I am new to the topic of copulas and my math is limited. I have different questions:

1. Is it correct to say, that $C: [0,1]^d \rightarrow [0,1]$ is a mapping from a multidimensional distribution to its one dimensional margins? Or that this is a mapping from the unit cube to the uniform margins?

2. I am often reading about the Fréchet-Hoeffding bounds, these seem to basically give the co-domain of the copula? Or what do they state and why do I need them? When are they reached? If I understood it correctly, than the lower bound is reached, when d=2 and the random variables are perfectly negatively dependent and the upper bound is reached, when the random variables are perfectly positively dependent?

3. In every book about copulas I read about the Sklar's Theorem. I do not understand the intuition of this mathematical theorem. What does it say and why do I need it in the context of copulas?

4. The definition of a copula is: It is a multivariate distribution with uniform marginal distributions, but I thought, that I can take any distribution for the marginal distributions and not only the uniform distribution?

I know these are beginners questions, but I could not find a good introductory book about copulas. Every book in this topic gives a lot of mathematical definitions and theorems and I am not on this level, unfortunately.

  • $\begingroup$ Just curious, which books have you been consulting? Mentioning a couple will do fine. $\endgroup$ Jun 30, 2013 at 16:56
  • 1
    $\begingroup$ @GraemeWalsh Embrechts Quantitative Risk Management and Härdles QF book. Embrechts sometimes summarizes the ideas shortly, but it is not sufficient for me. $\endgroup$ Jun 30, 2013 at 17:07

1 Answer 1


You might find Harry Joe's book easier going.

No, the mapping is not to the margins. It represents that the CDF is a function of its arguments - for every $(u_1,...,u_d)$ you get a single value for $P(U_1≤u_1,U_2≤u_2,...,U_d≤u_d)$; this is what $C$ is/does.

Q4 confuses copulas (which have uniform margins) with what you use them for (representing multivariate dependence more generally); you could transform the uniform margins of a copula to anything, but then the multivariate distribution is no longer itself a copula (but it still has the original copula)

Sklar's theorem is the thing that connects a distribution to its copula - in effect it explains what's going on in your Q4. Let's look at 2D. If distribution $H(x,y)$ has marginals $F(x)$ and $G(y)$ then Sklar says: $H(x,y)=C(F(x),G(y))$. Think about what $H$ is! What does the value of $H$ at some $(x,y)$ tell you?

Sklar's theorem is in effect telling you that if you take something with cdf $H$ and transform all its margins, you actually get $C$; the copula takes identical values to $H$ if you evaluate $C$ at $F(x)$ and $G(y)$.

Fréchet-Hoeffding bounds: Like any distribution function, $C$ is bounded between $0$ and $1$, but because of the additional restrictions on $C$, $C(\mathbf{0})=0$ and $C(\mathbf{1})=1$ (where $\mathbf{0} = (0,0,...,0)$ and similarly for $\mathbf{1}$). Fréchet-Hoeffding identifies the biggest and smallest values $C$ can have everywhere in between; the special restrictions on $C$ let us find bounds on $C$ for any $\mathbf{u}$. When $C$ takes its maximum value, yes, it corresponds to perfect monotonic dependence. In 2D, yes, the minimum value $C$ can take corresponds to perfect negative monotonic dependence.

Many of your confusions seem to stem from not really thinking about what $C$ or $H$ do. If you think about it in a very elementary sense, functions have a value; thinking about what it's the value of will help.

(Some more-or-less introductory papers and presentations that may help: 1 2 3 4 5 6 7 8)

  • 1
    $\begingroup$ +1 for your answer, but I don't know what H(x,y)=C(F(x),G(y)) tells me? But I can use any distribution for the margins right? I am not getting, why then the margins are uniform? Since they also can be normal or so? $\endgroup$ Jul 1, 2013 at 8:01
  • $\begingroup$ @StatTistician $C$ has uniform margins. The margins on $H$ can be anything. $C$ is the copula of $H$. Let's say I start with some known $H$, and compute $C$ via Sklar. Then I let $x_2 = F_2^{-1}(u_1)$ and $y_2 = G_2^{-1}(u_2)$, then $H_2$, the cdf of $(X_2,Y_2)$ is different from $H$... but they share the same $C$. $\endgroup$
    – Glen_b
    Jul 1, 2013 at 9:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.