{"id":"530582a7-55c5-44ac-bf5b-bb562c8259ae","arxiv_id":"2506.23473","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-phase cooperative sensing framework for cell-free massive MIMO ISAC, combining CRB-based AP placement and antenna allocation with continuous symbol-level fusion, reports 44% and 41.4% accuracy gains over a grid-based baseline.","lead":"This paper proposes a two-phase system for cell-free massive MIMO ISAC networks: first optimize access point placement and antenna counts to lower sensing error bounds, then fuse distributed radar-like measurements at the symbol level to estimate target locations and velocities. The simulations show 44% better location and 41.4% better velocity accuracy than a grid-based symbol-level baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) uses the CRB chain rule with mismatched Jacobian orientation: P in Eq. (13) is ∂tN/∂t (3L×2), but the formula applies P∈R^{2×3L} in (P J P^T)^{-1}; the Phase-1 CRB objective and the Fig. 9 'Theoretical performance' curves may be invalid.","rationale":"The reader's conditional verdict is appropriate. The strongest claim is the 44%/41.4% accuracy improvement, but that claim is anchored in a two-phase framework whose first phase supplies the theoretical lower bound and the system parameters for the second phase. The most load-bearing link is the CRB transformation in Lemma 1, because every optimization reward in Phase 1 and the benchmark 'Theoretical performance' line in Fig. 9 depends on it. I checked the displayed Jacobian against the text: Eq. (13) is the forward Jacobian G = ∂tN/∂t, while Eq. (5) needs a 2×3L matrix to make (P J P^T)^{-1} a 2×2 CRB. The two formulas are not equivalent for non-square Jacobians, so as written the Phase-1 objective is not the CRB of the target parameters. This is an internal-inconsistency issue rather than a disagreement with the consensus, and it directly affects the optimization claims. The second concern (known RCS, ignoring other targets) is a modeling simplification; it is real but secondary because it could be stated as a model assumption. The concrete test will settle whether the formula actually changes numerical results. If the corrected CRB changes the objective material, Phase 1 and the theoretical curves should be re-derived; Phase 2's relative comparison against MLE/lattice may survive, which is why CONDITIONAL/UNCHANGED rather than REJECT is the right call.","tokens_in":21565,"tokens_out":9207,"duration_ms":81038,"concrete_test":"With the Section V-A setup (L=4, N_A^max=32, ε=10, initial AP positions at (0,0), (350,0), (0,350), (350,350)), pick one target inside the area and compute J from Eqs. (7)–(12), G from Eq. (13), and also form a consistent 2×3L matrix P (either by correctly inverting the mapping or by using the implicit-function entries). Compare tr((G^T J G)^{-1}) with tr((P J P^T)^{-1}) from Eq. (5). If the traces differ, the Phase-1 CRB is mis-specified. Then re-run Algorithm 2 once with the corrected CRB and once with Eq. (5) and compare the optimized objectives and AP placements; a material difference would invalidate Fig. 5 and Fig. 10 as evidence for the two-phase gain. For Theorem 2, repeat the same comparison for T in Eq. (15) against ∂tN/∂v.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's two-phase claim depends on Phase 1 producing a genuine sensing lower bound and on the 'Theoretical performance' curves in Fig. 9. Lemma 1 (Eq. (5)) states CRBp = (P J(tN_u) P^T)^{-1} with P ∈ R^{2×3L}. But the matrix displayed in Eq. (13) contains the derivatives ∂τ/∂x, ∂sinθ/∂x, ∂f_D/∂x and ∂τ/∂y, ∂sinθ/∂y, ∂f_D/∂y — i.e., the 3L×2 Jacobian G = ∂tN/∂t. For this G the standard chain-rule CRB is (G^T J G)^{-1}, not (P J P^T)^{-1} with P = ∂t/∂tN. If P is intended to be ∂t/∂tN so that Eq. (5) is dimensionally consistent, the displayed numerical entries are the derivatives of the forward mapping and are not the entries of ∂t/∂tN. Thus Eq. (5) and Eq. (13) cannot both be correct as stated. This is not a notational quibble: for a non-square G, (G^T J G)^{-1} and (P J P^T)^{-1} are generally different, and the optimized locations/antenna allocations produced by Algorithm 2 (Fig. 5) and the Phase-2 gain demonstrated in Fig. 10 inherit whatever error is in the objective. Theorem 1's proof in Appendix A derives only the FIM and restates the chain rule; it does not resolve the orientation. Theorem 2 is asserted without derivation. Additionally, both CRBs condition on |α_u^l| as known and ignore the other U−1 targets and the randomness of β_S, so even a corrected formula would need to be tested against the actual multi-target likelihood before being used as a benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-phase cooperative sensing framework for cell-free massive MIMO ISAC systems with multiple dynamic targets. Phase 1 jointly optimizes AP placement and antenna allocation by minimizing weighted CRBs for target localization and absolute velocity estimation, using an ADMM-based decomposition with analytical updates and a truncated Newton step. Phase 2 proposes the SL-MDTS scheme, which separates and associates multi-target echoes via tensor decomposition and then performs continuous symbol-level fusion-based localization and velocity estimation through ABC-BFGS optimization. Simulations report that the proposed scheme improves localization and velocity estimation accuracy by 44% and 41.4% over a grid-based symbol-level fusion benchmark, and that the Phase 1 optimized parameters further reduce estimation error at high SNR.","tokens_in":22000,"tokens_out":6898,"duration_ms":76827,"significance":"If the CRB derivations are correct, the two-phase framework is a meaningful contribution to cooperative ISAC sensing: it addresses both system-level resource/placement optimization and a practical multi-target signal processing pipeline with continuous (non-grid) estimation. The SL-MDTS scheme is a useful design that combines tensor-based target separation with symbol-level fusion, and the reported gains are substantial. However, the Phase 1 objective and the 'theoretical performance' benchmarks in Fig. 9 rest entirely on the CRB formula, so the correctness of that formula is load-bearing. The paper also ships no code or dataset; the claims are simulation-based but reproducible in principle from the described algorithms.","major_comments":[{"comment":"The chain-rule formula in Eq. (5) is stated as CRBp = (P J(tN_u) P^T)^{-1} with P ∈ R^{2×3L}, but the matrix displayed in Eq. (13) is the Jacobian G = ∂tN_u/∂t_u ∈ R^{3L×2}: its rows are ∂τ/∂x, ∂sinθ/∂x, ∂f_D/∂x and then the corresponding derivatives with respect to y. For this G the standard CRB transformation is (G^T J G)^{-1}, not (P J P^T)^{-1}. If P is intended to be ∂t/∂tN, then the entries in (13) would need to be the derivatives of the inverse mapping, which they are not. The two interpretations cannot both be correct, and for a non-square G they produce different results. Since the Phase 1 objective (P0) and the 'Theoretical performance' curves in Fig. 9 use this CRB, the optimization results and benchmarks inherit the error. Appendix A only derives the FIM and restates Eq. (5); it does not resolve the orientation mismatch.","section":"Appendix A, Theorem 1 proof"},{"comment":"The velocity CRB is asserted without a derivation. The proof line 'Based on the relationship in (58), we derive the close form of T' does not show how the partial derivatives in (15) are computed, and the displayed T matrix has the same orientation problem as P in (13): its rows are derivatives with respect to the velocity parameters, so the formula should be (T^T J T)^{-1} if T = ∂tN_u/∂v_u. Because the velocity CRB enters the objective (P0) and the 'Theoretical performance' curve in Fig. 9(b), a complete and correct derivation is required.","section":"Appendix A, Eq. (53)–(57)"},{"comment":"The CRB is derived from a single-target likelihood, Eq. (53), in which the echo of target u is considered in isolation and the complex amplitude αu_l is treated as a known constant. The original signal model (1) contains U coupled targets and αu_l depends on the random RCS β_S ∼ CN(0, σβ^2). The FIM in (57) therefore is not the FIM of the actual observation model unless multi-target interference and the randomness of β_S are explicitly shown to be negligible or are integrated out. This assumption is not stated or justified in the manuscript, and it undermines the use of the CRB as a benchmark and as the Phase 1 objective.","section":"Eq. (1), Eq. (53)–(57)"}],"minor_comments":[{"comment":"The matrix element written as [R]_{l,l} should be [F]_{l,l}; this appears to be a typographical error in the block matrix notation of Theorem 1.","section":"§IV-B, problem (P3)"},{"comment":"The low-complexity MUSIC step uses a single-snapshot covariance Rcm = ˆaξ_d,l(ˆaξ_d,l)^H, which is rank-one; the noise subspace is then the null space of this rank-one matrix. The authors should clarify how this null space is formed reliably in the presence of noise and why the single-snapshot construction does not degrade the distance estimate, since a standard MUSIC would require averaging over snapshots.","section":"§IV-B"},{"comment":"The text in Section V-B3 refers to the 'SPO method' while the method is introduced as 'SFO' (symbol-level fusion-based optimization). Please unify the terminology.","section":"§V-C, Fig. 10"},{"comment":"The two-phase collaboration is demonstrated only at SNR 60–78 dB (Fig. 10), which is far outside the 0–18 dB range used in the rest of the paper. The authors justify this choice, but such a regime weakens the practical claim that Phase 1 optimization yields sensing gains; a lower-SNR demonstration or a qualitative argument that the ordering persists at realistic SNRs would strengthen the paper.","section":"§V-C"},{"comment":"The sentence 'The [29] prove that the factor matrices of Yl satisfy Kruskal's condition' is ungrammatical and the citation is ambiguous; please rephrase and specify which result in reference [29] is used.","section":"§IV-A1"}],"recommendation":"major_revision","confidential_remarks":"The Jacobian orientation issue in Eq. (5) versus Eq. (13) is a genuine load-bearing correctness problem, not a stylistic quibble: it directly affects the Phase 1 objective and the 'Theoretical performance' benchmarks in Fig. 9. That said, the SL-MDTS scheme and the simulation comparisons are largely independent of the CRB derivation, so the paper is repairable within its scope. I recommend major revision rather than rejection, and I suggest the editor ask the authors to (i) correct the chain-rule formula and the Jacobian definitions, (ii) provide a full derivation of Theorem 2, and (iii) justify or amend the single-target known-RCS assumptions in the CRB model. The authors should also verify their claim that absolute velocity estimation in CF-mMIMO ISAC cooperative sensing has not been addressed in prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2506.23473. The genuinely new piece is the Phase 2 estimator: continuous symbol-level fusion with ABC-BFGS for multi-dynamic-target CF-mMIMO, with tensor preprocessing. That part is credible and the simulation evidence against MLE and the lattice baseline is informative. The reported 44% and 41.4% accuracy gains are plausible for the fusion method itself.\n\nThe problem is Phase 1. Lemma 1 states the CRB as (P J P^T)^{-1} with P ∈ R^{2×3L}. But the matrix displayed in Eq. (13) is the forward Jacobian ∂tN/∂t, not ∂t/∂tN. The entries are derivatives of τ, sinθ, fD with respect to x,y – that's the transpose of what the chain rule needs given the stated P. If P is meant to be ∂t/∂tN, the displayed entries are wrong; if P is ∂tN/∂t, the formula should be (P^T J P)^{-1}. Either way, Eq. (5) and Eq. (13) together are not correct. This matters because the Phase 1 objective minimizes this CRB, so the optimized AP placements and antenna allocations in Fig. 5 inherit whatever error is in the formula. Theorem 2 is asserted without a proof, and the FIM is derived for a single target with known RCS amplitude, ignoring multi-target interference and random RCS. These are load-bearing issues for the theoretical contribution.\n\nThe Fig. 10 SNR range (60–78 dB) is odd but the authors disclose why they chose it; I'd call that a minor limitation, not a fatal one.\n\nNet: The Phase 2 fusion algorithm may survive a corrected Phase 1, but the paper as submitted cannot be accepted because the theoretical foundation for the joint optimization is shaky. The citations and self-citations look appropriate; no red flags there. The writing is clear enough.\n\nRecommendation: send to peer review, but make it conditional on fixing the CRB derivation, proving Theorem 2, and re-running Phase 1 after correction. A referee who checks the Jacobian math will catch this quickly.","headline":"Phase 2 symbol-level fusion is a plausible incremental advance, but the Phase 1 CRB has a Jacobian-orientation error that undermines the theoretical benchmarks; referee it conditionally.","tokens_in":22577,"tokens_out":5423,"would_cite":false,"duration_ms":52614,"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":"The paper claims that jointly optimizing AP placement and antenna allocation and then fusing echoes at symbol level improves localization accuracy by 44% and velocity accuracy by 41.4% in cell-free massive MIMO ISAC systems.","keywords":["cell-free massive MIMO","integrated sensing and communication","cooperative sensing","Cramer-Rao bound","symbol-level fusion","AP placement optimization","multi-target localization","OFDM ISAC"],"falsifier":"Run a Monte Carlo maximum-likelihood estimator on OFDM echo signals simulated with the optimized AP positions and antenna counts, and compare its empirical mean-squared error to the trace of the localization CRB in Eq. (6): if the estimator's MSE falls below the claimed bound, or if the Jacobian P in Eq. (13) does not have exactly two rows and 3L columns, the Phase-1 CRB derivation and the theoretical curves in Fig. 9 are not valid.","tokens_in":21303,"feed_emoji":"📡","tokens_out":11444,"duration_ms":103549,"temperature":0.7,"pith_summary":"The paper tries to show that cooperative sensing in a cell-free massive MIMO integrated sensing and communication (ISAC) network becomes more accurate when two design steps are coupled. The first step is a joint optimization of where the distributed access points (APs) are placed and how many antennas each one uses, with the objective of minimizing the Cramér-Rao bound (CRB), the lower bound on how precisely any unbiased estimator can locate a target and estimate its absolute velocity. The second step is a symbol-level fusion-based multi-dynamic target sensing (SL-MDTS) scheme that separates and associates multi-target echoes, combines the raw delay, angle, and Doppler measurements from all APs with noise-based weights, and solves continuous optimization problems to obtain target positions and velocities without grid search. If correct, the framework gives a concrete deployment-and-reception recipe for high-precision sensing in ISAC networks, with reported gains of 44% in localization accuracy and 41.4% in velocity accuracy over a grid-based symbol-level fusion baseline.","feed_headline":"Cooperative ISAC sensing gains 44% accuracy","feed_subtitle":"Joint AP placement and symbol-level fusion cut location and velocity errors in ISAC networks.","key_machinery":"The paper's machinery has two load-bearing parts. The first is the Cramér-Rao bound as a design objective: Lemma 1 uses a Jacobian chain rule to convert the Fisher information matrix of delay, angle, and Doppler into a CRB matrix for target location (P in Eq. (13)) and for absolute velocity (T in Eq. (15)), making the sensing precision a differentiable function of AP coordinates and antenna counts. The second is the SL-MDTS fusion engine: CP tensor decomposition separates multi-target echoes into rank-one delay/angle/Doppler components, a low-complexity MUSIC-assisted step associates them across APs, and the associated components are stacked into delay, angle, and Doppler matrices (S, K, T), weighted by inverse estimated noise variance in a maximal-ratio-combining style, and matched against continuous fusing matrices in optimization problems P3 and P4 solved by ABC-BFGS. The key effect of the second part is to avoid grid search, so estimation is continuous and immune to the off-grid loss the paper identifies in earlier symbol-level fusion schemes.","core_discovery":"The central claim is that the two phases reinforce each other: optimizing the system geometry and antenna resources against the CRB improves the achievable sensing lower bound, and the SL-MDTS signal-processing stage converts that theoretical gain into lower actual estimation error for multiple moving targets. The paper derives closed-form CRB expressions for location and absolute velocity in Section II (Theorems 1 and 2), formulates the joint AP placement and antenna-allocation problem as a minimax optimization (P0), and solves it by ADMM with analytic projections and truncated Newton updates. For the estimation stage, it decomposes each AP's echo tensor via CANDECOMP/PARAFAC (CP) decomposition, associates the rank-one components to physical targets using a low-complexity MUSIC-assisted search, and builds weighted delay, angle, and Doppler matrices that fuse all APs' symbol-level information before solving localization problem (P3) and velocity problem (P4) with an ABC-BFGS hybrid algorithm. The simulation section reports that this SL-MDTS scheme improves localization accuracy by 44% and absolute velocity estimation accuracy by 41.4% relative to the grid-based symbol-level lattice fusion method, and that the Phase-1 optimized parameters further reduce the estimation ARMSE compared with initial deployments.","pith_inferences":["A reader could extend the same Jacobian-and-CRB chain to 3D tracking by adding a height coordinate and elevation angle; the paper notes the extension but does not develop it.","The reported gains are tied to the specific grid-based baseline and simulation parameters, so a fair field or standards-compliant OFDM comparison against denser-grid or data-level fusion baselines would clarify how the gain behaves in practice.","Since the CRB derivation treats the radar cross-section amplitude as known and ignores the other U-1 targets, live multi-target tracking with fluctuating RCS might need an augmented CRB or online estimation of the cross-section; that is a natural robustness check rather than a result claimed by the paper."],"forward_implications":["The proposed Algorithm 2 outputs concrete AP locations and antenna counts that lower the sensing CRB over a target area, so the Phase-1 procedure can serve as a planning tool for CF-mMIMO ISAC deployments.","The SL-MDTS scheme estimates multiple dynamic targets by turning the multi-target problem into parallel single-target problems through tensor decoupling and data association, so the sensing pipeline scales with the number of targets approximately linearly.","The continuous ABC-BFGS optimization replaces grid-based search, which the paper argues removes the off-grid error that limits earlier symbol-level fusion schemes.","The two-phase results show that optimized system parameters further reduce actual estimation ARMSE compared with initial parameters, connecting deployment choices to end-to-end sensing accuracy.","The reported 44% and 41.4% gains are measured against the grid-based lattice fusion baseline and the MLE baseline, offering a quantitative benchmark for future cooperative ISAC sensing schemes."],"supporting_citations":[{"why":"Introduces symbol-level fusion for multi-BS cooperative sensing and supplies the grid-based lattice method that serves as the main benchmark for the 44% and 41.4% accuracy gains.","marker":"[5]"},{"why":"Provides the MLE-based localization method used as the second benchmark in the performance comparison.","marker":"[7]"},{"why":"Supplies the FIM/Jacobian chain-rule result that Lemma 1 uses to convert delay, angle, and Doppler information into position and velocity CRBs.","marker":"[23]"},{"why":"Provides the ADMM framework used to decompose the joint placement and antenna-allocation problem into the subproblems in Algorithm 2.","marker":"[25]"},{"why":"Provides the numerical optimization background and the truncated Newton method used to update the high-dimensional variable z in subproblem (P2.4).","marker":"[27]"},{"why":"Supplies the tensor CANDECOMP/PARAFAC decomposition framework used to separate multi-target echoes at each AP.","marker":"[28]"},{"why":"Provides the regularized alternating least-squares solution and uniqueness condition for the CP decomposition of the echo tensor.","marker":"[29]"},{"why":"Supplies the eigenvalue-based noise variance estimation used to build the maximal-ratio-combining weights in the symbol-level fusion matrices.","marker":"[30]"},{"why":"Supplies the artificial bee colony algorithm that, combined with BFGS, solves the non-convex localization and velocity estimation problems.","marker":"[32]"}],"fun_headline_variants":["Cooperative ISAC sensing boosts accuracy 44%","Cell-free MIMO ISAC cuts localization errors 44%","Joint AP placement and fusion improve ISAC 44%","Symbol-level fusion lifts ISAC sensing accuracy","ISAC cooperative sensing: 44% better location"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole Phase-1 performance benchmark rests on the chain-rule transformation that converts delay, angle, and Doppler measurements into target position and velocity being mathematically exact, with the radar cross-section amplitude treated as known; if that transformation or that model is wrong, the optimized CRB is not the true sensing lower bound, and the theoretical curves in Fig. 9 are not valid benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["Cooperative ISAC sensing boosts accuracy 44%","Cell-free MIMO ISAC cuts localization errors 44%","Joint AP placement and fusion improve ISAC 44%","Symbol-level fusion lifts ISAC sensing accuracy","ISAC cooperative sensing: 44% better location"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1430,"prompt_tokens":1080,"completion_tokens":350,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":273}},"tokens_in":696,"tokens_out":350,"duration_ms":3743,"temperature":1.0,"reasoning_tokens":273,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:43:22.209455+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a Monte Carlo maximum-likelihood estimator on OFDM echo signals simulated with the optimized AP positions and antenna counts, and compare its empirical mean-squared error to the trace of the localization CRB in Eq. (6): if the estimator's MSE falls below the claimed bound, or if the Jacobian P in Eq. (13) does not have exactly two rows and 3L columns, the Phase-1 CRB derivation and the theoretical curves in Fig. 9 are not valid.","supporting_citations":[{"cited_title":"Symbol-level integrated sensing and communication enabled multiple base stations cooperative sensing,","cited_arxiv_id":null,"evidence_quote":"Introduces symbol-level fusion for multi-BS cooperative sensing and supplies the grid-based lattice method that serves as the main benchmark for the 44% and 41.4% accuracy gains."},{"cited_title":"Device-free sensing in OFDM cellular network,","cited_arxiv_id":null,"evidence_quote":"Provides the MLE-based localization method used as the second benchmark in the performance comparison."},{"cited_title":"Target localization accuracy gain in MIMO radar-based systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the FIM/Jacobian chain-rule result that Lemma 1 uses to convert delay, angle, and Doppler information into position and velocity CRBs."},{"cited_title":"Distributed optimization and statistical learning via the alternating direction method of multipliers,","cited_arxiv_id":null,"evidence_quote":"Provides the ADMM framework used to decompose the joint placement and antenna-allocation problem into the subproblems in Algorithm 2."},{"cited_title":"Nocedal and S","cited_arxiv_id":null,"evidence_quote":"Provides the numerical optimization background and the truncated Newton method used to update the high-dimensional variable z in subproblem (P2.4)."},{"cited_title":"Tensor decompositions and applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor CANDECOMP/PARAFAC decomposition framework used to separate multi-target echoes at each AP."},{"cited_title":"Multipath Component-Aided Signal Processing for Integrated Sensing and Communication Systems,","cited_arxiv_id":null,"evidence_quote":"Provides the regularized alternating least-squares solution and uniqueness condition for the CP decomposition of the echo tensor."},{"cited_title":"Noise level estimation based on eigenvalue learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the eigenvalue-based noise variance estimation used to build the maximal-ratio-combining weights in the symbol-level fusion matrices."},{"cited_title":"On the performance of artificial bee colony (ABC) algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies the artificial bee colony algorithm that, combined with BFGS, solves the non-convex localization and velocity estimation problems."}],"review_version":1}