Extremal limit laws for a class of bivariate Poisson vectors (Q5953879)

Let \(U, V\) and \(W\) be independent Poisson r.v.'s with means \(\lambda_U , \lambda_V\) and \(\lambda_W\), respectively. Under a bivariate Poisson distributed random vector \((X,Y)\) with parameter vector \((\lambda_X, \lambda_Y, \lambda_W)\) the authors understand the vector defined by \(X=U+W\) and \(Y=V+W\). Then \((X,Y)\) has Poisson marginals with \(\lambda_X = \lambda_U + \lambda_W, \lambda_Y = \lambda_V + \lambda_W\) and \(\text{Cov}(X,Y) = \lambda_W\). For another approach to multivariate Poisson distribution see \textit{N. L. Ivanova} and \textit{Yu. S. Khokhlov} [Mosc. Univ. Comput. Math. Cybern. 2001, No.~1, 37-42 (2001); translation from Vestn. Mosk. Univ., Ser. XV 2001, No.~1, 33-37 (2001; Zbl 0996.60004)]. Consider a triangular array \((X_{n,i},Y_{n,i})\) of independent bivariate Poisson variables with parameter vector \((a_n, a_n, d_n)\) such that, for some integer \(r \geq 0\), \[ \log n = o(a_n^{(r+1)/(r+3)}), \quad (1 - d_n/a_n)\log n \rightarrow \lambda^2, \quad \lambda \geq 0 . \] Under these conditions the authors prove that there exists a suitable norming sequence \(u_n(.)\) such that for \(n \rightarrow \infty\), \(P(\max_{1\leq i\leq n}X_{n,i} \leq u_n(x) , \max_{1\leq i\leq n}Y_{n,i} \leq u_n(y))\) is weakly convergent to \[ H_{\lambda}(x,y) = \exp\Bigl\{ -\Phi\bigl( \lambda + \frac{x-y}{2\lambda}\bigr)e^{-y} - \Phi\bigl( \lambda - \frac{x-y}{2\lambda}\bigr)e^{-x} \Bigr\}. \] Here \(\Phi(.)\) denotes the standard normal df. The special cases of \(\lambda = 0\) and \(\lambda = \infty\) correspond to perfect (or total) dependence and independence, respectively. \textit{J. Hüsler} and \textit{R.-D. Reiss} [Stat. Probab. Lett. 7, 283-286 (1989; Zbl 0679.62038)] obtained \(H_{\lambda}(x,y)\) as limiting distribution for a suitably normalized sequence \((X_{n,i},Y_{n,i})\) of independent bivariate normal variables with standard marginals and correlation \(\rho (n)\) such that \((1-\rho (n))\log n \rightarrow \lambda^2\). Thus, one can observe that the extreme value behaviour of the bivariate Poisson family, suitably normalized, is identical to that of the bivariate normal distribution.