# Journal of Multivariate Analysis 101 (2010) 2571–2586

Contents lists available at ScienceDirect
Journal of Multivariate Analysis
journal homepage: www.elsevier.com/locate/jmva
A multivariate version of Hoeffding’s Phi-Square
Sandra Gaißer
a
, Martin Ruppert
b,∗
, Friedrich Schmid
a
a
Department of Economic and Social Statistics, University of Cologne, Albertus-Magnus-Platz, 50923 Köln, Germany
b
Graduate School of Risk Management, University of Cologne, Albertus-Magnus-Platz, 50923 Köln, Germany
a r t i c l e i n f o
Article history:
Available online 23 July 2010
AMS 2000 subject class:
Primary 62H20
60F05
62G20
Secondary 60G15
62G10
Keywords:
Multivariate measure of association
Copula
Nonparametric estimation
Empirical copula process
Weak convergence
Nonparametric bootstrap
Strong mixing
a b s t r a c t
A multivariate measure of association is proposed, which extends the bivariate copula-
based measure Phi-Square introduced by Hoeffding [22]. We discuss its analytical
properties and calculate its explicit value for some copulas of simple form; a simulation
procedure to approximate its value is provided otherwise. A nonparametric estimator for
multivariate Phi-Square is derived and its asymptotic behavior is established based on
the weak convergence of the empirical copula process both in the case of independent
observations anddependent observations fromstrictly stationary strong mixing sequences.
The asymptotic variance of the estimator can be estimated by means of nonparametric
bootstrap methods. For illustration, the theoretical results are applied to financial asset
return data.
1. Introduction
Measuring the degree of association between the components of a d-dimensional (d > 2) random vector X = (X
1
,
. . . , X
d
) has attracted much interest among scientists and practitioners in recent years. Naturally, such measures of
multivariate association are based on the copula of the random vector X, i.e., they are invariant with respect to the
marginal distributions of the random variables X
i
, i = 1, . . . , d. Wolff [49] introduces a class of multivariate measures
of association which is based on the L
1
- and L

-norms of the difference between the copula and the independence copula
(see [16] for various extensions). Other authors generalize existing bivariate measures of associationto the multivariate case.
For example, various multivariate extensions of Spearman’s Rho are considered by Nelsen [33] and Schmid and Schmidt
[38–40]. Blomqvist’s Beta is generalized by Úbeda-Flores [47] and Schmid and Schmidt [41], whereas a multivariate version
of Gini’s Gamma is proposed by Behboodian et al. [2]. Further, Joe [27] and Nelsen [33] discuss multivariate generalizations
of Kendall’s Tau. A multivariate version of Spearman’s footrule is considered by Genest et al. [19]. Joe [26,25] investigates
another type of multivariate measures whichis based onthe Kullback–Leibler mutual information. General considerations of
multivariate measures of association, in particular of measures of concordance, are discussed in [49,27,11,45]. For a survey of
copula-based measures of multivariate association, we refer to [42]. Despite their common basis on the copula of a random
vector, the aforementionedmeasures generally differ withregardto their analytical properties. We do not knowof a measure
that is superior in every respect and the choice of an appropriate measure depends essentially on the type of application.

Corresponding author.
E-mail addresses: sandra.gaisser@uni-koeln.de (S. Gaißer), martin.ruppert@uni-koeln.de (M. Ruppert), schmid@wiso.uni-koeln.de (F. Schmid).
doi:10.1016/j.jmva.2010.07.006
2572 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
In the present paper, we propose a multivariate extension of the bivariate measure of association which was suggested by
Wassily Hoeffding in his seminal doctoral dissertation (see [22,17]). Multivariate Hoeffding’s Phi-Square (which we denote
by Φ
2
) is based on a Cramér–von Mises functional and can be interpreted as the (normalized) average squared difference
between the copula of a random vector and the independence copula. It offers a set of properties which are advantageous
for many applications. For example, the measure is zero if and only if the components of X are stochastically independent.
This leads to the construction of statistical tests for stochastic independence based on Hoeffding’s Phi-Square (cf. [20]).
Multivariate Hoeffding’s Φ
2
is based on a L
2
-type distance between the copula C of a random vector and the independence
copula and attains its maximal value in the case of comonotonicity, i.e., when the copula equals the upper Fréchet–Hoeffding
bound. The concept of comonotonicity is of interest, e.g., in actuarial science or finance (see [9,10]). Further, multivariate
Hoeffding’s Phi-Square canbe estimatedwithlowcomputational complexity, evenfor large dimensiond, andnonparametric
statistical inference for the estimator can be established based on the empirical copula. The measure can thus be used to
quantify the degree of association of multivariate empirical data (e.g., financial asset returns).
The paper is organized as follows. Section 2 introduces relevant definitions and notation. Multivariate Hoeffding’s Φ
2
is
introduced in Section 3 and some of its analytical properties are investigated. We calculate the explicit value of multivariate
Hoeffding’s Φ
2
for some copulas of simple formand describe a simulation algorithmto approximate its value in cases where
the copula is of a more complicated form. In Section 4, a nonparametric estimator
¨
Φ
2
n
for Φ
2
based on the empirical copula
is derived. We establish its asymptotic behavior both in the case of independent observations and dependent observations
from strictly stationary strong mixing sequences. In general, two cases need to be distinguished: If Φ
2
> 0, asymptotic
normality of

n(
¨
Φ
2
n
− Φ
2
) can be shown using the functional Delta-method. The asymptotic variance can consistently be
estimated by means of a nonparametric (moving block) bootstrap method. When Φ
2
= 0, weak convergence of n
¨
Φ
2
n
to a
non-normal distribution follows. We show how the estimator can be adapted to account for small sample sizes. Section 5
presents an empirical study of financial contagion related to the bankruptcy of Lehman Brothers Inc. in September 2008
using multivariate Hoeffding’s Φ
2
. A brief conclusion is given in Section 6.
2. Notation and definitions
Let X = (X
1
, . . . , X
d
) be a d-dimensional random vector (d ≥ 2), defined on some probability space (Ω, F , P), with
distribution function F(x) = P(X
1
≤ x
1
, . . . , X
d
≤ x
d
) for x = (x
1
, . . . , x
d
) ∈ R
d
and marginal distribution functions
F
i
(x) = P(X
i
≤ x) whose survival functions are denoted by
¯
F
i
(x) = P(X
i
> x) for x ∈ R and i = 1, . . . , d. If not stated
otherwise, we always assume that the F
i
are continuous functions. According to Sklar’s theorem [44], there exists a unique
copula C : [0, 1]
d
→ [0, 1] such that
F(x) = C(F
1
(x
1
), . . . , F
d
(x
d
)) for all x ∈ R
d
.
The copula C is also referred to as the copula of the randomvector X(if an explicit reference to the randomvector is required,
we denote its copula by C
X
). It represents the joint distribution function of the random variables U
i
= F
i
(X
i
), i = 1, . . . , d,
i.e., C(u
1
, . . . , u
d
) = P(U
1
≤ u
1
, . . . , U
d
≤ u
d
) = P(U ≤ u) for all u = (u
1
, . . . , u
d
) ∈ [0, 1]
d
and random vector
U = (U
1
, . . . , U
d
). Moreover, C(u) = F(F
−1
1
(u
1
), . . . , F
−1
d
(u
d
)) for all u ∈ [0, 1]
d
. The generalized inverse function G
−1
is
defined by G
−1
(u) := inf{x ∈ R ∪ {∞}|G(x) ≥ u} for all u ∈ (0, 1] and G
−1
(0) := sup{x ∈ R ∪ {−∞}|G(x) = 0}.
The survival copula
˘
C of C is given by
˘
C(u) = P(U > 1−u) where 1−u = (1−u
1
, . . . , 1−u
d
). It represents the copula
of the randomvector −X = (−X
1
, . . . , −X
d
). The randomvector X = (X
1
, . . . , X
d
) is said to be marginally symmetric about
a ∈ R
d
if X
1
, . . . , X
d
1
, . . . , a
d
, respectively (i.e., F
k
(a
k
+ x) =
¯
F
k
(a
k
− x) for all x ∈ R, k = 1, . . . , d).
d
is given if, and only if, the random vector X is marginally symmetric and the copula
C
X
of X equals its survival copula
˘
C
X
, that is C
X
(u) = P(U ≤ u) = P(U > 1 − u) =
˘
C
X
(u) for all u ∈ [0, 1]
d
. Let α
k
be
a strictly decreasing transformation of the kth component X
k
of X, defined on the range of X
k
, and denote the transformed
random vector by α
k
(X) = (X
1
, . . . , X
k−1
, α
k
(X
k
), X
k+1
, . . . , X
d
), k = 1, . . . , d. The random vector X is then said to be
jointly symmetric about a ∈ R
d
if, and only if, X is marginally symmetric and its copula C
X
equals the copula C
α
k
(X)
of
α
k
(X), k = 1, . . . , d. Note that joint symmetry implies radial symmetry; see [34], Chapter 2.7.
Every copula C is bounded in the following sense:
W(u) := max{u
1
+ · · · + u
d
−(d − 1), 0}
≤ C(u) ≤ min{u
1
, . . . , u
d
} =: M(u) for all u ∈ [0, 1]
d
,
where M and W are called the upper and lower Fréchet–Hoeffding bounds, respectively. The upper bound M is a copula
itself and is also known as the comonotonic copula. It represents the copula of X
1
, . . . , X
d
if F
1
(X
1
) = · · · = F
d
(X
d
) with
probability one, i.e., if there is (with probability one) a strictly increasing functional relationship between X
i
and X
j
(i = j).
By contrast, the lower bound W is a copula only for dimension d = 2. Moreover, the independence copula Π with
Π(u) :=
d
¸
i=1
u
i
, u ∈ [0, 1]
d
,
describes the dependence structure of stochastically independent random variables X
1
, . . . , X
d
. For a detailed treatment of
copulas, we refer to [34,28]. Further, the space

([0, 1]
d
) is the collection of all uniformly bounded real-valued functions
defined on [0, 1]
d
. It is equipped with the uniform metric m(f
1
, f
2
) = sup
t∈[0,1]
d |f
1
(t) − f
2
(t)|.
S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2573
3. A multivariate version of Hoeffding’s Phi-Square
In his seminal dissertation, Hoeffding [22] suggests the following measure Φ
2
to quantify the amount of association
between the components of the two-dimensional random vector X with copula C
Φ
2
= 90

[0,1]
2
{C(u
1
, u
2
) − Π(u
1
, u
2
)}
2
du
1
du
2
. (1)
For the d-dimensional random vector X with copula C, we define a multivariate version of Φ
2
by
Φ
2
:= Φ
2
(C) = h(d)

[0,1]
d
{C(u) − Π(u)}
2
du, (2)
with normalization constant h(d) = [

[0,1]
d
{M(u) − Π(u)}
2
du]
−1
, whose explicit formis calculated later. In particular, Φ
2
can be regarded as a continuous functional on the subset of l

([0, 1]
d
) for which the integral on the right-hand side of Eq.
(2) is well-defined. This relationship is used in Section 4 where the statistical properties of Φ
2
are derived. An alternative
multivariate measure of association can be defined by
Φ := Φ(C) = +

Φ
2
(C).
This measure can be interpreted as the normalized average distance between the copula C and the independence copula Π
with respect to the L
2
-norm. Bivariate measures of association of this form based on the L
1
- and L

-norms are considered
by Schweizer and Wolff [43]. For the related multivariate case, we refer to [49,16]. If C is the copula of the random vector X,
we also refer to Φ
2
and Φ as Φ
2
X
and Φ
X
, respectively. Various analytical properties of Φ
2
are discussed next. They can also
be established for Φ.
Normalization: An important property of Φ
2
is that
Φ
2
= 0 if and only if C = Π.
In order to motivate the specific form of the normalization factor h(d) in Eq. (2), we calculate the value of the defining
integral for the lower and upper Fréchet–Hoeffding bounds, respectively. For C = M, we have
h(d)
−1
=

[0,1]
d
{M(u) − Π(u)}
2
du =
2
(d + 1)(d + 2)

1
2
d
d!
d
¸
i=0

i +
1
2

+

1
3

d
, (3)
and for C = W, we obtain
g(d)
−1
=

[0,1]
d
{W(u) − Π(u)}
2
du =
2
(d + 2)!
− 2
i=0
¸
d

d
i

(−1)
i
1
(d + 1 + i)!
+

1
3

d
. (4)
The calculations are outlined in Appendix A. Both expressions converge to zero as the dimension d tends to infinity. In
particular, direct calculations yield that
h(d)
−1
≥ g(d)
−1
for all d ≥ 2,
such that the range of Φ
2
as defined in (2) is restricted to the interval [0, 1], i.e.,
0 ≤ Φ
2
≤ 1 and Φ
2
= 1 iff C = M for d ≥ 3. (5)
Note that for dimension d = 2, it holds that Φ
2
= 1 iff C = M or C = W since g(2) = h(2).
Remark. In the bivariate case, Hoeffding’s Phi-Square thus represents a measure for strictly monotone functional
dependence. In consequence, a value of one of Hoeffding’s Phi-Square also implies that the random variables X
1
and X
2
are completely dependent, i.e., that there exists a one-to-one function ψ (which is not necessarily monotone) such that
P(X
2
= ψ(X
1
)) = 1 (cf. [23,30]). However, the converse does not hold. For example, two random variables X
1
and X
2
are
completely dependent if their copula is a shuffle of M. According to Nelsen [34], Theorem 3.2.2, we find shuffles of M which
(uniformly) approximate the independence copula arbitrarily closely (see also [32]). Hence, the value of bivariate Hoeffding’s
Phi-Square can be made arbitrarily small for completely dependent random variables.
Invariance with respect to permutations: For every permutation π of the components of X we have Φ
2
X
= Φ
2
π(X)
according to
Fubini’s Theorem.
Monotonicity: For copulas C
1
and C
2
with Π(u) ≤ C
1
(u) ≤ C
2
(u) ≤ M(u) for all u ∈ [0, 1]
d
, we have Φ
2
(C
1
) ≤ Φ
2
(C
2
).
For copulas C
3
and C
4
such that W(u) ≤ C
3
(u) ≤ C
4
(u) ≤ Π(u) for all u ∈ [0, 1]
d
, it follows that Φ
2
(C
3
) ≥ Φ
2
(C
4
). This
property generalizes a result of Yanagimoto [50] to the multivariate case.
Invariance under strictly monotonic transformations: The behavior of Φ
2
with respect to strictly monotonic transformations
is given in the next proposition.
2574 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
Proposition 1. Let X be a d-dimensional random vector with copula C.
(i) For dimension d ≥ 2, Φ
2
X
is invariant with regard to strictly increasing transformations of one or several components of X.
(ii) For dimension d = 2, Φ
2
X
is invariant under strictly decreasing transformations of one or both components of X.
For d ≥ 3, let α
k
be a strictly decreasing transformation of the kth component X
k
of X, k ∈ {1, . . . , d}, and let α
k
(X) =
(X
1
, . . . , X
k−1
, α
k
(X
k
), X
k+1
, . . . , X
d
). Then, Φ
2
X
= Φ
2
α
k
(X)
if one of the following three conditions holds:
• X
1
, . . . , X
d
are jointly symmetric about a ∈ R
d
, or
• (X
1
, . . . , X
k−1
, X
k+1
, . . . , X
d
) is stochastically independent of X
k
, or
• X
1
, . . . , X
k−1
, X
k+1
, . . . , X
d
are mutually stochastically independent.
The proof is outlined in Appendix C. If part (ii) of the above proposition is not satisfied, equality of Φ
2
X
and Φ
2
α
k
(X)
does not
hold in general. For example, let the copula C of X be the comonotonic copula, i.e., C = M, and let α
1
be a strictly decreasing
transformation of the first component X
1
of X. Then, C
α
1
(X)
(u) = M(u) for all u ∈ [0, 1]
d
, implying that Φ
2
α
1
(X)
< Φ
2
X
.
However, when applying the (strictly decreasing) inverse function α
−1
1
of α
1
to the first component of α
1
(X), we obtain that
Φ
2
α
−1
1

1
(X))
= Φ
2
X
> Φ
2
α
1
(X)
.
Duality: For dimension d = 2, Hoeffding’s Phi-Square satisfies duality, i.e. Φ
2
X
= Φ
2
−X
or equivalently Φ
2
(C) = Φ
2
(
˘
C), since
Φ
2
is invariant under strictly decreasing transformations of all components of X according to Proposition 1, part (ii). For
dimension d ≥ 3, duality does not hold in general except in the case that the random vector X is radially symmetric about
a ∈ R
d
. Taylor [45] discusses duality in the context of multivariate measures of concordance.
Continuity: If {C
m
}
m∈N
is a sequence of copulas such that C
m
(u) → C(u) for all u ∈ [0, 1]
d
, then Φ
2
(C
m
) → Φ
2
(C) as a direct
consequence of the dominated convergence theorem.
Examples. (i) Let C be the d-dimensional Farlie–Gumbel–Morgenstern copula defined by C(u
1
, . . . , u
d
) =
¸
d
i=1
u
i
+
θ
¸
d
i=1
u
i
(1 − u
i
), |θ| ≤ 1. Then, Hoeffding’s Phi-Square is given by
Φ
2
= h(d)θ
2

1
30

d
, d ≥ 2.
It follows that Φ
2
≤ 1/10 for d ≥ 2. This illustrates the restricted range of dependence modeled by the family of
Farlie–Gumbel–Morgenstern copulas.
(ii) Let C(u) = θM(u) +(1 −θ)Π(u) with 0 θ 1. Then
Φ
2
= θ
2
, d ≥ 2.
Note that for this family of copulas, the value of Φ
2
does not depend on the dimension d.
It is difficult to derive an explicit expression for Φ
2
if C is of a more complicated structure than in examples (i) and (ii).
In this case, the value of Φ
2
needs to be determined by simulation. The following equivalent representation of Φ
2
is useful
for this purpose:
Φ
2
= h(d)

[0,1]
d
{C(u) − Π(u)}
2
du = h(d)E
Π
¸
{C(U) − Π(U)}
2
¸
, (6)
where the randomvector U = (U
1
, . . . , U
d
) is uniformly distributed on [0, 1]
d
with stochastically independent components
U
i
, i = 1, . . . , d (which is indicated by the subscript Π). Thus, an approximation of Φ
2
is obtained by estimating the
expectation on the right-hand side of Eq. (6) consistently as follows:
ˆ
E
Π
¸
{C(U) − Π(U)}
2
¸
=
1
n
n
¸
i=1
{C(U
i
) − Π(U
i
)}
2
, (7)
with U
1
, . . . , U
n
being independent and identically distributed Monte Carlo replications fromU. For illustration we compute
the approximated value of Φ
2
for the equi-correlated Gaussian copula, given by
C
ρ
(u
1
, . . . , u
d
) = Φ
ρ

Φ
−1
(u
1
), . . . , Φ
−1
(u
d
)

(8)
with
Φ
ρ
(x
1
, . . . , x
d
) =

x
1
−∞
· · ·

x
d
−∞
(2π)

d
2

Σ
ρ

1
2
exp

1
2
x

Σ
−1
ρ
x

dx
d
· · · dx
1
and Σ
ρ
= ρ

11

+ (1 − ρ)I with −
1
d−1
< ρ < 1. The approximated values of Φ
2
and Φ for different choices of the
parameter ρ and for dimensions d = 2, d = 5, and d = 10 are displayed in Fig. 1.
The values of Φ
2
form a parabolic curve; see [22] for a power series representation of Φ
2
for the bivariate Gaussian
copula.
S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2575
Fig. 1. Gaussian copula. Approximated values of Φ
2
(left panel) and Φ (right panel) in the case of a d-dimensional equi-correlated Gaussian copula with
parameter ρ for dimension d = 2 (solid line), d = 5 (dashed line), and d = 10 (dotted line); calculations are based on n = 100, 000 Monte Carlo
replications.
4. Statistical inference for multivariate Hoeffding’s Phi-Square
Statistical inference for multivariate Hoeffding’s Phi-Square as introduced in formula (2) is based on the empirical copula
which was first discussed by Rüschendorf [37] and Deheuvels [8]. We derive a nonparametric estimator for multivariate
Hoeffding’s Phi-Square and establish its asymptotic behavior based on the weak convergence of the empirical copula
process. After illustrating our approach on the basis of independent observations, we generalize it to the case of dependent
observations from strictly stationary strong mixing sequences.
4.1. Nonparametric estimation
Let X
1
, . . . , X
n
be an (i.i.d.) random sample from the d-dimensional random vector X with distribution function F and
copula C. We assume that both F and C as well as the marginal distribution functions F
i
, i = 1, . . . , d, are completely
unknown. The copula C is estimated by the empirical copula
¨
C
n
, which is defined as
¨
C
n
(u) =
1
n
n
¸
j=1
d
¸
i=1
1
{
¨
U
ij,n
≤u
i
}
for u ∈ [0, 1]
d
(9)
with pseudo-observations
¨
U
ij,n
=
¨
F
i,n
(X
ij
) for i = 1, . . . , d and j = 1, . . . , n, and
¨
F
i,n
(x) =
1
n
¸
n
j=1
1
{X
ij
≤x}
for x ∈ R. As
¨
U
ij,n
=
1
n
(rank of X
ij
in X
i1
, . . . , X
in
), statistical inference is based on the ranks of the observations. For fixed n, we suppress
the subindex and refer to the pseudo-observations as
¨
U
ij
if it is clear from the context.
A nonparametric estimator for Φ
2
is then obtained by replacing the copula C in formula (2) by the empirical copula
¨
C
n
,
i.e.,
¨
Φ
2
n
:= Φ
2
(
¨
C
n
) = h(d)

[0,1]
d
¸
¨
C
n
(u) − Π(u)
¸
2
du.
The estimator is based on a Cramér–von Mises statistic and can explicitly be determined by
¨
Φ
2
n
= h(d)
¸

1
n

2 n
¸
j=1
n
¸
k=1
d
¸
i=1
(1 − max{
¨
U
ij
,
¨
U
ik
}) −
2
n

1
2

d n
¸
j=1
d
¸
i=1
(1 −
¨
U
2
ij
) +

1
3

d
¸
. (10)
The derivation is outlined in Appendix B. Obviously, an estimator for the alternative measure Φ is given by
¨
Φ
n
= +

¨
Φ
2
n
. The
asymptotic distributions of
¨
Φ
2
n
and
¨
Φ
n
can be deduced from the asymptotic behavior of the empirical copula process which
has been discussed, e.g., by Rüschendorf [37], Gänßler and Stute [18], Van der Vaart and Wellner [48], and Tsukahara [46].
The following result is shown in [14]:
Proposition 2. Let X
1
, . . . , X
n
be a random sample from the d-dimensional random vector X with joint distribution function F
and copula C. If the ith partial derivatives D
i
C(u) of C exist and are continuous for i = 1, . . . , d, the empirical process

n(
¨
C
n
−C)
converges weakly in

([0, 1]
d
) to the process G
C
which takes the form
2576 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
G
C
(u) = B
C
(u) −
d
¸
i=1
D
i
C(u)B
C
(u
(i)
). (11)
The process B
C
is a d-dimensional Brownian bridge on [0, 1]
d
with covariance function E{B
C
(u)B
C
(v)} = C(u ∧v) −C(u)C(v).
The vector u
(i)
corresponds to the vector where all coordinates, except the ith coordinate of u, are replaced by 1.
Then, asymptotic normality of
¨
Φ
2
n
and
¨
Φ
n
can be established.
Theorem 3. Under the assumptions of Proposition 2 and if C = Π, it follows that

n(
¨
Φ
2
n
− Φ
2
)
d
−→ Z
Φ
2 (12)
where Z
Φ
2 ∼ N(0, σ
2
Φ
2
) with
σ
2
Φ
2
= {2h(d)}
2

[0,1]
d

[0,1]
d
{C(u) − Π(u)}E{G
C
(u)G
C
(v)}{C(v) − Π(v)}dudv. (13)
Regarding the alternative measure Φ, we have

n(
¨
Φ
n
− Φ)
d
−→ Z
Φ
with Z
Φ
∼ N(0, σ
2
Φ
) and
σ
2
Φ
=
σ
2
Φ
2

2
= h(d)

[0,1]
d

[0,1]
d
{C(u) − Π(u)}E{G
C
(u)G
C
(v)}{C(v) − Π(v)}dudv

[0,1]
d
{C(u) − Π(u)}
2
du
.
The proof is given in Appendix C. Note that the assumption C = Π guarantees that the limiting random variable is
non-degenerated as implied by the form of the variance σ
2
Φ
2
. However, if C = Π, Proposition 2 and an application of the
continuous mapping theorem yield
n
¨
Φ
2
n
d
−→ h(d)

[0,1]
d
{G
Π
(u)}
2
du, as n → ∞, (14)
with
E
¸
h(d)

[0,1]
d
{G
Π
(u)}
2
du
¸
= h(d)
¸

1
2

d

1
3

d

d
6

1
3

d−1
¸
.
The asymptotic distribution of
¨
Φ
2
n
when C = Π is important for the construction of tests for stochastic independence
between the components of a multivariate random vector based on Hoeffding’s Phi-Square. In the bivariate setting, such
tests have been studied by Hoeffding [24] and Blum et al. [3]. Regarding the multivariate case, we mention Genest and
Rémillard [21] and Genest et al. [20] who consider various combinations of Cramér–von Mises statistics with special regard
to their asymptotic local efficiency. In our setting, a hypothesis test for H
0
: C = Π against H
1
: C = Π is performed
by rejecting H
0
if the value of n
¨
Φ
2
n
exceeds the (1 − α)-quantile of the limiting distribution in Eq. (14). The latter can be
determined by simulation; approximate critical values for the test statistic {h(d)}
−1
n
¨
Φ
2
n
are also provided in [20].
Remark. When the univariate marginal distribution functions F
i
are known, Hoeffding’s Phi-Square can also be estimated
using the theory of U-statistics. Consider the random variables U
ij
= F
i
(X
ij
), i = 1, . . . , d, j = 1, . . . , n with U
j
=
(U
1j
, . . . , U
dj
) having distribution function C. Since
Φ
2
= h(d)

[0,1]
d
{C(u) − Π(u)}
2
du
=

[0,1]
d

[0,1]
d

[0,1]
d
h(d)

d
¸
i=1
1
{x
i
≤u
i
}

d
¸
i=1
u
i

d
¸
i=1
1
{y
i
≤u
i
}

d
¸
i=1
u
i

dudC(x)dC(y),
an unbiased estimator of the latter based on the random sample U
1
, . . . , U
n
is given by the U-statistic
U
n
(ψ) =

n
2

−1 ¸
1≤j<k≤n
ψ(U
j
, U
k
)
S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2577
with kernel ψ of degree 2, defined by
ψ(x, y) = h(d)

[0,1,]
d

d
¸
i=1
1
{x
i
≤u
i
}

d
¸
i=1
u
i

d
¸
i=1
1
{y
i
≤u
i
}

d
¸
i=1
u
i

du, x, y ∈ [0, 1]
d
.
Results from the theory of U-statistics (see, e.g., Chapter 3 in [31]) and standard calculations yield that

n{U
n
(ψ) − Φ
2
} is
asymptotically normally distributed with mean zero and variance
σ
2
U
= Var{ψ
1
(X)}
= 4{h(d)}
2

[0,1]
d

[0,1]
d
{C(u) − Π(u)}{C(u ∧ v) − C(u)C(v)}{C(v) − Π(v)}dudv,
where ψ
1
(x) = E{ψ(X, Y)|X = x} with independent random vectors X, Y having distribution function C. The asymptotic
variance coincides with the asymptotic variance of

n(
¨
Φ
2
n
− Φ
2
) for known marginal distribution functions (cf. Eq. (13)).
In particular, U
n
(ψ) is degenerate when C = Π. The fact that both estimators have the same asymptotic distribution in the
case of known margins follows also from the relationship

n(
¨
Φ
2
n
− Φ
2
) =
1
n
3/2
n
¸
j=1
ψ(U
j
, U
j
) +

n
¸
n − 1
n
· U
n
(ψ) − Φ
2
¸
where the first term in the right equation converges to zero in probability for n → ∞. In the case of unknown marginal
distribution functions, the estimation of Φ
2
by means of U-statistics is more involved in comparison to the above approach
based on the empirical copula.
Now let {X
j
= (X
1j
, . . . , X
dj
)}
j∈Z
be a strictly stationary sequence of d-dimensional random vectors, being defined on a
probability space (Ω, F , P), with distribution function F, continuous marginal distribution functions F
i
, i = 1, . . . , d, and
copula C. In order to describe temporal dependence between the observations we use the concept of α-mixing or strong
mixing. Suppose Aand B are two σ-fields included in F and define
α(A, B) = sup
A∈A,B∈B
|P(A ∩ B) − P(A)P(B)|.
The mixing coefficient α
X
associatedwiththe sequence {X
j
}
j∈Z
is givenby α
X
(r) = sup
s≥0
α(F
s
, F
s+r
) where F
t
= σ{X
j
, j ≤
t} and F
t
= σ{X
j
, j ≥ t} denote the σ-fields generated by X
j
, j ≤ t, and X
j
, j ≥ t, respectively. The process {X
j
}
j∈Z
is said to
be strong mixing if
α
X
(r) → 0 for r → ∞.
Assume that our observations are realizations of the sample X
1
, . . . , X
n
and denote by
¨
Φ
2
n
the corresponding estimator for
Hoeffding’s Phi-Square calculated according to (10). The asymptotic behavior of
¨
Φ
2
n
is given next.
Theorem 4. Let X
1
, . . . , X
n
be observations from the strictly stationary strong mixing sequence {X
j
}
j∈Z
with coefficient α
X
(r)
satisfying α
X
(r) = O(r
−a
) for some a > 1. If the ith partial derivatives D
i
C(u) of C exist and are continuous for i = 1, . . . , d,
and C = Π, we have

n(
¨
Φ
2
n
− Φ
2
)
d
→ Z
Φ
2 ∼ N(0, σ
2
Φ
2
) with
σ
2
Φ
2
= {2h(d)}
2

[0,1]
d

[0,1]
d
E
¸
{C(u) − Π(u)}G

(u)G

(v){C(v) − Π(v)}
¸
dudv. (15)
The process G

has the same form as in Eq. (11) with process B
C
being replaced by the centered Gaussian process B

in [0, 1]
d
having covariance function
E{B

(u)B

(v)} =
¸
j∈Z
E
¸
{1
{U
0
≤u}
− C(u)}{1
{U
j
≤v}
− C(v)}
¸
,
where U
j
= (F
1
(X
1j
), . . . , F
d
(X
dj
)), j ∈ Z.
The proof is outlined in Appendix C. The result is based on the weak convergence of the empirical copula process to
the Gaussian process G

for strictly stationary strong mixing sequences; cf. also [15] for the asymptotic properties of the
smoothed empirical copula process in this setting. Note that the asymptotic variance in Eq. (15) depends not only on the
copula C as in the case of independent observations (cf. Eq. (13)) but also on the joint distribution of U
0
and U
j
, j ∈ Z.
Theorem 4 can be transferred to sequences with temporal dependence structures other than strong mixing; we refer, for
example, to [1,7,12].
Even though the asymptotic variance has a closed-form expression, it cannot be calculated explicitly in most cases but
has to be estimated adequately. For dependent observations, a bootstrap method, the (moving) block bootstrap, has been
2578 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
proposed by Künsch [29], which is briefly described in the following. Given the sample X
1
, . . . , X
n
, we define blocks of size
l, l(n) = o(n), of consecutive observations by
B
s,l
= {X
s+1
, . . . , X
s+l
}, s = 0, . . . , n − l.
The block bootstrap draws with replacement k blocks from the blocks B
s,l
, s = 0, . . . , n − l where we assume that n = kl
(otherwise the last block is shortened). With S
1
, . . . , S
k
being independent and uniformly distributed random variables on
{0, . . . , n − l}, the bootstrap sample thus comprises those observations from X
1
, . . . , X
n
which are among the k blocks
B
S
1
,l
, . . . , B
S
k
,l
, i.e.,
X
B
1
= X
S
1
+1
, . . . , X
B
l
= X
S
1
+l
, X
B
l+1
= X
S
2
+1
, . . . , X
B
n
= X
S
k
+l
.
The block length l is a function of n, i.e., l = l(n) with l(n) → ∞ as n → ∞. For a discussion regarding the choice of l(n);
see [29,5]. Denote by
¨
C
B
n
and
¨
F
B
n
the empirical copula and the empirical distribution function of the block bootstrap sample
X
B
1
, . . . , X
B
n
, respectively, and let
¨
Φ
2,B
n
be the corresponding estimator for Hoeffding’s Phi-Square. It follows that the block
bootstrap can be applied to estimate the asymptotic variance of

n(
¨
Φ
2
− Φ
2
).
Proposition 5. Let (X
B
j
)
j=1,...,n
be the block bootstrap sample from (X
j
)
j=1,...,n
, which are observations of a strictly stationary,
strong mixing sequence {X
j
}
j∈Z
of d-dimensional random vectors with distribution function F and copula C whose partial
derivatives exist and are continuous. Suppose further that

n(
¨
C
B
n

¨
C
n
) converges weakly in

([0, 1]
d
) to the same Gaussian
limit as

n(
¨
C
n
− C) in probability. If C = Π, the sequences

n(
¨
Φ
2
n
− Φ
2
) and

n(
¨
Φ
2,B
n

¨
Φ
2
n
) converge weakly to the same
Gaussian limit in probability.
The sequence

n(
¨
C
B
n

¨
C
n
) converges weakly in probability to the same Gaussian limit as

n(
¨
C
n
− C) if the (uniform)
empirical process

n(
¨
F
B
n

¨
F
n
) converges weakly in probability to the same Gaussian limit as

n(
¨
F
n
− F), provided that
all partial derivatives of the copula exist and are continuous. This can be shown analogously to the proof of Theorem 3.9.11
in [48] using the functional Delta-method (cf. also proof of Theorem4). The block bootstrap for empirical processes has been
discussed in various settings and for different dependence structures; for an overview see [35] and references therein. The
following sufficient conditions for

n(
¨
F
B
n

¨
F
n
) to converge weakly (in the space D([0, 1]
d
)) in probability to the appropriate
Gaussian process for strong mixing sequences are derived in [4]:

¸
r=0
(r + 1)
16(d+1)
α
1/2
X
(r) < ∞ and block length l(n) = O(n
1/2−ε
), ε > 0.
The results of a simulation study, which assesses the performance of the bootstrap variance estimator, are presented in
Section 4.2. Note that in the case of independent observations, the standard bootstrap, which draws with replacement n
single observations from the sample X
1
, . . . , X
n
, can be used to estimate the asymptotic variance of

n(
¨
Φ
2
n
− Φ
2
).
Theorem 4 together with Proposition 5 enables the calculation of an asymptotic (1 − α)-confidence interval for
Hoeffding’s Phi-Square Φ
2
∈ (0, 1), given by
¨
Φ
2
n
± Φ
−1

1 −
α
2

ˆ σ
B
Φ
2
n
/

n.
Further, an asymptotic hypothesis test for H
0
: Φ
2
= Φ
2
0
against H
1
: Φ
2
= Φ
2
0
with Φ
2
0
∈ (0, 1) can be constructed by
rejecting the null hypothesis at the confidence level α if

n
(
ˆ
Φ
2
n
− Φ
2
0
)
ˆ σ
B
Φ
2
n

> Φ
−1

1 −
α
2

.
Here, ( ˆ σ
B
Φ
2
n
)
2
denotes the consistent bootstrap variance estimator for σ
2
Φ
2
, obtained by the block bootstrap. Note that in the
case Φ
2
0
= 1, the copula corresponds to the upper Fréchet–Hoeffding bound M which does not possess continuous first
partial derivatives.
The above results can be extended to statistically analyze the difference of two Hoeffding’s Phi-Squares. In a financial
context, this may be of interest for assessing whether Hoeffding’s Phi-Square of one portfolio of financial assets significantly
differs fromthat of another portfolio (cf. Section 5). Suppose Φ
2
X
and Φ
2
Y
are multivariate Hoeffding’s Phi-Squares associated
with the strictly stationary sequences {X
j
}
j∈Z
and {Y
j
}
j∈Z
of d-dimensional random vectors with distribution functions F
X
and F
Y
, continuous marginal distribution functions, and copulas C
X
and C
Y
, respectively. Since the two sequences do not
have to be necessarily independent, consider the sequence {Z
j
= (X
j
, Y
j
)}
j∈Z
of 2d-dimensional random vectors with
joint distribution function F
Z
, continuous marginal distribution functions F
Z,i
, i = 1, . . . , 2d, and copula C
Z
such that
C
Z
(u, 1, . . . , 1) = C
X
(u) and C
Z
(1, . . . , 1, v) = C
Y
(v) for all u, v ∈ [0, 1]
d
.
Theorem 6. Let Z
1
= (X
1
, Y
1
), . . . , Z
n
= (X
n
, Y
n
) be observations of the strictly stationary, strong mixing sequence {Z
j
=
(X
j
, Y
j
)}
j∈Z
with strong mixing coefficient α
Z
satisfying α
Z
(r) = O(r
−a
) for some a > 1. If the ith partial derivatives of C
Z
exist
and are continuous for i = 1, . . . , 2d, and C
X
, C
Y
= Π, we have
S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2579

n
¸
ˆ
Φ
2
X
− Φ
2
X
−(
ˆ
Φ
2
Y
− Φ
2
Y
)
¸
d
−→ W ∼ N(0, σ
2
) as n → ∞,
where σ
2
= σ
2
Φ
2
X

2
Φ
2
Y
− 2σ
Φ
2
X

2
Y
with
σ
Φ
2
X

2
Y
= {2h(d)}
2

[0,1]
d

[0,1]
d
E
¸
{C
X
(u) − Π(u)}G

X
(u)G

Y
(v){C
Y
(v) − Π(v)}
¸
dudv,
and σ
2
Φ
2
X
= σ
Φ
2
X

2
X
and σ
2
Φ
2
Y
= σ
Φ
2
Y

2
Y
(cf. Eq. (15)). The processes G

X
and G

Y
are Gaussian processes on [0, 1]
d
with covariance
structure as given in Theorem 4.
The proof is given in Appendix C. Analogously to the discussion prior to Theorem6, an asymptotic confidence interval or
a statistical hypothesis test for the difference of two Hoeffding’s Phi-Squares can be formulated (cf. Section 5).
In order to reduce bias in finite samples, the independence copula Π in the definition of
¨
Φ
2
n
in Eq. (10) can be replaced by
its discrete counterpart
¸
d
i=1
U
n
(u
i
) where U
n
denotes the (univariate) distribution function of a randomvariable uniformly
distributed on the set
¸
1
n
, . . . ,
n
n
¸
. This has also been proposed by Genest et al. [20] in the context of tests for stochastic
independence (cf. Section 4.1). In order to ensure the normalization property of the estimator in small samples, we introduce
the normalization factor h(d, n) depending on both dimension d and sample size n. It is obtained by replacing the upper
Fréchet–Hoeffding bound M withits discrete counterpart M
n
(u) := min{U
n
(u
1
), . . . , U
n
(u
d
upper bound of the empirical copula for given sample size n. A small sample estimator for Φ
2
thus has the form
˜
Φ
2
n
= h(d, n)

[0,1]
d
¸
¨
C
n
(u) −
d
¸
i=1
U
n
(u
i
)
¸
2
du (16)
with
h(d, n)
−1
=

[0,1]
d
¸
M
n
(u) −
d
¸
i=1
U
n
(u
i
)
¸
2
du.
We obtain
˜
Φ
2
n
= h(d, n)
¸

1
n

2 n
¸
j=1
n
¸
k=1
d
¸
i=1
(1 − max{
¨
U
ij
,
¨
U
ik
})

2
n

1
2

d n
¸
j=1
d
¸
i=1
¸
1 −
¨
U
2
ij

1 −
¨
U
ij
n
¸
+

1
3

d
¸
(n − 1)(2n − 1)
2n
2
¸
d
¸
,
with
h(d, n)
−1
=

1
n

2 n
¸
j=1
n
¸
k=1

1 − max
¸
j
n
,
k
n
¸
d

2
n
n
¸
j=1
¸
n(n − 1) − j(j − 1)
2n
2
¸
d
+

1
3

d
¸
(n − 1)(2n − 1)
2n
2
¸
d
.
Note that
˜
Φ
2
n
and
¨
Φ
2
n
have the same asymptotic distribution, i.e. under the assumptions of Theorem 4

n(
˜
Φ
2
n
− Φ
2
)
d
−→ Z
Φ
2 ∼ N(0, σ
2
Φ
2
).
This can be shown analogously to the proof of Theorem 4 using the fact that lim
n→∞

n{h(d, n) − h(d)} = 0. Accordingly
it follows that the bootstrap to estimate the asymptotic variance of

n(
˜
Φ
2
n
− Φ
2
) works (cf. Proposition 5).
A simulation study is carried out in order to investigate the finite-sample performance of the block bootstrap estimator
for the asymptotic standard deviation σ
Φ
2 of

n(
˜
Φ
2
n
− Φ
2
).
For comparison, we additionally provide the corresponding simulation results when using a nonparametric jackknife
method to estimate the unknown standard deviation. For dependent observations, Künsch [29] introduces the delete-l
jackknife which, in contrast to the block bootstrap described in the previous section, is based on systematically deleting
one block B
s,l
of l consecutive observations each time from the original sample, s = 0, . . . , n − l. Let
˜
Φ
2,(s)
n
denote the
estimator of Hoeffding’s Phi-Square calculated from the original sample when we have deleted block B
s,l
, s = 0, . . . , n − l
and define
˜
Φ
2,(.)
n
= (n − l + 1)
−1
¸
n−l
s=0
˜
Φ
2,(s)
n
. The jackknife estimate of the standard deviation is then given by
ˆ σ
J
=

(n − l)
2
nl(n − l + 1)
n−l
¸
s=0

˜
Φ
2,(s)
n

˜
Φ
2,(.)
n

2
.
2580 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
Table 1
Gaussian copula (independent observations). Simulation results for estimating the asymptotic standard deviation of

n(
˜
Φ
2
n
− Φ
2
) by means of the
nonparametric block bootstrap with block length l and the delete-l jackknife (for l = 5): The table shows the empirical means m(·) and the empirical
standard deviations s(·) of the respective estimates, which are calculated based on 1000 Monte Carlo simulations of sample size n of a d-dimensional equi-
correlated Gaussian copula with parameter ρ and 250 bootstrap samples. The bootstrap estimates are labeled by the superscript B, jackknife estimates by
J .
ρ n Φ
2
m(
˜
Φ
2
n
) s(
˜
Φ
2
n
) m( ˆ σ
B
) m( ˆ σ
J
) s( ˆ σ
B
) s( ˆ σ
J
)
Dimension d = 2
0.2 50 0.032 0.077 0.047 0.055 0.049 0.019 0.026
100 0.032 0.054 0.034 0.035 0.033 0.012 0.015
500 0.032 0.035 0.015 0.015 0.015 0.003 0.003
0.5 50 0.197 0.231 0.095 0.089 0.094 0.017 0.022
100 0.197 0.218 0.070 0.067 0.069 0.010 0.011
500 0.197 0.202 0.032 0.031 0.032 0.003 0.002
−0.1 50 0.008 0.056 0.035 0.047 0.037 0.015 0.021
100 0.008 0.032 0.021 0.026 0.022 0.010 0.012
500 0.008 0.013 0.008 0.008 0.008 0.003 0.003
Dimension d = 5
0.2 50 0.028 0.044 0.023 0.026 0.022 0.010 0.010
100 0.028 0.036 0.016 0.017 0.015 0.005 0.005
500 0.028 0.030 0.007 0.007 0.007 0.001 0.001
0.5 50 0.191 0.208 0.065 0.063 0.062 0.013 0.014
100 0.191 0.202 0.048 0.045 0.046 0.007 0.008
500 0.191 0.196 0.022 0.021 0.021 0.002 0.002
−0.1 50 0.007 0.015 0.004 0.005 0.004 0.001 0.002
100 0.007 0.011 0.004 0.003 0.003 0.001 0.001
500 0.007 0.007 0.002 0.002 0.002 0.000 0.000
Dimension d = 10
0.2 50 0.007 0.014 0.009 0.012 0.008 0.007 0.005
100 0.007 0.011 0.005 0.007 0.005 0.003 0.003
500 0.007 0.008 0.002 0.002 0.002 0.001 0.001
0.5 50 0.098 0.111 0.046 0.049 0.043 0.017 0.016
100 0.098 0.107 0.033 0.035 0.033 0.009 0.009
500 0.098 0.100 0.015 0.015 0.015 0.002 0.002
−0.1 50 0.001 0.001 0.000 0.000 0.000 0.000 0.000
100 0.001 0.001 0.000 0.000 0.000 0.000 0.000
500 0.001 0.001 0.000 0.000 0.000 0.000 0.000
We consider observations froman AR(1)-process with autoregressive coefficient β (cf. Table 2) based on the equi-correlated
Gaussian copula as defined in (8). To generate these observations, we proceed as follows: Simulate n independent d-
dimensional random variates U
j
= (U
j1
, . . . , U
jd
), j = 1, . . . , n, from the equi-correlated Gaussian copula with parameter
ρ. Set ε
j
= (Φ
−1
(U
j1
), . . . , Φ
−1
(U
jd
)), j = 1, . . . , n. A sample (X
j
)
j=1,...,n
of the AR(1)-process is then obtained by setting
X
1
= ε
1
and completing the recursion X
j
= βX
j−1
+ ε
j
, j = 2, . . . , n. Additionally, we consider the case of independent
observations from the equi-correlated Gaussian copula (cf. Table 1). To ease comparison, the block bootstrap is used in this
case, too.
Tables 1 and 2 outline the simulation results for dimensions d = 2, 5, and 10, sample sizes n = 50, 100, and 500,
and different choices of the copula parameter ρ. Their calculation is based on 1000 Monte Carlo simulations of size n and
250 bootstrap replications, respectively. For simplicity, we set the block length l = 5 in all simulations. The autoregressive
coefficient β of the AR(1)-process equals 0.5. The third column of Tables 1 and 2 shows an approximation to the true value of
Φ
2
, which is calculated froma sample of size 100,000. Comparing the latter to m(
˜
Φ
2
n
) (column 4), we observe a finite-sample
bias which strongly depends on the dimension d and the parameter choices, and which decreases with increasing sample
size. The standard deviation estimations s(
˜
Φ
2
n
) and the empirical means of the block bootstrap estimations, m( ˆ σ
B
), as well
as the delete-l jackknife estimations, m( ˆ σ
J
), for the standard deviation are given in columns 5, 6, and 7. There is a good
agreement between their values, especially for the sample sizes n = 100 and n = 500, implying that the bootstrap and the
jackknife procedure to estimate the asymptotic standarddeviationof

n(
˜
Φ
2
n
−Φ
2
) performwell for the consideredGaussian
copula models. Further, the standard error s of the bootstrap standard deviation estimations is quite small (column 8) and
slightly smaller than the obtained jackknife estimates (column 9) in lower dimensions. For large sample size n, however,
the jackknife is of higher computational complexity.
5. Empirical study
For illustration purposes, we apply the theoretical results of Section 4 to empirical financial data. We analyze the
association between the daily (log-)returns of major S&P global sector indices before and after the bankruptcy of Lehman
S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2581
Table 2
Gaussian copula (dependent AR(1) observations). Simulation results for estimating the asymptotic standard deviation of

n(
˜
Φ
2
n
− Φ
2
) by means of the
nonparametric block bootstrap with block length l and the delete-l jackknife (for l = 5): The table shows the empirical means m(·) and the empirical
standard deviations s(·) of the respective estimates, which are calculated based on 1000 Monte Carlo simulations of sample size n of a d-dimensional
equi-correlated Gaussian copula with parameter ρ, AR(1)-processes with standard normal residuals in each margin (with coefficient β = 0.5 for the first
lag) and 250 bootstrap samples. The bootstrap estimates are labeled by the superscript B, jackknife estimates by J .
ρ n Φ
2
m(
˜
Φ
2
n
) s(
˜
Φ
2
n
) m( ˆ σ
B
) m( ˆ σ
J
) s( ˆ σ
B
) s( ˆ σ
J
)
Dimension d = 2
0.2 50 0.032 0.086 0.059 0.065 0.060 0.022 0.033
100 0.032 0.059 0.043 0.042 0.040 0.016 0.021
500 0.032 0.036 0.018 0.017 0.017 0.005 0.005
0.5 50 0.200 0.242 0.114 0.101 0.110 0.022 0.031
100 0.200 0.222 0.086 0.076 0.081 0.013 0.015
500 0.200 0.203 0.039 0.037 0.037 0.003 0.003
−0.1 50 0.008 0.067 0.045 0.058 0.050 0.019 0.029
100 0.008 0.037 0.025 0.032 0.028 0.012 0.016
500 0.008 0.014 0.010 0.010 0.009 0.004 0.005
Dimension d = 5
0.2 50 0.028 0.047 0.031 0.030 0.026 0.014 0.016
100 0.028 0.039 0.020 0.020 0.019 0.008 0.008
500 0.028 0.031 0.009 0.008 0.008 0.002 0.002
0.5 50 0.192 0.212 0.082 0.072 0.074 0.019 0.023
100 0.192 0.205 0.057 0.053 0.054 0.010 0.011
500 0.192 0.196 0.026 0.025 0.025 0.003 0.002
−0.1 50 0.007 0.015 0.005 0.006 0.005 0.002 0.002
100 0.007 0.011 0.004 0.004 0.004 0.001 0.001
500 0.007 0.008 0.002 0.002 0.002 0.000 0.000
Dimension d = 10
0.2 50 0.007 0.015 0.011 0.013 0.009 0.009 0.008
100 0.007 0.012 0.007 0.008 0.006 0.005 0.004
500 0.007 0.008 0.002 0.003 0.002 0.001 0.001
0.5 50 0.099 0.111 0.055 0.053 0.049 0.020 0.023
100 0.099 0.109 0.042 0.039 0.037 0.014 0.014
500 0.099 0.100 0.018 0.017 0.017 0.003 0.003
−0.1 50 0.001 0.001 0.000 0.000 0.000 0.000 0.000
100 0.001 0.001 0.000 0.000 0.000 0.000 0.000
500 0.001 0.001 0.000 0.000 0.000 0.000 0.000
Fig. 2. Evolution of the S&P global sector indices Financials, Energy, Industrials, and IT with respect to their value on January 1, 2008 (left panel).
Multivariate Hoeffding’s Phi-Square
˜
Φ
2
of the four indices’ returns series, where the estimation is based on a moving window approach with window
size 50 (right panel). The vertical line indicates the 15th of September 2008, the day of the bankruptcy of Lehman Brothers Inc.
Brothers Inc. using multivariate Hoeffding’s Phi-Square. The major S&P global sector indices considered are Financials,
Energy, Industrials, and IT during the period from 1st January 2008 to 8th April 2009 (331 observations).
Fig. 2 (left panel) shows the evolution of the four indices over the considered time horizon. To ease comparison, all series
are plotted with respect to their value on January 1, 2008. The vertical line indicates the 15th of September 2008, the day
of the bankruptcy of Lehman Brothers Inc. All series decrease in mid 2008 in the course of deteriorating financial markets;
they decline especially sharply after the bankruptcy of Lehman Brothers Inc. Table 3 reports the first four moments of the
2582 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
Table 3
First four moments (in %) and results of the Jarque–Bera (JB) test, calculated for the returns of the S&P indices, as well as results of the Ljung–Box (LB)
Q-statistics, calculated up to lag twenty from the squared returns.
Financials Energy Industrials IT
Mean −0.3140 −0.1732 −0.2213 −0.1637
Standard deviation 3.1696 3.0553 2.1199 2.1037
Skewness 0.1490 −0.2462 −0.1139 0.1148
Kurtosis 5.2523 6.6540 4.8467 5.2024
JB statistics 71.1875 187.4846 47.7495 67.6229
JB p-values 0.0000 0.0000 0.0001 0.0000
LB Q-statistics 141.8456 447.6130 386.8856 266.6008
LB p-values 0.0000 0.0000 0.0000 0.0000
daily returns of the four indices as well as the related results of the Jarque–Bera (JB) test, calculated over the entire time
horizon. In addition, the last two rows of the table display the results of the Ljung–Box (LB) Q-statistics, computed from the
squared returns of the indices up to lag twenty. All return series show skewness and excess kurtosis. The Jarque–Bera (JB)
test rejects the null hypothesis of normality at all standard levels of significance. Further, all squared returns showsignificant
serial correlation as indicated by the Ljung–Box (LB) test, which rejects the null hypothesis of no serial correlation. We fit
an ARMA–GARCH model to each return series (which cannot be rejected by common goodness-of-fit tests). The estimated
parameters are consistent with the assumption of strong mixing series (cf. [6]).
Fig. 2 (right panel) displays the evolution of multivariate Hoeffding’s Phi-Square of the indices’ returns, estimated on the
basis of a moving windowapproach with windowsize 50. Again, the vertical line indicates the day of Lehman’s bankruptcy.
We observe a sharp increase of Hoeffding’s Phi-Square after this date and, hence, an increase of the association between the
indices’ returns. In order to verify whether this increase is statistically significant, we compare Hoeffding’s Phi-Square over
the two distinct time periods before and after this date using the results on the statistical properties of the difference of two
Hoeffding’s Phi-Squares discussed in Section 4.1. Note that the test by Genest and Rémillard [21] (cf. Section 4.1) rejects the
null hypothesis of stochastic independence (i.e., C = Π) with a p-value of 0.0005 such that Theorem 6 can be applied. We
calculate the estimated values (based on 250 bootstrap samples with block length l = 5) of Hoeffding’s Phi-Square and the
asymptotic variances and covariance as stated in Theorem 6 for both time periods which comprise n = 100 observations
each:
˜
Φ
2
before
= 0.1982,
˜
Φ
2
after
= 0.7437,
( ˜ σ
B
before
)
2
= 0.1663,
( ˜ σ
B
after
)
2
= 0.2064,
˜ σ
B
before,after
= −0.0287.
The choice of the block length l = 5 is motivated by the results of the simulation study in Section 4.2. The corresponding
test statistic has the form

n
(
˜
Φ
2
after

˜
Φ
2
before
)

( ˜ σ
B
before
)
2
+( ˜ σ
B
after
)
2
− 2 ˜ σ
B
before,after

= 8.3190,
with corresponding p-value 0.0000. Hence, we conclude that there has been a significant increase in association between
the returns of the four indices after the bankruptcy of Lehman Brothers Inc.
6. Conclusion
A multivariate version is proposed for Hoeffding’s bivariate measure of association Phi-Square. A nonparametric
estimator for the proposed measure is obtained on the basis of the empirical copula process. Its asymptotic distribution is
established for the cases of independent observations as well as of dependent observations from a strictly stationary strong
mixing sequence. The asymptotic distribution can be approximated by nonparametric bootstrap methods. This allows for
the calculation of asymptotic confidence intervals for Hoeffding’s Phi-Square and the construction of hypothesis tests. The
results are derived under the general assumption of continuous marginal distributions.
Acknowledgments
We thank the associate editor and three anonymous referees for their comments which led to an improvement
of the paper. Moreover, we are grateful to the Regional Computing Center at the University of Cologne for providing
the computational resources required. Funding in support of the second author’s work was provided by the Deutsche
Forschungsgemeinschaft (DFG).
Appendix A. Derivation of the functions h(d)
−1
and g(d)
−1
We calculate the explicit form of the functions h(d)
−1
and g(d)
−1
, as stated in Eqs. (3) and (4). Regarding the function
h(d)
−1
, we have
S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2583
h(d)
−1
=

[0,1]
d
{M(u) − Π(u)}
2
du
=

[0,1]
d
{M(u)}
2
du − 2

[0,1]
d
M(u)Π(u)du +

[0,1]
d
{Π(u)}
2
du.
The first summand on the left-hand side of the above equation can be written as

[0,1]
d
{M(u)}
2
du = E

[min{U
1
, . . . , U
d
}]
2

= E

X
2

where U
1
, . . . , U
d
are i.i.d. from U ∼ R[0, 1] and X = min{U
1
, . . . , U
d
}. Therefore,
E

X
2

= d

1
0
x
2
(1 − x)
d−1
dx =
2
(d + 1)(d + 2)
. (A.1)
For the second summand, we obtain

[0,1]
d
M(u)Π(u)du =
1
2
d

[0,1]
d
min{u
1
, . . . , u
d
}
d
¸
i=1
2u
i
du =
1
2
d
E (min{V
1
, . . . , V
d
}) =
1
2
d
E(Y)
where V
1
, V
2
, . . . , V
d
are i.i.d. from V, which has density f
V
(v) = 2v for 0 v 1 and Y = min {V
1
, V
2
, . . . , V
d
}. Thus,
1
2
d
E(Y) =
1
2
d

1
0
xd(1 − x
2
)
d−1
2xdx =
1
2
d

1
0
(1 − x
2
)
d
dx =
1
2
d
1
2
Γ (d + 1)

π
Γ

d + 1 +
1
2

=
1
2
d
d!
d
¸
i=0

i +
1
2

. (A.2)
Combining Eqs. (A.1) and (A.2) and using that

[0,1]
d
{Π(u)}
2
du =

1
3

d
yields the asserted form of h(d)
−1
.
Regarding the function g(d)
−1
as defined in Eq. (4), we have
g(d)
−1
=

[0,1]
d
{W(u) − Π(u)}
2
du
=

[0,1]
d
{W(u)}
2
du − 2

[0,1]
d
W(u)Π(u)du +

[0,1]
d
{Π(u)}
2
du.
For the first summand, it follows that

[0,1]
d
{W(u)}
2
du =

1
0
· · ·

1
d−2−
d−2
¸
i=1
u
i

1
d−1−
d−1
¸
i=1
u
i

d
¸
i=1
u
i
− d + 1

2
du
d
· · · du
2
du
1
=
2
(d + 2)!
. (A.3)
Partial integration of the second term further yields

[0,1]
d
W(u)Π(u)du =

1
0
· · ·

1
d−2−
d−2
¸
i=1
u
i
u
d−1

1
d−1−
d−1
¸
i=1
u
i
u
d

d
¸
i=1
u
i
− d − 1

du
d
· · · du
2
du
1
=
d
¸
i=1

d
i

(−1)
i
1
(d + 1 + i)!
. (A.4)
Again, by combining Eqs. (A.3) and (A.4) and using that

[0,1]
d
{Π(u)}
2
du =

1
3

d
, we obtain the asserted form of g(d)
−1
.
Appendix B. Derivation of the estimator

Φ
2
n
We outline the derivation of the estimator
¨
Φ
2
n
as given in (10).
{h(d)}
−1
¨
Φ
2
n
=

[0,1]
d
¸
¨
C
n
(u) −
d
¸
i=1
u
i
¸
2
du =

[0,1]
d
¸
1
n
n
¸
j=1

d
¸
i=1
1
{
¨
U
ij
≤u
i
}

d
¸
i=1
u
i
¸
2
du
=

1
n

2 n
¸
j=1
n
¸
k=1

[0,1]
d

d
¸
i=1
1
{
¨
U
ij
≤u
i
}

d
¸
i=1
u
i

d
¸
i=1
1
{
¨
U
ik
≤u
i
}

d
¸
i=1
u
i

du
2584 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
=

1
n

2 n
¸
j=1
n
¸
k=1

[0,1]
d

d
¸
i=1
1
{max{
¨
U
ij
,
¨
U
ik
}≤u
i
}
+
d
¸
i=1
u
2
i

d
¸
i=1
u
i
1
{
¨
U
ij
≤u
i
}

d
¸
i=1
u
i
1
{
¨
U
ik
≤u
i
}

du
=

1
n

2 n
¸
j=1
n
¸
k=1
d
¸
i=1
(1 − max{
¨
U
ij
,
¨
U
ik
}) −
2
n

1
2

d n
¸
j=1
d
¸
i=1
(1 −
¨
U
2
ij
) +

1
3

d
.
Appendix C. Proofs
Proof of Proposition 1. Consider the d-dimensional random vector X with copula C.
(i) For all d ≥ 2, the copula C is invariant under strictly increasing transformations of one or several components of X
(see, e.g., [13, Theorem 2.6]). As direct functional of the copula, Φ
2
X
thus inherits this property.
(ii) Let α
k
be a strictly decreasing transformation of the kth component X
k
of X, k ∈ {1, . . . , d}, defined on the range of
X
k
. For dimension d = 2, we have
h(2)
−1
Φ
2

1
(X
1
),α
2
(X
2
))
=

[0,1]
2
¸
C

1
(X
1
),α
2
(X
2
))
(u
1
, u
2
) − Π(u
1
, u
2
)
¸
2
du
1
du
2
=

[0,1]
2
¸
C
(X
1
,X
2
)
(1, u
2
) − C
(X
1

2
(X
2
))
(1 − u
1
, u
2
) − u
1
u
2
¸
2
du
1
du
2
=

[0,1]
2
¸
u
1
+ u
2
− 1 − C
(X
1
,X
2
)
(1 − u
1
, 1 − u
2
) − u
1
u
2
¸
2
du
1
du
2
=

[0,1]
2
¸
1 − x + 1 − y − 1 + C
(X
1
,X
2
)
(x, y) −(1 − x)(1 − y)
¸
2
dxdy
=

[0,1]
2
¸
C
(X
1
,X
2
)
(x, y) − xy
¸
2
dxdy = h(2)
−1
Φ
2
X
,
see e.g. [13], Theorem 2.7, for the second equation. For dimension d ≥ 3, let α
k
(X) = (X
1
, . . . , X
k−1
, α
k
(X
k
), X
k+1
, . . . , X
d
)
denote the random vector where the kth component of X is transformed by the function α
k
. Without loss of generality set
k = 1. If a randomvector Xis jointly symmetric about a ∈ R
d
then its copula is invariant with respect to a strictly decreasing
transformation of any component. Hence, Φ
2
is invariant as well. For non-jointly symmetric random vectors, consider
h(d)
−1
Φ
2
α
1
(X)
=

[0,1]
d
¸
C
α
1
(X)
(u
1
, u
2
, . . . , u
d
) − Π(u
1
, u
2
, . . . , u
d
)
¸
2
du
=

[0,1]
d
{C
X
(1, u
2
, . . . , u
d
) − C
X
(1 − u
1
, u
2
, . . . , u
d
) − u
1
u
2
· · · · · u
d
}
2
du
=

[0,1]
d
[{C
X
(1, u
2
, . . . , u
d
) − u
2
· · · · · u
d
} − {C
X
(x, u
2
, . . . , u
d
) − xu
2
· · · · · u
d
}]
2
dxdu

where u

= (u
2
, . . . , u
d
)
=

[0,1]
d−1
{C
X
(1, u
2
, . . . , u
d
) − u
2
· · · · · u
d
}
2
du

+

[0,1]
d
{C
X
(x, u
2
, . . . , u
d
) − xu
2
· · · · · u
d
}
2
dxdu

[0,1]
d−1
{C
X
(1, u
2
, . . . , u
d
) − u
2
· · · · · u
d
}
¸
2

[0,1]
C
X
(x, u
2
, . . . , u
d
)dx − u
2
· · · · · u
d
¸
du

. (C.1)
According to Eq. (C.1), it holds that Φ
2
α
1
(X)
= Φ
2
X
if either
• C
X
(1, u
2
, . . . , u
d
) = u
2
· · · · · u
d
, meaning that X
2
, . . . , X
d
are independent or
• C
X
(1, u
2
, . . . , u
d
) = 2

[0,1]
C
X
(x, u
2
, . . . , u
d
)dx. This condition is fulfilled if X
1
is independent of (X
2
, . . . , X
d
) since
C
X
(x, u
2
, . . . , u
d
) = xC
X
(1, u
2
, . . . , u
d
) in this case.
Proof of Theorem 3. Note that Φ
2
= ϕ(C) represents a Hadamard-differentiable map ϕ on

([0, 1]
d
) of the copula C
(see [14,48] for the relevant definitions and background reading). Its derivative ϕ

C
at C ∈

([0, 1]
d
), a continuous linear
map on

([0, 1]
d
), is given by
ϕ

C
(D) = 2h(d)

[0,1]
d
{C(u) − Π(u)}D(u)du,
which can be shown as follows: For all converging sequences t
n
→ 0 and D
n
→ D such that C + t
n
D
n

([0, 1]
d
) for
every n, we have
S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2585
ϕ(C + t
n
D
n
) −ϕ(C)
t
n
=
h(d)

[0,1]
d
{C(u) − Π(u) + t
n
D
n
(u)}
2
du
t
n

h(d)

[0,1]
d
{C(u) − Π(u)}
2
du
t
n
=
2h(d)t
n

[0,1]
d
{C(u) − Π(u)}D
n
(u)du + t
2
n

[0,1]
d
D
2
n
(u)du
t
n
→ 2h(d)

[0,1]
d
{C(u) − Π(u)}D(u)du, (C.2)
for n → ∞, since the second integral in Eq. (C.2) is bounded for all D
n
. An application of the functional Delta-method given
in Theorem 3.9.4 in [48] together with Proposition 2 then implies

n(
¨
Φ
2
− Φ
2
) =

n{ϕ(
¨
C
n
) −ϕ(C)}
d
−→ ϕ

C
(G
C
), (C.3)
where ϕ

C
(G
C
) = 2h(d)

[0,1]
d
{C(u) − Π(u)}G
C
(u)du. Using the fact that G
C
(u) is a tight Gaussian process, Lemma 3.9.8
in [48], p. 377, implies that Z
Φ
2 = ϕ

C
(G
C
) is normally distributed with mean zero and variance σ
2
Φ
2
as stated in the theorem.
Another application of the Delta-method to (C.3) yields the weak convergence of

n{Φ(
¨
C
n
) −Φ(C)} to the randomvariable
Z
Φ
∼ N(0, σ
2
Φ
).
Proof of Theorem 4. Let
¨
C
n
denote the empirical copula based on the sample X
1
, . . . , X
n
. Given the weak convergence of
the empirical copula

n(
¨
C
n
−C) process to the Gaussianprocess G

, the asymptotic behavior of

n(
¨
Φ
2
n
−Φ
2
) as stated inthe
theorem follows by mimicking the proof of Theorem 3. Weak convergence of

n(
¨
C
n
−C) can be established analogously as
in [14] (see also [7]) and we outline the single steps in the following. First, by considering the transformed randomvariables
U
ij
= F
i
(X
ij
), i = 1, . . . , d, j = 1, . . . , n, it is possible to confine the analysis to the case where the marginal distributions
F
i
of F, i = 1, . . . , d, are uniform distributions on [0, 1] and thus, F has compact support [0, 1]
d
. Second, the functional
Delta-method is applied which is based on the following representation of the copula C as a map of its distribution function
F (cf. [48]):
C(u) = φ(F)(u) = F(F
−1
1
(u
1
), . . . , F
−1
d
(u
d
)), u ∈ [0, 1]
d
with map φ : D([0, 1]
d
) →

([0, 1]
d
). Here, the space D([0, 1]
d
) comprises all real-valued cadlag functions and C([0, 1]
d
)
the space of all continuous real-valued functions defined on [0, 1]
d
, both equipped with the uniform metric m. If all partial
derivatives of the copula exist and are continuous, the map φ is Hadamard-differentiable at F as a map from D([0, 1]
d
)
(tangentially to C([0, 1]
d
), cf. Lemma 2 in [14]) with derivative φ

F
. Further, Rio [36] shows that under the above assumptions
on the mixing coefficient α(r), it holds that

n{
¨
F
n
(u) − F(u)}
w
−→ B

(u)
in

([0, 1]
d
) with Gaussian process B

as defined in the theorem. Hence, an application of the functional Delta-method
yields the weak convergence of

n{φ(
¨
F
n
) −φ(F)} to the process φ

F
(B

) = G

. Since
sup
u∈[0,1]
d
|φ(
¨
F
n
)(u) −
¨
C
n
(u)| = O

1
n

cf. [14], the assertion follows by an application of Slutsky’s theorem.
Proof of Theorem 6. Let
¨
C
Z,n
denote the empirical copula based on the sample Z
1
, . . . , Z
n
. Under the assumption of the
theorem, weak convergence of the empirical copula process

n{
¨
C
Z,n
−C
Z
} to the Gaussian process G

Z
in

([0, 1]
2d
) follows
(cf. proof of Theorem 4). Since

Φ
2
X
Φ
2
Y

=

Φ
2
{C
Z
(u, 1, . . . , 1)}
Φ
2
{C
Z
(1, . . . , 1, v)}

= g(C
Z
),
the asymptotic behavior of (
¨
Φ
2
X
,
¨
Φ
2
Y
)

can be established analogously as in the proof of Theorem 3 using the Hadamard
differentiability of the map g at C
Z
whose derivative is denoted by g

C
Z
. Hence,

n{(
¨
Φ
2
X
,
¨
Φ
2
Y
)

− (Φ
2
X
, Φ
2
Y
)

} converges in
distribution to the multivariate normally distributed random vector g

C
Z
(G

Z
) given by
g

C
Z
(G

Z
) =

¸
¸
¸

[0,1]
d
{C
X
(u) − Π(u)}G

Z
(u, 1 . . . , 1)du

[0,1]
d
{C
Y
(v) − Π(v)}G

Z
(1, . . . , 1, v)dv

.
With G

X
(u) = G

Z
(u, 1, . . . , 1) and G

Y
(v) = G

Z
(1, . . . , 1, v), apply the continuous mapping theorem to conclude the
proof.
2586 S. Gaißer et al. / Journal of Multivariate Analysis 101 (2010) 2571–2586
References
[1] M.A. Arcones, B. Yu, Central limit theorems for empirical and U-processes of stationary mixing sequences, J. Theoret. Probab. 7 (1) (1994) 47–71.
[2] J. Behboodian, A Dolati, M. Úbeda-Flores, A multivariate version of Gini’s rank association coefficient, Statist. Papers 48 (2) (2007) 295–304.
[3] J.R. Blum, J. Kiefer, M. Rosenblatt, Distribution free tests of independence based on the sample distribution function, Ann. Math. Statist. 32 (2) (1961)
485–498.
[4] P. Bühlmann, The blockwise bootstrap in time series and empirical processes, Dissertation ETH Zürich, Diss. ETH No. 10354, 1993.
[5] P. Bühlmann, H.R. Künsch, Block length selection in the bootstrap for time series, Comput. Statist. Data Anal. 31 (3) (1999) 295–310.
[6] M. Carrasco, X. Chen, Mixing and moment properties of various GARCH and stochastic volatility models, Econom. Theory 18 (1) (2002) 17–39.
[7] J. Dedecker, P. Doukhan, G. Lang, J.R. León R., S. Louhichi, C. Prieur, Weak Dependence: With Examples and Applications, Springer Verlag, 2007.
[8] P. Deheuvels, La fonction de dépendance empirique et ses propriétés. Un test non paramétrique d’indépendance, Acad. Roy. Belg. Bull. Cl. Sci. 65 (5)
(1979) 274–292.
[9] J. Dhaene, M. Denuit, M.J. Goovaerts, R. Kaas, D. Vyncke, The concept of comonotonicity in actuarial science and finance: theory, Insur. Math. Econ. 31
(1) (2002) 3–33.
[10] J. Dhaene, M. Denuit, M.J. Goovaerts, R. Kaas, D. Vyncke, The concept of comonotonicity in actuarial science and finance: applications, Insur. Math.
Econ. 31 (2) (2002) 133–161.
[11] A. Dolati, M. Úbeda-Flores, On measures of multivariate concordance, J. Probab. Stat. Sci. 4 (2) (2006) 147–163.
[12] P. Doukhan, J.-D. Fermanian, G. Lang, An empirical central limit theorem with applications to copulas under weak dependence, Stat. Inference Stoch.
Process. 12 (1) (2009) 65–87.
[13] P. Embrechts, F. Lindskog, A. McNeil, Modelling dependence with copulas and applications to risk management, in: S. Rachev (Ed.), Handbook of Heavy
Tailed Distributions in Finance, Elsevier, 2003, pp. 329–384.
[14] J.-D. Fermanian, D. Radulović, M. Wegkamp, Weak convergence of empirical copula processes, Bernoulli 10 (5) (2004) 847–860.
[15] J.-D. Fermanian, O. Scaillet, Nonparametric estimation of copulas for time series, J. Risk 5 (2003) 25–54.
[16] B.F. Fernández, J.M. González-Barrios, Multidimensional dependency measures, J. Multivariate Anal. 89 (2) (2004) 351–370.
[17] N.I. Fisher, P.K. Sen (Eds.), The Collected Works of Wassily Hoeffding, Springer Verlag, 1994.
[18] P. Gänssler, W. Stute, Seminar on Empirical Processes, in: DMV-Seminar, vol. 9, Birkhäuser Verlag, 1987.
[19] C. Genest, J. Nešlehová, N. Ben Ghorbal, Spearman’s footrule and Gini’s gamma: a review with complements, J. Nonparametr. Stat. 22 (2010) in press
(doi:10.1080/10485250903499667).
[20] C. Genest, J.-F. Quessy, B. Rémillard, Asymptotic local efficiency of Cramér-von Mises tests for multivariate independence, Ann. Statist. 35 (1) (2007)
166–191.
[21] C. Genest, B. Rémillard, Tests of independence and randomness based on the empirical copula process, TEST 13 (2) (2004) 335–369.
[22] W. Hoeffding, Masstabinvariante Korrelationstheorie, Schrift. Math. Seminars Inst. Angew. Math. Univ. Berlin 5 (3) (1940) 181–233.
[23] W. Hoeffding, Stochastische Abhängigkeit und funktionaler Zusammenhang, Skand. Aktuarver. Tidskr. 25 (1942) 200–227.
[24] W. Hoeffding, A non-parametric test of independence, Ann. Math. Statist. 19 (4) (1948) 546–557.
[25] H. Joe, Estimation of entropy and other functionals of a multivariate density, Ann. Inst. Statist. Math. 41 (4) (1989) 683–697.
[26] H. Joe, Relative entropy measures of multivariate dependence, J. Amer. Statist. Assoc. 84 (405) (1989) 157–164.
[27] H. Joe, Multivariate concordance, J. Multivariate Anal. 35 (1) (1990) 12–30.
[28] H. Joe, Multivariate Models and Dependence Concepts, Chapman & Hall, 1997.
[29] H.R. Künsch, The jackknife and the bootstrap for general stationary observations, Ann. Statist. 17 (3) (1989) 1217–1241.
[30] H.O. Lancaster, Correlation and complete dependence of random variables, Ann. Math. Statist. 34 (4) (1963) 1315–1321.
[31] A.J. Lee, U-Statistics: Theory and Practice, Marcel Dekker, 1990.
[32] P. Mikusiński, H. Sherwood, M.D. Taylor, Shuffles of min, Stochastica 13 (1) (1992) 61–74.
[33] R.B. Nelsen, Nonparametric measures of multivariate association, in: Distributions with Fixed Marginals and Related Topics, in: IMS Lecture Notes—
Monograph Series, vol. 28, 1996, pp. 223–232.
[34] R.B. Nelsen, An Introduction to Copulas, 2nd ed., Springer Verlag, 2006.
[35] D. Radulović, On the bootstrap and empirical processes for dependent sequences, in: H. Dehling, T. Mikosch, M. Sørensen (Eds.), Empirical Process
Techniques for Dependent Data, Birkhäuser, 2002, pp. 345–364.
[36] E. Rio, Théorie Asymptotique des Processus Aléatoires Faiblement Dépendants, Springer Verlag, 2000.
[37] L. Rüschendorf, Asymptotic distributions of multivariate rank order statistics, Ann. Statist. 4 (5) (1976) 912–923.
[38] F. Schmid, R. Schmidt, Bootstrapping Spearman’s multivariate Rho, in: A. Rizzi, M. Vichi (Eds.), COMPSTAT—Proceedings in Computational Statistics,
Physika-Verlag, 2006.
[39] F. Schmid, R. Schmidt, Multivariate extensions of Spearman’s Rho and related statistics, Statist. Probab. Lett. 77 (4) (2007) 407–416.
[40] F. Schmid, R. Schmidt, Multivariate conditional versions of Spearman’s Rho and related measures of tail dependence, J. Multivariate Anal. 98 (6) (2007)
1123–1140.
[41] F. Schmid, R. Schmidt, Nonparametric inference on multivariate versions of Blomqvist’s beta and related measures of tail dependence, Metrika 66 (3)
(2007) 323–354.
[42] F. Schmid, R. Schmidt, T. Blumentritt, S. Gaißer, M. Ruppert, Copula-based measures of multivariate association, in: P. Jaworski, F. Durante, W. Härdle,
T. Rychlik (Eds.), Copula Theory and its Applications—Proceedings of the Workshop Held in Warsaw, 25–26 September 2009, Springer Verlag, 2010,
pp. 209–235.
[43] B. Schweizer, E.F. Wolff, On nonparametric measures of dependence for random variables, Ann. Statist. 9 (4) (1981) 879–885.
[44] A. Sklar, Fonctions de répartition à n dimensions et leurs marges, Publ. Inst. Statist. Univ. Paris 8 (1959) 229–231.
[45] M.D. Taylor, Multivariate measures of concordance, Ann. Inst. Statist. Math. 59 (4) (2007) 789–806.
[46] H. Tsukahara, Semiparametric estimation in copula models, Canad. J. Statist. 33 (3) (2005) 357–375.
[47] M. Úbeda-Flores, Multivariate versions of Blomqvist’s beta and Spearman’s footrule, Ann. Inst. Statist. Math. 57 (4) (2005) 781–788.
[48] A.W. Van der Vaart, J.A. Wellner, Weak Convergence and Empirical Processes, Springer Verlag, 1996.
[49] E.F. Wolff, N-dimensional measures of dependence, Stochastica 4 (3) (1980) 175–188.
[50] T. Yanagimoto, On measures of association and a related problem, Ann. Inst. Statist. Math. 22 (1) (1970) 57–63.

. we propose a multivariate extension of the bivariate measure of association which was suggested by Wassily Hoeffding in his seminal doctoral dissertation (see [22. . The concept of comonotonicity is of interest. u ∈ [0. . we refer to [34. Multivariate Hoeffding’s Phi-Square (which we denote by Φ 2 ) is based on a Cramér–von Mises functional and can be interpreted as the (normalized) average squared difference between the copula of a random vector and the independence copula. . . Xd ). . 1]d . Further. The random vector X = (X1 . Radial symmetry of X about a ∈ Rd is given if. . 1]d ) is the collection of all uniformly bounded real-valued functions defined on [0. .. . [20]). αk (Xk ).e. It represents the joint distribution function of the random variables Ui = Fi (Xi ). ad . the space ∞ ([0.. Further. . the lower bound W is a copula only for dimension d = 2. . . The generalized inverse function G−1 is −1 defined by G (u) := inf{x ∈ R ∪ {∞}|G(x) ≥ u} for all u ∈ (0.. ud } =: M (u) for all u ∈ [0. two cases need to be distinguished: If Φ 2 > 0. . We show how the estimator can be adapted to account for small sample sizes. when the copula equals the upper Fréchet–Hoeffding bound. if there is (with probability one) a strictly increasing functional relationship between Xi and Xj (i = j). . k = 1. k = 1. . . Multivariate Hoeffding’s Φ 2 is based on a L2 -type distance between the copula C of a random vector and the independence copula and attains its maximal value in the case of comonotonicity. 1]d and random vector − − U = (U1 . .e..e. . . It offers a set of properties which are advantageous for many applications. . . . .1]d |f1 (t) − f2 (t)|. The random vector X is then said to be jointly symmetric about a ∈ Rd if. . d. see [34]. . Notation and definitions Let X = (X1 . Fk (ak + x) = Fk (ak − x) for all x ∈ R. . . F . weak convergence of nΦn to a non-normal distribution follows. . 1]d → [0. asymptotic 2 normality of n(Φn − Φ 2 ) can be shown using the functional Delta-method. Xd are symmetric about a1 . . . C (u) = F (F1 1 (u1 ). Xk−1 . Ud ≤ ud ) = P (U ≤ u) for all u = (u1 . the measure is zero if and only if the components of X are stochastically independent. . . . . X is marginally symmetric and its copula CX equals the copula Cαk (X) of αk (X). in actuarial science or finance (see [9. 1]d . 1 − ud ). . It represents the copula of the random vector −X = (−X1 . The measure can thus be used to quantify the degree of association of multivariate empirical data (e. . .. We establish its asymptotic behavior both in the case of independent observations and dependent observations from strictly √ stationary strong mixing sequences.10]). Gaißer et al. . . According to Sklar’s theorem [44]. i. and only if. d. i. Xd . . If not stated otherwise. Let αk be a strictly decreasing transformation of the kth component Xk of X. the independence copula Π with d Π (u) := i=1 ui . . −Xd ). . . . . . . . i = 1. . .7. Section 5 presents an empirical study of financial contagion related to the bankruptcy of Lehman Brothers Inc. . When Φ 2 = 0. . . ud ) = P (U1 ≤ u1 . . We calculate the explicit value of multivariate Hoeffding’s Φ 2 for some copulas of simple form and describe a simulation algorithm to approximate its value in cases where 2 the copula is of a more complicated form. ud ) ∈ [0. Multivariate Hoeffding’s Φ 2 is introduced in Section 3 and some of its analytical properties are investigated. It is equipped with the uniform metric m(f1 .28]. . P ). Ud ). Moreover. . Note that joint symmetry implies radial symmetry. Section 2 introduces relevant definitions and notation. By contrast. . Xk+1 . The asymptotic variance can consistently be 2 estimated by means of a nonparametric (moving block) bootstrap method. d. The upper bound M is a copula itself and is also known as the comonotonic copula. . .. .2572 S. . . Xd ) be a d-dimensional random vector (d ≥ 2). For example. xd ) ∈ Rd and marginal distribution functions ¯ Fi (x) = P (Xi ≤ x) whose survival functions are denoted by Fi (x) = P (Xi > x) for x ∈ R and i = 1. . / Journal of Multivariate Analysis 101 (2010) 2571–2586 In the present paper. financial asset returns). . and only if. . .e. For a detailed treatment of copulas. . . . with distribution function F (x) = P (X1 ≤ x1 . This leads to the construction of statistical tests for stochastic independence based on Hoeffding’s Phi-Square (cf. In general. describes the dependence structure of stochastically independent random variables X1 . we always assume that the Fi are continuous functions. 1]d . . It represents the copula of X1 . . 1] such that F (x) = C (F1 (x1 ). the random vector X is marginally symmetric and the copula ˘ ˘ CX of X equals its survival copula CX . . . . Xd if F1 (X1 ) = · · · = Fd (Xd ) with probability one. 1]d .g. . . .g. . 0} ≤ C (u) ≤ min{u1 . Fd (xd )) for all x ∈ Rd . . 1] and G−1 (0) := sup{x ∈ R ∪ {−∞}|G(x) = 0}. d. . 2. that is CX (u) = P (U ≤ u) = P (U > 1 − u) = CX (u) for all u ∈ [0. . ˘ ˘ The survival copula C of C is given by C (u) = P (U > 1 − u) where 1 − u = (1 − u1 . k = 1. 1]d . . a nonparametric estimator Φn for Φ 2 based on the empirical copula is derived. . . . multivariate Hoeffding’s Phi-Square can be estimated with low computational complexity. d). . Every copula C is bounded in the following sense: W (u) := max{u1 + · · · + ud − (d − 1).17]). where M and W are called the upper and lower Fréchet–Hoeffding bounds. f2 ) = supt∈[0. The copula C is also referred to as the copula of the random vector X (if an explicit reference to the random vector is required. . . . . defined on some probability space (Ω . . and denote the transformed random vector by αk (X) = (X1 . defined on the range of Xk . we denote its copula by CX ). A brief conclusion is given in Section 6. Chapter 2. The paper is organized as follows. respectively. . e. Xd ) is said to be marginally symmetric about ¯ a ∈ Rd if X1 . there exists a unique copula C : [0. . . i. Xd ≤ xd ) for x = (x1 . in September 2008 using multivariate Hoeffding’s Φ 2 . . . Moreover. even for large dimension d. In Section 4. C (u1 . respectively (i. and nonparametric statistical inference for the estimator can be established based on the empirical copula. Fd 1 (ud )) for all u ∈ [0. .

we have Φ 2 (C1 ) ≤ Φ 2 (C2 ). 1]d .1]d {M (u) − Π (u)}2 du]−1 .1]d {M (u) − Π (u)}2 du = 2 (d + 1)(d + 2) − 1 2d d d! i+ i=0 1 2 + 1 3 d . respectively. In particular. we have h(d)−1 = [0. (2) with normalization constant h(d) = [ [0.e. 2 such that the range of Φ as defined in (2) is restricted to the interval [0. This relationship is used in Section 4 where the statistical properties of Φ 2 are derived. a value of one of Hoeffding’s Phi-Square also implies that the random variables X1 and X2 are completely dependent. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2573 3. it holds that Φ = 1 iff C = M or C = W since g (2) = h(2). (3) and for C = W . 1]d . Monotonicity: For copulas C1 and C2 with Π (u) ≤ C1 (u) ≤ C2 (u) ≤ M (u) for all u ∈ [0.16]. For copulas C3 and C4 such that W (u) ≤ C3 (u) ≤ C4 (u) ≤ Π (u) for all u ∈ [0. u2 )}2 du1 du2 . Gaißer et al. (5) Note that for dimension d = 2.S. According to Nelsen [34]. However. (2). Remark. Hence.e. Hoeffding’s Phi-Square thus represents a measure for strictly monotone functional dependence. and Φ 2 = 1 iff C = M 2 for d ≥ 3. Normalization: An important property of Φ 2 is that Φ 2 = 0 if and only if C = Π . They can also be established for Φ . For the related multivariate case. For example. Hoeffding [22] suggests the following measure Φ 2 to quantify the amount of association between the components of the two-dimensional random vector X with copula C Φ 2 = 90 [0. A multivariate version of Hoeffding’s Phi-Square In his seminal dissertation. i. respectively. we define a multivariate version of Φ 2 by Φ 2 := Φ 2 (C ) = h(d) [0. In consequence. direct calculations yield that h(d)−1 ≥ g (d)−1 0 ≤ Φ2 ≤ 1 for all d ≥ 2.2. If C is the copula of the random vector X. the value of bivariate Hoeffding’s Phi-Square can be made arbitrarily small for completely dependent random variables. two random variables X1 and X2 are completely dependent if their copula is a shuffle of M.30]). . 1]d ) for which the integral on the right-hand side of Eq. i. Both expressions converge to zero as the dimension d tends to infinity. For C = M.1]d {W (u) − Π (u)}2 du = 2 i=0 (d + 2)! −2 d d i (−1)i 1 (d + 1 + i)! + 1 3 d . 1].1]2 {C (u1 . (1) For the d-dimensional random vector X with copula C . we calculate the value of the defining integral for the lower and upper Fréchet–Hoeffding bounds. Φ 2 can be regarded as a continuous functional on the subset of l∞ ([0. that there exists a one-to-one function ψ (which is not necessarily monotone) such that P (X2 = ψ(X1 )) = 1 (cf. In order to motivate the specific form of the normalization factor h(d) in Eq. An alternative multivariate measure of association can be defined by Φ := Φ (C ) = + Φ 2 (C ). This property generalizes a result of Yanagimoto [50] to the multivariate case. [23.. Theorem 3.1]d {C (u) − Π (u)}2 du. we find shuffles of M which (uniformly) approximate the independence copula arbitrarily closely (see also [32]). (4) The calculations are outlined in Appendix A. Invariance under strictly monotonic transformations: The behavior of Φ 2 with respect to strictly monotonic transformations is given in the next proposition. (2) is well-defined. we obtain g (d)−1 = [0. In the bivariate case. 2 we also refer to Φ 2 and Φ as ΦX and ΦX . In particular. the converse does not hold. it follows that Φ 2 (C3 ) ≥ Φ 2 (C4 ). Various analytical properties of Φ 2 are discussed next. we refer to [49. This measure can be interpreted as the normalized average distance between the copula C and the independence copula Π with respect to the L2 -norm. whose explicit form is calculated later. u2 ) − Π (u1 . Bivariate measures of association of this form based on the L1 . 2 2 Invariance with respect to permutations: For every permutation π of the components of X we have ΦX = Φπ (X) according to Fubini’s Theorem.2.and L∞ -norms are considered by Schweizer and Wolff [43]..

e. C = M. given by Cρ (u1 . Cα1 (X) (u) = M (u) for all u ∈ [0. . d = 5.1]d {C (u) − Π (u)}2 du = h(d)EΠ {C (U) − Π (U)}2 . . . 1]d . Xd are jointly symmetric about a ∈ Rd . Then. . ΦX = Φα (X) if one of the following three conditions holds: • X1 . 2 (ii) For dimension d = 2. implying that Φα (X) < ΦX . Then. . ud ) = d i=1 ui + θ d i =1 ui (1 − ui ). . αk (Xk ). This illustrates the restricted range of dependence modeled by the family of Farlie–Gumbel–Morgenstern copulas. . k 2 2 The proof is outlined in Appendix C. . . Xk+1 . Xk+1 . The values of Φ 2 form a parabolic curve. . The approximated values of Φ 2 and Φ for different choices of the parameter ρ and for dimensions d = 2. . Then. i. . . or • X1 . part (ii). . . . . − However. ud ) = Φρ Φ −1 (u1 ). . (ii) Let C (u) = θ M (u) + (1 − θ )Π (u) with 0 θ 1. Examples. . In this case. 1]d . . .e. when applying the (strictly decreasing) inverse function α1 1 of α1 to the first component of α1 (X). or • (X1 . the value of Φ 2 does not depend on the dimension d. i. . . 2 (i) For dimension d ≥ 2. . Hoeffding’s Phi-Square is given by 1 30 d Φ 2 = h(d)θ 2 . we obtain that 2 2 2 = ΦX > Φα (X) . Un being independent and identically distributed Monte Carlo replications from U. since 2 Φ is invariant under strictly decreasing transformations of all components of X according to Proposition 1. d ≥ 2. . / Journal of Multivariate Analysis 101 (2010) 2571–2586 Proposition 1. Xd ). If part (ii) of the above proposition is not satisfied. then Φ 2 (Cm ) → Φ 2 (C ) as a direct consequence of the dominated convergence theorem. and d = 10 are displayed in Fig. Xk+1 .. Ud ) is uniformly distributed on [0. an approximation of Φ 2 is obtained by estimating the expectation on the right-hand side of Eq. . . Hoeffding’s Phi-Square satisfies duality. . see [22] for a power series representation of Φ 2 for the bivariate Gaussian copula. . For d ≥ 3. Xd ) is stochastically independent of Xk . ΦX = Φ−X or equivalently Φ 2 (C ) = Φ 2 (C ). xd ) = ··· −∞ −∞ (2π )− 2 Σρ d −1 2 1 and Σρ = ρ 11 + (1 − ρ)I with − d−1 < ρ < 1. let αk be a strictly decreasing transformation of the kth component Xk of X. i = 1. Note that for this family of copulas. (6) consistently as follows: ˆ EΠ {C (U) − Π (U)}2 = 1 n n i =1 {C (Ui ) − Π (Ui )}2 . Gaißer et al. . d}. . ΦX is invariant with regard to strictly increasing transformations of one or several components of X. . Xk−1 . . Φ −1 (ud ) with x1 xd (8) 1 − exp − x Σρ 1 x dxd · · · dx1 2 Φρ (x1 . . Continuity: If {Cm }m∈N is a sequence of copulas such that Cm (u) → C (u) for all u ∈ [0. The following equivalent representation of Φ 2 is useful for this purpose: Φ 2 = h(d) [0. . It follows that Φ 2 ≤ 1/10 for d ≥ 2. Xk−1 . For example. . . Xk−1 . Φ −1 α1 (α1 (X)) 1 1 2 2 ˘ Duality: For dimension d = 2. For dimension d ≥ 3. . . Let X be a d-dimensional random vector with copula C . . 1. . . . . . . . . . . . the value of Φ 2 needs to be determined by simulation. . . d (which is indicated by the subscript Π ). d ≥ 2. equality of ΦX and Φα (X) does not k hold in general. . 1]d with stochastically independent components Ui . For illustration we compute the approximated value of Φ 2 for the equi-correlated Gaussian copula. Thus. . k ∈ {1. let the copula C of X be the comonotonic copula. (i) Let C be the d-dimensional Farlie–Gumbel–Morgenstern copula defined by C (u1 . |θ | ≤ 1. . and let α1 be a strictly decreasing 2 2 transformation of the first component X1 of X. (6) where the random vector U = (U1 . Xd are mutually stochastically independent. . . Then Φ2 = θ 2. . Taylor [45] discusses duality in the context of multivariate measures of concordance. duality does not hold in general except in the case that the random vector X is radially symmetric about a ∈ Rd . (7) with U1 .2574 S. It is difficult to derive an explicit expression for Φ 2 if C is of a more complicated structure than in examples (i) and (ii). and let αk (X) = 2 2 (X1 . . . ΦX is invariant under strictly decreasing transformations of one or both components of X.

Let X1 . . We assume that both F and C as well as the marginal distribution functions Fi . . an estimator for the alternative measure Φ is given by Φn = + Φn . After illustrating our approach on the basis of independent observations. 1] ) to the process GC which takes the form . and Fi. The 2 asymptotic distributions of Φn and Φn can be deduced from the asymptotic behavior of the empirical copula process which has been discussed.1. . d = 5 (dashed line). If the ith partial derivatives Di C (u) of C exist and are continuous for i = 1. . . e.n = 1 n (rank of Xij in Xi1 . n. are completely unknown. Uik }) − j=1 k=1 i=1 2 n 1 2 d n d 2 (1 − Uij ) + 1 3 d . Nonparametric estimation Let X1 .n = Fi. . . The following result is shown in [14]: Proposition 2.n (x) = Uij. Gaussian copula.. (10) j=1 i=1 2 The derivation is outlined in Appendix B. statistical inference is based on the ranks of the observations.i.e. 4.g. Gaißer et al. . the empirical process n(Cn − C ) ∞ d converges weakly in ([0.n (Xij ) for i = 1. 000 Monte Carlo replications. . A nonparametric estimator for Φ 2 is then obtained by replacing the copula C in formula (2) by the empirical copula Cn . For fixed n. i = 1. We derive a nonparametric estimator for multivariate Hoeffding’s Phi-Square and establish its asymptotic behavior based on the weak convergence of the empirical copula process. . . Xn be a random sample from the d-dimensional random vector X with joint distribution function F √ and copula C . Xn be an (i.d. i. Xin ). 1. As with pseudo-observations Uij.. and d = 10 (dotted line). . Obviously. by Rüschendorf [37]. . The estimator is based on a Cramér–von Mises statistic and can explicitly be determined by 2 Φn = h(d) 1 n 2 n n d (1 − max{Uij . Statistical inference for multivariate Hoeffding’s Phi-Square Statistical inference for multivariate Hoeffding’s Phi-Square as introduced in formula (2) is based on the empirical copula which was first discussed by Rüschendorf [37] and Deheuvels [8]. calculations are based on n = 100. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2575 Fig.S. . The copula C is estimated by the empirical copula Cn . . and Tsukahara [46]. we generalize it to the case of dependent observations from strictly stationary strong mixing sequences. . 1]d 1 n n j =1 (9) 1{Xij ≤x} for x ∈ R. . . d. .1]d Cn (u) − Π (u) 2 du. . . 2 Φn := Φ 2 (Cn ) = h(d) [0. which is defined as Cn (u) = 1 n d n j=1 i=1 1{Uij. . we suppress the subindex and refer to the pseudo-observations as Uij if it is clear from the context.) random sample from the d-dimensional random vector X with distribution function F and copula C . . . 4. . Approximated values of Φ 2 (left panel) and Φ (right panel) in the case of a d-dimensional equi-correlated Gaussian copula with parameter ρ for dimension d = 2 (solid line). . .n ≤ui } for u ∈ [0. Van der Vaart and Wellner [48]. Gänßler and Stute [18]. d and j = 1. . d.

σΦ ) and 2 σΦ = 2 σΦ 2 4Φ 2 = h(d) [0. . are replaced by 1. it follows that √ 2 n(Φn − Φ 2 ) −→ ZΦ 2 d (12) 2 where ZΦ 2 ∼ N (0. Regarding the multivariate case. (14) with E h(d) [0.1]d [0. The latter can be 2 determined by simulation.1]d {C (u) − Π (u)}E {GC (u)GC (v)}{C (v) − Π (v)}dudv [0. as n → ∞. . . Proposition 2 and an application of the continuous mapping theorem yield 2 nΦn −→ h(d) d [0. 2 The asymptotic distribution of Φn when C = Π is important for the construction of tests for stochastic independence between the components of a multivariate random vector based on Hoeffding’s Phi-Square. Un is given by the U-statistic Un (ψ) = n −1 2 1≤j<k≤n ψ(Uj . . (14). Consider the random variables Uij = Fi (Xij ). Gaißer et al.1]d {C (u) − Π (u)}E {GC (u)GC (v)}{C (v) − Π (v)}dudv. except the ith coordinate of u.1]d [0. .1]d h(d) i =1 1{xi ≤ui } − i=1 ui i=1 1{yi ≤ui } − i=1 ui dudC (x)dC (y). . such tests have been studied by Hoeffding [24] and Blum et al. The vector u(i) corresponds to the vector where all coordinates.1]d {GΠ (u)}2 du. n with Uj = (U1j . 1]d with covariance function E {BC (u)BC (v)} = C (u ∧ v) − C (u)C (v). Theorem 3. [3].1]d [0. (11) The process BC is a d-dimensional Brownian bridge on [0. . When the univariate marginal distribution functions Fi are known. Hoeffding’s Phi-Square can also be estimated using the theory of U-statistics. Under the assumptions of Proposition 2 and if C = Π . In our setting. The proof is given in Appendix C. d. an unbiased estimator of the latter based on the random sample U1 . asymptotic normality of Φn and Φn can be established. . Uk ) . . . if C = Π . Remark. we have √ n(Φn − Φ ) −→ ZΦ d 2 with ZΦ ∼ N (0. However. (13) Regarding the alternative measure Φ . approximate critical values for the test statistic {h(d)}−1 nΦn are also provided in [20].1]d {GΠ (u)}2 du = h(d) 1 2 d − 1 3 d − d 6 1 3 d−1 . we mention Genest and Rémillard [21] and Genest et al. . σΦ 2 ) with 2 σΦ 2 = {2h(d)}2 [0. j = 1. Since Φ 2 = h(d) [0. Note that the assumption C = Π guarantees that the limiting random variable is 2 non-degenerated as implied by the form of the variance σΦ 2 . Udj ) having distribution function C . . [20] who consider various combinations of Cramér–von Mises statistics with special regard to their asymptotic local efficiency. / Journal of Multivariate Analysis 101 (2010) 2571–2586 GC (u) = BC (u) − i=1 Di C (u)BC (u(i) ). . 2 Then. . i = 1. .1]d [0. In the bivariate setting. .1]d {C (u) − Π (u)}2 du .2576 d S.1]d {C (u) − Π (u)}2 du d d d d = [0. a hypothesis test for H0 : C = Π against H1 : C = Π is performed 2 by rejecting H0 if the value of nΦn exceeds the (1 − α)-quantile of the limiting distribution in Eq.

Even though the asymptotic variance has a closed-form expression. . In order to describe temporal dependence between the observations we use the concept of α -mixing or strong mixing. σΦ 2 ) with d 2 σΦ 2 = {2h(d)}2 [0. cf. y ∈ [0. (15) depends not only on the copula C as in the case of independent observations (cf.g. The fact that both estimators have the same asymptotic distribution in the case of known margins follows also from the relationship √ 2 n(Φn − Φ 2 ) = 1 n3/2 n ψ(Uj . Gaißer et al. . defined by d d d d ψ(x. . (13)). and copula C .. (15) The process G∗ has the same form as in Eq. . the (moving) block bootstrap. where Uj = (F1 (X1j ). e. .1]d [0.S.12]. continuous marginal distribution functions Fi .]d 1{xi ≤ui } − i=1 i=1 ui i=1 1{yi ≤ui } − i=1 ui du. 1]d having covariance function E {B∗ (u)B∗ (v)} = j∈Z E {1{U0 ≤u} − C (u)}{1{Uj ≤v} − C (v)} . also [15] for the asymptotic properties of the smoothed empirical copula process in this setting. j ≤ t } and F t = σ {Xj . F s+r ) where Ft = σ {Xj . d. The proof is outlined in Appendix C.B∈B sup |P (A ∩ B) − P (A)P (B)|.1. d. x. F . Results from the theory of U-statistics (see. . the estimation of Φ 2 by means of U-statistics is more involved in comparison to the above approach based on the empirical copula. . / Journal of Multivariate Analysis 101 (2010) 2571–2586 2577 with kernel ψ of degree 2. (11) with process BC being replaced by the centered Gaussian process B∗ in [0. The asymptotic √ 2 variance coincides with the asymptotic variance of n(Φn − Φ 2 ) for known marginal distribution functions (cf. The process {Xj }j∈Z is said to be strong mixing if αX (r ) → 0 for r → ∞.1]d [0. j ≤ t. (13)) but also on the joint distribution of U0 and Uj . .7. to [1. Now let {Xj = (X1j . and Xj . . y) = h(d) [0. . Let X1 . . Xn and denote by Φn the corresponding estimator for 2 Hoeffding’s Phi-Square calculated according to (10). 1]d . For dependent observations. Note that the asymptotic variance in Eq. a bootstrap method. Theorem 4. Eq. If the ith partial derivatives Di C (u) of C exist and are continuous for i = 1. . 2 Assume that our observations are realizations of the sample X1 . where ψ1 (x) = E {ψ(X. Suppose A and B are two σ -fields included in F and define α(A. . we have √ 2 2 n(Φn − Φ 2 ) → ZΦ 2 ∼ N (0. with distribution function F . j ∈ Z. j ≥ t. Theorem 4 can be transferred to sequences with temporal dependence structures other than strong mixing. has been . The mixing coefficient αX associated with the sequence {Xj }j∈Z is given by αX (r ) = sups≥0 α(Fs . j ∈ Z. we refer. i = 1. B ) = A∈A. Uj ) + j =1 √ n n−1 n · Un (ψ) − Φ 2 where the first term in the right equation converges to zero in probability for n → ∞. for example. The asymptotic behavior of Φn is given next. . Chapter 3 in [31]) and standard calculations yield that asymptotically normally distributed with mean zero and variance 2 σU = Var{ψ1 (X)} √ n{Un (ψ) − Φ 2 } is = 4{h(d)}2 [0. it cannot be calculated explicitly in most cases but has to be estimated adequately. respectively. P ). In the case of unknown marginal distribution functions. Xdj )}j∈Z be a strictly stationary sequence of d-dimensional random vectors.1]d E {C (u) − Π (u)}G∗ (u)G∗ (v){C (v) − Π (v)} dudv. Xn be observations from the strictly stationary strong mixing sequence {Xj }j∈Z with coefficient αX (r ) satisfying αX (r ) = O(r −a ) for some a > 1. Un (ψ) is degenerate when C = Π . . In particular. . Fd (Xdj )). . being defined on a probability space (Ω . . j ≥ t } denote the σ -fields generated by Xj . Y)|X = x} with independent random vectors X. .1]d {C (u) − Π (u)}{C (u ∧ v) − C (u)C (v)}{C (v) − Π (v)}dudv. and C = Π . . . Eq. . . The result is based on the weak convergence of the empirical copula process to the Gaussian process G∗ for strictly stationary strong mixing sequences. Y having distribution function C . .

. and CX . s = 0. . consider the sequence {Zj = (Xj . Gaißer et al. n − l}. . 1) = CX (u) and CZ (1. .n .B − Φn ) converge weakly to the same Gaussian limit in probability. It follows that the block √ bootstrap can be applied to estimate the asymptotic variance of n(Φ 2 − Φ 2 ). . . This can be shown analogously to the proof of Theorem 3. Theorem 4 together with Proposition 5 enables the calculation of an asymptotic (1 − α)-confidence interval for Hoeffding’s Phi-Square Φ 2 ∈ (0. . . the standard bootstrap. 1) can be constructed by rejecting the null hypothesis at the confidence level α if 2 ˆ2 √ (Φn − Φ0 ) α n > Φ −1 1 − .5]. . i.2. Yj )}j∈Z with strong mixing coefficient αZ satisfying αZ (r ) = O(r −a ) for some a > 1. The above results can be extended to statistically analyze the difference of two Hoeffding’s Phi-Squares. also proof of Theorem 4). . j strong mixing sequence {Xj }j∈Z of d-dimensional random vectors with distribution function F and copula C whose partial √ B derivatives exist and are continuous.l . . which draws with replacement n √ 2 single observations from the sample X1 . Theorem 6.. . Xs+l }. we define blocks of size l. respectively.9. . . . continuous marginal distribution functions. Since the two sequences do not have to be necessarily independent. provided that all partial derivatives of the copula exist and are continuous. . which is briefly described in the following. . . l(n) = o(n). the sequences n(Φn − Φ 2 ) and n(Φn . .. i = 1. . . .2578 S. .l . Proposition 5. .. . . . . are presented in Section 4. Given the sample X1 . . ˆB n 2 2 2 Further. . and copula CZ such that CZ (u. XB = XS1 +1 . obtained by the block bootstrap.. which are observations of a strictly stationary. The block bootstrap draws with replacement k blocks from the blocks Bs.B be the corresponding estimator for Hoeffding’s Phi-Square. 1 l XB 1 = XS2 +1 . Sk being independent and uniformly distributed random variables on {0. . the bootstrap sample thus comprises those observations from X1 . 1]d ) to the same Gaussian √ √ √ 2 2 2 limit as n(Cn − C ) in probability. 2d. and copulas CX and CY . we have . . . Let (XB )j=1. Zn = (Xn . . . n − l. B The sequence n(Cn − Cn ) converges weakly in probability to the same Gaussian limit as n(Cn − C ) if the (uniform) √ B √ empirical process n(Fn − Fn ) converges weakly in probability to the same Gaussian limit as n(Fn − F ). Xn . . . . which assesses the performance of the bootstrap variance estimator.i . and let Φn . .n be the block bootstrap sample from (Xj )j=1. Section 5). . Xn . .e. 1.e. . Xn . . Note that in the case of independent observations. . XB = XS1 +l . 1. . If C = Π . In a financial context. . . . .11 in [48] using the functional Delta-method (cf. The block bootstrap for empirical processes has been discussed in various settings and for different dependence structures. B B see [29. respectively. can be used to estimate the asymptotic variance of n(Φn − Φ 2 ). . l = l(n) with l(n) → ∞ as n → ∞.. v ∈ [0. Xn which are among the k blocks BS1 . . . . Denote by Cn and Fn the empirical copula and the empirical distribution function of the block bootstrap sample B B 2 X1 . of consecutive observations by Bs. an asymptotic hypothesis test for H0 : Φ 2 = Φ0 against H1 : Φ 2 = Φ0 with Φ0 ∈ (0. strong mixing sequence {Zj = (Xj . / Journal of Multivariate Analysis 101 (2010) 2571–2586 proposed by Künsch [29]. .. l+ n The block length l is a function of n. For a discussion regarding the choice of l(n). . 1/2 ε > 0. i. given by 2 Φn ± Φ −1 1 − α 2 √ σΦ 2 / n. 2d. . n − l where we assume that n = kl (otherwise the last block is shortened). The √ B following sufficient conditions for n(Fn − Fn ) to converge weakly (in the space D([0. With S1 . Y1 ). 1]d . Let Z1 = (X1 . Note that in the 2 case = 1. . Suppose ΦX and ΦY are multivariate Hoeffding’s Phi-Squares associated with the strictly stationary sequences {Xj }j∈Z and {Yj }j∈Z of d-dimensional random vectors with distribution functions FX and FY . BSk . this may be of interest for assessing whether Hoeffding’s Phi-Square of one portfolio of financial assets significantly 2 2 differs from that of another portfolio (cf. 1). .l . . 1]d )) in probability to the appropriate Gaussian process for strong mixing sequences are derived in [4]: √ √ ∞ r =0 (r + 1)16(d+1) αX (r ) < ∞ and block length l(n) = O(n1/2−ε ). . the copula corresponds to the upper Fréchet–Hoeffding bound M which does not possess continuous first partial derivatives. . . ..l = {Xs+1 . CY = Π .. . Suppose further that n(Cn − Cn ) converges weakly in ∞ ([0. for an overview see [35] and references therein. v) = CY (v) for all u. (σ ˆ 2 Φ0 B 2 Φn 2 ) denotes the consistent bootstrap variance estimator for σΦ 2 .. The results of a simulation study. . .. continuous marginal distribution functions FZ. If the ith partial derivatives of CZ exist and are continuous for i = 1. 2 σΦ 2 ˆB n Here. Yj )}j∈Z of 2d-dimensional random vectors with joint distribution function FZ . . Yn ) be observations of the strictly stationary. s = 0. XB = XSk +l . .

we additionally provide the corresponding simulation results when using a nonparametric jackknife method to estimate the unknown standard deviation. For comparison. n) We obtain −1 = [0.1]d Mn (u) − i =1 U n ( ui ) du. ˜2 Φn = h(d. we introduce the normalization factor h(d. Un (ud )}. .1]d [0. n) − h(d)} = 0. Analogously to the discussion prior to Theorem 6. This has also been proposed by Genest et al. where σ 2 = σ 2 2 + σ 2 2 − 2σΦ 2 . A small sample estimator for Φ 2 thus has the form d 2 Φn 2 ˜ = h(d.Φ 2 with ΦX ΦY X Y σΦ 2 . (15)).(s) Φn . 4.1).1]d Cn (u) − i =1 U n ( ui ) du (16) with d 2 h(d. Accordingly √ ˜2 it follows that the bootstrap to estimate the asymptotic variance of n(Φn − Φ 2 ) works (cf. σ 2 ) d as n → ∞. i.l .1]d Y Y E {CX (u) − Π (u)}G∗ (u)G∗ (v){CY (v) − Π (v)} dudv. s = 0. .) = (n − l + 1)−1 n−l n −l s =0 ˜ 2. .2. . . n)−1 = 1 n 2 n n 1 − max j=1 k=1 j n n . σΦ 2 ).(. the independence copula Π in the definition of Φn in Eq. d This can be shown analogously to the proof of Theorem 4 using the fact that limn→∞ n{h(d. Small sample adjustments 2 In order to reduce bias in finite samples.l of l consecutive observations each time from the original sample. .(s) ˜ 2. The processes G∗ and G∗ are Gaussian processes on [0. Section 5). Section 4. .S. k d − 2 n n(n − 1) − j(j − 1) 2n2 d n j =1 + 1 3 d (n − 1)(2n − 1) 2n2 d .Φ 2 (cf. . It is obtained by replacing the upper Fréchet–Hoeffding bound M with its discrete counterpart Mn (u) := min{Un (u1 ). n) depending on both dimension d and sample size n. n . . 2 ˜2 Note that Φn and Φn have the same asymptotic distribution. Künsch [29] introduces the delete-l jackknife which. s=0 . . . under the assumptions of Theorem 4 √ 2 ˜2 n(Φn − Φ 2 ) −→ ZΦ 2 ∼ N (0. an asymptotic confidence interval or a statistical hypothesis test for the difference of two Hoeffding’s Phi-Squares can be formulated (cf. s = 0. is based on systematically deleting ˜ 2. (10) can be replaced by d ΦX X X ΦY its discrete counterpart i=1 Un (ui ) where Un denotes the (univariate) distribution function of a random variable uniformly n distributed on the set 1 . A simulation study is carried out in order to investigate the finite-sample performance of the block bootstrap estimator √ ˜2 for the asymptotic standard deviation σΦ 2 of n(Φn − Φ 2 ). / Journal of Multivariate Analysis 101 (2010) 2571–2586 2579 √ 2 2 ˆ2 ˆ2 n ΦX − ΦX − (ΦY − ΦY ) −→ W ∼ N (0. The proof is given in Appendix C. Proposition 5). .(. n − l √ ˜ and define Φn σJ = ˆ 2. in contrast to the block bootstrap described in the previous section. Eq.(s) denote the one block Bs. n) 2 n 1 2 1 n d n 2 n n d (1 − max{Uij .e. X Y and σ 2 2 = σΦ 2 . The jackknife estimate of the standard deviation is then given by 2 (n − l)2 nl(n − l + 1) ˜ 2. n − l. which represents an adequate upper bound of the empirical copula for given sample size n. . 1]d with covariance X Y structure as given in Theorem 4. . In order to ensure the normalization property of the estimator in small samples. Gaißer et al.) Φn − Φn . . For dependent observations. Let Φn estimator of Hoeffding’s Phi-Square calculated from the original sample when we have deleted block Bs. n) [0. j=1 i=1 h(d. Uik }) j=1 k=1 i=1 d 2 1 − Uij − − with 1 − Uij n + 1 3 d (n − 1)(2n − 1) 2n2 d . [20] in the context of tests for stochastic n independence (cf. .Φ 2 and σ 2 2 = σΦ 2 .Φ 2 = {2h(d)}2 X Y [0.

028 0. Ujd ).055 0.016 0. especially for the sample sizes n = √ and n = 500. Additionally.013 0.035 0. We analyze the association between the daily (log-)returns of major S&P global sector indices before and after the bankruptcy of Lehman . Set εj = (Φ −1 (Uj1 ).197 0.008 0.107 0.033 0. Their calculation is based on 1000 Monte Carlo simulations of size n and 250 bootstrap replications. .007 0.022 0.007 0. .191 0.033 0.002 0. and 7.000 0.017 0. . n.005 0.015 0.208 0. j = 1. .003 0.009 0.000 0.007 0. .001 0.015 0.001 0. .001 0.015 0.000 0. The autoregressive coefficient β of the AR(1)-process equals 0. we set the block length l = 5 in all simulations.010 0. . respectively. . Φ −1 (Ujd )).043 0.044 0.2 50 100 500 50 100 500 50 100 500 0. which is calculated from a sample of size 100. Table 2) based on the equi-correlated Gaussian copula as defined in (8).005 0.010 0.000 0. jackknife estimates by J.003 0.033 0.005 0.007 0.077 0.045 0.004 0. the jackknife is of higher computational complexity.000 0.001 0.004 0. sample sizes n = 50.1 Dimension d = 5 0.1 Dimension d = 10 0. For large sample size n.047 0. we apply the theoretical results of Section 4 to empirical financial data.098 0.100 0.035 0.032 0.022 0.015 0.003 0.000 0.022 0.098 0. . . / Journal of Multivariate Analysis 101 (2010) 2571–2586 ˜2 Gaussian copula (independent observations).002 0.049 0.002 0. we observe a finite-sample bias which strongly depends on the dimension d and the parameter choices. we consider the case of independent observations from the equi-correlated Gaussian copula (cf.002 0.191 0. m(σ ).008 0.002 0.005 0.001 0. m(σ B ).046 0. 5. and 500.014 0.002 0.062 0.011 0.2 50 100 500 50 100 500 50 100 500 0. .202 0.007 0.019 0.056 0. The bootstrap estimates are labeled by the superscript B. The third column of Tables 1 and 2 shows an approximation to the true value of ˜2 Φ 2 .003 n √ Φ2 ˜2 m(Φn ) ˜2 s(Φn ) m(σ B ) ˆ m(σ J ) ˆ s (σ B ) ˆ s(σ J ) ˆ 0.016 0.048 0.005 0.007 0.023 0.218 0.000.067 0.002 0.008 0. j = 1.005 0.000 0.012 0. the block bootstrap is used in this case.065 0.. ρ Dimension d = 2 0.008 0.197 0. Empirical study For illustration purposes.002 0.003 0.001 0. we proceed as follows: Simulate n independent ddimensional random variates Uj = (Uj1 .001 0.069 0.000 0.000 0.095 0.013 0.197 0.012 0.049 0. .098 0.007 0.001 0. too.008 0.026 0.094 0.2580 Table 1 S.008 0.015 0.196 0.002 0. Simulation results for estimating the asymptotic standard deviation of n(Φn − Φ 2 ) by means of the nonparametric block bootstrap with block length l and the delete-l jackknife (for l = 5): The table shows the empirical means m(·) and the empirical standard deviations s(·) of the respective estimates.009 0.003 0. To ease comparison.000 0.000 0.000 0.021 0.014 0. The standard deviation estimations s(Φn ) and the empirical means of the block bootstrap estimations. Tables 1 and 2 outline the simulation results for dimensions d = 2. There is a good ˆ agreement between their values.054 0.015 0.003 0. . n. To generate these observations.5.021 0.028 0. n.111 0. implying that the bootstrap and the 100 ˜2 jackknife procedure to estimate the asymptotic standard deviation of n(Φn −Φ 2 ) perform well for the considered Gaussian copula models.032 0.010 0. as well ˆ J as the delete-l jackknife estimations.000 0.046 0.5 −0.015 0.010 0.004 0.089 0.001 0.5 −0.011 0.015 0. 100.n of the AR(1)-process is then obtained by setting X1 = ε1 and completing the recursion Xj = β Xj−1 + εj .032 0.032 0. .037 0. which are calculated based on 1000 Monte Carlo simulations of sample size n of a d-dimensional equicorrelated Gaussian copula with parameter ρ and 250 bootstrap samples.2 50 100 500 50 100 500 50 100 500 0.001 0. and 10.002 0.070 0.028 0. the standard error s of the bootstrap standard deviation estimations is quite small (column 8) and slightly smaller than the obtained jackknife estimates (column 9) in lower dimensions. . .000 0.000 0.007 0.003 0.191 0.047 0.012 0.032 0.231 0. and different choices of the copula parameter ρ .017 0.001 0.002 0.022 0.015 0.008 0. A sample (Xj )j=1.008 0.026 0.031 0.5 −0.026 0.021 0.007 0.007 0. .007 0.035 0.001 0. however. . .011 0..063 0. and which decreases with increasing sample ˜2 size. from the equi-correlated Gaussian copula with parameter ρ . .021 0..008 0.034 0. Table 1). Gaißer et al.032 0. 5.000 0.007 0.015 0.017 0. Further.1 We consider observations from an AR(1)-process with autoregressive coefficient β (cf.009 0. j = 2. for the standard deviation are given in columns 5. For simplicity.007 0.030 0.035 0.002 0.001 0..000 0.036 0.202 0. Comparing the latter to m(Φn ) (column 4). 6.003 0.

2008 (left panel).008 0.032 0.026 0. The major S&P global sector indices considered are Financials.086 0.053 0. Table 3 reports the first four moments of the .004 0.016 0.008 0.001 0.011 0.242 0.017 0.000 0.000 0.029 0.054 0.000 0.002 0.196 0.025 0.060 0.109 0.059 0.5 −0.025 0. Industrials.000 0. Simulation results for estimating the asymptotic standard deviation of n(Φn − Φ 2 ) by means of the nonparametric block bootstrap with block length l and the delete-l jackknife (for l = 5): The table shows the empirical means m(·) and the empirical standard deviations s(·) of the respective estimates.005 0.2 50 100 500 50 100 500 50 100 500 0.007 0.003 0.5 for the first lag) and 250 bootstrap samples.019 0.037 0. using multivariate Hoeffding’s Phi-Square.072 0.114 0. Brothers Inc.022 0.203 0.010 0. The vertical line indicates the 15th of September 2008.004 0.039 0.018 0.004 0.002 0.008 0. and IT with respect to their value on January 1.003 0.050 0.005 n √ Φ2 ˜2 m(Φn ) ˜2 s(Φn ) m (σ B ) ˆ m(σ J ) ˆ s (σ B ) ˆ s (σ J ) ˆ 0.016 0.036 0.S. all series are plotted with respect to their value on January 1. 2008.020 0.014 0.002 0.5 −0. which are calculated based on 1000 Monte Carlo simulations of sample size n of a d-dimensional equi-correlated Gaussian copula with parameter ρ. ρ Dimension d = 2 0.032 0.111 0.040 0.058 0.005 0.008 0.020 0.014 0.014 0. ˜ Multivariate Hoeffding’s Phi-Square Φ 2 of the four indices’ returns series.001 0.002 0.020 0. All series decrease in mid 2008 in the course of deteriorating financial markets.001 0.047 0.032 0.005 0.000 0.031 0.009 0.031 0.000 0.002 0.099 0.025 0.001 0.008 0. The bootstrap estimates are labeled by the superscript B.1 Dimension d = 10 0.000 0.032 0.000 0.017 0.003 0.001 0. 2.101 0. the day of the bankruptcy of Lehman Brothers Inc.000 0.003 0.003 0.205 0.007 0.002 0.006 0.009 0.002 0.001 0.023 0.1 Dimension d = 5 0. 2 (left panel) shows the evolution of the four indices over the considered time horizon.003 0.1 Fig.2 50 100 500 50 100 500 50 100 500 0. To ease comparison.004 0.033 0.009 0.015 0.010 0.042 0.000 0. Industrials.015 0.110 0.005 0.192 0.026 0.008 0.016 0.013 0.018 0.042 0.017 0.019 0.008 0.000 0.212 0.022 0.049 0.009 0.000 0.007 0.039 0. Energy.001 0.200 0. AR(1)-processes with standard normal residuals in each margin (with coefficient β = 0.065 0.053 0.005 0.000 0.100 0.011 0.099 0.222 0.006 0.082 0.014 0.057 0.028 0. the day of the bankruptcy of Lehman Brothers Inc.192 0.002 0.000 0.200 0. and IT during the period from 1st January 2008 to 8th April 2009 (331 observations).5 −0.028 0.055 0. they decline especially sharply after the bankruptcy of Lehman Brothers Inc.086 0.021 0.001 0.067 0.004 0.015 0.010 0.017 0.007 0.045 0.028 0.192 0.013 0.099 0.000 0.037 0.007 0.002 0. The vertical line indicates the 15th of September 2008.007 0.028 0.059 0. Fig.008 0.037 0.037 0. Gaißer et al.043 0.008 0.000 0.019 0.081 0. / Journal of Multivariate Analysis 101 (2010) 2571–2586 Table 2 2581 ˜2 Gaussian copula (dependent AR(1) observations).000 0.076 0.008 0. jackknife estimates by J.002 0.011 0.008 0.200 0.001 0. Evolution of the S&P global sector indices Financials.2 50 100 500 50 100 500 50 100 500 0.030 0.012 0.023 0.031 0. Energy. where the estimation is based on a moving window approach with window size 50 (right panel).012 0.007 0.074 0.001 0.039 0.

6. (σbefore )2 = 0.7437. A nonparametric estimator for the proposed measure is obtained on the basis of the empirical copula process. the last two rows of the table display the results of the Ljung–Box (LB) Q -statistics.0553 −0. Appendix A.2462 6. In addition. Gaißer et al.1490 5. calculated for the returns of the S&P indices. The asymptotic distribution can be approximated by nonparametric bootstrap methods.2064.after ˜B ˜B ˜B = 8.1) rejects the null hypothesis of stochastic independence (i.2213 2.1148 5. we are grateful to the Regional Computing Center at the University of Cologne for providing the computational resources required.1732 3.2024 67. as stated in Eqs.0000. Its asymptotic distribution is established for the cases of independent observations as well as of dependent observations from a strictly stationary strong mixing sequence.3190. Funding in support of the second author’s work was provided by the Deutsche Forschungsgemeinschaft (DFG). ˜B σbefore.1696 0. We observe a sharp increase of Hoeffding’s Phi-Square after this date and.1875 0.1663.after = −0. C = Π ) with a p-value of 0.0001 386.6229 0. hence.7495 0.1637 2. computed from the squared returns of the indices up to lag twenty. The Jarque–Bera (JB) test rejects the null hypothesis of normality at all standard levels of significance. The corresponding test statistic has the form √ n ˜2 ˜2 (Φafter − Φbefore ) (σbefore )2 + (σafter )2 − 2σbefore. estimated on the basis of a moving window approach with window size 50. Financials Mean Standard deviation Skewness Kurtosis JB statistics JB p-values LB Q -statistics LB p-values Energy Industrials IT −0.0000 266.6130 0.0000 −0.1037 0.1982.3140 3. Section 4. an increase of the association between the indices’ returns. Fig.6008 0. we have .8467 47.0000 141. Conclusion A multivariate version is proposed for Hoeffding’s bivariate measure of association Phi-Square. We calculate the estimated values (based on 250 bootstrap samples with block length l = 5) of Hoeffding’s Phi-Square and the asymptotic variances and covariance as stated in Theorem 6 for both time periods which comprise n = 100 observations each: ˜2 Φbefore = 0. Again. all squared returns show significant serial correlation as indicated by the Ljung–Box (LB) test.0000 daily returns of the four indices as well as the related results of the Jarque–Bera (JB) test. All return series show skewness and excess kurtosis. Further. which rejects the null hypothesis of no serial correlation. In order to verify whether this increase is statistically significant. We fit an ARMA–GARCH model to each return series (which cannot be rejected by common goodness-of-fit tests).6540 187.0000 −0. / Journal of Multivariate Analysis 101 (2010) 2571–2586 Table 3 First four moments (in %) and results of the Jarque–Bera (JB) test. Hence.2. with corresponding p-value 0.1.2582 S. calculated over the entire time horizon. [6]). Derivation of the functions h(d )−1 and g (d )−1 We calculate the explicit form of the functions h(d)−1 and g (d)−1 . This allows for the calculation of asymptotic confidence intervals for Hoeffding’s Phi-Square and the construction of hypothesis tests.1199 −0.4846 0.0005 such that Theorem 6 can be applied. (3) and (4). 2 (right panel) displays the evolution of multivariate Hoeffding’s Phi-Square of the indices’ returns. ˜B The choice of the block length l = 5 is motivated by the results of the simulation study in Section 4. the vertical line indicates the day of Lehman’s bankruptcy.0287. we conclude that there has been a significant increase in association between the returns of the four indices after the bankruptcy of Lehman Brothers Inc..8456 0. calculated up to lag twenty from the squared returns. we compare Hoeffding’s Phi-Square over the two distinct time periods before and after this date using the results on the statistical properties of the difference of two Hoeffding’s Phi-Squares discussed in Section 4. Regarding the function h(d)−1 .8856 0. The estimated parameters are consistent with the assumption of strong mixing series (cf. Moreover.e.2523 71.1139 4. Note that the test by Genest and Rémillard [21] (cf. ˜2 Φafter = 0. ˜B (σafter )2 = 0. The results are derived under the general assumption of continuous marginal distributions.0000 −0.0000 447. Acknowledgments We thank the associate editor and three anonymous referees for their comments which led to an improvement of the paper. as well as results of the Ljung–Box (LB) Q -statistics.

S. .i. / Journal of Multivariate Analysis 101 (2010) 2571–2586 2583 h(d)−1 = [0. Ud }. . Ud are i. Derivation of the estimator Φn 2 We outline the derivation of the estimator Φn as given in (10). .1]d 1{Uij ≤ui } − i=1 1{Uik ≤ui } − i=1 i =1 ui du . Vd }) = 1 2d E (Y ) where V1 . (A. .4) and using that [0. we obtain M (u)Π (u)du = 1 2d [0.1) For the second summand.1]d d [0.1]d Cn (u) − i =1 n n ui d du = [0. . . .1]d {M (u) − Π (u)}2 du {M (u)}2 du − 2 M (u)Π (u)du + = [0.d. by combining Eqs. which has density fV (v) = 2v for 0 1 E (Y ) = d d 2 2 1 1 v 1 and Y = min {V1 . Thus.2) Combining Eqs. . 1 d . . (A. . ud } i =1 2ui du = 1 2d E (min{V1 . . For the first summand. (A. {W (u) − Π (u)}2 du {W (u)}2 du − 2 W (u)Π (u)du + = [0.1]d W (u)Π (u)du = 0 d ··· d−2− ud−1 ui d−1− ud ui i=1 ui − d − 1 dud · · · du2 du1 i=1 i=1 = i=1 d i (−1)i 1 (d + 1 + i)! . The first summand on the left-hand side of the above equation can be written as [0. . . (4). 1 E X2 = d 0 x2 (1 − x)d−1 dx = 2 (d + 1)(d + 2) . . Ud }]2 = E X 2 where U1 .4) we obtain the asserted form of g (d)−1 . . . . .1]d {M (u)}2 du = E [min{U1 .1]d [0. . d 2 {h(d)} −1 2 Φn = [0.1]d d 1 n d d 2 n j =1 ui i=1 1{Uij ≤ui } − i=1 d i=1 d ui du = 1 n 2 j=1 k=1 [0. . . .1]d [0. we have g (d)−1 = [0. 3 (A. xd(1 − x2 )d−1 2xdx = 0 1 2d 0 1 (1 − x2 )d dx = 1 1 Γ (d + 1) π 2d 2 Γ d + 1 + 1 2 √ = 1 2d d d! i+ i =0 1 2 . 1] and X = min{U1 .3) and (A. Gaißer et al. .i.1]d {Π (u)}2 du = Regarding the function g (d)−1 as defined in Eq.1) and (A.1]d [0. . V2 . (A. Vd }.1]d {Π (u)}2 du = 2 Appendix B.1]d {Π (u)}2 du. . Therefore. (A. from U ∼ R[0.1]d {Π (u)}2 du.1]d {W (u)}2 du = 0 ··· d−2− ui − d + 1 ui i=1 dud · · · du2 du1 = 2 ui i=1 d−1− i=1 (d + 2)! . it follows that 1 1 1 d−2 d−1 d 2 [0.3) Partial integration of the second term further yields 1 1 1 d−2 d d−1 [0. Vd are i.1]d [0. from V . V2 .2) and using that [0. . Again.1]d 1 d 3 yields the asserted form of h(d)−1 .d. .1]d min{u1 . .

consider 2 h(d)−1 Φα1 (X) = [0. Gaißer et al.X2 ) (1. u2 . ud ) − xu2 · · · · · ud }2 dxdu − [0. defined on the range of Xk . . . Consider the d-dimensional random vector X with copula C .1] CX (x. u2 . dxdy = [0. . the copula C is invariant under strictly increasing transformations of one or several components of X 2 (see. . Theorem 2. .1]d {C (u) − Π (u)}D(u)du. Xk−1 . let αk (X) = (X1 . we have .6]). . . k ∈ {1. . ud ) = u2 · · · · · ud . . Xd are independent or • CX (1. . . . . ud ) − u2 · · · · · ud } − {CX (x. (C. . for the second equation. . ud ) − u2 · · · · · ud }2 du + {CX (1. . . For dimension d ≥ 3. u2 ) − u1 u2 u1 + u2 − 1 − C(X1 . . . . . / Journal of Multivariate Analysis 101 (2010) 2571–2586 = 1 n 1 n 2 n n d d d d j=1 k=1 2 n n [0. . . . u2 ) − C(X1 . . . j=1 i=1 Appendix C. .α2 (X2 )) = [0.1]d {CX (x. . a continuous linear map on ∞ ([0.1]d Cα1 (X) (u1 . . ud )dx. which can be shown as follows: For all converging sequences tn → 0 and Dn → D such that C + tn Dn ∈ ∞ ([0. ud ) − CX (1 − u1 . . As direct functional of the copula. . we have 2 h(2)−1 Φ(α1 (X1 ). u2 . ud ) − u2 · · · · · ud } 2 [0. Hence. u2 ) − Π (u1 . y) − (1 − x)(1 − y) C(X1 . u2 . u2 . ΦX thus inherits this property. u2 ) 2 du1 du2 2 = [0. Φ 2 is invariant as well.X2 ) (x. . . u2 . . . . . ud ) − u1 u2 · · · · · ud }2 du [{CX (1. .7. . ud ) − xu2 · · · · · ud }]2 dxdu = [0. . . u2 . is given by ϕC (D) = 2h(d) [0. .. 1]d ) for every n. . . u2 .1]d where u = (u2 .1]d {CX (1. . . For dimension d = 2. . . Theorem 2. . (i) For all d ≥ 2. u2 .1]d−1 [0. . Xk+1 .48] for the relevant definitions and background reading). . If a random vector X is jointly symmetric about a ∈ Rd then its copula is invariant with respect to a strictly decreasing transformation of any component. ud ) = xCX (1. . . ud ) in this case. . . . u2 . . ud ) = [0. .1]d−1 {CX (1. . e. . d}. . This condition is fulfilled if X1 is independent of (X2 . For non-jointly symmetric random vectors.α2 (X2 )) (u1 . . meaning that X2 . . .1) 2 2 According to Eq. Proof of Theorem 3. Note that Φ 2 = ϕ(C ) represents a Hadamard-differentiable map ϕ on ∞ ([0.g.1).1]2 1 − x + 1 − y − 1 + C(X1 . ud ) 2 du = [0. ud ) = 2 [0. . . . . . it holds that Φα (X) = ΦX if either 1 • CX (1.Uik }≤ui } + i=1 i=1 u2 − i i=1 d n d ui 1{Uij ≤ui } − i=1 2 (1 − Uij ) + ui 1{Uik ≤ui } du 1 3 d = (1 − max{Uij . αk (Xk ). .1]2 C(α1 (X1 ). Proofs Proof of Proposition 1. (C. Xd ) denote the random vector where the kth component of X is transformed by the function αk . (ii) Let αk be a strictly decreasing transformation of the kth component Xk of X. . u2 .1]2 C(X1 . . 1]d ). .X2 ) (x. . u2 .2584 S. . . Without loss of generality set k = 1. 1 − u2 ) − u1 u2 du1 du2 du1 du2 2 = [0. . .1]2 2 = [0.α2 (X2 )) (1 − u1 . y) − xy 2 2 dxdy = h(2)−1 ΦX .1]2 see e. Uik }) − j=1 k=1 i=1 2 n 1 2 . .1]d d 1{max{Uij . . . u2 .g. . u2 . [13.1] CX (x. . ud )dx − u2 · · · · · ud du . [13]. Its derivative ϕC at C ∈ ∞ ([0. ud ) − Π (u1 . 1]d ). Xd ) since CX (x.X2 ) (1 − u1 . 1]d ) of the copula C (see [14. . u2 .

. First. . d. . it holds that √ n{Fn (u) − F (u)} −→ B∗ (u) w in ∞ ([0. both equipped with the uniform metric m. σΦ ). d (C.S. are uniform distributions on [0. 1] and thus. Another application of the Delta-method to (C. . v). 1.4 in [48] together with Proposition 2 then implies √ n( Φ 2 − Φ 2 ) = √ n{ϕ(Cn ) − ϕ(C )} −→ ϕC (GC ). proof of Theorem 4). 1]d .1]d {C (u) − Π (u)}2 du − tn tn 2 2h(d)tn [0. 1]d with map φ : D([0. . Hence. . . Φ 2 {CZ (1.2) is bounded for all Dn .n − CZ } to the Gaussian process G∗ in ∞ ([0. . 1]d . . 1)du Z   [0. 1.3) yields the weak convergence of 2 ZΦ ∼ N (0. 1. Fd 1 (ud )). 1]d ) (tangentially to C ([0. 1]d ). 1]d ) comprises all real-valued cadlag functions and C ([0. . the map φ is Hadamard-differentiable at F as a map from D([0. the space D([0. the assertion follows by an application of Slutsky’s theorem. weak convergence of the empirical copula process n{CZ. . . . the asymptotic behavior of n(Φn − Φ 2 ) as stated in the √ theorem follows by mimicking the proof of Theorem 3. . . Let CZ. . 1 . Here.1]d D2 (u)du n tn → 2h(d) [0. by considering the transformed random variables Uij = Fi (Xij ). the functional Delta-method is applied which is based on the following representation of the copula C as a map of its distribution function F (cf.9. . (C. . . . Gaißer et al. 1]d ) → ∞ ([0. j = 1.2) for n → ∞. √ n{Φ (Cn ) − Φ (C )} to the random variable Proof of Theorem 4. 1. Hence. Since sup |φ(Fn )(u) − Cn (u)| = O 1 n u∈[0. . . Lemma 3.1]d {C (u) − Π (u) + tn Dn (u)}2 du h(d) [0. Since 2 ΦX 2 ΦY = Φ 2 {CZ (u. . . . Rio [36] shows that under the above assumptions on the mixing coefficient α(r ). n{(ΦX .8 2 in [48]. n. . 1)} = g (CZ ). If all partial derivatives of the copula exist and are continuous. Proof of Theorem 6. u ∈ [0. . 1]d ) the space of all continuous real-valued functions defined on [0. 377. it is possible to confine the analysis to the case where the marginal distributions Fi of F . apply the continuous mapping theorem to conclude the Y Z proof. .1]d cf.1]d {C (u) − Π (u)}GC (u)du. i = 1. . Zn . 1. . . .n denote the empirical copula based on the sample Z1 . since the second integral in Eq. . p. . Second. gCZ (G∗ ) =  Z   {CY (v) − Π (v)}G∗ (1. . . . . . ΦY ) } converges in ∗ distribution to the multivariate normally distributed random vector gCZ (GZ ) given by  {CX (u) − Π (u)}G∗ (u. . 1]d ) with Gaussian process B∗ as defined in the theorem. ΦY ) can be established analogously as in the proof of Theorem 3 using the Hadamard √ 2 2 2 2 differentiability of the map g at CZ whose derivative is denoted by gCZ . Lemma 2 in [14]) with derivative φF . Further.√ Cn denote the empirical copula based on the sample X1 . / Journal of Multivariate Analysis 101 (2010) 2571–2586 2585 ϕ(C + tn Dn ) − ϕ(C ) tn = = h(d) [0. An application of the functional Delta-method given in Theorem 3.3) where ϕC (GC ) = 2h(d) [0. v)dv Z [0. i = 1. implies that ZΦ 2 = ϕC (GC ) is normally distributed with mean zero and variance σΦ 2 as stated in the theorem.1]d . .1]d {C (u) − Π (u)}Dn (u)du + tn [0. F has compact support [0. v)} 2 2 the asymptotic behavior of (ΦX . Weak convergence of n(Cn − C ) can be established analogously as in [14] (see also [7]) and we outline the single steps in the following. d.1]d  With GX (u) = GZ (u.9. cf. (C. 1) and G∗ (v) = G∗ (1. Using the fact that GC (u) is a tight Gaussian process. [14]. . ∗ ∗ .1]d {C (u) − Π (u)}D(u)du. 1]2d ) follows Z (cf. Under the assumption of the √ theorem. an application of the functional Delta-method √ yields the weak convergence of n{φ(Fn ) − φ(F )} to the process φF (B∗ ) = G∗ . . . Xn . ΦY ) − (ΦX . . . . 1]d ). Given the weak convergence of Let √ 2 the empirical copula n(Cn − C ) process to the Gaussian process G∗ . [48]): − − C (u) = φ(F )(u) = F (F1 1 (u1 ).

Nešlehová.K. Sen (Eds. [33] R. Bull. [19] C. 9 (4) (1981) 879–885. The Collected Works of Wassily Hoeffding. 33 (3) (2005) 357–375. 25–26 September 2009. Blum.-F.M. Math. Metrika 66 (3) (2007) 323–354. Masstabinvariante Korrelationstheorie. Seminars Inst. H.). Tsukahara. 22 (1) (1970) 57–63. Bühlmann. Rio. Multivariate Anal. León R. Stat. Kaas. Multivariate conditional versions of Spearman’s Rho and related measures of tail dependence. E. 4 (5) (1976) 912–923. [12] P. J. The concept of comonotonicity in actuarial science and finance: applications.A. Nonparametr. Data Anal. Goovaerts. Marcel Dekker. Angew. Dhaene. R. T. 345–364. J. Nonparametric estimation of copulas for time series. D. Statist. Statist. La fonction de dépendance empirique et ses propriétés. Math. Nonparametric measures of multivariate association. [18] P. [15] J. Aktuarver. 10354. The concept of comonotonicity in actuarial science and finance: theory. Wellner. H. Ben Ghorbal. Úbeda-Flores. Kiefer. Wolff. Behboodian. N. Ann. Schweizer. Künsch. [2] J. 2nd ed. pp. M. N-dimensional measures of dependence. Fernández. Inst. 1997. Bootstrapping Spearman’s multivariate Rho. Ann. Doukhan. in: Distributions with Fixed Marginals and Related Topics. [35] D. J.D. Nelsen. M. Theoret. [27] H. Multivariate versions of Blomqvist’s beta and Spearman’s footrule. Inst. O. pp. TEST 13 (2) (2004) 335–369. W. Math. vol. Block length selection in the bootstrap for time series. Springer Verlag. [38] F. Math. Joe. Härdle. Acad.2586 S. Dehling. Multivariate concordance. Diss. Radulović. Theory 18 (1) (2002) 17–39. Springer Verlag. Univ. Mikusiński. J. . M. [32] P. Statist. Genest. McNeil. [47] M. [20] C. F. Dhaene. Statist. J. Probab. Econ. Fonctions de répartition à n dimensions et leurs marges. Un test non paramétrique d’indépendance. Springer Verlag. in: DMV-Seminar. Hoeffding. 22 (2010) in press (doi:10. S. Ann. ETH No. Multivariate measures of concordance.R. On nonparametric measures of dependence for random variables. 35 (1) (2007) 166–191. vol. Inst. Papers 48 (2) (2007) 295–304. Sklar. Berlin 5 (3) (1940) 181–233. Risk 5 (2003) 25–54. An Introduction to Copulas. 32 (2) (1961) 485–498. On the bootstrap and empirical processes for dependent sequences. Multivariate Anal. Math. Fisher. Statist. Cl. Statist. [26] H. An empirical central limit theorem with applications to copulas under weak dependence. [45] M. 34 (4) (1963) 1315–1321. M. 57 (4) (2005) 781–788. 9. [42] F. in: IMS Lecture Notes— Monograph Series. Statist. Fermanian. Rémillard. [10] J. Weak Convergence and Empirical Processes. [39] F. 2006. Ann. Joe. Process. C. Inference Stoch. 84 (405) (1989) 157–164. Stochastische Abhängigkeit und funktionaler Zusammenhang. [8] P. M. in: A. [36] E.B. [37] L. in: S. Statist. M. Inst. M. J.J. Canad.). Elsevier. Rychlik (Eds. Lang. 12 (1) (2009) 65–87. Dedecker. Schmidt. Bühlmann. 89 (2) (2004) 351–370. Empirical Process Techniques for Dependent Data. Belg. Wolff. R. Publ. P. Tidskr. Stat. 1996. Statist. Stochastica 13 (1) (1992) 61–74. Scaillet. Amer.F. Gänssler. Doukhan. The jackknife and the bootstrap for general stationary observations. 2006. B. B. [14] J.A. On measures of multivariate concordance. Spearman’s footrule and Gini’s gamma: a review with complements. Chapman & Hall. [5] P. 35 (1) (1990) 12–30. Schmid. / Journal of Multivariate Analysis 101 (2010) 2571–2586 References [1] M. Lett. R. 223–232. Chen. [4] P. R. Seminar on Empirical Processes. Springer Verlag. Statist. Statist. [48] A. [40] F. 1987. Math. Quessy. 25 (1942) 200–227. 2010. 1996. J. [9] J. Prieur. Rüschendorf. Roy. in: P. 65 (5) (1979) 274–292. Mikosch. A multivariate version of Gini’s rank association coefficient. R. M. in: H. 2007. Lee. [24] W. Univ. M. Vichi (Eds. Multivariate Models and Dependence Concepts. COMPSTAT—Proceedings in Computational Statistics. Dissertation ETH Zürich. Vyncke. [25] H.R.B. [29] H. Van der Vaart. Statist. J. Copula-based measures of multivariate association. Copula Theory and its Applications—Proceedings of the Workshop Held in Warsaw. Springer Verlag. Statist. W. T.J. Relative entropy measures of multivariate dependence. Multidimensional dependency measures. Ruppert. Dolati. J. X. Sørensen (Eds.W.. J. [11] A. 1994. pp. Math. Kaas. [31] A. J. 2002. Birkhäuser. [49] E. Schmidt. Asymptotic local efficiency of Cramér-von Mises tests for multivariate independence. 1990. 329–384. 209–235. Carrasco. Genest. Nonparametric inference on multivariate versions of Blomqvist’s beta and related measures of tail dependence. Weak convergence of empirical copula processes. [17] N. Modelling dependence with copulas and applications to risk management. Ann.D. Mixing and moment properties of various GARCH and stochastic volatility models. S. 2003. Yanagimoto. Ann. Künsch. J. Lang. J. Schmid. 19 (4) (1948) 546–557. Lancaster. Durante. Úbeda-Flores.. Embrechts. D. Gaißer. Paris 8 (1959) 229–231. D. Distribution free tests of independence based on the sample distribution function. 31 (2) (2002) 133–161. Schmidt. Probab. Hoeffding. R.). Ann. Statist. J. Ann. Vyncke. T. Joe. Rosenblatt. 17 (3) (1989) 1217–1241. [7] J. U-Statistics: Theory and Practice. Schmid. 77 (4) (2007) 407–416.-D. Bernoulli 10 (5) (2004) 847–860. A non-parametric test of independence. Radulović. Math. pp. Louhichi. Handbook of Heavy Tailed Distributions in Finance. F.). Statist. [43] B. Taylor. González-Barrios. [3] J. [46] H. Central limit theorems for empirical and U-processes of stationary mixing sequences. [28] H. M. Ann. [23] W. Semiparametric estimation in copula models. 7 (1) (1994) 47–71. [6] M. Schmid. 1993. Nelsen. Jaworski. Úbeda-Flores. Shuffles of min. Statist. [41] F. Fermanian.O.F. Asymptotic distributions of multivariate rank order statistics. 4 (2) (2006) 147–163.R. Denuit. [30] H. The blockwise bootstrap in time series and empirical processes. 31 (3) (1999) 295–310. Assoc. Rémillard. Rizzi. 98 (6) (2007) 1123–1140. On measures of association and a related problem. 31 (1) (2002) 3–33. Insur. Stute.J. Fermanian. Joe. 41 (4) (1989) 683–697. Schmidt. Birkhäuser Verlag. [13] P. Springer Verlag.R. [50] T. Econ. Insur. A. Sci. G. A Dolati. Math. Weak Dependence: With Examples and Applications. 28. Arcones. Statist. Econom. Wegkamp. Probab. Physika-Verlag. R. Deheuvels. Math. P. Théorie Asymptotique des Processus Aléatoires Faiblement Dépendants.I. G. Schmidt. Comput. M. [44] A. Taylor. Denuit. Multivariate Anal. Schrift. Sci. Inst. Tests of independence and randomness based on the empirical copula process. Blumentritt.1080/10485250903499667). Rachev (Ed. Stat. Lindskog. [16] B.F. Stochastica 4 (3) (1980) 175–188.). Correlation and complete dependence of random variables. Schmid.-D. Goovaerts. B. M. J. [21] C. [34] R. J. Math. 2000. Sherwood.-D. [22] W. Ann. Ann. Multivariate extensions of Spearman’s Rho and related statistics. Yu. Genest. Gaißer et al. Hoeffding. Estimation of entropy and other functionals of a multivariate density. 59 (4) (2007) 789–806. Skand.