acceptodds
Withdrawn from ICLR 2027

Plug-and-Play is a Reverse SDE: A Unified Convergence Theory for Annealed Denoising Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Plug-and-Play (PnP) image reconstruction replaces the proximal step of an inverse-problem optimizer by an off-the-shelf denoiser, and has become the de facto standard when the prior is a deep denoiser or a diffusion score. Yet its convergence theory has lagged practice: classical fixed-point analyses require the denoiser to be non-expansive or Lipschitz, assumptions that modern annealed priors violate by construction. We show that the PnP iteration is, under a Gaussian-denoising (Tweedie) assumption, the Euler–Maruyama discretization of a time-inhomogeneous It\^o SDE whose drift is the sum of the data-fidelity step and the score of the noise-perturbed prior. When the denoiser is the MMSE denoiser and is a posterior-gradient step, this SDE is precisely the deterministic-drift reverse SDE of variance-exploding diffusion models. Building on this identification, we give a complete convergence chain: (i) a global-Lipschitz EM discretization error bound; (ii) geometric ergodicity of the limiting homogeneous SDE under dissipativity and a strictly positive terminal noise level ; and (iii) weak convergence of the PnP iterates to the invariant law , with a Wasserstein- bound that quantifies how score-approximation error controls the posterior bias. Boundedness of the denoiser alone is not sufficient: the guarantee is dissipativity a non-vanishing noise floor bounded drift. Experiments on real DRUNet and a pretrained ADM diffusion prior, across denoising, inpainting, super-resolution, and medical imaging, corroborate the EM rate, the threshold, and the posterior-distance bound.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Restoration, Plug and Play Imaging Processing, Stochastic Differential Equation, Diffusion Generative Prior, Anealed Posterior Sampling

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_3azzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops a convergence theory for modern Plug-and-Play (PnP) reconstruction by interpreting annealed PnP iterations as discretizations of stochastic differential equations.

The key observation is that, through the connection between denoiser residuals and score functions, a PnP update can be interpreted as an Euler--Maruyama discretization of a time-inhomogeneous SDE. This shifts the theoretical perspective from classical fixed-point analysis, which typically relies on non-expansiveness, Lipschitzness, or contraction properties of the denoiser, toward an SDE perspective based on dissipativity and ergodicity.

Under appropriate assumptions, the paper establishes discretization-error bounds, proves existence and uniqueness of an invariant distribution and geometric ergodicity of the limiting dynamics, and bounds the distance between the invariant distribution and a variance-inflated Bayesian posterior in terms of score approximation error. The resulting theory covers a regime with bounded but potentially non-contractive denoisers, dissipative data-consistency dynamics, and a non-vanishing noise floor.

The paper complements the theoretical analysis with experiments using diffusion priors and conventional denoisers across several inverse problems, examining convergence behavior, discretization error, stability conditions, and reconstruction performance.

Critical strengths
  1. I find the main conceptual shift of the paper interesting and potentially important. Viewing modern PnP reconstruction as a discretized stochastic process rather than primarily as a deterministic fixed-point iteration provides a natural way to analyze annealed denoisers and score-based priors that do not fit comfortably within classical contraction-based PnP theory.

  2. The paper provides a substantial theoretical development rather than only pointing out the connection between denoising and score estimation. It derives discretization-error bounds, establishes conditions for ergodicity and weak convergence, and quantitatively relates the invariant distribution of the resulting dynamics to the target posterior.

  3. The theory appears to cover a practically relevant regime that is difficult to handle using traditional PnP convergence analyses: bounded but non-contractive denoisers combined with dissipative data-consistency dynamics and a non-zero terminal noise level. I think this extension beyond contraction-based assumptions is one of the strongest aspects of the paper.

  4. I also appreciate the connection between the theoretical quantities and empirical experiments. The paper tests multiple denoisers and inverse problems and includes experiments targeting specific theoretical predictions, such as discretization error and the stability role of the terminal noise level.

  5. More broadly, I think the paper provides a useful conceptual bridge between the PnP and score/SDE perspectives. This may be valuable for understanding modern reconstruction methods in which the prior is increasingly implemented through diffusion or score-based models.

Critical weaknesses

My main concern is presentation and motivation rather than the technical direction of the work.

I found the central conceptual contribution unnecessarily difficult to extract, especially for a general ICLR reader who is not already familiar with PnP reconstruction, score-based models, Tweedie's identity, Euler--Maruyama discretization, and SDE ergodicity. The paper moves relatively quickly into the mathematical development, while the high-level conceptual shift is both simple and important.

In my understanding, the central message is that modern annealed PnP should not necessarily be analyzed as a deterministic map converging to a fixed point. Once the denoiser residual is interpreted as a score, the iteration can instead be viewed as a discretized stochastic process. The relevant theoretical question then changes from whether the iteration is contractive to whether the associated SDE is well-posed and ergodic, what invariant distribution it approaches, and how close that distribution is to the desired posterior. I think this conceptual transition should be made substantially more explicit near the beginning of the paper.

Relatedly, the paper would benefit from a much clearer explanation of the practical significance of the theoretical results. The authors establish a new convergence regime based on dissipativity, boundedness, and a non-vanishing noise floor rather than classical contraction assumptions. However, the introduction does not make sufficiently concrete which practically important modern PnP methods are excluded by previous theory but covered by the present results. Explicitly identifying representative modern denoisers or diffusion-based PnP procedures for which classical theory fails but the proposed theory applies would make the contribution much easier to appreciate.

I also strongly recommend adding a conceptual overview figure near the beginning of the paper. Currently, the first figure appears only in the experimental section. A simple figure contrasting the two theoretical perspectives could communicate most of the paper: classical PnP -> contraction/non-expansiveness -> fixed point, versus denoiser residual -> score -> Euler--Maruyama -> reverse SDE -> ergodicity -> invariant distribution close to the posterior. Such a figure would significantly lower the barrier to understanding the paper.

Finally, because the theoretical contribution depends on assumptions such as dissipativity, bounded drift/denoiser behavior, and a non-zero terminal noise level, I think the paper could do more to explain how restrictive these assumptions are in practical modern PnP systems. In particular, it would be helpful to distinguish assumptions that are primarily technical devices for the proof from assumptions that practitioners should actually enforce when designing an algorithm.

Questions for authors
  1. Could the authors state more concretely which practically important PnP methods are newly covered by the proposed theory but are not covered by existing contraction/non-expansiveness-based analyses? A few explicit examples would make the significance of the theoretical contribution much clearer.

  2. How restrictive is the dissipativity assumption in practice? For commonly used inverse problems and data-consistency operators, when should one expect it to hold, and when might it fail?

  3. How should practitioners interpret the requirement that sigma_min > 0? Is the non-zero noise floor fundamentally necessary for the stochastic interpretation and convergence result, or mainly a sufficient condition required by the current proof technique? What should one expect as sigma_min approaches zero?

  4. The theory assumes boundedness conditions on the denoiser/drift. To what extent are these conditions satisfied by practical neural denoisers such as DRUNet or diffusion-based score networks? Can they be verified or approximately measured empirically?

  5. I would encourage the authors to add a conceptual overview figure near the beginning of the paper. Could the main contribution be summarized explicitly as a transition from the classical “PnP -> contraction -> fixed point” perspective to a “PnP -> score -> SDE -> ergodicity -> posterior” perspective?

  6. More generally, could the authors provide a short intuitive explanation, before the formal theory, of why the SDE viewpoint enables convergence guarantees in cases where the classical fixed-point viewpoint fails? I think this would make the paper substantially more accessible to readers outside the PnP/SDE community.

Reviewer_jRC7OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the convergence of PnP image reconstruction by interpreting its annealed iterations as Euler-Maruyama discretizations of a reverse-time SDE. Using this framework, the authors establish theoretical guarantees for discretization error, geometric ergodicity, and Wasserstein bounds to the posterior distribution, corroborating these findings with extensive experiments across various denoisers and imaging tasks.

Critical strengths

This paper addresses the important theoretical problem of convergence in PnP methods when paired with modern denoisers, effectively bypassing the restrictive assumptions of classical fixed-point theory, and supports these theoretical insights with a broad experimental evaluation that spans multiple denoisers, reconstruction tasks, and detailed numerical studies of discretization rates and stationary distributions.

Critical weaknesses
  1. Unclear SDE representation. Proposition 3.1 interprets the PnP iteration as an Euler–Maruyama discretization. However, the original PnP update appears deterministic, while the proposed SDE introduces a Gaussian noise term whose origin is not clearly justified. This makes it difficult to determine whether the convergence results directly apply to the original algorithm.
  2. Strong theoretical assumptions. The convergence results rely on restrictive assumptions, including dissipativity and global Lipschitz continuity, which are not clearly verified for the denoisers used in the experiments.
Questions for authors
  1. Could the authors clarify how the Gaussian noise term in Proposition 3.1 arises from the original PnP iteration?
  2. In the proof of Theorem 5.4, the supremum over 1-Lipschitz functions appears to characterize (W_1) rather than (W_2). This step requires further justification, as it directly affects one of the paper's main theoretical guarantees. Is an additional argument needed to establish the claimed (W_2) bound?
Reviewer_ivioOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims an interpretation of annealed Plug-and-Play (PnP) as an Euler–Maruyama discretization of a reverse-time SDE. Based on this representation, they claim convergence to an invariant distribution under dissipativity and a positive terminal noise level. Experiments on image reconstruction tasks with learned denoisers and a pretrained diffusion model are presented to support these claims.

Critical strengths

The paper tries to address a relevant question: how annealed denoising priors relate to the convergence behavior of PnP reconstruction algorithms.

Critical weaknesses

The paper contains a fundamental confusion between deterministic Plug-and-Play (PnP) reconstruction algorithms and posterior sampling methods. These two families pursue different objectives: PnP reconstruction seeks a point estimate, typically through convergence to a fixed point or a stationary point, whereas posterior sampling seeks to generate samples from a target distribution. They can share the same denoiser or score model, but this does not make their algorithms equivalent. In particular, inserting a diffusion denoiser into a deterministic PnP algorithm does not turn it into a posterior sampler.

This confusion affects the central proposition and the subsequent convergence claims. My main concerns are the following.

  1. The main argument of the paper from Proposition 3.1 is false.

Equations (1)–(2) define the deterministic update

vt+1=D(h(vt;y),σt).v^{t+1}=D(h(v^t;y),\sigma_t).

Applying Assumption 2.1 gives

=h(v^t;y)-v^t +\sigma_t^2s_{\theta,\sigma_t}(h(v^t;y)) +r_{\sigma_t}(h(v^t;y)).$$ There is no Gaussian noise in this expression. Nevertheless, the proof introduces an additional term $\sigma_t\xi_t$, attributing it to Assumption 2.1. The Gaussian corruption model used to define an MMSE denoiser does not imply that evaluating that denoiser injects fresh noise. Appendix B identifies this perturbation with $z-v^t$n which is deterministic, with no justification for treating it as an independent Gaussian increment. Proposition 3.1 therefore does not establish the claimed Euler–Maruyama representation. Consequently, the subsequent SDE analysis does not provide the claimed guarantees for the stated PnP algorithm. 2. **Initialization is unspecified, although it is essential for sampling interpretation** With fixed measurements, a prescribed schedule, and a deterministic initialization, the stated PnP algorithm has deterministic iterates. Posterior samplers require a specified source of randomness. This remains true for probability-flow ODE or flow-matching samplers: their randomness comes from sampling the initial condition. The paper specifies neither the initial iterate nor its distribution, and does not explain how initialization is handled in the experiments. 3. **The algorithmic formulations are conflated.** Equations (1)–(2) describe PnP half-quadratic splitting, not ADMM. In addition, Proposition 3.2 replaces the proximal data step of Equation (1) with an explicit gradient step. 4. **Other unprecise statements.** Section 2.1 first interprets the denoising step as MAP estimation, then invokes the conditional mean and Tweedie’s identity. MAP and MMSE estimators differ. Also, there is confusion between prior and algorithm: a noise-dependent prior or score is a statistical object; it does not, by itself, specify a diffusion sampling algorithm. 5. **Theorem 5.3 does not establish its stated convergence conclusion.** Even setting aside Proposition 3.1, Equation (10) retains a nonzero error floor for fixed $\Delta$ and nonzero score error. Such a bound does not establish the announced weak convergence to $\pi$. Furthermore, Theorem 4.1 gives a finite-horizon estimate whose constant depends on $T$; using it to establish long-time convergence requires additional control. 6. The discussion of the three regimes also conflates convergence of deterministic iterates, strong approximation of SDE trajectories, and convergence of probability laws. Pathwise uniqueness of an SDE does not itself imply convergence of an algorithm to a fixed point. 7. **The presentation of existing PnP convergence theory is incomplete.** The introduction and abstract reduces prior theory to contraction arguments based on nonexpansive denoisers, overlooking results based on other structures. Relevant examples include [Xu et al. (2020)](https://arxiv.org/abs/2005.07685) on MMSE denoisers, [Hurault et al. (2021/ICLR 2022)](https://arxiv.org/abs/2110.03220) on gradient-step denoisers, and [Hurault et al. (ICML 2022)](https://arxiv.org/abs/2201.13256) on proximal denoisers and nonconvex regularization. 8. **The experimental diagnostics do not resolve the theoretical mismatch.** Stabilization of PSNR does not establish convergence in distribution to a posterior, as proved in the theorem. The experiments should also specify the exact implemented updates, initialization, and every source of randomness.
Questions for authors

See above

Reviewer_S4gtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at the convergence of plug-and-play methods with annealed denoising priors. It uses Tweedie’s identity to connect PnP updates to a reverse SDE, then develops claims about discretization error, convergence, and how close the limiting distribution is to the target posterior. The experiments include Gaussian toy examples and image restoration tasks such as denoising, inpainting, and super-resolution, along with an MRI example.

Critical strengths
  1. The paper asks a useful question: when annealed PnP converges, what does it actually converge to? I appreciate the effort to separate the effects of discretization, imperfect scores, and long-term behavior.
  2. Including Gaussian examples alongside image restoration experiments is a good choice. The Gaussian cases make it easier to check the theoretical predictions directly, while the restoration experiments show how the ideas relate to practical methods.
Critical weaknesses
  1. I do not see how the deterministic PnP update becomes the stochastic process in Proposition 3.1. The proposed SDE introduces Gaussian noise that is not part of the stated algorithm. The Gaussian noise assumed when training a denoiser does not automatically introduce fresh noise during inference, and Appendix B does not resolve this distinction. This leaves me unsure whether the subsequent theory actually applies to the PnP method being studied.
  2. I also have concerns about the main convergence proofs. Theorem 5.3 claims convergence to the continuous SDE’s invariant distribution, but its bound retains a discretization error that does not vanish at a fixed step size. In Theorem 5.4, the detailed-balance calculation appears to miss a factor of two, which affects the claimed stationary distribution. The proof also uses the dual formula for W1 to establish a W2 bound.
Questions for authors
  1. Where does the independent Gaussian noise in Proposition 3.1 enter the actual PnP algorithm? Is a noise-injection step missing from eqs (1)–(2), and was it used in the experiments?
  2. Is Theorem 5.3 meant to describe convergence to the discrete chain’s invariant distribution or the continuous SDE’s? If it is the latter, how does the discretization bias disappear at a fixed step size?
  3. How would you correct the detailed-balance calculation and the use of W_1 duality in Theorem 5.4? Would the claimed posterior and W_2 bound still hold under the current assumptions?

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

Related papers

Discussion (0)

Sign in to comment.