1. Introduction
Given a bivariate random vector
with joint distribution function
and marginal distribution functions
and
the random variable
describes the distance between
and
in a sense that depends on the dependence structure of the vector. Different structures assign different meanings to this random variable and lead, obviously, to different ways of computing the expectation
The distance can be applied to random variables with identical and non-identical distribution functions, and we consider both cases. If
then
and
are copies of the same random variable
X with distribution function
, and
reveals information about
An example is the case of independent and identically distributed random variables, in which
is the Gini’s mean difference of
a well-known measure of variability (see, for example, [
1]). We show in this work that, when
and
are dependent copies of the same random variable
is still a measure of variability of
X. A purpose of this paper is to study the properties of this functional in a general setting, where
and
are not necessarily independent.
If
and
are independent (or, more generally, if they are linked by a symmetric dependence structure),
treats symmetrically the events
and
However, sometimes, it is convenient to use a characteristic of proximity by treating them differently (in finance, for example, an investor evaluates differently gains and losses). The random excess of
over
where
denotes the positive part of
is useful if we are interested in measuring the extent to which one random variable exceeds the other, rather than the distance between them in a bidirectional sense. Note that the absolute value
can be split into two terms, each describing the excess of one random variable over the other, as follows:
If and are copies of the same variable then also reveals information about For example, if and are independent, is Gini’s mean semidifference.
In general, the functional
where
is a non-negative real function, has been largely studied in mathematics literature, mainly in the context of the Monge–Kantorovich problem (see [
2], and references therein). The interest on this and other functionals used to measure the degree of difference between two random quantities goes back at least to the 1930s and the important contributions by Gini (see [
3]) and Hoeffding [
4]. Given two random vectors
and
another purpose of this paper is to find conditions under which
where
is a subset of increasing real functions. Different choices of
give rise to different stochastic orderings between
and
This problem was addressed in [
5,
6,
7] for the case where
and
have independent components with the same marginal distribution functions (see
Section 2 below for details). Here, we are concerned with two random vectors
and
whose components are not necessarily independent nor are they required to have identical distribution functions. In this case, we explore conditions under which
where
and
are the usual stochastic order and the increasing convex order, respectively (these orders will be defined in
Section 2 below).
This work is organized as follows.
Section 2 contains preliminaries, such as definitions and background for the stochastic orders and dependence notions used in this paper, as well as a review of the properties that a variability measure should satisfy. In
Section 3, given a random variable
X with distribution function
F, we show that any functional of the form
where
and
are two copies of
X with any type of dependence structure, is a measure of variability of
More generally, the distribution function
of
is allowed to be a distortion of
F (we will explain the meaning of this below). In
Section 4, given two random vectors
and
we obtain conditions (both in terms of the marginals and the copulas) to make comparisons of the form (
2).
Section 5 contains two applications. In
Section 5.1, we define a general class of premium principles based on the class of variability measures studied in
Section 3. In
Section 5.2, in the context of portfolio risk management, we assess the inclusion of a new asset in a portfolio by using the results obtained in
Section 4. Finally,
Section 6 contains conclusions.
Throughout this paper, given two random vectors and we denote by and the respective marginal distribution functions. Given any other random variable we denote by its distribution function.
2. Preliminaries
Let
be a random vector with joint distribution function
and marginal distribution functions
and
respectively. According to the Sklar theorem, the joint distribution
can be written as
where
C is the copula of the random vector
that is, the joint distribution function of the vector-copula
(see [
8]). If
and
are continuous, then
C is unique. The copula contains the information about the structure of dependency of the random vector
For every copula
C and every
on
it is well-known that
where the copulas
and
are the Fréchet–Hoeffding bounds. Random variables with copula
M are called comonotonic and random variables with copula
W are called countermonotonic.
The motivation for the study of the properties of
where
and
are not necessarily independent, comes from the fact that some probability metrics and measures of variability that are sometimes better known under other expressions, take this form for different copulas between
and
To give some examples, note that
If
and
are two copies of a random variable
X with distribution function
(
4) becomes
If
and
are two independent copies of
then
is the Gini’s mean difference (GMD) of
a well-known index of variability (see, for example, [
1]). If
and
are comonotonic, then (see [
9] or [
10])
which is the Wasserstein distance, a well-known characteristic of proximity of two random variables (see [
11]). If
and
are countermonotonic, then
(see, for example, [
2]). It is easy to see that, if
and
are two copies of
(
7) can be rewritten as
where
is the median of
This measure is twice the median absolute deviation (MAD), another popular measure of variability.
In view of the above examples, it is natural to ask whether
where
and
are two copies of
X with a copula
fulfills the requirements to be considered as a measure of variability of
Recall that a measure of variability
is a map from the set of random variables to
, such that given a random variable
quantifies the variability of
Next, we list a set of properties that a measure of variability should reasonably satisfy (see, for example, [
12] and references therein):
- (P0)
Law invariance: if X and Y have the same distribution, then
- (P1)
Translation invariance: for all X and all constant k.
- (P2)
Positive homogeneity: and for all X and all .
- (P3)
Non-negativity: for all with if X is degenerated at .
Bickel and Lehmann [
13] also require
to be consistent with the dispersive order. Recall that two random variables
X and
Y are ordered in the dispersive order if the difference between any two quantiles of
X is smaller than the corresponding quantiles of
Y, where the quantile function of a random variable
X with distribution function
F is defined by
The formal definition is as follows.
Definition 1. Given two random variables X and Y with distribution functions F and G, respectively, we say that X is smaller than Y in the dispersive order (denoted by if for all
A functional
satisfying properties (P0) to (P3) is said to be a measure of variability or spread in the sense of Bickel and Lehmann if it satisfies in addition (see [
13]):
- (P4)
Consistency with dispersive order: if , then .
A measure of variability in the sense of Bickel and Lehmann considers the variability or spread of a random variable throughout its distribution. Sometimes, however, there is an interest in measuring only the variability of
X along the right tail of its distribution (in risk theory, for example, some popular measures focus on the variability of a risk
X beyond the value at risk). When this is the case, the requirement on
to be consistent with the dispersive order is too strong. A natural weaker requirement is to be consistent with the excess wealth order (see [
14]), which is defined as follows.
Definition 2. Given two random variables X and Y with distribution functions F and G, respectively, we say that X is smaller than Y in the excess wealth order (denoted by if where and are the tail (or survival) functions of X and respectively.
This allows us to consider the following property.
- (P5)
Consistency with excess wealth order: if , then .
Measures of variability have received great attention in the actuarial and financial literature (see [
12,
15,
16,
17,
18], among others). In actuarial science, for example, a variability measure sometimes is combined with a location measure to build a premium principle (see [
19]). For particular applications in this context, we may wish
to satisfy the following properties:
- (P6)
Comonotonic additivity: if X and Y are comonotonic, then
- (P7)
Subadditivity: for all X and
Furman et al. [
12] say that
is a coherent measure of variability if it satisfies (P0)–(P3) and (P7).
Next, we recall some other notions used in this paper. The sequence of inequalities (
3) induces the following definition (see [
8]).
Definition 3. Given two copulas C and we say that C is smaller than in the concordance order (and write ) if for all .
Obviously, for every copula The name of this order is due to the fact that some measures of concordance, such as Kendall’s tau and Spearman’s rho, are increasing with respect to
In
Section 4 and
Section 5, we will make use of the following stochastic orders. The reader may consult the books [
20,
21,
22] for properties and applications.
Definition 4. Let X and Y be two random variables with distribution functions F and G and finite expectations and respectively. Then, X is said to be smaller than
- (i)
in the usual stochastic order (denoted by if for all
- (ii)
in the increasing convex order (denoted by if
- (iii)
in the convex order (denoted by if and
- (iv)
in the increasing concave order (denoted by ) if
It can be shown that
(respectively
) if and only if
for all increasing (respectively convex, increasing convex, increasing concave) functions
for which the expectations exist. When
and
are independent copies of
and
respectively, it is well-known (see [
5] and [
6]) that
implies
implies
The result for the convex order was extended to the so-called s-convex order in [
7]. In
Section 3 and
Section 4, we extend these results to the case where
and
are not necessarily independent from (nor are they required to have identical distribution functions as)
and
respectively. For this, we need the following notions (see [
23,
24]).
Definition 5. Let be a random vector.
- (i)
We say that is stochastically increasing in denoted by , if is a nondecreasing function of for all
- (ii)
We say that is positively dependent through stochastic ordering (PDS) if and
Intuitively, if
is PDS, then its components are more likely simultaneously to have large values, compared with a vector of independent random variables with the same marginal distributions. For relationships between this and other dependence notions see, for example, Table 2 in [
25]. The negative dependence analog of Definition 5 is as follows (see [
24]).
Definition 6. Let be a random vector.
- (i)
We say that is stochastically decreasing in denoted by , if is a nonincreasing function of for all
- (ii)
We say that is negatively dependent through stochastic ordering (NDS) if and
Intuitively, if
is NDS, one component of the vector will tend to be large when the other component is small. It is easy to see that a random vector
with continuous marginals is PDS (respectively, NDS) if and only if
is componentwise concave (respectively, convex). It is also well-known (see [
26]) that a continuous random vector
with copula
C has the property PDS (resp. NDS) if and only if its copula
C is PDS (resp. NDS).
3. A Family of Measures of Variability
Let
and
be two random variables with respective continuous distribution functions
and
and finite expectations. If
and
are two independent copies of
it is well-known (see [
13]) that
is a measure of variability in the sense of Bickel and Lehmann (that is, it satisfies properties (P0) to (P4)). Let
h be a distortion function, that is, a non-decreasing function from
to
such that
and
(given two distribution functions
F and
if
we say that
G is a distortion of
F via
h). In this section, we show that any functional of the form
where
and
is a measure of variability of
In particular, if
h is the identity function (
for all
) and
and
have a NDS copula, this measure satisfies all the properties (P1 to P7) listed above.
The following theorem extends a result of [
5], stated as Theorem 3.B.42 in the book [
20], in two directions: first, we consider two random vectors with the same copula instead of two random vectors with independent components; and, second, we allow the first marginal of each vector to be a distortion of the other (via the same
h) instead of taking two copies of the same random variable.
Theorem 7. Let X and Y be two random variables with distribution functions F and respectively and let h be a distortion function. Let be a random vector with respective marginal distribution functions and Similarly, let be a random vector with marginal distribution functions and respectively. Suppose that and have the same copula If , then
Proof. Since the dispersive order is preserved by distortion functions (Theorem 13 in [
27]), we have
and
Since
and
have the same copula, it follows from Definition 2.1 in [
28] and Theorem 1 in [
29] that there exists a function
that maps stochastically
into
, i.e.,
, defined as
where
,
, is an increasing function that satisfies
It follows from the assumptions that
for all
x. Therefore,
where the first and second equality in (
9) follow from the fact that
and
, respectively. The inequality follows from (
8) by using Theorem 1.A.1 in [
20]. □
By taking in Theorem 7, we have the following corollary.
Corollary 8. Let and be two copies of and respectively, such that and have the same copula. If , then
Remark 9. Given two random vectors and with the same copula, the condition is equivalent to say that the bivariate random vectors and are ordered in a multivariate dispersion sense, see [30]. Now, we can prove the following result.
Theorem 10. Let X be a random variable with strictly increasing distribution function F and let h be a strictly increasing distortion function. Let and be two random variables with copula C and marginal distribution functions and respectively. Let
- (i)
If and then is a comonotonic additive measure of variability in the sense of Bickel and Lehmann, that is, it satisfies properties (P0)–(P4) and (P6).
- (ii)
If and the copula C is NDS, then satisfies all the properties (P0) to (P7).
Proof. We first prove (i). Let
C be the copula of
and
. From (
4), we have
Clearly,
if
X is degenerated at
. This, together with the fact that
is non-decreasing,
for all
k and
for all
(see [
31]), ensures that
satisfies properties (P0) to (P3). Since, given two random variables
and
the condition
implies that
property (P4) (consistency of
with respect to the dispersive order) is a direct consequence of Theorem 7. Property (P6) follows from the fact that, if
and
are comonotonic, then
for all
(see [
32]). Under the assumptions in (ii), we have
Standard arguments show that
Therefore, integrating (
11) by parts, we get
Since
C is componentwise convex, (P5) follows from Theorem 8. (ii) in [
33] and (P7) follows from Theorem 2.1 in [
12]. □
Example 11. Two functionals satisfying the assumptions of part (i) are and which is the Wasserstein distance between F and its distortion a variability measure introduced by [34]. Note that where and is the Fréchet–Hoeffding upper bound copula (see (6)). Example 12. Using (7), it follows from Theorem 10 (ii) that where is the median of satisfies all the properties (P0) to (P7) listed above. This measure can be written in the form where and where C is the Fréchet–Hoeffding lower bound copula (see (7) and the paragraph below it), which is an example of NDS copula (see [35] for this and other examples of NDS copulas). 4. Other Stochastic Comparisons
To begin this section, we consider two random vectors and with the same marginals. Denote by the space of bidimensional random vectors with marginal distribution functions and
Theorem 13. Let and be two random vectors with copulas C and respectively. If , then:
- (i)
- (ii)
Proof. Under the assumptions, it follows from Theorem 4 of [
36] that
This means that
for all increasing convex
. Since
is increasing and convex for any increasing convex function
it follows that
for all increasing and convex
which proves (i). The proof of (ii) is similar using the function
□
Remark 14. An alternative proof of Theorem 13 can be given by using Theorem 1 in [37], which provides conditions to ensure, under the above assumptions, that for certain classes of functions k. The proof is based on proving that the functions and with ϕ increasing and convex, satisfy those conditions. A more general type of comparison can be made between two random vectors with possibly different (but stochastically ordered) marginals. The following two results provide conditions to compare two random excesses. The first result is given in terms of the usual stochastic order and the second result in terms of the increasing convex order.
Theorem 15. Let be a random vector with respective marginal distribution functions and Similarly, let be a random vector with marginal distribution functions and respectively. If and have the same copula and , then
Proof. Since for all for all and increases in the first argument (since is the distribution function of the random variable ()). Therefore, for all which ends the proof. □
Theorem 16. Let and be two random vectors with copulas C and , and marginal distribution functions and respectively. If C is NDS, , and then
Proof. Let us consider a vector
with copula
C and such that
for
. From the assumptions, it follows that
which is equivalent to say that
(Theorem 4.A.1 in [
20]). Since
and
have the same copula
the random vectors
and
have the same copula
. Moreover, since
C is NDS (that is, componentwise convex), then
is PDS (that is, componentwise concave). It follows from Corollary 2.7 in [
38] that
Since
it follows from Theorem 13 that
The result follows by using the fact that the increasing convex order is transitive and is preserved by the increasing convex transformation
(see Theorem 4.A.8(a) in [
20]). □
Remark 17. In particular, Theorem 16 holds when and have the same NDS copula In this case, and imply
Lemma 18. Let X and Y be two random variables that are symmetric about 0. Then:
- (i)
If , then
- (ii)
If , then
Proof. Let
and
be the tail functions of
and
respectively. If
X and
Y are symmetric about 0, it is easy to see that
for all
where
h is the concave distortion function
Now (i) and (ii) follow, respectively, from Theorem 2.6 (i) and Theorem 2.6 (v) in [
39]. □
The following result follows immediately from Theorem 15 and Lemma 18.
Corollary 19. Let and be two random vectors with the same copula C and with marginal distribution functions and respectively. If and then
The following result is also an immediate corollary of Theorem 16 and Lemma 18.
Corollary 20. Let and be two random vectors with symmetric copulas C and , and marginal distribution functions and respectively. If C is NDS, , and then
Remark 21. In particular, Corollary 20 holds when and have the same symmetric NDS copula When this is the case, and imply
Since the independence copula is both NDS and PDS, a particular case of Corollaries 19 and 20 is the following.
Corollary 22. Let and be two random vectors with independent components and with marginal distribution functions and respectively.
- (i)
If and then
- (ii)
If and then
The following corollaries extend Lemma 2.2 in [
6] from the case of two random vectors with independent components to the case of two random vectors with the same symmetric NDS copula.
Corollary 23. Let and be two random vectors with the same symmetric NDS copula C and with marginal distribution functions and respectively. If and then
Proof. The assumption
implies
Since
holds if and only if
(Theorem 3.A.12 in [
20]), it follows
This is equivalent to write
therefore, the result follows from Corollary 20. □
Corollary 24. Let and be two random vectors with the same symmetric NDS copula, such that and , all variables having finite means. If , then
Proof. It is well-known (see (3.C.7) in [
20]) that
implies
. The result follows from Corollary 23. □