Current browse context:
math.ST
Change to browse by:
References & Citations
Mathematics > Statistics Theory
Title: Partial estimation of covariance matrices
(Submitted on 10 Aug 2010 (v1), last revised 11 Feb 2011 (this version, v2))
Abstract: A classical approach to accurately estimating the covariance matrix \Sigma of a p-variate normal distribution is to draw a sample of size n > p and form a sample covariance matrix. However, many modern applications operate with much smaller sample sizes, thus calling for estimation guarantees in the regime n << p. We show that a sample of size n = O(m log^6 p) is sufficient to accurately estimate in operator norm an arbitrary symmetric part of \Sigma consisting of m < n entries per row. This follows from a general result on estimating Hadamard products M.\Sigma, where M is an arbitrary symmetric matrix.
Submission history
From: Roman Vershynin [view email][v1] Tue, 10 Aug 2010 14:11:52 GMT (12kb)
[v2] Fri, 11 Feb 2011 22:18:17 GMT (13kb)
Link back to: arXiv, form interface, contact.