Cluster sets for a generalized law of the iterated logarithm in Banach spaces (Q1872231): Difference between revisions

Let \(X\), \(X_i\) be i.i.d. \(B\)-valued random variables, where \(B\) is a separable Banach space, and assume \(0 < E\|X\|< \infty\). The law of the iterated logarithm is extended to random vectors \(X\) that may fail to have a covariance function, generalizing results of \textit{U. Einmahl} [Ann. Probab. 21, 2012-2045 (1993; Zbl 0790.60034) and ibid. 23, 817-851 (1995; Zbl 0833.60029)] to infinite dimensions. It is shown that, with \(\gamma_n\) properly defined \((\gamma_n = \sqrt{2}\widetilde{K}(n/\log\log n)\log\log n\) where \(K\) is the sup of the \(K\) functions of \textit{M. J. Klass} [Z. Wahrscheinlichkeitstheorie Verw. Geb. 36, 165-178 (1976; Zbl 0319.60019)] corrresponding to the one-dimensional random variables \(f(X)\), \(f\in B^*\), \(\|f\|\leq 1\), \(E|f(X)|>0)\), if \(S_n/\gamma_n\to 0\) in probability and \(\sum\text{Pr}\{\|X\|>\gamma_n\}<\infty\), where \(S_n =\sum_{i=1}^n X_i\), then: a) \(\limsup \|S_n/\gamma_n\|=1\), b) almost surely, the cluster set of the sequence \(S_n/\gamma_n\) is a non-random set which is a closed subset of the unit ball of \(B\), non-empty, symmetric about zero and star-like at zero (and which does not have to contain vectors of length one if \(B\) is infinite-dimensional), and c) for any such set \(A\) in the unit ball of \(B\) there is a \(B\)-valued random variable \(X\) satisfying the stated conditions such that \(A\) is a.s. the cluster set of \(S_n/\gamma_n\).