Math 2.0

Resources / preprint

From Geometry to Generalization: Why Row Normalization Can Beat Adam and Muon TTEAQA1XXR

URIm2://res/TTEAQA1XXR
Kindpreprint
CreatorsJihwan Kim, Dogyoon Song, Chulhee Yun
Date2026-10-08
Subjectsarxiv cs.LG, arxiv cs.AI, arxiv math.OC, arxiv stat.ML
Scopeborder

Expressions

e1: version, preprint

Date2026-10-08
IdsarXiv 2610.11309

Items in this resource

The mentions of the inventory that name this resource. Each name links to the card of its inventory row.

ItemKindWhereSource
Normalization of a cumulative margin boundstatementlemma B.4arxiv
Uniform concentration of frozen exponential averagesstatementlemma E.3arxiv
Deterministic row-norm rigiditystatementproposition C.6arxiv
Uniform radius-to-margin ratiosstatementproposition D.4arxiv
Residual-exogenous optimizer limitsstatementlemma F.2arxiv
Stochastic optimizer accuracy orderingstatementproposition 3.4arxiv
All-aspect spectral inputstatementproposition C.12arxiv
Full-batch max--sum dominated margin convergencestatementtheorem B.6arxiv
Max-norm boundary separationstatementproposition C.10arxiv
Adam-step and signal stabilitystatementlemma E.2arxiv
All-aspect deterministic margin geometrystatementtheorem C.2arxiv
Uniform clipped-Gaussian angular barrierstatementlemma E.7arxiv
Row-normalized full-batch margin convergencestatementproposition B.14arxiv
Freezing one epochstatementlemma E.4arxiv
Geometry-adapted random-reshuffling momentum controlstatementproposition B.11arxiv
Geometry-adapted softmax and gradient stabilitystatementlemma B.9arxiv
Cumulative remainders for general learning ratesstatementlemma B.3arxiv
AdamProxy margin characterizationstatementproposition E.10arxiv
Population-error continuity along a direction setstatementlemma C.21arxiv
One-epoch \(\beta_1,\beta_2\to1\) RR-Adam approximationstatementproposition 3.5arxiv
Finite-dimensional geometry-to-error transferstatementproposition C.19arxiv
Deterministic optimizer accuracy orderingstatementtheorem 3.1arxiv
Opposite Row--Max orderings before and after a common Haar rotationstatementproposition G.7arxiv
RR-RowSGDM versus AdamProxy accuracy orderingstatementproposition 3.6arxiv
Selected mean geometrystatementlemma F.3arxiv
Uniform asymptotics of the clipped problemstatementlemma C.9arxiv
Full-batch Adam under the present learning-rate assumptionsstatementtheorem B.7arxiv
C2M1SV5VVJstatementtheoremarxiv
Uniform stochastic near-margin geometrystatementproposition D.5arxiv
Exact and near spectral minimizersstatementproposition C.15arxiv
Pairwise Gaussian error boundsstatementlemma C.17arxiv
Test-noise energies for selected directionsstatementlemma F.4arxiv
Random-reshuffling row-wise normalized SGDM margin convergencestatementproposition B.15arxiv
Population-error exponents of deterministic optimizersstatementproposition C.27arxiv
Weighted trace gap and hard-edge truncationstatementlemma C.13arxiv
Effective margin implies direction-set convergencestatementlemma B.1arxiv
Consequences of max--sum dominationstatementlemma B.5arxiv
Exact margin implies direction-set convergencestatementlemma B.2arxiv
Spectral-norm boundary separationstatementproposition C.16arxiv
Deterministic population-error exponentsstatementproposition C.26arxiv
Near-minimizer stabilitystatementproposition D.3arxiv
Accuracy ordering under anisotropy with independent orientationsstatementtheorem 4.2arxiv
Uniform row-norm alignmentstatementproposition C.7arxiv
Directed-pair exponent controls multiclass errorstatementlemma C.24arxiv
Stretched-exponential common inputstatementproposition C.4arxiv
Finite-population batch moments and relative scalestatementlemma E.1arxiv
Geometry-adapted cross-entropy boundsstatementlemma B.8arxiv
Selected-direction orderingstatementproposition F.5arxiv
Positive margin growth and a trajectory-level coordinate boundstatementproposition E.6arxiv
Quantitative sign statistics under a Haar rotationstatementlemma G.5arxiv
Capped AdamProxy trajectories and boundary separationstatementproposition E.8arxiv
Empirical-to-population polar transferstatementlemma C.23arxiv
Geometry-to-accuracy transferstatementtheorem C.20arxiv
Boundary metric under direction-set convergencestatementlemma C.22arxiv
Gaussian error transferstatementlemma F.1arxiv
Near-margin reductionstatementlemma D.1arxiv
Deterministic whole-face stabilitystatementlemma G.2arxiv
Scale-invariant reductionstatementlemma C.3arxiv
Continuity of the Adam ratiostatementlemma E.5arxiv
Quantitative whole-face geometry in aligned coordinatesstatementlemma G.3arxiv
Quantitative Gaussian hard-edge count at the square aspect ratiostatementlemma C.11arxiv
Sharp directed signals on the Max and Spectral facesstatementproposition C.25arxiv
Exact clipped problemstatementproposition C.8arxiv
Average normalization transferstatementlemma C.14arxiv
Normalization stabilitystatementlemma C.5arxiv
Full-batch Muon versus Adamstatementcorollary C.28arxiv
TA2S520HF9statementtheoremarxiv
Combined Gaussian-cloud eventstatementlemma D.2arxiv
Generic common rotation restores the Row coefficient advantagestatementlemma G.6arxiv
Geometry-adapted random-reshuffling momentum marginstatementcorollary B.12arxiv
Coordinate dependence of the RowGDM--Adam orderingstatementproposition 4.5arxiv
Quantitative Gaussian radial eventstatementlemma C.18arxiv
Gaussian coefficient comparisonstatementlemma G.1arxiv
An exact Row--Max transition under aligned test anisotropystatementproposition G.4arxiv
Spectral dispersion implies a transitionstatementcorollary G.8arxiv
High-momentum Row near-set controlstatementlemma E.9arxiv
Y48EP91BD4definitiondefinitionarxiv
Population-error exponentsstatementcorollary 3.3arxiv
Random-reshuffling noise and cancellationstatementlemma B.10arxiv
Row normalization is row-norm steepest descentstatementproposition B.13arxiv
Exact invariance under common feature rescalingstatementlemma C.1arxiv

Catalogue records