On the prediction performance of the Lasso (Q502891)

The prediction performance of the Lasso is studied for Gaussian multiple regression models with deterministic design. The term \textit{prediction performance} is understood as the magnitude of the risk measured by the prediction loss \(\frac{1}{n}\|X(\hat\beta-\beta)\|^2_2\), where \(n\) is the sample size, \(X\) is the design matrix, \(\hat\beta\) is the Lasso estimator, and \(\beta\) is the regression vector. In the paper under review, a simple measure of the correlations of the covariates is introduced. If this measure is incorporated in the choice of the tuning parameter, the Lasso prediction risk decays at a fast rate for a broad variety of settings including ones with strongly correlated covariates. However, for moderately correlated covariates, the prediction performance of the Lasso can be mediocre irrespective of the choice of the tuning parameter. The obtained results lead to near-optimal rates for the least-squares estimator with total variation penalty.