pith. sign in

arxiv: 2606.17252 · v1 · pith:T6HZKJVAnew · submitted 2026-06-15 · ❄️ cond-mat.dis-nn · cond-mat.stat-mech

Stochastic Thermodynamics of Score Matching in Diffusion Models

Xuehao Ding , H. T. Quan , Yuhai Tu This is my paper

Pith reviewed 2026-06-27 01:40 UTC · model grok-4.3

classification ❄️ cond-mat.dis-nn cond-mat.stat-mech
keywords diffusion modelsscore matchingstochastic thermodynamicsentropy productionfluctuation theoremsgenerative modelssampling diversitytime asymmetry
0
0 comments X

The pith

Time-asymmetry entropy production averages exactly to the score-matching objective in diffusion models.

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

The paper establishes a stochastic thermodynamic framework for score-based diffusion models by introducing a trajectory-dependent time-asymmetry entropy production (TAEP) constructed from the forward diffusion process and the learned reverse process. It demonstrates that TAEP obeys exact fluctuation theorems, that Hyvärinen's implicit score-matching kernel appears directly as the fluctuating part of TAEP, and that the average value of TAEP is exactly proportional to the score-matching objective used for training. This identification supplies a thermodynamic account of sampling unevenness, data-manifold coverage, and the preference of stochastic gradient descent for flatter minima. A sympathetic reader would care because the result supplies a concrete statistical-mechanical reason for the observed sampling diversity and generalization properties of these generative models.

Core claim

We introduce the time-asymmetry entropy production (TAEP) defined from the forward and reverse diffusion dynamics in score-based generative models. This quantity obeys exact fluctuation theorems from stochastic thermodynamics. Hyvärinen's implicit score-matching kernel emerges as a fluctuating component of TAEP, while the average TAEP is exactly proportional to the score-matching objective. Fluctuations of TAEP quantify sampling unevenness and provide a thermodynamic measure of data-manifold coverage. These results explain the superior sampling diversity of diffusion models and reveal a thermodynamic mechanism by which stochastic gradient descent favors flatter, more generalizable solutions.

What carries the argument

Time-asymmetry entropy production (TAEP), a trajectory-dependent quantity defined from the forward and reverse diffusion dynamics that obeys fluctuation theorems and whose average equals the score-matching objective.

If this is right

  • The score-matching objective equals the average of TAEP over trajectories.
  • Fluctuations of TAEP directly quantify unevenness in the model's sampling distribution.
  • TAEP supplies a thermodynamic measure of how completely the generated samples cover the data manifold.
  • The framework accounts for the observed sampling diversity of diffusion models relative to other generative approaches.
  • Stochastic gradient descent during training favors flatter minima because of the underlying thermodynamic structure.

Where Pith is reading between the lines

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

  • The same construction could be applied to other generative architectures by defining an analogous forward-reverse asymmetry measure.
  • One could check whether explicitly penalizing TAEP fluctuations during training improves sample diversity on benchmark datasets.
  • The fluctuation theorems might yield new bounds on the minimal number of reverse steps needed for accurate sampling.
  • This view suggests that the training dynamics themselves carry an entropic cost that could be monitored to detect overfitting.

Load-bearing premise

The learned reverse diffusion process accurately implements the probability flow reversal of the forward process.

What would settle it

A numerical check on a trained diffusion model (for example on a Gaussian mixture) in which the trajectory-averaged TAEP fails to equal the value of the score-matching loss.

read the original abstract

Score-based diffusion models are a powerful class of generative AI systems capable of sampling from complex, high-dimensional probability distributions. Their dynamics consist of a forward diffusion process that transforms data into noise and a learned reverse process that reconstructs data by reversing the probability flow. Here, we develop a stochastic thermodynamic framework for diffusion models and their score-matching objective. We introduce a trajectory-dependent quantity, time-asymmetry entropy production (TAEP), defined from the forward and reverse diffusion dynamics, and show that it obeys exact fluctuation theorems. Remarkably, Hyv\"{a}rinen's implicit score-matching kernel emerges naturally as a fluctuating component of TAEP, while the average TAEP is exactly proportional to the score-matching objective. We further show that fluctuations of TAEP quantify sampling unevenness and provide a thermodynamic measure of data-manifold coverage. These results yield a quantitative explanation for the superior sampling diversity of diffusion models and reveal a thermodynamic mechanism by which stochastic gradient descent favors flatter, more generalizable solutions. By uncovering the entropic nature of score matching, our work establishes fundamental statistical-mechanical principles underlying diffusion-based generative AI.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript develops a stochastic thermodynamic framework for score-based diffusion models. It introduces a trajectory-dependent time-asymmetry entropy production (TAEP) defined from the forward and reverse diffusion dynamics, claims that TAEP obeys exact fluctuation theorems, that Hyvärinen's implicit score-matching kernel emerges naturally as a fluctuating component of TAEP, and that the average TAEP is exactly proportional to the score-matching objective. It further claims that TAEP fluctuations quantify sampling unevenness and data-manifold coverage, providing a thermodynamic explanation for diffusion models' sampling diversity and for why SGD favors flatter solutions.

Significance. If the derivations establish the claimed exact relations without hidden dependence on the learned score or unstated assumptions about perfect reversal, the work would offer a novel statistical-mechanical interpretation of score matching and generative performance in diffusion models, with potential to connect fluctuation theorems to machine-learning objectives.

major comments (2)
  1. [Abstract; definition and fluctuation theorem for TAEP] The central claims of exact fluctuation theorems and exact proportionality between average TAEP and the score-matching objective rest on the assumption that the learned reverse process exactly implements the probability-flow reversal of the forward SDE. Because the reverse dynamics are obtained via score matching (an approximation), the time-reversal relation holds only in the zero-error limit; the manuscript must clarify whether the relations remain exact or become approximations and, if the latter, provide controlled error bounds. This assumption is load-bearing for all 'exact' statements in the abstract.
  2. [Definition of TAEP] The definition of TAEP is stated to be constructed from forward and reverse dynamics independently of the training objective, yet the claimed emergence of the score-matching kernel as a fluctuating component of TAEP requires explicit verification that no implicit dependence on the fitted score function enters the definition. The manuscript should display the explicit expression for TAEP (prior to any averaging) to confirm the claimed independence.
minor comments (1)
  1. Notation for the implicit score-matching kernel should be introduced once and used consistently; the abstract uses 'Hyvärinen's implicit score-matching kernel' while later text should avoid redefining it.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and valuable comments. We address each major point below and will revise the manuscript to improve clarity on the assumptions underlying our exact relations.

read point-by-point responses

  1. Referee: [Abstract; definition and fluctuation theorem for TAEP] The central claims of exact fluctuation theorems and exact proportionality between average TAEP and the score-matching objective rest on the assumption that the learned reverse process exactly implements the probability-flow reversal of the forward SDE. Because the reverse dynamics are obtained via score matching (an approximation), the time-reversal relation holds only in the zero-error limit; the manuscript must clarify whether the relations remain exact or become approximations and, if the latter, provide controlled error bounds. This assumption is load-bearing for all 'exact' statements in the abstract.

    Authors: We agree that the derivations rely on the reverse process being the exact time-reversal of the forward SDE. This holds precisely when the score is perfect (zero error), and under this assumption the fluctuation theorems and proportionality are exact. The manuscript presents the framework in this ideal setting. We will revise the abstract and main text to explicitly state this assumption and clarify that the exact relations refer to the zero-error limit. We do not derive explicit error bounds for the approximate case, as that would require a separate analysis of score approximation errors. revision: yes

  2. Referee: [Definition of TAEP] The definition of TAEP is stated to be constructed from forward and reverse dynamics independently of the training objective, yet the claimed emergence of the score-matching kernel as a fluctuating component of TAEP requires explicit verification that no implicit dependence on the fitted score function enters the definition. The manuscript should display the explicit expression for TAEP (prior to any averaging) to confirm the claimed independence.

    Authors: The TAEP is defined directly from the forward SDE and the reverse SDE (with its drift term), without reference to the particular optimization procedure used to obtain the score. The score-matching kernel appears upon algebraic rearrangement of the trajectory-dependent expression. To verify independence from the training objective, we will add the explicit unaveraged expression for TAEP in the revised manuscript, showing that it depends only on the forward and reverse dynamics. revision: yes

Circularity Check

0 steps flagged

No significant circularity; TAEP defined independently and proportionality presented as emergent

full rationale

TAEP is introduced as a trajectory-dependent quantity defined directly from the forward and reverse diffusion dynamics. The paper states that Hyvärinen's kernel emerges as a fluctuating component and that average TAEP is exactly proportional to the score-matching objective, but presents both as consequences of applying fluctuation theorems rather than as definitional identities. No equations or steps are shown that reduce the claimed proportionality to a fitted parameter or self-citation chain by construction. The modeling assumption that the learned reverse process implements exact probability-flow reversal is a standard idealization in the field and does not constitute a load-bearing self-referential step within the derivation itself. The central results therefore remain self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework introduces TAEP as a new trajectory-dependent quantity and relies on the applicability of fluctuation theorems to diffusion processes; no explicit free parameters are mentioned in the abstract.

axioms (1)
  • domain assumption Fluctuation theorems hold for the time-asymmetry entropy production defined on forward and reverse diffusion trajectories
    Invoked to obtain exact relations between TAEP statistics and the score-matching objective.
invented entities (1)
  • time-asymmetry entropy production (TAEP) no independent evidence
    purpose: Trajectory-dependent measure of irreversibility whose average equals the score-matching objective and whose fluctuations quantify sampling unevenness
    New quantity defined from the forward and reverse dynamics; no independent evidence outside the paper is provided in the abstract.

pith-pipeline@v0.9.1-grok · 5726 in / 1251 out tokens · 37906 ms · 2026-06-27T01:40:07.967742+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Estimating Free Energy Differences with Virtually Escorted Trajectories

    cond-mat.stat-mech 2026-06 unverdicted novelty 7.0

    Introduces virtually escorted trajectories with a virtual control field to generate multiple optimized work quantities satisfying the fluctuation theorem for improved free energy estimation convergence.

Reference graph

Works this paper leans on

53 extracted references · cited by 1 Pith paper

  1. [1]

    In: International Confer- ence on Machine Learning, pp

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: International Confer- ence on Machine Learning, pp. 2256–2265 (2015). pmlr

    2015

  2. [2]

    Physical Review Letters78(14), 2690 (1997)

    Jarzynski, C.: Nonequilibrium equality for free energy differences. Physical Review Letters78(14), 2690 (1997)

    1997

  3. [3]

    Physical review letters95(4), 040602 (2005)

    Seifert, U.: Entropy production along a stochastic trajectory and an integral fluctuation theorem. Physical review letters95(4), 040602 (2005)

    2005

  4. [4]

    Advances in neural information processing systems32(2019)

    Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems32(2019)

    2019

  5. [5]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  6. [6]

    In: International Conference on Learning Representations

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations

  7. [7]

    Stochastic Processes and their Applications12(3), 313–326 (1982)

    Anderson, B.D.: Reverse-time diffusion equation models. Stochastic Processes and their Applications12(3), 313–326 (1982)

    1982

  8. [8]

    Journal of Machine Learning Research6(4) (2005)

    Hyv¨ arinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research6(4) (2005)

    2005

  9. [9]

    Physical Review E111(1), 014111 (2025)

    Yu, Z., Huang, H.: Nonequilbrium physics of generative diffusion models. Physical Review E111(1), 014111 (2025)

    2025

  10. [10]

    Physical Review X15(3), 031031 (2025) 18

    Ikeda, K., Uda, T., Okanohara, D., Ito, S.: Speed-accuracy relations for diffusion models: Wisdom from nonequilibrium thermodynamics and optimal transport. Physical Review X15(3), 031031 (2025) 18

    2025

  11. [11]

    Journal of Statistical Mechanics: Theory and Experiment2023(9), 093402 (2023)

    Biroli, G., M´ ezard, M.: Generative diffusion in very large dimensions. Journal of Statistical Mechanics: Theory and Experiment2023(9), 093402 (2023)

    2023

  12. [12]

    Advances in Neural Information Processing Systems36, 66377–66389 (2023)

    Raya, G., Ambrogioni, L.: Spontaneous symmetry breaking in generative diffusion models. Advances in Neural Information Processing Systems36, 66377–66389 (2023)

    2023

  13. [13]

    Nature Communications15(1), 9957 (2024)

    Biroli, G., Bonnaire, T., De Bortoli, V., M´ ezard, M.: Dynamical regimes of diffusion models. Nature Communications15(1), 9957 (2024)

    2024

  14. [14]

    Entropy27(3), 291 (2025)

    Ambrogioni, L.: The statistical thermodynamics of generative diffusion models: Phase transitions, symmetry breaking, and critical instability. Entropy27(3), 291 (2025)

    2025

  15. [15]

    Lecture Notes in Physics (Springer, Berlin) 799(2010)

    Sekimoto, K.: Stochastic energetics. Lecture Notes in Physics (Springer, Berlin) 799(2010)

    2010

  16. [16]

    Reports on progress in physics75(12), 126001 (2012)

    Seifert, U.: Stochastic thermodynamics, fluctuation theorems and molecular machines. Reports on progress in physics75(12), 126001 (2012)

    2012

  17. [17]

    Physical review letters104(9), 090601 (2010)

    Esposito, M., Van den Broeck, C.: Three detailed fluctuation theorems. Physical review letters104(9), 090601 (2010)

    2010

  18. [18]

    Physica A: Statistical Mechanics and its Applications418, 6–16 (2015)

    Van den Broeck, C., Esposito, M.: Ensemble and trajectory thermodynamics: A brief introduction. Physica A: Statistical Mechanics and its Applications418, 6–16 (2015)

    2015

  19. [19]

    Journal of statistical physics 110(1), 269–310 (2003)

    Maes, C., Netoˇ cn` y, K.: Time-reversal and entropy. Journal of statistical physics 110(1), 269–310 (2003)

    2003

  20. [20]

    Progress of Theoretical Physics Supplement130, 29–44 (1998)

    Oono, Y., Paniconi, M.: Steady state thermodynamics. Progress of Theoretical Physics Supplement130, 29–44 (1998)

    1998

  21. [21]

    Physical review letters86(16), 3463 (2001)

    Hatano, T., Sasa, S.-i.: Steady-state thermodynamics of langevin systems. Physical review letters86(16), 3463 (2001)

    2001

  22. [22]

    Esposito, M., Van den Broeck, C.: Three faces of the second law. i. master equation formulation. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics82(1), 011143 (2010)

    2010

  23. [23]

    Van den Broeck, C., Esposito, M.: Three faces of the second law. ii. fokker-planck formulation. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics 82(1), 011144 (2010)

    2010

  24. [24]

    Physical Review E—Statistical, Nonlinear, and Soft Matter Physics76(1), 011123 (2007) 19

    Lau, A.W., Lubensky, T.C.: State-dependent diffusion: Thermodynamic con- sistency and its path integral formulation. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics76(1), 011123 (2007) 19

    2007

  25. [25]

    Journal of Statistical Mechanics: Theory and Experiment2006(08), 08001–08001 (2006)

    Chernyak, V.Y., Chertkov, M., Jarzynski, C.: Path-integral analysis of fluctuation theorems for general langevin processes. Journal of Statistical Mechanics: Theory and Experiment2006(08), 08001–08001 (2006)

    2006

  26. [26]

    In: Uncertainty in Artificial Intelligence, pp

    Song, Y., Garg, S., Shi, J., Ermon, S.: Sliced score matching: A scalable approach to density and score estimation. In: Uncertainty in Artificial Intelligence, pp. 574–584 (2020). PMLR

    2020

  27. [27]

    Physical Review E56(5), 5018 (1997)

    Jarzynski, C.: Equilibrium free-energy differences from nonequilibrium measure- ments: A master-equation approach. Physical Review E56(5), 5018 (1997)

    1997

  28. [28]

    Physical Review E98(4), 042122 (2018)

    Ding, X., Yi, J., Kim, Y.W., Talkner, P.: Measurement-driven single temperature engine. Physical Review E98(4), 042122 (2018)

    2018

  29. [29]

    Physical review letters120(19), 190602 (2018)

    Pietzonka, P., Seifert, U.: Universal trade-off between power, efficiency, and con- stancy in steady-state heat engines. Physical review letters120(19), 190602 (2018)

    2018

  30. [30]

    Advances in neural information processing systems29(2016)

    Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems29(2016)

    2016

  31. [31]

    The Annals of Statistics, 171–189 (2014)

    Li, K.: Second-order asymptotics for quantum hypothesis testing. The Annals of Statistics, 171–189 (2014)

    2014

  32. [32]

    In: NeurIPS 2021 Work- shop on Deep Generative Models and Downstream Applications

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Work- shop on Deep Generative Models and Downstream Applications

    2021

  33. [33]

    In: International Conference on Machine Learning, pp

    Naeem, M.F., Oh, S.J., Uh, Y., Choi, Y., Yoo, J.: Reliable fidelity and diversity metrics for generative models. In: International Conference on Machine Learning, pp. 7176–7185 (2020). PMLR

    2020

  34. [34]

    In: International Conference on Machine Learning, pp

    Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning, pp. 8162–8171 (2021). PMLR

    2021

  35. [35]

    Advances in neural information processing systems34, 8780–8794 (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

    2021

  36. [36]

    In: International Conference on Artificial Neural Networks (2018)

    Jastrzebski, S., Kenton, Z., Arpit, D., Ballas, N., Fischer, A., Bengio, Y., Storkey, A.: Three factors influencing minima in sgd. In: International Conference on Artificial Neural Networks (2018)

    2018

  37. [37]

    Proceedings of the National Academy of Sciences118(9), 2015617118 (2021)

    Feng, Y., Tu, Y.: The inverse variance–flatness relation in stochastic gradient descent is critical for finding flat minima. Proceedings of the National Academy of Sciences118(9), 2015617118 (2021)

    2021

  38. [38]

    Physical Review 20 Letters130(23), 237101 (2023)

    Yang, N., Tang, C., Tu, Y.: Stochastic gradient descent introduces an effec- tive landscape-dependent regularization favoring flat solutions. Physical Review 20 Letters130(23), 237101 (2023)

    2023

  39. [39]

    Nature Machine Intelligence 5(8), 908–918 (2023)

    Feng, Y., Zhang, W., Tu, Y.: Activity–weight duality in feed-forward neural net- works reveals two co-determinants for generalization. Nature Machine Intelligence 5(8), 908–918 (2023)

    2023

  40. [40]

    In: International Conference on Machine Learning, pp

    Zhu, Z., Wu, J., Yu, B., Wu, L., Ma, J.: The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects. In: International Conference on Machine Learning, pp. 7654–7663 (2019). PMLR

    2019

  41. [41]

    In: Proceedings of the 2020 SIAM International Conference on Data Mining, pp

    Li, X., Gu, Q., Zhou, Y., Chen, T., Banerjee, A.: Hessian based analysis of sgd for deep nets: Dynamics and generalization. In: Proceedings of the 2020 SIAM International Conference on Data Mining, pp. 190–198 (2020). SIAM

    2020

  42. [42]

    https://arxiv.org/abs/2602

    Zhang, Y., Yang, N., Tu, Y.: On the Superlinear Relationship between SGD Noise Covariance and Loss Landscape Curvature (2026). https://arxiv.org/abs/2602. 05600

    2026

  43. [43]

    Reviews of modern physics 81(4), 1665–1702 (2009)

    Esposito, M., Harbola, U., Mukamel, S.: Nonequilibrium fluctuations, fluctuation theorems, and counting statistics in quantum systems. Reviews of modern physics 81(4), 1665–1702 (2009)

    2009

  44. [44]

    Journal of Statistical Physics156(1), 55–65 (2014)

    Horowitz, J.M., Sagawa, T.: Equivalent definitions of the quantum nonadiabatic entropy production. Journal of Statistical Physics156(1), 55–65 (2014)

    2014

  45. [45]

    Physical Review X8(3), 031037 (2018)

    Manzano, G., Horowitz, J.M., Parrondo, J.M.: Quantum fluctuation theorems for arbitrary environments: Adiabatic and nonadiabatic entropy production. Physical Review X8(3), 031037 (2018)

    2018

  46. [46]

    Hutchinson, M.F.: A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines. Communications in Statistics-Simulation and Computation18(3), 1059–1076 (1989) 21 Supplementary Material for the paper ”Stochastic Thermodynamics of Score Matching in Diffusion Models” Xuehao Ding, 1 H. T. Quan, 2, 3, 4 and Yuhai Tu 1 1Flatiron I...

    1989

  47. [47]

    Suppose that there is a datasetPwith distributionp(x,0) and a related datasetQwith distribution q(x,0)

    Transfer learning, which is a machine learning paradigm that leverages knowledge acquired from a pre-trained source domain to enhance the learning efficiency and predictive performance of a model on a related but distinct task. Suppose that there is a datasetPwith distributionp(x,0) and a related datasetQwith distribution q(x,0). Suppose that a diffusion ...

  48. [48]

    Similarly to transfer learning, the model has been fully trained on the train setQ, and now we test the model on the test setP

    Generalization. Similarly to transfer learning, the model has been fully trained on the train setQ, and now we test the model on the test setP

  49. [49]

    Whenθis in the neighborhood of the optimumθ ∗,s θ(x, τ−t) can be well approximated as its projection onto the score-field space [2], which we denote by∇logq(x, t)

    Near-optimality. Whenθis in the neighborhood of the optimumθ ∗,s θ(x, τ−t) can be well approximated as its projection onto the score-field space [2], which we denote by∇logq(x, t). Notice that the evolution ofq(x, t) is governed by the same Fokker-Planck equation: ∂q(x, t) ∂t =−∇[µF(x)q(x, t)−µk BT∇q(x, t)].(S15) We substitutes θ(x, τ−t) =∇logq(x, t) into...

  50. [50]

    A. W. Lau and T. C. Lubensky, State-dependent diffusion: Thermodynamic consistency and its path integral formulation, Physical Review E—Statistical, Nonlinear, and Soft Matter Physics76, 011123 (2007)

    2007

  51. [51]

    C.-H. Lai, Y. Takida, N. Murata, T. Uesaka, Y. Mitsufuji, and S. Ermon, Fp-diffusion: Improving score-based diffusion models by enforcing the underlying score fokker-planck equation, inInternational Conference on Machine Learning(PMLR,

  52. [52]

    Csisz´ ar, On information-type measure of difference of probability distributions and indirect observations, Studia Sci

    I. Csisz´ ar, On information-type measure of difference of probability distributions and indirect observations, Studia Sci. Math. Hungar.2, 299 (1967)

    1967

  53. [53]

    J. Ho, A. Jain, and P. Abbeel, Denoising diffusion probabilistic models, Advances in neural information processing systems 33, 6840 (2020)

    2020