arxiv: 2603.19758 · v2 · submitted 2026-03-20 · 🧮 math.NA · cs.NA· math.OC· math.PR

Recognition: 2 theorem links

· Lean Theorem

Eigenvalue stability and new perturbation bounds for the extremal eigenvalues of a matrix

Phuc Tran , Van Vu

Authors on Pith no claims yet

Pith reviewed 2026-05-15 08:40 UTC · model grok-4.3

classification 🧮 math.NA cs.NAmath.OCmath.PR

keywords matrix perturbationsingular valuescondition numberDavis-Kahan theoremrandom noisecontour analysisregional stabilitynumerical linear algebra

0 comments

The pith

Regional stability and contour analysis yield tighter bounds on the smallest singular value under random noise.

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

The paper introduces regional stability as a framework for tracking how random perturbations affect the extremal singular values of a full-rank matrix. It first bounds changes in the associated singular spaces and then controls those changes through a contour-integral argument. The resulting estimates for the smallest singular value are sharper than classical ones precisely when the original matrix dominates the noise in norm. As a byproduct the same technique refines the Davis-Kahan theorem on subspace perturbation. These bounds directly inform how the condition number of a noisy matrix behaves in practice.

Core claim

By defining regional stability, the authors bound perturbations of extremal singular values through perturbations of singular spaces; the latter are controlled by a novel contour analysis that improves upon the classical Davis-Kahan theorem, producing new, more favorable estimates for the least singular value when the ground matrix A is large compared with the random noise matrix E.

What carries the argument

Regional stability, which links singular-value perturbations to singular-space perturbations, combined with a contour-integral argument that bounds the latter.

If this is right

Sharper a priori estimates for the condition number of noisy inputs to numerical algorithms.
Improved control on the smallest singular value when the signal matrix dominates additive random noise.
A strengthened version of the Davis-Kahan theorem that applies in additional regimes.
A modular framework that separates space perturbation from value perturbation for other extremal eigenproblems.

Where Pith is reading between the lines

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

The contour technique may extend to non-random structured perturbations if suitable contour choices can be made.
Numerical experiments on high-dimensional data matrices would test whether the predicted improvement in condition-number stability is observable in floating-point arithmetic.
The regional-stability notion could be adapted to track other matrix functionals such as the numerical range or pseudospectrum.

Load-bearing premise

The noise matrix is random, the ground matrix is full rank, and the contour analysis applies without extra restrictions on the spectrum.

What would settle it

Compute the ratio of the new bound to the classical bound on σ_n(A+E) for a sequence of matrices where ||A||/||E|| grows without bound and check whether the ratio remains below one.

read the original abstract

Let $A$ be a full ranked $ n\times n$ matrix, with singular values $\sigma_1 (A) \ge \dots \ge \sigma_n (A) >0$. The condition number $\kappa(A):= \sigma_1(A)/\sigma_n(A)=\|A\|\cdot \|A\|^{-1}$ is a key parameter in the analysis of algorithms taking $A$ as input. In practice, matrices (representing real data) are often perturbed by noise. Technically speaking, the real input would be a noisy variant $\tilde A =A +E$ of $A$, where $E$ represents the noise. The condition number $\kappa (\tilde A)$ will be used instead of $\kappa (A)$. Thus, it is of importance to measure the impact of noise on the condition number. In this paper, we focus on the case when the noise is random. We introduce the notion of regional stability, via which we design a new framework to estimate the perturbation of the extremal singular values and the condition number of a matrix. Our framework allows us to bound the perturbation of singular values through the perturbation of singular spaces. We then bound the latter using a novel contour analysis argument, which, as a co-product, provides an improved version of the classical Davis-Kahan theorem in many settings. Our new estimates concerning the least singular value $\sigma_n(A)$ complement well-known results in this area, and are more favorable in the case when the ground matrix $A$ is large compared to the noise matrix $E$.

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 introduces regional stability and contour integrals for tighter bounds on the smallest singular value when A dominates E, but the isolation step probably needs an explicit gap that the abstract downplays.

read the letter

The core new piece is the regional stability idea that routes perturbation bounds for extremal singular values through singular-space changes, with a contour argument that is supposed to sharpen classical Davis-Kahan estimates and give better control on sigma_n when the ground matrix is large relative to the noise. That focus on the large-A regime is the part that could actually matter for condition-number error analysis in practice. They also claim the same contour work yields an improved subspace bound as a side product. The abstract presents this as a clean framework rather than a rehash, and the emphasis on random noise plus full-rank A is standard but reasonable. The main soft spot is the contour isolation itself. Standard contour methods require the contour radius to exceed the perturbation norm while staying inside the gap to neighboring singular values; if the smallest singular values cluster, any fixed contour either misses the target or pulls in extras, and the claimed improvement collapses. The abstract says the argument works without additional spectrum restrictions, yet the large-A regime alone does not guarantee a gap bigger than ||E||. That tension needs explicit checking in the proofs. The paper is aimed at people who already work on perturbation bounds and condition numbers in numerical linear algebra. A reader who needs sharper sigma_n estimates under random noise would find the regional-stability lens worth testing, provided the gap issue is resolved. It is solid enough on its own terms to deserve a serious referee who can verify the contour derivations and see whether the improvements survive when singular values are not well separated.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the concept of regional stability for a full-rank matrix A perturbed by random noise E to derive new bounds on the extremal singular values σ_1(Ã) and σ_n(Ã) and the condition number κ(Ã). It frames perturbation of singular values through changes in singular subspaces and employs a contour-integral argument to bound those subspace perturbations, yielding an improved Davis-Kahan-type result in many settings and tighter estimates for the smallest singular value when ||A|| ≫ ||E||.

Significance. If the contour argument is valid without unstated gap restrictions, the work would supply practically useful refinements to classical perturbation theory for singular values under random noise, particularly in regimes where the ground matrix dominates the perturbation. The regional-stability framework offers a conceptual alternative to direct norm bounds and could inform stability analysis in numerical linear algebra and data-driven methods.

major comments (2)

[Abstract / contour analysis] Abstract and contour-analysis section: the claim that the novel contour argument applies 'without additional restrictions on the spectrum' is not supported by the standard requirements of contour isolation. Isolating the smallest singular value requires a contour radius larger than ||E|| yet strictly inside the gap to the next singular value; when singular values cluster near σ_n the isolation step fails and the claimed improvement over Davis-Kahan bounds does not hold.
[Title] Title versus content: the title announces results for eigenvalues, yet every statement, definition, and bound concerns singular values of A and Ã. This mismatch affects the central claim and must be resolved.

minor comments (2)

[Abstract] The abstract states that the new estimates 'complement well-known results' but does not cite the specific classical bounds (e.g., Weyl, Mirsky, or Davis-Kahan variants) being improved; explicit comparison would strengthen the contribution.
[Introduction] Notation for the perturbed matrix alternates between Ã and A+E without a single consistent definition; adopt one symbol throughout.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's careful reading of our manuscript and the constructive feedback provided. We respond to the major comments point by point below, and we will make the necessary revisions to the manuscript.

read point-by-point responses

Referee: [Abstract / contour analysis] Abstract and contour-analysis section: the claim that the novel contour argument applies 'without additional restrictions on the spectrum' is not supported by the standard requirements of contour isolation. Isolating the smallest singular value requires a contour radius larger than ||E|| yet strictly inside the gap to the next singular value; when singular values cluster near σ_n the isolation step fails and the claimed improvement over Davis-Kahan bounds does not hold.

Authors: We agree that standard contour integration requires the contour to isolate the relevant singular value, necessitating a sufficient gap. Our regional stability framework aims to provide bounds that are useful in the presence of random noise, but we acknowledge that the argument does depend on isolation. We will revise the abstract and the contour analysis section to explicitly state the required spectral gap condition and remove any implication that the method applies without restrictions on the spectrum. This will also better delineate the settings where our bounds improve upon Davis-Kahan. revision: yes
Referee: [Title] Title versus content: the title announces results for eigenvalues, yet every statement, definition, and bound concerns singular values of A and Ã. This mismatch affects the central claim and must be resolved.

Authors: We thank the referee for identifying this inconsistency. The work is indeed focused on singular values, and the title erroneously refers to eigenvalues. We will change the title to accurately describe the content, for example to 'Singular value stability and new perturbation bounds for the extremal singular values of a matrix'. revision: yes

Circularity Check

0 steps flagged

No circularity: new regional stability and contour framework adds independent content

full rationale

The paper defines regional stability as a new notion, then derives perturbation bounds on extremal singular values by first bounding singular-space perturbations via a novel contour-integral argument that also yields an improved Davis-Kahan variant. No step reduces by construction to a fitted parameter, self-defined quantity, or load-bearing self-citation; the contour isolation is presented as an original technical contribution whose applicability is asserted under the paper's stated assumptions on A and random E. The derivation chain therefore remains self-contained against external benchmarks and does not collapse to renaming or re-expressing its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the newly introduced regional-stability definition and the applicability of contour integration to singular-space perturbations; these are supplied by the paper rather than drawn from prior literature.

axioms (1)

domain assumption A is an n by n full-rank matrix with positive singular values σ1(A) ≥ … ≥ σn(A) > 0
Stated explicitly in the opening sentence of the abstract as the setup for the condition number.

invented entities (1)

regional stability no independent evidence
purpose: Framework to bound singular-value perturbations via singular-space perturbations
Newly defined notion introduced to organize the perturbation analysis

pith-pipeline@v0.9.0 · 5591 in / 1285 out tokens · 58799 ms · 2026-05-15T08:40:22.467532+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.

We introduce the notion of regional stability... bound the latter using a novel contour analysis argument... improved version of the classical Davis-Kahan theorem
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

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

Theorem 4.1 (Stability of projections)... contour double-jump strategy

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

32 extracted references · 32 canonical work pages

[1]

Benaych-Georges, A

F. Benaych-Georges, A. Guionnet, and M. Maida,Fluctuations of the extreme eigenvalues of finite rank deformations of random matrices, Electron. J. Probab.16(2011), no. 60, 1621–1662

work page 2011
[2]

Benaych-Georges and R

F. Benaych-Georges and R. Nadakuditi,The eigenvalues and eigenvectors of finite, low rank perturbations of large random matrices, Adv. Math.227(2011), no. 1, 494–521

work page 2011
[3]

Bhatia,Matrix analysis, Grad

R. Bhatia,Matrix analysis, Grad. Texts in Math., Springer, New York, NY, 2013

work page 2013
[4]

Bourgain,On a problem of farrell and vershynin in random matrix theory, Geometric aspects of func- tional analysis: Israel seminar (gafa) 2014–2016, 2017, pp

J. Bourgain,On a problem of farrell and vershynin in random matrix theory, Geometric aspects of func- tional analysis: Israel seminar (gafa) 2014–2016, 2017, pp. 65–69

work page 2014
[5]

Capitaine, C

M. Capitaine, C. Donati-Martin, and D. Feral,The largest eigenvalues of finite rank deformation of large wigner matrices: convergence and nonuniversality of the fluctuations, Ann. Prob.37(2009), no. 1, 1–47

work page 2009
[6]

Davis and W.M

C. Davis and W.M. Kahan,The rotation of eigenvectors by a perturbation, SIAM J. Numer. Anal7 (1970), 1–46

work page 1970
[7]

Farrell and R

B. Farrell and R. Vershynin,Smoothed analysis of symmetric random matrices with continuous distribu- tions, Proc. Amer. Math. Soc.144(2016), no. 5, 2257–2261

work page 2016
[8]

Golub and C.F

G.H. Golub and C.F. Van Loan,Matrix computations, JHU press, 2013

work page 2013
[9]

Higham,Accuracy and stability of numerical algorithms, SIAM, 2002

N.J. Higham,Accuracy and stability of numerical algorithms, SIAM, 2002

work page 2002
[10]

V. Jain, A. Sah, and M. Sawhney,On the smoothed analysis of the smallest singular value with discrete noise, Bull. Lond. Math. Soc.54(2022), no. 2, 369–388

work page 2022
[11]

Jirak and M

M. Jirak and M. Wahl,Perturbation bounds for eigenspaces under a relative gap condition, Proc. Amer. Math. Soc.148(2020), no. 2, 479–494

work page 2020
[12]

Math.412 (2023)

,Relative perturbation bounds with applications to empirical covariance operators, Adv. Math.412 (2023). Paper No. 108808, 59 pp

work page 2023
[13]

Kato,Perturbation theory for linear operators, Classics in Mathematics, Springer, 1980

T. Kato,Perturbation theory for linear operators, Classics in Mathematics, Springer, 1980

work page 1980
[14]

Knowles and J

A. Knowles and J. Yin,The isotropic semicircle law and deformation of wigner matrices, Comm. Pure Appl. Math.66(2013), no. 11, 1663–1749. EIGENVALUES STABILITY 23

work page 2013
[15]

Koltchinskii and K

V. Koltchinskii and K. Lounici,Asymptotics and concentration bounds for bilinear forms of spectral projectors of sample covariance, Ann. Inst. Henri Poincare Probab. Stat.52(2016), no. 4, 1976–2013

work page 2016
[16]

Koltchinskii and D

V. Koltchinskii and D. Xia,Perturbation of linear forms of singular vectors under gaussian noise, Progr. Probab.71(2016), 397–423

work page 2016
[17]

Livshyts, K

G.V. Livshyts, K. Tikhomirov, and R. Vershynin,The smallest singular value of inhomogeneous square random matrices, Ann. Prob.49(2021), no. 3, 1286–1309

work page 2021
[18]

O’Rourke, V

S. O’Rourke, V. Vu, and K. Wang,Matrices with gaussian noise: Optimal estimates for singular subspace perturbation, IEEE Trans. Inform. Theory70(2023), no. 3, 1978–2002

work page 2023
[19]

1, 26–59

Sean O’Rourke, Van Vu, and Ke Wang,Random perturbation of low rank matrices: Improving classical bounds, Linear Algebra Appl.540(2018), no. 1, 26–59

work page 2018
[20]

Pizzo, D

A. Pizzo, D. Renfrew, and A. Soshnikov,On finite rank deformations of wigner matrices, Ann. Inst. Henri Poincar´ e Probab. Stat.49(2013), no. 1, 64–94

work page 2013
[21]

Rudelson and R

M. Rudelson and R. Vershynin,The Littlewood–Offord problem and invertibility of random matrices, Adv. in Math.218(2008), no. 2, 600–633

work page 2008
[22]

Sankar, D.A

A. Sankar, D.A. Spielman, and S-H. Teng,Smoothed analysis of the condition numbers and growth factors of matrices, SIAM J. Matrix Anal. Appl.28(2006), no. 2, 446–476

work page 2006
[23]

Stein and R

E. Stein and R. Shakarchi,Complex analysis, Princeton Lectures in Analysis II, Princeton University Press, 2003

work page 2003
[24]

Tao,Topics in random matrix theory, Vol

T. Tao,Topics in random matrix theory, Vol. 132, American Mathematical Society, 2012

work page 2012
[25]

Tao and V

T. Tao and V. Vu,Random matrices: the circular law, Commun. Contemp. Math.10(2008), no. 02, 261–307

work page 2008
[26]

Comp.79(2010), no

,Smooth analysis of the condition number and the least singular value, Math. Comp.79(2010), no. 272, 2333–2352

work page 2010
[27]

Tikhomirov,Invertibility via distance for noncentered random matrices with continuous distributions, Random Structures & Algorithms57(2020), no

K. Tikhomirov,Invertibility via distance for noncentered random matrices with continuous distributions, Random Structures & Algorithms57(2020), no. 2, 526–562

work page 2020
[28]

P. Tran, N. K. Vishnoi, and V. Vu,Spectral perturbation bounds for low-rank approximation with ap- plications to privacy, Proceedings of the 39th Conference on Neural Information Processing Systems, 2025

work page 2025
[29]

Tran and V

P. Tran and V. Vu,Davis–Kahan theorem under a moderate gap condition, Commun. Contemp. Math. 28(2026), no. 1. Article no. 2550035

work page 2026
[30]

Trefethen and D

L.N. Trefethen and D. Bau,Numerical linear algebra, SIAM, Philadelphia, 1997

work page 1997
[31]

Vershynin,High dimensional probability, Cambridge Univ

R. Vershynin,High dimensional probability, Cambridge Univ. Press, 2019

work page 2019
[32]

Weyl,Das asymptotische verteilungsgesetz der eigenwerte linearer partieller differentialgleichungen, Mathematische Annalen71(1912), no

H. Weyl,Das asymptotische verteilungsgesetz der eigenwerte linearer partieller differentialgleichungen, Mathematische Annalen71(1912), no. 4, 441–479

work page 1912