A Unified Kullback--Leibler Divergence Analysis of Generative Diffusion Models via Entropy Production Rate

Han Wu; Zhiwen Zhang

REVIEW 3 major objections 5 minor 15 references

A Unified Kullback--Leibler Divergence Analysis of Generative Diffusion Models via Entropy Production Rate

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A single entropy-production identity decomposes the terminal KL error of any diffusion sampler into initialization, score, and discretization contributions, yielding an O(h²) step-size bound for Euler–Maruyama sampling.

desk verdict A genuinely useful entropy-production unification of sampler analyses, but the headline O(h^2) KL claim for the Euler–Maruyama chain is not proven and needs to be restated or repaired. read the letter →

arxiv 2608.02406 v1 pith:EUFTFIKK submitted 2026-08-03 math.NA cs.NAstat.ML

Han Wu , Zhiwen Zhang This is my paper

classification math.NAcs.NAstat.ML MSC 60J6065C3068T0794A17

keywords diffusionmodelsentropyproductionrateKullback–Leiblerdivergencescore-basedgenerativeEuler–Maruyamadiscretizationflowmatchingstochasticinterpolantsbackwarderroranalysis

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

This paper tries to establish that all sources of error in generative diffusion models—wrong initialization, imperfect score learning, and time discretization—can be tracked by one thermodynamic quantity, the entropy production rate of the forward–reverse process pair. If correct, an Euler–Maruyama–discretized score-based sampler produces a distribution whose KL divergence to the target is O(h²) in step size plus a term quadratic in the score error plus an exponentially decaying initialization error. The central mechanism is a closed-form identity expressing the entropy production rate through the current velocities of the two marginal flows, which holds even when the forward and reverse diffusions differ. Since every mainstream sampler is a choice of diffusion coefficient, one inequality covers score SDEs, probability-flow ODEs, flow matching, and stochastic interpolants, locating the optimal noise level and exposing the deterministic-versus-stochastic trade-off.

What carries the argument

The central object is the entropy production rate e_p(t)=−(d/dt) D_KL(ρ_t∥ρ′_{T−t}). The load-bearing identity expresses it in velocity form, e_p(t)=−∫ρ ∇log(ρ/ρ′) · (v_c+v′_c) dx, where v_c and v′_c are the current velocities of the two marginal flows. This identity reduces error accounting to comparing two velocity fields and requires only continuity equations, so it remains valid for state-dependent and unequal diffusion matrices. The authors then pair this identity with backward error analysis: the Euler–Maruyama chain tracks a modified continuous-time flow whose coefficients are O(h)-close to the original, and the marginal-level energy estimate converts that coefficient discrepancy into

What would settle it

Take a one-dimensional OU reverse SDE with the exact score, initialize exactly at the terminal distribution, discretize with Euler–Maruyama at step sizes h covering a range, and measure the KL divergence between the terminal law of the chain and the exact target. If, for a fixed T and non-Gaussian initial data, the KL error fails to scale as O(h²)—for example, if it scales as O(h) or O(h^{3/2})—then the transfer from the modified flow to the discrete chain is invalid. The paper's own Gaussian experiments show O(h²) (slopes near 2), so the decisive test would use a non-Gaussian target with know

Watch

Extended reading notes

Core claim

The paper claims that for a score-based diffusion model with score error ε and Euler–Maruyama step h, the generated distribution satisfies D_KL(X0 || X^{θ*,h,−}_T) ≤ C1(T) h² + C2(T) ε² + e^{−2T/C_LS} D_KL(μ0 || ρ∞). The O(h²) discretization rate, obtained by accumulating entropy production at the marginal level, improves on Girsanov-style O(h) rates. The proof rests on Theorem 3.3: for a pair of continuity-equation flows, the entropy production rate equals −∫ρ ∇log(ρ/ρ′) · (v_c+v′_c) dx, where v_c and v′_c are the current velocities. Integrating this identity decomposes terminal KL into initialization, score, and discretization parts. Early stopping replaces a global Fisher-information assu

Load-bearing premise

The proof of Theorem 4.9 transfers the KL bound from the modified continuous flow to the Euler–Maruyama chain by asserting that weak order O(h²) agreement is sufficient, but KL divergence is not controlled by weak test-function order and no density or total-variation estimate between the chain and the modified flow is provided.

Editorial extensions

If this is right

Euler–Maruyama–discretized score-based samplers have KL discretization error scaling as h², so halving the step size divides that error by four.
Score-based SDEs, probability-flow ODEs, flow matching, and stochastic interpolants all obey one bound controlled by two trainable quantities: current-velocity error and score error.
The optimal reverse noise equals the forward noise for score-parametrized samplers; for interpolants with score-tied velocity it is set by the schedule, so there is a finite optimal stochasticity that balances initialization forgetting, score error, and discretization error.
Early stopping, requiring only finite data second moment, yields finite KL bounds to the mollified target and Wasserstein bounds to the true singular target.
The entropy-production profile concentrates near the data end, giving a quantitative justification for non-uniform step schedules refined near the data end.

Reading between the lines

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

The O(h²) claim for the actual Euler–Maruyama chain rests on transferring the KL bound from the modified continuous flow via weak order O(h²); because KL is not controlled by weak test-function order and no density/total-variation estimate is given, the rigorous statement covers the modified flow. A direct density estimate between the chain and the modified flow would settle this gap.
If the unified inequality is correct, measuring e_p(t) on a deployed sampler could identify where discretization budget is wasted, turning schedule design into an empirical error-localization problem.
The error channels need not be independent: the paper's Experiment 5 shows a variance-type score perturbation can partially cancel the discretization bias, so the additive bound is conservative; quantifying that cancellation could yield sharper practical error estimates.

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

Short version: this paper gives you a nice identity and a useful way to view all the main samplers in one bound. But the abstract's headline claim, that the Euler–Maruyama sampler achieves terminal KL error O(h^2), is not actually derived. Theorem 4.9 bounds the continuous modified flow. The proof's bridge from that flow to the discrete chain is the sentence after (4.21): weak order O(h^2) agreement in test functions implies the same KL error. That implication is false without a density or total-variation estimate, and none is supplied. The stress-test note is right about this. The numerics support the scaling, but they compare the chain to the target, not the chain to the modified flow, so they cannot close the gap. The fix is straightforward: prove a chain-to-modified-flow relative-entropy estimate, or state Theorem 4.9 as a modified-flow bound and mark the discrete sampler claim as open. Given the paper's own honesty about priority, this would be an honest revision.

What is genuinely new and good: Theorem 3.3 is a clean, general entropy-production identity that works for state-dependent and unequal diffusions, and Section 5 is the real payoff. The single inequality in Theorem 5.4, plus Table 2, actually does unify score SDEs, PF-ODEs, flow matching, and stochastic interpolants, and it recovers the ODE/SDE trade-off from one mechanism. The early-stopping downgrade, using a Malliavin-calculus bound to replace global Fisher information by a finite second moment, is a useful practical extension. I also credit Experiment 5: perturbing the analytical score in a Gaussian model and checking the additive h^2 + eps^2 structure is the right kind of controlled test, and the near-perfect slope agreement is real evidence for the decomposition, even if it cannot certify the discrete-sampler claim.

Soft spots beyond the main gap: constants carry unspecified dimension and derivative-bound factors, acknowledged in the paper but still limiting; the log-Sobolev and Lipschitz assumptions on sampler marginals are load-bearing; and the deterministic-sampler bounds are basically Grönwall with Lipschitz growth, which is less novel. None of these are fatal. The authors are candid about what is theirs; the explicit disclaimer of priority is accurate rather than performative.

Who should read this: anyone working on diffusion model theory or sampler comparison will get value from Section 5 and the identity in Theorem 3.3. It deserves a serious referee, but only if the referee insists on the discrete-sampler claim being fixed or restated. My own verdict is conditional, not rejection.

Referee Report

3 major / 5 minor

Summary. The paper develops a unified error analysis for generative diffusion models based on the entropy production rate of a forward/reverse pair of diffusion processes. The central identity (Theorem 3.3) expresses the time-derivative of the marginal KL divergence in terms of current velocities, without requiring equal diffusion matrices. This identity is used to decompose the terminal KL error into initialization, score-approximation, and discretization contributions. The headline result is Theorem 4.9, which claims an O(h^2) terminal KL bound for the Euler–Maruyama sampler, and Section 5 extends the framework to PF-ODE, flow matching, and stochastic interpolants, recovering optimal-noise trade-offs. Numerical experiments on Gaussian and Beta targets report KL slopes close to 2 in the step size. The main technical gap is that Theorem 4.9 is proved for the continuous modified backward process, while the advertised O(h^2) rate for the discrete Euler–Maruyama chain relies on an unproved transfer from weak order to KL divergence.

Significance. The entropy-production-rate perspective is attractive and the identity of Theorem 3.3 is a clean, self-contained contribution. The decomposition into initialization, score, and discretization errors is natural, and the paper explicitly avoids fitted constants; the Gaussian experiments in Section 6 include exact closed-form checks of the entropy-production profile and the endpoint identity, which is a genuine strength. If the discrete-sampler gap were closed, the framework would provide a unified KL-based comparison of score SDEs, PF-ODEs, flow matching, and stochastic interpolants, with quantitative trade-offs. As it stands, the main advertised improvement over Girsanov-based O(h) bounds — namely an O(h^2) KL guarantee for the Euler–Maruyama sampler — is not established by the proof, because the bound in Theorem 4.9 applies to a modified continuous flow and no density/relative-entropy estimate links that flow to the actual chain marginals.

major comments (3)

[§4.4, proof of Theorem 4.9, after Eq. (4.21)] Theorem 4.9 bounds D_KL(X0 || X^{θ*,h,-}_T), where X^{θ*,h,-}_t is the continuous process with first-order modified coefficients, not the Euler–Maruyama chain. The only bridge to the EM iterates is the sentence: 'The Euler–Maruyama iterates themselves agree with the marginals of the modified flow ρ′ to weak order O(h^2) by Theorem 4.1; the theorem bounds the error of the modified flow, which is the continuous object that the scheme tracks at this order.' Weak order for smooth test functions does not control KL divergence, which is not a test function and requires a density, total-variation, or relative-entropy estimate between the chain marginal and the modified-flow marginal. No such estimate is provided. Consequently, the abstract's O(h^2) claim for the Euler–Maruyama sampler is unsupported; the numerical experiments do not fill the gap because they compare the chain to the target, not
[§4.1, Theorem 4.1 and its use in Theorem 4.9] Theorem 4.1 requires the drift f(x,t) to lie in C^k_b(R^n,[0,T]). In the reverse SDE of Theorem 4.9 the drift contains the score term 2A(T−t)s_{θ*}(x,T−t), which is only assumed to be an L^2 approximation of ∇logρ (Assumption 4.6). The proof asserts that b^{-}_{θ*} 'lies in C^k_b because b and s_θ* do', but no C^k_b bound on s_{θ*} is stated or derived. For compactly supported or boundary-singular data, the score can be unbounded even when the Fisher-information integral is finite (see Example 2.6 and Remark 4.5). This means the O(h^2) weak-order result may not apply to the learned reverse drift without additional smoothness assumptions, compounding the gap identified above. The authors should state explicit C^k_b hypotheses on the learned score or restrict the claim to settings where those hypotheses hold.
[§5.2, Proposition 5.2] The unified bounds (5.10)–(5.12) and the subsequent Theorem 5.4 assume that the sampler marginals satisfy a logarithmic Sobolev inequality with constant C_LS. This is not derived from the stated assumptions on the forward process or on the learned drift. The marginal ρ′(·,s) of the modified backward process with a misspecified drift need not inherit the LSI constant of the forward invariant measure ρ∞; the constant may degrade with the score/velocity error. Since the exponential forgetting rate and the optimal-noise conclusions (Corollaries 5.5 and 5.6) depend quantitatively on C_LS, the trade-off statements are conditional on an unverified regularity property. The authors should either prove that the relevant marginals satisfy LSI under their hypotheses, or state this as an explicit structural assumption on the sampler instead of on the target.

minor comments (5)

[§4.5, Corollary 4.12, Eq. (4.27)] The initialization term is written as e^{-2T/C_LS} D_KL(ρ_δ || ρ∞). Since the contraction is applied on the interval [δ,T], the exponent should presumably be -2(T−δ)/C_LS; as written the bound is stronger than what Proposition 4.8 justifies.
[§7] Typos: 'unifrom' in the first paragraph of the conclusions, and 'Kullback–Leibler' is misspelled in one place in Section 7.1.
[§4.1, Eq. (4.2)] The notation 'gg^T : ∇∇gg^T' in the formula for g_1 is ambiguous; specify the index contraction. Also state explicitly that g(t) is time-dependent (not state-dependent) in Theorem 4.1.
[§2.6, Definition 2.9] The definition of KL divergence first refers to probability mass functions and then to densities; the notation P(x) and Q(x) is used for both. This can be clarified.
[§2.3, Proposition 2.4] The proof of the Malliavin bound E|H_t|^2 ≤ C n/(r t) cites standard Meyer inequalities but does not track the constant; given the paper's stated interest in dimension dependence (Section 7.1), a more explicit statement would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's derivation is self-contained; the weak-order-to-KL transfer gap for the discrete sampler is a correctness issue, not a circular step.

full rationale

The central derivation chain is not circular. Theorem 3.3 derives the entropy-production identity directly from the continuity equations and the definition of the KL derivative; no target quantity is substituted back into the hypotheses. The error decomposition (2.25)/(4.1) is an exact identity, and the subsequent bounds in Theorem 4.9 are obtained by estimating the forcing terms through the modified-equation analysis of [1], the Fisher-information assumption, the score-error assumption, and the log-Sobolev initialization decay. The constants C1(T) and C2(T) are explicit and not fitted to data; no parameter is calibrated to the quantity being predicted. The numerical experiments confirm, rather than define, the predicted scaling. The self-citations [52,54,55,56] appear only as background examples of backward error analysis and are not load-bearing; the key modified-equation result, Theorem 4.1, is cited from the external reference [1]. The main weakness is the assertion after (4.21) that weak order O(h^2) agreement between the Euler–Maruyama iterates and the modified flow suffices to transfer the KL bound from the continuous modified process to the discrete chain. Weak convergence for smooth test functions does not control KL divergence, and no density or total-variation estimate is provided. This is a genuine gap in the proof of the abstract's discrete-sampler claim, but it is not a self-definitional, fitted-input, or self-citation circularity: the bound is not equivalent to its assumptions by construction. Under the specified circularity criteria, the honest finding is no circularity, score 0.

Assumptions & free parameters 4 free parameters · 9 assumptions · 0 invented entities

No parameters are fitted to data and no new physical entities are introduced. The central claim rests on smoothness/ellipticity, finite Fisher information (or finite second moment with early stopping), standard time-reversal theorems, modified-equation weak order analysis, and log-Sobolev constants. The free parameters listed are control or oracle quantities, not fitted values.

free parameters (4)

Early-stopping level δ = user-selected (not fitted)
Corollary 4.12 and Lemma 4.11 use δ∈(0,T) to truncate the Fisher integral and mollify singular data; the final Wasserstein gap is O(√δ). It is a control parameter, not fitted to data.
Score-error bound ε² = oracle quantity, not fitted
Assumption 4.6 posits ||s_θ* − ∇logρ||²_{L²ρ([δ,T])} ≤ ε²; the bound is linear in ε². It is assumed, not estimated from a trained network.
Log-Sobolev constant C_LS = 1 for standard Gaussian; otherwise assumed
Proposition 4.8 and Theorem 4.9 use the log-Sobolev constant of the invariant measure; for N(0,I) it is 1, otherwise assumed finite.
Fisher-information bound M_T = assumed finite, or bounded by C n/r (log(1/δ)+T) via Proposition 2.4
Assumption 2.2 bounds the integrated Fisher information and enters the constant C1(T) linearly in Theorem 4.9.

assumptions (9)

domain assumption Assumption 2.1: smooth coefficients and uniform ellipticity of the diffusion matrix
Used throughout to guarantee smooth positive marginals, time reversal, and the Malliavin bounds of Proposition 2.4.
domain assumption Assumption 2.2: finite global Fisher-information integral, or the weakened finite-second-moment assumption with early stopping (Prop 2.4/Cor 4.12)
The energy estimates in Lemma 4.3 and Theorem 4.9 require control of ∫∫|∇logρ|²ρ dx dt; the weakened form is used for singular data.
standard math Anderson time-reversal theorem (Proposition 2.3, citing [3,18,34])
The reverse drift formula b− = −b + 2ρ⁻¹∇·(Aρ) is the backbone of the score-based reverse process.
standard math Modified-equation weak order O(h²) for Euler–Maruyama (Theorem 4.1, citing [1])
Delivers the O(h) coefficient discrepancy α_h and hence the h² source term in the KL bound; the weak-order statement is imported from the literature.
standard math Logarithmic Sobolev inequality for the invariant measure (Prop 4.8, citing [4])
Gives the exponential decay of the initialization error e^{−2T/C_LS}.
domain assumption Assumption 4.6: L²ρ score approximation error ≤ ε²
The error bound is linear in ε²; the paper assumes the trained score satisfies this oracle condition.
domain assumption Log-Sobolev inequality for sampler marginals in Proposition 5.2
The stochastic-sampler bounds use the sampler's marginal log-Sobolev constant; this is assumed, not derived for general samplers.
domain assumption One-sided Lipschitz bound L on the learned reverse velocity for deterministic samplers (Theorem 5.4)
Wasserstein bounds for PF-ODE and flow matching use Grönwall with growth factor e^{LTδ}; L is assumed.
domain assumption Higher-order derivative and dimension bounds on the coefficients are implicit but not tracked (Remark 4.5, Section 7.1)
The constants C1, C2 depend on norms of drift derivatives that can blow up near the data end; explicit dimension and singularity dependence is left as future work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Unified Kullback--Leibler Divergence Analysis of Generative Diffusion Models via Entropy Production Rate." pith.science (2026). https://pith.science/paper/EUFTFIKK

@misc{pith2026260802406,
  author       = {Pith},
  title        = {Pith review of: A Unified Kullback--Leibler Divergence Analysis of Generative Diffusion Models via Entropy Production Rate},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUFTFIKK}},
  note         = {Machine review of arXiv:2608.02406}
}

abstract

We introduce a unified framework for the error analysis of generative models based on the entropy production rate of the forward-reverse diffusion process pair. For a pair of continuity equation flows, the rate admits a closed velocity form identity whose time integral decomposes the terminal Kullback--Leibler (KL) divergence into the sum of an initialization error, a score approximation error, and a time-discretization error. By analyzing the entropy production at the level of marginal distributions, rather than in path space, our framework yields a sharp convergence rate of $\mathcal{O}(h^2)$ for the Euler-Maruyama sampler, where $h$ is the step size. This improves upon the $\mathcal{O}(h)$ rates typically obtained from Girsanov's path-space analyses. Furthermore, our framework unifies the analysis of score-based SDEs, probability-flow ODEs, and stochastic interpolants by varying diffusion coefficients within a single inequality, revealing the trade-off between deterministic and stochastic sampling. Numerical experiments confirm the predicted scaling with step size and terminal time.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 1 canonical work pages

[1]

Abdulle, D

[1]A. Abdulle, D. Cohen, G. Vilmart, and K. C. Zygalakis,High weak order methods for stochastic differential equations based on modified equations, SIAM Journal on Scientific Computing, 34 (2012), pp. A1800–A1823, https://doi.org/10.1137/110846609. [2]M. S. Albergo, N. M. Boffi, and E. V anden-Eijnden,Stochastic interpolants: a unifying framework for flow...

arXiv 2012
[7]

Katsoulakis, Y

[23]M. Katsoulakis, Y. Pantazis, and L. Rey-Bellet,Measuring the irreversibility of numerical schemes for reversible stochastic differential equations, ESAIM: Mathematical Modelling and Numerical Analysis, 47 (2013), pp. 223–255. [24]D. P. Kingma, T. Salimans, B. Poole, and J. Ho,Variational diffusion models, in Advances in Neural Information Processing S...

2013
[9]

[32]C. Lu, K. Zheng, F. Bao, J. Chen, C. Li, and J. Zhu,Fast diffusion probabilistic model sampling through the lens of backward error analysis, arXiv preprint arXiv:2304.11446, (2023). [33]C. Maes and K. Neto ˇcn´y,Time-reversal and entropy, Journal of Statistical Physics, 110 (2003), pp. 269–310. [34]A. Millet, D. Nualart, and M. Sanz,Integration by par...

arXiv 2023
[12]

Premkumar,Generative diffusion from an action principle, arXiv preprint arXiv:2310.04490, (2023)

[41]A. Premkumar,Generative diffusion from an action principle, arXiv preprint arXiv:2310.04490, (2023). [42]A. Premkumar,Neural entropy, arXiv preprint arXiv:2409.03817, (2024). [43]H. Qian,Thermodynamics of the general diffusion process: equilibrium supercurrent and nonequilibrium driven circulation with dissipation, arXiv preprint arXiv:1412.5925, (201...

arXiv 2023
[14]

[47]B. P. Schaeffer, R. M. S. Rosa, and G. V alle,The effect of stochasticity in score-based diffusion sampling: a KL divergence analysis, arXiv preprint arXiv:2506.11378, (2025). [48]U. Seifert,Entropy production along a stochastic trajectory and an integral fluctuation theo- rem, Physical Review Letters, 95 (2005), p. 040602. [49]J. Sohl-Dickstein, E. A...

arXiv 2025
[15]

Sprekeler, H

[52]T. Sprekeler, H. Wu, and Z. Zhang,Numerical approximation of effective diffusivities in homogenization of nondivergence-form equations with large drift by a Lagrangian method, arXiv preprint arXiv:2506.14073, (2025). [53]P. Vincent,A connection between score matching and denoising autoencoders, Neural Com- putation, 23 (2011), pp. 1661–1674. [54]Z. W ...

arXiv 2025
[606]

[8]S. Chen, S. Chewi, H. Lee, Y. Li, J. Lu, and A. Salim,The probability flow ODE is provably fast, Advances in Neural Information Processing Systems (NeurIPS), 36 (2024). [9]S. Chen, S. Chewi, J. Li, Y. Li, A. Salim, and A. R. Zhang,Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions, in International Conf...

2024
[1141]

[45]R. Raqu ´epas,The large-time and vanishing-noise limits for entropy production in nondegen- DIFFUSION MODELS ANALYSIS VIA ENTROPY PRODUCTION37 erate diffusions, Annales de l’Institut Henri Poincar´ e (B) Probabilit´ es et Statistiques, 60 (2024), pp. 431–462. [46]T. Salimans and J. Ho,Progressive distillation for fast sampling of diffusion models, in ...

2024

Show all 15 references

[1967]

[37]A. Q. Nichol and P. Dhariwal,Improved denoising diffusion probabilistic models, in Inter- national Conference on Machine Learning (ICML), 2021, pp. 8162–8171. [38]D. Nualart,The Malliavin Calculus and Related Topics, Probability and its Applications, Springer-Verlag, Berli...

2021
[2006]

P ´erez-Cruz,Kullback–Leibler divergence estimation of continuous distributions, in 2008 IEEE International Symposium on Information Theory, IEEE, 2008, pp

[39]F. P ´erez-Cruz,Kullback–Leibler divergence estimation of continuous distributions, in 2008 IEEE International Symposium on Information Theory, IEEE, 2008, pp. 1666–1670. [40]E. Pierret and B. Galerne,Diffusion models for Gaussian distributions: exact solutions and Wassers...

2008
[2021]

Kusuoka and D

[25]S. Kusuoka and D. Stroock,Applications of the Malliavin calculus, Part II, Journal of the Faculty of Science, the University of Tokyo, Section IA, Mathematics, 32 (1985), pp. 1–76. [26]H. Lee, J. Lu, and Y. Tan,Convergence for score-based generative modeling with polynomia...

1985 arXiv
[2022]

Debussche and E

[14]A. Debussche and E. F aou,Weak backward error analysis for SDEs, SIAM Journal on Nu- merical Analysis, 50 (2012), pp. 1735–1752, https://doi.org/10.1137/110831544. [15]H. F ¨ollmer,Time reversal on Wiener space, in Stochastic Processes — Mathematics and Physics, vol. 1158 ...

2012 doi
[2023]

Cheng, J

[10]X. Cheng, J. Lu, Y. Tan, and Y. Xie,Convergence of flow-based generative models via proximal gradient descent in Wasserstein space, arXiv preprint arXiv:2310.17582, (2024). [11]G. Conforti, A. Durmus, and M. Gentiloni Silveri,KL convergence guarantees for score diffusion m...

2024 arXiv
[2024]

Cattiaux, G

[6]P. Cattiaux, G. Conforti, I. Gentil, and C. L ´eonard,Time reversal of diffusion processes under a finite entropy condition, arXiv preprint arXiv:2104.07708, (2022). [7]P.-H. Chavanis,On the connection between Nelson ’s stochastic quantum mechanics and Not- tale’s theory of...

2022 arXiv
[2025]

Gao and L

[17]X. Gao and L. Zhu,Convergence analysis for general probability flow ODEs of diffusion models in Wasserstein distances, arXiv preprint arXiv:2401.17958, (2023). [18]U. G. Haussmann and E. Pardoux,Time reversal of diffusions, The Annals of Probability, 14 (1986), pp. 1188–12...

2023 arXiv

Pith tools

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

[1] [1]

Abdulle, D

[1]A. Abdulle, D. Cohen, G. Vilmart, and K. C. Zygalakis,High weak order methods for stochastic differential equations based on modified equations, SIAM Journal on Scientific Computing, 34 (2012), pp. A1800–A1823, https://doi.org/10.1137/110846609. [2]M. S. Albergo, N. M. Boffi, and E. V anden-Eijnden,Stochastic interpolants: a unifying framework for flow...

arXiv 2012

[2] [7]

Katsoulakis, Y

[23]M. Katsoulakis, Y. Pantazis, and L. Rey-Bellet,Measuring the irreversibility of numerical schemes for reversible stochastic differential equations, ESAIM: Mathematical Modelling and Numerical Analysis, 47 (2013), pp. 223–255. [24]D. P. Kingma, T. Salimans, B. Poole, and J. Ho,Variational diffusion models, in Advances in Neural Information Processing S...

2013

[3] [9]

[32]C. Lu, K. Zheng, F. Bao, J. Chen, C. Li, and J. Zhu,Fast diffusion probabilistic model sampling through the lens of backward error analysis, arXiv preprint arXiv:2304.11446, (2023). [33]C. Maes and K. Neto ˇcn´y,Time-reversal and entropy, Journal of Statistical Physics, 110 (2003), pp. 269–310. [34]A. Millet, D. Nualart, and M. Sanz,Integration by par...

arXiv 2023

[4] [12]

Premkumar,Generative diffusion from an action principle, arXiv preprint arXiv:2310.04490, (2023)

[41]A. Premkumar,Generative diffusion from an action principle, arXiv preprint arXiv:2310.04490, (2023). [42]A. Premkumar,Neural entropy, arXiv preprint arXiv:2409.03817, (2024). [43]H. Qian,Thermodynamics of the general diffusion process: equilibrium supercurrent and nonequilibrium driven circulation with dissipation, arXiv preprint arXiv:1412.5925, (201...

arXiv 2023

[5] [14]

[47]B. P. Schaeffer, R. M. S. Rosa, and G. V alle,The effect of stochasticity in score-based diffusion sampling: a KL divergence analysis, arXiv preprint arXiv:2506.11378, (2025). [48]U. Seifert,Entropy production along a stochastic trajectory and an integral fluctuation theo- rem, Physical Review Letters, 95 (2005), p. 040602. [49]J. Sohl-Dickstein, E. A...

arXiv 2025

[6] [15]

Sprekeler, H

[52]T. Sprekeler, H. Wu, and Z. Zhang,Numerical approximation of effective diffusivities in homogenization of nondivergence-form equations with large drift by a Lagrangian method, arXiv preprint arXiv:2506.14073, (2025). [53]P. Vincent,A connection between score matching and denoising autoencoders, Neural Com- putation, 23 (2011), pp. 1661–1674. [54]Z. W ...

arXiv 2025

[7] [606]

[8]S. Chen, S. Chewi, H. Lee, Y. Li, J. Lu, and A. Salim,The probability flow ODE is provably fast, Advances in Neural Information Processing Systems (NeurIPS), 36 (2024). [9]S. Chen, S. Chewi, J. Li, Y. Li, A. Salim, and A. R. Zhang,Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions, in International Conf...

2024

[8] [1141]

[45]R. Raqu ´epas,The large-time and vanishing-noise limits for entropy production in nondegen- DIFFUSION MODELS ANALYSIS VIA ENTROPY PRODUCTION37 erate diffusions, Annales de l’Institut Henri Poincar´ e (B) Probabilit´ es et Statistiques, 60 (2024), pp. 431–462. [46]T. Salimans and J. Ho,Progressive distillation for fast sampling of diffusion models, in ...

2024