This 1960 paper by Masaaki Sibuya is a foundational work in multivariate extreme value theory. It introduces the "dependence function"—a precursor to the modern concept of copulas—to characterize how the maximum values of two random variables relate to one another as sample sizes grow to infinity.
TL;DR
Sibuya extends univariate extreme value theory to the bivariate case, introducing a "dependence function" to isolate the relationship between variables from their individual distributions. His most striking finding is that for many common distributions, including the bivariate normal, the maxima of the two variables become completely independent in the limit, unless they are perfectly correlated.
What Problem Does This Paper Address?
By 1960, statisticians well understood the "three types" of limit distributions for the maximum of a single variable (the Gumbel, Fréchet, and Weibull distributions). However, the bivariate case—where you track the maximum of two variables simultaneously, like the highest tide and the highest wind speed during a season—was much less clear.
The challenge was determining whether the dependence between two variables persists when you only look at their extreme values. If two variables are correlated, are their "once-in-a-century" events also correlated? Sibuya sought a mathematical framework to describe these joint limit distributions and identify the conditions under which extremes remain dependent or become independent.
Methods
Sibuya’s primary innovation is the dependence function, denoted as $\Omega(G, H)$. He defines the joint distribution $F(x, y)$ in terms of its marginal distributions $G(x)$ and $H(y)$ as: $$F(x, y) = \Omega(G(x), H(y)) \cdot G(x) \cdot H(y)$$
By focusing on $\Omega$, Sibuya could analyze the "pure" dependence between variables regardless of whether they followed a normal, exponential, or any other distribution. To find the limit behavior, he analyzed the function as the sample size $n$ tends toward infinity, specifically looking at the stability of $\Omega^n(G^{1/n}, H^{1/n})$.
Datasets
Not discussed in the paper. This is a theoretical statistics paper focused on mathematical proofs and derivations.
Key Insights and Ideas
- Asymptotic Independence: One of the paper's most famous results is that for a wide class of distributions, the maxima are "asymptotically independent." This means that even if two variables are strongly correlated in the "middle" of their distribution, their extreme values might behave as if they have nothing to do with each other.
- The Normal Distribution Surprise: Sibuya proves that for any bivariate normal distribution with a correlation coefficient $\rho < 1$, the maxima are asymptotically independent. Only when $\rho = 1$ (perfect correlation) do the extremes remain dependent. This has huge implications for risk management; you cannot assume that because two risks are correlated normally, their extreme "black swan" events will happen at the same time.
- Stable Distributions: Sibuya characterizes the functional form that a dependence function must take to be "stable" (meaning the distribution of the maximum of $n$ samples has the same dependence structure as a single sample). He shows this requires a specific convex function $\chi$, leading to what are now known as Sibuya’s stable extreme value distributions.
Key Results
- Theorem 2: Provides a formal condition for asymptotic independence based on the probability that both variables simultaneously exceed a high threshold.
- Theorem 5: Defines the necessary and sufficient form for a stable bivariate extreme value distribution: $\Omega(G, H) = G^{\chi(\log H / \log G)}$.
- Visualizing Dependence: Figure 1 illustrates the dependence function for bivariate normal distributions. The solid lines represent the dependence at a high correlation ($\rho = 0.8$), showing how the function shifts as we move toward the extremes.
Limitations and Future Work
The paper is primarily focused on the asymptotic properties (what happens as $n \to \infty$). In practice, researchers often work with finite samples where the "limit" hasn't been reached yet, meaning some residual dependence might still exist that this theory doesn't fully capture.
Sibuya also notes that this "Part I" focuses on the joint distribution of maxima; he suggests that further work is needed to explore other types of extreme statistics and practical applications, which he hints at for a subsequent paper. This work eventually paved the way for the extensive use of copulas in finance and hydrology.
Summary updated