A Necessary Condition for the Strong Oracle Property. (23rd November 2015)

• Record Type: Journal Article
• Authors: Kim, Yongdai; Jeon, Jong‐June; Han, Sangmi
• Is Part Of: Scandinavian journal of statistics. Volume 43:Number 2(2016:Jun.)
• Page Start: 610
• Page End: 624
• Publication Date: 2015-11-23
• Subjects: bridge penalty -- global optimizer -- non‐convex penalties -- oracle property; Statistics -- Periodicals; 310
• DOI: 10.1111/sjos.12195
• Languages: English
• ISSNs: 0303-6898

Abstract

Abstract: A problem of using a non‐convex penalty for sparse regression is that there are multiple local minima of the penalized sum of squared residuals, and it is not known which one is a good estimator. The aim of this paper is to give a guide to design a non‐convex penalty that has the strong oracle property. Here, the strong oracle property means that the oracle estimator is the unique local minimum of the objective function. We summarize three definitions of the oracle property – the global, weak and strong oracle properties. Then, we give sufficient conditions for the weak oracle property, which means that the oracle estimator becomes a local minimum. We give an example of non‐convex penalties that possess the weak oracle property but not the strong oracle property. Finally, we give a necessary condition for the strong oracle property.