Local Linear Convergence of ADMM on Quadratic or Linear Programs

In this paper, we analyze the convergence of the Alternating Direction Method of Multipliers (ADMM) as a matrix recurrence for the particular case of a quadratic program or a linear program. We identify a particular combination of the vector iterates in the standard ADMM iteration that exhibits almost monotonic convergence. We present an analysis which indicates the convergence depends on the eigenvalues of a particular matrix operator. The theory predicts that ADMM should exhibit linear convergence when close enough to the optimal solution, but when far away can exhibit slow “constant step” convergence. This is illustrated with a convergence trace from linear program.