11348e03e23b137d55d94464250a67a2-AuthorFeedback.pdf

The proposed algorithm inthis paper istoimprovethe solution efficiencyofthe sparse learning problems given by3 equation (1) in the main file. As discussed at the beginning of the supplemental file, Thunder outperforms existing4 solversismainly because ofthepassivefeature recruiting strategies, sampling method forfeature recruiting, andthe5 safestopcondition regardingfeature recruiting employedbythealgorithm. The correlation between features may affect the efficiencyofThunder,butitdoes not impact the algorithm'ssafety.15 Here safety means the algorithm final step active set does not miss any features in the optimal active set A of the16 problem. If the condition in Lemma 1is not18 met, Algorithm 2will notstop feature recruiting.