Yale University
Department of Statistics
ON OPTIMAL MATCHING AND BOOTSTRAP OF EMPIRICAL PROCESSES
Monday, October 14, 1996
Yongzhao Shao
Department of Statistics
Columbia University
New York, New York 10027
yshao@stat.columbia.edu
Seminar to be held in Room 107, 24 Hillhouse
Abstract
Matching problems have arisen in a number of interesting and seemingly
unrelated areas. In the last decade, there have been major advances in
optimal matching problems using probabilistic methods (see e.g., Statistical
Science, 1993, No.1). We first show that the problem of bootstrapping
empirical measures may be transformed into a problem of optimal matching.
Some recently obtained optimal matching results are applied to establish
rate of convergence results on bootstrapped empirical measures with
respect to the dual bounded Lipschitz metric as well as the Prohorov
distance (i.e. Glivenko-Cantelli type theorems) in all dimensions. The
connection between matching probelms and Efron's bootstrap is then used
through refining a representation method, introduced in Lo (1989), to obtain
rate of convergence results for bootstrapping empirical processes
exemplified by the empirical process indexed by Vapnik-Cervonenkis
classes of sets. For a particular example, the Donsker type central limit
theorem does not hold for the uniform empirical process indexed by
lower layers in the unit square, however the supremum distance between
a version of the empirical process and its bootstrapped counterpart
is shown to go to zero almost surely, with a rate faster than
``square-root n.'' The bootstrap results obtained here complement the
well-known results of Gin{\'e} and Zinn (1990) on bootstrapping
CLT of empirical processes.