Pathwise coordinate optimization for sparse learning: algorithm and theory (Q1747736)

In the modeling of high-dimensional data with number of variables greatly exceeding the sample size, it is often assumed that only a small subset of variables are relevant, and various regularized approaches have been proposed to estimate the regression coefficients for these variables. Specific optimization algorithms have been developed to find these estimates and, among them, the pathwise coordinate optimization framework has been shown to have a good empirical performance, but no theoretical guarantee has been established. In this paper, a new pathwise calibrated sparse shooting algorithm is proposed to improve the existing pathwise coordinate optimization framework and shown to attain a linear convergence with optimal statistical properties in parameter estimation and support recovery. The proposed algorithm is also compared with other related algorithms in numerical experiments.