关键词: 一阶梯度方法, 动量方法, 个体收敛速率, Heavy-ball方法, 稀疏性

Abstract: The momentum method is widely used as an acceleration technique to improve the convergence of the first-order gradient algorithms. So far, the momentum methods discussed in most literatures are only limited to the accelerated method proposed by Nesterov, but the Heavy-ball momentum method proposed by Polyak is seldom studied. In particular, in the case of non-smooth objective functions, the individual optimal convergence of Nesterov accelerated methods has been derived, and it has high performance in solving sparse optimization problems. In contrast, while it has been proved that the Heavy-ball momentum method has an optimal convergence rate,it is only in terms of the averaged outputs. To our best knowledge, whether it has optimal individual convergence or not still remains unknown. In this paper, we focus on the non-smooth optimizations. We prove that the Heavy-ball momentum method achieves the optimal individual convergence by skillfully selecting the time-varying step-size, which indicates that Heavy-ball momentum is an efficient acceleration strategy for the individual convergence of the projected subgradient methods. As an application, the constrained hinge loss function optimization problems within an l\-1-norm ball are considered. In comparison with other optimization algorithms, the experiments demonstrate the correctness of our theoretical analysis and performance of the proposed algorithms in keeping the sparsity.

Key words: first-order gradient methods, momentum methods, individual convergence rate, heavy-ball methods, sparsity