{"id":"287d6a5e-8dfc-4456-8043-07e2ae42b5dd","arxiv_id":"2412.06353","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Closed-form Cramer-Rao bounds for 3D extended-target kinematics in ISAC are derived and used to design CRB-aware beamforming, with a graph-neural-network low-complexity alternative.","lead":"This paper derives Cramer-Rao bounds for estimating the position, angles, and orientation of a 3D extended target in an integrated sensing and communication system, and uses those bounds to design transmit beamformers. It also introduces a graph-neural-network beamformer that learns communication and sensing objectives separately and then combines them for fast, scalable deployment.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The far-field small-target approximations in Appendix I-B are violated by the paper's own close-range simulation (d_o=8.7 m with a 5 m vehicle), so the closed-form CRB in Proposition 1 may not be the true CRB for the reported beamforming comparisons.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing issue I would stress-test: Appendix I-B's far-field small-target approximations are not satisfied by the simulation geometry, and Eq. (17) inherits them. This is not a stylistic objection; the approximations replace the actual scatterer-level Jacobian with direction-independent constants mu_2 and mu_3, and with simplified mu_1,k, so the closed-form EFIM may not equal the true Fisher information under the stated simulation parameters. Because every optimization objective and every learned loss uses this CRB, the numerical comparison of beamforming designs could be measuring a bound that is not the true CRB for the modeled target. The proposed concrete check is direct and decisive: recompute the FIM with exact geometry and compare. I did not find a stronger internal inconsistency. The algebraic structure of Eq. (17) is plausible under the stated assumptions, and the paper gives credit where due for a tractable extension to 3D and a modular GNN architecture. The absence of code and of error bars is secondary; the far-field validity check is the one that settles whether the central analytical claim supports the reported conclusions. Since the reader already conditioned acceptance on this issue, my finding does not move the verdict; it specifies the experiment that would test it.","tokens_in":19872,"tokens_out":3472,"duration_ms":39490,"concrete_test":"Numerically compute the exact FIM for the Section VI-A setup without the Appendix I-B approximations. For the same K=38 scatterers, TFS surface coefficients, and a beamformer produced by CRBmin Design-O, compute each scatterer's exact range, azimuth, and elevation from p_k = p_o + V rho_k, then obtain the exact Jacobian partial Theta_k / partial kappa by finite differences (or analytic spherical-coordinate derivatives). Build F_kappa and f_kappa,g using the exact Jacobians in Eqs. (52)-(54) and compare tr(J(kappa)^{-1}) with Eq. (17). Repeat at d_o = 8.7 m and at d_o = 100 m. If the relative difference at d_o = 8.7 m exceeds 10%, or if the design ranking in Fig. 6 changes when the exact CRB replaces Eq. (17), the paper's central comparisons are against the wrong bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Proposition 1, Eq. (17): a closed-form EFIM J(kappa) for the ET kinematic parameters, used as the sensing objective in both optimization problems (P1, P2) and in the GNN losses (49)-(50). The derivation depends on Appendix I-B, where partial derivatives of each scatterer's range, azimuth, and elevation with respect to kappa are replaced by the approximations |p_z,k| << d_perp,k, d_o approximately d_k, and 1/d_o -> 0, yielding the simple vectors mu_1,k, mu_2, mu_3. These approximations are load-bearing because they enter every term of Eq. (17) through the Jacobian of Theta_k with respect to kappa; if they fail, J(kappa) is not the Fisher information of the actual model. Section VI-A places a 5 m long, 2 m wide, 1.2 m high vehicle at d_o = 8.7 m. The front and rear scatterers are then roughly 6-11 m from the BS, so 1/d_o = 0.115 m^-1 is far from 0, and |p_z,k|/d_perp,k reaches about 0.2 for height 1.2 m at 6 m. These are not small numbers. Consequently, mu_2 and mu_3, which are independent of the scatterer index and ignore the dependence of each scatterer's angular coordinates on orientation and on the scatterer's local position, are suspect in exactly the simulated regime. Since all reported trade-off curves, beampatterns, and learning comparisons use Eq. (17) as the sensing performance measure, a mismatch between Eq. (17) and the true CRB at d_o = 8.7 m would mean the paper's quantitative claims are comparisons against a lower bound that does not apply to its own setup. This is the single most load-bearing concern because it directly undermines the validity of the main analytical result in the demonstrated operating regime, not just the optimization heuristics or simulation aesthetics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a monostatic ISAC system in which the radar target is a 3D extended target modeled by a second-order truncated Fourier series surface. The main theoretical contribution is Proposition 1, a closed-form effective Fisher information matrix (EFIM) and hence CRB for the target kinematic parameters (center range, azimuth, elevation, and orientation) as a function of the transmit covariance matrix and the target shape. On this basis, the paper formulates two beamforming problems: a CRB-minimization problem under per-user SINR constraints, solved by semidefinite relaxation, and a weighted CRB/rate problem, solved by successive convex approximation. A graph neural network, ISACBeam-GNN, is then proposed as a low-complexity alternative. Simulations compare the proposed designs with center and average beamforming baselines and evaluate the GNN's scalability.","tokens_in":20278,"tokens_out":11490,"duration_ms":112798,"significance":"The result is potentially valuable: if the EFIM formula is correct, it provides an explicit link between transmit beamforming and the geometry of a 3D extended target, enabling CRB-based joint design in a regime (near-range extended targets) that is often treated only with point-target models. The derivation is transparent and the optimization algorithms are standard SDR/SCA formulations, while the separate-then-integrate GNN architecture is a sensible modular approach to a low-complexity solver. These strengths are real. However, the closed-form CRB relies on far-field small-target approximations that are violated in the paper's own main simulation setting, and the correction term in Eq. (17) appears to miss a factor of t_s. Because the reported trade-off curves and learning comparisons all use Eq. (17) as the sensing objective, these issues directly affect the central quantitative claims.","major_comments":[{"comment":"The EFIM in Proposition 1 is derived under the assumptions |p_z,k| << d_perp,k, d_o ≈ d_k, and 1/d_o → 0, stated in Appendix I-B. The main simulation in Section VI-A uses d_o = 8.7 m with a 5 m × 2 m × 1.2 m vehicle, θ_o = 0, and φ_o = −23°. With the target center at about [0, 8.0, −3.4] m, scatterers on the visible side and top surfaces have p_x ∈ [−2.5, 2.5] m, p_y ∈ [7, 9] m, and p_z ∈ [−4.0, −2.8] m. Hence |p_z,k|/d_perp,k reaches roughly 0.54, 1/d_o = 0.115 m⁻¹ is of the same order as 1/d_k, and edge-scatterer azimuth angles differ from θ_o by as much as about 20°. These are not small numbers. Consequently, the simplified Jacobians µ2 and µ3 in Eqs. (24)–(25), which are constant across scatterers, and the overall EFIM (17) are not the exact CRB for the simulated scenario. The comparisons in Figs. 4–9 are therefore evaluated against an approximate bound of uncontrolled accuracy. The authors should either re-simulate in a regime where the stated assumptions provably hold, or derive the derivatives in Appendix I-B exactly and quantify the approximation error at d_o = 8.7 m.","section":null},{"comment":"The path-loss correction term appears to be missing a factor of the observation period t_s. From Eq. (53), f_κ,g = (2 g t_s N_r / σ_s²) µ4, and from Eq. (54), f_g = (2 t_s N_r / σ_s²) ν4⁻¹. Substituting these into J(κ) = F_κ − f_κ,g f_κ,gᵀ / f_g gives a correction term (2 g² N_r / σ_s²) · t_s · ν4 µ4 µ4ᵀ, not (2 g² N_r / σ_s²) · ν4 µ4 µ4ᵀ as printed in (17). Unless t_s is implicitly set to 1 and never stated to be general, the formula changes the dependence of the CRB on observation time. Since the simulations take t_s = 1 s, the numerical figures are unaffected by this particular typo, but Proposition 1 as a general closed-form statement is not correct as written.","section":null},{"comment":"The Center Design baseline is described as maximizing beam energy toward the ET center, but the displayed optimization problem is min a_oᴴ R_x a_o over {W_n}. As written, the optimum is to direct no energy toward the target, which would make the performance gap in Fig. 6 trivial rather than informative. The sign should be corrected to a maximization, and the constraint set (power, SINR, beam coverage) under which the baseline is solved should be stated explicitly so the comparison is fair.","section":null}],"minor_comments":[{"comment":"The outline says 'In Section IV' twice; the second occurrence, introducing the GNN-based design, should refer to Section V.","section":null},{"comment":"In the second sum of ρ_x(u,v), the term 'sin lv' appears; since the summation is over m and l is undefined there, this should likely be 'sin mv'.","section":null},{"comment":"The simulation setup states Γ = 0 dB for the SINR threshold in P1, but the caption of Fig. 5 specifies Γ = 4 dB for the CRB-min Design-O panels; please clarify which threshold is used in which figure.","section":null},{"comment":"The expressions '102/3' and '108/3' should be formatted as 10^{2/3} and 10^{8/3} to avoid confusion with integer powers.","section":null},{"comment":"There is a typo: 'max acheivable sum rate' should be 'max achievable sum rate'.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a direct extension of the authors' own 2D work [24], and the incremental novelty is acceptable for a journal if the technical issues are resolved. The main barrier to acceptance is the mismatch between the far-field assumptions used in the EFIM derivation and the close-range simulation geometry, together with the apparent missing t_s factor in Eq. (17). Both issues are fixable, but they affect the central contribution and require a substantive revision rather than a cosmetic one."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this is a genuine 3D extension of the authors' 2D ET CRB work, with a closed-form EFIM that depends on the transmit covariance and target shape, plus two beamforming designs and a GNN. The derivation is structured and I didn't find an algebraic contradiction. The 3D TFS surface model and the elevation/orientation terms are new relative to [24], and the separate-then-integrate GNN is a reasonable way to handle varying numbers of users and scatterers.\n\nThe soft spot is serious and it is in the paper's own simulation. Proposition 1 relies on Appendix I-B approximations: |p_z,k| << d_perp,k, d_o ≈ d_k, and 1/d_o → 0. The simulation puts a 5 m vehicle at d_o = 8.7 m with elevation -23°. Scatterers are 6–11 m away, so 1/d_o = 0.115 m^-1, and for a 1.2 m high scatterer at 6 m, |p_z|/d_perp ≈ 0.2. Those are not small numbers. The mu_2 and mu_3 vectors ignore per-scatterer angular dependence, so the EFIM in Eq. (17) is not the true CRB for the simulated geometry. All the beamforming comparisons and GNN losses use Eq. (17) as the sensing metric, so the quantitative claims rest on a bound that doesn't apply to the demonstrated regime.\n\nThis is fixable: either simulate at distances where the far-field assumption holds (tens of meters) or derive the exact Jacobian for near-field. As is, the paper's central theoretical contribution is still plausible under its stated assumptions, but the simulations need to match those assumptions.\n\nMinor issues: the surface discretization into K sections and the area weights S_k are somewhat arbitrary; the balance factors α, β and penalty weights λ1, λ2 are chosen ad hoc; no code or data released; plots have no error bars despite saying they average over 2000 realizations. Also a few typos (\"In Section IV\", \"The confirms\").\n\nBottom line: this is a solid framework paper for the ISAC community, worth a serious referee, but it needs a major revision to align the simulation regime with the approximations, or to lift the approximations. I'd send it to review.","headline":"Real 3D ET CRB extension, but the far-field approximations are violated by the paper's own close-range simulations, so the quantitative claims need rework.","tokens_in":20900,"tokens_out":2366,"would_cite":true,"duration_ms":22665,"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":"A closed-form Fisher information matrix ties a 3D extended target's Cramér-Rao bound to transmit covariance and target shape.","keywords":["integrated sensing and communication","extended target","Cramér-Rao bound","beamforming design","truncated Fourier series","graph neural network","MIMO radar","3D target"],"falsifier":"Compute the exact Fisher information matrix without the far-field approximations for the simulation geometry ($d_o=8.7$ m, 5 m-long vehicle, $\\phi_o=-23^\\circ$) and compare with Proposition 1; if the CRBs differ materially, or if a maximum-likelihood estimator's covariance does not touch the claimed bound, the closed form is not the true bound in that regime.","tokens_in":19607,"feed_emoji":"📡","tokens_out":10084,"duration_ms":92364,"temperature":0.7,"pith_summary":"The paper works out the fundamental accuracy limits for an integrated sensing and communication (ISAC) base station that locates a physically extended 3D target — a vehicle or drone — rather than a point scatterer. It models the target surface with a truncated Fourier series, so arbitrary shapes are represented by a limited set of coefficients, and derives a closed-form Cramér-Rao bound (CRB) for the kinematic parameters: center range, azimuth, elevation, and orientation. The bound is explicit in the transmit covariance matrix and target shape, which turns it into a design tool: the base station can shape its beams to lower the estimation limit while still serving communication users. The paper then formulates two beamforming problems — one minimizing the CRB under per-user SINR constraints, one trading CRB against sum rate — and a graph-neural-network approximation that runs far faster than the optimization solvers. If the bound is correct, ISAC designers get a differentiable, shape-aware sensing objective for joint beamforming.","feed_headline":"Closed-form bound links 3D target shape to ISAC sensing accuracy","feed_subtitle":"Shape-aware beams cut Cramér-Rao bounds for range, angle, orientation; a GNN makes them fast.","key_machinery":"The load-bearing object is the second-order truncated Fourier series (TFS) surface model $\\rho(u,v)=[\\rho_x(u,v),\\rho_y(u,v),\\rho_z(u,v)]^T$, whose coefficient vector $\\varrho$ is assumed known; rotating this surface by the orientation $\\varphi$ and placing it at center $p_o=d_o[\\sin\\theta_o\\cos\\phi_o,\\cos\\theta_o\\cos\\phi_o,\\sin\\phi_o]^T$ gives the 3D target. The visible surface is partitioned into $K$ non-overlapping scatterers with areas $S_k$, and each is assigned a range $d_k$, azimuth $\\theta_k$, and elevation $\\phi_k$. The derivation's engine is the chain-rule factorization of the Fisher information through $\\Theta_k=[d_k,\\theta_k,\\phi_k]^T$ together with the Schur-complement identity $J(\\kappa)=F_\\kappa-f_{\\kappa,g}f_g^{-1}f_{\\kappa,g}^T$ that removes the unknown path-loss $g$. Each term of the final formula has a physical reading: $\\nu_{1,k}$ carries bandwidth-range information, $\\nu_{2,k}$ and $\\nu_{3,k}$ carry azimuth and elevation information from array derivatives, $\\nu_{2,3,k}$ their cross-coupling, and $\\nu_4$ subtracts the information lost to path-loss uncertainty.","core_discovery":"On its own terms, the central discovery is Proposition 1: for a monostatic MIMO ISAC base station illuminating an arbitrarily shaped three-dimensional extended target whose surface is written as a second-order truncated Fourier series, the effective Fisher information matrix for the kinematic parameters $\\kappa=[d_o,\\theta_o,\\phi_o,\\varphi]^T$ is $$J(\\kappa)=\\frac{$2g^{2}$N_r}{\\$sigma_s^{2}$}\\left(\\sum_{k=1}^K[\\nu_{1,k}\\mu_{1,k}\\mu_{1,k}^T+\\nu_{2,k}\\mu_2\\mu_2^T+\\nu_{2,3,k}(\\mu_2\\mu_3^T+\\mu_3\\mu_2^T)+\\nu_{3,k}\\mu_3\\mu_3^T]-\\nu_4\\mu_4\\mu_4^T\\right),$$ where the $\\nu$ coefficients depend on the bandwidth, observation time, scatterer areas, and beam energies $a_k^H R_x a_k$, and the $\\mu$ vectors encode the scatterer geometry and target orientation. Consequently $\\mathrm{CRB}(\\kappa)=J(\\kappa)^{-1}$ gives the estimation limits for center range, azimuth, elevation, and orientation, and these limits depend explicitly on the transmit covariance and on the target's shape. The paper proves this by pushing each scatterer's Fisher information through the intermediate variable $\\Theta_k=[d_k,\\theta_k,\\phi_k]^T$ and then eliminating the unknown path-loss parameter via the Schur complement in (16). It also shows the 3D bound reduces to the point-target CRB only in the zero-elevation, effectively 2D case, so a true 3D extended target is not equivalent to a point.","pith_inferences":["A direct extension is to estimate the shape coefficients $\\varrho$ jointly with kinematics, adding a shape block to the Fisher matrix; the paper assumes $\\varrho$ known, so the joint bound would reveal whether shape uncertainty materially degrades kinematic accuracy.","Because $J(\\kappa)$ is differentiable in $\\mathbf{R}_x$, it could be used as an end-to-end loss for learning-based beamforming without the separate sensing and communication modules, although the paper does not test that.","For close-range large targets of the kind in the self-driving scenario, the far-field approximations in the proof suggest a corrected version of the bound with exact Jacobians; comparing the two at $d_o=8.7$ m would show where the closed form starts to fail.","The same EFIM structure could serve as the measurement-noise covariance for extended-object tracking filters, since it quantifies how shape, bandwidth, and beam energy shape the information available per observation."],"forward_implications":["Beamforming that spreads energy across the visible target surface, forming two distinct main lobes covering the front and rear, outperforms center-point and average-coverage baselines in both sensing CRB and communication trade-off.","Orientation estimation is intrinsically harder than direction estimation: the simulation shows $\\mathrm{CRB}(\\varphi)$ roughly three orders of magnitude above $\\mathrm{CRB}(\\theta_o)$ and $\\mathrm{CRB}(\\phi_o)$.","At fixed radar SNR, the ET CRBs decrease and then converge as distance grows, and they coincide with point-target CRBs only at zero elevation.","The weighted WIM design lets the operator trade sum rate against sensing CRB by a single weight $\\alpha$, allocating more power to strong users as communication weight increases.","ISACBeam-GNN reaches near-optimal beamformers in a fraction of the optimization time and tolerates changes in the numbers of users and scatterers without retraining within the tested range."],"supporting_citations":[{"why":"Supplies the 2D truncated-Fourier-series CRB derivation and CRB-based beamforming formulation that this paper extends to 3D.","marker":"[24]"},{"why":"Provides the parametric Fourier-surface representation adopted for the 3D target surface model.","marker":"[25]"},{"why":"Establishes CRB analysis for extended vehicular targets with known and unknown shape, the baseline approach this paper generalizes.","marker":"[16]"},{"why":"Introduces CRB-based joint radar-communication beamforming with SINR constraints, the optimization paradigm used in Problem P1.","marker":"[7]"},{"why":"Defines the average-coverage beamforming benchmark that the proposed designs are compared against in simulations.","marker":"[21]"},{"why":"Gives the semidefinite-relaxation reformulation with auxiliary matrix used to make the CRB-min problem convex.","marker":"[26]"},{"why":"Provides the first-order Taylor concave surrogate for the sum-rate term used in the SCA algorithm for the WIM problem.","marker":"[27]"},{"why":"Supplies the penalty-method loss formulation used to train the ISACBeam-GNN under constraints.","marker":"[31]"}],"fun_headline_variants":["Shape-aware ISAC beams sharpen 3D extended target sensing","Closed-form CRB ties 3D target shape to ISAC beam design","GNN-based beamformer matches optimal CRB for extended targets","Shape-aware MIMO beams improve ISAC sensing-communication trade-offs","CRB for 3D ETs sparks efficient beamforming with GNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the target is effectively in the far field — each scatterer's height is small compared with its horizontal distance from the base station, its range is nearly the center range, and $1/d_o$ is negligible — which the paper's own simulation of a 5 m vehicle at 8.7 m does not strictly satisfy.","fun_headline_variants_meta":{"raw":{"variants":["Shape-aware ISAC beams sharpen 3D extended target sensing","Closed-form CRB ties 3D target shape to ISAC beam design","GNN-based beamformer matches optimal CRB for extended targets","Shape-aware MIMO beams improve ISAC sensing-communication trade-offs","CRB for 3D ETs sparks efficient beamforming with GNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001009,"raw_usage":{"total_tokens":4399,"prompt_tokens":1216,"completion_tokens":3183,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":832,"completion_tokens_details":{"reasoning_tokens":3100}},"tokens_in":832,"tokens_out":3183,"duration_ms":22487,"temperature":1.0,"reasoning_tokens":3100,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:45:00.267884+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact Fisher information matrix without the far-field approximations for the simulation geometry ($d_o=8.7$ m, 5 m-long vehicle, $\\phi_o=-23^\\circ$) and compare with Proposition 1; if the CRBs differ materially, or if a maximum-likelihood estimator's covariance does not touch the claimed bound, the closed form is not the true bound in that regime.","supporting_citations":[{"cited_title":"Cram ´er-Rao bound analysis and beamforming design for integrated sensing and communication with extended targets,","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D truncated-Fourier-series CRB derivation and CRB-based beamforming formulation that this paper extends to 3D."},{"cited_title":"Boundary finding using Fourier surfaces of increasing order [simulated medical images],","cited_arxiv_id":null,"evidence_quote":"Provides the parametric Fourier-surface representation adopted for the 3D target surface model."},{"cited_title":"Cram ´er-Rao bound analysis of radars for extended vehicular targets with known and unknown shape,","cited_arxiv_id":null,"evidence_quote":"Establishes CRB analysis for extended vehicular targets with known and unknown shape, the baseline approach this paper generalizes."},{"cited_title":"Cram ´er-Rao bound optimization for joint radar-communication beamforming,","cited_arxiv_id":null,"evidence_quote":"Introduces CRB-based joint radar-communication beamforming with SINR constraints, the optimization paradigm used in Problem P1."},{"cited_title":"CRB minimization for RIS-aided mmWave integrated sensing and communications,","cited_arxiv_id":null,"evidence_quote":"Gives the semidefinite-relaxation reformulation with auxiliary matrix used to make the CRB-min problem convex."},{"cited_title":"Joint beamforming for IRS-aided multi-cell MISO system: Sum rate maximization and SINR balancing,","cited_arxiv_id":null,"evidence_quote":"Provides the first-order Taylor concave surrogate for the sum-rate term used in the SCA algorithm for the WIM problem."},{"cited_title":"Learning- based predictive beamforming for integrated sensing and communication in vehicular networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the penalty-method loss formulation used to train the ISACBeam-GNN under constraints."}],"review_version":1}