1: \begin{abstract}

2: In this paper, based on the limited memory techniques and subspace minimization conjugate gradient (SMCG)  methods, a regularized limited memory subspace minimization conjugate gradient method is proposed, which contains two types of iterations. In SMCG iteration, we obtain the search direction by minimizing the approximate quadratic model or approximate regularization model. In RQN iteration, combined with regularization technique and BFGS method, a modified regularized quasi-Newton method is used in the subspace to improve the orthogonality. Moreover, some simple acceleration criteria and an improved tactic for selecting the initial stepsize  to enhance the efficiency of the algorithm are designed. Additionally, an generalized nonmonotone line search is utilized and the global convergence of our proposed algorithm is established under mild conditions.  Finally, numerical results show that, the proposed algorithm has a significant improvement over ASMCG\_PR  and is superior to the particularly well-known limited memory conjugate gradient software packages CG\_DESCENT (6.8) and CGOPT(2.0) for the CUTEr library.

3: 

4: 

5: \keywords{ Limited memory \and Subspace minimization conjugate gradient method  \and Orthogonality \and Regularization model \and Quasi-Newton method }

6: \subclass{ 49M37 \and 65K05  \and 90C30 }

7: \end{abstract}