acceptodds
Under review as a conference paper at ICLR 2027

Principled MAP estimation for inverse problems: bridging the gap between convergence and performance

Abstract

Pretrained denoisers provide a powerful way to incorporate image priors into restoration algorithms. Plug-and-Play and RED approaches exploit fixed-noise-level denoisers within first-order optimization schemes, with convergence guarantees, but often struggle to achieve high-quality reconstruction on severely ill-posed inverse problems. In contrast, recent state-of-the-art approaches leverage denoisers derived from flow- or diffusion-based generative models and evaluate them along a sequence of decreasing noise levels. While these methods achieve strong empirical performance, their convergence theory remains limited. In this paper, we bridge this gap by specifically designing an algorithm that combines denoisers at decreasing noise levels with a schedule tailored to ensure convergence. From a Bayesian perspective, we prove that our method converges to a Maximum a Posteriori (MAP) estimate, under suitable assumptions. Subsequently, we apply our method to various ill-posed inverse problems and show that it surpasses convergent methods while competing with state-of-the-art empirical ones.

Primary area:
optimization
Keywords:
image restoration, inverse problems, generative models, optimization

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_SG9HOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper propose a MAP estimation for imaging inverse problems, based on noise annealing and averaging of MMSE denoiser (Gamma). The algorithm consists of two step (1) gradient step on adjusted data-fidelity $ (\lambda(\sigma^2)/2)||x − u||^²$, which adds a vanishing Tikhonov anchor toward a chosen reference point u (2) denoising the current step.

Three variants is proposed including deterministic Gamma (paper proves that the iterates are bounded and every cluster point is a MAP solution. If additionally σk2/λk→0\sigma_k^2/\lambda_k \to 0, the iterates converge to x†x^\dagger, the MAP solution closest to u.), renoised-GAMMA (the same result when the denoiser output is averaged over infinitely many noisy copies of the input.), stochastic-GAMMA (convergence in mean square for a single-sample version, but only if the injected noise is decoupled and much smaller than the denoiser level).

Experiments use annealed models on CelebA-128 and AFHQ-Cat-256 for denoising, deblurring, ×4 super-resolution, random inpainting and box inpainting. Baselines are PnP-Flow and Approx-PGD.

Critical strengths
  1. A single-loop algorithm that extends MMSE Averaging (Pesme 2025) from MAP denoising to general inverse problems.
  2. A unifying view where with λ = 0, GAMMA is annealed RED. Renoised-GAMMA with one sample is annealed SNORE.
  3. Convergence of the iterates themselves to a specific MAP point.
Critical weaknesses
  1. Novelty is thin. Some claims and prior work question the novelty of the work.
  • “First approach to bridge provably convergent denoiser methods and annealed generative methods”
  1. Shrestha & Fu, ICLR 2026: Score-based denoiser with ADMM, noise added to the denoiser and prove two convergence results (1) high-probability fixed-point-ball convergence with a constant step, and, (2) under weaker conditions, convergence with an adaptive step-size schedule via a bounded-denoiser argument.
  2. Chan IEEE TCI 2017: Their main theorem proves fixed-point convergence of PnP-ADMM for any “bounded denoiser”, where the continuation scheme drives the penalty to infinity, which results in annealing the denoiser strength.
  3. Park SGPnP: Similar algorithm with data-fidelity gradient + noise injected to the MMSE denoiser. Their regularizer is exactly the prior term of GAMMA’s renoised objective. Injected noise level differs from the conditional noise level denoiser is trained on like Gamma stochastic. SGPnP claims that under an annealed noise schedule its iterates converge to a stationary point of the exact, un-smoothed objective, (MAP objective).
  4. The vanishing Tikhonov anchor and selection of the MAP point closest to u is from classical convex optimization (Attouch & Cominetti (1996), Cominetti, Peypouquet & Sorin (2008)).
  • Assumptions: Authors justification on assumption H2 is not satisfactory. Log-concave is unimodal, and this matters here specifically, noiseless Gamma and Approx-PGD objective is MAP, where we see the cartoon like images , here rejoiced and relaxed version work fine. This shows that good results are consistent with not converging to MAP. The cited papers that use this assumption are Langevin sampling, very different setting.

  • Exact MMSE denoiser might create an issue in the assumed setting, while being fine in the other algorithms, because the contraction per step is only αkλk→0\alpha_k\lambda_k \to 0, there is nothing to offset the effect of inexact MMSE denoiser.

  • The proof contracts the distance to the moving minimizer by a factor 1−αkλk1 − \alpha_k \lambda_k per step. Since αk≤1\alpha_k \leq 1, we have ∑kαkλk≤∑kλk≈10⁻3× 1900≈1.9\sum_k \alpha_k \lambda_k \leq \sum_k \lambda_k \approx 10^{⁻³} × ~1900 \approx 1.9. So the total guaranteed contraction after 2000 iterations is at best e−1.9≈0.15e^{−1.9} \approx 0.15. The initial error shrinks by at most about 85%. The is not consistent with the theorem.

  • The selection condition σ2/λ→0\sigma^2/\lambda \to 0 is far from met. With σ0=10\sigma_0 = 10: σN≈10/41.4≈0.24\sigma_N \approx 10/41.4 \approx 0.24, so σN2≈0.058\sigma_{N^2} \approx 0.058 and σN2/λN≈0.058/0.00093≈63\sigma_{N^2}/\lambda_N \approx 0.058/0.00093 \approx 63. That ratio is supposed to go to zero, yet it is about 60. Even with σ₀ = 1 it is about 0.6. So the “converges to the MAP closest to u” result says nothing about these runs. The paper also never states what u is in the experiments.

  • Experimental resutls:

  1. The method aimed at the MAP is worse than the degraded input on deblurring (both AFHQ and Celeba). Deblurring is overall not a difficult inverse problem, around 10 dB worse seems to be something gone wrong like iterates not converging, denoiser far from exact MMSE, MAP is bad estimator.
  2. Could the authors report the convergence of iterates, like residual over iteration?
  3. Performance is bad, compared to PnP-Flow.
  4. Compute is not matched, this method seems to be comparably more expensive.
  5. Baselines are not included: Generative: only PnP-Flow is included. Missing are DiffPIR, DPS, RED-diff, and Flower, ICTM, D-Flow, SGPnP and annealed SNORE. Convergent methods: the abstract claims GAMMA “surpasses convergent methods,” but the only convergent baseline is Approx-PGD, which appears broken (point 1). The standard convergent PnP methods, GS-PnP and Prox-PnP (Hurault2022), are not compared.
  6. No evidence of convergence on real images
Questions for authors

how does an approximation error eke_k in the denoiser propagate through the analysis, and what decay of eke_k preserves convergence?

Are the schedule assumption same as reported runs? It says γ=0.98,β=0.01,N=2000,λ0=10−3\gamma = 0.98, \beta = 0.01, N = 2000, \lambda_0 = 10^{-3}, however, λN=λ0⋅2001−0.01≈0.93λ0\lambda_N =\lambda_0 · 2001^{−0.01} \approx 0.93 \lambda_0.

What is u in the experiments, and does changing it affect results?

3 comments

We thank the reviewer for their thorough reading and constructive feedback, and recognition of our paper's strengths (single-loop extension of MMSE Averaging to general inverse problems, unifying view of annealed RED/SNORE, and iterate convergence to a specific MAP estimator). Below, we first address the concern about the limited novelty of our method.

Novelty

Thank you for pointing out these interesting references, which we have carefully read. However, in our view, they do not diminish the novelty of our work.

  • "First approach to bridge...” (l.479): In Section 6, we wrote "convergent denoiser-based MAP methods, such as Pesme et al. (2025)", referring to convergence to a minimizer of the true unsmoothed objective FF. Our goal is to characterize the limit rather than showing that the iterates converge to an unknown limit.

  • Chan et al. (IEEE TCI 2017) & Shrestha & Fu (ICLR 2026) : The first paper is already cited in our Introduction (l.037–042), where we added the second one. In the two articles, it is proven that the iterates form a Cauchy sequence and thus converge. However, when using denoisers with noise levels going to 00, the limit point is not characterized and is not shown to minimize any underlying objective.

  • Vanishing Tikhonov (Attouch & Cominetti, 1996; Cominetti et al., 2008) As acknowledged in Section 3 ("Related works", l. 243–245, citing Attouch & László, 2024), vanishing Tikhonov is a classical convex optimization technique. We will add these pioneering references. Our contribution lies in coupling it with the non-standard Gaussian smoothing inside the logarithm (−log⁡pσ-\log p_\sigma) to make the algorithm work when the data-fidelity is not strongly convex, in the context of inverse problems.

  • Park et al. (SGPnP, 2026) : Thank you for pointing out this recent work. While SGPnP is indeed close to a single sample Renoised GAMMA without Tikhonov regularization, there are fundamental distinctions in what is proven, as well as unaddressed theoretical limitations and incompatible assumptions in its analysis, that our work overcomes.

    Main distinctions:
    The annealed convergence of SGPnP (Theorem 2) is proven (also in the setting where the denoisers are perfect MMSE denoisers) with a double-loop regime requiring IkI_k inner iterations for a fixed σk\sigma_k. By their Assumption 33 and Lemma 33, one can show that Vmax(σ)≥1/σ2V_{max}(\sigma)\geq 1/ \sigma^2, so that Ik→∞I_k \to \infty as σk→0\sigma_k \to 0. Moreover, no explicit step size schedule is given beyond the constraints γk≤1/Lk\gamma_k\leq 1/L_k and γkLkVmax(σk)→0\gamma_k L_k V_{max}(\sigma_k)\to 0. The proof of Theorem 2 assumes iterate compactness (Assumption 4). In contrast, for (Renoised-)GAMMA, we prove convergence of the single-loop iterates (Ik=1I_k = 1) to a uniquely characterized point x†=ProjArgmin F(u)x^\dagger = \text{Proj}_{\text{Argmin } F}(u), and we provide a family of explicit schedules αk=σk2σk2+τ\alpha_k = \frac{\sigma_k^2}{\sigma_k^2 + \tau}. We do not assume anything on the iterates.

    Theoretical limitations and incompatible assumptions: Several assumptions and proof steps are mathematically problematic.
    Assumption 1 requires ∇fσ=∇g+∇hσ\nabla f_\sigma=\nabla g+\nabla h_\sigma to be uniformely bounded on Rd\mathbb{R}^d, which fails for quadratic data-fidelity terms g(x)∝∥Ax−y∥2g(x)\propto \|Ax-y\|^2. Assumption 2 vs Assumption 3: By Assumption 2 and Tweedie's formula, Uσ(x,n)=−nσ−∇log⁡pσ(x+σn)U_\sigma(x,n)=-\frac{n}{\sigma}-\nabla \log p_\sigma(x+\sigma n). Requiring ∥Uσ(x,n)∥≤B\|U_\sigma(x,n)\|\leq B almost surely (Assumption 3) is therefore impossible for n∼N(0,Id)n\sim \mathcal{N}\left(0, Id\right). Assumption 4 and proof of Theorem 2: Assuming that the stochastic iterates remain within a compact set K\mathcal{K} (Assumption 4) is strong, and its justification in the first lines of the proof of Theorem 2 is invalid. Furthermore, the uniform bound Δ<+∞\Delta<+\infty on fσk(xk(0))−fσk⋆f_{\sigma_k}(x_k^{(0)})-f_{\sigma_k}^\star is unproved, and the use of Bolzano-Weierstrass in the end is improper. They do not show the convergence of the iterates, but rather that there exists a subsequence of iterates for which ∇f0\nabla f_0 goes to 00 almost surely.

We now address the remaining remarks and questions of the reviewer.

Discussion on the assumptions (H2)

"Authors justification on assumption H2 is not satisfactory."

We have edited the discussion below H2 (Section 4.1) to emphasize that, while log-concavity is a strong requirement, our proof is inherently local, which suggests that an extension to multimodal, locally log-concave distributions is possible. This perspective is supported by our experimental results on a Gaussian mixture prior and on natural images.

"noiseless Gamma and Approx-PGD objective is MAP, where we see the cartoon like images , here renoised and relaxed version work fine. This shows that good results are consistent with not converging to MAP. "

The artifacts of noiseless GAMMA and Approx-PGD do not come from converging to the MAP, and the good results of Renoised and Relaxed GAMMA do not come from diverging away from the MAP:

  • As now addressed in Section 5.1 (Figure 1, 2, 10, 11), when the MMSE denoiser is exact both noiseless GAMMA and Renoised-GAMMA converge to the same MAP solution, whereas PnP-Flow and DPS fail to converge.
  • We believe that the performance gap between GAMMA and Renoised/ Relaxed GAMMA on the natural images comes from the denoiser rather than from the MAP estimator itself. This diagnosis is consistent with Vert et al. (2026, already cited in Section 5.3), who show that the cartoon-like artifacts of algorithms targetting the MAP with Diffusion Models "seems to come from score errors" rather than from MAP estimation itself. Indeed, a neural denoiser being trained on noisy images, its evaluation on unnoised iterates is inaccurate.

Bad result for deblurring

"Deblurring is overall not a difficult inverse problem, around 10 dB worse seems to be something gone wrong"

Typo on our part in the Table 4: the super-resolution and deblurring entries of the GAMMA (noiseless) row were accidentally swapped, creating this 10dB gap. It is now fixed.

Cost of the method, guarantees for finite schedules

The slow convergence of GAMMA and Renoised-GAMMA is caused by the conservative choice of the schedule, not the method itself. Relaxed-GAMMA (Renoised-GAMMA with more aggressive schedules) converges in 100-500 iterations while producing high quality reconstructions for every inverse problem. We added Corollary 9 (Section 4.3) which establishes that, for a family of finite schedules, the final iterate xNx_N converges (at a given rate) to the MAP, when NN grows.

Additional baselines (DPS and DiffPIR)

"Baselines are not included"

We have added DPS and DiffPIR to our benchmarks on both CelebA (Table 1) and AFHQ-Cat (Table 4). Relaxed-GAMMA consistently outperforms DPS and DiffPIR while matching PnP-Flow.

"The standard convergent PnP methods, GS-PnP and Prox-PnP (Hurault2022), are not compared."

In our comparison, we restrict ourselves to methods relying on off-the-shelf denoisers, which excludes Gradient-Step (GS) denoisers. We have clarified the sentence in the abstract.

Evidence of convergence on images / choice of uu / contraction factor close to 1

We have added a new paragraph in Appendix D.2 with λ0=0.1\lambda_0 = 0.1, so that λk\lambda_k is 10210^2 times larger than in the main experiments. This regularization strength increases the impact of uu, makes (1−αkλk)(1 - \alpha_k \lambda_k) go to 1 less fast, and σ2/λ\sigma^2/\lambda go to 0 faster (e.g. contraction factor goes from e−1.9e^{-1.9} to e−190e^{-190}). We also run 4040k iterations instead of 22k, to get as close as possible to the asymptotic regime. For those examples, we added in Figures 7, 8 and 9 curves for the residual ∥xk−xk−1∥∥x0∥\frac{\| x_k - x_{k-1} \|}{\| x_0 \|} along the iterations kk, together with the MAP objective Fκ(xk)=f(xk)−κτlog⁡p1θ(xk)F^\kappa(x_k) = f(x_k) - \kappa \tau \log p_1^\theta(x_k) (where −log⁡p1θ-\log p_1^\theta is numerically estimated via the learned denoiser).

We observe the following:

  • Impact of u: in Figures 7, 8 and 9, an informative choice for uu (e.g. an average dataset image for inpainting that fills the mask, bicubic interpolation for superresolution) strongly improves the reconstruction, especially for the noiseless version of GAMMA. On superresolution in particular, we avoid the artifacts that originally appeared in the main experiments and achieve a PSNR far above the observation.
  • On convergence on images, both GAMMA and renoised GAMMA decrease the MAP objective and have low residual along the iterations. The convergence remains slow in these experiments, but the schedules used in the Relaxed GAMMA version suggest that accelerated convergence is achievable.
  • Renoised GAMMA reaches substantially lower MAP objective values than its noiseless version at equal iteration count, which is consistent with its better reconstruction metrics in our experiments.

Questions

Handling approximate MMSE

We added results in Appendix B.5. Similarly to what is shown in Pesme et al. 2025, when the datafit term is strongly convex and the score error ∥ek∥\|e_k\| over the iterates is bounded by ξ\xi, the iterates converge to a point at distance O(ξ) of the MAP. We believe a further study would deserve a separate work.

Are the schedule assumption same as reported runs

The schedules used in the runs are defined in Remark 8 and their parameters are indicated in Table 2&3. For GAMMA and Renoised-GAMMA, every schedule verifies the conditions of the theorems.

What is uu and does changing it affect the result ?

In the main experiments (Tables 1 and 4), we set u∼N(0,Id)u \sim \mathcal{N}(0, Id) (now explicitly stated at the beginning of Appendix D, l. 1682), which has a little impact on PSNR because λ0=10−3\lambda_0 = 10^{-3} is small. In contrast, when λ0\lambda_0 is larger (λ0=0.1\lambda_0 = 0.1, Appendix D.2, Figures 7–9), choosing an informative uu (dataset mean image or bicubic interpolation) significantly improves PSNR and removes artifacts, as detailed above.

Reviewer_tpK4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

This paper investigates Maximum A Posteriori (MAP) estimation for image inverse problems using generative priors. The authors propose GAMMA by leveraging the relationship between MMSE denoisers and the score functions of smoothed probability distributions. They derive principled schedules for gradient descent step sizes, noise annealing, and regularization, and establish convergence to a MAP solution under certain assumptions.

Critical strengths

Clear motivation and thorough theoretical analysis.

The paper is well structured and easy to follow. I was able to quickly understand the problem being addressed and how this work builds upon prior research. The theoretical analysis is relatively thorough: beyond proposing an algorithm, the authors establish convergence results by exploiting the strong convexity of smoothed objective functions, the regularity of the minimizer path, and carefully designed annealing schedules.

Critical weaknesses

1.Limited selection of baselines.

The experiments mainly compare GAMMA with PnP-Flow and Approx-PGD, while omitting representative diffusion-based inverse problem solvers such as DPS and DiffPIR. Although these methods are discussed or cited in the paper, they are not included in the quantitative comparisons. I suggest including these baselines under comparable experimental settings to provide a more comprehensive evaluation of GAMMA's reconstruction quality and computational efficiency, rather than focusing primarily on methods that are theoretically closely related.

2.Limited empirical improvements and an unclear trade-off between convergence and performance.

The results in Table 1 and Figure 2 show that GAMMA does not consistently outperform PnP-Flow, with some reported PSNR values being lower than those of PnP-Flow. In particular, Renoised-GAMMA, which satisfies the theoretical convergence conditions, does not consistently achieve better reconstruction quality than PnP-Flow. Meanwhile, the better-performing Relaxed GAMMA does not satisfy the convergence schedule conditions established in the paper. Therefore, the trade-off between the claimed convergence guarantees and practical reconstruction performance deserves further investigation.

Questions for authors

1.Given that the theoretically guaranteed GAMMA variants still exhibit a performance gap compared with PnP-Flow, what are the main factors contributing to this gap?

2.To what extent do the strict convergence conditions affect practical reconstruction quality?

1 comment

We thank the reviewer for their constructive assessment and their suggestions.

W1. Adding other baselines

We added the comparison with DPS and DiffPIR on both CelebA (Table 1) and AFHQ-Cat (Table 4). We find that neither of these methods consistently outperforms our approach across all five tasks: Renoised-GAMMA achieves performance comparable to DiffPIR and outperforms DPS, while Relaxed-GAMMA, using a comparable iteration budget to DiffPIR, matches or surpasses both DPS and DiffPIR across nearly all tasks.

W2. Trade-off between convergent and performant variants, and practical impact of schedule

"the trade-off between the claimed convergence guarantees and practical reconstruction performance deserves further investigation."

The theoretical guarantees that we obtain are based on worst-case analysis, which is by design pessimistic and most likely leads to overly conservative choices of schedules. In practice, using a more aggressive schedule that reaches 00 faster improves reconstruction quality. To address this concern, we have added Corollary 9 (Section 4.3), which analyzes a finite schedule with NN steps, relatively close to Relaxed-GAMMA, and shows that the final iterate xNx_N converges to the MAP estimator x†x^\dagger when N→+∞N\to +\infty, with an explicit convergence rate.

Our goal is to propose a method that simultaneously performs well in practice, and can be rigorously analyzed in tractable settings. We believe it represents a first step towards reconciling state-of-the-art empirical performance and theoretical convergence.

Q1. Performance gap with PnP-Flow

First, note that PnP-Flow does not come with convergence guarantees, even in controlled settings (e.g., log concave pp); it is purely designed for performance. In our new Figure 1, we show that in a simple log-concave setting, using the exact MMSE denoisers, neither PnP-Flow nor DPS converges to the MAP, whereas GAMMA and Renoised-GAMMA do.

Second, in our view, two factors explain the performance gap between (Renoised-)GAMMA and PnP-Flow on natural images. This is also discussed in Section 5.3.

  • Role of renoising (GAMMA vs Renoised GAMMA) : While GAMMA performs well on "simple" inverse problems (such as deblurring and random inpainting), it produces cartoon-like images on generative tasks (e.g., super-resolution and mask inpainting). Renoised-GAMMA performs much better in these cases and nearly reaches PnP-Flow performance. This is likely due to the fact that the neural denoisers used in the algorithms are learned more accurately on noisy images.
  • Role of schedules (Renoised-GAMMA vs Relaxed GAMMA): As mentioned in the answer to W2, GAMMA is designed to come with guarantees in a controlled setting, maybe at the price of overly conservative schedule choices: the decay of the noise σk\sigma_k is very slow (see Figure 5). Relaxed GAMMA (now theoretically supported by Corollary 9) is an intermediate variant that slightly deviates from GAMMA but reaches performance close to PnP-Flow across all tasks.

Q2. Impact of the schedule conditions on the reconstruction

Comparing Renoised-GAMMA and Relaxed-GAMMA in Tables 1 and 4 shows that the strict schedule conditions have a moderate impact on PSNR (∼\sim 1-1.5 dB), while their main cost is the number of iterations needed to obtain satisfactory results (N=2000N=2000 vs N⩽500N\leqslant 500). Indeed, in practice, relaxing conditions on the schedules to allow the noise level σk\sigma_k to go to 00 faster accelerates the convergence of the algorithms. Such a relaxation is now supported by our Corollary 9, which shows the convergence of the final iterate to the MAP x†x^\dagger as the schedule horizon grows.

Furthermore, to illustrate the impact of the schedule parameters (λ0,γ,β)(\lambda_0, \gamma, \beta) and of the center uu of our quadratic regularization, we added a detailed ablation on AFHQ and CelebA in Appendix D.2, with a higher λ0\lambda_0. Figures 7, 8 and 9 show that both GAMMA and Renoised-GAMMA consistently decrease the estimated MAP objective, and that choosing an informative uu substantially boosts PSNR.

Reviewer_yzDHOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an inverse problem solver that generalises MMSE Averaging from MAP denoising problems to general inverse problems. The authors show that under some conditions, convergence guarantees can be obtained for the proposed algorithms. The paper introduces two algorithm versions: Generalized Annealed MMSE Average (GAMMA) and Renoised-GAMMA. While GAMMA generalises the MMSE Averaging algorithm to any inverse problem by replacing the fixed term yy by a gradient descent step on a regularised version of ff, Renoised-GAMMA also replaces the denoising of the iterate by the average of its denoised noisy versions. The authors perform empirical experiments on CelebA and AFHQ.

Critical strengths
  • The paper proposes a generalisation of the MMSE Averaging algorithm that focussed on MAP denoising. The proposed algorithm extends this algorithm to general inverse problems.
  • The authors derive convergence guarantees for both variants of the proposed method.
Critical weaknesses

The log-concavity assumption on p that as stated in the paper is most likely not fulfilled for natural images. The experiments presented in the paper thus should not be labeled as “Theory: Yes” and the author’s claim that the gap between provably convergent methods and recent generative restoration methods can be substantially narrowed is not empirically proven in this work.

Questions for authors
  • Please could you clarify how the experiments performed on natural images can be interpreted towards closing the gap between provably convergent methods even though the log-concavity assumption does not hold in this case?
  • In the experimental part of the paper you show that your method does not necessarily only apply to imaging inverse problems. I suggest changing the phrasing in the introduction (where you focus on imaging inverse problems) to make it more consistent with the rest of the paper.
2 comments

We thank the reviewer for their insightful comments and suggestions.

W1 & Q1. Applicability of the convergence guarantees to natural images

We agree that the log-concavity assumption does not hold for natural image distributions. We have clarified this distinction right after introducing Assumption H2 (Section 4.1), and removed the "Theory" column from Tables 1 and 4.

As clarified in the revision, our goal is to propose a principled algorithm that simultaneously performs well in practice and admits rigorous MAP convergence guarantees in a tractable theoretical setting.

We added (Section 5.1) two experiments under exact log-concave priors, in which the MMSE denoiser is available exactly and convergence of GAMMA to the MAP can be measured.

  • Non-Gaussian log-concave prior (Figure 1). We consider a smooth Laplace-based prior, which satisfies the theoretical assumptions while being non-Gaussian. Its MMSE denoiser can be computed analytically, and the MAP solution can be obtained numerically with high accuracy. We observe that GAMMA converges to the MAP solution, in agreement with our theoretical results, whereas empirical solvers (PnP-Flow and DPS) fail to reach the MAP, even as the iteration budget grows to 10510^5.
  • Gaussianized CelebA (Figures 2, 10, 11). We construct a Gaussian prior using empirical mean and covariance estimated from the CelebA dataset. We observe convergence of GAMMA toward the MAP solution.

Finally, we added in Appendix D.2 (Figures 7, 8, 9) plots of convergence of GAMMA and Renoised-GAMMA when using a learned prior on natural images. Those plots show evidence of convergence even when the log-concavity assumption is not verified: both the estimated MAP objective and the relative iterate residuals consistently decrease across iterations. Without claiming that the convergence guarantees extend to arbitrary natural image distributions, the optimization dynamics seems to behave well, which is consistent with our intuition that the analysis could be extended in future work to weaker assumptions, such as local log-concavity.

Q2. Generality beyond imaging inverse problems

Indeed, while our experiments focused on imaging, our theoretical framework and GAMMA algorithms apply to general inverse problems in Rd\mathbb{R}^d ; we have reworded the abstract, the introduction and Section 2 in terms of general signals. We also added a non-imaging experiment on sparse signal recovery with a smooth Laplace prior (Section 5.1, Figure 1 and Appendix D.4).

Thank you very much for clarifying and making the suggested changes to the manuscript. My main concerns have been addressed and I will increase my score accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General answer to all reviewers

We thank the reviewers for their constructive feedback.

We added the new experiments and revisions in the new version of the manuscript in red.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.