{"id":"4f10fc7e-501c-47bf-97ca-535319fe9231","arxiv_id":"2501.13472","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A latent-domain plug-and-play ADMM algorithm estimates radio maps from sparse samples using pretrained image denoisers, with recoverability and convergence guarantees.","lead":"Radio map estimation rebuilds signal strength maps from sparse sensor measurements. This paper proposes a plug-and-play method that uses ordinary image denoisers, with no radio-map training data, and includes theoretical guarantees.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's recoverability proof silently assumes ground-truth SLFs lie in the range of the denoiser matrix W; if S♮_r is outside R(Q~_r), v♮_obj is infinite and the bound is unsupported.","rationale":"The reader's weakest-assumption analysis correctly identifies exact CP factorization with known R as a load-bearing modeling condition, but my stress test found a more pointed internal gap in the formal recoverability claim. The proof of Theorem 1 relies on the finite value v♮_obj, yet Lemma 1's induced regularizer includes an indicator that can make the objective at the ground truth infinite unless every S♮_r lies in the range of the denoiser matrix. This is not a matter of physical model mismatch; it is an unstated condition on the interaction between the ground truth and the chosen linear denoiser. Since Assumption 1 permits zero eigenvalues, this is not vacuous. The same issue affects Lemma 2 and the interpretation of Problem (10) as (26) in the KKT proof. I do not think this overturns the paper's empirical contribution: the method is reproducible, the experiments are diverse, and the training-free advantage is real. However, the headline 'formal guarantees' in the strongest claim is not fully supported as written. The appropriate remedy is a revision that either adds the range assumption, restricts to full-rank W, or weakens the theorem accordingly. Because the reader already issued a CONDITIONAL verdict and this concern reinforces that conditionality without invalidating the practical method, I recommend no change to the verdict. The empirical sections, public code, and the real-data comparison are independent support that should be credited even if the theoretical statement needs repair.","tokens_in":28162,"tokens_out":14106,"duration_ms":140529,"concrete_test":"Re-derive Lemma 2 and Theorem 1 keeping the indicator term 1[z ∈ R(Q~)] in v♮_obj. If S♮_r ∉ R(Q~_r) makes v♮_obj = ∞, the current proof collapses. As a numerical cross-check, take W = (1/MN)11^T, R = 1, non-constant nonnegative S♮, c♮ = 1, full observation (|Ω| = MN), and V = 0; run the proposed algorithm and verify that the recovered map has large RSE even though Theorem 1 as stated would predict small error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A's recoverability claim has an unstated feasibility assumption. Lemma 1 (Eq. 25) writes the regularizer induced by a linear denoiser W as r(Z) = (ρ/2λ) z^T Q(Λ^{-1}-I)Q^T z + 1[z ∈ R(Q~)], and the subsequent explicit problem (26) therefore constrains each S_r to R(Q~_r). However, Appendix A defines v♮_obj in Eq. (34) using only the quadratic term plus ζ∥c♮∥² and the noise term, omitting the indicator. If any ground-truth SLF S♮_r is not in R(Q~_r), the true objective at the ground truth is +∞, so the inequality vobj(S⋆) ≤ v♮_obj used to prove Lemma 2 and Theorem 1 is not established. Assumption 1 explicitly allows zero eigenvalues of W, so R(Q~) can be a proper subspace; for example, the rank-one projection W = (1/MN)11^T satisfies Assumption 1 yet only preserves constant SLFs. For non-constant S♮ the bound (30) is then either trivial or false. This gap is independent of whether the CP structure (9) holds exactly: even a perfectly rank-R radio map does not force the SLFs into the range of a given denoiser. The theorem needs an explicit assumption S♮_r ∈ R(Q~_r), or W full-rank, and the same issue propagates to Lemma 2 and to the KKT analysis when Problem (10) is identified with (26).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LaPnP, a plug-and-play ADMM algorithm for radio map estimation. It models the spatio-spectral radio map as a sum of R outer products of spatial loss fields (SLFs) and emitter PSDs, applies a grayscale image denoiser to the SLFs in the latent domain, and uses a fixed number J of HALS inner iterations for the factor updates. The authors provide a recoverability analysis for linear denoisers (Theorem 1), a fixed-point convergence result (Theorem 2), and a KKT convergence result for linear denoisers (Theorem 3). Experiments on statistical-model data, ray-tracing data, and the Mannheim real-world dataset compare LaPnP with interpolation, tensor-decomposition, learned, and data-domain PnP baselines, and show competitive accuracy with lower runtime and no radio-map training data.","tokens_in":28585,"tokens_out":15599,"duration_ms":137157,"significance":"If the theoretical results are made sound, the contribution is significant: LaPnP offers a training-free RME method that leverages powerful natural-image denoisers in a latent domain, reducing the number of denoising operations per ADMM iteration and improving noise robustness. The empirical study is fairly extensive, covering multiple datasets, sampling rates, shadowing levels, SNRs, and a sensitivity analysis for the number of emitters, and the code is publicly released. The recoverability and convergence results would address an important gap in PnP theory for tensor-structured inverse problems. However, the two theoretical gaps identified below affect the central formal claims and need to be resolved before the paper can be accepted.","major_comments":[{"comment":"Lemma 2 and Theorem 1 are not supported as stated because v♮_obj in Eq. (34) is infinite whenever a ground-truth SLF S♮_r is not in the range of Q~_r. Lemma 1 (Eq. (25)) defines r(·) with the indicator 1[z∈R(Q~)], and the explicit problem (26) therefore includes the constraint sr ∈ R(Q~_r). However, Eq. (34) evaluates vobj at {S♮_r,c♮_r} using only the quadratic term plus ζ∥c♮∥² and the noise term, omitting this constraint. Since Assumption 1 permits zero eigenvalues of W, R(Q~_r) can be a proper subspace; for example, the rank-one projection W=(1/MN)11^T satisfies Assumption 1 but maps every image to its constant mean, so any non-constant S♮_r gives v♮_obj=+∞ and the bounds (27a), (27b), and (30) are vacuous. The theorem needs an explicit assumption that S♮_r∈R(Q~_r) for all r (or that W is full rank), and the same issue propagates into Lemma 2 and Theorem 1.","section":"Sec. IV-A, Lemma 2, Eq. (34), Theorem 1"},{"comment":"The proof of Theorem 2 does not cover Algorithm 1 with J>1 HALS inner iterations. In Appendix C, the update of sr is analyzed through the single stationary equation (57), which corresponds to one exact block-coordinate minimization of the subproblem (17), and Eq. (59) then bounds s_r^{(t+1)}−s_r^{(t)} using that equation. But Algorithm 1 runs J=20 inner HALS sweeps before the outer iteration advances; the final s_r^{(t+1)} is the output of J block-coordinate steps, and its relation to s_r^{(t)} involves gradients evaluated at intermediate inner iterates, not at the fixed point of a single stationary equation. Thus the displayed bound (59) is not established for the actual algorithm. The authors should either state and prove Theorem 2 for one HALS step and then handle J>1 by a separate sub-iteration argument, or modify the proof to account for the cumulative effect of the J inner steps.","section":"Algorithm 1, Sec. IV-B, Appendix C (Eqs. (57)–(59))"},{"comment":"Assumption 3 is stated as a boundedness condition on partial gradients of f, but the only justification given ('a large error term will be penalized by f... while large values of cr will be prevented by ζc^T c') is not a proof and does not follow for the iterates of Algorithm 1, which are not shown to remain in a bounded set. Since the fixed-point convergence of Theorem 2 directly relies on the uniform bound L in Eq. (32), the theorem is at present conditional on an unverified property of the algorithm's trajectory. The authors should either prove that the iterates satisfy Assumption 3, or replace it by a verifiable condition such as boundedness of {c_r^{(t)}} and of the measured residuals.","section":"Sec. IV-B, Assumption 3"}],"minor_comments":[{"comment":"The probability space for the statement 'with probability 1−δ' is not specified; please state that Ω is drawn uniformly at random without replacement with fixed cardinality |Ω| and that the probability is over this choice.","section":"Sec. IV-A, Theorem 1"},{"comment":"The order of the HALS updates (18)–(20) is not specified; state whether sr(Ωvec), cr, and sr(Ω∁) are updated cyclically within each of the J inner iterations.","section":"Sec. III-B, Algorithm 1"},{"comment":"The Mannheim experiment sets R=7 for all methods following [10,11]; since the true number of emitters is unknown, a sensitivity analysis for R on the real data would help interpret the comparison.","section":"Sec. V-D"},{"comment":"The variables E and Z in Lemma 1 are used before their relation to the denoising argument is made explicit; please define E and Z clearly.","section":"Lemma 1, Eq. (24)"},{"comment":"The use of Serfling's inequality for sampling without replacement should include the exact statement or a citation with conditions, since the proof relies on it for the high-probability bound.","section":"Appendix B, Eq. (51)"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is solid and the code release is a plus. The main obstacles are in the theoretical sections: Theorem 1 is vacuous for a valid class of linear denoisers without an additional range assumption, and Theorem 2's proof does not match the J>1 HALS update in Algorithm 1. I would be willing to review a revised version addressing these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the central idea is clean and useful: instead of applying a natural-image denoiser to the whole 3D radio map, denoise only the spatial loss fields in a CP-type decomposition, using the PSD factors as multiplicative weights. This makes PnP computationally practical for spectrum cartography and avoids training on radio maps. Second, the empirical story is believable — LaPnP with BM3D/DRUnet is competitive with or better than trained baselines across synthetic, ray-tracing, and real data, and it is robust to noise and training-test mismatch. The code is public.\n\nWhat is genuinely new is the latent-domain formulation itself and the convergence/recoverability analysis that comes with it. The fixed-point analysis extends Chan et al. to a nonconvex, constrained setting and handles the HALS structure, which is not trivial. The recoverability bound is a useful attempt to give sample complexity for PnP without RIP.\n\nNow the soft spots, in order of seriousness. The recoverability proof in Theorem 1/Appendix A has a real gap. Lemma 1 states that the regularizer induced by a linear denoiser includes an indicator term 1[s ∈ R(Q̃)], and the explicit problem (26) carries that constraint. But v♯_obj in Eq. (34) drops the indicator. If a ground-truth SLF S♯_r is not in the range of the principal eigenvector matrix Q̃_r, then vobj(S♯) = +∞ and the displayed equality for v♯_obj is not correct. Since Assumption 1 explicitly allows zero eigenvalues of W, this is not a corner case. The bound then rests on an unstated feasibility assumption. This can be fixed by assuming S♯_r ∈ R(Q̃_r) or W full rank, but it has to be stated.\n\nSecond, Appendix C's proof of Theorem 2 treats the s_r update as an exact block-coordinate minimizer, while Algorithm 1 runs J=20 HALS steps. The proof uses an unconstrained minimizer u and then projects, which is exact for that subproblem, but HALS is only an approximate solver. The convergence of the inexact scheme needs either a tolerance argument or a different analysis.\n\nThird, the recoverability theory covers only linear denoisers, while the headline results use BM3D and DRUnet. That is acceptable, but the paper should be explicit that the theory is illustrative for the nonlinear cases. Minor: the tables lack error bars.\n\nOverall, the engineering contribution is solid and the experiments support the main empirical claims. The theory needs revision, not rejection. Send it to peer review; the right referee will want to see the range-space assumption addressed.","headline":"Latent-domain PnP for radio maps is a solid engineering contribution, but the recoverability proof has an unstated range-space assumption that should be fixed before the theory is trusted.","tokens_in":29037,"tokens_out":2659,"would_cite":true,"duration_ms":23969,"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":"Without radio-map training data, latent-domain PnP reconstructs full spatio-spectral radio maps from sparse sensor measurements.","keywords":["Radio map estimation","spectrum cartography","plug-and-play denoising","ADMM","latent domain","tensor completion","recoverability analysis","training-free"],"falsifier":"Generate a radio map whose shadowing fields differ across frequency bands (so the tensor cannot be written as $\\sum_r \\mathcal{S}_r \\circ \\mathbf{c}_r$ with one set of spatial fields), sample it at increasing rates with noiseless measurements, and run LaPnP with the true $R$: if the relative error does not monotonically decrease toward zero as the sampling rate grows, the exact-factorization assumption is indispensable and the central claim fails for such maps.","tokens_in":27985,"feed_emoji":"📡","tokens_out":9114,"duration_ms":762768,"temperature":0.7,"pith_summary":"The paper claims that a full spatio-spectral radio map can be reconstructed from sparse sensor measurements with no radio-map training data, by running a plug-and-play (PnP) ADMM algorithm that denoises only the latent spatial factors of the map. The key move is to factor the map as a sum of outer products of spatial loss fields and emitter power spectra, $\\mathcal{X} = \\sum_{r=1}^{R} \\mathcal{S}_r \\circ \\mathbf{c}_r$, and let a grayscale natural-image denoiser act as the implicit regularizer on each $\\mathcal{S}_r$. This latent-domain design needs only $R$ denoising calls per iteration instead of one per frequency band, has better noise tolerance than data-domain PnP, and avoids the training cost of deep-learning radio-map estimators. The paper proves that for linear symmetric denoisers every optimal solution is close to the ground truth with a bound that shrinks as the number of sensors grows, and that the ADMM iterates converge to a fixed point, or to a KKT point when the denoiser is linear. If correct, this makes sophisticated image denoisers directly transferable to spectrum cartography without retraining.","feed_headline":"No radio-map training data: sparse sensors still yield full 3D maps","feed_subtitle":"Plugging a natural-image denoiser into radio maps' latent spatial factors matches trained networks with fewer computations.","key_machinery":"The central object is the latent-factor decomposition $\\mathcal{X}(m,n,k) = \\sum_{r=1}^{R} \\mathcal{S}_r(m,n) \\mathbf{c}_r(k)$, where $\\mathcal{S}_r$ is the spatial loss field of emitter $r$ and $\\mathbf{c}_r$ is its power spectral density; this converts the 3D radio map into a small set of 2D images plus spectra. Around this, the paper builds an ADMM whose $\\mathcal{Z}_r$ update is literally a plug-and-play grayscale image denoiser applied to $\\mathcal{S}_r + \\Psi_r$, so the entire prior is supplied by an off-the-shelf denoiser such as DSG-NLM, BM3D, or DRUnet. The theoretical engine is Lemma 1, which identifies such denoisers with proximal operators of explicit quadratic regularizers when the denoising matrix is symmetric and satisfies a Perron-Frobenius-type spectral condition; that lemma turns the black-box algorithm into a tractable optimization problem and enables both the sample-complexity bound and the KKT characterization.","core_discovery":"This paper establishes that plug-and-play denoising can be carried out in the latent factor space of a radio map rather than in the data domain, and that this change is not merely a computational trick but preserves, and in some cases strengthens, theoretical guarantees. The authors model the radio map as $\\mathcal{X} = \\sum_{r=1}^{R} \\mathcal{S}_r \\circ \\mathbf{c}_r$, where $\\mathcal{S}_r$ is the spatial loss field (an $M \\times N$ image) and $\\mathbf{c}_r$ is the frequency spectrum of emitter $r$. Their LaPnP algorithm solves the resulting nonconvex problem by ADMM whose $\\mathcal{Z}_r$ update is exactly a denoising step applied to $\\mathcal{S}_r + \\Psi_r$, allowing any grayscale image denoiser to serve as the prior. Under a spectral condition on the denoising matrix, the implicit denoiser-regularization is identified with an explicit quadratic regularizer, which lets them prove that every optimal solution of the latent formulation lies within a bound of the ground truth that decreases with the number of sensors, and that the ADMM iterates converge to a fixed point, or a KKT point for linear denoisers. The reported experiments with statistical simulations, ray-traced maps, and real indoor measurements show the method matching deep-learning baselines that were trained on hundreds of thousands of radio-map examples.","pith_inferences":["If the exact-factorization premise holds only approximately, the same latent-domain scheme might still work with an overestimated $R$; the paper's own sensitivity experiment hints at this, but it is not proven in the paper.","The latent-domain idea should transfer to other tensor-completion problems whose physics provides a natural outer-product structure, such as hyperspectral imaging with endmember and abundance factors, or multi-channel biomedical arrays.","A testable extension would be to replace the fixed $R$ with an automatic rank-selection rule, since the recoverability analysis assumes $R$ is known.","The theory is established for linear symmetric denoisers; the positive results with BM3D and DRUnet are not yet supported by the same recoverability guarantee, so extending the explicit-regularizer argument to nonlinear denoisers is a natural next step."],"forward_implications":["If the central claim is correct, spectrum cartography can deploy sophisticated image denoisers with zero radio-map training, allowing instant adaptation when a sensing scenario changes.","The recoverability bound (Theorem 1) gives an RIP-free sample-complexity statement for PnP-based radio map estimation: for linear symmetric denoisers, the estimation error is controlled by the number of samples $|\\Omega|$, the noise, and a gap term that shrinks with more measurements.","The convergence results mean practitioners can trust the ADMM output to be a fixed point, and with linear denoisers actually a KKT point of the latent-regularized problem.","Because only $R$ denoising calls are needed per iteration instead of the number of frequency bands $K$, the method becomes substantially cheaper than data-domain PnP, especially when the number of frequency bands is large.","Overestimating the number of emitters $R$ degrades performance only mildly, while underestimating $R$ causes a clear drop, so calibration can err on the side of larger $R$."],"supporting_citations":[{"why":"Introduces plug-and-play priors, replacing the proximal update with a denoiser; the foundation of all PnP methods used here.","marker":"[15]"},{"why":"Surveys plug-and-play theory and algorithms; frames the PnP-ADMM paradigm the paper builds on.","marker":"[16]"},{"why":"Proves fixed-point convergence for data-domain PnP-ADMM; the proof architecture is extended to the latent-domain setting in Theorem 2.","marker":"[25]"},{"why":"Shows symmetric linear denoisers act as proximal operators and supplies the explicit regularizer form relied on by Lemma 1.","marker":"[27]"},{"why":"Derives the quadratic proximal-regularizer equivalence for linear denoisers; used to convert the PnP objective into explicit form (26).","marker":"[33]"},{"why":"Analyzes exact and robust recovery for PnP compressed sensing; the recoverability analysis here lifts that data-domain/RIP result to latent-domain RME.","marker":"[32]"},{"why":"Proves recovery bounds for PnP priors under a restricted eigenvalue condition; the closest prior recoverability result and the one this paper extends beyond data-domain settings.","marker":"[35]"},{"why":"Develops the radio-map factorization $\\mathcal{X} = \\sum_r \\mathcal{S}_r \\circ \\mathbf{c}_r$ and the SPA initialization for deep spectrum cartography; supplies both the model and initialization used by LaPnP.","marker":"[10]"},{"why":"Provides the DRUnet deep denoiser, one of the three denoisers tested in LaPnP.","marker":"[24]"},{"why":"Introduces the LL1 block-term tensor decomposition baseline for spectrum cartography; the strongest tensor-completion competitor in the experiments.","marker":"[7]"}],"fun_headline_variants":["Plug image denoisers into latent space to map radio signals","Latent PnP denoising creates radio maps without RF training data","Sparse sensors plus latent plug-and-play yields full radio maps","Image denoiser as radio prior: latent ADMM trims compute and error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true radio map is exactly a sum of $R$ outer products of a nonnegative spatial field and a nonnegative frequency profile, with $R$ known; if real shadowing is frequency-selective or $R$ is wrong, the denoiser regularizes a latent structure that does not match the data, and the recoverability and convergence guarantees no longer apply.","fun_headline_variants_meta":{"raw":{"variants":["Plug image denoisers into latent space to map radio signals","Latent PnP denoising creates radio maps without RF training data","Sparse sensors plus latent plug-and-play yields full radio maps","Image denoiser as radio prior: latent ADMM trims compute and error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1766,"prompt_tokens":1095,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":593}},"tokens_in":711,"tokens_out":671,"duration_ms":7779,"temperature":1.0,"reasoning_tokens":593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:54:41.711766+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a radio map whose shadowing fields differ across frequency bands (so the tensor cannot be written as $\\sum_r \\mathcal{S}_r \\circ \\mathbf{c}_r$ with one set of spatial fields), sample it at increasing rates with noiseless measurements, and run LaPnP with the true $R$: if the relative error does not monotonically decrease toward zero as the sampling rate grows, the exact-factorization assumption is indispensable and the central claim fails for such maps.","supporting_citations":[{"cited_title":"Plug-and-play priors for model based reconstruction,","cited_arxiv_id":null,"evidence_quote":"Surveys plug-and-play theory and algorithms; frames the PnP-ADMM paradigm the paper builds on."},{"cited_title":"Plug-and-play ADMM for image restoration: Fixed-point convergence and applications,","cited_arxiv_id":null,"evidence_quote":"Proves fixed-point convergence for data-domain PnP-ADMM; the proof architecture is extended to the latent-domain setting in Theorem 2."},{"cited_title":"Plug-and-play priors for bright field electron tomography and sparse interpolation,","cited_arxiv_id":null,"evidence_quote":"Shows symmetric linear denoisers act as proximal operators and supplies the explicit regularizer form relied on by Lemma 1."},{"cited_title":"On plug-and-play regularization using linear denoisers,","cited_arxiv_id":null,"evidence_quote":"Derives the quadratic proximal-regularizer equivalence for linear denoisers; used to convert the PnP objective into explicit form (26)."},{"cited_title":"Recovery analysis for plug-and-play priors using the restricted eigenvalue condition,","cited_arxiv_id":null,"evidence_quote":"Proves recovery bounds for PnP priors under a restricted eigenvalue condition; the closest prior recoverability result and the one this paper extends beyond data-domain settings."},{"cited_title":"Deep spectrum cartography: com- pleting radio map tensors using learned neural models,","cited_arxiv_id":null,"evidence_quote":"Develops the radio-map factorization $\\mathcal{X} = \\sum_r \\mathcal{S}_r \\circ \\mathbf{c}_r$ and the SPA initialization for deep spectrum cartography; supplies both the model and initialization used by LaPnP."},{"cited_title":"Plug-and-play image restoration with deep denoiser prior,","cited_arxiv_id":null,"evidence_quote":"Provides the DRUnet deep denoiser, one of the three denoisers tested in LaPnP."}],"review_version":1}