A Kernel Formula for Kinetic Fokker-Planck Equations

Fuqun Han; Wuchen Li

REVIEW 3 major objections 5 minor 32 references

A Kernel Formula for Kinetic Fokker-Planck Equations

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The kinetic Fokker-Planck equation can be advanced by one explicit kernel integral, with local error of order $h^2$ and conditional first-order finite-time convergence.

desk verdict Explicit kernel formula for kinetic Fokker-Planck is novel and locally consistent; the finite-time convergence claim is honestly conditional on an unproved moment bound, so referee it but demand work on (26). read the letter →

arxiv 2608.05770 v1 pith:PEXFOG7B submitted 2026-08-06 math.NA cs.NA

Fuqun Han , Wuchen Li This is my paper · ORCID

classification math.NAcs.NA MSC 35Q8465M1265M7560J6082C31

keywords kineticFokker-PlanckequationkernelformulaweakconsistencyhypoellipticdiffusionphasespacescoreestimationdeterministicparticlesamplingregularizedWassersteinproximal

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's aim is to show that density evolution for the kinetic Fokker-Planck equation can be advanced by a single explicit integral operator, rather than by solving a transport problem or an implicit variational step. The operator takes the explicit Gaussian transition kernel of a reference drift-free process, weights it with a function $\psi$, and normalizes by the semigroup action $G_h\psi$, so the drift enters through the normalization. In weighted phase spaces, the authors prove local weak consistency with residual $O(h^2)$ and, conditional on regularity and moment bounds stated in the paper, first-order finite-time weak convergence. A formal optimal-control derivation links the operator to a Fisher-information-regularized proximal step, and numerical experiments use the resulting score estimator in deterministic particle sampling.

What carries the argument

The machinery is the normalized ratio $G_h(\psi f)/(G_h\psi)$, built from the explicit Gaussian semigroup $G_h$ of the reference dynamics and the auxiliary weight $\psi=e^{-\varphi/2}$. Because $\gamma\beta^{-1}\nabla_v\varphi$ equals the full drift $U=\nabla_x V+\gamma v$, the logarithmic derivative of $\psi$ supplies exactly the missing drift term, so the ratio expands as $f+hLf+R$ with $R=O(h^2)$ in a weighted norm. The same Hopf-Cole substitution $\eta=e^{\Phi/2}$ factorizes the variational optimality system into a backward reference semigroup for $\eta$ and a forward reference semigroup for $\rho/\eta$, which is why the terminal density is exactly the kernel formula.

What would settle it

Take $d=1$, $V(x)=x^2/2$, $\beta=1$, $\gamma=5$, $\lambda=0.34$, and the nine-component Gaussian-mixture initial density from the experiments; evolve the exact kernel operator with high-accuracy quadrature on a large box at $h=0.1,0.05,0.025$ and compare with the exact Gaussian evolution at $T=0.35$. If the weak errors do not show the first-order decrease stated in Theorem 2.7, or if the computed discrete weighted moments $m_{W_{+,2m}}(\rho_k^h)$ grow without bound as $h$ decreases while the exact moments stay finite, then the paper's convergence claim or its key moment assumption is falsified in that setting.

Watch

Extended reading notes

Core claim

The central claim is that the operator $K_h$ defined by $(K_h\rho)(x,v)=\psi(x,v)\int G_h(x,v|y,w)\frac{\rho(y,w)}{(G_h\psi)(y,w)}\,dy\,dw$, with $\psi=\exp(-\varphi_{\text{ki}}^{\lambda}/2)$ and $G_h$ the Gaussian transition kernel of the reference dynamics $dx_t=v_t\,dt$, $dv_t=\sqrt{2\gamma\beta^{-1}}\,dB_t$, is a one-step weak approximation of the kinetic Fokker-Planck equation. For bounded-Hessian potentials and exponential weights, Lemma 2.4 and Proposition 2.6 give a one-step consistency residual of order $h^2$; Theorem 2.7 then yields first-order weak error over finite time, provided the stated backward regularity, duality, and uniform weighted-moment bounds hold. The paper presents the formula as the first closed-form kernel update of this type for a kinetic equation, derives the same formula from a Fisher-information-regularized optimal control problem via a Hopf-Cole transformation, and extends the construction to affine Fokker-Planck equations with constant diffusion.

Load-bearing premise

The finite-time convergence theorem assumes that the discrete kernel iterates keep their weighted moments bounded up to the final time and that the backward solutions obey the stated regularity and duality bounds; the paper does not prove these follow from its baseline potential assumptions.

Editorial extensions

If this is right

The kernel update preserves mass and positivity by construction, so it can serve as a mesh-free density surrogate that does not require solving a high-dimensional optimal transport problem at each step.
Differentiating the kernel mixture gives an explicit score function estimator, so the same formula supplies $\nabla\log\rho$ for deterministic particle and probability-flow methods.
Under the conditional assumptions, the exact-normalization scheme converges weakly at first order, making it a candidate building block for structure-preserving parabolic and kinetic solvers.
The same construction covers general affine Fokker-Planck equations whenever the Kalman rank condition makes the reference semigroup Gaussian, unifying degenerate and nondegenerate cases.

Reading between the lines

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

Going beyond the paper, one testable extension would be a Lyapunov estimate for $K_h$ on weighted spaces that proves the uniform moment bound (26) from the baseline assumptions, which would make Theorem 2.7 unconditional for strongly convex potentials.
The paper's sampling algorithm uses a mollified kernel and Laplace-approximated normalizers that lie outside the proved theory; a natural numerical check is whether one-step consistency remains $O(h^2)$ for fixed mollification width $\tau_x>0$ as $h\to0$.
The kernel's velocity-dependent normalization couples all particles globally at each sampling step, so the practical cost of the method depends on fast approximation of the normalizers; the paper does not address this computational scaling question.

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.

Desk Editor's Note

read the letter

This paper is worth a serious look, but read the headline claim carefully. The genuinely new thing is the explicit kernel operator (21): a mass-preserving one-step approximation to kinetic Fokker-Planck density evolution, built from a Gaussian reference semigroup and a weight that encodes the drift. The construction is clean, and the local weak consistency proof (Prop. 2.6) is real. The weighted Taylor estimates in Lemmas 2.2-2.4 are detailed and checkable, and the mass preservation argument is exact. The authors also state Theorem 2.7 as conditional on backward regularity and uniform discrete moment bounds, and they say plainly that (26) is not a consequence of Assumption 2.1. That honesty helps the paper rather than hurting it.

The soft spot is exactly the one the stress-test flags. The finite-time convergence theorem collapses if the iterates develop unbounded W_{+,2m}-moments, and (26) is only assumed, not proved. This is not a minor technical gap. K_h is an explicit integral operator with a Gaussian kernel and a weight that can be nearly flat in x for shallow potentials; whether repeated application preserves exponential moments uniformly in h and k is exactly what (26) asserts, and no argument is given. The local consistency result stands on its own, and the conditional theorem is a meaningful statement, but the paper does not prove unconditional first-order finite-time convergence. This is a limitation, not a deception, since it is openly acknowledged.

Minor issues: the sampling scheme in Section 4.3 uses a mollified kernel and Laplace-approximated normalizers, putting it outside the proved theory; the experiments are qualitative demonstrations, not evidence of convergence. The variational interpretation is formal, conditional on a classical solution and multiplier. None of this undercuts the core kernel formula.

Who should read this: anyone working on mesh-free density approximation for kinetic equations, deterministic particle sampling, or score estimation. The kernel formula itself is likely to be useful and citable. I would send it to a serious referee: the construction is novel, the local estimates are robust, and the gaps are clearly identified. The referee should push on the discrete moment bound (26); if that can be proved under reasonable assumptions, the paper becomes much stronger.

Referee Report

3 major / 5 minor

Summary. This manuscript proposes an explicit one-step kernel operator K_h, Eq. (21), for the kinetic Fokker-Planck equation (2). The construction takes the Gaussian transition density G_h of the velocity-diffusion reference dynamics (5)-(6) and the auxiliary weight ψ = exp(-φ^λ_ki/2), with φ^λ_ki chosen in (7) so that γβ^{-1}∇_v φ^λ_ki equals the drift U = ∇_x V + γv; K_h is then defined by normalizing the reference semigroup acting on ρ/(G_hψ). The paper proves that K_h preserves mass and positivity and, under Assumption 2.1, establishes weighted estimates (Lemmas 2.2-2.3) and the normalized expansion G_h(ψf)/G_hψ = f + hLf + O(h^2) in a weighted norm (Lemma 2.4). Proposition 2.6 gives the unconditional O(h^2) local weak consistency of K_h with L^*. Theorem 2.7 states an O(h) finite-time weak convergence result for the iterates ρ^h_{k+1} = K_h ρ^h_k, conditional on three unproved hypotheses (24)-(26): backward regularity of the exact semigroup, forward-backward duality, and uniform weighted-moment bounds for exact and discrete solutions. The paper explicitly and repeatedly declares these hypotheses not to be consequences of Assumption 2.1. A formal variational derivation of K_h via a Fisher-information-regularized kinetic optimal control problem (Propositions 2.8-2.9) and an abstract affine generalization (Section 3) are also given.

Significance. The construction is original and potentially useful: K_h is a genuinely explicit, mesh-free one-step update for a hypoelliptic Fokker-Planck equation, it preserves mass and positivity, and it yields an explicit score estimator. The strongest parts of the paper are the local weighted estimates: Lemmas 2.2-2.4 are proved in detail with explicit Gaussian computations, and Proposition 2.6 is an honest unconditional statement of local weak consistency. The derivation of ψ from the drift is parameter-free in the sense that no fitted constant enters (only the scalar λ remains free, and the theory tolerates it). The paper is also commendably candid in flagging its own limitations: the conditionality of Theorem 2.7, the formality of the variational derivation, the absence of Laplace-error estimates, and the heuristic status of the mollified sampling scheme are all stated in the text. The paper does not ship code, and the proofs are not machine-checked.

major comments (3)

[Section 2.1, Theorem 2.7 and Eq. (26)] The first-order global bound (27) rests entirely on the uniform discrete weighted-moment hypothesis (26). In the telescoping proof, both A_k and B_k are controlled by m_{W_{+,2m}}(ρ^h_k), and without (26) the O(h^2) local errors cannot be summed to O(h) uniformly in n. The paper states explicitly, before and after Theorem 2.7, that (24)-(26) are not consequences of Assumption 2.1; among the three, (26) is the least secure because it concerns the numerical iterates K_h^k ρ_0, not the exact solution, and the stress-test concern lands here. As written, the theorem applies only to initial data for which the iterates happen to satisfy the bound, with no sufficient condition, no example class, and no numerical check of (26) anywhere in the paper (the Section 4.2 experiment never measures the W_{+,2m}-moments of its iterates). Since first-order finite-time weak convergence appears as a headline result in the abstract, this gap is load-bearing. I recommend either (i) proving (26) for a nontrivial class, e.g. by establishing a Lyapunov-type estimate (K_h^* W_{+,2m})(z) ≤ (1+Ch)W_{+,2m}(z) for suitable V (say strongly convex with bounded Hessian) and λ, which would yield (1+Ch)^{T/h} ≤ e^{CT}; or (ii) presenting Theorem 2.7 explicitly as a formal/conjectural statement and re-focusing the abstract and Section 5 on the unconditional results: the explicit formula, mass preservation, local weak consistency (Proposition 2.6), and the variational identity (33).
[Section 4.3, Algorithm 2 and Eqs. (67)-(68)] The sampling scheme that the abstract cites as an application to deterministic particle schemes for kinetic sampling is not a numerical realization of the analyzed operator. With the reported parameters (τ_x = 0.16, γ = 1.5, β = 1, h = 0.02), the kernel width a_x = γh^3/(6β) + τ_x^2 is approximately 0.0256, of which γh^3/(6β) ≈ 2×10^{-6}; the mollification term therefore dominates, the positional coupling of G_mol_h has a fixed scale independent of h, and the score estimate in (66) is rescaled by replacing the coefficient 3β/(γh^2) with h/(2a_x) ≈ 0.39, several orders of magnitude smaller than the theoretical coefficient. The paper discloses that this scheme is not covered by Proposition 2.6 or Theorem 2.7, and that disclosure is appropriate; however, the abstract and the conclusion still present the sampling experiments as an application of the kernel approximation. Given that no consistency analysis for the mollified kernel is provided (consistency as h→0 would require a joint scaling of τ_x with h that is neither stated nor analyzed) and the Laplace-approximated normalizers (58) carry no error bound, the abstract and Section 5 should describe this part explicitly as a heuristic, numerically regularized variant whose connection to the proved theory is only formal.
[Section 4.2, Figure 1] The reported error study cannot be cleanly attributed to the kernel scheme. The quadrature grids 64×32, 158×40, 480×52, 1600×72 are changed together with h (0.175, 0.0875, 0.05, 0.025), the computation is a finite-box surrogate on [-7.5,7.5]×[-7,7], and the hybrid midpoint-kernel discretization with shifted interpolation is not specified in enough detail to assess its error. Quadrature error, box truncation, and interpolation error are thus mixed with the time-discretization error that the theory controls. Moreover, the relative denominator error ε_h from Section 4.1 is never estimated, so the one-step perturbation bound of O(ε_h) cannot be checked against the experiment. Please rerun the convergence study against a single fixed fine reference grid and report separately (a) the dependence on h at fixed quadrature resolution and (b) the size of the quadrature or denominator error; this separation is needed for the numerical section to function as evidence for the accuracy claim.

minor comments (5)

[Eq. (11) and Lemmas 2.3-2.4] The weight notation is easy to confuse: Lemma 2.3 defines W_m and W^+_m, while Lemma 2.4 defines W_{+,2m} := (1+|x|+|v|)^{2m} W^2 W^+, and Theorem 2.7 then uses m_{W_{+,2m}} without restating the definition. Please use one consistent family with a single index and restate it before Proposition 2.6. In Assumption 2.1, the sentence 'For 0<θ<θ̄, define the base and stronger phase space weights' should state explicitly that θ and θ̄ are fixed positive constants with θ<θ̄ and that W := exp(θ(1+|x|^2+|v|^2)), W^+ := exp(θ̄(1+|x|^2+|v|^2)) are the two weights used throughout.
[Section 4.2] The value λ = 0.34 is used without explanation; since λ is a free parameter of the construction and the theory permits a range of choices, a short sensitivity study or at least a statement of how λ was selected would strengthen the reported accuracy claim.
[Sections 4.1-4.3] No code or data repository is provided, and the shifted interpolation rule for under-resolved position Gaussian kernels is not spelled out; given that the implementation details matter for the claimed accuracy, please provide a precise description of the hybrid quadrature or make the code available.
[Eqs. (59)-(60) and Algorithm 2] The mass-normalized numerical operator $\hat{K}_h$ in (60) is nonlinear in ρ, and in Algorithm 2 the mass normalization is not applied because it cancels from the score; the mass error of the raw update is therefore uncontrolled in the experiments, and this should be stated at the point where Algorithm 2 is introduced.
[Figure 1] The rightmost panel's vertical axis label and the colorbar of the pointwise-error panel should be clarified: it is currently ambiguous whether the reported 'relative L2(ρ_ref) score error' uses reference-density weighting over the full box, and the pointwise-error colorbar seems to be labeled in units of 10^{-3}; please state the exact error definitions and scales in the caption.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: the kernel formula derives from an explicit semigroup expansion; finite-time convergence is honestly conditional on an unproved moment bound, a limitation rather than circularity.

full rationale

The one-step kernel operator (21) is constructed from the explicit reference Gaussian semigroup G_h and the weight psi = exp(-phi/2) chosen to satisfy gamma beta^{-1} nabla_v phi = U. Lemma 2.4 proves the normalized ratio G_h(psi f)/G_h psi equals f + h L f + R with weighted O(h^2) remainder via direct Taylor estimates (Lemmas 2.2-2.3 and Appendix A), so the local consistency statement is self-contained and does not rely on the cited kernel-formula papers [14], [22], or [29]. The finite-time Theorem 2.7 is explicitly conditional: it assumes the uniform discrete weighted-moment bound (26) and backward regularity bounds (24)-(25), and the paper states that (26) is assumed, not proved. This is an openly labeled limitation rather than a circular reduction: the telescoping proof uses those assumptions to convert local O(h^2) errors into O(h), which is the standard convergence argument. The variational interpretation in Section 2.2 is similarly formal and conditional, with no existence or uniqueness assertions, so the Hopf-Cole representation does not smuggle the formula in as evidence. Sampling experiments use a mollified kernel and Laplace-approximated normalizers outside the proved theory and are clearly labeled as numerical regularization. Therefore the central derivation is independent of fitted constants and self-citations; the main weakness is the unproved moment stability hypothesis, which is a genuine correctness or assumption gap but not circularity.

Assumptions & free parameters 2 free parameters · 6 assumptions · 2 invented entities

The central theoretical results rest on Assumption 2.1 plus, for the finite-time convergence theorem, unproved weighted regularity and uniform moment bounds. The variational interpretation requires additional integrability and multiplier existence. The numerical sampling scheme adds mollification and Laplace approximations outside the theory.

free parameters (2)

lambda (kernel weight) = 0.34 (density benchmark), 1.0 (sampling)
Appears in phi^lambda_ki and in every kernel update; no principled choice or sensitivity analysis is given, and the accuracy of the scheme can depend on it.
tau_x (mollification width) = 0.16
Introduced in G_mol^h (Eq. 67) for the deterministic sampling scheme; the sampling scheme using it is explicitly outside the convergence theory.

assumptions (6)

domain assumption Assumption 2.1: V in C^3, bounded below, bounded second derivative, polynomial third derivative
Used to prove Lemmas 2.2 and 2.3 and hence local consistency; excludes superquadratic potentials used in the sampling experiments.
domain assumption Backward regularity and duality assumptions (24)-(25) in Theorem 2.7
Assumed for the finite-time convergence proof; not established from Assumption 2.1.
domain assumption Uniform weighted-moment bound (26) for exact and discrete solutions
Assumed rather than proved; the paper explicitly flags it as not a consequence of Assumption 2.1.
domain assumption Integrability Z_lambda < infinity in Section 2.2
Needed for the terminal relative-entropy penalty in the variational interpretation; can fail for superquadratic potentials.
domain assumption Existence and regularity of a Lagrange multiplier and classical admissibility in Proposition 2.8
The variational interpretation is formal; no minimizer or multiplier existence is proved.
domain assumption Kalman rank condition (42) in Section 3
Ensures the reference Gaussian covariance Q_h is positive definite for the generalized affine class.

invented entities (2)

Auxiliary weight psi = exp(-phi^lambda_ki / 2)
purpose: Encodes the drift perturbation U through gamma beta^{-1} nabla_v phi = U in the normalized ratio formula.
Mathematical construction invented for the paper; it has no external evidence, but it is not a physical entity and is explicitly derived from V.
Regularized kinetic action cost C_{ki,h}
purpose: Provides a formal variational interpretation of the one-step kernel update.
The paper asserts no metric property; it is used only as a formal cost functional.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Kernel Formula for Kinetic Fokker-Planck Equations." pith.science (2026). https://pith.science/paper/PEXFOG7B

@misc{pith2026260805770,
  author       = {Pith},
  title        = {Pith review of: A Kernel Formula for Kinetic Fokker-Planck Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PEXFOG7B}},
  note         = {Machine review of arXiv:2608.05770}
}

read the original abstract

We derive an explicit one-step kernel formula for the kinetic Fokker-Planck equation. The construction begins with a reference dynamics admitting an explicit transition kernel and uses a carefully chosen auxiliary weight to encode the drift perturbation. The resulting kernel formula provides an explicit approximation of the density evolution. In a weighted phase-space setting, we establish local weak consistency and first-order finite-time weak convergence under suitable regularity and moment assumptions. We also give a formal variational interpretation of the formula in terms of a regularized Wasserstein-type proximal operator and present analogous constructions for a broader class of parabolic equations. Numerical experiments illustrate the accuracy of the kernel approximation and its application to deterministic particle schemes for kinetic sampling.

Figures

Figures reproduced from arXiv: 2608.05770 by the authors.

**Figure 2.** Rosenbrock-type (left pair) and six-peak (right p [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 28 canonical work pages

[1]

Adams, M

D. Adams, M. H. Duong, and G. dos Reis. Entropic regularis ation of non-gradient systems. SIAM J. Math. Anal. , 54(4):4495–4535, 2022

work page 2022
[2]

Albergo, N

M. Albergo, N. M. Boﬃ, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for ﬂows and diﬀusions. J. Mach. Learn. Res. , 26(209):1–80, 2025

work page 2025
[3]

Ambrosio, N

L. Ambrosio, N. Gigli, and G. Savar´ e. Gradient Flows: In Metric Spaces and in the Space of Probability Measures . Springer, 2005

work page 2005
[4]

Benamou and Y

J.-D. Benamou and Y. Brenier. A computational ﬂuid mecha nics solution to the Monge– Kantorovich mass transfer problem. Numer. Math. , 84(3):375–393, 2000

work page 2000
[5]

Betancourt

M. Betancourt. A conceptual introduction to Hamiltonia n Monte Carlo. arXiv preprint arXiv:1701.02434, 2017

arXiv 2017
[6]

N. M. Boﬃ and E. Vanden-Eijnden. Probability ﬂow solutio n of the Fokker–Planck equa- tion. Mach. Learn.: Sci. Technol. , 4(3):035012, 2023

work page 2023
[7]

Brigati, J

G. Brigati, J. Maas, and F. Quattrocchi. Kinetic optimal transport (OTIKIN)– part 1: Second-order discrepancies between probability me asures. arXiv preprint arXiv:2502.15665, 2025

arXiv 2025
[8]

Calvello, S

E. Calvello, S. Reich, and A. M. Stuart. Ensemble Kalman m ethods: A mean-ﬁeld per- spective. Acta Numer. , 34:123–291, 2025

work page 2025

Show all 32 references

[9]

J. A. Carrillo, L. Wang, W. Xu, and M. Yan. Variational asy mptotic preserving scheme for the Vlasov–Poisson–Fokker–Planck system. Multiscale Model. Simul. , 19(1):478–505, 2021

2021
[10]

Cheng, N

X. Cheng, N. S. Chatterji, P. L. Bartlett, and M. I. Jorda n. Underdamped Langevin MCMC: A non-asymptotic analysis. In Proceedings of the Conference on Learning Theory , pages 300–323. PMLR, 2018. 24

2018
[11]

Di Marino, E

S. Di Marino, E. Naldi, and S. Villa. Inexact JKO and prox imal-gradient algorithms in the Wasserstein space. arXiv preprint arXiv:2505.23517 , 2025

2025 arXiv
[12]

M. H. Duong, M. A. Peletier, and J. Zimmer. GENERIC forma lism of a Vlasov–Fokker– Planck equation and connection to large-deviation princip les. Nonlinearity, 26(11):2951– 2971, 2013

2013
[13]

M. H. Duong, M. A. Peletier, and J. Zimmer. Conservative -dissipative approximation schemes for a generalized Kramers equation. Math. Methods Appl. Sci. , 37(16):2517–2540, 2014

2014
[14]

F. Han, S. Osher, and W. Li. Convergence of noise-free sa mpling algorithms with regular- ized Wasserstein proximals. J. Mach. Learn. Res. , 27(85):1–66, 2026

2026
[15]

J. Ho, A. Jain, and P. Abbeel. Denoising diﬀusion probabi listic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.-F. Balcan, and H.-T. Lin, editors , Adv. Neural Inf. Process. Syst. 33 , 2020

2020
[16]

C. Huang. A variational principle for the Kramers equat ion with unbounded external forces. J. Math. Anal. Appl. , 250(1):333–367, 2000

2000
[17]

Jordan, D

R. Jordan, D. Kinderlehrer, and F. Otto. The variationa l formulation of the Fokker–Planck equation. SIAM J. Math. Anal. , 29(1):1–17, 1998

1998
[18]

Lasry and P.-L

J.-M. Lasry and P.-L. Lions. Mean ﬁeld games. Jpn. J. Math. , 2(1):229–260, 2007

2007
[19]

W. Lee, L. Wang, and W. Li. Deep JKO: Time-implicit parti cle methods for general nonlinear gradient ﬂows. J. Comput. Phys. , 514:113187, 2024

2024
[20]

Leli` evre and G

T. Leli` evre and G. Stoltz. Partial diﬀerential equations and stochastic methods in molecular dynamics. Acta Numer. , 25:681–880, 2016

2016
[21]

L´ eonard

C. L´ eonard. A survey of the Schr¨ odinger problem and some of its connections with optimal transport. Discrete Contin. Dyn. Syst. Ser. A , 34(4):1533–1574, 2014

2014
[22]

W. Li, S. Liu, and S. Osher. A kernel formula for regulari zed Wasserstein proximal opera- tors. Res. Math. Sci. , 10(43), 2023

2023
[23]

Lipman, R

Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le . Flow matching for generative modeling. In Int. Conf. Learn. Represent. , 2023

2023
[24]

H. C. ¨Ottinger. Beyond Equilibrium Thermodynamics . John Wiley & Sons, 2005

2005
[25]

Pinnau, C

R. Pinnau, C. Totzeck, O. Tse, and S. Martin. A consensus -based model for global opti- mization and its mean-ﬁeld limit. Math. Models Methods Appl. Sci. , 27(1):183–204, 2017

2017
[26]

H. Risken. The Fokker–Planck Equation: Methods of Solution and Applica tions. Springer, 2nd edition, 1996

1996
[27]

Ruthotto, S

L. Ruthotto, S. J. Osher, W. Li, L. Nurbekyan, and S. W. Fu ng. A machine learning framework for solving high-dimensional mean ﬁeld game and m ean ﬁeld control problems. Proc. Natl. Acad. Sci. USA , 117(17):9183–9193, 2020

2020
[28]

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Er mon, and B. Poole. Score- based generative modeling through stochastic diﬀerential e quations. In Int. Conf. Learn. Represent., 2021. 25

2021
[29]

H. Y. Tan, S. Osher, and W. Li. Accelerated regularized W asserstein proximal sampling algorithms. arXiv preprint arXiv:2601.09848 , 2026

2026
[30]

Tibshirani, S

R. Tibshirani, S. Fung, H. Heaton, and S. Osher. Laplace meets Moreau: Smooth approxi- mation to inﬁmal convolutions using Laplace’s method. J. Mach. Learn. Res. , 26(72):1–36, 2025

2025
[31]

C. Villani. Hypocoercivity. American Mathematical Society, 2009

2009
[32]

C. Xu, X. Cheng, and Y. Xie. Normalizing ﬂow neural netwo rks by JKO scheme. Adv. Neural Inf. Process. Syst. , 36:47379–47405, 2023. 26

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Adams, M

D. Adams, M. H. Duong, and G. dos Reis. Entropic regularis ation of non-gradient systems. SIAM J. Math. Anal. , 54(4):4495–4535, 2022

work page 2022

[2] [2]

Albergo, N

M. Albergo, N. M. Boﬃ, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for ﬂows and diﬀusions. J. Mach. Learn. Res. , 26(209):1–80, 2025

work page 2025

[3] [3]

Ambrosio, N

L. Ambrosio, N. Gigli, and G. Savar´ e. Gradient Flows: In Metric Spaces and in the Space of Probability Measures . Springer, 2005

work page 2005

[4] [4]

Benamou and Y

J.-D. Benamou and Y. Brenier. A computational ﬂuid mecha nics solution to the Monge– Kantorovich mass transfer problem. Numer. Math. , 84(3):375–393, 2000

work page 2000

[5] [5]

Betancourt

M. Betancourt. A conceptual introduction to Hamiltonia n Monte Carlo. arXiv preprint arXiv:1701.02434, 2017

arXiv 2017

[6] [6]

N. M. Boﬃ and E. Vanden-Eijnden. Probability ﬂow solutio n of the Fokker–Planck equa- tion. Mach. Learn.: Sci. Technol. , 4(3):035012, 2023

work page 2023

[7] [7]

Brigati, J

G. Brigati, J. Maas, and F. Quattrocchi. Kinetic optimal transport (OTIKIN)– part 1: Second-order discrepancies between probability me asures. arXiv preprint arXiv:2502.15665, 2025

arXiv 2025

[8] [8]

Calvello, S

E. Calvello, S. Reich, and A. M. Stuart. Ensemble Kalman m ethods: A mean-ﬁeld per- spective. Acta Numer. , 34:123–291, 2025

work page 2025