I-LAMM for sparse learning: simultaneous control of algorithmic complexity and statistical error (Q1750288)

This paper proposes a general computational framework for solving nonconvex optimisation problems such as the penalized M-estimator \(\mathrm{argmin}_{\beta\in{\mathbb R}^d}\{ {\mathcal L}(\beta) + {\mathcal R}_{\lambda}(\beta)\}\), where \({\mathcal L}(\beta)\) is a smooth loss function, \({\mathcal R}_{\lambda}(\beta)\) is a sparsity-inducing penalty with a regularization parameter \(\lambda\). The proposed strategy enables the simultaneous control of the algorithmic complexity and the statistical error when fitting high-dimensional models appearing in various problems including low rank matrix completion problems, high-dimensional graphical models and quantile regression.