acceptodds
Under review as a conference paper at ICLR 2027

Signal Denoising in Non-Gaussian Noise via Diffusion Regularization

Abstract

When the signal prior is known but the measurement-noise distribution is unknown and non-Gaussian, a key challenge is how to incorporate data-driven noise information into a principled denoising objective while explicitly preserving the signal prior. We develop a noise-domain variational framework for this setting. Specifically, a noise-domain optimization problem is rigorously derived from the finite-dispersion variational formulation of regularization by denoising diffusion process (RED-diff), with the learned noise distribution treated as the RED-diff prior and the signal structure retained explicitly. The resulting objective admits a maximum a posteriori (MAP) interpretation under a normalized effective noise likelihood. This framework is instantiated as Regularization by Denoising with an Atomic-Norm Prior (RED–Atomic) for gridless line spectral estimation, where the atomic norm explicitly encodes spectral sparsity. The corresponding optimization problem is solved by primal–dual splitting with a Toeplitz semidefinite representation. Numerical experiments demonstrate competitive or improved performance across a range of non-Gaussian noise models.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
signal denoising, non-Gaussian noise, diffusion model, variational regularization, gridless line spectral estimation.

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_8Yc9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RED–Atomic, a noise-domain diffusion regularization framework that combines a learned non-Gaussian noise model with an explicit atomic-norm signal prior for gridless line spectral estimation.

Critical strengths

The framework offers a principled variational and MAP interpretation while preserving known signal structure, with analytic calibration that usefully distinguishes score-learning error from diffusion-smoothing bias.

Critical weaknesses
  1. The experimental evidence is not sufficiently convincing. Evaluation is restricted to small synthetic line spectral estimation problems, without validation on real measurements. This leaves the practical benefits under realistic noise unclear. Beyond real data from the target application, extensions to image denoising benchmarks such as SIDD [1] and the Fluorescence Microscopy Denoising (FMD) dataset [2] would strengthen the evidence, with appropriate adaptation of the signal prior and noise model.

  2. The methodological novelty appears limited. RED/RED-diff and primal–dual splitting are established techniques. The main algorithmic contribution combines a diffusion-based noise model with a classical atomic-norm signal prior, which remains unchanged. Although the noise-domain variational derivation provides useful justification, the paper does not sufficiently demonstrate a substantial methodological advance beyond this combination of existing components.

  3. Optimization stability and convergence remain insufficiently established. The diffusion-based noise fidelity can be nonconvex, raising concerns about sensitivity to initialization and step sizes. Moreover, Appendix B.4 only characterizes the limiting fixed point if the iterates converge; it does not establish convergence for the general learned-score implementation. A nonconservative learned score field further weakens the connection between this fixed point and a MAP solution.

References

[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. “A High-Quality Denoising Dataset for Smartphone Cameras.” CVPR, 2018.

[2] Yide Zhang, Yinhao Zhu, Evan Nichols, Qingfei Wang, Siyuan Zhang, Cody Smith, and Scott Howard. “A Poisson-Gaussian Denoising Dataset with Real Fluorescence Microscopy Images.” CVPR, 2019.

Questions for authors

see weakness

3 comments

Response to Reviewer 8Yc9

General Response:

We sincerely thank the reviewer for the constructive feedback. We have revised the manuscript and added experiments to address the concerns regarding experimental evidence, methodological contributions, and optimization stability. If our responses have addressed your concerns, we would greatly appreciate your reconsideration of the overall rating. Below, we provide detailed responses to each comment.


Comment 1: The experimental evidence is not sufficiently convincing. Evaluation is restricted to small synthetic line spectral estimation problems, without validation on real measurements. This leaves the practical benefits under realistic noise unclear. Beyond real data from the target application, extensions to image denoising benchmarks such as SIDD [1] and the Fluorescence Microscopy Denoising (FMD) dataset [2] would strengthen the evidence, with appropriate adaptation of the signal prior and noise model.


Response:

We thank the reviewer for the suggestion to broaden the scope of our paper. First, SIDD [1] and FMD [2] are valuable datasets, but their noise models do not directly match ours. Our observation model assumes that the noise distribution is independent of the signal xx . The noise level function in SIDD is signal-dependent and satisfies β2(x)=β1x+β2\beta^2(x)=\beta_1x+\beta_2, whereas the noise in FMD satisfies ni=np(xi)+ngn_i=n_p(x_i)+n_g. Therefore, both noise distributions depend on the conditional likelihood p(n∣x)p(n\mid x), which is beyond the scope of our current model.

Appendix E adds a controlled RGB experiment on ten clean images (Kodak 02–08 [3] and CBSD68 0001–0003 [4]). Noise is sampled independently of the clean images from the pretrained RGB-MNIST diffusion model released by Stevens et al. [5]. The full-resolution noise images are assembled from a library of 64×6464\times64 patches.

We adopt the collaborative color total variation (CTV) prior T=TV∞,1,1T=\mathrm{TV}_{\infty,1,1} proposed in [6,7] as the image prior, which has been evaluated on the Kodak collection for various image restoration tasks [6]. Specifically, the prior is defined as

\max_{c\in\{R,G,B\}}|\Delta_d x_c(p)|.$$ It couples gradients across color channels and encourages shared edges. RED-TV and Gaussian+TV [6,7] use the same CTV prior and differ only in their data-fidelity terms. Specifically, Gaussian+TV uses the quadratic fidelity $D(r)=\|r\|_2^2/2$, whereas RED-TV uses the learned effective-noise fidelity. The denoising results are summarized in the following table. Full image denoising results are provided in Appendix E. RED-TV improves PSNR on all ten images, with an **average gain of 7.51 dB**. The results show that our effective-likelihood MAP framework also performs well in image denoising. **This also demonstrates the generality of the proposed framework.** | Metric | Observation | Gaussian+TV | RED-TV | |:---|---:|---:|---:| | PSNR (dB) | 18.89 | 18.91 | 26.42 | --- **References for Comment 1:** [1] Abdelhamed, A., Lin, S., and Brown, M. S. (2018). A High-Quality Denoising Dataset for Smartphone Cameras. *CVPR*. [2] Zhang, Y. et al. (2019). A Poisson-Gaussian Denoising Dataset With Real Fluorescence Microscopy Images. *CVPR*. [3] Kodak. Kodak Lossless True Color Image Suite. https://r0k.us/graphics/kodak/ [4] Martin, D., Fowlkes, C., Tal, D., and Malik, J. (2001). A Database of Human Segmented Natural Images and its Application to Evaluating Segmentation Algorithms and Measuring Ecological Statistics. *ICCV*, 416–423. CBSD68 dataset: https://github.com/clausmichele/CBSD68-dataset [5] Stevens, T. S. W. et al. (2025). Removing Structured Noise using Diffusion Models. *Transactions on Machine Learning Research*. Paper: https://openreview.net/forum?id=BvKYsaOVEn. Code: https://github.com/tristan-deep/joint-diffusion [6] Duran, J., Moeller, M., Sbert, C., and Cremers, D. (2016). Collaborative Total Variation: A General Framework for Vectorial TV Models. *SIAM Journal on Imaging Sciences*, 9(1), 116–151. https://doi.org/10.1137/15M102873X [7] Duran, J., Moeller, M., Sbert, C., and Cremers, D. (2016). On the Implementation of Collaborative TV Regularization: Application to Cartoon+Texture Decomposition. *Image Processing On Line*, 6, 27–74. https://doi.org/10.5201/ipol.2016.141

Comment 2: The methodological novelty appears limited. RED/RED-diff and primal–dual splitting are established techniques. The main algorithmic contribution combines a diffusion-based noise model with a classical atomic-norm signal prior, which remains unchanged. Although the noise-domain variational derivation provides useful justification, the paper does not sufficiently demonstrate a substantial methodological advance beyond this combination of existing components.


Response:

Our main contribution is the deterministic optimization framework in the noise domain and its theoretical justification.

First, our method is not a diffusion-based posterior-sampling method. We learn the noise score from noise-only samples, construct an effective noise likelihood, and combine it with an explicitly specified signal prior to obtain a deterministic structured estimator.

Second, we rigorously derive the zero-dispersion limiting noise-domain objective from the finite-dispersion RED-diff formulation. In RED-diff [1], the authors simplify the objective by setting δ=0\delta=0. At this point, the variational Gaussian distribution degenerates into a Dirac measure, whose KL divergence with respect to a continuous density is infinite. Our theoretical analysis rigorously establishes this limiting procedure, rather than treating the resulting optimization problem as a heuristic.

Furthermore, we prove that the resulting optimization problem is a MAP estimator under the effective likelihood, providing a clear statistical interpretation of the resulting optimization problem.

Third, the proposed framework is general. It can be applied to denoising in general linear inverse problems with an explicitly specified signal prior, rather than being restricted to gridless line spectral denoising. For example, it can be applied to denoising problems in compressed sensing [2] and low-rank matrix recovery [3].

Finally, we show how the proposed denoising framework can be combined with a spectral-sparsity prior and applied to gridless line spectral denoising through RED–Atomic. The primal–dual splitting method is only one of the algorithms for solving this problem. Therefore, the application of primal-dual splitting is merely an algorithmic choice, rather than the core contribution of this paper.

We hope this clarification helps the reviewer better understand the main contributions of our work.


References for Comment 2:

[1] Mardani, M., Song, J., Kautz, J., and Vahdat, A. (2024). A Variational Perspective on Solving Inverse Problems with Diffusion Models. ICLR 2024.

[2] Baraniuk, R. G., Cevher, V., Duarte, M. F., and Hegde, C. (2010). Model-Based Compressive Sensing. IEEE Transactions on Information Theory, 56(4), 1982–2001.

[3] Candès, E. J., and Recht, B. (2009). Exact Matrix Completion via Convex Optimization. Foundations of Computational Mathematics, 9, 717–772.

Comment 3: The diffusion-based noise fidelity can be nonconvex, raising concerns about sensitivity to initialization and step sizes. Appendix B.4 only characterizes the limiting fixed point if the iterates converge; it does not establish convergence for the general learned-score implementation. A nonconservative learned score field further weakens the connection between this fixed point and a MAP solution.


Response: We thank the reviewer for raising these concerns. We have conducted additional experiments and clarified the scope and generality of the proposed framework.

1. Initialization and step-size sensitivity.

We have added initialization and step-size experiments for RED–Atomic under nonconvex Bernoulli–Gaussian noise. We evaluate five initializations and three primal step-size factors (0.50, 0.75, and 1.00) at 5 and 10 dB. The maximum relative variations in frequency RMSE are summarized as follows.

Experiment5 dB10 dB
Five initializations0.002%0.010%
Three step-size factors (900 iterations)0.105%0.125%

The RMSE differences are negligible across the tested settings. Smaller step sizes, however, require more iterations to satisfy the stopping criteria. After extending the iteration budget to 2,000, the recovery accuracy becomes comparable across the tested step sizes.

2. Generality of the effective MAP framework and optimization algorithms.

Appendix B discusses the convergence guarantees of the primal–dual algorithm under standard convexity and step-size assumptions. These guarantees do not generally extend to nonconvex effective fidelities.

However, the proposed effective MAP framework can be applied to general additive-noise inverse problems with explicitly specified signal priors.

For example, in compressed sensing, an ℓ1\ell_1-norm prior gives

min⁡x Uπ(y−Ax)+λ∣x∣1.\min_x\ U_\pi(y-Ax)+\lambda|x|_1.

FISTA [1] and ADMM [2] can be applied for solving this problem. However, their convergence guarantees hold when UπU_\pi is convex. In nonconvex cases, their general convergence guarantees no longer follow without additional assumptions [3],[4].

In fact, establishing convergence guarantees for nonconvex optimization is inherently challenging, and existing results typically rely on stringent structural and algorithm-specific assumptions [5]. Moreover, sensitivity to initialization and step sizes is a well-recognized challenge in nonconvex optimization [6]. Therefore, the absence of general convergence guarantees, as well as the sensitivity of RED–Atomic to initialization and step sizes for arbitrary nonconvex problems is a broader challenge in the field rather than a limitation specific to our framework.

3. Estimation error bounds and scope of this work.

Appendix C characterizes the equilibrium conditions of the general learned-field implementation. We further establish an objective-gap bound when the effective likelihood is convex and a solution-distance bound under strong convexity. However, for nonconvex effective fidelities, no corresponding estimation-error guarantees can be established without additional structural assumptions.

Finally, our primary contribution is the effective MAP formulation and its integration with explicit priors, rather than a universal robust nonconvex solver. Developing optimization algorithms with stronger convergence guarantees for specific nonconvex problems is an important but separate research direction.

We hope that the additional experiments and clarifications adequately address the concerns raised and demonstrate the value of our work. Thank you again for your time and consideration.


References

[1] Amir Beck and Marc Teboulle. A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse Problems. SIAM Journal on Imaging Sciences, 2(1):183–202, 2009.

[2] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan Eckstein. Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers. Foundations and Trends in Machine Learning, 3(1):1–122, 2011.

[3] Hédy Attouch, Jérôme Bolte, and Benar Fux Svaiter. Convergence of Descent Methods for Semi-Algebraic and Tame Problems: Proximal Algorithms, Forward–Backward Splitting, and Regularized Gauss–Seidel Methods. Mathematical Programming, 137(1–2):91–129, 2013.

[4] Mingyi Hong, Zhi-Quan Luo, and Meisam Razaviyayn. Convergence Analysis of Alternating Direction Method of Multipliers for a Family of Nonconvex Problems. SIAM Journal on Optimization, 26(1):337–364, 2016.

[5] Jérôme Bolte, Shoham Sabach, and Marc Teboulle. Proximal Alternating Linearized Minimization for Nonconvex and Nonsmooth Problems. Mathematical Programming, 146(1–2):459–494, 2014.

[6] Rina Foygel Barber and Wooseok Ha. Gradient Descent with Non-Convex Constraints: Local Concavity Determines Convergence. Information and Inference: A Journal of the IMA, 7(4):755–806, 2018.

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

This paper studies signal denoising with an explicit signal prior and an unknown noise distribution learned from noise samples. It adapts RED-diff to the noise domain and derives a limiting objective with a MAP interpretation under an effective noise likelihood, subject to the stated assumptions. For gridless line spectral estimation, RED–Atomic combines this formulation with an atomic norm and uses primal–dual splitting. Experiments compare recovery performance across noise distributions and examine the use of multiple diffusion scales.

Critical strengths
  1. The framework learns the noise distribution with a diffusion model and retains an explicit signal prior for denoising under unknown non-Gaussian noise.
  2. RED–Atomic combines learned noise information with gridless spectral estimation. Experiments show better recovery across several non-Gaussian noise models.
Critical weaknesses
  1. Although Section 4.1 provides a theoretical motivation for inverse-SNR weighting, the experiments in Section 6.2 (Figure 2) only compare the proposed approach using eight diffusion scales with inverse-SNR weighting against a baseline using only one diffusion scale. This comparison simultaneously changes the number of diffusion scales and the weighting strategy, making it difficult to isolate the contribution of inverse-SNR weighting itself.
  2. The equivalence in Eq. (11) relies on identical terminal distributions, but the paper does not explain how this condition is reconciled with its finite-time VP process. In general, the identity includes an additional terminal KL term, DKL(qT∥pT)D_{\mathrm{KL}}(q_T\|p_T), which can depend on the variational mean and cannot automatically be discarded as an optimization-independent constant. The schedule in Appendix B.2 retains a nonzero αT\alpha_T.
Questions for authors
  1. Do you have results comparing inverse-SNR and uniform weights over the same eight scales, with the trained network and test observations held fixed?
  2. How is the condition qT=pTq_T = p_T in Eq. (11) reconciled with the finite-time VP schedule used in the experiments?
  3. Could you check the manuscript for repeated text and citations, including the Stevens et al. (2025) description in Lines 93–96 and the Song et al. (2023) citation in Lines 48–53?
1 comment

Response to Reviewer SYtH

General Response:

We sincerely appreciate the reviewer’s time and effort in evaluating our manuscript, the constructive feedback, and the favorable assessment of our work. We are pleased that the reviewer recognizes the contributions of this paper. We have carefully addressed the comments and revised the manuscript accordingly. We believe that these revisions have resolved the concerns raised and further improved the theoretical rigor, experimental evidence, and overall presentation of the paper. Our detailed responses are provided below.


Question 1: Do you have results comparing inverse-SNR and uniform weights over the same eight scales, with the trained network and test observations held fixed?

Response: We appreciate the reviewer’s insightful suggestion. We agree that the original comparison did not clearly distinguish the benefits of multiscale aggregation from those of inverse-SNR weighting. To address this concern, we have revised Section 6.2 and added a controlled ablation experiment in Figure 2, comparing three configurations under the same Bernoulli–Gaussian noise model:

  1. Single-scale: the original single-scale objective at t=0.002t = 0.002.
  2. Uniform-8: an eight-scale objective with uniform weights.
  3. Inverse-SNR-8: an eight-scale objective with inverse-SNR weights.

The two multiscale configurations use exactly the same eight geometrically spaced diffusion scales, the same trained score network, and the same test observations; they differ only in their weighting strategies. Specifically, the scales are fixed as τj∈geomspace⁡(0.01,0.6,8)\tau_j \in \operatorname{geomspace}(0.01, 0.6, 8), with normalized weights wj=1/8w_j = 1/8 for uniform weighting and wj∝τj2w_j \propto \tau_j^2 for inverse-SNR weighting.

The results demonstrate that inverse-SNR weighting consistently outperforms uniform weighting. For Ntrain=1,000N_{\mathrm{train}} = 1{,}000, the frequency RMSE decreases:

  • At 5 dB, from 1.593 × 10⁻³ to 0.997 × 10⁻³, a 37.4% reduction.
  • At 10 dB, from 0.891 × 10⁻³ to 0.566 × 10⁻³, a 36.5% reduction.

Furthermore, comparing the single-scale and uniform multiscale configurations demonstrates the benefit of multiscale aggregation itself. The frequency RMSE decreases from 2.559 × 10⁻³ to 1.593 × 10⁻³ at 5 dB, and from 0.956 × 10⁻³ to 0.891 × 10⁻³ at 10 dB. The same qualitative patterns are observed when Ntrain=5,000N_{\mathrm{train}} = 5{,}000.

These controlled comparisons separate the contributions of multiscale aggregation and inverse-SNR weighting. The former improves upon a single diffusion scale, while the latter provides additional gains over uniform weighting on the same multiscale support.

We believe that this additional experiment provides a clearer and more rigorous validation of the proposed weighting strategy.


Question 2: How is the condition qT=pTq_T = p_T in Eq. (11) reconciled with the finite-time VP schedule used in the experiments?

Response: We are grateful to the reviewer for identifying this mathematical issue. The reviewer is correct that, under a finite-time VP diffusion schedule with nonzero αT\alpha_T, the terminal distributions qTq_T and pTp_T are not necessarily identical. Consequently, the terminal KL term cannot generally be assumed to vanish.

We have corrected Eq. (11) by removing the restrictive assumption qT=pTq_T = p_T and retaining the terminal KL term:

+ \int_0^T \frac{\beta(t)}{2}\omega(t)I_t\\,dt.$$ The previous expression is recovered as a special case when $q_T = p_T$. This correction improves the mathematical rigor of the general decomposition. It is worth noting that **the correction does not affect the subsequent optimization formulation, theoretical results and numerical experiments.** We thank the reviewer for pointing out this issue, which has helped us improve the mathematical precision of the manuscript. --- **Question 3:** Could you check the manuscript for repeated text and citations, including the Stevens et al. (2025) description in Lines 93–96 and the Song et al. (2023) citation in Lines 48–53? **Response:** We thank the reviewer for carefully checking the manuscript and identifying these editorial errors. We have corrected both issues: 1. **Repeated text in Related Work:** We removed the duplicated description of Stevens et al. (2025) in Section 2. 2. **Repeated citation in the Introduction:** We corrected the duplicate citation to Song et al. (2023) in Section 1. We have also checked the manuscript for other repeated sentences and duplicate citations to improve its overall consistency and readability. Once again, we appreciate the reviewer’s constructive feedback, which has helped us further strengthen the manuscript.
Reviewer_cf15OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a noise-domain variational framework for denoising when the signal structure is explicitly known but the non-Gaussian noise distribution must be learned from samples. Starting from a finite-dispersion RED-diff formulation, the authors derive a zero-dispersion objective and establish its MAP interpretation under a normalized effective noise likelihood. The framework is instantiated as RED–Atomic for gridless line spectral estimation, combining a learned diffusion-based fidelity with atomic-norm regularization. A primal–dual implementation and experiments under several noise distributions are provided.

Critical strengths
  1. The paper is well-organized, logically progressing from motivation to application while clearly distinguishing explicit signal priors from learned noise models.
  2. The theoretical development provides a coherent foundation for the population objective, with Theorems 2 and 3 clarifying the statistical meaning of the proposed fidelity through normalization and Gaussian characterization.
Critical weaknesses
  1. The main text should explicitly clarify the connection between Theorem 2's exact-density MAP objective and the practical framework, acknowledging that the unconstrained score network may yield a learned-field equilibrium as noted in Appendix B.4.
  2. While Eq. (12) motivates inverse-SNR weighting via reduced score-error sensitivity, Table 5 shows that altering the effective likelihood can degrade estimation accuracy even under exact scores, underscoring the need for clear practical guidance on selecting weighting functions and smoothing ranges.
  3. While the experiments cover diverse noise distributions, the evaluation is limited to a narrow signal regime (M=32,K=3M=32, K=3 with frequency separation (4/M4/M), leaving performance across varying signal lengths, spectral complexities, and tighter separations unverified.
Questions for authors
  1. Could the authors incorporate the Appendix B.4 distinction into the main text and clarify how the practical learned field iteration approximates the population MAP formulation in Theorem 2?
  2. How were the default weighting function and smoothing range selected, and given Table 5's exact-score results, what principled guidance exists for balancing score-error sensitivity against effective-likelihood estimation bias?
  3. Do the experiments extend beyond the fixed M=32,K=3M=32, K=3 setting, and if not, what are the expected limitations under alternative signal lengths, component counts, or tighter frequency separations?

Comments on the paper

6 comments

Response to Reviewer cf15

General Response

We sincerely thank the reviewer for the positive assessment of our manuscript and the constructive suggestions. We have addressed the comments through additional theoretical clarification and experiments, and hope these revisions help the reviewer better understand the contribution and will be taken into account when reconsidering the overall rating.


Comment 1: Could the authors incorporate the Appendix B.4 distinction into the main text and clarify how the practical learned field iteration approximates the population MAP formulation in Theorem 2?

Response: We thank the reviewer for this helpful suggestion. We have incorporated the key discussion into Section 5.2 and expanded the theoretical analysis in Appendix C.

We clarify that the implemented primal-dual algorithm is a learned-field approximation to the population MAP formulation, not an exact MAP solver. Specifically, we distinguish three cases based on the learned field:

  1. Exact effective-likelihood gradient: The limiting fixed-point conditions reduce to the KKT conditions of the population MAP problem (Theorem 2).
  2. Conservative learned field: The conditions characterize the KKT conditions of the associated learned-fidelity optimization problem.
  3. General non-conservative learned field: The conditions characterize an equilibrium between the learned field and the PSD cone constraint.

We have further quantified the approximation error between the practical learned field and the population MAP formulation in Appendix C.

  • When UπU_\pi is convex and differentiable, both the first-order stationarity residual and the objective gap can be bounded by the learned-field approximation error επ=∥∣g^π−∇Uπ∥∣2.\varepsilon_\pi = \||\widehat g_\pi - \nabla U_\pi\||_2., where g^π\widehat g_\pi is the learned field.

  • When UπU_\pi is mm-strongly convex, the estimation error bound ∥∣x^−x⋆∥∣2≤επm.\||\widehat x - x^\star\||_2 \leq \frac{\varepsilon_{\pi}}{m}. can be further established.

  • For non-convex likelihoods, global optimization guarantees cannot generally be obtained.


Comment 2:: How were the default weighting function and smoothing range selected, and given Table 5's exact-score results, what principled guidance exists for balancing score-error sensitivity against effective-likelihood estimation bias?

Response: We thank the reviewer for this important question. The smoothing interval involves a trade-off between score-estimation error and effective-likelihood distortion. To investigate how the smoothing interval should be selected in practice, we conducted extensive experiments across four noise distributions, varying the training sample sizes and both interval endpoints. The results are reported in Appendix D. Our analysis leads to the following findings.

  1. Dependence on noise distribution and sample size. The preferred smoothing interval depends on both the noise distribution and the available training sample size. In particular, the upper endpoint should be selected according to these two factors, whereas the lower endpoint is considerably more stable. This is because inverse-SNR weighting assigns very little weight to the smallest diffusion scales, making the choice of $\tau_{\min}$ relatively insensitive to both factors.

  2. Structured noise. The low-dimensional structure of structured noise distributions allows the learned score field to be approximated accurately with relatively few training samples. Consequently, the preferred smoothing interval is relatively insensitive to training sample size. A conservative smoothing interval is therefore sufficient.

  3. Multi-state noise. For multi-state noise distributions, score estimation improves substantially as the training sample size increases, accompanied by significant improvements in frequency recovery. Accordingly, the preferred upper endpoint shifts toward smaller values as the training sample size increases.

  4. Practical selection. We recommend fixing τmin⁡\tau_{\min} at a small positive value and selecting τmax⁡\tau_{\max} according to the noise distribution and available training data. Based on the experimental results in Appendix D, the range 0.6-1.0 provides a useful starting point, with the upper endpoint reassessed as the training sample size increases.

We hope that these additional analyses provide useful guidance and address the reviewer's concern regarding the selection of the smoothing interval.

Comment 3: Do the experiments extend beyond the fixed M=32,K=3M=32, K=3 setting, and if not, what are the expected limitations under alternative signal lengths, component counts, or tighter frequency separations?

Response: We thank the reviewer for this helpful suggestion. We have conducted additional experiments and reported the results in Appendix B.5. We consider the following eight signal configurations.

  • M∈{32,64}M\in\lbrace 32,64\rbrace.
  • K∈{3,5}K\in \lbrace 3,5 \rbrace.
  • Minimum circular frequency separation: Δmin⁡∈{2/M,0.5/M}\Delta_{\min}\in \lbrace 2/M,0.5/M \rbrace.

For each configuration, we evaluate three noise distributions at five observation SNR levels (0, 5, 10, 15, and 20 dB). This results in a total of 120 evaluation cases.

RED–Atomic achieves the lowest pooled frequency RMSE in 91 out of the 120 evaluation cases. These results show that RED–Atomic maintains competitive performance across different signal lengths, spectral complexities, and frequency resolutions.

We have also further discussed the fundamental factors affecting recovery difficulty. First, the number of spectral components affects the identifiability and difficulty of the recovery problem. In the standard line spectral estimation setting, the usual sampling requirement for identifying KK spectral components from MM consecutive samples is M>2KM>2K [1]. The increase of KK will lead to greater spectral overlap and making frequency localization more challenging.

Moreover, the minimum frequency separation is closely related to the resolution capability of atomic-norm-based methods. Classical super-resolution theory establishes minimum-separation conditions on the order of O(1/M)O(1/M) for exact recovery guarantees [2]. When two frequencies become too close, neighboring atoms become increasingly coherent, making the recovery problem more ill-conditioned. The same trend is observed in our experiments.

We have directed readers to this additional evaluation in Section 6.1. We believe that this extension provides a more comprehensive assessment of the robustness and limitations of RED–Atomic beyond the default signal configuration.


Once again, we sincerely thank the reviewer for these constructive suggestions. We hope that our response has adequately addressed your concern.


References for Comment 3

[1] Zai Yang, Jian Li, Petre Stoica, and Lihua Xie. Sparse Methods for Direction-of-Arrival Estimation. In Academic Press Library in Signal Processing, Volume 7: Array, Radar and Communications Engineering, Chapter 11, pp. 509–581. Academic Press, 2018.

[2] Emmanuel J. Candès and Carlos Fernandez-Granda. Towards a Mathematical Theory of Super-resolution. Communications on Pure and Applied Mathematics, 67(6):906–956, 2014.

Reviewer_cf15OpenReview ↗

Comments

First, thank you for the authors’ detailed response. Regarding my first question, I consider it well addressed. For the second question, the response about the smoothing range is satisfactory, but the explanation on how to select the weights is somewhat inadequate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.