arxiv: 2604.11134 · v1 · submitted 2026-04-13 · 🧮 math.OC · math.PR

Recognition: unknown

Oscillating solutions to the mean-field Langevin descent-ascent flow

Jean-Christophe Mourrat , Loucas Pillaud-Vivien

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:43 UTC · model grok-4.3

classification 🧮 math.OC math.PR

keywords mean-field Langevin dynamicsdescent-ascent flowlimit cyclenon-convergenceentropic regularizationstochastic differential equationszero-sum gamesdouble-well potentials

0 comments

The pith

The mean-field Langevin descent-ascent flow on R squared can oscillate indefinitely instead of converging.

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

This paper constructs a counterexample to convergence claims for the mean-field Langevin descent-ascent flow. It uses payoff functions shaped as double wells in each coordinate so that the underlying deterministic dynamics has a limit cycle. When coupling between the coordinates is strong enough and entropic regularization is small enough, the stochastic mean-field version stays close to the cycle and keeps oscillating rather than reaching a stationary equilibrium.

Core claim

We present a counterexample to the statement of convergence of the mean-field Langevin descent-ascent flow on R^2. We consider payoff functions that are shaped as a double well in each coordinate, and for which the deterministic dynamics admits a limit cycle. When the coupling between the two coordinates is sufficiently strong and the entropic regularization sufficiently small, we show that the mean-field dynamics remains close to this cyclic behavior, and in particular, does not converge.

What carries the argument

The mean-field Langevin descent-ascent flow, a system of stochastic differential equations governing the evolution of two interacting measures under regularized gradient descent and ascent.

If this is right

The flow fails to converge and instead exhibits persistent oscillations near the deterministic limit cycle.
This non-convergence holds when inter-coordinate coupling exceeds a sufficient threshold.
The behavior appears when the entropic regularization parameter falls below a sufficient threshold.
The construction serves as a concrete counterexample to any claim of unconditional convergence for the flow.

Where Pith is reading between the lines

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

Convergence proofs for related mean-field game flows may need explicit bounds on coupling strength relative to regularization to exclude cycles.
Similar sustained oscillations could appear in discrete-time or finite-player approximations of the same payoff structure.
The counterexample raises the question of whether other regularizations or payoff shapes with deterministic cycles produce non-convergent mean-field behavior.

Load-bearing premise

The payoffs must be double wells in each coordinate so the deterministic dynamics has a limit cycle, and this cycle persists in the mean-field flow only for strong enough coupling and small enough regularization.

What would settle it

Run a numerical simulation of the mean-field dynamics with the double-well payoffs, strong coupling parameter, and small regularization parameter to check whether the measures sustain oscillations near the deterministic cycle or eventually converge to a fixed point.

Figures

Figures reproduced from arXiv: 2604.11134 by Jean-Christophe Mourrat, Loucas Pillaud-Vivien.

**Figure 2.** Figure 2: Particle snapshots for ε = 0.25 (left) and ε = 0.5 (right) at times t = 0, 5, 12.5, 20. The red points represent the positions of the N = 500 particles, the blue dot stands for the solution of the deterministic trajectory at time t. Its trajectory is displayed as a black dotted line, which merges with the limit cycle represented as a continuous black curve [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

read the original abstract

We present a counterexample to the statement of convergence of the mean-field Langevin descent-ascent flow on $\mathbb{R}^2$. We consider payoff functions that are shaped as a double well in each coordinate, and for which the deterministic dynamics admits a limit cycle. When the coupling between the two coordinates is sufficiently strong and the entropic regularization sufficiently small, we show that the mean-field dynamics remains close to this cyclic behavior, and in particular, does not converge.

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 claimed non-convergence for small positive regularization contradicts the convergence of the Fokker-Planck equation to its unique stationary measure.

read the letter

The paper gives an explicit family of double-well payoff functions on R^2 whose deterministic descent-ascent dynamics has a limit cycle. It then argues that when the coordinates are strongly coupled and the entropic regularization is small, the mean-field Langevin flow stays close to that cycle and therefore does not converge. The construction itself is the useful part. Having a concrete, parameter-free example where the zero-noise flow cycles is something people can check and build on. The difficulty is with the positive-regularization claim. The mean-field flow is the Fokker-Planck equation for the SDE with the given drift plus sqrt(2 reg) noise. For reg > 0 this is a linear parabolic PDE. Under the dissipativity conditions that the double wells plus strong coupling should satisfy, it has a unique stationary probability measure, and solutions converge to it. That rules out persistent closeness to a non-stationary cycle. The paper may be showing that the time to escape the cycle grows as reg goes to zero, but the stated conclusion that the dynamics does not converge does not follow for any fixed positive reg. This matters for anyone studying convergence of mean-field methods in games and optimization. The deterministic construction is worth seeing, but the main claim does not survive the standard long-time analysis of the Fokker-Planck equation. I would not bring the paper to a reading group and would not cite it. It does not look ready for serious refereeing; the authors would need to fix the statement about non-convergence or limit the result to the deterministic case.

Referee Report

2 major / 2 minor

Summary. The paper constructs explicit double-well payoff functions on R^2 for which the deterministic descent-ascent vector field admits a limit cycle. It claims that, for sufficiently strong inter-coordinate coupling and sufficiently small but positive entropic regularization, the associated mean-field Langevin descent-ascent flow remains close to this cycle for all time and therefore does not converge.

Significance. If the central claim holds, the explicit construction would provide a concrete counterexample to convergence statements for mean-field Langevin flows in min-max settings, showing that persistent oscillations are possible even with positive regularization. The use of standard SDE analysis on a parameter-free example is a strength.

major comments (2)

[Abstract] Abstract and the statement of the main result: the assertion that the flow 'does not converge' for any positive regularization contradicts the standard well-posedness theory for the Fokker-Planck equation associated to the SDE dZ_t = F(Z_t) dt + sqrt(2 reg) dW_t. Under the dissipativity induced by the double-well potentials plus strong coupling, this linear parabolic PDE admits a unique invariant probability measure to which every solution converges in total variation (or Wasserstein distance) as t → ∞; no non-constant time-periodic orbit in the space of measures can exist. The manuscript must either (i) show that its definition of the mean-field flow differs from this SDE or (ii) clarify what 'does not converge' means in light of this convergence.
[Main result / limiting argument] The limiting argument for closeness to the deterministic cycle (presumably in the section containing the main theorem): even if trajectories remain close to the cycle on finite-time horizons when reg is small, the long-time convergence of the law to the unique invariant measure (supported near the cycle) still occurs. The manuscript needs to reconcile the claimed perpetual closeness with this ergodicity; otherwise the non-convergence claim is not supported.

minor comments (2)

Notation for the regularization parameter should be introduced once and used consistently; avoid switching between 'reg', 'ε', etc.
If numerical illustrations of the oscillating behavior are included, add a brief description of the discretization scheme and time horizon used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and for identifying the need for greater precision in our statements about non-convergence. We address each major comment below.

read point-by-point responses

Referee: [Abstract] Abstract and the statement of the main result: the assertion that the flow 'does not converge' for any positive regularization contradicts the standard well-posedness theory for the Fokker-Planck equation associated to the SDE dZ_t = F(Z_t) dt + sqrt(2 reg) dW_t. Under the dissipativity induced by the double-well potentials plus strong coupling, this linear parabolic PDE admits a unique invariant probability measure to which every solution converges in total variation (or Wasserstein distance) as t → ∞; no non-constant time-periodic orbit in the space of measures can exist. The manuscript must either (i) show that its definition of the mean-field flow differs from this SDE or (ii) clarify what 'does not converge' means in light of this convergence.

Authors: We agree that the law of the process converges to the unique invariant measure of the associated Fokker-Planck equation. Our use of 'does not converge' is meant to indicate that the dynamics do not converge to a stationary Nash equilibrium (i.e., a Dirac measure supported at an equilibrium point of the game). Instead, for small regularization and strong coupling, the invariant measure is supported near the deterministic limit cycle, so that sample paths exhibit persistent oscillations around the cycle. We will revise the abstract and main-result statement to make this distinction explicit, thereby addressing option (ii). We do not assert that the underlying SDE differs from the standard form. revision: yes
Referee: [Main result / limiting argument] The limiting argument for closeness to the deterministic cycle (presumably in the section containing the main theorem): even if trajectories remain close to the cycle on finite-time horizons when reg is small, the long-time convergence of the law to the unique invariant measure (supported near the cycle) still occurs. The manuscript needs to reconcile the claimed perpetual closeness with this ergodicity; otherwise the non-convergence claim is not supported.

Authors: We concur that the law converges to the invariant measure. Because the invariant measure concentrates near the deterministic limit cycle when regularization is sufficiently small, the process remains close to the cycle for all sufficiently large times (in the sense that its distance to the cycle is small with high probability). Ergodicity governs mixing inside the thin tube surrounding the cycle but does not contradict uniform-in-time closeness to the cycle itself. We will add a clarifying paragraph in the main-result section that reconciles the limiting argument with ergodicity and reiterates that non-convergence refers to equilibria, not to the cycle. revision: yes

Circularity Check

0 steps flagged

No circularity; explicit counterexample from concrete functions and standard analysis

full rationale

The paper constructs explicit double-well payoff functions on R^2 for which the deterministic (zero-regularization) descent-ascent dynamics admits a limit cycle. It then invokes standard SDE perturbation and Fokker-Planck estimates to show that, for sufficiently strong coupling and sufficiently small positive regularization, trajectories remain close to that cycle for all time. No step defines a quantity in terms of itself, renames a fitted parameter as a prediction, or relies on a load-bearing self-citation whose content reduces to the present claim. The derivation is self-contained against external benchmarks (classical SDE theory and parabolic PDE estimates) and does not reduce any asserted result to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard existence results for mean-field SDEs and the explicit construction of the payoff functions; no free parameters are fitted and no new entities are postulated.

axioms (1)

standard math Existence and uniqueness of solutions to the mean-field stochastic differential equation
Invoked to ensure the dynamics is well-defined before analyzing its long-time behavior.

pith-pipeline@v0.9.0 · 5369 in / 1083 out tokens · 48415 ms · 2026-05-10T15:43:41.553136+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 3 canonical work pages · 1 internal anchor

[1]

A mean-field analysis of two-player zero-sum games.Advances in neural information processing systems, 33:20215–20226, 2020

Carles Domingo-Enrich, Samy Jelassi, Arthur Mensch, Grant Rotskoff, and Joan Bruna. A mean-field analysis of two-player zero-sum games.Advances in neural information processing systems, 33:20215–20226, 2020

2020
[2]

Zurich Lectures in Advanced Mathematics

Tomas Dominguez and Jean-Christophe Mourrat.Statistical mechanics of mean-field dis- ordered systems: a Hamilton-Jacobi approach. Zurich Lectures in Advanced Mathematics. European Mathematical Society (EMS), Zürich, 2024

2024
[3]

Freidlin-Wentzell LDP in path space for McKean-Vlasov equations and the functional iterated logarithm law.Ann

Gon¸calo dos Reis, William Salkeld, and Julian Tugaut. Freidlin-Wentzell LDP in path space for McKean-Vlasov equations and the functional iterated logarithm law.Ann. Appl. Probab., 29(3):1487–1540, 2019

2019
[4]

McKean-Vlasov SDE and SPDE with locally monotone coefficients.Ann

Wei Hong, Shanshan Hu, and Wei Liu. McKean-Vlasov SDE and SPDE with locally monotone coefficients.Ann. Appl. Probab., 34(2):2136–2189, 2024

2024
[5]

Symmetric mean-field Langevin dynamics for distributional minimax problems

Juno Kim, Kakei Yamamoto, Kazusato Oko, Zhuoran Yang, and Taiji Suzuki. Symmetric mean-field Langevin dynamics for distributional minimax problems. InInternational Conference on Learning Representations, 2024

2024
[6]

Two-scale gradient descent ascent dynamics finds mixed Nash equilibria of continuous games: a mean-field perspective

Yulong Lu. Two-scale gradient descent ascent dynamics finds mixed Nash equilibria of continuous games: a mean-field perspective. InInternational Conference on Machine Learning, pages 22790–22811. PMLR, 2023

2023
[7]

Provably convergent quasistatic dynamics for mean-field two-player zero-sum games

Chao Ma and Lexing Ying. Provably convergent quasistatic dynamics for mean-field two-player zero-sum games. InInternational Conference on Learning Representations, 2022

2022
[8]

PL conditions do not guarantee convergence of gradient descent-ascent dynamics

Jean-Christophe Mourrat. PL conditions do not guarantee convergence of gradient descent-ascent dynamics. Preprint, arXiv:2602.16517

work page internal anchor Pith review arXiv
[9]

Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality.J

Felix Otto and Cédric Villani. Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality.J. Funct. Anal., 173(2):361–400, 2000

2000
[10]

Springer-Verlag, New York, third edition, 2001

Lawrence Perko.Differential equations and dynamical systems, volume 7 ofTexts in Applied Mathematics. Springer-Verlag, New York, third edition, 2001

2001
[11]

Local exponen- tial stability of mean-field Langevin descent-ascent in Wasserstein space

Geuntaek Seo, Minseop Shin, Pierre Monmarché, and Beomjun Choi. Local expo- nential stability of mean-field langevin descent-ascent in Wasserstein space. Preprint, arXiv:2602.01564

work page arXiv
[12]

Open problem: Convergence of single-timescale mean-field Langevin descent-ascent for two-player zero-sum games

Guillaume Wang and Lénaïc Chizat. Open problem: Convergence of single-timescale mean-field Langevin descent-ascent for two-player zero-sum games. InThe Thirty Seventh Annual Conference on Learning Theory, pages 5345–5350. PMLR, 2024

2024
[13]

Local convergence of mean-field Langevin dy- namics: from gradient flows to linearly monotone games

Guillaume Wang and Lénaïc Chizat. Local convergence of mean-field Langevin dynamics: from gradient flows to linearly monotone games. Preprint, arXiv:2602.11999. (Jean-Christophe Mourrat)Department of Mathematics, ENS Lyon and CNRS, Lyon, France (Loucas Pillaud-Vivien)CERMICS, CNRS, ENPC, Institut Polytechnique de Paris, Marne-la-V allée, France

work page arXiv