Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Beyond-Diagonal RIS Under Non-Idealities: Learning-Based Architecture Discovery and Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read This paper argues that the optimal wiring topology of a beyond-diagonal reconfigurable intelligent surface changes once real hardware non-idealities are included, and that a two-tier learning framework can discover those topologies without

desk verdict A credible learning-based architecture discovery framework for non-ideal BD-RIS with a genuinely useful ideal-case validation, but the headline non-ideal insights rest on a single unvalidated ML pipeline and should be treated as suggestive, not established. read the letter →

arxiv 2510.15701 v2 pith:QMN52HMX submitted 2025-10-17 cs.IT cs.AIeess.SPmath.IT

classification cs.ITcs.AIeess.SPmath.IT
keywords beyond-diagonalRISarchitecturediscoverymutualcouplinglossycomponentsquantizationgraphneuralnetworkcircuitcomplexitymultiuserMIMO
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 tackles a design question: how many tunable interconnections a beyond-diagonal reconfigurable intelligent surface (BD-RIS) needs once real hardware effects—mutual coupling between elements, power losses in tunable components, and coarse phase quantization—are taken into account. It proposes a two-tier learning framework that generates candidate wiring topologies and optimizes their tunable component values, avoiding exhaustive search over an exponentially large architecture space. Against analytically known optimal architectures in ideal settings, the learned topologies match expected performance, validating the approach. The paper then reports three design insights: mutual coupling does not change the optimal architecture in multiuser MIMO, extra circuit complexity actively degrades lossy BD-RIS, and circuit complexity and quantization bits can compensate for each other. A sympathetic reader would care because these findings replace the usual assumption that fuller connectivity is always better with a concrete performance–complexity tradeoff for practical surfaces.

What carries the argument

The architecture characterization matrix A, a binary adjacency matrix whose off-diagonal entries mark which RIS elements are interconnected through tunable admittances, is the central object. The framework learns a probability vector over lower-triangular interconnections, selects the top K_cc entries to meet a circuit-complexity budget, and uses a straight-through estimator to keep the discrete selection differentiable. A graph neural network then treats each RIS element as a node and each interconnection as an edge, aggregates neighbor information across three graph-convolution layers, adds a residual branch to counter oversmoothing on dense graphs, and regresses the tunable component valu

What would settle it

For a small BD-RIS (e.g., 6 elements), exhaustively enumerate all reciprocal topologies at each circuit complexity under the paper's lossy and discrete-value models, optimize each exactly, and compare with LTTADF's chosen architectures; if any exhaustively better topology exists at a complexity level where the paper predicts decline or saturation, the central claim fails. Alternatively, rerun at different loss resistances R (e.g., R=0.1 and R=10) and check whether increasing complexity remains detrimental or the optimal complexity moves.

Watch

Extended reading notes

Core claim

The central discovery is that optimal BD-RIS architecture is not monotone in circuit complexity once non-idealities are modeled. For ideal surfaces the learned topologies reproduce the analytically optimal tree-connected (SU-SISO/SU-MISO) and band/stem-connected (MU-MIMO) architectures; under mutual coupling the same low-complexity MU-MIMO architecture remains optimal; but for lossy surfaces performance peaks at a moderate complexity and declines beyond it, while for discrete-value surfaces the number of quantization bits and the number of interconnections act as substitutes. The paper claims these conclusions are established by jointly training an architecture generator (which learns which

Load-bearing premise

The load-bearing premise is that the joint training drives the GNN performance optimizer to near-global optima for every candidate architecture, so the performance rankings behind the three conclusions reflect real architectural merit rather than where the optimizer happened to converge.

Editorial extensions

If this is right

  • Designers can pick band/stem-connected low-complexity BD-RIS for MU-MIMO without re-optimizing the topology when mutual coupling is present, since the optimal architecture is unchanged.
  • For lossy BD-RIS, there is an optimal circuit complexity: more tunable interconnections can reduce performance, so low-complexity learned topologies are both cheaper and better.
  • Quantization bits and circuit complexity are exchangeable resources: increasing one compensates for limits of the other, giving a two-dimensional cost frontier for discrete-value BD-RIS.
  • The framework reproduces analytically known optimal architectures where they exist, so the same learned search can be used to explore regimes where analysis is intractable.
  • The performance–complexity curves in ideal cases saturate, so the practical benefit of near-full connectivity is negligible once the analytically optimal complexity is reached.

Reading between the lines

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

  • The three headline insights are demonstrated at single operating points (loss resistance R=1, element spacing λ/2, 100 channel realizations). A natural extension would sweep R and spacing; the 'more complexity hurts' conclusion may sharpen or reverse outside that regime.
  • Because the architecture generator outputs a probability distribution over interconnections, the learned probabilities could be distilled into a closed-form wiring rule for non-ideal BD-RIS, just as tree/band/stem rules exist for ideal surfaces.
  • The discrete-value result suggests an information-theoretic view: topology edges and quantization bits are interchangeable degrees of freedom; the performance-complexity-bit tradeoff could be characterized as a rate-distortion frontier.
  • The framework's near-optimality claims are trust-based: without exhaustive search on small N_I or a convergence certificate, the 'optimal architectures' under non-idealities should be read as learned topologies whose ranking depends on the optimizer, not as proven global optima.
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

4 major / 5 minor

Summary. The paper proposes LTTADF, a learning-based two-tier framework (architecture generator + GNN-based performance optimizer) to discover BD-RIS architectures with a specified circuit complexity under four settings: ideal, mutual coupling, lossy, and discrete-value. The framework is validated on ideal SU-SISO/SU-MISO and MU-MIMO systems against analytically known tree-, band-, and stem-connected architectures. It is then used to make three physical claims: (i) mutual coupling does not change the optimal MU-MIMO architecture, (ii) for lossy BD-RIS, increasing circuit complexity can degrade performance, and (iii) circuit complexity and quantization resolution compensate for each other in discrete-value BD-RIS.

Significance. If the claims are correct, the paper provides a practical way to replace quadratic-complexity fully-connected BD-RIS with low-complexity learned topologies that preserve performance under non-idealities. The ideal-case validation in Figs. 3–5 is a genuine strength: the learned architectures reproduce known analytical optima at the same circuit complexity, which grounds the method's basic validity. The non-ideal insights are potentially valuable for both system design and hardware complexity trade-offs. However, those insights are not yet supported with the experimental rigor required to rule out optimizer or overfitting artifacts.

major comments (4)
  1. [Sec. III-C and Figs. 3–5] The ideal-case validation checks only the specific complexity points CC(Tree-conn.) and CC(Band/Stem-conn.). It does not validate the shape of the performance-complexity curve or the optimizer's convergence behavior at other complexities. The non-ideal conclusions (lossy decline, quantization/complexity substitution) are drawn from comparing learned performance at several K_cc values. Since the GNN optimizer is trained for each candidate architecture, a density-dependent optimization bias would directly alter those conclusions; the paper explicitly acknowledges the oversmoothing risk in Sec. III-B3. Please add multi-seed error bars, a brute-force or exhaustive search for small N_I, and a comparison with a conventional iterative optimizer for at least one non-ideal setting.
  2. [Sec. IV-C1, Fig. 7] The abstract and contribution list state that mutual coupling 'does not affect the optimal BD-RIS architecture in MU-MIMO systems.' This is presented as a general result, but the simulation uses only one inter-element distance (d=λ/2), one mutual-coupling realization procedure, and N=100 channel draws without a held-out split. The invariance claim requires at least a sweep over d (or over coupling strength) and a statistical evaluation over seeds. As written, the single operating point cannot support the general statement.
  3. [Sec. IV-C2, Figs. 8–9] The lossy-BD-RIS insight ('increasing circuit complexity can be detrimental') is demonstrated only for R=1 and for one set of channel realizations. R is a free parameter in the loss model and it controls the strength of the non-ideality. The performance-complexity concavity could be an artifact of the chosen operating point. Please sweep R (and, if feasible, L1/L2) and report the dispersion across random initializations and channel samples. This is load-bearing for one of the paper's three central claims.
  4. [Sec. III-C and Sec. IV-A] The joint learning loop optimizes the architecture generator and reports final performance on the same 100 channel realizations. No held-out split, cross-validation, or separate test set is described. Because architecture selection is driven by average objective on the training draws, the reported 'optimal architectures' and the rankings used for all non-ideal conclusions may be overfit to those draws. Please clarify the evaluation protocol and provide held-out results or a repeated random split analysis.
minor comments (5)
  1. [Sec. II-A] There is a typo 'the the transmitter'; the Introduction also spells 'mutiple' instead of 'multiple'.
  2. [Sec. IV-C2] 'assited' should be 'assisted'. In Eqs. (23)–(25), the notation eYi,j and eYij is confusing: it is unclear which symbol denotes the scalar admittance and which denotes the matrix element. Please introduce distinct notation, for example y_ij for the component value and Y_ij for the matrix entry.
  3. [Eq. (79)] The temperature τ is set to 0.1 but no sensitivity analysis is provided. Since τ controls the soft-quantization gradient, a brief ablation would increase confidence in the discrete-value results.
  4. [Figs. 8–11] Channel gain is plotted in dB, which visually compresses small gaps. The text should state whether the absolute differences are significant; consider reporting linear-scale values for the key comparison points.
  5. [Sec. IV-A] Please report the number of random initializations (or seeds) used for the learning runs and the variance across those runs, since the current text gives only 'N=100 channel realizations'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the non-ideal findings are numerical discoveries whose optimizer-convergence limitations are robustness concerns, not circular reductions.

full rationale

The paper's derivation chain is not circular. The LTTADF is validated against independent analytical optimality results [12], [14], [15] for ideal BD-RIS and mutual-coupling SU-SISO; these are mathematical results with stated assumptions and do not depend on any fitted values from this paper, so under the review rules they are independent support despite shared authorship. The non-ideal findings (MC invariance in MU-MIMO, lossy-complexity detriment, quantization/complexity compensation) are numerical observations obtained by evaluating the exact physical objectives (channel gain, sum rate) at the B or C matrices produced by the learned optimizer; no equation is defined in terms of the conclusion, and no fitted parameter is later relabeled as a prediction. The paper itself flags the main limitation: 'the LTTADF actually learns an approximation of the mapping between the input channel realizations and optimal solutions through a finite set of learnable parameters, leading to small approximation errors' (Sec. IV-B), and Sec. IV-A gives early stopping without a convergence certificate. These are correctness/robustness concerns about whether the optimizer ranks architectures correctly, not circular reductions; an honest non-finding is therefore appropriate.

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

The central claim rests on standard channel and BD-RIS models from the cited prior work and on the ML pipeline reaching near-global optima. Free parameters are mostly fixed inputs from prior papers ([15],[16]) plus the learned codebook and hyperparameters; they are disclosed, but the non-ideal conclusions are demonstrated at single operating points (R=1, d=λ/2, N_b∈{1,4}). Axioms 1–2 are externally sourced domain models; axioms 3–4 are the paper's own unverified assumptions about optimizer quality and data sufficiency.

free parameters (4)
  • Learnable quantization codebook B_Nb = not reported (learned during training)
    Used in discrete-value BD-RIS (Sec. III-B4, Eqs. 79–81); codebook values are fitted to data and directly shape the quantization–complexity tradeoff conclusion.
  • Soft-quantization temperature τ = 0.1
    Set by hand in Sec. IV-C3; controls the effective gradient of quantization and influences learned architecture quality.
  • Loss resistance R (lossy BD-RIS) = 1 ohm
    Fixed from [16] in Sec. IV-C2; the 'increasing complexity is detrimental' conclusion is demonstrated only for this value.
  • Mutual-coupling geometry: UPA inter-element spacing d = d = λ/2
    Fixed in Sec. IV-A (thin-wire dipoles, l=λ/4, f=28 GHz); the MC-invariance claim in MU-MIMO is only tested at this spacing.
assumptions (4)
  • domain assumption Lossy BD-RIS admittance model (Eqs. 23–25) with L1=6 nH, L2=0.7 nH, ĈĉĈµĈÅĈ·ij∈[0.35,3.20] pF
    Adopted from [16]; the loss–complexity tradeoff conclusions inherit the validity of this circuit model.
  • domain assumption Multiport network model for mutual coupling (Eqs. 10–16), assuming no coupling at Tx/user and Y_T=Y_R=Y_0 I
    From [15],[26]; the claim that MC does not change the optimal architecture is conditional on this modeling.
  • ad hoc to paper The GNN performance optimizer (3 GC layers + residual FC layers, Sec. III-B) reaches near-global optima of the non-convex objective for every candidate architecture
    Load-bearing for architecture ranking; no convergence or optimality certificate is provided (Sec. III-C, Sec. IV).
  • domain assumption Training on N=100 channel realizations (single set, no held-out split) suffices to characterize architecture performance
    All figures use N=100 realizations (Sec. IV-A); no generalization analysis is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond-Diagonal RIS Under Non-Idealities: Learning-Based Architecture Discovery and Optimization." pith.science (2026). https://pith.science/paper/QMN52HMX

@misc{pith2026251015701,
  author       = {Pith},
  title        = {Pith review of: Beyond-Diagonal RIS Under Non-Idealities: Learning-Based Architecture Discovery and Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMN52HMX}},
  note         = {Machine review of arXiv:2510.15701}
}
read the original abstract

Beyond-diagonal reconfigurable intelligent surface (BD-RIS) has recently been introduced to enable advanced control over electromagnetic waves to further increase the benefits of traditional RIS in enhancing signal quality and improving spectral and energy efficiency for next-generation wireless networks. A significant issue in designing and deploying BD-RIS is the tradeoff between its performance and circuit complexity. While existing studies have explored optimal architectures to minimize circuit complexity in ideal BD-RIS, architecture discovery for non-ideal BD-RIS remains uninvestigated. Consequently, how non-idealities and circuit complexity jointly affect the performance of BD-RIS remains unclear, making it difficult to achieve the performance-circuit complexity tradeoff in the presence of non-idealities. Essentially, architecture discovery for non-ideal BD-RIS faces challenges from both the computational complexity of global architecture search and the difficulty in achieving global optima. To tackle these challenges, we propose a learning-based two-tier architecture discovery framework (LTTADF) consisting of an architecture generator and a performance optimizer to jointly discover optimal architectures for non-ideal BD-RIS given specific circuit complexities, which can effectively explore over a large architecture space while avoiding getting trapped in poor local optima and thus achieving near-optimal solutions for the performance optimization. Numerical results provide valuable insights for deploying non-ideal BD-RIS considering the performance-circuit complexity tradeoff.

Figures

Figures reproduced from arXiv: 2510.15701 by the authors.

Figure 1
Figure 1. The BD-RIS aided MU-MIMO system. A. System Model We consider a BD-RIS aided MU-MIMO system. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The learning-based two-tier architecture discovery framework. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Channel gain versus circuit complexity of ideal BD-RIS in an SU [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Channel gain versus circuit complexity of ideal BD-RIS in an SU [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Sum rate versus circuit complexity of ideal BD-RIS in an MU-MIMO [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Channel gain versus circuit complexity of BD-RIS with mutual [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Sum rate versus circuit complexity of BD-RIS with mutual coupling [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 9
Figure 9. Figure 9: Sum channel gain versus circuit complexity of lossy BD-RIS with [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 12
Figure 12. Figure 12: Sum rate versus circuit complexity of discrete-value BD-RIS with [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Sum rate versus circuit complexity of discrete-value BD-RIS with [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 2 linked inside Pith

  1. [1]

    Reconfigurable intelligent surfaces for energy efficiency in wireless communication,

    C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in wireless communication,”IEEE Trans. Wireless Commun., vol. 18, no. 8, pp. 4157–4170, Aug. 2019

  2. [2]

    Intelligent reflecting surface enhanced wireless network via joint active and passive beamforming,

    Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless network via joint active and passive beamforming,”IEEE Trans. Wire- less Commun., vol. 18, no. 11, pp. 5394–5409, Nov. 2019

  3. [3]

    Modeling and architecture design of reconfigurable intelligent surfaces using scattering parameter network analysis,

    S. Shen, B. Clerckx, and R. Murch, “Modeling and architecture design of reconfigurable intelligent surfaces using scattering parameter network analysis,”IEEE Trans. Wireless Commun., vol. 21, no. 2, pp. 1229–1243, Feb. 2022

  4. [4]

    Reconfigurable intelligent surfaces 2.0: Beyond diagonal phase shift matrices,

    H. Li, S. Shen, M. Nerini, and B. Clerckx, “Reconfigurable intelligent surfaces 2.0: Beyond diagonal phase shift matrices,”IEEE Commun. Mag., vol. 62, no. 3, pp. 102–108, Mar. 2024

  5. [5]

    A tutorial on beyond- diagonal reconfigurable intelligent surfaces: Modeling, architectures, system design and optimization, and applications,

    H. Li, M. Nerini, S. Shen, and B. Clerckx, “A tutorial on beyond- diagonal reconfigurable intelligent surfaces: Modeling, architectures, system design and optimization, and applications,”arXiv preprint arXiv:2505.16504, May 2025

  6. [6]

    Reconfigurable intelligent surfaces relying on non-diagonal phase shift matrices,

    Q. Li, M. El-Hajjar, I. Hemadeh, A. Shojaeifard, A. A. M. Mourad, B. Clerckx, and L. Hanzo, “Reconfigurable intelligent surfaces relying on non-diagonal phase shift matrices,”IEEE Trans. V eh. Technol., vol. 71, no. 6, pp. 6367–6383, Jun. 2022

  7. [7]

    Beyond diagonal reconfigurable intelli- gent surfaces: From transmitting and reflecting modes to single-, group- , and fully-connected architectures,

    H. Li, S. Shen, and B. Clerckx, “Beyond diagonal reconfigurable intelli- gent surfaces: From transmitting and reflecting modes to single-, group- , and fully-connected architectures,”IEEE Trans. Wireless Commun., vol. 22, no. 4, pp. 2311–2324, Apr. 2023

  8. [8]

    Non-reciprocal beyond diagonal ris: Sum-rate maximization in full-duplex communications,

    Z. Liu, H. Li, and B. Clerckx, “Non-reciprocal beyond diagonal ris: Sum-rate maximization in full-duplex communications,”arXiv preprint arXiv:2411.18523, Dec. 2024

Show all 31 references
  1. [9]

    Closed-form global optimization of beyond diagonal reconfigurable intelligent surfaces,

    M. Nerini, S. Shen, and B. Clerckx, “Closed-form global optimization of beyond diagonal reconfigurable intelligent surfaces,”IEEE Trans. Wireless Commun., vol. 23, no. 2, pp. 1037–1051, Feb. 2024

  2. [10]

    Localized and distributed beyond diagonal reconfigurable intelligent surfaces with lossy intercon- nections: Modeling and optimization,

    M. Nerini, G. Ghiaasi, and B. Clerckx, “Localized and distributed beyond diagonal reconfigurable intelligent surfaces with lossy intercon- nections: Modeling and optimization,”IEEE Trans. Commun., vol. 73, no. 9, pp. 8140–8154, Sep. 2025

  3. [11]

    Pareto frontier for the performance- complexity trade-off in beyond diagonal reconfigurable intelligent sur- faces,

    M. Nerini and B. Clerckx, “Pareto frontier for the performance- complexity trade-off in beyond diagonal reconfigurable intelligent sur- faces,”IEEE Commun. Lett., vol. 27, no. 10, pp. 2842–2846, Oct. 2023

  4. [12]

    Beyond diagonal reconfig- urable intelligent surfaces utilizing graph theory: Modeling, architecture design, and optimization,

    M. Nerini, S. Shen, H. Li, and B. Clerckx, “Beyond diagonal reconfig- urable intelligent surfaces utilizing graph theory: Modeling, architecture design, and optimization,”IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 9972–9985, Aug. 2024

  5. [13]

    A novel q-stem connected architecture for beyond-diagonal reconfigurable intelligent surfaces,

    X. Zhou, T. Fang, and Y . Mao, “A novel q-stem connected architecture for beyond-diagonal reconfigurable intelligent surfaces,” inProc. ICC, Jun. 2025, pp. 6880–6885

  6. [14]

    Beyond-diagonal RIS in multiuser MIMO: Graph theoretic modeling and optimal architectures with low complex- ity,

    Z. Wu and B. Clerckx, “Beyond-diagonal RIS in multiuser MIMO: Graph theoretic modeling and optimal architectures with low complex- ity,”IEEE Trans. Inf. Theory, pp. 1–19, Sep. 2025

  7. [15]

    Global optimal closed-form solu- tions for intelligent surfaces with mutual coupling: Is mutual coupling detrimental or beneficial?

    M. Nerini, H. Li, and B. Clerckx, “Global optimal closed-form solu- tions for intelligent surfaces with mutual coupling: Is mutual coupling detrimental or beneficial?”IEEE Trans. Wireless Commun., pp. 1–15, 2025

  8. [16]

    Lossy beyond diagonal reconfigurable intelligent surfaces: Modeling and optimization,

    Y . Peng, H. Li, Z. Wu, and B. Clerckx, “Lossy beyond diagonal reconfigurable intelligent surfaces: Modeling and optimization,”arXiv preprint arXiv:2504.19744, Apr. 2025

  9. [17]

    Discrete-value group and fully connected architectures for beyond diagonal reconfigurable intelligent surfaces,

    M. Nerini, S. Shen, and B. Clerckx, “Discrete-value group and fully connected architectures for beyond diagonal reconfigurable intelligent surfaces,”IEEE Trans. V eh. Technol., vol. 72, no. 12, pp. 16 354–16 368, Dec. 2023

  10. [18]

    A multi- objective optimization approach accurately resolves protein domain architectures,

    J. Bernardes, F. Vieira, G. Zaverucha, and A. Carbone, “A multi- objective optimization approach accurately resolves protein domain architectures,”Bioinf., vol. 32, no. 3, pp. 345–353, Feb. 2016

  11. [19]

    Rethinking drug design in the artificial intelligence era,

    P. Schneider, W. P. Walters, A. T. Plowright, N. Sieroka, J. Listgarten, R. A. Goodnow, J. Fisher, J. M. Jansen, J. S. Duca, T. S. Rush, M. Zentgraf, J. E. Hill, E. Krutoholow, M. Kohler, J. Blaney, K. Funatsu, C. Luebkemann, and G. Schneider, “Rethinking drug design in the ar...

  12. [20]

    Pay less but get more: A dual-attention-based channel estimation network for massive MIMO systems with low-density pilots,

    B. Zhou, X. Yang, S. Ma, F. Gao, and G. Yang, “Pay less but get more: A dual-attention-based channel estimation network for massive MIMO systems with low-density pilots,”IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 6061–6076, Jun. 2024

  13. [21]

    Low-overhead channel estimation via 3D extrapolation for TDD mmWave massive MIMO systems under high-mobility scenarios,

    ——, “Low-overhead channel estimation via 3D extrapolation for TDD mmWave massive MIMO systems under high-mobility scenarios,”IEEE Trans. Wireless Commun., vol. 24, no. 4, pp. 2797–2813, Apr. 2025

  14. [22]

    Out-of-band modality synergy based multi-user beam prediction and proactive BS selection with zero pilot overhead,

    K. Li, B. Zhou, J. Guo, F. Gao, G. Yang, and S. Ma, “Out-of-band modality synergy based multi-user beam prediction and proactive BS selection with zero pilot overhead,”arXiv preprint arXiv:2506.15136, Jun. 2025

  15. [23]

    Unsupervised learning- based joint active and passive beamforming design for reconfigurable intelligent surfaces aided wireless networks,

    H. Song, M. Zhang, J. Gao, and C. Zhong, “Unsupervised learning- based joint active and passive beamforming design for reconfigurable intelligent surfaces aided wireless networks,”IEEE Commun. Lett., vol. 25, no. 3, pp. 892–896, Mar. 2021

  16. [24]

    Joint BS and beyond diagonal RIS beamforming design with DRL methods for mmWave 6G mobile communications,

    S. Sobhi-Givi, M. Nouri, H. Behroozi, and Z. Ding, “Joint BS and beyond diagonal RIS beamforming design with DRL methods for mmWave 6G mobile communications,” inProc. WCNC, Apr. 2024, pp. 1–6

  17. [25]

    D. M. Pozar,Microwave Engineering, 4th ed. Hoboken, NJ, USA: John Wiley & Sons, 2011

  18. [26]

    A universal framework for multiport network analysis of reconfigurable intelligent surfaces,

    M. Nerini, S. Shen, H. Li, M. Di Renzo, and B. Clerckx, “A universal framework for multiport network analysis of reconfigurable intelligent surfaces,”IEEE Trans. Wireless Commun., vol. 23, no. 10, pp. 14 575– 14 590, Oct. 2024

  19. [27]

    Ma and J

    Y . Ma and J. Tang,Deep Learning on Graphs. Cambridge, UK: Cambridge University Press, 2021

  20. [28]

    Optnet: Differentiable optimization as a layer in neural networks,

    B. Amos and J. Z. Kolter, “Optnet: Differentiable optimization as a layer in neural networks,” inProc. ICML, Aug. 2017, pp. 136–145

  21. [29]

    Learning to optimize,

    K. Li and J. Malik, “Learning to optimize,”arXiv preprint arXiv:1606.01885, Jun. 2016

  22. [30]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,”arXiv preprint arXiv:1609.02907, Sep. 2016

  23. [31]

    Deeper insights into graph convolutional networks for semi-supervised learning,

    Q. Li, Z. Han, and X.-m. Wu, “Deeper insights into graph convolutional networks for semi-supervised learning,”Proc. AAAI, vol. 32, no. 1, Apr. 2018

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.