|
Activity Number:
|
637
- Non-Standard Regression Models
|
|
Type:
|
Invited
|
|
Date/Time:
|
Thursday, August 3, 2017 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
IMS
|
|
Abstract #322007
|
View Presentation
|
|
Title:
|
Super-Efficiency with the Divide and Conquer Method
|
|
Author(s):
|
Moulinath Banerjee* and Cecile Durot and Bodhisattva Sen
|
|
Companies:
|
University of Michigan and University of Paris Ouest and Columbia University
|
|
Keywords:
|
divide and conquer ;
pooled estimate ;
cube root asymptotics
|
|
Abstract:
|
I will discuss how the divide and conquer method in non-standard problems (where natural estimates converge at non square root n rates and the limiting distributions are typically non-Gaussian) results in a super-efficiency phenomenon. I will focus mostly on problems with a cube-root convergence rate. For a fixed model, the asymptotic relative-efficiency of the divide-and-conquer-based pooled estimate to the global least squares estimate/global M-estimate goes to infinity, whereas its asymptotic relative efficiency in terms of maximal risk over a class of models (in a neighborhood of a fixed model) relative to the M-estimate converges to 0.
|
Authors who are presenting talks have a * after their name.
