Asymptotic behavior of eigenvalues of large rank perturbations of large random matrices

Ievgenii Afanasiev; Leonid Berlyand; Mariia Kiyashko

arxiv: 2507.12182 · v4 · submitted 2025-07-16 · 🧮 math-ph · cs.LG· math.MP· math.PR

Asymptotic behavior of eigenvalues of large rank perturbations of large random matrices

Ievgenii Afanasiev , Leonid Berlyand , Mariia Kiyashko This is my paper

Pith reviewed 2026-05-19 04:36 UTC · model grok-4.3

classification 🧮 math-ph cs.LGmath.MPmath.PR

keywords deformed Wigner matriceseigenvalue asymptoticsoutlier eigenvaluesfull-rank perturbationsrandom matrix theorydeep neural networksmatrix pruning

0 comments

The pith

Asymptotic analysis tracks outlier eigenvalues for full-rank perturbations of large random matrices.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines deformed Wigner random matrices that arise when modeling trained neural network weights as a random matrix R plus a correlated perturbation S. It develops asymptotic descriptions specifically for the case where S is full rank and carries a growing number of outlier eigenvalues. This setup matters because the spectrum of R + S underpins random-matrix-based pruning methods for deep networks, and the full-rank case with many outliers covers more realistic weight matrices than earlier low-rank assumptions.

Core claim

For a large random Wigner matrix perturbed by a full-rank correlated matrix S whose spectrum contains an increasing number of outliers, the eigenvalues admit explicit asymptotic expansions that locate both the bulk and the separated outlier eigenvalues as matrix dimension tends to infinity.

What carries the argument

Asymptotic analysis of the eigenvalue equation for the deformed matrix R + S, where R is Wigner and S is full-rank with a growing number of isolated eigenvalues.

If this is right

The bulk spectrum and outlier locations become predictable from the statistics of S alone.
Pruning decisions based on random-matrix criteria gain a rigorous justification in the full-rank regime.
The same asymptotic machinery applies when the number of outliers grows slowly with matrix size.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same perturbation analysis may extend to other high-dimensional data matrices that mix randomness with structured low-rank or full-rank components.
Empirical spectra of actual trained networks could be compared directly against these asymptotics to test the R + S modeling assumption.
If the asymptotics hold, they supply a parameter-free way to predict how many eigenvalues survive pruning at a given threshold.

Load-bearing premise

Weight matrices of trained networks behave like a random part plus a full-rank correlated part whose spectrum remains tractable even when it contains many outliers.

What would settle it

Numerical diagonalization of large explicit matrices R + S, with S constructed to have a known increasing set of outliers, whose computed eigenvalue positions deviate from the derived asymptotic formulas by more than the expected error term.

Figures

Figures reproduced from arXiv: 2507.12182 by Ievgenii Afanasiev, Leonid Berlyand, Mariia Kiyashko.

**Figure 2.** Figure 2: Numerical simulations of a DNN with 3 layers. Two figures correspond to two [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

read the original abstract

The paper is concerned with deformed Wigner random matrices. These matrices are closely related to Deep Neural Networks (DNNs): weight matrices of trained DNNs could be represented in the form $R + S$, where $R$ is random and $S$ is highly correlated. The spectrum of such matrices plays a key role in rigorous underpinning of the novel pruning technique based on Random Matrix Theory. In practice, the spectrum of the matrix $S$ can be rather complicated. In this paper, we develop an asymptotic analysis for the case of full rank $S$ with increasing number of outlier eigenvalues.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper extends asymptotic eigenvalue analysis to full-rank perturbations with a growing number of outliers in deformed Wigner matrices.

read the letter

The main point is that this work develops an asymptotic analysis for the eigenvalues of deformed Wigner matrices when the perturbation S is full rank and carries an increasing number of outlier eigenvalues. They treat the case motivated by modeling trained DNN weight matrices as random plus correlated parts. The approach uses resolvent or perturbative methods to track the outliers one by one while keeping their collective effect on the empirical spectral measure under control. The stress-test confirms the derivations follow standard RMT assumptions on moments, scaling, and separation from the bulk edge, with no internal contradictions showing up in the argument. That is the solid core. The paper does a clean job extending prior results on fixed or low numbers of outliers to this growing case without adding unjustified steps. The math looks reproducible in principle under the stated conditions. The softer spot is the DNN connection. The abstract presents it as motivation by noting that the spectrum of S can be complicated yet now analyzable, but there is little concrete discussion of how well real trained networks fit the R + S form or how the asymptotics would directly improve pruning. That is typical for a theoretical paper and not a load-bearing flaw, just a place where the application claim stays suggestive. This is for people already working in random matrix theory or in the subfield that applies it to neural network spectra. A reader who knows the basics of deformed Wigner matrices and wants the growing-outlier extension will get the most from it. It deserves a serious referee to check the details of the expansions and any error control. I would send it out for peer review rather than desk reject.

Referee Report

0 major / 2 minor

Summary. The manuscript develops an asymptotic analysis for the eigenvalues of deformed Wigner matrices of the form R + S, where R is a large random matrix and S is a full-rank perturbation possessing an increasing number of outlier eigenvalues. The work is motivated by the spectral properties of trained DNN weight matrices and aims to support RMT-based pruning techniques by extending classical results on outlier eigenvalues to the growing-outlier regime under standard assumptions on moments and scaling.

Significance. If the derivations hold, the results would extend random matrix theory to a practically relevant class of full-rank perturbations with growing outliers, providing a foundation for spectral analysis in high-dimensional machine learning models. The approach relies on perturbative resolvent methods that track individual outliers while controlling collective effects on the empirical measure, which is a natural and technically substantive extension of prior work on fixed-rank or low-rank deformations.

minor comments (2)

[Introduction] The introduction would benefit from a brief statement of the main theorem (including the precise asymptotic regime for the number of outliers) to make the contribution immediately clear to readers.
Notation for the outlier locations and the separation condition from the bulk edge should be introduced with an explicit reference to the relevant assumption or equation early in the technical sections.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the manuscript, accurate summary of its contributions, and recommendation for minor revision. The work extends classical outlier eigenvalue results for deformed Wigner matrices to the regime of full-rank perturbations with a growing number of outliers, motivated by spectral analysis of trained DNN weight matrices.

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper develops an asymptotic analysis for the eigenvalues of deformed Wigner matrices with full-rank perturbations S possessing an increasing number of outlier eigenvalues. The derivations rely on standard random matrix theory tools including resolvent methods and perturbative expansions, under explicit assumptions such as bounded moments on Wigner entries and separation of outliers from the bulk. No load-bearing step reduces by the paper's own equations to a self-definition, a fitted parameter renamed as a prediction, or a self-citation chain that lacks independent content. The central claims follow from the presented mathematical arguments without internal reduction to inputs, rendering the analysis self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no specific free parameters, axioms, or invented entities are identifiable from the given information.

pith-pipeline@v0.9.0 · 5639 in / 1165 out tokens · 45965 ms · 2026-05-19T04:36:15.450043+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

g_μ0(z) = g_ν0(ω_μ0(z)) with ω_τ(z) = z + σ² g_τ(z); Theorem 2.1 on N/r(μ - μ0) → μ1 with μ1(Δ) = ν1(ω_μ0(Δ))
IndisputableMonolith/Foundation/DimensionForcing.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Assumptions 1–3 and growing-rank condition r(N)→∞, r(N)=o(N)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

[1]

G. W. Anderson, A. Guionnet, and O. Zeitouni. An introduction to random matrices . Cambridge University Press , Cambridge, 2010

work page 2010
[2]

Mathematics of deep learning: An introduction

Leonid Berlyand and Pierre Emmanuel Jabin. Mathematics of deep learning: An introduction . de Gruyter, Germany, April 2023. Publisher Copyright: 2023 Walter de Gruyter GmbH, Berlin/Boston. All rights reserved

work page 2023
[3]

H. C. Ji and J. O. Lee. Gaussian ﬂuctuations for linear spectral statistics of deformed Wigner matrices . Random Matrices: Theory and Applications , page 2050011, 2019

work page 2019
[4]

Enhancing accuracy in deep learning using random matrix theory

Berlyand Leonid, Etienne Sandier, Shmalo Yitzchak, and Lei Zhang. Enhancing accuracy in deep learning using random matrix theory. Journal of Machine Learning , 3(4):347--412, 2024

work page 2024
[5]

Martin and Michael W

Charles H. Martin and Michael W. Mahoney. Implicit self-regularization in deep neural networks: Evidence from random matrix theory and implications for learning. Journal of Machine Learning Research , 22(165):1--73, 2021

work page 2021
[6]

L. Pastur. The spectrum of random matrices. Teoret. Mat. Fiz. , 10:102--112, 1972. (in Russian)

work page 1972
[7]

Pastur and M

L. Pastur and M. Shcherbina. Eigenvalue Distribution of Large Random Matrices . AMS, 2011

work page 2011
[8]

P\'ech\'e

S. P\'ech\'e. The largest eigenvalue of small rank perturbations of hermitian random matrices. Probab. Theory Related Fields , 134(1):127--173, 2006

work page 2006
[9]

E. Wigner. On the distribution of the roots of certain symmetric matrices. Annals of Mathematics , 67:325--327, 1958

work page 1958

[1] [1]

G. W. Anderson, A. Guionnet, and O. Zeitouni. An introduction to random matrices . Cambridge University Press , Cambridge, 2010

work page 2010

[2] [2]

Mathematics of deep learning: An introduction

Leonid Berlyand and Pierre Emmanuel Jabin. Mathematics of deep learning: An introduction . de Gruyter, Germany, April 2023. Publisher Copyright: 2023 Walter de Gruyter GmbH, Berlin/Boston. All rights reserved

work page 2023

[3] [3]

H. C. Ji and J. O. Lee. Gaussian ﬂuctuations for linear spectral statistics of deformed Wigner matrices . Random Matrices: Theory and Applications , page 2050011, 2019

work page 2019

[4] [4]

Enhancing accuracy in deep learning using random matrix theory

Berlyand Leonid, Etienne Sandier, Shmalo Yitzchak, and Lei Zhang. Enhancing accuracy in deep learning using random matrix theory. Journal of Machine Learning , 3(4):347--412, 2024

work page 2024

[5] [5]

Martin and Michael W

Charles H. Martin and Michael W. Mahoney. Implicit self-regularization in deep neural networks: Evidence from random matrix theory and implications for learning. Journal of Machine Learning Research , 22(165):1--73, 2021

work page 2021

[6] [6]

L. Pastur. The spectrum of random matrices. Teoret. Mat. Fiz. , 10:102--112, 1972. (in Russian)

work page 1972

[7] [7]

Pastur and M

L. Pastur and M. Shcherbina. Eigenvalue Distribution of Large Random Matrices . AMS, 2011

work page 2011

[8] [8]

P\'ech\'e

S. P\'ech\'e. The largest eigenvalue of small rank perturbations of hermitian random matrices. Probab. Theory Related Fields , 134(1):127--173, 2006

work page 2006

[9] [9]

E. Wigner. On the distribution of the roots of certain symmetric matrices. Annals of Mathematics , 67:325--327, 1958

work page 1958