A consistent algorithm to solve Lasso, elastic-net and Tikhonov regularization (Q2431337)

Recent approaches of supervised learning use tools which have been developed in the theory and practice of inverse problems. They all exploit penalized least squares methods partly under additional constraints. The new paper analyzes an iterative algorithm suggested by \textit{V.~Umanità} and \textit{S.~Villa} [``Elastic-net regularization: iterative algorithms and asymptotic behavior of solutions'', Numer.~Funct.~Anal.~31, No.~12, 1406--1432 (2010; Zbl 1213.47011)] with respect to the consistency of estimating relevant features of the regression function. The function to be determined is assumed to be sparse with respect to a fixed countable dictionary in the sense that the associated coefficient vector is in \(\ell^{1}\). In this context, the authors compare Lasso solutions, elastic-net and Tikhonov regularized solutions and their cross-connections. Moreover, they discuss the character of corresponding optimization problems and their approximate (iterative) solution.