acceptodds
Under review as a conference paper at ICLR 2027

A statistical theory for blind denoising: minimax estimation of the noise level from a single sample

Abstract

Motivated by blind denoising in diffusion models, we study estimation of an unknown Gaussian noise level from a single high-dimensional observation, assuming the signal law is known. We characterize the minimax mean-squared error under two structural assumptions on . For signals with covering complexity , the minimax rate is, up to logarithmic factors, , and the constrained MLE attains this rate. For -strongly log-concave signals, the rate is , attained up to constants by a -centered norm estimator. These results show that the structure of the signal law determines both the difficulty of blind noise estimation and the appropriate estimator.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
noise level estimation, minimax estimation, blind denoising, diffusion models, high-dimensional statistics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_r3toOpenReview ↗
Rating
2(+1)
Confidence
2(−1)
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper characterizes the minimax risk of estimating the noise level σ\sigma (or rather its precision λ=σ−2\lambda = \sigma^{-2}) from noisy images corrupted by additive white Gaussian noise. Two theorems are given depending on which assumptions are made on PP, the signal law (low intrinsic dimensionality or strong log-concavity). This paper is exclusively theoretical.

Critical strengths

The paper is mathematically well-founded, although I did not check all the proofs in Appendix. The literature review is well conducted.

Critical weaknesses

I recommend rejecting this paper because it is entirely theoretical and its conclusions are unlikely to provide any practical value to the research community.

Major:

  1. The paper is exclusively theoretical. Its goal is only to give minimax rates under two different assumptions on the signal distribution. The conclusion is that "the difficulty of blind noise estimation is governed by the structure of the signal law", which is an obvious conclusion to me. It is hard to envision any application for this paper. I do not think it will be useful to the community.

  2. The authors motivate their work in the context of blind denoising in diffusion models where blind denoising can improve image generation quality despite the missing noise input. However, not a single example or concrete application is given with diffusion models.

Minor:

  1. The sup⁡λ∈Λ\sup_{\lambda \in \Lambda} notation on line 199 initially confused me. Indeed, in machine learning, and particularly in the denoising literature, it is quite common to use the subscript of the expectation to indicate the random variables with respect to which the expectation is taken. Here however following the traditional minimax estimation framework, I understand that λ\lambda is a parameter rather than a random variable. I suggest clarifying this point to avoid any potential confusion for the reader.

  2. Please number each of your equations.

  3. Some references about single-image noise estimation are missing [a, b].

[a] Alessandro Foi, Mejdi Trimeche, Vladimir Katkovnik, and Karen Egiazarian. Practical Poissonian-Gaussian noise modeling and fitting for single-image raw-data. IEEE TIP, 2008.

[b] Stanislav Pyatykh, Jurgen Hesser, and Lei Zheng. Image noise level estimation by principal component analysis. IEEE TIP, 2013.

Questions for authors

In line 199, I have the intuition that λmin=arg⁡max⁡λ∈ΛEλ,P[(λ^P(Y)−λ)2]\lambda_{min} = \arg \max_{\lambda \in \Lambda} \mathbb{E}_{\lambda, P}[ (\hat{\lambda}_P(Y) - \lambda)^2]

so that R(P;Λ)\mathcal{R}(P; \Lambda) reads inf⁡λ^Eλ,P[(λ^P(Y)−λm)2]\inf_{\hat{\lambda}} \mathbb{E}_{\lambda, P}[(\hat{\lambda}_P(Y) - \lambda_m)^2]. Indeed, the more the strength of the noise is and the more difficult it will be to estimate the noise level. Do you agree with me?

3 comments
AuthorsOpenReview ↗

Response to Reviewer r3to

We thank the reviewer for carefully reading our paper and for the useful suggestions, particularly on practical relevance, presentation, and related literature. We address the concerns below.

Q1. Novelty and theoretical significance. Our contribution goes beyond the qualitative observation that signal structure matters. We characterize the optimal quantitative dependence of noise-estimation risk on signal structure.

Under low intrinsic dimensionality, the k2/d2k^2/d^2 term can dominate the usual 1/d1/d rate. Our lower-bound construction uses the same known PP at two noise levels, showing that latent signal variation can conceal a precision change of order k/dk/d. The MLE matches the lower bound up to logarithmic factors.

Under strong log-concavity, the centered norm estimator attains the minimax rate

\min\left\\{\Delta_\Lambda^2,\frac{(1+\alpha^{-1})^2}{d}\right\\},

with sharp dependence on α\alpha for fixed Λ\Lambda. This identifies a different estimation mechanism based on concentrated signal energy rather than low-dimensional support.

We will emphasize these quantitative insights and the same-PP construction in Sections 1.1 and 4.2.

Q2. Practical relevance and diffusion models. We appreciate the concern about experimental validation. Kadkhodaie et al. (2026) already provide extensive experiments demonstrating successful blind denoising in diffusion models. Their analysis relates this success to concentration of the posterior over the missing noise level, which allows blind and noise-conditioned denoisers to be close.

Our paper complements these experiments by showing that the corresponding low-dimensional noise-estimation rate is minimax optimal up to logarithmic factors. We aim to explain the statistical mechanism underlying blind denoising, rather than propose another practical algorithm. Our optimality result concerns noise estimation; it does not establish a lower bound on denoising or diffusion-sampling error.

We will clarify this connection in Sections 1.1 and 2.1 and explain the scope of our conclusions in Section 5.

Q3. Notation and presentation. We agree that the expectation notation could be clearer. The subscript in Eλ,P\mathbb E_{\lambda,P} specifies the observation law, with λ\lambda treated as a fixed unknown parameter:

Eλ,P[(λ^P(Y)−λ)2]=∫(λ^P(y)−λ)2pλ(y),dy.\mathbb E_{\lambda,P}[(\widehat\lambda_P(Y)-\lambda)^2]=\int(\widehat\lambda_P(y)-\lambda)^2p_\lambda(y)\\,dy.

The estimator may use PP, and the supremum over λ∈Λ\lambda\in\Lambda gives the worst-case risk. We will clarify this notation in Section 2.2 and number the displayed equations throughout the paper.

Q4. Missing references. We thank the reviewer for pointing out these references. We will incorporate both into the paper and clarify their relationship to our work.

All three works study noise estimation from a single noisy observation by exploiting signal structure. Foi et al. (2008) fit a signal-dependent Poissonian--Gaussian noise model to raw images, accounting for effects such as clipping. Pyatykh et al. (2013) estimate additive Gaussian noise using PCA of image patches, exploiting low-variance directions to distinguish noise from image content.

The difference lies in the goal and statistical setting. These papers develop practical image-based estimators, whereas we assume a known signal law PP and characterize minimax estimation risk under explicit structural assumptions. Our work provides a complementary theoretical account of how signal structure affects noise-estimation difficulty.

We will add both references to Section 1.2 and discuss their similarities to our work, as well as the differences in noise models, assumptions, and theoretical goals.

Q5. Dependence on noise strength. We appreciate this question. The dependence on noise strength depends on the parameterization and loss. We estimate precision λ=σ−2\lambda=\sigma^{-2}, so stronger noise corresponds to smaller λ\lambda.

For P=δ0P=\delta_0, let a=λ−1a=\lambda^{-1} and consider

a^=∣Y∣2d,λ~=d−2∣Y∣2.\widehat a=\frac{\\|Y\\|^2}{d},\qquad \widetilde\lambda=\frac{d-2}{\\|Y\\|^2}.

For d>4d>4, their exact mean-squared errors are

E(a^−a)2=2a2d,E(λ~−λ)2=2λ2d−4.\mathbb E(\widehat a-a)^2=\frac{2a^2}{d},\qquad \mathbb E(\widetilde\lambda-\lambda)^2=\frac{2\lambda^2}{d-4}.

Thus, stronger noise increases absolute variance-estimation error but decreases absolute precision-estimation error in this example. Both relative errors remain of order 1/d1/d. In particular, the worst-case precision risk need not occur at λmin⁡\lambda_{\min}.

Our minimax formulation uses squared precision loss over a fixed interval Λ\Lambda. Rate equivalence with standard-deviation estimation involves Λ\Lambda-dependent constants and does not imply identical pointwise dependence on noise strength.

We will clarify the parameterization and add this Gaussian example in Section 2.2.

I thank the authors for their response. In particular, I find the toy example in Q5 interesting and recommend adding it to the paper.

My initial inclination was to reject the paper, as I was skeptical that its conclusions would offer any practical value to the community. However, after carefully considering the authors' responses and the comments from the other reviewers, I have decided to slightly increase my overall rating. At the same time, I have lowered my confidence in my assessment, given my limited expertise in minimax estimation.

We thank the reviewer for reconsidering our responses. The Gaussian example discussed in Q5 has been added to Section 2.2 (Equation 13) of the revised manuscript.

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

This paper analyzes why "blind denoising" (denoising without the noise intensity as an input) is possible in diffusion models by framing the issue as a problem of pure statistical estimation. Specifically, it examines estimation efficiency in the scenario where one estimates the precision parameter λ\lambda from a single dd-dimensional observation YY—formed by adding Gaussian noise with unknown variance 1/λ1/\lambda to a signal XX generated from a known dd-dimensional distribution PP. The analysis assumes that PP is known and the noise is isotropic, with λ\lambda being the only unknown parameter. Under this setup, the paper identifies the structural properties of PP that enable blind denoising and evaluates the minimax rate of the Mean Squared Error (MSE) for the optimal estimator.

Critical strengths
  1. It demonstrates—through multiple examples—that high dimensionality alone is insufficient for blind noise estimation, and that the structure of the signal distribution PP is what matters.
  2. The mathematical techniques employed to prove each case are ingenious.
Critical weaknesses
  1. Although the relationship to diffusion models is cited as a motivation, what is actually proven has little to do with them. The results rely on strong assumptions: that PP is known and the noise follows an isotropic, zero-mean Gaussian distribution.
  2. The analysis applies to scenarios where the correct likelihood is obtainable, remaining within the scope of classical parameter estimation problems for known distribution families.
  3. Even though it involves a single sample, the components are i.i.d. with respect to dimensionality, so the fact that the MSE scales as O(1/d)O(1/d) is not particularly surprising.
Questions for authors
  1. Practical relevance of the known-PP assumption.
    The analysis assumes that the signal distribution PP is exactly known, so that the likelihood pλ(y)p_\lambda(y) is, at least in principle, fully specified by Gaussian convolution. In this setting, noise-level estimation reduces to a one-dimensional parametric estimation problem over a known family {Pλ}λ∈Λ\{P_\lambda\}_{\lambda\in\Lambda}. Could the authors clarify what aspect of the difficulty of blind denoising in learned diffusion models is captured by this oracle setting? In particular, how would the conclusions be expected to change when PP is only available through finite training samples or an approximate learned model?

  2. Necessity of strong log-concavity.
    The strong log-concavity result is presented as showing that low intrinsic dimensionality is not necessary for accurate blind noise estimation. However, the proposed norm estimator appears to rely more directly on concentration of the radial energy, in particular on controlling Var⁡P(∥X−μP∥2),\operatorname{Var}_P(\|X-\mu_P\|^2), while strong log-concavity is a substantially stronger sufficient condition. Can Theorem 3.2 be generalized under a weaker moment/concentration assumption directly controlling radial fluctuations? If so, would strong log-concavity still represent the appropriate structural message of the theorem?

  3. What is learned about blind denoising beyond noise-level identifiability?
    The paper establishes minimax rates for estimating the scalar noise level, but a blind denoiser need not explicitly estimate the noise level; it may directly learn a map from the noisy observation to the denoised signal. Is accurate estimation of λ\lambda necessary, rather than merely sufficient, for successful blind denoising? Can the authors provide a formal connection showing that the minimax limits derived here impose corresponding limitations on blind denoising or blind diffusion sampling?

2 comments
AuthorsOpenReview ↗

Response to Reviewer EziC

We thank the reviewer for carefully reading our paper and for the thoughtful questions about our assumptions and the connection to blind denoising. We address the three questions below.

Q1. Practical relevance of the known-PP assumption. We agree that this assumption deserves a clearer interpretation. It models an idealized diffusion setting in which abundant training data allow a learned blind denoiser to approach the population minimizer, which depends on PP. Our paper isolates the inference-time question: once the model is trained, can the missing noise level be inferred from one noisy state? Finite-data learning and network approximation introduce additional errors that our oracle analysis does not address.

Moreover, the dd coordinates of a noisy state need not be independent under a general PP; our k2/d2k^2/d^2 lower bound shows why a 1/d1/d rate is not universal. We will explain the population-minimizer interpretation in Section 2.1 and clarify the limits of the oracle model in Section 5.

Q2. Necessity of strong log-concavity. We agree that strong log-concavity is stronger than needed for the norm-estimator upper bound. The proof uses control of the signal's radial moments. Define

μP=EPX,m2(P)=EP∣X−μP∣2,v2(P)=Var⁡P(∣X−μP∣2).\mu_P=\mathbb E_P X,\quad m_2(P)=\mathbb E_P\\|X-\mu_P\\|^2,\quad v_2(P)=\operatorname{Var}_P(\\|X-\mu_P\\|^2).

For a=λ−1a=\lambda^{-1}, the corrected energy statistic

a~P=∣Y−μP∣2−m2(P)d\widetilde a_P=\frac{\\|Y-\mu_P\\|^2-m_2(P)}{d}

is unbiased for aa, and Lemma A.3 yields

Eλ,P(a~P−a)2=v2(P)+4a,m2(P)+2a2dd2.\mathbb E_{\lambda,P}(\widetilde a_P-a)^2=\frac{v_2(P)+4a\\,m_2(P)+2a^2d}{d^2}.

If m2(P)≤Adm_2(P)\le Ad and v2(P)≤CA2dv_2(P)\le CA^2d, projection onto the admissible variance interval followed by inversion gives

\sup_{\lambda\in\Lambda}\mathbb E_{\lambda,P}(\widehat\lambda_{\mathrm{norm},P}-\lambda)^2\lesssim_{\Lambda,C}\min\left\\{\Delta_\Lambda^2,\frac{(1+A)^2}{d}\right\\}.

Strong log-concavity implies these bounds with A=α−1A=\alpha^{-1} and C=4C=4. The underlying sufficient condition is control of radial fluctuations, not strong log-concavity itself. This observation concerns the norm estimator, not the MLE guarantee. We will clarify it after Theorem 3.2 without adding a separate theorem.

Q3. Noise-level estimation versus blind denoising. We agree that noise-estimation lower bounds do not automatically imply denoising lower bounds. Accurate noise identification is not universally necessary: if P=δ0P=\delta_0, the map g(Y)=0g(Y)=0 denoises perfectly at every noise level, although noise estimation still has positive risk.

For the noise-conditioned Bayes denoiser fλ,P(y)=Eλ,P[X∣Y=y]f_{\lambda,P}(y)=\mathbb E_{\lambda,P}[X\mid Y=y], the excess risk of any blind map gg satisfies

Eλ,P(g)=Eλ,P∣g(Y)−X∣2−Eλ,P∣fλ,P(Y)−X∣2=Eλ,P∣g(Y)−fλ,P(Y)∣2.\mathcal E_{\lambda,P}(g)=\mathbb E_{\lambda,P}\\|g(Y)-X\\|^2-\mathbb E_{\lambda,P}\\|f_{\lambda,P}(Y)-X\\|^2=\mathbb E_{\lambda,P}\\|g(Y)-f_{\lambda,P}(Y)\\|^2.

For two precisions with the same PP, writing pi=pλip_i=p_{\lambda_i} and fi=fλi,Pf_i=f_{\lambda_i,P} gives

inf⁡gmax⁡i∈0,1Eλi,P(g)≥12∫p0(y)p1(y)p0(y)+p1(y)∣f0(y)−f1(y)∣2,dy.\inf_g\max_{i\in\\{0,1\\}}\mathcal E_{\lambda_i,P}(g)\ge\frac12\int\frac{p_0(y)p_1(y)}{p_0(y)+p_1(y)}\\|f_0(y)-f_1(y)\\|^2\\,dy.

Thus, a positive denoising lower bound needs both overlap of the noisy marginals and separation of the conditional denoisers on that overlap. Our estimation lower bounds alone do not establish these conditions or a diffusion-sampling lower bound. We will clarify this distinction in Sections 2.1 and 5.

Thank you for the detailed and technically informative responses. I appreciate the authors' clarification of the scope of their results.

In particular, the response to Q2 confirms that the essential sufficient condition for the norm estimator is concentration of the signal's radial energy, rather than strong log-concavity itself. The response to Q3 also clarifies that minimax lower bounds for noise-level estimation do not directly imply lower bounds for blind denoising.

While these clarifications are helpful, they reinforce my concern about the significance of the results for blind diffusion models. The main theoretical contributions concern the minimax estimation of a scalar parameter in a known family of Gaussian-convolved distributions. The connection between these oracle estimation results and the behavior of learned blind denoisers remains indirect.

I recognize the technical value of the matching minimax bounds, particularly the lower-bound construction in the low-dimensional setting. However, the responses have not substantially changed my assessment of the paper's broader significance for generative modeling.

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

This paper considers the statistical model $ Y = X + \sigma N $ and analyzes the asymptotic order of the minmax estimation error of the noise strength $ \sigma $. In particular, the author(s) considered two classes of priors for XX: the distribution of low intrinsic dimensionality and strong log-concavity. Tight bounds are obtained for the estimation error for the two classes.

Critical strengths

This paper is clear and technically sound. The low-dimensional lower bound is sharp. The strongly log-concave result shows that low-dimensional support is unnecessary. A simple centered norm estimator achieves the optimal rate with sharp dependence on the curvature parameter. The comparison between likelihood-based and norm-based estimation explains why low covering complexity alone does not control the fluctuations of signal energy.

Critical weaknesses

From the reviewer's perspective, the main issue of this paper is that the discussion of the insights from the main results is not sufficient, and the motivation for considering the two classes of distributions is not clear. This limits the contribution of the paper. Please refer to the Questions for more details.

Questions for authors
  1. It is recommended to provide more discussion on the loss function R(P,Λ)R(P, \Lambda). Can this λ^P\widehat{\lambda}_P be given in a more explicit form? For example, the conditional expectation $ \mathbb{E} [\lambda | Y] $. Also, it seems λ\lambda has a prior distribution, but in the loss function, sup is taken over the interval Λ\Lambda, which is a fixed number. Please check.

  2. It is stated that the covering number N(supp(P),r0)≤2k−1N(\text{supp}(P), r_0) \leq 2^{k-1} and kk need not scale with dimension d. In such a definition, why is $ k$ called the intrinsic dimension? What measures satisfy that k does not scale with d? It would be better to give some specific examples and describe the geometry of the support of different classes of measures.

  3. It is recommended to give more discussion on why r0r_0 is called ‘resolution’ and its physical meaning.

  4. In Theorem 3.1, the right-hand side of the upper bound contains log⁡(2d)\log(2d) terms. However, in the following equation, i.e., the order of RLDR_{LD}, the log⁡\log terms are omitted. Please explain.

  5. The author(s) consider the order analysis of the MSE on two classes of Prob. Measures (prior distribution of the data). 1. The measure with low-dimensional geometry intrinsic dimensions, denoted by $ k $. 2. The class of measures with exponential decay, with log-concave parameter \alpha. However, the reviewer is expecting a more comprehensive comparison with these two classes. What is the relationship between these two classes, and what are typical examples for these two classes? The motivation for considering 1. and 2. is also not clear. The reviewer believes the quality of the paper would be improved if these were clarified.

  6. Another issue is that the paper mainly presents the technique, but the physical, mathematical, and machine learning insights behind the main results are insufficient. Also, it would be better to provide a discussion of the guidance/instructions that can be obtained from the main theoretical results to guide the design/application of the diffusion networks.

3 comments
AuthorsOpenReview ↗

Response to Reviewer PNXa

We thank the reviewer for carefully reading our paper and for the constructive suggestions on the motivation and interpretation of our results. We address the six questions below.

Q1. Loss function and parameterization. We agree that the risk formulation should be clearer. The precision parameter λ\lambda is fixed and unknown, while PP is the known signal law. The expectation is under the observation density pλp_{\lambda}:

Eλ,P[(λ^P(Y)−λ)2]=∫(λ^P(y)−λ)2pλ(y),dy.\mathbb E_{\lambda,P}[(\widehat\lambda_P(Y)-\lambda)^2]=\int(\widehat\lambda_P(y)-\lambda)^2p_{\lambda}(y)\\,dy.

The supremum over λ∈Λ\lambda\in\Lambda gives worst-case risk; the infimum is over estimators allowed to use PP. Examples are the MLE (Section 2.2) and centered norm estimator (Section 3.2). The training prior Π0\Pi_0 in Section 2.1 is distinct: our minimax formulation places no prior on λ\lambda. We will clarify these points after defining R(P;Λ)R(P;\Lambda) in Section 2.2.

Q2. Intrinsic dimension and examples. The condition 1+log⁡N(supp⁡(P),r0)≤k1+\log N(\operatorname{supp}(P),r_0)\le k bounds covering complexity at scale r0r_0, not necessarily linear or manifold dimension. Covering numbers are standard in statistics and also describe low-dimensional manifolds. For support on at most MM points, k=1+log⁡Mk=1+\log M suffices independently of dd. For support in an mm-dimensional radius-RR ball,

log⁡N(supp⁡(P),r0)≤mlog⁡(1+2R/r0)=O(mlog⁡d)\log N(\operatorname{supp}(P),r_0)\le m\log(1+2R/r_0)=O(m\log d)

for fixed R,ΛR,\Lambda. A fixed-radius sphere in a fixed-dimensional subspace is a manifold example. We will add these explanations and examples in Section 3.1.

Q3. Resolution. We agree this term deserves clarification. The quantity r0=λmin⁡/(λmax⁡d)r_0=\sqrt{\lambda_{\min}}/(\lambda_{\max}\sqrt d) is the scale at which covering complexity is measured. The assumption means that the support looks kk-dimensional at this resolution in the covering-number sense. We will clarify this in Section 3.1.

Q4. Rate simplification. For fixed Λ\Lambda, precision-dependent factors are absorbed into constants, while Θ~Λ\widetilde\Theta_{\Lambda} suppresses logarithmic factors. Specifically,

log⁡(2d)d+(k+log⁡(2d))2d2≲log⁡2(2d)(1d+k2d2).\frac{\log(2d)}d+\frac{(k+\log(2d))^2}{d^2}\lesssim\log^2(2d)\left(\frac1d+\frac{k^2}{d^2}\right).

Together with the lower bound, this gives the stated rate under 1≤k≲Λd1\le k\lesssim_{\Lambda}d and R2≳ΛkR^2\gtrsim_{\Lambda}k, with saturation at ΔΛ2\Delta_{\Lambda}^2. We will explain this simplification following Theorem 3.1.

Q5. Why these structural assumptions? We agree that the motivation should be clearer. Low intrinsic dimensionality, motivated by the manifold hypothesis, controls support geometry and is the setting studied in the BDDM analysis. Our result establishes its optimal noise-estimation scale up to logarithms. Strong log-concavity tests whether low-dimensional support is necessary; it is not. Full-dimensional examples such as N(μ,α−1Id)\mathcal N(\mu,\alpha^{-1}I_d) allow accurate estimation, with a centered norm estimator attaining the minimax rate. The two assumptions capture different mechanisms: geometric covering complexity and concentrated signal energy. We will clarify this comparison in Sections 1.1 and 3.2.

Q6. Insights and diffusion relevance. The k2/d2k^2/d^2 term shows why estimating one scalar noise parameter need not attain the usual 1/d1/d risk: even with the same known PP, latent signal variation can mask a precision change of order k/dk/d. Our construction gives two noisy marginals at TV distance at most 1−c1-c, so their optimal testing error under equal prior probabilities is at least c/2c/2.

In blind DDPMs, the population blind denoiser averages noise-conditioned denoisers over the posterior of the missing noise level. The BDDM analysis shows that this posterior concentrates under low intrinsic dimensionality. Our minimax result establishes the essential optimality of the corresponding noise-estimation scale. We do not claim a denoising or diffusion-sampling lower bound, which would require additional arguments. We will clarify the BDDM connection in Sections 1.1--2.1, the testing interpretation in Section 4.2, and the scope in Section 5.

Reviewer_PNXaOpenReview ↗

Reply to Author Response

I would like to thank the authors for the detailed response. Most of my concerns/condusion have been addressed, particularly regarding the loss function, the physical and geometric meanings of intrinsic dimension and resolution, as well as the applicability and practical rationality of the structural assumptions.

I believe that incorporating these clarifications into the revised manuscript could enhance the learning insights of the work. Therefore, I will keep my score.

We thank the reviewer for the positive follow-up and thoughtful suggestions. We have incorporated the clarifications into the revised manuscript, particularly regarding the risk formulation, intrinsic dimension, and structural assumptions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manuscript Revision

We thank all reviewers for their constructive feedback. We have uploaded a revised manuscript incorporating the clarifications outlined in our rebuttal.

The main revisions are:

  • Sections 1 and 5: Clarified the theoretical contributions, motivation, connection to blind diffusion, and scope of our results.
  • Section 2: Clarified the minimax risk and parameterization, added the Gaussian example, and explained the distinction between noise estimation and denoising.
  • Section 3: Expanded the discussion of intrinsic dimension, structural assumptions, and the sufficient conditions for norm-based estimation.
  • Section 4.2: Clarified the testing interpretation of the lower-bound construction.
  • Related work and presentation: Added the suggested references and numbered displayed equations.

The main theorems and their proofs remain unchanged.

We appreciate the reviewers' time and thoughtful comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.