Weak convergence of the least concave majorant of estimators for a concave distribution function (Q2412257)

The paper is concerned with the estimation of distribution functions \(F\) of non-negative random variables under the shape constraint that the distribution function \(F\) is concave. Let \(F_n\) be a sequence of estimators of \(F\), viewed as random elements in \(L^\infty ([0,\infty))\) such that \(G_n := \sqrt{n} (F_n - F)\) converges in distribution to a random element \(G\) in \(L^\infty([0,\infty))\), such that the paths of \(G\) are continuous and vanish at infinity. These assumptions are satisfied for example if \(F_n\) is the empirical distribution function of an independent and identically distributed sample of size \(n\) drawn from \(F\), or more generally for the empirical distribution function when certain mixing conditions are satisfied. Denote by \(\mathcal{M} : L^\infty([0,\infty)) \to L^\infty([0,\infty))\) the operator which assigns to each bounded measurable function on \([0,\infty)\) its least concave majorant. Let \(\widehat{F}_n := \mathcal{M} F_n\) and \(\widehat{G}_n := \sqrt{n} (\widehat{F}_n - F)\). The authors then establish the Hadamard directional differentiability of \(\mathcal{M}\) at each point \(\theta \in L^\infty([0,\infty))\), tangentially to the space of continuous functions vanishing at infinity. Using this, they show that \(\widehat{G}_n\) converges weakly to a process \(\widehat{G}\) given by \(\widehat{G} = \mathcal{M}_F' G\), where \(\mathcal{M}_F'\) is the Hadamard directional derivative at \(F\). An important point is that the authors require the distribution function \(F\) only to be concave, but not necessarily strictly concave. The authors then establish the asymptotic behaviour of bootstrap methods, with particular emphasis on the rescaled bootstrap and apply their results to construct reliable confidence bands for the true distribution. The authors also show how the used approach can be used for the study of other shape constrained estimators, at the example of isotonic regression estimators.