{"id":"af7bee10-f06c-4a3d-85b5-1efedca202e2","arxiv_id":"2412.05582","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DM-SBL jointly samples a sparse channel and structured interference using a diffusion-model score for the interference and an EM-updated sparse Bayesian prior for the channel, and it outperforms AWGN-based estimators at low signal-to-interference ratios in simulation.","lead":"Researchers built a channel estimator that uses a diffusion model to learn the pattern of structured interference and a sparse Bayesian prior to learn the channel, then samples both together from the received signal. In simulations with strong interference, it estimates the channel far more accurately than conventional estimators that treat all interference as Gaussian noise, which matters for coexistence of sonar or radar and communication in the same bandwidth.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq (13)'s K-sample factorization has no error bound or validation; if the approximation is poor, the reported 20 dB NMSE gains may be an artifact of the matched LFM simulation rather than a property of DM-SBL.","rationale":"The paper's central claim is empirical: DM-SBL significantly outperforms AWGN-only baselines under structured interference. For that claim to hold as a property of the method rather than of the specific simulation setup, the joint sampling procedure must produce samples close to the true posterior. The weakest link is exactly the K-sample factorization in Eq (13), which the reader identified. This factorization is load-bearing because the score decomposition in Eq (15) is built on it; if the approximation is biased, the final channel estimate is not a principled posterior mean and the reported NMSE gains may not generalize to other interference types or channel settings. I agree with the reader's assessment and do not find a more severe concern. The additional dimensional typo in Eq (25) reinforces the need for an independent implementation check, and the unconstrained EM update for gamma (which can go negative) is a further unvalidated heuristic, but these are secondary to the factorization's lack of correctness guarantee. The CONDITIONAL verdict remains appropriate: the empirical evidence is suggestive but not fully verified without code, error bars, or a tractable validation of the factorization.","tokens_in":16408,"tokens_out":10546,"duration_ms":98053,"concrete_test":"Construct a small tractable instance (e.g., M=L=8) with a known sparse channel h and interference drawn from a Gaussian mixture, so the true posterior p(h0,n0|y) can be computed by enumeration or MCMC. Run DM-SBL with the same K-sample factorization and compare the final mean of h0 samples to the true posterior mean and to the MMSE estimate, for K=1, K=64, and K=256. If the mean deviates beyond Monte Carlo error or depends strongly on K, the Eq (13) factorization introduces bias. Also rerun with gamma clamped to nonnegative values to isolate the effect of the unconstrained EM update.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reverse SDE's target distribution is defined by Eq (13), which factors p(bar_h_t, bar_n_t | y) as the product over all i,j of p(h_i,n_j|y)^(1/K), with no derivation or error bound. This heuristic, borrowed from [10,20], replaces the true posterior score for each h_i (which would require marginalizing over all other samples conditioned on y) with a geometric average over pairwise likelihoods. The resulting score in Eq (15) is not the score of any proper posterior for the K-sample generative model, and the paper provides no convergence analysis, no ablation varying K, and no validation that the approximation is accurate. The single-realization NMSE of -29.95/-30.47 dB versus -8.90 dB for SBL could therefore reflect the specific LFM interference (whose score the network was trained on) and the carefully tuned hyperparameters, rather than a generally valid joint sampling procedure. Additionally, Eq (25) as printed contains a dimensional inconsistency: the residual inside the parentheses uses A^H h_t where the model requires A h_t (A is M x L, so A^H h_t is L-dimensional and cannot be subtracted from y); an independent implementation following the paper literally would not run.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DM-SBL, a diffusion-model-based sparse Bayesian learning method for estimating a sparse channel from observations corrupted by both AWGN and structured interference. The interference score is learned by a diffusion model, while the channel is modeled as a zero-mean Gaussian with sparsity-controlling variance gamma updated via an EM-like procedure. The joint reverse SDE is run over K parallel channel and interference samples using a factorized posterior approximation, and two likelihood approximations (DMPS and PiGDM) are used. Simulations on an underwater acoustic channel with LFM interference report large NMSE improvements over MMSE, SBL, EM-BGGAMP, VAMP, and OMP, especially at low SIR.","tokens_in":16726,"tokens_out":9721,"duration_ms":90240,"significance":"If the central claims are correct, the paper offers a practical way to exploit a learned interference score without training a channel score, which is attractive for underwater acoustic and radar/communication coexistence scenarios. The analytic channel score, the use of two different perturbed-likelihood approximations, and the K-sample averaging are useful engineering contributions, and the paper reports computational times and a clear algorithmic framework. However, the validity of the method rests on an unproven K-sample factorization, the printed algorithm and score equations contain dimensional/indexing errors that block literal reproduction, and the evaluation is limited to LFM interference against AWGN-based baselines only. These gaps currently prevent the strong empirical claims from being fully supported.","major_comments":[{"comment":"Equation (13) defines the joint posterior targeted by the reverse SDE as a product over all pairs of K-sample posterior factors. This is called an 'implementation trick' and no derivation, error bound, or validation is provided. Since the score updates in Eq. (15) and the final channel estimate are the mean under this factored density, the central claim rests on this approximation. Please add either a derivation or a reference explaining the validity of this factorization for the K-sample joint posterior of h and n, or provide an empirical validation, e.g., a small-scale experiment comparing the factorized score with a Monte-Carlo estimate of the true conditional score, and an ablation over K (K = 1, 4, 16, 64, 256).","section":"Section III, Eqs. (13)-(15)"},{"comment":"Equation (25) contains a dimensional error: the DMPS score for h_t^{(i)} uses y - (1/alpha(t)) A^H h_t^{(i)} - ..., but A is M x L, so A^H h_t^{(i)} is L-dimensional and cannot be subtracted from y; the correct term, as in Eq. (22)-(24) and Table I, is A h_t^{(i)}. Similarly, Eq. (29) states that nabla_{h_t} \\hat{h}_t = (I_M + ...)/alpha(t), while the quantity is an L x L Jacobian and must be I_L. These errors make the algorithm impossible to execute exactly as printed and must be corrected.","section":"Section III-A, Eq. (25) and Eq. (29)"},{"comment":"The EM update gamma_new(l) = (1/alpha^2(t))(nu_h^t(l) + |\\hat{h}_t(l)|^2) - 2(1-alpha^2(t))/alpha^2(t) can be negative, because the subtracted term can dominate when the sample variance and mean are small. Since gamma is a variance, negative values are invalid and would break the subsequent score evaluation. Please modify the update to gamma <- max(0, expression), or derive the constrained EM solution, and verify that the reported results use such a clip.","section":"Section III-C, Eq. (36) and Algorithm 1 lines 16, 27"},{"comment":"The reverse-time discretization is not written consistently: dt is initialized to -1/T, but line 19 updates h^{(i)}_{t-dt} using h^{(i)}_{t-dt} on the right-hand side instead of h^{(i)}_t, and because dt < 0 the quantity t-dt = t + 1/T moves forward in time rather than backward. This is not the standard predictor-corrector discretization of the reverse SDE and cannot be reproduced as printed. Please rewrite the loop with an explicit decreasing time grid and correct the update equations.","section":"Algorithm 1, lines 2, 19-25"},{"comment":"The experimental section supports the central claim only for LFM interference and only against AWGN-based baselines, and it does not state the number of Monte Carlo realizations or provide error bars for Figs. 6 and 7. To substantiate the abstract's broad claim, add a comparison with a structured-interference-aware method such as the diffusion-based approach of [9], report results for at least one additional interference class (e.g., CW or HFM), and provide confidence intervals or error bars. Also state whether the mu and kappa values are tuned per SIR/SNR setting.","section":"Section IV"}],"minor_comments":[{"comment":"The phrase 'the complicate structure' should be 'the complicated structure'.","section":"Abstract"},{"comment":"The heading 'Updating γ' is preceded by the typo 'nusiance parameter' in the text; it should be 'nuisance parameter'.","section":"Section III-C heading"},{"comment":"The caption repeats labels (a), (b), (c) for the second row of subfigures; the lower row should be labeled (d), (e), (f).","section":"Fig. 5 caption"},{"comment":"The text refers to 'DM-SBL (GDM)' in several places (e.g., Fig. 3 caption and the discussion after Fig. 7); this should be 'DM-SBL (ΠGDM)' for consistency with the notation introduced in Section III-B.","section":"Fig. 3 caption and Section IV text"},{"comment":"The word 'elemenwise' should be 'elementwise'.","section":"Eq. (34) following text"},{"comment":"The pilot length is not reported, only the resulting M = 200; given M = N - L + 1, stating N for L = 200 and L = 300 would help reproducibility.","section":"Section IV, Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea and the simulations show a large improvement, but the unproven K-sample factorization and the several errors in the printed algorithm and score equations are blockers for a reader trying to reproduce or build on the work. I recommend asking the authors to treat the factorization validation and the algorithm corrections as mandatory, rather than as optional clarifications. The evaluation should also be broadened before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is not a paper that merely re-badges known components. The specific design—an analytic SBL channel prior whose variance is updated by EM inside a K-sample joint reverse sampler, with the interference score learned by a diffusion model—is genuinely new relative to the cited literature. The abstract's claim is borne out in the simulations: at SNR=30 dB, SIR=5 dB, both DM-SBL variants give NMSE around -30 dB while SBL sits at -8.9 dB, and the margin holds across the SNR/SIR grid. The mean of samples visibly converges to the true channel in the reverse process.\n\nThe paper is also honest about its debts: it builds on DMPS, PiGDM, and the multi-sample trick from [10,20]. The algebra in Section III is mostly correct and the two approximations are clearly separated. They report runtimes on an RTX 4090 and use public implementations of the baselines. That is solid workmanship.\n\nThe soft spots are real but not fatal. The biggest one is Eq (13), the K-sample factorization of the posterior. It is stated as an 'implementation trick' with no derivation, no error bound, and no ablation over K (only K=256 is used). If that approximation is poor, the sampling procedure is not sampling any well-defined posterior, and the NMSE gains could be partly an artifact of the matched LFM setting. A reviewer should demand at least an ablation over K and some validation against a case where the true posterior is known.\n\nThe second issue is a concrete typo: in Eq (25) the residual for the channel score uses A^H h_t in the parentheses, where the model requires A h_t. A is M x L, so A^H h_t is L-dimensional and cannot be subtracted from y. A literal implementation following the printed equations will not run. This is almost certainly a typo—the PiGDM version in Eq (30) uses A h_t correctly—but it must be fixed.\n\nThird, the EM update in Eq (36) can produce negative variance estimates for small alpha(t), and the paper does not clamp gamma or discuss the issue. This can destabilize the sampler, especially at early reverse steps. Minor but worth fixing.\n\nFinally, the Monte Carlo curves have no error bars, the weights kappa and mu are tuned per experiment (kappa differs between DMPS and PiGDM and between L=200 and L=300), and the closest diffusion baseline—learning scores for both h and n, as in [9]—is not compared. None of these sink the central claim, but they make the empirical evidence less airtight than the headline numbers suggest.\n\nWho is it for: people working on channel estimation under non-Gaussian interference and anyone using diffusion samplers for inverse problems with structured noise. It is a legitimately useful combination and the gains are large enough that I don't think they are a pure simulation artifact. I would send it to a serious referee, but with a request for the K-ablation, error bars, and fixing the equations.","headline":"New combination of EM-tuned SBL prior and diffusion posterior sampling for channel estimation under structured interference, with large reported gains but an unproven K-sample factorization and a dimensional typo that need fixing.","tokens_in":17224,"tokens_out":3557,"would_cite":true,"duration_ms":31907,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A diffusion-guided sparse Bayesian sampler estimates channels under structured interference.","keywords":["channel estimation","diffusion models","structured interference","sparse Bayesian learning","posterior sampling","underwater acoustic communication","score-based generative models","linear inverse problems"],"falsifier":"Run DM-SBL on a synthetic problem in which the true joint posterior is computable, for example a Gaussian channel prior with known gamma and Gaussian structured noise with known covariance, and compare the sampled channel mean with the optimal MMSE estimate; if the factorization approximation is poor at small K or at large t, the NMSE gap would reveal systematic bias instead of near-optimal estimates.","tokens_in":16193,"feed_emoji":"📡","tokens_out":7635,"duration_ms":63784,"temperature":0.7,"pith_summary":"DM-SBL sets out to solve channel estimation when the received signal is corrupted by both additive white Gaussian noise (AWGN) and structured interference, such as a radar or sonar signal sharing the same bandwidth. It jointly estimates the sparse channel and the interference: the interference score is learned once by a diffusion model, and the channel is described by a Gaussian prior whose variance enforces sparsity and is updated by expectation maximization during the reverse diffusion. Two posterior-sampling approximations, DMPS and PiGDM, are used for the noise-perturbed likelihood. The central numerical claim is that both variants dramatically outperform methods that treat interference as AWGN, with the single-realization NMSE at SNR = 30 dB and SIR = 5 dB reaching about $-30$ dB for DM-SBL versus $-8.90$ dB for SBL.","feed_headline":"Channel estimator hits -30 dB error under radar-style interference","feed_subtitle":"The method trains no channel network, only an interference score, yet beats AWGN-based estimators at low SIR.","key_machinery":"The central machinery is a joint reverse SDE run on K parallel samples of both channel and interference. The joint posterior is made tractable by the factorization $p(\\bar{h}_t, \\bar{n}_t \\mid y) = \\prod_{i,j} p(h_t^{(i)}, n_t^{(j)} \\mid y)^{1/K}$ in Eq. (13), which yields the score decomposition in Eq. (15): each channel sample is entangled with all interference samples through averaged likelihood terms, and vice versa. The channel-score half is analytic because $p(h_t^{(i)}; \\gamma)$ is Gaussian with covariance $\\Sigma_{h,t} = 2(1-\\alpha^2(t))I_L + \\alpha^2(t)\\mathrm{diag}(\\gamma)$, while the interference-score half comes from a trained network $s_\\theta(n_t, t)$. DMPS and PiGDM provide the two approximations to the noise-perturbed likelihood, and $\\gamma$ is updated by EM using the sample mean and variance of the K channel samples.","core_discovery":"On its own terms, the paper's discovery is that a channel-specific score network is unnecessary: only the structured interference's score is learned, while the sparse channel is carried by a Gaussian prior $p(h;\\gamma) = \\mathcal{CN}(h; 0, \\mathrm{diag}(\\gamma))$ whose variance $\\gamma$ is re-estimated at every reverse-time step by EM from the parallel channel samples. The joint reverse SDE separates $y = Ah + n + \\epsilon$ into a sparse signal and a structured noise component. At SNR = 30 dB and SIR = 5 dB, DM-SBL (DMPS) and DM-SBL (PiGDM) achieve single-realization channel NMSE of $-29.95$ dB and $-30.47$ dB, respectively, versus $-8.90$ dB for SBL, and at SIR = $-5$ dB DM-SBL remains informative while the baselines fail.","pith_inferences":["Editorial inference: a testable ablation would hold the learned interference score fixed and vary K, since the K-sample averaging in Eq. (13) is doing double duty as both a Monte Carlo estimate and an approximation to the joint posterior; if most of the gain persists even at K = 1, the factorization is less load-bearing than it appears.","Editorial inference: the same joint sampler could be applied to other inverse problems with structured artifacts, such as radar clutter suppression or image reconstruction with correlated sensor noise, provided the artifact class can be simulated for score training.","Editorial inference: the hyperparameters mu and kappa in Eq. (37) are set per scenario with no tuning rule, so a natural extension is to derive them from the relative scales of the prior and likelihood scores rather than fixing them manually.","Editorial inference: because the interference score is the only trained component, deployment cost is dominated by simulating realistic interference, not by collecting channel-specific training data."],"forward_implications":["At low SIR, performance stays stable across SNR, while AWGN-assuming baselines degrade sharply; at SIR = -5 dB DM-SBL still yields useful estimates and the baselines exceed 0 dB NMSE.","Using more parallel samples K and more sampling steps T lowers the NMSE of the averaged estimate, and the mean of samples is consistently better than any single sample.","Because no channel score is trained, the same learned interference score transfers across channel settings (p0 = 10, L = 200 and p0 = 15, L = 300) with only the EM-updated gamma adapting.","Both likelihood approximations work; PiGDM gives slightly better NMSE than DMPS at a modest increase in runtime (about 11 s versus 18 s for the tested setting).","The framework is presented as applicable to other linear inverse problems in which structured interference contaminates measurements."],"supporting_citations":[{"why":"Establishes joint conditional sampling of signal and structured noise by separating the posterior score, the setup DM-SBL builds on.","marker":"[9]"},{"why":"Supplies the K-sample product factorization used in Eq. (13) to make the joint posterior tractable.","marker":"[10]"},{"why":"Provides the DMPS approximation of the noise-perturbed likelihood score used in one DM-SBL variant.","marker":"[11]"},{"why":"Provides the pseudoinverse-guided (PiGDM) likelihood approximation and the denoising estimator used in the other variant.","marker":"[12]"},{"why":"Supplies the score-based SDE framework, the VP-SDE perturbation kernel, and the Predictor-Corrector sampler used in Algorithm 1.","marker":"[17]"},{"why":"Cited alongside [10] as the source of the implementation trick for the K-sample factorization in Eq. (13).","marker":"[20]"},{"why":"Supplies sparse Bayesian learning and its EM variance update, which motivates the channel prior and gamma update in DM-SBL.","marker":"[28]"}],"fun_headline_variants":["DM-SBL: Channel estimation without learning the channel","Only interference learned: DM-SBL hits -30 dB NMSE at low SIR","Channel estimation under radar interference: DM-SBL beats baselines","At SIR -5 dB, DM-SBL stays accurate where SBL fails"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the Eq. (13) factorization approximating the true joint posterior closely enough that the mean of the channel samples is trustworthy; no error bound or validation is given for that approximation.","fun_headline_variants_meta":{"raw":{"variants":["DM-SBL: Channel estimation without learning the channel","Only interference learned: DM-SBL hits -30 dB NMSE at low SIR","Channel estimation under radar interference: DM-SBL beats baselines","At SIR -5 dB, DM-SBL stays accurate where SBL fails"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1274,"prompt_tokens":983,"completion_tokens":291,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":213}},"tokens_in":599,"tokens_out":291,"duration_ms":3555,"temperature":1.0,"reasoning_tokens":213,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:34:27.479108+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DM-SBL on a synthetic problem in which the true joint posterior is computable, for example a Gaussian channel prior with known gamma and Gaussian structured noise with known covariance, and compare the sampled channel mean with the optimal MMSE estimate; if the factorization approximation is poor at small K or at large t, the NMSE gap would reveal systematic bias instead of near-optimal estimates.","supporting_citations":[{"cited_title":"Score-based source separation with applications to digital communication signals,","cited_arxiv_id":null,"evidence_quote":"Supplies the K-sample product factorization used in Eq. (13) to make the joint posterior tractable."},{"cited_title":"Diffusion model based posterior sampling for noisy linear inverse problems,","cited_arxiv_id":null,"evidence_quote":"Provides the DMPS approximation of the noise-perturbed likelihood score used in one DM-SBL variant."},{"cited_title":"Pseudoinverse-guided diffusion models for inverse problems,","cited_arxiv_id":null,"evidence_quote":"Provides the pseudoinverse-guided (PiGDM) likelihood approximation and the denoising estimator used in the other variant."},{"cited_title":"Score-based generative modeling through stochastic differential equations,","cited_arxiv_id":null,"evidence_quote":"Supplies the score-based SDE framework, the VP-SDE perturbation kernel, and the Predictor-Corrector sampler used in Algorithm 1."},{"cited_title":"Sparse bayesian learning and the relevance vector machine,","cited_arxiv_id":null,"evidence_quote":"Supplies sparse Bayesian learning and its EM variance update, which motivates the channel prior and gamma update in DM-SBL."}],"review_version":1}