References & Citations
Mathematics > Statistics Theory
Title: Consistent order estimation and minimal penalties
(Submitted on 5 Feb 2010 (v1), last revised 16 Feb 2012 (this version, v3))
Abstract: Consider an i.i.d. sequence of random variables whose distribution f* lies in one of a nested family of models M_q, q>=1. The smallest index q* such that M_{q*} contains f* is called the model order. We establish strong consistency of the penalized likelihood order estimator in a general setting with penalties of order \eta(q) log log n, where \eta(q) is a dimensional quantity. Moreover, such penalties are shown to be minimal. In contrast to previous work, an a priori upper bound on the model order is not assumed. The results rely on a sharp characterization of the pathwise fluctuations of the generalized likelihood ratio statistic under entropy assumptions on the model classes. Our results are applied to the geometrically complex problem of location mixture order estimation, which is widely used but poorly understood.
Submission history
From: Ramon Van Handel [view email][v1] Fri, 5 Feb 2010 19:57:08 GMT (53kb)
[v2] Thu, 16 Jun 2011 06:51:57 GMT (118kb)
[v3] Thu, 16 Feb 2012 00:53:52 GMT (24kb)
Link back to: arXiv, form interface, contact.