{"id":"acc78bc5-969e-4edf-9b7c-7bbbd0d667d7","arxiv_id":"2607.20909","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RadioTrace jointly estimates radio maps and transmitter locations by embedding Tx-coordinate gradient updates inside the reverse denoising loop of a frozen diffusion prior, without deployment-time fine-tuning.","lead":"RadioTrace estimates radio maps from sparse signal measurements while jointly locating the transmitters, using a frozen pre-trained diffusion model plus iterative Tx-coordinate refinement. It matters because it removes deployment-time retraining and improves reconstruction when measurements are missing from restricted areas.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multi-Tx validation rests on linear-superposition synthesis; true multi-Tx coupling could break the frozen prior's conditioning and the Tx update.","rationale":"I agree with the reader that the linear-superposition multi-Tx synthesis is the weakest load-bearing assumption. The central mechanism of RadioTrace is that a frozen single-Tx-trained diffusion prior can condition on an inferred multi-Tx map and thereby guide reconstruction; the RadioMapSeer experiments support this only under an additive-power construction that may not hold in real propagation. BART-Lab provides native multi-Tx data, but it is confounded by a large domain shift from the RadioDiff training distribution, so it does not cleanly resolve the concern. Other issues, such as missing error bars, the asymptotic lim-inf theorem with a finite 100-step loop, and the constant learning rate in Algorithm 1 versus the diminishing stepsize in Assumption 3, are real but secondary to the empirical validity of the central claim. A controlled native-versus-summation experiment would settle whether the multi-Tx evaluation is representative, and would directly test the generative prior's capacity that the Tx-update loop relies upon. The reader's CONDITIONAL verdict remains appropriate; nothing in this stress-test moves it further.","tokens_in":19607,"tokens_out":17727,"duration_ms":173406,"concrete_test":"Generate native multi-Tx RadioMapSeer scenes with the same Maxwell-based solver used for the dataset (same buildings and same 23 dBm Tx power, with two or three simultaneous sources), rather than summing single-Tx maps in linear power. Run RadioTrace with identical hyperparameters and masks on both the native scenes and the linear-summed composites built from the same Tx layouts. If NMSE on the native scenes is more than roughly 10% worse than on the summed scenes, or if the reconstruction error is concentrated in regions where summed maps are least physical, the linear-superposition evaluation overstates the multi-Tx capability of the frozen prior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the frozen RadioDiff prior, pre-trained on single-Tx RadioMapSeer maps, to condition on a multi-Tx binary map and generate the composite RSS field. Section V-A constructs the RadioMapSeer multi-Tx ground truth by converting single-Tx maps to linear power, summing them pixel-wise, and converting back to dBm. This assumes additive power with no inter-Tx coupling, no correlated shadowing, and no building-dependent interaction. The only native multi-Tx evaluation, BART-Lab, simultaneously changes the solver, building representation, Tx heights, and Tx power, and is reported only under restricted-area sampling, so it cannot cleanly separate multi-Tx prior capacity from domain shift. If the true multi-Tx field contains interactions the single-Tx-trained generator cannot express, the masked loss in Eq. (7) can be minimized while the Tx update in Eqs. (9)-(11) follows gradients of the wrong generative model, biasing both the RM estimate and the recovered Tx geometry. The linear-superposition test scenes therefore validate the method exactly under the assumption that needs testing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RadioTrace, a radio map (RM) estimation framework that reconstructs the full received-signal-strength (RSS) field from sparse measurements by embedding transmitter (Tx) coordinate refinement into the reverse denoising loop of a frozen, pre-trained diffusion prior. At each reverse step, the current Tx coordinates are converted into a binary Tx map, used to condition the frozen generator, and the masked reconstruction error between the generated clean estimate and the sparse observations is back-propagated through the soft Tx map to update the Tx coordinates. A propagation-guided K-means (PG-KMeans) initializer provides geometry-aware starting Tx positions, and a best-anchor proximal term stabilizes the refinement. The paper also provides a stochastic stability theorem for the Tx-coordinate update under time-varying objectives and reports experiments on RadioMapSeer and BART-Lab under random and restricted-area sampling, with ablations, noise robustness, localization accuracy, extreme sparsity, runtime, and scalability studies.","tokens_in":19768,"tokens_out":6572,"duration_ms":58156,"significance":"If the central claim holds, RadioTrace addresses a practically important limitation of prior-based RM estimation: it avoids deployment-time fine-tuning while explicitly incorporating the physical Tx geometry into the generative reconstruction process. The paper has clear strengths: it evaluates on two datasets across four sampling ratios, includes extensive ablations isolating the contributions of PG-KMeans, momentum GD, and the best-anchor mechanism, reports noise robustness and Tx localization accuracy, and provides the source code. The experimental trajectory of the refinement loss in Section V-L is a useful sanity check for the stability narrative. However, the significance is moderated by two weaknesses: the multi-Tx validation on RadioMapSeer is performed under a linear-superposition synthesis assumption that is exactly the assumption needing test, and the theoretical result is a lim-inf gradient bound for a time-varying objective whose relationship to the actual algorithm is only partially verified.","major_comments":[{"comment":"The RadioMapSeer multi-Tx test scenes are generated by converting each single-Tx map to linear power, summing the maps pixel-wise, and converting back to dBm (Section V-A). This linear-superposition synthesis validates RadioTrace exactly under the assumption that a single-Tx-trained prior can represent multi-Tx fields by additive power composition; it does not test the prior's capacity under inter-Tx coupling, correlated shadowing, or building-dependent interactions. The only native multi-Tx evaluation, BART-Lab, simultaneously changes the solver, building representation, Tx heights, and Tx power and is reported only under restricted-area sampling, so it cannot isolate multi-Tx prior capacity from domain shift. I recommend adding a controlled native multi-Tx experiment (for example, varying R on a dataset with true multi-Tx fields while holding other factors fixed) or narrowing the paper's multi-Tx claims accordingly.","section":"V-A"},{"comment":"Theorem 1 establishes lim inf_{k to infinity} E[||grad Lbar_k(Omega_k)||^2] = 0 for a time-varying stochastic objective. Because Lbar_k itself changes with k and the drift sequence dbar_k is summable, the lim-inf statement can be satisfied even if the algorithm never approaches a stationary point of any fixed objective, for instance if the objective becomes flat over time. The interpretation in Section IV that the refinement 'asymptotically enters a stationary regime' and the claim of 'stability' therefore overstate the result. Please state precisely what is guaranteed: at most, the expected squared gradient of the conditional mean loss vanishes along a subsequence for a sequence of objectives whose drift tends to zero. A corollary for a fixed objective, or a bound on the stationarity gap with respect to the final objective Lbar_T, would make the theoretical contribution substantive.","section":"IV, Theorem 1"},{"comment":"The theorem is conditional on Assumption 2, which requires summable objective drift sum dbar_k, summable gradient bias sum eta_k delta_k, and summable fluctuation sum eta_k sigma_k^2. These conditions are not verified for the actual RadioTrace recursion: the straight-through estimator gradient bias, the rounding error in the binary Tx map, and the temporal variation of Lbar_k due to the diffusion noise schedule are not bounded a priori. Section V-L's empirical curves illustrate behavior on one representative sample only. As written, the analysis is a conditional statement about a generic recursion under assumptions that may not hold for the proposed algorithm; please either prove the conditions for a stylized version of the update, such as with a fixed generator and Gaussian noise, or explicitly frame the theorem as a conditional stability result for a generic stochastic proximal-gradient recursion.","section":"IV, Assumption 2"},{"comment":"The best-anchor term in the objective is defined in Eq. (7) as Omega*_t = argmin over Omega_s, s >= t, of ||xhat_s0(Omega_s) circle M - O||_F^2, but Algorithm 1 updates the anchor only when the current loss improves (line 22). This online approximation makes the proximal target a function of the trajectory, which is not analyzed in the theory of Section IV, where Lbar_k is treated as an exogenous time-varying function. The mismatch should be flagged explicitly, and either the analysis should be extended to the online anchor or the definition in Eq. (7) revised to match the algorithm.","section":"III-C, Eq. (7); Algorithm 1"}],"minor_comments":[{"comment":"Algorithm 1 line 5 assigns samples using the absolute deviation | ||s_k - omega_i|| - d_k |, while Eq. (13) defines the range-consistency distortion with the squared residual (||s_k - omega_i|| - d_k)^2. These are equivalent for the per-sample argmin with fixed centers, but the relation should be stated to avoid confusion.","section":"III-D, Eq. (13); Algorithm 1"},{"comment":"The definition of the weight Otilde_k = log(1 + exp(O(m_k,n_k))) appears in Eq. (15), but the surrounding text refers to 'epsilon > 0 is a small constant' and the notation eOtilde_k is used without a clear definition. Please clean up the notation and define all symbols immediately after the equation.","section":"III-D, Eq. (15)"},{"comment":"Section V-B2 reports the Gaussian width as sigma = 10, while Fig. 9(a) plots PSNR against sigma^2 on a logarithmic axis. State explicitly whether the tuning parameter is sigma or sigma^2 so that the reported range is unambiguous.","section":"V-F, Fig. 9"},{"comment":"All results are reported as point estimates without variance. Since the metrics are averaged over test scenes and the baselines vary widely, especially at 1% restricted-area sampling, reporting standard deviations or compact box plots would substantially strengthen the comparison.","section":"Tables II-IV"},{"comment":"The relationship between this journal paper and the preliminary MLSP version [1] is mentioned only as 'a preliminary version'; please state explicitly which components are new here (PG-KMeans, best-anchor mechanism, stability analysis, and the BART-Lab evaluation) so the incremental contribution is clear.","section":"I, Footnote 1"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is plausible and the experimental work is extensive, but the multi-Tx validation gap and the mismatch between the theoretical analysis and the implemented algorithm are the main obstacles to acceptance. If the authors add a controlled native multi-Tx experiment or clearly scope the multi-Tx claims, and if they substantially revise the theoretical claims to match what Theorem 1 actually establishes, I would be willing to reconsider. The overlap with the authors' own MLSP preliminary version is disclosed; the journal version should more clearly delineate the incremental contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know: RadioTrace is a new mechanism, not a repackaging. It updates transmitter coordinates inside the reverse diffusion loop of a frozen pre-trained radio map generator, so sparse RSS measurements shape the conditioning signal for subsequent denoising steps. Previous prior-based methods like LaPnP and RadioDiff-Inverse only refine the RM-level estimate. That is a real distinction, and the paper is honest about it. The paper does a lot well. The experiments are extensive: two datasets, sampling rates from 1% to 20%, ablations for each component, noise robustness, localization accuracy, scalability, runtime. The ablations show each piece (PG-KMeans init, momentum GD, best anchor) contributes. Code is released, which helps a lot. The PG-KMeans initializer is a neat practical touch, mapping RSS to range circles and then a K-means-like assignment. No deployment-time fine-tuning is a genuine practical advantage. The soft spots are real but not disqualifying. The biggest one is the multi-Tx evaluation. RadioMapSeer only has single-Tx maps, so multi-Tx scenes are synthesized by summing single-Tx maps in linear power and converting back to dBm. That assumes linear superposition, no correlated shadowing, no building-dependent coupling between transmitters. The one native multi-Tx dataset, BART-Lab, is only used under restricted-area sampling and changes solver, building representation, Tx heights and power at once, so it cannot cleanly isolate multi-Tx prior capacity. If real multi-Tx fields contain interactions the single-Tx-trained prior cannot express, the Tx update would follow gradients from the wrong generative model. That is a genuine open question, and the current evidence does not resolve it. Also, there are no error bars or significance tests. The gaps are often large enough that they are likely real, but a few numbers are close (e.g., NMSE vs. RME-GAN under random sampling). The stability theorem is a lim inf bound on expected squared gradient of the conditional mean loss under unverified assumptions; that is okay as a stability statement, but it does not say anything about recovering the true map or Tx locations. Hyperparameters are tuned on the evaluation setting, and the number of transmitters is assumed known. Overall, the central claim holds up for single-Tx and linearly superimposed multi-Tx scenes. The paper is for anyone working on radio map estimation, inverse problems with diffusion priors, or physical latent variable inference. It deserves a serious referee. I would accept it for peer review and ask the referees to push on a cleaner multi-Tx test and some error bars.","headline":"RadioTrace is a genuinely new integration of Tx-coordinate refinement into the reverse diffusion loop for radio map estimation; empirically solid on single-Tx and linearly superposed multi-Tx scenes, but the linear-superposition assumption for multi-Tx validation needs a direct test.","tokens_in":736,"tokens_out":2472,"would_cite":true,"duration_ms":44706,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Radio maps can be rebuilt from sparse sensors by refining transmitter locations inside a frozen diffusion loop.","keywords":["radio map estimation","diffusion prior","transmitter localization","sparse RSS measurements","inverse problem","straight-through estimator","restricted-area sampling","propagation-guided K-means"],"falsifier":"Take a real measured multi-transmitter urban scene with known transmitter positions and strong correlated shadowing between transmitters, sample 1% of the map with a restricted-area mask covering one transmitter, run RadioTrace, and compare its reconstructed map and refined transmitter coordinates against the known ground truth; if the reconstruction error is no better than ordinary Kriging or the localized transmitter error exceeds the initialization error, the claim that the diffusion loop recovers geometry without fine-tuning is refuted. Alternatively, a synthetic experiment that breaks superposition, such as adding a non-linear interaction term between two transmitters' fields before sampling, and showing a sharp PSNR drop would isolate the weakness in the generative prior assumption.","tokens_in":19334,"feed_emoji":"📡","tokens_out":4880,"duration_ms":43690,"temperature":0.7,"pith_summary":"RadioTrace claims that a radio map can be reconstructed from very few received-signal-strength measurements, without retraining or fine-tuning any network, by using a frozen diffusion model as the map generator and treating the unknown transmitter locations as latent variables that are refined inside the denoising loop. The reason this matters is that real deployments face nonuniform and restricted-area sampling, where supervised estimators trained for one sampling pattern degrade; a method that adapts its own physical conditioning variables at inference could generalize without deployment-time data. On the RadioMapSeer and BART-Lab test scenes the paper reports that this loop stays competitive with learned methods under random sampling and maintains strong reconstruction quality when entire regions are unmeasured, while also shrinking transmitter localization error. The paper's central device is an alternating loop: generate a map conditioned on current transmitter positions, then back-propagate the mismatch between the generated map and the sparse observations into those positions, and repeat.","feed_headline":"Radio maps rise from sparse sensors via transmitter-guided diffusion","feed_subtitle":"A frozen diffusion prior rebuilds maps and locates hidden transmitters with no fine-tuning.","key_machinery":"The load-bearing object is the joint RM-generation and Tx-update loop defined by Eq. (7). The frozen diffusion generator $f_\\theta$ maps $(x_t, t, B, C_t)$ to a clean estimate $\\hat{x}_t^0$; the transmitter coordinates $\\Omega_t$ are optimized so that $\\hat{x}_t^0$ matches the masked observations $O$ at sampled cells, while a linearly scheduled proximal term $\\frac{\\kappa_t}{2}\\|\\Omega_t - \\Omega_t^\\star\\|_F^2$ pulls the iterates toward the best anchor found so far, suppressing oscillation as the noise level falls. The binary map is relaxed to a sum of Gaussian heatmaps so that gradients with respect to coordinates can be computed via a straight-through estimator, and the whole loop is seeded by propagation-guided K-means, which maps each RSS value to a range circle and clusters back-projected candidate transmitter positions. The mechanism's work is to make the sparse measurements influence not only the reconstructed map but the physical conditioning variables of the generative prior itself.","core_discovery":"The central claim is that embedding transmitter-coordinate refinement inside the reverse diffusion process of a frozen pre-trained radio-map generator turns sparse RSS measurements into a joint radio-map and transmitter-recovery problem that needs no deployment-time fine-tuning. At each reverse step $t$ the generator produces a clean estimate $\\hat{x}_t^0$ conditioned on a binary transmitter map built from continuous coordinates $\\Omega_t$, and the coordinates are updated by gradient descent on $L_t(\\Omega_t) = \\|\\hat{x}_t^0(\\Omega_t)\\odot M - O\\|_F^2 + \\frac{\\kappa_t}{2}\\|\\Omega_t - \\Omega_t^\\star\\|_F^2$, where $\\Omega_t^\\star$ is the best-anchor coordinate found so far. Because the soft transmitter map is a sum of Gaussian heatmaps, gradients flow through a straight-through estimator even though the conditioning map used by the generator is binary. A propagation-guided K-means initializer converts each measured RSS to a distance circle and clusters back-projected candidate positions to seed the loop. The stability analysis shows that, under bounded drift and perturbation assumptions, the expected squared gradient of the conditional-mean refinement loss tends to zero, so the refinement asymptotically enters a stationary regime.","pith_inferences":["The same \"refine the physical conditioning variable through the frozen generator\" pattern could extend to other inverse problems where a generative prior is conditioned on an unknown geometric object, such as source positions in imaging, antenna orientations, or scatterer locations; the paper does not claim this.","The multi-transmitter test scenes are synthesized by summing single-transmitter maps in linear power, so the method's practical ceiling on real multi-transmitter channels depends on how well propagation obeys superposition; testing on measured multi-transmitter data with correlated shadowing is a natural next step.","The best-anchor proximal schedule suggests a design principle: early diffusion steps explore geometry, late steps lock it in; similar anchor schedules could stabilize other latent-variable-guided diffusion inversions.","Because the method requires only a rough site-calibrated path-loss model for initialization, its accuracy in cluttered indoor or dense-urban environments may hinge on that calibration; an adaptive or learned initializer could remove this dependency."],"forward_implications":["A pre-trained radio-map generator can be reused for new deployment areas and sampling patterns with no retraining, because adaptation happens through the transmitter coordinates at inference.","Under restricted-area sampling, where no measurements exist inside certain regions that may contain transmitters, reconstruction quality is maintained rather than collapsing as it does for sampling-pattern-mismatched supervised models.","The same loop returns transmitter locations as a by-product, reducing mean localization error compared with the initialization across sampling rates.","The stability theorem implies that the coordinate refinement does not diverge as denoising progresses, provided the drift and stochastic perturbation are summable as assumed.","Fewer reverse steps can be used; the paper reports that $T=50$ roughly doubles throughput with small quality loss, and $T=10$ keeps usable quality, so accuracy can be traded against runtime."],"supporting_citations":[{"why":"Supplies the frozen pre-trained diffusion model RadioDiff that RadioTrace uses as the measurement-free radio-map generator.","marker":"[9]"},{"why":"Provides the RadioMapSeer dataset and the RadioUNet-style generative prior lineage used for training and evaluation.","marker":"[8]"},{"why":"LaPnP is the RM-level plug-and-play inverse baseline that RadioTrace contrasts with by updating transmitter conditioning instead of only the reconstructed map.","marker":"[17]"},{"why":"RadioDiff-Inverse is the RM-level diffusion posterior baseline that RadioTrace contrasts with by embedding transmitter coordinates into the reverse loop.","marker":"[18]"},{"why":"Supplies the BART-Lab dataset used for harder multi-transmitter restricted-area evaluation.","marker":"[21]"},{"why":"Provides the straight-through estimator technique that lets gradients flow through the non-differentiable binary transmitter map.","marker":"[19]"},{"why":"Reinforces the straight-through gradient estimator used for the discrete-to-continuous transmitter map relaxation.","marker":"[20]"},{"why":"The U-Net core architecture used inside the pre-trained RadioDiff backbone.","marker":"[22]"}],"fun_headline_variants":["Radio maps from sparse sensors, no fine-tuning needed","Transmitter-guided diffusion rebuilds radio maps on the fly","Sparse measurements, full radio map: diffusion without retraining","Radio map estimation that locates transmitters while denoising","Frozen diffusion prior plus transmitter tracking for radio maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a generator trained on single-transmitter radio maps can faithfully represent multi-transmitter scenes by pixel-wise summing single-transmitter maps in linear power and conditioning on a multi-transmitter binary map; if real multi-transmitter propagation involves non-linear interactions or correlated shadowing that this prior cannot express, both the generated maps and the transmitter-coordinate gradients would be systematically biased.","fun_headline_variants_meta":{"raw":{"variants":["Radio maps from sparse sensors, no fine-tuning needed","Transmitter-guided diffusion rebuilds radio maps on the fly","Sparse measurements, full radio map: diffusion without retraining","Radio map estimation that locates transmitters while denoising","Frozen diffusion prior plus transmitter tracking for radio maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000643,"raw_usage":{"total_tokens":3018,"prompt_tokens":1065,"completion_tokens":1953,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":681,"completion_tokens_details":{"reasoning_tokens":1872}},"tokens_in":681,"tokens_out":1953,"duration_ms":12373,"temperature":1.0,"reasoning_tokens":1872,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:30:04.174049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real measured multi-transmitter urban scene with known transmitter positions and strong correlated shadowing between transmitters, sample 1% of the map with a restricted-area mask covering one transmitter, run RadioTrace, and compare its reconstructed map and refined transmitter coordinates against the known ground truth; if the reconstruction error is no better than ordinary Kriging or the localized transmitter error exceeds the initialization error, the claim that the diffusion loop recovers geometry without fine-tuning is refuted. Alternatively, a synthetic experiment that breaks superposition, such as adding a non-linear interaction term between two transmitters' fields before sampling, and showing a sharp PSNR drop would isolate the weakness in the generative prior assumption.","supporting_citations":[{"cited_title":"RadioDiff: An effective generative diffusion model for sampling-free dynamic radio map construction,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen pre-trained diffusion model RadioDiff that RadioTrace uses as the measurement-free radio-map generator."},{"cited_title":"RadioUNet: Fast radio map estimation with convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the RadioMapSeer dataset and the RadioUNet-style generative prior lineage used for training and evaluation."},{"cited_title":"Radio map estimation via latent domain plug-and-play denoising,","cited_arxiv_id":null,"evidence_quote":"LaPnP is the RM-level plug-and-play inverse baseline that RadioTrace contrasts with by updating transmitter conditioning instead of only the reconstructed map."},{"cited_title":"RadioDiff-Inverse: Diffusion enhanced bayesian inverse estimation for isac radio map construction,","cited_arxiv_id":null,"evidence_quote":"RadioDiff-Inverse is the RM-level diffusion posterior baseline that RadioTrace contrasts with by embedding transmitter coordinates into the reverse loop."},{"cited_title":"Physics-inspired machine learning for radiomap estimation: Integration of radio propagation models and artificial intelligence,","cited_arxiv_id":null,"evidence_quote":"Supplies the BART-Lab dataset used for harder multi-transmitter restricted-area evaluation."},{"cited_title":"Neural discrete representation learning,","cited_arxiv_id":null,"evidence_quote":"Reinforces the straight-through gradient estimator used for the discrete-to-continuous transmitter map relaxation."}],"review_version":2}