathbb{E}[\lambda(\Xi\cap W)] = 1 - \exp\left(-\eta\mathbb{E}[\lambda(\Xi_0)]\right), \tag*{(8)}

for each Borel set W⊂R3W \subset\mathbb{R}^{3}, see Equation (3.15) in Chiu et al. [9]. The resulting set Ξ\Xi is stationary and mm-dependent where m=2rf(nb+2)m = 2r_f(n_b + 2), see Equation (3.18) in Chiu et al. [9] and thus fulfills weak dependence constraints required for asymptotic validity of the test, discussed in the Appendix.

For the power study, we draw realizations of the fiber model that are similar to the data set introduced in § 1.2 and represent the microstructure of a compressed paper material. Due to the compression, the fiber cross sections have elliptical shape instead, which is why we consider the transformed random closed set

Ξζ={(x1,x2,x3)∈R3:(x1,x2,ζx3)∈Ξ}\Xi_{\zeta} = \{(x_1, x_2, x_3) \in\mathbb{R}^{3} : (x_1, x_2, \zeta x_3) \in\Xi\}

for a compression factor ζ=0.4\zeta= 0.4. Note that Ξζ\Xi_{\zeta} is again stationary and has the same volume fraction as Ξ\Xi.

The simulations are performed on a voxel grid, where the voxels are cubic with an edge length of px=6.5 μmp_x = 6.5\ \mu\mathrm{m}. For simplicity, we simulate from Ξ\Xi instead of Ξ0.4\Xi_{0.4} with fiber length L=140L = 140 px and cross-sectional radius rf=2.5r_f = 2.5 px. Finally, we adapt the edge lengths of the voxels accordingly to obtain Ξ0.4\Xi_{0.4}. Moreover, note that our simulations model the interior part of fiber-based materials, while the image data of paper-based materials also show the boundaries of the paper, see Figure 1a. We further reduce the resolution of the discretized images by a factor of 5. Altogether, this means that Ξζ\Xi_{\zeta} is evaluated on a 400×200×30400 \times200 \times30 grid.

To determine the number of distinct balls nbn_b making up the fiber, we use

L=rfnb+42.L = r_f \frac{n_b + 4}{2}.

In order to imitate the volume fraction observed in image data of approximately ρ0=160/255\rho_0 = 160/255, we estimate the intensity η\eta by means of Equation (8), where the volume fraction of the left-hand side is estimated from image data and the expected volume fraction of an individual fiber is computed by means of Monte Carlo simulations.

By the spectral theorem we can decompose B=QΛQ′B = Q\Lambda Q^{\prime} in (6), where QQ is an orthogonal, and Λ\Lambda a diagonal matrix. We parametrize Q=QθQ = Q_{\theta} and define it to be the matrix representation of rotation around the yy-axis by angle θ\theta, and Λ=Λ(c)=diag⁡(100,100−c,0)\Lambda= \Lambda(c) = \operatorname{diag}(100, 100 - c, 0). For c=θ=0c = \theta= 0, the resulting Bingham distribution concentrates around the equator of the sphere. Increasing cc, the mass concentrates on two antipodal points on the equator. When θ\theta is chosen away from 0, this circular area of concentration tilts along the yy-axis.

We thus obtain a setting of planar isotropy for c=θ=0c = \theta= 0 and two ways of inducing anisotropy, where in the first situation, we increase cc from 0 up to a maximum of c=5c = 5, and for the second, θ\theta from 0 up to π/6\pi/6. The resulting densities for these cases as well as exemplary realizations of the stochastic fiber model are shown in Figure 5.

(a) Bingham density (c=0c = 0, θ=0\theta= 0).

Slices of 3-dimensional structures simulated with the stochastic fiber model

(b) Slices of 3-dimensional structures simulated with the stochastic fiber model (c=0c = 0, θ=0\theta= 0). Dashed lines indicate observation windows 100 × 50 (red), 200 × 100 (blue), 400 × 200 (green). Fibers run mostly horizontally, field fulfills planar isotropy.

(c) Bingham density (c=5c = 5, θ=0\theta= 0).

Slices of 3-dimensional structures simulated with the stochastic fiber model

(d) Slices of 3-dimensional structures simulated with the stochastic fiber model (c=5c = 5, θ=0\theta= 0). Fiber directions are concentrated to preferably extend in xx-direction.

(e) Bingham density (c=0c = 0, θ=π/6\theta= \pi/6).

Slices of 3-dimensional structures simulated with the stochastic fiber model

(f) Slices of 3-dimensional structures simulated with the stochastic fiber model (c=0c = 0, θ=π/6\theta= \pi/6). Fiber directions are mostly horizontal if they run in yy-direction, but have slope θ=π/6\theta= \pi/6 if they run in xx-direction.

Figure 5. Illustration of structures simulated with the stochastic fiber model in case of the null hypothesis (first row), and the extreme cases of the respective alternatives (second and third row). The first column illustrates the density of the Bingham distribution used to sample the main direction of each fiber. The second column shows one instance of the generated fibers, respectively, given by slicing the 3-dimensional image across the three cardinal directions. Units are given in pixels.

We investigate the power and size empirically on three sizes of observational windows: 100×50100 \times50, 200×100200 \times100, and 400×200400 \times200, the latter corresponding to the downscaled dimensions of the dataset. The extent in zz-direction is fixed to 30 px. Figure 5b illustrates these window sizes relatively to the generated fiber images.

Results

To transform the image to local directions and subsequently assess uniform distribution of those directions, we maximize the norm-difference of observations at each point xx and its neighbors H(x)\mathbb{H}(x). The lags over which the maximization is performed are chosen to be sufficiently long to reliably leave the fiber when they point orthogonally to its direction in the xx-yy plane or when the fiber is not purely horizontal, while remaining short enough to avoid frequently picking up neighboring fibers. Beyond their length, we may also choose their direction. We select one set of lags that measures anisotropy close to the cardinal directions (H1\mathbb{H}_1), one that measures close to the diagonals (H3\mathbb{H}_3), and one that is in between (H2\mathbb{H}_2):

H1:={(±2,±8),(±8,±2)},H2:={(±5,±10),(±10,±5)},\mathbb{H}_1 := \{(\pm2, \pm8), (\pm8, \pm2)\}, \qquad\mathbb{H}_2 := \{(\pm5, \pm10), (\pm10, \pm5)\},
H3:={(±8,±11),(±11,±8)}.\mathbb{H}_3 := \{(\pm8, \pm11), (\pm11, \pm8)\}.

Figure 6a illustrates these sets by overlaying them on a slice of a stochastic fiber model realization. Furthermore, we select a square sub-domain DD of size 16×1616 \times16. The empirical size under the null hypothesis remained fairly stable for window sizes varying between 10×1010 \times10 and 50×5050 \times50.

Figure 6. (a) Sets of lags used to evaluate isotropy of stochastically generated images. Lengths are upscaled for real dataset. (b) Proportion of selections and joint 95% asymptotic confidence intervals for each lag for paper data. Angle of radial lines corresponds to angle of lag. Length of radial lines is given by selection proportions. Red circle corresponds to the uniform distribution (1/8). Ring sections outline confidence intervals for the proportions (in radial direction). Generated by R package unisono [5]. Illustration of sets of lags H1_1 (purple), H2_2 (orange), and H3_3 (yellow) and selection proportions with confidence intervals for paper data.

Figure 7 depicts empirical size of the test, as well as power under both alternatives. Clearly, increasing sample size improves power. Furthermore, the alternative of tilting the circular region of concentration by increasing θ\theta is harder to detect and the test exhibits low power for small sample size.

Figure 7. (a) Alternative hypothesis: Fibers run mostly horizontally, but concentrate in a single direction. The concentration increases with cc. At c=0c = 0, the field is planar isotropic. (b) Alternative hypothesis: Fibers run mostly horizontally in yy-direction, but have slope θ\theta in xx-direction. At θ=0\theta= 0, the field is planar isotropic. Empirical power and size of UNI applied to the simulated paper material based on 1000 repetitions at significance level α=0.05\alpha= 0.05 (marked with a horizontal line in each plot). Kinks in the lines may be further explained by instabilities in the estimation of λ(Ξ0)\lambda(\Xi_0).

We even observe just negligible power up to a certain angle θ\theta. Suppose that the measurement of the paper sample is not perfectly aligned with the xyxy-plane, but instead has a certain tilt, then its interior (with respect to zz-direction) structure is related to that of a synthetically generated dataset with θ\theta equal to that angle of tilt. Hence, the result indicates that the testing procedure exhibits no strong sensitivity with respect to such tilts.

Evaluation of isotropy on the real data

Finally, we apply the testing procedure to the dataset outlined in § 1.2. To exclude influences caused of edge effects, the image is cropped in zz-direction to consider only the 60 interior pixels used to tune the stochastic fiber model. To be consistent with the choice of hyperparameters in § 5.1, we multiply all previously used lag sets and subwindow dimensions by 5.

Recall that our test requires a stationary field. Stationarity in the xyxy-plane is well expected, see e.g. § 3.1 of Neumann et al., [34]. Additionally, we test this assumption layer-by-layer (in zz-direction) using a stationarity test for textured images by Taylor et al., [45]. The test is implemented in the R package LS2Wstat (Taylor and Nunes, [46]). The outcome confirms that there is no signs for a violation of stationarity.

After non-parallelized execution time of approximately 10 seconds on a standard laptop (eleventh generation Intel i5-1135G7 CPU), the pp-value output is numerically 0 (< 2.2e-16). Hence, our results indicate that there is strong evidence of anisotropy present in the dataset. This is in line with considering commercially manufactured unbleached paper sheets subjected to hard-nip calendering [35]. For commercially manufactured paper, anisotropic effects naturally arise during the production process [36].

fig:6b further illustrates the respective proportions of how often each lag appears as a maximizer. Additionally, a joint hyperrectangular 95% asymptotic confidence region for the proportions is computed based on the asymptotic Gaussian distribution of SS and visualized using individual confidence intervals for each proportion. These are found using the R package mvtnorm ([15]). Note that they are thus not necessarily a superset of the retention region of the proposed hypothesis test. According to Figure 6b, it appears that the violation is more in shorter lag lengths (H1\mathcal{H}_1) or along the cardinal directions rather than longer lag lengths or along the diagonals. Figure 8 illustrates the process M0(x)M_0(x), defined through the lags H1\mathcal{H}_1 on D0\mathcal{D}_0, and compares it with our stochastic fiber models under three different parameterizations. The empirical data exhibit a pattern most similar to that of the fiber model with parameters c=5c = 5 and θ=0\theta= 0, which corresponds to a tendency for the fibers to be oriented vertically (parallel to the yy-axis).

Figure 8. Top row: Maximizing directions selected from H1\mathcal{H}_1 for the paper-microstructure dataset as well as exemplary realizations of the fiber model under the null hypothesis (c=θ=0c = \theta= 0) and the two extremal alternatives (c=5c = 5 and θ=π/6\theta= \pi/6, respectively). Bottom row: Legend indicating how colors used in top row correspond to elements of H1\mathcal{H}_1. The direction of the radial lines corresponds to that of the lag vector. The length of radial lines is given by the selection proportions. The red line is the length under uniform proportions (1/81/8).

Discussion and future work

In the present paper, we developed a hypothesis test for assessing the isotropy of abstract-valued random fields, with a particular focus on function-valued fields. We applied this methodology to evaluate the planar isotropy of microstructures in paper-based materials. Notably, the test applies to random fields in general spaces equipped with a semi-metric. Consequently, this approach provides a tool to justify simplifying assumptions regarding the spatial dependence structure for a very general class of fields.

While similar tests exist for spatio-temporal data, our setting differs fundamentally: we consider the temporal axis to be a fixed, non-expanding interval. Therefore, we do not require stationarity in time.

Furthermore, we provide graphical diagnostic tools in the form of simultaneous confidence bands. If the test rejects the null hypothesis of isotropy, these bands help identify which spatial lags are selected significantly more or less often than expected under a uniform distribution. The finite-sample performance of the test was investigated in both scalar and function-valued (specifically, high-dimensional multivariate) settings. In the scalar case, the proposed test exhibits computation times and empirical power comparable to the SIS test (Weller and Hoeting, 2020). While the GSC test [17] is generally more powerful, it requires substantially higher computation times. For large sample sizes, both our methodology and the SIS test achieve a power comparable to the GSC test.

The implementation of the test is publicly available in the R package unisono [5], which is currently being prepared for release on CRAN.

To support our methodology, we provide a theoretical framework in the appendix. In this paper, our approach is restricted to the gridded observation setting. A natural direction for future theoretical work is extending this methodology to fields observed on a continuum or on random locations. Additional methodological improvements include implementing edge corrections and developing more sophisticated procedures for handling missing data in the interior. Currently, this is only addressed implicitly by treating missing observations as part of the domain boundary. Finally, it may be beneficial to explore self-normalization techniques to bypass the estimation of the long-run covariance, see e.g., Zhang et al. [53].

Declarations

Declaration of generative AI and AI-assisted technologies in the writing process. During the preparation of this work, the authors used ChatGPT and Claude to assist with mathematical exploration, language editing, and consistency checks. The authors independently verified all mathematical claims, derivations, and proofs and reviewed and edited all AI-assisted text. The authors take full responsibility for the content of the article.

Acknowledgements. The authors thank Karin Zojer and Ingo Manke for providing 3D image data of a compressed paper sheet. The computational results have been achieved using the Austrian Scientific Computing (ASC) infrastructure.

Appendix

Necessary properties for H(0)

The following lemmas give necessary conditions for H(0)\mathbb{H}^{(0)} to be transitive under its orthogonal symmetry group.

Lemma 7.1 (Equal norms and finiteness). Let H⊂Zd\mathbb{H} \subset\mathbb{Z}^{d} be transitive under its orthogonal symmetry group. Then there exists r≥0r \ge0 such that ∥h∥=r\lVert h\rVert= r for all h∈Hh \in\mathbb{H}, i.e. H⊆S0d−1(r)∩Zd\mathbb{H} \subseteq S_{0}^{d-1}(r) \cap\mathbb{Z}^{d}. In particular, H\mathbb{H} is finite.

Proof. Let h,h~∈Hh,\widetilde{h} \in\mathbb{H}. By transitivity there exists Q∈O(d)Q \in O(d) with QH=HQ\mathbb{H} = \mathbb{H} and Qh=h~Qh = \widetilde{h}. Hence ∥h~∥=∥Qh∥=∥h∥\lVert\widetilde{h}\rVert= \lVert Qh\rVert= \lVert h\rVert. Thus H\mathbb{H} is a subset of the sphere S0d−1(r)S_{0}^{d-1}(r), which is bounded and therefore contains only finitely many points of Zd\mathbb{Z}^{d}.

Lemma 7.2 (Equal spherical Voronoi measures). Let d≥2d \ge2 and let H={h1,…,hm}⊂Zd\mathbb{H} = \{h_{1},\ldots,h_{m}\} \subset\mathbb{Z}^{d} be transitive under its orthogonal symmetry group, with common norm r>0r > 0 (Lemma 7.1). For i=1,…,mi = 1,\ldots,m define the spherical Voronoi cell

Ci={x∈S0d−1(r):∥x−hi∥≤∥x−hj∥ for all j=1,…,m}C_{i} = \left\{x \in S_{0}^{d-1}(r) : \lVert x-h_{i}\rVert\le\lVert x-h_{j}\rVert\text{ for all } j=1,\ldots,m\right\}

and let ν\nu denote the surface measure on S0d−1(r)S_{0}^{d-1}(r). Then

ν(Ci)=ν(S0d−1(r))m,i=1,…,m.\nu(C_{i}) = \frac{\nu\left(S_{0}^{d-1}(r)\right)}{m}, \qquad i=1,\ldots,m.

Proof. Let i,j∈{1,…,m}i,j \in\{1,\ldots,m\}. By transitivity there exists Q∈O(d)Q \in O(d) with QH=HQ\mathbb{H} = \mathbb{H} and Qhi=hjQh_{i} = h_{j}. Since QQ preserves distances and permutes H\mathbb{H}, it follows that QCi=CjQC_{i} = C_{j}. As ν\nu is invariant under O(d)O(d), ν(Ci)=ν(Cj)\nu(C_{i}) = \nu(C_{j}). Since all hih_{i} have norm rr, the intersection Ci∩CjC_{i} \cap C_{j}, i≠ji \ne j, lies in the hyperplane {x:⟨x,hi−hj⟩=0}\{x : \langle x,h_{i}-h_{j}\rangle= 0\} and is thus a ν\nu-null set. Hence the cells partition S0d−1(r)S_{0}^{d-1}(r) up to a ν\nu-null set, which proves the claim.

Lemma 7.3 (Identical distance profiles). Let H={h1,…,hm}⊂Zd\mathbb{H} = \{h_{1},\ldots,h_{m}\} \subset\mathbb{Z}^{d} be transitive under its orthogonal symmetry group. Then the multisets

{{∥hi−hj∥:j=1,…,m}}and{{⟨hi,hj⟩:j=1,…,m}}\{\{\lVert h_{i}-h_{j}\rVert: j=1,\ldots,m\}\} \quad\text{and} \quad\{\{\langle h_{i},h_{j}\rangle: j=1,\ldots,m\}\}

do not depend on ii.

Proof. Let i,i′∈{1,…,m}i,i' \in\{1,\ldots,m\} and choose Q∈O(d)Q \in O(d) with QH=HQ\mathbb{H} = \mathbb{H} and Qhi=hi′Qh_{i} = h_{i'}. Then hj↦Qhjh_{j} \mapsto Qh_{j} is a permutation of H\mathbb{H}, and

∥hi′−Qhj∥=∥Q(hi−hj)∥=∥hi−hj∥,j=1,…,m.\lVert h_{i'}-Qh_{j}\rVert= \lVert Q(h_{i}-h_{j})\rVert= \lVert h_{i}-h_{j}\rVert, \qquad j=1,\ldots,m.

Hence the multisets of norms coincide for ii and i′i'. The same holds true for the scalar products, since by Lemma 7.1, ∥hi−hj∥2=2r2−2⟨hi,hj⟩\lVert h_{i}-h_{j}\rVert^{2} = 2r^{2} - 2\langle h_{i},h_{j}\rangle.

Properties of M0(x)

Proof of Lemma 2.1. Let Q∈O(d)Q \in O(d) and let H(0)={h1,…,hm}\mathbb{H}^{(0)} = \{h_{1},\ldots,h_{m}\}. Due to motion invariance, we obtain

(d(ξ(hi),ξ(0)))i=1m=L(d(ξ(Qhi),ξ(0)))i=1m,\left(\mathrm{d}(\xi(h_{i}),\xi(0))\right)_{i=1}^{m} \overset{\mathcal{L}}{=} \left(\mathrm{d}(\xi(Qh_{i}),\xi(0))\right)_{i=1}^{m},

for Q∈O(d)Q \in O(d) so that H(0)\mathbb{H}^{(0)} is invariant under QQ. Note that this invariance property is also valid for Q′Q'. We may conclude

M0(x)=arg⁡max⁡h∈H(0) d(ξ(x+h),ξ(x))=Qarg⁡max⁡h∈H(0) d(ξ(x+Qh),ξ(x))M_{0}(x) = \underset{h \in\mathbb{H}^{(0)}}{\arg\max}\ \mathrm{d}(\xi(x+h),\xi(x)) = Q\underset{h \in\mathbb{H}^{(0)}}{\arg\max}\ \mathrm{d}(\xi(x+Qh),\xi(x))
=LQarg⁡max⁡h∈H(0) d(ξ(x+h),ξ(x))=QM0(x).\overset{\mathcal{L}}{=} Q\underset{h \in\mathbb{H}^{(0)}}{\arg\max}\ \mathrm{d}(\xi(x+h),\xi(x)) = QM_{0}(x).

Take any h,h~∈H(0)h,\tilde{h}\in\mathbb{H}^{(0)} and QQ such that Q′h=h~Q'h=\tilde{h} and H(0)\mathbb{H}^{(0)} is invariant under QQ, which exists by assumption. Then,

P(M0(x)=h)=P(QM0(x)=h)=P(M0(x)=h~).\mathbb{P}(M_0(x)=h)=\mathbb{P}(QM_0(x)=h)=\mathbb{P}(M_0(x)=\tilde{h}).

Example 7.1. Let ZZ be a stationary Gaussian random field on Z2\mathbb{Z}^2 with Cov⁡(Z(x),Z(y))=ψZ(∥x−y∥)\operatorname{Cov}(Z(x),Z(y))=\psi_Z(\lVert x-y\rVert), where ψZ(0)=1\psi_Z(0)=1 and ρ:=ψZ(1)∈(0,1)\rho:=\psi_Z(1)\in(0,1). Let Z00,Z01,Z10,Z11Z^{00},Z^{01},Z^{10},Z^{11} be independent copies of ZZ and define

ξ(x):=Zx mod 2(⌊x/2⌋),x∈Z2,\xi(x):=Z^{x\bmod2}\left(\left\lfloor x/2\right\rfloor\right),\qquad x\in\mathbb{Z}^2,

where  mod \bmod and ⌊⋅⌋\lfloor\cdot\rfloor act componentwise. Let H(0):={e1,−e1,e2}\mathbb{H}^{(0)}:=\{e_1,-e_1,e_2\}, which is not transitive under its orthogonal symmetry group. Then ξ\xi is motion invariant, but P(M0(x)=e2)>1/3\mathbb{P}(M_0(x)=e_2)>1/3, so M0(x)M_0(x) is not uniformly distributed on H(0)\mathbb{H}^{(0)}.

Proof. We first discuss motion invariance. It suffices to show that Cov⁡(ξ(x),ξ(y))\operatorname{Cov}(\xi(x),\xi(y)) depends only on ∥x−y∥\lVert x-y\rVert. If x≢y(mod2)x\not\equiv y\pmod2, then ξ(x)\xi(x) and ξ(y)\xi(y) stem from independent copies, and the covariance is zero. If x≡y(mod2)x\equiv y\pmod2, then ⌊x/2⌋−⌊y/2⌋=(x−y)/2\lfloor x/2\rfloor-\lfloor y/2\rfloor=(x-y)/2, and the covariance is ψZ(∥x−y∥/2)\psi_Z(\lVert x-y\rVert/2). Since squares of odd integers are ≡1\equiv1 and squares of even integers are ≡0(mod4)\equiv0\pmod4, the case x≡y(mod2)x\equiv y\pmod2 occurs if and only if ∥x−y∥2≡0(mod4)\lVert x-y\rVert^2\equiv0\pmod4. Hence the covariance depends only on ∥x−y∥\lVert x-y\rVert.

H(0)\mathbb{H}^{(0)} is not transitive under its orthogonal symmetry group due to Lemma 7.2. We subsequently show that M0M_0 is not uniformly distributed. By stationarity, we may take x=0x=0. Assume without loss of generality that ZZ is centered. The four relevant values are

ξ(0)=Z00(0),ξ(e2)=Z01(0),ξ(e1)=Z10(0),ξ(−e1)=Z10(−e1).\xi(0)=Z^{00}(0),\qquad\xi(e_2)=Z^{01}(0),\qquad\xi(e_1)=Z^{10}(0),\qquad\xi(-e_1)=Z^{10}(-e_1).

All four are standard normal and independent, except for the pair (ξ(e1),ξ(−e1))(\xi(e_1),\xi(-e_1)), which has correlation ρ\rho. Ties occur with probability zero, so M0(0)=e2M_0(0)=e_2 holds if and only if both ξ(e1)\xi(e_1) and ξ(−e1)\xi(-e_1) lie in the random interval

J:=(ξ(0)−R,ξ(0)+R),R:=∣ξ(e2)−ξ(0)∣.\mathcal{J}:=(\xi(0)-R,\xi(0)+R),\qquad R:=|\xi(e_2)-\xi(0)|.

The interval J\mathcal{J} is independent of (ξ(e1),ξ(−e1))(\xi(e_1),\xi(-e_1)), and R>0R>0 almost surely. Hence

P(M0(0)=e2)=E[pρ(J)],pρ(J):=P(ζ1∈J, ζ2∈J),\mathbb{P}(M_0(0)=e_2)=\mathbb{E}\left[p_\rho(\mathcal{J})\right],\qquad p_\rho(J):=\mathbb{P}(\zeta_1\in J,\ \zeta_2\in J),

where (ζ1,ζ2)(\zeta_1,\zeta_2) is jointly normal with correlation ρ\rho and has standard normal marginals. Note that the law of J\mathcal{J} does not depend on ρ\rho.

For ρ=0\rho=0, the four values are i.i.d., so by symmetry each of the three lags is the maximizer with probability 1/31/3. It therefore suffices to show that pρ(J)p_\rho(J) is strictly increasing in ρ∈[0,1)\rho\in[0,1) for every bounded interval JJ of positive length. By Mehler’s formula [7 Theorem 2.4.1],

pρ(J)=∑n=0∞ρnn!cn(J)2,cn(J):=∫JHe⁡n(s)φ(s) ds.p_\rho(J)=\sum_{n=0}^{\infty}\frac{\rho^n}{n!}c_n(J)^2,\qquad c_n(J):=\int_J \operatorname{He}_n(s)\varphi(s)\,\mathrm{d}s.

where φ\varphi is the standard normal density and He⁡n\operatorname{He}_n are the Hermite polynomials. All coefficients are nonnegative. Moreover, cn(J)≠0c_n(J)\ne0 for some n≥1n\ge1, since otherwise 1J\mathbf{1}_J would be orthogonal to all nonconstant Hermite polynomials and thus almost everywhere constant. Hence pρ(J)p_\rho(J) is strictly increasing in ρ∈[0,1)\rho\in[0,1), and P(M0(0)=e2)>1/3\mathbb{P}(M_0(0)=e_2)>1/3 for ρ∈(0,1)\rho\in(0,1).

Relation between H0H_0 and H0′H'_0

The following results support the claims in Example 2.4.

Proposition 7.1. Consider the setup described in Example 2.4. Then, for all x∈Zdx \in\mathbb{Z}^d,

sign⁡(P(M0(x)=h1)−12)=sign⁡(h1′Ah1−h2′Ah2).\operatorname{sign}\left(\mathbb{P}(M_0(x)=h_1)-\frac{1}{2}\right)=\operatorname{sign}(h'_1Ah_1-h'_2Ah_2).

Proof. Since Θ\Theta is orthogonal, we have for any u,v∈Zdu,v \in\mathbb{Z}^d that d(Θ′ξ(u),Θ′ξ(v))=d(ξ(u),ξ(v))\mathrm{d}(\Theta'\xi(u),\Theta'\xi(v))=\mathrm{d}(\xi(u),\xi(v)) when d\mathrm{d} is the Euclidean distance. We may therefore assume Θ=I\Theta=I, so that the components ξ1,…,ξp\xi_1,\ldots,\xi_p are independent and univariate fields as in Example 2.3, with covariance functions ψj(h′Ah)\psi_j(h'Ah). By stationarity it suffices to consider x=0x=0.

Define UjU_j and VjV_j as in Example 2.3, with ξ\xi replaced by ξj\xi_j. Then

Y:=∥ξ(h1)−ξ(0)∥2−∥ξ(h2)−ξ(0)∥2=∑j=1pUjVj,P(M0(0)=h1)=P(Y>0),Y:=\lVert\xi(h_1)-\xi(0)\rVert^2-\lVert\xi(h_2)-\xi(0)\rVert^2=\sum_{j=1}^{p}U_jV_j,\qquad\mathbb{P}(M_0(0)=h_1)=\mathbb{P}(Y>0),

where the pairs (Uj,Vj)(U_j,V_j) are independent and Cov⁡(Uj,Vj)=2(ψj(h2′Ah2)−ψj(h1′Ah1))\operatorname{Cov}(U_j,V_j)=2\bigl(\psi_j(h'_2Ah_2)-\psi_j(h'_1Ah_1)\bigr). Since every ψj\psi_j is strictly decreasing, all these covariances share the sign of h1′Ah1−h2′Ah2h'_1Ah_1-h'_2Ah_2.

Next, we write UjVjU_jV_j as a linear combination of χ12\chi^2_1 random variables. To this end, denote Zj,+:=Uj/Var(Uj)+Vj/Var(Vj)Z_{j,+}:=U_j/\sqrt{\mathop{\rm Var}(U_j)}+V_j/\sqrt{\mathop{\rm Var}(V_j)} and Zj,−:=Uj/Var(Uj)−Vj/Var(Vj)Z_{j,-}:=U_j/\sqrt{\mathop{\rm Var}(U_j)}-V_j/\sqrt{\mathop{\rm Var}(V_j)}, which are independent, centered Gaussian variables. Then it can be easily verified that

UjVj=14Var(Uj)Var(Vj)(Zj,+2−Zj,−2),U_jV_j=\frac{1}{4}\sqrt{\mathop{\rm Var}(U_j)\mathop{\rm Var}(V_j)}(Z_{j,+}^2-Z_{j,-}^2),

and consequently there exist constants aj,bj>0a_j,b_j>0 so that we may write

Y=∑j=1p(ajζj2−bjζ~j2).Y=\sum_{j=1}^{p}(a_j\zeta_j^2-b_j\widetilde{\zeta}_j^2).

where ζ12,ζ~12,…,ζp2,ζ~p2∼iidχ12\zeta_1^2,\widetilde{\zeta}_1^2,\ldots,\zeta_p^2,\widetilde{\zeta}_p^2\stackrel{\mathrm{iid}}{\sim}\chi_1^2. Moreover, it can readily be verified that aj−bj=12Cov⁡(Uj,Vj)a_j-b_j=\frac{1}{2}\operatorname{Cov}(U_j,V_j).

Suppose now h1′Ah1≥h2′Ah2h'_1Ah_1\ge h'_2Ah_2, so that aj≥bja_j\ge b_j for all jj. Set Y−:=∑j(bjζj2−ajζ~j2)Y^-:=\sum_j(b_j\zeta_j^2-a_j\widetilde{\zeta}_j^2). Exchanging ζj2\zeta_j^2 and ζ~j2\widetilde{\zeta}_j^2 shows Y−=L−YY^-\overset{\mathcal{L}}{=}-Y. Moreover, Y−Y−=∑j(aj−bj)(ζj2+ζ~j2)≥0Y-Y^-=\sum_j(a_j-b_j)(\zeta_j^2+\widetilde{\zeta}_j^2)\ge0. Hence

P(Y>0)−P(Y<0)=P(Y>0)−P(Y−>0)=P(Y−≤0<Y)≥0.\mathbb{P}(Y>0)-\mathbb{P}(Y<0)=\mathbb{P}(Y>0)-\mathbb{P}(Y^->0)=\mathbb{P}(Y^-\le0<Y)\ge0.

If h1′Ah1=h2′Ah2h'_1Ah_1=h'_2Ah_2, then aj=bja_j=b_j, so Y=Y−Y=Y^- and P(Y>0)=12\mathbb{P}(Y>0)=\frac{1}{2}. If h1′Ah1>h2′Ah2h'_1Ah_1>h'_2Ah_2, then at ζj2=ζ~j2=1\zeta_j^2=\widetilde{\zeta}_j^2=1 for all jj we have Y=∑j(aj−bj)>0>Y−Y=\sum_j(a_j-b_j)>0>Y^-. By continuity, this holds on an open neighborhood, which has positive probability. Hence P(Y>0)>12\mathbb{P}(Y>0)>\frac{1}{2}. The case h1′Ah1<h2′Ah2h'_1Ah_1<h'_2Ah_2 follows by exchanging h1h_1 and h2h_2.

Corollary 7.1. Let ξ\xi be as in Proposition 7.1 on Zd\mathbb{Z}^d, d≥2d\ge2, and for 1≤i<j≤d1\le i<j\le d define

Hij(0):={e1,ej},H~ij(0):={ei+ej,ei−ej}.\mathbb{H}^{(0)}_{ij}:=\{e_1,e_j\},\qquad\widetilde{\mathbb{H}}^{(0)}_{ij}:=\{e_i+e_j,e_i-e_j\}.

Each of these sets is transitive under its orthogonal symmetry group. If AA is not a positive multiple of the identity matrix, then M0(x)M_0(x) is not uniformly distributed with respect to at least one of these (d−1)(d+2)/2(d-1)(d+2)/2 lag sets. Conversely, if AA is a positive multiple of the identity, then ξ\xi is isotropic and M0(x)M_0(x) is uniform with respect to all of them. Hence H0=H0′H_0=H'_0 if uniformity of M0(x)M_0(x) is required with respect to all of these lag sets simultaneously.

Proof. As in Example 2.1 H(0)\mathbb{H}^{(0)} and H~(0)\widetilde{\mathbb{H}}^{(0)} are transitive under their respective orthogonal symmetry groups. By Proposition 7.1, M0(x)M_0(x) is uniform on {h1,h2}\{h_1,h_2\} if and only if h1′Ah1=h2′Ah2h'_1Ah_1=h'_2Ah_2. For the sets above,

e1′Ae1−ej′Aej=A1,1−Aj,j,(ei+ej)′A(ei+ej)−(ei−ej)′A(ei−ej)=4Ai,j.e'_1Ae_1-e'_jAe_j=A_{1,1}-A_{j,j},\qquad (e_i+e_j)'A(e_i+e_j)-(e_i-e_j)'A(e_i-e_j)=4A_{i,j}.

If M0(x)M_0(x) is uniform with respect to all of these sets, then all diagonal entries of AA coincide and all off-diagonal entries vanish, i.e. A=aIA=aI with a>0a>0 by positive definiteness, where II is the identity matrix. Conversely, if A=aIA=aI, then C(h)C(h) depends on hh only through ∥h∥\lVert h\rVert, so ξ\xi is isotropic and M0(x)M_0(x) is uniform by H0⊂H0′H_0\subset H'_0.

Asymptotic justification under the null

We assume a weak dependence structure for the field ξ\xi based on α\alpha-mixing. To this end, suppose I⊂DI \subset\mathcal{D} and denote the σ\sigma-algebra generated by {ξ(x):x∈I}\{\xi(x): x \in I\} by FI\mathcal{F}_I. Then we require

αξ(u):=sup⁡∣P(E∩E′)−P(E)P(E′)∣→0(u→∞),\alpha_{\xi}(u) := \sup\left|\mathbb{P}\left(E \cap E'\right) - \mathbb{P}(E)\mathbb{P}\left(E'\right)\right| \to0 \quad(u \to\infty),

at a certain rate, where the supremum is taken over all events E∈FIE \in\mathcal{F}_I and E′∈FI′E' \in\mathcal{F}_{I'} and all sets I,I′⊂DI,I' \subset\mathcal{D}, such that dist⁡(I,I′):=inf⁡x∈I,y∈I′∥x−y∥≥u\operatorname{dist}(I,I') := \inf_{x\in I,y\in I'} \lVert x-y \rVert\ge u. Controlling dependence decay by an α\alpha-mixing condition is common in isotropy testing and utilized by both Guan et al. [17] and Weller and Hoeting (2020). Furthermore, α\alpha-mixing is a comparably weak assumption within classical mixing conditions, see e.g. Bradley [6]. If we assume that V=R\mathcal{V} = \mathbb{R} and ξ\xi is a Gaussian random field on Zd\mathbb{Z}^d, then Cov⁡(ξ(0),ξ(x))=O(∣x∣−(3d+ε))\operatorname{Cov}(\xi(0),\xi(x)) = O(\lvert x\rvert^{-(3d+\varepsilon)}) implies the mixing condition αξ(u)=O(u−(2d+ε))\alpha_{\xi}(u) = O(u^{-(2d+\varepsilon)}), see Section 2.1.1 in Doukhan, [13]. Furthermore, mm-dependent fields fulfill the α\alpha-mixing assumption (at any rate).

For asymptotic statements, the observation domain is denoted by Dk\mathcal{D}^k and is assumed to grow with kk. We assume that Dk\mathcal{D}^k is sufficiently regular so that for most x∈Dkx \in\mathcal{D}^k, the full set of neighbors H(x)=x+H(0)\mathbb{H}(x) = x + \mathbb{H}^{(0)} is available. Analogously to D0\mathcal{D}_0, we denote the set of xx with this property by D0k\mathcal{D}_0^k. Formally we require the domain to grow regularly in the following sense (Nahapetian, [32], p. 154):

Assumption 1. The sets (Dk)k≥1(\mathcal{D}^k)_{k\ge1} (thus also (D0k)k≥1(\mathcal{D}_0^k)_{k\ge1}) form a Van Hove sequence in Zd\mathbb{Z}^d.

Note that Assumption 1 particularly ensures that lim⁡k→∞n(D0k)/n(Dk)=1\lim_{k\to\infty} n(\mathcal{D}_0^k)/n(\mathcal{D}^k) = 1. For a detailed discussion of Van Hove sequences in Rd\mathbb{R}^d and Zd\mathbb{Z}^d, we refer to Chapter 3.1 in Bulinski and Shashkin [8].

Theorem 7.1. Let ξ\xi be a motion-invariant V\mathcal{V}-valued random field on Zd\mathbb{Z}^d such that αξ(u)=O(u−(2d+ε))\alpha_{\xi}(u) = O(u^{-(2d+\varepsilon)}) for some ε>0\varepsilon> 0. Suppose that Assumption 1 holds and that H(0)\mathbb{H}^{(0)} is transitive under its orthogonal symmetry group. Then,

Sk:=1n(D0k)∑x∈D0k[I(x)−1m1m],S_k := \frac{1}{\sqrt{n(\mathcal{D}_0^k)}} \sum_{x\in\mathcal{D}_0^k} \left[\mathcal{I}(x) - \frac{1}{m}\mathbb{1}_m\right],

converges in law to a zero mean Gaussian vector,

Sk→k→∞LNm(0,Γ),Γ=∑h∈ZdCov⁡(I(0),I(h)).S_k \xrightarrow[k\to\infty]{\mathcal{L}} \mathcal{N}_m(0,\Gamma), \qquad\Gamma= \sum_{h\in\mathbb{Z}^d} \operatorname{Cov}(\mathcal{I}(0),\mathcal{I}(h)).

The matrix series defining Γ\Gamma is absolutely convergent.

To show Theorem 7.1, we use a central limit theorem by Nahapetian [32]. The following result is a special case of this theorem.

Lemma 7.4. Let YY be a real-valued centered, stationary random field on Zd\mathbb{Z}^d such that for some p>2p > 2, E[∣Y(0)∣p]<∞\mathbb{E}[|Y(0)|^p] < \infty that fulfills an α\alpha-mixing condition with αY(u)=O(u−(ρd+ε))\alpha_Y(u) = O(u^{-(\rho_d+\varepsilon)}) for some ε>0\varepsilon> 0, where

ρd=max⁡{pp−2,2d}.\rho_d = \max\left\{\frac{p}{p-2},2d\right\}.

Then for any sequence of sets Dk⊂Zd\mathcal{D}^k \subset\mathbb{Z}^d growing in the Van Hove sense, the central limit theorem

n(Dk)−1/2∑h∈DkY(h)→k→∞LN(0,σY2),n(\mathcal{D}^k)^{-1/2} \sum_{h\in\mathcal{D}^k} Y(h) \xrightarrow[k\to\infty]{\mathcal{L}} \mathcal{N}(0,\sigma_Y^2),

holds, where

0<σY2=∑h∈ZdE[Y(0)Y(h)]<∞.0 < \sigma_Y^2 = \sum_{h\in\mathbb{Z}^d} \mathbb{E}[Y(0)Y(h)] < \infty.

:::`

Proof of Theorem 7.1. The mixing condition of ξ\xi transfers to M0(x)M_0(x) and thus also to the process a′I(x)a^{\prime}\mathcal{I}(x), for any a∈Rma \in\mathbb{R}^{m}. In particular we obtain

αa′I(u+2r)≤αξ(u),\alpha_{a^{\prime}\mathcal{I}}(u+2r) \le\alpha_{\xi}(u),

for each u≥1u \ge1. Furthermore, stationarity of ξ\xi also implies stationarity of M0M_0 and a′Ia^{\prime}\mathcal{I}, which is an almost surely bounded random field, so the moment condition applies for arbitrary p>2p>2. By Lemma 7.4 we obtain that a′Sk→LN(0,a′Γa)a^{\prime}S_k \xrightarrow{\mathcal{L}} \mathcal{N}(0,a^{\prime}\Gamma a) when k→∞k \to\infty. The multivariate CLT follows by the Cramér-Wold device.

Corollary 7.2. Suppose that Γ^k\widehat{\Gamma}_k is a weakly consistent estimator of Γ\Gamma and that the assumptions of Theorem 7.1 hold. Then (4) is an asymptotically valid level- α\alpha test for H0′H'_0.

Asymptotic justification under the alternative

We establish consistency for the weaker null hypothesis

H0′: M0(x) is uniformly distributed.H'_0: \ M_0(x)\text{ is uniformly distributed.}

The asymptotic power under fixed and local alternatives is summarized in the following two propositions. The first is a consequence of the law of large numbers.

Proposition 7.2 (Consistency against fixed alternatives). Suppose that ξ\xi is an anisotropic random field in Zd\mathbb{Z}^{d} that fulfills the remaining conditions of Theorem 7.1, with non-uniformly distributed M0M_0 in the sense that

Δ:=E[I(0)]−1m1m≠0.\Delta:= \mathbb{E}[\mathcal{I}(0)]-\frac{1}{m}\mathbf{1}_m \ne0.

Let ξ\xi be observed on Dk\mathcal{D}^k according to Assumption 1. Then, the probability of rejecting H0′H'_0 tends to 1 as k→∞k \to\infty.

Next, we formulate a consistency result for local alternatives, i.e. settings, where anisotropy weakens with increasing sample size.

Proposition 7.3 (Power against local alternatives). Let (ξ(k))k≥1(\xi^{(k)})_{k\ge1} be a sequence of stationary random fields where ξ(k)\xi^{(k)} is observed on Dk⊂Zd\mathcal{D}^k \subset\mathbb{Z}^{d} following Assumption 1. Let I(k)\mathcal{I}^{(k)} be its associated index field. Suppose that

E[I(k)(0)]−1m1m=n(Dk)−1/2Δk,∥Δk∥=1,(9)\mathbb{E}[\mathcal{I}^{(k)}(0)]-\frac{1}{m}\mathbf{1}_m=n(\mathcal{D}^k)^{-1/2}\Delta_k,\quad\lVert\Delta_k\rVert=1, \tag*{(9)}

Let Γk\Gamma_k be the long-run covariance of I(k)\mathcal{I}^{(k)} and assume Γk→Γ0\Gamma_k \to\Gamma_0 as k→∞k \to\infty. Let further Γ^k\widehat{\Gamma}_k be weakly consistent estimators in the sense that (Γ^k−Γk)→0(\widehat{\Gamma}_k-\Gamma_k)\to0, in probability. Suppose sup⁡kαξ(k)(u)=O(u−(2d+ε))\sup_k\alpha_{\xi^{(k)}}(u)=O(u^{-(2d+\varepsilon)}) for some ε>0\varepsilon>0. Then, at significance level α\alpha,

lim inf⁡k→∞Pξ(k)(H0′ is rejected)>α.\liminf_{k\to\infty}\mathbb{P}_{\xi^{(k)}}(H'_0\text{ is rejected})>\alpha.

Proof. By (9),

1n(D0k)∑x∈D0k[I(k)(x)−1m1m]=1n(D0k)∑x∈D0k[I(k)(x)−EI(k)(0)]+n(D0k)n(Dk)Δk.\frac{1}{\sqrt{n(\mathcal{D}_0^k)}}\sum_{x\in\mathcal{D}_0^k}\left[\mathcal{I}^{(k)}(x)-\frac{1}{m}\mathbf{1}_m\right] = \frac{1}{\sqrt{n(\mathcal{D}_0^k)}}\sum_{x\in\mathcal{D}_0^k}\left[\mathcal{I}^{(k)}(x)-\mathbb{E}\mathcal{I}^{(k)}(0)\right] + \sqrt{\frac{n(\mathcal{D}_0^k)}{n(\mathcal{D}^k)}}\Delta_k.

The bounds sup⁡kαξ(k)(u)=O(u−(2d+ε))\sup_k\alpha_{\xi^{(k)}}(u)=O(u^{-(2d+\varepsilon)}) translate to the indicator fields and allow the triangular-array CLT of Jenish and Prucha [21] to be applied to every scalar projection. Hence, for every a∈Rma\in\mathbb{R}^{m},

a′1n(D0k)∑x∈D0k[I(k)(x)−EI(k)(0)]→LN(0,a′Γ0a),a^{\prime}\frac{1}{\sqrt{n(\mathcal{D}_0^k)}}\sum_{x\in\mathcal{D}_0^k}\left[\mathcal{I}^{(k)}(x)-\mathbb{E}\mathcal{I}^{(k)}(0)\right]\xrightarrow{\mathcal{L}}\mathcal{N}(0,a^{\prime}\Gamma_0a),

since Γ^k→Γ0\widehat{\Gamma}_k \to\Gamma_0. Note that n(D0k)/n(Dk)→1n(D_0^k)/n(D^k) \to1 by Assumption 1. By the Cramér–Wold device, along every subsequence Δkℓ→Δ∞\Delta_{k_\ell} \to\Delta_\infty,

1n(D0k)∑x∈D0k[I(k)(x)−1m1m]→LNm(Δ∞,Γ0).\frac{1}{\sqrt{n(D_0^k)}}\sum_{x\in D_0^k}\left[\mathcal{I}^{(k)}(x)-\frac{1}{m}\mathbf{1}_m\right]\xrightarrow{\mathcal{L}}\mathcal{N}_m(\Delta_\infty,\Gamma_0).

Since ∥Δk∥=1\lVert\Delta_k\rVert=1, every subsequence admits a further subsequence converging to a limit with norm 1. Moreover, we have Γ^k→PΓ0\widehat{\Gamma}_k \xrightarrow{\mathbb{P}} \Gamma_0. Thus, by Slutsky’s theorem, Tk=∥Sk∥2T_k=\lVert S_k\rVert^2 converges along every such subsequence to a specific weighted non-central χ2\chi^2-distribution, which stochastically strictly dominates the identically-weighted central χ2\chi^2-distribution. This implies our claim. □

Consistency of the long-run covariance estimator

To show consistency, we require the following conditions for the basis sub-domains and shifts.

Assumption 2. For each k∈Nk\in\mathbb{N}, let Dk⊂ZdD_k\subset\mathbb{Z}^d be a finite lattice set forming a Van Hove sequence, with n(Dk)→∞n(D_k)\to\infty. Let Dk;ℓ=Dk+sk;ℓ⊂D0kD_{k;\ell}=D_k+s_{k;\ell}\subset D_0^k, ℓ=1,…,Nk\ell=1,\ldots,N_k, be pairwise disjoint translates of DkD_k. Moreover, assume that n(Dk)/Nk→0n(D_k)/N_k\to0 as k→∞k\to\infty.

Theorem 7.2. Let ξ\xi be a stationary V\mathcal{V}-valued random field on Zd\mathbb{Z}^d that fulfills a mixing condition with αξ(u)=O(u−(d+ϵ))\alpha_\xi(u)=O(u^{-(d+\epsilon)}) for some ϵ>0\epsilon>0. Let ((Dk;ℓ)ℓ=1Nk)k≥1((D_{k;\ell})_{\ell=1}^{N_k})_{k\geq1} be a sequence of families of subdomains for which Assumption 2 holds. Then the estimator Γ^k\widehat{\Gamma}_k as described in Section 3.1 is weakly consistent.

Proof of Theorem 7.2. The estimator is positive semi-definite by the representation

Γ^k=n(Dk)Nk−1Uk′Uk,Uk;ℓ,i:=μ^k;ℓ,i−μ‾k;i,\widehat{\Gamma}_k=\frac{n(D_k)}{N_k-1}U_k^{\prime}U_k,\qquad U_{k;\ell,i}:=\widehat{\mu}_{k;\ell,i}-\overline{\mu}_{k;i},

for i=1,…,mi=1,\ldots,m and ℓ=1,…,Nk\ell=1,\ldots,N_k, where indices after the semicolon denote matrix and vector entries, respectively.

To show consistency, we adapt the approach of Pantle et al. [38] (specifically Lemma 2 and Theorem 3) to a random field observed on a lattice under a weak dependence condition. We establish entry-wise mean-square consistency by showing that the estimator is asymptotically unbiased and that the variances of its entries vanish as k→∞k\to\infty.

Throughout the proof, let I0(x):=I(x)−E[I(0)]\mathcal{I}_0(x):=\mathcal{I}(x)-\mathbb{E}[\mathcal{I}(0)] denote the centered index field. Since the estimator is invariant under a common deterministic shift of the block means, we may work with block means of I0\mathcal{I}_0. Write nk:=n(Dk)n_k:=n(D_k) and D‾k:=⋃ℓ=1NkDk;ℓ\overline{D}_k:=\bigcup_{\ell=1}^{N_k}D_{k;\ell}, so that n(D‾k)=Nknkn(\overline{D}_k)=N_kn_k.

Since I0\mathcal{I}_0 is a measurable transformation of ξ\xi, it satisfies the same mixing-rate bound. In particular, ∑h∈ZdαI0(∥h∥)<∞\sum_{h\in\mathbb{Z}^d}\alpha_{\mathcal{I}_0}(\lVert h\rVert)<\infty. Further, for every finite D,D′⊂ZdD,D'\subset\mathbb{Z}^d,

∑x∈D,y∈D′αI0(∥x−y∥)≤n(D)∑h∈ZdαI0(∥h∥).(10)\sum_{\substack{x\in D,y\in D'}}\alpha_{\mathcal{I}_0}(\lVert x-y\rVert)\leq n(D)\sum_{h\in\mathbb{Z}^d}\alpha_{\mathcal{I}_0}(\lVert h\rVert). \tag*{(10)}

First, decompose the estimator into

Γ^k=nkNk∑ℓ=1Nkμ^k;ℓμ^k;ℓ′−nkNk(Nk−1)∑ℓ≠ℓ′=1Nkμ^k;ℓμ^k;ℓ′′=:Sk,1+Sk,2.\widehat{\Gamma}_k=\frac{n_k}{N_k}\sum_{\ell=1}^{N_k}\widehat{\mu}_{k;\ell}\widehat{\mu}_{k;\ell}^{\prime}-\frac{n_k}{N_k(N_k-1)}\sum_{\ell\ne\ell'=1}^{N_k}\widehat{\mu}_{k;\ell}\widehat{\mu}_{k;\ell'}^{\prime}=:S_{k,1}+S_{k,2}.

By stationarity,

Γ−E[Sk,1]=∑h∈ZdCov⁡(I0(0),I0(h))n(Dk∖(Dk−h))nk.\Gamma-\mathbb{E}[S_{k,1}]=\sum_{h\in\mathbb{Z}^d}\operatorname{Cov}(\mathcal{I}_0(0),\mathcal{I}_0(h))\frac{n(D_k\setminus(D_k-h))}{n_k}.

::: By Bradley [6], the entries of Cov⁡(I0(0),I0(h))\operatorname{Cov}(\mathcal{I}_0(0),\mathcal{I}_0(h)) are absolutely bounded by 8αI0(∥h∥)8\alpha_{\mathcal{I}_0}(\lVert h\rVert) and are therefore absolutely summable. The Van Hove property and dominated convergence consequently yield E[Sk,1]→Γ\mathbb{E}[S_{k,1}] \to\Gamma.

For Sk,2S_{k,2}, the mixing covariance inequality and (10) give, entry-wise,

∣E[Sk,2;i,j]∣≤8Nk(Nk−1)nk∑ℓ≠ℓ′=1Nk∑x∈Dk;ℓy∈Dk;ℓ′αI0(∥x−y∥)≤8Nk(Nk−1)nk∑ℓ=1Nk∑x∈Dk;ℓy∈D‾kαI0(∥x−y∥)≤8Nk−1∑h∈ZdαI0(∥h∥).\begin{aligned} \left\lvert\mathbb{E}[S_{k,2;i,j}]\right\rvert\leq\frac{8}{N_k(N_k-1)n_k}\sum_{\substack{\ell\neq\ell'=1}}^{N_k}\sum_{\substack{x\in D_{k;\ell}\\ y\in D_{k;\ell'}}}\alpha_{\mathcal{I}_0}(\lVert x-y\rVert) \\ &\leq\frac{8}{N_k(N_k-1)n_k}\sum_{\ell=1}^{N_k}\sum_{\substack{x\in D_{k;\ell}\\ y\in\overline{D}_k}}\alpha_{\mathcal{I}_0}(\lVert x-y\rVert)\leq\frac{8}{N_k-1}\sum_{h\in\mathbb{Z}^d}\alpha_{\mathcal{I}_0}(\lVert h\rVert). \end{aligned}

which goes to 00 as k→∞k\to\infty.

Next, we show that the variances vanish. For Sk,1S_{k,1}, the mixing covariance inequality yields

V(Sk,1;i,j)≤4Nk2nk2∑ℓ,ℓ′=1Nk∑x1,y1∈Dk;ℓx2,y2∈Dk;ℓ′αI0({x1,y1},{x2,y2})∥I0(0)∥∞4.\mathbb{V}(S_{k,1;i,j})\leq\frac{4}{N_k^2n_k^2}\sum_{\ell,\ell'=1}^{N_k}\sum_{\substack{x_1,y_1\in D_{k;\ell}\\ x_2,y_2\in D_{k;\ell'}}}\alpha_{\mathcal{I}_0}(\{x_1,y_1\},\{x_2,y_2\})\lVert\mathcal{I}_0(0)\rVert_\infty^4.

where ∥I0(0)∥∞4≤1\lVert\mathcal{I}_0(0)\rVert_\infty^4\leq1. By monotonicity,

αI0({x1,y1},{x2,y2})≤∑x∈{x1,y1}y∈{x2,y2}αI0(∥x−y∥).\alpha_{\mathcal{I}_0}(\{x_1,y_1\},\{x_2,y_2\})\leq\sum_{\substack{x\in\{x_1,y_1\}\\ y\in\{x_2,y_2\}}}\alpha_{\mathcal{I}_0}(\lVert x-y\rVert).

Each summand on the right-hand side leaves at most nk2n_k^2 choices for the unused indices. After collapsing the sum, by (10),

V(Sk,1;i,j)≤16Nk2∑x,y∈D‾kαI0(∥x−y∥)≤16nkNk∑h∈ZdαI0(∥h∥)→0,\mathbb{V}(S_{k,1;i,j})\leq\frac{16}{N_k^2}\sum_{x,y\in\overline{D}_k}\alpha_{\mathcal{I}_0}(\lVert x-y\rVert)\leq\frac{16n_k}{N_k}\sum_{h\in\mathbb{Z}^d}\alpha_{\mathcal{I}_0}(\lVert h\rVert)\to0,

by Assumption 1. Similarly, for Sk,2S_{k,2},

V(Sk,2;i,j)≤16Nk2(Nk−1)2nk2∑x1,x2,y1,y2∈D‾kαI0(∥x1−x2∥)≤16Nknk(Nk−1)2∑h∈ZdαI0(∥h∥)→0.\begin{aligned} \mathbb{V}(S_{k,2;i,j})\leq\frac{16}{N_k^2(N_k-1)^2n_k^2}\sum_{x_1,x_2,y_1,y_2\in\overline{D}_k}\alpha_{\mathcal{I}_0}(\lVert x_1-x_2\rVert) \\ &\leq\frac{16N_kn_k}{(N_k-1)^2}\sum_{h\in\mathbb{Z}^d}\alpha_{\mathcal{I}_0}(\lVert h\rVert)\to0. \end{aligned}

Here, the first bound follows by expanding the covariance of the products, bounding the set-distance mixing coefficient by the sum of four pointwise coefficients as above, and dropping the restriction ℓ≠ℓ′\ell\neq\ell'.

We have thus shown that E[Sk,1]→Γ\mathbb{E}[S_{k,1}]\to\Gamma, E[Sk,2]→0\mathbb{E}[S_{k,2}]\to0, and the variances of all entries of Sk,1S_{k,1} and Sk,2S_{k,2} vanish. Consequently, Γ^k\widehat{\Gamma}_k converges entry-wise in mean square to Γ\Gamma, and is therefore a weakly consistent estimator of Γ\Gamma.

References

  1. [1]Abaach, M., Biermé, H., Di Bernardino, E., and Estrade, A. (2025). Pixel isotropy test based on directional perimeters. Spatial Statistics 65, 100869. DOI: 10.1016/j.spasta.2024.100869.DOI
  2. [2]Altendorf, H. and Jeulin, D. (2011). Random-walk-based stochastic modeling of three-dimensional fiber systems. Physical Review E 83, 041804. DOI: 10.1103/PhysREV E.83.041804.
  3. [3]Azaïs, J.-M., Dalmao, F., and De Castro, Y. (2025). Gaussian random field’s anisotropy using excursion sets. arXiv: 2512.11085.arxiv.org/abs/2512.11085
  4. [4]Banerjee, R., Hinebaugh, J., Liu, H., Yip, R., Ge, N., and Bazylak, A. (2016). Heterogeneous porosity distributions of polymer electrolyte membrane fuel cell gas diffusion layer materials with rib-channel compression. International Journal of Hydrogen Energy 41, 14885–14896. DOI: 10.1016/J.IJHYDENE.2016.06.147.DOI
  5. [5]Baumhakel, J., Hörmann, S., and Neumann, M. (2026). R package: Unisono. URL: www.github.com/baumhakel/unisono.
  6. [6]Bradley, R. C. (2007). Introduction to Strong Mixing Conditions. Kendrick Press.
  7. [7]Bryc, W. (1995). The Normal Distribution: Characterizations. Springer. DOI: 10.1007/978-1-4612-2560-7.DOI
  8. [8]Bulinski, A. and Shashkin, A. (2007). Limit Theorems for Associated Random Fields and Related Systems. World Scientific. DOI: 10.1142/6555.
  9. [9]Chiu, S. N., Stoyan, D., Kendall, W. S., and Mecke, J. (2013). Stochastic Geometry and its Applications. J. Wiley & Sons. DOI: 10.1002/9781118658222.DOI
  10. [10]Cressie, N. A. C. (2015). Statistics for Spatial Data. Revised edition. J. Wiley & Sons. DOI: 10.1002/9781119115151.
  11. [11]Davies, R. B. (1980). Algorithm AS 155: The distribution of a linear combination of χ² random variables. Journal of the Royal Statistical Society C 29, 323–333. DOI: 10.2307/2346911. URL: www.jstor.org/stable/2346911.DOI
  12. [12]Dodell, L., Neumann, M., Osenberg, M., Hilger, A., Studer, G., Esser, B., Manke, I., and Schmidt, V. (2025). Quantifying local heterogeneities in the 3D morphology of X-PVMPT battery electrodes based on FIB-SEM measurements. ACS Materials Letters 7, 1914–1921. DOI: 10.1021/acs materialeslett.4c01552.arxiv.org/abs/2502.19085
  13. [13]Doukhan, P. (1994). Mixing: Properties and Examples. Springer. DOI: 10.1007/978-1-4612-2642-0. URL: link.springer.com/10.1007/978-1-4612-2642-0.
  14. [14]Eckle, T., van Delft, A., and Dette, H. (2026). Balancing the edge effect and dimension of spectral spatial statistics under irregular sampling with applications to isotropy testing. Bernoulli 32, 442–466. DOI: 10.3150/25-bej1865.DOI
  15. [15]Genz, A. and Bretz, F. (2009). Computation of Multivariate Normal and t Probabilities. Springer.DOI
  16. [16]Gräfensteiner, P., Rodriguez, A., Leitl, P., Baikova, E., Fuchs, M., Machado-Charry, E., Hirn, U., Hilger, A., Manke, I., Schennach, R., Neumann, M., Schmidt, V., and Zojer, K. (2026). Predicting air flow in calendered paper sheets from micro-CT data: Combining physics with morphology. Transport in Porous Media 153, 15. DOI: 10.1007/s11242-025-02250-9.DOI
  17. [17]Guan, Y., Sherman, M., and Calvin, J. A. (2004). A nonparametric test for spatial isotropy using subsampling. Journal of the American Statistical Association 99, 810–821. DOI: 10.1198/016214504000001150.DOI
  18. [18]Haibel, A., Beckmann, F., Dose, T., Herzen, J., Ogurreck, M., Müller, M., and Schreyer, A. (2010). Latest developments in microtomography and nanotomography at PETRA III. Powder Diffraction 25, 161–164. DOI: 10.1154/1.3428364.DOI
  19. [19]Hosseinzadeh, E., Marco, J., and Jennings, P. (2018). The impact of multi-layered porosity distribution on the performance of a lithium ion battery. Applied Mathematical Modelling 61, 107–123. DOI: 10.1016/J.APM.2018.04.001.DOI
  20. [20]Hyvärinen, A., Karhunen, J., and Oja, E. (2001). Independent Component Analysis. J. Wiley & Sons. DOI: 10.1002/0471221317.DOI
  21. [21]Jenish, N. and Prucha, I. R. (2009). Central limit theorems and uniform laws of large numbers for arrays of random fields. Journal of Econometrics 150, 86–98. DOI: 10.1016/j.jeconom.2009.02.009.DOI
  22. [22]Jeulin, D. (2021). Morphological Models of Random Structures. Springer. DOI: 10.1007/978-3-030-75452-5.DOI
  23. [23]Jona-Lasinio, G. (2001). Modeling and exploring multivariate spatial variation: A test procedure for isotropy of multivariate spatial data. Journal of Multivariate Analysis 77, 295–317. DOI: 10.1006/jmva.2000.1938.DOI
  24. [24]Karafíátová, I., Møller, J., Pawlas, Z., Staněk, J., Seitl, F., and Beneš, V. (2023). Fitting the grain orientation distribution of a polycrystalline material conditioned on a Laguerre tessellation. Spatial Statistics 55, 100747. DOI: 10.1016/j.spasta.2023.100747.DOI
  25. [25]Karafíátová, I., Pawlas, Z., and Heller, L. (2022). Statistical assessment of stress redistribution in loaded polycrystals. Image Analysis & Stereology 41. DOI: 10.5566/ias.2642.DOI
  26. [26]Last, G. and Penrose, M. (2018). Lectures on the Poisson Process. Cambridge University Press. DOI: 10.1017/9781316104477.DOI
  27. [27]Leitl, P., Machado-Charry, E., Baikova, E., Neumann, M., Hirn, U., Schmidt, V., and Zojer, K. (2023). Joint distributions of local pore space metrics quantitatively explain simulated air flow variations in paper. Transport in Porous Media 148, 627–648. DOI: 10.1007/s11242-023-01964-y.DOI
  28. [30]Malyarenko, A. and Ostoja-Starzewski, M. (2023). Tensor- and spinor-valued random fields with applications to continuum physics and cosmology. Probability Surveys 20, 1–86. DOI: 10.1214/22-ps12.DOI
  29. [31]Mardia, K. V. and Jupp, P. E. (2000). Directional Statistics. J. Wiley & Sons. DOI: 10.1002/9780470316979.DOI
  30. [32]Nahapetian, B. (1991). Limit Theorems and Some Applications in Statistical Physics. Vieweg + Teubner Verlag. DOI: 10.1007/978-3-322-93432-1. URL: link.springer.com/10.1007/978-3-322-93432-1.DOI
  31. [33]Neumann, M., Gräfensteiner, P., Machado-Charry, E., Hirn, U., Hilger, A., Manke, I., Schennach, R., Schmidt, V., and Zojer, K. (2024). R-vine copulas for data-driven quantification of descriptor relationships in porous materials. Advanced Theory and Simulations 2024, 2301261. DOI: 10.1002/adts.202301261.DOI
  32. [34]Neumann, M., Machado-Charry, E., Zojer, K., and Schmidt, V. (2021a). On variability and interdependence of local porosity and local tortuosity in porous materials: a case study for sack paper. Methodology and Computing in Applied Probability 23, 613–627. DOI: 10.1007/s11009-019-09761-1.DOI
  33. [35]Neumann, M., Machado-Charry, E., Baikova, E., Hilger, A., Hirn, U., Schennach, R., Manke, I., Schmidt, V., and Zojer, K. (2021b). Capturing centimeter-scale local variations in paper pore space via μ-CT: A benchmark study using calendered paper. Microscopy and Microanalysis 27, 1305–1315. DOI: 10.1017/S1431927621012563.DOI
  34. [36]Niskanen, K. J. (2018). Distribution of Fibre Orientations in Paper. In: Fundamentals of Papermaking. Ed. by C. F. Baker and V. Punton. FRC, 275–308. DOI: 10.15376/frc.1989.1.275.
  35. [37]Nordhausen, K., Oja, H., Filzmoser, P., and Reimann, C. (2015). Blind source separation for spatial compositional data. Mathematical Geosciences 47, 753–770. DOI: 10.1007/s11004-014-9559-5.DOI
  36. [38]Pantle, U., Schmidt, V., and Spodarev, E. (2010). On the estimation of integrated covariance functions of stationary random fields. Scandinavian Journal of Statistics 37, 47–66. DOI: 10.1111/j.1467-9469.2009.00663.x.DOI
  37. [39]R Core Team (2023). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. URL: www.R-project.org/.
  38. [40]Rieder, P., Neumann, M., Fernandes, L. M., Mulard, A., Proudhon, H., Willot, F., and Schmidt, V. (2024). Stochastic 3D microstructure modeling of twinned polycrystals for investigating the mechanical behavior of γ-TiAl intermetallics. Computational Materials Science 238, 112922. DOI: 10.1016/j.commatsci.2024.112922.DOI
  39. [41]Rimmelé, G., Barlet-Gouédard, V., Porcherie, O., Goffé, B., and Brunet, F. (2008). Heterogeneous porosity distribution in Portland cement exposed to CO₂-rich fluids. Cement and Concrete Research, 1038–1048. DOI: 10.1016/J.CEMCONRES.2008.03.022.DOI
  40. [42]Sahoo, I., Guinness, J., and Reich, B. J. (2019). A test for isotropy on a sphere using spherical harmonic functions. Statistica Sinica 29, 1253–1276. DOI: 10.5705/ss.202017.0475.DOI
  41. [43]Sheng, N., Khazaie, S., Chevreuil, M., and Fréour, S. (2023). Statistical properties of effective elastic moduli of random cubic polycrystals. Mechanics & Industry 24, 33. DOI: 10.1051/meca/2023030.DOI
  42. [44]Sherman, M. (1996). Variance estimation for statistics computed from spatial lattice data. Journal of the Royal Statistical Society Series B 58, 509–523. DOI: 10.1111/J.2517-6161.1996.TB02097.X.DOI
  43. [45]Taylor, S., Eckley, I., and Nunes, M. (2014). A test of stationarity for textured images. Technometrics 56, 291–301. DOI: 10.1080/00401706.2013.823890.DOI
  44. [46]Taylor, S. and Nunes, M. (2023). R package: LS2Wstat: A multiscale test of spatial stationarity for LS2W processes. DOI: 10.32614/CRAN.package.LS2Wstat. URL: https://CRAN.R-project.org/package=LS2Wstat.
  45. [47]Weller, Z. D. (2018). SpTest: An R package implementing nonparametric tests of isotropy. Journal of Statistical Software 83, 1–24. DOI: 10.18637/jss.v083.i04.DOI
  46. [48]Weller, Z. D. and Hoeting, J. A. (2016). A review of nonparametric hypothesis tests of isotropy properties in spatial data. Statistical Science 31, 305–324. DOI: 10.1214/16-sts547. URL: www.jstor.org/stable/24780855.DOI
  47. [50]Wilde, F., Ogurreck, M., Greving, I., Hammel, J. U., Beckmann, F., Hipp, A., Lottermoser, L., Khokhriakov, I., Lytaev, P., Dose, T., Burmester, H., Müller, M., and Schreyer, A. (2016). Micro-CT at the imaging beamline P05 at PETRA III. In: AIP Conference Proceedings. Vol. 1741. AIP Publishing LLC, 030035. DOI: 10.1063/1.4952858.DOI
  48. [51]Wilson, W. A. (1931). On semi-metric spaces. American Journal of Mathematics 53, 361–373. DOI: 10.2307/2370790.DOI
  49. [52]Woudberg, S., Theron, F., Lys, E., and Le Coq, L. (2018). Investigating the influence of local porosity variations and anisotropy effects on the permeability of fibrous media for air filtration. Chemical Engineering Science 180, 70–78. DOI: 10.1016/J.CES.2018.01.035.DOI
  50. [53]Zhang, X., Li, B., and Shao, X. (2014). Self-normalization for spatial data. Scandinavian Journal of Statistics 41, 311–324. DOI: 10.1111/sjos.12028.DOI

Paper details

vlnc:W_mkrbvhaa
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT, Claude

View statement
Looking for related papers…
Contents