Iteration-complexity of first-order augmented Lagrangian methods for convex programming (Q5962727)

The authors consider the smooth convex programming problem containing linear equality and general convex constraints. They apply the usual augmented Lagrangian method with respect to linear equality constraints with an inexact solution of each inner problem obtained from the application of the Nesterov gradient method. The total complexity estimates are established for the case where the goal function has the Lipschitz gradient, they appear to be better than those for the utilization of the quadratic penalty method instead of the augmented Lagrangian one. Further improvement of estimates is obtained by the regularization of the goal function.