The Geometry of Constrained Optimization: Constrained Gradient Flows via Reparameterization: A-Stable Implicit Schemes, KKT from Stationarity, and Geometry-Respecting Algorithms

TL;DR

This paper introduces a geometry-respecting framework for constrained optimization using reparameterization, implicit schemes, and KKT conditions, leading to stable and efficient algorithms for various constraint types.

Contribution

It develops a novel reparameterization-based gradient flow framework for constrained problems, connecting stationarity with KKT conditions, and introduces implicit schemes with proven stability and efficiency.

Findings

Algorithms are stable and monotone with no step-size restrictions.

Numerical tests demonstrate improved stability and efficiency.

The framework unifies and extends constrained optimization methods.

Abstract

Gradient-flow (GF) viewpoints unify and illuminate optimization algorithms, yet most GF analyses focus on unconstrained settings. We develop a geometry-respecting framework for constrained problems by (i) reparameterizing feasible sets with maps whose Jacobians vanish on the boundary (orthant/box) or have rank $n-1$ on the simplex (the Fisher--Shahshahani operator), (ii) deriving flows in parameter space that induce feasible primal dynamics, (iii) discretizing with A-stable implicit schemes (backward Euler on vector domains; feasible Cayley on Stiefel) solved by robust inner loops (modified Gauss--Newton or a KL-prox/negative-entropy Newton--KKT solver), and (iv) proving that stationarity of the dynamics implies KKT, with complementary slackness arising from a simple kinematic mechanism (zero normal speed induced by a vanishing Jacobian or by the Fisher--Shahshahani operator on the simplex). We also treat the Stiefel manifold, where Riemannian stationarity coincides with KKT. The theory yields efficient, geometry-respecting algorithms for each constraint class, with monotone descent and no step-size cap. We include a brief A-stability discussion and present numerical tests (NNLS, simplex- and box-constrained least squares, and Stiefel) demonstrating stability, accuracy, and runtime efficiency of the implicit schemes.