New Donsker classes

DOI10.1214/AOP/1041903221MaRDI QIDQ674523zbMATH OpenOpenAlexFDO

Publication date 12 October 1997

Published in The Annals of Probability (Search for Journal in Brave)

Full work available at URL https://doi.org/10.1214/aop/1041903221

bracketing number Brownian bridge covering number Donsker class empirical central limit theorem empirical measure entropy

Mathematics Subject Classification ID

Functional limit theorems; invariance principles (60F17)

Review text

Let \(P\) be a probability measure on the measurable space \(({\mathcal H}, {\mathcal A})\) and \({\mathcal F}\) a class of measurable functions \(f\): \({\mathcal H} \to\mathbb{R}\) such that \(Pf^2 <\infty\) for every \(f\in {\mathcal F}\). The class \({\mathcal F}\) is called \(P\)-Donsker if the empirical process \(\{ \mathbb{G}_n f:f\in {\mathcal F}\}\) converges in distribution to a tight Brownian bridge process \(\mathbb{G}\) in the space \(l^\infty ({\mathcal F})\) of uniformly bounded functions on \({\mathcal F}\). Here \(\mathbb{G}_n= \sqrt n(\mathbb{P}_n -P)\) with \(\mathbb{P}_n= n^{-1} \sum^n_{i=1} \delta_{X_i}\) denoting the empirical measure of an i.i.d. sample from \(P\). Convergence in distribution is understood in the sense that \(E^*h (\mathbb{G}_n) \to Eh (\mathbb{G})\) for every continuous bounded function \(h: l^\infty ({\mathcal F}) \to\mathbb{R}\); compare \textit{R. M. Dudley} [in: Probability in Banach spaces V. Lect. Notes Math. 1153, 141-178 (1985; Zbl 0587.60008)]. Let \({\mathcal H} =\bigcup^\infty_{j=1} {\mathcal H}_j\) be a partition of \({\mathcal H}\) into measurable sets and let \({\mathcal F}_j\) be the class of functions \(f\mathbf{1}_{{\mathcal H}_j}\) when \(f\) ranges over \({\mathcal F}\). If the class \({\mathcal F}\) is Donsker, then each class \({\mathcal F}_j\) is Donsker. This is not obvious from the definition of a Donsker class, but can be proved by extending Corollary 14.8 of \textit{E. Giné} and \textit{J. Zinn} [in: Probability and Banach spaces. Lect. Notes Math. 1221, 50-113 (1986; Zbl 0605.60026)] with the measurability conditions taken care of along the lines of \textit{M. Talagrand} [Ann. Probab. 15, 204-212 (1987; Zbl 0622.60040)], because a restriction is a contraction in the sense that \(|(f{\mathbf{1}}_{{\mathcal H}_j}) (x)-(g {\mathbf{1}}_{{\mathcal H}_j}) (x)|\leq |f(x)- g(x)|\) for every \(x\). We shall not use this claim in this note, but be are interested in a converse of this statement. While the sum of infinitely many Donsker classes need not be Donsker, it is clear that if each \({\mathcal F}_j\) is Donsker and the classes \({\mathcal F}_j\) become suitably small as \(j\to\infty\), then \({\mathcal F}\) is Donsker. The following precise version of this principle enables us to deal with a number of interesting examples. Let \(|G|_{\mathcal F} =\sup \{|Gf|: f\in{\mathcal F}\}\) be the norm of the process \(G\). Let \(F\) be an envelope function of the class \({\mathcal F}\), a function such that \(|f|\leq F\) for every \(f\). Theorem 1.1 For each \(j\) let the class of functions \({\mathcal F}_j\) be Donsker and suppose that \({\mathbf E}^*_P|\mathbb{G}_n|_{{\mathcal F}_j} \leq Cc_j\) for a constant \(C\) not depending on \(j\) or \(n\). If \(\sum^\infty_{j=1} c_j<\infty\) and \(P^*F <\infty\), then the class \({\mathcal F}\) is \(P\)-Donsker.

Recommendations

Cites work

Cited in

(43)

This page was built for publication: New Donsker classes

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q674523)