{"id":"a926157e-68d7-41e0-b4c5-8dfcbb06f351","arxiv_id":"2607.18680","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A structured tensor-decomposition (SCPD) method jointly estimates tower clock offsets and multi-target delay/angle/Doppler parameters in networked ISAC, and fuses estimates across tower pairs to track 2D trajectories and velocities.","lead":"This paper proposes a tensor-decomposition algorithm for networked ISAC, where cell towers reuse their communication pilots as radar: the algorithm separates the radar echoes of different targets and, from the separated echoes, jointly estimates clock offsets between towers and target positions and velocities. Simulations show it beating three baselines and approaching the theoretical estimation bound — but the theoretical argument for reaching that bound has a gap.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CRB-attainment claim rests on a false MLE premise: the LS estimate feeding the Vandermonde projection has colored error covariance, so (24) is not the MLE of the original model.","rationale":"The reader's weakest assumption is exactly the load-bearing point. The paper's own Section IV-A limits the MLE statement to the projection step under an i.i.d. assumption, but the projection input is the output of (22), whose error covariance is \\sigma^2 (D^H D)^{-1} \\otimes I (up to vectorization), not i.i.d. This is an internal inconsistency in the argument, not merely a disagreement with consensus. The CRB derivation itself is a valid computation for the signal model, and Lemma 1's uniqueness conditions are plausible, but neither establishes that the alternating heuristic attains the CRB. The simulations are extensive and consistently show SCPD outperforming the baselines, which supports the practical contribution and should be credited. However, the abstract's theoretical claim should be revised to state that the exactly solved constrained subproblem would be MLE, or that SCPD empirically approaches the CRB in tested scenarios. Since the reader's CONDITIONAL verdict already reflects this gap, no verdict adjustment is needed.","tokens_in":25010,"tokens_out":4363,"duration_ms":50389,"concrete_test":"At a converged iterate of Algorithm 1 with the Fig. 2 parameters, form D = \\hat B_i \\odot \\hat C_i and compute \\Sigma = (D^H D)^{-1}. Check whether \\Sigma is diagonal with constant diagonal entries; if not, the error of the LS estimate (22) is not i.i.d. Gaussian, so the projection (24) is not the MLE of the original model. As a complementary check, re-run the Fig. 2 RMSE curves including failed trials (remove the success-conditioning) and compare against the CRB; if the curves separate substantially, the claimed CRB attainment is an artifact of conditioning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section I claim SCPD 'asymptotically achieves the Cramér–Rao bound,' and Section VI repeats this. The only bridge from the solver to the CRB is the assertion in Section IV-A that the projection step (24) is a maximum likelihood estimator under i.i.d. circular-Gaussian entries of the estimation error matrix. That premise is violated by the preceding step. In subproblem (21)-(22), \\bar X is the unconstrained LS estimate; with i.i.d. Gaussian noise, its error covariance is proportional to (D^H D)^{-1} \\otimes I, not to a scalar identity. For the D used in Algorithm 1 — Khatri-Rao products of Vandermonde factor matrices — (D^H D)^{-1} is generally non-diagonal, so the entries of \\bar X are correlated and non-identically distributed. Therefore the per-column projection (23)-(25) is not the MLE for the original measurement model, and the chain 'CALS → MLE → CRB' is not established. No theorem proves that the fixed point of the alternating projections is the joint MLE or that its variance reaches the CRB. The empirical curves in Figs. 2-6 cannot fill this gap because Section VI-A states that they are averaged only over successful trials, while the CRB is unconditional; this biases the comparison in SCPD's favor. The practical claim that SCPD outperforms the baselines in the simulated scenarios is plausible and supported by the figures, but the headline theoretical claim is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a tensor-based framework for joint time-frequency synchronization and multi-target sensing in networked ISAC. Each BS pair forms a third-order tensor from received MIMO-OFDM pilot signals, decomposes it via a structured canonical polyadic decomposition (SCPD) with Vandermonde factor constraints, and reads off offset and target parameters using bistatic reciprocity. The authors derive identifiability conditions for SCPD, compute a Slepian–Bangs CRB, and present a multi-target tracking algorithm with adaptive beamforming. Simulations compare SCPD against SOE-MP, ESPRIT-LS, and CP-VDM, reporting RMSE gains and success rates.","tokens_in":25307,"tokens_out":4482,"duration_ms":44039,"significance":"If the claims are upheld, the framework would be a useful contribution: it addresses inter-path interference in networked synchronization, exploits Vandermonde structure in a principled way, and fuses estimates from multiple BS pairs for trajectory and velocity tracking. The paper is explicit about the signal model, the reciprocity relations, and the CRB computation, which makes the derivations checkable. The simulation study is extensive and compares against three sensible baselines. However, the central theoretical claim—that SCPD asymptotically achieves the CRB—is not established by the arguments in the manuscript; the evidence offered is empirical and conditional on trial success. This is a load-bearing gap because the abstract and introduction foreground the CRB claim.","major_comments":[{"comment":"The paper claims that the projection step (24) 'constitutes a maximum likelihood estimator' under i.i.d. circular-Gaussian entries of the estimation error matrix, and uses this to assert SCPD asymptotically attains the CRB. This premise is not valid for the algorithm's own output. In subproblem (21)-(22), the unconstrained LS estimate has error covariance proportional to (D^H D)^{-1} ⊗ I, where D is a Khatri-Rao product of Vandermonde factors. Since (D^H D)^{-1} is generally non-diagonal, the entries of the estimate fed into (23)-(24) are statistically correlated and non-identically distributed. Thus (24) is not the MLE of the original measurement model, and the chain 'CALS → MLE → asymptotic CRB' is not established. The paper itself labels the solution to (21) as 'approximate yet efficient'; no theorem shows that the fixed point of the alternating projections is the joint MLE or that it","section":"Sec. IV-A, Eqs. (21)-(25)"},{"comment":"The RMSE curves are averaged only over successful Monte Carlo trials, while the CRB is an unconditional performance bound. The text states: 'the curves in Figs. 2-6 are smoothed by averaging only over successful trials and excluding the failed ones.' Table I reports success rates well below 100% (e.g., 93.73% for SCPD at L=2, SNR=-25 dB). Conditional averaging removes the estimation outliers that contribute to the mean squared error, so the comparison against the CRB is biased in favor of the proposed method. To support 'SCPD approaches the CRB,' the authors should report unconditional RMSE or explicitly frame the curves as conditional on successful detection/association, and discuss how the CRB comparison should be interpreted in that case.","section":"Sec. VI-A, Figs. 2-6"},{"comment":"The statement 'Under mild regularity conditions, maximum likelihood estimates are asymptotically unbiased and capable of reaching the CRB' is invoked to bridge the simulation results to the asymptotic claim. This inference depends entirely on the MLE status of SCPD, which is the same unproven premise as in Section IV-A. The citation [21] is to a tensor-based channel estimation paper and does not supply a theorem for the CALS algorithm here. The empirical curves alone cannot establish the asymptotic claim; they are consistent with the algorithm being a good estimator but not with CRB attainment in the unconditional sense.","section":"Sec. VI-A, text after Fig. 2"}],"minor_comments":[{"comment":"The stopping criterion |ε_{r+1}-ε_r|/ε_r < δ can be undefined when ε_r=0; a safeguard such as checking ε_r > 0 or using an absolute tolerance would make the algorithm description more robust.","section":"Sec. IV-A, Algorithm 1"},{"comment":"The notation M(·) is defined as the mean, but in (26) it is applied to a vector difference of frequency estimates. Since the text later uses V(·) for variance, it would help to state explicitly that the mean is taken over the L target indices.","section":"Eq. (26)"},{"comment":"The uniqueness conditions in (28) are stated as existence conditions on integer pairs {P_t,Q_t}. It would improve readability to give an example of how such pairs are chosen for the standard tensor dimensions, since the proof relies on the spatial smoothing construction without providing a practical selection rule.","section":"Sec. IV-B, Lemma 1"},{"comment":"The success-rate table is informative, but the definition of 'successful trial' is tied to the bistatic-range check in (51). This is a reasonable practical criterion, yet it is not the same as 'the estimator converged to the true parameter'; the conditional RMSE should be interpreted accordingly.","section":"Sec. VI-A, Table I"}],"recommendation":"major_revision","confidential_remarks":"The core algorithm and simulation study are likely of interest to the ISAC signal-processing community. The main obstacle is the unsupported CRB-attainment claim, which appears in the abstract and introduction. If the authors either provide a rigorous asymptotic analysis of the CALS estimator or, more realistically, rephrase the claim as an empirical property and add unconditional or detection-conditional RMSE results, the paper could become publishable. I would also encourage the editor to verify that the self-cited works [18], [19], [21] in fact contain the specific statistical results used to support the MLE/CRB reasoning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious look. It is the first to cast networked-ISAC time/frequency synchronization and multi-target parameter estimation as one structured CPD problem, and the tracking/fusion part with velocity recovery from geometric diversity is a genuine addition. The internal algebra checks out: the reciprocity-based CRB reduction in Section IV-C is coherent, and the estimator equations are consistent with the model. The simulations are extensive and the practical superiority over SOE-MP, ESPRIT-LS, and CP-VDM in these scenarios is credible.\n\nThe soft spot is real, and it is load-bearing. The abstract and Section VI repeat that SCPD 'asymptotically achieves the CRB.' The only bridge for that claim is Section IV-A's statement that the Vandermonde projection step (24) is a maximum likelihood estimator under i.i.d. Gaussian errors. But the input to that projection is the unconstrained LS estimate from (22), whose error covariance is proportional to (D^H D)^{-1} ⊗ I, not to a scalar identity. For the Khatri-Rao structured D used here, the columns of the LS estimate are correlated and non-identically distributed. So the i.i.d. premise is false, and the chain 'CALS → MLE → CRB' has no proof. The simulation curves cannot close that gap because they are averaged only over successful trials while the CRB is unconditional, which biases the comparison in SCPD's favor. That said, the success rates in Table I are high, so the bias is probably not the main reason SCPD looks good.\n\nMinor issues: no code was shipped, the free parameters J_o and η are not given any sensitivity analysis, and the authors' own tensor-based sync method [19] is a natural baseline that is missing. The disjoint-BWP assumption is standard but worth stating as a limitation.\n\nOverall, this is a competent engineering paper with a promising method and a solid simulation package. The theoretical headline is overclaimed. If the authors either prove the CRB attainment under the actual two-step procedure or tone down the claim to 'near-CRB in simulations,' the paper would be solid. I would send it to peer review with a request for major revision on the theory section, not desk-reject it.","headline":"Useful joint sync-and-sensing tensor framework, but the headline CRB claim is not actually established: the paper's own two-step solver breaks the MLE premise.","tokens_in":25883,"tokens_out":3279,"would_cite":true,"duration_ms":30006,"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":"SCPD, a structured tensor decomposition, jointly synchronizes networked ISAC and estimates target parameters, asymptotically reaching the Cramér–Rao bound.","keywords":["networked ISAC","synchronization","canonical polyadic decomposition","Vandermonde structure","Cramér–Rao bound","target tracking","adaptive beamforming","tensor decomposition"],"falsifier":"Compute the error covariance of the (22)-(25) two-step estimate for a single Vandermonde column in Gaussian noise and compare it with the Cramér–Rao bound; if the covariance exceeds the bound, the claim fails. Also, a Monte Carlo with artificially correlated residuals would reveal whether the empirical CRB tracking in Figs. 2–6 is an artifact of the i.i.d. noise assumption.","tokens_in":24819,"feed_emoji":"📡","tokens_out":6430,"duration_ms":53608,"temperature":0.7,"pith_summary":"Networked integrated sensing and communication (ISAC) lets multiple base stations cooperate to sense targets, but this requires accurate time and frequency synchronization. The paper proposes SCPD (structured canonical polyadic decomposition), which decomposes the received sensing tensor into Vandermonde-structured factor matrices so that each multipath component is separated from the others. From the separated components, SCPD jointly estimates timing and carrier-frequency offsets between each base-station pair, along with target delays, Doppler shifts, and angles. The paper further shows how estimates from different pairs can be associated and fused to track target trajectories and velocities, with an adaptive beamformer steering the next snapshot. The paper claims SCPD asymptotically attains the Cramér–Rao bound, the theoretical lower bound on unbiased estimator variance.","feed_headline":"One tensor step syncs and senses at the accuracy limit","feed_subtitle":"Separating multipath echoes lets one tensor decomposition synchronize base stations and track targets at the Cramér–Rao bound.","key_machinery":"The central object is the structured canonical polyadic decomposition (SCPD): a CPD of the received signal tensor in which every factor matrix is constrained to be Vandermonde, meaning each column is a geometric sinusoid with a distinct generator (frequency). The solver is a constrained alternating least-squares (CALS) algorithm that alternates between least-squares updates and a projection onto the Vandermonde set. The Vandermonde generators carry the physical parameters: delay plus offset, Doppler plus CFO, and angles; bistatic reciprocity between the two links of a base-station pair is used to cancel offsets and recover target parameters.","core_discovery":"The central claim is that enforcing Vandermonde structure on the factor matrices of a canonical polyadic decomposition of the received sensing signal separates the multipath components of the channel, and that this separation enables one tensor factorization to estimate both synchronization offsets (timing offset and carrier frequency offset per base-station pair) and target parameters (bistatic delay, Doppler shift, and angles) with an accuracy that asymptotically reaches the Cramér–Rao bound. The paper also establishes identifiability conditions under which the decomposition is unique, and it extends the pairwise estimates to a tracking algorithm that recovers 2D trajectories and velocitie","pith_inferences":["The CRB-optimality claim is only as strong as the MLE premise; the paper's two-step solver (unconstrained least squares followed by Vandermonde projection) is only shown to be an MLE under an i.i.d. Gaussian error assumption that the structured least-squares step does not obviously guarantee, so a direct covariance analysis of the solver is a natural next check.","The same Vandermonde-constrained tensor decomposition could apply to other multi-dimensional harmonic retrieval problems, such as MIMO channel estimation or monostatic radar with multiple snapshots, wherever the measurement tensor is a sum of sinusoids.","The paper assumes disjoint bandwidth parts per base station to avoid interference; an extension where base stations share the same bandwidth would need to handle cross-link interference inside the tensor model.","The tracking algorithm's outlier rejection and consistency check are heuristic; a probabilistic data-association or Bayesian filter could be tested against them in the same simulation setup."],"forward_implications":["Synchronization and sensing happen in a single decomposition step, removing the need for a separate calibration phase or dedicated synchronization pilots.","Because multipath components are separated, the inter-path interference that corrupts compressed-vector methods like SOE-MP is avoided.","If the Cramér–Rao bound is attained, no unbiased estimator of these parameters can achieve lower error variance, so the method is statistically optimal.","The tracking and adaptive beamforming loop keeps a network continuously focused on targets across snapshots, with a consistency check to discard outliers.","The identifiability conditions give a priori guarantees on when the decomposition is unique, and therefore when the parameter estimates are unambiguous."],"fun_headline_variants":["One tensor separates paths to sync and sense at the limit","Structured CP decomposition hits Cramér–Rao for ISAC","Joint sync and sensing solved by one tensor step","Tensor trick syncs base stations and tracks targets"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that SCPD asymptotically attains the Cramér–Rao bound rests on the assumption that the errors left after its approximate two-step (least-squares plus Vandermonde projection) update are independent and identically distributed Gaussian; if that assumption fails, the estimator is no longer maximum likelihood and the CRB argument does not follow.","fun_headline_variants_meta":{"raw":{"variants":["One tensor separates paths to sync and sense at the limit","Structured CP decomposition hits Cramér–Rao for ISAC","Joint sync and sensing solved by one tensor step","Tensor trick syncs base stations and tracks targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1069,"prompt_tokens":748,"completion_tokens":321,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":492,"tokens_out":321,"duration_ms":3867,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:40:14.738190+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the error covariance of the (22)-(25) two-step estimate for a single Vandermonde column in Gaussian noise and compare it with the Cramér–Rao bound; if the covariance exceeds the bound, the claim fails. Also, a Monte Carlo with artificially correlated residuals would reveal whether the empirical CRB tracking in Figs. 2–6 is an artifact of the i.i.d. noise assumption.","supporting_citations":[],"review_version":1}