Pith. sign in

REVIEW 2 major objections 4 minor 267 references

Score-matching gap for diffusion models tightens with decaying weights that put most weight on low-noise times.

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 →

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

T0 review · grok-4.5

2026-07-11 19:08 UTC pith:NVDKZ63D

load-bearing objection Clean continuous-time tightening of the score-matching gap with explicit decaying weights; solid math, useful qualitative message, regularity assumptions only lightly checked. the 2 major comments →

arxiv 2607.04442 v1 pith:NVDKZ63D submitted 2026-07-05 stat.ML cs.LG

Tightening the Score Matching Gap for Diffusion Models

Benjamin Dupuis , Tyler Farghly , Maxime Haddouche , Alain Durmus , Umut Simsekli This is my paper
classification stat.ML cs.LG MSC 60H1068T0562M45
keywords diffusion modelsscore matchingscore matching gapELBOlog-Sobolev inequalitiesentropy flowreflection couplingWasserstein distance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Diffusion models train by matching the score of a noising process, then reverse that process to generate samples. The usual ELBO bounds the KL between data and model by an integral of score error that is constant in time; that integral can be a loose proxy for sample quality. This paper shows that under regularity of the learned score (one-sided Lipschitz or dissipativity), the same KL can be bounded by a reweighted integral whose weight is strictly decreasing in forward time, plus a smaller initialization term. Parallel decaying bounds hold for reverse KL and Wasserstein distance. The message is that score accuracy at low noise matters more for closing the gap than accuracy at high noise, and that the contraction of the reverse diffusion is what produces the decay.

Core claim

Under one-sided Lipschitz (or dissipativity-at-distance) assumptions on the score network, the forward KL admits the tightened bound KL(μ ∥ ←μ_θ_T) ≲ ∫ λ_T(t) E[ε_θ_t(→X_t)^{2}] dt + K_T, where the explicit weight λ_T(t) < 1 is decreasing in t; analogous decaying weights hold for reverse KL and W_1/W_f.

What carries the argument

Entropy-flow identity for the pair of reverse Fokker–Planck equations, combined with time-dependent log-Sobolev constants of the reverse process (and reflection couplings for Wasserstein), which convert the usual constant-weight score integral into an exponentially or polynomially decaying weight.

Load-bearing premise

The learned score must obey a one-sided Lipschitz (or dissipativity-at-distance) condition that is uniform, or at worst mildly singular, for every noise level.

What would settle it

Train a score network that deliberately violates the one-sided Lipschitz condition at small noise and check whether the reweighted integral still upper-bounds the empirical KL (or W_1) of generated samples better than the classical ELBO; if the classical bound becomes tighter, the claimed decay fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Evaluation and early-stopping can safely reweight the score-matching loss toward low-noise times without losing the theoretical guarantee.
  • Existing continuous-time KL and Wasserstein convergence rates for diffusion models improve by replacing the factor T with an explicit decaying integral.
  • Training schedules that already down-weight high noise receive a theoretical justification: those times contribute less to the final gap.
  • Reverse-KL and Wasserstein versions of the same gap become available for analysis even when the data measure is singular.

Where Pith is reading between the lines

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

  • If modern U-Nets only satisfy the regularity assumptions at moderate noise, the practical gain of the decaying weight will be largest in the middle of the schedule rather than at the extreme low-noise end.
  • The same entropy-flow-plus-LSI argument should extend, with only notational changes, to variance-exploding and critically damped Langevin dynamics once their reverse LSI constants are controlled.
  • A discrete-time version of the decaying weights would immediately tighten many existing finite-step convergence proofs that currently carry a linear factor in the number of steps.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper analyzes the score-matching gap for continuous-time diffusion models (Ornstein–Uhlenbeck forward process). It first shows (Prop. 3.1) that the classical ELBO/Girsanov bound is essentially tight in the worst case without regularity on the score estimator. Under one-sided Lipschitz or dissipativity-at-distance assumptions on s_θ, it then derives tightened bounds on forward KL (Thms. 3.5–3.10), reverse KL (Thm. 4.3) and Wasserstein distances W_f / W_1 (Thm. 4.8, Cor. 4.9) that replace the constant weight 1/4 by explicit, time-decaying factors λ_T(t) or C_T(t) (Table 1). The technical engine is an entropy-flow identity (Lemmas 3.4, 4.2) combined with LSI estimates for the reverse process and reflection couplings. A short empirical section on a toy circle and a CIFAR-10 U-Net illustrates that low-noise score error tracks sample quality more closely and that the regularity assumptions hold at least qualitatively.

Significance. If the regularity hypotheses hold for practical score networks, the work supplies the first explicit, non-uniform weights that correctly prioritize low-noise regimes when converting score-matching loss into KL or Wasserstein guarantees. This both tightens existing continuous-time convergence analyses and offers a principled justification for the empirical practice of down-weighting small noise scales. The derivations are fully written out, the negative result via Schrödinger bridges is carefully stated, and the reflection-coupling argument for Wasserstein is new in the DM literature. The main limitation is that the load-bearing Lipschitz/dissipativity assumptions are only partially verified; if they fail for modern architectures the quantitative improvement disappears, but the qualitative message remains useful.

major comments (2)
  1. Theorems 3.7, 3.10 and 4.8 rest on a uniform (or 1+c/t) one-sided Lipschitz constant M (or the stronger (L,ρ,R)-dissipativity of Assumption 3.9) holding for every t∈[0,T]. Figure 4 only plots empirical κ(x,y) for a single CIFAR U-Net and shows that the local Lipschitz constant grows as t→0. Without a quantitative bound on M (or a proof that modern residual/attention blocks satisfy dissipativity at distance), the claimed exponential decay of λ_T(t) and C_T(t) remains conditional. A short discussion of how large M can be, or a counter-example network for which M>1, would make the scope of the theorems clearer.
  2. Section 5 experiments are purely illustrative: the toy circle uses a histogram KL estimate on 8 training points, and the CIFAR perturbation study uses a crude hand-tuned regularity constant. Neither experiment falsifies or quantitatively validates the decay rates of Table 1. At minimum the authors should report the numerical values of the estimated M (or c) used to produce the “Our weighting” curve in Figure 5, and state whether the same constant works across architectures.
minor comments (4)
  1. Table 1 lists decay rates but does not display the absolute prefactors that appear when α=1/2 is chosen; a short remark that the classical 1/4 is replaced by 1/2 would avoid confusion.
  2. Notation for the score error switches between ε_θ_t(x)^{2} and ∥s−s_θ∥^{2}; a single consistent definition early in §2 would help.
  3. The appendix proofs of the reflection-coupling argument (Prop. C.2) are long; a one-paragraph sketch in the main text of how the synchronous/reflection switch is handled would improve readability.
  4. References to concurrent Wasserstein analyses (Beyler & Bach 2025, Bruno & Sabanis 2025) are present but could be compared more explicitly on the dependence on T.

Circularity Check

0 steps flagged

No significant circularity: tightened score-matching bounds are derived from entropy-flow identities plus standard LSI/reflection-coupling estimates under explicit regularity assumptions.

full rationale

The central claims (Theorems 3.5–3.10, 4.3, 4.8 and Corollaries) are conditional mathematical upper bounds. They start from the Fokker–Planck equations of the true and approximate reverse processes, obtain the entropy-flow differential inequalities (Lemmas 3.4 and 4.2), replace Fisher information by KL via LSI (or control pathwise distance via Eberle-style reflection couplings), and close with Grönwall. The decaying weights λ_T(t) and C_T(t) are the explicit integrating factors that arise once a one-sided Lipschitz or dissipativity-at-distance hypothesis is imposed on s_θ (or an LSI hypothesis on μ for reverse KL). Those hypotheses are stated as assumptions, not fitted from the target quantity; the theorems do not claim to predict sample quality without them. Self-citations (Dupuis et al. 2025; Farghly et al. 2026) appear only in the optional Section D generalization extension and are not load-bearing for the main bounds. Empirical plots that insert a crude numerical M or c are illustrations of the already-proved functional form, not inputs that force the theorems. No step reduces the claimed bound to its own definition or to a fitted parameter renamed as a prediction.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The central claims rest on standard functional inequalities (LSI, Talagrand) and on regularity of the learned score that is not guaranteed for arbitrary neural networks. No free parameters are fitted into the theorems; the only free constants are those appearing in the dissipativity / Lipschitz assumptions themselves.

axioms (4)
  • domain assumption The reverse process densities satisfy the Fokker–Planck equation in C^{1,2} and the Fisher informations appearing in Assumptions 3.3 / 4.1 are finite and integrable.
    Standard mild regularity for diffusion models; invoked for all entropy-flow calculations (Lemma 3.4 onward).
  • ad hoc to paper Score network s_θ is one-sided Lipschitz with constant M (or M_t = 1 + c/t) or satisfies (L,ρ,R)-dissipativity at distance (Assumption 3.9).
    Needed to bound the LSI constant of ←μ_θ_t (Lemma B.4, Thm 3.7, 3.10) and the reflection-coupling rate (Thm 4.8); only partially verified empirically.
  • domain assumption Data distribution μ satisfies a logarithmic Sobolev inequality with constant ρ_0 (for reverse-KL results).
    Used in Theorem 4.3 and Corollary 4.6; can be relaxed to compact support or sub-Gaussian tails by cited convolution results.
  • standard math Standard Gaussian γ_d satisfies 1-LSI and the OU process contracts entropy at rate e^{-2t}.
    Classical Bakry–Émery / Gross theory; used throughout.

pith-pipeline@v1.1.0-grok45 · 47041 in / 2622 out tokens · 31274 ms · 2026-07-11T19:08:07.203146+00:00 · methodology

0 comments
read the original abstract

Diffusion models (DMs) are a state-of-the-art generative method to approximately sample from an unknown distribution. Their training and evaluation primarily rely on an Evidence Lower Bound (ELBO), which relates the Kullback-Leibler (KL) divergence of model samples to the score matching loss along the path, which serves as a tractable surrogate. The difference between sample quality and the score matching loss produced by this bound leads to the \emph{score matching gap}, which is known to be tight in the worst-case but not descriptive of sample quality in general. In this work, we provide a theoretical analysis of this gap, developing tighter bounds for three metrics: KL divergence, reverse KL divergence, and Wasserstein distance, effectively exploiting the regularity of the class of score estimators. Our results suggest that the quality of the score approximation has more impact on closing the score matching gap for low noise scales. To obtain these bounds, our key technical insight is to exploit the contraction properties of the backward processes. In particular, we rely on entropy flows, logarithmic Sobolev inequalities and reflection couplings, rigorously linking the ergodicity of the Langevin diffusion to the score matching gap problem.

Figures

Figures reproduced from arXiv: 2607.04442 by Alain Durmus, Benjamin Dupuis, Maxime Haddouche, Tyler Farghly, Umut Simsekli.

Figure 1
Figure 1. Figure 1: Decay term λT (t) of Theorem 3.7 for Mt ≡ M (full line) and for Mt = 1 + c/t (dotted line). Here T = 5. ones. We empirically investigate this in Section 5. Most importantly, our bounds decouple the time-weighting that should be used for training (Ho et al., 2020) and evaluation. Remark 3.8 (Connection to generalization bounds). Re￾cently, Farghly et al. (2026); Dupuis et al. (2025) uncovered links between … view at source ↗
Figure 2
Figure 2. Figure 2: Value of the decay term t 7→ λT (t) appearing in Theo￾rem 4.3 for different values of ρ0. with K¯ T := λ¯(T)KL γ d | −→µ T  , and, for t ∈ [0, T], λ¯(t) := r ρ0 ρ0 + e 2t − 1 ⩽ e −t/ max(1,ρ0) . (8) Remark 4.4. Similar to Remark 3.6 in the forward KL case, it follows from the joint convexity of the KL divergence that KL γ d | −→µ T  ⩽ 1 2 (e 2T − 1)W2(γ d , µ) 2 , when µ has finite moments of order 2. Th… view at source ↗
Figure 3
Figure 3. Figure 3: KL divergence and test score matching loss over training, with the SM loss split by forward-time interval. Earlier (low-noise) intervals track the KL divergence trajectory most closely. This is instrumental in our proofs, where it is used to control the distance between the two coupled SDEs (see Section C). To obtain Wasserstein contraction, a key insight from (Eberle, 2016) is to analyze the quantity f(rt… view at source ↗
Figure 6
Figure 6. Figure 6: CIFAR-10 samples with score perturbed at different times, tp. Small tp yields grainy, locally corrupted images and large tp alters global content. Score perturbations on CIFAR-10. Remaining in the same setting as the previous section, we consider a perturbed version of the score: s˜(x, t) = s(x, t) + 1tp⩽t⩽tp+δξ , where ξ is a Gaussian vector. We consider how changing tp impacts the KL divergence and score… view at source ↗
Figure 7
Figure 7. Figure 7: Weighted test score matching loss plotted against number of epochs during training. E.2. CIFAR-10 We also consider CIFAR-10, an implementation of the DDPM model from (Song et al., 2021b). We use the configuration titledvp.ddpm.cifar10 continuous which implements the DDPM model of (Ho et al., 2020) but for the continuous￾time variance preserving setting. The architecture is a U-Net (Ronneberger et al., 2015… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

267 extracted references · 4 linked inside Pith

  1. [1]

    Proceedings of the 2016

    Mart. Proceedings of the 2016

    2016

  2. [2]

    2025 , booktitle=

    Algorithm- and Data-Dependent Generalization Bounds for Score-Based Generative Models , author=. 2025 , booktitle=

    2025

  3. [3]

    Bernoulli , number =

    Patrick Cattiaux and Arnaud Guillin , title =. Bernoulli , number =

  4. [4]

    Annales de l'Institut Henri Poincaré, Probabilités et Statistiques , number =

    Feng-Yu Wang and Jian Wang , title =. Annales de l'Institut Henri Poincaré, Probabilités et Statistiques , number =

  5. [5]

    Bernoulli , number =

    Jean-Baptiste Bardet and Natha. Bernoulli , number =

  6. [6]

    2013 , author =

    Logarithmic Sobolev inequalities for mollified compactly supported measures , journal =. 2013 , author =

    2013

  7. [7]

    2001 , author =

    Hypercontractivity of Hamilton–Jacobi equations , journal =. 2001 , author =

    2001

  8. [8]

    Proceedings of The 25th International Conference on Artificial Intelligence and Statistics , pages =

    Convergence of Langevin Monte Carlo in Chi-Squared and Rényi Divergence , author =. Proceedings of The 25th International Conference on Artificial Intelligence and Statistics , pages =. 2022 , editor =

    2022

  9. [9]

    Forty-second International Conference on Machine Learning , year=

    Beyond Log-Concavity and Score Regularity: Improved Convergence Bounds for Score-Based Generative Models in W2-distance , author=. Forty-second International Conference on Machine Learning , year=

  10. [10]

    Generative Modeling by Estimating Gradients of the Data Distribution , volume =

    Song, Yang and Ermon, Stefano , booktitle =. Generative Modeling by Estimating Gradients of the Data Distribution , volume =

  11. [11]

    1983 , url=

    Partial Differential Equations of Parabolic Type , author=. 1983 , url=

    1983

  12. [12]

    Transactions on Machine Learning Research , year=

    An analysis of the noise schedule for score-based generative models , author=. Transactions on Machine Learning Research , year=

  13. [13]

    Annals of Mathematics , year=

    Note on the Derivatives with Respect to a Parameter of the Solutions of a System of Differential Equations , author=. Annals of Mathematics , year=

  14. [14]

    Train Faster, Generalize Better:

    Hardt, Moritz and Recht, Benjamin and Singer, Yoram , year =. Train Faster, Generalize Better:. 1509.01240 , primaryclass =

    Pith/arXiv arXiv

  15. [15]

    Holley, Richard and Stroock, Daniel , year =

  16. [16]

    2001 , author =

    Logarithmic Sobolev inequalities for some nonlinear PDE's , journal =. 2001 , author =

    2001

  17. [17]

    2024 , HAL_ID =

    Monmarch. 2024 , HAL_ID =

    2024

  18. [18]

    2002 , journal =

    Stability and Generalization , author =. 2002 , journal =

    2002

  19. [19]

    2025 , eprint=

    Wasserstein Bounds for generative diffusion models with Gaussian tail targets , author=. 2025 , eprint=

    2025

  20. [20]

    Diffusion Schr\"

    De Bortoli, Valentin and Thornton, James and Heng, Jeremy and Doucet, Arnaud , booktitle =. Diffusion Schr\"

  21. [21]

    Transactions on Machine Learning Research , year=

    On diffusion-based generative models and their error bounds: The log-concave case with full convergence estimates , author=. Transactions on Machine Learning Research , year=

  22. [22]

    2025 , eprint=

    Rate of convergence of the smoothed empirical Wasserstein distance , author=. 2025 , eprint=

    2025

  23. [23]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Smooth p -Wasserstein Distance: Structure, Empirical Approximation, and Statistical Applications , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

    2021

  24. [24]

    Proceedings of Thirty Fifth Conference on Learning Theory , pages =

    Improved analysis for a proximal algorithm for sampling , author =. Proceedings of Thirty Fifth Conference on Learning Theory , pages =. 2022 , editor =

    2022

  25. [25]

    The Eleventh International Conference on Learning Representations , year=

    Fast Sampling of Diffusion Models with Exponential Integrator , author=. The Eleventh International Conference on Learning Representations , year=

  26. [26]

    1982 , author =

    Reverse-time diffusion equation models , journal =. 1982 , author =

    1982

  27. [27]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Minimax Optimality of Score-based Diffusion Models: Beyond the Density Lower Bound Assumptions , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    2024

  28. [28]

    2023 , volume =

    Mbacke, Sokhna Diarra and Clerc, Florence and Germain, Pascal , booktitle =. 2023 , volume =

    2023

  29. [29]

    Advances in Neural Information Processing Systems , year=

    Maximum Likelihood Training of Score-Based Diffusion Models , author=. Advances in Neural Information Processing Systems , year=

  30. [30]

    Statistical Guarantees for Variational Autoencoders using

    Sokhna Diarra Mbacke and Florence Clerc and Pascal Germain , booktitle=. Statistical Guarantees for Variational Autoencoders using. 2023 , url=

    2023

  31. [31]

    Proceedings of Thirty Seventh Conference on Learning Theory , pages =

    Optimal score estimation via empirical Bayes smoothing , author =. Proceedings of Thirty Seventh Conference on Learning Theory , pages =. 2024 , editor =

    2024

  32. [32]

    2024 , eprint=

    From optimal score matching to optimal sampling , author=. 2024 , eprint=

    2024

  33. [33]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Improved Analysis of Score-based Generative Modeling: User-Friendly Bounds under Minimal Smoothness Assumptions , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  34. [34]

    2023 , eprint=

    Score Approximation, Estimation and Distribution Recovery of Diffusion Models on Low-Dimensional Data , author=. 2023 , eprint=

    2023

  35. [35]

    Huang, Chin-Wei and Lim, Jae Hyun and Courville, Aaron C , year =. A. Advances in

  36. [36]

    Score-Based

    Vahdat, Arash and Kreis, Karsten and Kautz, Jan , year =. Score-Based. Advances in

  37. [37]

    Advances in Neural Information Processing Systems , year=

    Convergence for score-based generative modeling with polynomial complexity , author=. Advances in Neural Information Processing Systems , year=

  38. [38]

    2022 , eprint=

    Convergence of score-based generative modeling for general data distributions , author=. 2022 , eprint=

    2022

  39. [39]

    Bogachev, V. I. and Rockner, M. and Shaposhnikov, S. V. , year =. Global. Theory of Probability & Its Applications , volume =

  40. [40]

    Advances in Neural Information Processing Systems , editor=

    Score-Based Generative Models Detect Manifolds , author=. Advances in Neural Information Processing Systems , editor=

  41. [41]

    Transactions on Machine Learning Research , year=

    Convergence of denoising diffusion models under the manifold hypothesis , author=. Transactions on Machine Learning Research , year=

  42. [42]

    2017 , eprint=

    Sharp asymptotic and finite-sample rates of convergence of empirical measures in Wasserstein distance , author=. 2017 , eprint=

    2017

  43. [43]

    Proceedings of the American Mat hematical Society , volume=

    The minimal spanning tree and the upper box dimension , author=. Proceedings of the American Mat hematical Society , volume=

  44. [44]

    2024 , eprint=

    Diffusion Models: A Comprehensive Survey of Methods and Applications , author=. 2024 , eprint=

    2024

  45. [45]

    Statistics and Computing , volume=

    Quantitative bounds of convergence for geometrically ergodic Markov chain in the Wasserstein distance with application to the Metropolis adjusted Langevin algorithm , author=. Statistics and Computing , volume=. 2015 , publisher=

    2015

  46. [46]

    2008 , journal =

    Collet, Jean-Fran. 2008 , journal =

    2008

  47. [47]

    Understanding

    Luo, Calvin , year =. Understanding. doi:10.48550/arXiv.2208.11970 , urldate =. arXiv , keywords =:2208.11970 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2208.11970

  48. [48]

    Diffusion Models Beat GANs on Image Synthesis , volume =

    Dhariwal, Prafulla and Nichol, Alexander , booktitle =. Diffusion Models Beat GANs on Image Synthesis , volume =

  49. [49]

    2017 , eprint=

    Adam: A Method for Stochastic Optimization , author=. 2017 , eprint=

    2017

  50. [50]

    Interpretation and

    Lyu, Siwei , year =. Interpretation and. ArXiv , urldate =

  51. [51]

    Divergence measures and message passing , author=

  52. [52]

    Li, Yingzhen and Turner, Richard E , year =. R. Advances in

  53. [53]

    Understanding

    Kingma, Diederik and Gao, Ruiqi , year =. Understanding. Advances in Neural Information Processing Systems , volume =

  54. [54]

    and Juergens, David and Bennett, Nathaniel R

    Watson, Joseph L. and Juergens, David and Bennett, Nathaniel R. and Trippe, Brian L. and Yim, Jason and Eisenach, Helen E. and Ahern, Woody and Borst, Andrew J. and Ragotte, Robert J. and Milles, Lukas F. and Wicky, Basile I. M. and Hanikel, Nikita and Pellock, Samuel J. and Courbet, Alexis and Sheffler, William and Wang, Jue and Venkatesh, Preetham and S...

    doi:10.1038/s41586-023-06415-8 2023

  55. [55]

    2023 , eprint=

    Convergence of the Inexact Langevin Algorithm and Score-based Generative Models in KL Divergence , author=. 2023 , eprint=

    2023

  56. [56]

    The Twelfth International Conference on Learning Representations , year=

    Neural Network-Based Score Estimation in Diffusion Models: Optimization and Generalization , author=. The Twelfth International Conference on Learning Representations , year=

  57. [57]

    2024 , eprint=

    A Good Score Does not Lead to A Good Generative Model , author=. 2024 , eprint=

    2024

  58. [58]

    Proceedings of Thirty Eighth Conference on Learning Theory , pages =

    Linear Convergence of Diffusion Models Under the Manifold Hypothesis , author =. Proceedings of Thirty Eighth Conference on Learning Theory , pages =. 2025 , editor =

    2025

  59. [59]

    and Shiryaev, Albert N

    Liptser, Robert S. and Shiryaev, Albert N. , year =. Statistics of. doi:10.1007/978-3-662-13043-8 , urldate =

    doi:10.1007/978-3-662-13043-8

  60. [60]

    Transactions of the American Mathematical Society , year=

    Quantitative Harris-type theorems for diffusions and McKean–Vlasov processes , author=. Transactions of the American Mathematical Society , year=

  61. [61]

    Convergence to Equilibrium in

    Bolley, Fran. Convergence to Equilibrium in. 2012 , journal =

    2012

  62. [62]

    Donsker, M. D. and Varadhan, S. R. S. , title =. Comm. Pure Appl. Math. , fjournal =

  63. [63]

    Fokker--

    Bogachev, Vladimir and Krylov, Nicolai and R. Fokker--. 2015 , series =

    2015

  64. [64]

    and Sturm, Karl-Theodor , year =

    Von Renesse, Max-K. and Sturm, Karl-Theodor , year =. Transport Inequalities, Gradient Estimates, Entropy and. Communications on Pure and Applied Mathematics , volume =

  65. [65]

    Lindvall, Torgny and Rogers, L. C. G. , year =. Coupling of. The Annals of Probability , volume =

  66. [66]

    Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, 1954–1955 , volume =

    Kac, Mark , title =. Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, 1954–1955 , volume =. 1956 , publisher =

    1954

  67. [67]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Diffusion Models are Minimax Optimal Distribution Estimators , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  68. [68]

    Bogachev, V. I. and R. Distances between Transition Probabilities of Diffusions and Applications to Nonlinear. 2016 , journal =

    2016

  69. [69]

    Logarithmic

    Gentil, Ivan and Imbert, Cyril , year =. Logarithmic. Asymptotic analysis , eprint =

  70. [70]

    Continuous

    Revuz, Daniel and Yor, Marc , year =. Continuous

  71. [71]

    2023 , eprint=

    Privacy Risk for anisotropic Langevin dynamics using relative entropy bounds , author=. 2023 , eprint=

    2023

  72. [72]

    Convergence of Smoothed Empirical Measures With Applications to Entropy Estimation , year=

    Goldfeld, Ziv and Greenewald, Kristjan and Niles-Weed, Jonathan and Polyanskiy, Yury , journal=. Convergence of Smoothed Empirical Measures With Applications to Entropy Estimation , year=

  73. [73]

    2023 , eprint=

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions , author=. 2023 , eprint=

    2023

  74. [74]

    The Fokker–Planck Equation: Methods of Solution and Applications , author =

  75. [75]

    2020 , MONTH = Oct, HAL_ID =

    Durmus, Alain and Eberle, Andreas and Guillin, Arnaud and Zimmer, Raphael , JOURNAL =. 2020 , MONTH = Oct, HAL_ID =

    2020

  76. [76]

    Eberle, Andreas , title =. Probab. Theory Relat. Fields , volume =

  77. [77]

    2024 , eprint=

    Convergence of Diffusion Models Under the Manifold Hypothesis in High-Dimensions , author=. 2024 , eprint=

    2024

  78. [78]

    Logarithmic Sobolev Inequalities Essentials , author =

  79. [79]

    Optimal Transport -

    Villani, C. Optimal Transport -. 2009 , publisher =

    2009

  80. [80]

    Analysis and

    Bakry, Dominique and Gentil, Ivan and Ledoux, Michel , year =. Analysis and

Showing first 80 references.