REVIEW 2 major objections 3 minor 2 cited by
The paper proves that for any C² payoff on a torus, the mean-field Langevin descent-ascent dynamics converges exponentially fast to the unique mixed Nash equilibrium, provided the initial strategies are sufficiently close in Wasserstein dis
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
T0 review · deepseek-v4-flash
2026-08-03 05:36 UTC pith:WKYHCY3W
load-bearing objection Real local stability theorem for MFL-DA, but the finite-particle claim in the title/abstract is not delivered anywhere in the paper. the 2 major comments →
Local exponential stability of mean-field Langevin descent-ascent and associated particle system
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
Theorem 1.1 asserts that, on the flat torus, for any C² payoff f, the unique mixed Nash equilibrium (μ*, ν*) of the entropy-regularized game is locally exponentially attracting for the MFL-DA flow. There exist δ, λ, C>0 such that if W₂(μ₀,μ*) + W₂(ν₀,ν*) < δ, then W₂(μ_t,μ*) + W₂(ν_t,ν*) ≤ C e^{-λt} times the initial distance, and the densities converge in C^{1,α} for every α∈(0,1). The proof shows that the second variation of the free energy at the equilibrium is uniformly coercive in the Wasserstein geometry, with a coercivity constant λ_gap > 0 coming from the spectral gap of the linearized operator; this local coercivity extends to a neighborhood, yielding a local evolution-variational i
What carries the argument
The key machinery is the linearized operator L acting on gradient vector fields, LΦ = -ρ*^{-1}∇·(ρ*∇Φ) + ∇²V_{ν*}·Φ, defined at the equilibrium density ρ* with effective potential V_{ν*}. A spectral gap argument shows its smallest eigenvalue λ_gap is strictly positive, which upgrades the merely non-negative second variation of the entropy-regularized payoff to a uniform coercivity on the Wasserstein tangent space. This coercivity, combined with a continuity argument (Lemma 3.3), yields a locally strong displacement convex–concave structure, and the resulting evolution variational inequalities imply exponential contraction.
Load-bearing premise
The proof requires the mixed Nash equilibrium to have a smooth, uniformly positive density, so that the linearized operator has a positive spectral gap; if the equilibrium density is not uniformly bounded away from zero, the local coercivity that drives exponential convergence is no longer guaranteed.
What would settle it
Compute the spectral gap λ_gap for a specific C² payoff on the torus and simulate MFL-DA from a small random perturbation of the equilibrium; if the observed rate is not at least close to λ_gap, or if trajectories escape for arbitrarily small perturbations, the theorem's conclusion would be refuted. A more direct test: construct a C² payoff whose equilibrium density vanishes somewhere; if the dynamics still shows exponential convergence, the uniform-positivity assumption is not necessary.
If this is right
- Local convergence with quantitative rate: any trajectory starting within a small Wasserstein neighbourhood of the equilibrium converges exponentially, and the rate λ can be chosen arbitrarily close to the spectral gap λ_gap.
- Stronger notions of convergence: since C⁰ convergence of densities controls the Nikaido–Isoda error, the result also gives convergence of the game-theoretic payoff gap, not just Wasserstein distance.
- Robustness to geometry: the proof extends to any compact Riemannian manifold without boundary, with the same coercivity identity, so the stability is not an artifact of the flat torus.
- Global result under convexity: if the payoff is λ-convex–concave, the local argument becomes global, giving exponential convergence from arbitrary initial data.
- Characterization of the rate: λ_gap is explicitly given as a Rayleigh quotient involving ∇²φ and the Hessian of the effective potential, so the convergence rate is computable in principle.
Where Pith is reading between the lines
- The spectral-gap mechanism suggests a definition of a 'condition number' for mean-field zero-sum games, analogous to strong-convexity constants; games with larger λ_gap should admit faster particle-based algorithms.
- The same coercivity framework could be applied to discrete-time and particle approximations, potentially yielding non-asymptotic convergence rates uniform in the number of particles, provided the mean-field limit is stable.
- The local convex-concave structure implies that heuristics from finite-dimensional GDA—such as momentum or adaptive step sizes—may transfer to the mean-field setting near equilibrium, offering a design principle for practical algorithms.
- If the equilibrium density degenerates (approaches zero), the spectral gap may shrink to zero; exploring such boundary cases could reveal sharp thresholds for local stability.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proves that the mean-field Langevin descent-ascent (MFL-DA) dynamics, on the flat torus and for a C^2 payoff f, is locally exponentially stable around its unique mixed Nash equilibrium: if the initial pair (μ0,ν0) is sufficiently close to (μ*,ν*) in W2, then the W2 distance of the solution decays exponentially at a quantitative rate, and the densities converge in C^{1,α}. The proof proceeds by establishing a positive spectral gap λ_gap for the linearized Wasserstein Hessian at equilibrium (Lemma 3.1), extending this to a local convex–concave coercivity in a W^{2,p} neighborhood (Lemma 3.3), deriving a local contraction estimate (Proposition 3.4), and bootstrapping this via smoothing and interpolation lemmas to obtain the final theorem. The abstract and title additionally claim that the finite-N particle system inherits this stability, but no such result is proved in the manuscript.
Significance. If the PDE-level theorem is correct—and the presented proofs are detailed and internally consistent—it gives a substantive positive answer to the local-stability and quantitative-rate parts of the Wang–Chizat open problem. The key mechanism, extracting a coercivity estimate for the entropy near equilibrium via spectral analysis of the linearized operator, is a useful and potentially transferable idea. The full appendix proofs are a strength: Lemmas 3.1, 3.3, 3.6, 3.7 and Proposition 3.4 are proved in detail, and the bootstrap argument in Theorem 1.1 is explicit. However, the advertised finite-N inheritance claim is not delivered anywhere in the paper, which significantly affects the paper's stated contribution as written.
major comments (2)
- [Abstract and Section 4] The abstract states: 'We further show that the finite-N particle system inherits this stability up to times exponential in N, with an N-independent exponential rate modulo a finite-particle error floor.' The title similarly advertises 'and associated particle system'. However, no theorem, proposition, or lemma in the main text or in Appendices A–B states or proves any finite-N inheritance result. Section 4 explicitly lists as the third open question 'whether the local stability proved here at the PDE level can be transferred to the finite-particle dynamics, yielding convergence rates that are uniform in the number of particles.' This is an explicit admission that the advertised finite-N result is not delivered. This is not a technical flaw in the PDE theorem, but it is a substantial overstatement of the paper's contributions. The authors must either supply a proof of the finite-N stabili
- [Theorem 1.1 / proof of (1.5)] The proof of Theorem 1.1 uses Lemma 3.5 with λ' possibly negative and then bootstraps from t≥1. In the displayed line before (1.5), the authors write W2^2(t) < e^{-2λ'} (initial) e^{-2λ(t-1)}, which is consistent only if the initial W2^2 is smaller than e^{2λ'} δ; this is stated earlier as W2^2(μ0,μ*)+W2^2(ν0,ν*) < e^{2λ'}δ. The argument is coherent, but the constant C in (1.5) depends on the possibly large factor e^{-2λ'}. The theorem requires only existence of some C, so this is not an error; however, the presentation could mislead a reader into thinking C is uniform in f. Please clarify that C is allowed to depend on f (including through λ'), as the abstract already implies.
minor comments (3)
- [Lemma 3.5 and footnote 7] The assumption '∇²yy f(x,y) ≤ −λ'I' combined with footnote 7 ('∇²yy f + λ'I is nonpositive definite') is confusing when λ' is allowed to be negative. Since the lemma is used for a possibly negative λ', the intended interpretation is that λ'-displacement concavity is meant in the sense of Definition A.25. Please state the hypothesis in the language of Definition A.25 or add a sentence clarifying the sign convention for negative λ'.
- [Remark 3.8] The claim W2(μt,μ*)+W2(νt,ν*) = O(e^{-λt}) for all λ ∈ (0, λ_gap) is stated as a remark following the theorem. This is a consequence of (1.5) with C depending on λ, but the asymptotic notation O(e^{-λt}) usually hides constants that can depend on λ; please make this dependence explicit to avoid ambiguity.
- [Section 3.1, Eq. (3.1)] The operator L is introduced formally as an object on gradient vector fields. In Remark B.1, the authors discuss a spectral decomposition and an L²μ*-orthogonal Hodge projection, but this is not used in the proof. This remark is interesting but somewhat speculative; consider shortening or marking it as a heuristic aside.
Circularity Check
No circular derivation: the mean-field stability proof is self-contained; minor self-citations are re-proved. The abstract's finite-N claim is unsupported and explicitly left open in the Conclusion, which is a claims mismatch, not circularity.
full rationale
The central chain (Lemma 3.1 -> Lemma 3.3 -> Proposition 3.4 -> Theorem 1.1) is internally consistent and does not reduce to its inputs. lambda_gap is defined as a variational Rayleigh quotient in (B.7) and proved positive by compactness/ellipticity; it is not fitted to the convergence data, and the exponential rate is then derived rather than assumed. The local convex-concavity lemma is obtained by continuity of the second variation, and the bootstrap uses external smoothing/interpolation facts proved in Appendix B.3. The only self-citations, mainly 'c.f. [CJS25]' in Lemmas A.20-A.21 and Lemma 3.6, are not load-bearing because the cited statements are re-proved in the appendices; hence they do not create circularity. The abstract's sentence 'We further show that the finite-N particle system inherits this stability up to times exponential in N...' has no matching theorem, and Section 4 explicitly lists as open 'whether the local stability proved here at the PDE level can be transferred to the finite-particle dynamics, yielding convergence rates that are uniform in the number of particles.' This is an overstatement of delivered results, not a circularity. Appendix A.10 similarly notes that R^n requires extra coercivity assumptions; this is an honest limitation. Overall no circular step is present; the low score reflects only the minor self-citation burden and the flagged claims-content mismatch.
Axiom & Free-Parameter Ledger
axioms (4)
- domain assumption Unique mixed Nash equilibrium exists and its densities are proportional to e^{-V}, with C², uniformly positive densities on the torus.
- standard math Second variation formula (B.4) and the integration-by-parts identity (B.6) for Gibbs measures.
- standard math Spectral compactness / Poincaré inequality for the weighted Laplacian; Holley–Stroock perturbation bound.
- domain assumption Sobolev stability of optimal transport maps (Lemma A.20) and parabolic smoothing/interpolation (Lemmas 3.6–3.7).
Cite this review
Pith. "Pith review of Local exponential stability of mean-field Langevin descent-ascent and associated particle system." pith.science (2026). https://pith.science/paper/WKYHCY3W
@misc{pith2026260201564,
author = {Pith},
title = {Pith review of: Local exponential stability of mean-field Langevin descent-ascent and associated particle system},
year = {2026},
howpublished = {\url{https://pith.science/paper/WKYHCY3W}},
note = {Machine review of arXiv:2602.01564}
}
read the original abstract
We study the mean-field Langevin descent-ascent (MFL-DA), a coupled optimization dynamics on the space of probability measures for entropically regularized two-player zero-sum games, together with its associated interacting particle system. For general nonconvex-nonconcave payoffs, Wang and Chizat (COLT 2024) asked whether the original single-timescale MFL-DA converges to the mixed Nash equilibrium and, if so, at what rate. We prove a local affirmative answer in Wasserstein space: if the initial datum is sufficiently close to the mixed Nash equilibrium, then the mean-field dynamics converges to it exponentially fast at a quantitative rate. We further show that the finite-$N$ particle system inherits this stability up to times exponential in $N$, with an $N$-independent exponential rate modulo a finite-particle error floor. Combined with the recent counterexample of Mourrat and Pillaud-Vivien for MFL-DA, which shows that global convergence cannot hold in general, our theorem completes the positive local counterpart of the Wang-Chizat question: the mixed Nash equilibrium has a robust basin of attraction, stable under both the mean-field flow and its finite-particle approximation.
Forward citations
Cited by 2 Pith papers
-
PL conditions do not guarantee convergence of gradient descent-ascent dynamics
There exists a function satisfying two-sided PL conditions for which GDA dynamics fail to converge by circling the saddle point.
-
Oscillating solutions to the mean-field Langevin descent-ascent flow
For certain double-well payoff functions with strong coupling and small noise, the mean-field Langevin descent-ascent flow stays near a limit cycle and fails to converge.
Reference graph
Works this paper leans on
-
[1]
Wasserstein generative adversarial networks
Martin Arjovsky, Soumith Chintala, and L \'e on Bottou. Wasserstein generative adversarial networks. In International conference on machine learning , pages 214--223. PMLR, 2017
2017
-
[2]
u rich. Birkh \
Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar \'e . Gradient Flows: In Metric Spaces and in the Space of Probability Measures . Lectures in Mathematics. ETH Z \"u rich. Birkh \"a user Basel, 2 edition, 2008
2008
-
[3]
Jing An and Jianfeng Lu. Convergence of two-timescale gradient descent ascent dynamics: finite-dimensional and mean-field perspectives. arXiv preprint arXiv:2501.17122 , 2025
arXiv 2025
-
[4]
Analysis and geometry of Markov diffusion operators , volume 348
Dominique Bakry, Ivan Gentil, and Michel Ledoux. Analysis and geometry of Markov diffusion operators , volume 348. Springer Science & Business Media, 2013
2013
-
[5]
V.S. Borkar. Stochastic Approximation: A Dynamical Systems Viewpoint . Cambridge University Press, 2008
2008
-
[6]
Functional Analysis, Sobolev Spaces and Partial Differential Equations
Haim Brezis. Functional Analysis, Sobolev Spaces and Partial Differential Equations . Universitext. Springer, New York, 2011
2011
-
[7]
On the global convergence of gradient descent for over-parameterized models using optimal transport
Lenaic Chizat and Francis Bach. On the global convergence of gradient descent for over-parameterized models using optimal transport. Advances in neural information processing systems , 31, 2018
2018
-
[8]
Mean-field langevin dynamics: Exponential convergence and annealing
L \'e na \" c Chizat. Mean-field langevin dynamics: Exponential convergence and annealing. Transactions on Machine Learning Research , 2022
2022
-
[9]
Coupled wasserstein gradient flows for min-max and cooperative games
Lauren Conger, Franca Hoffmann, Eric Mazumdar, and Lillian J Ratliff. Coupled wasserstein gradient flows for min-max and cooperative games. arXiv preprint arXiv:2411.07403 , 2024
Pith/arXiv arXiv 2024
-
[10]
Wasserstein- ojasiewicz inequalities and asymptotics of mckean-vlasov equation
Beomjun Choi, Seunghoon Jeong, and Geuntaek Seo. Wasserstein- ojasiewicz inequalities and asymptotics of mckean-vlasov equation. arXiv preprint arXiv:2511.23361 , 2025
arXiv 2025
-
[11]
Carrillo, Robert J
Jos\'e A. Carrillo, Robert J. McCann, and C\'edric Villani. Kinetic equilibration rates for granular media and related equations: entropy dissipation and mass transportation estimates. Rev. Mat. Iberoamericana , 19(3):971--1018, 2003
2003
-
[12]
On the convergence of min-max langevin dynamics and algorithm
Yang Cai, Siddharth Mitra, Xiuyuan Wang, and Andre Wibisono. On the convergence of min-max langevin dynamics and algorithm. In Proceedings of Thirty Eighth Conference on Learning Theory , volume 291 of PMLR , pages 677--754, 2025
2025
-
[13]
A mean-field analysis of two-player zero-sum games
Carlos Domingo-Enrich, Kilian Fatras, Nicolas Roux, and Gauthier Gidel. A mean-field analysis of two-player zero-sum games. Advances in Neural Information Processing Systems , 33:2020--2031, 2020
2020
-
[14]
Convergence rates of two-time-scale gradient descent-ascent dynamics for solving nonconvex min-max problems
Thinh Doan. Convergence rates of two-time-scale gradient descent-ascent dynamics for solving nonconvex min-max problems. In Learning for Dynamics and Control Conference , pages 192--206. PMLR, 2022
2022
-
[15]
Reducing exit-times of diffusions with repulsive interactions
Paul-Eric Chaudru de Raynal, Manh Hong Duong, Pierre Monmarch \'e , Milica Toma s evi \'c , and Julian Tugaut. Reducing exit-times of diffusions with repulsive interactions. ESAIM: Probability and Statistics , 27:723--748, 2023
2023
-
[16]
The complexity of constrained min-max optimization
Constantinos Daskalakis, Stratis Skoulakis, and Manolis Zampetakis. The complexity of constrained min-max optimization. In Proceedings of the 53rd Annual ACM SIGACT Symposium on Theory of Computing , pages 1466--1478, 2021
2021
-
[17]
Quantitative harris-type theorems for diffusions and mckean--vlasov processes
Andreas Eberle, Arnaud Guillin, and Raphael Zimmer. Quantitative harris-type theorems for diffusions and mckean--vlasov processes. Transactions of the American Mathematical Society , 371(10):7135--7173, 2019
2019
-
[18]
Lawrence C. Evans. Partial Differential Equations , volume 19 of Graduate Studies in Mathematics . American Mathematical Society, Providence, RI, 2 edition, 2010
2010
-
[19]
Do gans always have nash equilibria? In International Conference on Machine Learning , pages 3029--3039
Farzan Farnia and Asuman Ozdaglar. Do gans always have nash equilibria? In International Conference on Machine Learning , pages 3029--3039. PMLR, 2020
2020
-
[20]
A certain class of diffusion processes associated with nonlinear parabolic equations
Tadahisa Funaki. A certain class of diffusion processes associated with nonlinear parabolic equations. Zeitschrift f \"u r Wahrscheinlichkeitstheorie und Verwandte Gebiete , 67(3):331--348, 1984
1984
-
[21]
I. L. Glicksberg. A further generalization of the kakutani fixed point theorem, with application to nash equilibrium points. Proceedings of the American Mathematical Society , 3(1):170--174, 1952
1952
-
[22]
Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio
Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems 27 , pages 2672--2680, 2014
2014
-
[23]
Gilbarg and N
D. Gilbarg and N. S. Trudinger. Elliptic partial differential equations of second order . Classics in Mathematics. Springer-Verlag, Berlin, 2001. Reprint of the 1998 edition
2001
-
[24]
Xing Huang, Eva Kopfer, Pierre Monmarch \'e , and Panpan Ren. Exponential ergodicity in relative entropy and l^ 2 -wasserstein distance for non-equilibrium partially dissipative kinetic sdes. arXiv preprint arXiv:2507.02518 , 2025
Pith/arXiv arXiv 2025
-
[25]
Finding mixed N ash equilibria of generative adversarial networks
Ya-Ping Hsieh, Chen Liu, and Volkan Cevher. Finding mixed N ash equilibria of generative adversarial networks. In Proceedings of the 36th International Conference on Machine Learning , volume 97 of PMLR , pages 2810--2819, 2019
2019
-
[26]
Mean-field langevin dynamics and energy landscape of neural networks
Kaitong Hu, Zhenjie Ren, David S i s ka, and ukasz Szpruch. Mean-field langevin dynamics and energy landscape of neural networks. Annales de l'Institut Henri Poincare (B) Probabilites et statistiques , 57(4):2043--2065, 2021
2043
-
[27]
On gradient descent-ascent flows in metric spaces
Noboru Isobe and Sho Shimoyama. On gradient descent-ascent flows in metric spaces. arXiv preprint arXiv:2506.20258 , 2025
Pith/arXiv arXiv 2025
-
[28]
What is local optimality in nonconvex-nonconcave minimax optimization? In International conference on machine learning , pages 4880--4889
Chi Jin, Praneeth Netrapalli, and Michael Jordan. What is local optimality in nonconvex-nonconcave minimax optimization? In International conference on machine learning , pages 4880--4889. PMLR, 2020
2020
-
[29]
Symmetric mean-field langevin dynamics for minimax optimization
Minyoung Kim, Juho Lee, and Taiji Suzuki. Symmetric mean-field langevin dynamics for minimax optimization. In International Conference on Learning Representations , 2024
2024
-
[30]
N. V. Krylov. Lectures on Elliptic and Parabolic Equations in Sobolev Spaces , volume 96 of Graduate Studies in Mathematics . American Mathematical Society, Providence, RI, 2008
2008
-
[31]
Transformers learn nonlinear features in context: nonconvex mean-field dynamics on the attention landscape
Juno Kim and Taiji Suzuki. Transformers learn nonlinear features in context: nonconvex mean-field dynamics on the attention landscape. In Proceedings of the 41st International Conference on Machine Learning , pages 24527--24561, 2024
2024
-
[32]
Second order parabolic differential equations
Gary M Lieberman. Second order parabolic differential equations . World scientific, 1996
1996
-
[33]
Convergence of mean-field L angevin stochastic descent-ascent for distributional minimax optimization
Zhangyi Liu, Feng Liu, Rui Gao, and Shuang Li. Convergence of mean-field L angevin stochastic descent-ascent for distributional minimax optimization. In Proceedings of the 42nd International Conference on Machine Learning , volume 267 of PMLR , pages 38869--38893, 13--19 Jul 2025
2025
-
[34]
Yulong Lu and Pierre Monmarch \'e . Convergence of time-averaged mean field gradient descent dynamics for continuous multi-player zero-sum games. arXiv preprint arXiv:2505.07642 , 2025
Pith/arXiv arXiv 2025
-
[35]
Some geometric calculations on W asserstein space
John Lott. Some geometric calculations on W asserstein space. Communications in Mathematical Physics , 277:423--437, 2008
2008
-
[36]
Two-scale gradient descent ascent dynamics finds mixed N ash equilibria of continuous games: A mean-field perspective
Yulong Lu. Two-scale gradient descent ascent dynamics finds mixed N ash equilibria of continuous games: A mean-field perspective. In International Conference on Machine Learning , pages 22790--22811. PMLR, 2023
2023
-
[37]
A convexity principle for interacting gases
Robert J McCann. A convexity principle for interacting gases. Advances in mathematics , 128(1):153--179, 1997
1997
-
[38]
S. Mei, A. Montanari, and P.-M. Nguyen. A mean field view of the landscape of two-layer neural networks. Proceedings of the National Academy of Sciences , 115(33):E7665--E7671, 2018
2018
-
[39]
The numerics of gans
Lars Mescheder, Sebastian Nowozin, and Andreas Geiger. The numerics of gans. Advances in neural information processing systems , 30, 2017
2017
-
[40]
Local convergence rates for wasserstein gradient flows and mckean-vlasov equations with multiple stationary solutions
Pierre Monmarch \'e and Julien Reygner. Local convergence rates for wasserstein gradient flows and mckean-vlasov equations with multiple stationary solutions. Probability Theory and Related Fields , 2025
2025
-
[41]
Learning in games via reinforcement and regularization
Panayotis Mertikopoulos and William H Sandholm. Learning in games via reinforcement and regularization. Mathematics of Operations Research , 41(4):1297--1324, 2016
2016
-
[42]
On finding local nash equilibria (and only local nash equilibria) in zero-sum games
Eric Mazumdar, S Shankar Sastry, and Michael I Jordan. On finding local nash equilibria (and only local nash equilibria) in zero-sum games. ACM/IMS Journal of Data Science , 2(2):1--26, 2025
2025
-
[43]
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. arXiv preprint arXiv:2202.10947 , 2022
Pith/arXiv arXiv 2022
-
[44]
On elliptic partial differential equations
Louis Nirenberg. On elliptic partial differential equations. Annali della Scuola Normale Superiore di Pisa-Scienze Fisiche e Matematiche , 13(2):115--162, 1959
1959
-
[45]
Convex analysis of the mean field langevin dynamics
Atsushi Nitanda, Denny Wu, and Taiji Suzuki. Convex analysis of the mean field langevin dynamics. In International Conference on Artificial Intelligence and Statistics , pages 9741--9757. PMLR, 2022
2022
-
[46]
The geometry of dissipative evolution equations: the porous medium equation
Felix Otto. The geometry of dissipative evolution equations: the porous medium equation. Comm. Partial Differential Equations , 26(1-2):101--174, 2001
2001
-
[47]
Comparison between w2 distance and \.h - 1 norm, and localization of wasserstein distance
R \'e mi Peyre. Comparison between w2 distance and \.h - 1 norm, and localization of wasserstein distance. ESAIM: Control, Optimisation and Calculus of Variations , 24(4):1489--1501, 2018
2018
-
[48]
The Fokker-Planck Equation: Methods of Solution and Applications
Hannes Risken. The Fokker-Planck Equation: Methods of Solution and Applications . Springer Series in Synergetics. Springer, Berlin; New York, 2nd edition, 1989
1989
-
[49]
Population games and evolutionary dynamics
William H Sandholm. Population games and evolutionary dynamics . MIT press, 2010
2010
-
[50]
Convergence of mean-field langevin dynamics: time-space discretization, stochastic gradient, and variance reduction
Taiji Suzuki, Denny Wu, and Atsushi Nitanda. Convergence of mean-field langevin dynamics: time-space discretization, stochastic gradient, and variance reduction. Advances in Neural Information Processing Systems , 36:15545--15577, 2023
2023
-
[51]
Well-posedness of multidimensional diffusion processes with weakly differentiable coefficients
Dario Trevisan. Well-posedness of multidimensional diffusion processes with weakly differentiable coefficients. Electronic Journal of Probability , 21(22):1--41, 2016
2016
-
[52]
C. Villani. Optimal Transport: Old and New . Grundlehren der mathematischen Wissenschaften. Springer Berlin Heidelberg, 2008
2008
-
[53]
Open problem: Convergence of single-timescale mean-field langevin descent-ascent for two-player zero-sum games
Guillaume Wang and L \'e na\" i c Chizat. Open problem: Convergence of single-timescale mean-field langevin descent-ascent for two-player zero-sum games. In Proceedings of Thirty Seventh Conference on Learning Theory , volume 247 of PMLR , pages 5345--5350, 2024
2024
-
[54]
An exponentially converging particle method for the mixed nash equilibrium of continuous games
Guillaume Wang and L \'e na \" c Chizat. An exponentially converging particle method for the mixed nash equilibrium of continuous games. Open Journal of Mathematical Optimization , 6:1--66, January 2025
2025
-
[55]
Global convergence and variance reduction for a class of nonconvex-nonconcave minimax problems
Junchi Yang, Negar Kiyavash, and Niao He. Global convergence and variance reduction for a class of nonconvex-nonconcave minimax problems. Advances in neural information processing systems , 33:1153--1165, 2020
2020
-
[56]
Optimal epoch stochastic gradient descent ascent methods for min-max optimization
Yan Yan, Yi Xu, Qihang Lin, Wei Liu, and Tianbao Yang. Optimal epoch stochastic gradient descent ascent methods for min-max optimization. Advances in Neural Information Processing Systems , 33:5789--5800, 2020
2020
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.