Journal of the Operations Research Society of China ›› 2018, Vol. 6 ›› Issue (4): 485-506.doi: https://doi.org/10.1007/s40305-017-0186-y

Special Issue: Continuous Optimization

• Continuous Optimization •     Next Articles

Block-Wise ADMM with a Relaxation Factor for Multiple-Block Convex Programming

Bing-Sheng He1,2 ,Ming-Hua Xu3 ,Xiao-Ming Yuan4   

  1. 1 Department of Mathematics, South University of Science and Technology, Shenzhen 518055,Guangdong, China
    2 Department of Mathematics, Nanjing University, Nanjing 210093, China
    3 School of Mathematics and Physics, Changzhou University, Changzhou 213164, Jiangsu, China
    4 Department of Mathematics, The University of Hong Kong, Hong Kong, China
  • Online:2018-12-30 Published:2018-12-30
  • Supported by:

    Bing-Sheng He and Ming-Hua Xu were supported by the National Natural Science Foundation of China (No. 11471156).
    Xiao-Ming Yuan was supported by the General Research Fund from Hong Kong Research Grants Council (No. HKBU 12313516).

Abstract:

It has been shown that the alternating direction method of multipliers (ADMM) is not necessarily convergent when it is directly extended to a multiple-block linearly constrained convex minimization model with an objective function that is in the sum of more than two functions without coupled variables. Recently, we proposed the block-wise ADMM, which was obtained by regrouping the variables and functions of such a model as two blocks and then applying the original ADMM in block-wise. This note is a further study on this topic with the purpose of showing that a well-known relaxation factor proposed by Fortin and Glowinski for iteratively updating the Lagrangian multiplier of the original ADMM can also be used in the block-wise ADMM. We thus propose the block-wise ADMM with Fortin and Glowinski’s relaxation factor for the multiple-block convex minimization model. Like the block-wise ADMM, we also suggest further decomposing the resulting subproblems and regularizing them by proximal terms to ensure the convergence. For the block-wise ADMM with Fortin and Glowinski’s relaxation factor, its convergence and worst-case convergence rate measured by the iteration complexity in the ergodic sense are derived.

Key words: Convex programming ·, Operator splitting methods ·, Alternating direction method of multipliers ·, Fortin and Glowinski’s relaxation factor