Pith. sign in

REVIEW 5 major objections 7 minor 46 references

Robust Hybrid Precoding for Millimeter Wave MU-MISO System Via Meta-Learning

T0 review · 5 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A meta-learning loop can optimize hybrid precoders without pre-training, beating fully digital WMMSE.

desk verdict A plausible learning-aided hybrid precoding method, but the headline WMMSE comparison and 'free-of-training' framing are both overclaimed. read the letter →

arxiv 2411.15762 v1 pith:ZQFSCXRO submitted 2024-11-24 cs.IT eess.SPmath.IT

Yifan Guo This is my paper · ORCID
classification cs.ITeess.SPmath.IT
keywords hybridprecodingmillimeterwaveMU-MISOmeta-learningKolmogorov-ArnoldnetworksspectralefficiencyimperfectCSIalternatingoptimization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that hybrid analog-digital precoding in millimeter-wave multi-user systems can be solved by a meta-learning loop that needs no pre-training: feed the raw gradient of the spectral-efficiency objective into a small Kolmogorov-Arnold network, use its output as the next update of the digital or analog precoder, and update the network parameters from a global loss. The claim is that this alternating scheme out-achieves classical optimization baselines and data-driven CNN precoders, converges in about 200 iterations, and can surpass fully digital WMMSE once the number of RF chains exceeds six. A robust extension handles imperfect channel knowledge by converting the worst-case max-min problem into a geometric program over an effective channel. If correct, the framework offers a plug-and-play, low-complexity alternative to training-based deep learning for hybrid beamforming.

What carries the argument

The central mechanism is the gradient-guided meta-learning loop with two KAN subnetworks, the digital precoding network and the analog precoding network. In each outer iteration, the gradient of the loss with respect to the current digital precoder is fed to one KAN to produce an update, $\mathbf{D} = \mathbf{D}_0 + \operatorname{DPN}\big(\nabla_{\mathbf{D}_0}\mathcal{L}\big)$, followed by power rescaling; likewise, the analog precoder is updated as $\mathbf{F} = \mathcal{P}_{\mathcal{F}}\big(\mathbf{F}_0 + \operatorname{APN}\big(\nabla_{\mathbf{F}_0}\mathcal{L}\big)\big)$, where $\mathcal{P}_{\mathcal{F}}$ projects onto unit-modulus entries. The KANs use learnable B-spline activation functions, which the paper argues capture subtle structure in the gradient flow. A global perspective is imposed by updating the KAN parameters with an adaptive-moment optimizer on the loss $\mathcal{L} = -R + \beta\,\mathrm{Var}(R)$, coupling the two subproblems and guiding the iterations away from local optima. For imperfect CSI, Lemma 1 bounds the uncertain channel Gram matrix and Proposition 2 shows that the worst-case MSE is attained at that bound, so the robust problem reduces to the same GGML loop with an effective channel.

What would settle it

Run GGML on a larger or ill-conditioned system than those tested, such as N = 512 or K = 16 with near-degenerate scattering clusters, and check whether a fresh random seed still reaches the claimed spectral efficiency within 500 iterations; if some seeds stall below the manifold-optimization baseline or diverge, the free-of-training claim fails.

Watch

Extended reading notes

Core claim

The paper claims that the non-convex hybrid precoding problem can be solved by gradient-guided meta-learning (GGML): instead of learning a mapping from channels to precoding matrices, two KAN-based networks consume the objective gradient with respect to the current precoder and output update directions, while the network parameters themselves are updated online from a global loss. Digital and analog precoders are optimized alternately, with the analog update projected onto the constant-modulus phase-shifter constraint and the digital update rescaled to full transmit power. The authors report that this approach attains higher average spectral efficiency than manifold optimization, majorization-minimization, and CNN baselines, converges within about 200 iterations, runs about eight times faster than manifold optimization at 256 antennas, and can exceed the spectral efficiency of fully digital WMMSE when the number of RF chains is larger than six. The same framework is extended to imperfect CSI by replacing the estimated channel with a worst-case effective channel, turning the maximin problem into a geometric program over worst-case mean-square error.

Load-bearing premise

The method assumes that a randomly initialized small network, trained online for a fixed 500 iterations on a single channel realization, will always produce productive update directions and converge to a good precoder, for any channel and any system size.

Editorial extensions

If this is right

  • Hybrid precoders can be obtained with no offline training, since the KAN networks are freshly initialized and updated online for the given channel instance.
  • The per-iteration complexity scales as $O(LNMK^2)$, giving roughly eight times less CPU time than manifold optimization at 256 antennas in the reported simulations.
  • With more than six RF chains, the proposed hybrid precoder reportedly exceeds the spectral efficiency of a fully digital WMMSE system using the same antenna array.
  • The imperfect-CSI variant converts a worst-case max-min problem into geometric-program form and retains most of the spectral efficiency at channel-error levels up to 30 percent of the channel gain.
  • Because the loop only depends on gradients, projections, and power normalization, the framework is stated to generalize to MU-MIMO and other hybrid architectures.

Reading between the lines

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

  • Beyond the paper's setting, the same gradient-fed alternating loop should apply to other non-convex constrained optimization problems whose variables admit projection operators, such as constant-modulus, quantized, or unit-norm beamforming.
  • Because the KAN parameters are trained online per channel, the method's robustness rests on the optimizer rather than on transferable features; a natural extension is warm-starting the KAN parameters from a previous channel to accelerate adaptation in time-varying channels.
  • The reported ability to beat fully digital WMMSE suggests that the learned update path acts as an implicit regularizer adapted to the phase-shifter constraint; verifying whether the gap persists on measured or ray-traced channels would strengthen the deployment claim.
  • A testable sensitivity question is how the fixed budget of 500 iterations and random network initialization interact: measuring seed-to-seed variance in spectral efficiency would show how much of the gain is structural versus initialization luck.
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 / 7 minor

Summary. The paper proposes GGML, a gradient-guided meta-learning framework for hybrid analog-digital precoding in millimeter-wave MU-MISO systems. The digital and analog precoders are updated alternately using neural networks (KAN-based) that take the gradient of the objective as input, while the network parameters are updated globally with a loss that includes a variance penalty. The framework is extended to imperfect CSI by transforming a worst-case maximin problem into a geometric-program-like MMSE minimization. Simulation results claim that GGML outperforms manifold optimization, majorization-minimization, and CNN baselines, converges faster, and even surpasses fully digital WMMSE precoding when the number of RF chains exceeds 6.

Significance. If the claims were substantiated, the approach would offer a low-complexity, training-free hybrid precoding method that also handles imperfect CSI, which is of practical interest for mmWave systems. The paper provides a worst-case reformulation with a proof, a complexity analysis, and a simulation code link, which are strengths. However, the headline result that a hybrid scheme outperforms fully digital WMMSE is mathematically questionable, the loss function is not fully specified, and the statistical evidence lacks error bars and sensitivity analysis. The contribution is potentially useful but requires substantial revision and re-validation before the claims can be accepted.

major comments (5)
  1. [Section V-D, Fig. 6c] The claim that GGML outperforms fully digital WMMSE with the same number of antennas is inconsistent with a basic feasibility argument. For any hybrid precoder (F,D) satisfying the power constraint, the effective linear precoder W = F D in C^{N x K} is also feasible for a fully digital system with N antennas and N RF chains. Therefore the optimal value of the fully digital weighted sum-rate problem is an upper bound on the hybrid optimum. A correctly implemented and converged WMMSE baseline (refs. [15] and [42]) cannot be outperformed. The paper does not report the RF-chain count, initialization, convergence criterion, or power normalization used for the WMMSE baseline. This headline comparison must be corrected or removed; as it stands, the most striking result of the paper is unsupported.
  2. [Section III-B-3 and IV-B, Eqs. (19) and (34)] The loss function L = -R + beta * Var(R) is not well-defined because R in (5) is a scalar (the weighted sum rate), so Var(R) has no meaning unless it is defined as the variance of per-user rates R_k or over some other ensemble; no such definition is given. Moreover, the penalty coefficient beta is never specified in Section V, and no sensitivity analysis with respect to beta is reported. Since this loss directly drives the global update of the network parameters in Algorithm 1 (line 8) and Algorithm 2 (line 9), this ambiguity is load-bearing for reproducibility and for understanding the algorithm's behavior.
  3. [Algorithms 1 and 2] The claim that GGML is 'free-of-training' and 'plug-and-play' (abstract and Section I-B) is contradicted by the algorithm itself. Lines 8-10 of Algorithm 1 and lines 9-11 of Algorithm 2 update the network parameters theta_D and theta_F using Adam on the same channel instance that is later used for evaluation. There is no pre-training phase, no task distribution, and no separate adaptation set. At best, this is per-instance online adjustment of a randomly initialized network; it is not meta-learning in the standard sense of learning across a distribution of tasks, and it requires online computation during deployment. The terminology should be corrected and the computational cost of these online updates should be included in the complexity claims.
  4. [Section V-A and Figs. 4-8] The numerical results are averaged over 102 channel realizations, but no error bars, confidence intervals, or sensitivity analyses are reported. Some reported gains are small (e.g., 1.4% over MO at convergence in Section V-B), so without a measure of dispersion it is unclear whether the improvements are statistically meaningful. In addition, the KAN architecture (number of layers, width, grid size, spline order) and the Monte Carlo outage bound epsilon_k used for Fig. 8 are not disclosed. The paper provides a code link, which is good, but the manuscript itself should specify these essential simulation details to allow independent reproduction.
  5. [Section IV-A and V-A] The worst-case robust formulation (P2) assumes a bounded error set with sigma_max(e_k) <= epsilon_k, but the simulation model (Section V-A) draws e_k from an unbounded Gaussian distribution. The Monte Carlo outage construction only bounds the error with 5% probability, so the guarantee is probabilistic rather than the worst-case guarantee implied by the formulation. The paper should clarify whether the robust design is meant to satisfy the bound only with high probability, and it should report the outage level and the resulting epsilon_k values used in the figures.
minor comments (7)
  1. [Section III-B-2, Eq. (15)] The notation DPN is used both for the network and for its output; the text should clarify that DPN(.) denotes the neural network function, not the resulting precoding matrix.
  2. [Section II-B-2 and IV-A] The symbol e_k is used for both the channel estimation error in Eq. (8) and the MMSE error in Eq. (22); this is confusing and should be resolved by renaming one of them.
  3. [Section III-B-3, Eq. (19)] The parameter beta is described as a 'penalty rate'; it should be called a penalty coefficient, and its value in the simulations should be stated.
  4. [Section V-D, Fig. 6c] The caption should state the number of antennas N and the exact configuration of the WMMSE baseline (number of RF chains, iterations, initialization) so that the comparison is interpretable.
  5. [Section VI, Conclusion] The sentence 'only requires deploying lightweight neural networks at the user end' appears to be a typo; the proposed algorithm runs at the base station, not at the user equipment, and should be corrected.
  6. [Section V-B] The text mentions both '102 random initializations' and '102 independent channel realizations'; please clarify which quantity is averaged over in each figure.
  7. [Table I] The GGML complexity O(L N M K^2) does not include the complexity of the KAN forward and backward passes, which depends on the network width and number of layers; this should be acknowledged.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported spectral-efficiency gains are the online training objective evaluated on the same channel realizations used for training, so the 'free-of-training' meta-learning claim reduces to per-instance fitting; the WMMSE comparison is additionally infeasible as stated.

  1. fitted input called prediction [Algorithm 1 (lines 5-8), Eq. (19), Section V-A and V-D]
    "Calculate the L as (19); Update θD and θF according to (20) and (21); ... L = −R + β · Var(R) ... results averaged over 102 independent channel realizations."

    The loss L in Eq. (19) is the negative weighted sum spectral efficiency R plus a variance penalty, and R is exactly the average spectral efficiency reported in Section V. Algorithm 1 updates the network parameters θD and θF with Adam on the same channel realizations hk that are later used to compute the reported average SE, with no held-out set or task distribution. Therefore the claim that GGML 'significantly enhance[s] spectral efficiency' is not a prediction from a trained or free-of-training model but the training objective on the evaluation instances. The 'free-of-training' label is contradicted by Algorithm 1's per-instance parameter updates, and the reported gains reduce by construction to per-instance fitting of a KAN-augmented optimizer.

full rationale

The main circularity is the online training/evaluation overlap: Eq. (19) defines the training loss as -R + β·Var(R), Algorithm 1 updates θD and θF on each channel instance, and Section V-D reports R averaged over those same instances. Thus the headline performance numbers are the optimized training objective, not a held-out prediction; this is a fitted-input-called-prediction pattern. The comparison to MO, MM, and CNN remains meaningful as an optimization-algorithm benchmark, so the circularity is partial and mainly concerns the 'meta-learning' and 'free-of-training' framing. A separate, non-circular correctness risk is the claim in Section V-D that 'GGML even outperforms WMMSE precoding (simulated as a fully digital system with the same number of antennas)': since any hybrid precoder F·D is itself a feasible fully digital precoder, a correctly converged fully digital WMMSE cannot be worse; the absence of WMMSE initialization, convergence criterion, and RF-chain count makes the comparison unverifiable. This does not raise the circularity score but is a serious confidence risk. No load-bearing self-citations or imported uniqueness arguments were found; Proposition 1's citation to [35] and [42] is external and not authored by Yifan Guo, and the KAN architecture is cited to independent prior work.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The framework introduces no new physical entities. The free parameters are dominated by undisclosed hyperparameters (β, KAN architecture) and design choices (outage threshold, learning rates, iteration count). The most questionable axiom is the ad hoc assumption that the channel norm dominates the error norm, which is needed for the worst-case reformulation to hold.

free parameters (5)
  • Variance penalty coefficient β = Not specified
    Appears in the loss functions (19) and (34) as β·Var(R). The paper never states the value used in simulations, yet it directly affects the optimization and the reported spectral efficiency.
  • KAN architecture hyperparameters = Not specified
    The number of KAN layers, hidden widths, spline grid size, and B-spline order are not reported, despite being central to the method's performance.
  • Monte Carlo outage threshold and realizations = 5% outage, 10^5 realizations
    Used to set the channel error bound ε_k in the imperfect CSI framework (Section IV-B). This is a design choice that affects robustness and performance.
  • Learning rates αD and αF = αD = 1e-3, αF = 1.5e-3
    These are given in Section V-A, but they are tuned hyperparameters with no sensitivity analysis.
  • Number of iterations L = 500
    The algorithm runs for 500 iterations; convergence is claimed within 200 epochs. This fixed budget may not hold for other system sizes.
assumptions (6)
  • standard math KKT conditions and complementary slackness imply the full power property (Proposition 1).
    The proof is referenced to [35] and [42], relying on standard convex optimization results.
  • standard math Matrix calculus identities from [46], including ∇X^{-1} = -X^{-1}(∇X)X^{-1}.
    Used in the proof of Proposition 2 in Appendix A.
  • standard math Lemma 1 (adapted from Lemma 7.1 in [45]): XY^H + YX^H ≤ 2 σmax(X) σmax(Y) I_n.
    Invoked to bound the channel error term and derive the worst-case lower bound R_hk_WC in (30).
  • ad hoc to paper The channel error norm is bounded by the spectral norm, and σmax(ˆhk) ≥ ε_k.
    Assumed before (28) to ensure the worst-case bound is valid after truncation. This is a modeling assumption on the relative sizes of channel and error, not guaranteed in practice.
  • domain assumption Saleh-Valenzuela channel model with Nc clusters and Nray rays.
    Equation (2) models the mmWave channel; this is a standard assumption in the hybrid precoding literature.
  • domain assumption The channel is quasi-static and perfectly known in the perfect CSI problem P1.
    Standard assumption for precoding design, though in practice CSI is imperfect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Hybrid Precoding for Millimeter Wave MU-MISO System Via Meta-Learning." pith.science (2026). https://pith.science/paper/ZQFSCXRO

@misc{pith2026241115762,
  author       = {Pith},
  title        = {Pith review of: Robust Hybrid Precoding for Millimeter Wave MU-MISO System Via Meta-Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZQFSCXRO}},
  note         = {Machine review of arXiv:2411.15762}
}
read the original abstract

Thanks to the low cost and power consumption, hybrid analog-digital architectures are considered as a promising energy-efficient solution for massive multiple-input multiple-output (MIMO) systems. The key idea is to connect one RF chain to multiple antennas through low-cost phase shifters. However, due to the non-convex objective function and constraints, we propose a gradient-guided meta-learning (GGML) based alternating optimization framework to solve this challenging problem. The GGML based hybrid precoding framework is \textit{free-of-training} and \textit{plug-and-play}. Specifically, GGML feeds the raw gradient information into a neural network, leveraging gradient descent to alternately optimize sub-problems from a local perspective, while a lightweight neural network embedded within the meta-learning framework is updated from a global perspective. We also extend the proposed framework to include precoding with imperfect channel state information. Simulation results demonstrate that GGML can significantly enhance spectral efficiency, and speed up the convergence by 8 times faster compared to traditional approaches. Moreover, GGML could even outperform fully digital weighted minimum mean square error (WMMSE) precoding with the same number of antennas.

Figures

Figures reproduced from arXiv: 2411.15762 by the authors.

Figure 1
Figure 1. The architecture of hybrid precoding. II. SYSTEM MODEL AND PROBLEM FORMULATION In this section, we begin by presenting an overview of the MU-MISO system model with a hybrid structure. Subse￾quently, we formulate the hybrid precoding problem for both perfect and imperfect CSI scenarios. A. System Model As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. A block diagram of the proposed meta-learning algori [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Schematic diagram of the proposed GGML-ImCSI for rob [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Average spectral efficiency versus the iterations. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Average CPU time versus the BS antenna. framework can converge within 200 epochs. To be more specific, the performance of GGML steadily improves with increasing iterations, surpassing traditional methods at epoch 86. The essential reason is that, global optimization st…
Figure 6
Figure 6. Figure 6: The performance with different system parameters, ( [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Average spectral efficiency versus the epochs, with a [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Average spectral efficiency performance with imperfe [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 43 canonical work pages

  1. [35]

    Robust beamforming for ris - aided communications: Gradient-based manifold meta learn ing,

    F. Zhu, X. Wang, C. Huang, Z. Y ang, X. Chen, A. Alhammadi, Z. Zhang, C. Y uen, and M. Debbah, “Robust beamforming for ris - aided communications: Gradient-based manifold meta learn ing,” IEEE Transactions on Wireless Communications , pp. 1–1, 2024

    work page 2024

  2. [37]

    A lea rning-aided flexible gradient descent approach to miso beamforming,

    Z. Y ang, J.-Y . Xia, J. Luo, S. Zhang, and D. Gündüz, “A lea rning-aided flexible gradient descent approach to miso beamforming,” IEEE Wireless Communications Letters , vol. 11, no. 9, pp. 1895–1899, 2022

    work page 2022

  3. [15]

    An Iterativ ely Weighted MMSE Approach to Distributed Sum-Utility Maximization for a MIMO Interfering Broadcast Channel,

    Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, “An Iterativ ely Weighted MMSE Approach to Distributed Sum-Utility Maximization for a MIMO Interfering Broadcast Channel,” IEEE Trans. Signal Process. , vol. 59, no. 9, pp. 4331–4340, Sept. 2011

    work page 2011

  4. [42]

    Rethinking wmmse: Can its complexity scale linearly with the number of bs antennas ?

    X. Zhao, S. Lu, Q. Shi, and Z.-Q. Luo, “Rethinking wmmse: Can its complexity scale linearly with the number of bs antennas ?” IEEE Transactions on Signal Processing , vol. 71, pp. 433–446, 2023

    work page 2023

  5. [1]

    Multi-gigabit mi llimeter wave wireless communications for 5g: from fixed access to cellula r networks,

    P . Wang, Y . Li, L. Song, and B. Vucetic, “Multi-gigabit mi llimeter wave wireless communications for 5g: from fixed access to cellula r networks,” IEEE Communications Magazine , vol. 53, no. 1, pp. 168–178, 2015

    work page 2015

  6. [2]

    An overview of signal processing techniques for milli meter wave mimo systems,

    R. W. Heath, N. González-Prelcic, S. Rangan, W. Roh, and A . M. Say- eed, “An overview of signal processing techniques for milli meter wave mimo systems,” IEEE Journal of Selected Topics in Signal Processing , vol. 10, no. 3, pp. 436–453, 2016

    work page 2016

  7. [3]

    Millimeter-wave beamforming as an enabling t echnology for 5g cellular communications: theoretical feasibility a nd prototype results,

    W. Roh, J.-Y . Seol, J. Park, B. Lee, J. Lee, Y . Kim, J. Cho, K . Cheun, and F. Aryanfar, “Millimeter-wave beamforming as an enabling t echnology for 5g cellular communications: theoretical feasibility a nd prototype results,” IEEE Communications Magazine , vol. 52, no. 2, pp. 106–113, 2014

    work page 2014

  8. [4]

    What will 5g be?

    J. G. Andrews, S. Buzzi, W. Choi, S. V . Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang, “What will 5g be?” IEEE Journal on Selected Areas in Communications , vol. 32, no. 6, pp. 1065–1082, 2014

    work page 2014

Show all 46 references
  1. [5]

    Hybrid beamforming fo r 5g and beyond millimeter-wave systems: A holistic view,

    J. Zhang, X. Y u, and K. B. Letaief, “Hybrid beamforming fo r 5g and beyond millimeter-wave systems: A holistic view,” IEEE Open Journal of the Communications Society , vol. 1, pp. 77–91, 2020

    2020

  2. [6]

    Hybrid beamforming for massive mimo: A survey,

    A. F. Molisch, V . V . Ratnam, S. Han, Z. Li, S. L. H. Nguyen, L . Li, and K. Haneda, “Hybrid beamforming for massive mimo: A survey,” IEEE Communications Magazine, vol. 55, no. 9, pp. 134–141, 2017

    2017

  3. [7]

    Spatially sparse precoding in millimeter wave mimo system s,

    O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. He ath, “Spatially sparse precoding in millimeter wave mimo system s,” IEEE Transactions on Wireless Communications , vol. 13, no. 3, pp. 1499– 1513, 2014

    2014

  4. [8]

    Hybrid digital and analog beamform ing design for large-scale antenna arrays,

    F. Sohrabi and W. Y u, “Hybrid digital and analog beamform ing design for large-scale antenna arrays,” IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 501–513, 2016

    2016

  5. [9]

    Alternati ng minimization algorithms for hybrid precoding in millimeter wave mimo sys tems,

    X. Y u, J.-C. Shen, J. Zhang, and K. B. Letaief, “Alternati ng minimization algorithms for hybrid precoding in millimeter wave mimo sys tems,” IEEE Journal of Selected Topics in Signal Processing , vol. 10, no. 3, pp. 485–500, 2016

    2016

  6. [10]

    Frequency selective hybr id precoding for limited feedback millimeter wave systems,

    A. Alkhateeb and R. W. Heath, “Frequency selective hybr id precoding for limited feedback millimeter wave systems,” IEEE Transactions on Communications, vol. 64, no. 5, pp. 1801–1818, 2016

    2016

  7. [11]

    Hyb rid beamforming for millimeter wave systems using the mmse crit erion,

    T. Lin, J. Cong, Y . Zhu, J. Zhang, and K. Ben Letaief, “Hyb rid beamforming for millimeter wave systems using the mmse crit erion,” IEEE Transactions on Communications , vol. 67, no. 5, pp. 3693–3708, 2019

    2019

  8. [12]

    Hybrid mmse precodin g for millimeter wave mu-miso via trace maximization,

    P . R. Gautam, L. Zhang, and P . Fan, “Hybrid mmse precodin g for millimeter wave mu-miso via trace maximization,” IEEE Transactions on Wireless Communications , vol. 23, no. 3, pp. 1999–2010, 2024

    1999

  9. [13]

    Low-complexity hybrid pre coding in massive multiuser mimo systems,

    L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid pre coding in massive multiuser mimo systems,” IEEE Wireless Communications Letters, vol. 3, no. 6, pp. 653–656, 2014

    2014

  10. [14]

    Joint transceiver design with antenna selection for large -scale mu-mimo mmwave systems,

    X. Zhai, Y . Cai, Q. Shi, M. Zhao, G. Y . Li, and B. Champagne , “Joint transceiver design with antenna selection for large -scale mu-mimo mmwave systems,” IEEE Journal on Selected Areas in Communications , vol. 35, no. 9, pp. 2085–2096, 2017

    2017

  11. [16]

    Hyb rid mmse precoding and combining designs for mmwave multiuser syste ms,

    D. H. N. Nguyen, L. B. Le, T. Le-Ngoc, and R. W. Heath, “Hyb rid mmse precoding and combining designs for mmwave multiuser syste ms,” IEEE Access, vol. 5, pp. 19 167–19 181, 2017

    2017

  12. [17]

    Deep learning-bas ed channel estimation for beamspace mmwave massive mimo systems,

    H. He, C.-K. Wen, S. Jin, and G. Y . Li, “Deep learning-bas ed channel estimation for beamspace mmwave massive mimo systems,” IEEE Wireless Communications Letters , vol. 7, no. 5, pp. 852–855, 2018

    2018

  13. [18]

    Sparse channel esti mation and hybrid precoding using deep learning for millimeter wav e massive mimo,

    W. Ma, C. Qi, Z. Zhang, and J. Cheng, “Sparse channel esti mation and hybrid precoding using deep learning for millimeter wav e massive mimo,” IEEE Transactions on Communications, vol. 68, no. 5, pp. 2838– 2849, 2020

    2020

  14. [19]

    Model-driven deep learning based channel estimation and feedback for millime ter-wave massive hybrid mimo systems,

    X. Ma, Z. Gao, F. Gao, and M. Di Renzo, “Model-driven deep learning based channel estimation and feedback for millime ter-wave massive hybrid mimo systems,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 8, pp. 2388–2406, 2021

    2021

  15. [20]

    Deep learning for mas sive mimo csi feedback,

    C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for mas sive mimo csi feedback,” IEEE Wireless Communications Letters , vol. 7, no. 5, pp. 748–751, 2018

    2018

  16. [21]

    Model-driven deep learning for mimo detection,

    H. He, C.-K. Wen, S. Jin, and G. Y . Li, “Model-driven deep learning for mimo detection,” IEEE Transactions on Signal Processing , vol. 68, pp. 1702–1715, 2020

    2020

  17. [22]

    Adapt ive neural signal detection for massive mimo,

    M. Khani, M. Alizadeh, J. Hoydis, and P . Fleming, “Adapt ive neural signal detection for massive mimo,” IEEE Transactions on Wireless Communications, vol. 19, no. 8, pp. 5635–5648, 2020

    2020

  18. [23]

    Deep-l earning- based millimeter-wave massive mimo for hybrid precoding,

    H. Huang, Y . Song, J. Y ang, G. Gui, and F. Adachi, “Deep-l earning- based millimeter-wave massive mimo for hybrid precoding,” IEEE Transactions on V ehicular Technology, vol. 68, no. 3, pp. 3027–3032, 2019

    2019

  19. [24]

    Beamforming Design for Large-Scale A ntenna Arrays Using Deep Learning,

    T. Lin and Y . Zhu, “Beamforming Design for Large-Scale A ntenna Arrays Using Deep Learning,” IEEE Wirel. Commun. Lett. , vol. 9, no. 1, pp. 103–107, Jan. 2020

    2020

  20. [25]

    T wo- timescale end-to-end learning for channel acquisition and hybrid pre- coding,

    Q. Hu, Y . Cai, K. Kang, G. Y u, J. Hoydis, and Y . C. Eldar, “T wo- timescale end-to-end learning for channel acquisition and hybrid pre- coding,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 1, pp. 163–181, 2022

    2022

  21. [26]

    Hybrid precod ing for multiuser millimeter wave massive mimo systems: A deep lear ning approach,

    A. M. Elbir and A. K. Papazafeiropoulos, “Hybrid precod ing for multiuser millimeter wave massive mimo systems: A deep lear ning approach,” IEEE Transactions on V ehicular Technology, vol. 69, no. 1, pp. 552–563, 2020

    2020

  22. [27]

    Hybrid beamforming for mmwave massive mimo systems using conditio nal generative adversarial networks,

    B. Banerjee, R. C. Elliott, W. A. Krzymie ´n, and M. Medra, “Hybrid beamforming for mmwave massive mimo systems using conditio nal generative adversarial networks,” IEEE Transactions on V ehicular Tech- nology, vol. 73, no. 10, pp. 15 803–15 808, 2024

    2024

  23. [28]

    Atten tion- based hybrid precoding for mmwave mimo systems,

    H. Jiang, Y . Lu, X. Li, B. Wang, Y . Zhou, and L. Dai, “Atten tion- based hybrid precoding for mmwave mimo systems,” in 2021 IEEE Information Theory W orkshop (ITW) , 2021, pp. 1–6

    2021

  24. [29]

    Model-driven de ep learning for hybrid precoding in millimeter wave mu-mimo system,

    W. Jin, J. Zhang, C.-K. Wen, and S. Jin, “Model-driven de ep learning for hybrid precoding in millimeter wave mu-mimo system,” IEEE Transactions on Communications, vol. 71, no. 10, pp. 5862–5876, 2023

    2023

  25. [30]

    Learn to rapidly and robustl y optimize hybrid precoding,

    O. Lavi and N. Shlezinger, “Learn to rapidly and robustl y optimize hybrid precoding,” IEEE Transactions on Communications , vol. 71, no. 10, pp. 5814–5830, 2023

    2023

  26. [31]

    Deep unfolding hybrid beamfo rming designs for thz massive mimo systems,

    N. T. Nguyen, M. Ma, O. Lavi, N. Shlezinger, Y . C. Eldar, A . L. Swindlehurst, and M. Juntti, “Deep unfolding hybrid beamfo rming designs for thz massive mimo systems,” IEEE Transactions on Signal Processing, vol. 71, pp. 3788–3804, 2023

    2023

  27. [32]

    Transfer learning and meta learning-based fast downlink beamformin g adapta- tion,

    Y . Y uan, G. Zheng, K.-K. Wong, B. Ottersten, and Z.-Q. Lu o, “Transfer learning and meta learning-based fast downlink beamformin g adapta- tion,” IEEE Transactions on Wireless Communications , vol. 20, no. 3, pp. 1742–1755, 2021

    2021

  28. [33]

    Embedding model-based fast meta learning for downlink beamforming ad aptation,

    J. Zhang, Y . Y uan, G. Zheng, I. Krikidis, and K.-K. Wong, “Embedding model-based fast meta learning for downlink beamforming ad aptation,” IEEE Transactions on Wireless Communications, vol. 21, no. 1, pp. 149– 162, 2022

    2022

  29. [34]

    Meta lea rning- based mimo detectors: Design, simulation, and experimenta l test,

    J. Zhang, Y . He, Y .-W. Li, C.-K. Wen, and S. Jin, “Meta lea rning- based mimo detectors: Design, simulation, and experimenta l test,” IEEE Transactions on Wireless Communications , vol. 20, no. 2, pp. 1122– 1137, 2021

    2021

  30. [36]

    Metalearning-based alternating minimization algorithm for nonconvex optimization,

    J.-Y . Xia, S. Li, J.-J. Huang, Z. Y ang, I. M. Jaimoukha, a nd D. Gündüz, “Metalearning-based alternating minimization algorithm for nonconvex optimization,” IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 9, pp. 5366–5380, 2023

    2023

  31. [38]

    A deep learning-based fr amework for low complexity multiuser mimo precoding design,

    M. Zhang, J. Gao, and C. Zhong, “A deep learning-based fr amework for low complexity multiuser mimo precoding design,” IEEE Transactions on Wireless Communications , vol. 21, no. 12, pp. 11 193–11 206, 2022

    2022

  32. [39]

    A d ata and model-driven deep learning approach to robust downlink bea mforming optimization,

    K. Liang, G. Zheng, Z. Li, K.-K. Wong, and C.-B. Chae, “A d ata and model-driven deep learning approach to robust downlink bea mforming optimization,” IEEE Journal on Selected Areas in Communications , vol. 42, no. 11, pp. 3278–3292, 2024

    2024

  33. [40]

    Kan: Kolmogorov-arnold networks,

    Z. Liu, Y . Wang, S. V aidya, F. Ruehle, J. Halverson, M. So ljaˇci´c, T. Y . Hou, and M. Tegmark, “Kan: Kolmogorov-arnold networks,” 20 24. [Online]. Available: https://arxiv.org/abs/2404.19756

    arXiv

  34. [41]

    Kan or mlp: A fairer comparison ,

    R. Y u, W. Y u, and X. Wang, “Kan or mlp: A fairer comparison ,” 2024. [Online]. Available: https://arxiv.org/abs/2407.16674

    2024 arXiv

  35. [43]

    Hybrid beamforming for mil limeter wave multi-user mimo systems using learning machine,

    S. Huang, Y . Y e, and M. Xiao, “Hybrid beamforming for mil limeter wave multi-user mimo systems using learning machine,” IEEE Wireless Communications Letters , vol. 9, no. 11, pp. 1914–1918, 2020

    1914

  36. [44]

    Robust sum rate optimization for the downlink multiuser mimo systems: Wors t-case design,

    T. Endeshaw, B. K. Chalise, and L. V andendorpe, “Robust sum rate optimization for the downlink multiuser mimo systems: Wors t-case design,” in 2010 IEEE International Conference on Communications , 2010, pp. 1–5

    2010

  37. [45]

    A unified framework for communications t hrough mimo channels,

    D. P . Palomar, “A unified framework for communications t hrough mimo channels,” Ph.D. dissertation, Technical University of Ca talonia (UPC), Barcelona, Spain, 2003

    2003

  38. [46]

    fig1.png

    K. B. Petersen and M. S. Pedersen, The Matrix Cookbook . Lyngby, Denmark: Technical University of Denmark, 2008. This figure "fig1.png" is available in "png" format from: http://arxiv.org/ps/2411.15762v1

    2008 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.