acceptodds
Under review as a conference paper at ICLR 2027

FIRST-ORDER STATIONARITY OF REVERSE DIFFUSIONS

Abstract

Recent literature has shown a strong connection between optimization and sampling. We develop the corresponding first-order theory for diffusion models. First, the SDE-based reverse-time flows of overdamped and underdamped Langevin diffusions contract relative Fisher divergences at explicit exponential rates whenever the stationary potential of the forward process is strongly convex—a condition on the noising process one chooses, not on the data. This is a unique advantage of SDE-based reverse diffusion, absent in the reverse process based on ODEs. Second, we incorporate discretization and establish averaged first-order stationarity bounds—the sampling analog of averaged gradient-norm guarantees in nonconvex optimization—for samplers of both overdamped and underdamped diffusion models. As in nonconvex optimization, the convexity-free certificate is local: it guarantees score consistency, not global mode weights.

Primary area:
generative models
Keywords:
diffusion models, score-based generative models, sampling, convergence theory, underdamped Langevin dynamics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_59k4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the contraction of Fisher divergence for both overdamped and underdamped diffusions. Under the assumption of exact score functions and regularity conditions on the target distribution, such as score Lipschitzness and sub-Gaussian tails, the authors establish exponential contraction of Fisher divergence. Experiments on Gaussian mixtures illustrate this contraction.

Critical strengths
  • The paper provides a useful perspective on understanding the convergence of diffusion-based samplers through Fisher divergence. Its comparison between probability-flow ODEs and reverse-time SDEs helps clarify their different behaviors under initialization mismatch.
Critical weaknesses
  • The practical implications of Fisher divergence contraction for diffusion-based samplers remain unclear. Even in the Gaussian-mixture experiments, the contraction of Fisher divergence does not necessarily imply accurate recovery of the target distribution. Moreover, the theoretical and experimental results mainly focus on initialization error, which is typically small in practical diffusion models.
  • The theoretical setting is restrictive and lacks sufficient justification. The target distribution is assumed to have a Lipschitz score and sub-Gaussian tails, which seem stronger than the assumptions required in some existing KL/TV analyses[1,2] of diffusion models. The paper does not sufficiently discuss the practical relevance of these assumptions. Moreover, assuming exact scores or score errors controlled in L2(p^tk)L_2(\hat{p}_{t_k}) also appears difficult to justify in practice.

[1] Chen H, Lee H, Lu J. Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions [C]//International Conference on Machine Learning. PMLR, 2023: 4735-4763.

[2] Benton J, De Bortoli V, Doucet A, et al. Nearly $ d $-linear convergence bounds for diffusion models via stochastic localization[C]//International Conference on Learning Representations. 2024, 2024: 36916-36936.

Questions for authors
  • Could the authors provide more discussion on the practical implications of Fisher divergence contraction? In particular, what advantages does Fisher divergence analysis offer over existing KL/TV analyses?
  • Why do the experiments choose p0p_0 far from q0=ρTq_0=\rho_T? In practical diffusion models, the initial distribution for sampling is usually chosen to be close to the terminal distribution of the forward process.
4 comments

We thank the reviewer for the careful assessment and constructive suggestions. We address each point below (Part 1 of 2).

Weakness 1(a) and Question 1: Practical meaning of Fisher contraction and comparison with KL/TV

We agree that Fisher contraction does not imply accurate recovery of the target distribution. This limitation is a central message of the paper, not a claim we seek to avoid. The relative Fisher divergence measures the L2(p)L^2(p) discrepancy between the scores of pp and qq; it is a first-order, local certificate. It can be small even when probability mass is incorrectly allocated among well-separated modes. Accordingly, Fisher analysis complements rather than replaces KL/TV guarantees.

Its distinct contribution is to describe how initialization-induced score mismatch, or errors induced by discretization or score estimation as explained below, evolves under the reverse dynamics. The exact reverse SDE satisfies an entropy dissipation identity without a functional inequality on the data and, under strong convexity of the chosen forward potential, an exponential Fisher contraction estimate. General KL/TV data-processing inequalities alone do not provide this Fisher-level contraction rate.

In the Gaussian-mixture experiment, the reported terminal KL and TV remain around 0.400.40 and 0.410.41, respectively, while the terminal relative Fisher divergence is only about 4.5×10−34.5 \times 10^{-3}, down from an initial value of 100100. This demonstrates the gap between local stationarity and global fidelity. The CelebA-HQ experiment provides a high-dimensional illustration: visually plausible images can coexist with severe brightness-selection bias. Thus, the practical benefit of the Fisher viewpoint is a more precise characterization of what the stochastic dynamics can repair, and what they cannot certify.

Weakness 1(b) and Question 2: Why study nonstandard initializations?

We agree that standard generation typically starts from a tractable Gaussian chosen to approximate the terminal forward marginal, and that making this approximation accurate is important. The deliberately perturbed initializations in our experiments are stress tests designed to reveal the stability mechanism; they are not presented as representative of the default initialization used in every deployed model.

In the CelebA-HQ experiment, the mean shift is μ=±4⋅1/d\mu = \pm 4 \cdot \mathbf{1}/\sqrt{d}. Each coordinate is shifted by only about 0.0090.009 standard deviations for d≈2×105d \approx 2 \times 10^5. Its purpose is to expose the sensitivity of the sampling bias. The same phenomenon is known in deployed models (Lin et al., 2024) even with standard normal initialization. Our theory explains why the reverse SDE does not repair it.

Initialization mismatch also provides a controlled model for perturbation propagation. If the exact reverse dynamics is resumed from a perturbed law at an intermediate time, the dissipation and contraction identities apply to the remaining exact flow. This observation motivates studying perturbations caused by numerical or score errors, even though it does not automatically establish a quantitative bound for their continuously accumulated effects. Guided image-editing methods such as SDEdit (Meng et al., 2022) also intentionally initialize from laws different from the unconditional terminal marginal, making this stability question relevant beyond standard unconditional sampling.

References

Lin, S., Liu, B., Li, J., and Yang, X. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 5404-5411, 2024. https://arxiv.org/abs/2305.08891

Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.-Y., and Ermon, S. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2022. https://arxiv.org/abs/2108.01073

(Part 2 of 2, continued from our previous comment.)

Weakness 2(a): Strength of score regularity and tail assumptions

We have refined our proof techniques to remove the sub-Gaussian tail assumption, but the Lipschitz score assumption is still needed, and we agree that the discretized theorems have stronger data-side assumptions than several existing KL/TV analyses.

The reason is partly the quantity being controlled. Theorem 4.5 of the original submission (now Theorem 4.3 in the revision) bounds F(p^t∥qt)\mathcal{F}(\hat p_t \Vert q_t), an expectation under the numerical law. In contrast, many KL analyses of diffusion models use a direction of relative entropy and a path-space argument that place error expectations under the exact process and can exploit the OU structure under weaker score regularity. These are different targets and proof methods.

We will state that the current assumptions are sufficient rather than sharp, explain their exact role in Remark 4.5, and identify removing or weakening them as an open problem.

Weakness 2(b): Exact scores and learned-score error under the sampler law

This is a genuine limitation, and we agree it deserves greater prominence. The paper proves its principal results for exact scores. Remark 4.7 explains how an additional mean-square score error term can be incorporated when the error is controlled under the sampler law. However, the usual denoising-score-matching guarantee is under the exact noised-data marginal, and transferring that guarantee to the numerical law requires extra assumptions. An L2(q)L^2(q) bound alone cannot exclude large score errors in regions that the numerical sampler overweights.

We will clarify this distinction and describe possible sufficient conditions for the transfer (for example, bounded error, or suitable exponential-moment bounds). The pretrained-DDPM experiment is useful empirical evidence that initialization-dependent bias can persist with learned scores. A complete theorem simultaneously treating realistic score estimation, discretization, and Fisher stationarity remains future work.

Thank you for the detailed clarifications.

While I understand that Fisher contraction does not necessarily imply distributional recovery, the practical significance of the theoretical results remains unclear to me. As shown in the Gaussian-mixture experiment, Fisher divergence becomes very small while KL and TV remain substantial. In practice, diffusion models are primarily evaluated by the quality and fidelity of generated samples, which Fisher contraction alone does not guarantee.

Could the authors further clarify what meaningful insights or guarantees these theoretical results provide for practical diffusion sampling, particularly with learned scores and discretization errors?

Thank you for the thoughtful follow-up. We agree that the practical significance of Fisher contraction deserves clarification. We provide the following additional argument connecting Fisher divergence to within-mode distributional accuracy. These findings clarify what Fisher contraction guarantees even when global KL and TV remain substantial.

1. Quantitative within-mode accuracy from Fisher divergence

Our results distinguish within-mode fidelity from global mode coverage. Although small Fisher divergence does not imply global distributional recovery, it can control within-mode distributional discrepancies under suitable local regularity conditions.

Partition the sample space into regions A1,…,AKA_1,\ldots,A_K, representing modes or groups of similar samples. Let wip=p(Ai)w_i^p = p(A_i), wiq=q(Ai)w_i^q = q(A_i), and pi=p(⋅∣Ai)p_i = p(\cdot \mid A_i), qi=q(⋅∣Ai)q_i = q(\cdot \mid A_i). The KL chain rule yields

KL(p∥q)=KL(wp∥wq)⏟mode-weight error+∑iwipKL(pi∥qi)⏟within-mode error.\mathrm{KL}(p \Vert q) = \underset{\text{mode-weight error}}{\underbrace{\mathrm{KL}(w^p \Vert w^q)}} + \underset{\text{within-mode error}}{\underbrace{\sum_i w_i^p \mathrm{KL}(p_i \Vert q_i)}}.

Suppose each conditional target qiq_i satisfies a log-Sobolev inequality on AiA_i with constant CiC_i, namely KL(pi∥qi)≤(Ci/2)F(pi∥qi)\mathrm{KL}(p_i \Vert q_i) \le (C_i/2) \mathcal{F}(p_i \Vert q_i). Because pi/qip_i/q_i is a constant multiple of p/qp/q within each region, the relative scores agree there and ∑iwipF(pi∥qi)=F(p∥q)\sum_i w_i^p \mathcal{F}(p_i \Vert q_i) = \mathcal{F}(p \Vert q). Thus, writing Cloc=max⁡iCiC_{\mathrm{loc}} = \max_i C_i,

∑iwipKL(pi∥qi)≤Cloc2F(p∥q).\boxed{\sum_i w_i^p \mathrm{KL}(p_i \Vert q_i) \le \frac{C_{\mathrm{loc}}}{2} \mathcal{F}(p \Vert q).}

Pinsker's inequality and Cauchy–Schwarz also give

∑iwipTV(pi,qi)≤12ClocF(p∥q).\boxed{\sum_i w_i^p \mathrm{TV}(p_i, q_i) \le \frac{1}{2} \sqrt{C_{\mathrm{loc}} \mathcal{F}(p \Vert q)}.}

Hence Fisher divergence controls within-mode KL and TV when local log-Sobolev constants are moderate, even if the global distribution remains far from the target. Such local constants remain favorable for regions with a single mode (for an approximately Gaussian mode, the constant is of the order of the mode variance).

2. Concrete Gaussian-mixture evidence

In the Gaussian-mixture experiment, we can partition the real line at the midpoints between adjacent mode centers and divide it into 3 regions for 3 modes. Although terminal global KL and TV remain substantial (0.401 and 0.405), the weighted sum of conditional within-mode KL divergences is only approximately 0.005. Specifically,

0.401⏟global KL=0.396⏟mode-weight KL+0.005⏟weighted within-mode KL.\boxed{\underset{\text{global KL}}{\underbrace{0.401}} = \underset{\text{mode-weight KL}}{\underbrace{0.396}} + \underset{\text{weighted within-mode KL}}{\underbrace{0.005}}.}

Thus, approximately 99% of the residual global KL arises from incorrect mode weights, not from within-mode discrepancies. The small Fisher divergence therefore corresponds to substantial within-mode distributional accuracy even though global recovery is incomplete. Analogously, an unconditional image generator could produce convincing cats, dogs, and fish while sampling these categories in incorrect proportions.

3. Practical interpretation with learned scores and discretization

Our CelebA-HQ experiment (Appendix F.4) uses a pretrained DDPM with learned scores and discretization. An initialization shift of approximately 0.009 noise standard deviations per pixel yields visually realistic faces but noticeably biases the brightness distribution. This illustrates that plausible individual samples can conceal substantial global distributional bias.

It is empirically known that sample quality is decided mostly in the late, low-noise stages of generation, which is also where most samplers concentrate their steps. If a model is poorly trained or poorly scheduled at early stages, or the forward process is truncated too early, the resulting error acts as an initialization error for the rest of the process. Our theoretical results show that stochastic correction makes the generated samples locally indistinguishable from real data, even when this error is very large, and in doing so it can hide a severe bias in the sample proportions. This perspective complements global KL/TV convergence theory. It motivates evaluating both within-mode fidelity and global mode coverage when assessing diffusion samplers.

In summary, Fisher contraction can yield meaningful within-mode KL/TV accuracy, even when large errors are incurred at early generation stages. Together with our discretization and learned-score analysis, this offers a quantitative explanation of how local fidelity can coexist with global distributional bias.

Reviewer_A5KEOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper treats reverse diffusion as an analogy of optimization, where the relative Fisher divergence plays the role of the squared gradient norm. Along the exact reverse SDE, the Fisher divergence contracts exponentially whenever the forward potential is strongly convex, for both overdamped and underdamped dynamics. In comparison, the probability-flow ODE does not forget initialization error. For discretized samplers, the authors prove time-averaged Fisher bounds.

Critical strengths
  1. It provides a clear view of the first-order theoretical guarantee for diffusion samplers, including the overdamped and underdamped settings.

  2. For exact reverse flows, it proves contraction along the trajectory. For discrete samplers, it provides upper bounds for the averaged Fisher information.

  3. The paper is well written and organized.

Critical weaknesses
  1. I hold concerns about the problem's motivation. Although I understand the correspondence between optimization and reverse sampling, it remains unclear to me why such averaged Fisher bounds are meaningful. In my understanding, the averaged guarantee is meaningful for nonconvex optimization because it’s impossible to derive a last-iterate guarantee within that setting. This looks even stranger when Section 3 relies on the strong convexity condition (Assumption 3.1).

  2. Continuing on the last point, Section 2.1 defines the objects of the dictionary. However, it’s unclear why they are relevant to the main topic, especially Lemma 2.1. I do not see any reason to introduce it here, and I find it unrelated to the main results. The same issues hold for Section 3. I do not treat it as an important discovery despite it can be viewed as a analogy for nonconvex optimization.

  3. The results in Section 4-5 are not very satisfactory. First, score error has not been considered. Although a discussion is written in Remark 4.7, a score estimation error assumption with respect to the exact marginal is very essential for a diffusion theory. Thus, the score error has not been well addressed. I would recommend naming it as (non)logconcave sampling rather than diffusion for a theory that cannot deal with estimation error.

    In Section 5, the difference from the exact distribution is not addressed either. Thus, only the initialization error is considered. Remark 5.6 has not addressed this well, either.

  4. It misses several related work. And a comprehensive discussion should be provided.

    About Fisher contraction:

    (1) Fisher information dissipation for time-inhomogeneous stochastic differential equations. Feng et al. 2024

    (2) Almost sure contraction for diffusions on RdR^d. Application to generalised Langevin diffusions Monmarché 2020

    Error Contraction property of SDE

    (1) Restart Sampling for Improving Generative Processes Xu et al. 2023

    (2) Exploring the Optimal Choice for Generative Processes in Diffusion Models: Ordinary vs Stochastic Differential Equations. Cao et al. 2023

    Convergence theory of diffusion models

    (1) Improved Analysis of Score-based Generative Modeling: User-Friendly Bounds under Minimal Smoothness Assumptions. Chen et al. 2023

    (2) Nearly d-Linear Convergence Bounds for Diffusion Models via Stochastic Localization. Benton et al. 2024

    (3) Unified convergence analysis for score-based diffusion models with deterministic samplers Li et al. 2025

    (4)Wasserstein Convergence Guarantees for a General Class of Score-Based Generative Models. Gao et al. 2025

Questions for authors

Why does studying the averaged Fisher information guarantee in the reverse diffusion process matter for the real-life experiments?

3 comments

We thank the reviewer for the careful assessment and constructive suggestions. We address each point below (Part 1 of 2).

Weakness 1: Motivation for averaged Fisher bounds and apparent conflict with strong convexity

The key distinction is between the forward noising potential UU and the target law qtq_t. Assumption 3.1 concerns UU, which is chosen by the model designer (for the standard OU process, U(x)=∥x∥2/2U(x) = \Vert x \Vert^2/2). It does not assert that the data distribution, or the reverse-time marginal qtq_t, is log-concave. For multimodal data, the functional p↦KL(p∥qt)p \mapsto \mathrm{KL}(p \Vert q_t) need not have any convexity properties. Thus, the first-order perspective is meaningful precisely in settings where a global functional inequality for the data is unavailable.

For exact reverse SDEs, the stronger conclusion is indeed available: exponential contraction of the Fisher divergence under convexity of UU. For discretized samplers, the numerical score-freezing error prevents the present argument from yielding the same pointwise conclusion, and we obtain averaged stationarity instead. The two results are complementary rather than contradictory. Averaged Fisher control is meaningful as a certificate of local score agreement along the sampling trajectory.

Weakness 2: Relevance of Section 2.1, Lemma 2.1, and Section 3

We appreciate the request to clarify the logical structure. Lemma 2.1 is the starting point of the first-order analysis, not an unrelated analogy. For two evolving densities ptp_t and qtq_t satisfying continuity equations with velocities vtpv_t^p and vtqv_t^q, it expresses the derivative of KL(pt∥qt)\mathrm{KL}(p_t \Vert q_t) through their relative velocity. The same relative-velocity viewpoint explains why the probability-flow ODE preserves KL and other ff-divergences under mismatched initialization (Proposition 2.3 of the original submission, now Proposition 2.4). We have revised Section 2.1 to make this logic clearer.

Section 3 then strengthens dissipation to exponential contraction of the Fisher divergence. In optimization, exponential decay of the gradient norm requires a curvature condition on the objective (strong convexity, or at least a PL inequality), and the objective is given, not chosen. Its sampling counterpart is strong log-concavity, or a log-Sobolev inequality, of the target qtq_t, that is, of the data. Section 3 shows that the reverse SDE needs neither: the Hessian in (3.1) is that of the forward potential UU, so the relative score contracts at a rate fixed by the noising process alone, for arbitrary multimodal data. This is where the correspondence with optimization stops being an analogy: the "objective" is non-convex, yet the "gradient" contracts as if it were strongly convex. Together with Proposition 2.4, it separates the reverse SDE, which forgets earlier errors at an explicit rate, from the probability-flow ODE, which conserves them.

(Part 2 of 2, continued from our previous comment.)

Weakness 3(a): Score-estimation error measured under the exact marginals

We agree that learned-score error is essential for a comprehensive diffusion-model theory. Remark 4.7 gives a conditional extension: if the mean-square score error is controlled under the numerical sampler law p^tk\hat p_{t_k}, an additional score-error contribution enters the overdamped bound. This differs from the standard training assumption, which measures the score error under the exact marginal qtkq_{t_k}.

The measure mismatch is substantive: a small L2(qtk)L^2(q_{t_k}) error need not imply a small L2(p^tk)L^2(\hat p_{t_k}) error if the sampler visits regions assigned little probability by qtkq_{t_k}. We will elevate this limitation in the discussion.

Weakness 3(b): The kinetic theorem uses a twin chain rather than the exact reverse marginal

This distinction is correct. Theorem 5.5 bounds the full phase-space Fisher divergence F(p^i,τ∥q^i,τ)\mathcal{F}(\hat p_{i,\tau} \Vert \hat q_{i,\tau}), where both laws are propagated by the same numerical kernels but from different initializations. The difference between q^\hat q and the exact reverse law qq is a separate discretization component; Remark 5.6 discusses a path-space KL comparison based on Girsanov analysis. That KL comparison does not automatically yield a Fisher bound relative to qq. We also stress this in the discussion.

Weakness 4: Missing related literature

Thank you for identifying these references. We have expanded the related-work section to cover all the works you mentioned and to provide a more comprehensive discussion.

Question: Why is averaged Fisher stationarity relevant to practical experiments?

The main value is diagnostic: it separates local score consistency from global distributional fidelity. The CelebA-HQ experiment illustrates the phenomenon in a high-dimensional generative setting: outputs remain visually plausible under different initializations, but the distribution of generated brightness changes markedly. This reveals a potential caveat of diffusion models that people often overlook, as is empirically observed by Lin et al. (2024) for a deployed diffusion model with standard initialization. Our experiments show why a first-order certificate must be interpreted alongside global metrics.

The results also distinguish error propagation by stochastic reverse dynamics from deterministic probability flow. This provides a mathematical lens for understanding why additional stochastic correction can reduce certain perturbations.

Reference

Lin, S., Liu, B., Li, J., and Yang, X. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 5404-5411, 2024. https://arxiv.org/abs/2305.08891

Thank you for your detailed response. I will keep my score.

Reviewer_iwy4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper develops a first-order theory for reverse diffusion sampling, with relative Fisher divergence as the squared gradient norm and KL divergence as the objective. It shows that the reverse SDEs of the overdamped and underdamped Langevin diffusions contract relative Fisher divergence exponentially when the stationary potential of the forward process is strongly convex, and gives averaged stationarity bounds for the discretized samplers. Experiments on a one-dimensional Gaussian mixture and on CelebA-HQ show the divergence decaying as predicted while the mass on well-separated modes remains wrong.

Critical strengths

As far as I know, the underdamped half is new. It contracts a twisted Fisher divergence along the kinetic reverse flow (Theorem 3.4) and gives an averaged bound for the discretized kinetic sampler, stated against a twin chain, by a metric that moves inside each step (Theorem 5.5). The experiments confirm the predicted rates and the O(h)O(h) gap.

Critical weaknesses
  1. Theorems 4.5 and 5.5 add data-side assumptions, absent from Theorems 3.2 and 3.4, in the form of a uniform score Lipschitz constant LsL_s, sub-Gaussian data, and for Theorem 5.5 the constants LmapL_{\mathrm{map}} and LrefL_{\mathrm{ref}}. For the exact score the paper itself uses in Appendix F.3, the Lipschitz constant of sts_t at noise level ss grows like 1/(2s)1/(2s) or faster, so at smin⁡=10−5s_{\min}=10^{-5} the constant required by Assumption 4.1 is at least 5×1045\times10^{4}, and TLs2h≲cTL_s^2h\lesssim c needs h≲10−10ch\lesssim10^{-10}c on the uniform grid, while the CelebA run uses 150 geometric steps. The one-dimensional mixture, where Ls≈20L_s\approx20 is reported, is the only case in which the assumption is verified.

  2. The bounds of Theorems 4.5 and 5.5 control the average of FF over [0,T][0,T] rather than F(p^T∥qT)F(\hat p_T\|q_T) itself, and at t=Tt=T the only statement seems to be the relative entropy bound H0+dH_0+d of Theorem 4.5. Since what one cares about is the sample distribution finally returned, the certified quantity is not the one that determines it.

Questions for authors
  1. Can Theorem 4.5 be restated with Ls(s)L_s(s) and smin⁡s_{\min} explicit, and does the CelebA setting remain inside the resulting condition? What would it assert for a score like that of Appendix F.3 at smin⁡=10−5s_{\min}=10^{-5}?

  2. Is there any bound on F(p^T∥qT)F(\hat p_T\|q_T) itself, rather than on its average over [0,T][0,T]?

2 comments

We thank the reviewer for the careful assessment and constructive suggestions. We address each concern below.

Concern 1 and Question 1: Uniform score Lipschitzness, small noise levels, and the CelebA-HQ discretization

We have refined our proof techniques and revised the proof. As a result, the condition TLs2h<cT L_s^2 h \lt c and the sub-Gaussian tail assumption are no longer needed. The step-size threshold is now only h≤c/(1+Ls)h \le c/(1+L_s), so the condition has been relaxed.

Assumption 4.1 imposes a uniform spatial Lipschitz constant LsL_s over the entire reverse-time interval. For an empirical distribution convolved with Gaussian noise, the score can become increasingly steep as the forward noise level approaches zero; consequently, a uniform constant can be large near the endpoint. Early stopping at a positive noise level is a common way to avoid the singular endpoint, at the cost of a separate terminal approximation error. We agree that this is a limitation of the current theorem from a practical standpoint. We keep it in the main text for conceptual conciseness and have added a more practically relevant extension in the appendix.

The extension uses a nonuniform grid 0=t0<t1<⋯<tN≤T0 = t_0 \lt t_1 \lt \cdots \lt t_N \le T with hk=tk+1−tkh_k = t_{k+1} - t_k (so early stopping, tN<Tt_N \lt T, is covered), and local-in-time bounds LkL_k on the spatial Lipschitz constants over each interval. The freezing-error estimates can then be organized as sums of step-dependent contributions rather than replacing every LkL_k and hkh_k by their worst-case values. This permits finer steps in the low-noise region. We have added this version of Theorem 4.5 of the original submission (now Theorem 4.3 in the revision), together with its proof under a time-dependent Lipschitz score assumption, in Appendix C (Theorem C.7; see also Remark 4.6).

For the empirical score considered in Appendix F.3, the noise level smin⁡=10−5s_{\min} = 10^{-5} indeed produces a large Lipschitz constant. For such a score, the new extension provides the corresponding guarantee, requiring only that the step size near the endpoint is small enough. That is also why a diminishing step size is often used in practice. The conditions of Theorem 4.5 (now Theorem 4.3) are sufficient, not necessary, and we did not verify them for the CelebA-HQ experiment on real data.

Regarding the other assumptions, the distinction between the results is important. The exact reverse-flow contraction theorems (Theorems 3.2 and 3.4) impose convexity and smoothness on the chosen forward potential UU, not a uniform Lipschitz assumption on the data scores. The additional data-side assumptions arise in the discretization analysis, where the frozen score must be compared with the evolving score along a numerical step.

Concern 2 and Question 2: Time averages versus the returned terminal distribution

We agree that a last-iterate bound for F(p^T∥qT)\mathcal{F}(\hat p_T \Vert q_T) would be stronger and more directly connected to the returned samples. Theorem 4.5 of the original submission (now Theorem 4.3 in the revision) and Theorem 5.5 prove time-averaged bounds, not such a terminal bound; we do not have a general decaying terminal Fisher bound for these discretized schemes under the stated assumptions.

The averaged result nevertheless has a precise first-order interpretation: if a time is drawn uniformly from [0,T][0,T], the expected relative Fisher divergence at that time equals the time average. In particular, the bound guarantees the existence of at least one time with Fisher divergence no larger than the average. It does not guarantee that the final time is such a time. This is analogous to averaged stationarity guarantees in smooth nonconvex optimization, not a substitute for terminal KL/TV convergence theory.

By contrast, the exact overdamped reverse SDE satisfies the last-time estimate F(pT∥qT)≤e−2mTF(p0∥q0)\mathcal{F}(p_T \Vert q_T) \le e^{-2mT} \mathcal{F}(p_0 \Vert q_0) under Theorem 3.2; the underdamped case has the corresponding modified-Fisher contraction in Theorem 3.4. Extending these pointwise-in-time conclusions to discretized samplers would require additional control of the evolution of the Fisher divergence across numerical steps. We will state this limitation explicitly and identify last-iterate discrete Fisher guarantees as an open direction.

The revision resolves my main objection: Assumption 4.3 of the original submission (sub-Gaussian data) and the constant MM are gone, the second condition of (4.4) of the original submission is removed, and for the score of Appendix F.3 the requirement drops from h≲10−10ch\lesssim10^{-10}c to h≲c/(1+Ls)≈2×10−5ch\lesssim c/(1+L_s)\approx2\times10^{-5}c. Going further, Theorem C.7 of the revision, with its non-uniform grids, step-wise constants and early stopping, is the right form of an answer to my first question.

I keep my second weakness in a weaker form. The rebuttal is right that averaged stationarity is standard in smooth nonconvex optimisation. But a sampler returns the law at t=Tt=T, where the first-order quantity is controlled only on average; a bound on KL would not close the gap, since a log-Sobolev inequality only bounds FF from below. The paper certifies a time-averaged property of the marginals, not a property of the law that is returned, and the revised text does not say so.

I raise my rating from 2 to 3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.