I know that the Gaussian copula has a zero tail dependence (tail independence) due to the exponential behaviour at the tail. I am just wondering if there is a rigorous proof for this?

For simplicity sake, a bivariate Gaussian copula $C(X,Y)$ would do. In particular, there are no perfect correlations between $X$ and $Y$.

  • $\begingroup$ What definition of tail dependence do you have? $\endgroup$ – Glen_b Nov 13 '16 at 5:37
  • $\begingroup$ @Glen_b Limit as q->1 P[F(X) > q | F(Y) > q], where F is the standard normal marginal CDF. $\endgroup$ – NicTamHK Nov 13 '16 at 5:40
  • $\begingroup$ In what sense do you assume that the faster that exponential decaying normal distributions' tails are "exponential"? Where do you get that from? $\endgroup$ – Carl Nov 13 '16 at 5:50
  • $\begingroup$ @Carl I'm just assuming that might be the reason why the tail dependence is zero, since the normal PDF/CDF involves the exponential factor. $\endgroup$ – NicTamHK Nov 13 '16 at 5:52
  • $\begingroup$ @NicTamHK Tail heaviness is usually adjudicated by survival function ratios. See, stats.stackexchange.com/questions/86429/… as an example. With respect to Cupolas, tail heaviness may be irrelevant. Why do you think it relevant? $\endgroup$ – Carl Nov 13 '16 at 6:03

Consider a bivariate Gaussian copula $C(\cdot)$.

Because of the radial symmetry of a Gaussian copula we can consider just the lower tail dependence. We know that the lower tail dependence for this copula is:

$$\begin{align} \lambda&=\lim_{\,\,q\to 0^{+}} \frac{\partial C(q,q)}{\partial q}\\ &=\lim_{\,\,q\to 0^{+}} \text{Pr}(U_{2}\leq q\,|\,U_{1}=q)+ \lim_{\,\,q\to 0^{+}} \text{Pr}(U_{1}\leq q\,|\,U_{2}=q) \end{align}$$

Since a Gaussian copula is exchangeable, it follows that:

$$\lambda=2\lim_{\,\,q\to 0^{+}}\text{Pr}(U_{2}\leq q\,|\,U_{1}=q)$$

Now, let:


This means that $(X_{1},X_{2})$ has a bivariate normal distribution with standard marginals and correlation $\rho$. Now:

$$\begin{align} \lambda&=2\lim_{\,\,q\to 0^{+}}\text{Pr}(\Phi^{-1}(U_{2})\leq \Phi^{-1}(q)\,|\,\Phi^{-1}(U_{1})=\Phi^{-1}(q))\\ &=2\lim_{x\to -\infty}\text{Pr}(X_{2}\leq x\,|\, X_{1}=x) \end{align}$$

Finally, we know that $X_{2}\,|\,X_{1}\sim N(\rho x,1-\rho^{2})$, so:

$$\lambda=2\lim_{x\to -\infty}\Phi\Bigg(x\sqrt{\frac{(1-\rho)}{(1 +\rho)}}\Bigg)=0$$

  • $\begingroup$ I was wondering why in the partial derivative of the Gaussian copula there are two terms in the sum: $\lambda=\lim_{\,\,q\to 0^{+}} \text{Pr}(U_{2}\leq q\,|\,U_{1}=q)+ \lim_{\,\,q\to 0^{+}} \text{Pr}(U_{1}\leq q\,|\,U_{2}=q)$ $\endgroup$ – AlexMe Mar 5 '19 at 10:22
  • $\begingroup$ @AlexMe because both random variables have lower tails $\endgroup$ – epp Mar 5 '19 at 13:21
  • $\begingroup$ Thanks @StatsPlease, I'm just confused with how can one obtain the above mentioned sum as the partial derivative of the double integral (which is the form of the Gaussian copula). $\endgroup$ – AlexMe Mar 5 '19 at 13:55
  • 1
    $\begingroup$ @AlexMe Might be worth raising another question related to this. $\endgroup$ – epp Mar 5 '19 at 22:05

For a non-technical, intuitive view of what the tail index is telling you, we can look at simulation and compute sample estimates of the quantity $P[F(Y) > q | F(X) > q]$ as $q$ increases.

Plots of y vs x given the x values are in the upper tail, with proportion of y in the upper tail marked, for correlated Gaussian

Here the original correlation is $0.96$, but as we get further into the upper tail of $X$ the correlation decreases, and the proportion where $F(Y) > q$ given $F(X) > q$ decreases.

  • $\begingroup$ I think a formal proof of this observed convergence is given in my answer. $\endgroup$ – epp Jan 24 '17 at 6:56
  • 5
    $\begingroup$ Indeed, but often people looking at this also seek intuition about what's going on. I attempted to supply some. $\endgroup$ – Glen_b Jan 24 '17 at 7:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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