Copula (probability theory)
A copula is a multivariate cumulative distribution function whose one-dimensional marginal distributions are uniform on the unit interval. Copulas separate the marginal behavior of random variables from their dependence structure. The term derives from the Latin word for “link” and is analogous to the copula in grammar, since the function links marginal distributions to a joint distribution.
For a random vector ((X_1,\ldots,X_d)) with joint cumulative distribution function (H) and marginal cumulative distribution functions (F_1,\ldots,F_d), the associated copula (C) satisfies
[ H(x_1,\ldots,x_d)
C\bigl(F_1(x_1),\ldots,F_d(x_d)\bigr). ]
This decomposition is formalized by Sklar's theorem. It permits properties of the marginals, including their locations and tail shapes, to be studied separately from the dependence represented by (C).
Definition
A (d)-dimensional copula is a function
[ C:[0,1]^d\longrightarrow[0,1] ]
that is a cumulative distribution function with standard uniform marginals. Consequently, for every (u\in[0,1]) and each coordinate (k),
[ C(1,\ldots,1,u,1,\ldots,1)=u, ]
where (u) occupies the (k)-th coordinate.
The function is grounded, meaning that
[ C(u_1,\ldots,u_d)=0 ]
whenever at least one coordinate equals zero. It is also (d)-increasing: the (C)-volume assigned to every axis-aligned rectangle in ([0,1]^d) is nonnegative. This condition is the multivariate counterpart of monotonicity for an ordinary cumulative distribution function.
Every copula is coordinatewise nondecreasing and satisfies the Lipschitz bound
[ \left|C(\mathbf{u})-C(\mathbf{v})\right| \leq \sum_{j=1}^{d}|u_j-v_j|. ]
It is therefore continuous even when the distributions connected by it contain atoms.
Sklar's theorem
Sklar's theorem states that every (d)-dimensional cumulative distribution function (H), with marginals (F_1,\ldots,F_d), admits a copula (C) such that
[ H(x_1,\ldots,x_d)
C\bigl(F_1(x_1),\ldots,F_d(x_d)\bigr). ]
When all marginals are continuous, the copula is unique on the entire unit cube. If a marginal is discontinuous, the copula is uniquely determined only on the Cartesian product of the marginal ranges. Its values elsewhere can vary without changing (H).
Conversely, any copula combined with univariate cumulative distribution functions through the same formula produces a valid joint cumulative distribution function. For continuous marginals, the copula can be recovered from (H) through the generalized inverse functions (F_j^{-1}):
[ C(u_1,\ldots,u_d)
H\bigl(F_1^{-1}(u_1),\ldots,F_d^{-1}(u_d)\bigr). ]
Abe Sklar stated the theorem in 1959 while investigating probabilistic forms of metric spaces. In 1962, You Watanabe established the corresponding extension result for subcopulas defined on products of marginal ranges, clarifying how nonuniqueness arises when the marginals have discontinuities. The resulting formulation became the measure-theoretic version used for general distribution functions.
Dependence and invariance
A copula is unchanged by strictly increasing transformations of individual coordinates. If
[ Y_j=g_j(X_j) ]
for strictly increasing functions (g_j), then the random vectors ((X_1,\ldots,X_d)) and ((Y_1,\ldots,Y_d)) have the same copula. Copulas therefore describe dependence at the level of ranks rather than at the level of measurement units.
The independence copula is
[ \Pi(u_1,\ldots,u_d)=\prod_{j=1}^{d}u_j. ]
A random vector with continuous marginals has this copula exactly when its components are mutually independent. Independence is thus represented by a particular copula rather than by the absence of one.
Every (d)-dimensional copula satisfies the Fréchet–Hoeffding bounds:
[ \max\left{1-d+\sum_{j=1}^{d}u_j,0\right} \leq C(u_1,\ldots,u_d) \leq \min(u_1,\ldots,u_d). ]
The upper bound is itself a copula in every dimension and represents comonotonic dependence. The lower bound is a copula in dimension two, where it represents countermonotonic dependence, but it fails to be a copula in general dimensions greater than two. Maurice Fréchet derived the bounding structure for multivariate distributions, while Wassily Hoeffding connected these bounds with integral measures of statistical dependence.
Concordance
Copulas provide distribution-free expressions for several measures of concordance. For a continuous bivariate distribution with copula (C), Kendall's tau is
[ \tau
4\int_{[0,1]^2} C(u,v),dC(u,v)-1. ]
This quantity compares the probabilities of concordant and discordant pairs of independent observations from the same distribution.
Spearman's rho is given by
[ \rho_S
12\int_0^1\int_0^1 \bigl(C(u,v)-uv\bigr),du,dv. ]
The product (uv) is the bivariate independence copula, so the integral measures the average departure of (C) from independence across the unit square. Unlike ordinary Pearson correlation, these quantities depend only on the copula when the marginals are continuous.
Concordance does not impose a complete ordering on all copulas. A common partial order defines (C_1) as less concordant than (C_2) when
[ C_1(u,v)\leq C_2(u,v) ]
throughout the unit square, together with the corresponding inequality for the survival copulas. Many pairs of copulas remain incomparable under this relation.
Tail dependence
Tail dependence describes limiting association in extreme regions of a joint distribution. For a bivariate copula (C), the lower tail-dependence coefficient is
[ \lambda_L
\lim_{u\downarrow 0}\frac{C(u,u)}{u}, ]
provided that the limit exists. It represents the limiting conditional probability that one transformed variable lies below a low quantile when the other does so.
The upper tail-dependence coefficient is
[ \lambda_U
\lim_{u\uparrow 1} \frac{1-2u+C(u,u)}{1-u}. ]
Copulas with similar dependence near the center of the distribution can have different values of (\lambda_L) or (\lambda_U). Ordinary correlation therefore does not determine joint extreme behavior.
The Gaussian copula has zero upper and lower tail-dependence coefficients whenever its correlation parameter is strictly between (-1) and (1). The copula derived from the multivariate Student's t-distribution generally has positive symmetric tail dependence when its degrees of freedom are finite and its correlation is not at a degenerate boundary.
Principal constructions
An elliptical copula is obtained from an elliptical distribution by transforming each coordinate through its marginal cumulative distribution function. Its parameters generally encode a matrix describing the orientation of dependence, while the generating distribution determines additional features such as tail behavior.
An Archimedean copula is represented in the bivariate case by
[ C(u,v)
\varphi^{-1}\bigl(\varphi(u)+\varphi(v)\bigr), ]
where the generator (\varphi) is decreasing and convex under the standard bivariate conditions. Higher-dimensional validity requires stronger regularity, commonly expressed through complete monotonicity of an associated inverse generator. Archimedean constructions reduce a multivariate dependence model to a scalar generating function, although their basic exchangeable form treats coordinate permutations identically.
A vine copula decomposes a high-dimensional density into a product of bivariate copula densities and conditional distributions. The decomposition is organized by a sequence of linked trees, with later trees representing conditional dependence after variables from earlier trees have been fixed. This representation allows different pairs of coordinates to have distinct dependence structures.
Copula densities and conditional distributions
When (C) is sufficiently differentiable, its copula density is
[ c(u_1,\ldots,u_d)
\frac{\partial^d C(u_1,\ldots,u_d)} {\partial u_1\cdots\partial u_d}. ]
If the joint distribution and all marginals possess densities, then
[ f(x_1,\ldots,x_d)
c\bigl(F_1(x_1),\ldots,F_d(x_d)\bigr) \prod_{j=1}^{d} f_j(x_j). ]
The copula density is therefore the multiplicative component that distinguishes the joint density from the product of its marginal densities. For the independence copula, (c=1) almost everywhere.
In the bivariate absolutely continuous case, differentiation also yields conditional distributions. In particular,
[ \Pr(V\leq v\mid U=u)
\frac{\partial C(u,v)}{\partial u} ]
at points where the derivative and regular conditional distribution are defined in the corresponding density sense.
Statistical estimation
Parametric copula models separate marginal parameters from dependence parameters. Full maximum-likelihood estimation evaluates the joint density as a single model, whereas inference functions for margins estimate the marginal distributions before estimating the copula parameters from transformed observations.
Semiparametric estimation replaces fitted marginal distributions with empirical distribution functions. The resulting scaled ranks,
[ \widehat U_{ij}
\frac{R_{ij}}{n+1}, ]
are called pseudo-observations, where (R_{ij}) is the rank of observation (i) in coordinate (j). Their denominator keeps the transformed data away from the boundary of the unit cube, where several copula densities are unbounded.
The empirical copula is the joint empirical distribution of these pseudo-observations. Its asymptotic behavior supports goodness-of-fit statistics and rank-based estimators, subject to regularity conditions concerning the copula's partial derivatives. Tied observations require separate treatment because ordinary ranks no longer correspond to observations from continuous marginals.
Copula selection is not determined by linear correlation alone. Models sharing the same correlation or rank-correlation coefficient can differ in asymmetry, conditional behavior, and tail dependence. These distinctions arise from the full copula rather than from any single scalar summary.