Low-discrepancy perturbation design for numerically stable gradient estimation in feedback optimization
C. Jimenez Cortes, I. Monteiro, S. Coogan, M. Gamarra, M. Thits
IEEE Control Systems Letters, 2026
Abstract
We present a numerically stable gradient estimation algorithm for unknown objective functions in feedback-based optimization settings. Unlike classical extremum seeking control, which recovers gradient components sequentially via distinct demodulation frequencies, our approach estimates all components simultaneously, making it well-suited for high-dimensional and multi-agent systems. The core innovation lies in the use of carefully designed perturbations based on low-discrepancy sequences, which ensure structured and efficient probing. We provide theoretical guarantees of numerical stability, and numerical simulations validate the method's effectiveness in an integrated sensing and communication application.