{"id":"a0b2c33b-9e2c-4ffa-867c-8e3c64c45dd3","arxiv_id":"2501.02730","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A K-SVD-learned codebook operating in the wavenumber domain can precode for near-field and far-field users simultaneously, without region identification, and beats DFT and polar codebooks in simulations.","lead":"This paper proposes using a machine-learning method called K-SVD to learn a precoding codebook from wavenumber-domain channel data, so one base station codebook can serve both near-field and far-field users in 6G MIMO without knowing each user's region. Simulations suggest it outperforms standard DFT and polar codebooks in mixed-field multi-user scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unified codebook's load-bearing premise is that TR 38.901 near-field channels are sparse in the wavenumber domain; the paper asserts this via prior work but never validates it for the simulated near-field geometry.","rationale":"I read the paper as a proposal that a dictionary learned by K-SVD on wavenumber-domain channel estimates can serve both near- and far-field UEs without region labels. The strongest support is the simulation curves in Figs. 4-5, but those curves inherit an unvalidated representation. Section III-A literally asserts the wavenumber-domain sparse representation and cites prior work; it does not prove or verify sparsity for TR 38.901 near-field channels. The concrete check above would settle whether that premise holds. If sparsity holds, the central claim is plausible and the missing implementation details are addressable; if not, the proposed pipeline cannot deliver the claimed unification. The reader's CONDITIONAL verdict is therefore appropriate: the authors should either provide wavenumber-sparsity evidence or release the simulation code and channel model details. I found no more severe internal inconsistency; the main gap is an unvalidated load-bearing premise, not a mathematical error.","tokens_in":9184,"tokens_out":7434,"duration_ms":82062,"concrete_test":"Reproduce the Fig. 4b simulation with the exact TR 38.901 near-field extension (spherical-wave array response, 32x32 UPA, carrier frequency and user distances), compute the normalized wavenumber-domain coefficients of each channel realization using the transform from Sec. III-A, and measure the energy fraction captured by the top S coefficients (e.g., S=64 and S=256) for all near-field UEs. If the median energy fraction is below 90% for S=256, the sparse-recovery premise fails; additionally, retrain the K-SVD codebook on a disjoint training set and evaluate on held-out realizations to rule out overfitting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A states that Fourier harmonic expansion represents both far-field and near-field channel responses as a sparse superposition of plane waves [3],[14], and the rest of the paper (K-SVD learning in III-B, the 'no region identification' claim in the abstract) depends on that representation being accurate and sparse. The paper provides no derivation, no transform definition, and no sparsity measurement for the specific 3GPP TR 38.901 channels used in Figs. 4-5. A point source at finite distance has a continuous plane-wave spectrum; over a finite 32x32 UPA aperture, near-field spherical wavefronts generally produce a broad, non-sparse wavenumber footprint unless the channel is dominated by a few discrete scatterers. The asserted sparsity is an empirical property that must be checked under the simulated near-field distances, array geometry, and cluster angles. Fig. 2 compares OMP/MRF in wavenumber vs angular domains but omits the channel model, array parameters, and near-field configuration, so it cannot support the premise. If the wavenumber-domain representation is not (approximately) sparse for TR 38.901 near-field channels, the compressive-sensing channel estimation and the K-SVD dictionary update have no theoretical basis, and the claimed universal 'regression' codebook may simply be overfitting the simulation data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified codebook paradigm for serving coexisting near-field and far-field UEs in multi-user MIMO. The scheme combines wavenumber-domain channel estimation (imported from the authors' earlier work), K-SVD dictionary learning to generate a regression-based codebook, and standard PMI feedback/precoding, with constant-modulus projection for analog/hybrid implementations. The authors claim that the learned codebook 'removes the need for the base station to identify whether one specific UE stays in either near-field or far-field regions' and that simulations show spectral efficiency superior to DFT and polar codebooks in mixed and homogeneous near/far-field scenarios, while remaining compatible with 5G workflows and reference-signal structures.","tokens_in":9462,"tokens_out":3710,"duration_ms":40821,"significance":"If the central claim is substantiated, the contribution would be relevant to 3GPP Rel-19/20 discussions on near-field channel modeling and codebook design, since current Type I/Type II codebooks are far-field-centric and near-field coexistence is an acknowledged standardization gap. The idea of replacing fixed codebooks with a dictionary learned in the wavenumber domain is interesting and potentially practical, and the paper correctly identifies a real limitation of existing schemes. However, the evidence as presented is not yet sufficient: the wavenumber-domain sparsity premise is asserted rather than validated for the simulated geometry, the K-SVD optimization is not specified in enough detail to be reproduced, and the headline claim about avoiding region identification is not tested against a region-aware baseline. The paper is clearly written and the figures convey the qualitative behavior, but the technical depth is below what a journal contribution on codebook design normally requires.","major_comments":[{"comment":"The load-bearing premise that TR 38.901 near-field channels are sparse in the wavenumber domain is asserted via references [3], [14], but no derivation, transform definition, or sparsity measurement is provided for the specific array and channel geometry used in the simulations. Over a finite 32x32 UPA aperture, a near-field spherical wavefront generally has a broad wavenumber footprint; sparsity is an empirical property that must be checked under the simulated near-field distances, carrier frequency, array spacing, and cluster parameters. Since the compressive-sensing channel estimation and the K-SVD dictionary update both rely on this sparsity, the paper should either provide a rigorous justification for the wavenumber-domain representation or include a quantitative sparsity analysis (e.g., number of significant coefficients vs. distance/aperture) for the TR 38.901-based model.","section":"III-A"},{"comment":"The K-SVD regression is described only in words and by an illustrative figure; there is no equation stating the optimization objective, the sparsity constraint, the dictionary update rule, or the stopping criterion. The paper also does not report the training set size, number of atoms, sparsity level, number of iterations, or the exact procedure for the constant-modulus projection. Without these details, the results in Figs. 4-5 are not reproducible, and the performance could depend critically on untested hyperparameters. In particular, the claim that the constant-modulus projection 'has an acceptable impact' needs quantitative support, since analog precoding cannot use arbitrary complex coefficients.","section":"III-B"},{"comment":"The abstract's central claim that the proposed codebook removes the need to identify the near/far-field region is not tested against a region-aware baseline. A proper comparison would include a hybrid scheme that knows each UE's region and uses, for example, a polar codebook for near-field UEs and a DFT codebook for far-field UEs, or an optimal per-region codebook selection. Without such a baseline, the simulation only shows that the learned codebook outperforms two fixed codebooks that are each mismatched to the opposite region; it does not demonstrate that region identification is unnecessary. This is a load-bearing gap for the paper's headline contribution.","section":"III-C, Figs. 4-5"},{"comment":"The simulation setup is underspecified. The authors state only that the BS has a 32x32 UPA, serves 16 UEs, uses 4 clusters with 5 rays per cluster, and runs 500 simulations. Missing are the carrier frequency, array element spacing, the definition of Rayleigh distance, the distribution of near-field distances, cluster angular spreads, delay spreads, and the exact way in which the TR 38.901 model is modified to include near-field spherical wavefronts. This matters because the wavenumber-domain representation and the K-SVD generalization behavior are sensitive to these parameters. Additionally, TR 38.901 as currently published does not include a mature near-field spherical-wave model; the cited study item RP-241743 is still ongoing, so the paper should clarify whether it uses the legacy far-field model with a distance-dependent steering-vector modification or a newly developed near-field model.","section":"III-C"},{"comment":"The validation is performed on the full pipeline that includes the wavenumber-domain channel estimator proposed by the same authors (refs [2], [12], [13]), and the codebook is trained and tested on channels drawn from the same simulation model. This creates a circularity risk: the benefit of the learned codebook cannot be separated from the benefit of the channel estimator, and the results may not generalize to unseen channel statistics. The authors should report tests with perfect CSI, and should also evaluate the learned codebook on a channel realization drawn from a different scenario (e.g., different near-field distances, different number of UEs, or a different cluster configuration) than the one used for training, to demonstrate that the dictionary is not overfitting the training distribution.","section":"III-C, Figs. 4-5"}],"minor_comments":[{"comment":"Figure 2 lacks the channel-model parameters and array configuration used to generate the NMSE curves; adding this information is essential for interpreting the claimed advantage of the wavenumber domain.","section":"Fig. 2"},{"comment":"The spectral efficiency curves are presented as single average lines without variance or confidence intervals; given that only 500 simulations are mentioned and the results are the main evidence, error bars or at least a statement of variance should be included.","section":"Figs. 4-5"},{"comment":"There is a typo in 'sparse channel matirx' that should read 'matrix'.","section":"III-B"},{"comment":"The captions and axis labels contain typos and inconsistencies: 'Spectral efficieny' appears in Figs. 4 and 5a, and the axis labels of Fig. 5b appear garbled ('olar FT P D') and should be corrected to 'Polar' and 'DFT'.","section":"Figs. 4-5 captions"},{"comment":"The phrase 'proceding process' should be 'precoding process', and 'neccessary' in Section IV-D should be 'necessary'.","section":"III-C"},{"comment":"The description of Type II codebook as performing 'beam sweeping' is informal and potentially misleading; Type II CSI reporting in NR is based on linear combination of a set of beams quantized from the channel eigenvectors, which is different from the beam-sweeping procedure used for beam management. The paper should align its terminology with the actual 3GPP specification.","section":"II-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like a standards-motivated overview than a full technical paper: the wavenumber-domain channel estimation is taken from the authors' prior work, and the new K-SVD codebook component is described without mathematical formulation or ablation studies. The novelty is potentially sufficient for a communications journal, but the evidence is currently too thin to support the strong 'unified paradigm' claim. I would encourage the editor to require the authors to add the missing specification and validation, in particular the region-aware baseline and the sparsity analysis, before a second round of review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat's actually new: a single K-SVD-learned codebook, applied in the wavenumber domain, that serves mixed near-field and far-field UEs without the BS ever classifying users into regions. That is a genuine extension of the wavenumber-domain estimation developed in the authors' prior work (and Pizzo's group), and it speaks to a real 3GPP problem for the 7-24 GHz near-field study item. The paper is clearly written, and the motivation is sound: Type I/II DFT codebooks suffer power leakage for near-field UEs, and polar-domain codebooks tend to require region labels or heavy sampling. The mixture of offline dictionary learning with a standards-compatible PMI feedback flow is a sensible framing.\n\nThe soft spots are large, in proportion. The central claim is supported only by simulation curves. There are no equations for the K-SVD objective or its update rules, no hyperparameters (number of atoms, sparsity level, training set size), no variance across the 500 runs, and no code. The 'no region ID needed' claim is not tested against a region-aware baseline, so the reader cannot separate the benefit of learning from the benefit of knowing the region. The simulation description is thin: array size and cluster counts are given, but not the near-field distances or how TR 38.901 was extended to spherical wavefronts.\n\nThe deepest issue is the wavenumber-domain sparsity premise. The paper cites prior work and asserts that near-field TR 38.901 channels are sparse superpositions of plane waves. That is not automatic: a point source has a continuous plane-wave spectrum; over a 32x32 aperture, the wavenumber footprint can be broad unless the channel is dominated by a few isolated scatterers. With 4 clusters of 5 rays, each ray is a spherical wave, so discrete sparsity is not guaranteed. Fig. 2 does not settle it because it omits the channel model and the near-field configuration. The stress-test note slightly overstates by saying the method has no theoretical basis if sparsity fails—K-SVD could still learn a useful low-rank codebook—but the compressive-sensing channel estimation genuinely relies on sparsity, so the concern lands.\n\nWho is this for: people tracking 6G codebook proposals and the 3GPP near-field study. As a proposal sketch it is a reasonable candidate; as a rigorous evaluation it is not there yet. A serious referee can force disclosure of training details, a fair region-aware comparison, and a sparsity validation. I would send it to review on those terms, but I would not cite it for the main result as it stands. Cite the underlying wavenumber-domain papers instead.","headline":"Proposes a K-SVD-learned wavenumber-domain codebook to unify near/far-field precoding; the idea is worth a serious look, but the paper is a sketch that needs training details, fair baselines, and sparsity validation before the main claim is credible.","tokens_in":10003,"tokens_out":4614,"would_cite":false,"duration_ms":45796,"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":"A K-SVD-learned codebook trained on wavenumber-domain channel estimates can serve near-field and far-field users in multi-user MIMO with one precoding scheme, eliminating the need for the base station to classify any user's region.","keywords":["near-field communications","far-field communications","codebook design","wavenumber domain","K-SVD dictionary learning","multi-user MIMO precoding","compressive sensing channel estimation","6G MIMO"],"falsifier":"Rerun the paper's 32-by-32 array, 16-UE simulations with all users placed well inside the Rayleigh distance, for example at 0.1 times the Rayleigh distance, and check whether wavenumber-domain OMP or MRF still yields lower NMSE than angular-domain OMP; if the wavenumber representation is not sparser there, the K-SVD codebook's advantage over the polar codebook should disappear.","tokens_in":8937,"feed_emoji":"📡","tokens_out":7265,"duration_ms":66393,"temperature":0.7,"pith_summary":"This paper claims that 5G's DFT-based Type I and Type II codebooks fail when users sit inside the Rayleigh distance, and that replacing them with a K-SVD-learned codebook trained on wavenumber-domain channel estimates unifies near-field and far-field precoding. The wavenumber-domain representation expands the channel as a sparse superposition of plane waves, so one dictionary can serve both regimes. In simulations with a 1024-antenna array serving 16 users under the standardized TR 38.901 channel model, the learned codebook outperforms DFT and polar codebooks and approaches fully digital spectral efficiency, while keeping the existing reference-signal workflow. If this holds, 6G systems could precode mixed user populations without spending pilots on region classification.","feed_headline":"Learned codebook beats DFT and polar in mixed near/far-field MIMO","feed_subtitle":"Wavenumber-domain training removes the need to classify users and nears fully digital spectral efficiency.","key_machinery":"The machinery is the combination of a wavenumber-domain channel representation and K-SVD dictionary learning. The wavenumber-domain representation uses a Fourier harmonic expansion of the channel as a superposition of plane waves with different wavenumbers, which the paper argues makes both near-field and far-field channels sparse and avoids the angular power diffusion of DFT processing. K-SVD then alternates between solving sparse recovery problems for the coefficient matrix $\\bar{H}$ and updating the codebook matrix $A$ column by column with SVD, so the codebook is tuned to the statistics of the actual channel; a final constant-modulus projection renders the codewords feasible for hybrid analog-digital precoding.","core_discovery":"The central claim is that a codebook learned by K-SVD in the wavenumber domain ensures efficient precoding for all UEs in coexisting near-field and far-field multi-user MIMO, and removes the need for the base station to know which region any UE is in. The argument runs through two stages: first, the channel $H$ is estimated in the wavenumber domain, where both far-field planar waves and near-field spherical waves are represented as superpositions of plane waves via Fourier harmonic expansion, and the estimation is cast as a compressive-sensing sparse recovery problem; second, K-SVD learns a codebook matrix $A$ and a sparse coefficient matrix $\\bar{H}$ by alternating sparse coding and column-wise SVD updates, and a constant-modulus projection makes the codebook usable by analog phase shifters. Reported simulations over the standardized TR 38.901 channel model, with a $32\\times32$ uniform planar array serving 16 UEs in mixed, all-near-field, and all-far-field deployments, show the regression-based codebook exceeding DFT and polar codebooks and approaching the fully digital precoding bound.","pith_inferences":["I infer that the same K-SVD training pipeline could be run offline on site-specific or measured channel datasets, so the codebook would adapt to a deployment's actual scatterer geometry rather than to a fixed grid; the paper notes offline learning but does not develop it.","A testable extension is to scale the antenna aperture while holding the user distribution fixed: if the wavenumber-domain mechanism is what carries the gain, the regression codebook's margin over DFT should grow as the Rayleigh distance pushes more users into the near field.","The constant-modulus projection is acknowledged to erode the sparsity of the recovered channel; I infer that a systematic sweep of projection loss against array size and codebook size would reveal when hybrid precoding with a learned codebook stops being worth the learning overhead."],"forward_implications":["A single codebook trained in the wavenumber domain can precode mixed near-field and far-field UEs without the base station classifying any UE's region, eliminating the extra pilots and feedback that region-aware schemes require.","Because the workflow and reference signals stay aligned with current Type I and Type II procedures, the learned codebook is a candidate extension to existing standards rather than a new air interface.","Wavenumber-domain channel estimation reduces the angular power diffusion that near-field channels cause, which should weaken the domino effect where one user's bad CSI degrades every other user's precoding.","In the reported simulations, the K-SVD codebook outperforms both DFT and polar-domain codebooks and approaches the fully digital precoding bound under the standardized TR 38.901 channel model."],"supporting_citations":[{"why":"gives the Fourier harmonic expansion that represents channels as superpositions of plane waves.","marker":"[14]"},{"why":"supplies the wavenumber-domain unified channel model that the proposed scheme relies on.","marker":"[3]"},{"why":"provides the wavenumber-domain channel estimation and power-diffusion analysis justifying the approach.","marker":"[12]"},{"why":"provides the MRF-based sparse recovery used for wavenumber-domain CSI.","marker":"[13]"},{"why":"supplies the OMP compressive sensing method and near-field estimation baseline.","marker":"[2]"},{"why":"defines the channel model used in all simulations.","marker":"[8]"},{"why":"proposes the polar-domain codebook used as a comparison baseline.","marker":"[1]"},{"why":"supplies the hybrid precoding architecture used in the simulations.","marker":"[11]"},{"why":"supplies the constant-modulus projection applied to the learned codebook.","marker":"[7]"}],"fun_headline_variants":["Wavenumber-domain codebook unifies near and far field precoding","Codebook removes near-field/far-field user classification in MIMO","K-SVD codebook bridges near-field and far-field MIMO gap","Unified codebook approaches fully digital spectral efficiency in mixed MIMO","Learned codebook covers both near and far field without classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that near-field channels from the standardized TR 38.901 model are faithfully captured as sparse superpositions of plane waves in the wavenumber domain; if that representation loses or smears near-field energy, the compressive-sensing estimates and the K-SVD codebook trained on them lose their edge.","fun_headline_variants_meta":{"raw":{"variants":["Wavenumber-domain codebook unifies near and far field precoding","Codebook removes near-field/far-field user classification in MIMO","K-SVD codebook bridges near-field and far-field MIMO gap","Unified codebook approaches fully digital spectral efficiency in mixed MIMO","Learned codebook covers both near and far field without classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2632,"prompt_tokens":936,"completion_tokens":1696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1614}},"tokens_in":552,"tokens_out":1696,"duration_ms":12172,"temperature":1.0,"reasoning_tokens":1614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:05:38.931636+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the paper's 32-by-32 array, 16-UE simulations with all users placed well inside the Rayleigh distance, for example at 0.1 times the Rayleigh distance, and check whether wavenumber-domain OMP or MRF still yields lower NMSE than angular-domain OMP; if the wavenumber representation is not sparser there, the K-SVD codebook's advantage over the polar codebook should disappear.","supporting_citations":[{"cited_title":"Fourier plane-wave series expansion for holographic mimo communications,","cited_arxiv_id":null,"evidence_quote":"gives the Fourier harmonic expansion that represents channels as superpositions of plane waves."},{"cited_title":"Unifying far-field and near- field wireless communications in 6G MIMO,","cited_arxiv_id":null,"evidence_quote":"supplies the wavenumber-domain unified channel model that the proposed scheme relies on."},{"cited_title":"Unified Far-Field and Near-Field in Holographic MIMO: A Wavenumber-Domain Perspective","cited_arxiv_id":"2407.14815","evidence_quote":"provides the wavenumber-domain channel estimation and power-diffusion analysis justifying the approach."},{"cited_title":"Wavenumber domain sparse channel estimation in holographic MIMO,","cited_arxiv_id":null,"evidence_quote":"provides the MRF-based sparse recovery used for wavenumber-domain CSI."},{"cited_title":"Near-field channel estimation in dual- band XL-MIMO with side information-assisted compressed sensing,","cited_arxiv_id":null,"evidence_quote":"supplies the OMP compressive sensing method and near-field estimation baseline."},{"cited_title":"Study on Channel Model for Frequencies from 0.5 to 100 GHz,","cited_arxiv_id":null,"evidence_quote":"defines the channel model used in all simulations."},{"cited_title":"Limited feedback hybrid precoding for multi-user millimeter wave systems,","cited_arxiv_id":null,"evidence_quote":"supplies the hybrid precoding architecture used in the simulations."},{"cited_title":"Joint hybrid precoding and rate allocation for RSMA in near-field and far-field massive MIMO communications,","cited_arxiv_id":null,"evidence_quote":"supplies the constant-modulus projection applied to the learned codebook."}],"review_version":1}