Advances in Social Science, Education and Humanities Research, volume 123

2nd International Conference on Education, Sports, Arts and Management Engineering (ICESAME 2017)

Copula function theory and classification

Yueguang Hu 1,a and Zhigang Zhang1,b
Department of Information and Computing Science, University of Science and Technology Beijing
30 Xueyuan Road,Beijing,10083,China
a
[email protected], [email protected]

Keywords: copula, the joint distribution function, independence

Abstract: "copula" comes from Latin "copulare", meaning "together", and connecting to the
one-dimensional marginal distribution. Copula function is the multivariate distribution function on
[0,1], is also a usual method of measuring multivariate extreme value theory dependent function.
According to the statistics, the copula theory can be traced back to the multivariate non-Gauss
distribution in twentieth Century. Later In 1959, Professor Sklar has proposed the copula function,
he believes that a joint distribution is too complex, and it can be decomposed into k marginal
distribution, also, The copula function is not affected by the edge distribution，The copula function
describes the correlation structure between variables and studies the correlation between variables,
if the edge function is continuous, then the copula function is the only sure.

Introduction
With the continuous integration of the global economy, the risk and destruction of multiple factors
make people have to start to study the risk and to avoid it. The appearance of Copula function has
aroused people's attention. The application of Copula is also very extensive, in which the financial
risk measurement is the most. In fact, it is not surprising. With the deepening of economic
globalization, the capital flows in the world are more frequent, the relationship among financial
markets becomes increasingly complex, but also to strengthen the linkage of financial market. The
market interdependence greatly enhance the risk and destruction of financial market. Therefore,
financial risk management is particularly important. The copula function has played a very
important role in financial risk management.

The definition and theorem of Copula function

Definition of the N-copula function C:
1. C= 𝐼𝐼 𝑁𝑁 =[0，1]𝑁𝑁 ；
2. C for each variable is monotonically increasing, that is, a monotonically increasing function of
a single variable.
3. to meet the edge consistency，that is to say, 𝐶𝐶𝑛𝑛 (𝑢𝑢𝑛𝑛 ) = 𝐶𝐶(0, … 0, 𝑢𝑢𝑛𝑛 , 0 … 0) = 0；𝐶𝐶𝑛𝑛 (𝑢𝑢𝑛𝑛 ) =
𝐶𝐶(1, … 1, 𝑢𝑢𝑛𝑛 , 1 … 1) = 𝑢𝑢𝑛𝑛 ,𝑢𝑢𝑛𝑛 𝜖𝜖[0,1], 𝑛𝑛𝑛𝑛[1, 𝑁𝑁]
The function C satisfying the above properties is the copula function.
Given the definition, the corresponding should also be given Sklar's theorem.
(Sklar's theorem) set 𝑋𝑋1， … … 𝑋𝑋𝑛𝑛 are n random variables, K(𝑋𝑋1， … … ，𝑋𝑋𝑛𝑛 )are their joint
distribution functions，and if 𝐹𝐹1 (𝑥𝑥1 ), … … 𝐹𝐹𝑛𝑛 (𝑥𝑥𝑛𝑛 ) is its edge distribution function, there is a copula
function C(𝑢𝑢1 ，…𝑢𝑢𝑛𝑛 ), so that for any x, y, K(𝑥𝑥1， … … ，𝑥𝑥𝑛𝑛 )=C(𝐹𝐹1 (𝑥𝑥1 ), … … 𝐹𝐹𝑛𝑛 (𝑥𝑥𝑛𝑛 ))
If F(x), k(x) is continuous, the copula function C is unique.

The properties of copula function

Copula function has very perfect properties, a variety of multivariate distribution can be constructed
by copula function. We use copula theory to build financial models can study random variables and

Copyright © 2017, the Authors. Published by Atlantis Press.

This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). 1997
 Advances in Social Science, Education and Humanities Research, volume 123

their related structures independently, because the corollaries and relating measurements of the
copula function will be not changed for strictly monotone increasing transformation. In order to
facilitate the understanding. We use two variables Copula function C (U, V) as an example to
illustrate the nature of the copula function.
(1) set 𝑋𝑋1 , 𝑋𝑋2 and 𝑋𝑋3 ,𝑋𝑋4 are two independent random variables, 𝐹𝐹1 (𝑥𝑥), 𝐹𝐹2 (𝑥𝑥), 𝐹𝐹3 (𝑥𝑥), 𝐹𝐹4 (𝑥𝑥)
respectively, corresponding to their distribution functions. C(𝐹𝐹1 (𝑥𝑥), 𝐹𝐹2 (𝑥𝑥))C(𝐹𝐹3 (𝑥𝑥), 𝐹𝐹4 (𝑥𝑥))are still
their joint distribution functions.
It is easy to prove that our C(𝐹𝐹1 (𝑥𝑥), 𝐹𝐹2 (𝑥𝑥)) and C(𝐹𝐹3 (𝑥𝑥), 𝐹𝐹4 (𝑥𝑥)) are joint functions of 𝑋𝑋1 , 𝑋𝑋2
and 𝑋𝑋3 , 𝑋𝑋4 , Are all independent of each other, so the joint distribution multiplication is the joint
distribution function of 𝑋𝑋1 , 𝑋𝑋2 ，𝑋𝑋3 , 𝑋𝑋4
(2) For the binary copula function C(u, v), if u and v are independent, then C(u, v) = u · v;
that is, if the two variables are independent of each other, the copula function can be defined by the
edge function.
(3) C has two multiplicities, that is, the joint distribution increases with the increase of the
marginal distribution. Take the binary copula function as an example
∀u1 , u2 , v1 , v2 ∈ Iandu1 ≤ u2 , v1 ≤ v2 , thenc(u2 , v2 ) − c(u1 , v2 ) − c(u2 , v1 ) + c(u1 , v1 ) ≥ 0
(4) copula function has nonlinear transformation invariance. If the variables X and Y have
Copula function C, andh1 , h2 are incremental nonlinear continuous functions, then h1 (X) and
h1 (X) and also have the same Copula function C.
We all know that the commonly used correlation is only a linear correlation under the constant
change in a correlation index. If the correlation of the nonlinear function is involved, an erroneous
conclusion arises. But the copula function is not the same, it has a more extensive basis, than the
linear range is larger.
(5) Convex combination of Copula functions is also Copula function. 𝐶𝐶𝑖𝑖 (u,v) is a Copula
function, and its convex combination C(u, v) = ∑𝑛𝑛𝑖𝑖=1 𝜆𝜆𝑖𝑖 𝐶𝐶𝑖𝑖 (𝑢𝑢, 𝑣𝑣) for any n, when ∑𝑛𝑛𝑖𝑖=1 𝜆𝜆𝑖𝑖 = 1, 𝜆𝜆𝑖𝑖 ≥
0 时 C(u, v) is also a copula function.
(6) Monotonical increase: C(u, v) is incremented for variables u and v. If a marginal
distribution is constant and another marginal distribution increases, then the joint distribution will
be also increased.
(7) C(u, 0) = C(0, v) = 0, C(u, 1) = C(1, v) = v。Namely if the probability of any marginal
distribution is zero, the probability distribution of the corresponding joint occurrence is also zero; if
the probability of any marginal distribution is 1, then the joint distribution is determined by another
marginal distribution.
(8) （Frechet-Hosffding constraint） Assuming that C is a two – varieties copula function, for
each u ∈ [0,1], v ∈ [0,1]，W(u, v) ≤ C(u, v) ≤ M(u, v).
And，W(u, v) = max(u + v − 1,0) is the Frechet-Hosffding lower bound。
M(u, v) = min(u, v), is the Frechet-Hosffding upper bound。

Classification of Copula functions

The copula function classification have many ways. For example, according to the type of
distribution function, it can be classified into Elliptic copula and Archimedean copula are ;
According to the number of parameters, it can be classified into a single parameter copula, two
parameter copula and multi parameter copula. We still take two variables copula function as an
example to introduce some common forms of Copula function.
1、The product copula function:
Definition: satisfying the form of the copula function∏ 𝑢𝑢𝑢𝑢 = 𝑢𝑢. 𝑣𝑣 is product copula function.
The product copula function is the copula function of independent random variables, and it is
the simplest form of copula function.

1998
 Advances in Social Science, Education and Humanities Research, volume 123

Theorem: if U, V is continuous random variables, only when the random variable copula
function C = ∏ 𝑢𝑢. 𝑣𝑣 was satisfied, U, V is independent for each other.
2、Elliptic copula function:
Elliptic copula function includes normal copula function and t-copula function, they all belong
to the elliptic copula function. However, they have different properties.
normal copula function:The normal copula function has many advantages, such as random
sampling, few parameters, and the correlation structure is easy to calculate.
the form of Two variables normal copula function is:
φ −1 ( u ) φ −1 ( v ) 1  s 2 − 2 R12 st + t 2 
cGa (u, v ) ∫−∞ ∫ −∞ 1
exp  −
 2(1 − R122 ) 
 dsdt （1）
2p (1 − R12 )
2 2

And 𝑅𝑅12 is the correlation coefficient matrix Φ is the standard normal distribution function.
Normal distribution random variable 𝑋𝑋1 , … … , 𝑋𝑋𝑛𝑛 , its mean value is 𝜇𝜇1 , … … , 𝜇𝜇𝑛𝑛 And variance
𝑋𝑋 −𝜇𝜇
is 𝜎𝜎1 , … … , 𝜎𝜎𝑛𝑛 . The covariance matrix is R, then the random variable 𝑈𝑈𝑖𝑖 = Φ � 𝑖𝑖 𝑖𝑖 � , iϵI whose
𝜎𝜎𝑖𝑖
distribution function𝐶𝐶𝑅𝑅 (𝜇𝜇1 , … … , 𝜇𝜇𝑛𝑛 ) is normal copula function whose covariance matrix is R.
t-copula function: the form of Two variables t-copula function is:
𝑣𝑣+1
𝑣𝑣 −1 (𝑢𝑢) 𝑣𝑣 −1 (𝑣𝑣) 1 𝑠𝑠 2 −2𝜌𝜌𝜌𝜌𝜌𝜌+𝑡𝑡 2 2
𝐶𝐶𝑡𝑡 (𝑢𝑢, 𝑣𝑣, 𝛾𝛾, 𝜌𝜌) = ∫−∞ ∫−∞ 2𝜋𝜋(1−𝜌𝜌2 ) �1 + � 𝑑𝑑𝑑𝑑𝑑𝑑𝑡𝑡 （2）
𝑣𝑣(1−𝜌𝜌2 )
Normal distribution random variable 𝑋𝑋1 , … … , 𝑋𝑋𝑛𝑛 , its mean value is 0, the variance is 1, and the
covariance matrix is R. Y is 𝜒𝜒 2 random variable, with V degrees of freedom, is also independent of
√𝑣𝑣
（𝑋𝑋1 , … … , 𝑋𝑋𝑛𝑛 ）.The random variable 𝑈𝑈𝑖𝑖 = 𝑡𝑡𝑣𝑣 � 𝑋𝑋𝑖𝑖 ，i ∈ I� distribution function𝐶𝐶𝑣𝑣,𝑅𝑅 (𝜇𝜇1 , … … , 𝜇𝜇𝑛𝑛 ) is
√𝑌𝑌
the copula function, which is called the t-copula function with the V degree of freedom, the covariance
matrix R.
3、Archimedean Copula function:
The Archimedean Copula function has the characteristics of combination, easy to construct,
symmetry and easy calculation. Therefore, it is widely used in many fields.
Before defining the Archimedean Copula function, we first introduce the generating element of
the copula function.
Definition: the quasi inverse function: for continuous strictly monotone function: φ：[0,1] →
[0, ∞] satisfied φ(1) = 0. Then φ quasi inverse function 𝜑𝜑 −1 is:
𝜑𝜑−1 (𝑡𝑡), 0 ≤ 𝑡𝑡 ≤ 𝜑𝜑(0)
𝜑𝜑 −1 (t) = � （3）
𝜑𝜑(0) ≤ 𝑡𝑡 ≤ ∞
𝜑𝜑 −1 domain is [0, ∞], range is [0,1]. So in the case of multiple function 𝜑𝜑 is the generating
element of Archimedean Copula function, C(𝜇𝜇1 , … … , 𝜇𝜇𝑛𝑛 ) = 𝜑𝜑 −1 （𝜑𝜑(𝑢𝑢1 ) + ⋯ 𝜑𝜑(𝑢𝑢𝑛𝑛 ). Since the
generator is a strictly decreasing convex function on [0,1], each Archimedean Copula function
corresponds to a unique generator.
Set U, V is a uniform random variable on [0,1], and their joint distribution function is
generated by the generated Archimedean function C(u, v), in the case of two variable, the
distribution function of random variables is:
𝜑𝜑(𝑡𝑡)
𝐾𝐾𝐶𝐶 (𝑡𝑡) = 𝑃𝑃(C(U, V) ≤ t) = t − ′ (𝑡𝑡 +) ，t ∈ [0,1] （4）
𝜑𝜑
𝜑𝜑′ (𝑡𝑡 + )Represents the right inverse at t.
When the generator meets the necessary conditions, in accordance with the above approach,
you can get the form of multivariate Archimedean Copula function, such as Gumbel copula,
Clayton copula and Frank copula, etc.. The Gumbel copula function usually has only the upper tail
correlation, Clayton copula has only the lower tail correlation, and the Frank copula has neither the
tail dependence nor the tail dependence.
The following describes the kinds of Copula function of the form of the two element function.

1999
 Advances in Social Science, Education and Humanities Research, volume 123

（1） Gumbel copula：

φ(t) = (−𝑙𝑙𝑙𝑙𝑙𝑙)𝛼𝛼 ，αϵ[1, ∞)
1
𝐶𝐶𝐺𝐺𝐺𝐺 (𝑢𝑢, 𝑣𝑣, 𝛼𝛼) = exp{−[(−𝑙𝑙𝑙𝑙𝑙𝑙)𝛼𝛼 + (−𝑙𝑙𝑙𝑙𝑙𝑙)𝛼𝛼 ]𝛼𝛼 } （5）

（2） Clayton copula

1
φ(t) = (𝑡𝑡 −𝛼𝛼 − 1)，αϵ[−1, ∞)\{0}
𝛼𝛼
1
𝐶𝐶𝑐𝑐𝑐𝑐 (𝑢𝑢, 𝑣𝑣; 𝛼𝛼) = max{(𝑢𝑢−𝛼𝛼 + 𝑣𝑣 −𝛼𝛼 − 1)𝛼𝛼 , 0} （6）
（3） Frank copula
𝑒𝑒 −𝛼𝛼𝛼𝛼 − 1
φ(t) = 𝑙𝑙𝑙𝑙 𝛼𝛼 ，αϵ(−∞, ∞)\{0}
𝑒𝑒 − 1
1 (𝑒𝑒 −𝛼𝛼𝛼𝛼 −1)(𝑒𝑒 −𝛼𝛼𝛼𝛼 −1)
𝐶𝐶𝐹𝐹 (𝑢𝑢, 𝑣𝑣; 𝛼𝛼) = − ln(1 + ) （7）
𝛼𝛼 𝑒𝑒 −𝛼𝛼 −1
4、Two-parameter copula function:
The study of two-parameter copula function is relatively small, the relevant literature and
information is not much. The copula function has two parameters to represent the relationship, i.e. a
copula function can capture two related patterns, for example, a parameter to describe the tail
dependence, another parameter to describe the consistency; or, a description on the tail correlation,
a description of lower tail dependence and so on.
Its function form is:
−1 −1
C(𝑢𝑢1 , 𝑢𝑢2 ) = ψ(−logK(𝑒𝑒 −𝜓𝜓 (𝑣𝑣1) , 𝑒𝑒 −𝜓𝜓 (𝑣𝑣2) )) （8）
And, ψ is a Laplace transform, K is one of the biggest infinitely divisible copula. ψ and K
can take different forms, then the corresponding copula function and different in different forms, so
the two variables copula is flexible.
The product copula function is the most basic copula function. And The elliptic copula
function can be regarded as an indirect copula function, according to Skalr theorem and copula
definition, the degree of correlation among the basic copula function. The Archimedean Copula
function can be regarded as a direct copula function, because it is directly obtained by the concrete
generating function. It should be noted that these are a single parameter model, the general formula
for C(u, v; δ), where the parameters of the delta correlation. The same form of Copula function, the
greater the δ, the stronger the correlation. The single parameter copula parameter is used to
represent the consistency of the relevant metrics, such as Tau Kendall' and Spearman 's Rho, or tail
correlation metric, such as the Archimedean Copula function. The parameters of the two parameters
are used to represent the correlation parameters are two, so a copula function can capture the
correlation of the two models.
As the copula function is a new branch of science developed recently, the related theories are
still being studied and perfected. Therefore, in the near future, the copula function theory will be
improved and enriched.

References:
[1] Ming Han: Copula-a new econometric tools, statistics and information.Forum Vol.19
No.5(2004), p.93
[2] Yaoting Zhang: connection function (copula) technology and financial risk analysis , statistical
research No.4(2002)
[3] Yanhua Wei, Shiying Zhang, Lifeng Meng: the application of Copula theory in finance, Journal
of NSTUAF .Vol.3 No.5(2003), p.97
[4] Xiao Wang: Research on the properties of Copula function, Education space, p.139
[5] Shushan Li:Copula regression and prediction, statistics and decision 18(20150), p.70
[6] Ni Zhang, Yiwen Zhang: Research on the correlation between macroeconomic and stock market
based on Copula theory, Value Engineering, p.3

2000
 Advances in Social Science, Education and Humanities Research, volume 123

[7] Yan Mu, Zhongzhi Wang: A note on symmetric Copula, J. of Math.Vol.33(2013)No.6, p.1085
[8] Jialong Liu: application of Copula function in financial risk measure, graduate thesis(2012)
[9] Liqin Zhao: The Study of Financial Risk Measurement Based on Copula Function doctcoral
thesis(2009)

2001