Pith. sign in

REVIEW 4 major objections 5 minor 37 references

A learned blockwise port selector can cut fluid-antenna sidelobes by 3.26 dB while keeping sum rate.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

A learned blockwise port-activation scheme for fluid antenna arrays lowers average peak sidelobes by 3.26 dB over uniform sparse activation at a slightly higher simulated sum rate, and by 8–10 dB over channel-driven selectors.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A sensible, internally consistent combination of standard learning and geometry tools for FAA port selection, but the headline PSLL gain is largely trained-for and all numbers rest on one synthetic channel model; should go to peer review with a request for external-validity tests and code/data. the 4 major comments →

arxiv 2607.25365 v1 pith:HD2LTUFB submitted 2026-07-28 cs.IT math.IT

Learned Blockwise Port Activation for Real Time Beamforming in Fluid Antenna Arrays

classification cs.IT math.IT
keywords fluid antenna arrayport activationsparse array beamformingsidelobe suppressionregularized zero forcingneural network port selectionblockwise activationmultiuser downlink
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper is trying to establish that the hard combinatorial problem of choosing which ports of a fluid antenna array to switch on can be solved in real time by a small neural scorer plus a deterministic geometry rule, without iterative search. It claims this learned blockwise port activation (L-BPA) outperforms uniform sparse activation by 3.26 dB in average peak sidelobe level while slightly increasing sum rate, and beats greedy and gain-based selection by 8.13 dB and 10.10 dB in PSLL. The practical stakes: FAA downlinks could get channel-adaptive, sidelobe-aware port selection at the channel time scale, compatible with grouped-switch hardware.

Core claim

The central claim is that the rate–PSLL tradeoff in FAA downlink beamforming can be handled by decoupling the role of learning from the role of feasible selection. A lightweight convolutional network scores every candidate port from multiuser channel features, port coordinates, and power statistics; a block-wise constraint fixes the number of active ports per aperture block; a multi-scale repulsion term prevents local clustering; and the final mask feeds a regularized zero-forcing precoder. Trained end-to-end with straight-through masks and a differentiable peak sidelobe surrogate, the system reaches 98.06 bit/s/Hz average sum rate (about 76% of the full-port reference) at −15.64 dB average

What carries the argument

Blockwise activation structure (each aperture block contributes a fixed number of ports, enabling grouped-switch hardware and preventing clustering); a small Conv–GN–SiLU scorer producing channel-aware port logits; block-wise straight-through masks that keep the discrete budget during training; a differentiable log-sum-exp PSLL surrogate that gives gradient signal for sidelobes; and a multi-scale geometric repulsion rule at inference that subtracts a proximity penalty from learned scores before selecting top ports per block. Together they convert an NP-hard subset-selection problem into one CNN forward pass plus O(sum(N_b^2 + m_b N_b)) deterministic selection.

Load-bearing premise

The central numbers come only from channels drawn by one synthetic far-field generator; if real channels (with mutual coupling, near-field effects, or other geometry) behave differently, the reported sidelobe and rate gains may not survive.

What would settle it

Evaluate L-BPA on measured or full-wave-simulated fluid-antenna channels that include mutual coupling and near-field effects; if the PSLL gap to uniform sparse activation shrinks below a few decibels or average sum rate drops below Uniform, the paper's central claim is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • With only 24% of ports active, L-BPA keeps 76% of full-aperture sum rate while improving PSLL by 3.26 dB over a uniform sparse grid, so sparse FAA operation need not sacrifice sidelobe quality.
  • Because inference requires no iterative RZF re-evaluation over candidate masks, port activation can be recomputed at the channel time scale, compatible with real-time downlink scheduling.
  • The blockwise constraint matches grouped-switch hardware, meaning the method's mask is feasible to implement without per-port switches.
  • The ablation shows that removing either the learned scorer or the geometric PSLL surrogate degrades the PSLL/rate balance, indicating each component contributes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the gains transfer beyond the paper's synthetic far-field UMi channels, the same blockwise architecture could serve as a generic low-latency port-selection front end for any reconfigurable-aperture array, since the geometry rule is channel-agnostic.
  • The fixed repulsion parameters and block size are tuned for a 5λ×2.5λ aperture; a natural extension would be to learn them per aperture or per channel statistics, which the paper does not explore.
  • The paper's future-work mention of wideband beam-squint suggests that extending the PSLL surrogate to frequency-dependent patterns is a direct next test; one could check whether the 3.26 dB gain persists across subcarriers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes L-BPA, a learned blockwise port activation method for fluid antenna array (FAA) downlink beamforming. A lightweight convolutional scorer maps multiuser channel features, port coordinates, and power statistics to port logits; blockwise hard masks are trained with a straight-through Gumbel-softmax relaxation and a differentiable peak sidelobe level (PSLL) surrogate. At inference, learned scores are combined with a multiscale geometric repulsion rule, and the resulting sparse port set is used for regularized zero-forcing (RZF) precoding. In simulations on a synthetic UMi far-field channel model with N=1250 candidate ports, K=16 users, and M=300 active ports, L-BPA achieves 98.06 bit/s/Hz average sum rate and -15.64 dB average PSLL, versus 97.48 bit/s/Hz and -12.38 dB for Uniform, 87.60 bit/s/Hz and -7.51 dB for a Greedy baseline, and lower PSLL for gain-based selection. The paper claims this shows real-time, sidelobe-aware, hardware-compatible port activation without iterative online search.

Significance. If the reported performance persists under realistic channel mismatch, L-BPA is practically attractive: it enforces grouped-switch constraints, avoids iterative RZF search, and explicitly trades sum rate against sidelobe behavior. The paper's strengths are its transparent problem formulation, explicit online-complexity analysis (Eq. 28), and ablations (Table II) that isolate the contributions of the learned scorer and the geometric loss. The principal weakness is that the entire evaluation rests on one synthetic channel law; moreover, the PSLL improvement is largely an explicit training objective, and the Greedy baseline is prefiltered. These issues do not invalidate the method but materially limit the strength of the claims as written.

major comments (4)
  1. [Sec. IV-A / Eq. (2) / Fig. 7] All training, validation, and test channels are drawn from a single synthetic far-field sum-of-rays generator (Eq. 2 with UMi parameters). The no-retraining sweeps in Fig. 7 vary only scalar parameters of that same generator. No measured channels, mutual-coupling model (see refs. [22]–[24]), near-field, or different-geometry test is reported, despite the abstract's 'real time ... FAA downlink beamforming' deployment claim. For a data-driven method this is a load-bearing external-validity gap. Please add at least one out-of-distribution or measured-channel experiment, or substantially restrict the claims to the synthetic scenario.
  2. [Sec. IV-A / Table I] The Greedy baseline is restricted to a candidate pool 'prefiltered by aggregate channel power' before sequential RZF-based selection. The headline comparisons against Greedy (8.13 dB PSLL improvement, 10.46 bit/s/Hz rate improvement) are therefore against a handicapped baseline unless the same prefilter is applied to L-BPA or the pool is large enough to be inconsequential. Report the prefilter size and provide a comparison to an unfiltered greedy (possibly with a complexity cap) so that the comparison actually tests the proposed method against the advertised greedy algorithm.
  3. [Eq. (24) / Table I] The training loss explicitly includes the PSLL surrogate L_psll with target Γ_psll=-15 dB, and the reported average PSLL is -15.64 dB. Thus the '3.26 dB improvement over Uniform' is substantially a trained-for consequence of the objective rather than an emergent property. This is not a flaw per se, but the abstract's phrasing 'L-BPA reduces the average PSLL by 3.26 dB' overstates the finding. Please include a sweep of Γ_psll (or λ_psll) to show the achievable rate–PSLL tradeoff and clarify that the reported operating point is selected by the designer, not an unconstrained result.
  4. [Sec. II-C / Figs. 3–4] The PSLL metric is the equal-amplitude broadside array factor (Eq. 10), not the actual RZF-weighted multiuser beam pattern. Since the manuscript's title and abstract concern beamforming, the sidelobe claims should be validated on at least a few representative RZF-weighted beam patterns, or the proxy nature of the metric should be made prominent in the abstract and conclusions. As written, the 3.26 dB improvement may not translate directly to the radiated multiuser beams.
minor comments (5)
  1. [Sec. I] Typos: 'termedgeometry' and 'areal timeandhardware' are missing spaces.
  2. [Algorithm 1, step 11] The normalization of z to \bar z is not defined. Specify whether it is z-scoring, min-max, or another normalization.
  3. [Eq. (20)] The Gumbel-softmax relaxation uses clip(m_b softmax(...),0,1); after clipping the sum may not equal m_b. Clarify whether the exact cardinality constraint is enforced in the soft mask or only in the hard mask.
  4. [Table I] Report confidence intervals or paired significance tests for the 0.58 bit/s/Hz rate difference between L-BPA and Uniform; with 500 samples this difference could be within noise.
  5. [References] References [8] and [17] appear to be the same paper; please merge or differentiate.

Circularity Check

0 steps flagged

No significant circularity: PSLL is an explicit, disclosed training objective and the reported numbers are held-out evaluations; the rate result is emergent.

full rationale

The paper's chain is a conventional learning pipeline: Eq. (2) defines a far-field channel model; Eqs. (4)-(7) define RZF sum-rate; Eqs. (10)-(12) define PSLL; Eqs. (22)-(24) define a training loss that explicitly optimizes R(a) and a PSLL surrogate with target Γ_psll = -15 dB. Table I's average PSLL of -15.64 dB is therefore close to the target because the method was trained to minimize that objective, not because the metric is secretly an input. The paper discloses this: Sec. II-C calls PSLL a 'geometry proxy', Sec. III-C says the surrogate is used only for training, and Sec. IV-A gives Γ_psll and states that reported PSLL is recomputed on hard binary masks using a denser 121x121 grid. Reporting an optimized objective on a held-out 500-channel test set is a standard evaluation, not a circular derivation. The sum-rate improvement (98.06 vs 97.48 bit/s/Hz) is not forced by the PSLL target and is a genuine emergent result, although it is only demonstrated in-distribution. Self-citations to prior FAA work [18], [19], [22] are background and are not load-bearing; no uniqueness theorem or ansatz is imported from same-author work. The lack of measured or out-of-distribution channels is an external-validity limitation, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The claim introduces no invented physical entity and rests on standard precoding math. Its hidden cost is concentrated in hand-set geometry/training hyperparameters (σ1, σ2, ρ_g, μ_g, Γ_psll, λ_psll, β), an unspecified block partition, and three domain assumptions: the synthetic steering-channel model, the equal-amplitude PSLL proxy, and in-distribution train/test evaluation.

free parameters (7)
  • Inference repulsion scales σ1, σ2 (in λ) = σ1=0.20, σ2=0.35
    Hand-set in Sec. IV-A; directly determine how strongly the selector spreads ports, i.e., the PSLL/rate tradeoff being measured; no validation procedure is stated.
  • Repulsion weights ρ_g, μ_g = ρg=0.5, μg=0.15
    Hand-set secondary-scale weight and score-repulsion balance (Eqs. 26–27); tune the headlined tradeoff.
  • PSLL training target Γ_psll = −15 dB
    Target in the softplus surrogate loss (Eq. 23); the reported average PSLL of −15.64 dB closely tracks this target, so part of the 'PSLL improvement' headline is the chosen training anchor.
  • Surrogate weight λ_psll and sharpness β = λ_psll=1.2, β=80
    Set the strength and smooth-max sharpness of the PSLL gradient signal (Sec. IV-A).
  • Training hyperparameters (τ, lr, width, epochs) = 0.6, 1e-4, 48, 60
    Gumbel temperature, learning rate, hidden width, and epoch count (Sec. IV-A); standard choices affecting convergence but not the conceptual claim.
  • Block partition B_x × B_y and per-block budget m_b = not specified
    The core hardware constraint (Eq. 15) gives M = B_x·B_y·m_b = 300, but B_x, B_y, m_b are never given numerically. Many geometries realize M=300 (e.g., 5×5×12 vs 10×5×6) and they define different feasible sets, hence different results.
  • RZF regularization η_rzf and SNR perturbation Δ_SNR = η_rzf=0.1; Δ_SNR not given
    η_rzf is fixed for all compared methods to isolate port activation (shared choice). The training SNR is perturbed as U(ρ0−Δ, ρ0+Δ) but the range Δ is never specified.
axioms (6)
  • domain assumption The far-field steering sum-of-rays channel model (Eq. 2) with direction-cosine quantization is an adequate ground-truth channel for real FAA downlinks.
    Used to generate all CSI; contains no mutual coupling, near-field terms, or hardware errors, despite refs [22]–[24] citing coupling effects as an FAA capability.
  • standard math RZF SINR expression (Eqs. 4–7) with perfect CSI and total-power normalization is the correct rate objective.
    Standard linear-precoding analysis (ref [34]); defines what 'sum rate' means in the claim.
  • domain assumption Equal-amplitude broadside PSLL with the guard region (Eqs. 10–12) is the relevant sidelobe metric for the sparse aperture.
    Explicitly a 'geometry proxy' in Sec. II-C; actual RZF-weighted patterns are never evaluated, so headlined PSLL gains apply to this proxy only.
  • domain assumption Training and test channels are i.i.d. samples from the same generator, so a scorer trained on 1200 realizations transfers to the 500 test realizations.
    The entire evaluation (Table I, Fig. 6) is in-distribution; generalization sweeps in Fig. 7 vary only scalar parameters of the same generator.
  • ad hoc to paper The multi-scale repulsion rule (Eqs. 25–27) with hand-set widths/weights improves the true rate–PSLL tradeoff.
    This inference heuristic is introduced ad hoc; no derivation or sensitivity analysis shows why these σ1, σ2, ρ_g, μ_g values are appropriate beyond the chosen test distribution.
  • domain assumption UMi channel parameters (5 clusters, 12 rays, Rician K ~ N(7,4) dB) as configured in Sec. IV-A represent the deployment scenario.
    All quantitative claims are conditional on these generator settings; no alternate environment (indoor, dense urban, measured) is tested.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Learned Blockwise Port Activation for Real Time Beamforming in Fluid Antenna Arrays." pith.science (2026). https://pith.science/paper/HD2LTUFB

@misc{pith2026260725365,
  author       = {Pith},
  title        = {Pith review of: Learned Blockwise Port Activation for Real Time Beamforming in Fluid Antenna Arrays},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HD2LTUFB}},
  note         = {Machine review of arXiv:2607.25365}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Fluid antenna arrays (FAAs), support multiuser downlink transmission by activating a subset of reconfigurable ports. The activation mask jointly determines the effective channel and the sparse radiating aperture, which requires a balance among sum rate, sidelobe suppression, hardware constraints, and online complexity. Channel driven selection can cluster active ports and increase sidelobes, whereas sidelobe oriented synthesis is typically channel independent and can sacrifice sum rate. This paper proposes learned blockwise port activation (L-BPA), for real time sidelobe aware FAA downlink beamforming. L-BPA activates a fixed number of ports in each aperture block, which supports grouped switching hardware and limits port clustering. A lightweight convolutional network scores ports using multiuser channel features, port coordinates, and user power statistics. Training combines blockwise straight through masks with a differentiable peak sidelobe level (PSLL), surrogate. During inference, learned scores are combined with multiscale geometric repulsion, followed by regularized zero forcing precoding over the reduced effective channel. L-BPA reduces the average PSLL by 3.26 dB relative to uniform sparse activation while achieving a slightly higher sum rate. It also reduces the PSLL by 8.13 dB and 10.10 dB relative to greedy and gain based selection, respectively, without iterative online search.

Figures

Figures reproduced from arXiv: 2607.25365 by Hanjiang Hong, Hao Jiang, Kai-Kit Wong, Wenjun Zhang, Yin Xu, Yuanhui Wu, Zaichen Zhang, Zhentian Zhang.

Figure 1
Figure 1. Figure 1: Overall workflow of the proposed L-BPA method. The left panel corresponds to the multi-user FAA downlink model in Section [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: RZF sum-rate box plots over 500 test channels at 10 dB. Boxes show the interquartile range, center lines show medians, whiskers show the non-outlier [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Normalized broadside beam-power cuts for a representative test sample: (a) cut along the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Two-dimensional normalized broadside beam-power maps for the same sample as Fig. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Active-port layouts for the representative sample used in Figs. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Active-port tradeoff between average RZF sum-rate and equal-amplitude PSLL over 200 test samples. Each marker corresponds to one active-port [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: No-retraining generalization with M = 300 active ports. (a) Average RZF sum-rate versus desired SNR. (b) Average RZF sum-rate versus azimuth angular spread. (c) Average RZF sum-rate versus user angular-sector width. (d) Average PSLL versus desired SNR. (e) Average PSLL versus azimuth angular spread. (f) Average PSLL versus user angular-sector width, where a smaller sector means more correlated user directi… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 7 linked inside Pith

  1. [1]

    Performance limits of fluid antenna systems,

    K. K. Wong,et al., “Performance limits of fluid antenna systems,”IEEE Commun. Lett., vol. 24, no. 11, pp. 2469–2472, Nov. 2020

  2. [2]

    Fluid antenna systems,

    K. K. Wong,et al., “Fluid antenna systems,”IEEE Trans. Wireless Commun., vol. 20, no. 3, pp. 1950–1962, Mar. 2021

  3. [3]

    Cram´er–Rao bounds for activity detection in conventional and fluid antenna systems,

    Z. Zhang, K. K. Wong, H. Jiang, C. Masouros, and C.-B. Chae, “Cram´er–Rao bounds for activity detection in conventional and fluid antenna systems,”IEEE Wireless Commun. Lett., vol. 15, pp. 3059– 3063, 2026

  4. [4]

    On fundamental limits of slow-fluid antenna multiple access for unsourced random access,

    Z. Zhang,et al., “On fundamental limits of slow-fluid antenna multiple access for unsourced random access,”IEEE Wireless Commun. Lett., vol. 14, no. 11, pp. 3455–3459, Nov. 2025

  5. [5]

    On fundamental limits for fluid antenna-assisted integrated sensing and communications for unsourced random access,

    Z. Zhang, K. K. Wong, J. Dang, Z. Zhang, and C.-B. Chae, “On fundamental limits for fluid antenna-assisted integrated sensing and communications for unsourced random access,”IEEE J. Sel. Areas Commun., vol. 44, pp. 136–149, 2026

  6. [6]

    Fluid antenna systems enabling 6G: Principles, applica- tions, and research directions,

    T. Wu,et al., “Fluid antenna systems enabling 6G: Principles, applica- tions, and research directions,”IEEE Wireless Commun., early access, 2025, doi: 10.1109/MWC.2025.3629597

  7. [7]

    Fluid antenna system-assisted self-interference cancel- lation for in-band full duplex communications,

    H. Honget al., “Fluid antenna system-assisted self-interference cancel- lation for in-band full duplex communications,”IEEE Trans. Wireless Commun., vol. 25, pp. 7476-7489, 2026

  8. [9]

    FAS meets OFDM: Enabling wideband 5G NR,

    H. Honget al., “FAS meets OFDM: Enabling wideband 5G NR,”IEEE Trans. Commun., vol. 73, no. 11, pp. 12884–12898, Nov. 2025

  9. [10]

    Downlink OFDM-FAMA in 5G-NR systems,

    H. Honget al., “Downlink OFDM-FAMA in 5G-NR systems,”IEEE Trans. Wireless Commun., vol. 24, no. 12, pp. 10116–10132, Dec. 2025

  10. [11]

    Coded fluid antenna multiple access over fast fading channels,

    H. Hong, K. -K. Wong, K. -F. Tong, H. Shin and Y . Zhang, “Coded fluid antenna multiple access over fast fading channels,”IEEE Wireless Commun. Lett., vol. 14, no. 4, pp. 1249–1253, April 2025

  11. [12]

    Reimagining wireless connectivity: The FAS-RIS synergy for 6G smart cities,

    T. Wu,et al., “Reimagining wireless connectivity: The FAS-RIS synergy for 6G smart cities,”IEEE Commun. Mag., early access, 2025, doi: 10.1109/MCOM.001.2500802

  12. [13]

    Fluid antennas meet intelligent surfaces: Security analysis of NOMA systems under hardware impairments,

    T. Wu,et al., “Fluid antennas meet intelligent surfaces: Security analysis of NOMA systems under hardware impairments,”IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 7775–7788, 2026

  13. [14]

    Variable block-correlation modeling and optimization for secrecy analysis in fluid antenna systems,

    T. Wu,et al., “Variable block-correlation modeling and optimization for secrecy analysis in fluid antenna systems,”IEEE Trans. Wireless Commun., vol. 25, pp. 15069–15085, 2026

  14. [15]

    Fluid antenna multiple access for 6G: A holistic review,

    H. Honget al., “Fluid antenna multiple access for 6G: A holistic review,” IEEE Open J. Commun. Soc., vol. 7, pp. 2607–2633, 2026

  15. [16]

    A contemporary survey on fluid antenna systems: fundamentals and networking perspectives,

    H. Honget al., “A contemporary survey on fluid antenna systems: fundamentals and networking perspectives,”IEEE Trans. Netw. Sci. Eng., vol. 13, pp. 2305–2328, 2026

  16. [17]

    Scalable fluid antenna systems: A new paradigm for array signal processing,

    T. Wu,et al., “Scalable fluid antenna systems: A new paradigm for array signal processing,”IEEE J. Sel. Topics Signal Process., vol. 20, no. 3, pp. 389–406, Apr. 2026

  17. [18]

    Finite-aperture planar fluid antenna array,

    Z. Zhang,et al., “Finite-aperture planar fluid antenna array,”arXiv preprint, arXiv:2605.22040, 2026

  18. [19]

    Finite-aperture fluid antenna array design: Analysis and algorithm,

    Z. Zhang,et al., “Finite-aperture fluid antenna array design: Analysis and algorithm,”IEEE Wireless Commun. Lett., vol. 15, pp. 3199–3203, 2026

  19. [20]

    Fluid antenna-enhanced flexible beamforming,

    J. Xu,et al., “Fluid antenna-enhanced flexible beamforming,” inProc. IEEE Wireless Commun. Netw. Conf. (WCNC), Kuala Lumpur, Malaysia, Apr. 2026

  20. [21]

    Peak sidelobe suppression in planar fluid antenna array,

    H. Liang,et al., “Peak sidelobe suppression in planar fluid antenna array,” arXiv:2606.31149, 2026

  21. [22]

    Electromagnetic-aware fluid antenna array ,

    Z. Zhang,et al., “Electromagnetic-aware fluid antenna array ,”preprint, arXiv:2607.21375, 2026

  22. [23]

    Metasurface-based fluid antennas: From electromagnetics to communications model,

    P. Ram ´ırez-Espinosa,et al., “Metasurface-based fluid antennas: From electromagnetics to communications model,”arXiv preprint, arXiv:2507.17982, 2025

  23. [24]

    Implementing fluid antennas in the beamspace: Performance evaluation and codebook design,

    P. Ram ´ırez-Espinosa,et al., “Implementing fluid antennas in the beamspace: Performance evaluation and codebook design,”arXiv preprint, arXiv:2605.13286, 2026

  24. [25]

    Fluid antenna multiple access,

    K.-K. Wong and K.-F. Tong, “Fluid antenna multiple access,”IEEE Transactions on Wireless Communications, vol. 21, no. 7, pp. 4801– 4815, Jul. 2022

  25. [26]

    Fluid antenna system–Part I: Preliminaries,

    K.-K. Wong, W. K. New, X. Hao, K.-F. Tong, and C.-B. Chae, “Fluid antenna system–Part I: Preliminaries,”IEEE Communications Letters, vol. 27, no. 8, pp. 1919–1923, Aug. 2023

  26. [27]

    Port selection for fluid antenna systems,

    Z. Chai, K.-K. Wong, K.-F. Tong, Y . Chen, and Y . Zhang, “Port selection for fluid antenna systems,”IEEE Communications Letters, vol. 26, no. 5, pp. 1180–1184, May 2022

  27. [28]

    Antenna selection in MIMO systems,

    S. Sanayei and A. Nosratinia, “Antenna selection in MIMO systems,” IEEE Communications Magazine, vol. 42, no. 10, pp. 68–73, Oct. 2004

  28. [29]

    Joint port selection and beamforming design for fluid antenna assisted integrated data and energy transfer,

    L. Zhang, H. Yang, Y . Zhao, and J. Hu, “Joint port selection and beamforming design for fluid antenna assisted integrated data and energy transfer,”IEEE Wireless Communications Letters, vol. 13, no. 7, pp. 1833–1837, Jul. 2024

  29. [30]

    Near-field beamforming and port selection for fluid antenna systems,

    S. Chen, Z. Zhang, H. Jiang, K.-K. Wong, A. Li, and F. R. Ghadi, “Near-field beamforming and port selection for fluid antenna systems,” IEEE Communications Letters, early access, 2026

  30. [31]

    Finite-blocklength fluid an- tenna systems with spatial block-correlation channel model,

    Z. Zhang, K.-K. Wong, D. Morales-Jimenez, H. Jiang, P. Ramirez- Espinosa, C.-B. Chae, and C. Masouros, “Finite-blocklength fluid an- tenna systems with spatial block-correlation channel model,”IEEE Wireless Communications Letters, vol. 15, pp. 1911–1915, 2026

  31. [32]

    Joint activity detection and channel estimation for fluid antenna system exploiting geographical and angular informa- tion,

    Z. Zhang, J. Dang, D. Morales-Jimenez, H. Jiang, Z. Zhang, C. Ma- souros, and C.-B. Chae, “Joint activity detection and channel estimation for fluid antenna system exploiting geographical and angular informa- tion,” arXiv:2512.15342, 2025

  32. [33]

    Geometry-structured channel recon- struction for fluid antenna systems: Theoretical bounds and Bayesian inference,

    Z. Zhang, K.-K. Wong, K. Meng, D. Morales-Jimenez, H. Jiang, C. Masouros, H. Shin, and Z. Zhang, “Geometry-structured channel recon- struction for fluid antenna systems: Theoretical bounds and Bayesian inference,” arXiv:2606.04001, 2026

  33. [34]

    Large system analysis of linear precoding in correlated MISO broadcast channels under limited feedback,

    S. Wagner, R. Couillet, M. Debbah, and D. T. M. Slock, “Large system analysis of linear precoding in correlated MISO broadcast channels under limited feedback,”IEEE Transactions on Information Theory, vol. 58, no. 7, pp. 4509–4537, Jul. 2012

  34. [35]

    Estimating or propagat- ing gradients through stochastic neurons for conditional computation,

    Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propagat- ing gradients through stochastic neurons for conditional computation,” arXiv:1308.3432, 2013

  35. [36]

    Categorical reparameterization with Gumbel-Softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with Gumbel-Softmax,” inProc. International Conference on Learning Rep- resentations (ICLR), 2017

  36. [37]

    Group normalization,

    Y . Wu and K. He, “Group normalization,” inProc. European Conference on Computer Vision (ECCV), 2018, pp. 3–19

  37. [38]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inProc. International Conference on Learning Representations (ICLR), 2019

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.