COMBINED PRIMAL-DUAL AND PENALTY METHODS FOR CONSTRAINED MINIMIZATION.

A class of combined primal-dual and penalty methods often called methods of multipliers is considered. The analysis focuses mainly on the rate of convergence of these methods. It is shown that this rate is considerably more favorable than the corresponding rate for penalty function methods. Some efficient versions of multiplier methods are also considered whereby the intermediate unconstrained minimizations involved are approximate and only asymptotically exact. It is shown that such approximation schemes may lead to a substantial deterioration of the convergence rate, and a special approximation scheme is proposed which exhibits the same rate as the method with exact minimization. Finally, the properties of the step size rule of the multiplier method in relation to other possible step sizes are analyzed, and a modified step size rule is considered for the case of the convex programming problem.