Uniform Donsker classes of functions (Q1203665)

Given a probability space \((A,\mathbb{A},P)\) a class \({\mathcal F}\) of measurable functions is called a \(P\)-Donsker class if the sequence of processes \((\mathbb{X}^ P_ n(f))_{f\in{\mathcal F}}\), where \(\mathbb{X}^ P_ n=\sqrt n(\mathbb{P}_ n-P)\), converges weakly to a \(P\)-Brownian bridge \((G_ P(f))_{f\in{\mathcal F}}\). For fixed \(P\) the Donsker property of \({\mathcal F}\) is equivalent to the condition that the sequence of mappings \(f\mapsto X_ n(f)\), \(f\in{\mathcal F}\), is stochastically and asymptotically equicontinuous and either there exists a Brownian bridge \(G_ P\) on \({\mathcal F}\) or \({\mathcal F}\) is totally bounded w.r.t. \(\rho_ P(f,g)=(\text{Var}_ P(f(X)-g(X))^{1/2}\). Further equivalent conditions concern different forms of approximations of \(\mathbb{X}^ P_ n\) by sums \(\sum^ m_{i=1}\mathbb{Y}_ i\) where \(\mathbb{Y}_ i\) are independent copies of the Brownian bridge \(G_ P\). Another condition concerns the functional limit theorem for the processes \({1\over\sqrt n}\sum^{[ns]}_{i=1}(f(X_ i)-P(f))\). All these conditions except the last one are known from papers of Dudley, Philipp, van der Vaart, Wellner, Giné, Zinn and other authors. The aim of the present paper is to derive necessary and sufficient conditions for a class \({\mathcal F}\) to be a \({\mathcal P}\)-uniform Donsker class where \({\mathcal P}\) is a subcollection of all distributions. The case of all distributions was already studied by \textit{E. Giné} and \textit{J. Zinn} [Ann. Probab. 19, No. 2, 758-782 (1991; Zbl 0734.60007)]. To do this the authors give answer to the question how to introduce uniformities in the above conditions. If \({\mathcal F}\) has an envelope function which is \({\mathcal P}\)-uniformly square integrable, then Theorem 2.1 states conditions similar to the case of fixed \(P\) but now formulated uniformly w.r.t. \(P\in{\mathcal P}\) are equivalent to the \({\mathcal P}\)-uniform Donsker property of \({\mathcal F}\). The results of Theorem 2.1 are applied to the case where \({\mathcal P}\) is a sequence and to bootstrap resampling methods.