eCommons

 

High-Dimensional Inference By Unbiased Risk Estimation

Other Titles

Abstract

This thesis derives natural and efficient solutions of three high-dimensional statistical problems by exploiting unbiased risk estimation. They exemplify a general methodology that provides attractive estimators in situations where classical theory is unsuccessful, and that could be exploited in many other problems. First, we extend the classical James-Stein shrinkage estimator to the context where the number of covariates is larger than the sample size and the covariance matrix is unknown. The construction is obtained by manipulating an unbiased risk estimator and shown to dominate in invariant squared loss the maximum likelihood estimator. The estimator is interpreted as performing shrinkage only the random subspace spanned by the sample covariance matrix. Second, we investigate the estimation of a covariance and precision matrix, and discriminant coefficients, of linearly dependent data in a normal framework. By bounding the difference in risk over classes of interest using unbiased risk estimation, we construct interesting estimators and show domination over naive solutions. Finally, we explore the problem of estimating the noise coefficient in the spiked covariance model. By decomposing an unbiased risk estimator and minimizing its dominant part using calculus of variations, we obtain an estimator in closed form that approximates the optimal solution. Several attractive properties are proven about the proposed construction. We conclude by showing that the associated spiked covariance estimators possess excellent behavior under the Frobenius loss.

Journal / Series

Volume & Issue

Description

Sponsorship

Date Issued

2015-05-24

Publisher

Keywords

James-Stein; Covariance estimation; Spiked covariance

Location

Effective Date

Expiration Date

Sector

Employer

Union

Union Local

NAICS

Number of Workers

Committee Chair

Wells,Martin Timothy

Committee Co-Chair

Committee Member

Bien,Jacob
Bunea,Florentina

Degree Discipline

Statistics

Degree Name

Ph. D., Statistics

Degree Level

Doctor of Philosophy

Related Version

Related DOI

Related To

Related Part

Based on Related Item

Has Other Format(s)

Part of Related Item

Related To

Related Publication(s)

Link(s) to Related Publication(s)

References

Link(s) to Reference(s)

Previously Published As

Government Document

ISBN

ISMN

ISSN

Other Identifiers

Rights

Rights URI

Types

dissertation or thesis

Accessibility Feature

Accessibility Hazard

Accessibility Summary

Link(s) to Catalog Record