This article has been accepted for publication in IEEE Transactions on Signal Processing.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

A Robust Test for Elliptical Symmetry

Ilya Soloveychik
The Hebrew University of Jerusalem

Abstract—Most signal processing and statistical applications density generator function and scatter matrix1 , estimation of
heavily rely on specific data distribution models. The Gaussian the scatter (or covariance) matrix of the population becomes
distributions, although being the most common choice, are a prerequisite for almost any ellipticity GoF test. Below we
inadequate in most real world scenarios as they fail to account
for data coming from heavy-tailed populations or contaminated give a detailed exposition of the covariance estimation problem
by outliers. Such problems call for the use of Robust Statistics. in such scenarios, and provide a detailed explanation of our
The robust models and estimators are usually based on elliptical construction.
populations, making the latter ubiquitous in all methods of robust
statistics. To determine whether such tools are applicable in any
specific case, goodness-of-fit (GoF) tests are used to verify the A. Tests for Elliptical Symmetry
ellipticity hypothesis. Ellipticity GoF tests are usually hard to Numerous ellipticity GoF tests have been proposed in the
analyze and often their statistical power is not particularly strong.
statistical and signal processing literature, however, most of
In this work, assuming the true covariance matrix is unknown we
design and rigorously analyze a robust GoF test consistent against them lack statistical power as discussed in Section VII below
all alternatives to ellipticity on the unit sphere. The proposed test in detail. Such tests are also rarely supported by provable
is based on Tyler’s estimator and is formulated in terms of easily analysis, since such analysis often becomes infeasible when
computable statistics of the data. For its rigorous analysis, we the Gaussian assumptions are lifted. In addition, the compu-
develop a novel framework based on the exchangeable random
tational schemes of some of these testing procedures are so
variables calculus introduced by de Finetti. Our findings are
supported by numerical simulations comparing them to other complex that their usage becomes computationally infeasible
popular GoF tests and demonstrating the significantly higher even on modern machines.
statistical power of the suggested technique. In [1] Beran introduces a test based on marginal signs and
Index Terms—Goodness-of-fit test, elliptical population, Tyler’s ranks. That test is neither distribution-free within the family
estimator, robust statistics, exchangeable random variables. of elliptical populations nor affine-invariant. In addition, the
authors do not provide practical guidelines to the choice
of the basis functions involved in the test statistic making
its application difficult. Baringhaus [2] proposes a test for
I. I NTRODUCTION
spherical symmetry of Cramer-von Mises type based on the
independence between the norm of the samples and their
The majority of methods and techniques used by statistical
directions. The asymptotic distribution of this test is very hard
signal processing and data science heavily rely on various
to achieve and exploit. Dyckerhoff et al. [3] demonstrated
assumptions on the data such as independence of the sam-
empirically that this test can be used as a test for elliptical
ples, certain parametric families of possible distributions, etc.
symmetry in dimension 2. Koltchinskii and Sakhanenko [4]
Very rarely these assumptions are confirmed on the observed
design tests using the bootstrap methodology. These tests
samples and even if such verification attempt is made the data
are based on a class of functions closed under orthogonal
almost never agrees with the assumptions made. This leads to
transformations and have no known asymptotic distribution,
poor inference, or even to situations in which the researcher
thus requiring bootstrap to get the critical values. Manzotti
does not know the quality of the achieved results. The main
et al. [5] develop a test based on spherical harmonics to
reason for the lack of such tests is the technical complexity of
test whether the normalized vectors are uniformly distributed
their analysis especially when the data is far from being Gaus-
on the unit sphere. The test is computationally demanding
sian. The most popular substitute for Gaussian distributions in
and requires moments of order 4. Schott [6] builds a Wald-
such cases is the family of elliptical populations. Indeed, the
type test to compare the empirical fourth-order moments
latter is already quite flexible to reasonably approximate the
with the expected ones under elliptical symmetry. The test
heavy tails of the real-world populations or the outliers, while
is relatively simple in usage since it is based on the moments
still allowing rigorous analytical treatment. In this paper, we
of low order. However, it has very low power against several
focus on one of the fundamental questions accompanying any
alternatives. The authors of [7] propose a Pearson χ2 -type test
problem of signal processing or statistical inference. Namely,
with multidimensional cells. Its asymptotic distribution exists
how reliably does the elliptical family of distributions model
only for Gaussian scenario, otherwise bootstrap techniques are
the data at hand. In other words, we design a novel and
required. Cassart et al. [8] construct a pseudo-Gaussian test
easy to use Goodness-of-Fit (GoF) test that efficiently and
that is most efficient against a multivariate form of Fechner-
consistently validates the ellipticity assumptions. Given inde-
type asymmetry. The test requires finite moments of order 4.
pendently sampled data, such a test quantitatively and reliably
determines whether we can assume the data to be elliptically 1 We assume populations have zero mean, which is a very natural assump-
distributed. Since an elliptical distribution is determined by its tion in most applications.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

Tests based on Monte Carlo simulations can be found in [9] estimator can be viewed as an ML estimator of a certain
and [10]. The authors of [11] exploit graphical methods while distribution over a unit sphere. In elliptical populations the
[12] build conditional tests. We refer the reader to [13] and scatter matrix is equal to a positive multiple of the covariance
[14] for extensive surveys and performance analysis of the matrix when the latter exists.
aforementioned tests. The elliptical and generalized elliptical classes of distri-
butions are quite large to incorporate many known popula-
B. Covariance Estimation tions and they model the real non-Gaussian world behavior
Elliptical GoF tests are almost never possible without ex- much better [32, 33, 37, 38] than the Gaussian distributions.
plicit (e.g. in the plug-in form or through whitening the data) In particular, the GE family includes generalized Gaussian,
or implicit (e.g. spherical/isotropic GoF tests) estimation or compound Gaussian and many other widely used distributions
stipulation of the covariance structure. In all of these tests, [31]. Elliptical populations are commonly used to model radar
whenever the scatter matrix is unknown it must be estimated clutter [39], noise and interference in indoor and outdoor
from the available data. In this section we provide a brief mobile communication channels [40] and numerous other
survey of the non-Gaussian covariance estimation literature applications.
and focus attention on Tyler’s estimator which is later used in Other robust covariance estimation approaches were also
our construction. proposed, however, they have not become so much popular
Covariance estimation is a fundamental problem of its own as M -estimators. Among them is the Stahel-Donoho estimate
in multivariate statistical analysis. It arises in diverse applica- [41, 42], whose main idea is to detect and down-weight out-
tions such as signal processing [15, 16], geometric functional liers based on their one dimensional “outlyingness measure”.
analysis and computational geometry [17], genomics [18– Another method proposed by P. J. Rousseeuw [43] is the
20], functional MRI [21], modern social networks analysis Minimum Volume Ellipsoid estimate, whose name stems from
[22, 23], empirical finance [24, 25], classical problems of the fact that among all ellipsoids containing at least half of the
clustering and discriminant analysis [26], and many other data points, the one defined by the Minimum Volume Ellipsoid
fields. Application of structured covariance matrices instead estimate has the minimal volume. A more efficient approach,
of Bayesian classifiers based on Gaussian mixture densities or the so-called S-estimator, was later proposed by P. L. Davies
kernel densities proved to be very efficient for many pattern [44] and deeply investigated by H. P. Lopuhaa [45] and O.
recognition tasks, among them speech recognition, machine Hössjer [46]. The Minimum Covariance Determinant estimate
translation and object recognition [27]. [47] is another possibility of robust covariance estimation.
As mentioned earlier, in most real world applications the When prior knowledge on the estimator is available to the
Gaussian models become unacceptable and robust covariance research, a Bayesian covariance estimator [48–50] would
estimation methods that allow the populations to be heavy- become natural. Shrinkage estimators in the paradigm of
tailed or contain a small proportion of outliers are required James-Stein estimator are a particular case of the Bayesian
[28, 29]. In the 70-s through the analysis of elliptical pop- methodology. Shrinkage estimators of covariance matrices are
ulations and their Maximum Likelihood (ML) estimates, R. computed as a conical (often convex) combination of a certain
A. Maronna discovered a family of covariance M -estimators data statistic (e.g. sample covariance) and a constant matrix
[29]. These estimators turned out to be much more robust to (e.g. the identity matrix) representing the prior [18, 51]. Robust
outliers than the classical sample covariance which is the ML analogs of numerous shrinkage estimators were also recently
estimator in the Gaussian setup. The ideas of Maronna were developed and thoroughly studied [52, 53].
further developed by D. E. Tyler who derived a distribution- The behavior of Tyler’s estimator had been methodically
free robust covariance matrix estimator [30]. This estimator investigated in various asymptotic regimes and multiple high-
fits any population from the Generalized Elliptical (GE) family probability performance bounds have been developed for its
[31] and is also a member of the M -estimators family. Tyler’s analysis [34, 53–62]. However, all of these results only hold
estimator has become very widely used by engineers [32– if the samples are elliptically distributed, which is easily
35] since its discovery. Although, generally M -estimators are achievable in simulation studies but can hardly be guaranteed
given as solutions to optimization programs, Tyler showed that in real applications. Therefore a much more practical question
his estimator can be obtained as a solution to a simple fixed can be formulated as follows: Given the data, verify that the
point equation ellipticity assumptions can be applied to it and therefore the
n tools of robust statistics will yield meaningful and reliable
p X xi x> i
T= , (1) results when applied to it. This is the question we address in
n i=1 x>i T
−1 x
i
our work focusing on Tyler’s estimator.
where x1 , . . . , xn ∈ Rp are the collected sample vectors. To
avoid the obvious scaling ambiguity (for a solution T to (1),
c · T is also a solution whenever c > 0), it is common to fix C. Our Approach and Contribution
the scaling, e.g. by setting Tr (T) = p. When {xi }ni=1 are Many of the GoF tests method mentioned in Section I-A
i.i.d. (independent and identically distributed) elliptical [31], use plug-in estimates of the covariance matrix to whiten
their true scatter matrix Ω is positive definite and n > p, the samples before applying GoF test of uniformity over
Tyler’s estimator exists with probability one and is a consistent the unit sphere. Usually such plug-in estimates are based
estimator of Ω. In [36] Tyler also demonstrated that his on the sample covariance matrix (e.g. [4–7] and numerous

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

others) which creates an unbalanced situation. Indeed, such to the true covariance in elliptical populations maybe very
tests assume ellipticity of the population, thus allowing it to slow due to heavy tails [5], however, we emphasize that our
have heavy-tails, but the tools used to estimate the covariance technique allows the use of the sample covariance instead of
are not robust and therefore no performance guarantees in Tyler’s estimator. Our test is not limited to certain moments
finite samples can be offered by these tests. In other cases, which makes it more natural and less computationally
robust estimates are used, however, without rigorous studies demanding [6, 8]. Unlike Monte Carlo simulations-based
and claims because of the significant level of complexity of the techniques [9, 10], our methodology is rigorous and offers a
required analysis. When testing the ellipticity hypothesis, it is deterministic and computationally cheap algorithm. Finally,
common to separate the question of elliptical symmetry from unlike previously mentioned works (and articles referenced
the radial density [31, 36]. Such a standard approach enables in [71]) we believe that the methodology based on the
one to project the samples on the unit sphere as described exchangeability framework is the most suitable for the
in Sections II and III in detail and study the distribution of analysis of populations transformed by plug-in estimators.
their normalized values. Remarkably, modulo positive scaling, This type of analysis is usually technically more complex
such a transformation does not affect the scatter matrix of an but in our eyes it represents the natural approach to the
elliptical (or GE) population and it is the use of this very problem. Our theoretical studies are supported by extensive
technique that led to the discovery of Tyler’s estimator as an numerical simulations featuring the properties of the tests and
ML estimate of the scatter matrix [36]. Following these ideas - comparing them to other available tests mentioned in earlier
for the scenario of unknown scatter matrix - we exploit Tyler’s in Section I-A.
estimator to develop an asymptotically consistent GoF test The rest of the article is organized as follows. In Section
against all alternatives to ellipticity on the unit p-dimensional II, we introduce the setup and notation. The problem is
sphere.2 To enable analytical treatment of such a hypothesis formulated in Section III where we also present some of the
test, we reformulate it as an asymptotic uniformity test for existing tests for the known scatter case. In Section IV, we
a certain stochastically dependent sequence of unit random reformulate the problem and introduce necessary background
vectors. The main tool used in the construction and analysis on exchangeable random variables; Section IV-D provides
of the uniformity tests for i.i.d. scenario is the Central Limit some additional notation and auxiliary results. In Section
Theorem (CLT) [63–65] which is clearly not applicable when V, we formulate the main results and discuss them. Section
the measurements are not independent. For our setup, we VI provides numerical studies of the proposed tests while
develop a novel toolbox that allows verification of the null Section VII demonstrates their power in comparison to other
hypothesis by resorting to the concept of exchangeability. commonly used ellipticity tests. The conclusion is provided
A sequence of variables is called exchangeable if the joint in Section VIII. Some of the proofs are postponed to the
distribution of any finite subset of these variables is invariant Appendix.
under arbitrary permutations of their indices. Exchangeable
random variables were first introduced by de Finetti [66, 67]
as a direct and natural generalization of i.i.d. sequences. II. N OTATION AND S ETUP
Interestingly, exchangeable random variables serve as one of
the fundamental building blocks of the Bayesian statistics [68]. Definition 1 ([72]). A vector y ∈ Rp is elliptically distributed
Unlike the i.i.d. case, the behavior of exchangeable sequences with the scatter matrix Ω  0 and mean µ if there exists a
is much harder to analyze. We exploit certain versions of random vector w ∈ S p−1 uniformly distributed over the unit
CLT and the Strong Law of Large Numbers (SLLN) for p−1-dimensional sphere and an independent random variable
exchangeable variables to demonstrate asymptotic consistency r > 0, such that
of our test statistics built analogously to the generalized Ajne
and Giné statistics [63, 64, 69] developed for the i.i.d. case.3 y = µ + r · Ω1/2 w. (2)
Following Tyler, our approach becomes essentially q
distribution-free within the elliptical family since we do not For example, if r ∼ χ2p , then y ∼ N (µ, Ω). In what
focus on estimating the radial density function [71]. We follows we always assume that the data is centered, µ = 0.
offer a test which is consistent against all alternatives to Let us consider the normalized vector,
elliptical symmetry and not only certain classes of densities
[5, 8, 71]. We do not use the sample covariance matrix as y Ω1/2 w
a plug-in estimator in particular because its convergence x= = , (3)
kyk Ω1/2 w
2 The unknown distribution is assumed to be a Lebesgue measurable proba-
which can be equivalently viewed as disregarding the in-
bility measure over the p-dimensional Euclidean sphere as elaborated further.
By all alternatives, it is common to understand the set of all probability
formation stored in the scalar variable r but keeping the
measures except for the class under consideration, which in our case is the information provided by the scatter matrix. As we see below,
class of elliptical (or generalized elliptical) populations [4, 7, 14, 63]. the distribution of x contains all the information about the
3 The Ajne statistic was originally introduced for distributions on a circle
scatter matrix Ω. We are going to recover the scatter matrix
[69], the idea was extended by [70] to the 2-dimensional unit sphere and
later generalized by [64] for the p − 1-dimensional spheres. Similarly, Giné’s by sampling from the distribution of x. Denote by I = Ip the
statistic was originally defined for 1- and 2-dimensional spheres and later p-dimensional identity matrix.
extended by [64] to the general dimension.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

Definition 2 ([36]). The family of real Angular Central Gaus- Under H0 , w1 , . . . , wn ∼ U(I) and therefore the test (6)-(7)
sian (ACG) distributions on S p−1 is defined by the densities becomes actually a uniformity test on the unit sphere,
of the form i.i.d.
G0 : w1 , . . . , wn ∼ U(I), (9)
Γ(p/2) 1
p(x; Ω) = 1/2 > −1 p/2
, x ∈ S p−1 , (4) i.i.d.
G1 : w1 , . . . , wn  U(I). (10)
2π p/2 |Ω| (x Ω x)
for Ω  0 which is called the scatter matrix. Next, we summarize two uniformity tests on Sp−1 concluding
this section with Proposition 3 providing a uniformity test
When x is ACG distributed with the scatter matrix Ω, we consistent against all alternatives on the unit sphere. Based on
write it, we will develop an analogous test for (6)-(7) with unknown
x ∼ U (Ω) , (5) scatter matrix in the subsequent sections. Denote by
Z
in particular when Ω = I we get the uniform distribution 2π
Vp−1 = dx = p

(11)
over the unit sphere U (I). Note that ACG is not a member of x∈S p−1 Γ 2
the elliptical family but actually belongs to a wider class of the area of the unit sphere. In addition, by
generalized elliptical populations whose definition is identical
to Definition 1 except for weakened assumptions on r [72]. In ψij = arccos(x>
i xj ) (12)
generalized elliptical population, r does not have to be stochas-
we denote the angular separation (the shortest great circle
tically independent of w and does not have to be non-negative.
distance) between xi and xj and by
The following result allows us to reduce estimation of the
scatter matrices of elliptical populations to the estimation of N (y) = |{xi | y> xi > 0}|, y ∈ Sp−1 , (13)
the scatter matrices of ACG vectors.
the number of points falling into the hemisphere with the pole
Lemma 1 ([72]). For a random vector y sampled from a at y. Denote also
centered elliptical population with the scatter matrix Ω, x p
α = − 1, (14)
defined in (3) is ACG distributed with the same scatter matrix. 2
   
a+b−2 a+b−1
Now assume n > p i.i.d. random vectors x1 , . . . , xn ∈ ν(a, b) = + . (15)
a−1 a−1
Sp−1 are sampled from U (Ω), then as shown in [36] the ML
estimator of the scatter matrix exists almost surely and is given The following two popular statistics and detailed investiga-
by the fixed point equation (1). The solutions to this equation tion of their behavior can be found in [63, 69]. These results
form a ray since the latter is invariant under multiplication of were later generalized in [64] and summarized in [65].
the matrix T by a positive constant. To resolve the ambiguity Proposition 1 (Generalized Ajne Test, [64, 69]). Under the
we choose T to satisfy Tr (T) = p, however, we note that the uniformity hypothesis, the Ajne statistic
specific choice of the scaling does not affect any of the results Z
presented below. 1  n 2 n 1 X
tA = N (y) − dy = − ψij
nVp−1 y∈Sp−1 2 4 πn i<j
III. P ROBLEM F ORMULATION AND S TATE OF THE A RT is  asymptotically  distributed as
P∞ 2
A. Main Goal L a K ν(p−1,2q−1) , where K ξ are independent
q=1 2q−1
The problem considered in this article can be formulated as random variables distributed as χ2ξ and
follows. Given a sequence of vectors {xi }ni=1 ⊂ Sp−1 sampled
independently, we want to test two alternative hypotheses, (−1)q−1 2p−2 Γ(α + 1)Γ(q + α)(2q − 2)
a2q−1 = . (16)
π(q − 1)!(2q + p − 3)!
i.i.d.
H0 : x1 , . . . , xn ∼ U(Ω), for some Ω, (6)
Proposition 2 (Generalized Giné Test, [63, 64]). Under the
i.i.d.
H1 : x1 , . . . , xn  U(Ω), for any Ω, (7) uniformity hypothesis, the Giné statistic

!2
and in the case of H0 we want to estimate the scatter matrix n p − 1 Γ α + 21 X
Ω, as well. tG = − sin(ψij ) (17)
2 2n Γ(α + 1) i<j
The test (6)-(7) is a composite hypothesis since the scatter
P 
matrix is unknown. When the scatter matrix is known, the is asymptotically distributed as L
∞ 2
q=1 a2q Kν(p−1,2q) ,
problem can be equivalently reformulated as a uniformity test
where Kξ are independent random variables distributed as
on the sphere as shown below.
χ2ξ and
1 1

!2
B. Uniformity Tests on Sp−1


(p − 1)(2q − 1) Γ α + Γ q −
a22q = 2
2 . (18)
Assume the scatter matrix Ω in the hypothesis test (6)-(7) 8π(2q + p − 1) Γ q + α + 21
is known and introduce a derived i.i.d. sequence,
The following statement provides a concise and directly
Ω−1/2 xi applicable test for uniformity under the assumption that the
wi = , i = 1, . . . , n. (8) random vectors are sampled i.i.d. from U(I).
Ω−1/2 xi

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

Proposition 3 (Uniformity test, [63, 64]). Any weighted sum exchangeable. In his seminal works de Finetti [66, 67] demon-
of tA and tG is consistent against all alternatives to uniformity strated that in certain sense every (infinite) exchangeable
on Sp−1 . sequence can be represented as a composition of sequences
of i.i.d. variables. This result can be viewed as the analog of
In practice, one way to make the decision about accepting
Fourier decomposition in analysis, as it allows one to represent
or rejecting H0 or G0 is as follows. The statistician truncates
a more complicated exchangeable sequence as a superposition
the series mentioned in the last two propositions in a data-
of basic building blocks - independent sequences - objects
driven manner and compares the sample values of tA and tG
much easier accessible for analysis and reasoning.
with the tables (or explicit numerical approximations) of the
De Finetti [66, 67] and some of his followers focused
corresponding distributions. Another more general approach
on infinite exchangeable sequences. There exist, however,
consists in replacing tA and tG by statistics whose expansions
finite sets of exchangeable random variables which cannot
only have finite number of non-zero coefficients ak (see [63]
be embedded into infinite sequences, these are called finitely
for more details). An efficient data-driven approach to the
exchangeable or non-extendable. The analysis of extendable
design of the uniformity tests based on a modification of the
sequences can be reduced to the analysis of infinite sequences.
Bayesian Information Criterion was developed by [73]. Often
On the other hand, the non-extendable sequences require quite
in practice, the distributions of the statistics at hand under
different approaches [74]. Our sequence of samples {ti }ni=1
the null hypothesis are estimated empirically, by generating
is an example of a non-extendable exchangeable sequence of
samples from the latter. This is the approach adopted in
random vectors. Indeed, their order obviously does not matter
Section VII below.
since T is not affected by permutations of the measurements
In this paper we are interested in the case of unknown scatter
{xi }ni=1 . We can also see that this sequence is non-extendable,
matrix in (6)-(7). As we see below this makes the hypothesis
since addition of new random vectors xj without an amend-
test much more involved. In the next sections we develop
ment of T will turn the sequence into non-exchangeable. For
analogs of generalized Ajne and Giné uniformity tests for this
a detailed study of non-extendability we refer the reader to
scenario.
[74] and references therein.
The main result of our paper can be briefly summarized
IV. P ROBLEM R EFORMULATION AND E XCHANGEABILITY as follows. We demonstrate that the limiting behavior of the
A. Methodology samples {ti }ni=1 is in certain sense analogous to the behavior
From Theorem 3.1 from [30] we know that under H0 Tyler’s of the vectors uniformly distributed over the unit sphere and
estimator converges almost surely to the true scatter matrix therefore, we can apply similar tools for the hypothesis tests.
when n → ∞. This idea motivated our study of a new Below we show how to overcome the technical challenges on
sequence of vectors, defined as follows. Under H0 introduced this way.
in (6), we now consider the sequence
T−1/2 xi C. Limit Theorems for Exchangeable Variables
ti = ∈ Sp−1 , i = 1, . . . , n, (19) To illustrate the previous section and better describe the
T−1/2 xi
nature of the exchangeability phenomenon and its relation to
where T is defined in (1). The main challenge we face in the the stochastic independence, in this section we present analogs
study of {ti } is the lack of stochastic independence unlike the of the SLLN and CLT for triangular arrays of exchangeable
case of {wi } defined in (8). Indeed, most existing convergence variables.
results explicitly rely on independence in their derivations in
such a way that any deviation from this assumption ruins Lemma 2 (Strong Law of Large Numbers for Exchangeable
the performance analysis. For example, all the results of Arrays). Let {Xni }∞,n
n,i=1 be a triangular array of row-wise
Ajne, Giné, and Prentice utilize the CLT and thus require exchangeable random variables and {X∞i }∞i=1 be a sequence
independence as the most crucial assumption [63–65, 69]. of exchangeable random variables of bounded second moment
Next we include a brief summary of the exchangeability such that
a.s.
concept and the related toolbox. We then use it in Section 1) Xn1 −−→ X∞1 , n → ∞,
IV-D to overcome the loss of independence in our analysis of 2) var (Xn1 − X∞1 ) → 0, n → ∞,
the consistency of {ti } and their statistics. 3) E [Xn1 Xn2 ] → 0, n → ∞.
Then
n
B. Exchangeable Random Variables 1X a.s.
Xni −−→ 0, n → ∞. (20)
Definition 3. Given a sequence {Xi } (finite or infinite) of n i=1
random variables, we say that it is exchangeable if the joint
Proof. The proof can be found in the Appendix.
distribution of any finite subset of variables is invariant under
arbitrary permutations of their indices. Let kn < n be two sequences of natural numbers such that
In other words, exchangeability is our indifference to the kn
order of the measurements. This is clearly a much weaker hy- → γ ∈ [0, 1). (21)
n
pothesis than independence, as any i.i.d. sequence is obviously

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

Lemma 3 (Central Limit Theorem for Exchangeable Arrays, {tni }ni=1 for every fixed n > p. The following CLT-type result
Theorem 2 from [75]4 ). Let {Xni }∞,n
n,i=1 be a triangular array holds in our scenario.
of row-wise exchangeable random variables such that
Proposition 4. For the triangular array of vectors
1) E [Xn1 Xn2 ] → 0, n → ∞, {tni }∞,n
| P n=p+1,i=1 defined above,
2) max |X √ni −
n
→ 0, ∀n, n
16i6n
√ 1 X L
k
3) k1n i=1
P n
Xni2 P
−
→ 1, n → ∞. p· √ tni −
→ N (0, I), n → ∞. (27)
n i=1
Then
" kn n
# Proof. The proof can be found in the Appendix.
p 1 X 1X L
kn Xni − Xni −→ N (0, 1 − γ), n → ∞. Corollary 1. Under H0 , for any differentiable function
kn i=1 n i=1
f : Sp−1 → R,
As mentioned earlier this result provides an analog of the n
√ 1 X L 2
CLT for exchangeable sequences. However, it is important p· √ f (tni ) −
→ N (0, k∇f (0)k ), n → ∞. (28)
n i=1
to stress its distinction from the classical CLT-type claims
for the i.i.d. variables. Indeed, Lemma 3 only allows us to Proof. The proof follows the i.i.d. case verbatim using the
consider a subset of the sample of cardinality kn smaller than Maclaurin expansion of f .
the number of variables n in the row so that even their ratio
must not approach one. This is a reflection of the essential V. A SYMPTOTIC U NIFORMITY T ESTS FOR
difference between non-extendable exchangeable sequences E XCHANGEABLE V ECTORS
and their extendable counterparts that include i.i.d. sequences In Section III-B we introduced statistics tA and tG to test
as a particular case [74]. the null hypothesis of uniformity for independent samples
over the unit sphere Sp−1 . Our next statements constitute
D. Additional Notation and Auxiliary Results analogs of those result for the row-wise exchangeable array
Assume that an infinite i.i.d. sequence {xi }∞
i=1 is sampled
{tni }∞,n
n=p+1,i=1 . Denote
under the composite H0 with the true scatter matrix is un-
ψn,ij = arccos(t>
ni tnj ). (29)
known. For every n > p, let the sequence of corresponding
Tyler’s estimators be Proposition 5 (Generalized Ajne Test for tni ). Under H0 , the
n Ajne statistic
p X xi x> i
Tn = −1 , n = p + 1, . . . , (22) n 1 X
n i=1 x>
i Tn xi tA ({tni }) = − ψn,ij (30)
4 πn i<j
which exist almost surely for a random sample [37, 76].
Consider a triangular array of row-wise exchangeable random is  asymptotically  distributed as
P∞ 2
vectors L a
q=1 2q−1 K ν(p−1,2q−1) as n → ∞, where K ξ
2
Tn
−1/2
xi are independent random variables distributed as χξ and
tni = ∈ Sp−1 , i = 1, . . . , n, n = p + 1, . . . .
−1/2
Tn xi (−1)q−1 2p−2 Γ(α + 1)Γ(q + α)(2q − 2)
a2q−1 = . (31)
(23) π(q − 1)!(2q + p − 3)!
Note that by Definition 1, the sequence {xi }∞
i=1 can equiva- Proof. The proof follows [63] and [64] verbatim using Corol-
lently be defined as follows. Given a sequence {wi }∞ i=1 ∼ lary 1.
U(Ip ) of uniform i.i.d. random vectors, we look at their
transforms Definition 4. Random variable X is said to be first-order
Ω1/2 wi stochastically dominated by random variable Y if
xi = , (24)
Ω1/2 wi P[X ∈ A] 6 P[Y ∈ A], (32)
for some fixed but unknown Ω  0. Define also an auxiliary for any measurable set A.
sequence
t∞i = wi . (25) Proposition 6 (Generalized Giné Test for tnk ). Under H0 ,
the Giné statistic
Lemma 4. With the notation introduced above,
!2
a.s. n p − 1 Γ α + 21 X
tni −−→ t∞i , n → ∞. (26) tG ({tni }) = − sin(ψn,ij ) (33)
2 2n Γ(α + 1) i<j
Proof. The proof can be found in the Appendix.
is asymptotically first-order stochastically
P∞ 2 dominated by the
We are now interested in the empirical distributions of the random variable distributed as q=1 a2q Kν(p−1,2q) , where
rows of the obtained triangular array, which are the finite sets Kξ are independent random variables distributed as χ2ξ and
4 To simplify the notation we assume the number of the elements in the
!2
(p − 1)(2q − 1) Γ α + 21 Γ q − 12


n-th row to be n unlike the seemingly more general case of mn variables 2
a2q = , (34)
Γ q + α + 21


considered in [75]. 8π(2q + p − 1)

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

and
 
1 1 1
E [tG ({wi })]−E [tG ({tni })] ∼ + +O , n → ∞.
8 16p p2
Proof. The proof can be found in the Appendix.
It is important to emphasize the main difference between
Propositions 2 and 6. Indeed, in the former the asymptotic dis-
tribution is given by a sum of scaled χ2 -variables, while in the
second one the limiting distribution is first-order stochastically
dominated by the same distribution and is in fact significantly
thinner. This discrepancy is due to the fact that {tni } are
dependent in such a special way that their sample covariance
matrix is exactly I. The detailed reasoning and discussion can
be found in the Appendix.
Theorem 1 (Uniformity Test for tni ). Under H0 , any
Fig. 1: Comparison of the empirical distributions of the Ajne
weighted sum of tA ({tni }) and tG ({tni }) is consistent
and Giné test statistics computed for the sequences {wi } and
against all alternatives to the asymptotic uniformity of {tni }
{tni } defined in Section IV-D with the true scatter matrix
on Sp−1 .
being I.
Proof. The proof follows [63] and [64] verbatim using Propo-
sitions 5 and 6.
B. Criterion Performance for Alternatives
Remark 1. It is also important to emphasize that an asymp-
totic bound to the power of the uniformity test suggested by Following Theorem 1, to demonstrate the power of the sug-
Theorem 1 against all alternatives on the unit sphere can be gested methodology in our second experiment we compared
easily constructed using the asymptotic normality of the scaled the empirical distributions of the statistics
deviations of the Ajne and Giné statistics as shown in Section s ({tni }) = tA ({tni }) + tG ({tni }) , (36)
4 of [63]. These derivations are also valid in our case, and
therefore we omit them due to a lack of space. s ({wi }) = tA ({wi }) + tG ({wi }) (37)

VI. N UMERICAL S IMULATIONS under the null hypotheses H0 and G0 versus their distributions
under specific non-elliptical alternatives H1 and G1 . The
In this section, we investigate the behavior and advantages
alternatives were constructed as follows. We generated the
of the criterion proposed in Theorem 1 through numerical
uniform sequence {wi } as before, added a constant offset to
simulations.
all the obtained vectors and re-normalized them,
A. Distributions of the Statistics under the Null Hypothesis wi + a
w
ei = . (38)
In the first experiment, we compared the empirical distri- kwi + ak
butions of tA ({tni }) and tG ({tni }) with their counterparts Note that the distributions of w
e i and of xi constructed from
tA ({wi }) and tG ({wi }) for the independent samples playing it via (24)
the role of the benchmarks. In this simulation we took the true Ω1/2 wei
scatter matrix to be the identity Ω = I. Figure 1 demonstrates xi = , (39)
Ω1/2 wei
the anticipated in Section V difference in the behavior of
the Ajne and Giné statistics. More specifically, as claimed in are not ACG and therefore our test should be able to discrim-
Proposition 6 and discussed in detail in its proof, the statistic inate between the hypotheses. In this experiment we chose
tG ({tni }) is first-order stochastically dominated by tG ({wi }) p = 5, Ω = I and
due to the difference in the behavior of the quadratic term
 
1
in the expansion of the statistic (75) caused by dependencies 1 .
a = 0.05 · √  ..  . (40)
among {tni }. This is in contrast to the Ajne statistic whose p
distributions in both cases coincide since a similar expansion 1
into Gegenbauer polynomials involves odd degree polynomials Figure 2 demonstrates the 0.95-confidence bands for the dis-
only (for more details see the proof of Proposition 6). Note tributions of the statistics s ({tni }) and s ({wi }) as functions
also that the theoretically predicted by Theorem 1 difference of the number of measurements n in the sample. The bands
between the expected values for p = 8, in Figure 2 were averaged over 50000 independent trials. We
see from the graph that despite the small size of a, already
 
1 1 1
E [tG ({wi })] − E [tG ({tni })] ∼ + +O ≈ 0.133 with n = 15 measurements the criterion allows us to easily
8 16p p2
(35) discriminate between the hypotheses and its power is similar
is confirmed by the numerical simulation yielding the value for both the i.i.d. (know scatter matrix) and Tyler’s (unknown
of 0.131. scatter matrix) cases.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

(a) s ({tni }) (b) s ({wi })

Fig. 2: Comparison of the 0.95-confidence bands for H0 versus H1 designed in section VI-B for the i.i.d. {wi } samples and
their exchangeable {tni } counterparts.

VII. C OMPARISON TO OTHER T ESTS with the population means

For the purpose of this comparative study, we selected the

  
1 1
most popular tests used in the literature due to Koltchinskii and 1 −1
   
Sakhanenko [4, 14], Manzotti et al. (MPQ test) [5], Cassart 1 1
   
(PseudoGaussian test) [8], Schott [6], and Babic (SkewOptimal z1 = 5 · 1 ,
 
z2 = 5 · −1
 
(44)
test) [71]. Implementations of all these tests are available  ..   .. 
.  . 
online trough the ellipticalsymmetry5 package developed by    
1 1
[77].
1 −1
In our numerical experiment, we set the dimension to p = 5
and the number of samples in every batch n = 50. Each test
be all i.i.d. and define
from the aforementioned list and the tests proposed in this
paper (from Propositions 5 and 6 and Theorem 1) is invoked
to classify each such batch and determine whether it is coming Y i = X i + Z1 + Z2 , i = N/2 + 1, . . . , N. (45)
from an elliptical population or not. The goal is to compare
the powers of the tests using their ROC (Receiver Operating Clearly, for N/2 + 1, . . . , N the vectors in the corresponding
Characteristic) curves. To this end, we constructed N = 5000 batches Yi are non-elliptical and dont become GE even if
sample batches (each having n vectors of dimension p), half normalized by their Euclidean norms. Therefore, we would
of which were coming from the standard normal population expect all our tests to discriminate between elliptical and non-
and labeled as elliptical batch, and the rest form a non- elliptical batches.
elliptical population labeled as non-elliptical batch and defined Figure 3 shows the ROC (Receiver Operating Characteristic)
as follows. Let curves for all the available tests. Naturally, the highest (having
maximal power β) curve at each point on the horizontal axis
Xi,k ∼ N (0, Ip ), i = 1, . . . , N, k = 1, . . . , n, (41) corresponds to the most powerful test at this False Positive
Rate (α). We can see that the tests proposed in this article
be N batches of n vectors each with all Xi,k i.i.d. standard are more powerful than the rest. The distributions of the
normal. Now, we set statistics tA ({tni }) , tG ({tni }) , s ({tni }) under H0 derived
in Propositions 5 and 6 were estimated empirically for this
Yi = Xi , i = 1, . . . , N/2 (42) numerical study. We would like to emphasize a number of
points. First, the ROC curves of all the reference tests are very
for the first N/2 batches labeled as elliptical batch and another close to the bisector of the first quadrant, showing their very
N/2 batches marked as non-elliptical batch we sample from low statistical power and making them almost equivalent to
a non-elliptical population designed as follows. Let fair coin tossing. Second, every test involving tG ({tni }) has
a horizontal ROC curve and statistical power of 1 for every
Zj,i,k ∼ N (zj , Ip ), α, meaning that it correctly classified every batch. This is due
to the sharp concentration of the distribution of this statistic
j = 1, 2, i = N/2 + 1, . . . , N, k = 1, . . . , n, (43) around a constant value value as can be seen in Figure 1.
Because of that, any deviation from the expected value is a
5 https://cran.r-project.org/package=ellipticalsymmetry strong evidence against the ellipticity hypothesis.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

Proof of Proposition 4. As above, we can equivalently rewrite

tni as
−1/2
Tn Ω1/2 wi
tni = , i = 1, . . . , n, n = p + 1, . . . ,
−1/2
Tn Ω1/2 wi
which is just a useful representation as clearly Ω is not
revealed to the researcher. Fix a vector a ∈ Rp of unit norm
kak = 1 and consider the following triangular array of row-
wise exchangeable random variables
√
Xni = p · a> tni , i = 1, . . . , n, n = p + 1, . . . . (48)
Let us study the properties of {Xni }∞,n
n=p+1,i=1 . First, consider

E [Xn1 Xn2 ] = p a> E tn1 t>

 
n2 a, (49)
Lemma 4 implies that
Fig. 3: ROC curves of Koltchinskii-Sakhanenko, MPQ, Pseu- E tn1 t>
   >

n2 → E w1 w2 = 0, n → ∞, (50)
doGaussian, Schott, and SkewOptimal tests in comparison to
the test statistic s ({tni }) defined in (36) and justified by therefore,
Theorem 1. E [Xn1 Xn2 ] → 0, n → ∞. (51)
Next, note that
|Xni | √ |a> tni | √ kak ktni k
r
VIII. C ONCLUSION p
√ = p √ 6 p √ = → 0. (52)
In this paper we propose a novel elliptical symmetry test n n n n
based on the ideas of robust statistics, and specifically Tyler’s Finally, let us show that
estimator of covariance matrix. This is an easy to apply and
kn
computationally cheap test with provable performance guaran- 1 X P
X2 −
→ 1, n → ∞. (53)
tees. In our extensive comparative studies we demonstrate that kn i=1 ni
it surpasses all the other commonly exploited ellipticity tests
by a large margin when it comes to their statistical power. In Indeed,
kn kn
" kn
#
addition, based on the exchangeable random variables calculus 1 X 1 X 1 X
2 > > > >
introduced by de Finetti, we develop a natural mathematical X =p a tni tni a = p a tni tni a.
kn i=1 ni kn i=1 kn i=1
framework enabling rigorous analysis of our test and numerous
other tests and estimators based on dependent but exchange- For the sample covariance we obtain,
able sample measurements. kn kn −1/2 −1/2
1 X 1 X Tn Ω1/2 wi wi> Ω1/2 Tn
tni t>
ni = 2
A PPENDIX kn i=1 kn i=1 −1/2
Tn Ω1/2 wi
Proof of Lemma 2. Our proof is based on an analogous result  
in [78]. Both Lemmas 1 and 2 from [78] can be easily restated kn >
 1 X wi wi
= T−1/2
 1/2 −1/2
for our setup after replacing the Banach space E by R and n Ω1/2  2Ω Tn
kn i=1 −1/2 1/2
linear functionals by scalar multiplication. In addition, note Tn Ω wi
that our condition 1) immediately implies requirement (2.5) a.s. 1
from [78]. Now, the reasoning from the proof of Theorem 1 −−→ I, n → ∞, (54)
p
from [78] applies verbatim.
and therefore,
Proof of Lemma 4. As shown in Theorem 3.1 from [30], kn
a.s. 1 X P 2
Tn −−→ Ω  0, n → ∞, (46) → a> I a = kak = 1,
X2 − n → ∞. (55)
kn i=1 ni
therefore, starting from some n0 , Tn is almost surely invert-
ible for n > n0 and The rest of the proof is based on the argument proposed in
a.s.
[79]. Assume without loss of generality that n is an even
T−1/2
n Ω1/2 −−→ Ip , n → ∞. (47) number and set
n
Now the claim follows from the definition of the sequence kn = . (56)
2
{tni }n , Consider now the following sequence,
−1/2 −1/2
Tn xi Tn Ω1/2 wi
(
a.s. Xni , i 6 kn ,
tni = = −−→ wi , n → ∞. Yni = (57)
−1/2 −1/2
Tn xi Tn Ω1/2 wi −Xni , i > kn .

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

10

Clearly, the new sequence is exchangeable with the same joint Note that the Gegenbauer polynomials of odd/even order
distribution as the original sequence. Indeed, under H0 the involves monomials of only odd/even order, respectively.
joint distribution of {tni } is invariant under multiplication of In order to analyze the Giné statistic (33), we use the
any of the random vectors by −1 and Tn is an even function following expansion of sin θ into Gegenbauer polynomials in
of tni . Now all the conditions of Lemma 3 are satisfied for cos θ from [64],
the new sequence {Yni } and we obtain,
!2
1 p − 1 Γ α + 21
−
 
r n/2 n   sin θ
n 2 X 1X L 1 2 2 Γ(α + 1)
Yni − Yni  −
→ N 0, , n → ∞.
2 n i=1 n i=1 2 ∞
X (p − 1)(2q − 1)(4q + p − 2)
(58) =
q=1
(p − 2)(2q + p − 1)8π
Note that

!2
Γ α + 21 Γ q − 12


n/2 n α
2X 1X × C2q (cos θ),
Yni − Yni (59) Γ(q + α + 21 )
n i=1 n i=1
n/2

n/2 n
 where we remind the reader that
2X 1 X X
p
= Xni − Xni − Xni  α = − 1. (66)
n i=1 n i=1 2
i=n/2+1
n/2 n n Using (65) we can write,
1X 1 X 1X
= Xni + Xni = Xni ,
!2 ∞
n i=1 n n i=1 1 p − 1 Γ α + 12 X
i=n/2+1
− sin θ = γ2r (α, p) cos2r θ
2 2 Γ(α + 1)
to obtain r=0
∞
r n   X
n 1X
·
L
Xni −
→N 0,
1
, n → ∞, (60) = γ0 (α, p) + γ2 (α, p) cos2 θ + γ2r (α, p) cos2r θ.
2 n i=1 2 r=2

or Below we use the explicit form of γ2 (α, p),

n
1 X L ∞
√ Xni −
→ N (0, 1) , n → ∞. (61) X (p − 1)(2q − 1)(4q + p − 2)
n i=1 γ2 (α, p) =
q=1
(p − 2)(2q + p − 1)8π
By the definition of Xni we get
!2
Γ α + 21 Γ q − 12


n α
√ 1 X × ζ2q,2 , (67)
L
Γ q + α + 12


p · a> √ tni −
→ N (0, 1) , n → ∞. (62)
n i=1
where
Finally, recall that the vector a was chosen arbitrarily to α Γ(q + 1 + α)
ζ2q,2 = 2(−1)q−1 , (68)
conclude the proof. Γ(α)(q − 1)!
For a sequence of vectors {yi }ni=1 , denote their sample is the weight of z 2 in C2q
α
(z) defined as in (65). Thus, we
covariance by obtain
n
1X ∞
(p − 1)(2q − 1)(4q + p − 2)
Sy,n = yi yi> . (63) γ2 (α, p) =
X
(−1)q−1 (69)
n i=1 (p − 2)(2q + p − 1)4π
q=1
Lemma 5 (Theorem 6.2 from [80]). Let {yi }ni=1 ⊂ Sp−1 be
!2
Γ p2 − 12 Γ q − 12 Γ q + p2



a set of n > p vectors, then × .
Γ q + p2 − 21 Γ p2 − 1 (q − 1)!




1
Tr S2y,n > . (64) Since the last series is telescopic we conclude that in particular
p
Proof of Proposition 6. The difference in the behavior of the γ2 (α, p) > 0. (70)
Ajne and Giné statistics stems from the fact that the former
is a sum of Gegenbauer polynomials of odd orders involving Let us compute the first few terms of this series,
only monomials of odd powers, while the latter reads as a γ2 (α, p) (71)
sum of Gegenbauer polynomials of even orders involving only  
p(p + 2) 3p(p + 2)(p + 6) 1
monomials of even powers [63, 64]. Next we explain this in = − +O
more detail. 4(p + 1)(p − 1) 8(p − 1)(p + 1)2 (p + 3) p2
 
Gegenbauer (ultraspherical) polynomial [81] of index α and 1 1 1
= + +O .
order q > 2 is defined as 4 8p p2
bq/2c Recall that due to (29),
X Γ(q − k + α)
Cqα (z) = (−1)k (2z)q−2k . (65)
k=0
Γ(α)k!(q − 2k)! cos(ψn,ij ) = t>
ni tnj , (72)

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

11

and therefore, Recall that the limiting distribution of the spectrum of

p Sw,n is given by the Marchenko-Pastur law [82] whose
sin(ψn,ij ) = sin(ψn,ji ), (73)
second moment gives us the following asymptotic equivalence,
together with h 
2
i p2
sin(ψn,ii ) = 0. (74) E Tr [p Sw,n ] ∼ p + , n → ∞. (81)
n
Now we can see that the Giné statistic reads as As a consequence,

!2
n p − 1 Γ α + 12 1
n2 E Tr S2w,n − n2 E Tr S2t,n
X 
 


tG ({tni }) = − sin(ψn,ij ) 2n
2 2n Γ(α + 1) i<j n2 1
 
1 1 1

!2 n ∼ + − ∼ , n → ∞,
n p − 1 Γ α + 12 X 2n p n p 2
= − sin(ψn,ij )
2 4n Γ(α + 1) i,j=1
and from (75) we conclude,

!2 n 
n 1  p − 1 Γ α + 12 γ2 (α, p)
E [tG ({wi })] − E [tG ({tni })] ∼
X
= − sin(ψn,ij ) 2 
2 2n 2 Γ(α + 1) i,j=1 1 1 1
n
" ∼ + +O , n → ∞.
n 1 X 8 16p p2
= + γ0 (α, p) + γ2 (α, p) cos2 θ
2 2n i,j=1
∞
#
X
2r 1 R EFERENCES
+ γ2r (α, p) cos θ −
2
r=2
" [1] R. Beran, “Testing for ellipsoidal symmetry of a multi-
n variate density,” The Annals of Statistics, pp. 150–162,
n nγ0 (α, p) 1 X
2
= + + γ2 (α, p) t> ni tnj 1979.
4 2 2n i,j=1
[2] L. Baringhaus, “Testing for spherical symmetry of a
∞
#
X
>

2r multivariate distribution,” The Annals of Statistics, pp.
+ γ2r (α, p) tni tnj . (75) 899–917, 1991.
r=2
[3] R. Dyckerhoff, C. Ley, and D. Paindaveine, “Depth-based
Note that runs tests for bivariate central symmetry,” Annals of the
n p
" n #2 Institute of Statistical Mathematics, vol. 67, no. 5, pp.
X
2 X X (k) (l)
t>
ni tnj = tni tni (76) 917–941, 2015.
i,j=1 k,l=1 i=1 [4] V. Koltchinskii and L. Sakhanenko, “Testing for ellip-
soidal symmetry of a multivariate distribution,” High
 
2
= Tr [nSt,n ] = n2 Tr S2t,n ,


Dimensional Probability II, pp. 493–510, 2000.
where we denote [5] A. Manzotti, F. J. Pérez, and A. J. Quiroz, “A statistic
t(1)
 
for testing the null hypothesis of elliptical symmetry,”
t =  ...  . (77) Journal of Multivariate Analysis, vol. 81, no. 2, pp. 274–
 
(p) 285, 2002.
t
[6] J. R. Schott, “Testing for elliptical symmetry in
In our setup, the sample covariance matrix satisfies the fol- covariance-matrix-based analyses,” Statistics & Proba-
lowing relation, bility Letters, vol. 60, no. 4, pp. 395–404, 2002.
n n −1/2 −1/2 [7] F. W. Huffer and C. Park, “A test for elliptical symmetry,”
1X 1 X Tn xi x>
i Tn
St,n = tni t>
ni = 2 (78) Journal of Multivariate Analysis, vol. 98, no. 2, pp. 256–
n i=1 n i=1 −1/2 281, 2007.
Tn xi
n [8] D. Cassart, M. Hallin, and D. Paindaveine, “Optimal
1 −1/2 X xi x> 1 detection of Fechner-asymmetry,” Journal of Statistical
= Tn >
i
−1 T−1/2
n = I,
n x Tn xi
i=1 i
p Planning and Inference, vol. 138, no. 8, pp. 2499–2525,
2008.
and therefore,
[9] C. Diks and H. Tong, “A test for symmetries of multivari-
1 1 ate probability distributions,” Biometrika, vol. 86, no. 3,
Tr S2t,n = 2 Tr (I) = .


(79)
p p pp. 605–614, 1999.
By Lemma 5, for {wi }ni=1 i.i.d. uniformly distributed over [10] L. X. Zhu and G. Neuhaus, “Nonparametric Monte Carlo
Sp−1 with n > p, tests for multivariate distributions,” Biometrika, vol. 87,

1 no. 4, pp. 919–928, 2000.
Tr S2w,n > = Tr S2t,n .


(80) [11] R. Z. Li, K. T. Fang, and L. X. Zhu, “Some QQ
p probability plots to test spherical and elliptical symme-
Since γ2 (α, p) > 0, from (75) we infer that tG ({wi }) first- try,” Journal of Computational and Graphical Statistics,
order stochastically dominates tG ({tni }). vol. 6, no. 4, pp. 435–450, 1997.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

12

[12] L. X. Zhu and G. Neuhaus, “Conditional tests for ellipti- [30] D. E. Tyler, “A distribution-free M-estimator of multi-
cal symmetry,” Journal of Multivariate Analysis, vol. 84, variate scatter,” The Annals of Statistics, vol. 15, no. 1,
no. 2, pp. 284–298, 2003. pp. 234–251, 1987.
[13] R. J. Serfling, “Multivariate symmetry and asymmetry,” [31] G. Frahm, “Generalized elliptical distributions: theory
Encyclopedia of statistical sciences, vol. 8, pp. 5338– and applications,” Universität zu Köln, 2004.
5345, 2006. [32] Y. I. Abramovich, N. K. Spencer, and M. D. Turley,
[14] L. Sakhanenko, “Testing for ellipsoidal symmetry: A “Time-varying autoregressive (TVAR) models for mul-
comparison study,” Computational statistics and data tiple radar observations,” IEEE Transactions on Signal
analysis, vol. 53, no. 2, pp. 565–581, 2008. Processing, vol. 55, no. 4, pp. 1298–1311, 2007.
[15] E. J. Kelly, “An adaptive detection algorithm,” IEEE [33] F. Pascal, Y. Chitour, J. P. Ovarlez, P. Forster, and
Transactions on Aerospace and Electronic Systems, no. 2, P. Larzabal, “Covariance structure maximum-likelihood
pp. 115–127, 1986. estimates in compound Gaussian noise: Existence and
[16] I. Soloveychik and A. Wiesel, “Group symmetric robust algorithm analysis,” IEEE Transactions on Signal Pro-
covariance estimation,” IEEE Transactions on Signal cessing, vol. 56, no. 1, pp. 34–48, 2008.
Processing, vol. 64, no. 1, pp. 244–257, 2016. [34] F. Bandiera, O. Besson, and G. Ricci, “Knowledge-aided
[17] R. Adamczak, A. Litvak, A. Pajor, and N. Tomczak- covariance matrix estimation and adaptive detection in
Jaegermann, “Quantitative estimates of the convergence compound-Gaussian noise,” IEEE Transactions on Signal
of the empirical covariance matrix in log-concave ensem- Processing, vol. 58, no. 10, pp. 5391–5396, 2010.
bles,” Journal of American Mathematic Society, vol. 23, [35] A. Wiesel and T. Zhang, “Structured robust covariance
no. 2, pp. 535–561, 2010. estimation,” Foundations and Trends in Signal Process-
[18] J. Schäfer and K. Strimmer, “A shrinkage approach to ing, vol. 8, no. 3, pp. 127–216, 2015.
large-scale covariance matrix estimation and implications [36] D. E. Tyler, “Statistical analysis for the angular central
for functional genomics,” Statistical applications in ge- Gaussian distribution on the sphere,” Biometrika, vol. 74,
netics and molecular biology, vol. 4, no. 1, 2005. no. 3, pp. 579–589, 1987.
[19] J. Xie and P. M. Bentler, “Covariance structure models [37] A. Wiesel, “Geodesic convexity and covariance estima-
for gene expression microarray data,” Structural Equa- tion,” IEEE Transactions on Signal Processing, vol. 60,
tion Modeling, vol. 10, no. 4, pp. 566–582, 2003. no. 12, pp. 6182–6189, 2012.
[20] A. Hero and B. Rajaratnam, “Hub discovery in partial [38] T. Zhang, A. Wiesel, and M. S. Greco, “Multivariate
correlation graphs,” IEEE Transactions on Information generalized gaussian distribution: Convexity and graph-
Theory, vol. 58, no. 9, pp. 6064–6078, 2012. ical models,” IEEE Transactions on Signal Processing,
[21] G. Derado, F. D. Bowman, and C. D. Kilts, “Modeling vol. 61, no. 16, pp. 4141–4148, 2013.
the spatial and temporal dependence in fMRI data,” [39] E. Conte, M. Longo, and M. Lops, “Modelling and sim-
Biometrics, vol. 66, no. 3, pp. 949–957, 2010. ulation of non-Rayleigh radar clutter,” IEE Proceedings
[22] S. L. Lauritzen, “Graphical models,” Oxford University F on Radar and Signal Processing, vol. 138, no. 2, pp.
Press, 1996. 121–130, 1991.
[23] O. Banerjee, L. El Ghaoui, and A. d’Aspremont, “Model [40] D. Middleton, “Man-made noise in urban environments
selection through sparse maximum likelihood estimation and transportation systems: Models and measurements,”
for multivariate Gaussian or binary data,” The Journal of IEEE Transactions on Communications, vol. 21, no. 11,
Machine Learning Research, vol. 9, pp. 485–516, 2008. pp. 1232–1241, 1973.
[24] J. Bai and S. Shi, “Estimating high dimensional covari- [41] W. A. Stahel, “Breakdown of covariance estima-
ance matrices and its applications,” Annals of Economics tors,” Fachgruppe für Statistik, Eidgenössische Techn.
and Finance, vol. 12, no. 2, pp. 199–215, 2011. Hochsch., 1981.
[25] O. Ledoit and M. Wolf, “Improved estimation of the [42] D. L. Donoho, “Breakdown properties of multivariate lo-
covariance matrix of stock returns with an application cation estimates,” PhD qualifying paper, Dept. Statistics,
to portfolio selection,” Journal of Empirical Finance, Harvard Univ., 1982.
vol. 10, no. 5, pp. 603–621, 2003. [43] P. J. Rousseeuw, “Multivariate estimation with high
[26] J. H. Friedman, “Regularized discriminant analysis,” breakdown point,” Mathematical Statistics and Applica-
Journal of the American statistical association, vol. 84, tions, vol. 8, pp. 283–297, 1985.
no. 405, pp. 165–175, 1989. [44] P. L. Davies, “Asymptotic behaviour of S-estimates of
[27] J. Dahmen, D. Keysers, M. Pitz, and H. Ney, “Structured multivariate location parameters and dispersion matri-
covariance matrices for statistical image object recogni- ces,” Annals of Statistics, pp. 1269–1292, 1987.
tion,” Mustererkennung, pp. 99–106, 2000. [45] H. P. Lopuhaa, “On the relation between S-estimators and
[28] P. J. Huber, “Robust estimation of a location parameter,” M-estimators of multivariate location and covariance,”
The Annals of Mathematical Statistics, vol. 35, no. 1, pp. Annals of Statistics, pp. 1662–1683, 1989.
73–101, 1964. [46] O. Hössjer, “On the optimality of S-estimators,” Statistics
[29] R. A. Maronna, “Robust M-estimators of multivariate and Probability Letters, vol. 14, no. 5, pp. 413–419,
location and scatter,” The annals of statistics, pp. 51–67, 1992.
1976. [47] P. J. Rousseeuw and K. V. Driessen, “A fast algorithm for

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Signal Processing. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TSP.2023.3267993

13

the minimum covariance determinant estimator,” Techno- [65] E. Garcia-Portugués and T. Verdebout, “An overview
metrics, vol. 41, no. 3, pp. 212–223, 1999. of uniformity tests on the hypersphere,” arXiv preprint
[48] T. Leonard and J. S. J. Hsu, “Bayesian inference for arXiv:1804.00286, 2018.
a covariance matrix,” The Annals of Statistics, vol. 20, [66] B. De Finetti, “Funzione caratteristica di un fenomeno
no. 4, pp. 1669–1696, 1992. aleatorio,” Atti del Congresso Internazionale dei Matem-
[49] I. Alvarez, J. Niemi, and M. Simpson, “Bayesian atici: Bologna del 3 al 10 de settembre di 1928, pp.
inference for a covariance matrix,” arXiv preprint 179–190, 1929.
arXiv:1408.4050, 2014. [67] ——, “La prévision: ses lois logiques, ses sources sub-
[50] R. Yang and J. O. Berger, “Estimation of a covariance jectives,” Annales de l’institut Henri Poincaré, vol. 7,
matrix using the reference prior,” The Annals of Statistics, no. 1, pp. 1–68, 1937.
pp. 1195–1211, 1994. [68] D. M. Cifarelli and E. Regazzini, “De finetti’s contribu-
[51] O. Ledoit and M. Wolf, “Honey, i shrunk the sample co- tion to probability and statistics,” Statistical Science, pp.
variance matrix,” The Journal of Portfolio Management, 253–282, 1996.
vol. 30, no. 4, pp. 110–119, 2004. [69] B. Ajne, “A simple test for uniformity of a circular
[52] Y. Chen, A. Wiesel, and A. O. Hero, “Robust shrink- distribution,” Biometrika, vol. 55, no. 2, pp. 343–354,
age estimation of high-dimensional covariance matrices,” 1968.
IEEE Transactions on Signal Processing, vol. 59, no. 9, [70] R. J. Beran, “Testing for uniformity on a compact homo-
pp. 4097–4107, 2011. geneous space,” Journal of Applied Probability, vol. 5,
[53] R. Couillet and M. R. McKay, “Large dimensional anal- no. 1, pp. 177–195, 1968.
ysis and optimization of robust shrinkage covariance ma- [71] S. Babic, L. Gelbgras, M. Hallin, and C. Ley, “Optimal
trix estimators,” arXiv preprint arXiv:1401.4083, 2014. tests for elliptical symmetry: specified and unspecified
[54] I. Soloveychik and A. Wiesel, “Performance analysis location,” ULB–Universite Libre de Bruxelles, 2019.
of Tyler’s covariance estimator,” IEEE Transactions on [72] G. Frahm and U. Jaekel, “Tyler’s M-estimator, random
Signal Processing, vol. 63, no. 2, pp. 418–426, 2015. matrix theory, and generalized elliptical distributions
[55] O. Besson and Y. I. Abramovich, “On the fisher in- with applications to finance,” Technical report, Helmut
formation matrix for multivariate elliptically contoured Schmidt University, 2007.
distributions,” IEEE Signal Processing Letters, vol. 20, [73] P. E. Jupp, “Data-driven Sobolev tests of uniformity on
no. 11, pp. 1130–1133, 2013. compact Riemannian manifolds,” The Annals of Statis-
[56] M. Greco and F. Gini, “Cramér-Rao lower bounds on tics, vol. 36, no. 3, pp. 1246–1260, 2008.
covariance matrix estimation for complex elliptically [74] T. Konstantopoulos and L. Yuan, “On the extendibility
symmetric distributions,” IEEE Transactions on Signal of finitely exchangeable probability measures,” Transac-
Processing, vol. 61, no. 24, pp. 6401–6409, 2013. tions of the American Mathematical Society, vol. 371,
[57] L. Duembgen, “The asymptotic behavior of Tylers M- no. 10, pp. 7067–7092, 2019.
estimator of scatter in high dimension,” Ann. Inst. Statist. [75] N. C. Weber, “A martingale approach to central limit
Math, vol. 50, pp. 471–491, 1997. theorems for exchangeable random variables,” Journal of
[58] G. Frahm and K. Glombek, “Semicircle law of Tylers Applied Probability, vol. 17, no. 3, pp. 662–673, 1980.
M-estimator for scatter,” Statistics & Probability Letters, [76] A. Wiesel, “Unified framework to regularized covariance
vol. 82, no. 5, pp. 959–964, 2012. estimation in scaled gaussian models,” IEEE Transac-
[59] T. Zhang, X. Cheng, and A. Singer, “Marchenko-Pastur tions on Signal Processing, vol. 60, no. 1, pp. 29–38,
Law for Tyler’s and Maronna’s M-estimators,” arXiv 2012.
preprint arXiv:1401.3424, 2014. [77] S. Babić, C. Ley, and M. Palangetić, “Elliptical symmetry
[60] R. Couillet, F. Pascal, and J. W. Silverstein, “Robust tests in R,” arXiv preprint arXiv:2011.12560, 2020.
M-estimation for array processing: a random matrix [78] R. L. Taylor and R. F. Patterson, “Strong laws of large
approach,” IEEE Transactions on Signal Process, 2012. numbers for arrays of row-wise exchangeable random
[61] ——, “Robust estimates of covariance matrices in the elements,” International Journal of Mathematics and
large dimensional regime,” IEEE Transactions on Infor- Mathematical Sciences, vol. 8, no. 1, pp. 135–144, 1985.
mation Theory, vol. 60, no. 11, pp. 7269–7278, 2014. [79] I. Soloveychik, “Central limit theorem for symmet-
[62] F. Pascal, Y. Chitour, J.-P. Ovarlez, P. Forster, and ric exchangeable random variables,” arXiv preprint
P. Larzabal, “Covariance structure maximum-likelihood arXiv:2006.10819, 2020.
estimates in compound Gaussian noise: Existence and [80] J. J. Benedetto and M. Fickus, “Finite normalized
algorithm analysis,” IEEE Transactions on Signal Pro- tight frames,” Advances in Computational Mathematics,
cessing, vol. 56, no. 1, pp. 34–48, 2007. vol. 18, no. 2-4, pp. 357–385, 2003.
[63] E. Giné, “Invariant tests for uniformity on compact [81] M. Reimer, “Multivariate polynomial approximation,”
Riemannian manifolds based on Sobolev norms,” The Birkhäuser, vol. 144, 2012.
Annals of statistics, pp. 1243–1266, 1975. [82] V. A. Marcenko and L. A. Pastur, “Distribution of
[64] M. J. Prentice, “On invariant tests of uniformity for eigenvalues for some sets of random matrices,” Sbornik:
directions and orientations,” The Annals of Statistics, Mathematics, vol. 1, no. 4, pp. 457–483, 1967.
vol. 6, no. 1, pp. 169–176, 1978.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2023 at 02:30:15 UTC from IEEE Xplore. Restrictions apply.