Pith. sign in

REVIEW 4 major objections 4 minor 42 references

Replacing the diffusion denoiser with a closed-form conditional posterior mean yields gradient-free posterior sampling that reconstructs spectrum maps from sparse, noisy, low-bit measurements.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-03 14:28 UTC pith:L3QIRS3Q

load-bearing objection GSC is a plausible heuristic with a real application, but Theorem 1's exact-posterior claim does not hold, the main updates are essentially DMPS/QCS-SGM, and the active-sensing part is the genuinely new bit. the 4 major comments →

arxiv 2512.20108 v2 pith:L3QIRS3Q submitted 2025-12-23 cs.IT eess.SPmath.IT

Generative Spectrum Cartography: Unified Reconstruction and Active Sensing via Diffusion Models

classification cs.IT eess.SPmath.IT
keywords spectrum cartographydiffusion modelsinverse problemsquantized measurementsposterior samplingactive sensingradio environment mapsuncertainty estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that spectrum map reconstruction from sparse, noisy, and low-bit quantized measurements can be done by a diffusion model whose reverse process uses a closed-form conditional posterior mean instead of likelihood-gradient guidance. It claims that conditioning on measurements only changes the clean-signal estimator, so replacing the unconditional denoiser with the posterior mean produces a valid posterior reverse diffusion process. Theorems 2 and 3 provide explicit formulas for this posterior mean under linear and quantized observation models, enabling gradient-free, measurement-consistent sampling. If correct, this makes diffusion-based radio map recovery practical for satellite sensing with coarse ADCs and limited on-board computation, and it provides a principled uncertainty signal for active sensing.

Core claim

The central claim is Theorem 1: in the DDPM reverse step, replacing the unconditional clean-signal predictor E[x0|xt] with the measurement-conditioned posterior mean E[x0|xt,y] yields a reverse diffusion process whose final state follows the posterior p(x0|y). Theorem 2 gives the closed-form posterior mean for linear Gaussian observations as a LMMSE-style update with gain γ_t^2/(γ_t^2+σ_y^2) H^T, exploiting the idempotence of the binary masking operator. Theorem 3 gives the quantized case as an element-wise truncated-Gaussian correction, where the quantized likelihood is an interval probability and the posterior mean is expressed in terms of normal PDF and CDF ratios. Together they form a gr

What carries the argument

The load-bearing object is the conditional posterior mean E[x0|xt,y] substituted into the affine DDPM reverse mean μ_t = a_t x0 + b_t xt. The derivation assumes that, at each step, the clean signal given the noisy state is Gaussian with isotropic covariance: x0|xt ∼ N(x̂0|t, γ_t^2 I). Under that assumption, the linear observation likelihood yields a closed-form LMMSE update, and the scalar quantized likelihood becomes a Gaussian interval probability, whose posterior mean is computed analytically via the derivative of the log-partition function of a truncated Gaussian. This mechanism replaces per-step likelihood-gradient backpropagation with a direct analytical update.

Load-bearing premise

The whole posterior-sampling guarantee rests on the assumption that, at every diffusion step, the clean signal conditioned on the noisy state is exactly Gaussian with isotropic covariance γ_t^2 I, but the paper never specifies or estimates γ_t^2, and the proof of Theorem 1 keeps the unconditional variance when substituting the conditional mean, which is exact only if the conditional variance is zero.

What would settle it

Run GSC on a dataset whose clean-signal conditional is strongly non-Gaussian (e.g., heavy-tailed or structured interference maps) with a fixed γ_t^2, and compare the ensemble reconstructions against exact posterior samples or against a gradient-based solver where gradients are well-defined. If the ensemble's per-pixel variance does not match the true posterior uncertainty, or if reconstruction collapses, the Theorem 1 validity claim is falsified. Additionally, vary γ_t^2 over several orders of magnitude: if the reconstructed maps and uncertainty estimates change drastically, the unspecified γ_

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Reconstruction from 1-, 2-, and 3-bit quantized measurements becomes tractable with diffusion priors, since quantization is handled as interval probabilities rather than through undefined or zero gradients.
  • Removing per-step backpropagation through the measurement operator makes diffusion-based spectrum cartography fast enough for on-board satellite processing.
  • Posterior samples yield a pixel-wise uncertainty map, which can drive active sensing; the proposed K-means dispersion strategy avoids redundant sampling in clustered high-uncertainty regions.
  • A single pre-trained diffusion prior handles both linear and quantized observation models with the same framework, simplifying deployment across different sensing hardware.
  • The uncertainty-guided sampling loop can be closed: reconstructed uncertainty selects new measurement locations, and the added data further reduces uncertainty in a sense–reconstruct–plan cycle.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension would apply the same truncated-Gaussian update to other interval-censored or exponential-family likelihoods (e.g., Poisson or censored measurements), since the derivation relies on the derivative of a log-partition function.
  • Because Theorem 1 keeps the unconditional reverse variance while substituting the conditional mean, the resulting chain is not the exact posterior transition unless the conditional variance is zero; correcting this would likely shrink the sampling variance at observed pixels and sharpen the uncertainty map.
  • The framework leaves γ_t^2 unspecified; a testable check is whether reconstruction quality and the posterior ensemble vary strongly with the γ_t^2 schedule. If they do, γ_t^2 is a hidden control parameter that must be tuned per problem.
  • The active-sampling policy could be benchmarked against information-theoretic criteria such as mutual information or Bayesian experimental design; the variance heuristic may underperform when the posterior is skewed or multi-modal.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Generative Spectrum Cartography (GSC), a diffusion-based framework for reconstructing radio-environment maps from sparse, noisy, and low-bit quantized measurements. The method replaces the unconditional clean-signal estimate E[x0|xt] in the DDPM reverse step by a closed-form conditional posterior mean E[x0|xt,y], derived for linear (Theorem 2) and quantized (Theorem 3) observation models, and claims (Theorem 1) that this yields exact posterior sampling without likelihood-gradient guidance. The resulting posterior samples are used to estimate pixel-wise variance and to select new measurement locations via a K-means-based diversity-aware active sampling strategy (Algorithm 2). Experiments on simulated and satellite-simulated spectrum maps report consistent PSNR/LPIPS gains over IDW, Nasdac, Dowjons, DPS, and UNN baselines under sparse, noisy, and quantized observations.

Significance. If substantiated, the gradient-free closed-form updates would be practically attractive for low-bit quantized sensing, and the unification of reconstruction with uncertainty-driven active sensing is a useful systems contribution. The algebraic derivations in Theorems 2 and 3 are correct under the stated Gaussian-prior assumption, and the paper includes a fairly extensive empirical study, including a high-fidelity satellite simulation. However, the central theoretical claim of exact posterior sampling is not established: Theorem 1's proof drops a conditional-covariance term, and the assumed variance gamma_t^2 is never specified or estimated. Moreover, the linear update (14) and quantized update (22) coincide with previously published methods DMPS [30] and QCS-SGM [35,37] by the same group, so the novelty claims need substantial down-scoping, and the experiments omit those direct baselines. The paper could become a solid contribution after major revision.

major comments (4)
  1. [Section III, Theorem 1 proof (Eqs. 10–12)] The proof does not establish exact posterior sampling. Marginalizing q(xt-1|xt,x0) over p(x0|xt,y) gives a Gaussian with mean at E[x0|xt,y]+b_t xt and covariance sigma_tilde_t^2 I + a_t^2 Cov(x0|xt,y). The algorithm uses only sigma_tilde_t^2 I, dropping the covariance term. Unless Cov(x0|xt,y)=0 — which is neither argued nor generally true — the transition is not the true posterior backward kernel, and the claimed stationary distribution p(x0|y) is unsupported. The step 'summarize x0 by a point estimator' is a non sequitur; an MMSE point estimate is not a sufficient statistic for the next latent. This is the paper's central claim and must be corrected or explicitly reframed as an approximation.
  2. [Section III, Eq. (13) and Algorithm 1] Theorems 2 and 3 rest on the assumption x0|xt ~ N(xhat0|t, gamma_t^2 I), but gamma_t^2 is never specified, estimated, or reported. Algorithm 1 takes gamma_t^2 as an input, but no schedule, calibration procedure, or sensitivity analysis is given. For a trained diffusion prior, the conditional distribution is not exactly Gaussian isotropic. The closed-form 'posterior mean' is therefore at best an LMMSE/MMSE under an approximating Gaussian, and the behavior of the sampler depends critically on the unreported gamma_t^2. The paper must specify how gamma_t^2 is chosen and show sensitivity, or the exactness claims should be withdrawn.
  3. [Contributions (Section I) and Theorems 2–3] The claimed 'novel posterior sampling mechanism' and 'unified closed-form posterior inference' are overstated because the linear update (14) is algebraically the DMPS update [30], and the quantized update (22) is the MMSE update used in QCS-SGM [35,37]. These references are cited but not positioned as prior art. The paper should explicitly state that the closed-form posterior mean updates are known from these works, clarify the actual novelty (e.g., the system-level unification with active sensing), and benchmark against DMPS and QCS-SGM directly.
  4. [Section V, Tables I–VII] The empirical evaluation omits DMPS and QCS-SGM, which are the direct competitors with the same posterior-mean updates. Showing that GSC outperforms DPS mainly demonstrates that posterior-mean data consistency is more stable than likelihood-gradient guidance, which is a known property of DMPS/QCS-SGM. Additionally, no standard deviations across the 100 evaluation maps are reported, no gamma_t settings are given for the experiments, and the active-sensing ensemble size N is not stated. These omissions prevent reproduction and weaken the evidence for the claimed improvements.
minor comments (4)
  1. [Notation, Eqs. (2) and (24)] The observation model (2) uses noise variance sigma_y^2, while Theorem 3 and Eq. (24) use sigma_e; Algorithm 1 passes 'sigma_y^2 (or sigma_e^2)' without defining sigma_e. Please unify the notation.
  2. [Section II-A and Theorem 2 proof] H is defined as M x N, but Theorem 2's proof uses H H^T = H^2, which only holds for a square diagonal mask. Clarify whether H is a square masking matrix with zero rows/columns or an M x N selection matrix, and adjust the derivation accordingly.
  3. [Title/abstract] The abstract title ('Generative Spectrum Cartography: Unified Reconstruction and Active Sensing via Diffusion Models') differs from the full title ('Generative Bayesian Spectrum Cartography: Unified Reconstruction and Active Sensing via Diffusion Models'). Please align them.
  4. [Algorithm 1 and Section IV-A] Algorithm 1 outputs xhat_0^post at line 12, which is a posterior mean, not a posterior sample; Section IV-A uses multiple samples to estimate variance. Clarify how the mean and the samples are produced and used, and label V(i,j) as predictive variance rather than strictly 'epistemic uncertainty.'

Circularity Check

2 steps flagged

Theorem 1's 'valid posterior' claim is built into the update rather than derived; quantized/linear updates restate the same authors' prior MMSE formulas.

specific steps
  1. self definitional [Section III, Theorem 1 and its proof, Eqs. (10)-(12)]
    "When we summarize this latent variable by a single point estimator under mean-squared error, the optimal choice is precisely the posterior mean E[x0|xt,y]. Therefore, to construct a posterior reverse update that respects the DDPM affine structure (11), it is natural to replace the unknown x0 in atx0+btxt by the conditional posterior mean E[x0|xt,y] ... which completes the proof."

    The reverse update is defined by substituting E[x0|xt,y] into the affine mean, and then the theorem asserts that this 'yields a valid posterior reverse diffusion process whose final state follows p(x0|y).' But that validity is not obtained by marginalizing the true kernel q(xt-1|xt,x0) over p(x0|xt,y); doing so would add an at^2 Cov(x0|xt,y) term to the variance. The proof instead declares the point-estimator substitution 'natural' and treats the posterior-sampling property as an input to the construction. Thus the claimed conclusion is not derived from the diffusion equations; it is attached to the update rule by construction.

  2. renaming known result [Section I (Contributions) and Section III-B, Theorem 3, Eqs. (21)-(23)]
    "we propose a novel posterior sampling mechanism by deriving tractable closed-form solutions for the conditional expectation E[x0|xt,y] under both linear and quantized models ... Theorem 3 ... Delta_i = phi(a_i)-phi(b_i) / (Phi(b_i)-Phi(a_i))."

    The quantized MMSE update in Theorem 3 is the same correction used in the same authors' prior QCS-SGM works [35],[37], and the linear update in Theorem 2 is the DMPS [30] posterior mean. These are cited in the paper but presented as a new closed-form posterior inference mechanism. The derivation is algebraically self-contained and the external benchmarks are independent, so this is not a logical tautology; however, the 'first-principles' contribution is substantially a re-labeling of prior results from the same group.

full rationale

The main load-bearing mathematical claim, Theorem 1, is not a genuine derivation: it replaces x0 by E[x0|xt,y] in the DDPM reverse mean while keeping the unconditional variance, dropping the conditional-variance term that would appear in the true posterior backward kernel. The proof labels this substitution as 'natural' and then asserts posterior validity, so the theorem's output is effectively assumed in the design of the update. Theorems 2 and 3 are internally correct given the Gaussian ansatz x0|xt ~ N(xhat0|t, gamma_t^2 I), but gamma_t^2 is never specified or estimated; this is a missing assumption / tuning input rather than a fitted prediction called a prediction. The active-sensing component is a reasonable heuristic and does not reduce to its own inputs. The experimental comparisons are against external non-circular baselines, so the empirical portion is not circular. Overall, the circularity is partial: one central theorem is a proof-by-construction, and part of the claimed novelty restates prior work by the same authors. This yields a score of 4 rather than a higher score, because the closed-form updates and the experimental evaluation retain independent content once the Gaussian assumption and gamma_t^2 are fixed.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central derivation depends on one hand-specified variance γ_t^2 and on the assumption that the diffusion model's conditional clean-signal posterior is exactly Gaussian with isotropic covariance. These premises are load-bearing: without them the closed-form update is not the exact posterior mean. The remaining assumptions are standard diffusion and measurement-model assumptions.

free parameters (1)
  • γ_t^2 (conditional prior variance)
    Appears in Eqs. (13),(14),(20),(22) and Algorithm 1; controls the strength of the measurement update. The paper never gives its schedule or estimation method; if chosen to maximize PSNR on test maps it is a fitted hyperparameter. Theorems 2 and 3 are exact only for a fixed, correct γ_t^2.
axioms (5)
  • standard math DDPM forward/reverse Markov chain and Tweedie estimate E[x0|xt] are valid.
    Background diffusion formalism used throughout Section II-B; standard from Ho et al. [19] and Song et al. [21].
  • ad hoc to paper At each reverse step, x0|xt is exactly Gaussian with isotropic covariance γ_t^2 I.
    Assumed in Theorem 2 (Eq. 13) and Theorem 3; not derived from the diffusion model, and γ_t^2 is never specified or estimated. All closed-form updates (14) and (22) depend on this assumption.
  • domain assumption The observation operator H is a binary diagonal idempotent mask.
    Used to simplify the gain to Eq. (20). The text defines H as M×N but proves as if it were a diagonal square mask; assumes observed locations are non-overlapping and each measurement maps to one pixel.
  • domain assumption Measurement y is conditionally independent of the noisy state xt given the clean signal x0.
    Used in Theorem 1 and Theorem 2 via p(y|xt,x0)=p(y|x0); reasonable for the measurement models in Eqs. (1) and (2).
  • domain assumption Quantization intervals (li, ui] and noise variance σ_e^2 are known exactly.
    Needed for the normalized boundaries in Theorem 3 (Eq. 21); the paper never precisely defines σ_e in relation to σ_y.

pith-pipeline@v1.3.0-alltime-deepseek · 15809 in / 16389 out tokens · 153802 ms · 2026-08-03T14:28:55.491045+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Generative Spectrum Cartography: Unified Reconstruction and Active Sensing via Diffusion Models." pith.science (2026). https://pith.science/paper/L3QIRS3Q

@misc{pith2026251220108,
  author       = {Pith},
  title        = {Pith review of: Generative Spectrum Cartography: Unified Reconstruction and Active Sensing via Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L3QIRS3Q}},
  note         = {Machine review of arXiv:2512.20108}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

High-fidelity spectrum cartography is important for spectrum monitoring and wireless situational awareness, especially in satellite-based wide-area sensing scenarios where measurements are sparse, noisy, and often low-bit quantized. In such settings, two coupled challenges arise: accurate reconstruction from severely incomplete measurements and efficient allocation of additional sensing resources under a limited sensing budget. Existing methods usually address these problems separately, and, for reconstruction, they often rely on priors that are insufficiently expressive under sparse and quantized measurements. This paper proposes Generative Spectrum Cartography (GSC), a diffusion-based posterior inference framework for spectrum cartography with uncertainty-aware active sensing. Specifically, spectrum map recovery is formulated as a Bayesian inverse problem under a learned diffusion model prior, and closed-form posterior mean updates are derived for both linear and quantized measurement models. By embedding these updates into the reverse diffusion process, GSC enables gradient-free and measurement-consistent posterior sampling without relying on computationally costly likelihood-gradient guidance. The resulting posterior samples are further used to estimate spatial uncertainty and to guide diversity-aware selection of additional measurement locations for active sensing. Experiments on simulated electromagnetic maps and a high-fidelity simulated satellite monitoring scenario show that GSC achieves higher PSNR, lower LPIPS, and more efficient sensing than representative baseline methods under sparse, noisy, and low-bit quantized measurements.

Figures

Figures reproduced from arXiv: 2512.20108 by Linling Kuang, Sheng Wu, Xiangming Meng, Yuntong Gu, Zhiyuan Lin.

Figure 1
Figure 1. Figure 1: Illustration of the uncertainty-driven sampling procedure. (a) The estimated uncertainty map V(i, j) computed from posterior samples. (b) The top 5% pixels with the highest uncertainty, revealing significant spatial clustering. (c) Sampling locations selected via the proposed K-means strategy, which effectively disperses measurements to maximize spatial information gain. uncertainty of the reconstruction—i… view at source ↗
Figure 2
Figure 2. Figure 2: A 2D slice of the simulated electromagnetic map at a specific [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of sparse random sampling masks used in experiments. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reconstruction comparison under 15% sparse sampling (noise-free). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reconstruction comparison under 5% sparse sampling (noise-free). [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Reconstruction results of electromagnetic maps under 15% sparse [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Active learning performance under linear observations. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Reconstruction of Simulated Satellite radio map using 10% sampling. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 5 linked inside Pith

  1. [1]

    Informed spectrum usage in cognitive radio networks: Interference cartography,

    A. B. H. Alaya-Feki, S. B. Jemaa, B. Sayrac, P. Houze, and E. Moulines, “Informed spectrum usage in cognitive radio networks: Interference cartography,” in2008 IEEE 19th International Symposium on Personal, Indoor and Mobile Radio Communications, 2008, pp. 1–5

  2. [2]

    Deep spectrum cartography: Complet- ing radio map tensors using learned neural models,

    S. Shrestha, X. Fu, and M. Hong, “Deep spectrum cartography: Complet- ing radio map tensors using learned neural models,”IEEE Transactions on Signal Processing, vol. 70, pp. 1170–1184, 2022

  3. [3]

    Spec- trum cartography techniques, challenges, opportunities, and applications: A survey,

    Y . S. Reddy, A. Kumar, O. J. Pandey, and L. R. Cenkeramaddi, “Spec- trum cartography techniques, challenges, opportunities, and applications: A survey,”Pervasive and Mobile Computing, vol. 79, p. 101511, 2022

  4. [4]

    A low-complexity return-link beam-hopping scheduling for ngso mega-constellations with dynamic topology and uneven traffic,

    Z. Lin, L. Kuang, B. Liu, C. Jiang, and J. Jin, “A low-complexity return-link beam-hopping scheduling for ngso mega-constellations with dynamic topology and uneven traffic,”IEEE Internet of Things Journal, vol. 12, no. 11, pp. 16 695–16 710, Jun. 2025

  5. [5]

    Satellite-terrestrial coordinated multi-satellite beam hopping scheduling based on multi- agent deep reinforcement learning,

    Z. Lin, Z. Ni, L. Kuang, C. Jiang, and Z. Huang, “Satellite-terrestrial coordinated multi-satellite beam hopping scheduling based on multi- agent deep reinforcement learning,”IEEE Transactions on Wireless Communications, vol. 23, no. 8, pp. 10 091–10 103, 2024

  6. [6]

    Spectrum Map and Its Application in Resource Management in Cognitive Radio Networks,

    S. Debroy, S. Bhattacharjee, and M. Chatterjee, “Spectrum Map and Its Application in Resource Management in Cognitive Radio Networks,” IEEE Transactions on Cognitive Communications and Networking, vol. 1, no. 4, pp. 406–419, Jan. 2015

  7. [7]

    Reliability of a radio environment map: Case of spatial interpolation techniques,

    D. Denkovski, V . Atanasovski, L. Gavrilovska, J. Riihij ¨arvi, and P. M¨ah¨onen, “Reliability of a radio environment map: Case of spatial interpolation techniques,” in2012 7th international ICST conference on cognitive radio oriented wireless networks and communications (CROWNCOM). IEEE, 2012, pp. 248–253

  8. [8]

    Group-lasso on splines for spectrum cartography,

    J. A. Bazerque, G. Mateos, and G. B. Giannakis, “Group-lasso on splines for spectrum cartography,”IEEE Transactions on Signal Processing, vol. 59, no. 10, pp. 4648–4663, 2011

  9. [9]

    Non-parametric spectrum cartogra- phy using adaptive radial basis functions,

    M. Hamid and B. Beferull-Lozano, “Non-parametric spectrum cartogra- phy using adaptive radial basis functions,” in2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 3599–3603

  10. [10]

    Wire- less sensor network for spectrum cartography based on kriging interpola- tion,

    G. Boccolini, G. Hern ´andez-Pe˜naloza, and B. Beferull-Lozano, “Wire- less sensor network for spectrum cartography based on kriging interpola- tion,” in2012 IEEE 23rd International Symposium on Personal, Indoor and Mobile Radio Communications - (PIMRC), 2012, pp. 1565–1570

  11. [11]

    Compressive multispectral spectrum sensing for spectrum cartography,

    J. Mar ´ın Alfonso, J. I. Mart ´ınez Torre, H. Arguello Fuentes, and L. B. Agudelo, “Compressive multispectral spectrum sensing for spectrum cartography,”Sensors, vol. 18, no. 2, p. 387, 2018

  12. [12]

    Improved performance of spectrum cartography based on compressive sensing in cognitive radio networks,

    B. A. Jayawickrama, E. Dutkiewicz, I. Oppermann, G. Fang, and J. Ding, “Improved performance of spectrum cartography based on compressive sensing in cognitive radio networks,” in2013 IEEE International Con- ference on Communications (ICC). IEEE, 2013, pp. 5657–5661

  13. [13]

    Distributed spectrum sensing for cognitive radio networks by exploiting sparsity,

    J. A. Bazerque and G. B. Giannakis, “Distributed spectrum sensing for cognitive radio networks by exploiting sparsity,”IEEE Transactions on Signal Processing, vol. 58, no. 3, pp. 1847–1862, 2010

  14. [14]

    Spectrum cartog- raphy via coupled block-term tensor decomposition,

    G. Zhang, X. Fu, J. Wang, X.-L. Zhao, and M. Hong, “Spectrum cartog- raphy via coupled block-term tensor decomposition,”IEEE Transactions on Signal Processing, vol. 68, pp. 3660–3675, 2020

  15. [15]

    Spectrum cartography using quantized observations,

    D. Romero, S.-J. Kim, R. L ´opez-Valcarce, and G. B. Giannakis, “Spectrum cartography using quantized observations,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 3252–3256

  16. [16]

    Spatiotemporal correlation analysis and visualization of electromagnetic intensity based on multi- site and multi-time attention mechanism,

    Y . Xu, W. Zi, J. Song, R. Shao, and H. Chen, “Spatiotemporal correlation analysis and visualization of electromagnetic intensity based on multi- site and multi-time attention mechanism,”Journal of Nanjing University (Natural Sciences), vol. 57, no. 5, pp. 838–846, 2021

  17. [17]

    Sc- gan: A spectrum cartography with satellite internet based on pix2pix generative adversarial network,

    Z. Pan, Z. Bangning, W. Heng, M. Wenfeng, and G. Daoxing, “Sc- gan: A spectrum cartography with satellite internet based on pix2pix generative adversarial network,”China Communications, vol. 22, no. 2, pp. 47–61, 2025

  18. [18]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,”Advances in neural information processing systems, vol. 32, 2019

  19. [19]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

  20. [20]

    Improved denoising diffusion probabilis- tic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilis- tic models,” inInternational conference on machine learning. PMLR, 2021, pp. 8162–8171

  21. [21]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502, 2020

  22. [22]

    Diffusion posterior sampling for linear inverse problem solving: A filtering perspective,

    Z. Dou and Y . Song, “Diffusion posterior sampling for linear inverse problem solving: A filtering perspective,” inThe Twelfth International Conference on Learning Representations, 2024

  23. [23]

    Score-based generative modeling through stochastic differ- ential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,”arXiv preprint arXiv:2011.13456, 2020

  24. [24]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,”arXiv preprint arXiv:2209.03003, 2022

  25. [25]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,”arXiv preprint arXiv:2210.02747, 2022

  26. [26]

    Improving diffusion inverse problem solving with decoupled noise an- nealing,

    B. Zhang, W. Chu, J. Berner, C. Meng, A. Anandkumar, and Y . Song, “Improving diffusion inverse problem solving with decoupled noise an- nealing,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 20 895–20 905

  27. [27]

    Zero-shot image restoration using denoising diffusion null-space model,

    Y . Wang, J. Yu, and J. Zhang, “Zero-shot image restoration using denoising diffusion null-space model,”arXiv preprint arXiv:2212.00490, 2022. 12

  28. [28]

    Denoising diffusion restoration models,

    B. Kawar, M. Elad, S. Ermon, and J. Song, “Denoising diffusion restoration models,”Advances in neural information processing systems, vol. 35, pp. 23 593–23 606, 2022

  29. [29]

    Diffusion posterior sampling for general noisy inverse problems,

    H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye, “Diffusion posterior sampling for general noisy inverse problems,” inThe Eleventh International Conference on Learning Representations, 2023

  30. [30]

    Diffusion model based posterior sampling for noisy linear inverse problems,

    X. Meng and Y . Kabashima, “Diffusion model based posterior sampling for noisy linear inverse problems,” inAsian Conference on Machine Learning, 5-8 December 2024, Hanoi, Vietnam, ser. Proceedings of Machine Learning Research, V . Nguyen and H. Lin, Eds., vol. 260. PMLR, 2024, pp. 623–638

  31. [31]

    Pseudoinverse-guided diffusion models for inverse problems,

    J. Song, A. Vahdat, M. Mardani, and J. Kautz, “Pseudoinverse-guided diffusion models for inverse problems,” inInternational Conference on Learning Representations, 2023

  32. [32]

    Deep bayesian active learning with image data,

    Y . Gal, R. Islam, and Z. Ghahramani, “Deep bayesian active learning with image data,” inInternational conference on machine learning. PMLR, 2017, pp. 1183–1192

  33. [33]

    Uncertainty-guided perturbation for image super-resolution diffusion model,

    L. Zhang, W. You, K. Shi, and S. Gu, “Uncertainty-guided perturbation for image super-resolution diffusion model,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 980– 17 989

  34. [34]

    Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,

    X. Wang, K. Tao, N. Cheng, Z. Yin, Z. Li, Y . Zhang, and X. Shen, “Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 2, pp. 738–750, 2025

  35. [35]

    Quantized compressed sensing with score- based generative models,

    X. Meng and Y . Kabashima, “Quantized compressed sensing with score- based generative models,” inInternational Conference on Learning Representations, 2023

  36. [36]

    Compressed sensing with quan- tized measurements,

    A. Zymnis, S. Boyd, and E. Candes, “Compressed sensing with quan- tized measurements,”IEEE Signal Processing Letters, vol. 17, no. 2, pp. 149–152, 2009

  37. [37]

    Qcs-sgm+: Improved quantized com- pressed sensing with score-based generative models,

    X. Meng and Y . Kabashima, “Qcs-sgm+: Improved quantized com- pressed sensing with score-based generative models,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 13, 2024, pp. 14 341–14 349

  38. [38]

    Bayes-optimal joint channel-and-data estimation for massive mimo with low-precision adcs,

    C.-K. Wen, C.-J. Wang, S. Jin, K.-K. Wong, and P. Ting, “Bayes-optimal joint channel-and-data estimation for massive mimo with low-precision adcs,”IEEE Transactions on Signal Processing, vol. 64, no. 10, pp. 2541–2556, 2015

  39. [39]

    Reliability of a radio environment map: Case of spatial interpolation techniques,

    D. Denkovski, V . Atanasovski, L. Gavrilovska, J. Riihij ¨arvi, and P. M¨ah¨onen, “Reliability of a radio environment map: Case of spatial interpolation techniques,” in2012 7th International ICST Conference on Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM), 2012, pp. 248–253

  40. [40]

    Domain-factored untrained deep prior for spectrum cartography,

    S. Timilsina, S. Shrestha, L. Cheng, and X. Fu, “Domain-factored untrained deep prior for spectrum cartography,”IEEE Signal Processing Letters, vol. 32, pp. 3440–3444, 2025

  41. [41]

    Image quality assess- ment: from error visibility to structural similarity,

    Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assess- ment: from error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

  42. [42]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595