1: \begin{abstract}

2: Hard Thresholding Pursuit (HTP) is an iterative greedy selection

3: procedure for finding sparse solutions of underdetermined linear

4: systems. This method has been shown to have strong theoretical

5: guarantee and impressive numerical performance. In this paper, we

6: generalize HTP from compressive sensing to a generic problem setup

7: of sparsity-constrained convex optimization. The proposed algorithm

8: iterates between a standard gradient descent step and a hard

9: thresholding step with or without debiasing. We prove that our

10: method enjoys the strong guarantees analogous to HTP in terms of

11: rate of convergence and parameter estimation accuracy. Numerical

12: evidences show that our method is superior to the state-of-the-art

13: greedy selection methods in sparse logistic regression and sparse

14: precision matrix estimation tasks.

15: \end{abstract}

16: