A - Z of Copulas

An introduction to Copulas
Carlo Sempi
Dipartimento di Matematica Ennio De Giorgi
Universit del Salento
Lecce, Italy
carlo.sempi@unisalento.it
The 33rd Finnish Summer School on Probability Theory and
Statistics, June 6th10th, 2011
C. Sempi An introduction to Copulas. Tampere, June 2011.
Outline
1
Historical Introduction
2
Preliminaries
3
Copul
4
Sklars theorem
5
Copul and stochastic measures
The beginning of the story
The history of copulas may be said to begin with Frchet (1951).
Frchets problem: given the distribution functions F
j
(j = 1, 2, . . . , d) of d r.v.s X
1
, X
2
, . . . , X
d
dened on the same
probability space (, T, P), what can be said about the set
(F
1
, F
2
, . . . , F
d
) of the ddimensional d.f.s whose marginals are
the given F
j
?
H (F
1
, . . . , F
d
) H(+, . . . , +, t, +, . . . , +) = F
j
(t)
The set (F
1
, . . . , F
d
) is called the Frchet class of the F
j
s.
Notice (F
1
, . . . , F
d
) ,= since, if X
1
, X
2
, . . . , X
d
are independent,
then
H(x
1
, x
2
, . . . , x
d
) =
d
j =1
F
j
(x
j
).
But, it was not clear which the other elements of (F
1
, . . . , F
d
)
were.
Bibliography1
For Frchets work see, e.g.,
M. Frchet, Sur les tableaux de corrlation dont les marges
sont donns, Ann. Univ. Lyon, Science, 4, 1384 (1951)
G. DallAglio, Frchet classes and compatibility of distribution
functions, Symposia Math., 9, 131150 (1972)
In this latter paper DallAglio studies under which conditions there
is just one d.f. belonging to (F
1
, F
2
).
Enters Sklar
In 1959, Sklar obtained the most important result in this respect,
by introducing the notion, and the name, of a copula, and proving
the theorem that now bears his name.
Correspondence with Frchet
He and Bert Schweizer had been making progress in their work on
statistical metric spaces, to the extent that Menger suggested it
would be worthwhile to communicate their results to Frchet.
Frchet was interested, and asked to write an announcement for
the Comptes Rendus. This lead to an exchange of letters between
Sklar and Frchet, in the course of which Frchet sent Sklar several
packets of reprints, mainly dealing with the work he and his
colleagues were doing on distributions with given marginals. These
reprints were important for much of the subsequent work. At the
time, though, the most signicant reprint for Sklar was that of
Fron (1956).
Sklar2
Fron, in studying three-dimensional distributions had introduced
auxiliary functions, dened on the unit cube, that connected such
distributions with their one-dimensional margins. Sklar saw that
similar functions could be dened on the unit dcube for all d 2
and would similarly serve to link ddimensional distributions to their
onedimensional margins. Having worked out the basic properties
of these functions, he wrote about them to Frchet, in English.
Sklar3
Frchet asked Sklar to write a note about them in French. While
writing this, Sklar decided he needed a name for these functions.
Knowing the word copula as a grammatical term for a word or
expression that links a subject and predicate, he felt that this would
make an appropriate name for a function that links a
multidimensional distribution to its one-dimensional margins, and
used it as such. Frchet received Sklars note, corrected one
mathematical statement, made some minor corrections to Sklars
French, and had the note published by the Statistical Institute of
the University of Paris (Sklar, 1959).
A curiosity
Curiously, it should be noted that in that paper, the author Abe
Sklar is named as M. Sklar (should it be intended as
Monsieur?)
Lack of a proof
The proof of Sklars theorem was not given in (Sklar, 1959), but a
sketch of it was provided in (Sklar, 1973). (see also (Schweizer &
Sklar, 1974)), so that for a few years practitioners in the eld had
to reconstruct it relying on the handwritten notes by Sklar himself;
this was the case, for instance, of the present speaker. It should be
also mentioned that some indirect proofs of Sklars theorem
(without mentioning copula) were later discovered by Moore &
Spruill and Deheuvels.
For about 15 years, all the results concerning copulas were obtained
in the framework of the theory of Probabilistic Metric spaces
(Schweizer & Sklar, 1974). The event that arose the interest of the
statistical community in copulas occurred in the mid seventies,
when Bert Schweizer, in his own words (Schweizer, 2007),
quite by accident, reread a paper by A. Rnyi, entitled
On measures of dependence and realized that [he] could
easily construct such measures by using copulas.
The rst building blocks were the announcement by Schweizer &
Wol in the Comptes Rendus de lAcadmie des Sciences (1976)
and Wols Ph.D. Dissertation at the University of Massachusetts
at Amherst (1977). These results were presented to the statistical
community in (Schweizer & Wollf, 1981) (see also (Wol, 1980)).
However, for several other years, Chapter 6 of the 1983 book by
Schweizer & Sklar, devoted to the theory of Probabilistic metric
spaces, was the main source of basic information on copulas. Again
in Schweizers words from (Schweizer, 2007),
After the publication of these articles and of the book
. . . the pace quickened as more . . . students and colleagues
became involved. Moreover, since interest in questions of
statistical dependence was increasing, others came to the
subject from dierent directions. In 1986 the enticingly
entitled article The joy of copulas by C. Genest and R.C
MacKay (1986), attracted more attention.
Finance
At end of the nineties, the notion of copulas became increasingly
popular. Two books about copulas appeared and were to become
the standard references for the following decade. In 1997 Joe
published his book on multivariate models, with a great part
devoted to copulas and families of copulas. In 1999 Nelsen
published the rst edition of his introduction to copulas (reprinted
with some new results in 2006).
But, the main reason of this increased interest has to be found in
the discovery of the notion of copulas by researchers in several
applied eld, like nance. Here we should like briey to describe this
explosion by quoting Embrechtss comments (Embrechts, 2009).
Embrechts
. . . the notion of copula is both natural as well as easy
for looking at multivariate d.f.s. But why do we witness
such an incredible growth in papers published starting the
end of the nineties (recall, the concept goes back to the
fties and even earlier, but not under that name)? Here I
can give three reasons: nance, nance, nance. In the
eighties and nineties we experienced an explosive
development of quantitative risk management
methodology within nance and insurance, a lot of which
was driven by either new regulatory guidelines or the
development of new products . . . . Two papers more than
any others put the re to the fuse: the . . . 1998 RiskLab
report (Embrechts et al., 2002) and at around the same
time, the Li credit portfolio model (Li, 2001).
Today
The advent of copulas in nance originated a wealth of
investigations about copulas and, especially, applications of copulas.
At the same time, dierent elds like hydrology discovered the
importance of this concept for constructing more exible
multivariate models. Nowadays, it is near to impossible to give a
complete account of all the applications of copulas to the many
elds where they have be used.
Since the eld is still in eri, it is important from time to time to
survey the progresses that have been achieved, and the new
questions that they pose. The aim of this talk is to survey the
recent literature.
Today2
To quote Schweizer again:
The era of i.i.d. is over: and when dependence is
taken seriously, copulas naturally come into play. It
remains for the statistical community at large to recognize
this fact. And when every statistics text contains a
section or a chapter on copulas, the subject will have
come of age.
Preliminaries
Random variables and vectors
When a r.v. X = (X
1
, X
2
, . . . , X
d
) is given, two problems are
interesting:
to study the probabilistic behaviour of each one of its
components;
to investigate the relationship among them.
It will be seen how copulas allow to answer the second one of these
problems in an admirable and thorough way.
It is a general fact that in probability theory, theorems are proved in
the probability space (, T, P), while computations are usually
carried out in the measurable space (R
d
, B(R
d
)) endowed with the
law of the random vector X.
Preliminaries
Distribution functions
The study of the law P
X
is made easier by the knowledge of the
distribution function(=d.f.), as dened here.
Given a random vector X = (X
1
, X
2
, . . . , X
d
) on the probability
space (, T, P), its distribution function F
X
: R
d
I is dened by
F
X
(x
1
, x
2
, . . . , x
d
) = P
_
d
i =1
X
i
x
i
_
(1)
if all the x
i
s are in R, while:
F
X
(x
1
, x
2
, . . . , x
d
) = 0, if at least one of the arguments equals
F
X
(+, +, . . . , +) = 1.
Preliminaries
Cvolume
A dbox is a cartesian product
[a, b] =
d
j =1
[a
j
, b
j
],
where, for every index j 1, 2, . . . , d, 0 a
j
b
j
1.
For a function C : I
d
I, the Cvolume V
C
of the box [a, b] is
dened via
V
C
([a, b]) :=
v
sign(v) C(v)
where the sum is carried over all the 2
d
vertices v of the box [a, b];
here
sign(v) =
_
1, if v
j
= a
j
for an even number of indices,
1, if v
j
= a
j
for an odd number of indices.
Preliminaries
Properties of distribution functions
Theorem
The d.f. F
X
of the r.v. X = (X
1
, X
2
, . . . , X
d
) has the following
properties:
F is isotone, i.e. F(x) F(y) for all x, y R
d
, x y;
for all (x
1
, . . . , x
i 1
, x
i +1
, . . . , x
d
) R
d1
, the function
R t F
X
(x
1
, . . . , x
i 1
, t, x
i +1
, . . . , x
d
)
is rightcontinuous;
for every dbox [a, b], V
F
X
([a, b]) 0.
Preliminaries
Marginals
Let F be a ddimensional d.f. (d 2). Let = (j
1
, . . . , j
m
) a
subvector of (1, 2, . . . , d), 1 m d 1. We call marginal of F
the d.f. F
: R
m
I dened by setting d m arguments of F
equal to +, namely, for all x
1
, . . . , x
m
R,
F
(x
1
, . . . , x
m
) = F(y
1
, . . . , y
d
),
where, for every j 1, 2, . . . , d, y
j
= x
j
if j j
1
, . . . , j
m
, and
y
j
= + otherwise.
In particular, when = j , F
(j )
is usually called 1dimensional
marginal and it is denoted by F
j
.
If F is the d.f. of the r.v.

X = (X
1
, X
2
, . . . , X
d
), then the
marginal of F is the d.f. of the subvector (X
j
1
, . . . , X
j
m
).
Copul
The denition
Denition
For d 2, a ddimensional copula (shortly, a dcopula) is a
dvariate d.f. on I
d
whose univariate marginals are uniformly
distributed on I.
Each d-copula may be associated with a r.v. U = (U
1
, U
2
, . . . , U
d
)
such that U
i
|(I) for every i 1, 2, . . . , d and U C.
Conversely, any r.v. whose components are uniformly distributed on
I is distributed according to some copula.
The class of all dcopulas will be denoted by (
d
.
Copul
A characterization
Theorem
A function C : I
d
I is a copula if, and only if, the following
properties hold:
for every j 1, 2, . . . , d, C(u) = u
j
when all the
components of u are equal to 1 with the exception of the j th
one that is equal to u
j
I;
C is isotonic, i.e. C(u) C(v) for all u, v I
d
such that
u v;
C is dincreasing.
Copul
The special case d = 2
Explicitly, a bivariate copula is a function C : I
2
I such that
u [0, 1] C(u, 0) = C(0, u) = 0
u [0, 1] C(u, 1) = C(1, u) = u
for all u, u
, v, v
in I with u u
and v v
C(u
, v
) C(u
, v) C(u, v
) + C(u, v) 0
This last inequality is referred to as the rectangular inequality; a
function that satises it is said to be 2increasing.
Copul
Consequences
C(u) = 0 for every u I
d
having at least one of its
components equal to 0
(The 1Lipschitz property): for all u, v I
d
,
[C(u) C(v)[
d
i =1
[u
i
v
i
[.
(
d
is a compact set in the set C(I
d
, I) of all continuous
functions from I
d
into I equipped with the topology of
pointwise convergence.
Pointwise and uniform convergence are equivalent in (
d
.
Copul
Examples1
The independence copula
d
(u) = u
1
u
2
u
d
associated with
a random vector U = (U
1
, U
2
, . . . , U
d
) whose components are
independent and uniformly distributed on I.
The comonotonicity copula Min
d
(u) = minu
1
, u
2
, . . . , u
d
associated with a vector U = (U

1
, U
2
, . . . , U
d
) of r.v.s
uniformly distributed on I and such that U
1
= U
2
= = U
d
almost surely.
The countermonotonicity copula
W
2
(u
1
, u
2
) = maxu
1
+ u
2
1, 0 associated with a bivariate
vector U = (U
1
, U
2
) of r.v.s uniformly distributed on I and
such that U
1
= 1 U
2
almost surely.
Copul
Examples2: Convex combinations
Convex combinations of copulas: Let U
1
and U
2
be two
ddimensional r.v.s on (, T, P) distributed according to the
copulas C
1
and C
2
, respectively. Let Z be a Bernoulli r.v. such that
P(Z = 1) = and P(Z = 2) = 1 for some I. Suppose
that U
1
, U
2
and Z are independent. Now, consider the
ddimensional r.v. U
=
1
(Z) U
1
+
2
(Z) U
2
where, for i 1, 2,
i
(x) = 1, if x = i ,
i
(x) = 0, otherwise.
Then, U
is distributed according to the copula C

1
+ (1 ) C
2
.
Copul
Examples3
FrchetMardia family of copulas
C
FM
d
(u) =
d
(u) + (1 ) M
d
(u)
for every I. A convex sum of the copulas
d
and M
d
.
CuadrasAug family; for I,
C
CA
d
(u) = (
d
(u))
(M
d
(u))
1
,
Copul
The derivatives
Consider a bivariate copula C (
2
. For every v I, the functions
I t C(t, v)
I t C(v, t)
are increasing; therefore, their rst derivatives exists almost
everywhere with respect to Lebesgue measure and are positive,
where they exist. Because of the Lipschitz condition, they are also
bounded above by 1
0 D
1
C(s, t) 1 0 D
2
C(s, t) 1 a.e.
where
D
1
C(s, t) :=
C(s, t)
s
and D
2
C(s, t) :=
C(s, t)
t
Copul
A useful formula
The following integrationbyparts formula is sometimes useful in
the computation of statistical quantities.
Theorem
Let A and B be 2copul, and let the function : I R be
continuously dierentiable, i.e., C
1
. Then
_
[0,1]
2
AdB =
_
1
0
(t) dt
_
[0,1]
2
(A) D
1
AD
2
B du dv
=
_
1
0
(t) dt
_
[0,1]
2
(A) D
2
AD
1
B du dv
Copul
FrchetHoeding bounds
Theorem
For every C
d
(
d
and for every u I
d
,
W
d
(u) = max
_
d
i =1
u
i
d + 1, 0
_
C(u) M
d
(u).
These bounds are sharp:
inf
CC
d
C(u) = W
d
(u), sup
CC
d
C(u) = M
d
(u).
Notice that, while W
2
is a copula, W
d
is not a copula for d 3.
Copul
The marginals of a copula
A marginal of an dcopula C is obtained by setting some of its
argument equal 1. A kmarginal of C, k < d, is obtained by
setting exatly d k arguments equal to 1; therefore, there are
_
d
k
_
kmarginals of the dcopula C.
In particular, the d 1marginals are easily computed:
C(1, 1, . . . , 1, u
j
, 1, . . . , 1) = u
j
(j = 1, 2, . . . , d)
Sklars theorem
Sklars Theorem
Theorem
Given an ddimensional d.f. H there exists an dcopula C such
that for all (x
1
, x
2
, . . . , x
d
) R
n
H(x
1
, x
2
, . . . , x
d
) = C (F
1
(x
1
), F(x
2
), . . . , F
d
(x
d
)) (2)
The copula C is uniquely dened on

d
j =1
ran F
j
and is therefore
unique if all the marginals are continuous.
Conversely, if F
1
, F
2
,. . . , F
d
are d (1dimensional) d.f.s, then the
function H dened through eq. (2) is an ddimensional d.f..
Sklars theorem
How to obtain a copula from a joint d.f.
Given a dvariate d.f. F, one can derive a copula C. Specically,
when the marginals F
i
are continuous, C can be obtained by means
of the formula
C(u
1
, u
2
, . . . , u
d
) = F(F
1
1
(u
1
), F
1
2
(u
2
), . . . , F
1
d
(u
d
)),
where F
1
i
denoted the pseudoinverse of F
i
,
F
1
i
(s) = inft [ F
i
(t) s.
Thus, copul are essentially a way for transforming the r.v.
(X
1
, X
2
. . . , X
d
) into another r.v.
(U
1
, U
2
, . . . , U
d
) = (F
1
(X
1
), F
2
(X
2
), . . . , F
d
(X
d
))
having the margins uniform on I and preserving the dependence
among the components.
Sklars theorem
The uniqueness question
Sklars theorem immediately poses the question of the uniqueness
of the copula C:
If the r.v.s involved, or, equivalently, their d.f.s, are both
continuous, then the copula C is unique.
If at least one of the d.f.s has a discrete component, then the
copula C is uniquely dened only on the product of the ranges
ran F
1
ran F
2
ran F
d
, and there may well be more than
one copula extending C from this cartesian product to the whole
unit cube I
d
. In this latter case it is costumary to have recourse to
a procedure of bilinear interpolation in order to single out a unique
copula; this allow to speak of the copula of the pair (X, Y). See
Lemma 2.3.5 in (Nelsen, 2006) or (Darsow, Nguyen & Olsen, 1992)
Sklars theorem
Comments
Notice that in many papers where copul are applied there is
often hidden the assumption that the r.v.s involved are
continuous; this avoids the uniqueness question.
If all the d.f.s involved are continuous then to each joint d.f. in
the Frchet class (F
1
, F
2
, . . . , F
d
) there corresponds a unique
dcopula C (
d
; otherwise, to each H (F
1
, F
2
, . . . , F
d
)
there corresponds the set of copulas in (
d
that coincide on
d
j =1
ran F
j
Sklars theorem
Comments2
The second part of Sklars theorem is very easy to prove, but it is
extremely important for the applications; it is, in fact, the very
foundation of all the models built on copulas. Models are built
according to the following scheme:
the d rvs X
1
, X
2
, . . . , X
d
are individually described by their
1dimensional d.f.s F
1
, F
2
, . . . , F
d
then a copula C (
d
is introduced; this contains every piece
of information about the dependence relationship among the
r.v.s X
1
, X
2
, . . . , X
d
, independently of the choice of the
marginals F
1
, F
2
, . . . , F
d
.
In particular, copulas can serve for modelling situations where a
dierent distribution is needed for each marginal, providing a valid
alternative to several classical multivariate d.f.s such Gaussian,
Pareto, Gamma, etc.. This fact represents one of the main
advantage of the copulas idea.
Sklars theorem
Caution2
Sklars theorem should be used with some caution when the
margins have jumps. In fact, even if there exists a copula
representation for noncontinuous joint d.f.s, it is no longer
unique. In such cases, modelling and interpreting dependence
through copulas needs some caution. The interested readers should
refer to the paper (Marshall, 1996) and to the indepth discussion
by Genest and Nelehov (2007).
Sklars theorem
Survival copul
Sklars Theorem can be formulated in terms of survival functions
instead of d.f.s. Specically, given a r.v. X = (X
1
, X
2
, . . . X
d
) with
joint survival function F and univariate survival marginals F
i
(i = 1, 2, . . . , d), for all (x
1
, x
2
, . . . , x
n
) R
d
F(x
1
, x
2
, . . . , x
d
) =

C
_
F
1
(x
1
), F
2
(x
2
), . . . , F
d
(x
d
)
_
.
for some copula

C, usually called the survival copula of X (the
copula associated with the survival function of X).
Sklars theorem
Survival copul2
In particular, let C be the copula of X and let
U = (U
1
, U
2
, . . . , U
d
) be a vector such that U C. Then,
C(u) = C(1 u
1
, 1 u
2
, . . . , 1 u
d
),
where C(u) = P(U
1
> u
1
, U
2
> u
2
, . . . , U
d
> u
d
) is the survival
function associated with C, explicitly given by
C(u) = 1 +
d
k=1
(1)
k

1i
1
<i
2
<<i
k
n
C
i
1
i
2
i
k
(u
i
1
, u
i
2
, . . . , u
i
k
),
with C
i
1
i
2
,i
k
denoting the marginal of C related to (i
1
, i
2
, , i
k
).
Sklars theorem
Singular and absolutely continuous components
For simplicitys sake, we consider here only the case d = 2.
Every copula C (
2
may be expressed in the form
C = C
ac
+ C
s
where C
ac
is absolutely continuous and C
s
is singular.
For an absolutely continuous copula C one has a density c such that
C(u, v) =
_
I
2
c(s, t) ds dt =
_
1
0
ds
_
1
0
c(s, t) dt
The density c is found by dierentiation
c(u, v) = D
1
D
2
C(u, v) =

2
C(u, v)
u v
a.e.
Sklars theorem
Singular and absolutely continuous components2
The presence of a singular component in a copula often causes
analytical diculties. Nevertheless, there are specic applications in
which this presence is actually a useful feature; for instance, in
default models described by two random variables X and Y, the
fact that the event X = Y may have nonzero probability
implies, on the one hand, the existence of a singular component in
their copula, and, on the other hand, the possibility of joint defaults
of X and Y.
Sklars theorem
A special case
Notice, however, that, as a consequence of the Lipschitz condition,
for every copula C (
2
and for every v I, both functions
t C(t, v) and t C(v, t) are absolutely continuous so that
C(t, v) =
_
t
0
c
1,v
(s) ds and C(v, t) =
_
t
0
c
2,v
(s) ds
This latter representation has so far found no application.
Notice also that it possible to prove that, for a 2copula C,
D
1
D
2
C = D
2
D
1
C a.e.
Sklars theorem
Examples1
Both the copul W
2
and M
2
are singular:
M
2
uniformly spreads the probability mass on the main
diagonal v = u (u I) of the unit square;
W
2
uniformly spreads the probability mass on the opposite
diagonal v = u (u I) of the unit square.
The product copula
2
(u, v) := u v is absolutely continuous and its
density is given by
(u, v) = 1
I
2(u, v)
Sklars theorem
Rankinvariant property
Theorem
Let X = (X
1
, . . . , X
d
) be a r.v. with continuous d.f. F, univariate
marginals F
1
, F
2
,. . . , F
d
, and copula C. Let T
1
,. . . ,T
d
be strictly
increasing functions from R to R. Then C is also the copula of the
r.v. (T
1
(X
1
), . . . , T
d
(X
d
)).
the study of rank statistics insofar as it is the study
of properties invariant under such transformations may
be characterized as the study of copulas and
copula-invariant properties.
(Schweizer & Wol, 1981)
Sklars theorem
Independence
Theorem
Let (X
1
, X
2
, . . . , X
d
) be a r.v. with continuous joint d.f. F and
univariate marginals F
1
,. . . , F
d
. Then the copula of (X
1
, . . . , X
d
) is
d
if, and only if, X
1
,. . . , X
d
are independent.
Sklars theorem
Comonotonicity and countermonotonicity
Theorem
Let (X
1
, X
2
, . . . , X
d
) be a r.v. with continuous joint d.f. F and
univariate marginals F
1
,. . . ,F
d
. Then the copula of (X
1
, . . . , X
d
) is
M
d
if, and only if, there exists a r.v. Z and increasing functions
T
1
,. . . ,T
d
such that X = (T
1
(Z), . . . , T
d
(Z)) almost surely.
Theorem
Let (X
1
, X
2
) be a r.v. with continuous d.f. F and univariate
marginals F
1
, F
2
. Then (X
1
, X
2
) has copula W
2
if, and only if, for
some strictly decreasing function T, X
2
= T(X
1
) almost surely.
Sklars theorem
Stochastic measures
Denition
A measure on the measurable space (I
d
, B(I
d
)) will said to be
stochastic if, for every Borel set A and for every j 1, 2, . . . , d,
(I I
. .
j 1
A I I) = (A),
where denotes the (restriction to B(I) of the) Lebesgue measure.
Theorem
Every copula C (
d
induces a stochastic measure
C
on the
measurable space (I
d
, B(I
d
)) dened on the rectangles R = [a, b]
contained in I
d
, by
C
(R) := V
C
([a, b]) .
Conversely, to every stochastic measure on (I
d
, B(I
d
)) there
corresponds a unique copula C
(
d
dened by
C
(u) := ([0, u]) .

Markov operators
Denition
Given two probability spaces (
1
, T
2
, P
1
) and (
2
, T
2
, P
2
), a linear
operator T : L
(
1
) L
(
2
) is said to be a Markov operator if
T is positive, viz. Tf 0 whenever f 0;
T1 = 1 (here 1 denotes the constant function f 1);
E
2
(Tf ) = E
1
(f ) for every function f L
(
1
) (E
j
denotes
the expectation in the probability space (
j
, T
j
, P
j
) (j = 1, 2))
Theorem
Every Markov operator T : L
(
1
) L
(
2
) has an extension to
a bounded operator T : L
p
(
1
) L
p
(
2
) for every p 1.
Copul and Markov operators
Theorem
For every copula C (
2
the operator T
C
dened on L
1
(I) via
(T
C
f ) (x) :=
d
dx
_
1
0
D
2
C(x, t) f (t) dt
is a Markov operator on L
(I).
Conversely, for every Markov operator T on L
1
(I) the function C
T
dened on I
2
via
C
T
(x, y) :=
_
x
0
_
T1
[0,y]
_
(s) ds
is a 2copula.
Examples
(T
W
2
f ) (x) = f (1 x)
(T
M
2
f ) (x) = f (x)
(T
2
f ) (x) =
_
1
0
f d
Theorem
For the adjoint (T
C
)
of the Markov operator T

C
in the space L
p
with p ]1, +[ one has (T
C
)
= T
C
T , where the transpose C
T
of the copula C is dened by C
T
(x, y) := C(y, x).
The extension to the case d > 2
For d > 2, consider the factorization I
d
= I
p
I
q
, where
d = p + q. While for d = 2 there is only one possible factorization,
p = 1 and q = 1, this factorization is not unique when d > 2.
Let C (
d
be given; it induces a probability measure
C
on
(I
d
, B(I
d
)). Denote the marginals of
C
on (I
p
, B(I
p
)) and on
(I
q
, B(I
q
)) by
p
and
q
, respectively.
Given a decomposition d = p + q, there is a unique Markov
operator T : L
(I
p
) L
(I
q
) associated with
C
and, hence,
with the copula C. Therefore, to every copula C (
d
there
correspond as many Markov operators as there are solutions in
natural numbers p and q of the Diophantine equation p + q = d.
Since the number of these solutions is d 1, there are d 1
possible dierent Markov operators corresponding to a dcopula
when d 3.
Carlo Sempi
Lecce, Italy
Outline
1
Copul and Measurepreserving transformations
2
Construction of copulas
3
Shues of Min
4
Archimedean copul
5
How many Archimedean copul are there?
6
Copul and Brownian motion
Measurepreserving transformations
(, T, ) and (
, T
, ) two measure spaces.

f :
is a measurepreserving transformations (=MPT) if

B T
f
1
(B) T
B T

_
f
1
(B)
_
= (B)
From now on (, T, ) = (
, T
, ) = (I, B(I), )
B(I) the Borel sets I
the (restriction) of Lebesgue measure to B(I).
Copul and MPTs
Theorem
If f
1
, f
2
,. . . , f
d
are MPTs, the function C
f
1
,f
2
,...,f
d
: I
n
I dened
by
C
f
1
,f
2
,...,f
d
(x
1
, x
2
, . . . , x
d
) :=
_
f
1
1
[0, x
1
] f
1
d
[0, x
d
]
_
is a copula. Conversely, for every dcopula C (
d
, there exist d
MPTs f
1
, f
2
,. . . f
d
such that
C = C
f
1
,f
2
,...,f
d
.
This representation is not unique: if is another MPT on I, then
C
f
1
,f
2
,...,f
d
= C
f
1
,f
2
,...,f
d
.
Special MPTs
A transformation f is said to be ergodic if, for all measurable sets A
and B, one has
lim
n+
1
n
n1
k=0
_
f
k
A B
_
= (A) (B);
f is said to be strongly mixing iIf f satises the stronger property
lim
n+
_
f
n
A B
_
= (A) (B)
Two corollaries
Corollary
If f is strongly mixing, then, for all x, y [0, 1],
lim
n+
C
f
n
,g
(x, y) = xy =
2
(x, y).
Corollary
If f is ergodic, then, for all x, y [0, 1],
lim
n+
1
n
n1
j =0
C
f
j
,g
(x, y) = xy =
2
(x, y).
Two examples
For the copula M
2
one has
_
f
1
[0, x] f
1
[0, y]
_
=
_
f
1
([0, x] [0, y])
_
= ([0, x] [0, y]) = minx, y = M
2
(x, y).
for every measurepreserving transformation f .
As for the copula W
2
, recall that it concentrates all the probability
mass uniformly on the the diagonal (t) = 1 t of the unit square.
In this case =
1
, so that
1
[0, x] [0, y]
_
= ([1 x, 1] [0, y])
=
_
0, if x 1 y,
x + y 1, if x > 1 y;
therefore
W
2
(x, y) =
_
1
[0, x] [0, y]
_
.
The independence copula
Theorem
Let f and g be measurepreserving transformations. The following
conditions are equivalent for C
f ,g
(
2
:
(a) C
f ,g
=
2
(b) f and g, when regarded as random variables on the standard
probability space (I, B(I), ), are independent.
Patchwork
An at most countable family (S
i
)
i I
of closed and connected
subsets of I
2
S
i
S
j
S
i
S
j
C a copula
a continuous function F
i
: S
i
I
2
that is isotone in each place
and agrees with C (called background) on the the boundary
S
i
of S
i
, namely F
i
(u, v) = C(u, v) for every (u, v) S
i
The function F : I
2
I
F(u, v) :=
_
F
i
(u, v) , (u, v) S
i
,
C(u, v) , elsewhere,
is called the patchwork of (F
i
)
i I
into C.
Patchwork copul
Theorem
Given the family (R
i
)
i I
of rectangles, for the patchwork of the
family (F
i
)
i I
into the copula C the following statements are
equivalent:
(a) F is a copula;
(b) for every i I , F
i
is 2increasing on R
i
and coincides with C
on the boundary R
i
of R
i
.
Ordinal sums
J be a nite or countable subset of the natural numbers N
(]a
k
, b
k
[)
kJ
be a family of subintervals of I indexed by J. It
is required that any two of them have at most an endpoint in
common.
(C
k
)
kJ
a family of copulas also indexed by J
Denition
The ordinal sum C of (C
k
)
kJ
with respect to family of intervals
(]a
k
, b
k
[)
kJ
is dened, for all u = (u
1
, u
2
) I
2
by
C(u, v) :=
_
a
k
+ (b
k
a
k
) C
k
_
ua
k
b
k
a
k
,
va
k
b
k
a
k
_
, (u, v) [a
k
, b
k
]
2
,
minu, v , elsewhere.
Ordinal sums2
Theorem
The ordinal sum of the family of copulas (C
k
)
kJ
with respect to
the family of intervals (]a
k
, b
k
[)
kJ
is a copula.
An ordinal sum is a special case of the construction of patchwork
copulas; it suces to choose
the copula M
2
as the background copula;
S
k
= ]a
k
, b
k
[ ]a
k
, b
k
[ for every k J;
for every k J, F
k
is a version of the copula C
k
rescaled in
such a way as to meet the requirements of a patchwork
W
2
ordinal sums
Theorem
Let C (
2
be a copula for which there exists x
0
]0, 1[ such that
C(x
0
, 1 x
0
) = 0. Then there exist two 2copul C
1
(
2
and
C
2
(
2
such that
C(u, v) =
_
_
x
0
C
1
_
u
x
0
,
x
0
+v1
x
0
_
(u, v) [0, x
0
] [1 x
0
, 1]
(1 x
0
) C
2
_
ux
0
1x
0
,
v
1x
0
_
, (u, v) [x
0
, 1] [0, 1 x
0
]
W
2
(u, v), elsewhere.
Shues of Min
Shues of Min
A copula is said to be a shue of Min it is obtained through the
following procedure:
the probability mass is placed on the support of the copula
M
2
, namely on the main diagonal of the unit square;
then the unit square is cut into a nite number of vertical
strips;
these vertical strips are permuted (shued) and, possibly,
some of them are ipped about their vertical axes of symmetry;
nally the vertical strips are reassembled to form the unit
square again;
to the probability mass thus obtained there corresponds a
unique copula C, which is a shue of Min.
Shues of Min were introduced in (Mikusiski et al. (1992)).
Shues of Min
Shues of Min
A dierent presentation
Two continuous random variables X and Y have a shue of Min C
as their copula is if, and only if, one of them is an invertible
piecewise linear function of the other one.
The set of Shues of Min is dense in (
2
.
Shues of Min
Density of the shues
Theorem
Let X and Y be continuous random variables on the same
probability space (, T, P), let F and G be their marginal d.f.s and
H their joint d.f.. Then, for every > 0 there exist two random
variables X
and Y
on the same probability space and a piecewise

linear function : R R such that
(a) Y
= X
(b) F
:= F
X
= F and G
:= F
Y
= G
(c) |H H
<
where H
is the joint d.f. of X
and Y
, and | |
denotes the
L
norm on R
2
.
Shues of Min
A surprising consequence
The last result has a surprising consequence. Let X and Y be
independent (and continuous) random variables on the same
probability space, let F and G be their marginal d.f.s and
H = F G their joint d.f.. Then, according to the previous
theorem, it is possible to construct two sequences (X
n
) and (Y
n
) of
random variables such that, for every n N, their joint d.f. H
n
approximates H to within 1/n in the L
norm, but Y
n
is almost
surely a (piecewise linear) function of X
n
.
Shues of Min
A generalization; preliminaries1
(, T, ) a measure space
(
1
, T
1
) a measurable space
:
1
a measurable function
T the set of all measurepreserving transformations of
(I, B(I), )
T
p
the set of all measure-preserving permutations
(automorphisms) of this space
image measure of under
(A) = ( )(A) =
_
1
A
_
(A T
1
)
T equipped with the composition of mappings is a semigroup and
T
p
is a subgroup of T .
Shues of Min
Interval exchange transformations
J
1,i
(i = 1, 2, . . . , n) partition of I into the nondegenerate
intervals J
1,i
= [a
1,i
, b
1,i
[ and the singleton J
1,n
= 1.
J
2,i
(i = 1, 2, . . . , n) another such partition such that,
(J
1,i
) = (J
2,i
)
the interval exchange transformation
T(x) =
_
_
x a
1,1
+ a
2,1
, if x J
1,i
,
((I
n
i =1
J
1,i
) [0, x]) +
n
i =1
(b
2,i
a
2,a
) 1
[a
2
,1]
(x)
otherwise,
Shues of Min
A mapping on I
2
Given T : I I dene S
T
: I
2
I
2
via
S
T
(u, v) := (T(u), v) . ((u, v) I
2
)
J a (possibly degenerate) interval in I
the (vertical) strip J I
the partition of the unit square I
2
into possibly innitely many,
vertical strips.
Shues of Min
Generalized shuing
A shuing of a strip partition J
i
I
i I
(card J
0
) is any
permutation S of the unit square such that
(1
Sh
) admits the representation S = S
T
for some T : I I
(2
Sh
) is measurepreserving on the space
_
I
2
, B(I
2
),
2
_
(3
Sh
) the restriction S[
J
i
I
of S to every strip J
i
I is continuous
with respect to the standard product topology on I
2
Shues of Min
Generalized shuing2
Intuitively, shuing is just a reordering of the strips. This feature is
captured by the condition (1
Sh
), which represents the shuing by a
single transformation T of the unit interval. In particular, S
T
is a
permutation of I
2
if, and only if, T is a permutation of I. Because
of (2
Sh
) the single strips maintain their measure after shuing.
Finally, condition (3
Sh
) is just a technical tool for ensuring that,
during shuing, the integrity of strips is preserved.
Shues of Min
Shues: the new characterization
Lemma
Consider the image measure of a doubly stochastic measure
under S
T
. Then the following statements are equivalent:
(a)
S
T
is doubly stochastic
(b) T is in T .
Theorem
The following statements are equivalent:
(a) a copula C (
2
is a shue of Min;
(b) there exists a piecewise continuous T T such that
C
=
M
2
S
1
T
Shues of Min
Shues: the new denition
Denition
A copula C (
2
is a generalized shue of Min if
C
=
M
2
S
1
T
for some T T . Such a shue of Min is denoted by M
T
.
In this denition, T is allowed to have countably many
discontinuity points, which is a quite natural generalization of the
original notion of shue of Min.
Shues of Min
Shuing an arbitrary copula
Denition
Let C (
2
be a copula. A copula A is a shue of C if there exists
T T such that
A
=
C
S
1
T
. In this case, A is also called the
Tshue of C and denoted by C
T
.
If a copula C is represented by means of two measurepreserving
transformations f and g, C
f ,g
, then
(C
f ,g
)
T
= C
Tf ,g
Shues of Min
Orbits
The mapping which assigns to every T T and to every copula
C (
2
the corresponding shue C
T
denes an action of the group
T on the set of all copulas. The orbit of a copula C with respect to
this action is the set T (C) = C
T
[ T T constituted by all
shues of C. The general theory of group actions guarantees that
the classes of type T (C) form a partition of the set of all copulas.
The orbit of a copula is exactly the collection of all its shues.
Theorem
For a copula C (
2
the following statements are equivalent:
(a) C =
2
;
(b) T (C) = C.
Shues of Min
More on shues
Theorem
If C (
2
is absolutely continuous then so are all its shues.
Theorem
Every copula C (
2
other than
2
has a nonexchangeable shue.
Theorem
For every copula C (
2
, the independence copula
2
can be
approximated uniformly by elements of T (C).
Archimedean copul
Generators
A function : R
+
I is said to be an (outer additive) generator if
it is continuous, decreasing and (0) = 1, lim
t+
(t) = 0 and is
strictly decreasing on [0, t
0
], where t
0
:= inft > 0 : (t) = 0. If
the function is invertible, or, equivalently, strictly decreasing on
R
+
, then the generator is said to be strict. If is strict, then
(t) > 0 for every t > 0 (and lim
t+
(t) = 0).
Archimedean copul
Archimedan copul
A copula C (
d
is said to be Achimedean if a generator exists
such that
C(u) =
_
(1)
(u
1
) +
(1)
(u
2
) + +
(1)
(u
d
)
_
u I
d
.
Such a copula will be denoted by C
When is strict the copula C
is said to be strict; in this case, C
has the representation

C
(u) =
_
1
(u
1
) + +
1
(u
d
)
_
.
Archimedean copul
dmonotone functions
A function f : ]a, b[ R is called dmonotone in ]a, b[, where
a < b + if
it is dierentiable up to order d 2;
for every x ]a, b[, its derivatives satisfy the inequalities
(1)
k
f
(k)
(x) 0, (k = 0, 1, . . . , d 2)
(1)
d2
f
(d2)
is decreasing and convex in ]a, b[
f is 2monotone function i it is decreasing and convex. If f has
derivatives of every order and if
(1)
k
f
(k)
(x) 0,
for every x ]a, b[ and for every k Z
+
is said to be completely
monotonic.
Archimedean copul
Characterization of Archimedean copulas
Theorem
(McNeil & Nelehov) Let : R
+
I be a generator. Then the
following statements are equivalent:
(a) is dmonotone on ]0, +[;
(b) C
(u) :=
_
(1)
(u
1
) + +
(1)
(u
d
)
_
is a dcopula.
Corollary
Let : R
+
I be a generator. Then the following statements are
equivalent:
(a) is completely monotone on ]0, +[
(b) C
: I
d
I is a dcopula for every d 2
Archimedean copul
Examples
The copula
2
is Archimedean: take (t) = e
t
; since
lim
t+
(t) = 0 and (t) > 0 for every t > 0, is strict; then
1
(t) = ln t and
1
(u) +
1
(v)
_
= exp ((ln u ln v)) = uv =
2
(u, v).
Also W
2
is Archimedean; take (t) := max1 t, 0. Since
(1) = 0, is not strict. Its quasiinverse is
(1)
(t) = 1 t.
On the contrary, the upper FrchetHoeding bound M
2
is not
Archimedean.
Archimedean copul
The GumbelHougaard family
C
GH
(u) = exp
_
_
_
d
i =1
(log(u
i
))
_
1/
_
_
where 1. For = 1 we obtain the independence copula as a
special case, and the limit of C
GH
for + is the
comonotonicity copula. The Archimedean generator of this family
is given by (t) = exp
_
t
1/
_
. Each member of this class is
absolutely continuous.
Archimedean copul
The MardiaTakahasiClayton family
The standard expression for members of this family of dcopulas is
C
MTC
(u, v) = max
_
_
_
_
d
i =1
u
i
(d 1)
_
1/
, 0
_
_
_
where
1
d1
, ,= 0. The limiting case = 0 corresponds to the
independence copula.
The Archimedean generator of this family is given by
(t) = (max1 +t, 0)

1/
.
For every ddimensional Archimedean copula C and for every
u I
d
, C
L
u C(u) for
L
=
1
d1
.
Archimedean copul
Franks family
C
Fr
(u) =
1
log
_
1 +
d
i =1
_
e
u
i
1
_
(e
= 1)
d1
_
,
where > 0. The limiting case = 0 corresponds to
d
. For the
case d = 2, the parameter can be extended also to the case
< 0.
Copulas of this type have been introduced by Frank in relation with
a problem about associative functions on I. They are absolutely
continuous.
The Archimedean generator is given by
(t) =
1
log
_
1 (1 e
) e
t
_
Archimedean copul
EFGM copul1
For d 2 let o be the class of all subsets of 1, 2, . . . , d having
at least 2 elements; o contains 2
d
d 1 elements. To each
S o, we associate a real number
S
, with the convention that,
when S = i
1
, i
2
, . . . , i
k
,
S
=
i
1
i
2
...i
k
.
An EFGM copula can be expressed in the following form:
C
EFGM
d
(u) =
d
i =1
u
i
_
_
1 +
SS
j S
(1 u
j
)
_
_
,
for suitable values of the
S
s.
For the bivariate case EFGM copul have the following expression:
C
EFGM
2
u
1
, u
2
= u
1
u
2
(1 +
12
(1 u
1
)(1 u
2
)) ,
Archimedean copul
EFGM copul2
EFGM copul are absolutely continuous with density
c
EFGM
d
(u) = 1 +
SS
j S
(1 2u
j
).
As a consequence, the parameters
S
s have to satisfy the
following inequality
1 +
SS
j S
j
0
for every
j
1, 1. In particular, [
S
[ 1.
A necessary detour: associativity
Denition
A binary operation T on I is said to be associative if, for all s, t
and u in I,
T (T(s, t), u) = T (s, T(t, u))
Denition
The Tpowers of an element t I under the associative function
T are dened recursively by
t
1
:= t and n N t
n+1
:= T (t
n
, t) ,
tnorms
Denition
A triangular norm, or, briey, a tnorm T is a function T : I
2
I
that is associative, commutative, isotone in each place, viz., both
the functions
I t T(t, s) and I t T(s, t)
are isotone for every s I and such that T(1, t) = t for every
t I.
Denition
A tnorm T is said to be Archimedean if, for all s and t in ]0, 1[,
there is n N such that s
n
< t.
Copul and tnorms
Theorem
For a tnorm T the following statements are equivalent:
(a) T is a 2copula;
(b) T satises the Lipschitz condition:
T(x
, y) T(x, y) x
x x, x
, y I x x
Theorem
For an Archimedean tnorm T, which has as an outer additive
generator, the following statements are equivalent:
(a) T is a 2copula;
(b) either or
(1)
is convex.
Two important concepts
Denition
An element a ]0, 1[ is said to be a nilpotent element of the
tnorm T if there exists n N such that a
(n)
T
= 0.
Denition
A tnorm T is said to be strict if it is continuous on I
2
and is
strictly increasing on ]0, 1[; it is said to be nilpotent if it is
continuous on I
2
and every a ]0, 1[ is nilpotent.
The tnorm
2
(u, v) := uv is strict, while
W
2
(u, v) := maxu + v 1, 0 is nilpotent.
a ]0, 1[ a
n
W
2
= maxna (n 1), 0,
so that a
n
W
2
= 0 for n 1/(1 a).
Representation of tnorms
Under mild conditions the tnorm T has the following
representation
T(x, y) =
_
(1)
(x) +
(1)
(y)
_
x, y I,
where : R
+
I is continuous, decreasing and (0) = 1, while
(1)
: I R
+
is a quasiinverse of that is continuous, strictly
decreasing on I and such that
(1)
(1) = 0
Isomorphisms of generators
: R
+
I an Archimedean generator
a stricly increasing bijection on I, in particular, (0) = 0 and
(1) = 1. Then is also a generator.
If T
is the Archimedean tnorm generated by the outer generator

, then, as is immediately checked, is the generator of the
tnorm
T
(u, v) = ( )
_
(1)

1
(u) +
(1)

1
(v)
_
=
_
T
1
(u),
1
(v)
__
.
Isomorphisms of generators2
Denition
Two generators
1
and
2
are said to be isomorphic if there exists
a strictly increasing bijection : I I such that
2
=
1
.
Two tnorms T
1
and T
2
are said to be isomorphic if there exists a
strictly increasing bijection : I I such that, for all u and v in I,
T
2
(u, v) =
_
T
1
_
1
(u),
1
(v)
__
.
Two results on tnorms
Theorem
For a function T : I
2
I, the following statements are equivalent:
(a) T is a strict tnorm;
(b) T is isomorphic to
2
.
Theorem
For a function T : I
2
I, the following statements are equivalent:
(a) T is a nilpotent tnorm;
(b) T is isomorphic to W
2
.
Isomorphisms for copulas1
Theorem
For an Archimedean 2copula C (
2
, the following statements are
equivalent:
(a) C is strict;
(b) C is isomorphic to
2
;
(c) every additive generator of C is isomorphic to
2
(t) = e
t
(t R
+
)
Isomorphisms for copulas2
Theorem
For an Archimedean 2copula C (
2
, the following statements are
equivalent:
(a) C is nilpotent;
(b) C is isomorphic to W
2
;
(c) every outer additive generator of C is isomorphic to
W
2
(t) = max1 t, 0 (t R
+
)
An example
The copula
C(u, v) :=
uv
u + v uv
usually denoted by /( ) in the literature is strict; its
generator is
(t) =
1
1 + t
(t R
+
).
The isomorphism with
2
is realized by the function : I I
dened by
(s) =
1
1 ln s
.
Brownian motion
In a probability space (, T, P) let B
(1)
t
: t 0 and
B
(2)
t
: t 0 be two Brownian motions (=BMs). We explicitly
assume that the BM is continuous and consider, for every t 0,
the random vector
B
t
:=
_
B
(1)
t
, B
(2)
t
_
Then B
t
: t 0 denes a stochastic process with values in R
2
.
The literature deals mainly with the independent case, viz., B
(1)
t
and B
(2)
t
are independent for every t 0; this is usually called the
twodimensional BM.
Distribution functions
For every t 0, let F
(1)
t
and F
(2)
t
be the (rightcontinuous)
distribution functions (=d.f.s) of B
(1)
t
and B
(2)
t
, respectively; thus,
for every x R,
F
(j )
t
(x) = P
_
B
(j )
t
x
_
(j = 1, 2).
Actually, For every t 0, F
(1)
t
(x) = F
(2)
t
(x) = (x/
t), where
is the d.f. of the standard normal distribution N(0, 1).
Coupled BM1
For every t 0, let C
t
, which depends on t, be the bivariate copula
of the random pair (B
(1)
t
, B
(2)
t
). Then the d.f. H
t
: R
2
I of the
random pair B
t
, is given, for all x and y in R, by
H
t
(x, y) = C
t
_
F
(1)
t
(x), F
(2)
t
(y)
_
.
Since both B
(1)
t
and B
(2)
t
are normally distributed the copula C
t
is
uniquely determined for every t 0.
Coupled BM2
Through an abuse of notation we shall write
B
t
:= C
t
_
B
(1)
t
, B
(2)
t
_
Notice that, in principle, a dierent copula is allowed for every
t 0. The process B
t
: t 0 will be called the 2dimensional
coupled Brownian motion.
The traditional twodimensional BM is included in the picture; in
order to recover it, it suces to choose the independence copula
2
(u, v) := u v ((u, v) I
2
) and set C
t
=
2
for every t 0
H
t
(x, y) = F
(1)
t
(x) F
(2)
t
(y) ((x, y) R
2
).
Properties to be studied
The (onedimensional) BM is the example of a stochastic process
that has three properties
it a Markov process;
it is a martingale in continuous time;
it is a Gaussian process.
These three aspects will be examined for a coupled BM.
The Markov property
Since the Markov property for a ddimensional process
X
t
: t 0 disregards the dependence relationship of its
components at every t 0, but is solely concerned with the
dependence structure of the random vector X
t
at dierent times,
the traditional proof for the ordinary (independent) BM holds for
the coupled BM B
t
:= C
t
(B
(1)
t
, B
(2)
t
) : t 0. Therefore,
Theorem
A coupled Brownian motion B
t
:= C
t
(B
(1)
t
, B
(2)
t
) : t 0 is a
Markov process.
The coupled BM is a martingale
Theorem
The coupled Brownian motion B
t
:= C
t
(B
(1)
t
, B
(2)
t
) : t 0 is a
martingale.
Gaussian processes
One has rst to state what is meant by the expression Gaussian
process when a stochastic process with values in R
2
is considered.
We shall adopt the following denition.
Denition
A stochastic process X
t
: t 0 with values in R
d
is said to be
Gaussian if, for every n N, and for every choice of n times
0 t
1
< t
2
< < t
n
, the random vector (X
t
1
, X
t
2
, . . . , X
t
n
) has a
(d n)dimensional normal distribution.
Is a coupled BM a Gaussian process?
Let the copula C
t
coincide, for every t 0, with M
2
, i.e.,
M
2
(u, v) = minu, v, u and v in I. Then
H
t
(x, y)
=
1
2 t
min
__
x
expv
2
/(2t) du,
_
y
expu
2
/(2t) dv
_
=
_
minx, y
t
_
.
A simple calculation shows that
2
H
t
(x, y)
x y
= 0 a.e.
with respect to the Lebesgue measure
2
, so that H
t
is not even
absolutely continuous.
Example2
If the copula C
t
is given, for every t 0, by W
2
, where
W
2
(u, v) := maxu + v 1, 0,
then the d.f. H
t
of B
t
is given by
H
t
(x, y) = max
_
_
x
t
_
+
_
y
t
_
1, 0
_
,
which again leads, after simple calculations, to the conclusion that,
again, B
t
is not even absolutely continuous.
Singular copul
The two previous examples represent extreme cases; in fact, since
the d.f.s involved are continuous, the copula of two random
variables is M
2
if, and only if, they are comonotone, namely, each
of them is an increasing function of the other, while their copula is
W
2
if, and only if, they are countermonotone, namely, each of them
is a decreasing function of the other. In this sense both examples
are the opposite of the independent case, which is characterized by
the copula
2
.
We recall that a copula can be either absolutely continuous or
singular or, again, a mixture of the two types. In general, if the
copula C is singular, namely the d.f. of a probability measure
concentrated on a subset of zero Lebesgue measure
2
in the unit
square I
2
, then also B
t
is singular.
The absolutely continuous case
Now let the copula C
t
be absolutely continuous with density c
t
; a
simple calculation shows that B
t
is absolutely continuous and that
its density is given a.e. by
h
t
(x, y) =
1
2 t
exp
_
x
2
+ y
2
2t
_
c
t
_
_
x
t
_
,
_
y
t
__
As a consequence, B
t
has a normal law if, and only if, c
t
(u, v) = 1
for almost all u and v in I; together with the boundary conditions,
this implies C
t
(u, v) = u v =
2
(u, v).
The special position of independence
Theorem
In a coupled Brownian motion
_
B
t
= C
t
_
B
(1)
t
, B
(2)
t
_
: t 0
_
,
B
t
has a normal law if, and only if, C
t
=
2
, viz., if, and only if,
its components B
(1)
t
and B
(2)
t
are independent.
Carlo Sempi
Lecce, Italy
Outline
1
Construction of copulas2
2
Copul and stochastic processes
3
Measures of dependence
4
Quasicopul
The product
Denition
Given two copulas A and B in C
2
, dene a map via
(A B)(x, y) :=
_
1
0
D
2
A(x, t) D
1
B(t, y) dt .
Theorem
For all copulas A and B, A B is a copula, namely A B C
2
, or,
equivalently, : C
2
C
2
C
2
.
The product2
Lemma
For every pair A and B of 2copulas, one has
T
A
T
B
= T
AB
.
Continuity in one variable
Theorem
Consider a sequence (A
n
)
nN
of copulas and a copula B. If the
sequence (A
n
) converges (uniformly) to A C, A
n
A then both
A
n
B
n+
A B and B A
n

n+
B A,
in other words the product is continuous in each place with
respect to the uniform convergence of copulas.
A consequence
Theorem
The binary operation is associative, viz.
A (B C) = (A B) C, for all 2copulas A, B, and C.
Corollary
The set of copulas endowed with the product, (C
2
, ) is a
semigroup with identity.
However. . .
. . . the product is not commutative, so that the semigroup (C
2
, )
is not abelian.
Let C
1/2
be the copula belonging to the CuadrasAug family,
dened by
C
1/2
(u, v) =
_
u
v, u v,
u v, u v.
(W
2
C
1/2
)
_
1
4
,
1
2
_
=
1
4

2
8
=
1
2

3
4
= (C
1/2
W
2
)
_
1
4
,
1
2
_
Special cases
2
C = C
2
=
2
,
M
2
C = C M
2
= C,
(W
2
C)(u, v) = v C(1 u, v),
(C W
2
)(u, v) = u C(u, 1 v).
In particular, one has W
2
W
2
= M
2
.
Theorem
The copul
2
and M
2
are the (right and left) annihilator and the
identity of the product, respectively.
Copul and Conditional Expectations
Theorem
Let C be the copula of the continuous random variables X and Y
dened on the probability space (, F, P); then, for almost every
,
E
_
1
{Xx}
| Y
_
() = D
2
C (F
X
(x), F
Y
(Y()))
and
E
_
1
{Yy}
| X
_
() = D
1
C (F
X
(X()), F
Y
(y)) .
An important consequence
Corollary
Let X, Y and Z be continuous random variables on the probability
space (, F, P). If X and Z are conditionally independent given Y,
then
C
XZ
= C
XY
C
YZ
.
product and Markov processes
Theorem
Let (X
t
)
tT
be a real stochastic process, let each random variable
X
t
be continuous for every t T and let C
st
denote the (unique)
copula of the random variables X
s
and X
t
(s, t T). Then the
following statements are equivalent:
(a) for all s, t, u in T,
C
st
= C
su
C
ut
;
(b) the transition probabilities P(s, x, t, A) := P(X
t
A | X
s
= x)
satisfy the ChapmanKolmogorov equations
P(s, x, t, A) =
_
R
P(u, , t, A) P(s, x, u, d)
for every Borel set A, for all s and t in T with s < t, for every
u ]s, t[ T and for almost all x R.
The product
The ChapmanKolmogorov equation is a necessary but not a
sucient condition for a Markov process. This motivates the
introduction of another operation on copulas.
Denition
Let A C
m
and B C
n
; the product of A and B is the mapping
A B : I
m+n1
I dened by
(A B)(u
1
, . . . , u
m+n1
)
:=
_
x
m
0
D
m
A(u
1
, . . . , u
m1
, ) D
1
B(, u
m+1
, . . . , u
m+n1
) d.
Properties of the star product
(a) for all copulas A C
m
and B C
n
the product A B is an
(m + n 1)copula, viz. : C
m
C
n
C
m+n1
(b) the product is continuous in each place: if the sequence
(A
k
)
kN
converges uniformly to A C
m
, then, for every
B C
n
one has both
A
k
B
k+
A B and B A
k

k+
B A
(c) the product is associative:
(A B) C = A (B C)
Characterization of Markov processes
Theorem
For a stochastic process (X
t
)
tT
such that each random variable
X
t
has a continuous distribution the following statements are
equivalent:
(a) (X
t
) is a Markov process;
(b) for every choice of n 2 and of t
1
, t
2
,. . . , t
n
in T such that
t
1
< t
2
< < t
n
C
t
1
,t
2
,...,t
n
= C
t
1
t
2
C
t
2
t
3
C
t
n1
t
n
,
where C
t
1
,t
2
,...,t
n
is the unique copula of the random vector
(X
t
1
, X
t
2
, . . . , X
t
n
) and C
t
j
t
j +1
is the (unique) copula of the
random variables X
t
j
and X
t
j +1
.
The role of the ChapmanKolmogorov equations
It is now possible to see from the standpoint of copulas why the
ChapmanKolmogorov equations alone do not garantee that a
process is Markov. One can construct a family of ncopulas with
the following two requirements:
they do not satisfy the conditions of the equations
C
t
1
,t
2
,...,t
n
= C
t
1
t
2
C
t
2
t
3
C
t
n1
t
n
they do satisfy the conditions of the equations
C
st
= C
su
C
ut
and are, as a consequence, compatible with the 2copulas of
a Markov process and, hence, with the ChapmanKolmogorov
equations.
Construction of the example
Consider a stochastic process (X
t
) in which the random variables
are pairwise independent. Thus the copula of every pair of random
variables X
s
and X
t
is given by
2
. Since,
2

2
=
2
, the
ChapmanKolmogorov equations are satised. It is now an easy
task to verify that for every n > 2, the nfold product of
2
yields
(
2

2

2
)(u
1
, u
2
, . . . , u
n
) =
n
(u
1
, u
2
, . . . , u
n
) ,
so that it follows that the only Markov process with pairwise
indedependent (continuous) random variables is one where all nite
subsets of random variables in the process are independent.
Construction of the example2
On the other hand, there are many 3copul whose 2marginals
coincide with
2
; such an instance is represented by the family of
copulas
C
(u
1
, u
2
, u
3
) :=
3
(u
1
, u
2
, u
3
)+u
1
(1u
1
) u
2
(1u
2
) u
3
(1u
3
) ,
for ]1, 1[. Now consider a process (X
t
) such that
three of its random variables, call them X
1
, X
2
and X
3
, have
C
as their copula;
every nite set not containing all three of X
1
, X
2
and X
3
is
made of independent random variables;
the ncopula (n > 3) of a nite set containing all three of
them is given by
C
t
1
,...,t
n
(u
1
, . . . , u
n
) = C
(u
1
, u
2
, u
3
)
n3
(u
4
, . . . , u
n
) ,
where we set
1
(t) := t.
Construction of the example3
Such a process exists since it is easily veried that the resulting
joint distribution satisfy the compatibility of Kolmogorovs
consistency theorem; this ensures the existence of a stochastic
process with the specied joint distributions. Since any two random
variables in this process are independent, the
ChapmanKolmogorov equations are satised. However, the copula
of X
1
, X
2
and X
3
is inconsistent with the set of equations with the
product, so that the process is not a Markov process.
A comparison
It is instructive to compare the traditional way of specifying a
Markov process with the one due to Darsow, Olsen and Nguyen. In
the traditional approach a Markov process is singled out by
specifying the initial distribution F
0
a family of transition
probabilities P(s, x, t, A) that satisfy the ChapmanKolmogorov
equations. Notice that in the classical approach, the transition
probabilities are xed, so that changing the initial distribution
simultaneously varies all the marginal distributions. In the present
approach, a Markov process is specied by giving all the marginal
distributions and a family of 2copulas that satises
C
st
= C
su
C
ut
As a consequence, holding the copulas of the process xed and
varying the initial distribution does not aect the other marginals.
Copul and Conditional expectations2
Denition
A copula C will be said to be idempotent (with respect to the
product) if
C C = C,
or, equivalently if, for all (u, v) I
2
, it satises the
integrodierential equation
C(u, v) =
_
1
0
D
2
C(u, t) D
1
C(t, v) dt.
Both the copul
2
and M
2
are idempotent.
Pfanzagls characterization
Theorem
Let H be a subset of L
1
(, F, P) such that f H
(f H, R), 1 +H H (f H), f g H (f , g H) and
such that if (f
n
)
nN
is a decreasing sequence of elements of H that
tends to a function f L
1
, then f H. Then an operator
T : H H is the restriction to H of a conditional expectation if,
and only if, (a) Tf Tg whenever f g (f , g H); (b)
T(f ) = Tf ( R, f H; (c) T(1 + f ) = 1 + Tf (f H),
(d) E(Tf ) = E(f ) (f H), (e) T
2
:= T T = T. when these
conditions are satised, then T = E
G
, where
G = {A F : T 1
A
= 1
A
} .
Idempotent copul and Markov operators
Theorem
A Markov operator T : L
(I) L
(I) is the restriction to L
(I)
of a CE if, and only if, it is idempotent, viz. T
2
= T; when this
latter condition is satised, then T = E
G
, where
G := {A B(I) : T 1
A
= 1
A
}.
Theorem
A Markov operator T is idempotent with respect to composition
T
2
= T, if, and only if, the copula C
T
C
2
that corresponds to it
is idempotent, C
T
= C
T
C
T
.
Theorem
For a copula C, the following statements are equivalent:
(a) the corresponding Markov operator T
C
is a CE restricted to
L
(I, B(I), )
(b) the corresponding Markov operator T
C
is idempotent
(c) C is idempotent
Theorem
To every subeld G of B, the Borel eld of I, there
corresponds a unique idempotent copula C(G) such that
E
G
= T
C(G)
. Conversely, to every idempotent copula C there
corresponds a unique subeld G(C) of B such that T
C
= E
G(C)
.
T
2
f = E(f ) =
_
1
0
f (t) dt and T
M
2
f = f
for every f in L
1
(I). Therefore T
2
= E
N
, where N is the trivial
eld {, I}, and T
M
2
= E
B
; thus
2
and M
2
represent the
extreme cases of copulas corresponding to CEs.
Extreme copul
Denition
Given a copula C C
2
, a copula A C
2
will be said to be a left
inverse of C if A C = M
2
, while a copula B C
2
will be said to
be a right inverse of C if C B = M
2
.
Denition
A copula C C
2
is said to be extreme if the equality
C = A + (1 ) B with ]0, 1[ implies C = A = B.
Theorem
If a copula C C
2
possesses either a left or right inverse, then it is
extreme.
Inverses of copulas
Theorem
When they exist, left and right inverses of copulas in (C
2
, ) are
unique.
Theorem
For a copula C the following statements are equivalent:
(a) for every v I there exists a = a(v) ]0, 1[ such that
D
1
C(u, v) = 1
[a(v),1]
(u), for almost every u I;
(b) C has a left inverse;
(c) there exists a Borelmeasurable function : R R such that
Y = X a.e..
In either case the transpose C
T
of C is a left inverse of C.
Kendall distribution function
If X is a random variable on the probability space (, F, P) and if
its d.f. F is continuous, then the random variable F X = F(X) is
uniformly distributed on I. This is called the probability integral
transform (PIT for short)
Denition
Let (, F, P) be a probability space and on this let X and Y be
random variables with joinf d.f. given by H and with marginals F
and G, respectively. Then the Kendall distribution function of X
and Y is the d.f. of the random variable H(X, Y),
K
H
(t) := P(H(X, Y) t) =
H
__
(x, y) R
2
: H(x, y) t
__
.
Kendall distribution function2
K
H
depends only on the copula C of X and Y:
K
C
(t) := P(C(U, V) t) =
C
__
(u, v) I
2
: C(u, v) t
__
.
Consider an Archimedean copula with inner generator f ,
C
f
(u, v) = g (f (u) + f (v))
then
K
C
f
(t) = t
f (t)
f

(t)
A characterization of Kendall d.f.
Theorem
For every copula C C
2
, K
C
is a d.f. in I such that, for every t I,
(a) t K
C
(t) 1
(b)
K
C
(0) = 0
Moreover the bounds of (a) are attained, since K
M
2
(t) = t and
K
W
2
(t) = 1 for every t I.
For every d.f. F that satises properties (a) and (b) there exists a
copula C C
2
for which F = K
C
.
Kendalls tau
Let (X
1
, Y
1
) and (X
2
, Y
2
) be a pair of independent random vectors
dened on (, F, P) with joint d.f. H; then the population version
of Kendalls tau is dened as the dierence of the probabilities of
concordance and discordance, respectively, namely
X,Y
:= P[(X
1
X
2
) (Y
1
Y
2
) > 0]P[(X
1
X
2
) (Y
1
Y
2
) < 0] .
The concordance function
Theorem
Let X
1
, Y
1
, X
2
, Y
2
be continuous random variables on the
probability space (, F, P). Let the random vectors (X
1
, Y
1
) and
(X
2
, Y
2
) be independent, let H
1
and H
2
be their respective joint
d.f.s and let the marginals d.f.s satisfy F
X
1
= F
X
2
= F and
F
Y
1
= F
Y
2
= G, so that H
1
and H
2
both belong to the Frchet
class (F, G) and H
1
(x, y) = C
1
(F(x), G(y)) and
H
2
(x, y) = C
2
(F(x), G(y)), where C
1
and C
2
are the (unique)
copul of (X
1
, Y
1
) and (X
2
, Y
2
), respectively. Dene
Q := P[(X
1
X
2
) (Y
1
Y
2
) > 0] P[(X
1
X
2
) (Y
1
Y
2
) < 0] .
Then Q depends only on C
1
and C
2
and is given by
Q(C
1
, C
2
) = 4
_
I
2
C
2
(s, t) dC
1
(s, t) 1
Kendalls tau and copul
Corollary
The Kendalls tau of two continuous random variables X and Y on
the probability space (, F, P) depends only on the (unique) copula
C of X and Y and is given by
X,Y
= 4
_
I
2
C(s, t) dC(s, t) 1 .
In terms of the Kendall d.f.
(C) = 3
_
1
0
K
C
(t) dt
Examples
(M
2
) = 1 (W
2
) = 1 (
2
) = 0
For the FarlieGumbelMorgenstern copula C
=
2
9

2
9
,
2
9
_
For the Frchet family of 2copulas
C
,
= M
2
+ (1 )
2
+ W
2
,
where 0, 0 and + 1
(C
,
) =
1
3
( ) ( + + 2)
The case of Archimedean copulas
Theorem
The population version of Kendalls tau (C
f
) for an Archimedean
copula C
f
with inner additive generator f is given by
(C
f
) = 1 + 4
_
1
0
f (t)
f

(t)
dt
Spearmans rho
Let (X
1
, Y
1
), (X
2
, Y
2
) and (X
3
, Y
3
) there independent continuous
random vectors having a common joint distribution function H,
with marginals F and G and copula C. Then Spearmans rho
XY
is dened to be proportional to the dierence between the
probability of concordance and the probability of discordance for
the two vectors (X
1
, Y
1
) and (X
2
, Y
3
); notice that the distribution
function of the second vector is F G, since X
2
and Y
3
are
independent. Then
X,Y
:= 3 (P[(X
1
X
2
) (Y
1
Y
3
) > 0] P[(X
1
X
2
) (Y
1
Y
3
) < 0])
Spearmans rho and copul
Theorem
If C is the copula of two continuous random variables X and Y,
then the population version of Spearmans rho of X and Y depends
only on C, will be denoted indierently by
X,Y
or by
C
or by
(C), and is given by
X,Y
=
C
= 12
_
I
2
u v dC(u, v) 3 = 12
_
I
2
C(u, v) du dv 3
= 12
_
I
2
{C(u, v) u v} du dv
The SchweizerWol measure of dependence
Let X and Y be continuous random variables and let F and G be
their d.f.s, H their joint d.f., and C their (unique) connecting
copula. The graph of C is a surface over the unit square, which is
bounded above by the surface z = M
2
(u, v), and is bounded below
by the surface z = W
2
(u, v). If X and Y happen to be
independent, then the surface z = C(u, v) is the hyperbolic
paraboloid z = u v. The volume between the surfaces z = C(u, v)
and z = u v can be used as a measure of dependence. The
SchweizerWol measure of dependence
(X, Y) := 12
_
I
2
|C(u, v) u v| du du = 12
_
I
2
|C
2
| d
2
= 12
_
I
2
|H(u, v) F(u) G(v)| dF(u) dG(v)
Properties of then SW measure
(SW1) is dened for every pair of continuous random variables X
and Y dened on the same probability space (, F, P)
(SW2) (X, Y) = (Y, X)
(SW3) (X, Y) [0, 1]
(SW4) (X, Y) = 0 if, and only if, X and Y are independent;
(SW5) (X, Y) = 1 if either X = Y or Y = X for some
strictly monotone functions , : R R
(SW6) ( X, Y) = (X, Y) for strictly monotone if
, : R R
(SW7) (X, Y) = 6/ arcsin(||/2) for the bivariate normal
distribution with correlation coecient
(SW8) if (X
n
, Y
n
) has joint continuous d.f. H
n
and converges in law
to the random vector (X, Y) with continuous joint d.f. H
0
,
then (X
n
, Y
n
) (X, Y)
Rnyis axioms
(R1) R is dened for any pair of random variables X and Y that are
not a.e. constant
(R2) R is symmetric, R(X, Y) = R(Y, X)
(R3) for every pair of nonconstant random variables X and Y,
R(X, Y) belongs to [0, 1]
(R4) R(X, Y) = 0 if, and only if, X and Y are independent
(R5) R(X, Y) = 1 if either x = f Y or Y = g X for some Borel
measurable functions f and g
(R6) if f , g : R R are Borelmeasurable and onetoone, then
R(f X, g Y) = R(X, Y)
(R7) if the joint distribution of X and Y is a bivariate normal
distribution with correlation coecient , then R(X, Y) = ||
Other measures of dependence
the L
norm:
(X, Y) := k
C
2
= k
sup
(u,v)I
2
|C(u, v)
2
(u, v)| ;
the L
p
norm:
p
(X, Y) := k
p
__
I
2
|C(u, v)
2
(u, v)|
p
d
2
_
1/p
Measures of nonexchangeability
Let H(F) be the class of all random pairs (X, Y) such that X and
Y are identically distributed with continuous joint d.f. F.
Denition
A function : H(F) R
+
is called a measure of
nonexchangeability if
(A1) is bounded, (X, Y) K
(A2) (X, Y) = 0 if, and only if, (X, Y) is exchangeable
(A3) is symmetric: (X, Y) = (Y, X)
(A4) (X, Y) = (f (X), f (Y)) for every strictly monotone function
f
(A5) if (X
n
, Y
n
) and (X, Y) are pairs of random variables with joint
d.f.s H
n
and H, respectively, and if H
n
converges weakly to H,
then (X
n
, Y
n
) converges to (X, Y)
In the language of copulas
Denition
A function : C R
+
is called a measure of nonexchangeability
for H(F) if it satises the following properties:
(B1) (C) K
(B2) (C) = 0 if, and only if, C is symmetric;
(B3) (C) = (C
t
)
(B4) (C) = (
C)
(B5) if C
n

to+
C uniformly, then (C
n
)
to+
(C)
An explicit measure
Theorem
The mapping
p
: C R
+
dened by
p
(C) := d
p
(C, C
t
)
is a measure of nonexchangeability for every p [1, +].
Theorem
For every p [1, +[ and for every C C
2
, one has
p
(C)
_
2 3
p
(p + 1) (p + 2)
_
1/p
1
3
.
Quasicopul
Quasicopul
Denition
A track B in I
d
is a subset of unit cube I
d
that can be written in
the form
B := {(F
1
(t), F
2
(t), . . . , F
d
(t)) : t I}
where F
1
, F
2
, . . . , F
d
are continuous d.f.s such that F
j
(0) = 0
and F
j
(1) = 1 for j = 1, 2, . . . , d
Denition
A dquasicopula is a function Q : I
d
I such that for every track
B in I
d
there exists a dcopula C
B
that coincides with Q on B,
namely such that, for every point u B,
Q(u) = C
B
(u).
Quasicopul
An equivalent denition
Theorem
A dquasicopula Q satises the following properties:
(a) for every j {1, 2, . . . , d}, Q(1, . . . , 1, u
j
, 1, . . . , 1) = u
j
(b) Q is increasing in each place
(c) Q satises Lipschitz condition, if u and v are in I
d
, then
|Q(v) Q(u)|
d
j =1
|v
j
u
j
|
Conversely if Q : I
d
I satises properties (a), (b) and (b), then
it is a quasicopula.
Quasicopul
An immediate consequence
For d > 2 the function W
d
(u) := max{u
1
+ + u
d
d + 1, 0} is
a dquasicopula, but not a copula. For d > 2 consider the dbox
[1/2, 1] = [1/2, 1] [1/2, 1] [1/2, 1] .
Then W
d
volume of this dbox is, for d > 2,
V
W
d
([1/2, 1]) = 1
d
2
< 0,
so that W
d
cannot be a copula for d > 2, but is a proper
quasicopula.
Quasicopul
A surprising result
Let
Q
the real measure induced by the quasicopula Q on
(I
2
, B(I
2
)).
Theorem
For all given > 0 and M > 0, there exist a quasicopula Q and a
Borel subset S of I
2
such that
(a)
Q
(S) < M
(b) for all u and v in I, |Q(u, v)
2
(u, v)| <
Quasicopul
Quasiicopul form a lattice
Given a set S of functions from I
d
into I one denes
S(u) := inf{S(u) : S S}.

Theorem
Both the upper and the lower bounds,
_
Q and
_
Q of every set Q
of dquasicopulas are quasicopul,
_
Q Q
d
and
_
Q Q
d
.
Corollary
Both the upper and the lower bounds,
_
C and
_
C of every set C
of dcopulas are dquasicopul,
_
C Q
d
and
_
C Q
d
.
Quasicopul
An example
For I consider the copula
C
(s, t) =
_
_
min{s, t }, (s, t) [0, 1 ] [, 1] ,
min{s + 1, t}, (s, t) [1 , 1] [0, ] ,
W
2
(s, t), elsewhere,
If U and V are uniform rvs with V = U + (mod 1); then C
is
their copula. Set C =
_
C
1/3
, C
2/3
_
, then
_
C is given by
C(s, t) =
_
max{0, s 1/3, t 1/3, s + t 1}, 1/3 t s 2/3,
W
2
(s, t), elsewhere.
Notice
V
C
_
[1/3, 2/3]
2
_
= 1/3 < 0
Quasicopul
Q
d
as a lattice
A partially ordered set P = is said to be a lattice if both the join
x y and x y of every pair x and y of elements of P are in P. A
lattice P is said to be complete if both S and S belong to P for
every subset S of P.
Theorem
The set Q
d
of dquasicopulas is a complete lattice under
pointwise suprema and inma.
Theorem
Neither the family C
d
of copulas nor the family Q
d
\ C
d
of proper
quasicopulas is a lattice.
Carlo Sempi
Lecce, Italy
Outline
1
Construction of copulas: the geometric method
2
The compatibility problem
An example: the tent map
Choose in ]0, 1[ and consider the probability mass spread on the
segment joining the points (0, 0) and (, 1) and the probability
mass 1 spread on the segment joining the points (, 1) and
(1, 1). It is now easy to nd the expression for the copula C
of the
resulting probability distribution on the unit square:
C
(u, v) =
_
_
u, u [0, v] ,
v, u ] v, 1 (1 ) v[ ,
u + v 1, u [1 (1 ) v, 1] .
The diagonal of a copula
The diagonal section
C
of a copula C C
d
is the function
C
: I I, dened by
C
(t) := C(t, t, . . . , t).
The diagonal section has a probabilistic meaning. If U
1
, U
2
,. . . , U
d
are random variables dened on the same probability space
(, F, P), having uniform distribution on (0, 1) and C as their
(unique) copula, then
C
(t) = C(t, t, . . . , t) = P
_
_
d
j =1
{U
j
t}
_
_
= P (max{U
1
, U
2
, . . . , U
d
} t) = P
_
_
d
j =1
U
j
t
_
_
,
Then
C
is the d.f. of the random variable max{U
1
, U
2
, . . . , U
d
}
Properties of the diagonal section
Theorem
The diagonal section
C
of a copula C C
d
, or of a quasicopula
Q Q
d
, satises the following properties:
(D1)
C
(0) = 0 and
C
(1) = 1
(D2) t I
C
(t) t
(D3) the function I t
C
(t) is isotone;
(D4) |
C
(t
)
C
(t)| d |t
t| for all t and t
in I
The set of diagonals will be denoted by D
Questions
(Q.1) whether, given a diagonal D, there exists a copula C
whose diagonal section
C
coincides with , namely whether
the class C
is nonempty;
(Q.2) whether there exist bounds for the family C
; these, if they
exist, are necessarily sharper than the FrchetHoeding ones;
(Q.3) whether these bounds, when they exist, are the best possible.
Answer to (Q.1)
Theorem
For every D, the function K
: I
2
I dened by
K
(u, v) := min
_
u, v,
(u) + (v)
2
_
is a copula with diagonal , so that K
belongs to C
; it will be
called the diagonal copula associated with .
The probabilistic meaning
Theorem
Let X and Y be continuous random variables on the same
probability space (, F, P), with a common d.f. F and copula C.
Then the following statements are equivalent:
(a) The joint d.f. of the random variables min{X, Y} and
max{X, Y} is the FrchetHoeding upper bound
(b) C is a diagonal copula.
More probability
Lemma
For every diagonal and for every symmetric copula C C
one
has C K
.
Theorem
For a diagonal the following statements are equivalent:
(a) is the diagonal section of an absolutely continuous copula
C C
d
(b) the set {t I : (t) = t} has Lebesgue measure 0,
({t I : (t) = t}) = 0
The Bertino copula
For a given diagonal dened

(t) := t (t)
Theorem
For every diagonal D, the function B
: I
2
I dened by
B
(u, v) := min{u, v} min{
(t) : t [u v, u v]}
=
_
u min
t[u,v]
{t (t)}, u v,
v min
t[v,u]
{t (t)}, v u
is a symmetric 2copula having diagonal equal to , i.e., B
.
B
is called the Bertino copula of .

Bounds for copulas with given diagonal1
Theorem
For every diagonal D, the function A
: I
2
I dened by
A
(u, v) := min
_
u, v, max{u, v} max{
(t) : t [u v, u v]}
_
=
_
min
_
u, v max
t[u,v]
{t (t)}
_
, u v,
min
_
v, u max
t[v,u]
{t (t)}
_
, v u
is a symmetric 2quasicopula having diagonal equal to , i.e.,
A
.
Bounds for copulas with given diagonal2
Theorem
For every diagonal and for every copula C C
one has
B
C A
.
Theorem
The quasicopula A
is a copula if, and only if, A
= K
.
Theorem
For the quasicopula A
the following statements are equivalent:

(a) A
= K
(b) the graph of the function t (t) is piecewise linear; each

segment has slope equal to 0, 1 or 2 and has at least one of its
endpoints on the diagonal v = u.
Statement of the problem
In its most general form, the problem runs as follows. If k and d
with 1 < k d are natural numbers, the dcopula C has
_
d
k
_
kmarginals, which are obtained by setting d k of its arguments
equal to 1. In the other direction, if at most
_
d
k
_
kcopul are
given, there may not exist a dcopula of which the given kcopul
are the kmarginals. This may easily be seen in the case d = 3 and
k = 2; if, for instance, the three two copul are all equal to W
2
,
then, in view of the probabilistic meaning of the copula W
2
, there is
no 3copula C of which they are the marginals. On the other hand,
if an dcopula exists of which the given copul are the
kmarginals, then these are said to be compatible.
The special case d = 3 and k = 2
Let A and B be 2copul, A, B C
2
, and denote by D(A, B) the
set of all 2copulas that are compatible with A and B, in the sense
that, if C is in D(A, B), then there exists a 3copula

C such that,
for all (u, v, w) [0, 1]
3
,
C(u, v, 1) = A(u, v),

C(1, v, w) = B(v, w),

C(u, 1, w) = C(u, w).
Theorem
Given any two 2copulas A and B, there always exists a 2copula
C that is compatible with A and B, namely D(A, B) = , for
instance A B.
Examples
C
W
2
,W
2
(u, v, w) = max{0, v + (u w) 1},
C
M
2
,M
2
(u, v, w) = u v w = M
3
(u, v, w),
C
W
2
,M
2
(u, v, w) = max{0, u + (v w) 1},
C
M
2
,W
2
(u, v, w) = max{0, (u v) 1 + w},
C
2
,
2
(u, v, w) = uvw =
3
(u, v, w),
C
2
,M
2
(u, v, w) = u M
2
(v, w),
C
M
2
,
2
(u, v, w) = w M
2
(u, v).
Properties of D(A, B)
Theorem
The set D(A, B) of copulas that are compatible with two given
bivariate copulas A and B is convex and compact with respect to
the topology of uniform convergence in I
2
.
Minimality of D(A, B)
The class D(A, B) is said to be minimal when D(A, B) = {A B}.
It is worth asking: when is this the case? The following theorem
provides a sucient condition for this to happen.
Theorem
Let A and B be two bivariate copulas with A = C
f ,g
and B = C
p,r
,
where f , g, p and r are measurepreserving transformations from I
into I, and either pair (f , g) or (p, r ) is made of onetoone
transformations. Then D(A, B) is minimal.
Corollary
If either A or B (or both) is a shue of Min, then
D(A, B) = {A B}.
Gluing of two copulas1
Let A and B be dcopul, A, B C
d
, let i {1, 2, . . . , n}, and
choose in ]0, 1[. Dene the (u
i
= )gluing of A and B via
_
_
A
u
i
=
B
_
_
(u
1
, . . . , u
i 1
, u
i
, u
i +1
, . . . , u
d
)
:= A
_
u
1
, . . . , u
i 1
,
u
i
, u
i +1
, . . . , u
d
_
for u
i
[0, ]
Gluing of two copulas2
_
_
A
u
i
=
B
_
_
(u
1
, . . . , u
i 1
, u
i
, u
i +1
, . . . , u
d
)
:= A(u
1
, . . . , u
i 1
, 1, u
i +1
, . . . , u
d
)
+ (1 ) B
_
u
1
, . . . , u
i 1
,
u
i

1
, u
i +1
, . . . , u
d
_
for u
i
[, 1].
Theorem
For every pair A and B of dcopulas, for every index
i {1, 2, . . . , d}, and for every ]0, 1[, A

u
i
=
B is a dcopula.

A - Z of Copulas

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A - Z of Copulas

Uploaded by

Copyright:

Available Formats

An introduction to Copulas

associated with a vector U = (U

is distributed according to the copula C

(u) := ([0, u]) .

of the Markov operator T

, ) two measure spaces.

is a measurepreserving transformations (=MPT) if

on the same probability space and a piecewise

is the joint d.f. of X

When is strict the copula C

is said to be strict; in this case, C

has the representation

(t) = (max1 +t, 0)

is the Archimedean tnorm generated by the outer generator

(I) is the restriction to L

S(u) := inf{S(u) : S S}.

t| for all t and t

(u, v) := min{u, v} min{

is called the Bertino copula of .

is a copula if, and only if, A

the following statements are equivalent:

(b) the graph of the function t (t) is piecewise linear; each

C(u, v, 1) = A(u, v),

You might also like