On the non-parametric estimation of the bivariate extreme-value distributions (Q749091)

Let \((X_ 1,Y_ 1),(X_ 2,Y_ 2),...\), be an i.i.d. sequence of random vectors with a common bivariate extreme value distribution with standard Gumbel marginals, i.e. \[ \Lambda (x,y)=P[X_ 1\leq x,Y_ 1\leq y]=\exp [-(e^{-x}+e^{-y})]k(x-y) \] for \(-\infty <x,y<\infty\), where the dependence function k(.) satisfies suitable conditions and must be estimated. By the transformation \(U_ i=\exp (-X_ i)\), \(V_ i=\exp (- Y_ i)\), \(i=1,2\), we have: \[ H(u,v)=P[U_ 1>u,V_ 1>v]=\exp \{- (u+v)A[v(u+v)]\}, \] u\(>0\), \(v>0\), where \(\{\) A(s): \(0\leq s\leq 1\}\) is related to \(\{\) k(z): \(-\infty <z<\infty \}\) by the equation \(k(z)=A[1/(e^ z+1)].\) Necessary and sufficient conditions for k(.) to be the dependence function of an extreme value distribution can be written in a simple form in terms of A(.) [\textit{J. Pickands III}, Bull. Int. Stat. Inst. 49, No.2, 859-878 (1981; Zbl 0518.62045)]. They are as follows: 1) A(.) is convex on [0,1] and 2) max(s,1-s)\(\leq A(s)\leq 1\) for \(0\leq s\leq 1.\) The second author [Tech. Rep. 87-18, Dept. Stat., Iowa Stat. Univ. (1981)] introduced the following non-parametric estimator of A(s): \[ A_{n,\delta_ n}(s)=1-n^{-1}R(\delta_ n)\sum^{n}_{i=1}\min [(1-s)/U_ i^{\delta_ n},s/V_ i^{\delta_ n}], \] for \(0\leq s\leq 1\), where \(0<\delta_ n<1\) is a sequence of exponents such that \(\delta\nearrow 1\) as \(n\to \infty\), and where R(\(\delta\)) is a function of \(0<\delta <1\) such that \(R(\delta)/(1-\delta)\to 1\) as \(\delta\nearrow 1.\) In this paper the limiting behavior of \(A_{n,\delta_ n}\) as \(n\to \infty\) is studied. The main result is the following Theorem: \(\sup_{0<s<1}| A_{n,\delta_ n}(s)-A(s)| \to 0\) in probability as \(n\to \infty\), for any A(.), if and only if the sequence \(0<\delta_ n<1\) satisfies the condition that \(\delta_ n\to 1\) and \((1-\delta_ n)\log n\to \infty\) as \(n\to \infty\).