Sparsity-Aware Near-Field Beam Training via Multi-Beam Combination

Chien-Hua Chen; Jinesh Nair; Rama Kiran; Rui Zhang; Shawn Tsai; Tzu-Han Chou; Zijun Wang

REVIEW 4 major objections 5 minor 14 references

Sparsity-Aware Near-Field Beam Training via Multi-Beam Combination

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read One feedback beam reaches what 21 DFT beams achieve in near-field multipath channels.

desk verdict 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. read the letter →

arxiv 2505.08267 v1 pith:6OW42EF5 submitted 2025-05-13 eess.SP

Zijun Wang , Rama Kiran , Jinesh Nair , Chien-Hua Chen , Tzu-Han Chou , Shawn Tsai , Rui Zhang This is my paper · ORCID

classification eess.SP

keywords near-fieldcommunicationsbeamtrainingmulti-beamcombiningLASSOpolar-domaincodebookfeedbackoverheadoff-gridrefinementterahertz

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear colleague,

Quick 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.

What 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.

Where 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.

Also 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.

Verdict: 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.

I wouldn't cite it myself unless I needed a baseline for near-field beam training, but I'd happily send it to review.

Referee Report

4 major / 5 minor

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.

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 (4)

[Section III, Eq. (11) and Fig. 2] 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 IV, Figs. 3 and 4] 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 V, Fig. 9] 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.
[Eq. (15), Figs. 6 and 7] 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.

minor comments (5)

[Fig. 2 caption] 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 IV, numerical setup] 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.
[Eq. (17)] 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 I, paragraph 3] There is a typo, 'codeook', in the description of reference [7].
[Section IV, DFT scheme description] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the channel-reconstruction claims are validated against independent simulated ground truth and the near-field codebook is imported from external work [7].

full rationale

The paper's central claim is that selecting the K strongest received beams and reconstructing the near-field channel via linear combination or LASSO achieves lower feedback overhead than a DFT codebook. This is an estimation problem evaluated against simulated ground truth channels generated from Eqs. (1)-(4); the reconstruction error metrics in Figs. 3-9 compare the estimated channel to the true channel, so the results are not forced by definition. The K-largest-power beam selection in Section III is a heuristic sparsity assumption, not a circular step: it is an input to the estimator, and the paper does not define sparsity in terms of the final rate or overhead metric. The near-field codebook is adopted from reference [7], which is an external source rather than a self-citation, and no uniqueness theorem or prior same-author result is invoked to justify the approach. Equations (13), (15), and (17) are standard least-squares, LASSO, and off-grid refinement formulations, respectively; none of them presupposes the numerical claims they are used to support. The off-grid refinement improvement reported in Fig. 9 is an empirical simulation result measured against ground truth, not an analytical identity. Although the support-selection assumption could fail for highly coherent or off-grid channels, that is a correctness or generalization concern, not circular reasoning. No specific step in the derivation reduces by construction to its own inputs, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method's headroom comes from a favorable matched model (channel and codebook share the same parametric steering-vector family) and from error-free complex feedback. The unquantified design choices are beta, lambda, and the off-grid optimizer settings, all of which affect the claimed gains.

free parameters (3)

beta (near-field codebook distance sampling density) = 1.6; 2.384 for the compact codebook
Controls the number of distance samples per angle in the codebook from [7]. Chosen by the authors per experiment, and it affects coefficient sparsity and the feedback overhead needed.
lambda (LASSO regularization parameter) = not disclosed
Eq. (15) introduces lambda as a trade-off between fidelity and sparsity. The paper gives no selection rule or value, and the reported reconstruction quality depends on it.
Off-grid gradient descent hyperparameters (learning rate, iterations, initialization) = not disclosed
Eq. (17) is solved by gradient descent without implementation details. The 69.4% accuracy improvement claim depends on these choices.

assumptions (3)

domain assumption The true channel is a linear combination of steering vectors from the same parametric family used to construct the codebooks.
The channel model in Eqs. (2)-(4) uses near-field steering vectors b(theta, r); the codebook samples the same family, so the dictionary is matched to the generating process.
domain assumption Users feed back the selected received signals (amplitude and phase) without quantization or feedback error.
Section III assumes y_K is available exactly. Practical FDD feedback links quantize the values, which is not modeled.
domain assumption The polar-domain near-field codebook from [7] is taken as given, including its sampling rule.
The paper relies on the codebook's properties (low correlation between neighboring codewords) rather than reproducing or analyzing the construction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sparsity-Aware Near-Field Beam Training via Multi-Beam Combination." pith.science (2026). https://pith.science/paper/6OW42EF5

@misc{pith2026250508267,
  author       = {Pith},
  title        = {Pith review of: Sparsity-Aware Near-Field Beam Training via Multi-Beam Combination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6OW42EF5}},
  note         = {Machine review of arXiv:2505.08267}
}

read the original abstract

This paper proposes an adaptive near-field beam training method to enhance performance in multi-user and multipath environments. The approach identifies multiple strongest beams through beam sweeping and linearly combines their received signals - capturing both amplitude and phase - for improved channel estimation. Two codebooks are considered: the conventional DFT codebook and a near-field codebook that samples both angular and distance domains. As the near-field basis functions are generally non-orthogonal and often over-complete, we exploit sparsity in the solution using LASSO-based linear regression, which can also suppress noise. Simulation results show that the near-field codebook reduces feedback overhead by up to 95% compared to the DFT codebook. The proposed LASSO regression method also maintains robustness under varying noise levels, particularly in low SNR regions. Furthermore, an off-grid refinement scheme is introduced to enhance accuracy especially when the codebook sampling is coarse, improving reconstruction accuracy by 69.4%.

Figures

Figures reproduced from arXiv: 2505.08267 by the authors.

**Figure 2.** Coefficient power versus codebook beam index. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Influence of feedback overhead in a low SNR scenario. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Influence of feedback overhead at high SNR. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Influence of multipath number. (a) The L2 error versus SNR. (b) The achievable rate versus SNR [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Influence of SNR with 15 feedback overhead number. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 9.** Figure 9: Performance versus feedback overhead under grid [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 12 canonical work pages

[1]

Terahertz communications and sensing for 6g and beyond: How far are we?

J. M. Jornet, N. Yang, R. Nichols, S. Nie, C. Huang, and R. K. Mallik, “Terahertz communications and sensing for 6g and beyond: How far are we?” IEEE Wireless Communications, vol. 31, no. 1, pp. 8–9, 2024

work page 2024
[2]

Terahertz band communication: An old problem revisited and research directions for the next decade,

I. F. Akyildiz, C. Han, Z. Hu, S. Nie, and J. M. Jornet, “Terahertz band communication: An old problem revisited and research directions for the next decade,” IEEE Transactions on Communications , vol. 70, no. 6, pp. 4250–4285, 2022

2022
[3]

Near-field communications: A comprehensive survey,

Y . Liu, C. Ouyang, Z. Wang, J. Xu, X. Mu, and A. L. Swindlehurst, “Near-field communications: A comprehensive survey,” IEEE Commu- nications Surveys & Tutorials , pp. 1–42, 2024

work page 2024
[4]

Near- field communications: Research advances, potential, and challenges,

J. An, C. Yuen, L. Dai, M. Di Renzo, M. Debbah, and L. Hanzo, “Near- field communications: Research advances, potential, and challenges,” IEEE Wireless Communications, vol. 31, no. 3, pp. 100–107, 2024

work page 2024
[5]

Next generation terahertz communications: A rendezvous of sensing, imaging, and localization,

H. Sarieddeen, N. Saeed, T. Y . Al-Naffouri, and M.-S. Alouini, “Next generation terahertz communications: A rendezvous of sensing, imaging, and localization,” IEEE Communications Magazine , vol. 58, no. 5, pp. 69–75, 2020

work page 2020
[6]

Fraunhofer and fresnel distances: Unified derivation for aperture antennas,

K. T. Selvan and R. Janaswamy, “Fraunhofer and fresnel distances: Unified derivation for aperture antennas,” IEEE Antennas and Propa- gation Magazine, vol. 59, no. 4, pp. 12–15, 2017

work page 2017
[7]

Channel estimation for extremely large-scale mimo: Far-field or near-field?

M. Cui and L. Dai, “Channel estimation for extremely large-scale mimo: Far-field or near-field?” IEEE Transactions on Communications, vol. 70, no. 4, pp. 2663–2677, 2022

work page 2022
[8]

Two-stage hierarchical beam training for near-field communications,

C. Wu, C. You, Y . Liu, L. Chen, and S. Shi, “Two-stage hierarchical beam training for near-field communications,” IEEE Transactions on Vehicular Technology, vol. 73, no. 2, pp. 2032–2044, 2024

work page 2024

Show all 14 references

[9]

Fast near-field beam training for extremely large-scale array,

Y . Zhang, X. Wu, and C. You, “Fast near-field beam training for extremely large-scale array,” IEEE Wireless Communications Letters , vol. 11, no. 12, pp. 2625–2629, 2022

2022
[10]

Efficient hybrid near- and far-field beam training for xl-mimo communications,

J. Luo, J. Fan, K. Xie, and X. Shi, “Efficient hybrid near- and far-field beam training for xl-mimo communications,” IEEE Transactions on Vehicular Technology, pp. 1–6, 2024

2024
[11]

Near-field beam training: Joint angle and range estimation with dft codebook,

X. Wu, C. You, J. Li, and Y . Zhang, “Near-field beam training: Joint angle and range estimation with dft codebook,” IEEE Transactions on Wireless Communications, vol. 23, no. 9, pp. 11 890–11 903, 2024

2024
[12]

Csi type-ii codebook of code- books,

R. M. Dreifuerst and R. W. Heath, “Csi type-ii codebook of code- books,” in 2023 IEEE 24th International Workshop on Signal Process- ing Advances in Wireless Communications (SPAWC) , 2023, pp. 116– 120

2023
[13]

Qin and H

Z. Qin and H. Yin, A review of codebooks for csi feedback in 5g new radio and beyond , 2023. arXiv: 2302.09222 [cs.IT]. [Online]. Available: https://arxiv.org/abs/2302.09222

2023 arXiv
[14]

Plane wave expansion-based codebook design for 6g near-field mimo,

F. Wang, X. Hou, X. Li, and L. Chen, “Plane wave expansion-based codebook design for 6g near-field mimo,” in 2024 IEEE 29th Asia Pacific Conference on Communications (APCC) , 2024, pp. 536–541

2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Terahertz communications and sensing for 6g and beyond: How far are we?

J. M. Jornet, N. Yang, R. Nichols, S. Nie, C. Huang, and R. K. Mallik, “Terahertz communications and sensing for 6g and beyond: How far are we?” IEEE Wireless Communications, vol. 31, no. 1, pp. 8–9, 2024

work page 2024

[2] [2]

Terahertz band communication: An old problem revisited and research directions for the next decade,

I. F. Akyildiz, C. Han, Z. Hu, S. Nie, and J. M. Jornet, “Terahertz band communication: An old problem revisited and research directions for the next decade,” IEEE Transactions on Communications , vol. 70, no. 6, pp. 4250–4285, 2022

2022

[3] [3]

Near-field communications: A comprehensive survey,

Y . Liu, C. Ouyang, Z. Wang, J. Xu, X. Mu, and A. L. Swindlehurst, “Near-field communications: A comprehensive survey,” IEEE Commu- nications Surveys & Tutorials , pp. 1–42, 2024

work page 2024

[4] [4]

Near- field communications: Research advances, potential, and challenges,

J. An, C. Yuen, L. Dai, M. Di Renzo, M. Debbah, and L. Hanzo, “Near- field communications: Research advances, potential, and challenges,” IEEE Wireless Communications, vol. 31, no. 3, pp. 100–107, 2024

work page 2024

[5] [5]

Next generation terahertz communications: A rendezvous of sensing, imaging, and localization,

H. Sarieddeen, N. Saeed, T. Y . Al-Naffouri, and M.-S. Alouini, “Next generation terahertz communications: A rendezvous of sensing, imaging, and localization,” IEEE Communications Magazine , vol. 58, no. 5, pp. 69–75, 2020

work page 2020

[6] [6]

Fraunhofer and fresnel distances: Unified derivation for aperture antennas,

K. T. Selvan and R. Janaswamy, “Fraunhofer and fresnel distances: Unified derivation for aperture antennas,” IEEE Antennas and Propa- gation Magazine, vol. 59, no. 4, pp. 12–15, 2017

work page 2017

[7] [7]

Channel estimation for extremely large-scale mimo: Far-field or near-field?

M. Cui and L. Dai, “Channel estimation for extremely large-scale mimo: Far-field or near-field?” IEEE Transactions on Communications, vol. 70, no. 4, pp. 2663–2677, 2022

work page 2022

[8] [8]

Two-stage hierarchical beam training for near-field communications,

C. Wu, C. You, Y . Liu, L. Chen, and S. Shi, “Two-stage hierarchical beam training for near-field communications,” IEEE Transactions on Vehicular Technology, vol. 73, no. 2, pp. 2032–2044, 2024

work page 2024