Constrained optimization problems with nonconvex objectives and constraints are common in machine learning and engineering, but their primal-dual algorithms are often analyzed under strong assumptions. This paper studies a class where both the objective and inequality constraints are compositions of convex Lipschitz outer functions with smooth nonlinear mappings. The authors show that, contrary to prior work, multiplier regularity need not hold globally or along the entire iterate sequence; it is enough that it holds in a neighborhood of the feasible set.
The analysis proceeds through an auxiliary minimax problem with a compact dual set. Lyapunov estimates derived from this problem guarantee that iterates become nearly feasible when initialized feasibly and the penalty parameter is sufficiently large, without assuming multiplier boundedness. Local regularity then bounds the multipliers and ensures that dual truncation is inactive, yielding approximate KKT certificates for the original problem. For the smoothed prox-linear augmented Lagrangian method with dual regularization, the authors prove a KKT residual rate of O(K^{-1/3}).
When dual regularization is omitted, the paper shows that piecewise linear outer functions plus additional local structural assumptions lead to a local dual error bound and a sharper O(K^{-1/2}) rate. This distinction highlights a practical trade-off between algorithmic simplicity and convergence speed. The results extend the reach of first-order methods to a broader class of nonconvex constrained problems, while keeping the required regularity assumptions localized and verifiable.