Minimax Rates of Estimation for High-Dimensional Linear Regression Over $\ell_q$ -Balls

TLDR: In this paper, the authors studied the minimax rates of convergence for estimating β* in either l 2-loss and l2-prediction loss, assuming that β* belongs to an l q -ball \BBBq(Rq) for some q ∈ [0, 1].

Abstract: Consider the high-dimensional linear regression model y = X β* + w, where y ∈ \BBRn is an observation vector, X ∈ \BBRn × d is a design matrix with d >; n, β* ∈ \BBRd is an unknown regression vector, and w ~ N(0, σ2I) is additive Gaussian noise. This paper studies the minimax rates of convergence for estimating β* in either l2-loss and l2-prediction loss, assuming that β* belongs to an lq -ball \BBBq(Rq) for some q ∈ [0,1]. It is shown that under suitable regularity conditions on the design matrix X, the minimax optimal rate in l2-loss and l2-prediction loss scales as Θ(Rq ([(logd)/(n)])1-q/2). The analysis in this paper reveals that conditions on the design matrix X enter into the rates for l2-error and l2-prediction error in complementary ways in the upper and lower bounds. Our proofs of the lower bounds are information theoretic in nature, based on Fano's inequality and results on the metric entropy of the balls \BBBq(Rq), whereas our proofs of the upper bounds are constructive, involving direct analysis of least squares over lq-balls. For the special case q=0, corresponding to models with an exact sparsity constraint, our results show that although computationally efficient l1-based methods can achieve the minimax rates up to constant factors, they require slightly stronger assumptions on the design matrix X than optimal algorithms involving least-squares over the l0-ball.