{"id":"arxiv/2609.39114","url":"https://arxiv.org/abs/2609.39114","tags":["cs.LG","math.OC"],"score":0,"title":"The Row Normalization Puzzle in Muon","domain":"cs.LG","source":"arxiv","authors":["Jiayu Zhang","Tianyi Lin"],"reviews":0,"subject":"cs.LG","summary":"This paper examines how row-wise renormalization affects Muon, focusing on the gap between NorMuon's worst-case guarantees and its practical performance (Li et al.). Despite its growing adoption and promising performance in large language model (LLM) pretraining, NorMuon's worst-case guarantees remain poorly understood. One fundamental question is: Does row normalization yield provable convergence gains, potentially through its interaction with approximate polar computation and exponential moving-average momentum? Our results show that row normalization introduces a dimension-dependent…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-09-30","reactions":{}}