A Generalized $\ell_1$-Merit Function SQP Method Using Function Approximations with Tunable Accuracy (2507.06199v1)

Abstract: This paper develops a generalization of the line-search sequential quadratic programming (SQP) algorithm with $\ell_1$-merit function that uses objective and constraint function approximations with tunable accuracy to solve smooth equality-constrained optimization problems. The evaluation of objective and constraint functions and their gradients is potentially computationally expensive, but it is assumed that one can construct effective, computationally inexpensive models of these functions. This paper specifies how these models can be used to generate new iterates. At each iteration, the models have to satisfy function error and relative gradient error tolerances determined by the algorithm based on its progress. Moreover, bounds for the model errors are used to explore regions where the combined objective function and constraint models are sufficiently accurate. The algorithm has the same first-order global convergence properties as a line-search SQP algorithm with $\ell_1$-merit function, but only uses objective and constraint function models and the model error bounds. The algorithm is applied to a discretized boundary control problem in which the evaluation of the objective and constraint functions requires the solution of the Boussinesq partial differential equation (PDE). The models are constructed from projection-based reduced-order models of the Boussinesq PDE.