Ratio limit theorems for empirical processes (Q4445181)

Let \({\mathcal F}\) be a uniformly bounded class of \([0,1]\)-valued random variables defined on the probability space \((S,{\mathcal A},P)\) and let \((X_i)_{i\geq 1}\) be a sequence of i.i.d. \(S\)-random variables with distribution \(P\). Let \(\sigma_P:{\mathcal F}\to \mathbb{R}\) be a norming function satisfying the inequality \(\sigma^2_P(f)\geq\text{Var}(f(X_1))\). The authors investigate the asymptotic behavior of certain functionals of the empirical process, such as \[ \sup_{f\in{\mathcal F},\sigma_P(f)> \delta_n} {|\sum^n_{i=1} (f(X_i)- Ef(X_1))|\over \sqrt{n}\sigma_P(f)},\quad \sup_{f\in{\mathcal F},Ef(X_1)> \delta_n} \Biggl|{\sum^n_{i=1} f(X_i)\over nEf(X_1)}- 1\Biggr|, \] \[ \sup_{f\in{\mathcal F},\sigma_P(f)> \delta_n} {|\sum^n_{i=1} (f(X_i)- Ef(X_1))|\over \sqrt{n}\omega(\sigma_P(f))} \] for suitable modulus functions \(\omega\) and cutoff values \(\delta_n\). The authors use concentration inequalities to derive inequalities for the distribution of these functionals, which are in turn used to obtain limit theorems for the functionals. Both almost sure convergence as well as convergence in probability are studied. Most of the results have been obtained earlier by \textit{K. S. Alexander} [Probab. Theory Relat. Fields 75, 379--427 (1987; Zbl 0596.60029)], albeit with different and much harder proofs. As a new result a so-called oracle inequality for nonparametric regression is obtained. For the entire collection see [Zbl 1024.00031].