Copulas and Stochastic Processes

Von der Fakultät für Mathematik, Informatik und Naturwissenschaften

der Rheinisch-Westfälischen Technischen Hochschule Aachen
zur Erlangung des akademischen Grades eines Doktors
der Naturwissenschaften genehmigte Dissertation

vorgelegt von

Diplom-Mathematiker
Volker Schmitz
aus Düren

Berichter: Universitätsprofessor Dr. H.-H. Bock

Universitätsprofessor Dr. B. Rauhut

Tag der mündlichen Prüfung: 21. Januar 2003

Diese Dissertation ist auch als Verlagsveröffentlichung

(Shaker Verlag, Aachen, ISBN 3-8322-1278-7) erhältlich.
ii
Preface

This thesis was written during my time at the Institute of Statistics of Aachen University.
I would like to thank Professor Hans-Hermann Bock for giving me the opportunity and
freedom to choose this topic, his valuable comments, continuous interest and support for
this work. I am very grateful for having been given the opportunity to work in this highly
interesting and motivating field.
Many thanks to Professor B. Rauhut for also taking the effort to referee this thesis.
A special thankyou goes to Dr. E. Pier-Ribbert who drew my attention to the interesting
field of copulas during my stay at Dresdner Bank, Frankfurt.
I would also like to thank my colleagues at the Institute of Statistics of whom many con-
tributed to this thesis in one way or another. In particular, I have to mention Stefan Merx
who provided me with many useful hints and critical remarks and Eric Beutner for his
proof-reading of some of the chapters.
Last, I want to thank my wife Fiona for her patience and perseverance in supporting me
during the past years.

Aachen, February 2003 Volker Schmitz

iv
Contents

1 Introduction 1

2 Copulas and Their Properties 5

2.1 Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 The Role of Partial Derivatives . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Archimedean Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Survival Copula and Symmetry Concepts . . . . . . . . . . . . . . . . . . . 25
2.4.1 Symmetry Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.2 Survival Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3 Dependence Measures 33
3.1 Comonotonicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Concordance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 Kendall’s τ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Spearman’s ρ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Tail Dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6 Quadrant Dependence and Positive Likelihood Ratio Dependence . . . . . 39

4 Two Examples: Analysis of Spatial Dependence by Using Copulas 41

4.1 The Min-Max Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 The Copula of Brownian Motion and Its Supremum . . . . . . . . . . . . . 51

5 Stochastic Processes And Copulas 53

5.1 Stochastic Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.2 General Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2.1 Strict Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2.2 Equivalence of Stochastic Processes . . . . . . . . . . . . . . . . . . 58
5.2.3 Symmetric Processes . . . . . . . . . . . . . . . . . . . . . . . . . . 60
vi Contents

5.3 Continuous Time Stochastic Processes . . . . . . . . . . . . . . . . . . . . 60

5.3.1 Stochastic Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.2 EFGM processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Markov Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4.1 An Application: Structure of Records . . . . . . . . . . . . . . . . . 77
5.4.2 Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.5 Continuous Local Martingales . . . . . . . . . . . . . . . . . . . . . . . . . 86

A General Results from Probability Theory 93

A.1 Generalized Inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
A.2 Probability Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

B Simulation/Random Variate Generation 95

B.1 The Conditional Distribution Method . . . . . . . . . . . . . . . . . . . . . 95
B.2 Simulating from Archimedean Copulas . . . . . . . . . . . . . . . . . . . . 96
B.3 Generation of Time-Series from Bivariate Copulas . . . . . . . . . . . . . . 97

List of Symbols and Abbreviations 99

Bibliography 103

Index 109
1
Introduction

The modelling of dependence relations between random variables is one of the most widely
studied subjects in probability theory and statistics. Thus, a great variety of concepts for
dependence structures has emerged (see Jogdeo, 1982). However, without specific assump-
tions about the dependence, no meaningful stochastic model can be developed.
The dependence of random variables can basically be classified into spatial and temporal
dependencies where the former means the dependence between a number of variables at the
same time whereas the latter means the intertemporal dependence structure of a process.
The classic approach uses second order moments of the underlying random variables,
known as the covariance. It is well known that only linear dependence relationships can
be captured this way and that it is characterizing only for special classes of distributions,
for example the normal distribution. The question arises if there is a possibility to capture
the whole dependence structure without any disturbing effects coming from the marginal
distributions. This is exactly where copulas provide a beneficial approach.

As the details will be given in the next chapter we will only say that copulas al-
low to separate the effect of dependence from the effects of the marginal distributions,
i. e., if X1 , . . . , Xn are real-valued random variables on some probability space (Ω, A, P)
with joint distribution function H, i. e., H(x1 , . . . , xn ) := P(X1 ≤ x1 , . . . , Xn ≤ xn ),
xi ∈ , 1 ≤ i ≤ n, a copula C of (X1 , . . . , Xn ) is given by the relation H(x1 , . . . , xn ) =
C(F1 (x1 ), . . . , Fn (xn )) for all xi ∈ where the Fi are the univariate marginal distributions
of the Xi , i. e., Fi (x) := P(Xi ≤ x), x ∈ , 1 ≤ i ≤ n.
Simply put, a copula is a multivariate distribution function on the n-dimensional unit
cube with uniform marginal distributions.

The question of existence and uniqueness of such a copula was answered by Sklar (1959)
who also introduced the term “copula” which is now the most widely used term. (“Copula”
is derived from the Latin verb “copulare”, meaning “to join together”.) Also used are the
terms “dependence function” (e. g., Galambos, 1978, Definition 5.2.1) in the context of
2 1. Introduction

multivariate extreme value theory, “uniform representation” (e. g., Kimeldorf and Sampson,
1975) and “standard form” (e. g., Cook and Johnson, 1981).

Although the concept is rather old, the first volume of the Encyclopedia of Statistical
Sciences to contain the entry “copulas” was the Update Volume from 1997 (Fisher, 1997).
This may be due to the fact that from 1959 to 1974 results concerning copulas were ob-
tained in the course of the development of the theory of probabilistic metric spaces and
distributions with given marginals (for a historic account see Schweizer, 1991). Nowadays,
however, interest in copulas has grown especially in mathematical finance; spatial depen-
dence between assets is tried to be captured by copulas instead of the classic models such
as correlation structures. For an example of this approach see Embrechts et al. (1999) or
Schmitz and Pier-Ribbert (2001).

Most of the current research in copulas is done for spatial dependence due to great interest
in practice for new spatial dependence models, but the analysis of temporal dependence is
also possible by the copula approach. The first paper dealing exclusively with copulas and
stochastic processes was presented by Darsow et al. (1992), who established the connection
between copulas and Markov processes. Only few papers have appeared since then, and
most are working papers not yet published in a journal.

This thesis extends the relationship between copulas and stochastic processes into new di-
rections and tries to show that the copula approach can provide new structural insights into
the temporal dependence underlying the processes. It will become clear that many prop-
erties of stochastic processes can be analyzed by splitting conditions up into a “marginal
part” and a “dependence part”, thus providing a better insight into the laws governing the
temporal behaviour of the process.
Many known theorems can be given alternative and in most cases simpler proofs via
copulas.
Although the focus is laid upon univariate stochastic processes, we also will—in a small
excursion—deal with two special and interesting aspects of copulas for spatial dependence
(Chapter 4).

The structure of this thesis is as follows:

After this general Introduction, Chapter 2 (Copulas and Their Properties) will give
a basic introduction to the definition and theory of copulas. It will be a basic reference for
all other chapters to follow. A special focus should be laid upon Section 2.2 (The Role of
Partial Derivatives) in which the connection between conditional distribution functions and
partial derivatives of copulas will be developed in a mathematical rigorous way. Although
many of the results are well known, it is difficult to find mathematically satisfying proofs
for them.
As the family of Archimedean copulas is the most prominent of all currently used models,
we give a short introduction to them in Section 2.3.
As a preparation for the analysis of symmetric stochastic processes we need to introduce
the survival copula and some symmetry concepts for multivariate random variables in
 1. Introduction 3

Section 2.4. However, this section is of its own interest. The most important result is
that radial symmetry of a n-variate distribution function can be equivalently established
by radial symmetry of the underlying copula and (univariate) symmetry of the marginals
(Corollary 2.41).

As we learned above, copulas provide a means of dividing the multivariate distribution

function into separate models for the marginals and the dependence between the variables.
It is thus necessary to introduce some of the most widely used dependence concepts and
their relation to copulas in Chapter 3. The results (most of which are known) will be
used in the subsequent chapters. Especially the concepts of tail dependence and positive
likelihood ratio dependence will provide us with an elegant tool to compare Kendall’s τ
and Spearman’s ρ between the minimum and maximum of n independent and identically
distributed (iid) random variables in Section 4.1.

Chapter 4 (Two Examples: Analysis of Spatial Dependence With Copulas) consists

of two independent sections. In the first one we will take a closer look at the dependence
structure between the minimum and maximum of n iid random variables. By deriving their
copula we are able to give an alternative proof for the asymptotic independence which is
essentially shorter than the known ones and uses only elementary calculus.
We will further determine Kendall’s τ and Spearman’s ρ and show the relation 3τ ≥ ρ ≥
τ > 0 for them with the help of tail dependence and positive likelihood ratio dependence.
In the second section we will derive the copula of Brownian motion and its supremum
process. The rather astonishing result is that this copula is independent of time.

Chapter 5 deals with the main topic of this thesis: the connection between univariate
stochastic processes and copulas. As we will look at a rather wide field of different concepts,
an introduction to definitions used in the sequel is given in Section 5.1.
Section 5.2 introduces basic relations between copulas and stochastic processes. Concepts
such as strict stationarity, equivalence of stochastic processes and symmetry of processes
are characterized by copulas.
Section 5.3 will specialize on processes in continuous time, e. g., Brownian motion. We
develop sufficient conditions to the univariate marginals and the copula structure of a pro-
cess to ensure stochastic continuity (Section 5.3.1). These conditions are then used to give
an elegant proof of the findings of Cambanis (1991) for EFGM-processes. He showed that
a stochastic process based on multivariate EFGM-distributions cannot be stochastically
continuous. However, he established this result by a number of inequalities whereas we are
able to give a direct proof with the help of Theorem 5.30 in Theorem 5.35.
Section 5.4 (Markov Processes) introduces the results of Darsow, Nguyen, and Olsen
(1992) and extends the findings to a characterization of symmetric Markov processes by
radial symmetry of the bivariate copulas (Section 5.4.2). Interestingly, within the class of
Archimedean copulas, symmetric Markov processes can be characterized by the so-called
Frank copula.
Finally, the copula structure of continuous local martingales is derived in Section 5.5.
As continuous local martingales are not defined by their finite-dimensional distributions,
a different technique has to be used to determine the underlying copulas. We show that
4 1. Introduction

the copula of Brownian motion plays a central role as each continuous local martingale
is equivalent to a time change of Brownian motion, a result due to Dambis (1965) and
Dubins and Schwarz (1965). It is then possible, for example, to derive the copula of the
Ornstein-Uhlenbeck process and the Brownian bridge. These can be used to construct
and/or simulate processes which have the same intertemporal dependence structure but
with other marginal distributions.

In the Appendix we will provide some additional information about basic tools used
within this thesis which are not worth being dealt with in the main part. A special focus
is laid on the simulation of random variables and processes from copulas, since in practice
simulation from copulas is one of the main tools to analyze the underlying models.

The List of Symbols and Abbreviations tries to give the reader a central point
of access for all symbols used within the text. Where appropriate, the page number is
indicated where the corresponding symbol is defined or introduced.
I tried to used notations and symbols which seem to be standard in the field concerned,
the only speciality is the symbol “” which marks the end of an example or remark.

Concerning the Bibliography, two things should be mentioned. Firstly, the numbers at
the end of each entry refer to the page on which the work is cited in this thesis. Secondly,
it is clear that the average reader does not necessarily have access to every source in the
references. However, I decided to refer to the literature wherein the exact formulation of
the needed theorems can be found, but it is possible to read the text with a single reference,
e. g., Rogers and Williams (2000a,b).

Finally, the Index should certainly help the reader find what he or she is looking for. If
page numbers behind entries are emphasized (such as 4711 in contrast to 4711), they refer
to the most important pages concerning these entries. It usually means that the definition
can be found there.
2
Copulas and Their Properties

In this chapter we give a basic introduction to copulas and their properties. As there seems
to be a lack of literature connecting copulas and conditional probabilities in a mathematical
rigorous way, we also shed some light on this aspect of copulas which turns out to be a
very powerful tool when dealing with the modelling of stochastic processes with copulas.

2.1 Basic Properties

The basic idea underlying the concept of copulas can be described as follows: Assume a
pair of real-valued1 random variables X and Y on a common probability space (Ω, F , P)
with distribution functions F (x) := P(X ≤ x), x ∈ , and G(y) := P(Y ≤ y), y ∈ ,
respectively, and a joint distribution function H(x, y) := P(X ≤ x, Y ≤ y), x, y ∈ . If
we associate the three numbers F (x), G(y) and z := H(x, y) to every pair of real numbers
(x, y), we observe that each
of them lies in the unit interval I := [0, 1], i. e., each pair (x, y)
leads to a point F (x), G(y) in the unit plane I 2 and this ordered pair
in turn corresponds
to the number z ∈ I by the relation z = H(x, y) = C F (x), G(y) for some function
C : I 2 → I. The last correspondence turns out to be characterizing for copulas.

Let us define some notions first. We denote the extended real line by , i. e., :=
n
∪ {−∞} ∪ {∞}, analogously, denotes the extended n-dimensional real space.
n
For two vectors x = (x1 , . . . , xn ), y = (y1 , . . . , yn ) ∈ , we will write x ≤ y if xi ≤ yi ,
1 ≤ i ≤ n, and x < y if xi < yi for all 1 ≤ i ≤ n.

1 All random variables in this thesis will be assumed to be real-valued if not stated otherwise.
6 2. Copulas and Their Properties

2.1 Definition
n
A (halfopen) rectangle or interval in is the Cartesian product of n one-dimensional
intervals of the form
n
Rxy := ×(xi , yi ] = (x1 , y1 ] × · · · × (xn , yn ], x ≤ y,
i=1

n
where x = (x1 , . . . , xn ), y = (y1 , . . . , yn ) ∈ .
n
The set of all rectangles in will be denoted by Rn .
The vertices of a rectangle Rxy ∈ Rn (of which we clearly have 2n ) are the points
vert(Rxy ) := {(u1 , . . . , un ) : ui ∈ {xi , yi }, 1 ≤ i ≤ n}.
The unit cube I n is the product I × · · · × I (n times) where I = [0, 1] is the unit interval.
The domain and range of a function H will be denoted by dom(H) and ran(H), respec-
tively.
As we will use n-dimensional Lebesgue-(Stieltjes-)integration, we need a n-dimensional
analogue of an increasing2 function (i. e., some prototype of a measure generating function).
This can be established via the n-dimensional volume of a rectangle as we will see later.

2.2 Definition (H-volume)

n
Let S1 , . . . , Sn ⊂ be nonempty sets and H : → a function such that dom(H) ⊂
S1 × · · · × Sn . Let Ra = ×i=1 (ai , bi ] ∈ R , a, b ∈
b n n n
, be a rectangle with Rab ⊂ dom(H).
b
Then the H-volume VH of Ra is given by

VH (Rab ) := ∆ba H(x1 , . . . , xn ) (2.1)

with
(n) (n−1) (1)
∆ba H(x1 , . . . , xn ) := ∆(an ,bn ) ∆(an−1 ,bn−1 ) · · · ∆(a1 ,b1 ) H(x1 , . . . , xn ) (2.2)
where
(i)
∆(ai ,bi ) H(x1 , . . . , xn ) := H(x1 , . . . , xi−1 , bi , xi+1 , . . . , xn ) − H(x1 , . . . , xi−1 , ai , xi+1 , . . . , xn )
(2.3)
is the first order difference operator. ∆ba H(x1 , . . . , xn ) is therefore the n-th order difference
of H with regard to a and b.

2.3 Remark
The above notation is used to clarify on which coordinate the operator ∆ operates. As this
is notationally inconvenient, we will use the notation
(i)
∆baii ≡ ∆(ai ,bi )

where the coordinate is clear from the context. Where confusion could arise, we will use
the full notation. 
2 Note that the terms “increasing” and “decreasing” mean f (x) ≤ f (y), x ≤ y, for a function f and f (x) ≥ f (y), x ≤ y,

respectively. If a strict inequality holds, we will use the term “strictly in-/decreasing”.
 2.1 Basic Properties 7

Note that ∆ba H can also be written as

X Pn

∆ba H(x) = (−1) i=1 εi · H ε1 a1 + (1 − ε1 )b1 , . . . , εn an + (1 − εn )bn . (2.4)
(ε1 ,...,εn )∈{0,1}n

The proof of this is a straightforward induction and omitted.

The following definition is the n-dimensional analogue3 of the one-dimensional mono-
tonicity:

2.4 Definition (n-increasing)

n
Let H be a real-valued function with dom(H) ⊂ . H is said to be n-increasing if VH (R) ≥
n
0 for all rectangles R ∈ R with vert(R) ⊂ dom(H).
Some authors refer to n-increasing functions as quasi-monotone or ∆-monotone (e. g.,
Behnen and Neuhaus, 1995).

The next definition is intuitive in the sense that the notion of groundedness is a natural
requirement for a multivariate distribution function on n .

2.5 Definition (grounded)

Let H be a real-valued function with dom H = S1 × · · · × Sn , Si ⊂ , and let each Si have
a smallest element ai . H is called grounded if for all i = 1, . . . , n,

H(x1 , . . . , xi−1 , ai , xi+1 , . . . , xn ) = 0

for all xj ∈ Sj , 1 ≤ j ≤ n, j 6= i, ai ∈ Si .
We can now introduce the definition of a copula:

2.6 Definition (copula)

A copula (or n-copula) is a function C : I n → I with the following properties:

a) C is grounded, i. e., for every u = (u1 , . . . , un ) ∈ I n ,

C(u) = 0 if there exists an i ∈ {1, . . . , n} with ui = 0, (2.5)

and:
If all coordinates of u ∈ I n are 1 except uk , then C(u) = uk . (2.6)

b) C is n-increasing, i. e., for all a, b ∈ I n such that a ≤ b,

VC (Rab ) ≥ 0. (2.7)

The set of all n-copulas will be denoted by Cn .

3 In view of the property of being a measure inducing function

8 2. Copulas and Their Properties

We easily see that C(u1 , . . . , un ) = ∆u0 C(x1 , . . . , xn ) = VC ((0, u1 ] × · · · × (0, un ]) due to

(2.1)–(2.3) and (2.5), so one can think of C(u1 , . . . , un ) as an assignment of a number in I
to the rectangle (0, u1 ] × · · · × (0, un ].

We will now present some simple but basic properties of copulas. All these properties
hold for the general case n ≥ 2, although formulated and proved for the case
n = 2. The general case can be proved analogously, but a rigorous treatment would be
notationally complex.
2.7 Lemma
Let C ∈ C2 and 0 ≤ x1 ≤ x2 ≤ 1, then

y 7→ C(x2 , y) − C(x1 , y) (2.8)

is increasing on I.
Similarly, for 0 ≤ y1 ≤ y2 ≤ 1, the function

x 7→ C(x, y2 ) − C(x, y1 ) (2.9)

is increasing on I.
Proof. We only consider the former case as the latter one follows analogously. Let 0 ≤ y 1 ≤
y2 ≤ 1 and 0 ≤ x1 ≤ x2 ≤ 1. As C is 2-increasing (see Def. 2.4), we have (with x = (x1 , x2 ),
y = (y1 , y2 ))

VC (Rxy ) = C(x2 , y2 ) − C(x2 , y1 ) − C(x1 , y2 ) + C(x1 , y1 ) ≥ 0

(2.8*)
⇐⇒ C(x2 , y2 ) − C(x1 , y2 ) ≥ C(x2 , y1 ) − C(x1 , y1 ),

which is just the claimed monotonicity property.

2.8 Corollary
Copulas are Lipschitz-continuous (hence continuous), i. e., for any C ∈ C2 and 0 ≤ x1 ≤
x2 ≤ 1, 0 ≤ y1 ≤ y2 ≤ 1, and x, y ∈ I, it holds that

0 ≤ C(x2 , y) − C(x1 , y) ≤ x2 − x1 and (2.10 a)

0 ≤ C(x, y2 ) − C(x, y1 ) ≤ y2 − y1 so that (2.10 b)
|C(x2 , y2 ) − C(x1 , y1 )| ≤ |x2 − x1 | + |y2 − y1 |. (2.10)

Proof. Set y = 0, y = 1 and x = 0, x = 1, respectively, in Lemma 2.7, (2.8) and (2.9).

2.9 Corollary
Let C ∈ C2 , then y 7→ C(x, y) is increasing for all x ∈ I and analogously, x 7→ C(x, y) is
increasing for all y ∈ I.
Proof. Take x1 = 0 and y1 = 0, respectively, in Lemma 2.7.
 2.1 Basic Properties 9

2.10 Remark
The set Cn of all n-copulas is a compact and convex subset of the space of all continuous
real valued functions defined on I n under the topology of uniform convergence. It follows
that—in Cn —pointwise convergence implies uniform convergence.
The convexity follows directly by the definiton of a copula. The compactness follows by
noticing that the set of all continuous real-valued functions on I n is a compact metric space
and the subset of all n-copulas is closed. 

Two functions, of which only one is a copula for all dimensions (see Remark 2.12), are
special in that they provide bounds for all copulas:

2.11 Theorem (Fréchet bounds)

Let C be a 2-copula. Then for u, v ∈ I,

W (u, v) := max(u + v − 1, 0) ≤ C(u, v) ≤ min(u, v) =: M (u, v). (2.11)

The functions W and M are called lower and upper Fréchet bounds.

Proof. Due to monotonicity we have C(u, v) ≤ C(u, 1) = u and C(u, v) ≤
C(1, v) = v,

so C(u, v) ≤ M (u, v). C(u, v) ≥ 0 is trivially true and VC (u, 1] × (v, 1] ≥ 0 yields
C(1, 1) + C(u, v) − C(u, 1) − C(1, v) ≥ 0, so C(u, v) ≥ u + v − 1 from which the theorem
follows.

2.12 Remark
The multivariate version of Theorem 2.11 is given by

n
!
X
W (u) := max ui − n + 1, 0 ≤ C(u) ≤ min (ui ) =: M (u) (2.12)
1≤i≤n
i=1

for u = (u1 , . . . , un ) ∈ I n , n ≥ 2, C ∈ Cn .
Although the lower bound W is never a copula4 for n ≥ 3, the left-hand inequality in
(2.12) is “best-possible” in the sense that for any n ≥ 3 and any u ∈ I n , there is a n-copula
C such that C(u) = W (u) (see Nelsen, 1999, Theorem 2.10.12). 
Qn
By Π(u) := i=1 ui , u ∈ I n , we denote the so-called independence copula.

For a stochastic interpretation of the functions M , W and Π we need the following classic
result of Sklar (1959, 1973) which shows the relationship between copulas and random
variables and their distribution functions, respectively. This will explain the interest in the
study of copulas which also underlies this thesis.

4 This can be seen from the fact that V ((1/2, 1]) = 1 − n/2 with (1/2, 1] := (1/2, 1] × · · · × (1/2, 1] so that W is not
W
grounded.
 10 2. Copulas and Their Properties

W 1 1 1

M
0 0 0
1 1 1
v

v
1 1 1
00 u 00 u 00 u

1.0 1.0 1.0

Sfrag replacements
0.5 0.5 0.5

0.0 0.0 0.0

0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0

FIGURE 2.1. Lower Fréchet bound, independence copula and upper Fréchet bound; perspective
and related contour plots for the case n = 2.

2.13 Theorem (Sklar, 1959)

Let H be a n-dimensional distribution function with one-dimensional marginals F1 , . . . , Fn .
n
Then there exists a n-copula C such that for (x1 , . . . , xn ) ∈ ,


H(x1 , . . . , xn ) = C F1 (x1 ), . . . , Fn (xn ) . (2.13)
If F1 , . . . , Fn are all continuous, then C is unique; otherwise, C is uniquely determined on
ran F1 × · · · × ran Fn .
Conversely, if C is a n-copula and F1 , . . . , Fn are distribution functions on , then
the function H defined by (2.13) is a n-dimensional distribution function with marginals
F1 , . . . , Fn .
For a proof, see Nelsen (1999, Theorem 2.10.9) and the references therein.

In Figure 2.1, plots of the Fréchet bounds and the independence copula are given. The
geometric interpretation of Theorem 2.11 is that the surface of any copula of two random
variables must lie between the surfaces of the lower and the upper Fréchet bound. Every
kind of dependence between two random variables has a corresponding copula with a surface
which can be identified in the unit cube.
2.14 Remark (cf. Nelsen, 1999, Section 2.5)
With the theorem of Sklar, the copulas W , Π and M have the following stochastic inter-
pretations:
Two random variables U, V distributed uniformly on (0, 1) have the joint distribution
function
 2.1 Basic Properties 11

• M iff (if and only if) U is almost surely (a. s.) an increasing function of V :
U = f (V ) for some monotone, increasing function f : → ,

• W iff U is a. s. a decreasing function of V :

U = g(V ) for some monotone, decreasing function g : → ,

• Π iff U and V are independent.

These properties were firstly observed by Hoeffding (1940) and Fréchet (1957). 
We may now ask: Which are the copulas of monotone transformations of the underlying
random variables? The answer is given in the following theorem (see Nelsen, 1999, Theorems
2.4.2 and 2.4.3):
2.15 Theorem
Let X and Y be continuous random variables with copula CX,Y , and let f and g be real-
valued functions on ran(X) and ran(Y ), respectively.
a) If f and g are strictly increasing, then Cf (X),g(Y ) (u, v) = CX,Y (u, v) for all u, v ∈ I.
b) If f is strictly increasing and g is strictly decreasing, then Cf (X),g(Y ) (u, v) = u −
CX,Y (u, 1 − v) for all u, v ∈ I.
c) If f is strictly decreasing and g is strictly increasing, then Cf (X),g(Y ) (u, v) = v −
CX,Y (1 − u, v) for all u, v ∈ I.
d) If f and g are both strictly decreasing, then Cf (X),g(Y ) (u, v) = u + v − 1 + CX,Y (1 −
u, 1 − v) for all u, v ∈ I.
This theorem will be needed later to determine the copula of continuous local martin-
gales.

We will now see that property (2.7) guarantees that C induces a Lebesgue-Stieltjes-
measure on (I n , B n ∩ I n ), where B n denotes the Borel-σ-algebra.
2.16 Theorem
Let C be a n-copula. Then C induces a unique probability measure PC on (I n , B n ∩ I n ).
Proof. As C is a distribution function on I n , the fact follows easily from Shiryaev (1996,
Theorem II.3.2, p. 160) or Billingsley (1995, Example 1.1, p. 9).
However, it is not true that every probability measure P on I n is induced by a copula
C.5 For a measure to be induced by a copula, it must spread mass in a manner consistent
with the boundary conditions of a copula. Therefore, any probability measure on I n where
at least one one-dimensional margin is not uniformly distributed cannot be induced by a
copula.
In detail, we have:

5 This means that for a rectangle Ru being spanned by 0 and u (both in I n ), there is always a n-copula C such that
0
P(R0u ) = C(u) holds for all u ∈ I n .
12 2. Copulas and Their Properties

2.17 Corollary
A probability measure P on (I n , B n ∩ I n ) is induced by a copula C ∈ Cn if and only if


P (0, xi ] = xi , for all xi ∈ I, 1 ≤ i ≤ n,

with 0 := (0, . . . , 0), xi := (1, . . . , 1, xi , 1, . . . , 1).

Proof. If P is induced by a copula C we must have P (0, xi ] = C(1,
. . . , 1, xi , 1, . . . , 1) = xi
due to (2.6) for all xi ∈ I. On the other hand, if we have P (0, xi ] = xi , we can define the
copula C by C(u) := P((0, u]), u ∈ I n , and easily check the definition.

The following theorem concerning the denseness of the subset of absolutely continuous
copulas in the space of all copulas was firstly stated for the case n = 2 in Darsow et al.
(1992). For a proof, the authors refer to an oral presentation given at their institute. To
make things more rigorous, we state the theorem for the general case and give a detailed
proof.
As copulas are uniformly continuous on I n , the topology of uniform convergence and
its induced supremum metric dsup provide a natural framework for measuring distances
between copulas.
For notational convenience, λn will be used for the usual Lebesgue measure on ( n , B n )
and the Lebesgue measure restricted on I n , i. e., (I n , B n ∩ I n ). The meaning depends on
the context; where confusion could arise, additional remarks will be made.

2.18 Theorem
Let Cn denote the set of all n-copulas and Cλn ⊂ Cn the subset of all copulas which are
absolutely continuous with respect to (w. r. t.) the Lebesgue measure λn restricted on I n .
Then Cλn is dense in Cn w. r. t. the supremum metric dsup , i. e., the metric generating the
uniform topology on I n .

Proof. Let dsup denote the supremum metric on Cn , i. e., for C1 , C2 ∈ Cn ,

dsup (C1 , C2 ) := kC1 − C2 k := sup |C1 (x) − C2 (x)|,

x∈I n

where k · k denotes the corresponding supremum norm on I n .

As we know that copulas are continuous on the compact set I n , we could replace “sup”
by “max”.
Now, let C ∗ ∈ Cn be a copula and ε > 0 arbitrary. We need to construct a copula C ∈ Cn
with kC − C ∗ k < ε whose induced measure PC (see Theorem 2.16) is absolutely continuous
with respect to the Lebesgue measure λn |I n . Let PC ∗ denote the unique probability measure
on (I n , B n ∩ I n ) induced by C ∗ . For some N ∈ which we will choose later, let

   
i1 − 1 i1 in − 1 in
Si1 ,...,in := , × ··· × , , 1 ≤ ij ≤ N, ij ∈ , 1 ≤ j ≤ n, (2.14)
N N N N


 2.1 Basic Properties 13

be n-dimensional cubes (of which there are N n ) so that

X
(0, 1]n = Si1 ,...,in , (2.15)
1≤i1 ,...,in ≤N

and let X
RiN1 ,...,in := Sj1 ,...,jn (2.16)
1≤jk ≤ik
1≤k≤n

be the n-dimensional rectangle defined by the points (0, . . . , 0) and (i1 /N, . . . , in /N ).
N is chosen such that N > n/ε.
Now, for 1 ≤ i1 , . . . , in ≤ N , let C in a first step be defined on the points (i1 /N, . . . , in /N )
by    
i1 in N ∗ i1 in
C ,..., := PC ∗ (Ri1 ,...,in ) = C ,..., . (2.17)
N N N N
Note that the C-volume of any rectangle of the form R = (0, i1 /N ] × · · · × (0, in /N ] can
be determined by (2.16) from these values, i.e.,
X
VC (RiN1 ,...,in ) = VC (Sj1 ,...,jn ).
1≤jk ≤ik
1≤k≤n

n
Recall that using the notation for the n-th order difference operator ∆ba for a, b ∈ we
have
(i /N,...,in /N )
VC (Si1 ,...,in ) = ∆((i11 −1)/N,...,(i n −1)/N )
C(u1 , . . . , un ),
and this is well defined by (2.17).
In a second step, let us now define C by a λn -density in the following way which guar-
antees that the mass of C is spread uniformly on the cubes Si1 ,...,in :

dC X
n
(x , . . . , x ) ≡ c(x , . . . , x ) := N Si1 ,...,in (x) · VC (Si1 ,...,in ), (2.18)


1 n 1 n
dλn 1≤i ,...,i ≤N
1 n

where x := (x1 , . . . , xn ) ∈ I n . Obviously, c(x) ≥ 0 for all x ∈ I n and

Z Z X Z
dC n n n
dC = dλ = N VC (Si1 ,...,in ) Si1 ,...,in (x) dλ (x)


dλn 1≤i1 ,...,in ≤N

In In In
| {z }
=N −n
X (2.17)
= VC (Si1 ,...,in ) = VC (I n ) = VC ∗ (I n ) = 1.
1≤i1 ,...,in ≤N

Then C is absolutely continuous w. r. t. λn by definition.

As a third step, we still need to show that C indeed is a copula and kC − C ∗ k < ε.
That C is a copula can be seen as follows:
14 2. Copulas and Their Properties

• C is defined on I n , if xi = 0 for some i ∈ {1, . . . , n}, then, by (2.18) and xi not

belonging to any of the Si1 ,...,in , we get C(x1 , . . . , xn ) = 0. Therefore, C is grounded.

• Let ui = k/N for some 1 ≤ k ≤ N , then

(2.17)
C(1, . . . , 1, ui , 1, . . . , 1) = C ∗ (1, . . . , 1, ui , 1, . . . , 1) = ui

as C ∗ is a copula by assumption. Let ui 6= k/N for all 0 < k ≤ N . Then

ui := (1, . . . , 1, ui , 1, . . . , 1) ∈ SN,...,N,ji ,N,...,N for some 1 ≤ ji ≤ N,

and we have
 
N N N N
C(ui ) = C , . . . , , ui , , . . . ,
N N N N
Z1 Zui Z1 X
(2.18)
= ... ... Nn Si1 ,...,in (x) · VC (Si1 ,...,in ) dλn (x)


0 0 0 1≤i1 ,...,in ≤N

ji −1

X hZ1 ZN Z1
n n
= N VC (Si1 ,...,in ) ... ... Si1 ,...,in (x) dλ (x)


1≤i1 ,...,in ≤N 0 0 0
Z1 Zui Z1 i
n
+ ... ... Si1 ,...,in (x) dλ (x)


0 ji −1 0
N

X  n
n 1
= N VC (Sk1 ,...,kn ) ·
1≤k1 ,...,kn ≤N
N
ki ≤ji −1
X   n−1
n ji − 1 1
+ N VC (Sk1 ,...,kn ) ui −
1≤k1 ,...,kn ≤N
N N
ki =ji

X   X
ji − 1
= VC (Sk1 ,...,kn ) + N ui − VC (Sk1 ,...,kn )
1≤k1 ,...,kn ≤N
N 1≤k ,...,k 1 n ≤N
ki ≤ji −1 ki =ji
 
(2.17) ∗ ji − 1 ji − 1
= C (1, . . . , 1, , 1, . . . , 1) + N ui −
N N
 
∗ ji ∗ ji − 1
· C (1, . . . , 1, , 1, . . . , 1) − C (1, . . . , 1, , 1, . . . , 1)
N N
  
C ∗ is copula ji − 1 ji − 1 ji ji − 1
= + N ui − − = ui .
N N N N
 2.1 Basic Properties 15

• Trivially, the C-volume of any rectangle is non-negative by (2.18).

Therefore, C is a copula.

As a last step, let x = (x1 , . . . , xn ) ∈ I n with xi 6= 0 for all 1 ≤ i ≤ n. (Remember

that if xi = 0 for some i ∈ {1, . . . , n}, C ∗ (x) = C(x) = 0.) Then x ∈ Si1 ,...,in for some
(i1 , . . . , in ) ∈ {1, . . . , N }n . For v ∗ := argmin{kx − vk1 : v ∈ vert(Si1 ,...,in )}, we get
n
X n ε
kx − v ∗ k1 := |xi − vi∗ | ≤ < . (2.19)
| {z } 2N 2
i=1
≤(2N )−1

Therefore,

|C(x) − C ∗ (x)| = |C(x) − C(v ∗ ) + C(v ∗ ) − C ∗ (v ∗ ) + C ∗ (v ∗ ) − C ∗ (x)|

≤ |C(x) − C(v ∗ )| + |C(v ∗ ) − C ∗ (v ∗ )| +|C ∗ (v ∗ ) − C ∗ (x)|
| {z }
= 0 by construction

(2.10) (2.19)
≤ 2kx − v ∗ k1 < ε,

thus
kC − C ∗ k < ε,
which completes the proof.

2.19 Remark
The same construction idea is used for the bivariate case by Li et al. (1998) and Kulpa
(1999), who call it a “checkerboard approximation”. 
16 2. Copulas and Their Properties

2.2 The Role of Partial Derivatives

In this section we will take a closer look at the role played by the partial derivatives of a
copula C. We will see that there is a very strong connection between the partial derivatives
and conditional distributions related to C. To give a heuristic motivation of what may be
the interpretation of the partial derivatives, consider P(X ≤ x | Y = y) for two random
variables X and Y with distribution functions FX and FY where FY is continuous and
strictly increasing and with copula C and joint cumulative distribution function FXY .6 If
P(X ≤ x | Y = y) is assumed to be continuous from the right w. r. t. y, we have

P(X ≤ x | Y = y) = lim P(X ≤ x | y ≤ Y ≤ y + h)

h&0

FXY (x, y + h) − FXY (x, y)

= lim
h&0 FY (y + h) − FY (y)



C FX (x), FY (y + h) − C FX (x), FY (y)
= lim
h&0 FY (y + h) − FY (y)



C FX (x), FY (y) + ∆(h) − C FX (x), FY (y)
= lim
h&0 ∆(h)
∂ 


= C(u, v)
∂v FX (x),FY (y)

with ∆(h) := FY (y+h)−FY (y) wherever the derivative exists. This is the basic idea behind
the interpretation of partial derivatives which will be put in a more rigorous framework
hereafter.

For a n-copula C ∈ Cn , denote the n first partial derivatives of C by

∂
Dk C(u1 , . . . , un ) := C(u1 , . . . , un ), 1 ≤ k ≤ n, ui ∈ (0, 1), 1 ≤ i ≤ n. (2.20)
∂uk
Since monotone functions are differentiable almost everywhere, it follows that for any
k ∈ {1, . . . , n}, ui ∈ I, i 6= k, Dk C(u1 , . . . , un ) exists for almost all uk ∈ (0, 1), and
Corollaries 2.8 and 2.9 (for n dimensions) yield

0 ≤ Dk C(u1 , . . . , un ) ≤ 1 almost surely. (2.21)

Analogously, by Lemma 2.7 and Corollary 2.8, the functions in Lemma 2.7 have deriva-
tives almost everywhere, and
 
Dk C(u1 , . . . , un )u =y − Dk C(u1 , . . . , un )u =x ≥ 0 (2.22)
k k

if 0 ≤ x ≤ y ≤ 1, ui ∈ (0, 1).

6 We recall the definition P(X ≤ x | Y ) := E( | Y ).



{X≤x}
 2.2 The Role of Partial Derivatives 17

2.20 Notation
As we will need the operator Dk quite often, we will use the notation

∂ 

Dk C(u1 , . . . , uk−1 , x, uk+1 , . . . , un ) := C(u1 , . . . , un ) . (2.23)
∂uk uk =x

It follows that the functions x → Dk C(u1 , . . . , uk−1 , x, uk+1 , . . . , un ) are defined and
increasing almost everywhere.

Before formulating a lemma which relates the partial derivatives to conditional expecta-
tions, we need the following result from Elstrodt (1996) which—in connection with (2.21)
and the above remarks—gives us the integrability of the partial derivatives.
2.21 Theorem (Elstrodt, 1996, Satz 6.1, p. 150)
Let K ∈ { , } and f : [a, b] → K (a, b ∈ p , a < b, p ∈ ) be a bounded function.


Then f is Riemann-integrable if and only if the set of its points of discontinuity has λp -
measure 0. Then, the Riemann-integral and the Lebesgue-integral are identical.

We can now formulate the basic result connecting partial derivatives of copulas with
conditional distribution functions.
2.22 Lemma
Let X and Y be two real-valued random variables on the same probability space (Ω, A, P)
with corresponding copula C and continuous marginals FX , FY . Define the conditional
probabilities

P(X ≤ x | Y ) := E( | Y ) and P(Y ≤ y | X) := E( | X) (2.24)

 

{X≤x} {Y ≤y}

with x, y ∈ as usual 7 . Then,

D2 C FX (x), FY (Y ) is a version of P(X ≤ x | Y ) and

(2.25)
D1 C FX (X), FY (y) is a version of P(Y ≤ y | X).

Proof. (Cf. Darsow et al., 1992) We will only consider the first case, as the second one is
completely analogous. Let σ(Y ) denote the σ-Algebra generated by Y and consider the

Y
setting Ω, σ(Y ) −→ ( 1 , B 1 ). We have to show
Z Z Z
!

P(X ≤ x | Y ) dP = {X≤x} dP = D2 C FX (x), FY (Y ) dP (2.26)


A A A
R
R all A ∈ σ(Y ), x ∈
. With the definitions Qx (A) := A P(X ≤ x | Y ) dP and Q̃x (A) :=
for
A
D2 C FX (x), FY (Y ) dP for A ∈ σ(Y ) we therefore must show

Qx (A) = Q̃x (A) for all A ∈ σ(Y ), x ∈ . (2.27)

7 See Billingsley, 1995, Sections 34 and 35, or Rogers and Williams, 2000a, II.40–42.
18 2. Copulas and Their Properties

Note that Qx is well defined due to the definition of conditional expectation and Q̃x is
well defined due to (2.21) and Theorem 2.21. It is elementary to show that Qx and Q̃x are
σ-finite measures on σ(Y ).
It is known that E := {(−∞, a] : a ∈ } is a σ-finite generator of B 1 w. r. t. the
Lebesgue-measure λ1 which is closed under intersections. From Bauer (1992, Satz 5.4) or
Rogers and Williams (2000a, Corollary II.4.7) it follows that we only need to show (2.27)
for a σ-finite generator of σ(Y ) which is closed under intersections. We now have
(∗)
(∗∗)

σ(Y ) = Y −1 (B 1 ) = Y −1 σ(E) = σ Y −1 (E) ,

where (∗) is valid due to Billingsley (1995, Theorem 20.1, p. 255) and (∗∗) due to Elstrodt
(1996, Satz 4.4, p. 19).

So we see that E 0 := {Y −1 (E) : E ∈ E} is a generator of σ(Y ) which is closed under

intersections (as E10 ∩ E20 = Y −1 (E1 ) ∩ Y −1 (E2 ) = Y −1 (E1 ∩ E2 ) ∈ E 0 for some E1 , E2 ∈ E).


Therefore, we can assume the set A to be of the form A = Y −1 (−∞, a] for some a ∈ .
With Theorem 2.21, we have
Z Z



D2 C FX (x), FY (Y ) dP = D2 C FX (x), FY (ξ) dPY (ξ)
| {z }
A (−∞,a] dFY (ξ) in Lebesgue-Stieltjes sense
Z
(∗)

= D2 C FX (x), η dη
(0,FY (a)]
Z Z



= C FX (x), FY (a) − C FX (x), 0 = · dP = dP.
  

{X≤x} {Y ≤a} {X≤x}

| {z }
=0 Ω A

(∗) is due to Kamke (1960, Satz 1, p. 164) with η = FY (ξ) and the continuity of FY in
combination with Lemma A.2.
This is exactly (2.27) from which the assertion follows.

2.23 Remark
In view of Lemma A.3 we see that FX (X), FY (Y ) ∼ U(0, 1). Thus, (2.25) is equivalent to:


D2 C FX (x), V is a version of P(X ≤ x | Y ) and

(2.28)
D1 C U, FY (y) is a version of P(Y ≤ y | X),

where U, V ∼ U(0, 1) have copula C and are defined on the same probability space as X
and Y . This means that for P(X ≤ x | Y ), say, we only need the dependence information
via C and the marginal FX . 
A statement analogous to Lemma 2.22 holds for the general case of n random variables
and can be formulated as follows:
 2.2 The Role of Partial Derivatives 19

2.24 Corollary
Let X1 , . . . , Xn be real-valued random variables on the same probability space (Ω, A, P)
with corresponding copula C and continuous marginals F1 , . . . , Fn .8 Then, for any k ∈
{1, . . . , n},


Dk C F1 (x1 ), . . . , Fk−1 (xk−1 ), Fk (Xk ), Fk+1 (xk+1 ), . . . , Fn (xn )

is a version of (2.29)


P Xi ≤ xi , 1 ≤ i ≤ n, i 6= k | Xk

for all x1 , . . . , xn ∈ .

We will omit the proof as it corresponds to the one in Lemma 2.22.

The next goal should be to connect partial derivatives of n-copulas to expressions such
as P(Xn ≤ xn | X1 , . . . , Xn−1 ). This would be particularly helpful when dealing with
modelling conditionally specified time series by copulas. Indeed we can show more than
Lemma 2.22: If the higher partial derivatives of a n-copula exist and are continuous, they
are related to conditional probabilities in an analogous way as exemplified in Lemma 2.22.
Before stating the corresponding lemma let us introduce a notation for higher-order partial
derivatives of n-copulas as a generalization of (2.20):

2.25 Notation
Let C be a n-copula. For k ∈ {1, . . . , n}, we define

∂k 

D1,...,k C(x1 , . . . , xn ) := C(u1 , . . . , un ) (2.30)
∂u1 · · · ∂uk (x1 ,...,xn )

for all (x1 , . . . , xn ), (u1 , . . . , un ) ∈ (0, 1)n where the derivative exists. For k = n, we write

c1,...,n (x1 , . . . , xn ) := D1,...,n C(x1 , . . . , xn )

as this is the probability density function of (U1 , . . . , Un ) where the Ui ∼ U(0, 1) are con-
nected by the copula C.
More generally, let J ⊂ {1, . . . , n} be such that J = {j1 , . . . , jk }, ji 6= jl for i 6= l. Then,

∂k 

DJ C(x1 , . . . , xn ) := C(u1 , . . . , un ) . (2.31)
∂uj1 ∂uj2 · · · ∂ujk (x1 ,...,xn )

2.26 Remark
Due to our convention, expressions such as D1,2 C(0.2, 0.2, 0.4) make sense and are well
defined. 
8 Note that for a specific k we only need the continuity of the marginal Fk for the statement to hold.
20 2. Copulas and Their Properties

We can formulate the following theorem for uniformly distributed random variables. The
general expression will follow as a corollary.

2.27 Theorem
Let U = (U1 , . . . , Un ) be n real-valued random variables with uniform marginal distri-
butions U(0, 1), n ∈ , and let C1,...,n = ChU i, C1,...,n−1 = ChU1 , . . . , Un−1 i where Ch · i

denotes the corresponding copula. Assume that C1,...,n is (n − 1) times partially differen-
tiable w. r. t. the first (n − 1) arguments and the partial derivatives are continuous in these
arguments. Then,

D1,...,n−1 C1,...,n (U1 , . . . , Un−1 , un )

is a version of P(Un ≤ un | U1 , . . . , Un−1 ) (2.32)
D1,...,n−1 C1,...,n−1 (U1 , . . . , Un−1 )

for all un ∈ (0, 1).

Proof. The proof is essentially analogous to the one of Lemma 2.22. However, care has
to be taken of the fact that, e. g., D1,...,n−2 C1,...,n (0, u2 , . . . , un ) is not necessarily equal to
0.9 As we will see, we do not need this either. We will follow the outline of the proof of
Lemma 2.22.

Let σ(U1 , . . . , Un−1 ) denote the σ-algebra generated by U1 , . . . , Un−1 . We need to show
that

Z Z
D1,...,n−1 C1,...,n (U1 , . . . , Un−1 , un )
P(Un ≤ un | U1 , . . . , Un−1 ) dP = dP
D1,...,n−1 C1,...,n−1 (U1 , . . . , Un−1 )
A A

for all un ∈ (0, 1), A ∈ σ(U1 , . . . , Un−1 ). As En−1 := {(0, a] : a = (a1 , . . . , an−1 ) ∈ (0, 1)n−1 }
is a σ-finite generator of B n−1 ∩ (0, 1)n−1 which is closed under intersections, we can assume
A to be of the form



A = U1−1 (0, a1 ] ∩ · · · ∩ Un−1
−1
(0, an−1 ]

(the same theorems as in the previous proof apply). Then,

9 This
Qn
can be easily seen by Π(u1 , . . . , un ) = i=1 ui : We have D1,...,n−2 Π(0, u2 , . . . , un ) = un un−1 which is greater than
0 for u1 , u2 ∈ (0, 1).
 2.2 The Role of Partial Derivatives 21

Z
D1,...,n−1 C1,...,n (U1 , . . . , Un−1 , un )
dP
D1,...,n−1 C1,...,n−1 (U1 , . . . , Un−1 )
A
Z
D1,...,n−1 C1,...,n (u1 , . . . , un−1 , un )
= dC1,...,n−1 (u1 , . . . , un−1 )
D1,...,n−1 C1,...,n−1 (u1 , . . . , un−1 )
(0,a]
Z
(∗)
= D1,...,n−1 C1,...,n (u1 , . . . , un−1 , un ) du1 · · · dun−1
(0,a]
Z
Fubini 
= D1,...,n−2 C1,...,n (u1 , . . . , un−2 , an−1 , un )
(0,a1 ]×···×(0,an−2 ]

− D1,...,n−2 C1,...,n (u1 , . . . , un−2 , 0, un ) dun−2 · · · du1
Z

= D1,...,n−3 C1,...,n (u1 , . . . , un−3 , an−2 , an−1 , un )
(0,a1 ]×···×(0,an−3 ]

− D1,...,n−3 C1,...,n (u1 , . . . , un−3 , 0, an−1 , un )

− D1,...,n−3 C1,...,n (u1 , . . . , un−3 , an−2 , 0, un )

+ D1,...,n−3 C1,...,n (u1 , . . . , un−3 , 0, 0, un ) dun−3 · · · du1
X
= ··· = (−1)δ(b1 ,...,bn−1 ) C1,...,n (b1 , . . . , bn−1 , un )
bi ∈{0,ai }
1≤i≤n−1
Z n−1
Y Z
(∗∗)
= C1,...,n (a1 , . . . , an−1 , un ) = dP = dP
  

{Un ≤un } {Ui ≤ai } {Un ≤un }

Ω i=1 A

Pn−1
where δ(b1 , . . . , bn−1 ) := i=1 {0} (bi ) counts the number of zeros.


(∗) follows from Kamke (1960, Satz 8, p. 177) applied to n − 1 dimensions.

(∗∗) is due to the fact that C1,...,n (b1 , . . . , bn−1 , un ) = 0 if any of the bi is 0 (groundedness,
see (2.5)).

The following corollary is proven completely analogously so that we will omit its proof.

2.28 Corollary
Let X1 , . . . , Xn , n ∈ , be real-valued, continuous random variables with distribution

functions F1 , . . . , Fn and copula C1,...,n . For k ∈ {1, . . . , n}, assume that C1,...,n is k-times
partially differentiable w. r. t. the first k arguments and the partial derivative is continuous
in these arguments, i. e., D1,...,k C1,...,n (u1 , . . . , un ) exists and is continuous in u1 , . . . , uk .
Further, let C1,...,k be the copula of (X1 , . . . , Xk ) and c1,...,k the associated density func-
tion, i. e., c1,...,k (u1 , . . . , uk ) = D1,...,k C1,...,k (u1 , . . . , uk ). The marginal densities fi (x) :=
22 2. Copulas and Their Properties

d
F (x)
dx i
are assumed to exist for 1 ≤ i ≤ k. Then,


D1,...,k C1,...,n F1 (X1 ), . . . , Fk (Xk ), Fk+1 (xk+1 ), . . . , Fn (xn )

Qk
c1,...,k F1 (X1 ), . . . , Fk (Xk ) i=1 fi (Xi )
(2.33)
is a version of
P(Xk+1 ≤ xk+1 , . . . , Xn ≤ xn | X1 , . . . , Xk )
for all xk+1 , . . . , xn ∈ .
2.29 Remark
It would be possible to formulate an even more general form by allowing mixed partials.
I. e., for I ⊂ {1, . . . , n} we could express P(Xi ≤ xi , i ∈ I | Xj , j ∈ {1, . . . , n} \ I) with
partials of copulas (cf. equation (2.31) on page 19). As this is notationally inconvenient,
we will omit this. 

2.3 Archimedean Copulas

Archimedean copulas are a special class of copulas which find a wide range of applications
for a number of reasons: They can be easily constructed, a great variety of families of
copulas belongs to that class and they have many desirable properties.
Archimedean copulas originally appeared in the study of probabilistic metric spaces,
where they were part of the development of a version of the triangle inequality. The term
“Archimedean” seems to be derived from the fact that there is a very close connection
between these copulas and t-norms used to define the triangle inequality. These t-norms
induce a measure of distance between two points of the probabilistic metric space (see
Schweizer, 1991, for details and the references cited therein).
The source of all Archimedean copulas is a so-called generator which is a continuous,
strictly decreasing function ϕ from I to [0, ∞] such that ϕ(1) = 0.10 The following theorem
is elementary for the construction of Archimedean copulas:
2.30 Theorem (Nelsen, 1999, Theorem 4.1.4, p. 91)
Let ϕ be a continuous, strictly decreasing function from I to [0, ∞] such that ϕ(1) = 0,
and let ϕ[−1] be the generalized inverse of ϕ defined by

ϕ−1 (t), 0 ≤ t ≤ ϕ(0),
[−1]
ϕ (t) = (2.34)
0, ϕ(0) < t ≤ ∞.

Then the function C : I 2 → I given by

C(u, v) = ϕ[−1] ϕ(u) + ϕ(v) (2.35)

is a copula if and only if ϕ is convex. ϕ is called a generator of the copula.

10 Note that ϕ will also stand for the density of a standard normal random variable. As it is common to denote the generator

by ϕ we will stick to that convention as the meaning is always clear from the context.
 2.3 Archimedean Copulas 23

2.31 Remark
a) The original proof of this theorem was given by Alsina et al. (2001). As this mono-
graph is not yet available, we refer the reader to the proof in Nelsen (1999).
b) If ϕ(0) = ∞, ϕ is called a strict generator and ϕ[−1] = ϕ−1 , i. e., the usual inverse.
The induced copula is also called strict.
c) The generator ϕ of an Archimedean copula is not unique, as cϕ, c > 0, yields the
same copula as ϕ.


2.32 Example
a) For ϕ(t) = − ln(t), t ∈ (0, 1], we have ϕ(0) = limt&0 ϕ(t) = ∞, therefore ϕ is strict
with ϕ−1 (t) = exp(−t). The induced copula is given by
C(u, v) = exp(−[− ln u − ln v]) = uv = Π(u, v).
Thus, the independence copula is a strict Archimedean copula.
b) The lower Fréchet bound W (u, v) = max(0, u + v − 1), u, v ∈ I, is Archimedean:
Choosing ϕ(t) = 1 − t, 0 ≤ t ≤ 1, yields W .
However, the upper Fréchet bound M is not Archimedean. This is due to the result
that the diagonal δC (u) := C(u, u) of any Archimedean copula C meets the condition
δC (u) < u for all u ∈ (0, 1) (which in combination with associativity11 is characteriz-
ing for Archimedean copulas, see Ling, 1965). As δM (u) = u, this condition cannot
be met by M .
 
c) Setting ϕϑ (t) = − ln exp(−ϑt)−1
exp(−ϑ)−1
, ϑ ∈ \ {0}, yields ϕ(0) = ∞ so that a strict
copula family given by
 
1 (exp(−ϑu) − 1)(exp(−ϑv) − 1)
Cϑ (u, v) = − ln 1 + , ϑ ∈ \ {0}, (2.36)
ϑ exp(−ϑ) − 1
is generated. This is the Frank family which will prove to be important in the context
of symmetric Markov processes (see Section 5.4.2).

Many classes of Archimedean copulas and some of their properties are described in Nelsen
(1999, pp. 94–97).
[−1] −1 −1


For 0 ≤ t ≤ ϕ(0) we have

ϕ (t) = ϕ (t), so that C(u, v) = ϕ ϕ(u) + ϕ(v) which
is equivalent to ϕ C(u, v) = ϕ(u) + ϕ(v) for ϕ(u) + ϕ(v) ≤ ϕ(0). So we get
∂


ϕ C(u, v) = ϕ0 C(u, v) · D1 C(u, v) = ϕ0 (u) and
∂u
∂


ϕ C(u, v) = ϕ0 C(u, v) · D2 C(u, v) = ϕ0 (v),
∂v
11 A 2-copula C is associative if C(C(u, v), w) = C(u, C(v, w)) holds for all u, v, w ∈ I.
24 2. Copulas and Their Properties

thus

ϕ0 (u)
D1 C(u, v) =
,
ϕ0 ϕ−1 (ϕ(u) + ϕ(v))
ϕ0 (v)
D2 C(u, v) = 0 −1

ϕ ϕ (ϕ(u) + ϕ(v))

for ϕ(u) + ϕ(v) ≤ ϕ(0) (i. e., for all u, v if ϕ is strict).

If C is absolutely continuous, ϕ is twice differentiable almost everywhere, and we can
derive the density of C:


∂ ϕ0 (u) ϕ0 (u)ϕ00 C(u, v) D2 C(u, v)
c(u, v) = D1,2 C(u, v) = D2 D1 C(u, v) =
=−
∂v ϕ0 C(u, v) (ϕ0 (C(u, v)))2


ϕ0 (u)ϕ00 C(u, v) ϕ0 (v) ϕ0 (u)ϕ0 (v)ϕ00 (ϕ−1 (ϕ(u) + ϕ(v)))
=− = −
(ϕ0 (C(u, v)))3 (ϕ0 (ϕ−1 (ϕ(u) + ϕ(v))))3

for u, v ∈ (0, 1).

The level curve of a copula is the set Lt ≡ L(t) := {(u, v) ∈ I 2 : C(u, v) = t}, t ∈ I. For
an Archimedean copula, L(t) = {(u, v) ∈ I 2 : ϕ(t) = ϕ(u) + ϕ(v)} ⊂ I 2 which connects
the points (1, t) and (t, 1) (as ϕ(1) = 0). We can write the level curve as Lt (u) = v since
solving for v as a function of u yields



v = Lt (u) = ϕ[−1] ϕ(t) − ϕ(u) = ϕ−1 ϕ(t) − ϕ(u)

where the last step is justified because ϕ(t) − ϕ(u) is in the interval [0, ϕ(0)).
It can be shown (Nelsen, 1999, Theorem 4.3.2, p. 39) that the level curves Lt (u) are
convex for all t ∈ [0, 1).
An important theorem concerning applications such as estimation (cf. Genest and Rivest,
1993) and random variate generation (see Section B.2) can be obtained by finding the C-
measure of the region in I 2 lying on or below and to the left of each level curve.

2.33 Theorem (Nelsen, 1999, Theorem 4.3.4, pp. 101)

Let C be an Archimedean copula generated by ϕ ∈ Ω := {f : I → [0, ∞], f (1) = 0, f
continuous, strictly monotone decreasing and convex}.SFor t ∈ I, let KC (t) ≡ Kϕ (t) denote
2
the C-measure of the set {(u, v) ∈ I : C(u, v) ≤ t} = s∈[0,t] Ls , or equivalently, of the set
{(u, v) ∈ I 2 : ϕ(u) + ϕ(v) ≥ ϕ(t)}. Then, for t ∈ I,

ϕ(t)
Kϕ (t) = t − . (2.37)
ϕ0 (t+)

The following corollary presents a probabilistic interpretation of the above theorem which
will be useful in random variate generation (it is also useful for dealing with Kendall’s τ
for Archimedean copulas).
 2.4 Survival Copula and Symmetry Concepts 25

2.34 Corollary (Nelsen, 1999, Cor. 4.3.6, pp. 103)

Let U, V ∼ U(0, 1) whose joint distribution function is the Archimedean copula C generated
by ϕ ∈ Ω. Then the function Kϕ given by (2.37) is the distribution function of the random
variable C(U, V ). Furthermore, the joint distribution function of U and C(U, V ) is given
by
(
s, s ≤ t,
KC0 (s, t) = ϕ(t)−ϕ(s) (2.38)
t − ϕ0 (t+) , s > t.

The next theorem which is due to Genest and Rivest (1993) is an extension of Corol-
lary 2.34; an application is the algorithm for generating random variates from distributions
with Archimedean copulas (see Section B.2).

2.35 Theorem
Under the hypotheses of
Corollary 2.34, the joint distribution function H(s, t) of
S = ϕ(U )/ ϕ(U ) + ϕ(V ) and T = C(U, V ) is given by s · KC (t) for all s, t ∈ I. Hence S
and T are independent, and S is uniformly distributed on (0, 1).

2.36 Remark
a) The function Kϕ (z) = P(C(U, V ) ≤ z), z ∈ I, can be estimated by means of its em-
(n)
pirical distribution function Kϕ (z): Given a sample {(x1 , y1 ), . . . , (xn , yn )} of size
n from
P (X, Y ), construct the empirical bivariate distribution function Hn (x, y) :=
n−1 ni=1 (−∞,x] (xi ) · (−∞,y] (yi ), x, y ∈ , of the sample. Then, calculate Hn (xi , yi ),
 

1 ≤ i ≤ n, and use these values to construct the one-dimensional empirical distri-

(n)
bution function Kϕ (z). However, the first step is unnecessary, as Hn (xi , yi ) is the
proportion of sample observations which are less than or equal to (xi , yi ) componen-
twise. For details, see Genest and Rivest (1993, pp. 1036).

(n)
b) Kϕ (z) can be fitted by the distribution function Kϑ (z) of any family of Archimedean
copulas, where the parameter ϑ is estimated in such a manner that the fitted distri-
bution has a coefficient of concordance (Kendall’s τ ) equal to the empirical coefficient
(see Genest and Rivest, 1993).


2.4 Survival Copula and Symmetry Concepts

In this section we introduce the survival copula and its relation to a symmetry con-
cept called radial symmetry which will prove to be useful when characterizing symmetric
stochastic processes later on.
26 2. Copulas and Their Properties

2.4.1 Symmetry Concepts

In a univariate setting the meaning of symmetry of X about a ∈ is clear:
X − a ∼ a − X, i. e., P(X − a ≤ x) = P(a − X ≤ x) for all x ∈
so that
F (a − x) = F (a + x) for all x ∈ , F continuous. (2.39)

But what does it mean to say that (X, Y ) is “symmetric” about some point (a, b) ∈ 2 or
more generally that (X1 , . . . , Xn ) is “symmetric” about (a1 , . . . , an ) ∈ n ? In the multivari-
ate setting there is a number of possibilities in which sense random vectors are symmetric.
We will shortly introduce three of them for the bivariate case and then generalize the most
important one (radial symmetry) to n dimensions which will be a key tool for characterizing
symmetric processes.

2.37 Definition (e. g., Nelsen, 1999, p. 31)

2
Let X and Y be real-valued random variables and (a, b) ∈ .
a) (X, Y ) is marginally symmetric about (a, b) if X and Y are univariately symmetric
about a and b, respectively, according to (2.39).
b) (X, Y ) is radially symmetric about (a, b) if the joint distribution function of (X − a)
and (Y − b) is the same as the joint distribution function of (a − X) and (b − Y ),
i. e.,
(X − a, Y − b) ∼ (a − X, b − Y ). (2.40)

c) (X, Y ) is jointly symmetric about (a, b) if the following four pairs of random vectors
have a common joint distribution:
(X − a, Y − b), (X − a, b − Y ), (a − X, Y − b), (a − X, b − Y ).

2.38 Remark
a) For X and Y we can express radial symmetry in terms of the joint distribution and
the univariate survival functions of X and Y in a manner analogous to (2.39) (see
below).
b) In the above definition we have c)⇒b)⇒a). It can be shown that jointly symmet-
ric random variables must be uncorrelated when the second order moments exist
(Randles and Wolfe, 1979). Thus, the concept is too strong for many applications
and we will focus on radial symmetry.
c) Marginal symmetry does not imply an intuitive bivariate symmetry as can be seen
from Figure 2.2. Although the margins are symmetric N(0, 1) and t3 -variables, respec-
tively, the bivariate density does not seem to be symmetric due to the unsymmetric
dependence structure.

 2.4 Survival Copula and Symmetry Concepts 27

3 −3
2 −2
1 −1
0.10 0 0.10 0

x
0.05 −1 0.05 1
0.00
−3 −2 −2 0.00
−3 −2 2
−1 0 −1 0
x 1 2 −3 y 1 2 3
3 3

−3 3
−2 2
−1 1
0.10 0 0.10 0
y

x
0.05 1 0.05 −1
0.00
3 2 0.00
3 −2
2 1 2 1
0 −1 −2 3 0 −1 −2 −3
x −3 y −3
1.0
0.8
0.6
0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

FIGURE 2.2. Bivariate distribution based on the min-max copula with parameter 3 and N(0, 1)-
and t3 -distributed marginals (perspective and contour plots).
28 2. Copulas and Their Properties

H(a − x, b − y)
(a − x, a − y)

(a, b)
PSfrag replacements

(a + x, b + y)
H(a + x, b + y)

FIGURE 2.3. Regions of equal probability for radially symmetric random variables.

Let us generalize Definition 2.37 b):

2.39 Definition (radial symmetry)
Let X1 , . . . , Xn be real-valued random variables, n ∈ , and (a1 , . . . , an ) ∈ n . The vector

(X1 , . . . , Xn ) is said to be radially symmetric about (a1 , . . . , an ) if the joint distribution

function of X1 − a1 , . . . , Xn − an is the same as the one of a1 − X1 , . . . , an − Xn , i. e.,
(X1 − a1 , . . . , Xn − an ) ∼ (a1 − X1 , . . . , an − Xn ). (2.41)
The following theorem easily connects radial symmetry with the survival function of the
random vector (X1 , . . . , Xn ):
2.40 Theorem
A random vector (X1 , . . . , Xn ) with joint distribution function H is radially symmetric
about (a1 , . . . , an ) ∈ n iff
H(a1 + x1 , . . . , an + xn ) = H(a1 − x1 , . . . , an − xn ) for all xi ∈ , 1 ≤ i ≤ n, (2.42)
where H(x1 , . . . , xn ) := P(X1 > x1 , . . . , Xn > xn ) is the survival function.
The proof is obvious.

We see that “radial” means that (a1 + x1 , . . . , an + xn ) and (a1 − x1 , . . . , an − xn ) lie

on rays emanating in opposite directions from (a1 , . . . , an ). For the bivariate case this is
visualized in Figure 2.3.

2.4.2 Survival Copula

The univariate survival function of a random variable X is given by F (x) = 1 − F (x),
x ∈ . Analogously, the bivariate survival function of two random variables X and Y with
distribution functions F and G, respectively, and joint distribution function H and copula
C can be defined to be
H(x, y) : = P(X > x, Y > y) = 1 − F (x) − G(y) + H(x, y)


= F (x) + G(y) − 1 + C 1 − F (x), 1 − G(y) .
 2.4 Survival Copula and Symmetry Concepts 29

Note that the marginals of H are the univariate survival functions, i. e., H(x, −∞) = F (x)
and H(−∞, y) = G(y).
If we define
b v) := u + v − 1 + C(1 − u, 1 − v),
C(u, u, v ∈ I, (2.43)
we have

b F (x), G(y) .
H(x, y) = C
b is called the survival copula of C (it is easy to check that C
C b is indeed a copula).
Note that C b is not the joint survival function C of two U(0, 1) distributed random
variables with copula C as we have C(u, v) := P(U > u, V > v) = 1 − u − v + C(u, v) =
b − u, 1 − v).
C(1

Now we want to define the n-variate survival copula C b of (X1 , . . . , Xn ) such that


b F 1 (x1 ), . . . , F n (xn ) .
H(x1 , . . . , xn ) = P(X1 > x1 , . . . , Xn > xn ) = C (2.44)
We can derive it with the help of the inclusion-exclusion formula (Billingsley, 1995, p. 24):
Let (Ω, A, P) be a probability space and A1 , . . . , An ∈ A. Then,
n
! n
[ X X X
P Ak = P(Ak ) − P(Ai ∩ Aj ) + P(Ai ∩ Aj ∩ Ak ) + · · ·
k=1 k=1 1≤i<j≤n 1≤i<j<k≤n

+ (−1)n+1 P(A1 ∩ · · · ∩ An ). (2.45)

With Ai := {Xi ≤ xi } we have
n
!
[
H(x1 , . . . , xn ) := P(X1 > x1 , . . . , Xn > xn ) = 1 − P Ai
i=1
n
!
X X
=1− P(Ai ) − P(Ai ∩ Aj ) + · · · + (−1)n+1 P(A1 ∩ · · · ∩ An )
i=1 1≤i<j≤n
Xn X

=1− Fi (xi ) − Ci,j Fi (xi ), Fj (xj ) + · · ·
i=1 1≤i<j≤n


+ (−1)n+1 C1,...,n F1 (x1 ), . . . , Fn (xn )
n
X X

= F i (xi ) − (n − 1) + Ci,j 1 − F i (xi ), 1 − F j (xj ) + · · ·
i=1 1≤i<j≤n
n


+ (−1) · C1,...,n 1 − F 1 (x1 ), . . . , 1 − F n (xn )
so that
n
X X
b 1 , . . . , un ) :=
C(u ui − (n − 1) + Ci,j (1 − ui , 1 − uj )
i=1 1≤i<j≤n
X
− Ci,j,k (1 − ui , 1 − uj , 1 − uk ) + · · · + (−1)n · C1,...,n (1 − u1 , . . . , 1 − un ) (2.46)
1≤i<j<k≤n
30 2. Copulas and Their Properties

is the n-dimensional survival copula of (X1 , . . . , Xn ).

An equivalent representation due to the fact that all marginals Ci1 ,...,ik are derivable from
C = C1,...,n is given in Georges et al. (2001, Theorem 2), which states that
b 1 , . . . , un ) = C(1 − u1 , . . . , 1 − un )
C(u (2.47)

with
 
n
X X
C(u1 , . . . , un ) = (−1)k C(v1 , . . . , vn )
k=0 v∈Z(n−k,n,1;u)
X Pn

= i=1 εi
(−1) C ε1 u1 + (1 − ε1 ), . . . , εn un + (1 − εn ) .
(ε1 ,...,εn )∈{0,1}n
Pn
where Z(m, n, ε; u) := {v ∈ I n : vi ∈ {ui , ε}, 1 ≤ i ≤ n, {ε} (vk ) = m}.


k=1

For example, the three-variate version (C = C1,2,3 ) is given by

b 1 , u2 , u3 ) (2.46)
C(u = u1 + u2 + u3 − 2 + C1,2 (1 − u1 , 1 − u2 ) + C1,3 (1 − u1 , 1 − u3 )
+ C2,3 (1 − u2 , 1 − u3 ) − C1,2,3 (1 − u1 , 1 − u2 , 1 − u3 )
(2.47)
= C(1, 1, 1) − C(1 − u1 , 1, 1) − C(1, 1 − u2 , 1) − C(1, 1, 1 − u3 )
+ C(1 − u1 , 1 − u2 , 1) + C(1, 1 − u2 , 1 − u3 ) + C(1 − u1 , 1, 1 − u3 )
− C(1 − u1 , 1 − u2 , 1 − u3 ).

We are now ready to derive the following corollary from Theorem 2.40.
2.41 Corollary
A random vector (X1 , . . . , Xn ) with copula C and univariate survival functions F i , 1 ≤ i ≤
n, is radially symmetric about (a1 , . . . , an ) ∈ n iff
a) F i (ai − xi ) = Fi (ai + xi ) for all xi ∈ , 1 ≤ i ≤ n, i. e., all the Xi are symmetric
about ai , and
b 1 , . . . , un ) for all (u1 , . . . , un ) ∈ ran(F1 ) × · · · × ran(Fn ).
b) C(u1 , . . . , un ) = C(u
Proof. By Theorem 2.40, radial symmetry is equivalent to

H(a1 + x1 , . . . , an + xn ) = H(a1 − x1 , . . . , an − xn ) ⇐⇒



b F 1 (a1 − x1 ), . . . , F n (an − xn ) .
C F1 (a1 + x1 ), . . . , Fn (an + xn ) = C

Thus, the sufficiency of a) and b) is clear. For the necessity, we observe that in the case of
radial symmetry

Fi (ai + xi ) = H(∞, . . . , ∞, ai + xi , ∞, . . . , ∞)
radial symm.
= H(−∞, . . . , −∞, ai − xi , −∞, . . . , −∞) = F i (ai − xi ) (2.48)
 2.4 Survival Copula and Symmetry Concepts 31

for all xi ∈ .
Now, let (u1 , . . . , un ) ∈ ran(F1 ) × · · · × ran(Fn ). Then there exist xi ∈ such that
ui = Fi (ai + xi ), 1 ≤ i ≤ n. Thus,


C(u1 , . . . , un ) = C F1 (a1 + x1 ), . . . , Fn (an + xn ) = H(a1 + x1 , . . . , an + xn )
radial symm. (2.44)

= H(a1 − x1 , . . . , an − xn ) = C b F 1 (a1 − x1 ), . . . , F n (an − xn )
(2.48)

= C b F1 (a1 + x1 ), . . . , Fn (an + xn ) = C(u
b 1 , . . . , un )

which completes the proof.

This corollary will later be used to characterize symmetric stochastic processes.
32 2. Copulas and Their Properties
3
Dependence Measures

In this chapter we will give a short introduction to some dependence measures commonly
used in multivariate settings. Basic references are Nelsen (1999) and Joe (1997). It is obvious
that we are interested in dependence measures which are closely related to copulas.

3.1 Comonotonicity
Two random variables X and Y are said to be comonotonic if a random variable Z exists
such that
X = f (Z), Y = g(Z) a. s.,
where f and g are two increasing real-valued functions.
Comonotonicity is an extension of the concept of (perfect) positive correlation to vari-
ables with arbitrary distributions, i. e., while correlation of two random variables is often
understood as linear dependence, we may here have any monotone relation between the
variables.
We note that the upper Fréchet bound for copulas is reached if X and Y are comono-
tonic, the lower bound is reached if X and Y are countermonotonic, i.e., if X and −Y
are comonotonic (see Remark 2.14 and Embrechts et al., 1999). If X and Y are perfectly
positively correlated, then X and Y are comonotonic; the converse is generally not true.

3.2 Concordance
In the following we will need the notion of concordance. Informally, a pair of random
variables X and Y is concordant if “large” values of one tend to be associated to “large”
values of the other variable, and analogously for “small” values. To be precise, let (X, Y )
be a vector of two random variables and (x1 , y1 ), (x2 , y2 ) two samples from (X, Y ). We
will say that (x1 , y1 ) and (x2 , y2 ) are concordant if (x1 − x2 )(y1 − y2 ) > 0, the samples are
discordant if (x1 − x2 )(y1 − y2 ) < 0. It is clear that we could define concordance of two
34 3. Dependence Measures

vectors (x1 , y1 ) and (x2 , y2 ) without using random variables. However, in our context there
is no need for such a generalization.
Note that P(X1 = X2 ) = 0 if X1 and X2 are continuous. Therefore, if X and Y above are
continuous, the regions of concordance and discordance split the sample space as a subset
of 2 into two non-intersecting regions whose P-measure is 1.

3.3 Kendall’s τ
3.1 Definition (Kendall’s τ )
Let (X1 , Y1 ) and (X2 , Y2 ) be two independent and identically distributed random vectors on
some probability space (Ω, A, P). Kendall’s τ is defined as



τ ≡ τ (X1 , Y1 ) ≡ τX1 ,Y1 := P (X1 − X2 )(Y1 − Y2 ) > 0 − P (X1 − X2 )(Y1 − Y2 ) < 0 . (3.1)
We will also call this the population version of Kendall’s τ .
Thus, Kendall’s τ is just the difference between the probability of concordance and the
probability of discordance.
As one may suppose, there is also a“sample version”of this dependence measure, based on
a sample from (X, Y ) ∼ H. Let {(x1 , y1 ), . . . , (xn , yn )}, n ≥ 2, denote a sample of n obser-
vations from a vector (X, Y ) of continuous random variables. Each pair {(xi , yi ), (xj , yj )},
i, j ∈ {1, . . . , n}, i 6= j, is either discordant or concordant

as there are no ties in the sample
with probability one. There are obviously n2 distinct pairs of observations in the sam-
ple. Let c denote the number of concordant pairs, d the number of discordant pairs. Then
Kendall’s τ for the sample (the above mentioned “sample version”) is defined to be
c−d c−d
τ̂ ≡ t := = n
. (3.2)
c+d 2

If we choose a pair

of observations (xi , yi ) and (xj , yj ) randomly from the sample (for
n
which there are 2 possibilities which are equally likely) τ̂ can be interpreted as the
probability that we get a concordant pair minus the probability that we get a discordant
pair.
We can express concordance and discordance with indicator functions A (x) for Borel


sets A in the following way:

(xi , yi ) and (xj , yj ) are concordant ⇔ (0,∞) (xi − xj )(yi − yj ) = 1,


(xi , yi ) and (xj , yj ) are discordant ⇔ (−∞,0) (xi − xj )(yi − yj ) = 1.


When X or Y is continuous, we can also write

(Xi − Xj )(Yi − Yj ) = 0 [P]


(0,∞)


for discordance. Therefore, the number of concordant values c (x, y) in a sample (x, y) =
{(x1 , y1 ), . . . , (xn , yn )} is given by
n

1 X

c (x, y) = (xi − xj )(yi − yj ) .


(0,∞)
2 1≤i<j≤n
The next theorem is the basis for connecting copulas with Kendall’s τ .
 3.4 Spearman’s ρ 35

3.2 Theorem
Let (X, Y ) be a random vector with continuous marginals, i. e., X is a continuous random
variable with distribution function F and Y is continuous with distribution function G.
Let (X1 , Y1 ) and (X2 , Y2 ) be two random vectors with distribution function H1 and H2 ,
respectively, where Xi ∼ F and Yi ∼ G, i = 1, 2. Assume that all random variables are
defined on a probability space (Ω, A,
P).

Further, H1 (x, y) = C1 F (x), G(y) and H2 (x, y) = C2 F (x), G(y) , C1 and C2 denoting
the underlying copulas. If Q denotes the difference between the probabilities of concordance
and discordance of (X1 , Y1 ) and (X2 , Y2 ),



Q := P (X1 − X2 )(Y1 − Y2 ) > 0 − P (X1 − X2 )(Y1 − Y2 ) < 0 ,

we have: ZZ
Q = Q(C1 , C2 ) = 4 C2 (u, v) dC1 (u, v) − 1. (3.3)
I2

For a proof see Nelsen (1999, pp. 127).

Obviously, Q does not depend on the marginal distribution functions. Kendall’s τ can
then easily be expressed by the underlying copula of two random variables X and Y :
3.3 Theorem
Let X and Y be continuous real-valued random variables whose copula is C. Then Kendall’s
τ for X and Y is given by
ZZ
τX,Y ≡ τC = Q(C, C) = 4 C(u, v) dC(u, v) − 1 = 4 · EC (C(U, V )) − 1 (3.4)
I2

where U and V are U(0, 1) random variables with copula C and EC denotes the expectation
w. r. t. the measure PC induced by the copula.
Proof. Nelsen (1999, p. 129).

3.4 Spearman’s ρ
Spearman’s ρ is another nonparametric measure of dependence defined in terms of concor-
dance and discordance.
3.4 Definition (Spearman’s ρ)
Let (X1 , Y1 ), X2 , and Y2 be independent random variables and vector, respectively, where
Xi ∼ F , Yi ∼ G, i = 1, 2, and (X1 , Y1 ) has joint distribution function H. All random
variables are defined on (Ω, A, P).
The population version of Spearman’s ρ is defined to be proportional to the probability of
concordance minus the probability of discordance for the two vectors (X 1 , Y1 ) and (X2 , Y2 ):




ρ ≡ ρ(X1 , Y1 ) ≡ ρsX1 ,Y1 := 3 · P (X1 − X2 )(Y1 − Y2 ) > 0 − P (X1 − X2 )(Y1 − Y2 ) < 0 .
(3.5)
36 3. Dependence Measures

We can then formulate the following relationship between the underlying copula C and
Spearman’s ρ. For a proof see Nelsen (1999, pp. 135).

3.5 Theorem
Let X and Y be continuous random variables with copula C. Then
ZZ ZZ
s s
ρX,Y ≡ ρC = 12· uv dC(u, v)−3 = 12 C(u, v) du dv −3 = 12·E[C(U, V )]−3, (3.6)
I2 I2

where U and V are two iid U(0, 1) random variables.

3.5 Tail Dependencies

In this section we will give definitions of common measures of dependence of two random
variables X and Y w. r. t. their tail behaviour. Especially the concept of upper and lower
tail dependence is useful for modeling clustering of extreme values in time series (see Joe,
1997, pp. 249).

The following concept was firstly established in the works of Lehmann (1966) and
Esary et al. (1967).

3.6 Definition (LTD, RTI)

Let X and Y be two real-valued random variables on a probability space (Ω, A, P). X is
said to be left-tail decreasing in the variable Y , denoted LTD(X | Y ), if and only if

P(X ≤ x | Y ≤ y)

is a decreasing function in y for all x ∈ .

Likewise, X is declared to be right-tail increasing in Y , denoted RTI(X | Y ), if and only if

P(X > x | Y > y)

is an increasing function of y for all x ∈ .

3.7 Remark
Note that we have the following equivalent relations. They directly follow from elementary
calculations:

LTD(X | Y ) ⇐⇒ P(X > x | Y ≤ y) is increasing in y for all x, and

RTI(X | Y ) ⇐⇒ P(X ≤ x | Y > y) is decreasing in y for all x.

We can give a sufficient condition for ρ(X, Y ) ≥ τ (X, Y ) ≥ 0, which is a feature often
observed for positively dependent random variables.
 3.5 Tail Dependencies 37

3.8 Theorem (Capéraà and Genest, 1993)

Let X and Y be real-valued random variables with joint cumulative distribution function
H(x, y) and continuous marginals F and G, respectively. If P(X ≤ x | Y ≤ y) and
P(X ≤ x | Y > y) are decreasing functions of y for all x ∈ , then ρ(X, Y ) ≥ τ (X, Y ) ≥ 0.
Symbolically,

RTI(X | Y ) and LTD(X | Y ) =⇒ ρ(X, Y ) ≥ τ (X, Y ) ≥ 0, (3.7)

where ρ and τ denote Spearman’s ρ and Kendall’s τ , respectively.

It is now obvious to ask for a translation of these definitions and results into the “copula-
language”. We must observe that these definitions are not necessarily independent of the
marginals which can be seen from the form P(X ≤ x | Y ≤ y) = C(F (x), G(y))/G(y).
However, we are able to give sufficient conditions for two random variables X and Y to
be RTI and LTD based only on their copula(s)1 by using the next lemma and showing the
relations a) or b) for all u, v ∈ (0, 1).

3.9 Lemma
Let X and Y be two (not necessarily continuous) random variables on some probability
space (Ω, A, P) with marginals F and G, respectively, and copula C.2 Let IF := (0, 1) ∩
ran(F ) and IG := (0, 1) ∩ ran(G). Then we have the following relations:

a) LTD(X | Y ) ⇐⇒ C(u, v)/v is decreasing in v ∈ IG for all u ∈ IF .

b) RTI(X | Y ) ⇐⇒ (C(u, v) − u)/(1 − v) is increasing in v ∈ IG for all u ∈ IF .

Before we provide the proof, let us remark that we do not need to include the points
u = 0 or u = 1 as the copula is either 0 or v which results in constant and insofar increasing
functions.

Proof. We will only show the first part as the second one can be shown analogously. For
the “if”-part, assume C(u, v)/v to be a decreasing function in v ∈ IG for all u ∈ IF . Take
x ∈ , y 1 , y2 ∈ with y1 < y2 . Define u := F (x), v1 := G(y1 ) and v2 := G(y2 ). Then
v2 ≥ v1 due to the monotonicity of G. By assumption, we have



C(u, v1 ) C(u, v2 ) C F (x), G(y1 ) C F (x), G(y2 )
≥ ⇔ ≥
v1 v2 G(y1 ) G(y2 )
⇔ P(X ≤ x | Y ≤ y1 ) ≥ P(X ≤ x | Y ≤ y2 ).

As x, y1 and y2 are arbitrary, it follows that LTD(X | Y ).

For the “only if”-part, assume LTD(X | Y ) and u ∈ IF , v1 , v2 ∈ IG with v1 < v2 . Then
there exist x ∈ with F (x) = u and y1 , y2 ∈ with G(y1 ) = v1 , G(y2 ) = v2 . As G is

1 It turns out that we do not need uniqueness of the copula.

2 Note that no assumption of uniqueness of the copula is needed. Cf. Remark 3.10
38 3. Dependence Measures

increasing we must have y2 > y1 . (For, if y1 ≥ y2 , then v1 = G(y1 ) ≥ G(y2 ) = v2 due to

the monotonicity of G). But for y1 < y2 we have


C F (x), G(y2 ) C(u, v2 )
P(X ≤ x | Y ≤ y2 ) = = ≤ P(X ≤ x | Y ≤ y1 )
G(y2 ) v2


C F (x), G(y1 ) C(u, v1 )
= = .
G(y1 ) v1
Therefore, C(u, v)/v is decreasing in v ∈ IG for all u ∈ IF .

3.10 Remark
a) If the marginals F and G are continuous, we have ran(F ) = ran(G) = I and therefore
a unique copula C. This is due to the fact that in the proof of Sklar’s theorem, the
copula is defined on the points ran(F ) × ran(G) and interpolated for all other points.
However, the interpolation is not unique. Thus, if ran(F ) × ran(G) = I 2 , the copula
can be uniquely defined.

b) If C(u, v) is partially differentiable w. r. t. v for all u ∈ (0, 1), i. e., D2 C exists, we

have the following relations:

(i) C(u, v) ≥ vD2 C(u, v) for all u, v ∈ (0, 1) ⇐⇒ LTD(X | Y ),

(ii) D2 C(u, v)(1 − v) + C(u, v) − u ≥ 0 for all u, v ∈ (0, 1) ⇐⇒ RTI(X | Y ).


The next dependence concept relates to the amount of dependence in the upper quadrant
or lower quadrant tail of a bivariate distribution. It is particularly relevant to dependence
in extreme values and for the modeling of consecutive extremal events. In contrast to τ
and ρ, upper and lower tail dependence is a local dependence measure in the sense of being
defined for some interesting areas (the upper and lower quadrants) instead of being defined
globally.
Like Kendall’s τ and Spearman’s ρ tail dependence is a copula property. The amount of
tail dependence is therefore invariant under strictly increasing transformations of X and
Y.
3.11 Definition (upper and lower tail dependence)
Let X and Y be random variables with cumulative distribution functions (cdf ) F X and FY ,
respectively. The coefficient of upper tail dependence is then defined as


lim P Y > FY−1 (u) | X > FX−1 (u) =: λU (3.8)
u%1

if λU ∈ [0, 1] exists. If λU ∈ (0, 1], X and Y are said to be asymptotically dependent in the
upper tail. If C is the copula associated to X and Y we obtain

C(u, u) 1 − 2u + C(u, u)
λU = lim = lim , (3.9)
u%1 1 − u u%1 1−u
 3.6 Quadrant Dependence and Positive Likelihood Ratio Dependence 39

where C(u, v) = 1 − u − v + C(u, v) is the joint survival function. The coefficient of lower
tail dependence is analogously defined as

C(u, u)
λL := lim P Y ≤ FY−1 (u) | X ≤ FX−1 (u) = lim (3.10)
u&0 u&0 u
where the limit exists.
λ{L,U } is extensively used in extreme value theory for describing the property that one
variable is extreme given the other one is extreme. Note that λU = limu%1 P(U1 > u | U2 >
u) = limu%1 P(U2 > u | U1 > u).

3.6 Quadrant Dependence and Positive Likelihood Ratio Dependence

A very prominent dependence property between random variables is a “lack of dependence”
which means independence. Thus, it makes sense to compare the observed dependence with
that of independent variables.
The following definition will be needed for the result in Theorem 4.5:
3.12 Definition (PQD; Lehmann, 1966)
Let X and Y be real-valued random variables. X and Y are positively quadrant dependent
(PQD) if for all (x, y) ∈ 2
P(X ≤ x, Y ≤ y) ≥ P(X ≤ x) · P(Y ≤ y). (3.11)
We will write PQD(X, Y ).
In words, X and Y are PQD if the probability of them being small simultaneously (or
being large simultaneously) is at least as great as it would be were they independent.
It is clear that a sufficient condition for X and Y to be PQD is that
C(u, v) ≥ uv for all u, v ∈ I (3.12)
holds for their copula.
For two positively quadrant dependent random variables, the following theorem holds
for the relation between Kendall’s τ and Spearman’s ρ:
3.13 Theorem
Let X and Y be continuous random variables. If X and Y are PQD, then
3τX,Y ≥ ρX,Y ≥ 0. (3.13)
For a proof, see Nelsen (1999, Theorem 5.2.2).
A very strong dependence concept which will be sufficient for many other dependence
relationships to hold is “positive likelihood ratio dependence” which is closely related to
2
functions being “totally positive of order two”. A function f : → is totally positive of
0 0
order two (TP2 ) if f ≥ 0 and whenever x ≤ x and y ≤ y ,
 
 f (x, y) f (x, y 0 ) 
 
f (x0 , y) f (x0 , y 0 ) ≥ 0, (3.14)

where | · | denotes the determinant.

40 3. Dependence Measures

3.14 Definition (PLR; Lehmann, 1966)

Let X and Y be continuous random variables with joint density function h(x, y), x, y ∈ .
Then X and Y are positively likelihood ratio dependent (denoted by PLR(X, Y )) if h
satisfies
h(x, y) · h(x0 , y 0 ) ≥ h(x, y 0 ) · h(x0 , y) (3.15)
for all x, x0 , y, y 0 ∈ such that x ≤ x0 and y ≤ y 0 .
This property derives its name from the fact that the inequality in (3.15) is equivalent to
the requirement that the conditional density of Y given X = x has a monotone likelihood
ratio.
There are many relationships between all dependence concepts, however, we will only
make use of the following:
3.15 Theorem
Let X and Y be random variables with an absolutely continuous distribution function. If
PLR(X, Y ), then RTI(X | Y ), LTD(X | Y ) and PQD(X, Y ).
Proof. The assertion follows immediately with Theorem 5.2.19 and the remarks after Corol-
lary 5.2.17 in Nelsen (1999).
4
Two Examples: Analysis of Spatial Dependence by
Using Copulas

In this chapter we will give two examples of the use of copulas for the analysis of spatial
dependence, i. e., of (at least) two random variables at the same time.
The first application deals with the depencence between the minimum and the maximum
of n iid random variables. We will derive the corresponding copula and calculate Kendall’s τ
and Spearman’s ρ for (X(1) , X(n) ). We can finally show that the relation 3τn ≥ ρn ≥ τn > 0
holds in this situation.
As a second application, we will take a closer look at the joint distribution of Brownian
motion (Bt )t∈[0,∞) (see Definition 5.22) and its supremum process (St )t∈[0,∞) where St :=
sup0≤s≤t Bt . By deriving the copula we see that the dependence structure of Bt and St does
not depend on time, i. e., although we have to start with the derivation of Ct as a function
of time, the copula will prove to be time-independent.

4.1 The Min-Max Copula

In this section we are interested in the dependence structure of the minimum and the
maximum of n independent and identically distributed random variables X1 , . . . , Xn with
continuous distribution function F . Let X(i) denote the i-th order statistic of (X1 , . . . , Xn ),
i. e., X(1) ≤ X(2) ≤ . . . ≤ X(n) , then we would like to determine the copula of (X(1) , X(n) ).
If Fr,s (x, y) := P(X(r) ≤ x, X(s) ≤ y) denotes the joint distribution function of the r-th and
s-th order statistic (r < s), then it is well known (e. g., David, 1981, pp. 8–11) that
P P
 n j n! i j−i
[1 − F (y)]n−j , x < y,
j=s i=r i!(j−i)!(n−j)! F (x) [F (y) − F (x)]
Fr,s (x, y) = (4.1)
F (y), x ≥ y,
s

where Fs (x) := P(X(s) ≤ x) is the distribution function of the s-th order statistic:
X n  
n
n−i
Fs (x) = F i (x) 1 − F (x) . (4.2)
i=s
i
42 4. Two Examples: Analysis of Spatial Dependence by Using Copulas

For the marginal distributions of F1,n , we have in particular

F1 (x) = 1 − (1 − F (x))n , Fn (y) = F n (y), x, y ∈

so that in the special case of r = 1 and s = n we get

P

 n n F i (x)(F (y) − F (x))n−i , x < y,
i=1 i
F1,n (x, y) =
F (y), x ≥ y,
n

F n (y) − (F (y) − F (x))n , x < y,
=
F n (y), x ≥ y.


Solving the equation
C F 1 (x), Fn (y) = F1,n (x, y) for the function C by setting C(u, v) =


F1,n F1−1 (u), Fn−1 (v) , u, v ∈ (0, 1), with F1−1 (u) = F −1 1 − (1 − u)1/n and Fn−1 (v) =
F −1 (v 1/n ) yields the underlying copula Cn of (X(1) , X(n) ):


v − v 1/n + (1 − u)1/n − 1 n , 1 − (1 − u)1/n < v 1/n ,
Cn (u, v) = (4.3)
v, 1 − (1 − u)1/n ≥ v 1/n .

Note that the conditions for the domain of the copula stem from solving F1 (x) > Fn (y).
We thus found the family of copulas describing the dependence structure of the minimum
and maximum of n independent random variables. To get a visual impression we provide
some plots of these copulas for different values of n in Figure 4.1.
 4.1 The Min-Max Copula 43

n=3 n = 10 n = 50 n=∞
1.0

1.0

1.0

1.0
0.8

0.8

0.8

0.8
0.6

0.6

0.6

0.6
0.4

0.4

0.4

0.4
0.2

0.2

0.2

0.2
0.0

0.0

0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

5 5 5 5
1.0 1.0 1.0 1.0
y

y
c(x,y)

c(x,y)

c(x,y)

c(x,y)
0.5 0.5 0.5 0.5

0 0.5 x 1.0 0 0.5 x 1.0 0 0.5 x 1.0 0 0.5 x 1.0

1.0 1.0 1.0 1.0

1.0 1.0 1.0 1.0

0.5 0.5 0.5 0.5
y

0.5 0.5 0.5 0.5

0.5 x 1.0 0.5 x 1.0 0.5 x 1.0 0.5 x 1.0
1.0

1.0

1.0

1.0
0.8

0.8

0.8

0.8
0.6

0.6

0.6

0.6
0.4

0.4

0.4

0.4
0.2

0.2

0.2

0.2
0.0

0.0

0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

FIGURE 4.1. First row: contourplots of Cn (dotted line: contours of independence copula; main
diagonal added to provide better orientation), second row: corresponding densities; third and
fourth row: surface- and contourplots of (Cn − W )/(M − W ).
44 4. Two Examples: Analysis of Spatial Dependence by Using Copulas

We also note the interesting fact that this copula is closely related to a well known copula
family. From Theorem 2.15 we know that for a strictly decreasing function f and a strictly
increasing function g, the copula of (f (X(1) ), g(X(n) )) is given by

Cf (X(1) ),g(X(n) ) (u, v) = v − CX(1) ,X(n) (1 − u, v) = v − Cn (1 − u, v)

(
(v 1/n + u1/n − 1)n , v 1/n + u1/n − 1 > 0,
=
0, elsewhere.

However, this is a member of the family of copulas discussed by Kimeldorf and Sampson
(1975), Clayton (1978), Cook and Johnson (1981, 1986), Oakes (1982, 1986), and implicitly
already by Mardia (1962) which is usually called Clayton or Cook and Johnson copula. For
ϑ ∈ [−1, ∞) \ {0} it is defined as


CϑCl (u, v) := max (u−ϑ + v −ϑ − 1)−1/ϑ , 0 , u, v ∈ (0, 1). (4.4)

Note that it is an Archimedean copula with generator ϕϑ (t) = (t−ϑ − 1)/ϑ (see Nelsen,
1999, pp. 94).
This observation can also be interpreted stochastically: Let X1 , . . . , Xn be iid random
variables and define Yi = (Y1,i , Y2,i ) := (−Xi , Xi ), 1 ≤ i ≤ n, and

X(n) := ( max (Y1,i ), max Y2,i ).

1≤i≤n 1≤i≤n

Then



X(n) = max (−Xi ), max Xi = −X(1) , X(n) .
1≤i≤n 1≤i≤n


Note that X(n) = f (X(1) ), g(X(n) ) for f (x) = −x, g(x) = x. In view of the above remarks
the copula of the distribution of X(n) is given by

CX(n) (u, v) = v − Cn (1 − u, v) = CϑCl (u, v), ϑ = −1/n. (4.5)

For further investigations, set An := {(u, v) ∈ (0, 1)2 : 1 − (1 − u)1/n < v 1/n } and
Bn := {(u, v) ∈ (0, 1)2 : 1 − (1 − u)1/n > v 1/n }. On these open sets, Cn is partially
differentiable twice, and elementary calculation yields

∂ (v 1/n + (1 − u)1/n − 1)n−1 · (1 − u)(1/n)−1 , (u, v) ∈ An ,
D1 Cn (u, v) := Cn (u, v) =
∂u 0, (u, v) ∈ Bn ,
(4.6)

∂ 1 − (v 1/n + (1 − u)1/n − 1)n−1 · v (1/n)−1 , (u, v) ∈ An ,
D2 Cn (u, v) := Cn (u, v) =
∂v 1, (u, v) ∈ B ,
n
(4.7)
 4.1 The Min-Max Copula 45

and
∂2
D12 Cn (u, v) = Cn (u, v) (4.8)
∂u∂v

 n−1 (v(1 − u))(1/n)−1 (v 1/n + (1 − u)1/n − 1)n−2 , (u, v) ∈ An ,
n
= (4.9)
0, (u, v) ∈ Bn .

We thus have the following (Riemann-integrable) probability density for Cn :




 n−1 v(1 − u) 1/n−1 v 1/n + (1 − u)1/n − 1 n−2 , (u, v) ∈ An ,
n
cn (u, v) := (4.10)
0, elsewhere.

The next result has been known for a long time (Walsh, 1969, or David, 1981, p. 267),
but it is interesting to see its derivation with the help of copulas. Additionally, its proof is
simpler as, e. g., the ones in Reiss (1989, Section 4.2) or Galambos (1978, Theorem 2.9.1).
From an intuitive point of view, it is clear that the minimum and maximum of n iid random
variables should be asymptotically independent, i. e., we expect the limiting copula to be
the independence copula.
4.1 Lemma
Let Cn be as in (4.3). Then

lim Cn (u, v) = Π(u, v) = uv for all u, v ∈ I.

n→∞

Proof. For v = 0 or u = 0, Cn (u, v) = 0 = uv as Cn is a copula; analogously, for u = 1 or

v = 1 we have Cn (1, v) = v = uv and Cn (u, 1) = u = uv.
n→∞ n→∞
Let (u, v) ∈ (0, 1)2 . Since v 1/n −−−→ 1 and (1 − u)1/n −−−→ 1, there exists a N = N (u, v) ∈

such that 1−(1−u)1/n < v 1/n for all n ≥ N so that we only need to consider (u, v) ∈ An ,
i. e.,
n
Cn (u, v) = v − v 1/n + (1 − u)1/n − 1 .

x
Set f (x) := v 1/x + (1 − u)1/x − 1 for x ∈ (0, ∞). Then


1/x 1/x

ln v 1/x + (1 − u)1/x − 1
ln f (x) = x · ln v + (1 − u) − 1 =
x−1
whose limit for x → ∞ is of the form 0/0. With the rule of de l’Hospital (e. g., Walter,
1992, 10.11) we have


− ln v · v 1/x − ln(1 − u) · (1 − u)1/x x−2
lim ln f (x) = lim
x→∞ x→∞ −x−2 (v 1/x + (1 − u)1/x − 1)
v 1/x ln v + (1 − u)1/x ln(1 − u)

= lim = ln v + ln(1 − u) = ln v(1 − u) .
x→∞ v 1/x + (1 − u)1/x − 1
Therefore,

lim f (x) = exp ln(v(1 − u)) = v(1 − u),
x→∞
46 4. Two Examples: Analysis of Spatial Dependence by Using Copulas

and
lim Cn (u, v) = v − v(1 − u) = uv = Π(u, v) for (u, v) ∈ (0, 1)2 .
n→∞

4.2 Remark
The above proof is already short and non-technical in the sense that no other results
and only elementary calculus are used. However, it is known for the Clayton copula that
limϑ→0 CϑCl (u, v) = uv for all u, v ∈ (0, 1) (e. g., Cook and Johnson, 1981). As mentioned
Cl
above, v − Cn (1 − u, v) = C−1/n (u, v) so that

Cl
lim Cn (u, v) = v − lim C−1/n (1 − u, v) = v − lim CϑCl (1 − u, v)
n→∞ n→∞ ϑ→0
= v − (1 − u)v = uv.

This makes the proof even shorter. 

Based on these findings we can start investigating some of the properties of the above de-
fined min-max copula related to dependence. We will start with the calculation of Kendall’s
τ (see page 34) with the help of Theorem 3.3.

4.3 Theorem
Let X1 , . . . , Xn , n ∈
, be iid random variables. Then Kendall’s τ for X(1) and X(n) is
given by
1
τn (X(1) , X(n) ) ≡ τn = . (4.11)
2n − 1
Proof. From (4.3) we know that the corresponding copula for (X(1) , X(n) ) is given by Cn .
Thus, with Theorem 3.3 we know that
ZZ
τn = 4 · Cn (u, v) dCn (u, v) − 1.
I2

With (4.10) we have

ZZ
n−1
n 
(1/n)−1
τn = 4 · v − v 1/n + (1 − u)1/n − 1 v(1 − u)
n
I2

n−2

· v 1/n + (1 − u)1/n − 1 · An (u, v) du dv − 1


= 4 I1 (n) − I2 (n) − 1

with
ZZ
n − 1 1/n
n−2

I1 (n) := · v (1 − u)(1/n)−1 v 1/n + (1 − u)1/n − 1 · (u, v) du dv


An
n
I2
 4.1 The Min-Max Copula 47

and
ZZ
n − 1 (1/n)−1
2n−2

I2 (n) := ·v (1 − u)(1/n)−1 v 1/n + (1 − u)1/n − 1 · (u, v) du dv.


An
n
I2

The substitutions s = v 1/n and t = (1 − u)1/n yield for I1 (n):

Z1 Z1
(∗) n − 1 1/n
n−2
I1 (n) = v (1 − u)(1/n)−1 v 1/n + (1 − u)1/n − 1 dv du
0

n n
1−(1−u)1/n

Z1 Z1

n−2
= (n − 1)s(1 − u)(1/n)−1 s + (1 − u)1/n − 1 sn−1 ds du
0


1−(1−u)1/n

Z1 Z1
= (n − 1)nsn (s + t − 1)n−2 tn−1 t1−n ds dt
0 1−t
Z1 Z1 Z1 Z1
(∗)
= (n − 1)n sn (s + t − 1)n−2 ds dt = (n − 1)n sn (s + t − 1)n−2 dt ds
0 1−t 0 1−s
Z1 Z1
n
 t=1 1
=n s (s + t − 1)n−1 t=1−s ds = n s2n−1 ds = .
2
0 0

(∗) is valid by Fubini’s theorem (e. g., Rogers and Williams, 2000a, II. 12).
With the same substitutions, I2 (n) can be evaluated to
n−1
I2 (n) =
2(2n − 1)
so that  
1 n−1 1
τn = 4 − −1= .
2 2(2n − 1) 2n − 1

Analogously, we can derive the following result for Spearman’s ρ (see p. 35) based on
Theorem 3.5.
4.4 Theorem
Let X1 , . . . , Xn , n ∈
, be iid random variables. Then Spearman’s ρ for X(1) and X(n) is
given by
n  
12n X (−1)k 2n (n!)3
ρn (X(1) , X(n) ) ≡ ρn = 3 − 2n
+ 12 (−1)n . (4.12)
n k=0
2n − k n + k (3n)!
48 4. Two Examples: Analysis of Spatial Dependence by Using Copulas

Proof. With Theorem 3.5 and the sets An and Bn (see p. 44) we have
ZZ
ρn = 12 Cn (u, v) du dv − 3
2
I Z Z ZZ 
 1/n 1/n

n 
= 12 v− v + (1 − u) −1 du dv + v du dv −3
A Bn
 ZnZ ZZ 
1/n 1/n

n
= 12 v du dv − v + (1 − u) −1 du dv − 3.
An ∪Bn A
|n {z }
=:I(n)

It is easy to see that the first integral yields 1/2. With the substitutions s = (1 − u)1/n and
t = v 1/n , the integral I(n) can be calculated as follows:
1−(1−v 1/n )n
Z1 Z Z1 Z1

n
n
I(n) = v 1/n + (1 − u)1/n − 1 du dv = n v 1/n + s − 1 sn−1 ds dv
0 0 0 1−v 1/n

Z1 Z1 Z1 Z1
= n2 (t + s − 1)n sn−1 tn−1 ds dt = n2 tn−1 (t + s − 1)n sn−1 ds dt
0 1−t 0 1−t

n  Z
X
1 Z1
2 n k n−1
=n (t − 1) t sn−k sn−1 ds dt
k=0
k
0 1−t

n  Z
X
1  
2 n k n−1 1 2n−k


=n (t − 1) t 1 − (1 − t) dt
k=0
k 2n − k
0
 
n 
X  k Z1 Z1
n (−1) 
= n2 tn−1 (1 − t)k dt − tn−1 (1 − t)2n dt
k=0
k 2n − k
0 0

Xn  
2 n (−1)k
=n [B(n, k + 1) − B(n, 2n + 1)] ,
k=0
k 2n − k
R1
where B denotes the Beta function B(x, y) := 0 tx−1 (1 − t)y−1 dt. With Gould (1972,
formula (1.43)) the second sum is reduced to
Xn   n  
2 n (−1)k (n − 1)! (2n)! 2 X n (−1)k n (2n)! n! (−1)n
B(n, 2n + 1) n = n =

k=0
k 2n − k (3n)! k=0
k 2n − k (3n)! n 2n n
(n!)3
= (−1)n .
(3n)!
Using B(n, k + 1) = (n − 1)! k!/(n + k)! for the first sum proves the result.
 Sfrag replacements
4.1 The Min-Max Copula 49

1.0
0.8

Kendall’s τ
Spearman’s ρ n ρn τn ρn /τn
2 0.46 0.3 1.4
0.6
τ /ρ

3 0.2929 0.2 1.4643

5 0.1677 0.1 1.4897
0.4

10 0.0788 0.0526 1.4978

50 0.0151 0.0101 1.4999
0.2

100 0.0075 0.0050 1.499998

TABLE 4.1. Values for ρn , τn and ρn /τn .
0.0

0 10 20 30 40 50

FIGURE 4.2. Comparison of Kendall’s τ and

Spearman’s ρ for (X(1) , X(n) ).

Figure 4.2 and Table 4.1 reveal the fact that τn is smaller than ρn which is not surprising
in virtue of the results from Capéraà and Genest (1993) who showed that Spearman’s
ρ is generally larger than Kendall’s τ for positively dependent random variables in the
sense of left-tail decrease and right-tail increase (see Theorem 3.8). To be precise, they
showed that ρ ≥ τ ≥ 0 for two random variables X and Y whenever one of X or Y
is simultaneously left-tail decreasing (LTD) and right tail-increasing (RTI) in the other
variable (cf. Definition 3.6). Additionally, ρn /τn seems to converge to 3/2; however, this
has not yet been proven due to the unwieldy expression for ρn in equation (4.12). The
convergence to the value 3/2 of ρn /τn must be seen in a more general context: Nelsen (1991)
showed this relation for some copula families and posed the (yet unresolved) question, if
it is valid for all families of copulas {Cϑ (u, v)} indexed by a (possibly multidimensional)
parameter ϑ such that Cϑ0 = Π and Cϑ is a continuous function of ϑ at ϑ0 .
As it is not easy to see that ρn ≥ τn in the case of (X(1) , X(n) ) by the above results, we
will make use of Definition 3.14 and Theorem 3.15 to show that X(1) and X(n) are positively
likelihood ratio dependent so that this part of the the assertion follows.

4.5 Theorem
Let τn and ρn denote Kendall’s τ and Spearman’s ρ for (X(1) , X(n) ) in the above situation.
Then
3τn ≥ ρn ≥ τn > 0.

Proof. τn > 0 follows trivially from Theorem 4.3. If we show that PLR(X, Y ), the other
inequalities follow from Theorem 3.8, Theorem 3.13 and Theorem 3.15. In virtue of Def-
inition 3.14, let u1 , u2 , v1 , v2 ∈ (0, 1) and u1 ≤ u2 , v1 ≤ v2 . It is clear from (4.10) that
cn (x, y) = 0 for (x, y) ∈ \ {(0, 1)2 }. We have to show

cn (u1 , v1 )cn (u2 , v2 ) ≥ cn (u1 , v2 )cn (u2 , v1 ). (4.13)

50 4. Two Examples: Analysis of Spatial Dependence by Using Copulas

Recall that cn is given by




 n−1 v(1 − u) 1/n−1 v 1/n + (1 − u)1/n − 1 n−2 , (u, v) ∈ An ,
n
cn (u, v) =
0, elsewhere,

with An = {(u, v) ∈ (0, 1)2 : 1 − (1 − u)1/n < v 1/n }. We may therefore assume (u1 , v2 ),
(u2 , v1 ) ∈ An , for if this is not the case, the right hand side of (4.13) is zero and the
inequality trivially true. As u1 ≤ u2 and v1 ≤ v2 , we must also have (u1 , v1 ), (u2 , v2 ) ∈ An .
But then (4.13) is equivalent to

1/n−1 h 1/n
1/n
in
v1 (1 − u1 )v2 (1 − u2 ) v1 + (1 − u1 )1/n − 1 v2 + (1 − u2 )1/n − 1

1/n−1 h 1/n 1/n

1/n 1/n

in
≥ v1 (1 − u2 )v2 (1 − u1 ) v1 + (1 − u2 ) − 1 v2 + (1 − u1 ) − 1
1/n
1/n

⇔ v1 + (1 − u1 )1/n − 1 v2 + (1 − u2 )1/n − 1
1/n
1/n

≥ v1 + (1 − u2 )1/n − 1 v2 + (1 − u1 )1/n − 1 .
1/n 1/n
Setting x1 := u1 , x2 := u2 , y1 := 1 − (1 − v1 )1/n and y2 := 1 − (1 − v2 )1/n so that x1 ≤ x2
and y1 ≤ y2 , this inequality is equivalent to

(y1 − x1 )(y2 − x2 ) ≥ (y1 − x2 )(y2 − x1 )

⇔ −y1 x2 − x1 y2 ≥ −y1 x1 − x2 y2
⇔ y2 (x2 − x1 ) − y1 (x2 − y1 ) ≥ 0
⇔ (y2 − y1 )(x2 − x1 ) ≥ 0

which is clearly the case. The theorem is proved.

4.6 Remark
The question may arise if it is not possible to derive similar results for (X(r) , X(s) ), (r, s) 6=
(1, n). However, for the derivation of Cn it was of fundamental importance to be able to
determine the generalized inverses of F1 and Fn which could be done via (4.2). It does
not seem to be possible to determine the generalized inverses of Fs for 1 < s < n analyti-
cally. This is an unsolved problem and prevents the copula approach from being used for
(X(r) , X(s) ). 
 4.2 The Copula of Brownian Motion and Its Supremum 51

4.2 The Copula of Brownian Motion and Its Supremum

In this section we will shortly derive the copula of (standard) Brownian motion (Bt )t≥0 and
its supremum process (St )t≥0 with St := sup0≤s≤t Bt for t ≥ 0. Note that due to B0 = 0 we
have St ≥ 0 for all t ≥ 0.
Please refer to Rogers and Williams (2000a, Chapter I) or Karatzas and Shreve (2000,
Chapter 2) for a detailed introduction to Brownian motion.

We will be interested in the derivation of the copula CtBSup of (Bt , St ), t > 0. At this
stage we do not have any clues about its form; particularly, we do not have reason not to
expect the dependence structure to depend √ on the time-parameter−1t. √
As Bt ∼ N(0, t), we have FBt (x) = Φ(x/ t), x ∈ , such that FBt (u) = t · Φ−1 (u), 0 <
u < 1, where Φ denotes the cumulative distribution function of a standard normal random
variable. For the distribution function FSt of St , we have (e. g., Rogers and Williams, 2000a,
Corollary I.13.3 and the remark thereafter):
√
FSt (y) = P(St ≤ y) = 1 − 2 · P(Bt ≥ y) = 2 · Φ(y/ t) − 1, y ≥ 0, (4.14)

such that
 
√ v+1
FS−1 (v) = t · Φ−1 , v ∈ (0, 1). (4.15)
t
2

Let Ht (x, y) := P(Bt ≤ x, St ≤ y), x, y ∈ , denote the joint distribution function of

(Bt , St ), t > 0. From Rogers and Williams (2000a, Corollary I.13.3) (stating the reflection
principle) and using continuity of Bt and St , we have for x ≤ y:

Ht (x, y) = P(Bt ≤ x) − P(Bt ≤ x, St ≥ y)

√

= Φ(x/ t) − P Bt ≤ y − (y − x), St ≥ y
√
= Φ(x/ t) − P(Bt ≥ 2y − x)
    (4.16)
√ 2y − x
= Φ(x/ t) − 1 − Φ √
t
 
√ x − 2y
= Φ(x/ t) − Φ √
t
for the joint distribution of (Bt , St ). For the case x > y we obtain
√
Ht (x, y) = P(St ≤ y) = 2 · Φ(y/ t) − 1. (4.17)

Then the unique1 copula CtBSup is given by

CtBSup (u, v) = Ht FB−1
t
(u), FS−1
t
(v) , u, v ∈ (0, 1). (4.18)

This yields

1F and FSt are continuous.

Bt
52 4. Two Examples: Analysis of Spatial Dependence by Using Copulas
√
a) for the case u ≤ (v + 1)/2 (which comes from FB−1 (u) ≤ F S
−1
(v) ⇔ t · Φ−1 (u) ≤
√
t t

t · Φ−1 (v + 1)/2 ):
 
BSup 1 √ −1
Ct (u, v) = Φ √ · t · Φ (u)
t
 
1 √ −1
√ −1


(4.19)
−Φ √ t · Φ (u) − 2 · t · Φ (u + 1)/2
t



= u − Φ Φ−1 (u) − 2 · Φ−1 (v + 1)/2 ,

b) and for the case u > (v + 1)/2 :

 
√
1
CtBSup (u, v) =2·Φ t·Φ −1
(v + 1)/2 √ − 1 = v. (4.20)
t

We see that neither of these expressions depends on the point of time t. We can thus say
that the dependence between Brownian motion and its supremum is entirely determined
by the marginal distributions and the time-independent copula
(


−1 −1
u − Φ Φ (u) − 2 · Φ (v + 1)/2 , u ≤ (v + 1)/2,
C BSup (u, v) = (4.21)
v, u > (v + 1)/2.

What could be the use of these findings? One possible application could be found in
mathematical finance where geometric Brownian motion is a stochastic process (still) used
to model stock prices. Now, if this model were adequate, we could calculate the supremum
process of the stock price process and transform both processes by the probability integral
transform. For the resulting two uniform processes, we could develop a test for homogeneity
w. r. t. time. Under the null hypothesis that the underlying model is geometric Brownian
motion, we should not be able to detect dependencies on time; conservative tests could be
developed by testing for a trend in the data (as a trend is a special case of time dependence).
We will leave these aspects to future research.
5
Stochastic Processes And Copulas

In this chapter we will investigate the relationship between copulas and univariate stoch-
astic processes. Before we explore these relationships we first give the basic definitions and
properties used in the following sections concerning stochastic processes. For a detailed
introduction, please refer to the “standard literature”, e. g., Ross (1996) or Todorovic (1992)
for a general account to stochastic processes, Karatzas and Shreve (2000) or Revuz and Yor
(1991) for an emphasis on Brownian motion and stochastic calculus; Rogers and Williams
(2000a,b) may deal as a general account to diffusions, Markov processes and martingales.
The basic idea is that many properties of stochastic processes can be characterized by
their finite-dimensional distributions so that copulas can be used for analysis. However,
many concepts in the theory of stochastic processes are stronger than the finite-dimensional
distribution approach; examples are martingales and stochastic differential equations which
rely heavily on pathwise structures. Nevertheless, it is possible to derive the copula structure
in some special cases.

5.1 Stochastic Processes

We will always assume an underlying probability space (Ω, F , P) and a parameter set
T ⊂ (usually T = [0, ∞)).

5.1 Definition (stochastic process)

A n-dimensional stochastic process is a collection of random variables (X t )t∈T defined on
(Ω, F , P) and taking values in n , n ∈ . Symbolically, Xt : (Ω, F ) → ( n , B n ).

If the “n-dimensional” is omitted, we mean a one-dimensional process.

5.2 Definition (finite-dimensional distribution)
For a stochastic process X = (Xt )t∈T the set of finite-dimensional distributions is given by

DX := {PXt1 ,...,Xtn : t1 < · · · < tn , n ∈
, ti ∈ T }. (5.1)

54 5. Stochastic Processes And Copulas

From a stochastic point of view we are often only interested in the finite-dimensional
distributions of processes so that we declare two processes with the same finite-dimensional
distributions to be equivalent:

5.3 Definition (stochastic equivalence)

We will call two processes X = (Xt )t∈T and Y = (Yt )t∈T stochastically equivalent and
d
write X = Y if their finite-dimensional distributions are the same, that is, DX = DY .

It is important to realize that this is only one version of a number of equivalence concepts
for stochastic processes. Others are modification and indistinguishability which both are
stronger than the finite-dimensional distribution concept (cf. Karatzas and Shreve, 2000,
p. 2). However, as copulas are distribution functions, we will focus on the finite-dimensional
distributions.
A valuable source for the relationship between these concepts and sufficient conditions
for the existence of certain modified versions of a given process with desired properties
(such as continuity or differentiability) is Cramér and Leadbetter (1967).
A certain flow of information is revealed by the concept of filtrations and adapted pro-
cesses:

5.4 Definition (filtration, adapted process)

A filtration on (Ω, F ) is a family = S (Ft )t∈T of σ-algebras Ft ⊂ F such that for s < t


we have Fs ⊂ Ft . We set F∞ := σ( s∈T Fs ). A measurable space (Ω, F ) endowed with

a filtration = (Ft )t∈T is said to be a filtered space and will sometimes be denoted by


(Ω, F , , P).


A process X = (Xt )t∈T is said to be adapted to if Xt is Ft -measurable for all t ∈ T .



With a filtration, one can associate two other filtrations by setting

!
[ \
Ft− := σ Fs and Ft+ := Fs .
s<t s>t

The σ-algebra F0− is not defined and, by convention, we set F0− = F0 .

5.5 Definition (right-continuous filtration)

The filtration is right-continuous if Ft = Ft+ for all t ∈ T .


We will need right-continuity of a filtration for the “usual conditions” which will be used
for technical reasons in Section 5.5.

5.6 Definition (usual conditions)

A filtration (Ft ) is said to satisfy the usual conditions if it is right continuous and F0
contains all the P-negligible events in F .
 5.1 Stochastic Processes 55

5.7 Definition (stopping time)

Let be a filtration on (Ω, F , P) and assume T = [0, ∞). A random variable τ : Ω → [0, ∞)


is called stopping time w. r. t. if



{τ ≤ t} ∈ Ft for all t ≥ 0. (5.2)

The stopping time σ-algebra is defined to be

στ := {A ∈ F : A ∩ {τ ≤ t} ∈ Ft for all t ∈ T }. (5.3)

Stopping times play a prominent role in the technique of localizing a stochastic process
(see Revuz and Yor, 1991, pp. 117). I. e., many theorems in the context of stochastic anal-
ysis are valid if some conditions are met locally by the process, meaning essentially that
the stopped process has the desired properties.

5.8 Definition (stopped process)

For a stopping time τ and a process X = (Xt )t∈T , the stopped process X τ = (Xtτ )t∈T is
defined as (
Xt (ω), t ≤ τ (ω),
Xtτ (ω) := Xt∧τ (ω) (ω) = (5.4)
Xτ (ω) (ω), t > τ (ω).

A class of widely studied processes is the class of martingales. They may be thought of
as representing “fair games” as the expected value in the future based on “the present’s
point of view” is exactly the current value of the process.

5.9 Definition ((sub-/super-) martingale)

A real-valued process X = (Xt )t∈T adapted to (Ft )t∈T is a submartingale w. r. t. (Ft ) if

a) E(Xt+ ) < ∞ for every t ∈ T ,

b) E(Xt | Fs ) ≥ Xs almost surely for all s < t.

Xt+ denotes the positive part of Xt , i. e., Xt+ = max(Xt , 0).

A process X such that −X is a submartingale is called a supermartingale and a process
which is both a sub- and a supermartingale is a martingale.
56 5. Stochastic Processes And Copulas

5.2 General Properties

Although the main use of copulas is presently seen in modelling “vertical” or “spatial” de-
pendencies in the sense of the dependence between n random variables at the same time, it
is (in view of the following classic theorem) also possible and beneficial to use copulas when
constructing stochastic processes. The basic idea is that, under certain regularity condi-
tions, the finite-dimensional distributions of a stochastic process determine the probabilistic
behaviour, i. e., the distributional aspects, completely.
In the following, we will assume T ⊂ to be a parameter set. It often has the interpre-
tation of the time space for the process.
5.10 Theorem (Kolmogorov)
Let {Ft1 ,...,tn : ti ∈ T, n ∈ , 1 ≤ i ≤ n, t1 < t2 < · · · < tn } be a given family of

finite-dimensional distribution functions, satisfying the consistency condition

lim Ft1 ,...,tn (x1 , . . . , xn ) = Ft1 ,...,tk−1 ,tk+1 ,...,tn (x1 , . . . , xk−1 , xk+1 , . . . , xn ) for all 1 ≤ k ≤ n,
xk %∞
(5.5)
for all n ∈ , x1 , . . . , xn ∈ , t1 , . . . , tn ∈ T . Then there exist a filtered probability space

(Ω, F , , P) and a stochastic process X = (Xt )t∈T on (Ω, F , , P) such that

 

P(Xt1 ≤ x1 , . . . , Xtn ≤ xn ) = Ft1 ,...,tn (x1 , . . . , xn ) (5.6)

for all xi ∈ , ti ∈ T , 1 ≤ i ≤ n, n ∈
, and Xt is Ft -measurable for all t ∈ T .

For a proof, see Shiryaev (1996, Theorem II.9.1, pp. 246).
It is now easy to see that a collection of copulas and marginal distributions also define
a stochastic process.
5.11 Corollary
Let C := {Ct1 ,...,tn : ti ∈ T, t1 < · · · < tn , n ∈
} be a collection of copulas satisfying the
consistency condition

lim Ct1 ,...,tn (u1 , . . . , un ) = Ct1 ,...,tk−1 ,tk+1 ,...,tn (u1 , . . . , uk−1 , uk+1 , . . . , un ) (5.7)
uk %1

for all ui ∈ (0, 1), 1 ≤ k ≤ n, and D = {Ft : t ∈ T } a collection of one-dimensional

distribution functions.
Then there exist a filtered probability space (Ω, F , , P) and a stochastic process X =


(Xt )t∈T on (Ω, F , , P) such that



P(Xt1 ≤ x1 , . . . , Xtn ≤ xn ) = Ct1 ,...,tn Ft1 (x1 ), . . . , Ftn (xn ) (5.8)

for all xi ∈ , ti ∈ T , 1 ≤ i ≤ n, n ∈
, and Xt is Ft -measurable for all t ∈ T .

For a given process X = (Xt )t∈T , we will denote the associated family of copulas by
ChXi. In the same sense, ChXt1 , . . . , Xtn i ≡ CtX1 ,...,tn denotes a copula of (Xt1 , . . . , Xtn ).
Proof. The assertion follows directly from Theorem 5.10 and Theorem 2.13.
 5.2 General Properties 57

So, from a copula point of view, it is possible to interpret a stochastic process as a

process of uniform variables transformed by the marginals, i. e., we have a process (Ut )t∈T
where each Ut is U(0, 1)-distributed with an intertemporal dependence structure defined
by a family of copulas. We will call such a process a uniform process. This process is
then transformed by the family of marginals D via a quantile transformation, i. e., Xt :=
Ft−1 (Ut ), t ∈ T .
5.12 Example
a) Independence Q
If we take Ct1 ,...,tn (u1 , . . . , un ) := ni=1 ui (the independence copula family on I n )
we see that condition (5.7) is true. Therefore, for any family D = {Ft : t ∈ T }
of one-dimensional distribution functions, we have P(Xt1 ≤ x1 , . . . , Xtn ≤ xn ) =
Q n
i=1 Fti (xi ) so that the process (Xt )t∈T (which exists due to Corollary 5.11) consists
of independent random variables. This is exactly Corollary II.9.1 of Shiryaev (1996,
p. 247). The dependence structure is trivially homogeneous (in time), i. e., it does not
depend on the ti ’s.
b) Archimedean Family Pn

Let C := {Ct1 ,...,tn (u1 , . . . , un ) := ϕ[−1] i=1 ϕ(ui ) } be a set of Archimedean copulas
with generator ϕ. ϕ[−1] denotes the generalized inverse of ϕ (see Definition A.1). Note
that this is the most general case to ensure that C induces a stochastic process based
on an Archimedean family as it is not possible to use different generators ϕ1 and ϕ2
due to the consistency condition (5.7). Then
X
n 
[−1]
lim Ct1 ,...,tn (u1 , . . . , un ) = ϕ ϕ(ui )
uk %1
i=1
i6=k

= Ct1 ,...,tk−1 ,tk+1 ,...,tk (u1 , . . . , uk−1 , uk+1 , . . . , un ),

which is due to the fact that ϕ is continuous, strictly decreasing and convex with
ϕ(1) = 0. Note that ϕ[−1] is also continuous. Therefore, a uniform process (Ut ) exists
with
P(Ut1 ≤ u1 , . . . , Utn ≤ un ) = Ct1 ,...,tn (u1 , . . . , un ).


5.13 Remark
There is little overlap between the usual concepts of stochastic processes (Markov pro-
cesses, martingales, time series analysis or diffusion processes) and the concept of mod-
elling processes via copulas. For example, time series analysis heavily relies on the con-
cept of autocorrelations whereas it is of great importance to see that the copula ap-
proach tries to capture the full intertemporal dependence structure in the data. However,
Darsow, Nguyen, and Olsen (1992) managed to relate Markov processes to copulas and
even to derive necessary and sufficient conditions for a process being Markov based on its
copula family. We will come to this interesting point later. 
58 5. Stochastic Processes And Copulas

5.2.1 Strict Stationarity

Let us take a look at strict stationarity of a process which is a well known concept al-
though not as widely used as weak stationarity or stationarity of the second kind. Before
commenting on this fact let us give the definition.
5.14 Definition (strict stationarity)
Let X = (Xt )t∈T be a real-valued stochastic process on a probability space (Ω, F , P). Then
X is said to be strictly stationary, if for any n ∈ , {t1 , . . . , tn } ⊂ T and each h > 0 with

{t1 + h, . . . , tn + h} ⊂ T , we have
P(Xt1 +h ≤ x1 , . . . , Xtn +h ≤ xn ) = P(Xt1 ≤ x1 , . . . , Xtn ≤ xn ) (5.9)
for all x1 , . . . , xn ∈ .

5.15 Remark
It is clear (set n = 1 in the above definition) that all marginal distributions must be the
same, i. e., Xt ∼ F for all t ∈ T for some one-dimensional distribution function F . 

As an immediate consequence we can give the following equivalent definition of a strictly

stationary process:
5.16 Corollary
A real-valued stochastic process (Xt )t∈T is strictly stationary iff
• Xt ∼ F for all t ∈ T for some distribution function F and
• for all n ∈
and h > 0 with t1 , . . . , tn , t1 + h, . . . , tn + h ∈ T , we have
ChXt1 , . . . , Xtn i = ChXt1 +h , . . . , Xtn +h i on ran(F ), i. e.,
Ct1 ,...,tn (u1 , . . . , un ) = Ct1 +h,...,tn +h (u1 , . . . , un ) for all u1 , . . . , un ∈ ran(F ).

The processes in Example 5.12 clearly yield strictly stationary processes for Xt ∼ F for
some arbitrary distribution function F .

5.2.2 Equivalence of Stochastic Processes

We already raised the question when processes are equivalent from a stochastic point of
view. There are a few equivalence concepts providing possible answers to this question, such
as stochastic equivalence in the wide sense or indistinguishability. As stochastic equivalence
relies on the finite-dimensional distributions of a process, it should be related to copulas.
We will assume (Xt )t∈T and (Yt )t∈T to be stochastic processes on the same common
probability space (Ω, F , P) with state space (S, S) and parameter space T .
The definition of stochastic equivalence was given in Definition 5.3. It means that
P(Xt1 ∈ B1 , . . . , Xtn ∈ Bn ) = P(Yt1 ∈ B1 , . . . , Ytn ∈ Bn ) (5.10)
for all n ∈
, {t1 , . . . , tn } ⊂ T and B1 , . . . , Bn ∈ S.
 5.2 General Properties 59

5.17 Remark
If (S, S) = ( 1 , B 1 ), (5.10) is equivalent to

P(Xt1 ≤ x1 , . . . , Xtn ≤ xn ) = P(Yt1 ≤ x1 , . . . , Ytn ≤ xn ) (5.11)

for all n ∈ , xi ∈ , 1 ≤ i ≤ n.

A fortiori, Xt and Yt must have the same distribution function Ft for all t ∈ T . 

If CtX1 ,...,tn and CtY1 ,...,tn denote the copulas of (Xt1 , . . . , Xtn ) and (Yt1 , . . . , Ytn ), respectively,
we can give an equivalent characterization of stochastically equivalent processes:

5.18 Theorem
Let (Xt ) and (Yt ) be real-valued stochastic processes on (Ω, F , P) with Xt ∼ Ft , Yt ∼ Gt ,
t ∈ T , where Ft and Gt denote the distribution functions of Xt and Yt . For t1 < · · · < tn ,
n ∈ , set CtX1 ,...,tn = ChXt1 , . . . , Xtn i and CtY1 ,...,tn = ChYt1 , . . . , Ytn i.

Then, (Xt ) and (Yt ) are stochastically equivalent if and only if

a) Ft = Gt for all t ∈ T ,

b) CtX1 ,...,tn (u1 , . . . , un ) = CtY1 ,...,tn (u1 , . . . , un ) for all n ∈
, ui ∈ ran(Fti ), 1 ≤ i ≤ n.

Proof. For the “only if”-part, Remark 5.17 yields Ft = Gt for all t ∈ T . Further, assume
n
(u1 , . . . , un ) ∈ ran(Ft1 ) × · · · × ran(F
tn ) so that there exists (x1 , . . . , xn ) ∈ such that
(u1 , . . . , un ) = Ft1 (x1 ), . . . , Ftn (xn ) . Thus,


CtX1 ,...,tn (u1 , . . . , un ) = CtX1 ,...,tn Ft1 (x1 ), . . . , Ftn (xn )
equivalence
= P(Xt1 ≤ x1 , . . . , Xtn ≤ xn ) = P(Yt1 ≤ x1 , . . . , Ytn ≤ xn )

Ft =Gt Y

= CtY1 ,...,tn Gt1 (x1 ), . . . , Gtn (xn ) = Ct1 ,...,tn Ft1 (x1 ), . . . , Ftn (xn )
= CtY1 ,...,tn (u1 , . . . , un ).

For the “if”-part, we have

P(Xt1 ≤ x1 , . . . , Xtn ≤ xn ) = CtX1 ,...,tn Ft1 (x1 ), . . . , Ftn (xn )
a),b)

= CtY1 ,...,tn Gt1 (x1 ), . . . , Gtn (xn ) = P(Yt1 ≤ x1 , . . . , Ytn ≤ xn ).

It should be clear that stronger concepts of equivalence do not relate to copulas in the
same fashion: For example, (Yt ) is a modification of (Xt ) if P(Xt = Yt ) = 1 for all t ∈ T , but
there is no connection to copulas for this relation. The same is true for indistinguishability,
i. e., P(Xt = Yt for all t ∈ T ) = 1, which is the strongest form of equivalence (almost all
sample paths coincide). For details, cf. Todorovic (1992, pp. 5).
60 5. Stochastic Processes And Copulas

5.2.3 Symmetric Processes

In this section we will use the results of Section 2.4 to characterize symmetric stochastic
processes. Let us first give a slightly more general definition of a symmetric process. Let
T = [0, ∞) be the parameter space (w. l. o. g.).
5.19 Definition
Let X = (Xt )t∈T be a real-valued stochastic process on (Ω, F , P) and µt ≡ µ(t) : T → a
deterministic function. We will say that the process X is symmetric about µ t for all t ∈ T
if
d
(Xt − µt )t∈T = (µt − Xt )t∈T , (5.12)
i. e., all finite-dimensional distributions coincide.
5.20 Remark
The usual definition is the case µ(t) = 0 for all t. It is easy to see that (Xt ) is symmetric
d
about (µt ) iff (Xt ) = (µt + Yt ) and (Yt ) is symmetric about 0 (as (Xt − µt ) = (µt − Xt ) =
−(Xt − µt )). µ(t) is usually called the “drift” of the process.
We can therefore assume µ(t) = 0 for all t and immediately give a characterization of a
symmetric process. 
5.21 Theorem
Let X = (Xt )t∈T be a stochastic process, Xt ∼ Ft , and denote the copulas by Ct1 ,...,tn =
ChXt1 , . . . , Xtn i for all t1 < · · · < tn , n ∈ .

Then X is symmetric about 0 iff

a) Ft (x) = F t (−x) for all x ∈ (“marginals are symmetric”), and
bt1 ,...,tn (u1 , . . . , un ) for all t1 < · · · < tn , ti ∈ T , ui ∈ ran(Ft ),
b) Ct1 ,...,tn (u1 , . . . , un ) = C i
1 ≤ i ≤ n, n ∈ (“all copulas are radially symmetric”).

d
Proof. We must have (Xt ) = (−Xt ). Take t1 < · · · < tn arbitrarily. Then the proof is
completely analogous to the one of Corollary 2.41.

5.3 Continuous Time Stochastic Processes

In this section we will look at some properties relating to stochastic processes in continuous
time, i. e., we assume T to be an interval in .
Let us start with an example how to extract the copula of well-known processes such as
Brownian motion. There is a vast amount of literature dealing with Brownian motion, and
we take the definition of the process from a very popular one:
5.22 Definition (Karatzas and Shreve, 2000, Definition 1.1, p. 47)
A (standard, one-dimensional) Brownian motion is a continuous, adapted 1 process X =
(Xt )t≥0 , defined on a filtered probability space (Ω, F , , P) with the following properties:


1 Recall that a process (Xt ) is adapted to a filtration  = (Ft )t∈T if each Xt is Ft -measurable for all t.
 5.3 Continuous Time Stochastic Processes 61

a) X0 = 0 almost surely,

b) Xt − Xs is independent of Fs for all 0 ≤ s < t,

c) Xt − Xs ∼ N(0, t − s), 0 ≤ s < t.

It follows from Karatzas and Shreve (2000, Theorem 5.12, p. 75) that this process is a
Markov process. We can now extract the underlying bivariate copula family ChXi.

5.23 Example (copula of Brownian motion)

Let (Bt )t∈ + be a standard Brownian motion. In the sequel, let Φ denote the distribution


function of a standard normal random variable and ϕ the corresponding density.

The transition probabilities for a standard Brownian motion are given by
 
y−x
P(x, s; y, t) := P(Bt ≤ y | Bs = x) = Φ √ , t > s, x, y ∈ . (5.13)
t−s
From Lemma 2.22 we have
B


P(x, s; y, t) = D1 Cs,t Fs (x), Ft (y) ,
B
where Cs,t denotes the copula of Bs and Bt , Fs and Ft are the corresponding marginals.
Thus, we get
Zx
B

B


Cs,t Fs (x), Ft (y) = D1 Cs,t Fs (z), Ft (y) dFs (z)
−∞
  (5.14)
Zx
y−z
= Φ √ dFs (z) for 0 < s < t.
t−s
−∞
√
√ B0−1= 0 we have Bt − B0 ∼ N(0, t) so that Ft (x) = Φ(x/ t) which is
From the assumption
equivalent to x = t · Φ (Ft (x)). Substitution into (5.14) yields
Zu √ √ 
B tΦ−1 (v) − sΦ−1 (w)
Cs,t (u, v) = Φ √ dw. (5.15)
t−s
0

It is now easy to derive the corresponding partial derivatives and the copula density. Ele-
mentary calculus yields
 √ −1 √ 
B tΦ (v) − sΦ−1 (u)
D1 Cs,t (u, v) = Φ √ , (u, v) ∈ (0, 1)2 , (5.16)
t−s
r Zu √ √ 
B 1 t tΦ−1 (v) − sΦ−1 (w)
D2 Cs,t (u, v) =
ϕ √ dw, (u, v) ∈ (0, 1)2 .
ϕ Φ−1 (v) t−s t−s
0
(5.17)
62 5. Stochastic Processes And Copulas

For the last result, an interchange of integration and differentiation is necessary which
holds from the “differentiation lemma” (e. g., Bauer, 1992, Lemma 16.2, p. 102).
Differentiation of (5.16) yields the density
r √ √ √

B t ϕ ( tΦ−1 (v) − sΦ−1 (u))/ t − s
cs,t (u, v) =
, (u, v) ∈ (0, 1)2 , (5.18)
t−s ϕ Φ−1 (v)
for the Brownian copula. 
After having derived the copula we are free to generate a new stochastic process (again
of Markovian nature as the Markov structure is characterized by the bivariate copulas, see
Theorem 5.44) with the same intertemporal dependence structure as Brownian motion but
with different marginals. An empirical example for this is given in Figure 5.1. Every path
is generated the following way:
• Choose a time grid (t1 , . . . , tn ) where ti > 0, ti < ti+1 .

• A realization (u1 , . . . , un ) from the Brownian copula process is generated on the time
grid (cf. Section B.3).
√
• For the path of standard Brownian motion, set bn := tn Φ−1 (un ) such that bn can
be interpreted as a realization of a normal random variable with µ = 0 and σ 2 = tn .

• For
q the path of the generalized Brownian motion with scaled tν -marginals, set dn :=
tn ν−2
ν
· Fν−1 (un ) where Fν−1 is the quantile function of a tν -distributed random
variable (cf. Section A.2, p. 94) and take ν = 3. The dn can then be interpreted as
realizations from a rescaled t3 distributed random variable Dn such that E(Dn ) = 0,
Var(Dn ) = tn .
Note that it is not possible to distinguish the standard Brownian motion from the general-
ized one only considering first and second moments of the marginals and the dependence
structure. This is a widely unrecognized danger in practice.

5.24 Remark
a) It is now straightforward to see that the copula of any strictly monotone increasing
transformation of a Brownian motion is exactly the Brownian copula (see Theo-
rem 2.15).
For example, a geometric Brownian motion X = (Xt )t∈T satisfies the stochastic
integral equation
Zt Zt
Xt = X0 + µ Xs ds + ν Xs dBs (5.19)
0 0

whose solution is given by

 
1
Xt = X0 · exp (µ − ν 2 )t + νBt (5.20)
2
 5.3 Continuous Time Stochastic Processes 63

(Karatzas and Shreve, 2000, pp. 349) where µ ∈ , ν > 0, X0 > 0. Thus, Xt
is just
a strictly increasing transformation of Bt by f (z) := X0 · exp (µ − 12 ν 2 )t + νz . The
copula of geometric Brownian motion is therefore given by equation (5.15).

However, this is only a special case of a related question: How can we derive the
family of copulas for general transformations of diffusions or—more general—local
(semi-)martingales? We will deal with this interesting point in Section 5.5.

b) As Brownian motion is also a Gaussian process2 , Z := (Bt1 , . . . , Btn ) has a mul-

tivariate normal-distribution. Therefore, the n-dimensional copula must be the n-
dimensional Gaussian copula which is derived from the multivariate normal distribu-
tion.


2A Gaussian process is a process whose finite-dimensional distributions are all multivariate normal distributions.
 64 5. Stochastic Processes And Copulas

Standard Brownian motion Copula process underlying the Brownian motion

1.0
4

0.8
3

0.6
2
B(t)

Ct

0.4
1
0

0.2
−1

0.0

0 2 4 6 8 10 0 2 4 6 8 10

t t

Generalized Brownian motion with t3 -marginals Difference between standard and t3 Brownian motion
0.5
4
3

0.0
2
1

−0.5
0

g replacements
−1

−1.0

0 2 4 6 8 10 0 2 4 6 8 10

t t

FIGURE 5.1. Standard Brownian motion and underlying copula process (upper row); “generalized
Brownian motion” with scaled marginals such that E(Xt ) = 0, Var(Xt ) = t as in the Brownian
case.
 5.3 Continuous Time Stochastic Processes 65

5.3.1 Stochastic Continuity

In this section we will investigate the concept of stochastic continuity of processes and how
it is related to copulas. When constructing processes from copulas it is essential to have
sufficient conditions for the copulas so that the constructed process is continuous in a sense
yet to be defined.
One possible concept is stochastic continuity which is an analogue to stochastic conver-
gence of random variables. For the following, let (Xt )t∈T be a stochastic process on (Ω, F , P)
where the index set T is a dense subset of . Ft denotes the distribution function of Xt ,
Hs,t is the bivariate distribution function of (Xs , Xt ) and Cs,t = ChXs , Xt i the associated
copula.

We will make the general assumption that the marginal distributions Ft are
continuous and strictly increasing for all t ∈ T .

5.25 Definition (stochastic continuity)

(Xt ) is said to be stochastically continuous (or continuous in probability) at a point t0 ∈ T
if, for any ε > 0,

lim P( |Xt − Xt0 | > ε ) = 0. (5.21)

t→t0
t∈T

If (5.21) holds for all t ∈ T , we say that the process is stochastically continuous on T .

Following Todorovic (1992, Remark 1.9.1), we note that “stochastic continuity is a reg-
ularity condition on bivariate marginal distributions of the process”. Having in mind the
goal of constructing a stochastically continuous process, we will examine the relationship
between (5.21) and copulas.
First, note that

P( |Xt − Xt0 | > ε ) = 1 − P( |Xt − Xt0 | ≤ ε )

Z
= 1 − P(|Xt − Xt0 | ≤ ε | Xt0 = x) dFt0 (x)


(5.22)
Z Zx+ε
=1− dPXt |Xt0 =x (y) dPXt0 (x),

x−ε

and
ZZ ZZ


P( |Xt − Xt0 | > ε) = Ht0 ,t ( dx, dy) = Ct0 ,t Ft0 ( dx), Ft ( dy) .
|x−y|>ε |x−y|>ε
66 5. Stochastic Processes And Copulas

Denoting the conditional distribution function by Ft|t0 (y | x) := P(Xt ≤ y | Xt0 = x)

and using Lebesgue-Stieltjes notation, we can write (5.22) as

Z Zx+ε
P( |Xt − Xt0 | > ε) = 1 − dFt|t0 (y | x) dFt0 (x)
x−ε
Z 

 
=1− Ft|t0 (x + ε | x) − Ft|t0 (x − ε | x) dFt0 (x)
Z


(5.23)



=1− D1 Ct0 ,t Ft0 (x), Ft (x + ε) − D1 Ct0 ,t Ft0 (x), Ft (x − ε) dFt0 (x)


Z



=1− D1 Ct0 ,t u, Ft (Ft−1
0
(u) + ε) − D1 Ct0 ,t u, Ft (Ft−1
0
(u) − ε) du.
(0,1)

For the process to be stochastically continuous in t0 , this expression must tend to 0 for
t → t0 , which is equivalent to
Z



lim D1 Ct0 ,t u, Ft (Ft−1
0
(u) + ε) − D1 Ct0 ,t u, Ft (Ft−1
0
(u) − ε) du = 1.
t→t0 | {z }
(0,1) =:f (u;t0 ,t,ε)

Now, from (2.22) we have 0 ≤ f (u; t0 , t, ε) ≤ 1 for all u ∈ (0, 1). Further, assume that
the limit limt→t0 f (u; t0 , t, ε) exists for almost all u ∈ (0, 1) and is a measurable function.
With Lebesgue’s dominated convergence theorem (e. g., Billingsley, 1995, Theorem 16.4)
we obtain the condition
Z Z
lim f (u; t0 , t, ε) du = lim f (u; t0 , t, ε) du = 1.
t→t0 t→t0
(0,1) (0,1)

But then necessarily

lim f (u; t0 , t, ε) = 1 [λ1 |(0,1) ]
t→t0

which follows from Bauer (1992, Theorem 13.2). Moreover, we see that the following is
equivalent to stochastic continuity in t0 :


lim D1 Ct0 ,t u, Ft (Ft−1
0
(u) + ε) = 1 [λ], and
t→t0

(5.24)
lim D1 Ct0 ,t u, Ft (Ft−1
0
(u) − ε) = 0 [λ]
t→t0

for any ε > 0. This
is due to the fact
that 0 ≤ D1 C(u, v) ≤ 1 for any copula (see (2.21))
and Ft Ft−1
0
(u) + ε ≥ Ft Ft−10
(u) − ε for all ε > 0.
The next goal should be to provide sufficient conditions for (5.24) to hold.
Let us remark that for the upper Fréchet bound M (u, v) = min(u, v), u, v ∈ (0, 1), we
have D1 M (u, v) = (0,v) (u) for v ∈ (0, 1). If we compare this with (5.24), we see that the


limit should yield strictly monotone dependence. This will be formalized in the following
Lemma.
 5.3 Continuous Time Stochastic Processes 67

5.26 Lemma
Let (Cn )n∈ ⊂ C2 be a sequence of 2-copulas and M (u, v) = min(u, v), u, v ∈ (0, 1), the


upper Fréchet bound. Assume that Cn (u, v) is differentiable on (0, 1) w. r. t. u and

 ( 
1, u < v,
lim D1 Cn (u, v) = D1 M (u, v) = (0,v) (u) = for u ∈ (0, 1) \ {v},


n→∞ 0, u > v,
(5.25)
for all v ∈ (0, 1).
Then limn→∞ Cn (u, v) = M (u, v) for all u, v ∈ (0, 1).
Proof. First, assume u < v. Then,
(2.11)
|Cn (u, v) − M (u, v)| = M (u, v) − Cn (u, v) = min(u, v) − Cn (u, v) = u − Cn (u, v)
(∗)
= u − [Cn (u, v) − Cn (0, v)] = u − D1 Cn (ξn , v) · u
| {z }
=0

for some ξn ∈ (0, u). (∗) is valid due to the existence of the partial derivative and the
continuity of Cn (mean value theorem of differential calculus, e. g., Walter, 1992, Theorem
10.10). As ξn < u < v for all n ∈ , the assumption yields limn→∞ D1 Cn (ξn , v) = 1 so that

lim |Cn (u, v) − M (u, v)| = u − lim D1 Cn (ξn , v) · u = u − u = 0.

n→∞ n→∞

The case u > v follows analogously:

|Cn (u, v) − M (u, v)| = v − Cn (u, v) = v − Cn (u, v) + Cn (1, v) −v

| {z }
=v
= (1 − u)D1 Cn (ξn , v)
n→∞
for some u < ξn < 1. As v < u < ξn , the assumption yields D1 Cn (ξn , v) −−−→ 0.
The remaining case u = v follows from
Cor. 2.9 n→∞
0 ≤ M (u, u) − Cn (u, u) ≤ u − Cn (u − ε, u) −−−→ u − (u − ε) = ε

for all 0 < ε < u so that limn→∞ Cn (u, u) = M (u, u) for all u ∈ (0, 1).

If we know that Cn converges to M , we can provide the following lemma:

5.27 Lemma
n→∞
Let (Cn )n∈ ⊂ C2 and M (u, v) := min(u, v). If Cn (u, v) −−−→ M (u, v) on (0, 1)2 ,


limn→∞ D1 Cn (u, v) exists for all u ∈ (0, 1) \ {v} and limn→∞ D1 Cn (u, v) is continuous
in u for all v, then, for any v ∈ (0, 1),
(
1, u < v,
lim D1 Cn (u, v) = D1 M (u, v) = for u ∈ (0, 1) \ {v}.
n→∞ 0, u > v,
68 5. Stochastic Processes And Copulas

Proof. Let v ∈ (0, 1) and u ∈ (0, 1) \ {v}. W. l. o. g. u < v. Assume

lim D1 Cn (u, v) = q(u, v) < 1 = D1 M (u, v). (5.26)

n→∞

Due to the continuity of the limit, there exists an ε > 0 such that limn→∞ D1 Cn (x, v) < 1
for all x ∈ (u − ε, u + ε) ∩ (0, 1). Choose δ = min{ε, 1 − u, u, v − u} so that (u − δ, u +
δ) ⊂ (0, v). Then also limn→∞ D1 Cn (x, v) < 1 for all x ∈ (u − δ, u + δ). Now, observe
Cn (u + δ, v) − Cn (u, v). First,

lim [Cn (u + δ, v) − Cn (u, v)] = lim Cn (u + δ, v) − lim Cn (u, v)

n→∞ n→∞ n→∞
(∗)
= M (u + δ, v) − M (u, v) = u + δ − u = δ,

as u, u + δ < v. But from the mean value theorem, we also have

Cn (u + δ, v) − Cn (u, v) = D1 Cn (ξn , v) · δ

for some ξn ∈ (u, u + δ) so that

lim [Cn (u + δ, v) − Cn (u, v)] = lim D1 Cn (ξn , v) ·δ < δ,

n→∞ n→∞
| {z }
<1

which is clearly a contradiction to the above equation (∗). The assertion follows.

From these two Lemmas, we can immediately derive the following corollary.

5.28 Corollary
Let (Cn )n∈ ⊂ C2 be a sequence of 2-copulas and M (u, v) = min(u, v). For any v ∈ (0, 1),


let limn→∞ D1 Cn (u, v) exist for all u ∈ (0, 1) \ {v} and be continuous in u. Then the
following properties are equivalent:

a) lim Cn (u, v) = M (u, v) for all u, v ∈ (0, 1).

n→∞

b) lim D1 Cn (u, v) = D1 M (u, v) = (0,v) (u) for any v ∈ (0, 1).



n→∞

Proof. a)⇒b) follows from Lemma 5.27, b)⇒a) from Lemma 5.26.

5.29 Remark
Lemmas 5.26 and 5.27 and Corollary 5.28 could equivalently have been formulated for a
family {Cs,t : s, t ∈ T } of bivariate copulas where limt→s Cs,t (u, v) must be considered. This
is an important aspect when dealing with continuous time stochastic processes. 

With these preparations, we are able to give sufficient conditions for the stochastic con-
tinuity of a process generated by a family of copulas:
 5.3 Continuous Time Stochastic Processes 69

5.30 Theorem
Let (Xt )t∈T be a stochastic process, Ft the cumulative distribution function of Xt , t ∈ T ,
and Cs,t = ChXs , Xt i the copula of (Xs , Xt ), s, t ∈ T .
If

a) lim Ft−1 (u) = Fs−1 (u) for almost all u ∈ (0, 1), and
t→s

b) lim D1 Cs,t (u, v) = (0,v) (u) exists for all u ∈ (0, 1) \ {v},


t→s

then (Xt ) is stochastically continuous at time s.

Proof. Let ε > 0 be given. Let N ∈ B 1 be a Lebesgue null-set such that

lim Ft−1 (u) = Fs−1 (u) for all u ∈ N c .

t→s

For a fixed u ∈ N c , there exists δ = δ(ε, u) such that

Fs−1 (u) − ε < Ft−1 (u) < Fs−1 (u) + ε if |t − s| < δ.

Now, limt→s D1 Cs,t (u, v) = (0,v) (u), and therefore



 

= 0 and 
= 1.
lim D1 Cs,t (u, v) lim D1 Cs,t (u, v)
t→s u,Ft (Fs−1 (u)−ε) t→s u,Ft (Fs−1 (u)+ε)

The assertion follows with (5.24).

5.31 Remark
a) We cannot expect to enhance the above result w. r. t. stronger continuity concepts as
they are usually based on other concepts than finite-dimensional distributions.

b) The finite-dimensional distributions alone do not give all the information regarding

the continuity properties of a process. For example, let (Ω, F , P) = [0, ∞), B, µ
where B denotes the Borel σ-algebra on [0, ∞), and µ is a probability measure on
[0, ∞) with no mass on single points. It is easy to see that Xt (ω) = {t} (ω) and


Yt (ω) = 0 for all (t, ω) ∈ [0, ∞)×[0, ∞) have the same finite-dimensional distributions
and that (Xt ) is a version of (Yt ),3 but t → Yt (ω) is continuous for all ω while
t → Xt (ω) is discontinuous for all ω.

c) An important consequence of stochastic continuity is given by the following fact

(Todorovic, 1992, Proposition 1.11.1, p. 29): If (Xt )t∈T is a real-valued process and
T ⊂ a compact interval such that (Xt ) is stochastically continuous (or continuous
in probability) on T , then there exists a version (X̃t ) of (Xt ) which is separable and
measurable.

3 Recall that (Xt ) is a version of (Yt ) if for all t ∈ T : P(Xt = Yt ) = 1.
70 5. Stochastic Processes And Copulas

Let us give an example for the use of the theorem. We have already determined the copula
of a Brownian motion (see Example 5.23). It is known that Brownian motion almost surely
has continuous sample paths, from which the stochastic continuity on + follows. We
should be able to check the latter property via the copula approach:

5.32 Example (stochastic continuity of Brownian motion)

Let (Bt )t∈ + be a standard Brownian motion (i. e., B0 = 0). Then Bt ∼ N(0, t) so that
√


√
Ft (x) = Φ(x/ t), x ∈ , which is equivalent to Ft−1 (y) = tΦ−1 (y), y ∈ (0, 1). Then
√ √ t→s
Ft−1 (y) − Fs−1 (y) = ( t − s)Φ−1 (y) −−→ 0
B
for y ∈ (0, 1). From Example 5.23 we know that the copula Cs,t is given by

Zu √ √ 
B tΦ−1 (v) − sΦ−1 (w)
Cs,t (u, v) = Φ √ dw,
t−s
0

i. e.,  √ −1 √ 
B tΦ (v) − sΦ−1 (u)
D1 Cs,t (u, v)
=Φ √ , t > s.
t−s
√ √ √

(Note that the copula would be Φ ( sΦ−1 (u) − tΦ−1 (v))/ s − t for s > t.) Clearly,
B
D1 Cs,t (u, v) is continuous in u. Now,
√ √ −1 t→s √

tΦ−1 (v) − sΦ (u) −−→ s Φ−1 (v) − Φ−1 (u)
√ t→s
and t − s −−→ 0; therefore
√ −1 √ (
tΦ (v) − sΦ−1 (u) ∞, u < v,
lim √ =
t→s t−s −∞, u > v,

B
so that limt→s D1 Cs,t (u, v) = (0,v) (u). The stochastic continuity now follows from Theo-


rem 5.30. 
We will show in the next section that Theorem 5.30 can be used advantageously to
shorten existing proofs.

5.3.2 EFGM processes

In the following, we will give an example of a family of copulas whose induced uniform
process cannot be stochastically continuous. We will use a particulary simple class of
multivariate distributions, the so-called Eyraud-Farlie-Gumbel-Morgenstern distributions
(EFGM-distributions), which were discussed by Morgenstern (1956), Gumbel (1958) and
Farlie (1960), but it seems that the earliest publication of the functional form is Eyraud
(1938).
Although their form makes them natural candidates for multivariate distributions, it
was discovered that there are fundamental problems inhibiting the definition of smooth
 5.3 Continuous Time Stochastic Processes 71

continuous-parameter stationary EFGM processes. Cambanis (1991) showed that these pro-
cesses are not stochastically continuous and even not measurable. To be precise, he showed
this for strictly stationary EFGM processes. He did not use an approach via the underlying
copulas but considered the complete parametric form (including the marginals), showing
the discontinuity of any such process by using a set of inequalities. These results were
obtained under a weak symmetry-type condition on the support of the marginal distribu-
tion4 . Due to this approach, Cambanis was not able to derive results for the non-stationary
case. (Note: “non-stationarity” can mean time-varying marginals, different copulas C s,t and
Cs+h,t+h , or both!)

In the usual form, a bivariate EFGM-distribution H with marginal distributions F1 and

F2 is given by



H(x, y) = F1 (x)F2 (y) 1 − α 1 − F1 (x) 1 − F2 (y) , α ∈ [αmin , αmax ], x, y ∈ . (5.27)

When both marginals are absolutely continuous, αmin = −1, αmax = 1, but generally,

αmin = − min (M1 M2 )−1 , ((1 − m1 )(1 − m2 ))−1 ,
 (5.28)
αmax = min (M1 (1 − m2 ))−1 , ((1 − m1 )M2 )−1 ,
where

mk = inf ({Fk (x) : x ∈ } \ {0, 1}), Mk = sup ({Fk (x) : x ∈ } \ {0, 1}), k = 1, 2.

For a proof of these assertions, see Cambanis (1977, Theorem 1).

A simple multivariate version discussed by Johnson and Kotz (1975) is
n
!
Y X


H(x1 , . . . , xn ) = Fi (x) 1 + αk,j 1 − Fk (xk ) 1 − Fj (xj ) , (5.29)
i=1 1≤k<j≤n

where X
1+ εk εj αk,j ≥ 0 for all εk ∈ {−Mk , 1 − mk } (5.30)
1≤k<j≤n

must be valid.
Based on this multivariate distribution, one can now introduce EFGM random processes
(Xt )t∈T by requiring that all finite-dimensional distributions are multivariate EFGM, i. e.,
for the distribution function of (Xt1 , . . . , Xtn ), t1 < · · · < tn , ti ∈ T , n ∈ , we have

n
!
Y X


Ht1 ,...,tn (x1 , . . . , xn ) = Fti (xi ) 1 + α(tk , tj ) 1 − Ftk (xk ) 1 − Ftj (xj ) ,
i=1 1≤k<j≤n
(5.31)
where Ft is the cdf of Xt .5 The coefficient function α(s, t), s 6=P
t, is assumed to be symmetric
(α(s, t) = α(t, s)) and the admissible values are given by 1 + 1≤k<j≤n εk εj α(tk , tj ) ≥ 0 for

4 Note that we deliberately write “marginal distribution” instead of “marginal distributions” as Cambanis considers strictly

stationary processes.
5 The consistency condition (5.5) can be easily checked.
72 5. Stochastic Processes And Copulas

all εk ∈ {−Mtk , 1−mtk } according to (5.30), or εk ∈ {−1, 1} if Ftk is absolutely continuous.

The dependence structure is fully determined by α(·, ·) via (5.31) which will prove to be
the main limitation of these processes w. r. t. stochastic continuity.
In the case of strict stationarity, we have Ft ≡ F for all t ∈ T on the one hand and on
the other hand, α(s, t) = α̃(t − s) for all t 6= s for a function α̃.
With the help of several inequalities, Cambanis (1991, Proposition 4) was able to show:

5.33 Theorem
Let (Xt )t∈T be a real-valued, strictly stationary EFGM process with nondegenerate mar-
ginal distribution F and dependence function α. If M (F ) = 1 − m(F ), where m(F ) :=
inf({F (x) : x ∈ } \ {0, 1}), M (F ) := sup({F (x) : x ∈ } \ {0, 1}), then the process is
not stochastically continuous.6
From Section 5.3 we know that stochastic continuity is a matter of bivariate distributions.
It is therefore clear and advantageous for the analysis of stochastic continuity that we only
need the bivariate distributions of the process, i. e., a stochastic process form of (5.27):




Hs,t (xs , xt ) = Fs (xs )Ft (xt ) 1 + α(s, t) 1 − Fs (xs ) 1 − Ft (xt ) , xs , xt ∈ , (5.32)

where Hs,t is the distribution function of (Xs , Xt ), s < t. We can easily identify the copula
as

Cs,t (u, v) = uv 1 + α(s, t)(1 − u)(1 − v) , u, v ∈ [0, 1]. (5.33)
Note that the conditions for α(s, t) yield that Cs,t is 2-increasing for all s < t; generally,
only α(s, t) ∈ [−1, 1] is required for Cs,t to be a copula. Now,

D1 Cs,t (u, v) = v + α(s, t)(1 − v)(1 − 2u), u, v ∈ (0, 1). (5.34)

From this, we can easily formulate the following lemma:

5.34 Lemma
A uniform process (Ut )t∈T generated from the EFGM family is not continuous in probability
in any point t ∈ T .

Proof. From (5.24) and (5.34), we have (Ft (u) = Ft−1 (u) = u for all u ∈ [0, 1]) for all ε > 0:

lim D1 Cs,t (u, u + ε) = lim[u + ε + α(s, t)(1 − (u + ε))(1 − 2u)] = 1

t→s t→s
1−u−ε
⇐⇒ lim α(s, t) = for all ε < 1 − u, u 6= 1/2 [λ].
s→t (1 − u − ε)(1 − 2u)

This can clearly not be the case as α(s, t) is independent from u. Thus, the process cannot
be continuous in t.

In the following theorem we will present a result similar to Theorem 5.33.

6 In the sense of: “For any t ∈ T , X is not stochastically continuous in t.”

 5.4 Markov Processes 73

5.35 Theorem
Let (Xt )t∈T be a real-valued, strictly stationary EFGM process with continuous marginal(s)
F . Then (Xt ) is not stochastically continuous in any t ∈ T .

Proof. For the process to be continuous in t, (5.24) must hold for all ε > 0 and almost all
u ∈ (0, 1). Taking the second equation and (5.34), we must particularly have



 t→s
D1 Cs,t u, F (F −1 (u)−ε) = F F −1 (u)−ε +α(s, t) 1−F F −1 (u)−ε (1−2u) −−→ 0 [λ1 ].

Without loss of generality, assume u 6= 1/2, as {1/2} is a λ1 null set. Let ε > 0 be given.
The case F (F −1 (u) − ε) = 1 cannot occur as this is equivalent to F −1 (u) ≥ ω(F ) + ε.7
As u ∈ (0, 1), this is impossible. For the same reason, g(u, ε) := F (F −1 (u) − ε) < 1 for all
u ∈ (0, 1). So the above equation is equivalent to

s→t g(u, ε) 1
α(s, t) −−→ · [λ1 ],
g(u, ε) − 1 1 − 2u

which evidently cannot be the case, as any Borel set with positive measure consists of
uncountably many points. (The left side is independent of the right side.) Therefore, (Xt )
is not stochastically continuous in t.

5.4 Markov Processes

Markov processes are a certain class of stochastic processes with a special kind of depen-
dence structure: The “future” is independent of the “past” given the “present”. They are
important in many disciplines. For example, in mathematical finance using Markov pro-
cesses to model stock price behaviour is compatible with the assumption of (weak) market
efficiency. This states that the present price of a stock impounds all the information con-
tained in a record of past prices.

+
We assume T ⊂ .

5.36 Definition (Markov process)

A real-valued process (Xt )t∈T on some probability space (Ω, F , P) is called a Markov process
if, for any n ∈ , t1 < · · · < tn , t1 , . . . , tn ∈ T , and B ∈ B (where B is the Borel σ-algebra

on ), we have

P(Xtn ∈ B | Xt1 , . . . , Xtn−1 ) = P(Xtn ∈ B | Xtn−1 ) [P]. (5.35)

5.37 Remark
(5.35) is equivalent to P(Xtn ≤ x | Xt1 , . . . , Xtn−1 ) = P(Xtn ≤ x | Xtn−1 ) [P] for all x ∈
as {(−∞, x] : x ∈ } is a generator of B closed under intersections. 

7 ω(F ) is the right endpoint of the support of F ; see p. 94.

74 5. Stochastic Processes And Copulas

We will use the previously introduced notation for the conditional distribution functions:

P(x, s; y, t) := P(Xt ≤ y | Xs = x), x, y ∈ , s, t ∈ T, s < t. (5.36)

It is straightforward to see that we are actually dealing with transition kernels:

5.38 Definition (transition probability)

Let P(x, s; y, t) be a version of P(Xt ≤ y | Xs = x) having the properties:

a) For all s < t, x ∈ , P(x, s; · , t) is a distribution function on (and therefore

induces a unique probability measure on B);

b) for all s < t, y ∈ , P( · , s; y, t) is B-measurable.

Then P(x, s; y, t) is called the transition probability of the Markov process.

As a consequence of the Markov property, the conditional probabilities satisfy the
Chapman-Kolmogorov equations (e. g., Meyn and Tweedie, 1996, Theorem 3.4.2, for the
form used here), which relate the state of the process at time t with that at an earlier time
s through an intermediate time u:
Z
P(x, s; y, t) = P(z, u; y, t) P(x, s; dz, u) (s < u < t, x, y ∈ ). (5.37)


If the conditional densities p(x, s; y, t) := P(Xt = y | Xs = x) exist, we can write

Z
p(x, s; y, t) = p(x, s; z, u)p(z, u; y, t) dz (s < u < t, x, y ∈ ).


This can be interpreted as a continuous version of the law of total probability, modified by
the Markov property.
We will now give an outline of the work of Darsow, Nguyen, and Olsen (1992) who
were the first to relate copulas to Markov processes. The key to the understanding of this
relationship was the introduction of a product for 2-copulas which essentially captures the
dependence model of the Chapman-Kolmogorov equations.

5.39 Definition (product of copulas)

Let C1 , C2 ∈ C2 be 2-copulas. The product of C1 and C2 is the function C1 ∗ C2 : I 2 → I,
given by Z
(C1 ∗ C2 )(u, v) := D2 C1 (u, z) · D1 C2 (z, v) dz. (5.38)
(0,1)

Note that the ∗-product is well-defined due to (2.21).

The next theorem shows that C2 is closed under the ∗-operation. We will give a detailed
version as the proof in Darsow et al. (1992) is only sketched.
 5.4 Markov Processes 75

5.40 Theorem (Darsow et al., 1992, Theorem 2.1)

For C1 , C2 ∈ C2 , C1 ∗ C2 ∈ C2 , i. e., C1 ∗ C2 is again a copula.
Proof. For the boundary conditions (2.5) and (2.6) we have
Z
(C1 ∗ C2 )(0, v) = D2 C1 (0, z) · D1 C2 (z, v) dz,
(0,1)

but D2 C1 (0, z) = limh&0 (C1 (0, z + h) − C1 (0, z))/h = 0 as copulas are grounded so that
(C1 ∗ C2 )(0, v) = 0 for v ∈ I; the case (C1 ∗ C2 )(u, 0) = 0 for all u ∈ I follows analogously.
Further,
Z Z
(C1 ∗ C2 )(1, v) = D2 C1 (1, z) · D1 C2 (z, v) dz = D1 C2 (z, v) dz = C2 (1, v) − C2 (0, v)
(0,1) (0,1)

=v−0=v
for v ∈ I, as D2 C1 (1, z) = limh&0 (C1 (1, z + h) − C1 (1, z))/h = limh&0 (z + h − z)/h = 1.
The case (C1 ∗ C2 )(u, 1) = u for all u ∈ I is analogous.
It remains to show that C1 ∗ C2 is 2-increasing, i. e., the (C1 ∗ C2 )-volume of any rectangle
(u, v] (u = (u1 , u2 ), v = (v1 , v2 ), ui , vi ∈ I, u1 ≤ u2 , v1 ≤ v2 ) must be non-negative:


VC1 ∗C2 (u, v]
= (C1 ∗ C2 )(u2 , v2 ) − (C1 ∗ C2 )(u1 , v2 ) − (C1 ∗ C2 )(u2 , v1 ) + (C1 ∗ C2 )(u1 , v1 )
Z

= D2 C1 (u2 , z)D1 C2 (z, v2 ) − D2 C1 (u1 , z)D1 C2 (z, v2 )
(0,1)

− D2 C1 (u2 , z)D1 C2 (z, v1 ) + D2 C1 (u1 , z)D1 C2 (z, v1 ) dz
Z
= [D2 C1 (u2 , z) − D2 C1 (u1 , z)] [D1 C2 (z, v2 ) − D1 C2 (z, v1 )] dz ≥ 0.
| {z }| {z }
(0,1) ≥0, see (2.22) ≥0, see (2.22)

Thus, C1 ∗ C2 is a copula.

The ∗-product of copulas can be seen as a continuous analogue of the multiplication

operator for transition matrices. Some algebraic properties are preserved:
5.41 Lemma (Darsow et al., 1992, pp. 605)
Let C ∈ C2 be any 2-copula, Π the independence copula and M , W the upper and lower
Fréchet bounds. Then
Π ∗ C = C ∗ Π = Π,
M ∗ C = C ∗ M = C,
(W ∗ C)(x, y) = y − C(1 − x, y), x, y ∈ I,
(C ∗ W )(x, y) = x − C(x, 1 − y), x, y ∈ I.
In particular, Π is a null element in C2 and M is an identity w. r. t. the ∗-operator.
76 5. Stochastic Processes And Copulas

Additionally, we have the following continuity property of “∗”:

5.42 Theorem (Darsow et al., 1992, Theorems 2.3 and 2.4)

n→∞
Let A, B, C ∈ C2 , (An )n∈ ⊂ C2 with An −−−→ A (pointwise in I 2 ). Then the following


hold:
n→∞ n→∞
a) An ∗ B −−−→ A ∗ B and B ∗ An −−−→ B ∗ A.

b) “∗” is associative, i. e.,

A ∗ (B ∗ C) = (A ∗ B) ∗ C.

The main result of Darsow, Nguyen, and Olsen is the following:

5.43 Theorem (Darsow et al., 1992, Theorem 3.2)

Let (Xt )t∈T be a real-valued stochastic process, and for each s, t ∈ T , let Cs,t denote the
copula of (Xs , Xt ). Then the following are equivalent:
a) The conditional distribution functions P(x, s; y, t) satisfy the Chapman-Kolmogorov
conditions (5.37) for all s < u < t, s, u, t ∈ T , and almost all x, y ∈ .

b) For all s, u, t ∈ T with s < u < t,

Cs,t = Cs,u ∗ Cu,t . (5.39)

Satisfaction of the Chapman-Kolmogorov equations is a necessary but not sufficient

condition for (Xt )t∈T to be a Markov process. If we introduce an additional operator (a
generalized version of the ∗-operator), we can give equivalent conditions in terms of copulas.

Let C1 ∈ Cm , C2 ∈ Cn , m, n ≥ 2. Define C1 ? C2 : I m+n−1 → I by

Z
(C1 ? C2 )(u1 , . . . , um+n−1 ) := Dm C1 (u1 , . . . , um−1 , z)D1 C2 (z, um+1 , . . . , um+n−1 ) dz.
(0,um )
(5.40)
If m = n = 2, the ∗ and ? products are related by

(C1 ∗ C2 )(u, v) = (C1 ? C2 )(u, 1, v), u, v ∈ I.

In analogy to the above theorems, C1 ? C2 is a (m + n − 1)-copula, the ?-product is

associative and continuous in each coordinate.
We can finally give the following characterization of Markov processes:

5.44 Theorem (Darsow et al., 1992, Theorem 3.3)

A real-valued stochastic process (Xt )t∈T is a Markov process if and only if for all n ≥ 3,
t1 , . . . , tn ∈ T with tk < tk+1 , k = 1, . . . , n − 1,

Ct1 ,...,tn = Ct1 ,t2 ? Ct2 ,t3 ? · · · ? Ctn−1 ,tn (5.41)

where Ct1 ,...,tn = ChXt1 , . . . , Xtn i and Ctk ,tk+1 = ChXtk , Xtk+1 i.
 5.4 Markov Processes 77

These findings create a new approach to the theory of Markov processes and yield a new
technique for constructing such processes. We cite Darsow et al. (1992):

In the conventional approach, one specifies a Markov process by giving the

initial distribution Ft0 and a family of transition probabilities P(x, s; y, t) sat-
isfying the Chapman-Kolmogorov equations. In our approach, one specifies a
Markov process by giving all of the marginal distributions and a family of 2-
copula satisfying (5.41). Ours is accordingly an alternative approach to the
study of Markov processes which is different in principle from the conventional
one. Holding the transition probabilities of a Markov process fixed and varying
the initial distribution necessarily varies all of the marginal distributions, but
holding the copulas of the process fixed and varying the initial distribution does
not affect any other marginal distribution.

5.4.1 An Application: Structure of Records

It might be interesting to see the use of the findings of Section 5.4. Let us therefore consider
the following setting:
Let Xi , i ∈ , be a sequence of iid random variables with common continuous distribu-

tion function F , and denote the maximum up to time n by

n
_
Mn := Xi := max Xi ,
1≤i≤n
i=1

n∈
. Thus, Mn ∼ F n . If n is such that Mn > Mn−1 , n is called “record time”. Define

L(1) := 1, L(n + 1) := inf{j > L(n) : Mj > ML(n) }
with the convention that L(n) = ∞ if the set is empty. {L(n) : n ≥ 1} are the times where
the Markov process8 (Mn ) jumps.
The set {XL(n) : n ≥ 1} = {ML(n) : n ≥ 1} is called the “record values”.
For t1 , t2 ∈ , t1 < t2 , the bivariate distribution function of (Mt1 , Mt2 ) is given by

Ht1 ,t2 (x1 , x2 ) = F t1 (x1 ∧ x2 ) · F t2 −t1 (x2 ), x 1 , x2 ∈ ,

(see Resnick, 1987, p. 165) so that the copula and its partial derivatives are given by
1/t 1/t
1/t

Ct1 ,t2 (u1 , u2 ) = F t1 F −1 (u1 1 ) ∧ F −1 (u2 2 ) · F t2 −t1 F −1 (u2 2 )
(
 t 1  t2 −t1 1−t /t 1/t 1/t
u1 · u 2 1 2 , u1 1 ≤ u 2 2 , (5.42)
1/t1 1/t2 1/t2
= u1 ∧ u 2 · u2 = 1/t 1/t
u2 , u1 1 > u 2 2 ,

(
1−t1 /t2 1/t 1/t
u2 , 0 < u1 1 < u2 2 < 1,
D1 Ct1 ,t2 (u1 , u2 ) = 1/t 1/t (5.43)
0, 0 < u2 2 < u1 1 < 1,

8 This is a known fact (see Resnick, 1987, Proposition 4.1); we will derive this result with the help of the new tools based

on the theory of copulas.

78 5. Stochastic Processes And Copulas

and
(  
t1 −t1 /t2 1/t1 1/t2
u1 1 − t2
u2 , 0 < u1 < u2 < 1,
D2 Ct1 ,t2 (u1 , u2 ) = 1/t 1/t
(5.44)
1, 0< u2 2 < u1 1 < 1.

Note that “∧” and “∨” denote the minimum and maximum, respectively. For t1 = t2 we
have (clearly) C(u, v) = M (u, v) = min(u, v), u, v ∈ I, so this case is consistent with the
findings. For t2 → ∞, we have Ct1 ,t2 (u, v) → Π(u, v), u, v ∈ I, as can be easily seen. Mk
and Ml are asymptotically independent.
To derive the finite-dimensional distributions, i. e. for (Mt1 , . . . , Mtk ), we note that for
t1 < t 2 < · · · < t k ,

k
! k
!
^ ^
Ht1 ,...,tk (x1 , . . . , xk ) = F t1 xi · F t2 −t1 xi · · · F tk −tk−1 (xk )
i=1 i=2

(see Resnick, 1987, p. 165). The copula is derived to be

Ct1 ,...,tk (u1 , . . . , uk )

^ k  ^
k   
t1 −1 1/ti t2 −t1 1/t 1/t
=F F (ui ) · F F −1 (ui i ) · · · F tk −tk−1 F −1 (uk k )
i=1 i=2
^
k t 1 ^
k t2 −t1  tk −tk−1 (5.45)
1/ti 1/ti 1/t
= ui · ui · · · uk k
i=1 i=2
Yk ^
k tj −tj−1
t0 :=0 1/ti
= ui .
j=1 i=j

Let us see if the Chapman-Kolmogorov equations hold. Assume t1 < t2 < t3 , u1 , u3 ∈

(0, 1). Then, by (5.42)–(5.44),

Z
(Ct1 ,t2 ∗Ct2 ,t3 )(u1 , u3 ) := D2 Ct1 ,t2 (u1 , v) · D1 Ct2 ,t3 (v, u3 ) dv
(0,1)
t /t
u12 1
Z Z1
= D1 Ct2 ,t3 (v, u3 ) dv + u1 (1 − t1 /t2 )v −t1 /t2 D1 Ct2 ,t3 (v, u3 ) dv,
| {z }
0 t /t1
u12
t /t3
=0 for v>u32

t /t3
2
uZ
3
t /t1 1−t2 /t3
= Ct2 ,t3 (u12 , u3 ) + u1 (1 − t1 /t2 ) · v −t1 /t2 · u3 dv
t /t1 t /t
u12 ∧u32 3
 5.4 Markov Processes 79

h iu3t2 /t3
t /t1 1−t2 /t3
= Ct2 ,t3 (u12 , u3 ) + u 1 u3 v 1−t1 /t2 t /t1 t /t
u12 ∧u32 3

 t 2  t3 −t2 h i
1/t 1/t 1/t 1−t /t t /t −t /t 1/t 1/t
t −t
= u1 1 ∧ u 3 3 u3 3 + u1 u3 2 3 u32 3 1 3 − u1 1 ∧ u3 3 2 1

1/t1 1/t3
t2 1−t2 /t3 1−t1 /t3 1−t2 /t3 1/t1 1/t3
t2 −t1
= u1 ∧ u3 u3 + u 1 u3 − u 1 u3 u1 ∧ u3


ut12 /t1 u1−t
3
2 /t3 1−t /t 1−t /t (t −t )/t 1−t /t 1/t 1/t
+ u 1 u3 1 3 − u 1 u3 2 3 u1 2 1 1 = u 1 u3 1 3 , u1 1 < u 3 3 ,
=
u + u u1−t1 /t3 − u u1−t1 /t3 = u , u1
1/t1 1/t
≥ u3 3 .
3 1 3 1 3 3

As we had expected, this yields Ct1 ,t2 ∗ Ct2 ,t3 = Ct1 ,t3 when comparing with equation (5.42).

We can now prove the above mentioned theorem within the copula framework:
5.45 Theorem
Let the situation be as above. Then {XL(n) : n ∈
} is a Markov process.
Proof. We will use the characterization provided by Theorem 5.44.
For t1 < t2 < · · · < tn , ti ∈ , n ≥ 3, we have to show

Ct1 ,...,tn = Ct1 ,t2 ? Ct2 ,t3 ? · · · ? Ctn−1 ,tn .

We use induction w. r. t. n starting with n = 3. Take u1 , u2 , u3 ∈ (0, 1). Then,

Z


Ct1 ,t2 ? Ct2 ,t3 (u1 , u2 , u3 ) := D2 Ct1 ,t2 (u1 , z) · D1 Ct2 ,t3 (z, u3 ) dz
| {z }
(0,u2 ) =0 for z>u32
t /t3

2 t /t3
u2 ∧u
Z3 h iz=u2 ∧ut32 /t3
1−t /t 1−t /t
= D2 Ct1 ,t2 (u1 , z) · u3 2 3 dz = u3 2 3 · Ct1 ,t2 (u1 , z)
z=0
0
 t 1  t2 −t1
1−t /t 1/t t /t
1/t2 t /t
1/t2
= u3 2 3 u1 1 ∧ u2 ∧ u22 3 u2 ∧ u32 1
1−t2 /t3 1/t1 1/t2 1/t3
t1 1/t2 1/t3
t2 −t1
= u3 u1 ∧ u2 ∧ u3 u2 ∧ u3
(5.45)
= Ct1 ,t2 ,t3 (u1 , u2 , u3 ).
80 5. Stochastic Processes And Copulas

Assuming the validity for some n ∈
, n ≥ 3, we get the validity for the case n + 1:


Ct1 ,t2 ? · · · ? Ctn−1 ,tn ? Ctn ,tn+1 (u1 , . . . , un , un+1 )
“?” associative

= Ct1 ,t2 ? · · · ? Ctn−1 ,tn (u1 , . . . , un ) ? Ctn ,tn+1 (un , un+1 )
assumption
= Ct1 ,...,tn (u1 , . . . , un ) ? Ctn ,tn+1 (un , un+1 )

Z
Def.
= Dn Ct1 ,...,tn (u1 , . . . , un−1 , z) · D1 Ctn ,tn+1 (z, un+1 ) dz
(0,un )
tn /t
un ∧un+1 n+1
Z
1−t /tn+1
= un+1n Dn Ct1 ,...,tn (u1 , . . . , un−1 , z) dz
0
h iz=un ∧un+1
tn /tn+1
1−t /t
= un+1n n+1 Ct1 ,...,tn (u1 , . . . , un−1 , z)
z=0
n−1
^
1/tn t1
1−t /tn+1 1/ti t /t
= un+1n ui n n+1
∧ un ∧ un+1
i=1
n−1
^
1/tn t2 −t1
1−tn−1 /tn
1/ti n n+1 t /t t /t
n n+1
· ui ∧ un ∧ un+1 · · · un ∧ un+1
i=2

n+1
^ n+1
^ n+1
^
1−t /t 1/t
t 1/t
t −t 1/ti
tn −tn−1
= un+1n n+1 ui i 1 ui i 2 1 ··· ui
i=1 i=2 i=n
= Ct1 ,...,tn+1 (u1 , . . . , un+1 ).
This finishes the proof.

5.46 Remark
This result easily yields Proposition 4.1 in Resnick (1987, p. 165) which says that the
stationary transition probabilities of {XL(n) } are given by
(
1−F (y)

, y > x,
Π x, (y, ∞) := P(XL(n+1) > y | XL(n) = x) = 1−F (x) (5.46)
1, y ≤ x,



as Π x, (y, ∞) = 1 − D1 CL(n),L(n+1) F L(n) (x), F L(n+1) (y) . F denotes the distribution
function of the Xi as usual. 
However, for absolutely continuous cdf F we have
r−1  j

X 1 1
FXL(r) (x) = 1 − 1 − F (x) log
j=0
j! 1 − F (x)

(Kamps, 1995, p. 32). Thus, an analytic derivation of the generalized inverse of FXL(r) and
therefore the copula of (XL(r) , XL(s) ) does not seem to be possible.
 5.4 Markov Processes 81

5.4.2 Symmetry
d
A well-known property of standard Brownian motion is its symmetry in the sense of (Bt ) =
(−Bt ). We can now ask under which condition(s) a “generalized Brownian motion”, which
is a process (Xt ) with Ch(Xt )i = Ch(Bt )i (copulas coincide but possibly other marginals)
will be symmetric. Recalling that Brownian motion is a Markov process we can generalize
this question:
Which condition(s) must be fulfilled by the copulas and the marginals of the process for
Xt to be symmetric? In Section 5.2.3 a characterization for symmetry of general processes
was given so that we can make use of it in the current context.
We will tackle the problem by considering a general Markov process (Xt )t∈T where T =
[0, ∞) without loss of generality.
To this end, the Markov property yields
P(Xt1 ≤ x1 , . . . ,Xtn ≤ xn )
Yn
= P(Xti ≤ xi | Xt1 ≤ x1 , . . . , Xti−1 ≤ xi−1 ) · P(Xt1 ≤ x1 )
i=2
n
Y
Markov property (5.47)
= P(Xti ≤ xi | Xti−1 ≤ xi−1 ) · P(Xt1 ≤ x1 )
i=2
Qn

i=2 Cti−1 ,ti Fti−1 (xi−1 ), Fti (xi )
= Qn−1
i=2 Fti (xi )

for all t1 , . . . , tn ∈ T , n ≥ 2, x1 , . . . , xn ∈ . We see that the finite-dimensional distributions

of the process are given by the bivariate copulas and the marginals.
On the other hand, we have
P(Xt1 > x1 , . . . ,Xtn > xn )
Yn
= P(Xti > xi | Xt1 > x1 , . . . , Xti−1 > xi−1 ) · P(Xt1 > x1 )
i=2
n
Y
Markov property (5.48)
= P(Xti > xi | Xti−1 > xi−1 ) · P(Xt1 > x1 )
i=2
Qn

b
i=2 Cti−1 ,ti F ti−1 (xi−1 ), F ti (xi )
= Qn−1 .
i=2 F t i
(x i )
In view of Theorem 5.21 it is clear that the conditions for symmetry of a Markov process
only depend on the marginals and the bivariate copulas Cs,t . Thus, the process is symmetric
iff

Ft (x) = F t (−x) (5.49)

for all t ∈ T , x ∈ (marginal symmetry) and for all s < t:
! bs,t (u, v) = u + v − 1 + Cs,t (1 − u, 1 − v) for all u, v ∈ (0, 1),
Cs,t (u, v) = C (5.50)
82 5. Stochastic Processes And Copulas

bs,t is the “survival copula” of (Xs , Xt ) (see Section 2.4.2).

where C

5.47 Example
For the EFGM copula (see Section 5.3.2)

Cα (u, v) = uv(1 − α(1 − u)(1 − v)), u, v ∈ I, α ∈ [−1, 1],

bα so that this family generates

it is easyily shown by elementary calculations that Cα = C
symmetric Markov processes. 

It is well known that Brownian motion is symmetric about 0 (Karatzas and Shreve,
2000, Lemma 9.4, p. 104). In the following example (which ends on page 84) we will give
an alternative proof for the symmetry with the help of Theorem 5.21. Note that we only
need to check condition (5.50) for the bivariate copulas due to the above remarks.

5.48 Example (symmetry

√ of Brownian motion )
We have Ft (x) = Φ(x/ t) and with Φ(x) = 1−Φ(−x) property a) of Theorem 5.21 follows,
√ √
i. e., Φ(x/ t) = Φ(−x/ t).
To check condition b), we need the following result from Heuser (1988, Satz 167.5, p. 279):

5.49 Theorem
Let G ⊂ p be an open connected set and f = (f1 , . . . , fq ) : G → q , p, q ∈ . If the

partial derivatives of all component functions f1 , . . . , fq of f exist and vanish on G, then f

is constant.

We can easily deduce:

5.50 Corollary
Let G ⊂ p be an open connected set and p, q ∈ . If f = (f1 , . . . , fq ), g = (g1 , . . . , gq ) :

G → q and all partial derivatives are the same, i. e., ∂x∂ i fj (x1 , . . . , xp ) = ∂x∂ i gj (x1 , . . . , xp )
for all 1 ≤ i ≤ p, 1 ≤ j ≤ q, (x1 , . . . , xp ) ∈ G, and there exists x0 ∈ G with f (x0 ) = g(x0 ),
then f ≡ g on G.

Proof. Set h = f − g. Then ∂x∂ j h(x1 , . . . , xp ) = 0, 1 ≤ j ≤ p so that by Theorem 5.49

f (x) = g(x) + c for some c ∈ q for all x ∈ p . As f (x0 ) = g(x0 ) we have c = 0.

5.51 Remark
If f and g are continuous on G (the closure of G), then f (x0 ) = g(x0 ) for some x0 ∈
∂G := G \ int G (the boundary, “int” denotes the interior of G) will be sufficient for the
above corollary to hold. 
 5.4 Markov Processes 83

Due to Corollary 5.50 and the following Remark 5.51 it is sufficient for proving symmetry
B
to check if the partial derivatives of Cs,t b B coincide on (0, 1)2 where C B is the copula
and C s,t s,t
of Brownian motion (see Example 5.23).9
Let (u, v) ∈ (0, 1)2 , 0 < s < t < ∞. It was shown in (5.16) and (5.17) that
√ √ 
B tΦ−1 (v) − sΦ−1 (u)
D1 Cs,t (u, v) =Φ √
t−s

and
r Zu √ √ 
B t 1 tΦ−1 (v) − sΦ−1 (w)
D2 Cs,t (u, v) = ·
ϕ √ dw.
t − s ϕ Φ−1 (v) t−s
0

Now, from (5.50), we have

 √ −1 √ 
bs,t
B B tΦ (1 − v) − sΦ−1 (1 − u)
D1 C = 1−
(u, v) − u, 1 − v) = 1 − Φ
D1 Cs,t (1 √
t−s
 √ −1 √ −1 
−1 −1
Φ (u)=−Φ (1−u) tΦ (v) − sΦ (u)
= 1−Φ − √
t−s
 √ −1 √ −1 
tΦ (v) − sΦ (u) B
=Φ √ = D1 Cs,t (u, v) as required.
t−s

For the second partial derivative we have

bB (u, v) = 1 − D2 C B (1 − u, 1 − v)
D2 C s,t s,t
r Z1−u  √ −1 √ 
t 1 tΦ (1 − v) − sΦ−1 (w)
=1− ·
ϕ √ dw
t − s ϕ Φ−1 (1 − v) t−s
0
r Z1−u  √ −1 √ 
Φ−1 (1−v)=−Φ−1 (v)
and ϕ symmetric t 1 − tΦ (v) + sΦ−1 (1 − w)
= 1− ·
ϕ √ dw
t − s ϕ Φ (v)
−1 t−s
0
r Z1−u √ √ −1 
t 1 tΦ−1 (v) − sΦ (1 − w)
=1− ·
ϕ √ dw
t − s ϕ Φ (v)
−1 t−s
0
r Z1 √ √ 
1−w=x t 1 tΦ−1 (v) − sΦ−1 (x)
= 1− ·
ϕ √ dx.
t − s ϕ Φ (v)
−1 t−s
u

9 Note that both C B and C b B are copulas so that they are continuous on I 2 and all values C B (u, v) and C
b B (u, v) are
s,t s,t s,t s,t
the same for u ∈ {0, 1} or v ∈ {0, 1}.
84 5. Stochastic Processes And Copulas

Thus,
r Z1 √ √ 
B bB (u, v) t 1 tΦ−1 (v) − sΦ−1 (x)
D2 Cs,t (u, v) − D2 C s,t = ·
ϕ √ dx − 1
t − s ϕ Φ−1 (v) t−s
0
Z1
(5.18)
= cB
s,t (x, v) dx − 1 = (0,1) (v) −1=0


0
(5.51)

for v ∈ (0, 1).

We have shown that b) of Theorem 5.21 holds so that we have an alternative proof for
the symmetry of Brownian motion. (End of Example 5.48) 

The question arises if there is a characterization of those classes of bivariate copulas, i. e.,
of a set C sym ⊂ C2 , which generate symmetric Markov processes by solving the functional
equation C = C. b
For a wide class of copulas, the Archimedean family (see Section 2.3), we can give a
positive answer due to the work of Frank (1979). The answer leads to a special class
of Archimedean copulas, the Frank family which is the only Archimedean family with
b v) = C(u, v) for all u, v ∈ I. This is exactly the desired property.
C(u,

5.52 Definition (Frank, 1979, or Nelsen, 1999, p. 94)

For ϑ ∈ define
 h i
 1 (exp(−ϑu)−1)(exp(−ϑv)−1)

 − ϑ
ln 1 + exp(−ϑ)−1
, ϑ∈ \ {0},


Π(u, v), ϑ = 0,
Cϑ (u, v) := (5.52)

 M (u, v), ϑ = ∞,



W (u, v), ϑ = −∞.

Then this family of copulas is called Frank family, each member is called a Frank copula.
Please refer to Figure 5.2 for a visual impression of the density when the margins are
N(0, 1).

5.53 Remark  
e−ϑt −1
a) The characteristic generator of Cϑ for ϑ ∈ \ {0} is ϕϑ (t) = − ln e−ϑ −1
(see
Section 2.3).

b) The definition of C−∞ , C0 and C∞ is motivated by the fact that these copulas are the
limits of Cϑ , e. g., limϑ→0 Cϑ (u, v) = Π(u, v). They have been added for completeness
as they also generate symmetric Markov processes.

c) Some of the statistical properties of this family were discussed in Nelsen (1986) and
Genest (1987).
 5.4 Markov Processes 85

θ = −8 θ = −2

1 2 3

1 2 3
−1

−1
−3

−3
−3 −1 1 2 3 −3 −1 1 2 3
θ = −0.5 θ = 0.5
1 2 3

1 2 3
−1

−1
−3

−3 −1 1 2 3 −3 −3 −1 1 2 3
θ=2 θ=8
1 2 3

1 2 3
−1

−1
−3

−3

−3 −1 1 2 3 −3 −1 1 2 3

FIGURE 5.2. Contour plots of the Frank copula density with standard normal margins.

d) Cϑ is radially symmetric for all ϑ ∈ . This could be checked tediously, but we get
this from Theorem 5.21 and the following Theorem 5.54.


By investigating associativity of copulas and certain functional equations, Frank (1979)

derived the following theorem:

5.54 Theorem (Frank, 1979, Theorem 4.1)

The functions {Cϑ : ϑ ∈ \ {0}} are the only bivariate Archimedean copulas which satisfy
b
C = C.

Thus, this result yields a characterization of symmetric Markov processes whose copulas
are in the Archimedean class.
86 5. Stochastic Processes And Copulas

5.5 Continuous Local Martingales

In this section we will use the technique of time transformation to show how the copula
of continuous local martingales can be determined. The idea of time transformation is
a rather old one, already used by Dambis (1965) and Dubins and Schwarz (1965). Their
results provide us with a very elegant way to derive the copulas of the process. We have to
assume that the usual conditions hold.
5.55 Definition (continuous local martingale)
An adapted, right continuous process X is a (Ft , P)-local martingale if there exist stopping
times Tn , n ≥ 1, such that

a) Tn+1 ≥ Tn for all n and limn→∞ Tn = ∞ almost surely, and

b) for every n, the stopped process10 X Tn is a (Ft , P)-martingale.

We write X ∈ Mloc for a local martingale and X ∈ Mc,loc if it is also continuous.

To understand and apply the following Theorem 5.58 we need the “quadratic varia-
tion” hXi of a process X. However, neither is it possible nor does it make sense to de-
velop the complete theory here. Many excellent books have been written on the subject of
stochastic integrals and differential equations. As already mentioned in the introduction the
reader should take monographs such as Øksendal (1998), Rogers and Williams (2000a,b),
Revuz and Yor (1991) or Karatzas and Shreve (2000) as a reference. Only the necessary
definitions and results will be stated here.
5.56 Definition (square integrable)
Let X = (Xt ) be a right-continuous martingale. We say that X is square-integrable if
E(Xt2 ) < ∞ for all t ∈ T . If, in addition, X0 = 0 a. s., we write X ∈ M2 (or X ∈ Mc2 if
X is also pathwise continuous).

5.57 Definition (quadratic variation, cross variation)

For X ∈ M2 , the quadratic variation (process) hXi = (hXit )t∈T of X is defined to be the
unique (up to indistinguishability) adapted, increasing, natural process for which hXi 0 = 0
a. s. and X 2 − hXi is a martingale.
Here, a process is natural if for P-a. e. ω ∈ Ω the process is increasing and right-
continuous.

For any two martingales X, Y ∈ M2 , we define their cross-variation process hX, Y i by

1
hX, Y it := [hX + Y it − hX − Y it ] , t ∈ T. (5.53)
4
For a martingale M we clearly have
1
hM, M it := [hM + M it − hM − M it ] = hM it (5.54)
4
10 see Definition 5.8
 5.5 Continuous Local Martingales 87

as h0it ≡ 0 and h2M it = 4hM it for all t ≥ 0 due to

(2Mt )2 − h2M it = 4 (Mt2 − hM it ) .

| {z }
martingale

One can think of hXi as the increasing process of the Doob-Meyer-decomposition of X

(see Karatzas and Shreve, 2000, Theorem 4.10, pp. 24).

It can be shown that cross variation is bilinear.

5.58 Theorem (Dambis, 1965; Dubins and Schwarz, 1965)
t→∞
If M is a (Ft , P)-local martingale vanishing at 0 and such that hM it −−−→ ∞ a. s., and if
we set
Tt := inf{s : hM is > t}, (5.55)
then Bt := MTt is a (FTt )-Brownian motion and Mt = BhM it for all t ∈ T .
For a proof, see Karatzas and Shreve (2000, Theorem 4.6, pp. 174) or Revuz and Yor
(1991).
This theorem states that any continuous local martingale can be time transformed into a
Brownian motion. Thus, the problem of determining the copula is changed into the problem
of determining hM it . As we have already determined the copula of Brownian motion in
Example 5.23, the next Corollary can be given without a proof as it is obvious from the
above Theorem and (5.15).

5.59 Corollary
t→∞
If M is a (Ft , P)-local martingale vanishing at 0 and such that h(t) := hM it −−−→ ∞, the
bivariate copulas of the process are given by
Zu p p !
h(t)Φ−1 (v) − h(s)Φ−1 (w)
Cs,t (u, v) = Φ p dw, 0 ≤ s < t. (5.56)
h(t) − h(s)
0

Let us give some examples for the application of Corollary 5.59. Bt always denotes
standard Brownian motion.

5.60 Example (Ornstein-Uhlenbeck process)

Let α, σ > 0 and X0 be a random variable on (Ω, F ). Then the Ornstein-Uhlenbeck process
is given by the stochastic differential equation (SDE)

dXt = −αXt dt + σ dBt , t > 0, (5.57)

(see, e. g., Karatzas and Shreve, 2000, p. 358, or Øksendal, 1998, p. 74). The solution is
given by
Zt
Xt = X0 · exp(−αt) + σ exp(−α(t − s)) dBs , 0 ≤ t < ∞. (5.58)
0
88 5. Stochastic Processes And Copulas

However, (Xt ) is not a martingale if X0 6= 0 on a set A ∈ F with P(A) > 0. This is due to
Zt


E(Xt | Fs ) = X0 · exp(−αt) + σ · E exp(−α(t − u)) dBu | Fs
0
Zs
= X0 · exp(−αt) + σ exp(−α(t − u)) dBu
0
= Xs + X0 · exp(−α(t − s)), 0 ≤ s < t,
Rt
where the second last equality is valid because 0 exp(−α(t−u)) dBu is a martingale w. r. t.
(Ft ) (see Øksendal, 1998, Corollary 3.2.6 or Karatzas and Shreve, 2000, Proposition 3.2.10,
p. 139). If we set
Zt
Yt := exp(αt)Xt = X0 + σ exp(αs) dBs (5.59)
0

we see that (Yt )t≥0 is a martingale and

Mt := Yt − X0 , t ≥ 0,

is a martingale vanishing at 0. We then get

Zt Zt
(5.54)
hM it = hY − X0 , Y − X0 it = hσ exp(αs) dBs , σ exp(αs) dBs i
0 0
Zt Zt Zt
h·, ·i bilinear 2 (∗) 2 (5.60)
= σ h exp(αs) dBs , exp(αs) dBs i = σ exp(2αs) ds
0 0 0
2
σ
= (exp(2αt) − 1),
2α
where (∗) is due to the so-called “Itô-isometry” (see Karatzas and Shreve, 2000, equation
t→∞
(2.19) on page 138 and Proposition 2.10 thereafter). Note that hM it −−−→ ∞ as required
by Theorem 5.58.
As Xt = exp(−αt)Yt is a strictly monotone increasing transformation of Yt the copulas
of (Xs , Xt ) and (Ys , Yt ) are equal (cf. Theorem 2.15). The same applies to Yt = f (Mt ) with
f (x) := x + X0 . The bivariate copulas for the Ornstein-Uhlenbeck process are therefore
given by
Zu s s !
OU h(t) h(s)
Cs,t (u, v) = Φ · Φ−1 (v) − · Φ−1 (w) dw
h(t) − h(s) h(t) − h(s)
0
Zu s s !
exp(2αt) − 1 exp(2αs) − 1
= Φ · Φ−1 (v) − · Φ−1 (w) dw
exp(2αt) − exp(2αs) exp(2αt) − exp(2αs)
0
(5.61)
 5.5 Continuous Local Martingales 89

for 0 ≤ s < t and where h(t) ≡ h(t; α, σ) = σ 2 (2α)−1 (exp(2αt) − 1).

As in the case of the Brownian copula, we have

Zu
OU


lim Cs,t (u, v) = Φ Φ−1 (v) dz = uv = Π(u, v),
t→∞
0

i. e., Xs and Xt are asymptotically independent. Note that Lebesgue’s dominated conver-
gence theorem must be used to interchange integration and the taking of the limit.
The copula is also independent of σ. This means that α alone determines the dependence
between Xs and Xt .
For the parameter α, we have the following limiting copulas:
α→∞:

We have
e2αt − 1 1 − e−2αt α→∞
= −−−→ 1
e2αt − e2αs 1 − e−2α(t−s)

and

e2αs − 1 1 − e−2αs α→∞

= −−−→ 0
e2αt − e2αs e2α(t−s) − 1

from which we get

Zu
OU
lim Cs,t (u, v) = Φ(Φ−1 (v)) ds = uv = Π(u, v),
α→∞
0

the independence copula.

α→0:

In this case, we use the rule of de l’Hospital (case “0/0”) to obtain

1 − e−2αt 2te−2αt t
lim −2α(t−s)
= lim −2α(t−s)
=
α→0 1 − e α→0 2(t − s)e t−s

and

1 − e−2αs 2se−2αs s
lim 2α(t−s)
= lim −2α(t−s)
=
α→0 e − 1 α→0 2(t − s)e t−s
90 5. Stochastic Processes And Copulas

so that
OU B
lim Cs,t = Cs,t
α→0
B
where Cs,t is the Brownian copula. This result is not surprising in view of the defining SDE
(5.57) and its solution (5.58): Letting α → 0 in (5.58), we get
Zt
Xt = X 0 + σ dBs = X0 + σBt .
0

This again is only a strictly monotone increasing transformation of the Brownian motion
(Bt ), so the copulas must be the same. The interchange of the integral and limit is justified
by monotone convergence; it is standard to show it for elementary functions and then
generalize it by monotone class theorems. 

The next example is somewhat different as the time parameter t is bounded.

5.61 Example (Brownian bridge)

Let a, b, T ∈ , a < b, T > 0, and consider the one-dimensional SDE

b − Xt
dXt = dt + dBt , 0 ≤ t < T, X0 = a. (5.62)
T −t
It can be shown (Karatzas and Shreve, 2000, pp. 358 or Øksendal, 1998, p. 75) that the
solution is given by
  Zt
t t dBs
Xt = a 1 − + b + (T − t) , 0 ≤ t < T. (5.63)
T T T −s
0

The process can be thought of to be a “linear bridge” between the points (0, a) and (T, b)
which is disturbed R t by sa time-scaled Brownian motion.
Setting Yt := 0 TdB −s
, we see that, again, Xt = f (Yt ) for f (x) ≡ f (x; a, b, T, t) = a(1 −
t t
T
) + b( T ) + (T − t)x so that the copulas must coincide. (Yt )0≤t<T is a continuous martingale
with Y0 = 0. In analogy to Example 5.60 we calculate the quadratic variation process to
be
Zt t
1 −1  1 1
hY it = 2
ds = (T − s)  = − . (5.64)
(T − s) 0 T −t T
0
t→∞
One problem remains: We do not have hY it −−−→ ∞, but we do not need this requirement.
Taking a look at the proof of Theorem 5.58 in Karatzas and Shreve (2000, pp. 174) we see
that in the case of a bounded time interval [0, T ), we need the condition limt→T hY it = ∞
which is met by (5.64). For h(t) = T 1−t − T1 , 0 ≤ t < T , we get
s s s s
h(t) t(T − s) h(s) s(T − t)
= and = .
h(t) − h(s) T (t − s) h(t) − h(s) T (t − s)
 5.5 Continuous Local Martingales 91

The copula of the Brownian bridge is therefore given by

Zu p p !
BB t(T − s)Φ−1 (v) − s(T − t)Φ−1 (w)
Cs,t (u, v) = Φ p dw (5.65)
T (t − s)
0

for 0 ≤ s < t < T . 

5.62 Remark
Note that h(t) = hM, M it is usually a stochastic process, i. e., one should write h(t, ω).
However, in the above examples, h(t) turned out to be a deterministic function which is
due to the fact that the quadratic variation of Brownian motion is only a function of time
which is a consequence of the Itô-isometry.
It is this fact which makes the theory of time transformation applicable to the derivation
of copulas, for, if the time transformation were stochastic, we would end up with a copula
with stochastic time. With reference to equation (5.56) we had to write Cs,t (u, v; ω) and
h(t, ω), yielding a random copula. This would be of only theoretical interest.
Nonetheless, from a theoretical point of view, the central result is that of Monroe (1978)
who showed that a process is equivalent to a time change of Brownian motion if and only if
it is a local semimartingale. This is the most general result in that direction. Note that—for
example—Lévy processes fall into this class. 
It is well-known that stochastic integrals are Markov processes, and therefore, the above
examples also yield bivariate copulas generating Markov processes. In accordance with the
remarks on page 62 we are now free to construct new processes with the intertemporal
dependence given by the above copulas but with arbitrary marginals.
92 5. Stochastic Processes And Copulas
Appendix A
General Results from Probability Theory

A.1 Generalized Inverse

A.1 Definition (generalized inverse)
Let F : → [0, 1] be a distribution function. Then the generalized inverse or quantile
−1
function F : [0, 1] → of F is defined as
(
sup{x ∈ : F (x) = 0}, t = 0,
F −1 (t) := (A.1)
inf{x ∈ : F (x) ≥ t}, 0 < t ≤ 1,
where inf(∅) := ∞, sup(∅) := −∞.
The generalized inverse has the following properties:
A.2 Lemma
Let F , F −1 be as in Definition A.1.
a) F −1 is increasing on (0, 1) and left-continuous.


b) F F −1 (y) ≥ y for all 0 < y < 1.


c) If F is continuous in F −1 (y) (0 < y < 1), then F F −1 (y) = y.


d) F −1 F (x) ≤ x, x ∈ .


e) If F −1 is continuous in F (x) (x ∈ ) and 0 < F (x) < 1, then F −1 F (x) = x.
Proof. See Pfeifer (1989, Lemma 1.1).
A useful consequence of Lemma A.2 is
A.3 Lemma
Let X be a real-valued random variable with continuous distribution function F . Then
F (X) ∼ U(0, 1). If F is an arbitrary distribution function and U ∼ U(0, 1), then F −1 (U )
has distribution function F .
94 Appendix A. General Results from Probability Theory

For a proof, see Pfeifer (1989, Lemma 1.2).

We introduce the following notation:
A.4 Notation
Let F be a distribution function. Then

α(F ) ≡ αF := inf{x ∈ : F (x) > 0},

ω(F ) ≡ ωF := sup{x ∈ : F (x) < 1},

(with inf(∅) := ∞, sup(∅) := −∞) denote the right and the left endpoint of the support
of F .

A.2 Probability Distributions

Uniform Distribution
Symbol: U(a, b), a, b ∈ ,a<b
Density: f (x) = (b − a)−1 · (a,b) (x)


Dominating measure: Lebesgue measure on [a, b]

Mean: (a + b)/2
Variance: (b − a)2 /12

Normal Distribution
Symbol: N(µ, σ 2 ), µ ∈ , σ > 0
√  2

Density: f (x) = ( 2πσ)−1 exp − (x−µ)
2σ 2

Dominating measure: Lebesgue measure on

Mean: µ
Variance: σ2

t-Distribution
Symbol: tν (µ, σ 2 ); tν for µ = 0, σ = 1; µ ∈ , σ > 0, ν ∈

 − ν+1
Γ( ν+1 ) (x−µ)2 2
Density: f (x) = νπΓ( ν )σ 1 + νσ2
√ 2
2

Dominating measure: Lebesgue measure on

Mean: µ, if ν > 1
ν
Variance: σ 2 ν−2 , if ν > 2
Appendix B
Simulation/Random Variate Generation

In most cases of the application of copulas no closed form solutions exist (e. g., portfolio
analysis and option pricing in mathematical finance) so that there is a great interest in
simulation and Monte Carlo studies.
We will therefore give a short introduction to the generation of random samples from
a given joint distribution. We assume that procedures for generating independent uniform
variates are known to the reader (e. g., Lange, 1998, pp. 269, and Devroye, 1986, for al-
ternative methods). From these uniform samples we can generate samples x of a random
variable X with distribution function F by the following two steps:
1. Generate u from U(0, 1).
2. Set x = F −1 (u) where F −1 is the generalized inverse of F (see page 93).
Here we will focus on generating random samples from a copula. No proofs will be given
as they can be found in Nelsen (1999) or follow directly as in the case of the generation of
time-series from bivariate copulas.

By virtue of Sklar’s theorem, we only need to generate a pair (u, v) of U(0, 1) variables
(U, V ) with copula C (which is also the copula of X := F −1 (U ) and Y := G−1 (V )) and
then transform this pair by the quantile transformation.

B.1 The Conditional Distribution Method

For this method we need the conditional distribution function of V given U = u which we
denote by cu (v) as a shorthand notation, cu (v) := P(V ≤ v | U = u), u, v ∈ (0, 1). From
Lemma 2.22 we have
cu (v) = D1 C(u, v)
so that the following algorithm yields a sample (u, v) from a copula C (recall that cu (v)
exists almost everywhere for any copula C):
96 Appendix B. Simulation/Random Variate Generation

1. Generate u and t independently from U(0, 1)

(−1) (−1)
2. Set v = cu (t) where cu denotes the generalized inverse of cu

3. The desired pair is (u, v)

More generally, for a n-copula we have:

Let C1,...,k (u1 , . . . , uk ) be the copula of (U1 , . . . , Uk ), 2 ≤ k ≤ n, and set C1 (u1 ) := u1 ,
Cn (u1 , . . . , un ) := C(u1 , . . . , un ). From Theorem 2.27 we then have

Ck (uk | u1 , . . . , uk−1 ) := P(Uk ≤ uk | U1 = u1 , . . . , Uk−1 = uk−1 )

D1,...,k−1 C1,...,k (u1 , . . . , uk )
= .
D1,...,k−1 C1,...,k−1 (u1 , . . . , uk−1 )

We can deduce the following algorithm:

1. Simulate u1 from U(0, 1)

2. Simulate u2 from the conditional distribution function C2 (u2 | u1 )

3. Simulate u3 from the conditional distribution function C3 (u3 | u1 , u2 )

..
.

n. Simulate un from Cn (un | u1 , . . . un−1 )

The vector (u1 , . . . , un ) is the desired sample from C

Simulating uk from Ck (uk | u1 , . . . , uk−1 ) can be done by simulating q from U(0, 1) from
(−1)
which uk = Ck (q | u1 , . . . , uk−1 ) can be obtained by solving q = Ck (uk | u1 , . . . , uk−1 )
w. r. t. uk 1 if the inverse Ck−1 of Ck does not exist in closed form.
Because of the potential numeric root finding, this general algorithm is often not suit-
able. However, other and more efficient algorithms exist only in special cases such as the
Archimedean family, as we will see in the next section.

B.2 Simulating from Archimedean Copulas

Recall (cf. Section 2.3) that an Archimedean copula is defined by its generator ϕ and that
Kϕ (t) := t − ϕ(t)/ϕ0 (t+), t ∈ (0, 1), is the cumulative distribution function of the variable
Z = C(U, V ), U, V ∼ U(0, 1) (see Corollary 2.34 and Genest and Rivest, 1993). We can
then use the following algorithm for random variate generation from C:

1. Simulate s and q independently from U(0, 1)

1 E. g., by numerical root finding.

 B.3 Generation of Time-Series from Bivariate Copulas 97

2. Set t = λ−1 −1 0
ϕ (q) where λϕ is the generalized inverse of λϕ (v) = ϕ(v)/ϕ (v+), 0 < v ≤
1.



3. Set u = ϕ−1 sϕ(t) , v = ϕ−1 (1 − s)ϕ(t)

The desired pair is (u, v).

B.3 Generation of Time-Series from Bivariate Copulas

The following algorithm is a consequence of the conditional distribution method presented
in Section B.1.

Let 0 ≤ t0 < t1 < t2 < · · · < tn < ∞, n ∈ , be a (time-)grid on which we want to

simulate a uniform process (Ut ).2 Let Cti ,ti+1 (ui , ui+1 ) be a bivariate copula of (Uti , Uti+1 )
and D1 Cti ,ti+1 the partial derivative w. r. t. the first argument.
Then the following algorithm yields a time-series (u0 , u1 , . . . , un ) generated from
(Ut0 , Ut1 , . . . , Utn ) with starting value ut0 (which may be a realization of an U(0, 1)-variable
or simply Ft−1 0
(x0 ) for some starting value x0 ∈ ) with bivariate copulas Cti ,ti+1 :

1. Choose the starting value u0 ∈ [0, 1]

2. Set qu (v; s, t) := D1 Cs,t (u, v) and let qu−1 (z; s, t) denote the generalized inverse, z ∈
(0, 1)

3. Generate n iid U(0, 1) samples (r1 , . . . , rn )

4. For each i in {1, . . . , n} set ui := qu−1

i−1
(ri ; ti−1 , ti )

The vector (u0 , u1 , . . . , un ) contains the sample including the starting value.

A library for the computer language R is available from the author on request (email:
[Link]@[Link]). R is an integrated suite of software facilities for data
manipulation, calculation and graphical display not unlike STM (see Ihaka and Gentleman,
1996) and freely available under the GNU General Public License for many platforms from
[Link]
The copula-library contains a set of fully documented functions for dealing with copulas
and can be easily installed within a working R-environment by the command
R CMD INSTALL copula*.tgz (Linux)
or within the graphical user interface. Please consult the documentation.

2 Recall that we get a general process from this by applying the quantile transformation Ft−1
i
to the sample.
98 Appendix B. Simulation/Random Variate Generation
List of Symbols and Abbreviations

× cartesian product
<, ≤ less than (or equal to) componentwise
hXi quadratic variation process of X, page 86
d
= equality of finite-dimensional distributions
∼ “is distributed as”
∗ product of copulas, see equation (5.37), page 74
? generalized product of copulas, see equation (5.39), page 76
, A (x) indicator function (of a set A): A (x) = 1 if x ∈ A and otherwise 0
  

a. s. “almost surely”
α(F ), ω(F ) left and right endpoint of the support of F
Bn n-dimensional Borel-σ-algebra
Bt Brownian motion
cdf “cumulative distribution function”
C, Ci , Ct , Ct1 ,...,tn , etc. copulas
Ch · i copula (family) of the argument
Cn set of all n-copulas
b
C survival copula of C, see equation (2.43)
∆ba H(x1 , . . . , xn ) n-th difference of a function H w. r. t. (a, b), see equation (2.2),
page 6
δC (u) := C(u, u) diagonal of the copula C
(i)
∆(ai ,bi ) H(x1 , . . . , xn ) first difference of the function H in the i-th component w. r. t.
(ai , bi ), see equation (2.3), page 6
100 List of Symbols and Abbreviations

Dk , Di,j,k , etc. partial derivatives of a copula, page 17

dom(H) domain of the function H
DX finite-dimensional distributions of X
F =1−F univariate survival function
F, G, Fi , etc. cumulative distribution functions (cdf)
Φ, ϕ cdf and pdf of a standard normal random variable
n
I, I unit interval, n-dimensional unit cube
iff “if and only if”
iid “independent and identically distributed”
inf, sup infimum, supremum
λn n-dimensional Lebesgue measure
λU , λL coefficients of upper and lower tail dependence, page 38
Lt level curve, page 24
LTD “left-tail decreasing”, page 36
M upper Fréchet bound, see equation (2.11), page 9
M2 set of square-integrable martingales with M0 = 0
min, max minimum, maximum
2
N(µ, σ ) normal distribution with expecation µ and variance σ 2
(Ω, A, P) probability space
P, Q probability measures
PC probability measure induced by a copula C
pdf “probability density function”
Π independence copula, see equation (2.12), page 9
PLR “positive likelihood ratio dependent”
PQD “positively quadrant dependent”
+
, , real line, extended real line, non-negative real line
ran(H) range of the function H
n
R set of all n-dimensional rectangles
Rxy n-dimensional rectangle induced by x and y, page 6
ρ, ρs , ρX,Y Spearman’s ρ
RTI “right-tail increasing”, page 36
SDE “stochastic differential equation”
σ( · ) generated σ-algebra
τ , τX,Y Kendall’s τ
 List of Symbols and Abbreviations 101

U(a, b) uniform distribution on (a, b)

vert(Rxy ) vertices of the rectangle
VH (Rxy ) H-volume of Rxy , see equation (2.1), page 6
W lower Fréchet bound, see equation (2.11), page 9
w. l. o. g. “without loss of generality”
w. r. t. “with respect to”
X, Y, Xi , etc. random variables
X(i) i-th order statistic
 end of example or remark
 end of proof
102 List of Symbols and Abbreviations
Bibliography

Numbers at the end of each entry refer to the page(s) on which the work is cited.

Alsina, Claudi, Maurice J. Frank, and Berthold Schweizer (2001). Associative functions on
intervals: A primer of triangular norms. Monograph, to appear in 2001. 23

Bauer, Heinz (1992). Maß- und Integrationstheorie. de Gruyter, Berlin, second edition.
18, 62, 66

Behnen, Konrad and Georg Neuhaus (1995). Grundkurs Stochastik. B. G. Teubner,

Stuttgart, third edition. 7

Billingsley, Patrick (1995). Probability and Measure. Wiley, New York, third edition. 11,
17, 18, 29, 66

Cambanis, Stamatis (1977). Some properties and generalizations of multivariate Eyraud-

Gumbel-Morgenstern distributions. Journal of Multivariate Analysis, 7, 551–559. 71

Cambanis, Stamatis (1991). On Eyraud-Farlie-Gumbel-Morgenstern random processes.

In: G. Dall’Aglio, S. Kotz, and G. Salinetti (editors), Advances in Probability Distri-
butions with Given Marginals: Beyond the Copulas, volume 67 of Mathematics and Its
Applications. Kluwer Academic Publishers, Dordrecht – Boston – London. 3, 71, 72

Capéraà, Philippe and Christian Genest (1993). Spearman’s ρ is larger than Kendall’s τ
for positively dependent random variables. Nonparametric Statistics, 2, 183–194. 37, 49

Clayton, David G. (1978). A model for association in bivariate life tables and its application
in epidemiological studies of familial tendency in chronic disease incidence. Biometrika,
65(1), 141–151. 44
104 Bibliography

Cook, R. Dennis and Mark E. Johnson (1981). A family of distributions for modelling
non-elliptically symmetric multivariate data. Journal of the Royal Statistical Society B,
43(2), 210–218. 2, 44, 46

Cook, R. Dennis and Mark E. Johnson (1986). Generalized Burr-Pareto-logistic distri-

butions with applications to a uranium exploration data set. Technometrics, 28(2),
123–131. 44

Cramér, Harald and M. Ross Leadbetter (1967). Stationary and Related Stochastic Pro-
cesses: Sample Function Properties and Their Applications. Wiley, New York. 54

Dambis, Karl E. (1965). On the decomposition of continuous submartingales. Theor.

Probab. Appl., 10, 401–410. 4, 86, 87

Darsow, William F., Bao Nguyen, and Elwood T. Olsen (1992). Copulas and Markov
processes. Illinois Journal of Mathematics, 36(4), 600–642. 2, 3, 12, 17, 57, 74, 75, 76,
77

David, Herbert Aron (1981). Order Statistics. Wiley, New York. 41, 45

Devroye, Luc (1986). Non-Uniform Random Variate Generation. Springer, New York. 95

Dubins, Lester E. and Gideon Schwarz (1965). On continuous martingales. Proc. Nat.
Acad. Sciences U.S.A., 53, 913–916. 4, 86, 87

Elstrodt, Jürgen (1996). Maß- und Integrationstheorie. Springer, Berlin. 17, 18

Embrechts, Paul, Alexander McNeil, and Daniel Straumann (1999). Correlation and
dependence in risk management: Properties and pitfalls. Working Paper. Risklab:
[Link] 2, 33

Esary, James D., Frank Proschan, and David W. Walkup (1967). Association of random
variables, with applications. The Annals of Mathematical Statistics, 38(2), 1466–1474.
36

Eyraud, Henri (1938). Les principes de la mesure des corrélations. Annales de l’Université
de Lyon, Series A, 1, 30–47. 70

Farlie, D. J. G. (1960). The performance of some correlation coefficients for a general bi-
variate distribution. Biometrika, 47, 307–323. 70

Fisher, Nicholas I. (1997). Copulas. In: Samuel Kotz, Campbell B. Read, and David L.
Banks (editors), Encyclopedia of Statistical Sciences, Update Volume 1. Wiley, New York,
pp. 159–163. 2

Frank, Maurice J. (1979). On the simultaneous associativity of f (x, y) and x + y − f (x, y).
Aequationes Mathematicae, 19, 194–226. 84, 85
 Bibliography 105

Fréchet, Maurice (1957). Les tableauxs de corrélations dont les marges sont données.
Annales de l’Université de Lyon, Sciences Mathématiques et Astronomie, Serie A(4),
13–31. 11
Galambos, János (1978). The Asymptotic Theory of Extreme Order Statistics. Wiley,
New York. 1, 45
Genest, Christian (1987). Frank’s family of bivariate distributions. Biometrika, 74(3),
549–555. 84
Genest, Christian and Louis-Paul Rivest (1993). Statistical inference procedures for bi-
variate Archimedean copulas. Journal of the American Statistical Association, 88(423),
1034–1043. 24, 25, 96
Georges, P., A.-G. Lamy, E. Nicolas, G. Quibel, and Thierry Roncalli
(2001). Multivariate survival modelling: A unified approach with copu-
las. working paper, Groupe de Recherche Opérationelle Crédit Lyonnais,
[Link] 30
Gould, Henry W. (1972). Combinatorial Identities. A Standardized Set Of Tables Listing
500 Binomial Coefficient Summations. Rev. Ed. Morgantown, W. Va. 48
Gumbel, Emile J. (1958). Distributions à plusieurs variables dont les marges sont données.
C. R. Acad. Sci. Paris, 246, 2717–2719. 70
Heuser, Harro (1988). Lehrbuch der Analysis, Teil 2. Teubner, Stuttgart, fourth edition.
82
Hoeffding, Wassily (1940). Maßstabinvariante Korrelationstheorie. Schriften des Mathe-
matischen Seminars und des Instituts für Angewandte Mathematik der Universität
Berlin, 5(3), 179–233. 11
Ihaka, Ross and Robert Gentleman (1996). R: A language for data analysis and graphics.
Journal of Computational and Graphical Statistics, 5(3), 299–314. 97
Joe, Harry (1997). Multivariate Models and Dependence Concepts. Chapman & Hall,
London. 33, 36
Jogdeo, Kumar (1982). Dependence, concepts of. In: Samuel Kotz, Norman L. Johnson,
and Campbell B. Read (editors), Encyclopedia of Statistical Sciences, volume 2. Wiley,
New York, pp. 324–334. 1
Johnson, Norman L. and Samuel Kotz (1975). On some generalized Farlie-Gumbel-
Morgenstern distributions. Communications in Statistics, 4(5), 415–427. 71
Kamke, Erich (1960). Das Lebesgue-Stieltjes-Integral. B. G. Teubner, Leipzig, second
edition. 18, 21
Kamps, Udo (1995). A Concept of Generalized Order Statistics. Teubner Skripten zur
Mathematischen Stochastik. Teubner, Stuttgart. 80
106 Bibliography

Karatzas, Ioannis and Steven E. Shreve (2000). Brownian Motion and Stochastic Calculus.
Graduate Texts in Mathematics. Springer, New York, second edition. 51, 53, 54, 60, 61,
63, 82, 86, 87, 88, 90

Kimeldorf, George and Allan Sampson (1975). Uniform representations of bivariate distri-
butions. Communications in Statistics, 4(7), 617–627. 2, 44

Kulpa, Tomasz (1999). On approximation of copulas. International Journal of Mathematics

& Mathematical Sciences, 22(2), 259–269. 15

Lange, Kenneth (1998). Numerical Analysis for Statisticians. Springer, New York. 95

Lehmann, Erich L. (1966). Some concepts of dependence. The Annals of Mathematical

Statistics, 37(2), 1137–1153. 36, 39, 40

Li, Xin, Piotr Mikusiński, and Michael D. Taylor (1998). Strong approximation of copulas.
Journal of Mathematical Analysis and Applications, 225, 608–623. 15

Ling, C. H. (1965). Representation of associative functions. Publ. Math. Debrecen, 12,

189–212. 23

Mardia, Kanti V. (1962). Multivariate Pareto distributions. Annals Math. Statist., 33,
1008–1015. 44

Meyn, Sean P. and Richard L. Tweedie (1996). Markov Chains and Stochastic Stability.
Springer, London, third edition. 74

Monroe, Itrel (1978). Processes that can be embedded in Brownian motion. The Annals
of Probability, 6(1), 42–56. 91

Morgenstern, Dietrich (1956). Einfache Beispiele zweidimensionaler Verteilungen. Mit-

teilungsblatt für Mathematische Statistik, 8, 234–235. 70

Nelsen, Roger B. (1986). Properties of a one-parameter family of bivariate distributions

with given marginals. Comm. Statist. – Theory and Methods, 15, 3277–3285. 84

Nelsen, Roger B. (1991). Copulas and association. In: Giorgio Dall’Aglio, Samuel Kotz, and
Gabriella Salinetti (editors), Advances in Probability Distributions with Given Marginals:
Beyond the Copulas, volume 67 of Mathematics and Its Applications. Kluwer Academic
Publishers, Dordrecht – Boston – London. 49

Nelsen, Roger B. (1999). An Introduction to Copulas. Lecture Notes in Statistics. Springer,

New York. 9, 10, 11, 22, 23, 24, 25, 26, 33, 35, 36, 39, 40, 44, 84, 95

Oakes, David (1982). A model for association in bivariate survival data. Journal of the
Royal Statistical Society B, 44(3), 414–422. 44

Oakes, David (1986). Semiparametric inference in a model for association in bivariate

survival data. Biometrika, 73(2), 353–361. 44
 Bibliography 107

Øksendal, Bernt (1998). Stochastic Differential Equations. Springer, New York, fifth edi-
tion. 86, 87, 88, 90

Pfeifer, Dietmar (1989). Einführung in die Extremwertstatistik. Teubner, Stuttgart. 93,

94

Randles, Ronald H. and Douglas A. Wolfe (1979). Introduction to the Theory of Nonpara-
metric Statistics. Wiley, New York. 26

Reiss, Rolf-Dieter (1989). Approximate Distributions of Order Statistics. Springer,

New York. 45

Resnick, Sidney I. (1987). Extreme Values, Regular Variation, and Point Processes.
Springer, New York. 77, 78, 80

Revuz, Daniel and Marc Yor (1991). Continuous Martingales and Brownian Motion.
Springer, Berlin. 53, 55, 86, 87

Rogers, L. Chris G. and David Williams (2000a). Diffusions, Markov Processes and Martin-
gales, Volume 1: Foundations. Cambridge University Press, Cambridge, second edition.
4, 17, 18, 47, 51, 53, 86

Rogers, L. Chris G. and David Williams (2000b). Diffusions, Markov Processes and Martin-
gales, Volume 2: Itô Calculus. Cambridge University Press, Cambridge, second edition.
4, 53, 86

Ross, Sheldon M. (1996). Stochastic Processes. Wiley, New York, second edition. 53

Schmitz, Volker and Erwin Pier-Ribbert (2001). Copulas in finance. derivatives week,
10(6), 6–7. 2

Schweizer, Berthold (1991). Thirty years of copulas. In: G. Dall’Aglio, S. Kotz, and
G. Salinetti (editors), Advances in Probability Distributions with Given Marginals, Math-
ematics and Its Applications. Kluwer Academic Publishers, Dordrecht. 2, 22

Shiryaev, Albert N. (1996). Probability. Springer, New York, second edition. 11, 56, 57

Sklar, Abe (1959). Fonctions de répartition à n dimensions et leurs marges. Publ. Inst.
Statist. Univ. Paris, 8, 229–231. 1, 9, 10

Sklar, Abe (1973). Random variables, joint distribution functions, and copulas. Kyber-
netika, 9(6), 449–460. 9

Todorovic, Petar (1992). An Introduction to Stochastic Processes and Their Applications.

Springer, New York. 53, 59, 65, 69

Walsh, John E. (1969). Asymptotic independence between largest and smallest of a set of
independent observations. Ann. Inst. Statist. Math., 21, 287–289. 45

Walter, Wolfgang (1992). Analysis 1. Springer, Berlin, third edition. 45, 67

108 Bibliography
Index

Adapted process, 54 Dependence

measures, 33–40
Brownian bridge, 90 tail-
Brownian motion, 51–52, 60, 70 lower-, 38
geometric-, 62 upper-, 38
symmetry of-, 82 Difference operator, 6
Discordance, see Concordance
Chapman-Kolmogorov equation, 74, 76
Domain, 6
Checkerboard approximation, 15
Coefficient of lower tail dependence, 39 EFGM process, see Process, EFGM
Coefficient of upper tail dependence, 39 Equivalence
Comonotonicity, 33 of processes, 58–59
Concordance, 33
Copula, 7 Filtration, 54
Archimedean-, 22–25, 57 Finite-dimensional distribution, 56
diagonal-, 23 Fréchet bound
Frank-, 23, 84 lower, 9, 23
independence-, 9, 23, 57 upper, 9, 23, 67
induced measure, 11–12 Fréchet bounds, 9, 10, 33, 75
min-max-, 27, 41–50 stochastic interpretation, 10
of Brownian motion, 61 Generalized inverse, 93
of Brownian motion and its supremum, Generator, 22
51 Grounded, groundedness, 7
of geometric Brownian motion, 63
of record times, 78 Increasing
strict, 23 n-increasing, 7
survival-, 29, 25–31 Indistinguishability, 59
Countermonotonicity, 33 Interval
Cross variation, 86 halfopen, 5
110 Index

unit, see Unit interval equivalence, 54

Symmetry
Kendall’s τ , 34 joint-, 26
for min-max copula, 46 marginal-, 26
Level curve, 24 of Markov processes, 81–85
LTD, 36, 49 radial-, 26, 28, 60

Martingale, 55, 57 Tail dependence, 36

continuous local-, 86 Time series, 57
Modification, 59 Transition probability, 74

n-increasing, see increasing Uniform process, see Process, uniform-

Unit cube, 6
Order statistic, 41 Unit interval, 5
Usual conditions, 54
Partial derivatives, 16–22
Positive likelihood ratio dependent, 40 Volume of a rectangle, 6
Positively quadrant dependent, 39
Process
diffusion-, 57
EFGM-, 70–73
Markov-, 57, 73–77
Ornstein-Uhlenbeck-, 87
stochastic-, 53
stopped-, 55
symmetric-, 60
uniform-, 57
Product of copulas, 74

Quadratic variation, 86
Quasi-monotone, see Increasing, n-increasing

Radial symmetry, see Symmetry, radial

Random variate generation, 95
Range, 6
Records, 77–80
Rectangle
halfopen, 5
RTI, 36, 49

Sklar, theorem of, 10

Spearman’s ρ, 35
Square-integrable martingale, 86
Stationarity
strict, 58
Stochastic
continuity, 65–70
Lebenslauf Volker Schmitz

14. Mai 1973 In Düren geboren als Sohn von Irene und Klaus Schmitz

1979–1983 Besuch der Katholischen Grundschule in Kreuzau

1983–1992 Besuch des Burgau Gymnasiums in Düren

27. Juni 1992 Erwerb der allgemeinen Hochschulreife (Abitur) am Burgau

Gymnasium in Düren

1992–1997 Studium der Mathematik mit Nebenfach Betriebswirtschafts-

lehre an der RWTH Aachen

1995–1997 Studentische Hilfskraft am Institut für Statistik und

Wirtschaftsmathematik der RWTH Aachen

1997–1998 Ableisten des Zivildienstes

seit 1. Juli 1998 Wissenschaftlicher Angestellter am Institut für Statistik und

Wirtschaftsmathematik der RWTH Aachen