16
17

Out-of-sample error estimate for robust M-estimators with convex penalty

Abstract

A generic out-of-sample error estimate is proposed for robust MM-estimators regularized with a convex penalty in high-dimensional linear regression where (X,y)(X,y) is observed and p,np,n are of the same order. If ψ\psi is the derivative of the robust data-fitting loss ρ\rho, the estimate depends on the observed data only through the quantities ψ^=ψ(yXβ^)\hat\psi = \psi(y-X\hat\beta), Xψ^X^\top \hat\psi and the derivatives (/y)ψ^(\partial/\partial y) \hat\psi and (/y)Xβ^(\partial/\partial y) X\hat\beta for fixed XX. The out-of-sample error estimate enjoys a relative error of order n1/2n^{-1/2} in a linear model with Gaussian covariates and independent noise, either non-asymptotically when p/nγp/n\le \gamma or asymptotically in the high-dimensional asymptotic regime p/nγ(0,)p/n\to\gamma'\in(0,\infty). General differentiable loss functions ρ\rho are allowed provided that ψ=ρ\psi=\rho' is 1-Lipschitz. The validity of the out-of-sample error estimate holds either under a strong convexity assumption, or for the 1\ell_1-penalized Huber M-estimator if the number of corrupted observations and sparsity of the true β\beta are bounded from above by sns_*n for some small enough constant s(0,1)s_*\in(0,1) independent of n,pn,p. For the square loss and in the absence of corruption in the response, the results additionally yield n1/2n^{-1/2}-consistent estimates of the noise variance and of the generalization error. This generalizes, to arbitrary convex penalty, estimates that were previously known for the Lasso.

View on arXiv
Comments on this paper