Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

Deep Learning Optimization of Two-State Pinching Antennas Systems

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A graph-neural-network policy can learn which pinching antennas to activate along a waveguide and reach 93% of the optimal rate on 50 antennas, staying above 90% on arrays up to 1000.

desk verdict A solid engineering contribution with a clean formulation, but the empirical claims need a defined metric and verified optimality before they can be trusted. read the letter →

arxiv 2507.06222 v1 pith:FIKY46CD submitted 2025-07-08 cs.LG

Odysseas G. Karagiannidis , Victoria E. Galanopoulou , Panagiotis D. Diamantoulakis , Zhiguo Ding , Octavia Dobre This is my paper
classification cs.LG MSC 90C2090C2790C3268T07
keywords pinchingantennasantennaactivationquadraticfractional0-1programminggraphneuralnetworksdistributedattentionpolicyuserlocationuncertaintyratemaximizationnetworkoptimization
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 show that a graph neural network with attention can learn, in real time, which fixed-position pinching antennas to activate along a waveguide so that a user's communication rate is nearly maximized. The activation task is inherently combinatorial—each antenna contributes a phase-shifted signal while also splitting transmit power—and the paper models it as a quadratic fractional 0-1 program. Trained only on 50-antenna instances, the proposed GNN+DisPN model is claimed to reach 93% of the optimal SNR on 50 antennas and over 90% on arrays of 100, 200, 500, and 1000 antennas, without retraining. It also keeps accuracy above 82% when user location estimates are noisy, where a single-shot optimizer degrades sharply. If the claim holds, it offers a scalable, data-driven way to run this new class of flexible-antenna systems.

What carries the argument

The central object is the effective-gain vector $B\in\mathbb{C}^N$, whose entries $B_n = h_n e^{-j\theta_n}$ combine free-space path loss and phase with waveguide-induced phase shift $\theta_n$ from the feed point. From it the paper forms $Q = \operatorname{Re}(B B^H)$, a real symmetric positive-semidefinite matrix of rank at most 2 whose entry $Q_{ij}$ measures the phase-aligned correlation between antennas $i$ and $j$; the objective reduces to maximizing $a^\top Q a / (1^\top a)$. The learning machinery is a graph neural network over a user-plus-antennas graph with edge features $(|B_n|, \angle B_n)$, followed by a two-stage attention policy: the user embedding acts as a query against antenna key/value embeddings, and a sharpened tanh scoring produces per-antenna activation probabilities. Training uses labels from an iterative fractional-programming solver plus a composite loss that upweights active antennas and penalizes deviations of the soft-antenna SNR from the optimal SNR.

What would settle it

Use the same dataset-generation procedure but have the solver report its optimality gap at termination, or re-solve small instances (where brute force is feasible) to certify the labels; then recompute the GNN+DisPN SNR accuracy against the certified optima. If the 90%+ accuracy on 100-antenna arrays does not survive certified labels, the central claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that activation vectors produced by a message-passing graph network with a distributed attention policy achieve SNR within a few percent of the optimum: 87% on 50 antennas with classification loss, 93% with an augmented SNR-aware loss, and 91–95% across 100 to 1000 antennas. The optimization surface is the fractional objective $\max_a \frac{|a^\top B|^2}{\|a\|_0}$, where $a$ is the binary activation vector and $B_n = h_n e^{-j\theta_n}$ absorbs each antenna's free-space channel and the waveguide-induced phase. The same learned policy beats the nearest-antenna heuristic (about 7%) by a wide margin, outperforms a conventional co-located MIMO baseline by more than 3 dB in achievable SNR, and under position uncertainty a Monte-Carlo mean-activation rule keeps accuracy above 82%.

Load-bearing premise

The accuracy numbers are measured against activation labels produced by an iterative fractional-programming solver, and the paper assumes those labels are globally optimal for every instance size up to 1000 antennas, even though no optimality gaps or solver termination tolerances are reported.

Editorial extensions

If this is right

  • A fixed pinching-antenna array can be operated by a learned policy at inference time, avoiding NP-hard combinatorial optimization in each scheduling slot.
  • A model trained on 50-antenna systems transfers to 100, 200, 500, and 1000 antennas with no retraining, so the approach scales with array size.
  • Attention scores give an interpretable per-antenna importance ranking; a top-K post-processing step keeps SNR accuracy above 90% of optimal for selection ratios between 0.7 and 1.0.
  • The Monte-Carlo mean-activation policy makes user-location uncertainty a handled input rather than a failure mode, and automatically activates fewer antennas as uncertainty grows.
  • Pinching-antenna systems operating under the learned policy can exceed a conventional co-located MIMO system with the same antenna count by more than 3 dB in achievable SNR.

Reading between the lines

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

  • Our inference: because Q has rank at most 2, the optimal activation decision is effectively governed by a two-dimensional subspace of antenna phase alignments; that structural simplicity likely explains why the graph model generalizes across array sizes.
  • Our inference: the same graph-plus-attention pipeline should transfer to other quadratic fractional 0-1 resource-selection problems, since the paper shows that a surrogate B can be reconstructed from the spectral decomposition of Q when direct channel features are unavailable.
  • Our inference: a natural next experiment is to vary the transmit SNR during training and testing; the current study fixes transmit SNR at 40 dB, so the model's ability to adapt to changing link budgets remains untested.
  • Our inference: the reported accuracies are relative to solver labels, so an independent validation against certified optima on small instances would cleanly separate the model's approximation error from any label suboptimality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The manuscript studies the selection of activated pinching antennas in a single-waveguide downlink. It formulates the rate-maximizing binary antenna selection as a fractional 0-1 quadratic program with Q = Re(BB^H), then trains supervised MLP, GNN+MLP, and GNN+DisPN policies on Gurobi-generated optimal labels. The central claim is that the attention-augmented GNN achieves 87% SNR accuracy at N=50, 93% with the proposed augmented loss, and over 90% for N=100, 200, 500, and 1000 antennas, while remaining robust to user-location uncertainty through Monte Carlo aggregation. The paper also reports near-Gurobi SNR with substantially lower inference cost.

Significance. If substantiated, the claim that a lightweight GNN can replace an expensive combinatorial solver for pinching-antenna activation, generalize across antenna counts, and provide robustness to location uncertainty would be a useful engineering contribution to ML-based combinatorial wireless optimization. The graph/attention design is natural for this problem, the formulation in Section III is sound, and the inclusion of a distance heuristic baseline plus complexity measurements is a strength. However, the quantitative support is not falsifiable as written: the headline accuracy metric is never defined, the Gurobi reference solutions are uncertified, and the evaluation loop uses the same solver as teacher and benchmark. Once these issues are addressed with explicit definitions, optimality gaps, and an external validation method, the contribution could be significant.

major comments (5)
  1. [Section VI, Tables II and IV] The quantity "SNR accuracy" is never defined. It is the headline metric in Tables II and IV and in Figures 8 and 9, but no equation specifies whether it is the average ratio gamma_model/gamma_opt, the fraction of instances where gamma_model exceeds some fraction of gamma_opt, or something else. Similarly, "accuracy in terms of achievable rate" and "bitwise accuracy" in Table II lack definitions. Without a precise definition, the aggregation rule, and the test-set size, the central quantitative claims cannot be reproduced or compared with future work. Please define the metric explicitly and report the number of test instances.
  2. [Appendix B and Section IV-D] Appendix B states that the Dinkelbach subproblems in Eq. (62) are "concave maximizations (since Q is positive semi-definite)". This is incorrect: for PSD Q, x^T Q x is convex, so Eq. (62) maximizes a convex quadratic over binary variables, which is not a concave maximization and is NP-hard in general. The claim that Dinkelbach's method "ensures that the final solution is globally optimal" therefore depends on Gurobi solving these hard subproblems exactly. Section IV-D reports no MIPGap, time limit, or termination criterion for dataset generation, especially for N up to 1000. Please correct the characterization in Appendix B, report the Gurobi settings, and provide optimality gaps or an independent certificate of global optimality.
  3. [Section IV-D and Section VI] The evaluation loop is circular in its current form: the supervised training labels and the reference solutions in Tables II and IV and Figures 6, 7, 8, and 10 are all produced by the same Gurobi-based procedure. The reported 87-95% accuracies therefore measure agreement with the solver's output, not with the true optimum of problem (18), unless the solver is certified to be exact. Please add a solver-independent benchmark, for example exhaustive search for small N or an upper bound on the fractional objective, so that both Gurobi and the learned policies are compared against a reference that does not depend on the training-label generator.
  4. [Section V-B and Figure 8] Figure 8 compares the GNN+DisPN model with Monte Carlo aggregation over M noisy position samples against a single Gurobi run on one noisy input realization. This comparison conflates the benefit of ensembling over M samples with the quality of the learned policy. To make the robustness claim meaningful, either provide Gurobi with the same M-sample aggregation procedure (for example, optimizing each sampled position and then aggregating the resulting activation vectors) or compare both methods at fixed M and report M. As written, the comparison is not like-for-like.
  5. [Tables II-IV and Figures 8-9] All reported accuracies are point estimates with no standard deviations, no number of random seeds, and no test-set sizes for N=50, 100, 200, 500, and 1000. The claimed improvements (87% to 93% from the augmented loss, and 81% to 91% for GNN+MLP versus GNN+DisPN at N=100) may be within run-to-run variability. Please report multiple training seeds, error bars or confidence intervals, and the number of test instances for each antenna count.
minor comments (6)
  1. [Section II-A] The channel definitions are inconsistent: Eq. (5) and Eq. (7) both define h_n but with different expressions, Eq. (7) contains an undefined psi_m term, and Eq. (11) uses h_{n,m}. Please harmonize the notation.
  2. [Eq. (18)] The optimization problem in Eq. (18) uses "max_x" while the decision variable is a; the subscript should be a.
  3. [Section IV-E, Eqs. (46)-(47)] The signal-aware loss uses a differentiable soft SNR estimate gamma_hat, but no equation specifies how the sigmoid probabilities replace the binary activations in Eq. (16), especially how the normalization by N_a = ||a||_0 is handled. Please state this approximation explicitly.
  4. [Figure 6] The y-axis label reads "Accuracy" while the text describes the mean SNR ratio; the caption and axis label should be aligned.
  5. [Eq. (50)] The position uncertainty is modeled as epsilon ~ N(0, sigma_p^2 I_3), but earlier the user position is described with z in [0,1] and only x,y vary on the ground plane. Please clarify whether z is also uncertain and how this affects the channel computation.
  6. [Table V] Learning rates and loss coefficients are given as ranges ("10^-4 -> 10^-5", "0.5 -> 0.3") rather than schedules; please state the scheduling rule.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learning-to-optimize pipeline is a standard supervised benchmark loop, and the central problem derivation is self-contained.

full rationale

I walked the derivation chain from the system model to the learning results. The objective in Eq. (17), max_a |a^T B|^2 / ||a||_0, is converted algebraically in Proposition III.1 to the fractional 0-1 quadratic program max_a a^T Q a / (1^T a) with Q = Re(B B^H). This is a direct equivalence, not a definition of the target in terms of the learned output. The neural networks are trained on Gurobi-generated optimal activation labels and then evaluated against Gurobi-derived SNR references; this is a teacher-benchmark loop, but it is not circular in the logical sense because the test instances are held out, the model never sees the labels at inference, and the reported SNR values are computed from the physical equation (16) using the model's predicted activations. No fitted parameter is renamed as a prediction, and no equation reduces the reported accuracy to the training loss by construction. The citation of Dinkelbach's algorithm [33] is an external, standard method, and the GNN+DisPN architecture citation [32] is not load-bearing for the physical derivation. I note two non-circular concerns that belong under correctness and reproducibility rather than circularity: (i) Appendix B's claim that the binary quadratic subproblems are 'concave maximizations' is mathematically incorrect, and no optimality gaps or solver termination criteria are reported, so the Gurobi references are not certified global optima; (ii) the 'SNR accuracy' metric is never formally defined, making the headline percentages hard to reproduce. These issues weaken the empirical claims but do not make the derivation circular, because the model's output is not defined in terms of the benchmark it is compared against. Therefore the paper's central derivation is self-contained and no circular step is exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The work adds no new physical entities or fitted physical constants. The free parameters are only training loss weights. The load-bearing assumptions are the channel model, global optimality of the Gurobi labels, transferability across array sizes, and the Gaussian uncertainty model.

free parameters (1)
  • Loss weighting coefficients (lambda_1, lambda_gamma, lambda_c), positive-class weight alpha, sharpening c = lambda_1: 0.5 to 0.3; lambda_gamma: 2 to 8; lambda_c: 100 to 20; alpha=1.6; c=10
    Chosen by hand in Table V to balance classification and SNR terms; the reported accuracies depend on this tuning.
assumptions (4)
  • domain assumption Free-space spherical-wave channel h_n = exp(-j2*pi/lambda * ||psi-psi_n||)/||psi-psi_n|| and waveguide phase theta_n = 2*pi/lambda_g * ||psi_n-psi_0||
    Section II Eqs. (5)-(9); no experimental validation is provided.
  • standard math Dinkelbach's algorithm with exactly solved subproblems finds the global optimum of (18), and Gurobi solves each subproblem exactly
    Appendix B; this is the basis for all ground-truth labels.
  • domain assumption A policy trained on N=50 instances transfers to N=1000 instances without retraining
    Section VI Table IV; no theoretical guarantee, only reported test accuracy.
  • domain assumption User position errors are i.i.d. Gaussian with covariance sigma_p^2 I, and averaging M forward passes approximates the robust decision
    Section V Eqs. (50)-(53); M is not specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Learning Optimization of Two-State Pinching Antennas Systems." pith.science (2026). https://pith.science/paper/FIKY46CD

@misc{pith2026250706222,
  author       = {Pith},
  title        = {Pith review of: Deep Learning Optimization of Two-State Pinching Antennas Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FIKY46CD}},
  note         = {Machine review of arXiv:2507.06222}
}
read the original abstract

The evolution of wireless communication systems requires flexible, energy-efficient, and cost-effective antenna technologies. Pinching antennas (PAs), which can dynamically control electromagnetic wave propagation through binary activation states, have recently emerged as a promising candidate. In this work, we investigate the problem of optimally selecting a subset of fixed-position PAs to activate in a waveguide, when the aim is to maximize the communication rate at a user terminal. Due to the complex interplay between antenna activation, waveguide-induced phase shifts, and power division, this problem is formulated as a combinatorial fractional 0-1 quadratic program. To efficiently solve this challenging problem, we use neural network architectures of varying complexity to learn activation policies directly from data, leveraging spatial features and signal structure. Furthermore, we incorporate user location uncertainty into our training and evaluation pipeline to simulate realistic deployment conditions. Simulation results demonstrate the effectiveness and robustness of the proposed models.

Figures

Figures reproduced from arXiv: 2507.06222 by the authors.

Figure 1
Figure 1. Pinching antenna system PASs arising from location estimation errors, affecting the constructive superposition of delayed signal components at the receiver. To enhance the model’s robustness, we introduce a modified activation policy that accounts for this uncertainty and incorporate it into the simulations. This approach enables reliable antenna activation even in the presence of imprecise location estimates. C. St… view at source ↗
Figure 2
Figure 2. Architecture overview of the GNN + MLP model for antenna activation [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Graph representation of a user connected to [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: GNN-DisPN architecture 3) Advantages and Capabilities: This attention-based for￾mulation provides several advantages: (i) it enables distributed reasoning where the user embedding dynamically interacts with all antennas; (ii) it allows the model to focus selectively on…
Figure 5
Figure 5. Figure 5: Accuracy Convergence. Given that the GNN+DisPN architecture achieved the high￾est accuracy among the evaluated models, we further focused our investigation on this model. To improve its robustness and performance, we trained it with the augmented loss function presente…
Figure 7
Figure 7. Figure 7: Antennas activation percentage. puted as the number of selected antennas divided by the total number N, for both the model and the optimal Gurobi-based solution, across increasing antenna counts. As N grows, both curves exhibit a decreasing trend that stabilizes beyond…
Figure 8
Figure 8. Figure 8: Accuracy in uncertain user position coordinates [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 10
Figure 10. Figure 10: Achievable SNR vs N available antennas. the scalability of the proposed method. VII. CONCLUSION In this work, we investigated the challenging problem of optimizing antenna activation in PASs in order to maximize the user’s rate. We formulated the task as a QF01P probl…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Joint Transmit and Pinching Beamforming Optimization in Pinching Antenna-Assisted Symbiotic Radio Systems

    eess.SP 2025-08 reject novelty 6.0 of 10

    Jointly optimizing transmit beamforming and pinching-antenna positions in a backscatter-assisted symbiotic radio system is claimed to raise achievable sum rate by 17-35% over fixed-antenna baselines.

  2. EGS-SLAM: RGB-D Gaussian Splatting SLAM with Events

    cs.RO 2025-08 unverdicted novelty 5.0 of 10

    EGS-SLAM claims to fuse events with RGB-D in Gaussian Splatting SLAM to beat blur, but the supplied full text is an unrelated paper, so the claim is unverifiable.

Reference graph

Works this paper leans on

33 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    Towards smart and reconfigurable environment: Intelligent reflecting surface aided wireless network,

    Q. Wu and R. Zhang, “Towards smart and reconfigurable environment: Intelligent reflecting surface aided wireless network,”IEEE Commun. Mag., vol. 58, no. 1, pp. 106–112, 2020

    2020

  2. [2]

    Fluid antenna systems,

    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, 2021

    1950

  3. [3]

    Fluid antenna systems enabling 6g: Principles, applications, and research directions,

    T. Wu, K. Zhi, J. Yao, X. Lai, J. Zheng, H. Niu, M. Elkashlan, K.-K. Wong, C.-B. Chae, Z. Dinget al., “Fluid antenna systems enabling 6g: Principles, applications, and research directions,”arXiv preprint arXiv:2412.03839, 2024

    arXiv 2024

  4. [4]

    Modeling and performance analysis for movable antenna enabled wireless communications,

    L. Zhu, W. Ma, and R. Zhang, “Modeling and performance analysis for movable antenna enabled wireless communications,”IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 6234–6250, Jun. 2024

    2024

  5. [5]

    Mimo capacity characterization for movable antenna systems,

    W. Ma, L. Zhu, and R. Zhang, “Mimo capacity characterization for movable antenna systems,”IEEE Trans. Wireless Commun., vol. 23, no. 4, pp. 3392–3407, 2024. 13

    work page 2024

  6. [6]

    A survey of integrating generative artificial intelligence and 6g mobile services: Architectures, solutions, technologies and outlooks,

    Y .-J. Liu, H. Du, X. Xu, R. Zhang, G. Feng, B. Cao, D. Niyato, D. I. Kim, A. Jamalipour, K. B. Letaief, and R. Tafazolli, “A survey of integrating generative artificial intelligence and 6g mobile services: Architectures, solutions, technologies and outlooks,”IEEE Trans. Cogn. Commun. Netw., pp. 1–1, 2025

    work page 2025

  7. [7]

    Toward generative 6g simulation: An experimental multi-agent llm and ns-3 integration,

    F. Rezazadeh, A. A. Gargari, S. Lagen, H. Song, D. Niyato, and L. Liu, “Toward generative 6g simulation: An experimental multi-agent llm and ns-3 integration,”arXiv preprint arXiv:2503.13402, 2025. [Online]. Available: https://arxiv.org/abs/2503.13402

    arXiv 2025

  8. [8]

    Pinching antenna,

    NTT DOCOMO, INC., “Pinching antenna,” https://www.docomo.ne.jp/ english/info/media center/event/mwc21/pdf/06 MWC2021 docomo Pinching Antenna en.pdf, 2021, accessed: Apr. 7, 2025

    work page 2021

Show all 33 references
  1. [9]

    Pinching antennas: Principles, applications and challenges,

    Z. Yang, N. Wang, Y . Sun, Z. Ding, G. K. Karagiannidis, V . Wong, and O. A. Dobre, “Pinching antennas: Principles, applications and challenges,”IEEE Commun. Mag., 2023, submitted, under review

    2023

  2. [10]

    Holographic mimo for leo satellite communications aided by reconfigurable holographic surfaces,

    R. Deng, B. Di, H. Zhang, H. V . Poor, and L. Song, “Holographic mimo for leo satellite communications aided by reconfigurable holographic surfaces,”IEEE J. Sel. Areas Commun., pp. 1–1, 2023

    2023

  3. [11]

    Flexible-antenna systems: A pinching-antenna perspective,

    Z. Ding, R. Schober, and H. Vincent Poor, “Flexible-antenna systems: A pinching-antenna perspective,”IEEE Trans. Commun., pp. 1–1, 2025

    2025

  4. [12]

    Los blockage in pinching-antenna systems: Curse or blessing?

    Z. Ding and H. V . Poor, “Los blockage in pinching-antenna systems: Curse or blessing?”arXiv preprint, vol. arXiv:2503.08554, 2025

    2025 arXiv

  5. [13]

    Array gain for pinching- antenna systems (pass),

    C. Ouyang, Z. Wang, Y . Liu, and Z. Ding, “Array gain for pinching- antenna systems (pass),”IEEE Commun. Lett., pp. 1–1, 2025

    2025

  6. [14]

    Pinching antennas: Principles, applica- tions and challenges,

    Z. Yang, N. Wang, Y . Sun, Z. Ding, R. Schober, G. K. Karagiannidis, V . W. Wong, and O. A. Dobre, “Pinching antennas: Principles, applica- tions and challenges,”arXiv preprint arXiv:2501.10753, 2025

    2025 arXiv

  7. [15]

    Resource allocation for pinching-antenna systems: State-of-the-art, key techniques and open issues,

    M. Zeng, J. Wang, O. A. Dobre, Z. Ding, G. K. Karagiannidis, R. Schober, and H. V . Poor, “Resource allocation for pinching-antenna systems: State-of-the-art, key techniques and open issues,”arXiv preprint arXiv:2506.06156, 2025

    2025 arXiv

  8. [16]

    Minimum data rate maximization for uplink pinching-antenna systems,

    S. A. Tegos, P. D. Diamantoulakis, Z. Ding, and G. K. Karagiannidis, “Minimum data rate maximization for uplink pinching-antenna systems,” IEEE Wireless Commun. Lett., vol. 14, no. 5, pp. 1516–1520, 2025

    2025

  9. [17]

    Rate maximization for downlink pinching-antenna systems,

    Y . Xu, Z. Ding, and G. K. Karagiannidis, “Rate maximization for downlink pinching-antenna systems,”IEEE Wireless Commun. Lett., vol. 14, no. 5, pp. 1431–1435, 2025

    2025

  10. [19]

    Performance analysis of pinching- antenna systems,

    D. Tyrovolas, S. A. Tegos, P. D. Diamantoulakis, S. Ioannidis, C. K. Liaskos, and G. K. Karagiannidis, “Performance analysis of pinching- antenna systems,”IEEE Trans. Cogn. Commun. Netw., 2025

    2025

  11. [20]

    Secrecy rate maximization with artificial noise for pinching-antenna systems,

    P. P. Papanikolaou, D. Bozanis, S. A. Tegos, P. D. Diamantoulakis, and G. K. Karagiannidis, “Secrecy rate maximization with artificial noise for pinching-antenna systems,”arXiv preprint arXiv:2504.10656, 2025

    2025 arXiv

  12. [21]

    Pinching- antenna systems (pass)-enabled secure wireless communications,

    G. Zhu, X. Mu, L. Guo, S. Xu, Y . Liu, and N. Al-Dhahir, “Pinching- antenna systems (pass)-enabled secure wireless communications,” arXiv preprint arXiv:2504.13670, 2025. [Online]. Available: https: //doi.org/10.48550/arXiv.2504.13670

    2025 doi

  13. [22]

    Integrated sensing and communications for pinching-antenna systems (pass),

    Z. Zhang, Z. Wang, X. Mu, B. He, J. Chen, and Y . Liu, “Integrated sensing and communications for pinching-antenna systems (pass),”arXiv preprint arXiv:2504.07709, 2025. [Online]. Available: https://arxiv.org/abs/2504.07709

    2025 arXiv

  14. [23]

    Cram´er-rao bounds for integrated sensing and communications in pinching-antenna systems,

    D. Bozanis, V . K. Papanikolaou, S. A. Tegos, and G. K. Karagiannidis, “Cram´er-rao bounds for integrated sensing and communications in pinching-antenna systems,”arXiv preprint arXiv:2505.01333, 2025. [Online]. Available: https://arxiv.org/abs/2505.01333

    2025 arXiv

  15. [24]

    Graph neural network enabled pinching antennas,

    X. Xie, Y . Lu, and Z. Ding, “Graph neural network enabled pinching antennas,”arXiv preprint arXiv:2502.05447, 2025

    2025 arXiv

  16. [25]

    Antenna activation and resource allocation in multi- waveguide pinching-antenna systems,

    K. Wang, Z. Ding, and G. K. Karagiannidis, “Antenna activation and resource allocation in multi- waveguide pinching-antenna systems,” arXiv, 2025

    2025

  17. [27]

    Antenna activation for noma assisted pinching-antenna systems,

    K. Wang, Z. Ding, and R. Schober, “Antenna activation for noma assisted pinching-antenna systems,”IEEE Wireless Commun. Lett., vol. 14, no. 5, pp. 1526–1530, 2025. [Online]. Available: https://ieeexplore.ieee.org/document/10912473

    2025

  18. [28]

    Available: https://arxiv.org/abs/2502.14250

    [Online]. Available: https://arxiv.org/abs/2502.14250

    arXiv

  19. [29]

    Achieving network resilience through graph neural network-enabled deep reinforcement learning,

    X. Li, T. Zhang, J. Wang, Z. Han, J. Liu, J. Kang, D. Niyato, and A. Jamalipour, “Achieving network resilience through graph neural network-enabled deep reinforcement learning,”IEEE Netw., pp. 1–1, 2025

    2025

  20. [30]

    Energy- efficient resource allocation for noma-assisted uplink pinching-antenna systems,

    M. Zeng, X. Li, J. Wang, G. Huang, O. A. Dobre, and Z. Ding, “Energy- efficient resource allocation for noma-assisted uplink pinching-antenna systems,”arXiv preprint arXiv:2505.07555, 2025

    2025 arXiv

  21. [31]

    Gpass: Deep learning for beamforming in pinching-antenna systems (pass),

    J. Guo, Y . Liu, and A. Nallanathan, “Gpass: Deep learning for beamforming in pinching-antenna systems (pass),”arXiv preprint arXiv:2502.01438, 2025. [Online]. Available: https://doi.org/10.48550/ arXiv.2502.01438

    2025 doi arXiv

  22. [32]

    Spatio-temporal gnn-based cell-free massive mimo network with maximal benefit-cost ratio,

    J. Jiang, Y . Li, Y . Ye, D. Feng, J. Zhang, W. Sutthiphan, and D. Niy- ato, “Spatio-temporal gnn-based cell-free massive mimo network with maximal benefit-cost ratio,”IEEE Trans. V eh. Technol., pp. 1–5, 2025

    2025

  23. [33]

    On nonlinear fractional programming,

    W. Dinkelbach, “On nonlinear fractional programming,”Management Science, vol. 13, no. 7, pp. 492–498, 1967

    1967

  24. [34]

    A reinforcement learning approach for optimizing multiple traveling salesman problems over graphs,

    Y . Hu, Y . Yao, and W. S. Lee, “A reinforcement learning approach for optimizing multiple traveling salesman problems over graphs,” Knowledge-Based Systems, vol. 204, p. 106244, 2020

    2020

  25. [2025]

    Available: https://arxiv.org/abs/2502.12365

    [Online]. Available: https://arxiv.org/abs/2502.12365

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.