Research map / Learning theory and algorithms / Combinatorial and search optimization
A Penalty Approach For Differentiation Through Black-box Quadratic Programming Solvers
Yuxuan Linghu, Zhiyuan Liu, Qi Deng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace KKT-based backward pass (dQP) with smoothed-penalty primal SPD solve (dXPP) | probability simplex projection, size 1000000 · backward runtime | backward time (ms): 950.64 -> 226.21 (4.2x speedup) | Improved |
| replace dQP backward with dXPP | chain projection, size 1000000 · backward runtime | backward time (ms): 3071.18 -> 333.42 (9.2x speedup) | Improved |
| replace dQP with dXPP in dense mode multi-period portfolio optimization | horizon H=200 · per-epoch backward runtime | backward time (ms): 39105.75 -> 113.97 | Improved |
| dXPP versus OptNet on multi-period portfolio | horizon H=200 · per-epoch backward runtime | backward time (ms): 605.89 -> 113.97 (5.3x) | Improved |
| dXPP versus OptNet at very small problem size | simplex projection, size 20 · backward runtime | backward time (ms): OptNet 0.54 vs dXPP 0.72 | Hurt |
| use dXPP gradients instead of dQP gradients on random QPs | random strictly convex QPs · gradient accuracy | relative gradient difference: 1.91 × 10−7 (10x5) to 5.14 × 10−4 (5000x2000) | No clear effect |
| choose intermediate smoothing scale rather than very large or very small (ablation) | dXPP · random QPs gradient accuracy | relative gradient difference: 5000x2000: 3.18×10−6 (1e-5), 2.62×10−7 (1e-7), 1.63×10−5 (1e-9) | Mixed |
| prune inactive-constraint terms (omit E_delta and F_delta) in the backward pass (ablation) | dXPP · performance | gradient quality | No clear effect |
| use dXPP forward pass with Gurobi instead of dQP in dense mode | multi-period portfolio · total runtime | total time (ms): 40066.11 -> 977.98 at H=200 | Improved |
Closest papers by meaning
- Differentiation Through Black-Box Quadratic Programming Solvers (NeurIPS 2025 · similarity 0.92)
- Leveraging augmented-Lagrangian techniques for differentiating over infeasible quadratic programs in machine learning (ICLR 2024 · similarity 0.85)
- Deep Distributed Optimization for Large-Scale Quadratic Programming (ICLR 2025 · similarity 0.84)
- Accelerating Optimization via Differentiable Stopping Time (NeurIPS 2025 · similarity 0.84)
- Automatic Differentiation of Optimization Algorithms with Time-Varying Updates (ICML 2025 · similarity 0.84)
- BayeSQP: Bayesian Optimization through Sequential Quadratic Programming (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
