KYBERNETI KA MANUS CRI P T P REVI EW

CONSTRUCTION OF MULTIVARIATE COPULAS

IN N-BOXES - PART II
JOS

E M. GONZ

ALEZ-BARRIOS AND MAR

IA M. HERN

ANDEZ-CEDILLO
In this paper we give an alternative proof of the construction of n-dimensional ordinal sums given in
Mesiar and Sempi [16], we also provide a new methodology to construct n-copulas extending the patch-
work methodology of Durante, Saminger-Platz and Sarkoci in [6] and [7]. Finally, we use the gluing
method of Siburg and Stoimenov [19] to give an alternative method of patchwork construction of n-copulas,
which can be also used in composition with our patchwork method.
Keywords: n-copulas, modular functions, rectangular patchwork
Classication: 60A10, 60E05
1. INTRODUCTION
In this paper we we will use the same notation and denitions as in the rst part of Gonz alez-
Barrios and Hern andez-Cedillo [11]. In the rest of this manuscript we will refer to this paper
simply by the rst part. We will start this section by recalling the main theoretical results
whose proofs can be found in the rst part. We start with a generalization of De Baets and De
Meyer [2]
Theorem 1.1. Let C : [0, 1]
n
[0, 1] be an n-copula, let R =
n
i=1
[u
i
, v
i
] [0, 1]
n
be a non
trivial n-box. Let D : R [0, 1] be a function. Dene Q : [0, 1]
n
[0, 1] by
Q(x
1
, . . . , x
n
) =
_
D(x
1
, . . . , x
n
) if x
1
, . . . , x
n
R,
C(x
1
, . . . , x
n
) if x
1
, . . . , x
n
[0, 1]
n
\R.
(1)
Then, Q is an n-copula if and only if D = C on (R) and D is n-increasing.
Using Aczel and Dhombres [1], we give a characterization of modular functions and a useful
Lemma.
Lemma 1.2. Let F : D R I be a function where D R I
n
for some n 2. Then F is modular
if and only if there exist n functions G
i
: R I
n1
R I such that for every x = x
1
, x
2
, . . . , x
n
D
F(x) = G
1
(x
2
, x
3
, . . . , x
n
) + + G
i
(x
1
, . . . , x
i1
, x
i+1
, . . . , x
n
) + + G
n
(x
1
, . . . , x
n1
). (2)
2 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
Lemma 1.3. Let n 2, let R =
n
i=1
[u
i
, v
i
] [0, 1]
n
be an n-box, let D : R R I be an
n-increasing function and let E : R R I be a modular function. If we dene F : R R I by
F(x) = D(x) + E(x). (3)
Then F is an n-increasing function.
The following Theorem is the main result in the patchwork construction of 2-copulas given in
Durante et al [7], and an alternative proof is given in the rst part.
Theorem A Let C be a copula, let {C
j
}
jJ
be a family of copulas and let {R
j
= [u
j
1
, v
j
1
]
[u
j
2
, v
j
2
]}
jJ
be a family of 2-boxes, in this case rectangles in [0, 1]
2
, such that R
j
R
k
(R
j
)
(R
k
) for every j, k J with j k. Dene for every j J,
j
= V
C
(R
j
), and for every
x [u
j
1
, v
j
1
] and for every y [u
j
2
, v
j
2
], R
j,x
= [u
j
1
, x] [u
j
2
, v
j
2
] and R
j,y
= [u
j
1
, v
j
1
] [u
j
2
, y]. Let

C : [0, 1]
2
[0, 1] dened by

C(x, y) =
_

j
C
j
_
V
C
(R
j,x
)

i
,
V
C
(R
j,y
)

j
_
+
C
j
(x, y) if (x, y) R
j
and
j
> 0,
C(x, y), otherwise,
(4)
where
C
j
(x, y) = h
C
u
j
2
(x) + v
C
u
j
1
(y) h
C
u
j
2
(u
j
1
). Then

C is a copula.
An important result about n-increasing functions that will be used in this paper is the following
Lemma 1.4. Let n 2 and let R =
n
i=1
[u
i
, v
i
] [0, 1]
n
be a non trivial n-box, let C : [0, 1]
n

[0, 1] and D : [0, 1]

n
[0, 1] be two n-copulas. Let = V
C
(R) and assume that > 0. Dene
E : R [0, ] by
E(x) = D
_
V
C
(R
x
1
)

, . . . ,
V
C
(R
x
n
)

_
for every x R, (5)
where for every i {1, . . . , n}, u
i
x
i
v
i
and
R
x
i
= [u
1
, v
1
] [u
i1
, v
i1
] [u
i
, x
i
] [u
i+1
, v
i+1
] [u
n
, v
n
]. (6)
Then E is an n-increasing function on R.
Besides, if x = v
1
, . . . , v
j1
, x
j
, v
j+1
, . . . , v
n
, for some j {1, . . . , n} and u
j
x
j
< v
j
, then
E(x) = V
C
(R
x
j
), in particular E(v
1
, v
2
, . . . , v
n
) = .
In Section 2 we use our methodology to give an alternative proof of the ordinal sum construction
of n-copulas for n 3 rst proved in Mesiar and Sempi [16].
In Section 3 we proposed a patchwork construction method for n-copulas with n 3. We
analyze in detail the dierence between the cases n = 2 and n 3 using the multivariate ordinal
sum construction. We also provide some examples that show the importance of this construction
in the multivariate case.
In section 4 we use the gluing method of Siburg and Stoimenov [19] to provide an alternative
method of construction of copulas in n-boxes, with n 3, and we show that this methodology
can be composed with the patchwork construction given in Section 3 to provide new n-copulas.
In Section 5 we give some nal observations.
Multivariate Copulas in n-Boxes - Part II 3
2. MULTIVARIATE ORDINAL SUMS OF COPULAS
In this section we will use our results to prove that the construction of ordinal sums can be
extended to n 3. This result was rst proved in Mesiar and Sempi [16]. We start with a
general Proposition.
Proposition 2.1. Let C
1
be an n-copula for some n 2, and let 0 a
1
< b
1
1, dene
R =
n
i=1
[a
1
, b
1
] = [a
1
, b
1
]
n
an n-box in [0, 1]
n
. Dene C : [0, 1]
n
[0, 1] by
C(x) =
_
a
1
+ (b
1
a
1
)C
1
_
min{x
1
,b
1
}a
1
b
1
a
1
, . . . ,
min{x
n
,b
1
}a
1
b
1
a
1
_
if min{x
1
, . . . , x
n
} [a
1
, b
1
],
min{x
1
, . . . , x
n
} elsewhere.
(7)
Then C is an n-copula.
Proof: We will proceed by induction. Let n = 2, and let 0 a
1
< b
1
1. Dene C as in
equation (7). Let A
1
= {x
1
, x
2
[0, 1]
2
| min{x
1
, x
2
} [a
1
, b
1
]}, then it is clear that
A
c
1
= {x
1
, x
2
[0, 1]
2
| x
1
< a
1
or x
2
< a
1
}{x
1
, x
2
[0, 1]
2
| x
1
> b
1
and x
2
> b
1
} = A
2
A
3
.
Observe that A
1
is the union of the 3 = 2
2
1 rectangles R
1
= [a
1
, b
1
]
2
, R
2
= [a
1
, b
1
] [b
1
, 1]
and R
3
= [b
1
, 1] [a
1
, b
1
], see Figure 1.
Figure 1.- Regions A
1
, A
2
and A
3
1 0 a
1
b
1
x
1
a
1
b
1
1
x
2
T
E
A
2
A
1
R
1
R
2
R
3
A
3
. .
.
.
Dene D : A
1
R I by
D(x
1
, x
2
) = a
1
+ (b
1
a
1
)C
1
_
min{x
1
, b
1
} a
1
b
1
a
1
,
min{x
2
, b
1
} a
1
b
1
a
1
_
. (8)
If x
1
, x
2
R
2
then D(x
1
, x
2
) = a
1
+ (b
1
a
1
)C
1
((x
1
a
1
)/(b
1
a
1
), 1) = a
1
+ (x
1
a
1
) = x
1
=
min{x
1
, x
2
}. Similarly, if x
1
, x
2
R
3
then D(x
1
, x
2
) = x
2
= min{x
1
, x
2
}. Using Theorem 1.1
we have to see that D is 2-increasing on R
1
and that D and M(x
1
, x
2
) = min{x
1
, x
2
} coincide on
(R
1
). We know that C
1
is a 2-copula, so, C
1
is 2-increasing. Dene a function h : R
1
[0, 1]
2
by:
h(x
1
, x
2
) =
_
x
1
a
1
b
1
a
1
,
x
2
a
1
b
1
a
1
_
. (9)
4 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
Then it is clear that h is a bijection which takes R
1
onto [0, 1]
2
, and also h is increasing in each
coordinate. Let S = [x
1,1
, x
1,2
] [x
2,1
, x
2,2
] where a
1
x
i,1
x
i,2
b
1
for i = 1, 2. Then S is a
rectangle included in R
1
, and the function h in equation (9), takes S onto
h[S ] =
_
x
1,1
a
1
b
1
a
1
,
x
1,2
a
1
b
1
a
1
_

_
x
2,1
a
1
b
1
a
1
,
x
2,2
a
1
b
1
a
1
_
,
and using Lemma 1.3 with E(x
1
, x
2
) = a
1
, we have from equation (8) that
V
D
(S ) = (b
1
a
1
)V
C
1
(h[S ]) 0. (10)
Therefore, from equation (10), D is 2-increasing on R
1
. Now, we prove that D coincides with M
on (R
1
). Let x = (x
1
, a
1
) or x = (a
1
, x
2
), where a
1
x
1
, x
2
b
1
. Then, since C
1
is a 2-copula
D(x
1
, a
1
) = a
1
+ (b
1
a
1
)C
1
_
x
1
a
1
b
1
a
1
,
a
1
a
1
b
1
a
1
_
= a
1
= M(x
1
, a
1
) (11)
and
D(a
1
, x
2
) = a
1
+ (b
1
a
1
)C
1
_
a
1
a
1
b
1
a
1
,
x
2
a
1
b
1
a
1
_
= a
1
= M(a
1
, x
2
). (12)
Finally, if x = (x
1
, b
1
) or x = (b
1
, x
2
), where a
1
x
1
, x
2
b
1
. Then, since C
1
is a 2-copula
D(x
1
, b
1
) = a
1
+(b
1
a
1
)C
1
_
x
1
a
1
b
1
a
1
,
b
1
a
1
b
1
a
1
_
= a
1
+(b
1
a
1
)
x
1
a
1
b
1
a
1
= x
1
= M(x
1
, b
1
) (13)
and
D(b
1
, x
2
) = a
1
+(b
1
a
1
)C
1
_
b
1
a
1
b
1
a
1
,
x
2
a
1
b
1
a
1
_
= a
1
+(b
1
a
1
)
x
2
a
1
b
1
a
1
= x
2
= M(b
1
, x
2
). (14)
From equations (11), (12), (13) and (14) D and M coincide on (R
1
) and C in equation (7) is a
2-copula according to Theorem 1.1.
Let n > 2 and let C
1
be an n-copula, let R
1
= [a
1
, b
1
]
n
be an n-box in [0, 1]
n
and let C dened
as in equation (7). Dene D : A
1
R I , where A
1
= {x
1
, . . . , x
n
[0, 1]
n
| min{x
1
, . . . , x
n
}
[a
1
, b
1
]}, then it is clear that
A
c
1
= {x
1
, . . . , x
n
[0, 1]
n
| there exists i {1, . . . , n} such that x
i
< a
1
}
{x
1
, . . . , x
n
[0, 1]
2
| x
i
> b
1
for every i {1, . . . , n}}
= A
2
A
3
. (15)
We will observe that in this case A
1
is the union of 2
n
1 n-boxes with disjoint interiors, which
include R
1
. Let I
1,i
= [a
1
, b
1
] and I
2,i
= [b
1
, 1] for i {1, . . . , n} then as in equation (24) of the
rst part,
[a
1
, 1]
n
= ([a
1
, b
1
] [b
1
, 1])
n
=
j
1
,..., j
n
}{1,2}
n
n
i=1
I
j
i
,i
.
Therefore,
A
1
= [a
1
, 1]
n
\[b
1
, 1]
n
=
j
1
,..., j
n
}{1,2}
n
n
i=1
I
j
i
,i
\
n
i=1
I
2,i
,
Multivariate Copulas in n-Boxes - Part II 5
which is a union of 2
n
1 n boxes with disjoint interiors. Dene for every x
1
, . . . , x
n
A
1
D(x
1
, . . . , x
n
) = a
1
+ (b
1
a
1
)C
1
_
min{x
1
, b
1
} a
1
b
1
a
1
, . . . ,
min{x
n
, b
1
} a
1
b
1
a
1
_
. (16)
Using the same ideas as above, and the fact that C
1
is n-increasing, it is not dicult to see that
D is n-increasing on A
1
. So, by a natural generalization of Theorem 1.1, we only have to see
that M(x
1
, . . . , x
n
) = min{x
1
, . . . , x
n
} and D coincide on (A
1
) (A
2
A
3
).
Using equation (15), if x
1
, . . . , x
n
(A
1
) (A
2
A
3
), then there exists i {1, . . . , n} such
that x
i
= a
1
and for every j {1, . . . , n}\{i}, x
j
[a
1
, 1], or there exists i {1, . . . , n} such that
x
i
= b
1
and for every j {1, . . . , n}\{i}, x
j
[b
1
, 1]. In the rst case using equation (16) and the
frontier conditions of C
1
, we have that
D(x
1
, . . . , x
n
) = a
1
+ (b
1
a
1
) 0 = a
1
= min{x
1
, . . . , x
n
},
and in the second case
D(x
1
, . . . , x
n
) = a
1
+ (b
1
a
1
)C(1, . . . , 1,
b
1
a
1
b
1
a
1
, 1 . . . , 1) = b
1
= min{x
1
, . . . , x
n
}.
Therefore, M and D coincide on (A
1
)(A
2
A
3
) and C dened in equation (7) is an n-copula.
Remark 2.2. In the previous Proposition we dene R = [a
1
, b
1
]
n
. Let us denote by = V
M
(R),
where M(x
1
, . . . , x
n
) = min{x
1
, . . . , x
n
}, we will see that = b
1
a
1
. We know that using formula
(2) in Denition 2.2 of the rst part, we have that
V
M
(R) =

cVert(R)
sgn(c)M(c). (17)
We observe that if c = c
1
, . . . , c
n
Vert(R), then for every i {1, . . . , n} c
i
= a
1
or c
i
= b
1
. So,
if there exists i {1, . . . , n} such that c
i
= a
1
then M(c) = a
1
, and the only vertex of R such that
M(c) = b
1
is c = b
1
, b
1
, . . . , b
1
=: b
1
. Then using (16), we have that
= V
M
(R) = b
1
+ a
1

cVert(R)\{b
1
}
sgn(c). (18)
Observe that
_
cVert(R)
sgn(c) = 0, this follows using the binomial expansion of 0 = ((1) +1)
n
.
But, in this case
0 = = ((1) + 1)
n
=
n

k=0
_
n
k
_
(1)
k
(1)
nk
= 1 +
n

k=1
_
n
k
_
(1)
k
(1)
nk
.
Therefore,
_
n
k=1
_
n
k
_
(1)
k
(1)
nk
= 1, and using (18) we get that = V
M
(R) = b
1
a
1
.
6 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
Remark 2.3. It is very important to observe that in the case n = 2, the construction of ordinal
sums is made by modifying the copula M only on squares that have opposite vertices on the main
diagonal, namely, if R = [a, b]
2
[0, 1]
2
and D is a 2-copula, then the function C : [0, 1]
2

[0, 1] given by
C(x
1
, x
2
) =
_
a + (b a)D
_
x
1
a
ba
,
x
2
b
ba
_
if x
1
, x
2
[a, b]
2
M(x
1
, x
2
) = min{x
1
, x
2
} elsewhere
is a copula, see for example Nelsen (2006). If we try to extend directly this idea to larger dimen-
sions the result is false, that is, if we take any n 3, D an n-copula and R = [a, b]
n
[0, 1]
n
an
n-box with opposite vertices on the main diagonal and we dene a function C : [0, 1]
n
[0, 1]
by
C(x
1
, . . . , x
n
) =
_
a + (b a)D
_
x
1
a
ba
, ,
x
n
b
ba
_
if x
1
, . . . , x
n
[a, b]
n
M(x
1
, . . . , x
n
) = min{x
1
, . . . , x
n
} elsewhere.
Then C is not necessarily an n-copula. To see this, we give an easy example with n = 3, D =
and R = [0, 1/3]
3
. If we dene C as in the above equation, we have that
C(x
1
, x
2
, x
3
) =
_
1
3
(3x
1
, 3x
2
, 3x
3
) if x
1
, x
2
, x
3
[0, 1/3]
3
M(x
1
, x
2
, x
3
) = min{x
1
, x
2
, x
3
} elsewhere.
(19)
In this case, if we take x
1
, x
2
, x
3
= 1/3, 1/4, 1/4, then x
1
, x
2
, x
3
is a point on an upper face
of the 3-box R, but
1
3

_
3
1
3
, 3
1
4
, 3
1
4
_
=
3
16

1
4
= min
_
1
3
,
1
4
,
1
4
_
.
Therefore, by Proposition 2.7 in the rst part, C in equation (19) is not a 3-copula.
By Proposition 2.1, we know that if we dene
C
1
(x
1
, x
2
, x
3
) =
_
1
3

_
min{x
1
,1/3}
1/3
,
min{x
2
,1/3}
1/3
,
min{x
3
,1/3}
1/3
_
if min{x
1
, x
2
, x
3
} [0, 1/3],
min{x
1
, x
2
, x
3
} elsewhere.
(20)
Then C
1
is a 3-copula. Observe that the big dierence between equations (19) and (20) is that
in the rst line of (19) the region is R = [0, 1/3]
3
, and in the rst line of equation (20) the region
is [0, 1]
3
\(1/3, 1]
3
, that is, the complement of the 3-box (1/3, 1]
3
.
Of course we can extend Proposition 2.1 to obtain the multivariate version of ordinal sums, as in
Mesiar and Sempi [16]. See also the application given in Durante and Fern andez-S anchez [8].
Theorem B Let {C
j
}
jJ
be a family of n-copulas, let {[a
j
, b
j
]}
jJ
where J = {1, . . . , n} or
J = {1, 2, . . .}. Assume that for every j J, 0 a
j
< b
j
1, and even more for every
j, j + 1 J, b
j
a
j+1
. Dene C : [0, 1]
n
[0, 1] by
C(x) =
_

_
a
j
+ (b
j
a
j
)C
j
_
min{x
1
,b
j
}a
j
b
j
a
j
, . . . ,
min{x
n
,b
j
}a
j
b
j
a
j
_
if min{x
1
, . . . , x
n
} [a
j
, b
j
] for j J
M(x) = min{x
1
, . . . , x
n
} elsewhere.
(21)
Then C is an n-copula.
Proof: The proof of Theorem B is an easy induction that uses Proposition 2.1, equation (21),
and the same arguments that we use on the proof of Theorem A in the rst part.
Multivariate Copulas in n-Boxes - Part II 7
3. A MULTIVARIATE PATCHWORK CONSTRUCTION
In this section we provide a multivariate patchwork construction of n-copulas in n-boxes by
using the regions determined in multivariate ordinal sums. We will start by taking a 3copula
and a 3-box R with 1, 1, 1 as one of its vertices.
Theorem 3.1. Let C and C
1
be two 3-copulas and let R = [u
1
, 1] [u
2
, 1] [u
3
, 1] where
0 < u
i
< 1 for i {1, 2, 3} and dene 0 = 0, 0, 0. Assume that = V
C
(R) > 0, and for every
x
1
[u
1
, 1], for every x
2
[u
2
, 1] and for every x
3
[u
3
, 1], dene R
x
1
= [u
1
, x
1
][u
2
, 1][u
3
, 1],
R
x
2
= [u
1
, 1] [u
2
, x
2
] [u
3
, 1] and R
x
3
= [u
1
, 1] [u
2
, 1] [u
3
, x
3
]. Let

C : [0, 1]
3
[0, 1] be
dened in x = x
1
, x
2
, x
3
by

C(x) =
_

_
C
1
_
V
C
(R
x
1
)

,
V
C
(R
x
2
)

,
V
C
(R
x
3
)

_
+ V
C
([0, x] \ [u, x]) if x R
C(x), otherwise,
(22)
where [a, b] = [a
1
, b
1
] [a
2
, b
2
] [a
3
, b
3
] for a = [a
1
, a
2
, a
3
], b = [b
1
, b
2
, b
3
]. Then

C is a
3-copula.
Remark 3.2. The 3-box R can be written as R = [u, 1] where 1 = 1, 1, 1.
Proof: Let D(x) = E(x) + F(x) with E(x) = C
1
_
V
C
(R
x
1
)

,
V
C
(R
x
2
)

,
V
C
(R
x
3
)

_
and F(x) = V
C
([0, x] \
[u, x]). By Lemma 1.4 E is a 3-increasing function.
To see that D is also 3-increasing, using Lemma 1.3, we just need to prove that F is a modular
function. Let x R then,
F(x) = V
C
([0, x]) V
C
([u, x])
= C(x)

cVert([u,x])
sgn(c)C(c)
= C(x
1
, x
2
, x
3
) {C(x
1
, x
2
, x
3
)
C(u
1
, x
2
, x
3
) C(x
1
, u
2
, x
3
) C(x
1
, x
2
, u
3
)
+C(x
1
, u
2
, u
3
) + C(u
1
, x
2
, u
3
) C(u
1
, u
2
, x
3
)
C(u
1
, u
2
, u
3
)}
= C(u
1
, x
2
, x
3
) + C(x
1
, u
2
, x
3
) + C(x
1
, x
2
, u
3
)
C(x
1
, u
2
, u
3
) C(u
1
, x
2
, u
3
) C(u
1
, u
2
, x
3
)
+C(u
1
, u
2
, u
3
), (23)
but equation (23) is a modular function by the observation just above Lemma 1.3.
Now we will prove that

C is 3-increasing. Let x be a point in one of the lower faces of R.
Without loss of generality let x = x
1
, x
2
, u
3
(R) with u
1
x
1
1, u
2
x
2
1. Then
V
C
(R
u
3
) = 0 and D(x) = C
1
_
V
C
(R
x
1
)

,
V
C
(R
x
2
)

,
0

_
+ V
C
([0, x
1
] [0, x
2
] [0, u
3
]) 0 = C(x). So,
D = C in the lower faces of R.
Using the second part of the proof of Theorem 1.1, namely, Theorem 2.12 and Lemma 2.11
in the rst part, we can see that V

C
(S ) = V
C
(S {[0, 1]
3
\R}) + V
D
(S R) 0 for any 3-box
S [0, 1]
3
and so

C is 3-increasing.
8 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
Finally, we prove that

C satises the boundary conditions of a copula.
First

C(x
1
, x
2
, x
3
) = C(x
1
, x
2
, x
3
) = 0 if any of the x
i
= 0 for i {1, 2, 3}, and

C(1, 1, 1) =
C(1, 1, 1)+V
C
([0, 1]
3
)V
C
([u, 1]) = +1V
C
(R) = 1 by the denition of . Then, since

C = C
in [0, 1]
3
\R we just need to see that

C(x) = D(x) = x
i
, for x = x
1
, 1, 1, 1, x
2
, 1, 1, 1, x
3

with x
i
[u
i
, 1], i = 1, 2, 3. By the second part of Lemma 1.4 we have E(x
1
, 1, 1) = V
C
(R
x
1
),
E(1, x
2
, 1) = V
C
(R
x
2
), E(1, 1, x
3
) = V
C
(R
x
3
) for x
i
[u
i
, 1], i = 1, 2, 3. Without losing generality
let us assume x = x
1
, 1, 1, where u
1
x
1
1, then
D(x) = E(x
1
, 1, 1) + F(x
1
, 1, 1)
= V
C
(R
x
1
) + V
C
([0, x
1
] [0, 1] [0, 1]) V
C
([u
1
, x
1
] [u
2
, 1] [u
3
, 1])
= V
C
(R
x
1
) + C(x
1
, 1, 1) V
C
(R
x
1
)
= x
1
. (24)
Similar results as (24) hold if x = 1, x
2
, 1 or if x = 1, 1, x
3
. On the other hand if x = x
1
, 1, 1
where 0 x
1
u
1
, then

C(x) = C(x) = x
1
, since C is a 3-copula. Similar results are obtained
for x = 1, x
2
, 1 and x = 1, 1, x
3
when 0 x
2
u
2
and 0 x
3
u
3
. Therefore,

C in equation
(22) is a 3-copula.
Remark 3.3. If we let u
i
= 0 for some i {1, 2, 3} in the previous Theorem the result still holds.
For example if u
1
= 0, then R = [0, 1] [u
2
, 1] [u
3
, 1], and if we take x = 0, x
2
, x
3
where
u
2
x
2
1 and u
3
x
3
1, then by denition (22) we have that

C(x) = C
1
_
V
C
(R
x
1
=0
)

,
V
C
(R
x
2
)

,
V
C
(R
x
3
)

_
+ V
C
([0, x]) V
C
([u, x])
= 0 + V
C
([0, 0] [0, x
2
] [0, x
3
]) V
C
([0, 0] [u
2
, x
2
] [u
3
, x
3
])
= 0.
Clearly, Theorem 3.1 can be generalized easily to larger dimensions.
Theorem 3.4. For every n 3 let C and C
1
be two n-copulas and let R = [u
1
, 1] [u
2
, 1]
[u
n
, 1] where 0 u
i
< 1 for i {1, . . . , n}. Assume that = V
C
(R) > 0, and for every
i {1, . . . , n} and for every x
i
[u
i
, 1] dene R
x
i
= [u
1
, 1] [u
i1
, 1] [u
i
, x
i
] [u
i+1
, 1]
[u
n
, 1]. Let (C
_
u
C
1
) : [0, 1]
n
[0, 1] be dened in x = x
1
, . . . , x
n
by
_
C
_
u
C
1
_
(x) =
_

_
C
1
_
V
C
(R
x
1
)

, . . . ,
V
C
(R
xn
)

_
+ V
C
([0, x] \ [u, x]) if x R
C(x), otherwise,
(25)
where [a, b] = [a
1
, b
1
] [a
n
, b
n
] for a = [a
1
, . . . , a
n
], b = [b
1
, . . . , b
n
]. Then (C
_
u
C
1
) is
an n-copula.
Proof: It follows the same steps as the proof of Theorem 3.1
Remark 3.5. For every n 3 and for every C n-copula we can obtain from equation (25) every
n-copula C
1
. Let C and C
1
arbitrary n-copulas and let R = [0, 1]
n
, then V
C
(R) = 1 = , and for
every i {1, . . . , n} and for every x
i
[0, 1], R
x
i
= [0, 1] [0, 1] [0, x
i
] [0, 1] [0, 1].
So, V
C
(R
x
i
) = x
i
for every i {1, . . . , n} and u = 0. Therefore, from equation (25) we have that
(C
_
u
C
1
)(x) = C
1
(x) for every x [0, 1]
n
.
Multivariate Copulas in n-Boxes - Part II 9
Using Theorem 3.4 we can construct many dierent new n-copulas. Observe that in the con-
struction of the copula (C
_
u
C
1
) on R = [u, 1], given in equation (25), the copula C remains
xed on [0, 1]
n
\R, and on R we have a rescaled version of the copula C
1
. Using Theorem 3.4 we
have the following
Denition 3.6. Let C
n
be the family of all n-copulas for some n 3, for every xed u [0, 1)
n
,
the function
_
u
: C
n
C
n
C
n
dened by
_
u
(C, C
1
) = (C
_
u
C
1
) (26)
is an operator.
Lemma 3.7. Let n = 3, u (0, 1)
3
and the product 3-copula and let C
1
and C
2
be 3-copulas.
Then
_
C
1
_
u
C
2
) = (27)
if and only if C
1
= C
2
= .
Proof: Let n = 3, u (0, 1)
3
and R = [u, 1]. First, assume that C
1
= C
2
= , then = V

(R) =
(1u
1
)(1u
2
)(1u
3
). Now, we will see that V

(R
x
i
)/ = (x
i
u
i
)/(1u
i
) for every i {1, 2, 3}.
Without losing generality we will assume that i = 1. Since R
x
1
= [u
1
, x
1
] [u
2
, 1] [u
3
, 1] then
V

(R
x
1
)

=
(x
1
u
1
)(1 u
2
)(1 u
3
)
(1 u
1
)(1 u
2
)(1 u
3
)
=
(x
1
u
1
)
(1 u
1
)
.
So, using equations (22), (23) and Denition 3.6, we have that for x R
(C
1
_
u
C
2
)(x) =
_
V

(R
x
1
)

,
V

(R
x
2
)

,
V

(R
x
3
)

_
+ V

([0, x] \ [u, x])

=
_
(x
1
u
1
)
(1 u
1
)
,
(x
2
u
2
)
(1 u
2
)
,
(x
3
u
3
)
(1 u
3
)
_
+ V

([0, x] \ [u, x])

= (x
1
u
1
)(x
2
u
2
)(x
3
u
3
) + x
1
x
2
x
3
(x
1
u
1
)(x
2
u
2
)(x
3
u
3
)
= x
1
x
2
x
3
= (x).
Conversely, assume that equation (27) holds for some C
1
and C
2
3-copulas, then from equation
(22) it is clear that C
1
must be in [0, 1]
3
\R. Then = (1 u
1
)(1 u
2
)(1 u
3
) and V
C
1
([0, x] \
[u, x]) = x
1
x
2
x
3
(x
1
u
1
)(x
2
u
2
)(x
3
u
3
). But, from equations (22) and (27) this implies that
C
2
= .
Of course, Lemma 3.7 also holds for n > 3.
Example 3.8. Let n = 3, let C
1
(x) = exp
_

_
(ln(x
1
))

+ (ln(x
2
))

+ (ln(x
3
))

_
1/
_
for some
1, which is a member of the Gumbel-Hougaard Archimedean family, see Example 4.23 in
Nelsen [17], and let C = . Let R = [1/2, 1] [1/2, 1] [1/2, 1], then = V

(R) = 1/8, and

V

(R
x
i
) = (1/4)(x
i
1/2) for every i {1, 2, 3} and for every x
i
[1/2, 1]. Besides, V

([0, x]) =
10 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
x
1
x
2
x
3
and since u = 1/2, 1/2, /1/2 then V

([u, x]) = (x
1
1/2)(x
2
1/2)(x
3
1/2). Therefore,
using Theorem 3.1 if we dene

C(x) =
_
1
8
C
1
(2x
1
1, 2x
2
1, 2x
3
1) + x
1
x
2
x
3
(x
1
1/2)(x
2
1/2)(x
3
1/2) if x R
x
1
x
2
x
3
otherwise.
Then

C is a 3-copula, which behaves like a rescaled version of the Gumbel-Hougaard family on
the upper 3-box R and on the rest is the product copula. It is also clear that

C is not an ordinal
sum.
Example 3.9. We will see that even for n = 3 and a simple 3-box of the form R = [0, v
1
]
[0, v
2
] [0, v
3
] I
3
, if we take two 3-copulas C and D such that = V
C
(R) > 0 and we dene
R
x
i
for x
i
[0, v
i
] as in equation (22) of Theorem 3.1, for every i {1, 2, 3}, and taking
E(x) = D
_
V
C
(R
x
1
)

,
V
C
(R
x
2
)

,
V
C
(R
x
3
)

_
,
which is 3-increasing by Lemma 1.4. We can nd functions F
1,2
(x
1
, X
2
), F
1,3
(x
1
, x
3
), F
2,3
(x
2
, x
3
)
from [0, 1]
2
into R I , and functions H
1
(x
1
), H
2
(x
2
) and H
2
(x
2
) from [0, 1] into R I , such that if we
dene for every x R
(x) = F
1,2
(x
1
, x
2
) + F
1,3
(x
1
, x
3
) + F
2,3
(x
2
, x
3
) + H
1
(x
1
) + H
2
(x
2
) + H
3
(x
3
),
then the function Q : [0, 1]
3
[0, 1] dened by
Q(x) =
_
E(x) + (x) if x R
C(x) if x [0, 1]
3
\R
satises that Q is continuous. However Q is not in general a 3-copula.
We know that the function dened above is a modular function by Lemma 1.2, and by Lemma
1.3 the rst row in the denition of Q is also 3-increasing. Besides, since C and D are 3-copulas
we also know that E(x) is continuous. The idea now is try to nd an appropriate continuous
function , such that it makes Q continuous.
First we observe that (R) is given by the union of {x R| there exists i {1, 2, 3} such that x
i
=
0} with {x R| there exists i {1, 2, 3} such that x
i
= v
i
}. Since C is continuous we only have
to nd such it makes coincide the rst row with the second row of Q on the upper faces of R,
that is, on
{x R| there exists i {1, 2, 3} such that x
i
= v
i
}.
In this case we want to nd adequate functions in the denition of such that the following
three conditions hold
Q(v
1
, x
2
, x
3
) = C(v
1
, x
2
, x
3
) for every x
2
, x
3
[0, v
2
] [0, v
3
],
Q(x
1
, v
2
, x
3
) = C(x
1
, v
2
, x
3
) for every x
1
, x
3
[0, v
1
] [0, v
3
],
and
Q(x
1
, x
2
, v
3
) = C(x
1
, x
2
, v
3
) for every x
1
, x
2
[0, v
1
] [0, v
2
].
Multivariate Copulas in n-Boxes - Part II 11
We dene for x
1
, x
2
[0, v
1
] [0, v
2
]
F
1,2
(x
1
, x
2
) = D
_
V
C
(R
x
1
)

,
V
C
(R
x
2
)

, 1
_
+ C(x
1
, x
2
, v
3
),
for x
1
, x
3
[0, v
1
] [0, v
3
]
F
1,3
(x
1
, x
3
) = D
_
V
C
(R
x
1
)

, 1,
V
C
(R
x
3
)

_
+ C(x
1
, v
2
, x
3
),
and for x
2
, x
3
[0, v
2
] [0, v
3
]
F
2,3
(x
2
, x
3
) = D
_
1,
V
C
(R
x
2
)

,
V
C
(R
x
3
)

_
+ C(v
1
, x
2
, x
3
).
We also dene for x
1
[0, v
1
], x
2
[0, v
2
] and x
3
[0, v
3
]
H
1
(x
1
) = V
C
(R
x
1
) C(x
1
, v
2
, v
3
), H
2
(x
2
) = V
C
(R
x
2
) C(v
1
, x
2
, v
3
)
and
H
3
(x
3
) = V
C
(R
x
3
) C(v
1
, v
2
, x
3
).
Then if we take v
1
, x
2
, x
3
R, we observe that R
v
1
= R, = V
C
(R) = C(v
1
, v
2
, v
3
), and using
the frontier properties of D, we have that
Q(v
1
, x
2
, x
3
) = D
_
1,
V
C
(R
x
2
)

,
V
C
(R
x
3
)

_
D
_
1,
V
C
(R
x
2
)

, 1
_
+ C(v
1
, x
2
, v
3
)
D
_
1, 1,
V
C
(R
x
3
)

_
+ C(v
1
, v
2
, x
3
)
D
_
1,
V
C
(R
x
2
)

,
V
C
(R
x
3
)

_
+ C(v
1
, x
2
, x
3
)
+V
C
(R
v
1
) C(v
1
, v
2
, v
3
) + V
C
(R
x
2
) C(v
1
, x
2
, v
3
) + V
C
(R
x
3
) C(v
1
, v
2
, x
3
)
= V
C
(R
x
2
) V
C
(R
x
3
) + C(v
1
, x
2
, x
3
) + + V
C
(R
x
2
) + V
C
(R
x
3
)
= C(v
1
, x
2
, x
3
).
Therefore, the rst condition above holds. Analogously it is easy to see that the other two
conditions also hold with the same denition of which is clearly continuous. Hence, Q is
continuous.
Now, if x R is such that x
i
= 0 for some i {1, 2, 3}, then it is clear that V
C
(R
x
i
) = 0 and
since D and C are 3-copulas then E(x) = 0 = C(x). However, if we take x = 0, x
2
, x
3
with
x
2
(0, v
2
] and x
3
(0, v
3
], then E(x) = 0, F
1,2
(0, x
2
) = 0, F
1,3
(0, x
3
) = 0, F
2,3
(x
2
, x
3
) =
D(1, V
C
(R
x
2
)/, V
C
(R
x
3
)/) + C(v
1
, x
2
, x
3
), H
1
(0) = 0, H
2
(x
2
) = V
C
(R
x
2
) C(v
1
, x
2
, v
3
) =
C(v
1
, x
2
, v
3
)C(v
1
, x
2
, v
3
) = 0 and H
3
(x
3
) = V
C
(R
x
3
)C(v
1
, v
2
, x
3
) = C(v
1
, v
2
, x
3
)C(v
1
, v
2
, x
3
) =
0. So, (x) = D(1, C(v
1
, x
2
, v
3
)/, C(v
1
, v
2
, x
3
)/)+C(v
1
, x
2
, x
3
), which in general is not zero.
Therefore, Q is not a 3-copula even if it continuous.
12 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
This last example shows that it is not easy to nd a modular function , such that the function
Q satises being a 3-copula. The problem is to nd a modular function such that the rst and
second rows in the denition of Q coincide on the lower and upper faces of R. It seems that we
can make them coincide if we take only the lower faces or only the upper faces, but not both at
the same time, in order to get a 3-copula. Maybe if we dene a dierent function we could
obtain a 3-copula, but this still remains an open problem.
3.1. Patchwork Construction Method for Dimensions Larger Than or Equal to Three
We rst observe that the construction of new copulas in Theorem 3.4 is restricted to n-boxes
of the form R = [u, 1] [0, 1]
n
where u = u
1
, . . . , u
n
and 1 = 1, . . . , 1. The question here
is, how to do a construction of a new copula when we want to modify a base n-copula C on an
arbitrary n-box R = [u, v] [0, 1]
n
by another rescaled n-copula D?
In order to answer this question we will propose a new methodology which is based on Theorem
3.4.
Let us assume that R = [u, v] [0, 1]
n
is a non trivial n-box such that for every i {1, . . . , n},
0 < u
i
< v
i
< 1. Let Vert(R) be the set of vertices of R. We will rst establish an order in this set
of vertices using the same idea given in Proposition 2.10 in the rst part. Let c Vert(R), dene
a bijection f : Vert(R) {1, 2}
n
given by f (c) = l
1
, l
2
, . . . , l
n
{1, 2}
n
, where l
i
= 1 if c
i
= u
i
and l
i
= 2 if c
i
= v
i
. Dene Q
f (c)
= {k {1, . . . , n} | f (c)
k
= 2}, where f (c)
k
is the k
th
coordinate
of f (c). Now we dene the composition : Vert(R) {1, 2, . . . , 2
n
} given by
( f (c)) = 1 +

kQ
f (c)
2
nk
for every c Vert(R).
Of course, Q
f (c)
= if and only if c = u
1
, u
2
, . . . , u
n
and in this case ( f (u
1
, u
2
, . . . , u
n
)) = 1.
Also observe that if c = v
1
, v
2
, . . . , v
n
, then ( f (v
1
, v
2
, . . . , v
n
)) = 1+
_
n
k=1
2
nk
= 1+
_
n1
j=0
2
j
=
2
n
. It is easy to see that is a bijection which establishes an order among the vertices of R, in
fact, this order gives the number one to the lowest vertex and the number 2
n
to the highest
vertex.
Let C be an n-copula which we will call base copula, let D be an n-copula which we will call
modifying copula, and let R = [u, v] [0, 1]
n
be a non trivial n-box. Then inductively dene
Let C
1
= (C
_
u
D) where 1 = (u).
Given C
k
for 1 k < 2
n
, let c Vert(R) such that (c) = k+1 and dene C
k+1
= (C
k
_
c
C).
The n-copula C
2
n is our target copula.
Observe that if we follow this construction, then in the rst step in C
1
we introduce the rescaled
version of the n-copula D on [u, 1], and in the remaining steps we keep unaltered C
1
at least in
the semi open n-box [u, v) =
n
i=1
[u
i
, v
i
).
Example 3.10. Let us assume that n = 3 and that we want to construct a 3-copula C which has
a desired behavior in each of the eight vertices of I
3
:= [0, 1]
3
. We will use the 3-box given by
Multivariate Copulas in n-Boxes - Part II 13
R = [0, 1/2]
3
as an auxiliary tool. We rst establish a bijection h : Vert(I
3
) Vert(R), among
the vertices of I
3
and the vertices of R. If c Vert(I
3
) then the coordinates of h(c) are given by
h(c)
i
=
_
0 if c
i
= 0
1/2 if c
i
= 1,
for i = 1, 2, 3. Using the order established at the beginning of this subsection we have Table 1
Table 1.- Order of the vertices of I
3
c f (c) Q
f (c)
( f (c))
0, 0, 0 1, 1, 1 1
0, 0, 1 1, 1, 2 {3} 1 + 2
33
= 2
0, 1, 0 1, 2, 1 {2} 1 + 2
32
= 3
0, 1, 1 1, 2, 2 {2, 3} 1 + 2
32
+ 2
33
= 4
1, 0, 0 2, 1, 1 {1} 1 + 2
31
= 5
1, 0, 1 2, 1, 2 {1, 3} 1 + 2
31
+ 2
33
= 6
1, 1, 0 2, 2, 1 {1, 2} 1 + 2
31
+ 2
32
= 7
1, 1, 1 2, 2, 2 {1, 2, 3} 1 + 2
31
+ 2
32
+ 2
33
= 8
Of course, the order of the vertices of R is the same, that is,
( f (h(c))) = ( f (c)) for every c Vert(I
3
).
Assume that we have selected eight 3-copulas {C
j
}
8
j=1
, such that if c Vert(I
3
) and it saties that
( f (c)) = j, then C
j
has the desired behavior near the vertex c, for every j {1, . . . , 8}. Now
we proceed with the construction of a 3-copula which satises the desired properties using the
3-box R:
Let j = 1, R
1
= [0, 1]
3
and dene D
1
= C
1
on R
1
, then D
1
is exactly C
1
near c = 0
Vert(R), 0 Vert(I
3
) and ( f (0)) = 1.
Let j = 2, R
2
= [0, 1] [0, 1] [1/2, 1], u
2
= 0, 0, 1/2 Vert(R), then ( f (h(u
2
))) =
2 = j. Dene D
2
= (D
1
_
u
2 C
2
), then from equation (25), D
2
behaves like C
1
near 0 and
like C
2
near 0, 0, 1 = c, where ( f (c)) = 2 = j.
Inductively, given D
j1
for 3 j 8, let R
j
= [c
j
1
, 1] [c
j
2
, 1] [c
j
3
, 1], where u
j
=
c
j
1
, c
j
2
, c
j
3
Vert(R) is such that ( f (h(u
j
))) = j. Dene D
j
= (D
j1
_
u
j C
j
), then from
equation (25), D
j
behaves like C
k
near c
k
Vert(I
3
) for every k = {1, 2, . . . , j}.
Then if we dene C := D
8
, C has the desired properties.
The last example has of course generalizations to larger dimensions, and it is of great importance
because it allows to model tail dependence, see for example Joe [12] and Nelsen [17]. In fact, in
Finance and Risk Theory one of the biggest problems is to model tail dependence for economic
variables using copulas, see for example Embrechts et al [9], Cherubini et al [3], McNeil et al
[15], Malevergne et al [14], Zhang [20], just to mention some references. The use of copulas for
modeling dependence has been also used in other areas such as ecology, hidrology, medicine,
14 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
etc., see for example Dorey and Joubert [4], Erdely and Daz-Vieira [10] or Salvadori et al [18].
The last example provides a method for modeling the tail dependence in dimensions larger than
or equal to three.
Of course this methodology also allows us to study the multidimensional versions of the hori-
zontal and vertical sections of 2-copulas and the construction of 2-copulas with given horizontal
or vertical sections, see for example see Nelsen [17], Durante et al [5] or Klement et al [13]. For
n 3, we can analyze the structure of the n 1 dimensional faces of an n-copula, of the form
C
i
(x) = {C(x
1
, . . . , x
ii
, a
i
, x
i+1
, . . . , x
n
) | x
j
[0, 1] for every j {1, . . . , n}\{i}},
where a
i
[0, 1] is xed and i {1, . . . , n}; which are clearly equivalent to the horizontal and
vertical sections of a 2-copula.
Example 3.11. Let us assume that n = 3 and R = [0, 1/2]
3
in Example 3.10, and let us select
the rst six copulas as C
1
= = C
6
= , C
7
= M and C
8
is a Gumbel-Hougaard copula with
parameter = 2, that is,
C
8
(x) = exp
_

_
(ln x
1
)
2
+ (ln x
2
)
2
+ (ln x
3
)
2
_
1/2
_
.
If we proceed with the order given in Table 1 and the steps of Example 3.10, we get that D
1
=
D
2
= = D
6
= by Lemma 3.7, so, we can start at the seventh step.
Let u
7
= 1/2, 1/2, 0 Vert(R), R
7
= [1/2, 1] [1/2, 1] [0, 1] and dene D
7
= (D
6
_
u
7 C
7
),
after some calculations we have that
7
= V

(R
7
) = 1/4, V

(R
7
x
1
) = x
1
/2 1/4, x
2
/2
1/4, V

(R
7
x
3
) = x
3
/4, V

([u
7
, x]) = x
1
x
2
x
3
(1/2)(x
2
x
3
+ x
1
x
3
) + x
3
/4 and
D
7
(x) =
_

_
1
4
min{2x
1
1, 2x
2
1, x
3
}
+
1
2
(x
2
x
3
+ x
1
x
3
)
1
4
x
3
if x R
7
x
1
x
2
x
3
otherwise.
For the last step we have u
8
= 1/2, 1/2, 1/2 Vert(R), R
8
= [1/2, 1] [1/2, 1] [1/2, 1].
Dene D
8
= (D
7
_
u
8 C
8
), then we get
8
= V
D
7
(R
8
) = 1/8, V
D
7
(R
8
x
1
) = x
1
/2 1/4 C
7
(2x
1

1, 1, 1/2)/4, V
D
7
(R
8
x
2
) = x
2
/21/4C
7
(1, 2x
2
1, 1/2)/4, V
C
7
(R
8
x
3
) = x
3
/41/8, V
D
7
([u
8
, x]) =
D
7
(x
1
, x
2
, x
3
) C
7
(2x
1
1, 2x
2
1, 1/2)/4 + x
3
/4 (x
2
x
3
+ x
1
x
3
)/2 and
D
8
(x) =
_

_
1
8
C
8
_
4x
1
2 2 min{2x
1
1,
1
2
}, 4x
2
2 2 min{2x
2
1,
1
2
}, 2x
3
1
_
+
1
4
min{2x
1
1, 2x
2
1,
1
2
}
1
4
x
3
+
1
2
(x
2
x
3
+ x
1
x
3
) if x R
8
1
4
min{2x
1
1, 2x
2
1, x
3
}
+
1
2
(x
2
x
3
+ x
1
x
3
)
1
4
x
3
if x [1/2, 1] [1/2, 1] [0, 1/2]
x
1
x
2
x
3
otherwise.
We can see that this last copula behaves similar to the copula M in the vertex 1, 1, 0 and similar
to a Gumbel-Hougaard copula near the vertex 1, 1, 1.
4. AN ALTERNATIVE PATCHWORK USING GLUING COPULAS
In Siburg and Stoimenov [19] a new methodology of constructing n-copulas is proposed. The
main idea is to glue two rescaled copulas on adjacent n-boxes of [0, 1]
n
whose union is [0, 1]
n
,
see Denition 2.9 in the rst part. Their main result is the following:
Multivariate Copulas in n-Boxes - Part II 15
Theorem C Let n 2 and let C
1
, C
2
be two n-copulas. Let 0 1 and dene R
l
i,
= [0, 1]
[0, 1] [0, ] [0, 1] [0, 1], where the interval [0, ] is located on the i
th
coordinate,
for some i {1, 2, . . . , n}, similarly dene R
u
i,
= [0, 1] [0, 1] [, 1] [0, 1] [0, 1].
Dene for every x = (x
1
, . . . , x
i
, . . . , x
n
) [0, 1]
n
_
C
1

x
i
=
C
2
_
(x)=
_

_
C
1
_
x
1
, . . . ,
x
i

, . . . , x
n
_
if x R
l
i,
(1 )C
2
_
x
1
, . . . ,
x
i

1
, . . . , x
n
_
+ C
1
(x
1
, . . . , 1, . . . , x
n
) if x R
u
i,
.
(28)
Then C
1

x
i
=
C
2
is an n-copula.
Proof: Observe that the rst row in equation (28) it is simply C
1
rescaled on R
l
i,
and the second
row is C
2
rescaled on R
u
i,
plus the value of the rescaled C
1
in x
1
, . . . , 1, . . . x
n
. It is clear that
C
1

x
i
=
C
2
satises the frontier conditions of an n-copula. Using that R
l
i,
and R
u
i,
are adjacent
n-boxes with common face R
i,
= {x [0, 1]
n
| x
i
= }, and observing that on R
i,
both rows
of equation (28) coincide, by Lemma 2.11 in the rst part, it is clear that C
1

x
i
=
C
2
is an
n-increasing function. Therefore, C
1

x
i
=
C
2
is an n-copula.
Of course, the binary operation
x
i
=
, where i {1, . . . , n} and [0, 1], is dened on the family
C
n
of all n-copulas. Of course,
x
i
=
is not a commutative operation. In Proposition 3.1 of Siburg
and Stoimenov, it is shown that if is the n-product copula, i {1, . . . , n} and [0, 1]. Then
C
1

x
i
=
C
2
= if and only if C
1
= C
2
= . (29)
This result follows directly from equation (28).
Remark 4.1. For dimension n = 2 the constructions proposed in Theorem A in Section 1 and
the one given in Theorem C are quite dierent. Let C = M, J = {1}, R
1
= [0, 1/2, 1] [0, 1]
and C
1
= in Theorem A. Then
1
= V
M
(R
1
) = 1/2, R
1,x
= [0, x] [0, 1], R
1,y
= [0, 1/2], [0, y]
for every x [0, 1/2] and for every y [0, 1]. So, V
M
(R
1,x
) = x and V
M
(R
1,y
) = min{1/2, y} and

M
1
(x, y) = h
M
0
(x) + v
M
0
(y) h
M
0
(0) = 0. Therefore, using equation (59) in the rst part

C(x, y) =
_
1
2
(2x, 2 min{1/2, y}) if x, y [0, 1/2] [0, 1]
min{x, y} if x, y [1/2, 1] [0, 1].
=
_
2x min{1/2, y} if x, y [0, 1/2] [0, 1]
min{x, y} if x, y [1/2, 1] [0, 1].
On the other hand, if we let C
1
= , C
2
= M and = 1/2 in Theorem C, then using equation
(28), we get that
_

x=1/2
M
_
(x, y) =
_
1
2
(2x, y) if x, y [0, 1/2] [0, 1]
1
2
min{2x 1, y} +
1
2
min{1, y} if x, y [1/2, 1] [0, 1].
=
_
xy if x, y [0, 1/2] [0, 1]
1
2
min{2x + y 1, 2y} if x, y [1/2, 1] [0, 1].
Of course,

C
_

x=1/2
M
_
.
16 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
Example 4.2. We will see that even for n = 3 the patchwork construction given in Theorem 3.1
is also dierent from the gluing proposal in Theorem C. In the case where is the base copula
they may coincide, but in general they are dierent. Let us glue with some 3-copula C
1
in
x
1
= 1/2, using equation (28) we get
_

x
1
=1/2
C
1
_
(x)=
_

_
1/2
_
x
1
1/2
, x
2
, x
3
_
if x [0, 1/2] [0, 1] [0, 1]
(1 1/2)C
1
_
x
1
1/2
11/2
, x
2
, x
3
_
if x [1/2, 1] [0, 1] [0, 1]
+1/2(1, x
2
, x
3
)
=
_

_
x
1
x
2
x
3
if x [0, 1/2] [0, 1] [0, 1]
1
2
C
1
(2x
1
1, x
2
, x
3
) if x [1/2, 1] [0, 1] [0, 1]
+
1
2
x
2
x
3
(30)
Now, using Remark 3.3, let us patch the 3-copula C
1
in R = [1/2, 1] [0, 1] [0, 1] with the
3-copula . In this case, u =1/2, 0, 0, = V

(R) = 1/2, V

(R
x
1
) = x
1

1
2
, V

(R
x
2
) =
x
2
2
, V

(R
x
3
) =
x
3
3
, V

([0, x]) = (x
1
, x
2
, x
3
) and V

([u, x]) = V

([1/2, x
1
] [0, x
2
] [0, x
3
]) =
(x
1
1/2)x
2
x
3
, for every x
1
[1/2, 1] and for every x
2
, x
3
[0, 1]. Then using equation (22) in
Theorem 3.1 we have that
_

_
u
C
1
_

_
(x)=
_

_
(x
1
, x
2
, x
3
) if x [0, 1/2] [0, 1] [0, 1]
1/2C
1
_
x
1
1/2
1/2
,
x
2
/2
1/2
,
x
3
/2
1/2
_
+ (x
1
, x
2
, x
3
) if x R = [1/2, 1] [0, 1] [0, 1]
V

([1/2, x
1
] [0, x
2
] [0, x
3
])
=
_

_
x
1
x
2
x
3
if x [0, 1/2] [0, 1] [0, 1]
1
2
C
1
(2x
1
1, x
2
, x
3
) if x R = [1/2, 1] [0, 1] [0, 1]
+
1
2
x
2
x
3
(31)
Then from equations (30) and (31), clearly, (
x
1
=1/2
C
1
) = (
_
u
C
1
). But if we reverse the
order and we glue some copula C
1
, say C
1
=
M+
2
, with in x = 1/2 we get from equation (28)
that
_
C
1

x
1
=1/2

_
(x)=
_

_
1/2C
1
_
x
1
1/2
, x
2
, x
3
_
if x [0, 1/2] [0, 1] [0, 1]
(1 1/2)
_
x
1
1/2
11/2
, x
2
, x
3
_
if x [1/2, 1] [0, 1] [0, 1]
+1/2C
1
(1, x
2
, x
3
)
=
_

_
1
4
min{2x
1
, x
2
, x
3
} +
1
2
x
1
x
2
x
3
if x [0, 1/2] [0, 1] [0, 1]
x
1
x
2
x
3

x
2
x
3
4
+
min{x
2
,x
3
}
4
if x [1/2, 1] [0, 1] [0, 1],
(32)
and using Theorem 3.1 and Remark 3.3 again, if we patch the same copula C
1
in R = [1/2, 1]
[0, 1] [0, 1] with we have that u = 1/2, 0, 0, = V
C
1
(R) = C
1
(1, 1, 1) C
1
(1/2, 1, 1) = 1/2,
V
C
1
(R
x
1
) = C
1
(x
1
, 1, 1)C
1
(1/2, 1, 1) = x
1
1/2, V
C
1
(R
x
2
) = C
1
(1, x
2
, 1)C
1
(1/2, x
2
, 1) = (3x
2

min{1, 2x
2
})/4, V
C
1
(R
x
3
) = C
1
(1, 1, x
3
)C
1
(1/2, 1, x
3
) = (3x
3
min{1, 2x
3
})/4, V
C
1
([0, x]\[u, x])
= C
1
(x
1
, x
2
, x
3
) (C
1
(x
1
, x
2
, x
3
) C
1
(1/2, x
2
, x
3
)) = C
1
(1/2, x
2
, x
3
), for every x
1
[1/2, 1] and
for every x
2
, x
3
[0, 1]. So, by equation (22) in Theorem 3.1 we have that
Multivariate Copulas in n-Boxes - Part II 17
_

_
C
1
_
u

_
(x)=
_

_
C
1
(x
1
, x
2
, x
3
) if x [0, 1/2] [0, 1] [0, 1]
1
2

_
x
1
1/2
1/2
,
(3x
2
min{1,2x
2
})/4
1/2
,
(3x
3
min{1,2x
3
})/4
1/2
_
elsewhere
+C
1
(1/2, x
2
x
3
)
=
_

_
1
2
min{x
1
, x
2
, x
3
} +
1
2
x
1
x
2
x
3
if x [0, 1/2] [0, 1] [0, 1]
1
4
_
x
1

1
2
_
(3x
2
min{1, 2x
2
}) (3x
3
min{1, 2x
3
}) if x [1/2, 1] [0, 1] [0, 1]
+
1
2
min{1/2, x
2
, x
3
} +
1
4
x
2
x
3
(33)
If x = 1/4, 1/4, 1/4 then from equations (32) and (33), (C
1

x
1
=1/2
)(x) = 9/128, but
(C
1
_
u
)(x) = 9/64. So, gluing and patching copulas do not always yield the same result.
This example also shows that the operators gluing and patching are not commutative.
4.1. Proposed Methodology Using Gluing Copulas
Let R =
n
i=1
[u
i
, v
i
] [0, 1]
n
be a non trivial n-box, we want to dene an n-copula C, such that
on R it behaves as a rescaled version of another n-copula C
1
. We propose to take as a background
copula (x
1
, . . . , x
n
) = x
1
x
2
x
n
the n-product copula and to give a rescaled version of C
1
on
R. In order to do so, we can use the gluing method as follows:
Since R is non trivial then for every i {1, . . . , n}, 0 u
i
< v
i
1, then there exists
a
i
[0, 1) such that a
i
v
i
= u
i
.
Dene T
1
=
x
1
=a
1
C
1
, then T
1
is an n-copula such that for x
1
[0, a
1
] behaves like ,
and for x
1
[a
1
, 1] behaves like a rescaled version of C
1
.
Inductively, let T
k
=
x
k
=a
k
T
k1
for every k {2, . . . , n}, then T
k
is an n-copula such
that for x
k
[0, a
k
] behaves like .
Dene inductively T
n+j
= T
n+j1

x
j
=v
j
for every j {1, 2, . . . , n}.
Let C = T
2n
, then C has the desired properties.
Example 4.3. We will see that this proposed methodology does not yield to the ordinal sum. In
order to do so we will consider the copula . Let us dene
T
1
(x
1
, x
2
) =
_
M
x
1
=1/2

_
(x
1
, x
2
) =
_
min
_
x
1
,
1
2
x
2
_
if x
1
, x
2
[0, 1/2] [0, 1]
x
1
x
2
if x
1
, x
2
[1/2, 1] [0, 1].
In the next step we glue M with the copula T
1
at x
2
= 1/2, that is, we dene
T
2
(x
1
, x
2
) =
_
M
x
2
=1/2
T
1
_
(x
1
, x
2
) =
_

_
min
_
1
2
x
1
, x
2
_
if x
1
, x
2
[0, 1] [0, 1/2]
min
_
x
1
,
2x
1
+2x
2
1
4
_
if x
1
, x
2
[0, 1/2] [1/2, 1]
x
1
x 2 if x
1
, x
2
[1/2, 1] [1/2, 1].
The resulting copula T
2
behaves like on [1/2, 1]
2
, on [0, 1] [0, 1/2] has support the line
joining 0, 0 and 1, 1/2 with mass 1/2 and on [0, 1/2, 1] [1/2, 1] has support the line joining
0, 1/2 and 1/2, 1 with mass 1/4, see Figure 2 below.
18 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
Figure 2.- Support of the gluing copula T
2
1 0 1/2 x
1
1/2
1
x
2
T
E

.
.. . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
On the other hand the ordinal sum of on [1/2, 1]
2
is given by
C(x
1
, x
2
) =
_
2x
1
x
2
x
1
x
2
+ 1 if x
1
, x
2
[1, 2, 1]
2
min{x
1
, x
2
} otherwise.
Of course, the gluing copula T
2
is dierent from C the ordinal sum.
5. FINAL REMARKS
In this paper we give a new multivariate patchwork construction of n-copulas using a base n-
copula C and an n-box R [0, 1]
n
and a function D : R [0, 1] which is easily proved to be
n-increasing on R, then we try to construct a modular function E on R such that C = D + E on
the frontier of R, (R). Of course, this procedure can be carried out in many dierent ways, and
in this paper we explore just a few possibilities. However, other proposals may lead to dierent
solutions that the one given in this paper.
The construction described in Section three for arbitrary n-boxes is sequential and it is more
complicated than the one given in Durante et al in [7]. The reason for this complication is related
to the dimension n, and this problem is very common with multidimensional generalizations.
Of course we can construct new n-copulas composing our methodology with the gluing copulas
of Siburg and Stoimenov [19], because as we noticed in our examples both constructions are not
necessarily equivalent.
The study of tail behavior and the multivariate generalization of horizontal and vertical sections
of dimension two can also be studied using our methodology, but it goes beyond the goal of this
article.
(Received ????)
REF ERENCES
[1] J. Acz el and J. Dhombres: Functional Equations in Several Variables. Cambridge University Press,
Cambridge, UK 1989.
[2] B. De Baets and H. De Meyer: Orthogonal grid constructions of copulas. IEEE Trans. Fuzzy Syst.,
15, No. 6, [2007], 10531062.
Multivariate Copulas in n-Boxes - Part II 19
[3] U. Cherubini, E. Luciano and W. Vecchiato: Copula Methods in Finance. John Wiley & Sons,
Chichester, 2004.
[4] M. Dorey and P. Joubert: Modelling copulas: An overview. The Staple Inn Actuarial Society.
London, 2005.
[5] F. Durante, A. Koles arov a, R. Mesiar and C. Sempi: Copulas with given diagonal sections: novel
constructions and applications. Int. J. Uncertain. Fuzziness Knowledge-Based Syst., 15, [2007],
397410.
[6] F. Durante, S. Saminger-Platz and P. Sarkoci: On representations of 2-increasing binary aggregation
functions. Information Sciences, 178, [2008], 45344541.
[7] F. Durante, S. Saminger-Platz and P. Sarkoci: Rectangular patchwork for bivariate copulas and tail
dependence. Comm. Statist. Theory and Methods, 38, [2009], 25152527.
[8] F. Durante and J. Fern andez-S anchez: Multivariate shues and approximation of copulas. Statist.
Probab. Lett., 80, [2010], 18271834.
[9] P. Embrechts, A. H oing and A. Juri: Using copulae to bound the value-at-risk for functions of
dependent risks. Finance Stochat., 7, [2003], 145167.
[10] A. Erdely and M. Daz-Vieira: Nonparametric and semiparametric bivariate modeling of petrophys-
ical porosity-permeability dependence from well log data. In P. Jaworski, F. Durante, W. K. H ardle
and T. Rychlik, eds.: Copula Theory and Its Applications, Lect. Notes in Statist., Springer, [2010],
267278.
[11] J.M. Gonz alez-Barrios and M.M. Hern andez-Cedillo: Construction of Multivariate Copulas in n-
Boxes - Part I. Preprint [2012].
[12] H. Joe: Multivariate Models and Dependence Concepts. Chapman & Hall, London, 1997.
[13] E.P. Klement, A. Koles arov a, R. Mesiar and C. Sempi: Copulas constructed from the horizontal
section. Comm. Statist. Theor. Meth., 36, [2007], 29012911.
[14] Y. Malevergne and D. Sornette: Extreme Financial Risk. From Dependence to Risk Management.
Springer-Verlag, Berlin, 2006.
[15] A.J. McNeil, R. Frey and P. Embrechts: Quantitative Risk Management. Princeton University Press,
Princeton, 2005.
[16] R. Mesiar and C. Sempi: Ordinal sums and idempotent of copulas. Aequat. Math., 79, No. 1-2,
[2010], 3952.
[17] R.B. Nelsen: An Introduction to Copulas. 2nd ed., Springer, New York, 2006.
[18] G. Salvadori, C. De Michele, N.T. Kottegoda and R. Rosso: Extremes in Nature. An Approach Using
Copulas. Series: Water Science and Technology Library, Vol. 56, Springer, Amsterdam, 2007
[19] K.F. Siburg and P.A. Stoimenov: Gluing copulas. Comm. Statist. Theory and Methods, 37, [2008],
31243134.
[20] M.H. Zhang: Modelling total tail dependence along diagonals. Insur. Math. Econ., 42, [2008],
7380.
Jose M. Gonzalez-Barrios, Universidad Nacional Autonoma de Mexico, Instituto de Investiga-
ciones en Matem aticas Aplicadas y en Sistemas, Departamento de Probabilidad y Estadstica,
Circuito Escolar s/n, Ciudad Universitaria, C.P. 04510, Mexico D.F., Mexico
e-mail: gonzaba@sigma.iimas.unam.mx
20 J.M. GONZ

ALEZ-BARRIOS AND M.M. HERN

ANDEZ-CEDILLO
Mara M. Hern andez-Cedillo, Universidad Nacional Aut onoma de Mexico, Instituto de Investi-
gaciones en Matem aticas Aplicadas y en Sistemas, Departamento de Probabilidad y Estadstica,
Circuito Escolar s/n, Ciudad Universitaria, C.P. 04510, Mexico D.F., Mexico
e-mail: magdaz@gmail.com