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

A generic out-of-sample error estimate is proposed for robust -estimators regularized with a convex penalty in high-dimensional linear regression where is observed and are of the same order. If is the derivative of the robust data-fitting loss , the estimate depends on the observed data only through the quantities , and the derivatives and for fixed . The out-of-sample error estimate enjoys a relative error of order in a linear model with Gaussian covariates and independent noise, either non-asymptotically when or asymptotically in the high-dimensional asymptotic regime . General differentiable loss functions are allowed provided that is 1-Lipschitz. The validity of the out-of-sample error estimate holds either under a strong convexity assumption, or for the -penalized Huber M-estimator if the number of corrupted observations and sparsity of the true are bounded from above by for some small enough constant independent of . For the square loss and in the absence of corruption in the response, the results additionally yield -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