Self Equivalence of the Alternating Direction Method of Multipliers
Abstract
In this paper, we show interesting self equivalence results for the alternating direction method of multipliers (ADM or ADMM). Specifically, we show that ADM on a primal problem is equivalent to ADM on its Lagrange dual problem; ADM is equivalent to a primal-dual algorithm applied to a saddle- point formulation of the problem; when one of the two objective functions is quadratic with an a ne domain, we can swap the update order of the two variables in ADM and obtain an equivalent algorithm. An example in extended monotropic programming is given to demonstrate that the primal-dual algorithm may be preferable over the other equivalent algorithms for its lower per-iteration complexity and, in the setting of distributed computation, better load balancing.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 11, 2014
- Accession Number
- ADA610274
Entities
People
- Ming Yan
- Wotao Yin
Organizations
- University of California, Los Angeles