A sharp factor d/2 for diagonal preconditioning of single-sample SGD

Consider single-sample stochastic gradient descent with a constant step on noiseless Gaussian least squares with covariance Σ, preconditioned by a positive diagonal matrix P. Let γ* be the exact asymptotic decay rate of the expected loss at the best constant step, and let R(Σ) be the ratio of the rates of the two Gauss–Newton diagonal preconditioners diag(Σ)^(−1/2) and diag(Σ)^(−1). We prove that for every dimension d ≥ 3 and every Σ, R(Σ) < d/2, and that sup_Σ R(Σ) = d/2; the supremum is not attained. Equivalently, no diagonal preconditioner improves the exact rate of diag(Σ)^(−1) (Jacobi scaling) by a factor d/2 or more, and the factor d/2 is sharp. This is in contrast with the spectral condition number, where the corresponding worst-case factor of Jacobi scaling equals n/2 only for 2 ≤ n ≤ 5. The proof is computer-assisted. The reduction to a secular system and the lower-bound family are proved by hand, and so is the outer regime for d ≥ 26, up to two one-variable inequalities between rational functions that are proved in exact rational arithmetic. The regimes λ_min(C) → 0, λ_min(C) → 1 and d → ∞ are reduced by hand arguments to finitely many exact or interval checks on boxes that contain the limit; one of them, a dimension-free certificate for a floor on the fourth eigenvalue, serves every d ≥ 4. The remaining compact regions, in dimensions d ≤ 59, are certified with outward-rounded interval arithmetic on branch-and-bound trees whose leaves are checked to partition the domain by an exact volume count, together with exact rational identities and root counts. For d ≥ 60 we also prove the quantitative bound R(Σ) < (d/2)(1 − (1/20)·√min(λ_min(C), 1/(5d))), whose order √λ_min(C) is exact, and we give an explicit family showing that for large mini-batches the factor d/2 fails from d = 13 on. Files: the paper (PDF) and sources_and_supplement.zip (LaTeX source, bibliography, build script, the interval-arithmetic certificates with their logs and the exact hand checks). Companion papers: "On two questions of Shapiro on Jacobi scaling", doi:10.5281/zenodo.24315181, and version 4 of the WEDGE preprint, doi:10.5281/zenodo.24315153.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24315182
Citations
1
Primary Topic
Stochastic Gradient Optimization Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A sharp factor d/2 for diagonal preconditioning of single-sample SGD

Gabriel Costa de Mello Paiva
1 citations
Zenodo (CERN European Organization for Nuclear Research)
Stochastic Gradient Optimization Techniques
preprint

A sharp factor d/2 for diagonal preconditioning of single-sample SGD

Gabriel Costa de Mello Paiva
preprint en
1 citations

Abstract

Consider single-sample stochastic gradient descent with a constant step on noiseless Gaussian least squares with covariance Σ, preconditioned by a positive diagonal matrix P. Let γ* be the exact asymptotic decay rate of the expected loss at the best constant step, and let R(Σ) be the ratio of the rates of the two Gauss–Newton diagonal preconditioners diag(Σ)^(−1/2) and diag(Σ)^(−1). We prove that for every dimension d ≥ 3 and every Σ, R(Σ) < d/2, and that sup_Σ R(Σ) = d/2; the supremum is not attained. Equivalently, no diagonal preconditioner improves the exact rate of diag(Σ)^(−1) (Jacobi scaling) by a factor d/2 or more, and the factor d/2 is sharp. This is in contrast with the spectral condition number, where the corresponding worst-case factor of Jacobi scaling equals n/2 only for 2 ≤ n ≤ 5. The proof is computer-assisted. The reduction to a secular system and the lower-bound family are proved by hand, and so is the outer regime for d ≥ 26, up to two one-variable inequalities between rational functions that are proved in exact rational arithmetic. The regimes λ_min(C) → 0, λ_min(C) → 1 and d → ∞ are reduced by hand arguments to finitely many exact or interval checks on boxes that contain the limit; one of them, a dimension-free certificate for a floor on the fourth eigenvalue, serves every d ≥ 4. The remaining compact regions, in dimensions d ≤ 59, are certified with outward-rounded interval arithmetic on branch-and-bound trees whose leaves are checked to partition the domain by an exact volume count, together with exact rational identities and root counts. For d ≥ 60 we also prove the quantitative bound R(Σ) < (d/2)(1 − (1/20)·√min(λ_min(C), 1/(5d))), whose order √λ_min(C) is exact, and we give an explicit family showing that for large mini-batches the factor d/2 fails from d = 13 on. Files: the paper (PDF) and sources_and_supplement.zip (LaTeX source, bibliography, build script, the interval-arithmetic certificates with their logs and the exact hand checks). Companion papers: "On two questions of Shapiro on Jacobi scaling", doi:10.5281/zenodo.24315181, and version 4 of the WEDGE preprint, doi:10.5281/zenodo.24315153.

Zenodo (CERN European Organization for Nuclear Research)
Elizabethtown Community Hospital (US)
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