A Penalty Approach for Differentiation Through Black-Box Quadratic Programming Solvers
Researchers have introduced dXPP, a novel penalty-based framework designed to improve the differentiation process through black-box quadratic programming (QP) solvers. This method decouples the QP solving from the differentiation step, offering greater computational efficiency and numerical robustness compared to traditional approaches that rely on the Karush--Kuhn--Tucker (KKT) system. Empirical evaluations on various QP tasks, including portfolio optimization, show dXPP to be competitive with existing methods while achieving significant speedups on large-scale problems. AI