{"id":"55877dec-d67c-4a07-a09e-51334fdc9f06","arxiv_id":"2505.08267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-beam feedback scheme using LASSO reconstruction on a near-field codebook achieves the same rate as DFT beam training with up to 95% less feedback overhead, plus an off-grid refinement that cuts reconstruction error by 69.4%.","lead":"This paper proposes a beam training scheme for near-field wireless systems that combines the few strongest beams using sparse regression, cutting feedback overhead by up to 95% in simulations. It is a practical candidate for 6G and THz links, where huge antenna arrays make near-field effects unavoidable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"K-largest received-power selection is not proven to be a valid support oracle for the coherent overcomplete near-field codebook, so the overhead-reduction claim rests on an unverified assumption.","rationale":"The reader's weakest assumption is that the channel is sufficiently K-sparse in the polar-domain codebook and that the K strongest beams suffice for LASSO recovery. My concern is the same underlying condition, made more precise: even if the channel is K-sparse in the full codebook, the specific rule 'select the K largest received powers' must also return a set of atoms that contains the true support. For a coherent overcomplete dictionary, this is not automatically true. A strong path can mask a weaker distinct path, and adjacent codewords can receive large projections from the same path, wasting K on redundant atoms. Since the reconstruction in Eq. (13) and Eq. (15) is restricted to the preselected V_K, the selection step is load-bearing. The paper gives only one illustrative coefficient plot (Fig. 2) and an aggregate overhead-versus-L curve (Fig. 5), neither of which verifies support recall. This does not make the paper wrong; the method may work well in typical random geometries, and the simulations are consistent with that. But it means the central claim is conditionally accepted pending a direct test of the support-selection step, especially under unequal path powers and off-grid positions. I therefore leave the reader's CONDITIONAL verdict unchanged, with the added requirement that the reproduction and disclosure include a support-recall analysis rather than only end-to-end rate curves.","tokens_in":8179,"tokens_out":8095,"duration_ms":93036,"concrete_test":"Run the following diagnostic with the same system parameters as Fig. 5: fix N=512 and L=2 with two equal-power paths, one on-grid LoS at (theta=0, r=5 m) and one off-grid NLoS at (theta=0.3, r=4.37 m) not lying on the 1890-codeword near-field grid. In a noiseless setting, sweep all beams, select the K=2 beams with the largest |y_i|, and check whether the projection residual ||h - V_K (V_K^H V_K)^dagger V_K^H h|| is as small as the residual achieved when the true two-path support is known. Repeat for 100 random scatterer positions and report the fraction of trials in which the weaker path is included in the selected support. If that fraction is not close to 1, the K-largest selection rule is not a support oracle and the feedback-overhead claim is conditional on favorable path geometries.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central overhead-reduction result depends on the claim that the K beams with the largest received powers are sufficient to reconstruct the near-field channel. This enters in Section III just before Eq. (10)-(11), where the paper states that 'both schemes exhibit a certain level of sparsity' and then selects the K highest-power beams. For the near-field polar-domain codebook, the columns are non-orthogonal and coherent: a single propagation path at an off-grid angle/distance produces large inner products with many adjacent codewords, while a weaker but distinct path may produce received powers smaller than several redundant neighbors of the strong path. Consequently, selecting the K largest |y(v_i)| can include multiple atoms from one physical path and omit another path entirely. Because Eq. (13) and Eq. (15) restrict reconstruction to the preselected span V_K, any omitted true path is unrecoverable regardless of how well LASSO suppresses noise. The paper's only support-selection evidence is one illustrative example (Fig. 2), with no disclosure of the path parameters used. Fig. 5 shows total required overhead grows with path number L, but it does not report whether the selected set actually contains the true support, nor how often a weaker path is dropped. Without such a check, the headline comparison (K=1 for NF+LASSO vs. K=21 for DFT) may not generalize beyond the specific simulated channel realizations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive near-field beam training framework for multi-user, multipath FDD systems. A base station sweeps a codebook, users feed back the indices plus amplitude/phase of the K strongest received beams, and the BS reconstructs the channel by linear combination (LS or LASSO). Two codebooks are compared: a conventional 512-codeword DFT codebook and a 1890-codeword polar-domain near-field codebook. A continuous off-grid refinement scheme is added to mitigate discretization errors. The central claimed results are that the near-field codebook reduces feedback overhead by up to 95% compared with the DFT codebook, that LASSO provides robustness at low SNR, and that off-grid refinement improves reconstruction accuracy by 69.4% with a compact codebook.","tokens_in":8487,"tokens_out":4941,"duration_ms":54944,"significance":"If the claims hold, the work would be a useful engineering contribution to near-field beam training: it extends the Type-II multi-beam combining idea to the polar domain, uses LASSO to handle the over-complete and coherent near-field codebook, and proposes a plausible off-grid refinement. The paper is clearly written, the signal model is internally consistent, and the plotted trends support the qualitative behavior of the three schemes. The main limitation is that the headline quantitative claims rest on single-point examples and on unverified support-selection assumptions, so the significance is currently conditional on additional validation.","major_comments":[{"comment":"The K-largest received-power selection is not established as a valid support oracle for the coherent over-complete near-field codebook. A single strong propagation path can produce high received power across many neighboring codewords, potentially causing a weaker but distinct path to be excluded from the selected set K. Since the reconstruction in Eq. (13) and Eq. (16) is restricted to the span of V_K, any omitted path is unrecoverable. The paper's only evidence is a single illustrative example in Fig. 2, for which the channel parameters are not disclosed, and Fig. 5 does not report how often the true support is contained in the selected set. To support the central overhead-reduction claim, the authors should provide support-recovery statistics (e.g., probability that all significant paths are selected) over many channel realizations, or compare against an oracle support selector.","section":"Section III, Eq. (11) and Fig. 2"},{"comment":"The reported overhead comparison is not complexity-normalized. The NF and NF+LASSO schemes sweep 1890 codewords while the DFT scheme sweeps 512 codewords, so the two schemes incur different beam-sweeping costs before any feedback is sent. The quantity plotted as 'feedback overhead' counts only the number K of fed-back beam indices, not the total training cost (beam sweeps plus feedback). The headline example 'to attain 7.4 bps/Hz, the NF schemes need K=1 whereas the DFT scheme needs K=21' therefore overstates the overall overhead reduction by ignoring the fourfold difference in M. The authors should report a common complexity metric, for example the total number of beam measurements plus fed-back coefficients, or clearly separate sweep cost from feedback cost.","section":"Section IV, Figs. 3 and 4"},{"comment":"The 69.4% reconstruction-accuracy improvement is a single-point comparison at one feedback overhead value and one SNR, using a specific compact codebook (520 codewords with β=2.384). No Monte Carlo confidence intervals, channel-realization details, or sensitivity to the gradient-descent hyperparameters (learning rate, number of iterations, initialization) are provided. Moreover, the objective in Eq. (17) is nonconvex, yet no convergence analysis is given. The improvement percentage should be reported as a distribution over channel realizations and codebook configurations, and the solver settings should be specified so the result is reproducible.","section":"Section V, Fig. 9"},{"comment":"The LASSO regularization parameter λ is never specified, although it directly controls the trade-off between data fidelity and sparsity and hence the claimed noise robustness. The reported behavior across SNR at fixed feedback overhead could depend on λ being tuned per scenario. The authors should state the λ values used in each figure and provide a sensitivity analysis over a range of λ (and over the near-field codebook density parameter β) to show that the qualitative conclusions do not hinge on particular tunings.","section":"Eq. (15), Figs. 6 and 7"}],"minor_comments":[{"comment":"Please specify the channel parameters used to generate the coefficient-power plots: number of paths L, angles, distances, SNR, and whether the positions are on-grid or off-grid.","section":"Fig. 2 caption"},{"comment":"The simulation section does not state the number of independent channel realizations or the distribution of user/scatterer locations; adding this and showing confidence intervals or error bars would materially improve the reliability of the plotted comparisons.","section":"Section IV, numerical setup"},{"comment":"The notation V_K^H V_K(θ,r) is ambiguous: please clarify that the left factor is the fixed on-grid sub-codebook used for beam sweeping while the right factor is the continuously parameterized codebook, since the received signal y_K was measured with the fixed beams.","section":"Eq. (17)"},{"comment":"There is a typo, 'codeook', in the description of reference [7].","section":"Section I, paragraph 3"},{"comment":"The DFT scheme description says the codebook size is 512 and the beam sweep uses N=512 antennas; it would be helpful to state explicitly that the number of antennas N equals 512 and that the near-field codebook also has 512 angular samples.","section":"Section IV, DFT scheme description"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a reasonable engineering paper, not a breakthrough. The new bit is feeding back the K strongest beams' amplitudes and phases (the 5G Type II idea) and then reconstructing the near-field channel from a polar-domain codebook with LASSO, plus an off-grid refinement. That combination does not appear in the cited prior work, and the simulations are consistent with the qualitative story. The off-grid refinement is the most original piece and seems to help when the codebook is coarse.\n\nWhat the paper does well: the signal model is standard and internally consistent, the LASSO formulation is sensible for an overcomplete non-orthogonal codebook, and the experiments cover low/high SNR, multipath order, and grid mismatch. The figures support the qualitative claims: near-field codebooks behave sparser than DFT, and LASSO helps at low SNR.\n\nWhere it's soft. The headline overhead numbers (e.g., K=1 vs 21) come from single-point examples with asymmetric codebook sizes (1890 vs 512 codewords) that aren't normalized for codebook size or beam-sweeping cost. More importantly, the central mechanism—that the K beams with largest received power are the right support for a coherent overcomplete near-field codebook—is taken as given. For a polar-domain codebook, adjacent codewords are highly correlated; a single strong path can dominate several neighboring beams and crowd out a weaker but distinct path. The paper's only support-selection evidence is one illustrative example (Fig. 2), with no path parameters disclosed. Fig. 5 shows required overhead grows with L, but it doesn't report whether the selected set actually contains the true support. So the overhead-reduction claim may not generalize much beyond the simulated channels.\n\nAlso missing: no code, no error bars, no comparison to the closest prior multi-beam near-field scheme, and the LASSO λ and off-grid hyperparameters are not disclosed. These are addressable gaps, not fatal flaws.\n\nVerdict: the paper is worth a serious referee. The idea is plausible and the problem (near-field beam training overhead in FDD/XL-MIMO) is real. But the authors should be pushed to verify the support-selection assumption directly (e.g., probability of missing a true path), to complexity-normalize the codebook comparison, and to release code or at least detailed parameter settings. If those checks come back positive, the paper would be a solid contribution to an engineering venue.\n\nI wouldn't cite it myself unless I needed a baseline for near-field beam training, but I'd happily send it to review.","headline":"A plausible engineering extension of Type II beam combining to near-field, with a real overhead-reduction claim that currently rests on an unverified support-selection assumption.","tokens_in":9003,"tokens_out":2546,"would_cite":false,"duration_ms":22981,"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":"One feedback beam reaches what 21 DFT beams achieve in near-field multipath channels.","keywords":["near-field communications","beam training","multi-beam combining","LASSO","polar-domain codebook","feedback overhead","off-grid refinement","terahertz communications"],"falsifier":"Simulate or measure a near-field multipath environment with many significant paths, say L at least 20, or with a deliberately coarse near-field codebook: if the feedback overhead needed by NF + LASSO to reach 99% of the perfect-CSI rate approaches or exceeds the DFT scheme's overhead, the sparsity premise is violated.","tokens_in":8027,"feed_emoji":"📡","tokens_out":4475,"duration_ms":44165,"temperature":0.7,"pith_summary":"This paper claims that a base station can train near-field beams in multipath environments by asking each user to report only the K strongest received beams together with their amplitudes and phases, then reconstructing the channel as a weighted combination of those beams. With a polar-domain near-field codebook, this multi-beam combining reaches a given achievable rate with dramatically less feedback than the standard DFT codebook, for example one feedback beam instead of 21 for 7.4 bps/Hz at low SNR. Because near-field codebooks are over-complete and non-orthogonal, the authors add a LASSO step that selects sparse coefficients and suppresses noise, and an off-grid refinement that continues optimizing angles and distances when the codebook grid is coarse.","feed_headline":"One feedback beam reaches what 21 DFT beams achieve","feed_subtitle":"Selecting the K strongest near-field beams and combining them with LASSO cuts feedback overhead up to 95 percent.","key_machinery":"The load-bearing object is the polar-domain near-field codebook, whose codewords are parameterized by both angle and distance to match spherical wavefronts, together with the LASSO regression in Eq. (15) that solves for sparse combination coefficients $\\boldsymbol{\\alpha}_K$ under an $\\ell^1$ penalty. The K selected beams form a sub-codebook $\\mathbf{V}_K$, the measured amplitudes and phases $\\mathbf{y}_K$ provide the data, and LASSO identifies dominant paths instead of blindly inverting the non-orthogonal Gram matrix $\\mathbf{V}_K^H \\mathbf{V}_K$. The off-grid refinement of Eq. (17) then treats angles and distances as continuous variables and optimizes them jointly with the sparse weights, mitigating the discretization error of a finite codebook.","core_discovery":"The central claim is that the near-field channel, expressed in a polar-domain codebook that samples both angle and distance, is sparse enough that the K beams with the largest received powers carry the information needed to reconstruct the channel, and that combining those beams with estimated complex coefficients achieves near-optimal rates. In the simulations this makes the near-field codebook far more feedback-efficient than the DFT codebook: reaching 7.4 bps/Hz requires overhead 1 instead of 21 at 4 dB SNR, and the overall feedback reduction is up to 95%. The LASSO formulation prevents the noise amplification that the direct pseudo-inverse combination suffers from at low SNR, and the off-grid refinement improves reconstruction accuracy by 69.4% with a compact 520-codeword codebook.","pith_inferences":["The K-sparse premise suggests an adaptive protocol that estimates each channel's effective sparsity and chooses K accordingly could cut feedback further, since required overhead grows with path count L.","Because the feedback format is beam indices plus complex coefficients, the scheme could likely be extended to multi-user MIMO where one beam sweep serves several users and the base station assigns beams from the combined set.","The continuous angle-distance optimization in the off-grid refinement points toward a natural extension for beam tracking of moving users, where the previous estimate initializes the next refinement.","A testable prediction is that in rich-scattering environments with dozens of significant paths, the overhead advantage of the near-field codebook narrows or reverses; the paper only simulates up to L=9 paths."],"forward_implications":["For a fixed achievable-rate target, the near-field codebook cuts feedback overhead by up to 95% relative to DFT; at 7.4 bps/Hz only one beam index plus amplitude and phase needs to be reported.","LASSO-based combining keeps channel reconstruction accurate at low SNR, where the plain pseudo-inverse combination degrades because the inverse operation amplifies noise.","The required feedback overhead grows with the number of propagation paths, but the near-field codebook still needs up to 43.75% fewer beams than DFT to reach 99% of the perfect-CSI rate.","When the near-field codebook is made compact (520 codewords), grid mismatch costs about 2.3% of achievable rate, and the off-grid refinement recovers the loss, improving reconstruction accuracy by 69.4%.","The off-grid refinement also closes the gap for coarse DFT codebooks, though the gain is small when the DFT angular grid is already fine."],"supporting_citations":[{"why":"Supplies the polar-domain near-field codebook construction that samples both angle and distance and is used throughout the paper for beam sweeping and sparse channel representation.","marker":"[7]"},{"why":"Provides the DFT-codebook near-field beam training baseline and motivates the energy-split problem that the near-field codebook is designed to avoid.","marker":"[11]"},{"why":"Defines the Fresnel and Rayleigh distances used to set the near-field region in the system model.","marker":"[6]"},{"why":"Describes the 5G Type II codebook framework of linearly combining selected DFT codewords, which the proposed multi-beam combining extends to near-field conditions.","marker":"[12]"},{"why":"Reviews codebooks for CSI feedback in 5G NR and beyond, giving context for the feedback-overhead comparison.","marker":"[13]"},{"why":"Proposes a plane-wave-expansion near-field codebook that serves as a prior extension of Type II combining to near-field and a point of comparison for the DFT-based approach.","marker":"[14]"}],"fun_headline_variants":["95% less feedback: near-field beams with LASSO","Near-field beam training: one overhead beats 21 DFT beams","LASSO combines sparse near-field beams: 95% less feedback, 69% better accuracy","Near-field sparsity cut feedback 95% and boosted accuracy 69%","Sparse near-field beams: LASSO combo cuts feedback by 95%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The near-field channel must be genuinely sparse in the polar-domain codebook, so that the K strongest received beams contain enough information for LASSO to reconstruct the channel; if many paths carry comparable power or the codebook grid is too coarse, the feedback savings disappear.","fun_headline_variants_meta":{"raw":{"variants":["95% less feedback: near-field beams with LASSO","Near-field beam training: one overhead beats 21 DFT beams","LASSO combines sparse near-field beams: 95% less feedback, 69% better accuracy","Near-field sparsity cut feedback 95% and boosted accuracy 69%","Sparse near-field beams: LASSO combo cuts feedback by 95%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0006,"raw_usage":{"total_tokens":2760,"prompt_tokens":860,"completion_tokens":1900,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":1802}},"tokens_in":476,"tokens_out":1900,"duration_ms":12377,"temperature":1.0,"reasoning_tokens":1802,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:58:43.350348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or measure a near-field multipath environment with many significant paths, say L at least 20, or with a deliberately coarse near-field codebook: if the feedback overhead needed by NF + LASSO to reach 99% of the perfect-CSI rate approaches or exceeds the DFT scheme's overhead, the sparsity premise is violated.","supporting_citations":[{"cited_title":"Channel estimation for extremely large-scale mimo: Far-field or near-field?","cited_arxiv_id":null,"evidence_quote":"Supplies the polar-domain near-field codebook construction that samples both angle and distance and is used throughout the paper for beam sweeping and sparse channel representation."},{"cited_title":"Near-field beam training: Joint angle and range estimation with dft codebook,","cited_arxiv_id":null,"evidence_quote":"Provides the DFT-codebook near-field beam training baseline and motivates the energy-split problem that the near-field codebook is designed to avoid."},{"cited_title":"Fraunhofer and fresnel distances: Unified derivation for aperture antennas,","cited_arxiv_id":null,"evidence_quote":"Defines the Fresnel and Rayleigh distances used to set the near-field region in the system model."},{"cited_title":"Csi type-ii codebook of code- books,","cited_arxiv_id":null,"evidence_quote":"Describes the 5G Type II codebook framework of linearly combining selected DFT codewords, which the proposed multi-beam combining extends to near-field conditions."},{"cited_title":"Plane wave expansion-based codebook design for 6g near-field mimo,","cited_arxiv_id":null,"evidence_quote":"Proposes a plane-wave-expansion near-field codebook that serves as a prior extension of Type II combining to near-field and a point of comparison for the DFT-based approach."}],"review_version":1}