arxiv: 2604.20008 · v1 · submitted 2026-04-21 · 🧮 math.PR

Recognition: unknown

Mixing times of Langevin dynamics for spiked matrix models

Reza Gheissari , Curtis Grant , Tianmin Yu

Authors on Pith no claims yet

Pith reviewed 2026-05-10 01:01 UTC · model grok-4.3

classification 🧮 math.PR

keywords mixing timesLangevin dynamicsspiked matrix modelsmetastabilityfree energyWigner matriceshigh-dimensional samplingspherical prior

0 comments

The pith

Langevin dynamics for large-signal spiked matrices mix in O(log N) from uniform spherical starts even below the critical temperature.

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

The paper examines Langevin dynamics on the sphere for Wigner matrices with a rank-one spherical spike when the signal-to-noise ratio θ stays large but fixed. It establishes a sharp transition in worst-case mixing time at the critical inverse temperature β = 1/θ: logarithmic for α < 1 and exponential in N for α > 1. Symmetric initializations, including the uniform spherical prior, evade the exponential slowdown and still mix in O(log N) time for α > 1. The work then identifies the precise exponential rate of the slow (worst-case) mixing as the free-energy difference between the spiked and null models.

Core claim

In the regime of large but order-one signal-to-noise ratio θ, the mixing time of Langevin dynamics transitions sharply at β = 1/θ: for α <1 in β=α/θ it is O(log N), for α>1 it is exp(N) in worst case. However, from the uniform spherical prior or any initialization symmetric wrt the top eigenvector, the mixing remains O(log N) even for α>1. The worst-case exponential rate equals the difference of free energies of the spiked and null models.

What carries the argument

Symmetry with respect to the top eigenvector of the spiked matrix, which lets the dynamics avoid the metastable null-model basin and reach the spiked equilibrium in logarithmic time; the free-energy gap then sets the precise escape rate for asymmetric starts.

If this is right

For α < 1 the mixing time remains O(log N) from any reasonable initialization.
Symmetric initializations achieve O(log N) mixing for all α > 1, removing the exponential barrier.
The worst-case mixing time for α > 1 is exactly exponential with rate given by the free-energy difference between spiked and null models.
The metastability picture holds uniformly for any initialization symmetric about the top eigenvector.

Where Pith is reading between the lines

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

In practice, drawing an initial point uniformly on the sphere may suffice for fast sampling in this and similar spiked models even at low temperature.
The free-energy gap may govern escape rates in other high-dimensional diffusions or Glauber dynamics on spiked structures.
Small perturbations away from exact symmetry could still preserve fast mixing for moderate N, providing a testable robustness check.

Load-bearing premise

The signal-to-noise ratio θ must be large yet remain order one, and the initial distribution must be symmetric with respect to the leading eigenvector.

What would settle it

A direct simulation of the dynamics from a symmetric initialization at α slightly larger than 1 and moderate N that shows an escape or mixing time growing exponentially with N rather than logarithmically.

read the original abstract

We investigate the Langevin dynamics for Wigner matrices with a spherical spike, in the regime where the signal-to-noise ratio $\theta$ is large, but order one. For large, order-$1$, signal-to-noise, the (worst-case) mixing time undergoes a sharp transition around the critical inverse temperature $\beta_c(\theta) = \frac{1}{\theta}$. Namely, if $\beta = \alpha/\theta$, and $\alpha<1$ then at large $\theta$ the mixing time is $O(\log N)$, and if $\alpha>1$ it is exponential in $N$. We show that initialized from the uniform-at-random spherical prior, however, the mixing time in the low-temperature $\alpha>1$ regime circumvents the exponential bottleneck and the mixing time is $O(\log N)$. In fact, this fast mixing holds for any initialization that is symmetric with respect to the top eigenvector of the spiked matrix. Using this, we are able to show a low-temperature metastability picture, pinning down the exact exponential rate of the (worst-case initialization) mixing time for low temperatures, showing it is given by the difference of the free energies of the spiked and null models.

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

Symmetric starts give O(log N) mixing for Langevin on spiked Wigner matrices even below the critical temperature, with the worst-case exponential rate exactly the free-energy gap.

read the letter

The core result is that Langevin dynamics on a Wigner matrix with a spherical spike mixes in O(log N) time from the uniform spherical prior or any initialization symmetric under sign flip of the top eigenvector, even when the inverse temperature satisfies α > 1 and generic starts require exponential time. The paper also pins the precise exponential rate for the worst-case mixing time to the free-energy difference between the spiked and null models. This combination of the sharp transition at β_c(θ) = 1/θ, the bypass via symmetry, and the exact metastability rate looks like a solid addition to the existing work on dynamics for spiked models. The claims are stated cleanly for θ large but order one, and the distinction between symmetric and asymmetric starts is useful for understanding when MCMC succeeds in these inference settings. The analysis appears to rest on standard free-energy calculations and overlap processes rather than circular definitions. The main soft spot is that the abstract gives the asymptotics but the error bounds, technical lemmas, and control on the effective potential are not visible here, so the constants and the range of validity for moderate θ need verification in the proofs. The symmetry assumption is explicit and necessary for the fast-mixing claim, which is fine but means the result does not apply to arbitrary starts. Overall the paper is aimed at people working on mixing times and metastability for high-dimensional statistical models. It is precise enough and addresses a concrete algorithmic question to deserve a serious referee, even if the proofs require close checking.

Referee Report

0 major / 2 minor

Summary. The paper studies mixing times of Langevin dynamics for spherical spiked Wigner matrices in the regime of large but order-one signal-to-noise ratio θ. It identifies a sharp transition at the critical inverse temperature β_c(θ)=1/θ: when β=α/θ with α<1 the mixing time is O(log N) at large θ, while for α>1 the worst-case mixing time is exponential in N. However, for initializations symmetric with respect to the top eigenvector (including the uniform spherical prior), the mixing time remains O(log N) even in the low-temperature regime α>1. The paper further establishes a low-temperature metastability result in which the exact exponential rate of the worst-case mixing time equals the free-energy difference between the spiked and null models.

Significance. If the derivations hold, the results give a precise characterization of mixing and metastability for Langevin dynamics on a non-convex landscape arising from a canonical spiked random-matrix model. The distinction between symmetric and generic initializations clarifies how symmetry bypasses the exponential barrier, while the exact free-energy rate strengthens the link to statistical-mechanics metastability theory. These findings are relevant to sampling algorithms in high-dimensional statistics and machine learning.

minor comments (2)

The abstract and introduction would benefit from an explicit statement of the precise error terms or uniformity requirements in the O(log N) bounds (e.g., dependence on θ and α).
Notation for the spherical prior and the symmetry condition with respect to the top eigenvector could be introduced earlier and used consistently throughout the metastability section.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their careful reading of the manuscript, for the accurate summary of our results, and for the positive recommendation to accept. We are pleased that the distinction between symmetric and generic initializations, as well as the precise free-energy rate for metastability, were viewed as strengthening the connection to statistical-mechanics theory.

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The paper derives mixing-time bounds for Langevin dynamics on the spiked Wigner model by combining symmetry of the uniform spherical prior (which places the initialization at the saddle of the overlap potential) with standard metastability estimates. The O(log N) mixing from symmetric initializations follows directly from the absence of a tunneling requirement in the low-temperature regime. The worst-case exponential rate is identified with the free-energy difference between the spiked and null models; these free energies are defined independently via the respective partition functions and are not fitted to the mixing-time conclusion. No equation reduces a prediction to a fitted parameter by construction, no uniqueness theorem is imported from self-citation, and no ansatz is smuggled in. The central claims therefore rest on the model's explicit definitions and classical metastability techniques rather than on any circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The analysis relies on standard domain assumptions of random matrix theory for Wigner matrices and spherical priors; no explicit free parameters or new invented entities are introduced in the abstract.

axioms (1)

domain assumption Standard properties of Wigner matrices with spherical spike in the large-N limit with fixed θ
Invoked to define the model and the critical temperature β_c(θ)=1/θ.

pith-pipeline@v0.9.0 · 5510 in / 1321 out tokens · 99633 ms · 2026-05-10T01:01:16.493406+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

37 extracted references · 4 canonical work pages

[1]

Fundamental limits of detection in the spiked Wigner model.The Annals of Statistics, 48(2):863 – 885, 2020

Ahmed El Alaoui, Florent Krzakala, and Michael Jordan. Fundamental limits of detection in the spiked Wigner model.The Annals of Statistics, 48(2):863 – 885, 2020

2020
[2]

Symmetric langevin spin glass dynamics.The Annals of Probability, 25(3):1367– 1422, 1997

G Ben Arous and Alice Guionnet. Symmetric langevin spin glass dynamics.The Annals of Probability, 25(3):1367– 1422, 1997

1997
[3]

Online stochastic gradient descent on non-convex losses from high-dimensional inference.Journal of Machine Learning Research, 22(106):1–51, 2021

Gerard Ben Arous, Reza Gheissari, and Aukosh Jagannath. Online stochastic gradient descent on non-convex losses from high-dimensional inference.Journal of Machine Learning Research, 22(106):1–51, 2021

2021
[4]

Langevin dynamics for high-dimensional optimization: the case of multi-spiked tensor pca, 2024

G´ erard Ben Arous, C´ edric Gerbelot, and Vanessa Piccolo. Langevin dynamics for high-dimensional optimization: the case of multi-spiked tensor pca, 2024

2024
[5]

Necessary and sufficient conditions for almost sure convergence of the largest eigenvalue of a wigner matrix.The Annals of Probability, pages 1729–1741, 1988

Zhi-Dong Bai and Yong-Qua Yin. Necessary and sufficient conditions for almost sure convergence of the largest eigenvalue of a wigner matrix.The Annals of Probability, pages 1729–1741, 1988

1988
[6]

Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices.Ann

Jinho Baik, G´ erard Ben Arous, and Sandrine P´ ech´ e. Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices.Ann. Probab., 33(5):1643–1697, 2005

2005
[7]

Spherical spin glass model with external field.Journal of Statistical Physics, 183(2):31, 2021

Jinho Baik, Elizabeth Collins-Woodfin, Pierre Le Doussal, and Hao Wu. Spherical spin glass model with external field.Journal of Statistical Physics, 183(2):31, 2021

2021
[8]

Ben Arous, A

G. Ben Arous, A. Dembo, and A. Guionnet. Aging of spherical spin glasses.Probability Theory and Related Fields, 120(1):1–67, 2001

2001
[9]

Cugliandolo-Kurchan equations for dynamics of spin- glasses.Probab

G´ erard Ben Arous, Amir Dembo, and Alice Guionnet. Cugliandolo-Kurchan equations for dynamics of spin- glasses.Probab. Theory Related Fields, 136(4):619–660, 2006

2006
[10]

Spectral gap estimates in mean field spin glasses.Communications in Mathematical Physics, 361(1):1–52, 2018

G´ erard Ben Arous and Aukosh Jagannath. Spectral gap estimates in mean field spin glasses.Communications in Mathematical Physics, 361(1):1–52, 2018

2018
[11]

Stochastic gradient descent in high dimensions for multi-spiked tensor pca, 2025

G´ erard Ben Arous, C´ edric Gerbelot, and Vanessa Piccolo. Stochastic gradient descent in high dimensions for multi-spiked tensor pca, 2025

2025
[12]

Benaych-Georges, A

F. Benaych-Georges, A. Guionnet, and M. Maida. Large deviations of the extreme eigenvalues of random defor- mations of matrices.Probability Theory and Related Fields, 154(3):703–751, Dec 2012

2012
[13]

Rank-one matrix estimation: analytic time evolution of gradient descent dynamics

Antoine Bodin and Nicolas Macris. Rank-one matrix estimation: analytic time evolution of gradient descent dynamics. InConference on Learning Theory, pages 635–678. PMLR, 2021

2021
[14]

Springer, Cham, , 2015

Anton Bovier and Frank den Hollander.Metastability, volume 351 ofGrundlehren der Mathematischen Wis- senschaften [Fundamental Principles of Mathematical Sciences]. Springer, Cham, , 2015. A potential-theoretic approach

2015
[15]

A note on the isoperimetric constant

Peter Buser. A note on the isoperimetric constant. InAnnales scientifiques de l’ ´Ecole normale sup´ erieure, vol- ume 15, pages 213–230, 1982

1982
[16]

The largest eigenvalues of finite rank defor- mation of large wigner matrices: convergence and nonuniversality of the fluctuations.The Annals of Probability, 37(1):1–47, 2009

Mireille Capitaine, Catherine Donati-Martin, Delphine F´ eral, et al. The largest eigenvalues of finite rank defor- mation of large wigner matrices: convergence and nonuniversality of the fluctuations.The Annals of Probability, 37(1):1–47, 2009

2009
[17]

A lower bound for the smallest eigenvalue of the laplacian

Jeff Cheeger. A lower bound for the smallest eigenvalue of the laplacian. In R. C. Gunning, editor,Problems in Analysis: A Symposium in Honor of Salomon Bochner, pages 195–199. Princeton University Press, Princeton, NJ, 1970

1970
[18]

Crisanti, H

A. Crisanti, H. Horner, and H. J. Sommers. The spherical p-spin interaction spin-glass model.Zeitschrift f¨ ur Physik B Condensed Matter, 92(2):257–271, Jun 1993

1993
[19]

Cugliandolo and Jorge Kurchan

Leticia F. Cugliandolo and Jorge Kurchan. Analytical solution of the off-equilibrium dynamics of a long-range spin-glass model.Phys. Rev. Lett., 71:173–176, Jul 1993

1993
[20]

Rigidity of eigenvalues of generalized wigner matrices.Advances in Mathematics, 229(3):1435–1515, 2012

L´ aszl´ o Erd˝ os, Horng-Tzer Yau, and Jun Yin. Rigidity of eigenvalues of generalized wigner matrices.Advances in Mathematics, 229(3):1435–1515, 2012

2012
[21]

Metastability in glauber dynamics for heavy-tailed spin glasses.Communica- tions in Mathematical Physics, 406(4):84, 2025

Reza Gheissari and Curtis Grant. Metastability in glauber dynamics for heavy-tailed spin glasses.Communica- tions in Mathematical Physics, 406(4):84, 2025

2025
[22]

On the spectral gap of spherical spin glass dynamics.Annales de l’Institut Henri Poincar´ e, Probabilit´ es et Statistiques, 55(2):756 – 776, 2019

Reza Gheissari and Aukosh Jagannath. On the spectral gap of spherical spin glass dynamics.Annales de l’Institut Henri Poincar´ e, Probabilit´ es et Statistiques, 55(2):756 – 776, 2019

2019
[23]

Local semicircle law under moment conditions

Friedrich G¨ otze, Alexey Naumov, and Alexander Tikhomirov. Local semicircle law under moment conditions. part i: The stieltjes transform.arXiv preprint arXiv:1510.07350, 2015. 26 REZA GHEISSARI, CURTIS GRANT, AND TIANMIN YU

work page arXiv 2015
[24]

Local semicircle law under moment conditions

Friedrich G¨ otze, Alexey Naumov, and Alexander Tikhomirov. Local semicircle law under moment conditions. part ii: Localization and delocalization.arXiv preprint arXiv:1511.00862, 2015

work page arXiv 2015
[25]

Brice Huang, Sidhanth Mohanty, Amit Rajaraman, and David X. Wu. Weak poincar´ e inequalities, simulated annealing, and sampling from spherical spin glasses, 2024

2024
[26]

Elsevier, 2014

Nobuyuki Ikeda and Shinzo Watanabe.Stochastic differential equations and diffusion processes, volume 24. Elsevier, 2014

2014
[27]

Department of Statistics, Stanford Uni- versity, 2000

Iain Johnstone.On the distribution of the largest principal component. Department of Statistics, Stanford Uni- versity, 2000

2000
[28]

Fundamental limits of symmetric low-rank matrix estimation

Marc Lelarge and L´ eo Miolane. Fundamental limits of symmetric low-rank matrix estimation. InConference on Learning Theory, pages 1297–1301. PMLR, 2017

2017
[29]

American Mathematical Soc., , 2017

David A Levin and Yuval Peres.Markov chains and mixing times, volume 107. American Mathematical Soc., , 2017

2017
[30]

High-dimensional asymptotics of langevin dynamics in spiked matrix models.Information and Inference: A Journal of the IMA, 12(4):2720–2752, 10 2023

Tengyuan Liang, Subhabrata Sen, and Pragya Sur. High-dimensional asymptotics of langevin dynamics in spiked matrix models.Information and Inference: A Journal of the IMA, 12(4):2720–2752, 10 2023

2023
[31]

Large deviations for the largest eigenvalue of rank one deformations of gaussian ensembles

Myl` ene Maida. Large deviations for the largest eigenvalue of rank one deformations of gaussian ensembles. Electronic Journal of Probability, 12:1131–1150, 2007

2007
[32]

A statistical model for tensor pca.Advances in neural information pro- cessing systems, 27, 2014

Andrea Montanari and Emile Richard. A statistical model for tensor pca.Advances in neural information pro- cessing systems, 27, 2014

2014
[33]

P´ ech´ e

S. P´ ech´ e. The largest eigenvalue of small rank perturbations of hermitian random matrices.Probability Theory and Related Fields, 134(1):127–173, Jan 2006

2006
[34]

Wein, Afonso S

Amelia Perry, Alexander S Wein, Afonso S Bandeira, and Ankur Moitra. Optimality and sub-optimality of pca for spiked random matrices and synchronization.arXiv preprint arXiv:1609.05573, 2016

work page arXiv 2016
[35]

Dynamic theory of the spin-glass phase.Physical Review Letters, 47(5):359, 1981

Haim Sompolinsky and Annette Zippelius. Dynamic theory of the spin-glass phase.Physical Review Letters, 47(5):359, 1981

1981
[36]

Relaxational dynamics of the edwards-anderson model and the mean- field theory of spin-glasses.Physical Review B, 25(11):6860, 1982

Haim Sompolinsky and Annette Zippelius. Relaxational dynamics of the edwards-anderson model and the mean- field theory of spin-glasses.Physical Review B, 25(11):6860, 1982

1982
[37]

Analyzing dynamics and average case complexity in the spherical sherrington-kirkpatrick model: a focus on extreme eigenvectors.arXiv preprint arXiv:2401.03668, 2024

Tingzhou Yu. Analyzing dynamics and average case complexity in the spherical sherrington-kirkpatrick model: a focus on extreme eigenvectors.arXiv preprint arXiv:2401.03668, 2024. AppendixA.Deferred equilibrium estimates for the spiked matrix model In this section we shall prove Lemmas 4.1 and 4.2, as well as Lemma 1.4. Throughout we will use the shorthand...

work page arXiv 2024