Some limit theorems for empirical processes (with discussion) (Q802198): Difference between revisions

The invited paper on the CLT for empirical processes indexed by classes of functions starts with a neat description of the basic tools of this field. Among others it introduces functional P-Donsker classes, random entropies, metric entropies, Vapnik-Červonenkis classes as well as K. Alexander's measurability concepts and adaptions of relevant inequalities from probability theory in Banach spaces. The following sections 3 to 8 are concerned with conditions which ensure the CLT for empirical processes indexed by functions, which modulo some measurability conditions ensure that the function class is a P-Donsker class [using a result of \textit{R. M. Dudley} and \textit{W. Philipp}, Z. Wahrscheinlichkeitstheor. Verw. Geb. 62, 509-552 (1983; Zbl 0488.60044)]. One interesting result for uniformly bounded function classes is that a P-Donsker class is a P-pregaussian class satisfying some 'reduced' randomized equicontinuity condition. The main result of the paper (Theorem 5.7) shows (using an idea of Le Cam) that the equicontinuity condition may be replaced by some tail condition for random entropies for the case of set classes. Applications of these interesting results are made to uniformly bounded processes in C[0,1] and D[0,1] as well as to classes of unbounded functions. Some of the technical lemmas of this paper might be useful in other contexts than the CLT. The paper is followed by a discussion in which \textit{K. Alexander} emphasizes the usefulness of Le Cam's square root trick' (Lemma 5.2), \textit{R. M. Dudley} illuminates the relation with the extensive literature on the CLT in Banach spaces, whereas \textit{P. Gaenssler} points out application in non parametric statistics estimating spatial patterns. \textit{W. Philipp} mentions the connection with the law of the iterated logarithm, \textit{D. Pollard} comments on various inequalities, \textit{R. Pyke} emphasizes unsolved problems in connection with computer aided sampling and \textit{W. Stute} comments on the fact the paper of Giné and Zinn does not apply directly to the invariance principle for the empirical characteristic function.