Stability of maxima of random variables with multidimensional indices (Q2488448): Difference between revisions

Let \(\underline m=(m_1,\dots,m_r)\) and \(\underline{n}=(n_1,\dots,n_r)\) be multiindeces, \(X_{\underline{m}}\) are i.i.d. r.v.s with CDF \(F\), \(M_{\underline{n}}=\max_{\underline{m}:\;m_i\leq n_i,i=1,\dots,r} X_{\underline m}\). The authors consider conditions under which there exists a set of real numbers \(\{a_{\underline{n}}\}\) such that \(M_{\underline{n}}/a_{\underline{n}}\to 1\) in probability, a.s. or completely. (The convergence \(a_{\underline{n}}\to a\) means that \(\forall\varepsilon>0\) \(\exists \underline{n}(\varepsilon)\) such that \(| a_{\underline{n}}-a| <\varepsilon \) for all \(\underline{n}\) with \(n_i>n_i(\varepsilon)\) for some \(i\).) E.g., assume that \(F(x)<1\) for all \(x\) and the sequence \(b_k=k(1-F((1-\varepsilon)F^{-1}(1-k^{-1}))\) is nondecreasing for every \(\varepsilon\). Then \(M_{\underline{n}}/a_{\underline{n}}\to 1\) a.s. iff \[ \int_1^\infty\left(\log{1\over 1-F(\delta x)}\right)^{r-1} {dF(x)\over 1-F(\delta x)}<\infty \] for all \(0<\delta<1\).