Research map / Optimization / Nonconvex and smooth optimization
Convergence of Muon with Newton-Schulz
Gyu Yeol Kim, Min-hwan Oh
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 |
|---|---|---|---|
| use Muon with Newton-Schulz q=1 instead of SGD with momentum | CifarNet on CIFAR-10 · CIFAR-10 | training and test loss | Improved |
| increase Newton-Schulz steps q from 1 to 2 or 3 (ablation) | Muon · CIFAR-10, CifarNet | train/test loss per epoch | Improved |
| replace exact SVD polar factor with Newton-Schulz (q=2 or 3) (ablation) | Muon · CIFAR-10, CifarNet | wall-clock time to a given test loss | Improved |
| increase Newton-Schulz polynomial degree kappa from 1 to 5 (q=3) (ablation) | Muon · CIFAR-10, CifarNet | loss per epoch vs time per step | Mixed |
| orthogonalize momentum (Muon) instead of SGD with momentum (ablation) | optimizer · rank sweep r in {16,...,216} | nuclear gradient norm vs rank: SGD-M slope approx 0.3; Muon nearly flat | Improved |
Closest papers by meaning
- Denoise First, Orthogonalize Later: Understanding Momentum in Muon via Spectral Filtering (NeurIPS 2026 · similarity 0.85)
- Phases of Muon: When Muon Eclipses SignSGD (NeurIPS 2026 · similarity 0.84)
- Muon Dynamics as a Spectral Wasserstein Flow (NeurIPS 2026 · similarity 0.83)
- Muon Does Not Converge on Convex Lipschitz Functions (NeurIPS 2026 · similarity 0.83)
- On the Convergence Analysis of Muon (NeurIPS 2026 · similarity 0.82)
- Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives (NeurIPS 2026 · similarity 0.82)
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.
