[Paper Review] Implementing a smooth exact penalty function for equality-constrained nonlinear optimization

Ron Estrin, Michael P. Friedlander+2 more|arXiv (Cornell University)|Oct 9, 2019

Matrix Theory and Algorithms47 references18 citations

TL;DR

This paper presents a computationally efficient implementation of Fletcher's smooth exact penalty function for equality-constrained nonlinear optimization. By leveraging structured linear system solves and adaptive regularization, it achieves R-quadratic convergence while demonstrating practicality on both standard nonlinear programs and PDE-constrained problems, challenging the long-held belief that such penalty functions are computationally prohibitive.

ABSTRACT

We develop a general equality-constrained nonlinear optimization algorithm based on a smooth penalty function proposed by Fletcher (1970). Although it was historically considered to be computationally prohibitive in practice, we demonstrate that the computational kernels required are no more expensive than other widely accepted methods for nonlinear optimization. The main kernel required to evaluate the penalty function and its derivatives is solving a structured linear system. We show how to solve this system efficiently by storing a single factorization each iteration when the matrices are available explicitly. We further show how to adapt the penalty function to the class of factorization-free algorithms by solving the linear system iteratively. The penalty function therefore has promise when the linear system can be solved efficiently, e.g., for PDE-constrained optimization problems where efficient preconditioners exist. We discuss extensions including handling simple constraints explicitly, regularizing the penalty function, and inexact evaluation of the penalty function and its gradients. We demonstrate the merits of the approach and its various features on some nonlinear programs from a standard test set, and some PDE-constrained optimization problems.

Motivation & Objective

• To challenge the long-standing perception that Fletcher's smooth exact penalty function is computationally infeasible for practical use.
• To demonstrate that the core computational kernels—solving a structured saddle-point system—are no more expensive than those in widely used methods like sequential quadratic programming.
• To enable application of the penalty function to large-scale problems, especially PDE-constrained optimization, via efficient direct and iterative linear solvers.
• To develop practical enhancements such as regularization for rank-deficient constraints and inexact evaluation strategies.
• To establish global and local convergence properties under mild assumptions, including R-quadratic convergence with adaptive penalty parameter update.

Proposed method

• Proposes a smooth exact penalty function φσ(x) derived from minimizing a quadratic approximation of the Lagrangian, with multiplier estimates yσ(x) computed via a least-squares problem.
• Identifies the solution of a structured saddle-point system as the key computational kernel, which can be efficiently solved using a single factorization per iteration when matrices are explicit.
• Adapts the method for factorization-free algorithms by solving the linear system iteratively, enabling use with Krylov subspace methods and preconditioners.
• Introduces a regularized variant φσ(x; δ) to handle degenerate constraint Jacobians and improve numerical stability.
• Employs an adaptive penalty parameter strategy where δk is updated based on gradient norms, ensuring R-quadratic convergence.
• Uses inexact Newton methods with trust-region-like globalization, allowing inexact evaluation of gradients and Hessian-vector products.

Experimental results

Research questions

• RQ1 Can Fletcher's smooth exact penalty function be implemented efficiently enough to be practical for large-scale nonlinear optimization?
• RQ2 Is the computational cost of evaluating the penalty function and its derivatives comparable to that of established methods like sequential quadratic programming?
• RQ3 Can the penalty function be adapted for use in factorization-free optimization frameworks, particularly for PDE-constrained problems with efficient preconditioners?
• RQ4 Does the proposed regularization strategy ensure convergence and stability when the constraint Jacobian is rank-deficient?
• RQ5 Can R-quadratic convergence be achieved with an adaptive penalty parameter update rule that avoids exact Hessian computation?

Key findings

• The penalty function φσ(x) is exact: local minimizers of the original problem are minimizers of φσ for all σ > σ∗, ensuring global convergence.
• The main computational cost—solving a structured saddle-point system—is no more expensive than standard methods when matrices are explicitly available and factored once per iteration.
• For large-scale problems, iterative solution of the linear system enables efficient implementation, especially when good preconditioners exist, such as in PDE-constrained optimization.
• The regularized penalty function φσ(x; δ) ensures C2 smoothness and positive definiteness of the Hessian near the solution, enabling superlinear convergence.
• An adaptive penalty parameter update strategy achieves R-quadratic convergence, with ∥xk − x∗∥ = O(δk) and δk+1 = O(δk²), even when gradients are evaluated inexactly.
• Numerical experiments on standard test problems and PDE-constrained optimization problems confirm the method’s robustness, efficiency, and fast local convergence.