Electronic Journal of Statistics

Vol. 13 (2019) 536–581

ISSN: 1935-7524
https://doi.org/10.1214/19-EJS1530

Lipschitz-Killing curvatures of excursion

sets for two-dimensional random fields
Hermine Biermé
LMA UMR CNRS 7348, Université de Poitiers,
11 bd Marie et Pierre Curie, 86962 Chasseneuil, France
e-mail: hermine.bierme@math.univ-poitiers.fr

Elena Di Bernardino
Conservatoire national des arts et métiers, Paris, EA4629,
292 rue Saint-Martin, Paris Cedex 03, France
e-mail: elena.di bernardino@cnam.fr

Céline Duval
MAP5 UMR CNRS 8145, Université Paris Descartes
45 rue des Saints-Pères, 75270 Paris Cedex 06, France
e-mail: celine.duval@parisdescartes.fr

and

Anne Estrade
MAP5 UMR CNRS 8145, Université Paris Descartes
45 rue des Saints-Pères, 75270 Paris Cedex 06, France
e-mail: anne.estrade@parisdescartes.fr

Abstract: In the present paper we study three geometrical characteris-

tics for the excursion sets of a two-dimensional stationary isotropic random
field. First, we show that these characteristics can be estimated without
bias if the considered field satisfies a kinematic formula, this is for instance
the case of fields given by a function of smooth Gaussian fields or of some
shot noise fields. By using the proposed estimators of these geometric char-
acteristics, we describe some inference procedures for the estimation of the
parameters of the field. An extensive simulation study illustrates the perfor-
mances of each estimator. Then, we use the Euler characteristic estimator
to build a test to determine whether a given field is Gaussian or not, when
compared to various alternatives. The test is based on a sparse informa-
tion, i.e., the excursion sets for two different levels of the field to be tested.
Finally, the proposed test is adapted to an applied case, synthesized 2D
digital mammograms.

MSC 2010 subject classifications: Primary 60G60, 62F12; Secondary

62F03, 60G10.
Keywords and phrases: Test of Gaussianity, Gaussian and shot-noise
fields, excursion sets, level sets, Euler characteristic, crossings.

Received May 2018.

536
 LK curvatures of excursion sets for two-dimensional random fields 537

1. Introduction

In this paper we are interested in using simple geometrical objects to build

inference and testing procedures to recover global characteristics of a two-
dimensional stationary isotropic random field by using a sparse information.
Assume that we only have access to the excursion set of one realization of a
field over one (or two) level, namely we observe a black-and-white image de-
picting where the field is above or below some level u (see, e.g., the second and
third images of the first row of Figures 1, 3, 5, 7 and 8). From this image, it is
possible to compute several Lipschitz-Killing (LK) curvatures of the given ex-
cursion set. In our bivariate framework, there are three LK curvatures available.
Roughly speaking, they are the following quantities: surface area, half perime-
ter and Euler characteristic. Each of these three functionals carries a different
information on the geometry or topology of the excursion set. Intuitively, the
surface area is related to its occupation density, the perimeter to its regularity
and the Euler characteristic to its connectivity. In the present paper we have
multiple objectives. Firstly, we aim at proposing a unified definition of these
quantities for a large class of stationary and isotropic random fields that can
be either smooth or present discontinuities. We jointly use these quantities to
build unbiased estimators for the expected values of the LK curvatures of the
excursion sets by solving a triangular system. In the whole paper, the use of
the term unbiased means that our estimators take into account the observation
bias due to the fact that the excursion set is observed on some finite (large)
window and not on the whole plane. Moreover, we construct a test procedure to
determine whether the excursion sets of a given field come from the realization
of a Gaussian field.
The Lipschitz-Killing curvatures, depending on the scientific domain under
consideration, are also called intrinsic volumes or Minkowski functionals, they
can also be assimilated to genus types. Considering real-life data as a realization
of a random field and analyzing them with the help of LK curvatures is an effec-
tive approach that has already been successfully applied in various disciplines
in the past decades. For instance, in cosmology, the question of Gaussianity and
anisotropy of the Cosmic Microwave Background (CMB) radiation has been
tackled in a huge amount of publications (see [10] for a recent overview) and
part of them are explicitly based on Minkowski functionals (see, e.g., [31] or
[17]). The methodology of these studies is often the following: choose a spe-
cific parametric model for the CMB signal and fit the parameters so that the
observed Minkowski functional matches the theoretical one. A similar method-
ology is also applied for analyzing the distribution of galaxies (see, for instance
[18]), where again the best-fitting model is the target. Another domain where
the LK curvatures are used as methodological tools is in brain imaging (see [2],
Section 5, and references therein). In this context, the main purpose is to find
the locations of high brain activity assuming the brain is a manifold with very
complex geometry. The tools that are involved concern both the LK curvatures
of high level excursions of the signal and the LK curvatures of the brain itself.
As seen in the aforementioned domains of application, the field of interest can
538 Biermé et al.

be either a smooth (e.g. Gaussian or transformations of Gaussian fields) or a dis-

continuous object (e.g. shot noise fields). A shot noise field is a multidimensional
extension of one dimensional models such as compound Poisson processes. It is
obtained by throwing random shapes at random locations and counting how
many shapes cover each point of the space with eventually random weights.
Those models are very popular since explicit computations can (sometimes) be
performed and simulations are rather simple. They are natural alternatives to
Gaussian fields both for theoretical and applicative purposes. This underlines
the need for a framework that remains robust if the nature of the field changes.
That is why we focus our study on providing tools and methods that can be
applied to both Gaussian or transformations of Gaussian and shot noise fields.
The same objective can be found in [6] and [7], which have also inspired the
present study. Note that all the random fields that we will consider are defined
on the continuous space R2 .
There is clearly an abundant literature on parameter estimation and tests for
multidimensional stationary random fields. Inference methods and tests usually
rely on the estimation of the covariance function and/or on the estimation of the
finite dimensional distributions of the field (see [14], [28] or [27] among others).
This usually requires the observation of the entire field and/or independent
copies of the field. In the present paper, we have a completely different approach
based on the observation of the geometry of a single excursion set above a fixed
level, considering the random field as a random surface. Let us mention some
famous precursors, for instance [25], [37] or [24], who mainly worked on the
modeling of sea waves considering them as realizations of a two-dimensional
random field. Let us also mention two major books on this topics, [1] and [3].
In the mathematical domain, the LK curvatures are the subject of several
works studying their probabilistic and statistical properties. We mainly refer to
[32] and [35] for a precise definition of these geometrical and topological objects.
They have been applied to the excursion sets of random fields in many situations.
For instance, in [9], [6] and [5], the length of the level sets (i.e. the perimeter
of the excursion sets) is taken into account, while several limit theorems are
obtained for the area in [8] and [33]. In [12] and [16], the Euler characteristic of
excursion sets is used to test Gaussianity or isotropy. In the second reference,
looking simultaneously at the intersections of excursion sets with rectangles and
segments is a way to get additional information. Although the studies often focus
on one of the LK curvatures, considering all the LK curvatures of the excursion
sets at the same time is the purpose of recent papers such as [19], [20], [26] or
[7]. With the same objective -joint study of all Minkowski functionals-, but in
another context -excursion sets of random fields defined on a discrete space-
, one can also quote the recent papers [29] and [13]. The present paper lies
within the same scope since we provide three unbiased estimators, one for each
LK curvature, valid for a wide class of stationary isotropic random fields. The
estimator of the LK curvature corresponding to the Euler characteristic allows
us to construct a test to determine if two observed excursion sets result from
a Gaussian field or not. The strength of this test is that it is independent of
a specific choice of the correlation function of the considered random field, in
 LK curvatures of excursion sets for two-dimensional random fields 539

particular it does not depend on its second spectral moment. Since this function
is unknown in many applications, it is an important and very useful property
of the proposed test. In this sense, it can be considered as more efficient than
the Gaussianity test introduced in [12].
To summarize, the originality of the present study is threefold. First, we
use a very sparse information on the field, namely excursion sets above one or
two levels, to recover global characteristics of the considered random field. Our
second asset is to gather in the same study smooth Gaussian type fields and
shot noise fields, which are archetypes of non continuous fields. The third key
point of our paper is the joint use of several LK curvatures for a given excursion
set.
Outline of the paper. The paper is organized as follows. In the remaining
of this section we define the three objects of interest, i.e., the Lipschitz-Killing
densities for the excursion set of a two-dimensional random field. Section 2 is
devoted to the study of unbiased estimators with edge correction of these LK
densities from the observation of one excursion set. We examine a wide range of
random fields, namely fields of Gaussian type and shot noise fields. Moreover,
we show how the knowledge of the LK densities permits to recover and to infer
on parameters of the considered fields. The problem of the consistency of the
proposed estimators is studied, when it is possible. Furthermore, their perfor-
mances are numerically analyzed. In Section 3 we build a test to detect whether
a given field is Gaussian or not based on the knowledge of two excursion sets
corresponding to two different levels. A variant of this test is put into practice
in Section 4 on synthesized 2D digital mammograms provided by GE Health-
care France (department Mammography). Finally, Appendix section gathers the
proofs of the technical results.

1.1. Lipschitz-Killing densities of excursion sets

There are three additive functionals, Lj for j = 0, 1, 2, defined on subsets of

Borelians in R2 that are extensively used in the literature. Depending on the
mathematical domain, they are called either intrinsic volumes, Minkowski func-
tionals or Lipschitz-Killing curvatures. Roughly speaking, for A a Borelian set in
R2 , L0 (A) stands for the Euler characteristic of A, L1 (A) for the half perimeter of
the boundary of A and L2 (A) is equal to the area of A, i.e. the two-dimensional
Lebesgue measure.
All over the paper, we will denote by | · | the two-dimensional Lebesgue
measure of any Borelian set in R2 and by | · |1 its one-dimensional Hausdorff
measure. In particular, when T is a bounded rectangle in R2 with non empty
interior,
1
L0 (T ) = 1, L1 (T ) = |∂T |1 , L2 (T ) = |T |, (1)
2
where ∂T stands for the frontier of the set T . The rest of this section is dedicated
to precisely define the Lipschitz-Killing (LK) curvatures of excursion sets as well
as the associated LK densities.
540 Biermé et al.

Let T be a bounded rectangle in R2 with non empty interior and u be a real

number. For X a real-valued stationary random field defined on R2 , we consider
the excursion set within T above level u:

{t ∈ T : X(t) ≥ u} = T ∩ EX (u), where EX (u) := X −1 ([u, +∞)).

We are interested in the mean of the three LK curvatures of the excursion

set, assuming they are well defined. The case of the area L2 is particularly
easy to handle with. Actually, since X is stationary, Fubini’s Theorem gives
immediately
E [ |T ∩ EX (u)| ] = |T | P(X(0) ≥ u).
This formula is valid without any further assumption on X, whereas the
existence and the exact value of the two other LK curvatures is more involving.
In particular it is often needed to make strong regularity assumptions on the
field in order that one can consider the LK curvatures of its excursion sets.
In the following definitions, we introduce positive reach sets and curvature
measures for those sets (see [35], based on the seminal work of Federer [15]).
Definition 1.1 (Positive reach sets). For a set A ⊂ R2 and r a positive real
number, let Ar = {x ∈ R2 : dist(A, x) ≤ r}, with dist the Euclidean distance.
Then, the reach of A is defined as

reach(A) := sup{r ≥ 0 : ∀ y ∈ Ar , ∃ a unique x ∈ A nearest to y}.

The set A is said to have positive reach, if reach(A) > 0.

Intuitively, A is a positive reach set if one can roll a ball of positive radius
along the exterior boundary of A keeping in touch with A. Let us remark that
positive reach sets are necessarily closed sets. Moreover, convex sets in R2 and
compact C 2 submanifolds of R2 have positive reach, see e.g., Proposition 14 in
[35]).
Definition 1.2 (Curvature measures). Let A be a positive reach set. Its cur-
vature measures Φi (A, ·), for i = 0, 1, 2, are defined for any Borel set U ⊂ R2
by
1 1
Φ0 (A, U ) = TC(∂A, U ), Φ1 (A, U ) = |∂A ∩ U |1 and Φ2 (A, U ) = |A ∩ U |,
2π 2
where TC(∂A, U ) denotes the integral over U of the curvature along the posi-
tively oriented curve ∂A.
For a compact positive reach set A such that A ⊂ U , with U an open set,
let us note that Φ0 (A, U ) coincides with the Euler characteristic of A. For more
details on the total curvature TC, we refer to Definition 9 and Theorem 31 in
[35] or to Definition 1 and Theorem 1 in [7].
Since the curvature measures Φi (A, ·) are additive functionals with respect to
A, it is natural to deal with unions of positive reach sets. Therefore, we introduce
the next definition (see [35] again).
 LK curvatures of excursion sets for two-dimensional random fields 541

Definition 1.3 (UPR class). Let UPR be the class of locally finite unions of
sets with positive reach.
In the sequel, we are interested in the curvature measures of the excursions
above the level u of the random field X, namely we consider A = T ∩ EX (u).
Note that the random set T ∩ EX (u) belongs to the UPR class a.s. when,
for instance, the random field X is of class C 2 a.s., (actually, in this case the
random set T ∩ EX (u) has positive reach, as EX (u) is a C 2 submanifold of
R2 and its intersection with the rectangle T provides compactness and positive
reach property), or when EX (u) is locally given by a finite union of disks.
Definition 1.4 (LK curvatures and associated densities for UPR excursion
sets). Let X be a stationary random field defined on R2 and let T be a bounded
rectangle in R2 with non empty interior. Assuming that T ∩ EX (u) is a UPR
set, define the LK curvatures of the excursion set EX (u) within T by

Ci (X, u, T ) := Φi (T ∩ EX (u), T ), for i = 0, 1, 2.

Define the normalized LK curvatures by

/T Ci (X, u, T )
Ci (X, u) := , for i = 0, 1, 2, (2)
|T |
and, assuming the limits exist, the associated LK densities are

Ci∗ (X, u) := lim 2 E[Ci (X, u)], for i = 0, 1, 2,

/T
(3)
T R

where lim 2 stands for the limit along any sequence of bounded rectangles that
T R
grows to R2 .
As already noticed, the case i = 2 in the above definition boils down to

C2∗ (X, u) = E[C2 (X, u)] = P(X(0) ≥ u).

/T
(4)

2. Kinematic formulas and inference

From now on, all the rectangles T in R2 are bounded with non empty interior.
Notation T R2 stands for the limit along any sequence of bounded rectangles
that grows to R2 . Then, |∂T |1
|T | always goes to 0 as T R2 .
In this section, we build unbiased estimators of the LK densities in (3), for
i ∈ {0, 1}. The problem is more involved than it appears. Actually, a naive
approach is to consider (2), that can be easily computed from T ∩ EX (u), as an
estimator of (3). But, the main difference between theses quantities comes from
the boundary terms. Indeed, we can write

Ci (X, T, u) = Φi (T ∩ EX (u), T ) = Φi (EX (u), T̊ ) + Φi (T ∩ EX (u), ∂T ),

 
where, loosely speaking, |T |−1 E Φi (EX (u), T̊ ) contains all the information on
Ci∗ (X, u) and boundary terms involving ∂T is blurring the estimation for fixed
542 Biermé et al.

|T |. However, as |∂T |1 /|T | → 0 when T R2 , this bias term vanishes. Hereafter,

we derive the exact relations between (2) and (3) which are valid for a wide range
of fields, that are called standard in the sequel, and permit to build unbiased
estimators of Ci∗ (X, u), i ∈ {0, 1} with an edge correction. Starting from those
estimators, we are able to infer on some parameters of the random field X.

2.1. Kinematic formulas

We use kinematic formulas to get unbiased estimators of Ci∗ (X, u), i ∈ {0, 1}
as it is usual to proceed in convex geometry. To this aim, we first introduce the
definition of a standard random field. We borrow the adjective standard to [32]
because the fields we call standard have excursion sets that are standard in the
sense of Definition 9.2.1 in [32].
Definition 2.1 (Standard random field). Let X be a stationary isotropic ran-
dom field defined on R2 . We say that X is standard at level u ∈ R if T ∩ EX (u)
is UPR for any rectangle T in R2 , if Ci∗ (X, u), for i = 0, 1 in (3) exist, and if

1 ∗ |∂T |1 1
E[C0 (X, u)] = C0∗ (X, u) + + C2∗ (X, u)
/T
C (X, u) ,
π 1 |T | |T |
1 |∂T |1
E[C1 (X, u)] = C1∗ (X, u) + C2∗ (X, u)
/T
.
2 |T |

By using (4), no edge correction is necessary for estimating C2∗ (X, u) whatever
the field X is. It explains why we do not include a third constraint in the
definition of a standard field. The first example of a standard field is given by
the large class of fields of Gaussian type defined below.
Definition 2.2 (Fields of Gaussian type). We call field of Gaussian type any
random field X = F (G), where for some k ∈ N∗ , F : Rk → R is a C 2 function
and G = (G1 , . . . , Gk ) is a family of i.i.d. Gaussian random fields defined on R2
that are C 3 , stationary, isotropic, centered, with unit variance, and such that
Var Gi (0) = λI2 for some λ > 0, with Gi denoting the gradient of Gi for all
i ∈ {1, . . . , k} and I2 the 2 × 2 identity matrix.
Proposition 2.3. Let X be a field of Gaussian type as in Definition 2.2, then
X is standard at any level u ∈ R.
Proof. Following Definition 2.2, we write X = F (G) and we denote by λ the
second spectral moment of the Gaussian vectorial field G. The Gaussian kine-
matic formula provides the mean LK curvatures of excursion sets of X within
a rectangle T (see e.g. Theorem 15.9.5 in [1] or Theorem 4.8.1 in [2]), for u ∈ R
and i = 0, 1, 2,
2−i 
 
i+l
E [Ci (X, T, u)] = (2π)−l/2 λl/2 Ml (X, u) Li+l (T ) (5)
l
l=0
 LK curvatures of excursion sets for two-dimensional random fields 543
   
i+l i + l ωl+i
where Lj (T ), j = 0, 1, 2 are defined in (1), = with ωk the
l l ωl ωi
Lebesgue measure of the k-dimensional unit ball (w0 = 1, w1 = 2 and w2 = π),
and, following Formula (3.5.2) in [2], the coefficients Ml (X, u), l = 0, 1, 2 are
obtained having an expansion in ρ at order 2 of the probability that G(0) belongs
to
T ube(F, ρ) := {x ∈ Rk such that dist(x, F −1 ([u, ∞))) ≤ ρ},
as ρ → 0+ . Namely, the expansion is given by
1
P G(0) ∈ T ube(F, ρ) = M0 (X, u) + ρM1 (X, u) + ρ2 M2 (X, u) + O(ρ3 ).
2
(6)

Dividing Equation (5) by |T | and letting T grow to R2 imply the existence of

the LK densities and
λ 1 πλ
C0∗ (X, u) = M2 (X, u), C1∗ (X, u) = M1 (X, u), C2∗ (X, u) = M0 (X, u).
2π 2 2
(7)

Injecting these in (5) and dividing by |T | we obtain that X is standard at level

u.
Another way to get standard fields relies on convex geometry tools especially
developed for Boolean fields with convex grains.
Proposition 2.4. Let X be a stationary isotropic random field defined on R2
such that for u ∈ R, its excursion set EX (u) satisfies:
• a.s. T ∩ EX (u) is the union of a finite number N (T ∩ EX (u)) of compact
convex
 sets with non empty interior,
• E 2N (T ∩EX (u)) < +∞.
Then, X is standard at level u.
Proof. First let us note that by stationarity and isotropy of X we have that
EX (u) is a stationary and isotropic random set and the assumptions imply that
EX (u) is a standard random set in the sense of Definition 9.2.1 of [32]. Following
the notations of this book, Ci (X, u, T ) = Vi (T ∩ EX (u)) with Vi the so-called
intrinsic volume and Ci∗ (X, u) = Vi (EX (u)) with Vi the intrinsic density. It
holds according to Theorem 9.4.1 of [32], for i ∈ {0, 1, 2},


2
E(Ci (X, u, T )) = cik,2−k+i Ck∗ (X, u)L2−k+i (T ),
k=i

where, by Equation (5.5) of [32], c0,2 2,0 2,1 1,2 1,1

0 = c0 = c1 = c1 = 1 and c0 =
2
π. We
deduce that X is standard at level u.
When X is a random field that is standard at level u, we are now able to
build unbiased estimators of Ci∗ (X, u), for i = 0, 1, 2.
544 Biermé et al.

Proposition 2.5 (Unbiased estimators for LK densities). Let u ∈ R and let X

be a stationary isotropic random field defined on R2 that is standard at level u.
Assume we observe T ∩ EX (u) for T a rectangle in R2 . The following quantities
are unbiased estimator of Ci∗ (X, u), i = 0, 1, 2,
  2 
/T |∂T |1 /T 1 |∂T |1 1 /T
C0,T (X, u) = C0 (X, u) − C (X, u) + − C2 (X, u),
π|T | 1 2π |T | |T |
(8)
/T |∂T |1 /T
C1,T (X, u) = C1 (X, u) − C (X, u), (9)
2|T | 2
/T
C2,T (X, u) = C2 (X, u). (10)

We highlight that the term unbiased in Proposition 2.5 does not refer to the
pixelization error that arise numerically due to the discretized representation of
images. Here, the use of the term unbiased means that our estimation of the LK
densities takes into account the observation bias due to the intersection of the
excursion set with the observation window T . Other common edge correction
methods include for instance the toroidal correction, where the edge on one side
can be thought of as being wrapped around to the opposite edge. Numerically
for large data sets, the minus-sampling methods are often used (see e.g., [34];
[11]), where the points near the edges are ignored for the estimation but are still
considered as neighbours for interior points. However, it is statistically inefficient
because it discards a substantial amount of data.
In the following two sections we analyse how the unbiased estimators in
Proposition 2.5 for standard fields can be used for parameter inference pur-
poses. Section 2.2 is devoted to fields of Gaussian type whereas Section 2.3
concerns shot-noise fields.

2.2. Fields of Gaussian type

2.2.1. Gaussian field

We start by considering the simplest case of Gaussian type field, i.e. with no-
tation of Definition 2.2, k = 1 and F : x → x, then X = G1 := G. To compute
the LK densities, we use expansion (6) to write out
1
P G(0) ≥ u − ρ = ψ(u − ρ) = ψ(u) − ρψ  (u) + ρ2 ψ  (u) + O(ρ3 ),
2
with ψ being the Gaussian tail distribution with zero mean and unit variance.
Hence, one easily gets
1 1/2 −u2 /2
C0∗ (G, u) = (2π)−3/2 λ u e−u C1∗ (G, u) = C2∗ (G, u) = ψ(u),
2
/2
, λ e ,
4
(11)
where λ denotes the second spectral moment of G as in Definition 2.2.
 LK curvatures of excursion sets for two-dimensional random fields 545

Parameter estimation. Suppose we observe T ∩EG (u) for a given level u and
a given bounded rectangle T in R2 . Then we can use the unbiased estimators of
Proposition 2.5 to propose a consistent estimator for the second spectral moment
of the Gaussian field G. Each of the quantities Ci (G, u, T ) satisfies a Central
Limit Theorem (CLT) with normalizing term equals to |T |1/2 (see, e.g., [20] or
[26]). Concerning various levels and/or various disjoint domains, a joint CLT
for the Euler characteristic is proved in [12]. To get the asymptotic normality of
the estimator of λ, we moreover assume that the Gaussian field G satisfies the
following condition.
(A) For any fixed x in R2 , the covariance matrix of the random vector (G(x),
 
G (x), G (x)) has full rank and r, the covariance function of G, is such
that,

Mr (x) → 0 when x → +∞ and Mr ∈ L1 (R2 ) ,

∂kr
where Mr (x) = max ∂xk
(x) ; k = (k1 , k2 ) ∈ N2 , k1 + k2 ≤ 4 .

Proposition 2.6 (CLT for the spectral moment estimator). Assume that G
satisfies condition (A). Consider C0,T (G, u) the estimator defined in (8) built
on the observation T ∩ EG (u), u = 0 being fixed. By using (11), we define the
estimator of λ as

(2π)3/2 u2 /2
λT (u) := e C0,T (G, u), for u = 0.
u
Then, it holds that
 d
|T | λT (u) − λ −−−−→ N (0, Σ(u)), for some Σ(u) < +∞, (12)
2T R

d
where −
→ stands for the convergence in distribution.
Remark that for u = 0, the Euler density in (11) always vanishes, indepen-
dently of λ. The estimator is therefore not defined for u = 0 in Proposition
2.6.
   d
Proof. It is sufficient to show that |T | C0,T (G, u)−C0∗ (G, u) −−−−→ N (0, V (u))
2 T R
for some V (u) < +∞. Decompose this quantity as follows
   
|T | C0,T (G, u) − C0∗ (G, u) = |T | C0 (G, u) − E[C0 (G, u)]
/T /T

1 /T /T |∂T |1
− |T | C1 (G, u) − E[C1 (G, u)]
π |T |
 /T /T 1 |∂T |21 1
+ |T | C2 (G, u) − E[C2 (G, u)] −
2π |T |2 |T |
:= I0 (T ) + I1 (T ) + I2 (T ).
546 Biermé et al.

d
Theorem 4 in [12] leads to I0 (T ) −−−−→ N (0, V (u)) with a finite asymptotic
T R
2

variance V (u) given by Equation (9) in [12]. Furthermore, the CLT in [20]
  /T /T 
gives that |T | C1 (G, u) − E[C1 (G, u)] admits a Gaussian centered limit
P
distribution and therefore I1 (T ) −−−−→ 0. Similar arguments hold for I2 (T ).
T R
2

u2
Finally the convergence of λT (u) is established with Σ(u) = V (u)(2π)3 eu2 .

Numerical illustration. Figure 1 displays a realization of a Gaussian ran-

dom field, two excursion sets and an illustration for the performance of the three
estimators C0,T (G, u), C1,T (G, u) and C2,T (G, u) in Proposition 2.5, for G with
covariance function r(x) = e−κ x and second spectral moment λ = 2κ2 . Let
2 2

us remark that the parameter κ is chosen such that |T |κ2 remains bounded,
/T
which explains small values of κ and λ in Figure 1. The quantities C0 (G, u),
/T /T
C1 (G, u) and C2 (G, u) are computed with the Matlab functions bweuler,
bwperim and bwarea, respectively. When it is required to specify the connectiv-
ity, we average between the 4th and the 8th connectivity. Since C1∗ is defined as
the average half perimeter, we divide by 2 the output derived from bwperim.
From a numerical point of view, bweuler and bwarea functions seem very
precise contrary to the bwperim function which performs less well. It was ex-
pected due to the pixelisation effect. This behavior will be observed also in other
random field cases (see, for instance, Figures 7 and 8 in some shot-noise cases).
/T
Figure 1 (center) illustrates that C1 (G, u) (green dashed line) does not well
∗
approximate C1 (G, u) (blue plain line), especially for small levels u and that
the correction induced by (9) (red stars) improves the approximation. In Figure
1 (left), we provide an analogous bias correction for the Euler characteristic
by using (8). However in this case, the discrepancy is less evident than in the
perimeter case.

In Figure 2 (left), we observe the unbiased estimator λT (u) for different values
of u. The asymptotic variance Σ(u) in Proposition 2.6 is empirically estimated
on M = 100 sample simulations (Figure 2, right). This allows us to identify some
choices of levels u where the variance is minimum. Furthermore, we remark that
for small and large values of u, less statistics is available than for intermediate
values of |u|.

2.2.2. Chi-square field

We consider a chi-square random field with k degrees of freedom Zk defined

with Zk = F (G) for F : x ∈ Rk → x 2 , with notations of Definition 2.2. In
order to deal with a random field with zero mean and unit variance, we also
introduce Zk as
1
Zk (t) := √ (Zk (t) − k), t ∈ R2 .
2k
 LK curvatures of excursion sets for two-dimensional random fields 547

2 2
Fig 1. Gaussian random field as in Section 2.2.1 with covariance r(x) = e−κ x , for
κ = 100/210 in a domain of size 210 × 210 pixels. First row: A realization of a Gaussian
random field (left) and two excursion sets for u = 0 (center) and u = 1 (right). Second row:
C0,T (G, u) (left), C1,T (G, u) (center) and C2,T (G, u) (right) as a function of the level u. We
display the averaged values on M = 100 sample simulations (red stars) and the associated
empirical intervals (vertical red lines). Theoretical u → C0∗ (G, u), C1∗ (G, u) and C2∗ (G, u) in
/T /T
(11) are drawn in blue lines. We also present u → C0 (G, u) and C1 (G, u) in green dashed
lines. These samples have been obtained with Matlab using circulant embedding matrix.

Hence, Zk is C 3 (R2 ), stationary, isotropic, centered, with unit variance and
Var Zk (0) = 2λI2 .
We assume that Zk is observed on a rectangle T ⊂√R2 through its excursion
set above a fixed level u; T ∩ EZk (u) = T ∩ EZk (k + u 2k). Let us remark that

the above excursion is a proper subset only for a level u such that u > − k/2.
Moreover, the LK densities satisfy
√
Cj∗ (Zk , u) = Cj∗ (Zk , k + u 2k), for j = 0, 1, 2.

The Cj∗ (Zk , ·) can be computed using (7) in the framework of Gaussian type
fields described previously. Indeed, we have Zk = F (G) with G = (G1√ , . . . , Gk )
−1
and F : x ∈ Rk → ||x||2 . Hence, √ F ([u, +∞)) = {x ∈ R k
; ||x|| ≥ u} and
T ube(F, ρ) = {x ∈ Rk ; ||x|| ≥ u − ρ}, so that expansion in (6) has to be
written with
 √
P G(0) ∈ T ube(F, ρ) = P χ2k ≥ u − ρ ,
548 Biermé et al.

2 2
Fig 2. Gaussian random field with covariance r(x) = e−κ x , for κ = 100/210 in a
domain of size 210 × 210 pixels. Estimate λT (u) with associated confidence intervals for
M = 100 sample simulations as prescribed by the CLT in (12), for different values of u (left).
Theoretical value λ = 0.019 is represented by the horizontal line. The empirically estimated
variance Σ(u) is displayed for different values of u in the right panel.

where χ2k stands for a chi-square random variable with k degrees of freedom.
We recover the following formulas that can be found, for instance, in Theorem
15.10.1 in [1] or in Section 5.2.1 in [2]:
 √ 
∗  λ √ (k−2)/2 √ k + u 2k
C0 (Zk , u) = k + u 2k u 2k + 1 exp − ,
π2k/2 Γ(k/2) 2
(13)
√  √ 
∗  πλ √ (k−1)/2 k + u 2k
C1 (Zk , u) = (k+1)/2 k + u 2k exp − ,
2 Γ(k/2) 2
√
C2∗ (Zk , u) = P χ2k ≥ k + u 2k .

These formulas were originally established in [36]. Figure 3 displays a realization

of a normalized chi-square random field Zk with k = 2 degrees of freedom, two
excursion sets and an illustration of the performance of the three estimators
C0,T (Zk , u), C1,T (Zk , u) and C2,T (Zk , u).

Parameter estimation. Suppose we observe T ∩ EZ (u), the excursion of a

centered chi-square field Z with unit variance, unknown degree of freedom K
and unknown Gaussian second spectral moment λ. Using that for a centered chi-
 u) only depends
square field with unit variance and k degrees of freedom C2∗ (Z,
on k, we propose the following estimator of K:

K(u) = arg min C2∗ (Zk , u) − C2,T (Z,

 u) , (14)
k ∈ [1, kmax ]

with C2,T as in (10) and kmax a large positive integer. This estimator can be
 u) to derive an estimator of the Gaussian second spectral
plugged in C0∗ (Z,
 LK curvatures of excursion sets for two-dimensional random fields 549

Fig 3. Chi-square field as in Section 2.2.2 with 2 degrees of freedom and λ = 0.019 in a do-
main of size 210 ×210 pixels. First row: A realization of a normalized chi-square random field
(left) and two excursion sets for u = 0 (center) and u = 1 (right). Second row: C0,T (Z k , u)
k , u) (center) and C2,T (Z
(left), C1,T (Z k , u) (right) as a function of the level u. We display
the averaged values on M = 100 sample simulations (red stars) and the associated empirical
intervals (vertical red lines). Theoretical u → C0∗ (Zk , u), C ∗ (Z
k , u) and C ∗ (Z
k , u) are draw
1 2
in blue lines.


moment of Z:
 u)
C0,T (Z,
λT,K(u) (u) = , (15)
M2,Z (K(u), u)

with C0,T given by (8), K(u) as in (14) and where M2,Z (k, u) does not depend
on λ and is derived from (13)
 √ 
1 √ (k−2)/2 √ k + u 2k
M2,Z (k, u) = k + u 2k u 2k + 1 exp − .
π2k/2 Γ(k/2) 2

An illustration of this inference procedure is provided in Figure 4 for a centered

chi-square field with K = 2 degrees of freedom.

2.2.3. Student field

Following Definition 2.2, let k ≥ 3 be an integer

and consider F : R
k+1
→ R as
follows, x = (x1 , y) ∈ R × R → F (x) := x1 /
k
y /k, then Tk = F (G) is a
2
550 Biermé et al.

Fig 4. Chi-square field as in Section 2.2.2 with K = 2 degrees of freedom and λ = 0.019,
kmax = 15. Estimated K(u) as in Equation (14) (left) and associated λT,K(u) (u) (right) as
in Equation (15). Theoretical values represented by horizontal lines.

Student random field with k degrees of freedom. Assumption k ≥ 3 ensures that

all the Gi (t), for i = 1, . . . , k, cannot vanish at some point t ∈ R2 with positive
probability. Then, Tk is C 3 (R2 ), stationary, isotropic, centered and Var Tk (0) =
k/(k − 2). Furthermore, Tk and −Tk have the same distribution. In order to deal
with a centered and unit variance field, we introduce

k (t) := (k − 2)/k Tk (t) , t ∈ R2 .
T

We again assume that T k is observed on a rectangle T ⊂ R2 through its excursion


set above a fixed level u, i.e., T ∩ETk (u) = T ∩ETk (u k/(k − 2)). Let us remark
that the symmetry property of the distribution of Tk allows us to only consider
levels u that are non-negative. Moreover

Cj∗ (Tk , u) = Cj∗ (Tk , u k/(k − 2)), j = 0, 1, 2,

and the LK densities are given in the following result.

Proposition 2.7 (Student LK densities). Let k ≥ 3 and Tk be a centered
Student random field with unit variance and k degrees of freedom. Then, it holds
that
   1−k
λ(k − 1) u Γ k−1 u2 2

C0∗ (Tk , u) = √ k

2
1+ ,
3
4π 2 k−2 Γ 2 k−2
√   1−k
λ u2 2

C1∗ (Tk , u) = 1+ and

4 k−2

C2∗ (Tk , u) = P(Student(k) ≥ u k/(k − 2)).

The proof of Proposition 2.7 is postponed to Appendix (see Section A.1) as

well as the proof of the following Lemma.
 LK curvatures of excursion sets for two-dimensional random fields 551

Lemma 2.8. If k ≥ 5, the second spectral moment λStu (k) of Tk is finite and
⎧
⎨λ k−2
k
  if k is even,
λStu (k) = √ (16)
⎩λ k k−2
1
+ π (k−5)!  else,
2k−3 ( k−5 )!Γ k
2 2

where λ is the spectral moment of the underlying Gaussian fields. Note that
λStu (k) → λ as k → ∞.
Figure 5 displays a realization of a normalized Student random field Tk with
k = 4 degrees of freedom, two excursion sets and an illustration for the perfor-
mance of the three estimators C0,T (Tk , u), C1,T (Tk , u) and C2,T (Tk , u).

Fig 5. Student field as in Section 2.2.3 with 4 degrees of freedom and λ = 0.019 in a domain
of size 210 × 210 pixels. First row: A realization of a normalized Student random field (left)
and two excursion sets for u = 0 (center) and u = 1 (right). Second row: C0,T (Tk , u) (left),
C1,T (Tk , u) (center) and C2,T (Tk , u) (right) as a function of the level u. We display the
averaged values on M = 100 sample simulations (red stars) and the associated empirical
intervals (vertical red lines). Theoretical u → C0∗ (Tk , u), C1∗ (Tk , u) and C2∗ (Tk , u) are draw
in blue lines.

Parameter estimation. Suppose we observe T ∩ET (u) the excursion at level

u of T a centered Student field with unit variance, unknown degree of freedom
K and unknown second spectral moment λ of the underlying Gaussian fields.
As in Section 2.2.2, using that for a centered Student field with unit variance
552 Biermé et al.

with k degrees of freedom C2∗ (Tk , u) only depends on k, we propose the following
estimator of K:

K(u) = arg min C2∗ (Tk , u) − C2,T (T , u) , (17)

k ∈ [3, kmax ]

with C2,T as in (10) and C2∗ (Tk , u) given by Proposition 2.7. Furthermore, the
Gaussian second spectral moment of the considered Student field can be esti-
mated by
C0,T (T , u)
λT,K(u) (u) = ,
M2,Tk (K(u), u)

with C0,T as in (8), K(u) as in (17) and where M2,Tk (k, u) does not depend on
λ and is given by
   1−k
(k − 1) u Γ k−1 u2 2

M2,Tk (k, u) = √ k

2
1+ .
4π 2
3
k−2 Γ 2 k−2

An illustration of the performance of this inference procedure is given in Figure

6 for a centered Student field with K = 4 degrees of freedom.
Moreover by using Lemma 2.8 and the two previous estimators one can derive
an estimator for λStu (k).

Fig 6. Student field as in Section 2.2.3 with K = 4 degrees of freedom, λ = 0.019 and
kmax = 20. Estimated K(u) as in Equation (17) (left) and associated λT,K(u) (u) (right).
Theoretical values are represented by horizontal lines.

2.3. Shot-noise field

In this section, we consider a shot-noise field SΦ : R2 → R defined as in [6] and

[7]. It is prescribed by

SΦ (t) = bi 1ri D (t − xi ) , for t ∈ R2 ,
(xi ,bi ,ri )∈Φ
 LK curvatures of excursion sets for two-dimensional random fields 553

where Φ is a stationary Poisson point process on R2 × R∗ × R+ with intensity

measure ν LebR2 ⊗ dFB ⊗ dFR where ν > 0, FB is a probability measure on
R∗ , FR is a probability measure on R+ and D is the unit disc in R2 . Under the
condition

|b|1rD (x) dx FB (db)FR (dr) = π E[|B|] E[R2 ] < +∞,
R2 ×R∗ ×R+
with B and R two random variables with respective distributions FB and FR ,
the field SΦ is well-defined, stationary and isotropic. It is moreover integrable
and, for any t ∈ R2 ,
E[SΦ (t)] = ν ā E[B],
with ā = π E[R2 ]. Furthermore, if E[B 2 ] < +∞, SΦ is square integrable and
Var(SΦ (t)) = ν ā E[B 2 ]. We also introduce p̄ = 2πE(R).
In the case B = 1 a.s., the marginals of SΦ are Poisson distributed with
parameter νā. In the case where B is uniformly distributed in {−1, +1}, the
field SΦ is symmetric and its marginals have a Skellam distribution with both
parameters equal to νā/2 (see [30]), that is to say that they coincide with the
difference of two independent Poisson random variables with parameter νā/2.
In that case, the marginal distribution of SΦ is given by k ∈ Z → e−νā I|k| (νā),
where I. denotes the modified Bessel function of the first kind, defined as follows:
∞ (x/2)2m+n
In (x) = m=0 m!Γ(m+n+1) , for any n ∈ Z+ and any x ∈ R.
When B is positive a.s., we prove in the following result that the considered
shot-noise field SΦ is standard as in Definition 2.1.
Proposition 2.9. Let Rmax > 0. If B > 0 and R ≤ Rmax a.s., the random
field SΦ is standard at any level u ∈ R.
Proof. Since B > 0 a.s., T ∩ ESΦ (u) = T , for any u ≤ 0. Hence we may assume
that u > 0. Since R ≤ Rmax a.s. the number of discs intercepting T is bounded
by NΦ (T ) := #{xi ; d(xi , T ) ≤ Rmax } that is a Poisson random variable of
finite intensity. It follows that T ∩ ESΦ (u) must be given by the union of a
finite number N (T ∩ ESΦ (u)) ≤ NΦ (T ) of some finite intersections of discs that
are convex bodies. Since NΦ (T ) is a Poisson random variable, we clearly have
E(2N (T ∩ESΦ (u)) ) < +∞ so that Proposition 2.4 allows us to conclude that SΦ
is standard at level u.
If B changes sign, excursion sets T ∩ESΦ (u) are no more a.s. closed, nor locally
convex, nor necessarily of positive reach, however they are elementary sets in
the sense of Definition 1 in [7], with boundaries included in the discontinuity set
of SΦ so that one can compute length and total curvature. Therefore, we still
define Φi and Ci as in Definitions 1.2 and 1.4. We state the following lemma, in
the same vein as Theorem 6 in [7].
Lemma 2.10. Let B be a discrete random variable with values in Z\{0}. Then,
for any level u, the LK densities C1∗ (SΦ , u) and C0∗ (SΦ , u) are well defined and
their Fourier transforms are given, for t = 0, by

1 1
eiut C1∗ (SΦ , u)du = ν p̄ ϕS(0) (t) (ϕB + (t) − ϕB − (t)) , (18)
R 2 it
554 Biermé et al.
  
1 ν p̄2
e iut
C0∗ (SΦ , u)du = ν ϕS(0) (t) ϕB (t) − 1 + (ϕ + (t) − ϕB − (t)) ,
2
R it 4π B
(19)

where ϕZ (t) stands for the characteristic function E[eitZ ] of a random variable
Z, B + = max(B, 0) and B − = min(B, 0).
The proof of Lemma 2.10 is established in the Appendix Section A.3. Turn-
ing back to the special cases where B = 1, a.s. or B uniformly distributed in
{−1, +1} and following the same inverse Fourier procedure as Theorem 6 in [7],
we obtain the explicit formulas below.
Proposition 2.11. If B = 1 and u ∈ R+ \ Z+ , it holds that
 
(νā) u p̄2 p̄2
C0∗ (SΦ , u) = e−νā ν 1−ν + u ,
u! 4π 4πā
(νā) u  (νā)k
C1∗ (SΦ , u) = e−νā ν p̄ and C2∗ (SΦ , u) = e−νā .
2u! k!
k>u

If B is uniformly distributed in {−1, +1} and u ∈ R \ Z, it holds that


∗ ν −νā
C0 (SΦ , u) = e (I| u | − I| u +1| )(νā)
2

ν p̄2
+ (I| u −1| + I| u | − I| u +1| − I| u +2| )(νā) ,
8π
ν p̄ −νā 
C1∗ (SΦ , u) = e (I| u | + I| u +1| )(νā) and C2∗ (SΦ , u) = e−νā I|k| (νā),
4
k>u

where · denotes the greatest integer less than u.

Figure 7 (resp. Figure 8) displays two excursion sets and an illustration of the
shape of Ci∗ (SΦ , u), for i ∈ {0, 1, 2} when B = 1, a.s. (resp. when B is uniformly
distributed in {−1, +1}). Remark that if B is uniformly distributed in {−1, +1}
we do not know if SΦ is a standard random field in the sense of Definition 2.1.
Therefore in Figure 8, we do not apply the unbiased formulas of Proposition 2.5
/T
and we use Ci (SΦ , u), for i ∈ {0, 1, 2}.
Let us quote general results of [21] that can be used to obtain CLT for the
volume, the perimeter and the Euler characteristic of shot noise excursion sets
(see especially Proposition 4.1 in [21]). In particular, in our setting, by Theorem
2.4 of [21], one can get for R ≤ Rmax and B ∈ R∗ that

|T | C2,T (SΦ , u) − C2∗ (SΦ , u) −−−−→
d
N (0, σ22 (u)), with 0 < σ22 (u) < ∞.
2 T R

A similar question is the purpose of [19] where a joint CLT is established for
all the intrinsic volumes of a Boolean model. Note that the occupied domain of
a Boolean model is nothing but the excursion set above level 1 of a shot-noise
field with B = 1, a.s.
 LK curvatures of excursion sets for two-dimensional random fields 555

Fig 7. Shot-noise field as in Section 2.3 with B = 1, a.s., ν = 5 × 10−4 with random disks
of radius R = 50 or R = 100 (each with probability 0.5) in a domain of size 210 × 210 pixels.
First row: Shot-noise field random field (left) and two excursion sets for u = 7.5 (center)
and u = 14.5 (right). Second row C0,T (SΦ , u) (left), C1,T (SΦ , u) (center) and C2,T (SΦ , u)
(right) as a function of the level u. We display the averaged values on M = 100 iterations (red
stars) and the associated empirical intervals (vertical red lines). Theoretical u → C0∗ (SΦ , u),
C1∗ (SΦ , u) and C2∗ (SΦ , u) are draw with blue dots.

In our framework, we provide a non-asymptotic result to control the variance

on a finite domain T and prove consistency of Ci,T (SΦ , u), for i = 0, 1, 2.
Theorem 2.12. If B takes values in Z \ {0} and R ≤ Rmax a.s., there exists
a constant C, depending on ν, p, a and Rmax only, such that, for i = 0, 1, 2,
 
 2  1/2 1 |∂T |1
E Ci,T (SΦ , u) − Ci∗ (SΦ , u) ≤C + .
|T |1/2 |T |

|∂T |1
We recover that choosing T such that |T | → +∞ and |T | → 0, leads
to the consistency of Ci,T (SΦ , u). The proof of Theorem 2.12 is postponed to
Section A.4. To establish this result we first derive variance bounds for the weak
versions of the perimeter and the total curvature (see Proposition A.1, Section
A.2) which present interest on its own and is related to the framework of [21].

Parameter estimation. Suppose we observe T ∩ ESΦ (u) for SΦ a shot-noise

with B = 1, a.s. and for u in R+ \ Z+ with unknown parameters ν, ā and p̄.
556 Biermé et al.

Fig 8. Shot-noise field as in Section 2.3 with B uniformly distributed in {−1, +1}, ν =
3 × 10−5 and random disks of radius R = 150 or R = 100 (each with probability 0.5) in
a domain of size 210 × 210 pixels. First row: Shot-noise field random field (left) and two
/T
excursion sets for u = −0.5 (center) and u = 1.5 (right). Second row C0 (SΦ , u) (left),
/T /T
C1 (SΦ , u) (center) and C2 (SΦ , u) (right) as a function of the level u. We display the
averaged values on M = 100 iterations (red stars) and the associated empirical intervals
(vertical red lines). Theoretical u → C0∗ (SΦ , u), C1∗ (SΦ , u) and C2∗ (SΦ , u) are draw with blue
dots.

Define the estimator of νā, for some positive A,

νa(u) = arg min C2∗ (SΦ , u) − C2,T (SΦ , u) , (20)

ν a ∈ (0, A]

with C2,T as in (10) and νā → C2∗ (SΦ , u) given by Proposition 2.11. In the same
spirit, an estimator of ν p̄ is obtained using Proposition 2.11
 −1
−νa(u) (νa(u)) u
νp(u) = C1,T (SΦ , u) e , (21)
2u!

with C1,T as in (9) and νa(u) as in (20). Then, we define the following estimator
for ν
   
C0,T (SΦ , u) (νp(u))2 u(νp(u))2
ν(u) = + − , (22)

e−νa(u) (ν a(u))
u
4π 4 π νa(u)
u !
 LK curvatures of excursion sets for two-dimensional random fields 557

where C0,T (SΦ , u) is as in Equation (8), νa(u) as in (20) and νp(u) as in (21).
Finally, the obtained ν(u) can be used to isolate a(u) and p(u) in Equations (20)
and (21), respectively. An illustration of this inference procedure is provided in
Figure 9 (first row).
As mentioned earlier, the unsatisfactory quality of the Matlab function bwperim
for the estimation of C1∗ (SΦ , u) strongly impacts the final performance of the
estimation of νp(u) (see Figure 9, center panel). For this reason, we also present
2
here the simplified case where R is constant. In this particular case νa = νp 4π .
Then, we estimate νa(u) as in (20) and ν(u) by using the simplified version of
(22), i.e.,
 ⎛ ⎞
1 ⎝ C 0,T (S Φ , u) ⎠.
ν(u) = (23)
1 + u − νa(u) e − a(u) (νa(u))u
ν
u !

An illustration of the performance of these inference procedures is provided in

Figure 9 (second row).

Fig 9. First row: Shot-noise field as in Section 2.3 with B = 1, a.s., ν = 5×10−4 with random
disks of radius R = 50 or R = 100 (each with probability 0.5) in a domain of size 210 × 210
pixels. Estimated νa(u) as in Equation (20) for A = 40 (left), νp(u) as in Equation (21)
(center) and ν(u) as in (22) (right). Theoretical values are represented by horizontal lines.
Second row: Shot-noise field with B = 1, a.s., ν = 5 × 10−4 and R = 100 a.s., in a domain
of size 210 × 210 pixels. Estimated νa(u) as in Equation (20) with A = 40 (left) and ν(u) as
in (23) (right). Theoretical values are represented by horizontal lines.
558 Biermé et al.

3. Using LK densities for testing Gaussianity

Let X be a stationary, isotropic, centered with unit variance random field. We

observe an excursion of X above some levels u on the domain T , namely T ∩
EX (u) and we aim at testing whether X is Gaussian.

3.1. Testing against a Student field

In this section we are interested in testing

H0 : X is Gaussian versus
H1 : ∃ k ≥ 3, X is Student with k degrees of freedom and unit variance,
where under H0 and H1 the fields are defined through Definition 2.2, where F :
 
x ∈ R → x, under H0 and F : (x1 , y) ∈ Rk+1 → (k − 2)/k x1 / y 2 /k ,
under H1.
Under H0 and H1 the field X is symmetric and then centered. Suppose that
for 1 ≤ u1 < u2 we observe T ∩ EX (u), for u ∈ {u1 , u2 }. Notice that the fact
that under H0 and H1 the fields have unit variance enables the comparison of
the LK densities of their excursions at the same meaningful levels. Assuming
u1 and u2 are positive and greater than 1 is not necessary. Indeed the test can
be easily modified for negative values of u1 or u2 by using the symmetry of X
under H0 and H1. Furthermore, under H1 we do not assume that k is known.
Unsurprisingly, the performance of the test will depend on the unknown k.
Moreover, we underline that under H0 and H1 we do not impose any constraint
on the shape of the covariance function nor on the spectral moment other than
(A). In particular, the spectral moments of X under H0 and H1 can be different.

A test statistic. Our test statistics is built from the previously studied LK
/T
densities. As the perimeter C1 is hard to evaluate in practice (see the earlier
comment on the Matlab function bwperim), we do not use it to build a test.
Moreover, Gaussian and Student distributions differ in the tails, which would
lead to consider large values for u2 . But for large levels we do not observe many
excursion sets which deteriorates the estimation of C2∗ . Therefore to work with
intermediate values of u1 and u2 , we consider C0∗ to build our test statistic: it
is well evaluated in practice at these levels. Finally, in order to get statistic free
in λ, we take the ratio of C0∗ between to different levels u1 and u2 . This is why
we need to observe two distinct excursion sets.
Without loss of generality assume that u2 = γ u1 , for some γ > 1. For the
following empirically accessible ratios, we derive from Equation (11) and from
Proposition 2.7 that
C0∗ (X, u2 ) u21 2
= γ e 2 (1−γ ) =: v(γ, ∞) under H0, (24)
C0∗ (X, u1 )
  k−1
C0∗ (X, u2 ) (γ 2 − 1)u21 2

= γ 1 − =: v(γ, k) under H1. (25)

C0∗ (X, u1 ) k − 2 + γ 2 u21
 LK curvatures of excursion sets for two-dimensional random fields 559

Quantities v(γ, ∞) and v(γ, k) in (24) and (25), for values of u1 and γ not too
large, are quite different. Furthermore, for all k it holds that v(γ, k) > v(γ, ∞)
and k → v(γ, k) is decreasing. We build a non symmetric test where we reject
H0 whenever the associated empirical ratio is too large compared to its expected
behavior under H0. It is important to observe that v(γ, ∞) does not depend on
the spectral moment of X nor on its covariance function. The choice of u1 ≥ 1
and γ > 1 are left to the practitioner, or might be imposed by the data-set.
We expect that the larger γ is and the smaller k is the better the test will be.
However, in practice γ should not be too large as we need to estimate C0∗ (X, u2 ),
which requires to have sufficiently enough excursions above the level u2 = γ u1 .
Let T1 and T2 be two rectangles in R2 such that dist(T1 , T2 ) > 0 and |T1 | =
(N )
|T2 | > 0. For any positive integer N , we define Ti = {N t : t ∈ Ti }, for
i = 1, 2. Consider the statistics

C0,T (N ) (X, u2 )
2
Rγ,N := , (26)
C0,T (N ) (X, u1 )
1

with C0,T defined in (8).

Test with asymptotic level α. We can establish the asymptotic normality

of the statistics Rγ,N .
(N ) (N )
Proposition 3.1. Assume that X satisfies condition (A). Let T1 and T2
defined as before and Rγ,N as in (26). Then, under H0 it holds that

(N ) d
|T1 | Rγ,N − v(γ, ∞) −−−−→ N (0, Σ(u1 , u2 )),
N →∞

where Σ(u1 , u2 ) < ∞.

(N ) (N ) (N ) (N )
Proof. As the sets T1 and T2 are such that dist(T1 , T2 ) → +∞ and
(N ) (N )
|T1 | = |T2 | → +∞, from Proposition 5a in [12] it holds that
⎛ ⎞
   
(N )
 C
/T1
(N ) (X, u1 ) − E[C0 (X, u )]
(N ) ⎜ 0,T1 1 ⎟ d,H0 0 V (u1 ) 0
|T1 | ⎝ (N ) ⎠ −−−−→ N , ,
/T2 N →∞ 0 0 V (u2 )
C0,T (N ) (X, u2 ) − E[C0 (X, u2 )]
2

where u → V (u) is defined in Equation (9) of [12]. Moreover, one can prove
that the limit of
  (N ) (N ) 
/T1 /T
(N )| C1 (X, u1 ) − E[C1 1 (X, u1 )]
|T1 /T
(N )
/T
(N )
C1 2 (X, u2 ) − E[C1 2 (X, u2 )]
is non degenerate (see [20] or [26]). Then, the same decomposition as in the
proof of Proposition 2.6 leads to
⎛ ⎞
 C0,T (N ) (X, u2 ) − C0∗ (X, u2 )    
0 V (u2 ) 0
|⎝ ⎠ −−
(N ) d,H0
|T1 2 −−→ N , . (27)
C0,T (N ) (X, u1 ) − C0∗ (X, u1 ) N →∞ 0 0 V (u1 )
1
560 Biermé et al.
 
V (u2 )+v(γ,∞)2 V (u1 )
Applying the delta method, we get the result with Σ(u1 , u2 ) = C0∗ (X,u1 )2
.

 (N )
Consider the consistent empirical estimator Σu1 ,u2 := V(Rγ,N ) |T1 | of Σ(u1 , u2 ),

where V(Rγ,N ) is the empirical variance of the considered ratio. Then, from
Proposition 3.1, it holds that
" "
# (N )
# |T
$ 1 | R  #
# 1   d,H0
γ,N − v(γ, ∞) = $ Rγ,N − v(γ, ∞) −−−−→ N (0, 1).
Σu1 ,u2  N →∞
V(Rγ,N )

Take a confidence level α ∈ (0, 1) and set q1−α such that P(N (0, 1) ≥ q1−α ) = α.
We define the test φT (N ) with asymptotic level α as
φT (N ) = 1%   &. (28)

1
Rγ,N −v(γ,∞) ≥q1−α
V(Rγ,N )

Discussion on consistency. Suppose that X is a Student field with K de-

grees of freedom (K is fixed). Then, the test statistic can be decomposed as
follows
"
# (N ) 
# |T
$ 1 | R  −1/2 (N )  
γ,N − v(γ, ∞) = Σu1 ,u2 |T1 | Rγ,N − v(γ, K)
Σu1 ,u2

(N )  
+ |T1 | v(γ, K) − v(γ, ∞)

: = Σ−1/2
u1 ,u2 (I1 + I2 ),

P, H1 −1/2
where I2 −−−−→ +∞. Then, if we were able to establish that Σu1 ,u2 is bounded
N →∞
 /T /T 
away from 0 and if a joint central limit theorem for Ci (X, u1 ), Ci (X, u2 ) ,
i = 0, 1 for X a Student fields was known this would entail the consistency of
the test, i.e., PH1 (φT (N ) = 1) −→ 1. Remark that at least numerically, in our
N →∞
simulations studies it seems to be the case (see Figure 10 below).

Numerical illustrations. In Figure 10 we show the empirical PH1 (φT (N ) =

1) with φT (N ) as in (28) for two values of γ. We observe that when γ is larger the
power of the test increases. Remark that to perform this test, it is not necessary
to know the degree of freedom of the Student alternative. However, as illustrated
in Figure 10, if k gets too large or γ is too small the test fails to distinguish H1
from H0. It is easily explained by the following expansion as K → ∞,
γ−1
v(γ, K) = v(γ, ∞) 1 + O .
K
Additionally, under H0 several experiments have been reproduced to illustrate
that the test φT (N ) with empirical Σu1 ,u2 , has the desired asymptotic level α.
However, for the sake of brevity, we do not display these results here.
 LK curvatures of excursion sets for two-dimensional random fields 561

Fig 10. Student random field with unit variance and different degrees of freedom in a domain
of size 210 × 210 pixels. We display the empirical PH1 (φT (N ) = 1) on M = 100 iterations.
Left: u1 = 1, γ = 2 and, under H0, R2,N = 0.4463. Right: u1 = 1, γ = 3 and, under H0,
R3,N = 0.0549. Threshold α = 0.05 is displayed by a dashed horizontal line.

3.2. Testing against a power of a Gaussian field

Consider the alternative

H1(η) : ∃ η > 0, η = 1 such that X = sign(G) |G|η σ −1 (η),

where G is a centered Gaussian random field with unit variance and σ 2 (η) :=
21−η Γ(2η)
Γ(η) , such that the field X has unit variance under H1 (η). It is centered by
construction and it is unnecessary to impose any assumption on the covariance
1
function of G. It is straightforward to get C0∗ (X, u) = C0∗ (G, sign(u) |u| η σ(η))
and the statistics in (25) for these alternatives H1 becomes, for u2 = γu1 and
γ > 1,
 
C0∗ (X, u2 )
2/η
1 (γ 2/η − 1)u1 σ 2 (η)
= γ exp −
η := v(γ, η). (29)
C0∗ (X, u1 ) 2

Note that under H0, η = 1 and v(γ, ∞) previously defined in (24) coincides
with v(γ, 1). For γ > 1, the quantity in (29) is either smaller or larger than
(24) depending on η. Symmetrizing the previous test, it follows that to test
H0 : η = 1 against H1(η) : η = 1 we may consider the test with asymptotic
level α
φT (N ) := 1% &


1
Rγ,N −
v (γ,1) ≥q1− α
V(Rγ,N ) 2

where P(N (0, 1) ≤ q1− α2 ) = 1 − α2 . In this case we have a better understanding

of the behavior of the test statistic under H1. Under (A), applying (27) and
562 Biermé et al.

using that EX (u) = EG (sign(u) |u|1/η σ(η)), we derive ∀ η > 0

⎛ ⎞
 /T
(N )    
C 2 (X, u2 ) − C0∗ (X, u2 )⎠ 0 V (fη (u2 )) 0
|T1 | ⎝ 0 (N )
(N ) d
−−−−→ N , ,
/T N →∞ 0 0 V (fη (u1 ))
C0 1 (X, u1 ) − C0∗ (X, u1 )
(30)

with fη (u) := sign(u)|u|1/η σ(η). If we suppose that X = sign(G) |G|η σ −1 (η)

with η = 1 it holds that
"
# (N ) 
# |T
$ 1 | R 
(γ, 1) = Σ−1/2
(N )  
γ,N − v u1 ,u2 |T1 | Rγ,N − v(γ, η)
Σu1 ,u2

(N )  
+ |T1 | v(γ, η) − v(γ, 1)

:= Σ−1/2  
u1 ,u2 (I1 + I2 ),

P,H1
where from (30) we get I1 −−−−→ N (0, Σ) and I2 −−−−→ ∞. Then, if Σu1 ,u2 is
d
N →∞ N →∞
bounded from below, which is not easy to establish, we would have PH1 (φT (N ) =
1) −→2 1, which would ensure the consistency of the test φT (N ) .
T R

4. Illustration on 2D digital mammograms

In this section we consider images from a recent solid breast texture model
inspired by the morphology of medium and small scale fibro-glandular and adi-
pose tissue observed in clinical breast computed tomography (bCT) images (UC
Davis database). Each adipose compartment is modeled as a union of overlap-
ping ellipsoids and the whole model is formulated as a spatial marked point
process. The contour of each ellipsoid is blurred to render the model more re-
alistic (for details see [22], Section 2.2 and Figure 1). Finally, considered mam-
mograms images were simulated by x-ray projection. Evaluation provided in
[22] has shown that simulated mammograms and digital breast tomosynthesis
images are visually similar, according to medical experts.

Description of data-set. We consider 15 simulated 2D digital images gener-

ated by this texture model. The images were kindly provided by GE Healthcare
France, department Mammography. From a clinical point of view, radiologists
use the Breast Imaging Reporting and Data System (or BI-RADS) to classify
breast density into four categories. They go from almost all fatty tissue to ex-
tremely dense tissue with very little fat. In this latter category, it can be hard
to see small tumors in or around the dense tissue. The images we studied belong
to first three morphologic situation groups:
(A) Almost entirely adipose breasts;
(B) Scattered fibro-glandular dense breasts;
 LK curvatures of excursion sets for two-dimensional random fields 563

(C) Heterogeneously dense breasts.

One image from each group is reported in Figure 11. As remarked in Section
3 in [22], the considered simulated digital mammograms from groups (A), (B)
and (C) show a high visual realism compared to real images in these 3 different
clinical situations.

Fig 11. One image from group (A) (left), group (B) (center) and group (C) (right). Image
size: 251 × 251.

For a mathematical point of view, one difference between the considered 3

groups is the chosen intensity of the homogeneous spatial point process in [22]’s
model: νA = 3.5 × 10−3 mm−3 for group (A), νB = 5 × 10−3 mm−3 for group
(B) and νC = 2.3 × 10−3 mm−3 for the (C) one.

Gaussian test based on C0∗ . We perform the test described in Section 3.1
for these 3 data-sets composed by N = 5 images and for u1 = 1. The test
is more difficult as it may seem, due to the blurring step in the procedure to
produce the images. Then, the excursion sets have more irregular contour than
a standard shot noise field as studied in Section 2.3. Moreover, due to the small
size of the considered images (251 × 251) and of the simulated samples (N = 5),
we relax the conditions to reject H0. First, we consider larger levels for the
test, i.e., α ∈ {0.2, 0.1, 0.05}. Second, as we have access to the entire image, we
consider 1000 different tests corresponding to different values of u2 ∈ [−3, 3] for
the excursion sets. Then, we have to symmetrize the test in Section 3.1 since
we consider both negative and positive values of u2 . The test statistic in (28)
becomes
φT (N ) = 1'  −1/2 (
V(Rγ,N ) Rγ,N −
v (γ,1) ≥q1− α
2

with q1− α2 the (1 − α2 )-quantile of a N (0, 1). As the 5 images of each group
are random generations of the same parameter setting model, we estimate the
empirical variance of Rγ based on the 5 images of each group.
We compute for each image in each group the 1000 p−values associated to
the considered u2 . In Table 1, we display the number of the obtained p−values
smaller than the significant α-levels (α ∈ {0.2, 0.1, 0.05}). Furthermore, we high-
light in bold text the numbers that are larger than α × 1000 and for which H0
can be put in default.
564 Biermé et al.

Table 1
Number of p−values associated to the 1000 different values of u2 ∈ [−3, 3] that are smaller
than the significant α-levels. In bold text the numbers larger than α × 1000 for which H0 is
rejected.
Group Level Image
α 1.A 2.A 3.A 4.A 5.A
0.2 84 76 248 683 651
A 0.1 41 41 136 613 565
0.05 27 8 57 491 467

α 1.B 2.B 3.B 4.B 5.B

0.2 65 119 58 43 900
B 0.1 19 71 28 12 858
0.05 10 35 15 6 797

α 1.C 2.C 3.C 4.C 5.C

0.2 389 230 347 575 468
C 0.1 267 164 210 411 312
0.05 190 126 142 288 242

We can remark the discrepancy with the Gaussianity hypothesis for all images
of group (C) and for all considered significant α-levels. The same consideration
holds true for the last three images of group (A). Conversely, some simulated
2D digital mammograms seem to be not so far from the Gaussianity in term
of the studied Rγ,N ratio. Roughly speaking, images of group (B) are closer
to Gaussinity than group (A), which is itself closer to Gaussianity than (C).
This might be viewed in parallel with the intensities considered for each group:
νC < νA < νB . This chosen parameter setting may explain why our test rejects
Gaussianity more easily in group C than in group B. The interested reader is re-
ferred for instance to the first four images of group (B); a very different behavior
is realized by the last image of this group where the Gaussianity hypothesis is
rejected for almost all considered levels u2 . Finally, the robustness with respect
to the chosen significant α-levels can be observed in Table 1.

5. Conclusions and discussion

We have presented new statistical tools for inferring parameters and testing
Gaussianity when only a sparse observation of a 2D random field is available,
namely only the excursion set(s) above one or two level(s) within a large window.
These tools are based on the three LK curvatures of the excursion sets, which
are loosely given by the Euler characteristic, the half perimeter and the area.
The idea of considering the statistical characteristics of the excursion sets has
been originally developed for one-dimensional processes with the powerful theory
of crossings. Let us comment how our two-dimensional results can be adapted
to dimension one. First, we recall that only two LK curvatures are available for
 LK curvatures of excursion sets for two-dimensional random fields 565

Borelian subsets of R: the one-dimensional Lebesgue measure (i.e. the length)

and the Euler characteristic C0 (i.e. the number of connected components). A
similar statement as Proposition 2.5 provides an estimator of C0∗ (X, u) for any
standard stationary process X.
In particular, when X is a Gaussian process defined on R, an adapted version
of Proposition 2.6 allows to build a consistent estimator of the second spectral
moment of X, which is asymptotically normal with known asymptotic variance.
Indeed, the latter variance can be computed from the asymptotic variance V (u)
appearing in [12] with an explicit formula in dimension one (see Proposition
8 in [12]). Let us again insist on the fact that the consistent estimator of the
second spectral moment that we propose only relies on the observation of a
single excursion of X. It would be interesting to compare it with more usual
estimators based for instance on the estimation of the covariance function of X,
which requires the observation of the whole process, or with the naive empirical
estimator based on the number of crossings at different levels that is described
in [23] and used in [12], Section 4.1.2.
In the same vein, for a chi-square process defined on R, the finiteness of the
asymptotic variance of the Euler characteristic of excursion sets is proved in
Proposition 7 in [12]. Using this result, one could expect to obtain the con-
sistency of the estimator K(u) of the degrees of freedom and the estimator
λT,K(u) (u) of the Gaussian second spectral moment, which are obtained from
Equations (14) and (15) adapted to the univariate framework.
Turning back to the bivariate framework, let us discuss about the potential
improvement provided by the joint observations of several LK densities. In Sec-
tion 4, in order to analyse the simulated 2D digital mammograms data-set, we
have used C0∗ (·, u). Let us now investigate the use of C2∗ (·, u) and the associated
unbiased estimator in Proposition 2.5. In Figure 12 we display C2,T (·, u) for each
image in the 3 groups compared with the tail distribution with zero mean and
unit variance ψ(u) (see Equation (11)). In order to improve the readability for
extreme values of u, in Figure 12 we chose a logarithmic scale for the y−axis.

Fig 12. C2,T (·, u) in (10) for the 3 groups for different values u. The full black line represents
ψ(u) in Equation (11), i.e., the Gaussian tail distribution with zero mean and unit variance.
We take here a logarithmic scale for the y−axis.
566 Biermé et al.

A clear assessment is that the two LK densities C0∗ and C2∗ bring two different
points of view on the 15 images. For instance, for extreme levels u, say above
level 2.5, the C2,T (·, u) of the mammograms clearly have a different behavior
than the Gaussian one (see Figure 12). Moreover, the difference is greater for
images of type (C) than for images of type (B), which is already greater than
images of type (A). A similar behavior was not observed with the C0,T (·, u). This
point clearly deserves to be studied. As it involves larger levels, it is certainly
related to extreme values theory.
Following the idea of taking advantage of the joint observation of several LK
densities, our study can be continued with the development of efficient numerical
tools adapted to different sorts of medical images. For instance, considering 2D
x-ray bone images in order to detect osteoporosis, one can expect getting infor-
mation on the bone density through C2∗ and on the bone connectivity through
C0∗ . Both quantities really affect the mechanical bone resistance and should be
both evaluated with the target of reducing the number of misclassified patients.
Another direction that merits to be explored in the use of 3D images that are
now quite common for medical diagnoses. Hence, one should improve our meth-
ods by including now four LK curvatures that are directly linked to the bone
tissue itself instead of projected images.
In this article, we have not fully used the joint estimation of two (nor three)
LK densities. Indeed, we are not aware of any joint central limit theorem for
the LK densities of excursion sets of stationary random fields on R2 , except in
the case of a Boolean model in [19] or in the case of pixelated images in [29]
and binary images in [13]. Those results do not apply in our context. If it exists,
such a theorem would enable to build asymptotic confident regions that take
into account the whole information of LK densities of the observed excursion
set.
Finally, in the present study, we became aware that the Matlab function
bwperim seems to not perform very well on the excursion sets of the two-
dimensional fields we considered, yielding an overestimation of C1∗ . The observed
pixelization errors arise numerically due to the discretized representation of the
smooth level set from a pixelated image, we mentioned this drawback in Section
2. It should be interesting to link our study with the numerous literature on
pixelization effects. The already mentioned papers [29] and [13] should provide
good tools for such an approach.

Appendix A: Technical proofs

A.1. Proof of Proposition 2.7

Computation of the LK densities for Tk . We use (7) after observing that

Tk = F (G) with G = (G0 , G1 , . . . , Gk ) and F : (x, y) ∈ R × Rk → ||y||/
x√
k
. Then

) √ *
−1 k
F ([u, ∞))) = (x, y) ∈ R × R : ||y|| ≤
k
x , if u > 0. (31)
u
 LK curvatures of excursion sets for two-dimensional random fields 567

Fig 13. Case k = 2. The domain F −1 ([u, ∞)) in (31) (area delimited by plain lines) and
k
the tube T ube(F, ρ) around (area delimited by dotted lines), where ||y||2 = 2
i=1 yi , x0 :=
√
k  √
ρ sin(θk ) =  u k and ρ := cosρθ = ρ 1 + uk2 with θk := arctan uk .
1+ 2 k
u

Denote by Fk the cumulative chi distribution (the square root of a chi-square

random variable) with k degrees of freedom. Using (31) and u > 0, we have that
 √
e−x /2
2
k
P G(0) ∈ T ube(F, ρ) = √ Fk x + ρ dx + O(ρk ),
x≥−x0 2π u
where we have neglected the circular part of T ube(F, ρ). Then, using x0 =
√
 −1 
u
k
1 + k
u2 , ρ = ρ 1 + uk2 and that Fk (0) = Fk (0) = 0, we get

 √
e−x /2
2
k
P G(0) ∈ T ube(F, ρ) = √ Fk x dx
x≥0 2π u
 √
e−x /2 
2
k k
+ρ 1+ 2 √ Fk x dx
u x≥0 2π u
 √
e−x /2 
2
1 2 k k
+ ρ 1+ 2 √ Fk x dx + O(ρk ).
2 u x≥0 2π u
The latter formula is established for u > 0. The case u < 0 is derived using
symmetry arguments. If u = 0, F −1 ([0, ∞)) = R+ × Rk and Equation (6)
becomes

e−x /2
2
1 1 ρ2
P G(0) ∈ T ube(F, ρ) = √ × 1dx = + ρ √ + × 0 + O(ρ3 ).
x≥−ρ 2π 2 2π 2
Then, we derive the following formulas, valid for k ≥ 3,
⎧, √
⎪ +∞ e−x2 /2
⎪
⎨ 0 √
2π
F k ( k
u x) dx if u > 0,
M0 (Tk , u) = P(Student(k) ≥ u) = 12 if u = 0,
⎪
⎩1 − , +∞ e−x
⎪ 2 /2 √
√
0
F ( k x) dx
2π k −u else;
568 Biermé et al.
% , +∞ 2 √
e−x /2
(1 + k 1/2
u2 )
√ Fk ( |u|k x) dx if u = 0,
M1 (Tk , u) = 0 2π
√1 if u = 0;
2π
% , +∞ −x2 /2
√
sign(u)(1 + k
u2 )
e√
Fk ( |u|k x) dx if u = 0,
M2 (Tk , u) = 0 2π
0 if u = 0.

These quantities can be simplified using the following result E[|N (0, 1)|k ] =
k  
√1 2 2 Γ k+1 . It follows from simple computations that
π 2
1−k
1 u2 2
M0 (Tk , u) = P(Student(k) ≥ u), M1 (Tk , u) = √ 1+ ,
2π k
 k−1 
k−1 u Γ u 2 1−k
2
M2 (Tk , u) = √ √ 2 1 + .
2 π k Γ k2 k

Then, it is straightforward to derive the Student LK densities from (7). The

above quantities do not depend on the parameter dimension of the Student field
Tk , which is equal to 2 in our bivariate context. 

Computation of the second spectral moment of Tk (Formula (16)).

The Student field is isotropic, the second spectral moment is
λStu := [V(Tk (0) )](1,1) = V(∂1 Tk (0))
(N ) (N ) ∂f
where for t = (T1 , T2 ), ∂1 f (t) = (N ) (t). Note that
∂T1

√ G0 (t)
Tk (t) = k , t ∈ R2 ,
Zk (t)
k
where Zk = i=1 Gi (t)2 is a chi-square random field with k degrees of freedom
independent of the Gaussian field G0 . We have
  
√ ∂1 G0 (t) Zk (t) − G0 (t) ∂√
1 Zk (t)
√
2 Zk (t) ∂1 G0 (t) 1 ∂1 Zk (t)
∂1 Tk (t) = k = k  − G0 (t) 3/2
,
Zk (t) Zk (t) 2 Zk (t)

which is centered, as G0 and ∂1 G0 are independent, centered and independent

of (Zk , ∂1 Zk ). Therefore, we get
 
1 1 ∂1 Zk (0)2
λStu = E[(∂1 Tk (0))2 ] = k λ E + E ,
Zk (0) 4 Zk (0)3
where we used that G0 and ∂1 G0 are independent, centered, with unit variance
and independent of (Zk , ∂1 Zk ) together with E[∂1 G0 (0)2 ] := λ. To complete the
computation of λStu , we first have that, for k ≥ 3, E Zk1(0) = k−2 1
. Second,
using the definition of a chi-square distribution, it holds that
k
∂1 Zk (0)2 i,j=1 Gi (0)Gj (0)∂1 Gi (0)∂1 Gj (0)
=4  k  .
Zk (0) 3 2 3
i=1 Gi (0)
 LK curvatures of excursion sets for two-dimensional random fields 569
 
Taking expectation and using that ∂1 Gi is independent of (∂1 Gj , Gj )j=i , Gi
and centered, together with the definition of λ, we derive that
     
∂1 Zk (0)2 G1 (0)2 G1 (0)2
E = 4kλ E  k  = 4kλ E  3 ,
Zk (0)3 2 3 G1 (0)2 + Zk−1 (0)
i=1 Gi (0)

where Zk−1 is a chi-square random field with k − 1 degrees of freedom and

independent of G1 . Then,
 
G1 (0)2
I(k) := E  k 
2 3
i=1 Gi (0)
 
z 2 −1
k−1
1 x2 −(x2 +z)/2
= √ e k−1   dzdx
2π R R+ (x2 + z)3 2 2 Γ k−1 2
 
1 x2
e−x (1+y)/2 xk−3 y 2 −1 x2 dydx
2 k−1
= √ k−1  k−1  6 3
2π2 2 Γ 2 R R+ x (1 + y)
 k−3  
1 y 2 (1 + y)1/2 k−5 −x2 (1+y)/2
= k−1   5/2
√ x e dx dy,
2 2 Γ k−1 2 R+ (1 + y) R 2π

where we made the change of variable z = x2 y. We recognize the (k − 5)-

th moment of a N (0, (1 + y)−1 ) random variable in the last line. Let μ() :=
E[N (0, 1) ] with  ∈ N, it holds that
 k−3  
μ(k − 5) u 2 μ(k − 5) k−1 1
I(k) = k−1   k/2
du = k−1   B , ,
2 2 Γ k−1 R+ (1 + u) 2 2 Γ k−1 2 2
2 2

where B(x, y) denotes the beta function. It simplifies in

  1
μ(k − 5) Γ k−1 Γ 2 √ 1
I(k) = k−1   2
 k
 = πμ(k − 5) k−1   .
k−1
2 2 Γ 2 Γ 2 2 2 Γ k2
Finally, for k ≥ 5,
 
1 √ E[N (0, 1)k−5 ]
λStu = λ k + π  
k−2 k−1
2 2 Γ k
⎧ 2
⎨λ k
k−2 if k is even,
√
= 1 π (k−5)! 
⎩λ k k−2 + 2k−3 ( k−5 )!Γ k
else.
2 2

A.2. An auxiliary result

In the following we state and prove an auxiliary result which is crucial in the
derivation of Proposition 2.11 and Theorem 2.12. In the sequel we will make
570 Biermé et al.

extensive use of notations and definitions introduced in [7], some of them will
be redefined, for the others the reader can refer to [7]. As in Section 2.3, we
consider a shot-noise field SΦ prescribed by

SΦ (t) = bi 1ri D (t − xi ) , t ∈ R2 ,
(xi ,bi ,ri )∈Φ

where Φ is a stationary Poisson point process on R2 × R∗ × R+ with intensity

measure ν LebR2 ⊗ dFB ⊗ dFR where ν > 0, FB is a probability measure on R∗ ,
FR is a probability measure on R+ and D is the unit disc in R2 , well defined
under the condition E[|B|] E[R2 ] < +∞, with B and R two random variables
with respective distributions FB and FR .
Since excursion sets of SΦ are not always UPR, we need to introduce the level
perimeter LP and the level total curvature LTC integrals as introduced in [7].
Using this setting, the excursions ESΦ (u) are elementary sets (see Definition 1
[7]). Then, for a bounded open set U and h a continuous bounded function, one
can define

LPSΦ (h, U ) := 2 h(u)Φ1 (ESΦ (u), U )du
 R 
LTCSΦ (h, U ) := h(u)TC(∂ESΦ (u), U )du = 2π h(u)Φ0 (ESΦ (u), U )du.
R R

Let us consider the distribution F (dm) = FB (db) × FR (dr) for m = (b, r) ∈

R × (0, +∞) ⊂ R2 and denote gm (x) = b1rD (x) with D the unit disk. The
function gm is an elementary function on R2 (see Definition 3 in [7]) with dis-
continuity set given by
Sgm = Rgm = ∂Dr ,
where ∂Dr is the circle of radius r. According to Theorem 4 of [7], a.s. for any
U bounded open set, the shot noise field SΦ is an elementary function on U
with discontinuity set given by U ∩ SSΦ with

SS Φ = ∪ τx ∂Dr = RSΦ ∪ ISΦ ,

(x,m)∈Φ

with
=
RS Φ = ∪ τx ∂Dr \ISΦ for ISΦ = ∪ τx1 ∂Dr1 ∩ τx2 ∂Dr2 ,
(x,m)∈Φ (x1 ,m1 ),(x2 ,m2 )∈Φ

where τx y = x + y.
By Proposition 2 and Theorem 4 of [7], for H a primitive of h and H1 the
one dimensional Hausdorff measure (recall that H1 (A) = |A|1 ), it follows that
 
LPSΦ (h, U ) = [H(SΦ\{(x,m)} (z) + b+ )
(x,m)∈Φ τx ∂Dr ∩U

− H(SΦ\{(x,m)} (z) + b− )]H1 (dz)

 LK curvatures of excursion sets for two-dimensional random fields 571

Fig 14. Illustration with an elementary function given by f = 1Dr1 + 1Dr2 with 0 < r1 < r2 .
Left: its discontinuity set Sf = Rf ∪ If with If given by the red crosses. Center and Right:
Ef (1) and Ef (2): the boundaries are regular except at corner points given by If .

where we use the fact that H1 (τx ∂Dr ∩ τx ∂Dr ) = 0 a.s. for all different
points of Φ and denote b+ = max(b, 0) and b− = min(b, 0). Besides, we have
LTCSΦ (h, U ) = RSΦ (h, U ) + ISΦ (h, U ), where
 
RSΦ (h, U ) = [H(SΦ\{(x,m)} (z) + b+ )
(x,m)∈Φ τx ∂Dr ∩U

− H(SΦ\{(x,m)} (z) + b− )] sign(b)κ∂Dr (z)H1 (dz),

with κ∂Dr (z) = 1

r the curvature of the curve ∂Dr at point z ∈ ∂Dr , and
=
 
1
ISΦ (h, U ) = Δb1 ,b2 H(SΦ\{(x1 ,m1 ),(x2 ,m2 )} (z))βb1 ,b2 (z),
2 (x1 ,m1 ) τx1 ∂Dr1
∈Φ z∈ ∩τ
(x2 ,m2 ) x2 ∂Dr2 ∩U

 − − −
where Δb1 ,b2 H(t) = H(t + b+ 1 + b2 ) + H(t + b1 + b2 )-H(t + b1 + b2 ) − H(t +
+ +

b−
1 + b2 ) and for z ∈ τx1 ∂Dr1 ∩ τx2 ∂Dr2 ,
+

 
x1 − z x2 − z
βb1 ,b2 (z) = dist sign(b1 ) , sign(b2 ) ,
r1 r2
with dist the geodesic distance on the circle.
We can now state the following result that has interest on its own: it allows to
obtain uniform bounds for the integral functionals, these bounds being usually
hard to obtain when considering an excursion set at a fixed level.
Proposition A.1 (Moments of LPSΦ and LTCSΦ ). Assume that E(R2 ) < +∞
and E(|B|) < +∞ and recall that p = 2πE(R) and a = πE(R2 ). Then, for any
U bounded open set and any continuous function h with primitive H, one has
,
i). E (LPSΦ (h, U ))) = νp|U | R E ([H(SΦ (0)+b+ )−H(SΦ (0)+b−)]) FB (db);
,
E (LTCSΦ (h, U )) = 2πν|U | R E [H(SΦ (0)+b+ )−H(SΦ (0)+b− )] sign(b)FB (db)
ii). 2 ,
+ (νp)
2
|U | R2 E (Δb1 ,b2 H(SΦ (0))) FB (db1 )FB (db2 ).

Moreover, when R ≤ Rmax a.s. for some positive constant Rmax and H is
bounded, one has
2
iii). Var (LPSΦ (h, U ))) ≤ CH |U | (Rmax νp) + νa ;
572 Biermé et al.
)   *
iv). Var (LTCSΦ (h, U )) ≤ CH |U | (Rmax ν)2 + ν + (νp)2 (Rmax νp)2 + νa + 1 ,

for a constant CH that only depends on H ∞ and such that CH ≤ 27 π H 2

∞.

Proof. i). Let us introduce


[H(SΦ (z) + b+ ) − H(SΦ (z) + b− )]H1 (dz)
(1)
fh ((x, m), Φ) =
τx ∂Dr ∩U

 (1)
such that LPSΦ (h, U ) = (x,m)∈Φ fh ((x, m), Φ\{(x, m)}). It follows from the
reduced Campbell formula (see e.g. [4] Theorem 1.4.3),

  (1)
E LPSΦ (h, U ) = E fh ((x, m), Φ) νdxF (dm),
R2 ×R2

with E fh ((x, m), Φ) = E ([H(SΦ (0)+b+ )−H(SΦ (0)+b− )]) H1 (τx ∂Dr ∩ U )
(1)

by stationarity of SΦ . Hence,
 
ELPSΦ (h, U )) = ν|U | H1 (∂Dr )FR (dr) E([H(SΦ (0) + b+ ) − H(SΦ (0) + b− )])FB (db)
R R

 
= |U |νp E [H(SΦ (0) + b+ ) − H(SΦ (0) + b− )] FB (db).
R

iii). It holds that

2
 (1)
(LPSΦ (h, U )) = fh ((x1 , m1 ), Φ\{(x1 , m1 )})2
(x1 ,m1 )∈Φ
=
 (2)
+ fh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x2 , m2 )}),
(x1 ,m1 )
(x2 ,m2 )
∈Φ

where
(2)
fh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x2 , m2 )})
(1) (1)
= fh ((x1 , m1 ), Φ\{(x1 , m1 )})fh ((x2 , m2 ), Φ\{(x2 , m2 )})
 
= Kh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x2 , m2 )})(z1 , z2 )H1 (dz2 )H1 (dz1 )
τx1 ∂Dr1 ∩U τx2 ∂Dr2 ∩U

for Kh ((x1 , m1 ), (x2 , m2 ), Φ)(z1 , z2 ) equals to

−
1 ) − H(SΦ (z1 ) + τx2 gm2 (z1 ) + b1 )]
[H(SΦ (z1 ) + τx2 gm2 (z1 ) + b+
−
2 ) − H(SΦ (z2 ) + τx1 gm1 (z2 ) + b2 )].
× [H(SΦ (z2 ) + τx1 gm1 (z2 ) + b+

Since R ≤ Rmax a.s, for any |z − z  | > 2Rmax , the variables SΦ (z) and SΦ (z  )
are independent.
 LK curvatures of excursion sets for two-dimensional random fields 573

Firstly, it holds
⎛ ⎞

⎜ 
=
(2) ⎟
E⎝ fh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x2 , m2 )})⎠
(x1 ,m1 )
(x2 ,m2 )
∈Φ
 
(2)
= E(fh ((x1 , m1 ), (x2 , m2 ), Φ)ν 2 dx1 F (dm1 )dx2 F (dm2 ) := I1 + I2 ,
R2 ×R2 R2 ×R2

where we split the integral over the domains |x1 −x2 | ≤ 4Rmax and |x1 −x2 | >
4Rmax . Indeed, for |x1 − x2 | > 4Rmax we have |zj − xi | > 3Rmax for any
{i, j} = {1, 2} and zj ∈ τxj ∂Drj so that τxi gmi (zj ) = 0. By independence,
we get that
 
 2 (1)
I2 = E(LPSΦ (h, U )) − 1|x1 −x2 |≤4Rmax E(fh ((x1 , m1 ), Φ)×
R2 ×R2 R2 ×R2
(1)
× E(fh ((x2 , m2 ), Φ)ν 2 dx1 F (dm1 )dx2 F (dm2 ).

Hence  2
I1 + I2 = E(LPSΦ (h, U )) + ν 2 Δh (U ),
with
 
Δh (U ) = 1|x1 −x2 |≤4Rmax Δh (U, (x1 , m1 ), (x2 , m2 ))dx1 F (dm1 )dx2 F (dm2 ),
R2 ×R2 R2 ×R2

for Δh (U, (x1 , m1 ), (x2 , m2 )) equals to

(2) (1) (1)
E(fh ((x1 , m1 ), (x2 , m2 ), Φ)) − E(fh ((x1 , m1 ), Φ))E(fh ((x2 , m2 ), Φ)).
(1)
As H is bounded, we have |fh ((x, m), Φ)| ≤ 2 H ∞H
1
(τx ∂Dr ∩ U ), so that

|Δh (U, (x1 , m1 ), (x2 , m2 ))| ≤ 8 H ∞ H (τx1 ∂Dr1

2 1
∩ U )H1 (τx2 ∂Dr2 ∩ U ).

It follows that, for CH := 27 π H 2

∞,
 2
Δh (U ) ≤ 8 H ∞ |U |
2
× π(4Rmax ) 2
H (∂Dr )F (dr)
1
≤ CH Rmax
2
|U |p2 .
R

Secondly,
 
(1) (1)
E fh (x1 , m1 ), Φ\{(x1 , m1 )})2 = E fh ((x, m), Φ)2 νF (dm)dx
(x1 ,m1 )∈Φ
R2 ×R2

≤ ν|U |4H2∞ H1 (∂Dr )2 F (dr) ≤ 16πH2∞ |U |νa.
R

Hence, we obtain
 2 
Var (LPSΦ (h, U )) ≤ CH |U | Rmax ν 2 p2 + νa .
574 Biermé et al.

ii). iv). Recall that LTCSΦ (h, U ) = RSΦ (h, U ) + ISΦ (h, U ). Then, we bound
the expectation and variance of RSΦ (h, U ) and ISΦ (h, U ) separately.
First consider RSΦ (h, U ), we prove similarly
E (RSΦ (h, U ))
 
H1 (∂Dr )  
= ν|U | F (dr) E [H(SΦ (0) + b+ ) − H(SΦ (0) + b− )] sign(b)FB (db),
R r R

 
= 2πν|U | E [H(SΦ (0) + b+ ) − H(SΦ (0) + b− )] sign(b)FB (db),
R

and
 
 h (U ), with |Δ
E((RSΦ (h, U ))2 ) = E(RSΦ (h, U ))2 +Δ  h (U )| ≤ CH |U | R2 ν 2 +ν .
max

Now, let us focus on ISΦ (h, U ). We have

=

1 (2)
ISΦ (h, U ) = fh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x2 , m2 )}),
2
(x1 ,m1 ),(x2 ,m2 )∈Φ

where
(2)

fh ((x1 , m1 ), (x2 , m2 ), Φ) = Δb1 ,b2 H(SΦ (z))βb1 ,b2 (z).
z∈τx1 ∂Dr1 ∩τx2 ∂Dr2 ∩U

It holds that

E (ISΦ (h, U ))
 
1 (2)
= E fh ((x1 , m1 ), (x2 , m2 ), Φ) ν 2 dx1 F (dm1 )dx2 F (dm2 )
2 R4 R4
  
1
= E(Δb1 ,b2 H(SΦ (z)))βb1 ,b2 (z)ν 2 dx1 F (dm1 )dx2 F (dm2 )
2 R4 R4 τx ∂Dr ∩
z∈ τ 1 1
x2 ∂Dr2 ∩U

ν 2 |U |
= E(Δb1 ,b2 H(SΦ (0)))×
2 R2
⎛ ⎞
 
⎜ −z x−z ⎟
⎜
×⎝ dist(sign(b1 ) , sign(b2 ) ))dxdr1 dr2 ⎟
r1 r2 ⎠ FB (db1 )FB (db2 )
4 R ∂Dr1
z∈ ∩τ
x ∂Dr2

with, by kinematic formula (see Lemma 1 in [7] for instance),

  −z x−z
dist(sign(b1 ) , sign(b2 ) ))dx = H1 (∂Dr1 )H1 (∂Dr2 ).
R2 r1 r2
z∈∂Dr1 ∩τx ∂Dr2

It follows that

(νp)2 |U |
E (ISΦ (h, U )) = E(Δb1 ,b2 H(SΦ (0)))FB (db1 )FB (db2 ).
2 R2
 LK curvatures of excursion sets for two-dimensional random fields 575

Let us write
2
(ISΦ (h, U ))
⎛ ⎞2
=

1⎝
fh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x2 , m2 )})⎠
(2)
=
4
(x1 ,m1 ),(x2 ,m2 )∈Φ
=

1 (4)
= fh ((x1 , m1 ), . . . , (x4 , m4 ), Φ\{(x1 , m1 ), . . . , (x4 , m4 )}
4
(x1 ,m1 ),...,(x4 ,m4 )∈Φ
=
 (3)
+ fh ((x1 , m1 ), (x2 , m2 ), (x3 , m3 ), Φ\{(x1 , m1 ), (x2 , m2 ), (x3 , m3 )}
(x1 ,m1 )
(x2 ,m2 ) ∈Φ
(x3 ,m3 )

=

1 2
(2)
+ fh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x2 , m2 )})
4
(x1 ,m1 ),(x2 ,m2 )∈Φ

:= I1 (Φ) + I2 (Φ) + I3 (Φ),

where
(4)
fh ((x1 , m1 ), . . . , (x4 , m4 ), Φ\{(x1 , m1 ), . . . , (x4 , m4 )})
(2)
= fh ((x1 , m1 ), (x2 , m2 ), Φ\{(x1 , m1 ), (x1 , m2 )})
(2)
× fh ((x3 , m3 ), (x4 , m4 ), Φ\{(x3 , m3 ), (x4 , m4 )})
 
= Δb1 ,b2 H(SΦ\{(x1 ,m1 ),...,(x4 ,m4 )} (z) + τx3 gm3 (z) + τx4 gm4 (z))
τx1 ∂Dr1 τx3 ∂Dr3
z∈ z ∈
∩τx2 ∂Dr2 ∩U ∩τx4 ∂Dr4 ∩U

× βb1 ,b2 (z)Δb3 ,b4 H(SΦ\{(x1 ,m1 ),...,(x4 ,m4 )} (z ) + τx1 gm1 (z ) + τx2 gm2 (z ))βb3 ,b4 (z ),

and, by symmetry,
(3)
fh ((x1 , m1 ), (x2 , m2 ), (x3 , m3 ), Φ\{(x1 , m1 ), (x2 , m2 ), (x3 , m3 )})
 
= Δb1 ,b2 H(SΦ\{(x1 ,m1 ),...,(x3 ,m3 )} (z) + τx3 gm3 (z))
τx1 ∂Dr1 τx3 ∂Dr3
z∈ ∩τ z  ∈ ∩τ
x2 ∂Dr2 ∩U x1 ∂Dr1 ∩U

× βb1 ,b2 (z)Δb3 ,b1 H(SΦ\{(x1 ,m1 ),...,(x3 ,m3 )} (z  ) + τx1 gm1 (z  ))βb1 ,b3 (z  ).

Hence, we can also write

 2 
 h (U ),
E (I1 (Φ)) = E (ISΦ (h, U )) + ν 4 Δ


 h (U ) equals to
with Δ

1 
 h (U, (x1 , m1 ), . . . , (x4 , m4 ))dx1 F (dm1 ) . . . dx4 F (dm4 ),
1|x1 −x3 |≤4Rmax Δ
4 R16
576 Biermé et al.

for

 h (U, (x1 , m1 ), . . . , (x4 , m4 ))
Δ
 
≤ 2(4 H ∞ )2 βb1 ,b2 (z)βb3 ,b4 (z  )dx1 F (dm1 ) . . . dx4 F (dm4 ),
τx1 ∂Dr1 τx3 ∂Dr3
z∈ ∩τ z  ∈ ∩τ
x2 ∂Dr2 ∩U x4 ∂Dr4 ∩U


 h (U )| ≤ CH |U |R2 p4 . Similar arguments lead to
Therefore, |Δ max

|E(I2 (Φ))| ≤ CH ν 3 |U |p2 a and |E(I3 (Φ))| ≤ CH ν 2 |U |p2 ..

 2
Finally, since Var (ISΦ (h, U )) = E(I1 (Φ))+E(I1 (Φ))+E(I1 (Φ))− E(ISΦ (h, U )) ,
we obtain
 
Var (ISΦ (h, U )) ≤ CH |U |(νp)2 (Rmax νp)2 + νa + 1 .

A.3. Proof of Lemma 2.10

From now on we assume that B is a discrete random variable with values in Z∗ .

Then, the random field SΦ has values in Z, and it follows that for all u ∈ R,
ESΦ (u) = ESΦ (u). For u ∈ Z, ,choosing h a continuous function with compact
support in (0, 1) and such that R h(t)dt = 1 one has

1 1
Φ1 (ESΦ (u), U ) = LPSΦ (τu−1 h, U ) and Φ0 (ESΦ (u), U ) = LTCSΦ (τu−1 h, U ),
2 2π
where τu−1 h(t) = h(t − (u − 1)) vanishes outside (u − 1, u). Now, considering for
t = 0, the function ht (u) = eiut we can compute the Fourier transforms of u →
|U | E (Φj (ESΦ (u), U )) as 2|U | E (LPSΦ (ht , U )) for j = 1 and 2π|U | E (LTCSΦ (ht , U ))
1 1 1

for j = 0. We obtain the right members of (18) and (19) using Proposition A.1
i) and ii), with Ht (u) = e it−1 . The link with Cj∗ (SΦ , u) is given in the next
itu

proposition.
Proposition A.2. For all u ∈ Z,
 /T Φ (E (u),T̊ ) 2 1/2
|∂T |1
• E C1 (SΦ , u) − 1 S|TΦ | ≤ |T | ;
 /T Φ (E (u),T̊ ) 2 1/2
• E C0 (SΦ , u) − 0 S|TΦ | ≤ 2νp |∂T |1
|T | , as soon as νp|∂T |1 > 1.

E(Φj (ESΦ (u),U ))

As a consequence Cj∗ (SΦ , u) exists for j ∈ {0, 1} and is equal to |U | ,
for any bounded open set U .
Proof. Let us remark that the boundary contributions (see Figure 2 of [7]) may
be controlled using the fact that a.s.

C1 (SΦ , u, T ) − Φ1 (ESΦ (u), T̊ ) ≤ |∂T |1

 LK curvatures of excursion sets for two-dimensional random fields 577

and
1
C0 (SΦ , u, T ) − Φ0 (ESΦ (u), T̊ ) ≤ # (SSΦ ∩ ∂T ) + 1,
2
where we use that T being a rectangle the curvature is 0 on its boundary except
at corner points where it is π2 . Dividing by |T | we immediately get the first point.
For the second one, since SSΦ = ∪i τxi ∂Dri , with τxi ∂Dri ∩ τxj ∂Drj ∩ ∂T = ∅
a.s. as soon as i = j, we have

# (SSΦ ∩ ∂T ) = # (τxi ∂Dri ∩ ∂T ) a.s.
i

It follows that

E (# (SSΦ ∩ ∂T )) = # (τx ∂Dr ∩ ∂T ) νdxFR (dr).
R2 ×R+

But, again kinematic formula (see Lemma 1 of [7]) yields


# (τx ∂Dr ∩ ∂T ) dx = 4r|∂T |1 ,
R2

so that
2
E (# (SSΦ ∩ ∂T )) = νp|∂T |1 .
π
Moreover,

2 4
Var (# (SSΦ ∩ ∂T )) = (#(τx ∂Dr ∩ ∂T )) νdxFR (dr) ≤ νp|∂T |1 ,
R2 ×R+ π

since # (τx ∂Dr ∩ ∂T ) ≤ 2. Therefore,

E((#(SSΦ ∩ ∂T ))2 ) =Var(#(SSΦ ∩ ∂T )) + E(#(SSΦ ∩ ∂T ))2

4 1 2
≤ νp|∂T |1 (1 + νp|∂T |1 ) ≤ 2 (νp|∂T |1 ) ,
π π

as soon as νp|∂T√|1 ≥ 1. The second bound is obtained dividing by |T | and

using the fact that 2/2 + 1 ≤ 2. Now we have
 
/T E(Φj (ESΦ (u), T̊ )) /T Φ1 (ESΦ (u), T̊ )
|E Cj (SΦ , u) − |≤E Cj (SΦ , u) −
|T̊ | |T |
|∂T |1
≤ max(1, νp) ,
|T |

by the previous bound. This allows us to prove that Cj∗ (SΦ , u), j = 0, 1 exist
/T
as the limit of E(Cj (SΦ , u)) when T R2 .
578 Biermé et al.

A.4. Proof of Theorem 2.12

We work under the same assumptions as in Sections A.2 and A.3. We further-
 1/2
more assume that R ≤ Rmax a.s. Denote by Z 2 := E[Z 2 ] , then we write
- -
- Φj (ESΦ (u), T̊ ) -
- /T -
Cj∗ (SΦ , u) 2
/T
Cj (SΦ , u) − ≤ -Cj (SΦ , u) − -
- |T | -
- -2
- Φ (E (u), T̊ ) -
- j SΦ -
+- − Cj∗ (SΦ , u)- .
- |T | -
2

The first term is bounded by max(2νp, 1) |∂T |1

|T | by Proposition A.2. Moreover,
using R ≤ Rmax a.s. and choosing h, a non-negative continuous function with
compact support in (0, 1) such that R h(t)dt = 1,
- -2
- Φ (E (u), T̊ ) -
- 1 SΦ ∗ - 1 C  2 
- −C1 (SΦ , u)- = Var LPSΦ (τu−1 h, T̊ ) ≤ Rmax ν 2 p2 +νa ,
- |T | - 4|T |2 4|T |
2

by Proposition A.1, since we can choose a primitive of τu−1 h uniformly bounded

by 1. Similarly, we obtain
- -2
- Φ0 (ESΦ (u), T̊ ) - 1
- −C ∗
(S , u) - = Var LTCSΦ (τu−1 h, T̊ )
- |T | 0 Φ - 4π 2 |T |2
2
C 2  
≤ (Rmax ν) + ν + (νp)2 (Rmax νp)2 + νa + 1 .
2π |T |
2

Finally,

1   2
4πRmax
C2 (SΦ , u) − C2∗ (SΦ , u)22 =
/T
Cov 1SΦ (x)≥u , 1SΦ (y)≥u dxdy ≤ ,
|T |2 T ×T |T |

using the fact that SΦ (x) and SΦ (y) are independent as soon as |x−y| > 2Rmax .
By the triangle inequality, we obtain the result for C.
j,T (SΦ , u), j ∈ {0, 1, 2}.

Acknowledgments

The authors acknowledge Z. Li, (GE Healthcare France, department Mammogra-

phy) for simulated 2D digital mammograms data-sets and the related discussion.
This work has been partially supported by the project LEFE-MANU MULTI-
RISK. The authors sincerely express their gratitude to the referees for their
valuable comments that helped to motivate the study, make the paper clearer
and correct a normalization constant in the previous version of this paper.
 LK curvatures of excursion sets for two-dimensional random fields 579

References

[1] R. J. Adler and J. E. Taylor. Random fields and geometry. Springer Mono-
graphs in Mathematics. Springer, New York, 2007. MR2319516
[2] R. J. Adler and J. E. Taylor. Topological complexity of smooth random
functions, volume 2019 of Lecture Notes in Mathematics. Springer, Hei-
delberg, 2011. Lectures from the 39th Probability Summer School held in
Saint-Flour, 2009, École d’Été de Probabilités de Saint-Flour. [Saint-Flour
Probability Summer School]. MR2768175
[3] J. M. Azaı̈s and M. Wschebor. Level sets and extrema of random processes
and fields. John Wiley & Sons, 2009. MR2478201
[4] F. Baccelli and B. Blaszczyszyn. Stochastic Geometry and Wireless Net-
works, Volume I - Theory, volume 3, No 3–4 of Foundations and Trends in
Networking. NoW Publishers, 2009.
[5] C. Berzin. Estimation of Local Anisotropy Based on Level Sets. ArXiv
e-prints:1801.03760, January 2018.
[6] H. Biermé and A. Desolneux. On the perimeter of excursion sets of
shot noise random fields. The Annals of Probability, 44(1):521–543, 2016.
MR3457393
[7] H. Biermé and A. Desolneux. Mean Geometry for 2D random fields: level
perimeter and level total curvature integrals. Preprint hal-01370902, July
2017.
[8] A. Bulinski, E. Spodarev, and F. Timmermann. Central limit theorems for
the excursion set volumes of weakly dependent random fields. Bernoulli,
18(1):100–118, 2012. MR2888700
[9] E. M Cabaña. Affine processes: a test of isotropy based on level sets. SIAM
Journal on Applied Mathematics, 47(4):886–891, 1987. MR0898839
[10] B Casaponsa, B Crill, L Colombo, L Danese, J Bock, A Catalano,
A Bonaldi, S Basak, L Bonavera, A Coulais, et al. Planck 2015 results:
XVI. Isotropy and statistics of the CMB. 2016.
[11] S.N. Chiu, D. Stoyan, W.S. Kendall, and J. Mecke. Stochastic Geometry
and Its Applications. Wiley Series in Probability and Statistics. Wiley,
2013. MR3236788
[12] E. Di Bernardino, A. Estrade, and J. R. León. A test of Gaussianity based
on the Euler Characteristic of excursion sets. Electronic Journal of Statis-
tics, 11(1):843–890, 2017. MR3629017
[13] B. Ebner, N. Henze, M. A. Klatt, and K. Mecke. Goodness-of-fit tests
for complete spatial randomness based on Minkowski functionals of binary
images. Electronic Journal of Statistics, 12(2):2873–2904, 2018. MR3855358
[14] T.W. Epps. Testing that a stationary time series is Gaussian. The Annals
of Statistics, pages 1683–1698, 1987. MR0913582
[15] H. Federer. Curvature measures. Transactions of the American Mathemat-
ical Society, 93(3):418–491, 1959. MR0110078
[16] J. Fournier. Identification and isotropy characterization of deformed ran-
dom fields through excursion sets. Advances in Applied Probrobability,
50:706–725, 2018. MR3877250
580 Biermé et al.

[17] J. R. Gott, W. N. Colley, C-G Park, C. Park, and C. Mugnolo. Genus

topology of the cosmic microwave background from the WMAP 3-year
data. Monthly Notices of the Royal Astronomical Society, 377(4):1668–
1678, 2007.
[18] J. R. Gott, D. C. Hambrick, M. S. Vogeley, J. Kim, C. Park, Y-Y. Choi,
R. Cen, J. P. Ostriker, and K. Nagamine. Genus topology of structure in
the Sloan Digital Sky Survey: Model testing. The Astrophysical Journal,
675(1):16, 2008.
[19] D. Hug, G. Last, and M. Schulte. Second-order properties and central
limit theorems for geometric functionals of Boolean models. The Annals of
Applied Probability, 26(1):73–135, 2016. MR3449314
[20] M. Kratz and S. Vadlamani. Central limit theorem for Lipschitz–Killing
curvatures of excursion sets of Gaussian random fields. Journal of Theo-
retical Probability, 31(3):1729–1758, 2017. MR3842168
[21] R. Lachièze-Rey. Shot-noise excursions and non-stabilizing Poisson func-
tionals. ArXiv e-prints 1712.01558, December 2017.
[22] Z. Li, A. Desolneux, S. Muller, and A. K. Carton. A novel 3D stochastic
solid breast texture model for x-ray breast imaging. In Anders Tingberg,
Kristina Lång, and Pontus Timberg, editors, Breast Imaging, pages 660–
667, Cham, 2016. Springer International Publishing.
[23] G. Lindgren. Spectral moment estimation by means of level crossings.
Biometrika, 61(2):401–418, 1974. MR0426265
[24] G. Lindgren. Wave analysis by slepian models. Probabilistic engineering
mechanics, 15(1):49–57, 2000.
[25] M. S Longuet-Higgins. The statistical analysis of a random, moving surface.
Philosophical Transactions of the Royal Society A, 249(966):321–387, 1957.
MR0087257
[26] D. Müller. A central limit theorem for Lipschitz–Killing curvatures of
gaussian excursions. Journal of Mathematical Analysis and Applications,
452(2):1040–1081, 2017. MR3632692
[27] A. Nieto-Reyes, J. A. Cuesta-Albertos, and F. Gamboa. A random-
projection based test of Gaussianity for stationary processes. Computa-
tional Statistics & Data Analysis, 75:124–141, 2014. MR3178362
[28] U. Pantle, V. Schmidt, and E. Spodarev. On the estimation of integrated
covariance functions of stationary random fields. Scandinavian Journal of
Statistics, 37(1):47–66, 2010. MR2675939
[29] T. R. Reddy, S. Vadlamani, and D. Yogeshwaran. Central limit theorem for
exponentially quasi-local statistics of spin models on cayley graphs. Journal
of Statistical Physics, 173:941-984, 2018. MR3876914
[30] K. Sato. Lévy processes and infinitely divisible distributions, volume 68 of
Cambridge Studies in Advanced Mathematics. Cambridge University Press,
Cambridge, 2013. Translated from the 1990 Japanese original, Revised
edition of the 1999 English translation. MR3185174
[31] J. Schmalzing and K. M. Górski. Minkowski functionals used in the
morphological analysis of cosmic microwave background anisotropy maps.
Monthly Notices of the Royal Astronomical Society, 297(2):355–365, June
 LK curvatures of excursion sets for two-dimensional random fields 581

1998.
[32] R. Schneider and W. Weil. Stochastic and integral geometry. Probability
and its Applications. Springer-Verlag, Berlin, 2008. MR2455326
[33] E. Spodarev. Limit theorems for excursion sets of stationary random
fields. In Modern stochastics and applications, volume 90 of Springer Op-
tim. Appl., pages 221–241. Springer, Cham, 2014. MR3236077
[34] D. Stoyan and H. Stoyan. Fractals, random shapes, and point fields: meth-
ods of geometrical statistics. Wiley series in probability and mathematical
statistics: Applied probability and statistics. Wiley, 1994. MR1297125
[35] C. Thäle. 50 years sets with positive reach - a survey. Surveys in Mathe-
matics and its Applications, 3:123–165, 2008. MR2443192
[36] K. J. Worsley. Local maxima and the expected Euler characteristic of excur-
sion sets of χ2 , F and t fields. Advances in Applied Probability, 26(1):13–42,
1994. MR1260300
[37] M. Wschebor. Surfaces aléatoires: mesure géométrique des ensembles de
niveau, volume 1147. Springer, 2006. MR0871689