J. Appl. Prob.

46, 925–937 (2009)

Printed in England
© Applied Probability Trust 2009

TAIL DEPENDENCE FOR HEAVY-TAILED SCALE

MIXTURES OF MULTIVARIATE DISTRIBUTIONS

HAIJUN LI ∗ ∗∗ and
YANNAN SUN,∗ ∗∗∗ Washington State University

Abstract
The tail dependence of multivariate distributions is frequently studied via the tool of
copulas. In this paper we develop a general method, which is based on multivariate regular
variation, to evaluate the tail dependence of heavy-tailed scale mixtures of multivariate
distributions, whose copulas are not explicitly accessible. Tractable formulae for tail
dependence parameters are derived, and a sufficient condition under which the parameters
are monotone with respect to the heavy tail index is obtained. The multivariate elliptical
distributions are discussed to illustrate the results.
Keywords: Tail dependence; heavy tail; multivariate regular variation; copula; hazard rate
ordering; multivariate elliptical distribution
2000 Mathematics Subject Classification: Primary 62H20; 62P05

1. Introduction
Tail dependence parameters describe the amount of dependence in the upper tail or lower tail
of a multivariate distribution and can be used to analyze the dependence among extreme values.
This paper focuses on the tail dependence of multivariate distributions of random vectors of the
form
(X1 , . . . , Xd ) := (RT1 , . . . , RTd ), (1.1)
where (T1 , . . . , Td ) has the joint distribution G(t1 , . . . , td ) with some finite moments, and the
scale variable R, independent of (T1 , . . . , Td ), has a regularly varying right tail at infinity with
survival function (see, e.g. [3, pp. 17–21] for details),
L(r)
F̄ (r) := 1 − F (r) = , r > 0, α > 0, (1.2)
rα
where L is a slowly varying function; that is, L is a positive function on (0, ∞) with the property
that
L(cr)
lim = 1 for every c > 0. (1.3)
r→∞ L(r)
The class of distributions of the form (1.1) has a variety of interpretations in different appli-
cations, including, for example, multivariate elliptical distributions and various multivariate
Pareto distributions as special cases.
It is possible to define and analyze tail dependence via copulas; however, the copula method’s
effectiveness diminishes in higher dimensions and in the case of (1.1), where the copulas are

Received 8 March 2007; revision received 14 September 2009.

∗ Postal address: Department of Mathematics, Washington State University, Pullman, WA 99164, USA.
∗∗ Email address: [email protected]
Research supported in part by NSF grant CMMI 0825960.
∗∗∗ Email address: [email protected]

925

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 926 H. LI AND Y. SUN

often not explicitly accessible. A copula C is a distribution function, defined on the unit cube
[0, 1]d , with uniform one-dimensional margins. Given a copula C, if we define

F (t1 , . . . , td ) := C(F1 (t1 ), . . . , Fd (td )), (t1 , . . . , td ) ∈ Rd , (1.4)

then F is a multivariate distribution with univariate margins F1 , . . . , Fd . Given a distribution

F with margins F1 , . . . , Fd , there exists a copula C such that (1.4) holds. If F1 , . . . , Fd are all
continuous then the corresponding copula C is unique and can be written as

C(u1 , . . . , ud ) = F (F1−1 (u1 ), . . . , Fd−1 (ud )), (u1 , . . . , ud ) ∈ [0, 1]d .

Thus, for multivariate distributions with continuous margins, the univariate margins and the
multivariate dependence structure (as described by their copulas) can be separated [17], [23].
The survival copula is defined similarly. Consider a random vector (X1 , . . . , Xd ) with
continuous margins F1 , . . . , Fd and copula C. Observe that F̄i (Xi ) = 1 − Fi (Xi ), 1 ≤ i ≤ d,
is also uniformly distributed over [0, 1]; thus

Ĉ(u1 , . . . , ud ) := Pr{F¯1 (X1 ) ≤ u1 , . . . , F¯d (Xd ) ≤ ud }

is a copula. We call Ĉ the survival copula of (X1 , . . . , Xd ). The joint survival function of the
random vector (X1 , . . . , Xd ) is expressed as

F̄ (t1 , . . . , td ) := Pr{X1 > t1 , . . . , Xd > td } = Ĉ(F̄1 (t1 ), . . . , F̄d (td )), (t1 , . . . , td ) ∈ Rd .

It also follows that, for any (u1 , . . . , ud ) ∈ [0, 1]d ,

C̄(u1 , . . . , ud ) := Pr{F1 (X1 ) > u1 , . . . , Fd (Xd ) > ud } = Ĉ(1 − u1 , . . . , 1 − ud ), (1.5)

where C̄ is the joint survival function of copula C.

The bivariate tail dependence parameter has been discussed extensively in [11]. Moreover,
various multivariate tail dependence parameters can be introduced. Definition 1.1, below,
details two notions of tail dependence which have been discussed in [14] and [21].
Definition 1.1. Let X = (X1 , . . . , Xd ) be a random vector with continuous margins F1 , . . . ,Fd
and copula C. Let N denote the index set {1, . . . , d}.
1. X is said to be upper-orthant tail dependent if, for some subset ∅  = J ⊂ N , the
following limit exists and is positive:

τJ = lim Pr{Fj (Xj ) > u for all j ∈ N \ J | Fi (Xi ) > u for all i ∈ J } > 0. (1.6)
u→1−

If, for all ∅  = J ⊂ N , τJ = 0, then we say that X is upper-orthant tail independent.

2. X is said to be upper extremal dependent if the following limit exists and is positive:

γ = lim Pr{Fj (Xj ) > u for all j ∈ N | Fi (Xi ) > u there exists i ∈ N } > 0. (1.7)
u→1−

If γ = 0 then we say that X is upper extremal independent.

The lower tail dependence of a copula is defined as the upper tail dependence of its survival
copula in (1.5). The limits τJ and γ are called the upper tail and extremal dependence
parameters, respectively. Obviously, tail dependence is a copula property; these parameters do

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 Tail dependence for heavy-tailed scale mixtures of multivariate distributions 927

not depend on the marginal distributions. If X1 , . . . , Xn are independent then the corresponding
upper tail (extremal) dependence parameters are all zeros. Clearly, τJ ≥ γ for all nonempty
J ⊂ N. Thus, the extremal dependence parameter provides a lower bound for orthant tail
dependence parameters. The multivariate tail dependence parameters, the τJ s, have been
used in [9] to analyze the contagion risk in banking systems, and the extremal dependence
parameter γ in the bivariate case has been used in [4] to analyze the extremal dependence in
financial return data.
Definition 1.1 implies that the tail dependence parameters of a distribution can be derived
directly from its copula, and this has been done for the bivariate case [5], [11]. In [14] and [15],
using the copula method, Li derived explicit expressions of the orthant tail dependence for
Marshall–Olkin distributions as well as multivariate extreme value distributions and their scaled
mixtures. It is already evident in [14] that even if the copula is explicitly available, the direct
copula method has to equalize the distribution margins by taking componentwise marginal
transforms (see (1.4)); this method becomes really cumbersome in higher dimensions. The
copula method is obviously ineffective for (1.1), whose copulas are generally not explicitly
accessible. This paper develops an alternative method to derive tractable tail dependence
formulae for (1.1), which belong to the family of multivariate regularly varying distributions
[1], [2]. Our method is rooted in the theory of multivariate regular variation [19], [20],
and, thus, unlike the copula method, our method avoids taking marginal transforms on the
entire distribution. It is well known that the bivariate normal distribution is asymptotically
tail independent if its correlation coefficient ρ is less than 1. Bivariate elliptical distributions,
a special case of (1.1), possess the upper- (and lower-) orthant tail dependence property if
their generating random variable is regularly varying [10], [21]. The formula for the extremal
dependence γ of bivariate elliptical distributions is also derived in [8]. Our method not only
yields similar results for multivariate elliptical distributions, but also captures a more detailed
relationship between various tail/extremal dependence parameters and the theory of multivariate
regular variation.
The paper is organized as follows. In Section 2 we first discuss the tail dependence functions,
which can be used to express various tail dependence parameters. Furthermore, it is shown
that, for a multivariate regularly varying distribution, the upper tail dependence function and the
intensity measure are equivalent in describing its extremal dependence structure. In Section 3 we
derive explicit expressions for the tail dependence of (1.1). Multivariate elliptical distributions
are also discussed as examples. Finally, some comments in Section 4 conclude the paper.
Throughout this paper, the terms ‘increasing’ and ‘decreasing’ are used in the weak sense, and
the measurability of functions is assumed. The notation a ∨ b and a ∧ b mean the maximum
and the minimum of a and b, respectively.

2. Tail dependence function of multivariate regular variation

Let F be the distribution function (DF) of a d-dimensional random vector X = (X1 , . . . , Xd )
with continuous margins F1 , . . . , Fd and copula C. The lower and upper tail dependence
functions, denoted by b(·; C) and b∗ (·; C), respectively, are introduced in [12], [13], and [18]
as follows:
C(uwj for all j ∈ N )
b(w; C) := lim for all w = (w1 , . . . , wd ) ∈ Rd+ ,
u→0 + u
C̄(1 − uwj for all j ∈ N )
b∗ (w; C) := lim for all w = (w1 , . . . , wd ) ∈ Rd+ , (2.1)
u→0 + u

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 928 H. LI AND Y. SUN

provided that the limits exist. Since b(w; Ĉ) = b∗ (w; C), where Ĉ(u1 , . . . , ud ) = C̄(1 −
u1 , . . . , 1 − ud ) is the survival copula in (1.5), this paper focuses only on the upper tail
dependence. The explicit expression of b∗ for elliptical distributions was obtained in [13].
A theory of tail dependence functions was developed in [12] and [18], based on Euler’s
homogeneous representation:

d
b∗ (w; C) = wj tj (wi , i  = j | wj ) for all w = (w1 , . . . , wd ) ∈ Rd+ , (2.2)
j =1

where

tj (wi , i  = j | wj )
:= lim Pr{Fi (Xi ) > 1 − wi u for all i  = j | Fj (Xj ) = 1 − wj u}, j ∈ N.
u→0+

The tj s are called upper conditional tail dependence functions. For copulas with explicit
expressions, the tail dependence functions are obtained directly from the copulas with relative
ease. For copulas without explicit expressions, the tail dependence functions can be obtained
from (2.2) by exploring closure properties of the related conditional distributions. In [18], for
example, the tail dependence function of the multivariate t distribution is obtained by (2.2).
It follows from (1.6) and (2.1) that the upper tail dependence parameters can be expressed
as
b∗ (1, . . . , 1; C)
τJ = ∗ for all ∅  = J ⊂ N, (2.3)
b (1, . . . , 1; CJ )
where CJ is the multivariate margin of C with component indexes in J . It was shown in
[12] that b∗ (w; C) > 0 for all w ∈ Rd+ if and only if b∗ (1, . . . , 1; C) > 0. Unlike the τJ ,
however, the tail dependence function provides all the extremal dependence information of the
copula C as specified by its extreme value copula (EV copula). The upper EV copula of C,
1/n 1/n
denoted by C UEV , is defined as C UEV (u1 , . . . , ud ) := limn→∞ C n (u1 , . . . , ud ) for any
(u1 , . . . , ud ) ∈ [0, 1] if the limit exists [11].
d

Define the upper exponent function a ∗ ( · ; C) of copula C as

a ∗ (w; C) := (−1)|S|−1 bS∗ (wi , i ∈ S; CS ), (2.4)
S⊆N
S=∅

where bS∗ (wi , i ∈ S; CS ) denotes the upper tail dependence function of the margin CS of C
with component indexes in S. It follows from (1.7) and (2.4) that the extremal dependence
parameter can be expressed as

b∗ (1, . . . , 1; C)
γ = . (2.5)
a ∗ (1, . . . , 1; C)
Similar to tail dependence functions, the exponent function has the following homogeneous
representation:

d
a ∗ (w; C) = wj t j (wi , i  = j | wj ) for all w = (w1 , . . . , wd ) ∈ Rd+ , (2.6)
j =1

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 Tail dependence for heavy-tailed scale mixtures of multivariate distributions 929

where

t j (wi , i  = j | wj )
= lim Pr{Fi (Xi ) ≤ 1 − wi u for all i  = j | Fj (Xj ) = 1 − wj u}, j ∈ N.
u→0+

It was shown in [12] that tail dependence functions {bS∗ (wi , i ∈ S; CS )} and the exponent
function a ∗ (w; C) are uniquely determined from each other.
Theorem 2.1. ([12] and [18].) Let C be a d-dimensional copula. Then the upper EV copula
is defined as

C UEV (u1 , . . . , ud ) = exp{−a ∗ (− log u1 , . . . , − log ud )} for all (u1 , . . . , ud ) ∈ [0, 1]d ,

where a ∗ is given by either (2.4) or (2.6).

Using the language of extreme value theory (see, e.g. Chapter 7 of [16]), if the upper tail
dependence function b∗ (·; C) exists then the copula C is in the copula domain of attraction of
the EV copula C UEV , and C UEV is uniquely determined by b∗ (·; C). Next, we will show that
the upper tail dependence function and the intensity measure for multivariate regular variation
are also uniquely determined from
each other. For notational convenience, we denote hereafter
d
by [a, b] the Cartesian product di=1 [ai , bi ], where a, b ∈ R and ai ≤ bi for each i. The
following definition can be found in [7, pp. 185–286], [19], and [20].
Definition 2.1. A random vector X is said to be multivariate regularly varying (MRV) if there
d
exists a Radon measure µ (i.e. finite on compact sets), called the intensity measure, on R \ {0},
d
where 0 is the origin in R , such that

Pr{X ∈ tB}
lim = µ(B) (2.7)
t→∞ Pr{||X|| > t}
d
for any relatively compact set B ⊂ R \ {0} that satisfies µ(∂B) = 0, where || · || denotes a
d d
norm on Rd . (Here R = [−∞, ∞]d is compact and the punctured version R \ {0} is modified
via one-point uncompactification (see, e.g. [20, pp. 168–170]).)
If X is MRV with µ([0, 1]c ) > 0, where 1 denotes the vector of 1s and [0, 1]c denotes the
d
complement of [0, 1] in Rd , then, for every relatively compact set B ⊂ R \ {0} that satisfies
µ(∂B) = 0,
Pr{X ∈ tB}
lim = µ̃(B), (2.8)
t→∞ Pr{X ∈ t[0, 1]c }

where µ̃(B) = µ(B)/µ([0, 1]c ). In particular, any MRV DF with support in Rd+ = [0, ∞]d
admits the following spectral representation.
Theorem 2.2. ([7], [19], and [20].) If X is nonnegative with MRV distribution F then there
exists a Radon measure µ̃ on Rd+ such that

1 − F (tx) Pr{X/t ∈ [0, x]c }

lim = lim = µ̃([0, x]c ) (2.9)
t→∞ 1 − F (t1) t→∞ Pr{X/t ∈ [0, 1]c }

for all continuous points x of µ̃, where µ̃([0, x]c ) = c Sd−1 max1≤j ≤d (uj /xj )α S(du) for some
+
constant c and a probability measure S on Sd−1 + := {x ∈ Rd+ : ||x|| = 1}.

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 930 H. LI AND Y. SUN

There is a well-known relation (see Proposition 5.15 of [19] for details) between MRV
distributions and multivariate extreme value (MEV) distributions with identical Fréchet margins
H (x; 1) := exp{−x −1 } for x > 0. In general, the margins of an MEV distribution are expressed
in terms of the generalized extreme value family,
H (x; γ ) := exp{−(max{1 + γ x, 0})−1/γ }, x ∈ R, γ ∈ R.
Note, however, that the parametric feature enjoyed by the univariate EV distributions vanishes
in the multivariate context.
Theorem 2.3. Consider a nonnegative random vector X = (X1 , . . . , Xd ) with MRV DF F
and continuous margins F1 , . . . , Fd . Let C F and µ respectively denote the copula and the
intensity measure of F . If the margins are tail equivalent (i.e. F̄i (x)/F¯j (x) → 1 as x → ∞
for any i  = j ) then the upper tail dependence function b∗ (·; C F ) exists and
 −1/α  −1/α c
∗ µ( di=1 [wi , ∞]) ∗ µ(( di=1 [0, wi ]) )
1. b (w; C ) = F
d−1
, a (w; C F
) = d−1 c
,
µ([1, ∞] × R+ ) µ(([0, 1] × R+ ) )
µ([w, ∞]) b∗ ((w1−α , . . . , wd−α ); C F ) µ([0, w]c ) a ∗ ((w1−α , . . . , wd−α ); C F )
2. = , = .
µ([0, 1]c ) a ∗ ((1, . . . , 1); C F ) µ([0, 1]c ) a ∗ ((1, . . . , 1); C F )
Proof. Since each Fi , i ∈ N , is regularly varying, from (1.2) and (1.3), we have F̄i (x) =
Li (x)/x α for x ≥ 0. To estimate F̄i−1 (wi u) when u → 0+ for fixed wi > 0, consider
1/α
Li (wi x)
= wi−1 F̄i (x)gi (wi , x),
1/α
F̄i (wi x) =
wi x α
1/α
where gi (wi , x) := Li (wi x)/Li (x) → 1 as x → ∞. Substituting F̄i (x) = wi u into the
above expression and taking F̄i−1 on both sides, we obtain
−1/α
F̄i−1 (wi u) = wi F̄i−1 (ugi (wi , F̄i−1 (wi u))).
−1/α
Asymptotically, F̄i−1 (wi u) ≈ wi F̄i−1 (u) as u → 0+ .
For any fixed w = (w1 , . . . , wd ) with wi > 0, i ∈ N , consider
Pr{Fi (Xi ) > 1 − wi u for all i ∈ N }
b∗ (w; C F ) = lim
u→0+ Pr{F1 (X1 ) > 1 − u}
−1
Pr{Xi > F̄i (wi u) for all i ∈ N }
= lim −1
u→0+ Pr{X1 > F¯1 (u)}
−1/α
Pr{Xi > wi F̄i−1 (ugi (wi , F̄i−1 (wi u))) for all i ∈ N }
= lim −1
. (2.10)
u→0+ Pr{X1 > F¯1 (u)}
−1 −1
Since F̄i (x)/F¯1 (x) → 1 as x → ∞, we have from [20, Proposition 2.6] F̄i (u)/F¯1 (u) → 1
as u → 0+ . For any small ε > 0, when u is sufficiently small,
1 − ε < gi (wi , F̄i−1 (wi u)) < 1 + ε for all i ∈ N.
Thus, when u is sufficiently small,
F̄i−1 (u(1 − ε)) F̄i−1 (ugi (wi , F̄i−1 (wi u))) F̄i−1 (u(1 + ε))
−1
≥ −1
≥ −1
. (2.11)
F¯1 (u) F¯1 (u) F¯1 (u)

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 Tail dependence for heavy-tailed scale mixtures of multivariate distributions 931

−1
Since F¯1 (u) is regularly varying at 0 (see, e.g. Proposition 2.6 of [20]—this result is stated
in [20] in terms of increasing regularly varying functions, but is easily verified to be true for
−1 −1
decreasing regularly varying functions) or, more precisely, F¯1 (uc)/F¯1 (u) → c−1/α as
u → 0+ for any c > 0, then by taking the limits in (2.11) we have

F̄i−1 (ugi (wi , F̄i−1 (wi u)))

(1 − ε)−1/α ≥ −1
≥ (1 + ε)−1/α as u → 0+
F¯1 (u)
for any small ε > 0. That is, when u is sufficiently small,
−1
(1 − ε)−1/α F¯1 (u)
≥ F̄i−1 (ugi (wi , F̄i−1 (wi u)))
−1
≥ (1 + ε)−1/α F¯1 (u) for all ε > 0, i ∈ N.

Combining these inequalities with (2.10), we have, for any ε > 0,

−1/α −1
Pr{Xi > wi (1 − ε)−1/α F¯1 (u) for all i ∈ N }
lim −1
u→0+ Pr{X1 > F¯1 (u)}
∗
≤ b (w; C ) F

−1/α −1
Pr{Xi > wi (1 + ε)−1/α F¯1 (u) for all i ∈ N }
≤ lim −1
,
u→0+ Pr{X1 > F¯1 (u)}
−1
which implies that, after substituting t = F¯1 (u),
−1/α
Pr{Xi > wi t, i ∈ N }
b∗ (w; C F ) = lim . (2.12)
t→∞ Pr{X1 > t}
d −1/α d−1
Set A = i=1 [wi , ∞] and B = [1, ∞] × R+ . Then (2.7) implies that
 −1/α
∗ µ(A) µ( di=1 [wi , ∞])
b (w; C ) = F
= d−1
.
µ(B) µ([1, ∞] × R+ )

Similar to (2.10), it follows from (2.4) that

Pr{Fi (Xi ) > 1 − wi u there exists i ∈ N }
a ∗ (w; C F ) = lim
u→0+ Pr{F1 (X1 ) > 1 − u}
−1/α
Pr{Xi > wi F̄i−1 (ugi (wi , F̄i−1 (wi u))) there exists i ∈ N }
= lim −1
.
u→0+ Pr{X1 > F¯1 (u)}
Borrowing from the derivation of (2.12), then using (2.7), we have
−1/α  −1/α c
∗ Pr{Xi > wi t there exists i ∈ N } µ(( di=1 [0, wi ]) )
a (w; C ) = lim
F
= d−1
.
t→∞ Pr{X1 > t} µ([1, ∞] × R+ )

The expressions in part 2 can be easily obtained from part 1.

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 932 H. LI AND Y. SUN

Observe that the rescaled intensity measure µ̃(B) = µ(B)/µ([0, 1]c ) in part 2 of Theo-
rem 2.3 (also see (2.8) and (2.9)) satisfies that µ̃(([0, 1])c ) = 1. The intensity measures µ and
µ̃ are uniquely determined from each other.
In general, F̄i (x)/F¯j (x) → rij as x → ∞ for any i  = j [20, p. 174]. If 0 < rij < ∞
then Theorem 2.3 still holds by properly adjusting the marginal scaling constants. If rij = 0 or
rij = ∞, then some margins have heavier tails than others and more subtle separate marginal
scalings are needed to derive the limiting results.

3. Tail dependence of heavy-tailed scale mixtures of multivariate distributions

As shown in Theorem 2.3, the upper tail dependence function represents a rescaled version
of the intensity measure. The link presented in Theorem 2.3 allows people, on the one hand, to
develop, via tail dependence functions, tractable parametric models for the intensity measure,
and, on the other hand, to derive, via multivariate regular variation, tail dependence functions
in the situations where copulas have neither explicit analytic expressions nor closure properties
for their conditional distributions. We illustrate the latter part in this section for the random
vectors of the form (1.1).
Let Ti+ := Ti ∨ 0, i ∈ N, and assume in this section that 0 < E(Ti+ α+ε
) < ∞ for some
ε > 0. Since R is regularly varying with survival function (1.2), then, by Breiman’s theorem
(see, e.g. Proposition 7.5 of [20]), RTi+ , i ∈ N , is also regularly varying with survival function
Pr{RTi+ > r} = E(Ti+ α )L (r)/r α , where L (r)/L (r) → 1 as r → ∞ for i  = j . Consider
i i j

(Y1 , . . . , Yd ) := D[(E(T1+
α
))−1/α T1+ , . . . , (E(Td+
α
))−1/α Td+ ](R, . . . , R) , (3.1)

where D[a1 , . . . , ad ] is a d ×d diagonal matrix with main diagonal entries a1 , . . . , ad (here and
hereafter ‘ ’ denotes the matrix transpose). Since Yi is strictly increasing in Xi ≥ 0, i ∈ N ,
(Y1 , . . . , Yd ) in (3.1) and (X1 , . . . , Xd ) in (1.1) have the same upper tail dependence function.
It follows from Proposition A.1 of [2] that (Y1 , . . . , Yd ) is regularly varying, and it is also easy
to verify that Pr{Yi > r}/ Pr{Yj > r} → 1 as r → ∞. Thus, by Theorem 2.3, the expression
of the upper tail dependence function of (X1 , . . . , Xd ) boils down to the determination of the
intensity measure of (Y1 , . . . , Yd ) in (3.1).
Proposition A.1 of [2] presents a general formula for the intensity measure of a random
affine transform of a random vector that is regularly varying. Applying this formula to (3.1),
d
the intensity measure µ of (3.1) is given by, for any Borel measurable subset B ⊆ R+ ,
−1 −1
µ(B) = E(ν(D[(E(T1+
α
))1/α T1+ α
, . . . , (E(Td+ ))1/α Td+ ](B))), (3.2)

where ν is the intensity measure of (R, . . . , R). Using (2.9), we have, for any nonnegative
w = (w1 , . . . , wd ),

Pr{R > t di=1 wi }  −α
d
ν([0, w]c ) = lim = wi .
t→∞ Pr{R > t}
i=1
 |S|−1
 
Using the inclusion–exclusion relation and the fact that ∅=S⊆N (−1) i∈S wi = i∈N wi
for all nonnegative w1 , . . . , wd , we also have


d
ν([w, ∞]) = wi−α .
i=1

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 Tail dependence for heavy-tailed scale mixtures of multivariate distributions 933

Substituting these two expressions into (3.2), then


d
wi−α Ti+
α 
d
wi−α Ti+
α
µ([0, w]c ) = E α) and µ([w, ∞]) = E α) ,
E(Ti+ E(Ti+
i=1 i=1
which lead to the expression of the upper tail dependence function.
In summary, we give the following theorem.
Theorem 3.1. Consider a random vector X = (RT1 , . . . , RTd ) in the form of (1.1) with DF F
α+ε
and continuous margins F1 , . . . , Fd . Assume that 0 < E(Ti+ ) < ∞, i ∈ N , for some ε > 0.
Then the upper tail dependence and exponent functions are given by

d α
wi Ti+ 
d α
wi Ti+
b∗ (w; C F ) = E α) and a ∗ (w; C F ) = E α) .
E(Ti+ E(Ti+
i=1 i=1
Applying Theorem 3.1 to (2.3) and (2.5), we obtain the following corollary.
Corollary 3.1. Making the same assumptions as Theorem 3.1, the upper tail and extremal
dependence parameters are given by
 α ))−1 T α )  α ))−1 T α )
E( di=1 (E(Ti+ i+ E( di=1 (E(Ti+ i+
τJ =  α ))−1 T α ) for ∅  = J ⊆ N and γ =  .
E( i∈J (E(Ti+ E( d
(E(T α ))−1 T α )
i+ i=1 i+ i+
We illustrate our main results using multivariate elliptical distributions. Let
be a d × d
positive semidefinite matrix, and let U = (U1 , . . . , Um ) be uniformly distributed on the
unit sphere (with respect to Euclidean distance) in Rm . Consider the following stochastic
representation:
(X1 , . . . , Xd ) = (µ1 , . . . , µd ) + RA(U1 , . . . , Um ) ,
where A is a d × m matrix with AA =
and R is a nonnegative random variable independent
of U . The distribution of (X1 , . . . , Xd ) is known as an elliptical contoured distribution and
is one of most widely used radially symmetric multivariate distributions. The examples of
elliptical distributions include the multivariate normal, t-, and logistic distributions. We refer
the reader to [6, Chapter 2] for details on properties of elliptical distributions.
Let (T1 , . . . , Td ) = A(U1 , . . . , Um ) . Then we have
(X1 , . . . , Xd ) = (µ1 , . . . , µd ) + R(T1 , . . . , Td ),
which is a scale mixture of (T1 , . . . , Td ). The random vector (X1 , . . . , Xd ) is regularly varying
if and only if R has a regularly varying right tail [10], [21]. A formula for the bivariate
tail dependence parameter has also been established in [10] and [21], and the upper tail
dependence function of a multivariate elliptical distribution was explicitly derived in [13].
The multivariate upper tail dependence and exponent functions, and the upper tail and extremal
dependence parameters for multivariate elliptical distributions are now given by Theorem 3.1
and Corollary 3.1. We discuss the bivariate case in detail. Let
2
σ11 σ12 σ11 0

= and A = ,
σ12 2
σ22 σ12 /σ11 σ22 1 − ρ 2
where ρ = σ12 /σ11 σ22 . Thus,
σ12
T1 = σ11 U1 , T2 = U1 + σ22 1 − ρ 2 U2 .
σ11

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 934 H. LI AND Y. SUN

Observe that, marginally, RT1 and RT2 have one-dimensional elliptical distributions, and
−1 −1
σ11 RT1 and σ22 RT2 have the same distribution (see [6, Chapter 2]). Assume that R has
a regularly varying right tail with heavy tail index α > 0. Since (X1 , X2 ) and (RT1 , RT2 ) have
the same tail dependence parameter, we have
−1 −1
τ1 = lim Pr{σ22 RT2 > t | σ11 RT1 > t}.
t→∞

Obviously, (T1 , T2 ) has a bounded support; thus, it follows from Corollary 3.1 that
−1 −1
E(σ11 T1+ ∧ σ22 T2+ )α E(U1+ ∧ (ρU1 + 1 − ρ 2 U2 )+ )α
τ1 = −1
= , (3.3)
E(σ11 T1+ )α E(U1+ )α

where U1+ := U1 ∨ 0 and (ρU1 + 1 − ρ 2 U2 )+ := (ρU1 + 1 − ρ 2 U2 ) ∨ 0.

If ρ = 1 then, trivially, τ1 = 1. Suppose that ρ < 1. Transferring to polar coordinates, we
have

U1 = cos , U2 = sin , and ρU1 + 1 − ρ 2 U2 = sin( + θ0 ),

where is uniformly distributed on [0, 2π] and θ0 = tan−1 (ρ/ 1 − ρ 2 ). Clearly,

 π/2  π/2  1
1 1 1 uα
α
E(U1+ )= cosα θ dθ = cosα θ dθ = √ du.
2π −π/2 π 0 π 0 1 − u2
To calculate the quantity in the numerator of (3.3), consider the boundary case when

u1 = ρu1 + 1 − ρ 2 u2 ,

which leads to the solution θ1 = tan−1 (u2 /u1 ) = tan−1 ((1 − ρ)/ 1 − ρ 2 ). That is,

cos θ ≥ sin(θ + θ0 ) ≥ 0 if − θ0 ≤ θ ≤ θ1 ,
0 ≤ cos θ ≤ sin(θ + θ0 ) if θ1 ≤ θ ≤ π/2.

Simple trigonometric arguments show that θ0 + 2θ1 = π/2. Thus,

 θ1  π/2
1 1
E(U1+ ∧ (ρU1 + 1 − ρ 2 U 2 )+ )α = sinα (θ + θ0 ) dθ + cosα θ dθ
2π −θ0 2π θ1
 π/2−θ1  π/2
1 1
= sinα θ dθ + cosα θ dθ.
2π 0 2π θ1
√
Since cos θ1 = 1 − ρ 2 / 2 − 2ρ = ((1 + ρ)/2)1/2 , we obtain
 ((1+ρ)/2)1/2
1 uα
E(U1+ ∧ (ρU1 + 1 − ρ 2 U 2 )+ )α = √ du.
π 0 1 − u2
Hence,
 ((1+ρ)/2)1/2  1
uα uα
τ1 = √ du √ du,
0 1 − u2 0 1 − u2
which is identical to the one obtained in [10] and [21].

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 Tail dependence for heavy-tailed scale mixtures of multivariate distributions 935

An intriguing issue is whether or not these tail dependence parameters are monotone in
response to changes in the heavy tail index α. It was demonstrated by the numerical results in
[21] that the tail dependence parameter τ1 is decreasing in α. We show that this is indeed the
case using the ratio-of-moments expressions obtained in Corollary 3.1.
A nonnegative random variable X is said to be smaller than a nonnegative random variable
Y in the hazard rate ordering, denoted by X ≤hr Y , if the hazard rate of X is larger than that
of Y . A detailed discussion on the hazard rate ordering can be found, for example, in Chapter 1
of [22], from which, the following result also holds.
Lemma 3.1. Let X and Y be nonnegative random variables. If X ≤hr Y then E Xα /E Y α is
decreasing in α.
Proof. Theorem 1.B.12 of [22] states that if X ≤hr Y then
E g2 (X) E g1 (Y ) ≤ E g1 (X) E g2 (Y ) (3.4)
for all nonnegative real functions g1 and g2 satisfying the conditions that g1 (·) is increasing
and g2 (·)/g1 (·) is increasing. Consider the situation in which α1 ≤ α2 , g1 (x) = x α1 , and
g2 (x) = x α2 for x ≥ 0. Then (3.4) reduces to
E Xα2 E Y α1 ≤ E Xα1 E Y α2 .
The monotonicity of E Xα /E Y α follows.
Proposition 3.1. Let RTi be the ith component of (1.1), where Ti and R satisfy the regularity
α ) = E(T α ), i  = j .
conditions specified in Theorem 3.1. Suppose that E(Ti+ j+
d 
1. If i=1 Ti+ ≤hr i∈J Ti+ then τJ is decreasing in α.
 
2. If di=1 Ti+ ≤hr di=1 Ti+ then γ is decreasing in α.
α ) = E(T α ), i  = j , then, from Corollary 3.1,
Proof. If E(Ti+ j+
d 
E( i=1 Ti+ )α E( di=1 Ti+ )α
τJ =  for all ∅  = J ⊂ N and γ = d .
E( i∈J Ti+ )α E( i=1 Ti+ )α
The monotone properties follow from Lemma 3.1.
Note that inequality (3.4) resembles the property of total positivity of order 2, and, in fact,
Proposition 3.1 can be established directly by using the theory of total positivity.
To show that τ1 of a bivariate elliptical distribution is decreasing in α, we need, by virtue
of Lemma 3.1, to establish that U1+ ≥hr U1+ ∧ (ρU1 + 1 − ρ 2 U2 )+ . From [22, Condi-
tion 1.B.3, p. 16], it is sufficient to show that
Pr{U1+ > t}
is increasing in t ∈ [0, s], (3.5)
Pr{U1+ ∧ (ρU1 + 1 − ρ 2 U2 )+ > t}
where s ≤ 1 is the right endpoint of the support of U1+ ∧ (ρU1 + 1 − ρ 2 U2 )+ . Again, using
the polar coordinate system, we have, for any t ∈ [0, s],
cos−1 t
Pr{U1+ > t} = ,
π
(cos−1 t − sin−1 t)+
Pr{U1+ ∧ (ρU1 + 1 − ρ 2 U2 )+ > t} = .
2π

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 936 H. LI AND Y. SUN

It is easy to verify that 21 (cos−1 t/(cos−1 t − sin−1 t)) is increasing in t ∈ [0, s], and then (3.5)
follows. Therefore, τ1 is decreasing in α.

4. Concluding remarks
In this paper we have developed a general method based on multivariate regular variation to
derive tail dependence parameters for heavy-tailed scale mixtures of multivariate distributions.
Our method properly rescales the marginal distributions so that, asymptotically, the scaled
random variables have the same tail marginal distributions; thus, the calculation of their tail
dependence avoids taking the marginal transforms on the entire distribution. Our tail analysis
leads to tractable expressions of tail dependence parameters which depend on joint moments of
the random variables being mixed and the heavy tail index of the mixing random variable. Our
method also establishes the link between multivariate tail dependence parameters and the theory
of multivariate regular variation. The structural properties of multivariate tail dependence, such
as monotonicity of tail dependence with respect to structural parameters of the distribution, have
also been discussed for bivariate elliptical distributions. Yet, a general theory remains to be
developed.

Acknowledgement
The authors would like to sincerely thank an anonymous referee for his/her insightful
comments, which led to significant improvements in the results, contents, and presentation
of this paper.

References
[1] Basrak, B., Davis, R. A. and Mikosch, T. (2002). A characterization of multivariate regular variation. Ann.
Appl. Prob. 12, 908–920.
[2] Basrak, B., Davis, R. A. and Mikosch, T. (2002). Regular variation of GARCH processes. Stoch. Process.
Appl. 99, 95–115.
[3] Bingham, N. H., Goldie, C. M. and Teugels, J. L. (1987). Regular Variation. Cambridge University Press.
[4] Buhl, C., Reich, C. and Wegmann, P. (2002). Extremal dependence between return risk and liquidity risk: an
analysis for the Swiss Market. Tech. Rep., Department of Finance, University of Basel.
[5] Embrechts, P., Lindskog, F. and McNeil, A. (2003). Modeling dependence with copulas and applications to
risk management. In Handbook of Heavy Tailed Distributions in Finance, ed. S. Rachev, Elsevier, pp. 329–384.
[6] Fang, K. T., Kotz, S. and Ng, K. W. (1990). Symmetric Multivariate and Related Distributions. Chapman and
Hall, London.
[7] Finkenstädt, B. and Rootzén, H. (2004). Extreme Values in Finance, Telecommunications, and the
Environment. Chapman and Hall/CRC, New York.
[8] Frahm, G. (2006). On the extreme dependence coefficient of multivariate distributions. Statist. Prob. Lett. 76,
1470–1481.
[9] Hartmann, P., Straetmans, S., and de Vries, C. (2006). Banking system stability. A cross-Atlantic
perspective. In The Risks of Financial Institutions, University of Chicago Press, pp. 133–192.
[10] Hult, H. and Lindskog, F. (2002). Multivariate extremes, aggregation and dependence in elliptical distributions.
Adv. Appl. Prob. 34, 587–608.
[11] Joe, H. (1997). Multivariate Models and Dependence Concepts. Chapman and Hall, London.
[12] Joe, H., Li, H. and Nikoloulopoulos, A. K. (2009). Tail dependence functions and vine copulas. To appear
in J. Multivariate Anal.
[13] Klüppelberg, C., Kuhn, G. and Peng, L. (2008). Semi-parametric models for the multivariate tail dependence
function—the asymptotically dependent. Scand. J. Statist. 35, 701–718.
[14] Li, H. (2008). Tail dependence comparison of survival Marshall–Olkin copulas. Methodology Comput. Appl.
Prob. 10, 39–54.
[15] Li, H. (2009). Orthant tail dependence of multivariate extreme value distributions. J. Multivariate Anal. 100,
243–256.
[16] McNeil, A. J., Frey, R., Embrechts, P. (2005). Quantitative Risk Management. Princeton University Press.

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press

 Tail dependence for heavy-tailed scale mixtures of multivariate distributions 937

[17] Nelsen, R. B. (2006). An Introduction to Copulas, 2nd edn. Springer, New York.
[18] Nikoloulopoulos, A. K., Joe, H. and Li, H. (2009). Extreme value properties of multivariate t copulas.
Extremes 12, 129–148.
[19] Resnick, S. I. (1987). Extreme Values, Regular Variation, and Point Processes. Springer, New York.
[20] Resnick, S. I. (2007). Heavy-Tail Phenomena. Springer, New York.
[21] Schmidt, R. (2002). Tail dependence for elliptically contoured distributions. Math. Meth. Operat. Res. 55,
301–327.
[22] Shaked, M. and Shanthikumar, J. G. (2007). Stochastic Orders. Springer, New York.
[23] Sklar, A. (1959). Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Statist. Univ. Paris 8,
229–231.

https://doi.org/10.1239/jap/1261670680 Published online by Cambridge University Press