Fresh Sketches Beat Power Iteration in Randomized Low-Rank Muon

Randomized low-rank variants of Muon orthogonalize a Gaussian sketch of the momentum and, following randomized linear algebra, sharpen it with a power iteration. In language-model pretraining at high momentum, we show that this step makes training worse, give a partial account of why, and remove the wall-clock overhead of randomized Muon on nanoGPT. (1) Removing power iteration improves training. On nanoGPT (80M parameters in the Muon-updated matrices, three seeds) the final perplexity drops from 29.43 to 28.67 with fewer matrix products; in a separate reimplementation on the GPT-2 Medium track (about 200M, two seeds) against a numerically exact h = 1 (subspace iteration in FP32), h = 0 is again better, by 0.018 and 0.020 in final validation loss (pre-registered). On nanoGPT and in one toy setting, a wider sketch is a better use of the same compute. (2) About half of the 80M gap is bf16 rounding. The rest is not a per-step effect: in exact arithmetic, power iteration always improves the single-step alignment of the update with the momentum. (3) We propose an account based on the update averaged over fresh sketches, a diagonal spectral filter that power iteration makes less flat (proved for one sketch column and rank two; certified by interval arithmetic in a deterministic-equivalent model). Consistent with it, freezing the sketch reverses the ranking (three seeds) and averaging each update over sketches favours no power iteration in 31 of 32 toy settings; but a mean-field toy favours power iteration, so flatness alone does not decide the winner. (4) The overhead comes from batched Householder QR of the sketch; a compiled, shifted CholeskyQR3 removes about 100 ms per step. On nanoGPT the link between flatness and loss is correlational, and why the advantage persists at low momentum is open.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24182873
Primary Topic
Stochastic Gradient Optimization Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Fresh Sketches Beat Power Iteration in Randomized Low-Rank Muon

創規 福田
Zenodo (CERN European Organization for Nuclear Research)
Stochastic Gradient Optimization Techniques
preprint

Fresh Sketches Beat Power Iteration in Randomized Low-Rank Muon

創規 福田
preprint en

Abstract

Randomized low-rank variants of Muon orthogonalize a Gaussian sketch of the momentum and, following randomized linear algebra, sharpen it with a power iteration. In language-model pretraining at high momentum, we show that this step makes training worse, give a partial account of why, and remove the wall-clock overhead of randomized Muon on nanoGPT. (1) Removing power iteration improves training. On nanoGPT (80M parameters in the Muon-updated matrices, three seeds) the final perplexity drops from 29.43 to 28.67 with fewer matrix products; in a separate reimplementation on the GPT-2 Medium track (about 200M, two seeds) against a numerically exact h = 1 (subspace iteration in FP32), h = 0 is again better, by 0.018 and 0.020 in final validation loss (pre-registered). On nanoGPT and in one toy setting, a wider sketch is a better use of the same compute. (2) About half of the 80M gap is bf16 rounding. The rest is not a per-step effect: in exact arithmetic, power iteration always improves the single-step alignment of the update with the momentum. (3) We propose an account based on the update averaged over fresh sketches, a diagonal spectral filter that power iteration makes less flat (proved for one sketch column and rank two; certified by interval arithmetic in a deterministic-equivalent model). Consistent with it, freezing the sketch reverses the ranking (three seeds) and averaging each update over sketches favours no power iteration in 31 of 32 toy settings; but a mean-field toy favours power iteration, so flatness alone does not decide the winner. (4) The overhead comes from batched Householder QR of the sketch; a compiled, shifted CholeskyQR3 removes about 100 ms per step. On nanoGPT the link between flatness and loss is correlational, and why the advantage persists at low momentum is open.

Zenodo (CERN European Organization for Nuclear Research)
The University of Tokyo (JP)
Stochastic Gradient Optimization Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon