Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Graph Neural Network Enabled Fluid Antenna Systems: A Two-Stage Approach

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A two-stage graph neural network, trained unsupervised on the system utility, jointly chooses fluid-antenna positions and beamforming to beat an equidistant-antenna convex baseline in sum rate and energy efficiency while inferring in…

desk verdict Plausible two-stage GNN for FAS, but the printed feasibility equations invert the power scaling and allow negative spacings, so the central guarantee and all numerical tables are built on an internal inconsistency. read the letter →

arxiv 2502.03922 v1 pith:66RDLZV6 submitted 2025-02-06 eess.SP

Changpeng He , Yang Lu , Wei Chen , Bo Ai , Kai-Kit Wong , Dusit Niyato This is my paper · ORCID
classification eess.SP
keywords fluidantennasystemsgraphneuralnetworksunsupervisedlearningMU-MISOsum-ratemaximizationenergyefficiencybeamformingpositionoptimization
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

The paper tries to establish that the antenna positions of a fluid antenna system (FAS) are optimizable variables that can be learned together with the beamforming vectors. The authors formulate sum-rate and energy-efficiency maximization for a multi-user MISO downlink and solve both with a single two-stage graph neural network (GNN), trained without labels by taking the inverse of the system utility as the loss. The claim is that this approach beats a fixed-antenna convex-optimization baseline in both objectives, generalizes to user counts unseen in training, and cuts inference time from seconds to tens of microseconds. If correct, it would make FAS a practical real-time option and replace per-scenario convex solvers with one trained network.

What carries the argument

The load-bearing object is the two-stage complex GNN built from complex graph attention layers and complex fully-connected layers, with a virtual node in the first stage to produce graph-level antenna positions and residual connections to reduce over-smoothing. Feasibility is enforced inside the architecture: positions are derived from unconstrained auxiliary variables through a normalized readout that guarantees $\delta_n \ge 0$ and $\sum_n \delta_n \le \delta_{\max}$, while beamformers are parameterized as $w_k = \sqrt{p_k}\,\bar{w}_k(\alpha_k)$ with $\alpha_k \in (0,1)$ blending MRT and ZF directions. The unsupervised loss $L(\Theta) = \frac{1}{M}\sum_m 1/U^{(m)}$ back-propagates through the SINR expressions, coupling the two stages during training.

What would settle it

Run the same two-stage GNN under a standard Rayleigh or Rician fading model, or on measured channels, and compare sum rate and energy efficiency with the equidistant-antenna CVX baseline. If the GNN-FAS advantage shrinks or reverses, the claim that the learned positions deliver real-world gains is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that the joint antenna-position and beamforming problem for a MU-MISO FAS can be mapped onto a fully connected graph and solved end-to-end by a two-stage complex-valued GNN. The first stage outputs feasible antenna positions through an auxiliary-variable parameterization that enforces ordering, minimum spacing, and total aperture; the second outputs beamforming vectors through a hybrid maximum-ratio-transmission/zero-forcing decomposition of each beam into a power coefficient and a normalized direction. Both stages share a single unsupervised loss equal to the inverse of the chosen utility, so no labeled optimal solutions are needed. In simulation, the two-stage GNN achieves higher sum rate and energy efficiency than the equidistant-antenna CVX baseline, keeps working for user counts not seen during training, and runs in roughly $0.05$ ms per inference.

Load-bearing premise

The numerical gains rest on a deterministic line-of-sight channel model with only large-scale path loss, so the learned antenna positions are not tested under the small-scale multipath fading that fluid antennas are meant to mitigate.

Editorial extensions

If this is right

  • The antenna-position stage carries most of the gain: given identical CVX beamforming, first-stage positions improve sum rate and EE over equidistant antennas in the (8,4) scenario.
  • The beamforming stage can run alone: for equidistant antennas it reaches near-CVX sum rate and EE with orders-of-magnitude faster inference, so it can serve as a standalone or warm-start beamformer.
  • Because the number of learnable parameters is independent of $K$, the same trained network applies to user counts it never saw during training.
  • The full system runs at about 0.05 ms per inference, which is fast enough for real-time fluid antenna control.

Reading between the lines

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

  • The line-of-sight-only channel model leaves open whether the learned positions help under multipath fading; testing the architecture on fading channels would directly extend the paper's claim.
  • Because the loss only needs a differentiable utility, the same two-stage scheme could be retargeted to max-min fairness, outage probability, or other objectives without changing the architecture.
  • The graph formulation suggests applicability to multi-cell or user-centric networks where users and access points are nodes, although the paper only simulates a single-cell MU-MISO setting.
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.

Referee Report

4 major / 4 minor

Summary. The paper considers a MU-MISO fluid antenna system (FAS) and formulates sum-rate and energy-efficiency maximization with constraints on transmit power and antenna spacing. It proposes a two-stage graph neural network (GNN) in which the first stage outputs antenna positions and the second stage outputs beamforming vectors, trained with an unsupervised loss equal to the inverse of the system utility. Numerical results report higher sum rate and EE than an equidistant-antenna CVX baseline and an MLP, with very low inference times and scalability to unseen numbers of users (Tables II–III), and the authors claim that the two stages can function separately (Table IV).

Significance. If the feasibility and performance claims are correct, the two-stage GNN is a useful contribution to real-time FAS optimization, extending graph-based learning to the new dimension of antenna positions. The paper is honest in comparing its second stage against CVX given identical antenna positions (Table IV), which isolates the beamforming loss, and the unsupervised loss is the natural utility objective rather than a fitted surrogate. However, the printed feasibility mappings contain internal errors that directly affect the reported numerical results, so the significance can only be assessed after those are corrected.

major comments (4)
  1. [§III.A.2, Eq. (14)] The power-normalization activation is inverted. When Σp_i > Pmax, the printed formula multiplies each p_k by (Σp_i)/Pmax, which is greater than 1, so the total output power becomes (Σp_i)^2/Pmax > Pmax. This violates constraint (3b) and contradicts the adjacent statement that the formula ensures Σ∥w_i∥² ≤ Pmax. The intended correction is p_k ← p_k · Pmax/Σp_i. As printed, every numerical result in Tables II–IV is obtained with a model that can exceed the power budget, making the reported rates/EE potentially unattainable.
  2. [§III.A.1, Eqs. (9), (23), (24)] The claim that δ_n in Eq. (9) satisfies (8) “regardless of the values of {ξ_i} and ξmax” is false. The variables ξ_i are recovered in Eq. (23) as the real part of an unconstrained complex-layer output, so they can be negative or sum to zero. If any ξ_i is negative, the corresponding δ_n is negative, violating the minimum-spacing constraint (3d); if the sum is zero, Eq. (9) divides by zero. Even when all ξ_i are positive, there is no built-in guarantee of a positive sum. A positivity-preserving parameterization (e.g., exp or softplus followed by normalization) is needed. This error undermines the central claim that the two-stage GNN outputs feasible antenna positions.
  3. [§III.D, Eq. (28)] The unsupervised loss in Eq. (28) is exactly the inverse of the system utility and contains no penalty for infeasibility. This is acceptable only if the architecture guarantees feasibility by construction. Given the errors in Eqs. (9) and (14), the training procedure can accept and reward solutions that violate the power budget or antenna-spacing constraints. The authors should either confirm the printed equations are typos and provide the actual implemented formulas, or add a projected/reparameterized model that provably satisfies (3b)–(3d) before the numerical comparisons can be trusted.
  4. [§II, Eq. (1)–(2) and §IV] The channel model is a deterministic line-of-sight steering vector with only large-scale path loss; no small-scale multipath fading is simulated. Since a primary motivation of FAS is to mitigate deep fading, the reported gains over equidistant antennas may not persist in the fading channels for which the system is designed. I recommend either simulating a fading channel (e.g., correlated Rayleigh or Rician over the fluid aperture) or explicitly limiting the conclusions to LoS scenarios.
minor comments (4)
  1. [Eqs. (25)–(26)] “Sigmod” should be “Sigmoid” in the text.
  2. [Tables II–III] The use of the † symbol for KTe values is confusing because the column header already separates KTr and KTe; please clarify in the caption that † denotes a test-only (unseen) user count rather than a different performance metric.
  3. [§III.A.1, Eq. (9)] Even with positive ξ_i, the denominator Σ_{i∈N\{1}} ξ_i should be explicitly normalized to avoid ambiguity; a softmax-style parameterization would make the feasibility claim cleaner and easier to verify.
  4. [§III.A.1, after Eq. (8)] The statement that setting x1 = 0 causes no optimality loss should be justified in one sentence, e.g., by noting that a common shift of all antenna positions multiplies each steering vector by a common phase factor that cancels in the SINR expression.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the two-stage GNN is a self-contained learning-to-optimize pipeline whose loss is the system utility itself, not a fitted parameter disguised as a prediction.

full rationale

The paper's derivation chain is self-contained with respect to the kinds of circularity defined here. The two-stage GNN maps the steering-angle inputs {θ_i} to auxiliary variables {ξ_i}, ξ_max, {p_i}, and {α_i}, then recovers antenna positions through Eq. (9)-(10) and beamformers through Eq. (11)-(12). The unsupervised loss in Eq. (28) is simply the reciprocal of the system utility, which is the standard learning-to-optimize setup rather than a circular derivation: the network is trained to maximize the same objective that is later reported, and no fitted constant is relabeled as a prediction. The cited prior works, including the authors' own GNN papers [5]-[11], are used for architectural components (multi-head attention, residual connections, complex graph attention layers), not as load-bearing evidence that forces the numerical conclusions; the central numerical comparison is against an independently solved CVX baseline. The use of [4] to justify setting x1=0 is a cited external result, and the HZF learning scheme is cited from [12], neither being a self-citation chain that imports a contested uniqueness claim. Consequently, there is no step in which a claimed output reduces by construction to an input, and no self-citation is load-bearing for the main result. Note that the feasibility-guarantee equations (9) and (14) contain apparent mathematical errors that would violate constraints as written, but those are internal correctness issues rather than circularity, because they do not make an output equal to its input by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper states a deterministic LOS channel as the simulation model (Eq. 1-2), invokes an optimality-preserving normalization x1=0 from [4], and assumes HZF beamforming does not materially reduce optimality. The only hand-chosen numbers are the neural network's architectural hyperparameters, which are not fully reported. No new physical entities are introduced.

free parameters (1)
  • GNN architecture hyperparameters (layer counts, feature dimensions, attention heads) = not specified
    The numbers G1-G4, F1-F4, Vg, and Z are chosen by hand but not reported; these affect achieved performance.
assumptions (3)
  • domain assumption The FAS channel is modeled as a deterministic line-of-sight steering vector with no small-scale fading.
    Eq. (1) defines h(x, theta_k) as a plane-wave steering vector; Eq. (2) uses only large-scale path loss d_k. The deep-fading scenario motivating FAS is not present in the numerical evaluation.
  • domain assumption Setting x1=0 entails no optimality loss.
    Invoked in Section III-A following [4]; relies on global phase compensation by beamforming, which is plausible but not proven in this paper.
  • domain assumption The HZF structure (hybrid MRT/ZF) does not significantly degrade the optimality of beamforming.
    Invoked in Section III-A2; the paper admits it may slightly degrade optimality, and Table IV quantifies a small gap for one scenario.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Neural Network Enabled Fluid Antenna Systems: A Two-Stage Approach." pith.science (2026). https://pith.science/paper/66RDLZV6

@misc{pith2026250203922,
  author       = {Pith},
  title        = {Pith review of: Graph Neural Network Enabled Fluid Antenna Systems: A Two-Stage Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66RDLZV6}},
  note         = {Machine review of arXiv:2502.03922}
}
read the original abstract

An emerging fluid antenna system (FAS) brings a new dimension, i.e., the antenna positions, to deal with the deep fading, but simultaneously introduces challenges related to the transmit design. This paper proposes an ``unsupervised learning to optimize" paradigm to optimize the FAS. Particularly, we formulate the sum-rate and energy efficiency (EE) maximization problems for a multiple-user multiple-input single-output (MU-MISO) FAS and solved by a two-stage graph neural network (GNN) where the first stage and the second stage are for the inference of antenna positions and beamforming vectors, respectively. The outputs of the two stages are jointly input into a unsupervised loss function to train the two-stage GNN. The numerical results demonstrates that the advantages of the FAS for performance improvement and the two-stage GNN for real-time and scalable optimization. Besides, the two stages can function separately.

Figures

Figures reproduced from arXiv: 2502.03922 by the authors.

Figure 1
Figure 1. Illustration of MU-MISO FAS. GAL 。。。 GAL FL 。。。 FL i GALs FLs GALs FLs i max  Antenna position learning stage  i x GALs FLs GALs FLs i max  Antenna position learning stage  i x GALs FLs GALs FLs i  pi wi GALs FLs GALs FLs i  pi wi Beamforming learning stage (1) hx( , )  k  Loss function (28) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The structure of the two-stage GNN. from the FAS to the k-th (k ∈ K ≜ {1, . . . , K}) user as θk ∈ [0, π), and the corresponding steer vector is given by h(x, θk) = h e i 2π λ x1 cos(θk) , ei 2π λ x2 cos(θk) , . . . , ei 2π λ xN cos(θk) iT , (1) where x ≜ [x1, x2, . . . , xN ] T represents the antenna position vector and λ denotes the wavelength. The BS transmits K different information symbols for the K users over … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. EM-Guided Graph Learning for Fluid Antenna Beamforming under Current-Domain Constraints

    cs.IT 2026-08 conditional novelty 5.0 of 10

    A mutual-impedance-aware graph network that predicts fluid-antenna port layouts and precoders under current-domain constraints comes within about 0.7 bit/s/Hz of an iterative optimizer while cutting configuration late...

Reference graph

Works this paper leans on

14 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    K. -K. Wong, A. Shojaeifard, K. -F. Tong and Y . Zhang, ”Fluid antenna systems,” IEEE Trans. Wireless Commun., vol. 20, no. 3, pp. 1950-1962, Mar. 2021

    work page 1950

  2. [2]

    K. -K. Wong, K. -F. Tong and C. -B. Chae, ”Fluid antenna system—part II: Research opportunities,” IEEE Commun. Lett. , vol. 27, no. 8, pp. 1924-1928, Aug. 2023

    work page 1924

  3. [3]

    C. Wang, Z. Li, K. -K. Wong, R. Murch, C. -B. Chae and S. Jin, ”AI- empowered fluid antenna systems: Opportunities, challenges, and future directions,” IEEE Wireless Commun. , vol. 31, no. 5, pp. 34-41, Oct. 2024

    work page 2024

  4. [4]

    Deep learning enabled multicast beamforming with mov- able antenna array,

    J.-M. Kang, “Deep learning enabled multicast beamforming with mov- able antenna array,” IEEE Wireless Commun. Lett. , vol. 13, no. 7, pp. 1848-1852, Jul. 2024

    work page 2024

  5. [5]

    Graph neural networks for wireless networks: Graph representation, architecture and evaluation,

    Y . Lu, Y . Li, R. Zhang, W. Chen, B. Ai, and D. Niyato, “Graph neural networks for wireless networks: Graph representation, architecture and evaluation,” early accessed in IEEE Wireless Commun. , 2024

    work page 2024

  6. [6]

    Deep learning for energy efficient beamforming in MU-MISO networks: A GAT-based approach,

    Y . Li, Y . Lu, R. Zhang, B. Ai, and Z. Zhong, “Deep learning for energy efficient beamforming in MU-MISO networks: A GAT-based approach,” IEEE Wireless Commun. Lett. , vol. 12, no. 7, pp. 1264-1268, Jul. 2023

    work page 2023

  7. [7]

    GNN-based beamforming for sum-rate maximization in MU-MISO networks,

    Y . Li, Y . Lu, B. Ai, O. A. Dobre, Z. Ding, and D. Niyato, “GNN-based beamforming for sum-rate maximization in MU-MISO networks,” IEEE Trans. Wireless Commun. , vol. 23, no. 8, pp. 9251-9264, Aug. 2024

    2024

  8. [8]

    GNN-enabled max-min fair beamforming,

    Y . Li, Y . Lu, B. Ai, Z. Zhong, D. Niyato, and Z. Ding, “GNN-enabled max-min fair beamforming,” IEEE Trans. V eh. Technol., vol. 73, no. 8, pp. 12184-12188, Aug. 2024

    work page 2024

Show all 14 references
  1. [9]

    Q. Wang, Y . Lu, W. Chen, B. Ai, Z. Zhong and D. Niyato, ”GNN- enabled optimization of placement and transmission design for UA V communications,” early accessed in IEEE Trans. V eh. Technol, 2024

    2024

  2. [10]

    ICNet: GNN-enabled beamforming for MISO interference channels with statistical CSI,

    C. He et al., “ICNet: GNN-enabled beamforming for MISO interference channels with statistical CSI,” IEEE Trans. V eh. Technol, vol. 73, no. 8, pp. 12225-12230, Aug. 2024

    2024

  3. [11]

    Song et al., ”A deep learning framework for physical-layer secure beamforming,” early accessed in IEEE Trans

    Z. Song et al., ”A deep learning framework for physical-layer secure beamforming,” early accessed in IEEE Trans. V eh. Technol, 2024

    2024

  4. [12]

    Hybrid MRT and ZF learning for energy-efficient transmission in multi-RIS-assisted networks,

    W. Guo et al., “Hybrid MRT and ZF learning for energy-efficient transmission in multi-RIS-assisted networks,” IEEE Trans. V eh. Technol., vol. 73, no. 8, pp. 12247-12251, Aug. 2024

    2024

  5. [13]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and S. Christian, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in Proc. ICML, 2015

    2015

  6. [14]

    Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification,” in Proc. ICCV, pp. 1026-1034, 2015

    2015

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.