{"id":"f0e26819-42b8-4456-ae8d-5d319461b298","arxiv_id":"2412.01029","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Cramér-Rao bounds and a ConvNeXt-based scheme are proposed for near-field user localization in wideband XL-MIMO, reporting simulated distance RMSE near 5 cm.","lead":"A simulated study claims centimeter-level user localization in wideband XL-MIMO systems by combining controllable beam squint with a ConvNeXt deep network, and derives Cramér-Rao bounds for angle and distance under spatial non-stationarity. A generalist might read it to see how 6G base stations could locate users with very large antenna arrays and wide bandwidths.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Centimeter-level claim rests on an unspecified mixed LoS/NLoS simulator: Table I cites only Eq. (7) of [25], without path delays, angles, gains, or training data sampling, so the headline results cannot be reproduced or trusted.","rationale":"The reader's weakest assumption — that the synthetic channel simulator is a faithful proxy for real XL-MIMO propagation — is exactly the load-bearing point I would stress. The paper's analytic CRB results (Eqs. (14)–(15)) are self-contained under the stated LoS model, but the headline DL results and the CHISEL comparisons in Table I cannot be reproduced from the information given. My concern is not that the authors acted in bad faith; it is that the claim of centimeter-level accuracy in mixed LoS/NLoS scenarios is empirically unanchored, and an independent ray-tracing or measured-channel check could easily overturn it. I therefore keep the reader's CONDITIONAL verdict: the contribution is plausible but conditional on releasing code, the full multipath simulator, and validation on independent channels. I additionally noted a concrete reproducibility typo in Algorithm 1, lines 20–21: the update assigns r^j_start twice, and the second assignment should be r^j_end, and the index \\hat{l}^{(j+1)} should be \\hat{l}^{(j-1)}. That typo does not change the central claim about ConvNeXt, but it is another concrete gap in the proposed CBS-BT scheme worth fixing in a revision.","tokens_in":26106,"tokens_out":12567,"duration_ms":103458,"concrete_test":"Obtain the authors' exact mixed LoS/NLoS simulator (or reconstruct it from the missing parameters: delay, angle, and complex gain of the NLoS and diffraction paths, plus the VR mask of Eq. (7) of [25]), retrain ConvNeXt-L and CHISEL-7 on that simulator, then evaluate the fixed trained models on an independent ray-tracing channel generator (e.g., Wireless InSite or a full-wave solver) with the same N=512, fc=100 GHz, B=6 GHz, user range 5–50 m, and angle range ±60°. If the location RMSE at 20 and 30 dB changes by more than a factor of two, or the margin over CHISEL-7 collapses, the claimed centimeter-level accuracy is an artifact of the unverified simulator rather than a physical result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claimed result — ConvNeXt achieves centimeter-level localization accuracy, specifically 0.0746 m location RMSE at 30 dB in Table I for mixed LoS/NLoS non-stationary scenarios — depends entirely on the correctness of a channel simulator that is never specified. Section IV-D states only that the mixed scenario has one LoS path, one NLoS path, and one diffraction path, and that spatial non-stationarity is modeled by Eq. (7) of [25]. However, Eq. (7) of [25] is a visibility-region mask, not a complete multipath generator. The paper does not give the NLoS and diffraction path delays, angles, complex attenuations, their dependence on user position, or how the 50k–100k training samples are drawn from those distributions. Because the ConvNeXt is trained and tested on this same unstated generator, the network could be memorizing simulator idiosyncrasies rather than learning physically robust localization. The surprising result in Table I that ConvNeXt-L performs better in mixed LoS/NLoS (0.0746 m) than in pure LoS (0.0958 m) underscores that the reported behavior is not obviously physical. If the simulator deviates from real XL-MIMO propagation, the abstract's \"centimeter-level accuracy\" claim and the gains over CHISEL are simulation artifacts rather than established performance. This is load-bearing because the paper provides no experimental or independent ray-tracing validation to anchor the simulation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies user localization in wideband near-field XL-MIMO systems under beam squint and spatial non-stationarity. It derives Cramér-Rao bounds (CRBs) for angle and distance estimation, proposes a controllable-beam-squint beam training (CBS-BT) algorithm, and designs a ConvNeXt-based deep learning localization scheme. Simulations show CRB trends with respect to subcarrier count, bandwidth, and SNR, and claim centimeter-level localization accuracy, with comparisons against the CBS method and the CHISEL network.","tokens_in":26444,"tokens_out":6684,"duration_ms":56095,"significance":"The CRB derivation is a useful contribution: it extends prior narrowband or single-carrier bounds to a wideband multi-subcarrier model with a visibility-region mask, and the appendix provides explicit partial derivatives. The CBS-BT idea of grouping subcarriers to refine angle and distance estimates is reasonable, and the FLOPs comparison with CHISEL is a positive feature. The deep-learning results are promising but are entirely simulation-based; the mixed LoS/NLoS scenario used for Table I is not specified, so the headline accuracy claim is currently not reproducible. If the missing model details and the Algorithm 1 correction are provided, the paper could become a solid contribution.","major_comments":[{"comment":"The mixed LoS/NLoS/diffraction channel model used for Table I is not specified. The text states only that there is one LoS path, one NLoS path, and one diffraction path, and that spatial non-stationarity follows Eq. (7) of [25]. The paper does not provide the path delays, angles, complex gains, their dependence on user position, or the distribution used to draw the training and test samples. Without this information, the reported location RMSEs (e.g., 0.0746 m for ConvNeXt-L at 30 dB) and the gains over CHISEL cannot be reproduced or independently verified. This is load-bearing because the abstract's 'centimeter-level accuracy' claim rests on these numbers. Please provide the complete stochastic channel model and the training/test data generation procedure, or replace this section with results on a publicly available channel simulator or measured data.","section":"Section IV-D, Table I"},{"comment":"The distance update in Stage III assigns rj_start twice: 'rj_start = rj−1 c, ˆl(j−1) max −1 and rj_start = rj−1 c, ˆl(j−1) max +1'. The second assignment should be to rj_end. As written, the search interval does not shrink, so the while-loop condition 'rj_end − rj_start ≥ ε' cannot lead to convergence, and the CBS-BT distance estimates in Fig. 8 and Fig. 9 are not obtained from the algorithm as stated. Please correct the typo and confirm that the reported CBS-BT results were generated with the corrected update.","section":"Algorithm 1, line 21"},{"comment":"The observation that ConvNeXt-L performs better in mixed LoS/NLoS (0.0746 m) than in pure LoS non-stationary (0.0958 m) is counterintuitive and is not explained by any physical argument in the text. Since the mixed-path model is itself unspecified, this result may be an artifact of the simulator's particular NLoS/diffraction parameterization. Please either provide a physical explanation supported by the specified model or temper the claim.","section":"Section IV-D, Table I"}],"minor_comments":[{"comment":"In the FIM expression, the second term should contain ∂u_k/∂ρ_{k,j} rather than ∂u^H_k/∂ρ_{k,j}, and the trace term should use indices i,j rather than i,i. The subsequent specialization to R=σ²I yields the correct CRBs, so this appears to be a typographical error, but the displayed general formula is not the standard Slepian-Bangs form.","section":"Section III-A, Eq. (12)"},{"comment":"The RMSE definitions are incorrect as written because they do not square the errors and do not include the averaging factor: for example, RMSE_{k,θ} should be sqrt( (1/Niter) Σ_i (θ̂_{k,i}-θ_{k,i})² ). The reported numbers presumably use the standard definition, but the displayed equations should be corrected.","section":"Section IV-A"},{"comment":"The loss function L = sqrt((ˆr2_k - r_k)²) + sqrt((ˆθ2_k - θ_k)²) is an absolute-error sum, not an RMSE; please correct the formula or the description.","section":"Section III-C.4"},{"comment":"The text uses 'TDD lines' in the paragraph after Eq. (7); this should be 'TTD lines'.","section":"Section II-B"},{"comment":"There are typographical errors in the captions of Fig. 2 and Fig. 3 ('shwon' should be 'shown'), and the caption of Fig. 2 is missing a closing parenthesis.","section":"Figure captions"},{"comment":"The notation rj_start and rj_end is introduced for j starting at 1, but line 14 defines r0_start and r0_end; please clarify the iteration indices and the base case.","section":"Algorithm 1, lines 14-16"}],"recommendation":"major_revision","confidential_remarks":"The CRB part is the strongest contribution and appears internally consistent. The main risk is the DL section's reproducibility: Table I relies on an unspecified multipath simulator, and the Algorithm 1 distance-update typo undermines the CBS-BT results as stated. I would encourage the editor to request the full channel model and, ideally, release of code or data before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the CRB derivation in Eqs. (14)-(15) for the VR-masked, multi-subcarrier beam-squint model. That derivation is internally consistent; the partial derivatives in the appendix check out, and the FIM setup is standard. The paper is also honest that the CBS-BT scheme builds directly on [20] — the subcarrier-grouping refinement is a real, if modest, engineering tweak, and the ConvNeXt postprocessor is a new combination of known pieces. I agree with the reader that this is a solid subfield contribution, not a field reorganization.\n\nThe soft spots are real and mostly about reproducibility. The most serious is Table I. For the mixed LoS/NLoS scenario, the paper says only that there is one LoS, one NLoS, and one diffraction path, and that spatial non-stationarity follows Eq. (7) of [25]. That equation is a visibility-region mask, not a multipath generator. No path delays, angles, gains, or sampling procedure are given. The training and test data come from this same unstated simulator, so the 0.0746 m location RMSE at 30 dB could be a simulator artifact. The surprising result that ConvNeXt-L does better in mixed LoS/NLoS (0.0746 m) than in pure LoS non-stationary (0.0958 m) makes me suspect exactly that. This is load-bearing for the abstract's \"centimeter-level\" claim.\n\nThere are also smaller correctness blemishes. Algorithm 1 line 21 assigns rj_start twice and never updates rj_end; the distance refinement loop cannot work as written. The FIM formula in (12) has a conjugate-transpose typo in the second term (∂u^H_k/∂ρ_j should be ∂u_k/∂ρ_j). The loss function in III-C4 is written as a sum of absolute errors, not RMSE. None of these are fatal, but they add up to a paper that was not carefully proofread.\n\nOn the plus side, the CRB analysis itself does not depend on the unstated simulator, and the comparison with CHISEL is fair in the sense that they matched FLOPs and trained both on the same synthetic data. The paper does not pretend to have measured data; it is a simulation study. The problem is that the simulation's key parameters are missing.\n\nNet: the CRB part deserves publication, the DL part is refereeable but only with the simulator specified and the algorithm fixed. I would send this to peer review, but I would ask for the mixed-scenario channel model, training data generation details, and code or at least full hyperparameters and a corrected Algorithm 1 before it is accepted. The centimeter-level claim should be tempered unless the simulator is independently validated or the data are released.","headline":"A sound CRB extension for near-field wideband XL-MIMO with visibility regions, wrapped in a DL localization paper whose headline centimeter-level claim rests on a simulator the paper never fully specifies.","tokens_in":27012,"tokens_out":2290,"would_cite":true,"duration_ms":23035,"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":"This paper claims that in wideband XL-MIMO, controllable beam squint can be turned from a nuisance into a localization tool, with new Cramér-Rao bounds and a ConvNeXt postprocessor yielding centimeter-level user positions.","keywords":["near-field localization","XL-MIMO","beam squint","controllable beam squint","Cramér-Rao bound","spatial non-stationarity","visibility region","deep learning localization"],"falsifier":"Measure the same N=512, 100 GHz, 6 GHz setup in an anechoic chamber or a calibrated over-the-air testbed with a user at known positions, run the controllable-beam-squint beam training and ConvNeXt schemes, and compare the RMSE against the Cramér-Rao bounds in Eqs. (14)-(15); if the measured RMSE falls below the CRB, or the simulated gains over the baseline vanish with measured channels, the central claims fail.","tokens_in":25887,"feed_emoji":"📡","tokens_out":3800,"duration_ms":37476,"temperature":0.7,"pith_summary":"The paper argues that beam squint, normally a nuisance in wideband XL-MIMO, can be controlled and used as a localization resource. It derives Cramér-Rao bounds for joint angle and distance estimation in a near-field channel that includes both spherical-wave propagation and spatial non-stationarity, where each user sees only part of the array. These bounds show that spatial non-stationarity raises the angle bound noticeably but leaves the distance bound nearly unchanged, and that more subcarriers and wider bandwidth improve both. The authors then propose a two-part localization scheme: a controllable-beam-squint beam training method that groups subcarriers to sweep angles and distances, and a ConvNeXt neural network that refines those estimates. Simulated results put distance RMSE at 0.049 m at 20 dB and location RMSE at 0.0746 m at 30 dB in a mixed LoS and NLoS scenario.","feed_headline":"Beam squint becomes a localization tool, hitting centimeter accuracy","feed_subtitle":"A ConvNeXt scheme turns wideband beam squint into sub-decimeter user positions, with new CRBs to back it.","key_machinery":"The central object is the near-field controllable beam squint effect: true-time-delay-based beamforming lets the first and last subcarriers focus at chosen positions, so all M subcarriers trace a controlled curve in angle-distance space. This is combined with the near-field array response vector masked by a visibility-region indicator in Eq. (4), and the Cramér-Rao bounds in Eqs. (14)-(15) are the inverse of the Fisher information matrix built from the partial derivatives in Eqs. (19) and (22). The ConvNeXt network then maps the received subcarrier powers and the coarse beam-squint estimates to refined angle and distance predictions.","core_discovery":"The paper establishes closed-form Cramér-Rao bounds for angle and distance estimation in a downlink wideband XL-MIMO system under near-field beam squint and visibility-region spatial non-stationarity. The bounds show that spatial non-stationarity degrades angle estimation but has a negligible effect on distance, and that increasing the number of subcarriers and the bandwidth lowers both bounds. On top of this, the paper demonstrates a subcarrier-grouping beam-training scheme that outperforms the prior single-beam-sweep controllable-beam-squint method, and a ConvNeXt-based postprocessor that reaches centimeter-level localization accuracy while comparing favorably with a deep-learning localization baseline at similar or lower computational cost.","pith_inferences":["If the simulator fidelity holds up in real channels, the method points to positioning without extra infrastructure by reusing wideband communication beams, but the learned gains would likely require retraining when the distribution of visibility regions or user angles shifts.","The asymmetric Cramér-Rao behavior suggests a resource-splitting design: allocate bandwidth and subcarriers to distance refinement while dedicating antenna aperture and TTD control to angular resolution.","The observed angle accuracy improving near the array broadside implies coverage quality varies strongly with user angle, so sectorized or multi-array deployments could equalize localization performance across the cell.","A weighted loss that normalizes angle and distance by their respective Cramér-Rao bounds could further improve Cartesian localization accuracy beyond the equal-weight RMSE loss used in the paper."],"forward_implications":["If the Cramér-Rao bounds are correct, system designers can predict localization accuracy directly from the number of antennas, bandwidth, subcarrier count, and visibility region without running Monte Carlo simulations.","Because spatial non-stationarity hurts angle estimation much more than distance estimation, practical designs should spend additional resources on angular refinement rather than on ranging.","The subcarrier-grouping beam-training scheme achieves better angle and distance accuracy than the prior single-sweep controllable-beam-squint method at the same or lower beam-sweeping overhead.","The ConvNeXt postprocessor converts coarse beam-squint estimates into centimeter-level distance accuracy, indicating that learned refinement can substitute for additional beam sweeping in wideband XL-MIMO.","The comparison against the deep-learning baseline suggests that the same beam-squint inputs carry more localization information than model-based processing alone extracts, particularly in mixed LoS and NLoS scenarios."],"supporting_citations":[{"why":"Supplies the controllable beam squint localization concept, the trajectory formulas (8)-(9), and the CBS scheme whose angle and distance estimates serve as inputs to the proposed methods.","marker":"[20]"},{"why":"Provides the visibility-region spatial non-stationary channel model used in Eq. (4) and referenced for the mixed LoS/NLoS and diffraction path modeling in Table I.","marker":"[25]"},{"why":"Gives the Fisher information matrix and Cramér-Rao bound formulation for near-field estimation that the paper adapts to its wideband multi-subcarrier signal model.","marker":"[44]"},{"why":"Supplies the ConvNeXt architecture, including large convolutional kernels and block structure, that the proposed deep-learning localization scheme is built on.","marker":"[39]"},{"why":"Defines the CHISEL deep-learning localization baseline whose channel-scaled variants are compared against ConvNeXt in Table I and Fig. 10.","marker":"[45]"},{"why":"Provides the near-field wideband beam training and TTD-based beamforming background that supports the controllable beam squint signal model.","marker":"[22]"},{"why":"Supplies the delay-phase precoding and TTD-based architecture that underlies the frequency-dependent beamforming vectors in Eq. (7).","marker":"[24]"}],"fun_headline_variants":["Beam squint now locates users to centimeter-level","Turning beam squint into a precise positioning tool","Deep learning tames beam squint for near-field localization","Squint effect becomes a feature for XL-MIMO positioning","Centimeter-accurate user localization using beam squint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed centimeter-level accuracy rests on the assumption that the synthetic channel simulator is a faithful stand-in for real XL-MIMO propagation, including the mixed LoS/NLoS and diffraction paths that the paper does not specify.","fun_headline_variants_meta":{"raw":{"variants":["Beam squint now locates users to centimeter-level","Turning beam squint into a precise positioning tool","Deep learning tames beam squint for near-field localization","Squint effect becomes a feature for XL-MIMO positioning","Centimeter-accurate user localization using beam squint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1563,"prompt_tokens":969,"completion_tokens":594,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":514}},"tokens_in":585,"tokens_out":594,"duration_ms":5354,"temperature":1.0,"reasoning_tokens":514,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:45:48.600786+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same N=512, 100 GHz, 6 GHz setup in an anechoic chamber or a calibrated over-the-air testbed with a user at known positions, run the controllable-beam-squint beam training and ConvNeXt schemes, and compare the RMSE against the Cramér-Rao bounds in Eqs. (14)-(15); if the measured RMSE falls below the CRB, or the simulated gains over the baseline vanish with measured channels, the central claims fail.","supporting_citations":[{"cited_title":"Beam squint assist ed user localization in near-field integrated sensing and co mmunications systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the controllable beam squint localization concept, the trajectory formulas (8)-(9), and the CBS scheme whose angle and distance estimates serve as inputs to the proposed methods."},{"cited_title":"Spat ial non-stationary near-field channel modeling and validat ion for massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Provides the visibility-region spatial non-stationary channel model used in Eq. (4) and referenced for the mixed LoS/NLoS and diffraction path modeling in Table I."},{"cited_title":"Conditi onal and unconditional Cram´ er-Rao bounds for near-field lo calization in bistatic MIMO radar systems,","cited_arxiv_id":null,"evidence_quote":"Gives the Fisher information matrix and Cramér-Rao bound formulation for near-field estimation that the paper adapts to its wideband multi-subcarrier signal model."},{"cited_title":"A ConvNet for the 2020s,","cited_arxiv_id":null,"evidence_quote":"Supplies the ConvNeXt architecture, including large convolutional kernels and block structure, that the proposed deep-learning localization scheme is built on."},{"cited_title":"CHISEL: Compression -aware high-accuracy embedded indoor localization with de ep learning,","cited_arxiv_id":null,"evidence_quote":"Defines the CHISEL deep-learning localization baseline whose channel-scaled variants are compared against ConvNeXt in Table I and Fig. 10."},{"cited_title":"Near-field rai nbow: Wideband beam training for XL-MIMO,","cited_arxiv_id":null,"evidence_quote":"Provides the near-field wideband beam training and TTD-based beamforming background that supports the controllable beam squint signal model."},{"cited_title":"Delay-phase prec oding for wideband THz massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Supplies the delay-phase precoding and TTD-based architecture that underlies the frequency-dependent beamforming vectors in Eq. (7)."}],"review_version":1}