aiqbench analysis
This research introduces dXPP, a penalty-based framework that decouples quadratic programming solving from differentiation to overcome the scalability and robustness limits of traditional KKT methods. By implicitly differentiating through smooth approximations rather than explicit systems, it achieves significant speedups on large-scale tasks like sparse projections and portfolio optimization while remaining solver-agnostic. The approach advances core machine learning infrastructure by enabling more efficient training pipelines for problems involving constrained convex optimization.
A Penalty Approach for Differentiation Through Black-Box Quadratic Programming Solvers
arXiv:2602.14154v3 Announce Type: replace Abstract: Differentiating through the solution of a quadratic program (QP) is a central problem in differentiable optimization. Most existing approaches differentiate through the Karush--Kuhn--Tucker (KKT) system, but their computational cost and numerical robustness can degrade at scale. To address these limitations, we propose dXPP, a penalty-based differentiation framework that decouples QP solving from differentiation. In the solving step (forward pass), dXPP is solver-agnostic and can leverage any black-box QP solver. In the differentiation step (backward pass), we map the solution to a smooth approximate penalty problem and implicitly differentiate through it, requiring only the solution of a much smaller linear system in the primal variables. This approach bypasses the difficulties inherent in explicit KKT differentiation and significantly improves computational efficiency and robustness. We evaluate dXPP on various tasks, including randomly generated QPs, large-scale sparse projection problems, and a real-world multi-period portfolio optimization task. Empirical results demonstrate that dXPP is competitive with KKT-based differentiation methods and achieves substantial speedups on large-scale problems. Our implementation is open source and available at https://github.com/mmmmmmlinghu/dXPP.
Abstract:Differentiating through the solution of a quadratic program (QP) is a central problem in differentiable optimization. Most existing approaches differentiate through the Karush--Kuhn--Tucker (KKT) system, but their computational cost and numerical robustness can degrade at scale. To address these limitations, we propose dXPP, a penalty-based differentiation framework that decouples QP solving from differentiation. In the solving step (forward pass), dXPP is solver-agnostic and can leverage any black-box QP solver. In the differentiation step (backward pass), we map the solution to a smooth approximate penalty problem and implicitly differentiate through it, requiring only the solution of a much smaller linear system in the primal variables. This approach bypasses the difficulties inherent in explicit KKT differentiation and significantly improves computational efficiency and robustness. We evaluate dXPP on various tasks, including randomly generated QPs, large-scale sparse projection problems, and a real-world multi-period portfolio optimization task. Empirical results demonstrate that dXPP is competitive with KKT-based differentiation methods and achieves substantial speedups on large-scale problems. Our implementation is open source and available at this https URL.
Machine Learning (cs.LG); Optimization and Control (math.OC)
arXiv:2602.14154 [cs.LG]
(or arXiv:2602.14154v3 [cs.LG] for this version)
Full text under the publisher's open licence · 186 words
Related coverage
The Elements of Differentiable Programming
arXiv · Aug 5, 2026
Compiling to recurrent neurons
arXiv · Jul 24, 2026
Compile Once, Differentiate Everywhere: A Differentiable Meta-Circular Interpreter
arXiv · Jun 10, 2026
Recursive Improvement of a Differentiable Scientific Software Ecosystem
arXiv · Oct 7, 2026