Pith. sign in

REVIEW 2 major objections 5 minor 300 references

Diffusion models cannot overfit training data and still generalize unless the sample size grows exponentially with dimension.

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

2026-07-12 07:49 UTC pith:HOIBCMWX

load-bearing objection Clean architecture-independent impossibility of simultaneous small train/test score-matching losses, plus exact linear asymptotics that isolate why score matching lacks the alignment that makes regression benignly overfit. the 2 major comments →

arxiv 2607.02671 v1 pith:HOIBCMWX submitted 2026-07-02 stat.ML cs.LG

Benign Overfitting Does Not Occur in Diffusion Models

Tyler Farghly , Benjamin Dupuis , Alain Durmus , Umut Simsekli This is my paper
classification stat.ML cs.LG MSC 68T0762G0560H10
keywords diffusion modelsbenign overfittingdouble descentscore matchingimplicit regularizationrandom feature modelsgeneralization
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 reuse the same neural nets and regression-style losses that, in ordinary deep learning, famously allow perfect training fit to coexist with good test performance (benign overfitting and double descent). This paper shows that score matching does not work that way. Unless the number of samples grows exponentially with the intrinsic dimension of the data, the empirical and population score-matching losses cannot both be small. As a result the test risk forms a classical U-shaped curve in model size rather than a second descent. The authors isolate the mechanical reason: ordinary regression can exploit an alignment between the target and the empirical covariance that cancels residual error in the null space of the data; the score target has no such alignment, so the same residual explodes as noise vanishes. Time-smoothness of a single network trained across noise levels and early stopping act as built-in regularizers that keep models away from this harmful regime, which the authors confirm on high-dimensional image models. The upshot is that the generalization story for diffusion is governed by different mechanisms than classical regression and needs its own theory.

Core claim

Unless the sample size grows exponentially with the data's lower Rényi dimension, the population and empirical score-matching losses cannot be simultaneously small; consequently the population loss follows a classical U-shaped curve in model complexity rather than double descent. The structural reason is that score matching admits no target-covariance alignment of the kind that makes residual error small in ordinary regression.

What carries the argument

A Fisher-information / total-variation lower bound (Lemma 3.1 and Theorem 3.1) that forces the empirical and population measures at small noise to be close, together with an exact asymptotic comparison of linear random-feature score matching versus multi-output regression that isolates the missing alignment.

Load-bearing premise

The quantitative exponential sample-size lower bound needs the time-weighting density to be non-decreasing on an interval and both horizon and ambient dimension to be large enough; the rate also collapses if the data measure has zero lower Rényi dimension.

What would settle it

Train a single score network on a data distribution of known positive intrinsic dimension d' with sample size well below the exponential threshold of Theorem 3.1, drive the empirical score-matching loss to near zero, and check whether the population ESM loss (or reverse-process KL) remains bounded away from zero as claimed; a simultaneous drop of both losses would refute the impossibility result.

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

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 / 5 minor

Summary. The paper argues that benign overfitting and double descent, which are central to modern deep-learning theory, do not occur for diffusion models under practical sample sizes. Architecture-independent impossibility results (Lemma 3.1, Theorem 3.1) show that the empirical and population explicit score-matching losses cannot be simultaneously small unless n grows exponentially with the lower Rényi dimension of the data; consequently the population risk is U-shaped in model complexity. A linear random-feature analysis (Proposition 4.1) makes the mechanism precise by contrasting score matching with ordinary regression: the latter benefits from alignment of the target with the empirical covariance, while the former does not. Time-smoothness of the score and early stopping are identified as implicit regularizers that keep the population loss controlled (Propositions 5.1–5.2). High-dimensional U-Net experiments on a CIFAR-10 subset corroborate the U-shaped curves and the regularizing effect of time range and early stopping.

Significance. If correct, the result cleanly separates the generalization theory of score-based generative models from the benign-overfitting paradigm that has dominated supervised deep learning. The architecture-free lower bound of Theorem 3.1 is especially valuable: it rests only on Girsanov, Pinsker and Rényi dimension, applies to any score network, and already covers the ELBO and variance weightings used in practice. The linear-RFNN asymptotics supply an explicit, Marchenko–Pastur-based explanation of why the null-space contribution explodes as t→0, and the identification of time-smoothness and early stopping as built-in regularizers is both theoretically clean and experimentally supported. Together these contributions motivate a distinct theoretical program for diffusion models and give practitioners concrete guidance on when memorization becomes inevitable.

major comments (2)
  1. [Section 3.2, Theorem 3.1] Theorem 3.1 and Remark 3.1: the exponential lower bound is stated for weightings whose density is non-decreasing on [ϵ,T]. While ELBO and variance weightings satisfy this, many practical schedules (e.g., cosine or truncated-SNR) do not. A short extension or counter-example for non-monotone ϖ would clarify the scope of the impossibility claim.
  2. [Section 4, Proposition 4.1] Proposition 4.1 and the surrounding comparison with regression: the argument that score matching admits “no alignment” relies on the isotropic Gaussian assumption (Assumption 4.1) and on the exact ERM solution AW=−Σ̂t−1. For non-Gaussian or low-rank data the residual on ker(Σ̂) need not explode as σt−2; a brief remark on the robustness of this mechanism would strengthen the claim that the phenomenon is generic rather than Gaussian-specific.
minor comments (5)
  1. [Figure 1, Appendix C.1] Figure 1 caption and Appendix C.1: the random-feature diffusion experiment uses a Monte-Carlo approximation of the DSM expectation (Ng=10). A one-sentence statement of the resulting bias would help readers assess the plotted train/test curves.
  2. [Appendix A.1] Definition A.1 and Remark A.2: the lower Rényi dimension is introduced without a pointer to the more common Hausdorff or Minkowski dimensions used in the manifold-hypothesis literature; a short cross-reference would improve accessibility.
  3. [Proposition 4.1] Equation (7) and the O(1/ψd) remainder: the asymptotic is stated for ψd,ψp>1; the transition regime ψp≈ψd is left implicit. A sentence noting continuity of the limiting expressions would remove a minor ambiguity.
  4. [Figure 2] Figure 2 (right panel): the nearest-train-image comparison is shown only for NF=8. Displaying the same panel for a larger NF (e.g., 64) would make the memorization claim visually complete.
  5. [Title page and Appendix A] Typographical: “Umut S ¸ims ¸ekli” and several occurrences of “R ´enyi” retain LaTeX spacing artifacts; these should be cleaned for the camera-ready version.

Circularity Check

0 steps flagged

No significant circularity: impossibility bounds and linear asymptotics are derived from first principles without self-definitional or fitted-input reductions.

full rationale

The central claims rest on Lemma 3.1 / Theorem 3.1 (architecture-independent lower bounds via relative Fisher information, Girsanov, Pinsker and the lower Rényi dimension) and on Proposition 4.1 / 5.1 / 5.2 (exact Marchenko–Pastur asymptotics of an explicitly solvable linear ERM problem). None of these steps defines a quantity in terms of the target phenomenon and then re-labels it as a prediction; no free parameter is fitted to data and then used to “predict” a closely related observable; and the few self-citations (Farghly et al. 2026a,b; Dupuis et al. 2025) appear only as contextual motivation or complementary generalization bounds, never as load-bearing uniqueness theorems that force the present conclusions. The modeling idealizations (Gaussian data, linear features, non-decreasing density of ϖ) are stated explicitly and the resulting U-shaped excess-risk curves follow by direct calculation, not by construction. Consequently the derivation chain is self-contained and non-circular.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central impossibility claim rests on standard concentration and information-theoretic inequalities plus the definition of lower Rényi dimension; the fine-grained linear analysis rests on the Marchenko-Pastur theorem and Gaussian data. No free parameters are fitted to produce the qualitative U-shape; experimental hyper-parameters affect only the illustrative figures.

free parameters (2)
  • ridge λ in random-feature experiments
    Set by hand to 10^{-4} for numerical stability; does not enter the theoretical claims.
  • time-horizon T and early-stop ϵ
    Chosen large/small enough for the asymptotic statements; concrete numerical values appear only in experiments.
axioms (4)
  • standard math Marchenko-Pastur theorem for empirical covariance spectra under i.i.d. entries with 4+ε moments
    Invoked in the proofs of Propositions 4.1 and 5.1 to obtain almost-sure limits of the ESM and DSM losses.
  • standard math Girsanov theorem relating path-space KL to integrated score error
    Used to convert small ESM losses into small KL(p_δ || q_{T-δ}) in the proof of Theorem 3.1.
  • domain assumption Data distribution ν has finite second moments and positive lower Rényi dimension d'
    Required for the quantitative sample-size lower bound of Theorem 3.1; holds for compact manifolds and many practical image distributions.
  • ad hoc to paper Score network is a two-layer linear random-feature model with Gaussian weights (Assumption 4.1)
    Simplifying modeling choice that enables exact asymptotics; the impossibility results of Section 3 do not rely on it.

pith-pipeline@v1.1.0-grok45 · 33734 in / 2678 out tokens · 28725 ms · 2026-07-12T07:49:02.216349+00:00 · methodology

0 comments
read the original abstract

Benign overfitting and double descent have come to shape our understanding of generalization in deep learning, establishing that overfitting is not only compatible with good generalization but can actively benefit it. Diffusion models share much of the machinery of standard deep learning, so it is natural to assume that they also exhibit these properties. In this work, we show that this assumption is largely incorrect. We first establish fundamental impossibility results showing that, unless the sample size grows exponentially with the data dimension, overfitting and good generalization cannot occur simultaneously. Consequently, the population loss follows a classical U-shaped curve in model complexity rather than exhibiting double descent. Analyzing a simplified setting, we identify a key difference between regression and score matching: regression benefits from an alignment between the target and the empirical covariance; score matching admits no such alignment, leaving overfitting irreparably harmful. We further identify implicit regularization stemming from time-smoothness of the score and early stopping during training as mechanisms that prevent such overfitting and verify our findings with high-dimensional image generation experiments. Our results reveal that generalization in diffusion models is governed by mechanisms distinct from those of traditional regression, motivating the development of new theory.

Figures

Figures reproduced from arXiv: 2607.02671 by Alain Durmus, Benjamin Dupuis, Tyler Farghly, Umut Simsekli.

Figure 1
Figure 1. Figure 1: (Left) 2-layer random feature network regression, displaying the classical double descent curve. (Right) 2-layer random feature network diffusion, showing the absence of benign overfitting. Details on the setup and hyperparameters for these experiments are available in Section C.1. known as benign overfitting (Bartlett et al., 2020). This is explained by theoretical analysis showing that generalization is … view at source ↗
Figure 2
Figure 2. Figure 2: We train a DDPM U-Net model on a subset of CIFAR10 to convergence. We observe train and test error for different configurations of the model, varying the number of features (NF) in the U-Net. (Left) Test and train error at convergence varying NF and thus, varying the number of parameters in the model. (Center) The same setting but for the population negative log-likelihood. (Right) Generated samples for di… view at source ↗
Figure 3
Figure 3. Figure 3: Linear random features under Assumption 4.1. (Left) time is fixed at t = 10−1 and ψp varies. (Center) ψd is fixed at 102 and t varies. (Right) time integrated losses (as in Section 5.1) with π the uniform distribution on [0, T]. Train error denotes the empirical DSM loss with the weighting used in training in practice. Test error is the population ESM loss with ELBO weighting. Theorem 3.1. Suppose that ϖ i… view at source ↗
Figure 5
Figure 5. Figure 5: The same setting as [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] 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

300 extracted references · 68 linked inside Pith

  1. [1]

    Random Features for Large-Scale Kernel Machines , volume =

    Rahimi, Ali and Recht, Benjamin , booktitle =. Random Features for Large-Scale Kernel Machines , volume =

  2. [2]

    The Eleventh International Conference on Learning Representations , year=

    Statistical Efficiency of Score Matching: The View from Isoperimetry , author=. The Eleventh International Conference on Learning Representations , year=

  3. [3]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  4. [4]

    EurIPS 2025 Workshop on Principles of Generative Modeling (PriGM) , year=

    Generalization Dynamics of Linear Diffusion Models , author=. EurIPS 2025 Workshop on Principles of Generative Modeling (PriGM) , year=

    2025

  5. [5]

    Marchenko, V. A. and Pastur, L. A. , year=. Distribution of Eigenvalues for some Sets of Random Matrices , volume=. Mathematics of the USSR-Sbornik , publisher=

  6. [6]

    2023 , eprint=

    Six Lectures on Linearized Neural Networks , author=. 2023 , eprint=

    2023

  7. [7]

    ICLR 2026 Workshop on Geometry-grounded Representation Learning and Generative Modeling , year=

    Manifold Generalization Provably Proceeds Memorization in Diffusion Models , author=. ICLR 2026 Workshop on Geometry-grounded Representation Learning and Generative Modeling , year=

    2026

  8. [8]

    Comment on: ``

    Otto, Felix and Villani, C. Comment on: ``. 2001 , journal =

    2001

  9. [9]

    Procaccia and P

    I. Procaccia and P. Grassberger and H. G. E. Hentschel , title =. Dynamical Systems and Chaos , year =

  10. [10]

    1993 , month = may, journal =

    On Rigorous Mathematical Definitions of Correlation Dimension and Generalized Spectrum for Dimensions , author =. 1993 , month = may, journal =. doi:10.1007/BF01058436 , urldate =

    doi:10.1007/bf01058436 1993

  11. [11]

    The Fourteenth International Conference on Learning Representations , year=

    Provable Separations between Memorization and Generalization in Diffusion Models , author=. The Fourteenth International Conference on Learning Representations , year=

  12. [12]

    Denoising

    George, Anand Jerry and Veiga, Rodrigo and Macris, Nicolas , year =. Denoising

  13. [13]

    Reversible

    Aldous, David and Allen Fill, James , year =. Reversible

  14. [14]

    Explicit Error Bounds for

    Rudolf, Daniel , year =. Explicit Error Bounds for. Dissertationes Mathematicae , volume =. 1108.3201 , primaryclass =

    Pith/arXiv arXiv

  15. [15]

    Nakkiran, Preetum and Kaplun, Gal and Bansal, Yamini and Yang, Tristan and Barak, Boaz and Sutskever, Ilya , year =. Deep. ICLR 2020 , eprint =

    2020

  16. [16]

    Fractional

    Tristani, Isabelle , year =. Fractional. Commun. Math. Sci. 13 , eprint =

  17. [17]

    Proceedings of the Thirty-Second Conference on Learning Theory , pages =

    Normal Approximation for Stochastic Gradient Descent via Non-Asymptotic Rates of Martingale CLT , author=. Proceedings of the Thirty-Second Conference on Learning Theory , pages =. 2019 , volume =

    2019

  18. [18]

    The Cifar-10 Dataset , author =

  19. [19]

    2009 , url=

    Learning Multiple Layers of Features from Tiny Images , author=. 2009 , url=

    2009

  20. [20]

    , year =

    Hodgkinson, Liam and Mahoney, Michael W. , year =. Multiplicative Noise and Heavy Tails in Stochastic Optimization , booktitle =. 2006.06293 , primaryclass =

    Pith/arXiv arXiv 2006

  21. [21]

    Journal of Machine Learning Research , volume=

    Topology of deep neural networks , author=. Journal of Machine Learning Research , volume=

  22. [22]

    Very Deep Convolutional Networks for Large-Scale Image Recognition , booktitle =

    Karen Simonyan and Andrew Zisserman , editor =. Very Deep Convolutional Networks for Large-Scale Image Recognition , booktitle =

  23. [23]

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , booktitle =

  24. [24]

    North American Chapter of the Association for Computational Linguistics , year=

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , author=. North American Chapter of the Association for Computational Linguistics , year=

  25. [25]

    Language

    Brown, Tom and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared D and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and. Language. Advances in. 2020 , volume =

    2020

  26. [26]

    and Monro, S

    Robbins, H. and Monro, S. , journal =

  27. [27]

    Bottou, L. The. Advances in. 2007 , volume =

    2007

  28. [28]

    , year =

    Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey E. , year =. Communications of the ACM , volume =

  29. [29]

    Foundations of

    Kallenberg, Olav , year =. Foundations of. doi:10.1007/978-3-030-61871-1 , urldate =

    doi:10.1007/978-3-030-61871-1

  30. [30]

    2023 IEEE 6th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) , pages=

    Deep neural networks architectures from the perspective of manifold learning , author=. 2023 IEEE 6th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) , pages=. 2023 , organization=

    2023

  31. [31]

    arXiv preprint arXiv:2204.08624 , year=

    Topology and geometry of data manifold in deep learning , author=. arXiv preprint arXiv:2204.08624 , year=

    Pith/arXiv arXiv

  32. [32]

    Understanding Machine Learning: From Theory to Algorithms , publisher=

    Shalev-Shwartz, Shai and Ben-David, Shai , year=. Understanding Machine Learning: From Theory to Algorithms , publisher=

  33. [33]

    Understanding Machine Learning: From Theory to Algorithms , publisher=

    Bach, Francis , series =. Understanding Machine Learning: From Theory to Algorithms , publisher=

  34. [34]

    Theory of Probability and its Applications , year =

    Vladimir Vapnik and Alexey Chervonenkis , title =. Theory of Probability and its Applications , year =

  35. [35]

    Doklady Akademii Nauk USSR , year =

    Vladimir Vapnik and Alexey Chervonenkis , title =. Doklady Akademii Nauk USSR , year =

  36. [36]

    Differentially

    Kuru, Nurdan and Birbil,. Differentially. 2022 , month = jun, journal =. 2008.01989 , primaryclass =

    Pith/arXiv arXiv 2022

  37. [37]

    and Blei, David M

    Mandt, Stephan and Hoffman, Matthew D. and Blei, David M. , year =. A. International

  38. [38]

    2023 , eprint=

    Equivalences of Geometric Ergodicity of Markov Chains , author=. 2023 , eprint=

    2023

  39. [39]

    Pertti Mattila , title =

  40. [40]

    John Mackay and Jeremy Tyson , title =

  41. [41]

    Estimating

    Herrera, Calypso and Krach, Florian and Teichmann, Josef , year =. Estimating. Estimating Full Lipschitz Constants of Deep Neural Networks , eprint =

  42. [42]

    2018 , month = jan, journal =

    Stochastic Gradient Descent Performs Variational Inference, Converges to Limit Cycles for Deep Networks , author =. 2018 , month = jan, journal =. 1710.11029 , primaryclass =

    Pith/arXiv arXiv 2018

  43. [43]

    , year =

    Chaudhuri, Kamalika and Monteleoni, Claire and Sarwate, Anand D. , year =. Differentially. Journal of Machine Learning Research , volume =. 0912.0071 , primaryclass =

    Pith/arXiv arXiv

  44. [44]

    Annals of Mathematics and Artificial Intelligence , volume=

    Topological measurement of deep neural networks using persistent homology , author=. Annals of Mathematics and Artificial Intelligence , volume=. 2022 , publisher=

    2022

  45. [45]

    and Tweedie, Richard L

    Meyn, Sean P. and Tweedie, Richard L. , year =. Markov

  46. [46]

    and Braun, H

    Riedmiller, M. and Braun, H. , booktitle=. A direct adaptive method for faster backpropagation learning: the RPROP algorithm , year=

  47. [47]

    Fashion-

    Xiao, Han and Rasul, Kashif and Vollgraf, Roland , year =. Fashion-. 1708.07747 , primaryclass =

    Pith/arXiv arXiv

  48. [48]

    High-dimensional limit theorems for SGD: Effective dynamics and critical scaling , booktitle=

    Gerard Ben Arous and Reza Gheissari and Aukosh Jagannath , year=. High-dimensional limit theorems for SGD: Effective dynamics and critical scaling , booktitle=. 2206.04030 , archivePrefix=

    Pith/arXiv arXiv

  49. [49]

    Evolution of the

    Alfonsi, Aur. Evolution of the. 2016 , month = dec, number =. doi:10.48550/arXiv.1606.02994 , urldate =. 1606.02994 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.1606.02994 2016

  50. [50]

    and Orbanz, Peter , year =

    Zhou, Wenda and Veitch, Victor and Austern, Morgane and Adams, Ryan P. and Orbanz, Peter , year =. Non-Vacuous. International

  51. [51]

    2024 , keywords =

    Alquier, Pierre , journal=. 2024 , keywords =

    2024

  52. [52]

    Thinking Outside the Ball: Optimal Learning with Gradient Descent for Generalized Linear Stochastic Convex Optimization , volume =

    Amir, Idan and Livni, Roi and Srebro, Nati , booktitle =. Thinking Outside the Ball: Optimal Learning with Gradient Descent for Generalized Linear Stochastic Convex Optimization , volume =

  53. [53]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Ron Amit and Baruch Epstein and Shay Moran and Ron Meir , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  54. [54]

    Topological

    Andreeva, Rayna and Dupuis, Benjamin and Sarkar, Rik and Birdal, Tolga and Simsekli, Umut , year =. Topological. 38th. 2407.08723 , primaryclass =

    Pith/arXiv arXiv

  55. [55]

    ICML 2023 Workshop on Topological, Algebraic and Geometric Learning , year =

    Rayna Andreeva and Katharina Limbeck and Bastian Rieck and Rik Sarkar , title =. ICML 2023 Workshop on Topological, Algebraic and Geometric Learning , year =

    2023

  56. [56]

    Metric Space Magnitude for Evaluating the Diversity of Latent Representations , volume =

    Limbeck, Katharina and Andreeva, Rayna and Sarkar, Rik and Rieck, Bastian , booktitle =. Metric Space Magnitude for Evaluating the Diversity of Latent Representations , volume =. doi:10.52202/079017-3937 , editor =

    doi:10.52202/079017-3937

  57. [57]

    Deep Learning in Optical Metrology: A Review , shorttitle =

    Zuo, Chao and Qian, Jiaming and Feng, Shijie and Yin, Wei and Li, Yixuan and Fan, Pengfei and Han, Jing and Qian, Kemao and Chen, Qian , year =. Deep Learning in Optical Metrology: A Review , shorttitle =. Light: Science & Applications , volume =

  58. [58]

    1999 , publisher =

    Neural Network Learning: Theoretical Foundations , author =. 1999 , publisher =

    1999

  59. [59]

    , year =

    Andrieu, Christophe and Lee, Anthony and Power, Sam and Wang, Andi Q. , year =. Poincar. 2208.05239 , publisher =

    Pith/arXiv arXiv

  60. [60]

    On Logarithmic

    An. On Logarithmic. 2000 , journal =

    2000

  61. [61]

    Estimating

    Aristoff, David , year =. Estimating. 1205.2400 , primaryclass =

    Pith/arXiv arXiv

  62. [62]

    and Abbe, Emmanuel and Verd

    Asadi, Amir R. and Abbe, Emmanuel and Verd. Chaining. 2019 , month = jul, number =. 1806.03803 , primaryclass =

    Pith/arXiv arXiv 2019

  63. [63]

    Analysis and

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

  64. [64]

    Probability Measures on Groups , author =

    Functional. Probability Measures on Groups , author =. 2004 , pages =

    2004

  65. [65]

    2024 , eprint=

    Stochastic Differential Equations models for Least-Squares Stochastic Gradient Descent , author=. 2024 , eprint=

    2024

  66. [66]

    and Larchev, M

    Schertzer, D. and Larchev, M. and Duan, J. and Yanovsky, V. V. and Lovejoy, S. , year =. Fractional. Journal of Mathematical Physics , volume =. math/0409486 , pages =

    Pith/arXiv arXiv

  67. [67]

    Freitag, Eberhard and Rusam, Rolf , year =. Complex

  68. [68]

    Lang, Serge , year =. Complex

  69. [69]

    Distributions and

    Grubb, Gerd , year =. Distributions and

  70. [70]

    International Conference on Learning Representations , year=

    A Diffusion Theory For Deep Learning Dynamics: Stochastic Gradient Descent Exponentially Favors Flat Minima , author=. International Conference on Learning Representations , year=

  71. [71]

    Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics , pages =

    Stable behaviour of infinitely wide deep neural networks , author =. Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics , pages =. 2020 , editor =

    2020

  72. [72]

    Stochastic

    Wojtowytsch, Stephan , year =. Stochastic. Journal of Nonlinear Science , volume =

  73. [73]

    Modified Log-

    Adamczak, Rados. Modified Log-. 2022 , month = apr, journal =. 2007.10209 , primaryclass =

    Pith/arXiv arXiv 2022

  74. [74]

    International Conference on Learning Representations , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations , year=

  75. [75]

    SIAM Journal on Mathematics of Data Science , volume=

    KL convergence guarantees for score diffusion models under minimal data assumptions , author=. SIAM Journal on Mathematics of Data Science , volume=. 2025 , publisher=

    2025

  76. [76]

    1982 , issn =

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

    1982

  77. [77]

    Jung, Paul and Lee, Hoil and Lee, Jiho and Yang, Hongseok , year =. alpha-. Advances in Applied Probability , Volume 55 , Issue 4 , eprint =

  78. [78]

    2020 , journal =

    A Fractal Dimension for Measures via Persistent Homology , author =. 2020 , journal =. 1808.01079 , primaryclass =

    Pith/arXiv arXiv 2020

  79. [79]

    Frontiers in Artificial Intelligence , volume=

    Topology applied to machine learning: From global to local , author=. Frontiers in Artificial Intelligence , volume=. 2021 , publisher=

    2021

  80. [80]

    Stability

    Banerjee, Arindam and Chen, Tiancong and Li, Xinyan and Zhou, Yingxue , year =. Stability. arXiv.org , urldate =

Showing first 80 references.