Finite sample performance of linear least squares estimation

TL;DR

This paper provides finite sample bounds for linear least squares estimation, including tail bounds and convergence rates, especially under correlated noise and martingale difference sequences, enhancing understanding of its accuracy with limited data.

Contribution

It introduces new finite sample bounds for linear least squares, extending analysis to correlated noise and martingale differences, and demonstrates their tightness through simulations.

Findings

Bounds on the tail of the estimator's distribution are accurate.

Fast exponential convergence of sample size for desired accuracy.

Bounds outperform previous results in $L_{ extinfty}$ norm accuracy.

Abstract

Linear Least Squares is a very well known technique for parameter estimation, which is used even when sub-optimal, because of its very low computational requirements and the fact that exact knowledge of the noise statistics is not required. Surprisingly, bounding the probability of large errors with finitely many samples has been left open, especially when dealing with correlated noise with unknown covariance. In this paper we analyze the finite sample performance of the linear least squares estimator. Using these bounds we obtain accurate bounds on the tail of the estimator's distribution. We show the fast exponential convergence of the number of samples required to ensure a given accuracy with high probability. We analyze a sub-Gaussian setting with a fixed or random design matrix of the linear least squares problem. We also extend the results to the case of a martingale difference noise sequence. Our analysis method is simple and uses simple $L_{\infty}$ type bounds on the estimation error. We also provide probabilistic finite sample bounds on the estimation error $L_2$ norm. The tightness of the bounds is tested through simulation. We demonstrate that our results are tighter than previously proposed bounds for $L_{\infty}$ norm of the error. The proposed bounds make it possible to predict the number of samples required for least squares estimation even when the least squares is sub-optimal and is used for computational simplicity.