Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Generating Samples of Stationary Distributions of Weakly Interacting Diffusion Models Without Finite Particle Truncation: A Weak Generative Approach

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes a generative sampler that computes and samples stationary distributions of McKean–Vlasov diffusion processes directly from the weak form of the stationary nonlinear Fokker–Planck equation, with no finite-particle simulat

desk verdict Useful extension of WGS to nonlinear McKean-Vlasov stationary problems with credible low-dimensional results, but the core identifiability of the adaptive Gaussian test-function loss is unproved and the high-dimensional validation leans on the finite-particle simulations the method claims to avoid. read the letter →

arxiv 2509.12841 v2 pith:FGWQAZDV submitted 2025-09-16 physics.comp-ph

Zhiqiang Cai , Chengyu Liu , Xiang Zhou This is my paper · ORCID
classification physics.comp-ph MSC 35Q8460H1065C3082C31 PACS 05.10.Gg
keywords McKean-VlasovprocessstationaryFokker-Planckequationnormalizingflowsphasetransitionsmean-fieldinteractingparticlesweakgenerativesamplerinvariantmeasurehigh-dimensionalsampling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper extends the Weak Generative Sampler (WGS) to McKean–Vlasov processes, whose stationary law solves a nonlinear Fokker–Planck equation. It trains a normalizing flow so that pushed-forward samples satisfy the weak form of that equation, with the mean-field interaction either treated implicitly (II-WGS) or frozen through a Picard iteration (PI-WGS). The goal is to compute and sample invariant measures of the infinite-particle mean-field limit directly, without simulating N particles, and to do so even when multiple stationary states—including unstable ones—coexist. If correct, the method gives a particle-free route to the long-time statistics of interacting diffusion systems and resolves the ambiguity that finite-particle simulations face near phase transitions.

What carries the argument

The machinery is the weak form of the stationary nonlinear Fokker-Planck equation, ⟨L*_p φ, p⟩=0 for all smooth test functions φ, together with a normalizing-flow transport map Gθ that pushes a base density to p. In practice the test functions are Gaussians centered at generated samples with width κ, and the loss (18)/(19) evaluates the adjoint operator on these functions using samples from the flow. II-WGS threads the current generative map into the interaction term; PI-WGS holds a frozen approximation of the law in the interaction while updating the map. This combination avoids density derivatives, positivity constraints, and any N-body simulation.

What would settle it

Compare II-WGS/PI-WGS output against a reliable long-time N-particle simulation in a regime where uniform propagation of chaos is known to hold (e.g., the linear Example 1). If the generated samples and the empirical one-particle distribution differ by more than Monte Carlo error (in a chosen metric such as 1-Wasserstein distance), the method is not computing the true stationary law. Alternatively, evaluate the weak residual ⟨L*_p φ_ζ, p̂θ⟩ on a dense grid of Gaussian centers ζ; if the residual is large for some ζ while the empirical loss (14) is small, the test-function family is insufficient

Watch

Extended reading notes

Core claim

For a McKean–Vlasov process, invariant measures solve the stationary nonlinear Fokker-Planck equation (SNFPE), a self-consistent PDE because the drift depends on the law through a convolution. The paper discovers that this equation can be turned into a training objective for a normalizing flow: plug transported samples into the adjoint equation, evaluate against Gaussian test functions, and minimize the squared violations. The mean-field interaction is either included implicitly in the loss (II-WGS) or frozen from the previous iterate (PI-WGS). The resulting generative maps produce i.i.d. samples that follow a true stationary measure of the mean-field process, including, with II-WGS, all sta

Load-bearing premise

The load-bearing assumption is that the chosen finite family of Gaussian test functions, with a manually fixed width κ, is rich enough that minimizing the empirical weak loss forces the generated distribution to satisfy the true stationary nonlinear Fokker-Planck equation; this sufficiency is not proved, and the paper only discusses the limits κ→∞ and κ→0.

Editorial extensions

If this is right

  • Stationary measures of McKean–Vlasov systems can be sampled as i.i.d. draws directly from the mean-field limit, with no need to simulate N interacting particles.
  • In phase-transition regimes, II-WGS can resolve multiple coexisting stationary distributions, including unstable ones that finite-particle simulations typically miss or average over.
  • PI-WGS provides a branch-selective alternative that converges to stable stationary solutions, matching the contraction behaviour of fixed-point iterations.
  • The method extends to non-gradient drift, parametric interaction kernels, and high-dimensional state spaces, with numerical evidence in 10 and 30 dimensions.
  • Conditioning the normalizing flow on a physical parameter (e.g., the colored-noise correlation time) yields a single model that approximates stationary distributions over a range of parameter values, including extrapolation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the weak-loss construction only needs the adjoint of the linearized stationary operator, the same recipe could be applied to other nonlinear stationary PDEs with a transport-map representation, such as Vlasov–Fokker–Planck systems or fractional-diffusion limits.
  • The ability of II-WGS to converge to unstable stationary branches suggests it could be used to trace complete bifurcation diagrams for McKean–Vlasov equations, a task where fixed-point and Monte Carlo methods are systematically biased.
  • The unresolved gap between the practical Gaussian test-function loss and the exact weak formulation implies that a quantitative guarantee would require either a completeness proof for the test family or an adaptive κ-schedule that provably controls a stationarity residual; this is an open problem the paper does not settle.
  • Parametric conditional flows of this type could be combined with continuation or annealing in temperature to detect hysteresis in mean-field phase transitions, a direct extension the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper extends the Weak Generative Sampler (WGS) of Cai et al. to stationary distributions of McKean-Vlasov processes. It proposes two training schemes (II-WGS and PI-WGS) for a normalizing flow: minimize an empirical weak-form residual of the stationary nonlinear Fokker-Planck operator using adaptive Gaussian test functions, with the mean-field convolution evaluated either from current samples (implicit) or from a previous frozen model (Picard). Numerical experiments cover a linear Gaussian test case, a double-well model with phase transition, a parameterized active-particle model, a 10D non-gradient system, and a 30D truncated-Coulombic system. The paper claims the approach is free from finite-particle truncation and can capture all stationary branches when multiple invariant measures exist.

Significance. The problem is important: finite-particle simulation can be unreliable in metastable mean-field systems, and a PDE-based generative sampler would be a useful tool. The weak-form derivation in Section III is standard and clean; the Gaussian example provides a closed-form check; the double-well comparison against a fixed-point iteration is a genuinely external validation that supports the existence of multiple stationary branches; and the parametric active-particle experiment is a nice demonstration of generalization. However, the central identifiability step is unproved, and the high-dimensional validation relies on the very finite-particle simulations the method claims to avoid. These gaps must be addressed before the central claim is supported. If the identifiability gap can be closed, the method would be a significant contribution to computational mean-field models.

major comments (4)
  1. [Sec. III.A, Eqs. (14)-(15); Sec. III.B, Eqs. (18)-(19)] The empirical loss is a finite projection of the stationary residual onto N_phi Gaussians centered at generated samples. The paper's own discussion of kappa->infinity (loss vanishes) and kappa->0 (strong residual at centers) does not cover the intermediate kappa values used in the experiments. No completeness or identifiability result shows that a small value of this loss implies L_p p is small as a distribution. Moreover, centers are drawn from the current approximate measure, so a wrong density that matches the residual near sampled centers but is arbitrary elsewhere can have small loss. This is load-bearing: the output is claimed to be i.i.d. samples from an invariant measure. Please supply an identifiability/sufficiency result, or a quantitative residual bound, or at least out-of-sample tests on independent test functions not adapted to the current samples.
  2. [Sec. IV.D and Sec. IV.E] The 'true' distributions in the high-dimensional examples are obtained by Euler-Maruyama simulation of N=1000 particles over T=1e5 with burn-in T0=100. This is exactly the finite-particle truncation that the introduction argues can fail for metastable systems (Sec. I). No convergence study in N and T is reported, and no time-uniform propagation of chaos is available for these models. Thus these experiments do not validate the central claim of computing mean-field stationary distributions without finite-particle truncation. The double-well example, validated by fixed-point iteration, is the only external PDE-based check; the high-dimensional examples need a comparable reference or careful N/T convergence analysis.
  3. [Appendix A, Eqs. (A7)-(A10); Sec. IV.B] The contraction condition beta * theta * Var_{p_bar_x}(x) < 1 is derived for the scalar fixed-point iteration x_{n+1} = F(x_n) in one dimension. This is used to conclude that PI-WGS 'resembles' a non-contractive fixed-point iteration and therefore cannot reach the unstable branch. However, no formal link is established between Algorithm 3 and the gradient-descent training dynamics of the normalizing flow in PI-WGS. The reduction of the 2D Example 2 to two independent 1D marginals is also assumed rather than proved. Please either provide a precise relationship or present the contraction argument as a heuristic explanation rather than a conclusion about PI-WGS.
  4. [Sec. III.A, Eq. (15)] The ball penalty L_b restricts all generated samples to B_r(x0). For the models considered (linear, quadratic, or harmonic confining potentials), the true stationary measures have unbounded Gaussian tails, so this penalty introduces a truncation bias. The paper does not report r for most experiments or quantify the effect of the truncation. Because L_b can be minimized by concentrating mass inside the ball, it also interacts with the identifiability issue of the weak loss. Please report r for all experiments and provide a sensitivity analysis, or replace L_b with a regularization that is consistent with the decay of the true stationary solution.
minor comments (5)
  1. [Sec. I, III.A, Algorithm 2] Typos: 'proorogation' (Sec. I), 'avarage' (Sec. III.A), 'generaive' (Algorithm 2 output), 'normalizatoin' (Sec. II.B).
  2. [Sec. III.A, Eqs. (7)-(13)] The weak form is stated for C_c^infinity test functions, while the implemented Gaussian test functions have unbounded support. This is likely justified by Gaussian decay, but the growth/decay conditions on f, K, and p needed for integration by parts should be stated.
  3. [Sec. IV] The values of the ball radius r and penalty parameters lambda, c are not reported for Examples 4 and 5. Without these, the experiments are not fully reproducible.
  4. [Sec. IV.D and IV.E] The high-dimensional examples report only visual histogram comparisons. Quantitative errors such as relative L2 distance or Wasserstein distance are reported only for Examples 1 and 3.
  5. [Sec. II.A, Eq. (4)] The sum index in the individual potential term is written as n while the sum elsewhere uses i; please standardize the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the weak-form derivation is self-contained, self-citation is not load-bearing, and external benchmarks validate the method.

full rationale

The paper's central derivation is the weak-form stationary Fokker-Planck equation (8)-(12), which is exact and not defined in terms of the method's outputs. The loss functions (18) and (19) are direct Monte Carlo approximations of a self-consistent weak residual; no quantity is fitted to a target and then renamed as a prediction. The Gaussian test-function family (13) is an acknowledged ansatz inherited from [CCHZ24], but the paper does not use that citation to prove correctness; it reports numerical validation against an analytical Gaussian solution (Example 1), a fixed-point iteration for the double-well ground truth (Example 2), and SDE simulations in Examples 4-5. Example 3 is a genuine out-of-sample test: training on epsilon in [0.3,0.6] and evaluating on [0.1,0.8]. The Appendix stability condition beta*theta*Var is derived analytically from the fixed-point map, not from the trained model. The self-citation of [CCHZ24] is a methodological precursor, not a load-bearing uniqueness theorem or a substitute for verification. The paper does leave an identifiability gap: it is not proved that the finite, adaptively centered Gaussian test functions are rich enough to force the residual to vanish at intermediate kappa. That is a limitation in the mathematical guarantee, but it is not circular because the loss is not constructed so that its minimizer is the answer by definition. No prediction reduces to an input by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on standard weak-form theory plus three ad hoc premises: that Gaussian-kernel test functions centered on generated samples are rich enough to enforce the stationary equation, that the ball penalty does not bias the result, and that the neural training of PI-WGS respects the scalar fixed-point stability condition. No convergence theorem for the optimization-loss connection is provided.

free parameters (3)
  • Gaussian kernel width kappa = 1.0 (Ex1); 0.3/0.5 (Ex2); 1.0 to 0.8 (Ex3); 1.2 to 0.8 (Ex4); 10 and 11 to decay (Ex5)
    Width of the test functions in the weak loss; controls the richness of the test family and the effective residual. Chosen per example without a principled criterion.
  • ball penalty radius r and strength lambda, c = not reported
    Lb confines generated samples to B_r(x0); if r is too small it truncates the true distribution. Values are not given in the paper.
  • center noise scale gamma = 0.8 to 0.08 (Ex5); others unreported
    Adds Gaussian noise to test function centers to aid exploration. Not reported for Examples 1 through 4.
assumptions (4)
  • standard math The stationary Fokker-Planck equation (2) is well-posed with the stated boundary and normalization conditions, and its weak form (7) is equivalent to the strong form for C_c^infinity test functions.
    Used implicitly in Section III A to formulate the loss and to justify the adjoint form (8)-(9).
  • ad hoc to paper A finite family of Gaussian kernels (13), with centers on generated samples and width kappa, suffices to determine the stationary distribution.
    This is the methodological assumption that makes the loss practical; no completeness or representativeness proof is given (Section III A, III B).
  • ad hoc to paper The ball constraint Lb does not bias the solution, i.e., all stationary distributions have support inside B_r(x0).
    Introduced in Eq. (15) to prevent divergence; the paper does not analyze the resulting bias on the learned density.
  • ad hoc to paper The training dynamics of PI-WGS follow the scalar fixed-point iteration (A7), so the local contraction condition (A10) determines which stationary solution is reached.
    The paper justifies PI-WGS's stable-branch behavior by analogy to Algorithm 3 (Appendix A), but no proof connects the neural network training to the fixed-point map.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generating Samples of Stationary Distributions of Weakly Interacting Diffusion Models Without Finite Particle Truncation: A Weak Generative Approach." pith.science (2026). https://pith.science/paper/FGWQAZDV

@misc{pith2026250912841,
  author       = {Pith},
  title        = {Pith review of: Generating Samples of Stationary Distributions of Weakly Interacting Diffusion Models Without Finite Particle Truncation: A Weak Generative Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FGWQAZDV}},
  note         = {Machine review of arXiv:2509.12841}
}
read the original abstract

Computing the stationary probability density and generating corresponding samples for the mean-field model of an infinite number of weakly interacting diffusion particles pose significant numerical challenges, particularly in the phase transition regime where the interchangeability of infinite-time and infinite-particle limits breaks down. Traditional approaches, such as direct simulation of finite-particle systems, often fail to accurately pinpoint multiple stationary distributions in the mean-field meta-stable setting. On the other hand, solving the high-dimensional McKean-Vlasov partial differential equation using neural networks typically yields only the density function, limiting its utility for estimating statistical quantities from generating samples. In this work, we propose a novel generative framework based on the weak PDE formulation of the mean-field model to address these challenges. Our approach simultaneously computes the stationary distributions of McKean-Vlasov processes and generates independent and identically distributed samples that satisfy these distributions. This integrated approach not only reveals the true stationary distributions without the random perturbation of finite particle truncation, but also offers deeper insight into the system's behavior in the mean-field limit. Extensive numerical experiments demonstrate the effectiveness of the proposed method, showcasing its ability to accurately approximate stationary distributions, capture intricate phase transitions, and handle high-dimensional complex systems.

Figures

Figures reproduced from arXiv: 2509.12841 by the authors.

Figure 1
Figure 1. FIG. 1. (Example 1) Convergence of II-WGS and PI-WGS for the linear system. Left: Training loss computed using (18) and (19) versus [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. (Example 1) Solving the linear drift system with quadratic potential by II-WGS and PI-WGS. Contour plots of the stationary distribu [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. (Example 2 with [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: FIG. 4. (Example 2) Comparison of the true stationary distributions for [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. (Example 3) Comparison of the true stationary distributions for [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. (Example 3) Relative [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. (Example 4) The upper and lower panels show histogram plots of the sample data points generated by the II-WGS (upper) and [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. (Example 5) The upper and lower panels show histogram plots of the sample data points generated by the II-WGS (upper) and [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. (Example 5) The left and right panels display the estimated means (red stars) with the true means (black points) in each dimension by [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10. The left panels display the stationary distributions computed using Algorithm 3 with various initializations ¯x [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 3 linked inside Pith

  1. [1]

    merlin.mbs aapmrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aapmrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translat...

    2010

  2. [2]

    merlin.mbs aipauth4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aipauth4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translat...

    2010

  3. [3]

    merlin.mbs aipnum4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aipnum4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  4. [4]

    merlin.mbs apsrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs apsrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  5. [5]

    merlin.mbs apsrmp4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs apsrmp4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  6. [6]

    Uniform convergence to equilibrium for granular media

    Fran c ois Bolley, Ivan Gentil, and Arnaud Guillin. Uniform convergence to equilibrium for granular media. Archive for Rational Mechanics and Analysis , 208:429--445, 2013

    2013

  7. [7]

    Fundamentals of plasma physics

    Jos \'e A Bittencourt. Fundamentals of plasma physics . Springer Science & Business Media, 2013

    2013

  8. [8]

    Galactic dynamics

    James Binney and Scott Tremaine. Galactic dynamics . Princeton university press, 2011

    2011

Show all 47 references
  1. [9]

    Probability flow solution of the F okker-- P lanck equation

    Nicholas M Boffi and Eric Vanden-Eijnden. Probability flow solution of the F okker-- P lanck equation. Machine Learning: Science and Technology , 4(3):035012, 2023

    2023

  2. [10]

    Weak Generative Sampler to Efficiently Sample Invariant Distribution of Stochastic Differential Equation

    Zhiqiang Cai, Yu Cao, Yuanfei Huang, and Xiang Zhou. Weak Generative Sampler to Efficiently Sample Invariant Distribution of Stochastic Differential Equation . arXiv preprint arXiv:2405.19256 , 2024

    2024 arXiv

  3. [11]

    Propagation of chaos: A review of models, methods and applications

    Louis-Pierre Chaintron and Antoine Diez. Propagation of chaos: A review of models, methods and applications. I . Models and methods . Kinetic and Related Models , 15(6):895--1015, 2022

    2022

  4. [12]

    Propagation of chaos: A review of models, methods and applications

    Louis-Pierre Chaintron and Antoine Diez. Propagation of chaos: A review of models, methods and applications. II . Applications . Kinetic and Related Models , 15(6):1017--1173, 2022

    2022

  5. [13]

    Long-time behaviour and phase transitions for the McKean--Vlasov equation on the torus

    Jose Antonio Carrillo, Rishabh S Gvalani, Grigorios A Pavliotis, and Andre Schlichting. Long-time behaviour and phase transitions for the McKean--Vlasov equation on the torus . Archive for Rational Mechanics and Analysis , 235(1):635--690, 2020

    2020

  6. [14]

    Kinetic equilibration rates for granular media and related equations: entropy dissipation and mass transportation estimates

    Jos \'e A Carrillo, Robert J McCann, and C \'e dric Villani. Kinetic equilibration rates for granular media and related equations: entropy dissipation and mass transportation estimates. Revista Matematica Iberoamericana , 19(3):971--1018, 2003

    2003

  7. [15]

    Donald A. Dawson. Critical dynamics and fluctuations for a mean-field model of cooperative behavior. Journal of Statistical Physics , 31(1):29--85, 1983

    1983

  8. [16]

    u rgen G\

    Donald A. Dawson and J\" u rgen G\" a rtner. Large deviations, free energy functional and quasi-potential for a mean field model of interacting diffusions , volume 78. Memoirs of the American Mathematical Society, 1989

    1989

  9. [17]

    Mean field limit for coulomb-type flows

    Mitia Duerinckx and Sylvia Serfaty. Mean field limit for coulomb-type flows. Duke Mathematical Journal , 169(15):2887--2935, 2020

    2020

  10. [18]

    Density estimation using Real NVP

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP . In International Conference on Learning Representations , 2017

    2017

  11. [19]

    T.D. Frank. Nonlinear Fokker-Planck Equations: Fundamentals and Applications . Springer Series in Synergetics. Springer Berlin Heidelberg, 2005

    2005

  12. [20]

    Uniform in time propagation of chaos for the 2D vortex model and other singular stochastic systems

    Arnaud Guillin, Pierre Le Bris, and Pierre Monmarch \'e . Uniform in time propagation of chaos for the 2D vortex model and other singular stochastic systems . Journal of the European Mathematical Society , pages 1--28, 2024

    2024

  13. [21]

    Uniform Poincar \'e and logarithmic Sobolev inequalities for mean field particle systems

    Arnaud Guillin, Wei Liu, Liming Wu, and Chaoen Zhang. Uniform Poincar \'e and logarithmic Sobolev inequalities for mean field particle systems . The Annals of Applied Probability , 32(3):1590--1614, 2022

    2022

  14. [22]

    Mean field limits for interacting diffusions with colored noise: phase transitions and spectral numerical methods

    Susana N Gomes, Grigorios A Pavliotis, and Urbain Vaes. Mean field limits for interacting diffusions with colored noise: phase transitions and spectral numerical methods. Multiscale Modeling & Simulation , 18(3):1343--1370, 2020

    2020

  15. [23]

    Gvalani and André Schlichting

    Rishabh S. Gvalani and André Schlichting. Barriers of the McKean–Vlasov energy via a mountain pass theorem in the space of probability measures. Journal of Functional Analysis , 279(11):108720, 2020

    2020

  16. [24]

    Score-based physics-informed neural networks for high-dimensional fokker--planck equations

    Zheyuan Hu, Zhongqiang Zhang, George E Karniadakis, and Kenji Kawaguchi. Score-based physics-informed neural networks for high-dimensional fokker--planck equations. SIAM Journal on Scientific Computing , 47(3):C680--C705, 2025

    2025

  17. [25]

    A review of the mean field limits for Vlasov equations

    Pierre-Emmanuel Jabin. A review of the mean field limits for Vlasov equations . Kinetic and Related models , 7(4):661--711, 2014

    2014

  18. [26]

    The variational formulation of the fokker--planck equation

    Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker--planck equation. SIAM Journal on Mathematical Analysis , 29(1):1--17, 1998

    1998

  19. [27]

    Random batch methods ( RBM ) for interacting particle systems

    Shi Jin, Lei Li, and Jian-Guo Liu. Random batch methods ( RBM ) for interacting particle systems. J. Comput. Phys. , 400(108877):108877, January 2020

    2020

  20. [28]

    Foundations of kinetic theory

    M Kac. Foundations of kinetic theory. In Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability , volume 3, pages 171--197. University of California Press Berkeley, Los Angeles, California, 1956

    1956

  21. [29]

    Counting the number of stationary solutions of partial differential equations via infinite dimensional sampling

    Martin Kolodziejczyk, Michela Ottobre, and Gideon Simpson. Counting the number of stationary solutions of partial differential equations via infinite dimensional sampling. Philosophical transactions. Series A, Mathematical, physical, and engineering sciences , 383(2298):202402...

    2025

  22. [30]

    Model for chemotaxis

    Evelyn F Keller and Lee A Segel. Model for chemotaxis. Journal of theoretical biology , 30(2):225--234, 1971

    1971

  23. [31]

    Rhythms and turbulence in populations of chemical oscillators

    Yoshiki Kuramoto. Rhythms and turbulence in populations of chemical oscillators. Physica A: Statistical Mechanics and its Applications , 106(1-2):128--143, 1981

    1981

  24. [32]

    Computing the invariant distribution of randomly perturbed dynamical systems using deep learning

    Bo Lin, Qianxiao Li, and Weiqing Ren. Computing the invariant distribution of randomly perturbed dynamical systems using deep learning. Journal of Scientific Computing , 91(3):77, 2022

    2022

  25. [33]

    Solving Stationary Nonlinear Fokker--Planck Equations via Sampling

    Lei Li, Yijia Tang, and Jingtong Zhang. Solving Stationary Nonlinear Fokker--Planck Equations via Sampling . SIAM Journal on Applied Mathematics , 85(1):249--277, 2025

    2025

  26. [34]

    Score-based transport modeling for mean-field Fokker-Planck equations

    Jianfeng Lu, Yue Wu, and Yang Xiang. Score-based transport modeling for mean-field Fokker-Planck equations . Journal of Computational Physics , 503:112859, 2024

    2024

  27. [35]

    A random-batch Monte Carlo method for many-body systems with singular kernels

    Lei Li, Zhenli Xu, and Yue Zhao. A random-batch Monte Carlo method for many-body systems with singular kernels . SIAM Journal on Scientific Computing , 42(3):A1486--A1509, 2020

    2020

  28. [36]

    Analysis of 1 + 1 dimensional stochastic models of liquid crystal polymer flows

    Tiejun Li, Pingwen Zhang, and Xiang Zhou. Analysis of 1 + 1 dimensional stochastic models of liquid crystal polymer flows. Comm. Math. Sci. , 2(2):295--316, 2004

    2004

  29. [37]

    H. P. McKean . A class of markov processes associated with nonlinear parabolic equations. Proceedings of the National Academy of Sciences , 56(6):1907--1911, 1966

    1907

  30. [38]

    H. P. McKean . Propagation of chaos for a class of nonlinear parabolic equations. Lecture Series in Differential Equations , 7:41--57, 1967

    1967

  31. [39]

    Long-time propagation of chaos and exit times for metastable mean-field particle systems

    Pierre Monmarch \'e . Long-time propagation of chaos and exit times for metastable mean-field particle systems. arXiv preprint arXiv:2503.00157 , 2025

    2025 arXiv

  32. [40]

    \"O ttinger

    H.C. \"O ttinger. Stochastic Processes in Polymeric Fluids: Tools and Examples for Developing Simulation Algorithms . Springer-Verlag, Berlin Heidelberg, 1996

    1996

  33. [41]

    Normalizing flows for probabilistic modeling and inference

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research , 22(57):1--64, 2021

    2021

  34. [42]

    From microscopic to macroscopic scale equations: mean field, hydrodynamic and graph limits

    Thierry Paul and Emmanuel Tr \'e lat . From microscopic to macroscopic scale equations: mean field, hydrodynamic and graph limits . arXiv e-prints , page arXiv:2209.08832, September 2022

    2022 arXiv

  35. [43]

    Variational inference with normalizing flows

    Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning , pages 1530--1538. PMLR, 2015

    2015

  36. [44]

    Entropy-dissipation Informed Neural Network for McKean-Vlasov Type PDEs

    Zebang Shen and Zhenfu Wang. Entropy-dissipation Informed Neural Network for McKean-Vlasov Type PDEs . Advances in Neural Information Processing Systems , 36, 2024

    2024

  37. [45]

    Topics in propagation of chaos

    Alain-Sol Sznitman. Topics in propagation of chaos. In Paul-Louis Hennequin, editor, Ecole d'Et \'e de Probabilit \'e s de Saint-Flour XIX --- 1989 , volume 1464 of Lecture Notes in Mathematics , pages 165--251. Springer Berlin Heidelberg, 1991

    1989

  38. [46]

    Phase transitions of McKean--Vlasov processes in double-wells landscape

    Julian Tugaut. Phase transitions of McKean--Vlasov processes in double-wells landscape . Stochastics An International Journal of Probability and Stochastic Processes , 86(2):257--284, 2014

    2014

  39. [47]

    Adaptive Deep Density Approximation for Fractional Fokker--Planck Equations

    Li Zeng, Xiaoliang Wan, and Tao Zhou. Adaptive Deep Density Approximation for Fractional Fokker--Planck Equations . Journal of Scientific Computing , 97(3):68, 2023

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.